Handbook of Applied Hydrology, Second Edition by Vijay P. Singh

Handbook of Applied Hydrology
00_Singh_FM_pi-lvi.indd 1 8/26/16 4:35 PM

Editor-in-Chief
Vijay P. Singh, Ph.D., D.Sc., D. Eng. (Hon.), Ph.D. (Hon.), D. Sc. (Hon.), P.E., P.H., Hon.
D. WRE, Academician (GFA), is a Distinguished Professor and Caroline & William N. Lehrer
Distinguished Chair in Water Engineering in the Department of Biological and Agricultural
Engineering and Zachry Department of Civil Engineering at Texas A&M University. He holds
a B.Tech degree from U.P. University of Agriculture and Technology, a Master’s Degree from
the University of Guelph, a Ph.D. from Colorado State University, and a D.Sc. from the
University of the Witwatersrand. One of today’s leading experts in the field of hydrology, Dr.
Singh specializes in surface water hydrology, groundwater hydrology, hydraulics, irrigation
engineering, environmental quality, and water resources. He has published 25 books and has
edited over 58 books, and has published hundreds of journal articles. He has been the Editor-
in-Chief of the Journal of Hydrologic Engineering, ASCE; is currently serving as Editor-in-
Chief of Open Agriculture, and Journal of Agricultural research, and Journal of Groundwater
Research and is on the editorial boards of numerous journals. He is also serving as Editor-in-
Chief of Water Science and Technology Book Series as well as World Water Resources Book
series. He has received more than 75 national and international awards.

Handbook of Applied Hydrology
EDITED BY VIJAY P. SINGH
Second Edition to replace the classic 1963 edition edited by Ven Te Chow
New York Chicago San Francisco Athens London

Madrid Mexico City Milan New Delhi
Singapore Sydney Toronto

Cataloging-in-Publication Data is on file with the Library of Congress.
McGraw-Hill Education books are available at special quantity discounts to use as premiums and sales pro-
motions, or for use in corporate training programs. To contact a representative please visit the Contact Us
page at www.mhprofessional.com.
Handbook of Applied Hydrology, Second Edition
Copyright ©2017 by McGraw-Hill Education. All rights reserved. Printed in the United States of America.
Except as permitted under the United States Copyright Act of 1976, no part of this publication may be
reproduced or distributed in any form or by any means, or stored in a data base or retrieval system, without
the prior written permission of the publisher.
1 2 3 4 5 6 7 8 9 DOW 21 20 19 18 17 16
ISBN 978-0-07-183509-1
MHID 0-07-183509-1
The pages within this book were printed on acid-free paper.
Sponsoring Editor Copy Editor

Lauren Poplawski Cenveo® Publisher Services
Editorial Supervisor Proofreader
Donna M. Martone Cenveo® Publisher Services
Acquisitions Coordinator Indexer
Lauren Rogers Ariel O. Tuplano
Project Manager Production Supervisor
Raghavi Khullar, Cenveo® Pamela A. Pelton
Publisher Services
Composition Art Director, Cover
Jeff Weeks
Information contained in this work has been obtained by McGraw-Hill Education from sources believed to be
reliable. However, neither McGraw-Hill Education nor its authors guarantee the accuracy or completeness of any
information published herein, and neither McGraw-Hill Education nor its authors shall be responsible for any
errors, omissions, or damages arising out of use of this information. This work is published with the understanding
that McGraw-Hill Education and its authors are supplying information but are not attempting to render engineer-
ing or other professional services. If such services are required, the assistance of an appropriate professional should
be sought.

Dedicated to
Hydrologists and Water Scientists

Contents in Brief
Part 1 Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1
Part 2 Data Collection and Processing . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
Part 3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-1
Part 4 Hydrologic Processes and Modeling . . . . . . . . . . . . . . . . . . . . . 37-1
Part 5 Sediment and Pollutant Transport . . . . . . . . . . . . . . . . . . . . . . . 63-1
Part 6 Hydrometeorologic and Hydrologic Extremes . . . . . . . . . . . . 72-1
Part 7 Systems Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-1
Part 8 Hydrology of Large Rivers and Lake Basins . . . . . . . . . . . . . . . 93-1
Part 9 Applications and Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-1
Part 10 Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151-1
vii

For online supplements and color versions of images, please go to www.mhprofessional.com/
handbookofappliedhydrology

Contents
Preface xxxv
Foreword xxxix
Acknowledgments xli
Contributors xliii
International Advisory Board liii
Practitioner Advisory Board lv
Part 1. Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1
Chapter 1. The Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3

1.1 Characteristics of Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3
1.2 Definition of Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-4
1.3 Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-4
1.4 Components of the Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
1.5 Schematic Representation of the Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
1.6 Scales in Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-6
1.7 Impact of Climate Change on the Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . 1-6
1.8 Influence of Human Activities and Land Use Changes on Hydrologic Cycle . . . 1-6
1.9 Relation Between Hydrologic Cycle and Carbon and Nitrogen Cycles . . . . . . . . 1-7
1.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-9
Chapter 2. Watersheds, River Basins, and Land Use . . . . . . . . . . . . . . . . 2-1

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1
2.2 Components of Watersheds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1
2.3 Delineation of a Watershed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4
2.4 Watershed Hydrological Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4
2.5 Characteristics of a Watershed That Impact on Hydrological Processes . . . . . . 2-4
2.6 River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
2.7 River Basin Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
2.8 Major River Basins in the World . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2.9 Land Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8
Chapter 3. Water Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.2 Hydrologic Fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.3 Water on the Earth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.4 Water Balance Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
3.5 Natural and Anthropogenic Effects on the Water Balance . . . . . . . . . . . . . . . . . . . 3-7
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
PART 2. Data Collection and Processing . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
Chapter 4. Hydrometeors and Quantitative Precipitation Estimation 4-3

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
4.2 Types of Hydrometeorological Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
4.3 Remote Sensing of Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
4.4 Hydrometeorological Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-4
4.5 Hydrometeorological Data Quality Assurance and Control . . . . . . . . . . . . . . . . . . 4-5
4.6 Quantitative Precipitation Estimate, Data Use, Archiving, and Accessibility . . 4-6
ix

x Contents
Chapter 5. Streamflow Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1

5.1 Streamflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1
5.2 Types of Streamflow Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1
5.3 Streamgage Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1
5.4 Quality Assurance of Streamgage Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-6
5.5 Derived Streamflow Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-7
Chapter 6. Streamflow Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1
6.2 Rating Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1
6.3 Simple Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-4
6.4 Complex Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
6.5 Slope Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-8
6.6 Rate of Change of Stage Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-8
6.7 Dynamic-Flow Model Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
6.8 Index-Velocity Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
6.9 Shifting-Control Method for Dealing with Rating Complexities . . . . . . . . . . . . 6-11
6.10 Uncertainty in Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-13
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-13
Chapter 7. Hydrologic Information Systems . . . . . . . . . . . . . . . . . . . . . . 7-1

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-1
7.2 Hydrologic Data Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-1
7.3 Service-Oriented Architectures for Integrating Distributed Hydrologic
Data and Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-2
7.4 The CUAHSI Hydrologic Information System as an Example HIS . . . . . . . . . . . 7-3
7.5 HydroShare as a Next-Generation HIS Software . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5
7.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
Chapter 8. Remote Sensing Techniques and Data Assimilation for

Hydrologic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-1
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-1
8.2 Remote Sensing Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-1
8.3 Remote Sensing in Hydrological Sciences: A Historical Perspective . . . . . . . . 8-1
8.4 Remote Sensing: Methods and Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.5 Data Assimilation: Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-3
8.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
Chapter 9. Geographic Information Systems . . . . . . . . . . . . . . . . . . . . . . 9-1

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-1
9.2 Basic Principles of GIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-1
9.3 Data Sources and Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
9.4 Representation of Model Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
9.5 Model/GIS Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-7
9.6 Current Status and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-8
Chapter 10. Design of Hydrologic Networks . . . . . . . . . . . . . . . . . . . . . . 10-1

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-1
10.2 Hydrologic Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-1
10.3 Necessity of Hydrologic Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-1
10.4 Impact of Hydrologic Network Density on Streamflow Estimates . . . . . . . . . . 10-2
10.5 Design Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-2
10.6 Design of Hydrologic Networks: Methodologies . . . . . . . . . . . . . . . . . . . . . . . . . . 10-2
10.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4
Part 3. Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-1
Chapter 11. Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-3

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-3
11.2 Historical Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-3
11.3 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-4
11.4 ANN Training and Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-4

Contents xi
11.5 Drawbacks of ANN Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-5

11.6 Shortcomings in ANN Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-5
11.7 Future Direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-6
Chapter 12. Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-1

12.1 Fuzzy Logic Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-1
12.2 Function of Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-2
12.3 Fuzzy Rule-Based Modeling (Fuzzy Inference) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-3
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-5
Chapter 13. Evolutionary Computing: Genetic Algorithms . . . . . . . . . 13-1

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-1
13.2 Evolutionary Computing in Hydrology: An Overview . . . . . . . . . . . . . . . . . . . . . . 13-1
13.3 Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-2
13.4 GA Applications in Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-3
13.5 Conclusion and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-3
Chapter 14. Relevance Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-1

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-1
14.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-1
14.3 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-2
14.4 Application of RVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-3
14.5 Examples from Hydrology and Future Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-4
14.6 Miscellaneous Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-6
Chapter 15. Harmonic Analysis and Wavelets . . . . . . . . . . . . . . . . . . . . . 15-1

15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-1
15.2 The Continuous Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-1
15.3 Discrete Time Wavelet Transform and Multiresolution Analysis . . . . . . . . . . . . 15-3
15.4 Signal Energy Repartition in the Wavelet Frame . . . . . . . . . . . . . . . . . . . . . . . . . . 15-3
15.5 Wavelet Analysis of the Time-Scale Relationship Between Two Signals . . . . . 15-4
15.6 Wavelet Cross Spectrum and Coherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-4
15.7 Applications of Wavelet Transforms in Hydrology and Earth Sciences . . . . . . 15-5
15.8 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-5
Chapter 16. Outlier Analysis and Infilling of Missing Records

in Hydrologic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-1
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-1
16.2 Concepts and Methods for Outlier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-2
16.3 Concepts and Methods for Handling Missing Records . . . . . . . . . . . . . . . . . . . . . 16-3
16.4 Discussion and Concluding Remarks on Methods for Outliers and Infilling
of Missing Records . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-5
16.5 Future Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16-6
Chapter 17. Linear and Nonlinear Regression . . . . . . . . . . . . . . . . . . . . . 17-1

17.1 Linear and Nonlinear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-1
17.2 Measures for Goodness of Fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-4
17.3 Multiple Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-5
17.4 Nonlinear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17-9
Chapter 18. Time Series Analysis and Models . . . . . . . . . . . . . . . . . . . . . 18-1

18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-1
18.2 Properties of Hydrological Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-1
18.3 Time-Series Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-2
18.4 Modeling of Continuous Time Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-2
18.5 Univariate Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-2
18.6 Univariate Periodic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-4
18.7 Multivariate Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-6
18.8 Disaggregation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-7
18.9 Nonparametric Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-7

xii Contents
18.10 Stochastic Simulation, Forecasting, and Uncertainty . . . . . . . . . . . . . . . . . . . . . . 18-8

18.11 Conceptual Stochastic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-8
18.12 Final Remarks and Future Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18-9
Chapter 19. Statistical Detection of Nonstationarity: Issues

and Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-1
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-1
19.2 Exploratory Methods for Detection of Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-2
19.3 Statistical Exploration of Nonstationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-2
19.4 Effect of Nonconstant Error Variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-3
19.5 Effect of a Priori Filtering of Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-4
19.6 Distribution of a Breakpoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-4
19.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19-5
Chapter 20. Spatial Analysis and Geostatistical Methods . . . . . . . . . . 20-1

20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-1
20.2 Data Types and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-1
20.3 Spatial Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-1
20.4 Heterogeneous Field Estimation and Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 20-3
20.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-7
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20-7
Chapter 21. Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-1

21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-1
21.2 Discrete Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-1
21.3 Classification of Continuous Frequency Distributions . . . . . . . . . . . . . . . . . . . . . 21-2
21.4 Continuous Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-2
21.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-10
Chapter 22. Calibration, Parameter Estimation, Uncertainty, Data

Assimilation, Sensitivity Analysis, and Validation . . . . . . . . 22-1
22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-1
22.2 Parameter Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-1
22.3 Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-3
22.4 Data Assimilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-5
22.5 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-10
22.6 Validation Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-11
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22-15
Chapter 23. Bayesian Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-1

23.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-1
23.2 The Bayesian Inference Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-1
23.3 Computational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-4
23.4 Diagnostics to Scrutinize Model Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-6
23.5 Applications in Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-7
23.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-9
Chapter 24. Optimization Approaches for Integrated Water Resources

Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-1
24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-1
24.2 Trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-2
24.3 Challenges and Research Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-4
24.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-5
24.5 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-5
24.6 Appendix: Literature Trend Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24-5
Chapter 25. Nonparametric Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-1

25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-1
25.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-1
25.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-2
25.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-3
25.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-5

Contents xiii
Chapter 26. Predictive Uncertainty Assessment

and Decision Making . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-1
26.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-1
26.2 Forecasting in Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-1
26.3 Motivations for Converting Deterministic to Stochastic Prediction . . . . . . . . . 26-1
26.4 Predictive Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-2
26.5 Techniques Aimed at Assessing Predictive Uncertainty . . . . . . . . . . . . . . . . . . . . 26-4
26.6 Verification of the Estimated Predictive Density . . . . . . . . . . . . . . . . . . . . . . . . . . 26-9
26.7 Major Reasons Undermining the Operational Use of Predictive Uncertainty 26-9
26.8 Examples of Proper Use of Predictive Uncertainty to Improve Decisions . . . . 26-13
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26-15
Chapter 27. Risk-Reliability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-1

27.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-1
27.2 Measures of Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-1
27.3 Performance Function and Reliability Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-2
27.4 Direct Integration Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-2
27.5 First-Order Second-Moment Reliability Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 27-3
27.6 Time-Dependent (Dynamic) Reliability Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-5
27.7 Time-to-Failure Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-5
27.8 Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-9
Chapter 28. Scaling and Fractals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-1

28.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-1
28.2 Scale-Invariant Sets and Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-1
28.3 Some Properties of H-SSSI Processes and MF Cascades . . . . . . . . . . . . . . . . . . . . 28-2
28.4 Inference of Scaling for Stationary Multifractal Measures . . . . . . . . . . . . . . . . . 28-4
28.5 Processes with Limited Scale Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-4
28.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-5
Chapter 29. Nonlinear Dynamics and Chaos . . . . . . . . . . . . . . . . . . . . . . 29-1

29.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-1
29.2 Chaos Theory: A Brief History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-2
29.3 Chaos Concepts and Identification Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-2
29.4 Issues in Chaos Identification and Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-5
29.5 Hydrologic Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-7
29.6 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29-8
Chapter 30. Copula Modeling in Hydrologic Frequency Analysis . . . 30-1

30.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-1
30.2 Description of Copula Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-1
30.3 Overview of Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-2
30.4 Multivariate Quantile and Return Period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-5
30.5 An Illustration: The Fraser River at Hope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-5
30.6 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-7
30.7 Resources and Further Specific References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30-9
Chapter 31. Entropy Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-1

31.1 Origin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-1
31.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-1
31.3 Forms of Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-2
31.4 Directional Information Transfer Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-3
31.5 Entropy under Transformation of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-3
31.6 Informational Correlation Coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-3
31.7 Total Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-3
31.8 Theory of Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-4
31.9 Methodology for Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-4
31.10 Hydrologic Modeling Using Entropy Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-7
31.11 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-8

xiv Contents
Chapter 32. Entropy Production Extremum Principles . . . . . . . . . . . . . 32-1

32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32-1
32.2 Background and Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32-1
32.3 Maximum Entropy Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32-4
32.4 Review of Applications in Hydrology and Hydraulics . . . . . . . . . . . . . . . . . . . . . . 32-5
32.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32-5
Chapter 33. Data-Based Mechanistic Modeling . . . . . . . . . . . . . . . . . . . 33-1

33.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33-1
33.2 The Main Stages of DBM Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33-2
33.3 Linear DBM Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33-3
33.4 Time Variable and State-Dependent Parameter Models . . . . . . . . . . . . . . . . . . . 33-6
33.5 Hypothetico-Inductive DBM Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33-7
33.6 DBM Emulation Modeling of High-Order Simulation Models . . . . . . . . . . . . . . . 33-8
33.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33-11
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33-11
Chapter 34. Decomposition Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34-1

34.1 Introduction: Adomian’s Decompositions Method . . . . . . . . . . . . . . . . . . . . . . . . 34-1
34.2 Regional Flow in an Unconfined Aquifer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34-2
34.3 Propagation of Nonlinear Kinematic Flood Waves in Rivers . . . . . . . . . . . . . . . . 34-2
34.4 Nonlinear Infiltration in Unsaturated Soils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34-4
34.5 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34-5
Chapter 35. Network Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-1

35.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-1
35.2 Network Theory: Concept and History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-2
35.3 Network Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-3
35.4 Network Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-4
35.5 Applications in Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-6
35.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35-8
Chapter 36. Hydroeconomic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36-1

36.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36-1
36.2 Estimating the Economic Value of Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36-1
36.3 Water Demand Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36-2
36.4 Considerations in the Design of Hydroeconomic Analysis Studies . . . . . . . . . . 36-3
36.5 Applications and Implementation of Hydroeconomic Analysis for
Management and Decision Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36-5
36.6 Discussion of Challenges, Limitations, and Future Directions . . . . . . . . . . . . . . 36-6
36.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36-7
Part 4. Hydrologic Processes and Modeling . . . . . . . . . . . . . . . . . . . . . . . 37-1
Chapter 37. Weather and Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37-3

37.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37-3
37.2 Hydrologic Engineering and Intersection with Weather and Climate . . . . . . . 37-3
37.3 Weather . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37-5
37.4 Observing Weather . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37-7
37.5 Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37-7
37.6 Climate Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37-10
Chapter 38. Hydroclimatology: Global Warming and Climate Change 38-1

38.1 Introduction: The Ambiguity of Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38-1
38.2 Natural and Human Influences on Present-Day Climate . . . . . . . . . . . . . . . . . . . 38-1
38.3 Impacts of Climate Change on the Hydrological Cycle in the Twentieth and
Twenty-First Centuries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38-2
38.4 Global Climate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38-3
38.5 Working with Climate Model Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38-4
38.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38-6

Contents xv
Chapter 39. Spatial and Temporal Estimation and

Analysis of Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39-1
39.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39-1
39.2 Estimates of Mean Areal Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39-1
39.3 Missing Precipitation Data Estimation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 39-4
39.4 Limitations of Estimation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39-5
39.5 New Methods for Missing Data Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39-5
39.6 Summary of Issues for Missing Precipitation Data Estimation . . . . . . . . . . . . . . 39-6
39.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39-8
Chapter 40. Snow Distribution and Snowpack Characteristics . . . . . 40-1

40.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40-1
40.2 Processes Controlling Snow Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40-1
40.3 Spatial Patterns of Snow at Various Scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40-2
40.4 Snowpack Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40-2
40.5 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40-3
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40-4
Chapter 41. Time-Space Modeling of Precipitation . . . . . . . . . . . . . . . . 41-1

41.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41-1
41.2 Stochastic Modeling of Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41-1
41.3 Deterministic Numerical Modeling of Time-Space Precipitation . . . . . . . . . . . . 41-3
41.4 Remote Sensing for the Modeling of Time-Space Precipitation . . . . . . . . . . . . 41-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41-10
Chapter 42. Evapotranspiration and Evaporative Demand . . . . . . . . . 42-1

42.1 Introduction and History of Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42-1
42.2 Relevant Concepts and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42-3
42.3 Outstanding Problems and Directions for Future Work . . . . . . . . . . . . . . . . . . . . 42-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42-14
Chapter 43. Rainfall Interception, Detention, and

Depression Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43-1
43.1 Canopy Interception Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43-1
43.2 Forest Floor Interception Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43-3
43.3 Detention and Depression Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43-3
43.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43-3
Chapter 44. Watershed Geomorphological Characteristics . . . . . . . . . 44-1

44.1 Introduction and Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44-1
44.2 Watersheds and Drainage Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44-2
44.3 Outstanding Problems and Directions for Future Work . . . . . . . . . . . . . . . . . . . . 44-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44-11
Chapter 45. Infiltration Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45-1

45.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45-1
45.2 Basic Equations for Vertical Infiltration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45-1
45.3 Classical Models for Point Infiltration into Vertically Homogeneous Soils . . . 45-2
45.4 Modeling of Point Infiltration into Vertically Nonuniform Soils . . . . . . . . . . . . 45-4
45.5 Models for Rainfall Infiltration over Heterogeneous Areas . . . . . . . . . . . . . . . . . 45-5
45.6 Soil Conservation Service Runoff Curve Number Model . . . . . . . . . . . . . . . . . . . 45-6
45.7 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45-7
Chapter 46. Soil Moisture and Vadose Zone Modeling . . . . . . . . . . . . . 46-1

46.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46-1
46.2 Continuum-Scale Models for Partially Saturated Flow in the Vadose Zone . . 46-2
46.3 Numerical Vadose Zone and Land Surface Models . . . . . . . . . . . . . . . . . . . . . . . . 46-3
46.4 Soil Moisture across Spatial-Temporal Scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46-5
46.5 Inverse Modeling—Soil Hydraulic Properties at the Model Grid Scale . . . . . . 46-8
46.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46-11

xvi Contents
Chapter 47. Hydrogeologic Characterization . . . . . . . . . . . . . . . . . . . . . 47-1

47.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47-1
47.2 Borehole Samples and Groundwater Monitoring Wells . . . . . . . . . . . . . . . . . . . 47-1
47.3 Investigation of Borehole Drilling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47-1
47.4 Cone Penetrometry, Permeametry, and Electrical Conductivity Logging . . . . 47-2
47.5 Electrical Resistivity Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47-2
47.6 Hydraulic Methods for in Situ Conductivity Measurement . . . . . . . . . . . . . . . . . 47-2
47.7 Characterization of the Hydraulic Gradient and Flow Rates . . . . . . . . . . . . . . . . 47-3
47.8 Recharge Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47-3
47.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47-3
Chapter 48. Groundwater Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-1

48.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-1
48.2 Groundwater Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-1
48.3 Grid Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-2
48.4 Model Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-3
48.5 Case Study: Groundwater Modeling in Baton Rouge,
Southeastern Louisiana . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-3
48.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48-7
Chapter 49. Watershed Runoff, Streamflow Generation,

and Hydrologic Flow Regimes . . . . . . . . . . . . . . . . . . . . . . . . 49-1
49.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49-1
49.2 Dominant Runoff Generation Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49-3
49.3 Infiltration Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49-6
49.4 Factors Affecting Runoff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49-6
49.5 Poorly Understood Factors Affecting Runoff Generation . . . . . . . . . . . . . . . . . . 49-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49-11
Chapter 50. Snowmelt Runoff Generation and Modeling . . . . . . . . . . 50-1

50.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50-1
50.2 Snow Accumulation Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50-1
50.3 Energy Budget of Snow Pack and Snowmelt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50-2
50.4 Simulation of Snow Accumulation Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50-3
50.5 Heat and Water Transfer in Melting Snow Cover . . . . . . . . . . . . . . . . . . . . . . . . . . 50-3
50.6 Spatial Variability of Snow Cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50-4
50.7 Constructing General Model of Snowmelt Runoff Generation . . . . . . . . . . . . . . 50-4
50.8 Case studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50-9
Chapter 51. Glacial Melting and Runoff Modeling . . . . . . . . . . . . . . . . . 51-1

51.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51-1
51.2 Remote Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51-1
51.3 Application of Remote Sensing in Glacier Quantification . . . . . . . . . . . . . . . . . . 51-1
51.4 Glaciated Versus Nonglaciated Watersheds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51-3
51.5 Application in Streamflow Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51-3
51.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51-4
Chapter 52. Reservoir and Channel Routing . . . . . . . . . . . . . . . . . . . . . . 52-1

52.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-1
52.2 Reservoir Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-1
52.3 River Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-2
52.4 The Classical Muskingum Flood Routing Method . . . . . . . . . . . . . . . . . . . . . . . . . 52-3
52.5 Nash Cascade Model for River Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-4
52.6 Other Linear Storage Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-4
52.7 Linear Diffusion Analogy Routing Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-5
52.8 Nonlinear Routing Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-5
52.9 Flow Routing Using Hydraulic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-6
52.10 Basis for the Development of Simplified Momentum Equations . . . . . . . . . . . 52-7
52.11 Simplified Hydraulic Flood Routing Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-7
52.12 Kalinin–Milyukov Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-9
52.13 Variable Parameter Muskingum Stage Routing Method . . . . . . . . . . . . . . . . . . . 52-12

Contents xvii
52.14 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-14
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-14
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52-15
Chapter 53. Waterlogging and Salinzation . . . . . . . . . . . . . . . . . . . . . . . . 53-1

53.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-1
53.2 Salinity Features and Impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-1
53.3 Irrigation Induced Rises of the Watertable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-2
53.4 Irrigation Induced Land Salinization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-3
53.5 Land Salinization Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-4
53.6 Salt Balance and Leaching Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-4
53.7 Monitoring and Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-5
53.8 Remedial Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-5
53.9 New Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-6
53.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53-7
Chapter 54. Surface Water–Groundwater Interactions:

Integrated Modeling of a Coupled System . . . . . . . . . . . . 54-1
54.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-1
54.2 Surface Water Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-2
54.3 Subsurface Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-4
54.4 Soil Plant Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-5
54.5 Coupling Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-6
54.6 Scale Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-7
54.7 Data Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-9
54.8 Integrated Models and Watershed Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 54-10
54.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54-12
Chapter 55. Seawater Intrusion in Coastal Aquifers: Concepts,

Mitigation, and Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 55-1
55.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55-1
55.2 Assumptions and Approaches for Modeling Seawater Intrusion . . . . . . . . . . . 55-1
55.3 Mitigation of Seawater Intrusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55-3
55.4 Case Study: The Nile Delta Aquifer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55-3
55.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55-9
Chapter 56. Regional Land Subsidence Caused by the Compaction

of Susceptible Aquifer Systems Accompanying
Groundwater Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56-1
56.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56-1
56.2 Detection and Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56-2
56.3 Analysis and Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56-9
Chapter 57. Hydraulic Fracturing and Hydrologic Impacts . . . . . . . . . 57-1

57.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57-1
57.2 Hydraulic Fracturing Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57-2
57.3 Risk of Groundwater Contamination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57-4
57.4 Potential for Induced Seismicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57-6
57.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57-8
Chapter 58. Catchment Classification and Regionalization . . . . . . . . . 58-1

58.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58-1
58.2 Catchment Classification: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58-2
58.3 Classification Based on Chaos Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58-2
58.4 Classification Based on Network Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58-4
58.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58-7

xviii Contents
Chapter 59. Rainfall-Runoff Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-1

59.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-1
59.2 A Brief History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-1
59.3 Computation of Runoff Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-1
59.4 Determination of Peak Discharge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-3
59.5 Runoff Hydrograph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-3
59.6 Computation of Runoff Hydrograph by Hydraulic Approaches . . . . . . . . . . . . . 59-6
59.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59-8
Chapter 60. Continuous Watershed Modeling . . . . . . . . . . . . . . . . . . . . . 60-1

60.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60-1
60.3 Concepts and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60-3
60.4 Outstanding Problems/Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60-7
60.5 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60-10
Chapter 61. Calibration and Evaluation of Watershed Models . . . . . . 61-1

61.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-1
61.2 Calibration and Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-2
61.3 Elements of a Calibration/Evaluation Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-2
61.4 Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-2
61.5 Model Calibration and Evaluation Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-3
61.6 Strategies for Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-3
61.7 Desirable Properties of a Successful Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . 61-4
61.8 Preparation Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-4
61.9 Calibration Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-8
61.10 Evaluation Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-8
61.11 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61-9
Chapter 62. Feasibility, Engineering, and Operations Models: Using

the Decision Environment to Inform the Model Design 62-1
62.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-1
62.2 The Decision Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-1
62.3 USACE Decision Environment Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-2
62.4 Precision Dimensions of Decision Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-2
62.5 Process Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-3
62.6 Information Content Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-5
62.7 Case Study: Picayune Strand-Restoration Project . . . . . . . . . . . . . . . . . . . . . . . . . 62-5
62.8 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-6
62.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62-7
Part 5. Sediment and Pollutant Transport . . . . . . . . . . . . . . . . . . . . . . . . . 63-1

Chapter 63. Water Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63-3
63.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63-3
63.2 Water Pollution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63-3
63.3 Water Quality Standards and Water Use Designations . . . . . . . . . . . . . . . . . . . . . 63-3
63.4 Restoration of Water Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63-5
63.5 Sensor-Based Water Quality Monitoring Technologies . . . . . . . . . . . . . . . . . . . . 63-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63-9
Chapter 64. Soil Erosion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-1

64.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-1
64.2 Erosion by Wind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-1
64.3 Erosion by Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-1
64.4 Gravity-Induced Erosion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-2
64.5 Tillage Erosion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-3
64.6 Snowmelt Erosion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-3
64.7 Irrigation-Induced Erosion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-3
64.8 Erosion by Wind-Driven Rain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-3
64.9 Erosion Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-4
64.10 Erosion Assessment—Field and Laboratory Measurements . . . . . . . . . . . . . . . 64-4

Contents xix
64.11 Erosion Assessment—Equations and Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-5

64.12 Erosion Impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-5
64.13 Erosion Control and Soil Conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-5
64.14 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64-6
Chapter 65. Channel Erosion and Sediment Transport . . . . . . . . . . . . . 65-1

65.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65-1
65.2 Sediment Production and Loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65-1
65.3 Partitioning of Sediment Loads in Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65-2
65.4 Bank Profile and Erosional Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65-3
65.5 Modes of Sediment Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65-5
65.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65-10
Chapter 66. Sedimentation of Floodplains, Lakes, and Reservoirs . . . . . . 66-1

66.1 Floodplain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-1
66.2 Lake . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-1
66.3 Reservoir . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-2
66.4 The Sedimentation Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-5
66.5 Prediction of Sediments in Floodplains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-7
66.6 Reservoir Trap Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-7
66.7 Estimation of Sediment in Lakes and Reservoirs . . . . . . . . . . . . . . . . . . . . . . . . . . 66-7
66.8 Protective Measure Against Sedimentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-9
66.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-9
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66-10
Chapter 67. Pollutant Transport in Surface Water . . . . . . . . . . . . . . . . . 67-1

67.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67-1
67.2 Fundamental Processes Controlling Pollutant Transport . . . . . . . . . . . . . . . . . . 67-1
67.3 Pollutant Transport in Rivers and Streams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67-3
67.4 Pollutant Transport in Lakes and Reservoirs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67-4
67.5 Pollutant Transport in Coastal Waters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67-9
Chapter 68. Pollutant Transport in Vadose Zone . . . . . . . . . . . . . . . . . . 68-1

68.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68-1
68.2 Water Potential in the Unsaturated Zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68-1
68.3 Governing Equation of Flow in Vadose Zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68-1
68.4 Deterministic Approach to Solute Transport in the Vadose Zone . . . . . . . . . . . 68-3
68.5 Codes for Numerical Solution of Vadose Zone Flow and Transport . . . . . . . . . 68-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68-6
Chapter 69. Pollutant Transport in Groundwater . . . . . . . . . . . . . . . . . . 69-1

69.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69-1
69.2 Learn from the Field Work—A Case of
Saltwater Intrusion Observation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69-3
69.3 Reactive Solute Transport Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69-3
69.4 Dispersion Processes in Groundwater
(Sato and Lwasa, 2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69-5
69.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69-9
Chapter 70. Salinization and Salinity Management in Watersheds . . . . 70-1

70.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70-1
70.2 Salinization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70-1
70.3 Salinity-Related Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70-2
70.4 Salinity Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70-2
70.5 Salinity Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70-4
Chapter 71. Transport of Biochemicals and Microorganisms . . . . . . . 71-1

71.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71-1
71.2 Biochemicals and Microorganisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71-1

xx Contents
71.3 Mathematical Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71-2

71.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71-6
Part 6. Hydrometeorologic and Hydrologic Extremes . . . . . . . . . . . . . . 72-1
Chapter 72. Atmospheric Rivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72-3

72.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72-3
72.2 Large-Scale Circulation and Moisture Sources and Pathways . . . . . . . . . . . . . . 72-4
72.3 Precipitation and Flooding Associated with ARs . . . . . . . . . . . . . . . . . . . . . . . . . . 72-4
72.4 Modeling of ARs and Associated Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72-5
72.5 Projection of Future Changes in ARs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72-6
72.6 Future Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72-8
Chapter 73. Hydrometeorological Extremes

(Hurricanes and Typhoons) . . . . . . . . . . . . . . . . . . . . . . . . . . 73-1
73.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73-1
73.2 Climatology of Heavy Rainfall and Flooding from Tropical Cyclones . . . . . . . . 73-1
73.3 Remote Rainfall Associated with Tropical Cyclones: Predecessor
Rain Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73-2
73.4 Projected Increases in Rainfall Associated with Tropical Cyclones . . . . . . . . . . 73-2
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73-3
Chapter 74. Extreme Rainfall: Global Perspective . . . . . . . . . . . . . . . . . 74-1

74.1 Introduction: The Importance of Studying Extreme Rainfall and
Related Difficulties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74-1
74.2 A Global Survey of Record Rainfall Depths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74-3
74.3 Approaches in Estimating Extreme Rainfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74-7
74.4 The Concept of Probable Maximum Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . 74-10
74.5 Probabilistic Approach to Extreme Rainfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74-12
74.6 Ombrian (Intensity-Duration-Frequency) Curves . . . . . . . . . . . . . . . . . . . . . . . . . . 74-13
74.7 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74-15
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74-16
Chapter 75. Floods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75-1

75.1 Introduction to Floods and Flooding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75-1
75.2 Flood Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75-1
75.3 Flood Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75-2
75.4 Estimation of Flood Magnitudes and Design Floods . . . . . . . . . . . . . . . . . . . . . . . 75-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75-5
Chapter 76. Flood Frequency Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-1

76.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-1
76.2 Describing the Chance of Flood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-1
76.3 Looking at the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-3
76.4 Fitting Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-4
76.5 Index Flood Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-5
76.6 Method of Moments, Bulletin 17B, and Bulletin 17C,
with the LP3 Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-6
76.7 Uncertainty Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-6
76.8 Estimation at Ungaged Sites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-7
76.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76-7
Chapter 77. Regional Flood Frequency Modeling . . . . . . . . . . . . . . . . . 77-1

77.1 Introduction: The Regional Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77-1
77.2 Regional Estimation Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77-1
77.3 Delineation of Homogeneous Regions and Homogeneity Testing . . . . . . . . . . 77-1
77.4 Regional Transfer Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77-3
77.5 One-Step Regional Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77-4
77.6 Nonlinear Models in Regional Flood Frequency Modeling . . . . . . . . . . . . . . . . . 77-4
77.7 Multivariate Regional Flood Frequency Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 77-5

Contents xxi
77.8 Nonstationary Regional Flood Frequency Approaches . . . . . . . . . . . . . . . . . . . . 77-5

77.9 Regional Flood Frequency Analysis Based on Seasonality Measures . . . . . . . 77-6
77.10 Combination of Local and Regional Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 77-6
77.11 Daily Streamflow Estimation at Ungauged Sites . . . . . . . . . . . . . . . . . . . . . . . . . . 77-6
77.12 Discussion and New Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77-7
Chapter 78. Risk, Reliability, and Return Periods and

Hydrologic Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78-1
78.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78-1
78.2 Probabilistic- and Risk-Based Approaches to Hydrologic Design . . . . . . . . . . 78-2
78.3 Multivariate Probabilistic- and Risk-Based Approaches to
Hydrologic Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78-9
Chapter 79. Drought Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79-1

79.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79-1
79.2 Drought, Aridity, and Water Scarcity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79-2
79.3 Drought Occurrences in the World . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79-3
79.4 Drought Properties Based on Statistical Techniques . . . . . . . . . . . . . . . . . . . . . . 79-3
79.5 Numerical Characterization of Drought Properties . . . . . . . . . . . . . . . . . . . . . . . 79-4
79.6 Indices for Drought Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79-6
79.7 Outstanding Problems and Direction for Future Work . . . . . . . . . . . . . . . . . . . . . 79-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79-10
Chapter 80. Low Flow and Drought Analysis . . . . . . . . . . . . . . . . . . . . . . 80-1

80.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80-1
80.2 Need for Low Flow Hydrology Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80-1
80.3 Factors Affecting Low Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80-2
80.4 Low Flow Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80-2
80.5 Methods of Low Flow Estimation in Ungaged Catchments . . . . . . . . . . . . . . . . . 80-4
80.6 Drought Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80-4
80.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80-9
Part 7. Systems Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-1
Chapter 81. Isotope Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-3

81.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-3
81.2 Isotopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-3
81.3 Groundwater Dating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-5
81.4 Sampling Methods and Isotope Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-7
81.5 Isotope Applications in Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-12
Chapter 82. Lake Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-1

82.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-1
82.2 Origin of Lakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-1
82.3 Water Balance of Lakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-1
82.4 Thermal Regime of Lakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-2
82.5 Ice Growth on Lakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-4
82.6 Circulation Processes in Lakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82-5
Chapter 83. Urban Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83-1

83.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83-1
83.2 The Effects of Urbanization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83-2
83.3 Other Aspects of Urban Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83-3
83.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83-4
Chapter 84. Agricultural Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84-1

84.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84-1
84.2 Water Movement in the Root Zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84-1
84.3 Evaporation and Transpiration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84-4

xxii Contents
Chapter 85. Forest Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85-1

85.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85-1
85.3 Principles of Forest Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85-2
85.4 Research Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85-3
85.5 Key Findings in Forest-Stream Water Quantity and Quality Relationships . . . 85-4
85.6 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85-5
Chapter 86. Coastal Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86-1

86.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86-1
86.2 Overview of Coastal Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86-1
86.3 Movement of Water and Sediment in Coastal Areas . . . . . . . . . . . . . . . . . . . . . . . 86-2
86.4 Mathematical Models of Water and Sediment Movement . . . . . . . . . . . . . . . . . . 86-6
86.5 Summary and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86-8
Chapter 87. Wetland Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-1

87.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-1
87.2 Importance of Hydrology on Wetland Functioning . . . . . . . . . . . . . . . . . . . . . . . . 87-1
87.3 Hydroperiod . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-2
87.4 Wetland Hydrologic Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-2
87.5 Wetland Water Budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-3
87.6 Wetland Hydraulics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-5
87.7 Modeling Groundwater-Surface Water Interactions . . . . . . . . . . . . . . . . . . . . . . . 87-6
87.8 Wetland Hydrology at the Watershed Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-7
87.9 Anthropogenic and Climate Change Impacts on Wetlands . . . . . . . . . . . . . . . . . 87-7
87.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87-8
Chapter 88. Arid Zone Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-1

88.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-1
88.2 Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-1
88.3 Condensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-1
88.4 Infiltration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-1
88.5 Runoff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-2
88.6 Transmission Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-3
88.7 Change in Storage (ΔS)-Groundwater Recharge . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-3
88.8 Evapotranspiration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-4
88.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88-6
Chapter 89. Karst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89-1

89.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89-1
89.2 Investigation and Characterization of Karst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89-4
89.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89-10
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89-11
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89-11
Chapter 90. Cryospheric Hydrology: Mountainous Environment . . . 90-1

90.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-1
90.2 Alpine Runoff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-1
90.3 Monsoon Dominated Asian Mountain Ranges . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-1
90.4 Glacier Runoff as a Resource . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-1
90.5 Glacier Melting Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-4
90.6 Glacier Melt Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-5
90.7 Drainage and Storage Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-5
90.8 Diurnal and Seasonal Variations in Glacier Meltwater . . . . . . . . . . . . . . . . . . . . . 90-6
90.9 Cryospheric Hydrology and Climate Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-7
90.10 Glacier Lake Outburst Floods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-9
90.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90-10

Contents xxiii
Chapter 91. Hydrology of Transportation Systems . . . . . . . . . . . . . . . . 91-1

91.1 Pathways in Nature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91-1
91.2 Scales of Movement and Accumulation Processes . . . . . . . . . . . . . . . . . . . . . . . . . 91-2
91.3 Soil and Groundwater . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91-3
91.4 Streams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91-4
91.5 Lakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91-6
Chapter 92. Large-Scale and Global Hydrology . . . . . . . . . . . . . . . . . . . 92-1

92.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92-1
92.2 The Distribution of Water on Earth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92-1
92.3 The Global Water Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92-1
92.4 Numerical Modeling and Data Assimilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92-2
92.5 Global Water Cycle Variability, Predictability, and Change . . . . . . . . . . . . . . . . . 92-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92-6
Part 8. Hydrology of Large River and Lake Basins . . . . . . . . . . . . . . . . . 93-1
Chapter 93. Amazon River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93-3

93.1
Main Geographical Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93-3
93.2
Amazon Hydrological Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93-3
93.3
Environment, Economics and Potentialities of the Basin . . . . . . . . . . . . . . . . . . . 93-5
93.4
Impact of Anthropic Activities in the Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93-6
93.5
Amazon Basin in the Context of Its Water Footprint
and Environmental Role . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93-6
Chapter 94. Paraná (Rio de la Plata) River Basin . . . . . . . . . . . . . . . . . . . 94-1

94.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94-1
94.2 Geographical Features and Main Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94-1
94.3 Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94-1
94.4 Hydrological Features and Water Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94-2
94.5 Variability and Trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94-2
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94-4
Chapter 95. Orinoco River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-1

95.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-1
95.2 Regional Geological and Topographical Setting . . . . . . . . . . . . . . . . . . . . . . . . . . 95-1
95.3 Hydroclimatic Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-1
95.4 The Main Stem and Its Major Tributaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-2
95.5 Floodplains and Seasonal Sediment Regime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-2
95.6 Delta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-3
95.7 Regional Vegetation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-4
95.8 Human Impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-4
95.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95-4
Chapter 96. Nile River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96-1

96.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96-1
96.2 Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96-3
96.3 Hydrology of the Nile Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96-9
Chapter 97. Congo River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-1

97.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-1
97.2 The State of Hydrological Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-1
97.3 Climate Regimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-2
97.4 Catchments Characteristics and Hydrological Similarities . . . . . . . . . . . . . . . . . 97-2
97.5 Hydrogeochemical Processes and Sediment Transport . . . . . . . . . . . . . . . . . . . . 97-4
97.6 Hydrological Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-4
97.7 Climate and Land Use Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-5
97.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-5
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97-5

xxiv Contents
Chapter 98. Zambezi River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-1

98.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-1
98.2 Physical Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-1
98.3 Main Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-1
98.4 Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-3
98.5 Runoff Regime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-3
98.6 Past Hydrological Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-4
98.7 Hydrological Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-5
98.8 Current Concerns and Future Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98-6
Chapter 99. Euphrates and Tigris River Basin . . . . . . . . . . . . . . . . . . . . . 99-1

99.1 General Characteristics of the Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99-1
99.2 Hydrology and Water Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99-1
99.3 Water Resources Developments in the Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99-3
99.4 Environmental Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99-3
99.5 Climate Change Impacts on Basin Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99-3
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99-4
Chapter 100. Yangtze River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100-1

100.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100-1
100.2 Climate and Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100-2
100.3 Station Network and Water Conservancy Projects . . . . . . . . . . . . . . . . . . . . . . . 100-4
100.4 Significant Water Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100-5
100.5 Research on the Yangtze River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100-5
100.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100-6
Chapter 101. Yellow River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101-1

101.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101-1
101.2 Climate and Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101-3
101.3 Station Network and Water Conservancy Projects . . . . . . . . . . . . . . . . . . . . . . . 101-4
101.4 Significant Water Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101-5
101.5 Research on the Yellow River basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101-5
101.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101-6
Chapter 102. Mekong River . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102-1

102.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102-1
102.2 Upper Mekong River (Langcang Jiang) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102-3
102.3 Lower Mekong River . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102-5
102.4 Floods and Flood Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 6
102.5 Mekong Delta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 8
102.6 Biodiversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 8
102.7 Agriculture and Aquaculture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 8
102.8 Mekong River Commission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 9
102.9 Environmental Threats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 9
102.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102- 9
Chapter 103. Yenisei River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103-1

103.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103-1
103.2 Central and Lower Sections of the Yenisei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103-1
103.3 Upper Section of the Yenisei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103-1
103.4 Fluvial System of the Angara River . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103-1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103-4
Chapter 104. Lena River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104-1

104.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104-1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104-4
Chapter 105. Brahmaputra River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . 105-1

105.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105-1
105.2 Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105-1

Contents xxv
105.3 Drainage Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105-2

105.4 Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105-4
105.5 Problems Faced, Projects Undertaken, and Future Scope . . . . . . . . . . . . . . . . 105-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105-6
Chapter 106. Ganga River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-1

106.1 Ganga River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-1
106.2 Ganga River . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-1
106.3 Climate and Hydrology of the Ganga Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-2
106.4 Floods and Droughts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-2
106.5 Groundwater Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-2
106.6 Hydropower . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-2
106.7 Sediments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-3
106.8 Water Quality Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-3
106.9 Major Water Resources Development Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-3
106.10 Social and Environmental Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-3
106.11 Future Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106-4
Chapter 107. Narmada Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-1

107.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-1
107.2 Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-2
107.3 Major Tributaries and Sub-Basins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-2
107.4 Climate in Narmada Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-3
107.5 Soils and Land Use in Narmada Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-4
107.6 Water Resources of Narmada Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-4
107.7 Major Water Resources Projects in Narmada Basin . . . . . . . . . . . . . . . . . . . . . . . 107-5
107.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107-6
Chapter 108. Indus River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-1

108.1 Introduction to Indus River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-1
108.2 Physiography of Indus River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-1
108.3 River Network and Principal Hydrologic Units of Indus River Basin . . . . . . . 108-1
108.4 Climate of Indus River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-2
108.5 Hydrological Characteristics of Indus River Basin . . . . . . . . . . . . . . . . . . . . . . . . 108-3
108.6 Indus Water Treaty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-5
108.7 Major Water Resources Development Projects in Indus River Basin . . . . . . . 108-5
108.8 Groundwater Resources of Indus River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-9
108.9 Climate Change in Indus River Basin and Its Hydrologic Consequences . . . 108-9
108.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108-10
Chapter 109. The Mississippi River Basin . . . . . . . . . . . . . . . . . . . . . . . . . 109-1

109.1 Mississippi River Basin Physiography and Hydrology . . . . . . . . . . . . . . . . . . . . 109-1
109.2 Mississippi River Basin Climatology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109-1
109.3 Anthropogenic Changes in the River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109-2
109.4 Future Critical Challenges in the Mississippi River Basin . . . . . . . . . . . . . . . . . . 109-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109-5
Chapter 110. Colorado River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110-1

110.1 Introduction and Basin Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110-1
110.2 History of Water Resources Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110-1
110.3 Reservoirs and Other Water Management Facilities . . . . . . . . . . . . . . . . . . . . . . 110-2
110.4 Development and Use of the Colorado River Simulation System . . . . . . . . . . 110-3
110.5 Hydrologic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110-3
110.6 Generating Projected Future Flow Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110-4
110.7 Future Priorities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110-4
Chapter 111. Columbia River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111-1

111.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111-1
111.2 Basin History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111-1
111.3 River Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111-2

xxvi Contents
111.4 Current and Future Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111-4

111.5 Future Opportunities and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111-5
Chapter 112. St. Lawrence River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112-1

112.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112-1
112.2 Characteristics of the St. Lawrence River and Its Basin . . . . . . . . . . . . . . . . . . . 112-1
112.3 Hydrological Characteristics of the River . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112-5
Chapter 113. River Rhine Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-1

113.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-1
113.2 Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-1
113.3 Water Balance in the Rhine Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-4
113.4 Long-Term Variability of Hydrometeorological Variables in
the Rhine Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-4
113.5 Discharge Characteristics in Longitudinal Profile . . . . . . . . . . . . . . . . . . . . . . . . 113-5
113.6 The Runoff Regime of the Rhine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-6
113.7 Changes in the Runoff Characteristics of the Rhine Since the Beginning
of the Twentieth Century . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-9
113.8 Changes in the Runoff Regime of the Upper Rhine . . . . . . . . . . . . . . . . . . . . . . . 113-10
113.9 Changes in the Runoff Regime of the Middle and Lower Rhine . . . . . . . . . . . 113-10
113.10 Development in Extreme Runoff Situations: Flood . . . . . . . . . . . . . . . . . . . . . . . 113-10
113.11 Development in Extreme Runoff Situations: Low Water . . . . . . . . . . . . . . . . . . 113-10
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113-10
Chapter 114. Danube River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114-1

114.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114-1
114.2 History of the River System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114-1
114.3 Climate, Drainage Characteristics, and Hydrology . . . . . . . . . . . . . . . . . . . . . . . 114-2
114.4 Problem Faced . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114-3
114.5 Scope of Future Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114-5
Chapter 115. Ob River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115-1

115.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115-1
115.2 Ob River Basin and Discharge Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115-1
115.3 Streamflow Characteristics and Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115-1
115.4 Water Temperature and Thermal Regime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115-5
Chapter 116. Po River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116-1

116.1
River Basin Morphology and Geology, Geometry of the River Network . . . . 116-1
116.2
Climate and Meteorology. Genesis of Extreme Events and Droughts . . . . . . 116-1
116.3
Monitoring Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116-3
116.4
Hydrological Balance in the Po River Basin: Exploitation
of Water Resources and Sustainability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116-3
116.5 The River Regime: Variability, Seasonality, Long-Term Patterns . . . . . . . . . . . 116-3
116.6 History of Po River Floods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116-3
116.7 Flood Hazard Mitigation Along the Course of the Po River . . . . . . . . . . . . . . . 116-4
116.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116-4
Chapter 117. River Thames Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-1

117.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-1
117.2 The Thames Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-1
117.3 Hydrometric Measurement in the Thames Basin . . . . . . . . . . . . . . . . . . . . . . . . . 117-1
117.4 Droughts and Alleviation Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117- 2
117.5 Floods and Flood Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117- 3
117.6 Research and Forecasting Initiatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-4
117.7 Trends in Runoff Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-4
117.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-4

Contents xxvii
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-6
Useful Websites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117-6
Chapter 118. Managing Water in an Arid Land: The Murray

Darling Basin, Australia . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-1
118.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-1
118.2 The Murray Darling Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-2
118.3 The Changing Hydrology of the Murray Darling Basin . . . . . . . . . . . . . . . . . . . . 118-3
118.4 Potential for Ecosystem Collapse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-5
118.5 The Policy Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-5
118.6 What Future Do We Want for the Basin Anyway? . . . . . . . . . . . . . . . . . . . . . . . . . 118-6
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-7
Useful Websites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118-9
Chapter 119. The Great Lakes System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119-1

119.1 Introduction to the Great Lakes Hydrological System . . . . . . . . . . . . . . . . . . . . 119-1
119.2 The Great Lakes Water Budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119-2
119.3 Great Lakes Water levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119-5
Chapter 120. The East African Great Lakes . . . . . . . . . . . . . . . . . . . . . . . . 120.1

120.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 1
120.2 Lake Victoria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 1
120.3 Lake Tanganyika . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 3
120.4 Lake Malawi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 4
120.5 Lake Albert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 5
120.6 Lake Turkana . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 6
120.7 Lake Kivu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 7
120.8 Lake Edward . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. 9
Chapter 121. Aral Sea Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121-1

121.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121-1
121.2 Subsurface Flux Changes and Interactions with the Shrinking Sea . . . . . . . . 121-1
121.3 Surface Flux Changes and Interactions with Climate Change . . . . . . . . . . . . . 121-2
121.4 Opportunities and Challenges for Water Quantity and
Quality Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121-3
121.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121-4
Chapter 122. Baltic Sea Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-1

122.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-1
122.2 Physiography and Hydroclimatology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-1
122.3 Hydrology and Water Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-4
122.4 Water Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-6
122.5 Future Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-8
122.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122-9
Chapter 123. Black Sea Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123-1

123.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123-1
123.2 Geographical Location and Basic Morphometric Characteristics . . . . . . . . . . 123-1
123.3 Hydrological Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123-1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123-5
Chapter 124. The Caspian Sea Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-1

124.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-1
124.2 The Caspian Sea Watershed Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-1
124.3 The Volga River . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-1
124.4 Physicogeographical Conditions of the Caspian Sea . . . . . . . . . . . . . . . . . . . . . 124-3
124.5 Hydrometeorology and Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-3
124.6 Physical Oceanography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-3
124.7 Sea Level Variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-4

xxviii Contents
124.8 Marine Chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-4

124.9 Marine Biology and Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-5
124.10 Ecological Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-5
124.11 Oil Pollution of the Sea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-6
124.12 Seismic Activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-7
124.13 Desertification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-7
124.14 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124-7
PART 9. Applications and Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-1
Chapter 125. Design Rainfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-3

125.1 Purpose of Design Rainfalls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-3
125.2 Constructing Databases of Rainfall to Derive Design Rainfalls . . . . . . . . . . . . 125-4
125.3 AMS Versus PDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-5
125.4 Appropriate Probability Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-5
125.5 Regionalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-6
125.6 Deriving Sub-Daily and Sub-Hourly IDF Relationships . . . . . . . . . . . . . . . . . . . 125-7
125.7 Scaling Relationships in Design Rainfalls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-7
125.8 Developing Design Rainfall Grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-8
125.9 Uncertainty in Design Rainfall Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-8
125.10 Design Temporal Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-8
125.11 Design Spatial Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-9
125.12 Implications of Temperature Linked Non-Stationarity on the Design Rain . 125-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-11
Chapter 126. Probable Maximum Precipitation . . . . . . . . . . . . . . . . . . . 126-1

126.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-1
126.2 Deterministic Method of Estimation of PMP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-1
126.3 Probabilistic Method of Estimation of PMP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-10
126.4 Generalized Versus Basin-Specific PMP Estimates . . . . . . . . . . . . . . . . . . . . . . . . 126-13
126.5 All-Season Versus Seasonal PMP Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-13
126.6 Orographic Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-13
126.7 Spatial Variation of PMP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-13
126.8 Temporal Distribution of PMP—Development of PMS . . . . . . . . . . . . . . . . . . . . 126-14
126.9 Seasonal Variation of PMP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-16
126.10 Cautionary Notes on the Procedures for Estimation of New Set of PMP . . . . 126-16
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126-17
Chapter 127. Runoff Prediction in Ungauged Basins . . . . . . . . . . . . . . 127-1

127.1 The Prediction in Ungauged Basins Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127-1
127.2 Best Practice Recommendations for Predicting Runoff in
Ungauged Basins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127-1
127.3 Prediction of Floods in Ungauged Basins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127-2
127.4 Prediction of Low Flows in Ungauged Basins . . . . . . . . . . . . . . . . . . . . . . . . . . . 127-3
127.5 Prediction of Runoff Hydrographs in Ungauged Basins . . . . . . . . . . . . . . . . . 127-4
127.6 Where to Go from Here . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127-7
Chapter 128. Stochastic Streamflow Simulation and Forecasting . . . 128.1

128.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 1
128.2 Stochastic Simulation of Streamflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 1
128.3 Nonparametric Time Series Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 2
128.4 Multisite Streamflow Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 3
128.5 Nonstationary Streamflow Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 3
128.6 Streamflow Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 4
128.7 Stochastic Weather Generators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 4
128.8 Software and Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128. 5

Contents xxix
Chapter 129. Flood Forecasting and Flash Flood Forecasting—

Special Considerations in Hydrologic Modeling for
the Expressed Purpose of Flood and Flash
Flood Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-1
129.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-1
129.2 Real-Time Data Requirements and Forecasts and Data Quality Control . . . . 129-1
129.3 Computational Efficiency and Latency Requirements . . . . . . . . . . . . . . . . . . . . . 129-3
129.4 Data Assimilation and/or Adjusting Model Inputs, States, and Outcomes . 129-3
129.5 Use of Future Weather . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-3
129.6 Requirements for Regulation Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-3
129.7 Flood-Control and Water-Supply Reservoirs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.8 Reliability and Stability Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.9 Understanding and Conveying Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.10 Lead Time Considerations and Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.11 Temporal (Time Step) Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.12 Dissemination and Coordination with Customers, Partners, and the
General Public . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.13 Applications Related to Flash Flooding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.14 Special Temporal (Time Step) Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-4
129.15 Automated Data-Analysis Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-5
129.16 Automated Data Assimilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-5
129.17 Ungaged Watershed Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-5
129.18 Identification of Highly Vulnerable/at Risk Locations . . . . . . . . . . . . . . . . . . . . . 129-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129-5
Chapter 130. Reservoir Operation Design . . . . . . . . . . . . . . . . . . . . . . . . 130-1

130.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130-1
130.2 Reservoir Planning and Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130-1
130.3 Reservoir Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130-3
130.4 Future Trends in Reservoir Operation Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130-6
130.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130-7
Chapter 131. Floodplain Management . . . . . . . . . . . . . . . . . . . . . . . . . . . 131-1

131.1 Responses to Flood Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131-1
131.2 Evolution of Floodplain Management Practices . . . . . . . . . . . . . . . . . . . . . . . . . . 131-2
131.3 Concepts and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131-2
131.4 Certification in Floodplain Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131-3
131.5 Summary of Issues and Needed Advances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131-3

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131-3
Chapter 132. Storm Water Management, Best Management

Practices, and Low-Impact Development . . . . . . . . . . . . 132-1
132.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132-1
132.2 The Need for Stormwater Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132-2
132.3 Specific BMPs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132-3
132.4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132-4
132.5 Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132-4
Chapter 133. Flood Proofing and Infrastructure Development . . . . . 133-1

133.1 Introduction to Flood Proofing and Infrastructure Development . . . . . . . . . 133-1
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133-9
Chapter 134. Environmental Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-1

134.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-1
134.2 Evolution of Environmental Flow Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-2
134.3 Trade-offs in Development and Conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-2
134.4 Estimation of Environmental Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-2
134.5 Methodologies for Assessment of Environmental Flow Requirement . . . . . 134-3
134.6 Implementation of EF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-6
134.7 Environmental Flows in IWRM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-7
134.8 Future Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134-7

xxx Contents
Chapter 135. Drainage and Culvert Design . . . . . . . . . . . . . . . . . . . . . . . 135-1

135.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135-1
135.2 Fundamentals of Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135-1
135.3 Design Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135-4
135.4 Potential Impact of Climate Change on Culvert Design and Operation . . . . 135-6
135.5 Sustainable Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135-6
Chapter 136. Wetland and River Restoration . . . . . . . . . . . . . . . . . . . . . . 136-1

136.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136-1
136.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136-1
136.3 The Restoration Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136-2
136.4 Approaches to Wetland and River Restoration . . . . . . . . . . . . . . . . . . . . . . . . . . . 136-3
136.5 Advancing the Science and Practice of Stream and Wetland
Restoration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136-6
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136-7
Chapter 137. Institutional Framework for Water Management . . . . . 137-1

137.1 Sustainable Integrated Water Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-1
137.2 Water Management in Texas: An Illustrative Case Study . . . . . . . . . . . . . . . . . . 137-1
137.3 Water Management Communities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-1
137.4 Federal Agency Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-3
137.5 State Water Resources Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-4
137.6 Water Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-4
137.7 Environmental Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-5
137.8 Flood Mitigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-6
137.9 Institutional Aspects of Computer Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-6
137.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137-7
Chapter 138. Peak Water, Virtual Water, and Water Footprints:

New Definitions and Tools for Water Research
and Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138-1
138.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138-1
138.2 Peak Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138-2
138.3 Comparison of Peak Production in Oil and Water . . . . . . . . . . . . . . . . . . . . . . . . 138-3
138.4 Water Transfers and the Concept of Virtual Water . . . . . . . . . . . . . . . . . . . . . . . 138-4
138.5 Water Footprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138-6
138.6 Soft Water Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138-7
138.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138-8
Chapter 139. Transboundary Water Management . . . . . . . . . . . . . . . . . 139-1

139.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-1
139.2 Water Conflict . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-1
139.3 Sources of Water Conflict . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-2
139.4 International Water Conflicts Versus National Water Conflicts . . . . . . . . . . . . 139-2
139.5 Resolving Water Conflicts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-2
139.6 The Importance of Institutional Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-3
139.7 Hydro-Hegemony . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-3
139.8 International Water Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-3
139.9 Third Party Involvement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-3
139.10 Future Directions and Ways to Address New Problems . . . . . . . . . . . . . . . . . . . 139-4
139.11 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-4
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139-4
Chapter 140. Integrated River Basin Management . . . . . . . . . . . . . . . . 140-1

140.1 Integrated River Basin Management: A Framework and Process . . . . . . . . . . 140-1
140.2 Elements of IRBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140-1
140.3 Historical Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140-2
140.4 Institutional Arrangements for IRBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140-2
140.5 Technical Concepts and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140-4
140.6 Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140-5
140.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140-7

Contents xxxi
Chapter 141. Conflict Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-1

141.1 Conflict Exists Everywhere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-1
141.2 Game Theoretic Models for Conflict Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-1
141.3 Graph Model for Conflict Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-1
141.4 Fair Water Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-2
141.5 Compliance Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-4
141.6 Agent-Based Modeling and Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-5
141.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141-5
Chapter 142. Long-distance Water Transfers . . . . . . . . . . . . . . . . . . . . . . 142-1

142.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142-1
142.2 Transfers Among Basins, Regions and
Countries—Achievements and Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142-1
Further Thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142-6
Chapter 143. The Indian River-Linking Program . . . . . . . . . . . . . . . . . . . 143-1

143.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143-1
143.2 India’s Water Resources and the River-Linking Plan . . . . . . . . . . . . . . . . . . . . . . 143-1
143.3 Prognosis and Implication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143-5
143.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143-8
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143-8
Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143-8
Chapter 144. Irrigation Scheduling and Management . . . . . . . . . . . . . 144-1

144.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144-1
144.2 Soil-Plant-Atmosphere Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144-2
144.3 On-Farm Irrigation Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144-2
144.4 Canal Irrigation Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144-7
144.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144-8
144.6 Research Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144-8
Chapter 145. Rainwater Harvesting and Groundwater Recharge . . . 145-1

145.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145-1
145.2 Ancient Methods of Water Harvesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145-1
145.3 Watershed Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145-2
145.4 Rainwater-Harvesting Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145-3
145.5 Assessment of Rainwater Harvesting Using Remote Sensing and GIS . . . . . 145-7
145.6 Rainwater Harvesting Studies a Global Perspective . . . . . . . . . . . . . . . . . . . . . 145-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145-9
Chapter 146. Reuse-Reclaimed Water in Managed

Aquifer Recharge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-1
146.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-1
146.2 Treatment Mechanisms in Natural Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-1
146.3 Managed Aquifer Recharge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-2
146.4 Water Quality Considerations for Managed Aquifer Recharge . . . . . . . . . . . . 146-3
146.5 Surface Spreading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-3
146.6 Injection Wells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-4
146.7 Recovery of Reclaimed Water through Aquifer Storage and Recovery . . . . . 146-5
146.8 Subsurface Geochemical Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-6
146.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146-6
Chapter 147. River Bank Filtration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-1

147.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-1
147.2 River Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-1
147.3 Potential of River Bank Filtration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-2
147.4 Simulating RBF Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-3
147.5 Optimizing Distance of Well from the River . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-6
147.6 River Bank Filtration Sites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-6
147.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147-7

xxxii Contents
Chapter 148. Assessment of Climate Change Impacts on

Water Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-1
148.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-1
148.2 Climate Change Observations and Model-Based Projections . . . . . . . . . . . . . 148-1
148.3 Observations and Projections of Climate Change Impact on Water
Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-2
148.4 Aim of Modeling of Climate Change Impact on Freshwater Resources . . . . 148-3
148.5 Methodology of Modeling Climate Change Impacts on Water Resources . . 148-3
148.6 Hydrological Models for Climate Change Impact Assessment . . . . . . . . . . . . 148-4
148.7 Model Selection, Calibration, and Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-6
148.8 Examples of Applications of Hydrological Models for Climate
Change Impact Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-7
148.9 Propagation of Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-7
148.10 Gaps and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148-10
Chapter 149. Human Impacts on Hydrology . . . . . . . . . . . . . . . . . . . . . . 149-1

149.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149-1
149.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149-2
149.3 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149-4
149.4 Future Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149-7
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149-8
Chapter 150. Climate Change and Its Impacts on

Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-1
150.1
Climate Change: What Does It Mean? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-1
150.2
Causes of Climate Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-1
150.3
Measure of Climate Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-3
150.4
Impacts of Climate Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-3
150.5
Impacts of Climate Change on
The Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-3
150.6 Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150-7
Part 10. Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151-1
Chapter 151. Human-Hydrology Systems Modeling . . . . . . . . . . . . . . . 151-3

151.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151-3
151.2 Evolution of Human-Hydrology Systems Modeling . . . . . . . . . . . . . . . . . . . . . . 151-3
151.3 Methods for Modeling Human-Hydrology Systems . . . . . . . . . . . . . . . . . . . . . . 151-4
151.4 Applications of Human-Hydrology Systems Modeling . . . . . . . . . . . . . . . . . . . 151-6
151.5 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151-9
151.6 Future Directions and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151-11
Chapter 152. Variability of Hydrological Processes and

Systems in a Changing Environment . . . . . . . . . . . . . . . . 152-1
152.1 Climate Change and the Water Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152-1
152.2 Human Activities and the Water Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152-1
152.3 Intensification of the Hydrological Cycle in a Changing Environment . . . . . 152-2
152.4 Sustainability, Hydrologic Risk, and Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . 152-3
152.5 Tracking Hydrologic Change: Trend and Predictability . . . . . . . . . . . . . . . . . . . 152-3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152-4

Contents xxxiii
Chapter 153. Extraterrestrial Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153-1

153.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153-1
153.2 The Origin of the Earth’s Water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153-1
153.3 Water in Our Solar System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153-1
153.4 Water Beyond Our Solar System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153-3
153.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153-4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153-4
Chapter 154. Water Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-1

154.1 Availability of Freshwater Around the World . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-1
154.2 Current State of Water Affairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-1
154.3 Water Security—the Discourse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-3
154.4 Global Availability of Groundwater . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-3
154.5 Water Security as Part of Economic Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-4
154.6 Managing Water in a Changing World . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-4
154.7 Meeting the Water Gap: Unlocking the Potential of Green Water . . . . . . . . . . 154-4
154.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-5
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154-5
Chapter 155. Social Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-1

155.1
Water and Human Beings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-1
155.2
What is Social Hydrology? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-3
155.3
Great Hydraulic Mission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-4
155.4
Nature’s Talk Back . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-4
155.5
Evolution of Social Hydrology
as a New Discipline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-6
155.6 Challenges ahead in Water Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-7
155.7 Recent Trends and Developments
in Social Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-8
155.8 Climate Change and Future Issues
in Social Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-8
155.9 Future Direction in Social Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155-10
Chapter 156. Grand Challenges Facing the Hydrologic Sciences . . . . 156-1

156.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-1
156.2 Quality-Differentiated Water Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-1
156.3 Conversion Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-2
156.4 Scoping the Greater Hydrologic System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-2
156.5 Scoping Water Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-3
156.6 Emerging Infrastructure Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-3
156.7 Information Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-4
156.8 Prioritization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-4
156.9 Selected Priority Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-5
156.10 From Research to Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-6
156.11 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156-6
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I-1

Preface
The original Handbook of Applied Hydrology, edited by V.T. Chow, was published in 1964.
The Handbook has had a major influence on hydrology. Since then it has been widely used as
a reference book and has inspired numerous hydrologists by its depth and breadth of coverage.
During the intervening period of over 50 years hydrology has witnessed phenomenal develop-
ment in both depth and breadth, with the result that it has become a full-fledged geophysical
science while still serving as basis for engineering applications. The development has occurred
along four main lines. First, the 1960s witnessed the birth of a digital revolution and the com-
puting prowess has since expanded exponentially. Now-a-days portable laptops are manifold
more powerful than a large computer that then filled an entire room. The enormous comput-
ing capability permitted the birth of digital or numerical hydrology. Beginning with the devel-
opment of the Stanford Watershed Model in 1966, a large number of watershed models,
simulating virtually the entire hydrologic cycle, were developed. One can easily count in the
world over one hundred watershed hydrology models in use these days. The development of
watershed models is still continuing, encompassing simulation of other aspects that are associ-
ated with hydrology, such as atmospheric or hydrometeorological processes, hydraulics, cli-
mate change, ecosystems, geo-biochemistry, and human interactions. Another area has been
the solution of hydrodynamic equations, be those of surface water, subsurface water, ground-
water, water quality, and so on. Because of computers, it is now possible to simulate any com-
ponent of the hydrologic cycle in great detail and do inverse modeling as well.
Second, a tremendous growth in the development of new methods of solution, both deter-
ministic and stochastic, has occurred. On the deterministic side, new transform techniques,
optimization techniques, artificial intelligence methods, data mining, and numerical schemes
were developed. On the stochastic side, developments have been even greater. Examples are
copula theory, entropy theory, chaos theory, network theory, fractals, scaling theory, Monte
Carlo simulation, and so on.
Third, techniques for novel data collection, processing, storage, archiving, retrieval, and
sharing have reached unprecedented heights. Remote sensing methods, such as radar and
satellite technologies are now providing spatial data that were not even imagined prior to 1964.
Information technology is providing means for dissemination of information in seconds, per-
mitting hydrology to get closer to people as for example is happening in applications at farm-
ing scalwe. Another example is social hydrology, an emerging area that is gaining wide
recognition these days.
Fourth, new theories and new concepts have been developed, facilitated greatly by the afore-
mentioned three areas. We have witnessed the evolution of new concepts in geomorphology
and, more importantly, application of these concepts in watershed hydrology, hydrogeological
characterization, hydrometeorology, cryosphere, glaciology, geo-bio-chemical transport, land
use change, climate change, ecosystems science, medical hydrology, and the list goes on. In
some areas, developments have been so large that together they have given rise to new branch-
es of hydrology that were not even heard of prior to 1964.
It is, therefore, now appropriate to pause and take a stock of where we were, where we have
been and where the multiple disciplines that make hydrology should be going. This philosophy
served as the motivation for preparing the layout for the second version of the Handbook of
Applied Hydrology. The word “Applied” is important here. The emphasis in this version has
been on the application at the expense of deep science and mathematics. With these consider-
ations in mind, the subject matter of the handbook is divided into ten parts each containing a
xxxv

xxxvi Preface
number of chapters. The First Part is on fundamentals and contains three chapters that deal
with the hydrologic cycle, watersheds, river basins, land use, and water balance. This part
attempts to set the context for the handbook. For the science and engineering of hydrology to
progress, data and their collection and accessibility are the foundation. Therefore, the Second
Part of this handbook deals with data collection and processing. It comprises seven chapters
that discuss hydrometeorological data, streamflow data, rating curves, hydrologic information
systems, remote sensing and geographical information systems. This part is concluded with a
discussion on design of hydrologic networks.
Hydrologic analyses and syntheses entails tools from mathematics, statistics, physics, biol-
ogy, chemistry, information science, and social-economic science. Like hydrology, these areas
have witnessed enormous developments. Fortunately, hydrologists have been quite savvy and
open minded to borrow techniques from other disciplines and apply them in hydrology. Thus,
Part Three spans sixteen chapters and presents these methods. These chapters deal with arti-
ficial neural networks; fuzzy logic; evolutionary computing; relevance vector machines; har-
monic analysis and wavelets; outlier analysis; infilling of missing records; regression analysis;
time series analysis; change detection and nonstationarity; spatial analysis and geostatistical
methods; frequency distributions; model calibration and validation; sensitivity analysis;
Bayesian methods; optimization methods; nonparametric methods; uncertainty analysis and
decision making; risk-reliability assessment; scaling and fractals; nonlinear dynamics and
chaos; copula modeling; entropy theory; entropy production extremum principles; data-based
mechanistic modeling; decomposition methods; network theory; and hydroeconomic analysis.
Experimental and field data and a variety of methods to analyze them have helped uncover
hydrologic mysteries and better understand and model hydrologic processes. Therefore, Part
Four containing twenty six chapters is focused on hydrologic processes and modeling.
Beginning with a discussion of weather and climate, these chapters go on to discuss hydrocli-
matology; spatial and temporal analysis of precipitation; snowpack characteristics; precipita-
tion modeling; evapotranspiration; interception; detention and depression storage;
geomorphological characteristics; infiltration; soil moisture and the vadose zone; hydrogeo-
logic characterization; groundwater modeling; watershed runoff; streamflow generation;
snowmelt runoff generation; glacial melting; reservoir and channel routing; water logging and
salinization; surface water-groundwater interaction; saltwater intrusion; land subsidence;
hydraulic fracturing; catchment classification; rainfall-runoff modeling; continuous watershed
modeling; and calibration and evaluation of watershed models. The concluding chapter deals
with feasibility, engineering, and process models.
With growing awareness of the environment, water, air and soil quality started to receive
increasing attention in the 1970s and onwards. To that end, sediment and pollutant transport
are dealt with in Part Five that comprises nine chapters. Beginning with a discussion of water
quality, the chapters go on to discuss soil erosion; channel erosion and sediment transport;
sedimentation of floodplains, lakes and reservoirs; pollutant transport in surface water, vadose
zone and ground water; and salinization and salinity management in watersheds. The con-
cluding chapter discusses transport of biochemicals and microorganisms.
The availability of data and appropriate statistical methods facilitated analysis of extremes.
Part Six, therefore, includes nine chapters that discuss hydrometeorological and hydrologic
extremes. The subject matter includes atmospheric rivers; hurricanes and typhoons; extreme
rainfall; floods; flood frequency analysis; regional hydrological modeling and regional fre-
quency analysis; risk, reliability, return periods and hydrologic design; drought characteristics;
and analysis of low flow and drought.
In hydrology, many schools of thought are applied. One school emphasizes systems
approach where intrinsic system details are not considered in detail. This school of thought
was pervasive in the 1950s through the 1970s but the subject is still of great importance.
Therefore, systems hydrology constitutes the theme of Part Seven which contains twelve chap-
ters. These chapters discuss isotope hydrology; lake hydrology; urban hydrology; agricultural
hydrology; forest hydrology; coastal hydrology; wetland hydrology; arid zone hydrology; karst
hydrology; cryospheric hydrology; hydrology for transportation systems; and large-scale and
global hydrology.

Preface xxxvii
Part Eight deals with the hydrology of large river basins and basins of lakes and seas span-
ning thirty-two chapters. The large river basins included are those of the Amazon River;
Parana (Rio de la Plata) River; Orinoco River; Nile River; Congo River; Zambezi River;
Euphrates and Tigris Rivers; Yangtze River; Yellow River; Mekong River; Yenisei River; Lena
River; Brahmaputra River; Ganga River; Narmada River; Indus River; Mississippi River;
Colorado River; Columbia River; St. Lawrence River; Rhine River; Danube River; Ob River;
Po River; Thames River; and Murray Darling River. The basins of the Great Lakes System and
the African Great Lakes are described in two chapters. The last four chapters on the basins of
the Aral Sea; the Baltic Sea; the Black Sea; and the Caspian Sea conclude this part.
Because the Handbook emphasizes application, Part Nine comprising twenty four chapters
is devoted to applications and design. Beginning with a discussion of design rainfall, the chap-
ters go on to discuss probable maximum precipitation; prediction in ungaged basins; stochas-
tic streamflow simulation and forecasting; flood forecasting and flash and real-time flood
forecasting; reservoir design, regulation and operation; floodplain management; stormwater
management and low impact development; flood proofing and infrastructure development;
environmental flows; drainage and culvert design; wetland and river restoration; institutional
framework for water management; virtual water, water footprint, and peak water; transbound-
ary water management; integrated river basin management; conflict resolution; long distance
water transfer; the Indian river linking program; irrigation scheduling and management;
rainwater harvesting and groundwater recharge; reuse-reclaimed water in managed aquifer
recharge; and river bank filtration. This part is concluded with a discussion on the assessment
of climate change impacts and water resources.
Where the multiple disciplines of hydrology progress to is partly conjectured in the con-
cluding Part Ten that deals with the future in seven chapters. These chapters discuss human
impacts on hydrology, climate change and its impact on the hydrologic cycle, human-hydrol-
ogy systems modeling, variability of hydrological processes and systems in a changing envi-
ronment, extraterrestrial water, and water security. This part is concluded with a discussion of
grand challenges.
It is hoped that the Handbook will become useful to college faculty, graduate students,
and researchers as well as practitioners in hydrology, water science and engineering, water
resources management, urban development, hydrometeorology, geosciences, environmental
and ecological sciences, and agricultural and forest sciences.
Vijay P. Singh
Texas A&M University
College Station, Texas

Foreword
Looking back when Ven Te Chow published his Handbook of Applied Hydrology in 1964, it was
without any doubt a landmark event for the water sector. Using his personal status as one of
the world’s leading water experts, he successfully managed to convince leading world authori-
ties on different aspects of water to contribute chapters in their areas of expertise. After the
book was published, it was universally acknowledged to be the most definitive and authorita-
tive text available on water resources management for well over a decade. This mammoth book
at 1,467 pages summarized and elaborated all the latest thinking and developments in different
areas of water.
It was a tribute to Chow’s encyclopaedic knowledge and versatility that he wrote several
chapters on very different topics himself. Later on, he confided to me that a few of the chapters
did not meet his high standards and thus he had to substantially rewrite them. Looked from
any direction, the depth and quality of the Handbook was a major contribution which ensured
that almost a generation of professionals used the text whenever they needed reliable informa-
tion on any aspect of water.
The Handbook was directly responsible for forging a lifelong friendship between Chow and
myself. When the book was published, I was starting my career at the University of Strathclyde,
Glasgow, as a lecturer. One of the British water journals requested me to write a review of all
the water texts available and suggest 3 to 4 that I considered the best and most useful. In the
mid-Sixties, it was possible to read all the texts on water that were available since the number
was rather limited.
After I published my assessment in late 1964, I received a letter from a somewhat annoyed
Chow since this list did not include his latest tome Handbook of Applied Hydrology. I explained
to him that I did not omit the book deliberately. I fact, I knew the book was published and
available in the United States. Thus, I had written to McGraw-Hill asking if I could get a copy
of the book for my assessment purposes. McGraw-Hill responded by saying that the book was
indeed available in the United States but it would not be available in the United Kingdom for
another six months.
Chow, being the gentleman he was, immediately apologized for his letter. Shortly thereafter,
I moved to Canada. In the sixties, when I was a Director in the Government of Canada, Chow
invited me to come to the University of Illinois and give a talk during a meeting of the hydrol-
ogy professors in Urbana.
In early 1970, when Chow was considering the possibility of creating an interdisciplinary
and multisectoral professional water association, the International Water Resources
Association, he contacted me and requested me if I could be a founding member and help him
to establish a Canadian chapter of the new Association. We remained very good friends until
his untimely death.
There have been exponential scientific and technological developments in all water-related
areas, especially during the post-1980 period. Thus, advances in knowledge had made the
Handbook obsolete by the early 1980s. It is indeed a tribute to Chow’s genius that the book had
remained useful for such a long period.
xxxix

xl Foreword
I am delighted that Prof. Vijay P. Singh, Caroline and William N. Lehrer Distinguished
Chair in Water Engineering, Texas A&M University, College Station, Texas, is completely
rewriting and updating Chow’s Handbook. Prof. Singh is a prominent international authority
in hydrologic sciences, and, in my view, he is easily within the top 0.1% of all global water-
related researchers. Only a person of Prof. Singh’s stature can undertake such a Herculean task.
Following Chow’s footsteps mean these are big shoes to fill. From the manuscript, it is obvi-
ous to me that Prof. Singh has not only filled the big shoes completely but he has achieved
almost the impossible with panache and style.
Even with knowledge advancing at a significantly higher rate at present, compared to when
Chow first published the book in 1964, I have no doubt that the new and updated version of
this Handbook will remain the authoritative text for at least a decade. The new text is a fitting
memory and tribute to Chow’s enduring legacy, and also a fitting confirmation of Prof. Singh’s
global stature.
Singapore
October 18, 2015
Asit K. Biswas
Distinguished Visiting Professor
Lee Kuan Yew School of Public Policy
National University of Singapore

Acknowledgments
This Handbook is a result of the collective effort of the authors who have contributed the
chapters. It has been my privilege and honor to have known most of the authors for a long
time. They have long been at the forefront of hydrologic research and teaching or practice and
their rich experience and knowledge are hopefully reflected in the chapters they have written.
Their contribution symbolizes their love of labor and their desire to contribute to advancing
hydrology, for their time and effort are ad honorem. I am much grateful to the authors for
accepting my invitation to write the chapters and their cooperation during the preparation of
the Handbook. Without them, this Handbook would not have come to fruition, and it clearly
belongs to them.
I would like to express my deep gratitude to the members of two advisory boards: aca-
demic and practitioners. When the Handbook was being conceptualized, I often turned to
them for their advice on the chapter layout and identifying potential chapter authors. They
were more than willing to help and responded to my request promptly.
The chapters were peer reviewed and finding three or more reviewers for each of 156 chap-
ters was a challenge. They are too numerous to mention here, but I am grateful for the review-
ers’ constructive reviews.
I would like to acknowledge my family (wife Anita, son Vinay, daughter Arti, and grandsons
Ronin, Kayden and Davin) for their support and affection while I worked on this handbook
project. Without them, my task of completing the handbook would have been much more
difficult.
Finally, I want to express my gratitude to the McGraw-Hill staff, particularly Lauren
Poplawski, who were always willing to help and provide advice and suggestions.
xli

Contributors
Benjamin Abban M. Babbar-Sebens

Department of Civil and Environmental Engineering School of Civil and Construction Engineering
University of Tennessee Oregon State University
Knoxville, Tennessee Corvallis, Oregon
M. C. Acreman Daniel W. Baker

Centre for Ecology and Hydrology Department of Civil and Environmental Engineering
Crowmarsh Gifford Colorado State University
Wallingford, United Kingdom Fort Collins, Colorado
Emanuele Baratti
Lauren Adams
Department DICAM
Center for Watershed Sciences
University of Bologna
University of California, Davis
Bologna, Italy
Davis, California
Donald E. Barbe
Pradeep Aggarwal Department of Civil and Environmental Engineering
Isotope Hydrology Section University of New Orleans
Division of Physical and Chemical Sciences New Orleans, Louisiana
Department of Nuclear Sciences and Applications
International Atomic Energy Agency (IAEA) Henry Barousse
Vienna, Austria Louisiana Department of Transportation and Development, Retired
Robert Bastian
Mohammad Z. Al-Hamdan
U.S. Environmental Protection Agency
Universities Space Research Association
Washington, D.C.
NASA/Marshall Space Flight Center
National Space Science and Technology Center Peter Bauer-Gottwein
Huntsville, Alabama Department of Environmental Engineering
Technical University of Denmark
Doug Alsdorf Kongens Lyngby, Denmark
Byrd Polar and Climate Research Center
The Ohio State University Hiroko Kato Beaudoing
Columbus, Ohio Earth System Science Interdisciplinary Center
University of Maryland
Devendra M. Amatya College Park, Maryland
Center for Forested Wetlands Research Beijing Normal University
Southern Research Station, USDA Forest Service Research and Development Centre
Cordesville, South Carolina Nippon Koei Co. Ltd.
(Consulting Engineers)
Daniel P. Ames Tsukuba, Japan
Department of Civil and Environmental Engineering
Brigham Young University Kati Bell
Provo, Utah Water Reuse Global Practice Leader, MWH Global
Nashville, Tennessee
Michael Anderson
Joerg Uwe Belz
State Climatologist, California Department of Water Resources
Department of Hydrometry and Hydrological Survey Federal Institute of Hydrology
Sacramento, California
Koblenz, Germany
Mazdak Arabi Lars Bengtsson
Department of Civil and Environmental Engineering Department of Water Resources Engineering
Colorado State University Lund Institute of Technology
Fort Collins, Colorado Lund University
Lund, Sweden
Luis J. Araguas-Araguas
Isotope Hydrology Section Ronny Berndtsson
Division of Physical and Chemical Sciences Department of Water Resources Engineering
Department of Nuclear Sciences and Applications and Center for Middle Eastern Studies
International Atomic Energy Agency (IAEA) Lund University
Vienna, Austria Lund, Sweden
Mustafa M. Aral Daniela Biondi

School of Civil and Environmental Engineering Institute of Hydraulic Construction
Georgia Institute of Technology University of Bologna
Atlanta, Georgia Bologna, Italy
Bhavna Arora Brian P. Bledsoe

Earth Sciences Division College of Engineering
Lawrence Berkeley National Laboratory University of Georgia
Berkeley, California Athens, Georgia
xliii

xliv Contributors
Günter Blöschl Yang Cheng

Centre for Water Resource Systems Department of Civil and Environmental Engineering
Institut für Wasserbau und Ingenieurhydrologie Syracuse University
Technische Universität Wien Syracuse, New York
Wien, Austria
Nannan Cheng
Istvan Bogardi State Key Laboratory of Soil Erosion and Dryland Farming on the Loess Plateau
Professor, Department of Meteorology Institute of Soil and Water Conservation
Eotvos University Chinese Academy of Sciences and Ministry of Water Resources
Budapest, Hungary Northwest Agriculture and Forest University
Department of Civil Engineering Yangling, China
University of Nebraska Lincoln
Lincoln, Nebraska Ekaterina V. Chuprina
Water Problems Institute of the Russian Academy of Sciences
John D. Bolten Moscow, Russia
Hydrological Sciences Branch
NASA Goddard Space Flight Center Theodore G. Cleveland
Associate Professor, Department of Civil, Environmental, and Construction Engineering
André Bouchard Texas Tech University
Program Manager, National Hydrological Service Meteorological Service of Canada Lubbock, Texas
Environment and Climate Change Canada
Québec, Canada Walter Collischonn
Institute of Hydraulic Research
Armando Brath Universidade Federal do Rio Grande do Sul
Department of Civil, Chemical, Environmental and Materials Engineering - DICAM IPH-UFRGS
ALMA MATER STUDIORUM - Università di Bologna Porto Alegre, Brazil
Bologna, Italy
Zhentao Cong
Juan Martín Bravo Department of Hydraulic Engineering
Instituto de Pesquisas Hidráulicas Tsinghua University
Universidade Federal do Rio Grande do Sul Beijing, China
IPH-UFRGS
Porto Alegre, Brazil James Connaughton
Richmond, California
Dmitry A. Burakov
Head of the Department of Environmental Engineering, Corrado Corradini
Krasnoyarsk State Agrarian University Department of Civil and Environmental Engineering
Krasnoyarsk, Russia University of Perugia
Perugia, Italy
Donald H. Burn
Alva Couch
Department of Computer Science
University of Waterloo
Tufts University
Waterloo, Canada
Medford, Massachusetts
Aaron R. Byrd
Research Civil Engineer, Coastal and Hydraulics Laboratory Norman H. Crawford
U.S. Army Engineer Research and Development Center President, Hydrocomp, Inc.
Vicksburg, Mississippi Menlo Park, California
James F. Cruise
Ximing Cai
University of Alabama in Huntsville
University of Illinois at Urbana-Champaign
Huntsville, Alabama
Urbana, Illinois
Allegra da Silva
Darryl Carlyle-Moses
Water Reuse Practice Leader - Rocky Mountain Region, MWH Global
Department of Geography and Environmental Studies
Denver, Colorado
Thompson Rivers University
Kamloops, Canada Baptiste Dafflon
Earth Sciences Division
Attilio Castellarin Lawrence Berkeley National Laboratory
Department of Civil, Chemical, Environmental and Materials Engineering - DICAM Berkeley, California
ALMA MATER STUDIORUM - Università di Bologna
Bologna, Italy Kumer Pial Das
Department of Mathematics
Serena Ceola
Lamar University
Department of Civil, Chemical, Environmental and Materials Engineering - DICAM
Beaumont, Texas
ALMA MATER STUDIORUM - Università di Bologna
Bologna, Italy Claus Davidsen
Department of Environmental Engineering
Cem P. Cetinkaya
Technical University of Denmark
Dokuz Eylül University
Kongens Lyngby, Denmark
Water Resources Management Research & Application Center (SUMER)
Izmir, Turkey Rodrigo Cauduro Dias de Paiva
Maria A. Charina Institute of Hydraulic Research
Water Problems Institute of the Russian Academy of Sciences Federal University of Rio Grande do Sul
Moscow, Russia Porto Alegre, Brazil
Fateh Chebana Zhi-Qiang Deng

Institut national de la recherche scientifique Department of Civil and Environmental Engineering
Centre Eau-Terre-Environnement Louisiana State University
Québec, Canada Baton Rouge, Los Angeles
Ji Chen Georgia Destouni

Department of Civil Engineering Department of Physical Geography
The University of Hong Kong Stockholm University
Hong Kong, China Stockholm, Sweden

Contributors xlv
Michael DeWeese Donald K. Frevert

National Weather Service Hydraulic Engineer and Team Leader, Retired
North Central River Forecast Center Technical Services Center
Chanhassen, Minnesota Bureau of Reclamation
Lakewood, Colorado
Daniel H. Doctor
Eastern Geology and Paleoclimate Science Center Devin L. Galloway
U.S. Geological Survey Water Mission Area
Reston, Virginia U.S. Geological Survey
Indianapolis, Indiana
Alessio Domeneghetti
Department of Civil, Chemical, Environmental and Materials Engineering - DICAM Timothy S. Gambrell
ALMA MATER STUDIORUM - Università di Bologna Executive Director, Mississippi River Commission U.S. Army Corps of Engineers
Bologna, Italy Mississippi Valley Division
Vicksburg, Mississippi
A. S. Donigian, Jr.
Principal Consultant, AQUA TERRA Consultants Christian Genest
Mountain View, California Department of Mathematics and Statistics
Charles W. Downer McGill University
Research Civil Engineer, Coastal and Hydraulics Laboratory Montréal, Canada
U.S. Army Engineer Research and Development Center
Timothy R. Ginn
Washington Sate University
André Dozier Pullman, Washington
Jorge Gironás
Colorado State University
Departamento de Ingeniería Hidráulica y Ambiental
Fort Collins, Colorado
Pontificia Universidad Católica de Chile
Jim Dumont Santiago, Chile
Water Infrastructure Specialist
Peter H. Gleick
Salt Spring Island, Canada
Pacific Institute
Dipankar Dwivedi Oakland, California
Lawrence Berkeley National Laboratory Narendra Kumar Gontia
Berkeley, California Principal and Dean, College of Agricultural Engineering and Technology, Junagadh
Junagadh Agricultural University, Junagadh
Walter L. Ellenburg Gujarat, India
University of Alabama in Huntsville Jonathan L. Goodall
Huntsville, Alabama Department of Civil and Environmental Engineering
University of Virginia
A. Ercan Charlottesville, Virginia
Hydrologic Research Laboratory
Department of Civil and Environmental Engineering David C. Goodrich
University of California USDA-ARS
Davis, California Southwest Watershed Research Center
Tucson, Arizona
Alvar Escriva-Bou
Research Fellow, Water Policy Center R. S. Govindaraju
Public Policy Institute of California Delon and Elizabeth Hampton Hall of Civil Engineering
San Francisco, California Purdue University
West Lafayette, Indiana
James S. Famiglietti
NASA Jet Propulsion Laboratory Robert E. Griffin
California Institute of Technology Atmospheric Science Department
Pasadena, California Earth System Science Program
University of Alabama in Huntsville
Koren Fang
Huntsville, Alabama
School of Civil and Environmental Engineering
The University of New South Wales Neil S. Grigg
Sydney, Australia Professor, Department of Civil and Environmental Engineering
Liping Fang
Fort Collins, Colorado
Department of Mechanical and Industrial Engineering
Faculty of Engineering and Architectural Science Andrew D. Gronewold
Ryerson University NOAA (Great Lakes Environmental Research Laboratory)
Toronto, Canada University of Michigan (Civil and Environmental Engineering)
Ann Arbor, Michigan
Bonifacio Fernandez
Departamento de Ingeniería Hidráulica y Ambiental Georges Gulemvuga
Pontificia Universidad Católica de Chile Commission Internationale du bassin Congo-Oubangui-Sangha
Santiago, Chile Kinshasa-Gombe, Democratic Republic of the Congo
Stefan Finsterle Orhan Gunduz

Earth Sciences Division Dokuz Eylul University
Lawrence Berkeley National Laboratory Department of Environmental Engineering
Berkeley, California Izmir, Turkey
Faith A. Fitzpatrick Hoshin V. Gupta

Research Hydrologist, U.S. Geological SurveyWisconsin Department of Hydrology and Water Resources
Water Science Center University of Arizona
Tucson, Arizona
Darrell Fontane
Department of Civil and Environmental Engineering Semyon M. Guziy
Colorado State University Engineer, Krasnoyarsk Hydroproject Institute
Fort Collins, Colorado Krasnoyarsk, Russia

xlvi Contributors
Alan F. Hamlet Justin Huntington

Department of Civil and Environmental Engineering and Earth Sciences Desert Research Institute
University of Notre Dame Division of Hydrologic Sciences
Notre Dame, Indiana Western Regional Climate Center
Reno, Nevada
Cameron Handyside
Earth System Science Center John C. Imhoff
University of Alabama in Huntsville Senior Scientist, AQUA TERRA Consultants
Huntsville, Alabama Ouray, Colorado
K. Ishida
Mohamed M. Hantush
Research Hydrologist, National Risk Management Research Laboratory
U.S. Environmental Protection Agency
University of California
Cincinnati, Ohio
Davis, California
Umesh Haritashya Sharad K. Jain
Department of Geology Water Resources Systems Division
University of Dayton National Institute of Hydrology
Dayton, Ohio Roorkee, India
Bekki Ward Harjo L. Douglas James
Senior Hydrologist, National Weather Service Fairfax, Virginia
Arkansas-Red Basin River Forecast Center
A. W. Jayawardena
Tulsa, Oklahoma
Department of Civil Engineering
Nilgun B. Harmancioglu The University of Hong Kong
Dokuz Eylul University Hong Kong, China
Water Resources Management Research & Application Center (SUMER) Research and Development Center
Izmir, Turkey Nippon Koei Co. Ltd. (Engineering Consultants)
Tsukuba, Japan
Julien J. Harou
School of Mechanical, Aerospace and Civil Engineering Shanhu Jiang
The University of Manchester Lecturer, College of Hydrology and Water Resources
Manchester, United Kingdom Hohai University
Nanjing, China
Robert Hartman
Kenji Jinno
Hydrologist in Charge, California-Nevada River Forecast Center
Department of Urban & Environmental Engineering
National Weather Service
Graduate School of Engineering
Sacramento, California
Kyushu University
Hongming He Fukuoka, Japan
Institute of Soil and Water Conservation Fiona Johnson
Chinese Academy of Sciences and Ministry of Water Resources School of Civil and Environmental Engineering
Yangling, China The University of New South Wales
Janet S. Herman Sydney, Australia
Professor, Department of Environmental Sciences V. Jothiprakash
University of Virginia Department of Civil Engineering
Charlottesville, Virginia Indian Institute of Technology Bombay
Mumbai, India
Kith W. Hipel
Department of Systems Design Engineering Tijana Jovanovic
University of Waterloo Department of Civil and Environmental Engineering
Waterloo, Canada The Pennsylvania State University
State College, Pennsylvania
Yoshinari Hiroshiro
Department of Urban & Environmental Engineering Latif Kalin
Graduate School of Engineering Professor, School of Forestry and Wildlife Sciences
Kyushu University Auburn University
Fukuoka, Japan Auburn, Alabama
Jagath Kaluarachchi
Michael T. Hobbins
Professor of Civil and Environmental Engineering
Research Hydrologist, Earth System Research Laboratory & Cooperative Institute
College of Engineering
for Research in Environmental Sciences
Utah State University
NOAA-Earth System Research Laboratory
Logan, Utah
Boulder, Colorado
Raghupathy Karthikeyan
Robert R. Holmes, Jr. Department of Biological and Agricultural Engineering
Hydrologist and National Flood Hazard Coordinator, Office of Surface Water Texas A&M University
U.S. Geological Survey College Station, Texas
Rolla, Missouri
Laila Kasuri
Jeffery S. Horsburgh Lead Analyst and Senior Lecturer, Centre for Water Informatics and Technology
Department of Civil and Environmental Engineering Lahore University of Management Sciences
Utah State University Lahore, Pakistan
Logan, Utah Dmitri Kavetski
Chi Hua Huang School of Engineering
National Soil Erosion Research Laboratory University of Newcastle
West Lafayette, Indiana Callaghan, Australia
M. Levent Kavvas
Denis Arthur Hughes
Institute for Water Research
Rhodes University
Grahamstown, South Africa
Davis, California
Rui Hui Akira Kawamura
Center for Watershed Sciences Department of Civil and Environmental Engineering
University of California, Davis Tokyo Metropolitan University
Davis, California Tokyo, Japan

Contributors xlvii
Soksamnang Keo Zbigniew W. Kundzewicz

State Key Laboratory of Soil Erosion and Dryland Farming on the Loess Plateau Institute of Agricultural and Forest Environment
Institute of Soil and Water Conservation Polish Academy of Sciences
Chinese Academy of Sciences and Ministry of Water Resources Poznan, Poland
Northwest Agriculture and Forest University
John Labadie
Yangling, China
Greg Kerr Colorado State University
Office of Water Programs Fort Collins, Colorado
University of Wyoming David Labat
Laramie, Wyoming Géosciences Environnement Toulouse
Toulouse, France
Martin Keulertz
Agricultural and Biological Engineering Venkataraman (Venkat) Lakshmi
Purdue University Department of Earth and Ocean Sciences
West Lafayetter, Indiana University of South Carolina
Columbia, South Carolina
Jongho Keum Upmanu Lall
Department of Civil Engineering Department of Earth and Environmental Engineering and
McMaster University Department of Civil Engineering and Engineering Mechanics
Hamilton, Canada Columbia Water Center
International Research Institute for Climate and Society
C. Prakash Khedun
Columbia University
Department of Biological and Agricultural Engineering
New York, New York
College Station, Texas Alain Laraque
GET - UMR CNRS / IRD / UPS - UMR 5563 du CNRS, UMR234 de l'IRD
Max Kigobe Toulouse, France
College of Engineering, Design, Art and Technology (CEDAT) Magnus Larson
Makerere University Department of Water Resources Engineering
Kampala, Uganda Lund Institute of Technology
Lund University
Thomas R. Kjeldsen Lund, Sweden
Department of Architecture and Civil Engineering Se-Yeun Lee
University of Bath Climate Impacts Group, College of the Environment
Bath, United Kingdom University of Washington
Seattle, Washington
Barbara A. Kleiss
Director, Mississippi River Science & Technology Program U.S. Army Corps of Engineers Stanley A. Leake
Mississippi Valley Division Water Mission Area
Vicksburg, Mississippi U.S. Geological Survey
Tucson, Arizona
Randal Koster
Chiara Lepore
Global Modeling and Assimilation Office
Ocean and Climate Physics
NASA Goddard Space Flight Center
Lamont-Doherty Earth Observatory of Columbia University
Greenbelt, Maryland
Palisades, New York
Andrey Kostianoy L. Ruby Leung
Chief Scientist, Ocean Experimental Physics Laboratory Atmospheric Sciences and Global Change Division
P.P. Shirshov Institute of Oceanology Pacific Northwest National Laboratory
Moscow, Russia Richland, Washington
Benjamin Lord
Demetris Koutsoyiannis
Raleigh, North Carolina
Department of Water Resources and Environmental Engineering
School of Civil Engineering Yajie Lu
National Technical University of Athens State Key Laboratory of Soil Erosion and Dryland Farming on the Loess Plateau
Zographou, Greece Institute of Soil and Water Conservation
Chinese Academy of Sciences and Ministry of Water Resources
Peter Krahe Northwest Agriculture and Forest University
Contact person of subject group River Basin Modelling Department Yangling, China
Water Balance, Forecasting and Predictions Federal Institute of Hydrology
Jay R. Lund
Rhineland-Palatinate, Germany
Valentina Krysanova
Potsdam Institute for Climate Impact Research
Davis, California
Potsdam, Germany
Gil Mahe
Lev Kuchment IRD, Laboratoire Hydro Sciences
Laboratory of Water Cycle Montpellier, France
Water Problems Institute
D. R. Mailapalli
Russian Academy of Sciences
Agricultural and Food Engineering Department
Moscow, Russia
Indian Institute of Technology Kharagpur
George Kuczera Kharagpur, India
School of Engineering Terry Marsh
University of Newcastle Centre for Ecology & Hydrology
Callaghan, Australia Wallingford, United Kingdom
Bhishm Kumar Philip Marsh

Ex Professional Staff, Isotope Hydrology Section Water Science Wilfrid Laurier University
International Atomic Energy Agency (IAEA) Waterloo, Canada
Vienna, Austria José Pedro Matos
Ex Head, Hydrological Investigations Division Laboratoire de constructions hydrauliques (LCH)
National Institute of Hydrology Ecolepolytechniquefédérale de Lausanne (EPFL)
Roorkee, India Lausanne, Switzerland

xlviii Contributors
Larry W. Mays Dauren Mussabek

Arizona State University Department of Water Resources Engineering
Tempe, Arizona Lund University
Lund, Sweden
Richard H. McCuen
The Ben Dyer Professor of Civil & Environmental Engineering Kei Nakagawa
Civil and Environmental Engineering Department Nagasaki University
University of Maryland Graduate School of Fisheries Science and Environmental Studies
College Park, Maryland Nagasaki University
Nagasaki, Japan
Steven G. McNulty
Mohamed K. Nassar
Eastern Forest Environmental Threat Assessment Center
Southern Research Station, USDA Forest Service
University of California Davis
Raleigh, North Carolina
Davis, California
Alfonso Mejia Environmental Studies and Research Institute
Department of Civil and Environmental Engineering University of Sadat City
The Pennsylvania State University Sadat, Egypt
State College, Pennsylvania Grey S. Nearing
Tarek Merabtene NASA/GSFC
Department of Civil and Environmental Engineering Hydrological Sciences Branch
College of Engineering Greenbelt, Maryland
University of Sharjah WaiWah Ng
Sharjah, United Arab Emirates Department of Civil Engineering
Lakehead University
Venkatesh Merwade
Thunder Bay, Canada
Lyles School of Civil Engineering
Purdue University John Nielsen-Gammon
West Lafayette, Indiana Department of Atmospheric Sciences
Guziy Semyon Mikhailovich College Station, Texas
Institute “Krasnoyarskgidroproekt”
Krasnoyarsk State Agricultural University Jun Niu
Krasnoyarsk branch of JSC “SibENTC” College of Water Resources and Civil Engineering
Krasnoyarsk, Russia China Agricultural University
Beijing, China
Pavol Miklanek
Robert K. Niven
Institute of Hydrology
School of Engineering and Information Technology
Slovak Academy of Sciences
The University of New South Wales
Bratislava, Slovakia
Canberra, Australia
Srikanta Mishra P. E. O’Connell
Institute Fellow & Chief Scientist, Energy & Environment School of Civil Engineering and Geosciences
Battelle Memorial Institute Newcastle University
Columbus, Ohio Newcastle upon Tyne, United Kingdom
Ashok K. Mishra Greg O’Donnell
Glenn Department of Civil Engineering School of Civil Engineering and Geosciences
Clemson University Newcastle University
Clemson, South Carolina Newcastle upon Tyne, United Kingdom
Natalia N. Mitina Robert Occhipinti

Water Problems Institute of the Russian Academy of Sciences United States Army Corps of Engineers, Retired
Moscow, Russia Fred L. Ogden
Department of Civil and Architectural Engineering and
Iliana E. Mladenova
Haub School of Environment and Natural Resources
Hydrological Sciences Branch
University of Wyoming
Laramie, Wyoming
Greenbelt, Maryland
N. Ohara
Binayak P. Mohanty Department of Civil and Architectural Engineering
Biological and Agricultural Engineering University of Wyoming
Texas A&M University Laramie, Wyoming
C. S. P. Ojha
Rabi H. Mohtar Department of Civil Engineering
Biological and Agricultural Engineering and Indian Institute of Technology Roorkee
Zachry Department of Civil Engineering Roorkee, India
Taha B. M. J. Ouarda
iWATER Center,
Kazuro Momii Masdar Institute of Science and Technology
Professor, Faculty of Agriculture Abu Dhabi, United Arab Emirates
Kagoshima University Hydro-Climate Modeling Lab
Korimoto, Japan INRS-ETE
Quebec, Canada
Alberto Montanari
Department of Civil, Chemical, Environmental and Materials Engineering - DICAM Hisashi Ozawa
ALMA MATER STUDIORUM - Università di Bologna Graduate School of Integrated Arts and Sciences
Bologna, Italy Hiroshima University
Higashi-Hiroshima, Japan
R. Morbidelli
Umed S. Panu
Perugia University
Lakehead University
Perugia, Italy
Thunder Bay, Canada
Biswajit Mukhopadhyay Simon Michael Papalexiou
National Practice Leader, Water Resources Department of Water Resources and Environmental Engineering
North American Infrastructure School of Civil Engineering
Jacobs Engineering Group, Inc. National Technical University of Athens
Dallas, Texas Zographou, Greece

Contributors xlix
Thanos Papanicolaou Manuel Pulido-Velazquez

Department of Civil and Environmental Engineering Research Institute of Water and Environmental Engineering (IIAMA)
University of Tennessee Universitat Politècnica de València
Knoxville, Tennessee Valencia, Spain
Sandra Pavlovic Narendra Singh Raghuwanshi

University Corporation for Atmospheric Research Agricultural and Food Engineering Department
Boulder, Colorado Indian Institute of Technology Kharagpur
Kharagpur, India
Pavla Pekarova
Ataur Rahman
Institute of Hydrology
Water and Environmental Engineering
Slovak Academy of Sciences
School of Computing, Engineering and Mathematics
Bratislava, Slovakia
Western Sydney University
Angela Pelle Penrith, Australia
Balaji Rajagopalan
University of Alabama
Department of Civil, Environmental and Architectural Engineering
Tuscaloosa, Alabama
University of Colorado
Mauri Pelto Boulder, Colorado
Department of Environmental Science Harji D. Rank
Nichols College Professor, Department of Soil and Water Engineering, College of Agricultural Engineering
Dudley, Massachusetts and Technology, Junagadh
Junagadh Agricultural University, Junagadh
Silvio J. Pereira-Cardenal
Gujarat, India
COWI A/S
Lyngby, Denmark Liliang Ren
Professor, State Key Laboratory of Hydrology, Water Resources and Hydraulic Engineering
Magnus Persson
Hohai University
Department of Water Resources Engineering
Nanjing, China
Lund Institute of Technology
Lund University Ben Renard
Lund, Sweden School of Engineering
University of Newcastle
Muthiah Perumal Callaghan, Australia
Professor, Department of Hydrology
Indian Institute of Technology Roorkee Pedro J. Resptrepo
Roorkee, India National Weather Service
North Central River Forecast Center
Jacob Petersen-Perlman Chanhassen, Minnesota
Oregon State University
Corvallis, Oregon Mehdi Rezaeianzadeh
School of Forestry and Wildlife Sciences
Christa D. Peters-Lidard Auburn University
Hydrological Sciences Laboratory Auburn, Alabama
Niels Riegels
Greenbelt, Maryland
DHI
Hai V. Pham Hørsholm, Denmark
Postdoctoral Fellow, Department of Civil and Environmental Engineering
Matthew Rodell
Louisiana State University
Laboratory Chief, Hydrological Sciences Laboratory
Baton Rouge, Louisiana
Mikołaj Piniewski Greenbelt, Maryland
Warsaw University of Life Sciences Renata del Giudice Rodriguez
Warsaw, Poland CAPES Foundation
Paulo Rógenes Monteiro Pontes Ministry of Education of Brazil, Brasilia,
Instituto de Pesquisas Hidráulicas Distrito Federal, Brazil
Universidade Federal do Rio Grande do Sul A. Charles Rowney
IPH-UFRGS Manager, ACR, llc Longwood, Florida
Porto Alegre, Brazil
Albert. I. Rugumayo
Erik C. Porse Faculty of Engineering
UCLA Institute of the Environment and Sustainability Ndejje University
Los Angeles, California Kampala, Uganda
College of Engineering, Design Art and Technology
Nawa R. Pradhan
Makerere University
Research Civil Engineer, Coastal and Hydraulics Laboratory
Kampala, Uganda
U.S. Army Engineer Research and Development Center
Vicksburg, Mississippi Udisha Saklani
Former Research Assistant at the Institute of Water Policy,
James Prairie Lee Kuan Yew School of Public Policy,
Hydrologic Engineer, Upper Colorado Region
National University of Singapore,
Bureau of Reclamation
Singapore
Boulder, Colorado
Jose D. Salas
K. S. H. Prasad Professor Emeritus, Department of Civil and Environmental Engineering
Department of Civil Engineering Colorado State University
Indian Institute of Technology Roorkee Fort Collins, Colorado
Roorkee, India
Samuel Sandoval-Solis
Roland K. Price University of California Davis
Emeritus Professor, UNESCO-IHE, Institute for Water Education Department of LAWR
Delft, The Netherlands Davis, California
Fernando Falco Pruski Sankar Sarkar
Agricultural Engineering Department Physics and Applied Mathematics Unit
Federal University of Viçosa Indian Statistical Institute
Viçosa, Brazil Kolkata, India

l Contributors
Arup K. Sarma Jery R. Stedinger

Professor, Department of Civil Engineering Dwight C. Baum Professor of Engineering
Indian Institute of Technology Guwahati School of Civil and Environmental Engineering Cornell
Guwahati, India University Ithaca
Ithaca, New York
Shinji Sato
Department of Civil Engineering R. Subbaiah
University of Tokyo Department of Soil and Water Engineering
Tokyo, Japan College of Agricultural Engineering and Technology
Junagadh Agricultural University
John Schaake
Junagadh, India
Annapolis, Maryland
William A. Scharffenberg Caroline A. Sullivan

CEIWR-HEC School of Environment, Science and Engineering
Hydrologic Engineering Center Lismore Campus Marine Ecology Research Centre
Davis, California Southern Cross University
New South Wales, Australia
Anton Schleiss
Laboratoire de constructions hydrauliques (LCH) Ge Sun
Ecolepolytechniquefédérale de Lausanne (EPFL) Research Hydrologist, Eastern Forest Environmental Threat Assessment Center
LCH - ENAC - EPFL Southern Research Station, USDA Forest Service
Lausanne, Switzerland Raleigh, North Carolina
Sergio E. Serrano Liqun Sun

Department of Civil and Environmental Engineering Department of Civil Engineering
Temple University The University of Hong Kong
Philadelphia, Pennsylvania Hong Kong, China
Yury V. Shan’ko Óli G. B. Sveinsson

Scientific Researcher, Institute of Computational Modeling SB RAS Landsvirkjun (The National Power Company of Iceland)
Krasnoyarsk, Russia Reykjavík, Iceland
Ashish Sharma David G. Tarboton

School of Civil and Environmental Engineering Department of Civil and Environmental Engineering
The University of New South Wales Utah State University
Sydney, Australia Logan, Utah
Jack Sharp Gokmen Tayfur
University of Texas Department of Civil Engineering
Austin, Texas Izmir Institute of Technology
Mohsen Sherif Izmir, Turkey
Civil and Environmental Engineering Department
Charles J. Taylor
College of Engineering
Water Resources Section
UAE University
Kentucky Geological Survey
United Arab Emirates
University of Kentucky
Haiyun Shi Lexington, Kentucky
State Key Laboratory of Hydroscience and Engineering
Ramesh S. V. Teegavarapu
Tsinghua University
Associate Professor, Department of Civil, Environmental and
Beijing, China
Geomatics Engineering
Joel S. Sholtes Florida Atlantic University
Department of Civil and Environmental Engineering Boca Raton, Florida
Fort Collins, Colorado A. K. Thakur
Darbhanga College of Engineering
Lucas Siegfried Darbhanga, India
Ahtna Engineering Services, LLC
Marina, California Reed Thayer
Vijay P. Singh University of California
Department of Biological and Agricultural Engineering & Davis, California
Zachry Department of Civil Engineering
Texas A&M University Mark Thyer
College Station, Texas School of Engineering
University of Newcastle
Bellie Sivakumar Callaghan, Australia
School of Civil and Environmental Engineering
The University of New South Wales Ezio Todini
Sydney, Australia Retired Professor, Dipartimento di Scienze Biologiche, Geologiche e Ambientali
Department of Land, Air and Water Resources University of Bologna
University of California Bologna, Italy
Davis, California
Glenn Tootle
Brian E. Skahill Associate Professor, The University of Alabama
Research Civil Engineer, Coastal and Hydraulics Laboratory Department of Civil, Construction and Environmental Engineering (CCEE)
U.S. Army Engineer Research and Development Center Tuscaloosa, Alabama
Cecilia Tortajada
Lambert K. Smedema Institute of Water Policy
Independent Irrigation/Drainage Consultant Lee Kuan Yew School of Public Policy
Arnhem, The Netherlands National University of Singapore
Singapore
Soroosh Sorooshian
Center for Hydrometeorology and Remote Sensing Shivam Tripathi
Department of Civil and Environmental Engineering Department of Civil Engineering
University of California Irvine Indian Institute of Technology Kanpur
Irvine, California Kanpur, India

Contributors li
Frank T.-C. Tsai Ralph A. Wurbs

Professor, Department of Civil and Environmental Engineering Zachry Department of Civil Engineering
Louisiana State University Texas A&M University
Baton Rouge, Los Angeles College Station, Texas
Raphael M. Tshimanga Yi Xiao

Faculty of Agronomic Sciences, Department of Natural Resources Management & Department of Systems Design Engineering
CB-HYDRONET University of Waterloo
University of Kinshasa Waterloo, Canada
Kinshasa, Democratic Republic of the Congo
Dawen Yang
Yeou-Koung Tung Department of Hydraulic Engineering
Chair Professor, Disaster Prevention & Water Environment Research Center Tsinghua University Beijing
National Chiao Tung University Beijing, China
Hsinchu, Taiwan
Daqing Yang
Kamshat Tussupova National Hydrology Research Center
Department of Water Resources Engineering Environment Canada
Lund University Saskatoon, Canada
Lund, Sweden
Xiaoli Yang
Julie A. Vano
Associate Professor, College of Hydrology and Water Resources
Research Applications Laboratory
Hohai University
National Center for Atmospheric Research Boulder, Colorado
Nanjing, China
Jennifer C. Veilleux
Sooyeon Yi
Oregon State University
Corvallis, Oregon
Davis, California
Daniele Veneziano
Department of Civil and Environmental Engineering Peter C. Young
Massachusetts Institute of Technology Professor Emeritus, Systems and Control Group, Lancaster Environment Centre & Integrated
Cambridge, Massachusetts Catchment Assessment and Management Centre
Fenner School of Environment & Society
R. Vignesh Lancaster University, UK & Australian National University, Canberra, ACT
Department of Civil Engineering Australia
Indian Institute of Technology Bombay
Mumbai, India S. Yu
Witte Moscow University
Gabriele Villarini Moscow, Russia
IIHR-Hydroscience and Engineering
The University of Iowa Fei Yuan
Civil and Environmental Engineering Associate Professor, State Key Laboratory of Hydrology
Iowa City, Iowa Water Resources and Hydraulic Engineering
Hohai University
Richard M. Vogel Nanjing, China
Tufts University Edith Zagona
Medford, Massachusetts Research Professor, Department of Civil, Environmental and Architectural Engineering
University of Colorado
Haruko M. Wainwright
Boulder, Colorado
Lawrence Berkeley National Laboratory Kaveh Zamani
Berkeley, California Department of Mechanical and Aerospace Engineering
Glenn Warner
San Diego, California
Department of Natural Resources and the Environment
The University of Connecticut Ilya Zaslavsky
Storrs, Connecticut Spatial Information Systems Laboratory
Wallace A. Wilson San Diego Supercomputer Center
Advisor, ASFPM Foundation University of California, San Diego
Association of State Floodplain Managers, Inc. La Jolla, California
Madison, Wisconsin
Qiang Zhang
Fitsum M. Woldemeskel Department of Water Resources and Environment
School of Civil and Environmental Engineering Sun Yat-sen University
The University of New South Wales Guangzhou, China
Sydney, Australia
Lan Zhang
Aaron T. Wolf Assistant Professor, Department of Civil Engineering
Oregon State University University of Akron
Corvallis, Oregon Akron, Ohio
Ming-ko Woo Igor S. Zonn

School of Geography and Earth Sciences Director, Engineering Research Production Center for Water Management,
McMaster University Land Reclamation and Ecology “Soyuzvodproject”
Hamilton, Canada Moscow, Russia

International Advisory Board
Chair: Richard H. McCuen
The Ben Dyer Professor of Civil & Environmental
Engineering
Department of Civil & Environmental Engineering
University of Maryland
Maryland
Andras Bardossy P. E. O’Connell

Lehrstuhl für Hydrologie und Geohydrologie School of Civil Engineering and Geosciences
Institute für Wasser- und Umweltsystemmodellierung Newcastle University
Universität Stuttgart New Castle upon Tyne, United Kingdom
Stuttgart, Germany
C. S. P. Ojha
Ronny Berndtsson Department of Civil Engineering
Department of Water Resources Engineering & Center Indian Institute of Technology Roorkee
for Middle Eastern Studies Roorkee, India
Lund University
Lund, Sweden Umed S. Panu
Corrado Corradini Department of Civil Engineering
Department of Civil and Environmental Engineering Lakehead University
University of Perugia Thunder Bay, Canada
Perugia, Italy
Fernando Falco Pruski
James F. Cruise Agricultural Engineering Department
Department of Civil and Environmental Engineering Federal University of Viçosa
University of Alabama Viçosa, Brazil
Huntsville, Alabama
João Pedroso de Lima Liliang Ren

Department of Civil Engineering Professor, State Key Laboratory of Hydrology, Water Resources and
University of Coimbra Hydraulic Engineering
Coimbra, Portugal Hohai University
Nanjing, China
M. Fiorentino
Department of Environmental Engineering and Physics Dan Rosbjerg
University of Basilicata Department of Environmental Engineering
Potenza, Italy Technical University of Denmark
R. S. Govindaraju Kongens Lyngby, Denmark
Delon and Elizabeth Hampton Hall of Civil Engineering
Purdue University Jose D. Salas
West Lafayette, Indiana Professor Emeritus, Department of Civil and Environmental Engineering
Nilgun B. Harmancioglu Fort Collins, Colorado
Dokuz Eylul University
Water Resources Management Sergio E. Serrano
Research & Application Center (SUMER) Department of Civil & Environmental Engineering
Izmir, Turkey Temple University
Sharad K. Jain Philadelphia, Pennsylvania
Water Resources Systems Division
National Institute of Hydrology Ashish Sharma
Roorkee, India School of Civil and Environmental Engineering
The University of New South Wales
M. Levent Kavvas Sydney, Australia
Department of Civil and Environmental Engineering Witold Strupczewski
University of California Water Resources Department
Davis, California Institute of Geophysics
Lev Kuchment Polish Academy of Sciences
Laboratory of Hydrological Cycle Warsaw, Poland
Water Problems Institute of the Russian Academy of Sciences,
Ezio Todini
Moscow, Russia
Retired Professor, Dipartimento di Scienze Biologiche, Geologiche e Ambientali
Venkat Laxmi University of Bologna
Department of Earth and Ocean Sciences Bologna, Italy
University of South Carolina
Columbia, South Carolina Dawen Yang
Department of Hydraulic Engineering
Upmanu Lall Tsinghua University
Department of Earth and Environmental Engineering and Beijing, China
Department of Civil Engineering and Engineering Mechanics
Columbia Water Center Qiang Zhang
International Research Institute for Climate and Society Department of Water Resources and Environment
Columbia University Sun Yat-sen University
New York, New York Guangzhou, China
liii

Practitioner Advisory Board
Chair: R. D. Singh
National Institute of Hydrology
Roorkee, India
Aaron R. Byrd Mohamed M. Hantush

Research Hydraulic Engineer & Branch Chief Research Hydrologist
Hydrologic Systems Branch National Risk Management Research Laboratory
Coastal and Hydraulics Laboratory U.S. Environmental Protection Agency
Engineer Research Development Center Cincinnati, Ohio
U.S. Army Corps of Engineers
Vicksburg, Mississippi T. Moramarco
National Research Council, Institute for Hydrogeological Protection,
Anthony S. Donigian, Jr. Perugia, Italy
Principal Consultant
AQUA TERRA Consultants John W. Mueller
Mountain View, California State Conservation Engineer
USDA-NRCS
Donald K. Frevert Temple, Texas
Hydraulic Engineer and Team Leader, Retired
Technical Services Center Biswajit Mukhopadhyay
Bureau of Reclamation National Practice Leader | Water Resources
Lakewood, Colorado North American Infrastructure
Jacobs Engineering Group, Inc.
Randall (Randy) W. Gentry Dallas, Texas
Argonne National Laboratory
Environmental Science Division Dallas, Texas J. Obeysekara
Argonne, Illinois Hydrologic & Environmental Systems Modeling
South Florida Water Management District
lv

PART 1
FUNDAMENTALS
01_Singh_ch01_p1.1-1.10.indd 1 8/22/16 10:25 AM

01_Singh_ch01_p1.1-1.10.indd 2 8/22/16 10:25 AM
Chapter 1
The Hydrologic Cycle
BY
VIJAY P. SINGH
ABSTRACT of hydronium ions that are highly mobile and can wreak havoc. A positively
charged water molecule is a hydronium ion.
Hydrology can be called as the study of the hydrologic cycle. There are sev-
If we look at the structure of the water molecule, then it is found that the
eral biogeochemical cycles in nature that are interactive, but the dominant
hydrogen atoms are tied to the oxygen atom in a V-shaped arrangement at an
among these cycles is probably the hydrologic cycle. This chapter discusses
angle of 105° resulting in an asymmetrical configuration. This also means that
the hydrologic cycle and its association with carbon and nitrogen cycles. It
the shared electrons are closer to oxygen than to hydrogen and the charges are
also discusses the impact of climate change on the hydrologic cycle. Before
not evenly divided. This phenomenon is characterized as polarity. Thus, there
discussing the hydrologic cycle, it will be useful to discuss the characteristics
are two sides of the configuration, one side on which hydrogen atoms are
of water at a more fundamental level that impact the behavior of water at
located and the other opposite side. The hydrogen side tends to be electro-
large scales.
positive and the opposite side electronegative. When water changes from liq-
uid to solid or vapor or vice versa, the water molecules arrange themselves in
1.1 CHARACTERISTICS OF WATER
distinctly different patterns during the phase change. When water becomes ice,
Water is the most common substance on earth and covers more than 70% of the molecular arrangement of ice results in an increase in volume and a
its surface. It is vital for life on earth and for that matter anywhere in the decrease in density. The pattern of molecules is highly organized and is in a
universe. Water is a simple compound where each water molecule is com- rigid geometric form. When water freezes to ice, there occurs an expansion of
posed of two hydrogen (H) atoms and one oxygen (O) atom. This atomic molecules, causing ice to float on water. In the case of liquid water, water mol-
structure causes the water molecules to have unique electrochemical proper- ecules organize themselves in small groups of joint particles. It is this property
ties. Each hydrogen atom shares its single electron with the oxygen atom, and that permits water to move and flow. The molecules of water vapor tend not to
thus these atoms are bonded together as H-O-H. If we consider two water form bonds among themselves and are in a state of high energy, causing mol-
molecules, then it is seen that there occurs an attraction between a hydrogen ecules to be moving.
atom of one water molecule and the oxygen atom of another molecule. The The polarity characteristic is the cause for the attraction of water molecules
bonding between these molecules has low energy and is often referred to as to electrostatically charged ions and colloidal surfaces. This is how hydration
hydrogen bonding (Brady and Weil, 2008). This bonding is responsible for occurs. For example, cations, such as Na+, K+, and Ca2+, get hydrated because
polymerization of water as well as water’s high boiling point, specific heat, and of their attraction to the negative or oxygen end of water molecules. Likewise,
viscosity in comparison with other hydrogen-containing compounds. It is the negatively charged clay surfaces attract water through the positive or hydro-
structure of water molecule that determines its ability to influence soil pro- gen end of molecules. Salt gets dissolved in water because its ionic compo-
cesses as well as its existence as a liquid at temperatures observed on Earth, nents have more attraction for water molecules than for each other.
and makes it a powerful solvent. The bulk water, what is usually discussed in Water is found to have three phases: solid (as snow and ice), liquid, and
hydrology, is the collection of water molecules. The arrangement of these vapor. It is found in the liquid phase and the solid phase on the land surface
molecules, that constitutes the bulk water, is still being debated. and beneath, in the liquid phase only in the oceans and seas, and in the vapor
The retention and movement of water in soils are partly caused by two phase in the atmosphere. Recently, another phase, which is beyond these
forces: cohesion and adhesion. Cohesion is the attraction of water molecules three phases, has been discovered (Pollack, 2013). This is the fourth phase
for each other, and adhesion is the attraction of water molecules for solid and is referred to the “exclusion zone” (EZ).
surfaces. In soil water, adhesion is also called adsorption. In the soil, the water The EZ forms next to the many submerged materials and is unexpectedly a
molecules are attached to solid surfaces by virtue of adhesion. These water large zone of water. It derives its name from the fact that it excludes practically
molecules are attached to other water molecules by virtue of cohesion, and everything. It contains a lot of charge and its characteristics are different from
these are attached to other molecules farther away from the solids, and so on. those of the bulk water. There are many water-related mysteries or phenomena
Thus, through forces of cohesion and adhesion, the water is retained and it in nature that we observe but we do not understand them well, because they
moves through the soil. The forces of cohesion and adhesion are also respon- comprise the EZ that we do not quite understand. In his masterpiece book,
sible for another property of water, called surface tension, commonly Pollack (2013) details such phenomena and the social behavior of water.
observed at the liquid-air interfaces. Surface tension together with adhesion Examples of such mysteries include water serving as a glue when building sand
causes capillarity (the ability of a liquid to flow in narrow spaces without the castles from wet sand, tsunami waves traveling very long distances before
assistance of, and in opposition to, external forces like gravity). petering out, slipperiness of ice, swelling upon bruising or breakage, freezing
It is known that electrons and protons are the elementary units of charge warm water, flow of water upward from plant roots through narrow columns,
and they have opposite charges, that is, one has a positive charge, while the cracking of concrete by upwelling tree roots, spreading of droplets on surfaces,
other has a negative charge. Electrons and protons play a central role in deter- walking on water, formation of isolated clouds, floating of ice, consistency of
mining how water behaves. Since the oxygen atom has a charge of –2 and each yogurt, migration of microspheres away from the center in a beaker of water,
hydrogen atom has a plus charge, the water molecule is reduced to have no the bridge made up of water, and floating water droplets, among others. These
charge or is neutral, since the negative and positive units of charge cancel each phenomena entail crowds of water molecules, not water at the molecular level.
other. When protons latch on to water molecules, the result is the formation Nevertheless, they defy easy explanation and show that we know little about
1-3
01_Singh_ch01_p1.1-1.10.indd 3 8/22/16 10:25 AM

1-4 The Hydrologic Cycle
the interaction of water molecules with other water molecules or water’s social 1.3 HYDROLOGIC CYCLE
behavior. Surprisingly, we understand little about Earth’s most common sub-
There are two systems through and between which the water moves: Earth
stance—water. Water has some unique properties that no other liquid on Earth
and atmosphere. Earth can be divided into oceans and seas, and the land part.
seems to enjoy. It has high specific heat, conducts heat more easily, has a neural
Oceans and seas form what is called the hydrosphere. The land part, also
pH, is a universal solvent, has high surface tension, exists in all three phases,
called land sphere, is divided into the land surface part and subsurface part
occupies larger volume upon freezing, and has maximum density at 4°C.
that includes unsaturated (soils) and saturated parts (geologic formations).
The land surface also includes water bodies, such as lakes and reservoirs,
1.2 DEFINITION OF HYDROLOGY
wetlands, canals and rivers, channels, and lagoons. The unsaturated part is
referred to as pedosphere and the saturated part as lithosphere. The charac-
Hydrology deals with bulk water primarily in liquid, gaseous, and solid teristics of water movement, distribution, occurrence, and storage are differ-
phases. This means that hydrology incorporates the study of snow and gla- ent in these spheres, and to understand these characteristics, this partitioning
ciers. It deals with both quality and quantity that vary in space and time, so is essential. The interconnectivity of these spheres is essential for water mov-
hydrology deals with how water is distributed. Water occurs over the land ing through the hydrologic cycle.
surface, in soil, and in geological formations. Water is dynamic, so hydrology On the continents, the water evaporates from the hydrosphere, pedo-
deals with the movement of water (surface and subsurface). Water can be sphere, and lithosphere, and is transpired by vegetation and plants into the
stored in lakes and reservoirs as well as in subterranean environment, so atmosphere. The water vapor moves in the atmosphere, and under suitable
hydrology deals with the storage of water. The occurrence, distribution, conditions, it condenses and precipitates over the land surface and over
movement, and storage of water quality and quantity are not always entirely oceans and seas. The liquid part of the precipitated water (or melted snow
deterministic, so hydrology encompasses their statistical characteristics. and ice) runs off in part over the land surface and in part into the pedosphere
Thus, hydrology can be defined as the science that deals with the space, time from where water either transpires/evaporates back to the atmosphere or
and frequency characteristics of the occurrence, distribution, movement, and percolates to groundwater (or lithosphere). Evaporation takes place from
storage of the quantity and quality of the waters of the Earth. open surfaces (rivers, lakes, wetlands, etc.) as well. The water that is not
Frequently, hydrology is partitioned into surface-water hydrology, vadose reevaporated back to the atmosphere continues to the oceans, where a net
zone hydrology, groundwater hydrology, and snow and glacier hydrology. In evaporation takes place. Conversely, a net precipitation occurs over the con-
all of these partitions, both water quantity and quality are dealt with. To tinents. This endless circulation of water from the hydrosphere and land
emphasize quality aspects, water quality hydrology is sometimes considered a sphere to the atmosphere and back is called the hydrologic cycle. It has no
separate branch. Likewise, other partitions of hydrology are agricultural beginning and no end.
hydrology, forest hydrology, wetland hydrology, environmental hydrology, The hydrologic cycle can be viewed as a natural machine, a constantly run-
ecosystems hydrology, atmospheric hydrology, arid lands hydrology, coastal ning, distillation and pumping system. The primary source of energy for the
hydrology, urban hydrology, and geohydrology. These partitions point to the operation of this machine is the Sun that supplies heat energy. In Fig. 1.1, the
broad scope and the interdisciplinary nature of hydrology. Also, these branch- spatial and temporal averages of these energy components are illustrated. Of
es are not entirely insulated from each other. Hydrology should be viewed in particular importance for hydrology is the loss of heat energy from the Earth’s
a broader context, for water is sometimes on the surface and at other times, it surface. The 29% share of heat loss can be further subdivided into 22% loss as
is below the surface but then again it reappears on the surface. latent heat and 7% loss as sensible heat. It should be emphasized that there are
Depending upon the methods of analysis and synthesis, hydrology is also huge variations from the average picture both in time (diurnal, seasonal, and
divided into different branches, such as mathematical hydrology, numerical yearly) and in space. Together with the force of gravity, this energy keeps the
hydrology, digital hydrology, systems hydrology, parametric hydrology, water moving as evaporation and transpiration from the earth to the atmo-
empirical hydrology, statistical hydrology, and stochastic hydrology. Based on sphere, as condensation and precipitation from the atmosphere to the earth,
the emphasis of scientific concepts, hydrology can be divided into physical and as streamflow and groundwater from the Earth to the oceans. Thus, the
hydrology, chemical hydrology, and biological hydrology. hydrologic cycle encompasses three major systems with the hydrosphere as
Space Shortwave solar radiation Longwave radiation and heat transfer
30% reflected
and scattered 70% radiated
(342 W/m2)
100%
26% reflected
and scattered 65% radiated
109% absorbed
96% radiated
26% absorbed 5% lost back down
Earth’s
to space
atmosphere
29% lost
as latent
4% reflected Greenhouse and
effect sensible
heat
Back
radiation
Earth’s
surface 47% absorbed
Figure 1.1 The global energy balance.
01_Singh_ch01_p1.1-1.10.indd 4 8/22/16 10:25 AM

Schematic Representation of the Hydrologic Cycle 1-5
Storage in Moisture over land

ice and snow
Condensation
Precipitation
on land
Surface Evaporation from land
runoff Precipitation
Evapotranspiration on ocean
Freshwater
storage Evaporation
Lake Sub
Infiltration sur
fac
e flow Evaporation from ocean
Lake
Surface outflow
Groundwater flow
Groundwater outflow
Ocean
Groundwater
storage
Ocean
storage
Figure 1.2 The hydrologic cycle [Source: http://water.usgs.gov].
the major source of water, the atmosphere as the deliverer of water, and the land surface, replenishing soil moisture, and percolating down to recharge
land sphere as user of water. groundwater. Part of the infiltrated water includes subsurface runoff or inter-
It may be noted that water is transported, temporarily stored, and may flow, and groundwater runoff or baseflow. Part of rainfall and snowmelt runs
change state in each sphere. Consider, for example, the atmosphere where off over the land surface, joins streamflow, and eventually joins the sea.
water occurs as vapor flow, stored as vapor storage in the atmosphere, and Streamflow includes surface runoff as well as subsurface runoff. It may be
condenses and precipitates under change of state from vapor to either that of noted that these different components have different time scales and their
liquid or solid. Much more water stays in storage for longer periods of time significance is different at different scales.
than in movement through the hydrologic cycle.
1.5 SCHEMATIC REPRESENTATION OF THE
1.4 COMPONENTS OF THE HYDROLOGIC CYCLE HYDROLOGIC CYCLE
The major components of the hydrologic cycle are precipitation, evaporation, The hydrologic cycle can be illustrated as shown in Fig. 1.2, depicting the
infiltration, groundwater, and streamflow. Precipitation includes all forms of various components and their principal directions of flow. This, however, does
water that falls from the atmosphere, including rainfall, snowfall, hail, sleet, not provide any information on the relative significance of these components
drizzle, dew, and fog. Rainfall and snowfall constitute the main forms of pre- which varies with the space (watershed, land system, and global) and time
cipitation. Snow is a form of stored water that remains where it falls until scales. The cycle is truly an endless circulation at the global scale, but not
melting occurs. Each year, approximately 505,000 km3 (121,000 mi3) of water necessarily so at smaller scales. At the global scale, the atmosphere, the hydro-
falls as precipitation of which about 78% [approximately 398,000 km3 (95,000 sphere, and the land sphere are connected and interactive, as shown in Fig. 1.3.
mi3)] falls over the oceans. The precipitation falling as rain is about 107,000 The major components of the cycle at this spatial scale are precipitation,
km3 (26,000 mi3) and as snow is about 1000 km3 (240 mi3). Precipitation is evaporation and transpiration, and streamflow. The time scale is also much
partitioned into four parts: (1) interception, (2) evaporation, (3) infiltration, larger and hence more transient components, such as infiltration, surface
and (3) runoff. Part of precipitation is intercepted by trees, buildings, and other runoff, interception, etc., become embedded in other components. The global
abstract objects. Most of this portion is evaporated to the atmosphere. A small scale hydrologic cycle is a closed system and it is normally neither modeled
portion of snow also evaporates. Oceans contribute about 86% of global nor used for water resources planning and management at this scale. The
evaporation that reduces oceanic temperature through evaporative cooling. general circulation models are, however, employed for climate impact studies.
Part of rainfall fills surficial depressions, forming small ponds, where some Of greater significance is the hydrologic cycle at the watershed scale which
water infiltrates and some evaporates. Another part of rainfall infiltrates the can be quite small or very large. At this spatial scale, all components of the
Precipitation
Precipitation River flow

Atmospheric
system Earth system Tidal flow, sea Ocean system
ET
rise, GW
Evaporation
Figure 1.3 A global schematic of the hydrologic cycle [Source: Singh, 1992].
01_Singh_ch01_p1.1-1.10.indd 5 8/22/16 10:25 AM

Precipitation Global
Global Global
ET weather CO2
systems variations
10,000 Km
Development
Land system Surface of major
runoff river rasins
Soil
1000 Km formation
Runoff
Exfiltration Infiltration cycle
Space
Mesoscale Drainage
100 Km weather Mesoscale network
systems soil Soil formation
Subsurface (floods) moisture erosion
Subsurface system variation Shallow
runoff 10 Km
ground
Stream Nutrient water
flow cycles circulation
Upward
Percolation 1 Km Thunder
moisture
(recharge) storm
movement
Time Log/
Local
Ground water 2 4 6 8 10 12 14 16 second
Aquifer system runoff
Second Minute Day Year Century One million One billion
years years
Figure 1.4 A schematic of the hydrologic cycle of the earth system [Source: Singh, 1992].
Figure 1.5 Spatial and temporal scales of hydrologic processes and cycle
[Source: National Researach Council, 1991].
hydrologic cycle play their role, but the significance of their role depends on
the time scale. For example, at small time scales, evapotranspiration may not
be significant, but infiltration plays a dominant role. The reverse happens at The spatial and temporal scales of some hydrologic processes and cycles are
the large time scale. If the watershed is forested or highly urbanized, then shown in Fig. 1.5.
interception may be significant. At a large watershed scale, the hydrologic
cycle is illustrated in Fig. 1.4. 1.6.2 Spatial Scales
Spatial scales that are important for hydrologic investigations are hillslope,
catchment, basin, regional, continental, and global. At the global scale, the
1.6 SCALES IN HYDROLOGIC CYCLE
atmosphere, the hydrosphere, and the land and lithosphere are strongly con-
The significance of the components of the hydrologic cycle is closely connected through the processes of precipitation, evaporation, and runoff. Global
nected with their space and time scales. Further, mathematical relationships fluxes and circulation patterns are investigated at this scale and are important
describing the different components depend on the scale. Therefore, these for assessing the impact of climate change on the hydrologic cycle. Figure 1.2
scales play a critical role when simulating the cycle or the components shows a global view of the hydrologic cycle.
thereof. The catchment scale has a very broad range, ranging from a few square
kilometers to tens of thousands of kilometers. For small catchments, the pro-
1.6.1 Time Scales cesses of precipitation, runoff, infiltration, soil moisture, and runoff are most
The components of the hydrologic cycle have vastly different time scales of important. As catchment becomes large, groundwater in most cases also plays
movement and residence times. The residence time of a reservoir is the an important role.
average time the water will spend in the reservoir (i.e., the storage divided
by the flux through the reservoir), and thus is also a measure of the age of
1.7 IMPACT OF CLIMATE CHANGE ON THE
water of that reservoir. For example, the rate at which surface runoff moves HYDROLOGIC CYCLE
is very high as compared with that of groundwater. Likewise, the residence
time of surface water is a small fraction of that of groundwater. Groundwater It is now widely accepted that as a result of anthropogenic influences, there is
may reside for over 10,000 years below the earth’s surface before leaving, global warming and the climate is changing. The increased emission of green-
whereas water spends about 2–6 months in rivers, about 1–2 months in the house gases is considered to be the principal cause of increase in Earth’s tem-
soil, and only a little over a week in the atmosphere. These times play a perature. This is supported by measured temperature data series as well as
critical role in numerical simulation of the hydrologic cycle. The time step paleoclimatological records. The implications of climatic change for the hydro-
should be sufficiently small so that the variations in the component pro- logic cycle are many. The increase in temperature would cause higher evapo-
cesses are captured in sufficient detail without putting undue burden on transpiration; changes in patterns, timing, intensity, and distribution of
data collection and computational efforts. Shiklomanov (1999) has pre- precipitation; melting of polar ice caps and recession of glaciers; more fre-
sented the estimated periods of renewal of water resources on the earth, as quently occurring droughts; increased frequency of flooding; higher melting
shown in Table 1.1. of polar ice and glaciers; sea water level rise; inundation of islands of low eleva-
tions as well as cities adjacent to seas; changes in vegetation dynamics; ecosys-
tem health; and so on. These changes are already being witnessed. Nevertheless,
Table 1.1 Periods of Water Resources Renewal on the Earth it is not understood with acceptable degree of certainty as to what, where, and
how much changes are or will occur. On the other hand, an increase in tem-
Water of hydrosphere and land sphere Period of renewal
perature may mean greater precipitation, some of which may occur in the form
World Ocean 2,500 years of snow at the poles, leading to an additional accumulation of ice.
Groundwater 1,400 years
Since the three spheres constitute a continuum, changes in the land sphere
and lithosphere would cause a change in the climate. For example, changes in
Polar ice 9,700 years vegetation mean changes in evapotranspiration, soil moisture, albedo and
Mountain glaciers 1,600 years radiation balance; burning of fossil fuels; increased use of water day-to-day
Ground ice of the permafrost zone 10,000 years
needs; increased irrigation; increased industrial activities; and large-scale
water transfer between basins would contribute to climate change. Climate
Lakes 17 years change manifests itself in changed patterns of spatial and temporal variability
Bogs 5 years in the components of the hydrologic cycle.
Soil moisture 1 year
Channel network 16 days 1.8 INFLUENCE OF HUMAN ACTIVITIES AND LAND USE
CHANGES ON HYDROLOGIC CYCLE
Atmospheric moisture 8 days
There are a multitude of changes, minor or major, caused in watersheds by
Biological water Several hours
human activities. These changes influence virtually all components of the
[Source: Shiklomanov, 1999] hydrologic cycle. The watershed changes can be either point or nonpoint
01_Singh_ch01_p1.1-1.10.indd 6 8/22/16 10:25 AM

Relation between Hydrologic cycle and Carbon and Nitrogen Cycles 1-7
changes. Examples of point changes are structural changes, such as dam con- peak of outflow from the reservoir and the volume of flow downstream may
struction, channel improvement, and detention storage, and these changes be considerably reduced if the reservoir water is diverted elsewhere. Channel
affect watershed response in terms of evaporation, seepage, residence time, improvements include straightening of the channel, removal of vegetation,
etc. Examples of nonpoint land use changes, that affect catchment response, lining of the channel, maintenance of bends, or increasing slope. These
include forestry, agriculture, mining, and urbanization. The seriousness of changes may translate into decreased channel roughness increasing flow
hydrologic consequences depends on the magnitude of watershed changes. velocity and hence peak discharge. Depending upon the bed material, infiltra-
There is growing need to quantify the impact of man-made changes on the tion through the bed and banks also modifies flow characteristics.
hydrologic cycle in order to anticipate and minimize the potential environ-
mental detriment and to satisfy water resources requirements of the society. 1.9 RELATION BETWEEN HYDROLOGIC CYCLE AND
Agricultural operations exercise a significant influence on the hydrologic CARBON AND NITROGEN CYCLES
response of watersheds. Agriculture alters land cover, influencing evapotrans-
piration; changes upper soil layers, altering infiltration, soil moisture, leach- The hydrologic cycle is a biogeochemical cycle. The movement of water over
ing and evaporation; irrigating land, causing recharge and changing water and below the Earth is fundamental to the cycling of biogeochemicals. This
table; applying fertilizers, causing soil and water pollution; and tilling the soil, suggests that the hydrologic cycle not only impacts other biogeochemical
causing soil erosion by wind and water. These changes lead to changes in cycles, such as the carbon cycle and the nitrogen cycle, but is also impacted
evapotranspiration, overland flow, channel flow, and infiltration. by them in a feedback system. Consider, for example, runoff and streamflow.
Urbanization transforms forested and agricultural land into urban areas Eroded sediment and phosphorus are carried by surface runoff to streams
where houses, roads, schools, buildings, shopping malls, parks, parking lots, and rivers and further on to water bodies. Phosphorus, applied to agricul-
sewers, etc., are constructed. Urbanized lands are dominated by paved or tural lands through fertilizer in excess of what plants consume, is carried by
impervious surfaces that reduce infiltration and evapotranspiration and surface runoff to lakes and reservoirs, and that causes eutrophication.
increase storm water. The hydrologic effects of urbanization include increases Likewise, nitrogen is transported by surface runoff and groundwater and
in the volume and peak of direct runoff for a given rainfall; reduction in time streamflow to water bodies. As an example, the water has become hypoxic in
of travel resulting in lower lag time and lower time of concentration; diminu- the Gulf of Mexico as a result of nitrates from fertilizer being transported by
tion of baseflow; reduction in infiltration; increase in the withdrawal of runoff from agricultural lands to rivers and then by the Mississippi. Transport
groundwater; increase in pollution of rivers and aquifers; endangering the of eroded sediment and dissolved salts causes salinity of the oceans and seas.
ecology; and increase in temperature of urban areas, causing a change in
1.9.1 Carbon Cycle
microclimate.
The principal forest activities are afforestation (planting trees), deforesta- Carbon is ubiquitous in all living things as well as is part of nonliving things,
tion (cutting trees), and management. Forest management includes road con- such as ocean, air, and earth. Fossil fuels, such as coal and oil, are made of
struction, erosion control and water management, electrification, and chemical carbon. By joining oxygen, it forms carbon dioxide that is present in the
sprays. The immediate effect of these activities is the change in vegetal cover. atmosphere. The movement of carbon back and forth among biosphere,
Deforestation virtually eliminates the interception of precipitation. Forest litter atmosphere, hydrosphere, and geosphere defines the carbon cycle, as shown
removal changes infiltration capacity of the soil and has a pronounced effect in Fig. 1.6. It is one of the biogeochemical cycles and consists of various
on raindrop impact and the resulting soil erosion. With the loss of vegetative sinks and stores of carbon and processes by which sources and sinks
cover, evapotranspiration decreases. These changes increase direct runoff, exchange carbon.
reduce surface roughness, and decrease recharge to groundwater. The carbon cycle is a closed system, suggesting a fixed amount of carbon in
Typical structural changes include a dam, a weir, levees, spurs, dykes, dredg- the world. The processes in the carbon cycle are interactive. Fossil fuel burn-
ing, channel improvement works, etc. A dam reservoir, in general, reduces the ing and land use change are two of the important processes impacting the
Sunlight
Auto and
factory
CO2 cycle emissions
Photosynthesis
Plant
respiration
Animal
respiration
Organic carbon
Root
Decay
respiration
organisms Dead organisms
and waste products
Ocean
Fossils and fossil fuels uptake
Figure 1.6 Carbon cycle [Source: https://eo.ucar.edu].
01_Singh_ch01_p1.1-1.10.indd 7 8/22/16 10:25 AM

Figure 1.7 Nitrogen cycle [Source: http://cnx.org and USGS].
balance of carbon. Oceans contain the greatest amount of actively cycled nitrogen into compounds, such as nitrate or ammonia, which can be used by
carbon in the world. Only the lithosphere stores more carbon than oceans. plants. This “fixed” nitrogen (or reactive nitrogen) is a limiting factor for plant
The geologic component of the carbon cycle operates more slowly than other growth in both managed and wild environments.
parts of the global carbon cycle. Carbon leaves the geosphere in several ways, The nitrogen cycle, as shown in Fig. 1.7, is the process by which nitrogen is
for example, during the metamorphosis of carbonate rocks. transformed to various chemical forms through biological and physical pro-
In the atmosphere, carbon exists primarily in gaseous form as carbon diox- cesses. Important processes in the nitrogen cycle include fixation, ammonifi-
ide (CO2) and methane. Both these gases absorb and retain heat in the atmo- cation, nitrification, and denitrification. The nitrogen cycle constitutes an
sphere, and are partly responsible for the greenhouse effect. In comparison important component of the ecosystem. The nitrogen cycle has hydrologic
with carbon dioxide, methane produces a large greenhouse effect per volume, significance because the availability of nitrogen affects the rate of key ecosys-
but it exists in much lower concentration and is more short-lived. This tem processes, including primary production and decomposition. When
explains why carbon dioxide is a more important greenhouse gas than meth- plant and animal wastes decompose, they add nitrogen to the soil. Bacteria in
ane. Carbon dioxide traps heat in the atmosphere, and were it not for it and the soil convert those forms of nitrogen into forms that plants can use. Plants
other greenhouse gases, the earth would be a frozen planet. Because of so use the nitrogen in the soil to grow. People and animals eat plants; then ani-
much more emission of carbon dioxide to the atmosphere, we are experienc- mal and plant residues return nitrogen to the soil again, completing the cycle.
ing global warming. Human activities, such as fossil fuel combustion, use of artificial nitrogen
Plants employ carbon dioxide and sunlight to make their own food fertilizers, and release of nitrogen in wastewater, significantly impact the
through the process of photosynthesis, and carbon becomes part of plants. On global nitrogen cycle.
burning fossil fuels, most of the carbon enters the atmosphere as carbon Nitrogen enters the nitrogen cycle from different sources: (1) air through
dioxide. Living beings release CO2 back to the atmosphere during respiration. several unique types of microorganisms that can convert N2 gas to inorganic
Large exchange of CO2 also takes place between the hydrosphere and the forms usable by plants; (2) manure and decaying of plant materials; (3) appli-
atmosphere. The dissolved CO2 in the oceans is used by marine biota for cation of commercial nitrogen fertilizers; and (4) as inorganic nitrogen from
photosynthesis, thus carbon dioxide is entering the oceanic biosphere as well. the atmosphere and factories. It leaves the cycle in four ways: (1) denitrification,
Also, from the atmosphere carbon dioxide directly dissolves into water bodies (2) bacteria change nitrate in the soil to atmospheric nitrogen, (3) volatilization,
(oceans, lakes, reservoirs, etc.) as well as dissolves in precipitation as rain- and (4) turning urea fertilizers and manures on the soil surface into gases.
drops. When dissolved in water, it reacts with water molecules and forms Nitrogen reenters the cycle through one of the aforementioned processes or
carbonic acid that contributes to the ocean acidity. Carbon leaves the terres- through other processes. As shown in Fig. 1.7, nitrogen changes from organic
trial biosphere in different ways and at different time scales. For example, matter in the soil, to bacteria, to plants, and back to the organic matter.
combustion releases organic carbon rapidly into the atmosphere. Carbon Nitrogen is present in the environment in a variety of chemical forms:
stored in soil can remain there for up to thousands of years before being organic nitrogen, ammonium (NH4+), nitrite (NO2−), nitrate (NO3−), nitrous
washed into rivers by erosion or released into the atmosphere through soil oxide (N2O), nitric oxide (NO), or inorganic nitrogen gas (N2). Figure 1.7
respiration. shows how these processes fit together to form the nitrogen cycle.
The main component of the nitrogen cycle starts with the element nitrogen
1.9.2 Nitrogen Cycle in the air. Upon interaction with oxygen, several compounds, such as nitric
Nitrogen (N2) is a colorless, odorless, nontoxic gas, and is essential for all oxide (NO) and nitrogen dioxide (NO2) are formed. Eventually nitrogen
known forms of life on Earth. Being close to 78%, it is the largest constituent dioxide may react with water in rain to form nitric acid (HNO3), which may
of the atmosphere, but this form is relatively nonreactive and unusable by be utilized by plants as a nutrient. Nitrogen in the air becomes part of the
plants. Nitrogen in plants is used in chlorophyll molecules, which are essential biological matter through bacterial action and algae through a process called
for photosynthesis and plant growth. Natural fixation converts gaseous nitrogen fixation. Legume plants, such as clover, alfalfa, and soybeans, form
01_Singh_ch01_p1.1-1.10.indd 8 8/22/16 10:25 AM

REFERENCES 1-9
nodules on the roots, where nitrogen-fixing bacteria take nitrogen from the 1.10 CONCLUSION
air and convert it into ammonia (NH3). The main source of nitrogen in soils
The hydrologic cycle is receiving a lot of attention these days from various
is from organic matter, which largely arises from plant and animal residues.
quarters of the society, partly because of climate change; frequently occurring
The conversion of ammonia to nitrate is performed primarily by soil-living
natural disasters, such as floods and droughts; increasing demand for water;
bacteria and other nitrifying bacteria. Onsite sewage facilities, such as septic
and the growing recognition of water-energy-food nexus. Even laymen are
tanks, release large amounts of nitrogen into the environment by discharging
now talking about the specter of climate change and the hydrologic cycle.
through drain fields into the ground.
Fortunately, in primary and junior high schools, the hydrologic cycle is being
Ammonia (NH3) is highly toxic to fish, and the level of ammonia discharged
taught and this is causing an interest in water issues and climate change at the
from wastewater-treatment facilities must be closely monitored. Nitrogen and
grass root level. The hydrologic cycle is fundamental to other cycles and it is
hydrogen react under great pressure and temperature in the presence of a
important that the cycle is managed properly. The fate of human civilization
catalyst to make ammonia. The reaction of ammonia and nitric acid produces
and economic prosperity will depend on how well water is managed. Ancient
ammonium nitrate, which may be used as a fertilizer. Animal wastes when
civilizations understood this dependence well. It is now time to learn from
decomposed also return to the Earth as nitrates. Nitrate is a concern for water
them and begin a concerted effort to respect nature and its fundamental ele-
quality when it is converted to the nitrate (NO3–) form because nitrate is very
ments, especially air, water, and soil.
mobile and easily moves with water in the soil. High levels of nitrates can be
toxic to newborns, causing anoxia or internal suffocation.
REFERENCES
Denitrification is the reduction of nitrates back into the largely inert N2 gas,
completing the nitrogen cycle. This process is performed by some bacterial Brady, N. C. and R. R. Weil, The Nature and Properties of Soils, Pearson/
species in anaerobic conditions. Nitrous oxide gas (N2O) is a side product of Prentice Hall, Columbus, OH, 2008.
this reaction. It is also produced as a result of agricultural fertilization, bio- PhysicalGeography.net, Introduction to the Hydrosphere, http:www.
mass burning, cattle and feedlots, and industrial sources. Nitrous oxide is also physicalgeography.net/fundamentals/8n.html. Accessed on April 20, 2015.
a greenhouse gas and is currently the third largest contributor to global National Research Council. Opportunities in the Hydrologic Sciences,
warming, after carbon dioxide and methane. Although not as abundant in the National Academy of Sciences, National Academy Press, Washington, D.C.,
atmosphere as carbon dioxide, for an equivalent mass, it is nearly 300 times 1991.
more potent in its ability to warm the planet. Pollack, G. H., The Fourth Phase of Water: Beyond Solid, Liquid and Vapor,
In the atmosphere, ammonia (NH3) acts as an aerosol, decreases air quality Ebner & Sons Publishers, Seattle, Washington, D.C., 2013, p. 357.
and clings to water droplets, eventually resulting in nitric acid (HNO3) that Shiklomanov, I. A., World Water Resources: Modern Assessment and
produces acid rain. Ammonia and nitrous oxides actively alter atmospheric Outlook for the 21st Century (Prepared in the framework of IHP, UNESCO),
chemistry. They are precursors of tropospheric ozone production, which con- State Hydrology Institute, St. Petersburg, Russia, 1999.
tributes to smog and acid rain, damages plants, and increases nitrogen inputs Singh, V. P., Elementary Hydrology, Prentice Hall, Englewood Cliffs,
to ecosystems. New Jersey, 1992.
01_Singh_ch01_p1.1-1.10.indd 9 8/22/16 10:25 AM

Chapter 2
Watersheds, River Basins,
and Land Use
BY
DAWEN YANG, A.W. JAYAWARDENA, AND ZHENTAO CONG
ABSTRACT however, different terminologies used in different countries and regions that
convey the same meaning. For example, “watershed” and “drainage basin” are
This Chapter describes the components, characteristics, hydrological pro-
used synonymously in North America. In the United Kingdom, and in coun-
cesses including the characteristics that impact them, river basins and their
tries that follow the UK terminology the term “catchment area” is used. A
management, and land use, in the context of the basin scale hydrological cycle
second usage of the term “catchment area” is to mean an area covered/served
which may be considered as at the lower end of the hierarchically nested
by an institution such as a school or a hospital.
hydrological cycle. It further describes the basic components of a watershed
A watershed (or a drainage basin or a catchment area) is the basic unit
and its delineation, topography, geology, soil type, hillslope, stream network,
within which the processes of the basin scale take place and are of interest to
wetlands, artificial canals and waterways including how they impact the
hydrologists and water engineers. It is an open system in contrast to the global
hydrological processes. The Chapter ends with a brief description of river
hydrological cycle which is a closed system. Inputs to the watershed include
basin management including the water policies in a few selected countries/
precipitation (rainfall, snowfall, and other minor forms) and solar energy for
regions, and the effect of land use on the basin scale hydrological cycle includ-
evaporation. Outputs include runoff into the seas or to other water bodies/
ing the link between land use, runoff and water quality.
watersheds, infiltration into the subsoil, which may percolate further into
deeper subsoils and finally to aquifers, evaporation, and transpiration from
2.1 INTRODUCTION vegetation (evapotranspiration) and moisture replenishment. Water storage
in the watershed may include interception storage, depression storage, soil
Hydrological cycle refers to the continuous circulation of earth’s water
moisture storage, water absorbed through roots by osmosis and stored in
resources in solid, liquid, and vapor phases, linking the atmosphere, land, and
vegetation, water stored in natural or artificial water bodies, groundwater
the oceans, and powered by the energy of the sun. It can be viewed as a nested
storage, and channel storage.
system at different scales. At the top of the hierarchy is the global hydrological
Processes of water movement in a watershed include precipitation (from
cycle, in which there is no water added or lost and is a closed system. There
atmosphere to land surface), stemflow (intercepted water flowing along the
are no inputs or outputs. At the next level is the regional or continental scale
stems of vegetation), infiltration (from land surface to subsoil), surface runoff
hydrological cycle, which is an open system, which can be nested into the
(along the land surface), soil moisture movement (in the partially saturated
global hydrological cycle. It may include the atmosphere, land surface, and/or
soil zone), percolation (gravity flow from subsoil to deeper soil layers and
the oceans. At the lower end is the basin scale hydrological cycle, which is a
finally to aquifers), and interflow or throughflow (flow below the surface in
local open system with inputs and outputs and is of primary importance to
the downslope direction).
hydrologists and water professionals. It can be defined as the hydrological
A drainage basin is usually exoreic, meaning that the water finds its way
cycle in a river basin bounded by its watershed and its outlet which may
finally to one of the major oceans. However, there are some endorheic drain-
sometimes be the ocean. This chapter describes the components, characteris-
age basins, particularly in arid areas without any rivers where the water does
tics, hydrological processes including the characteristics that affect them,
not find its way to an ocean and ends up underground, or in inland lakes
river basins and their management, and land use, in the context of the basin
within the basin or returned back to the atmosphere by evaporation.
scale hydrological cycle.
Examples of basins that do not drain into an ocean include those that drain
The name “watershed” has its origin in the early nineteenth century from
into Caspian sea, Aral Sea, and other inland lakes. In the context of ground-
the combination of “water” and “shed” in the sense “ridge of high ground”
water flow, the watershed below the land surface may not always coincide
suggested by German Wasserscheide. The Oxford English Dictionary defines
with that above. There can be watersheds within watersheds in a nested man-
watershed as an area or ridge of land that separates waters flowing to different
ner and arranged hierarchically. All land is part of one watershed or another.
rivers, basins or seas. Merriam Webster Dictionary defines it as a region or
area bounded peripherally by a divide and draining ultimately to a particu-
lar watercourse or a body of water. United States Environmental Protection 2.2 COMPONENTS OF WATERSHEDS
Agency defines it as the area of land where all of the water that is under it or
drains off of it goes into the same place. Collins English Dictionary defines it In general, the components of a watershed include hill slopes, a stream net-
as the dividing line between two adjacent river systems, such as a ridge. work, water bodies, wetlands, swamps, bogs, artificial canals and waterways,
Watershed also has another definition in common usage. Oxford Dictionary hydraulic structures, floodplains, riparian zones, gullies and rills, vegetation,
defines this second usage as an event or period marking a turning point in a soils, etc. Underlying a surface watershed is a subsurface groundwater storage
situation. Vocabulary.com defines it as an event marking a unique or impor- and seepage system, the boundaries of which may not coincide with the sur-
tant historical change of course or one on which important developments face boundaries. The subsurface may include aquifers which may be uncon-
depended. Collins English Dictionary defines it as an important period or fined or confined. Underlying the aquifers, or separating confined and
factor that serves as a dividing line. unconfined aquifers lies the impermeable bedrock forming the base of a
Similar definitions of both usages can be found in other dictionaries too. In watershed. Some of the key components are described in more detail in the
the context of hydrology, it is the first definition that matters. There are, following sections.
2-1
02_Singh_ch02_p2.1-2.8.indd 1 8/22/16 10:38 AM

2-2 Watersheds, River Basins, and Land Use
2.2.1 Hill Slope will have more streams implying a higher risk of flooding in a low bifurcation
The simplest—and somewhat ideal—watershed has the shape of an inverted stream network. Typical values range from 2 to 4.
two-sided roof top. Water falling on the two sides, which may have equal or Another characteristic is the fractal dimension of an individual stream that
unequal areas drain into a stream which connects the two hill slopes. Natural is a measure of its irregularity as seen by its meandering. In the case of a
watersheds can be of different sizes and shapes. Contributions to watershed stream network, the fractal dimension is a measure of the network’s ability to
runoff come mainly from hill slopes. They may be in the form of overland fill space (or a surface area in this case). The more branches and more sinuos-
flow which consists of precipitated water in excess of the infiltration capacity ity the network has, the greater is its ability to fill such space. A network that
which for modeling purposes is assumed to flow down the hill slope in the completely fills a surface like space has a fractal dimension of two, whereas
form of a thin sheet, and throughflow, which for modeling purposes is often normal networks which are not completely spacefilling, would have fractal
assumed to flow beneath the surface of the hill slope in a saturated form. The dimensions less than two, and which vary from place to place over the surface.
latter type of flow exists in watersheds with highly porous surface soil layers The geometrical pattern of a stream network can be viewed as a fractal with
or with subsoils consisting of soil pipes in which the saturated and partially a fractal dimension, D, between 1 and 2. For irregularly shaped surface areas,
saturated flows through the subsoil take place. Hill slope hydrology is an the perimeter (length)–area relationship is (Mandelbrot, 1982)
important area of interest to hydrologists as the first stage of water movement
after precipitation has taken place along hill slopes. In modeling hill slope ( P(ε))1/D α (A(ε))1/2 (2.1)
hydrology, the kinematic wave approximation to the St. Venant’s equations is
where P(ε) is the perimeter of the irregular shape measured with a measuring
normally made since the slope terms (physical slope as well as the friction
stick of length ε, A(ε) is the area measured in units of ε2, and α is a symbol
slope) in the momentum equation are at least an order of magnitude larger
to denote the proportionality relationship. This relationship can be extended
than the other terms. The kinematic wave approximation has been first
to stream networks too. Hack (1957) has obtained the following relationship
applied to flood flows (Lighthill and Whitham, 1955), and subsequently used
for rivers in Virginia and Maryland in USA:
in several other catchment hydrological studies (e.g., Woolhiser and Ligget,
1967; White and Jayawardena, 1975; Jayawardena and White, 1977, 1979). An P = 1.4 A D (2.2)
advantage in using the kinematic wave equations is that a downstream
boundary condition is not required for solving the constitutive equations. The where P is measured along the longest stream, A is measured from topo-
flow is upstream driven. graphical maps, and the fractal dimension D = 1.2. Other researchers (e.g.,
Rosso et al., 1991) have found slightly different values for D ranging from
2.2.2 Stream Network 1.036 to 1.29. Equation (2.2) establishes a link between fractal dimensions and
The stream network in a watershed transports water and sediments down- Horton’s law of stream numbers.1 It can be shown (Xie, 1993, p. 46) that
stream to a main stream, river, or to another water body. The denser the Du
stream network, the faster the surface waters flow down to the main stream. A2
Some streams in some watersheds are perennial whereas some are only P = P0   (2.3)
ephemeral and become dry after the rain stops.  A0 
Stream networks are usually characterized by their Horton-Strahler num-
bers (Horton, 1945; Strahler, 1952) in which each segment of a stream is where
treated as a node in a tree with the first order at the upstream highest level.
For example, most upstream segments of the stream will be order-1. When ln(rb )
Du = (2.4)
two order-1 segments join together, they form an order-2 stream. When two ln(rL )
order-2 segments join together, they form an order-3 stream. However, if an
order-2 stream intersects an order-1 stream, it still retains order-2 status. That and rb and rL, respectively, are the bifurcation and length ratios of streams
is, a lower order stream joining a higher order stream does not change the defined as:
order of the higher order stream. The lowest order stream (order-1) has no
tributaries. The highest order recorded is in Amazon River with order-12 at Ni
rb = (2.5a)
the mouth. Most streams are of orders 1–3. A typical stream network for a N i +1
catchment in China is shown in Fig 2.1. Associated with Horton-Strahler
number is another index called the bifurcation ratio which is the ratio of the Pi
number of streams of any order to the number of streams of the next higher rL = (2.5b)
Pi−1
order. A lower bifurcation ratio indicates fewer streams whereas a higher one
In Eqs. (2.5a) and (2.5b), Ni is the number of streams of order i, and Ni+1 is
the number of streams of order i+1, Pi is the average length of streams of order
i, Pi–1 is the average length of streams of order i - 1, and P0, A0 are the average
length and area of the smallest streams. The bifurcation and length ratios rb
and rL can respectively be obtained as the gradients of the plots of Ni vs. Ni+1
and Pi vs. Pi –1. The overall bifurcation ratio may also be obtained by averaging
the bifurcation ratios of different orders. However, it is to be noted that rivers
and streams are not exactly self-similar.
2.2.3 Water Bodies
Water bodies include lakes, reservoirs, detention ponds, polders, glaciers,
icecaps, lagoons, moats, fjords, bays, swamps, wetlands, marshes, rivers,
waterways, canals, and other areas which collect and store water. A water
body can be as small as a puddle and as large as an ocean, and can be natural
or artificial. They have a significant role to play in the basin scale hydrologi-
cal cycle by way of increasing the time of travel of water from origin to des-
tination. Water bodies may have different functions such as providing for
drinking water, irrigation, recreation, hydropower generation, floodwater
detention, aquatic life, etc., with different water quality requirements for
different functions.
1
Horton’s law of stream numbers states that there exists a geometric relationship
between the number of streams of a given order Ni and the corresponding order, i. The
parameter of this geometric relationship is the Bifurcation Ratio, rb. Horton’s law of
stream lengths states that there exists a geometric relationship between the average
length of streams of a given order and the corresponding order, i. The parameter of this
Figure 2.1 Sketch of a typical drainage basin of the Wuli River in Northern China. relationship is the Length Ratio, rL. Horton’s law of stream areas states that there exists
(The numbers in each stream segment refer to the Horton-Strahler numbers as a geometric relationship between the average areas drained by streams of a given order
explained in Sec. 2.2.2.) and the corresponding order i. The parameter of this relationship is the area ratio, RA
02_Singh_ch02_p2.1-2.8.indd 2 8/22/16 10:39 AM

COMPONENTS OF WATERSHEDS 2-3
2.2.4 Wetlands, Swamps, and Bogs “Noria,” was invented circa 750 BC to lift water from a flowing river by clay
Under the Ramsar international wetland conservation treaty, wetlands are buckets arranged at the circumference of a water wheel. This device enabled
defined as areas of marsh, fen, peatland or water, whether natural or artificial, water to be lifted to a higher elevation using the energy of the flowing water.
permanent or temporary, with water that is static or flowing, fresh, brackish or An underground canal system known as “qanats” (in Arabic) was used in
salt, including areas of marine water the depth of which at low tide does not Mesopotamia circa 550 BC. Such systems can be found even at the present
exceed 6 m. They may incorporate riparian and coastal zones adjacent to the time in the Xinjiang Province of China where the snow melt water from the
wetlands, and islands or bodies of marine water deeper than 6 m at low Tianshan mountains are conveyed to the “grape valley” area near Turpan and
tide lying within the wetlands. They are lands where the soil is saturated with in the city of Kumul via a network of underground canals.
water, which determines the type of plant and animal communities that can live Ancient irrigation works in Sri Lanka which date back to circa 300 BC dur-
in the soil. They generally include marshes, swamps, bogs, and similar areas. ing the reign of King Pandukabaya are considered to be some of the most
Wetlands play important roles in the environment. They purify the water complex irrigation systems of the ancient world. They include but are not
through their biofilters; they assist in flood control by detaining the floodwa- limited to many artificial reservoirs (known as “tanks” in local terminology),
ters temporarily; they support shoreline stability; they are biologically diverse canals, and sluices, a few of which still remain. Other engineering marvels of
ecosystems that support many types of plants and animals. Plants may be the time include the construction of “Yoda Ela” (meaning giant canal), which
completely submerged, floating, or partially submerged (emergent). Fish are maintained a gradient of approximately 10–20 cm/km for a distance of about
the most abundant biotic species, but other amphibians such as frogs, reptiles, 87 km during the reign of King Dathusena in fifth century AD to transfer
alligators, and crocodiles also make wetlands their home. Wetlands occur in excess waters from an upstream reservoir known as “Kala Wewa” in the
almost all countries. The largest may be in the Amazon River Basin. Wetlands Polonnaruwa district to a downstream reservoir known as “Tissa Wewa” in the
can be tidal or non-tidal. There are also artificial (man-made) wetlands, but Anuradhapura district, “Minipe Yoda Ela,” another giant canal to divert water
they take a long time (of the order of 10-100 years) to attain the same vegeta- from the longest river in Sri Lanka to provide water for irrigation on the left
tion composition as in natural wetlands. bank, also circa fifth century AD, which has now been extended and expanded
Wetlands are known to be carbon sinks thereby mitigating the effect of and in operation, and a giant reservoir known as “Parakrama Samudraya”
greenhouse gases. However, they are sources of nitrous oxide from the soils (meaning Parakrama sea; Parakrama is the name of the king who built it)
through denitrification.and nitrification processes. Nitrous oxide is a more which still provides water for irrigation and built during the reign of King
potent greenhouse gas than carbon dioxide. Forested wetlands are generally Parakramabahu I, circa twelfth century, AD. An important feature of irrigation
known as swamps. The largest swamps are found in the Amazon, Mississippi, canals in Sri Lanka is that they have been, and are still being, constructed along
and Congo River basins. A bog (sometimes called peatland) occurs where the topographical contours with only one embankment thereby receiving the
ground surface water is acidic and is a collection of dead plant material. They overland runoff from the other side.
are carbon sinks, if left undisturbed. The Romans built aqueducts to carry water from snowmelt in the Alps to
cities in the valleys below. The first aqueduct, the “Aqua Appia” was built in
2.2.5 Artificial Canals and Waterways 312 BC and was followed by many others throughout the Roman Empire,
In addition to the natural stream network, many watersheds have artificial some below ground surface and some elevated. Over a period of 500 years,
canals and waterways. They may run in the same direction as the natural 11 aqueducts were built, the last one being “Aqua Alexandrina” built in 226 AD.
waterway or may cut across a watershed and connect to another watershed. A Roman aqueducts consisted of infiltration galleries, steep chutes or drop
waterway is used to transport goods and people or for connecting places such shafts, settling tanks, tunnels, covered trenches, bridges to support the aque-
as cities. Examples include the waterways in cities such as Amsterdam, duct, siphons, and a distribution system at the destination. Prehistoric devel-
Bangkok, Suzhou, and Venice. Aqueducts, or conveyancing canals, on the opments of irrigation canals also have taken place in Mexico circa 500 BC.
other hand are used to transport water from a source to a destination. They Canals have also been constructed to connect seas. Two of the well-known
transport water for domestic use, irrigation and hydropower generation. The sea-level waterways are the Suez canal in Egypt that connects the
longest, and still existing, is perhaps the “Grand Canal” (also known as the Mediterranean Sea and the Red Sea opened in November 1869 after 10 years
Beijing-Hangzhou Grand Canal) of China. It passes through the city of Tianjin of construction, and the Panama canal managed by the Panama Canal
and the provinces of Hebei, Shandong, Jiangsu, and Zhejiang for a distance of Authority that connects the Atlantic Ocean and the Pacific Ocean opened in
about 1800 km. The oldest parts of the canal date back to the fifth century BC, August 1914 after more than 30 years since construction began due to techni-
although the various sections were eventually combined during the Sui cal and health issues. These two canals cut the respective sea voyages by about
dynasty by emperor Yang Guang (581–618 AD). It has been reconstructed 7000 km and 12,875 km.
several times and is now (since June 2014) a UNESCO World Heritage site. A Design of Alluvial Canals
more recent Mega project, also in China, is the South to North Water Transfer
Project aimed at transferring water from the Yangtze River to Beijing, Tianjin, Most, if not all, artificial canals for irrigation are constructed on natural forma-
and other northern cities of China via two routes known as the middle and the tions and are therefore prone to erosion and deposition and should be designed
eastern routes. The eastern route, which has a length of 1156 km with a flow to maintain channel stability. A “stable channel,” is defined as an unlined chan-
capacity of 150 m3/s, generally follows the “Grand Canal,” and will divert water nel for carrying water, the banks and beds of which are not scoured by the
from the lower reach of Yangtze River to the eastern Huang-Huai-Hai Plain moving water and in which objectionable deposits of sediments do not occur.
and terminates in Tianjin. The middle route, which has a length of 1267 km The widely used methods of design include the method of maximum permis-
with a flow capacity of 130 m3/s, will divert water from the Danjiangkou sible velocity, the regime method, and the tractive force method. Design for-
Reservoir on the Han River, a tributary of Yangtze River, to Beijing, and mulae for the first method, which aims at the maximum velocity which will not
Tianjin. This project, which commenced in 2002, and part of which is already cause erosion or deposition, are based on measured velocities in stable chan-
completed is expected to be fully operational in the year 2015 at a cost of about nels in various places and under various flow conditions. The second method
350 billion RMB (approximately US$ 57 billion). is empirical and based on observations made in the field, with notable field
There have been irrigation canals built from ancient times. They originate data from India, Egypt, and Pakistan. The design approach is to relate the depth
from artificial reservoirs or natural lakes, or from rivers across which hydraulic of flow, bed width, friction slope, and velocity empirically to the discharge and
structures such as low dams and weirs are built to divert the water flow. In soil particle size. The third method, which requires a relationship between
almost all irrigation canals, the flow of water is by gravity, but in some arid roughness coefficient and grain size and has some theoretical basis, was devel-
regions, where the source is groundwater, water has to be pumped to the sur- oped by the United States Bureau of Reclamation. Significant contributions to
face at a high energy cost. Irrigation started with the dawn of civilization as it this topic can be found in the studies by Kennedy (1895), Lacey (1930), Shields
was necessary for human beings to ensure food security throughout the year. (1936), and Lane (1955) among others.
The earliest recorded attempts to harness the waters of natural rivers may have
been in Egypt where cuts across embankments of the River Nile were made to 2.2.6 Hydraulic Structures
divert water to irrigable areas circa 6000 BC, in Mesopotamia (present Iraq and There are many types of hydraulic structures built from ancient times to serve
Iran), where the waters of the Tigris and Euphrates rivers were utilized circa many functions. They include dams, embankments, spillways, sluices, gates,
4000 BC, in India where irrigation canals were constructed circa 2600 BC, in locks, weirs, flumes, siphons, culverts, and storm surge barriers among others.
China where the Zhibo channel was constructed circa 613 BC in the pre-Qin The main function of a hydraulic structure is to change the natural flow in a
dynasty period (Xu, 2006), and the Dujiangyan irrigation system in Sichuan water body by diverting, controlling, or completely stopping. In some
Province constructed circa 256 BC that still supplies water for irrigation. instances, a hydraulic structure may be built to measure the flow rate. Dams,
Nilometers, invented circa 3500 BC, were used to measure the rise and fall of which store water behind the upstream face, may be classified according to
the river stage in the Nile, which helped to understand the timing of flooding several criteria such as the purpose, material used to build the dam and struc-
and in turn the timing of irrigation. The “Egyptian bucket wheel,” called tural design. The purposes of a storage dam may be for flood mitigation,
02_Singh_ch02_p2.1-2.8.indd 3 8/22/16 10:39 AM

Table 2.1 Ten Tallest Dams in The World

Name of dam Height of dam (m) Type Country River
Jinping-I 305 Concrete arch China Yalong

Nurek 304 Earthfill embankment Tajikistan Vakhsh
Xiaowan 292 Concrete arch China Lancang Jiang
Xiluodu 285.5 Concrete arch China Jinsha Jiang
Grande Dixence 285 Concrete arch Switzerland Dixence
Inguri 271.5 Concrete arch Georgia Inguri
Vajont (disused) 261.6 Concrete arch Italy Vajont
Manuel Moreno Torres 261 Earthfill embankment Mexico Grijalva
Nuozhadu 261 Embankment China Lancang Jiang
Tehri 260.5 Rockfill and earthfill India Bhagirathi
Mauvoisin 250 Concrete arch Switzerland Val de Bagnes
[Source: http://en.wikipedia.org/wiki/List_of_tallest_dams_in_the_world]
irrigation, hydro power generation, water supply, fisheries, and recreation. A triangular irregular network,3 the latter being usually measured whereas the
detention dam, which temporarily stores water during high flows, can also act former is computed. Data for determining a DEM can be acquired by direct
as a storage dam. A diversion dam, on the other hand, diverts the flow in a river land surveying or remote sensing. Remote sensing employs aerial surveys by
or canal, to a different area for purposes such as irrigation or water supply. flying objects, Interferometric Synthetic Aperture Radar, with two passes of a
They are generally low in height. A dam (or part of it) may also be of an radar satellite, Shuttle Radar Topography Mission (SRTM) with a single pass
overflow type or nonoverflow type. Structurally, a dam may be a gravity type, with satellites equipped with two antennas, Advanced Spaceborne Thermal
or arch type, or buttress type, or embankment type. A gravity dam resists the Emission and Reflection Radiometer (ASTER), etc. They may also be
external forces by its weight, whereas an arch dam resists the external forces obtained by interpolation between digital contour maps obtained by direct
by the arch action. The dead weight of an arch dam is much less than that of land surveying. Publicly available DEM’s of the entire world include
a gravity dam and requires less material to construct, but needs very strong GTOPO30 and ASTER, but their resolutions and quality in some areas vary.
abutments. Buttress dams are “mini” gravity dams with inclined upstream Higher quality and higher resolution DEM’s are available commercially.
faces supported by struts on the downstream side. Dams are constructed With the availability of DEM’s, watershed delineation can be done using
using concrete, rock, earth, and sometimes their combinations. Embankment- several packages of Geographical Information Systems (GIS) software.
type dams are constructed using soil, gravel, and other natural material and Examples of software packages include ArcGIS distributed by Environmental
generally not as high as concrete dams. They resist the external forces by the Systems Research Institute (ESRI), Geographical Resources Analysis Support
shear strength of the fill material, and, may be earthfill or rockfill. Components System (GRASS), Topographic PArameteriZation (TOPAZ) tools, and River
of a dam include a spillway, a control structure to control and measure the Tools from River Systems, Inc. (RSI). In the widely used ArcGIS environment,
outflow, a discharge channel, and energy dissipaters. Table 2.1 gives a list of the “Watershed” tool contained in the “Hydrology toolset” of the “Spatial
the 10 tallest dams in the world until 2015. Analysis Toolbox” can be used to demarcate the catchment boundary and the
Storm surge barriers are a relatively new type of hydraulic structures catchment area for a specified outlet point. However, the use of DEM’s for
designed and constructed across estuaries to prevent the storm surges from hydrological analysis has a problem with depressions. A group of raster cells
the sea penetrating into the river under extreme weather conditions. They are completely surrounded by cells of higher elevation may represent a real
usually gated structures that allow the passage of ships under normal condi- depression or an imaginary depression arising from interpolation errors dur-
tions. Among them are the world’s largest, the Oosterschelde, located ing DEM generation. It is important to arrive at a depressionless DEM before
between the islands Schouwen-Duiveland and Noors-Bevelanand, and the it can be used for any hydrological analysis. Algorithms for removing spuri-
world’s largest movable structure, the Maeslantkering, across the New ous depressions (Jenson and Domingue, 1988) are embedded into GIS soft-
Waterway (Nieuwe Waterweg, a canal that connects to River Rhine forming ware such as ArcView.
the artificial mouth of the River Rhine) that connects Rotterdam with North
Sea, both in the Netherlands, the world’s second largest movable structure, 2.4 WATERSHED HYDROLOGICAL PROCESSES
the Thames Barrier, located downstream of central London to protect
London from North Sea flooding, and the Inner Harbour Navigation Canal The primary hydrological process in a watershed is the capture of precipita-
Lake Borgne Surge Barrier constructed near the confluence of and across the tion which may be in the form of rain, snow or other forms, and the release
Gulf Intracoastal Waterway and the Mississippi River Gulf Outlet near New of part of the water gathered into a network of streams downslope to a water-
Orleans in United States. shed outlet and to store the remainder either above or below the ground
The design of hydraulic structures involves two stages: firstly, the hydraulic surface. Part of the precipitation intercepted by the watershed surface may
design to meet the flow requirement criteria and the second stage, the struc- also be returned back into the atmosphere by direct evaporation and evapo-
tural design to resist the external forces. Design guidelines, procedures as well transpiration. In addition to the water-related functions, a watershed is also
as examples can be found in most hydraulic engineering text books. the home to many types of biotic species including humans, plants, animals,
and microbes. There are designated watersheds for drinking water supply
2.3 DELINEATION OF A WATERSHED
purposes where no developments and/or settlements are permitted as well as
nondesignated watersheds where other types of development activities can
2.3.1 Use of Contour Maps take place. Other development activities may include agriculture, grazing for
For many years, delineation of a watershed was done manually using topo- cattle, and other domesticated animal species, industry, and residential. When
graphical contour maps. The procedure involves demarcating a curve starting there are water bodies in a watershed, they may also be used for recreation
from a specified outlet of the watershed and following it upstream at right purposes and for fisheries.
angles to the contour lines. It can be done by drawing arrows representing the
flow directions drawn perpendicular to each contour and in the direction of 2.5 CHARACTERISTICS OF A WATERSHED THAT IMPACT
the steepest gradient. The location of a catchment divide is taken to be where ON HYDROLOGICAL PROCESSES
flow directions diverge or where the arrows point in opposite directions.
2.5.1 Topography
2.3.2 Use of Digital Elevation Models Topography, which is formed over millions of years of geomorphological pro-
A digital elevation model (DEM) gives a three-dimensional representation of cesses, plays an important role in water movement in watersheds. Steep
a terrain surface. It can be represented as a RASTER2 or as a vector-based
3
A TIN is a vector-based representation of a surface made up of irregularly distrib-
2
A RASTER is a grid of squares which is a commonly used dataset in Geographical uted nodes and lines with three-dimensional coordinates (x, y, and z) that are arranged
Information Systems. in a network of nonoverlapping triangles.
02_Singh_ch02_p2.1-2.8.indd 4 8/22/16 10:39 AM

CHARACTERISTICS OF A WATERSHED THAT IMPACT ON HYDROLOGICAL PROCESSES 2-5
catchments have a faster time of concentration4 leading to a relatively quick state of sediment equilibrium, a condition when the overall erosion is
response to rainfall. In the catchment scale, topography may be quantified by balanced by the overall deposition. This condition is quite common when a
an average slope of the overland flow plane and an average slope of the main dam is constructed across a sediment-carrying river.
stream. These characteristics have been used to define certain empirical rela- Attempts to incorporate geomorphological features into hydrograph computa-
tionships for estimating the time of concentration and other process parame- tions have been introduced by defining a geomorphologically-based unit hydro-
ters. Topography also determines the flow patterns and pathways, and storages, graph (Rodrigues-Itube, 1979; Gupta et al, 1980) which has subsequently been
which influences the residence time of water within the catchment and has critically reviewed (Shamseldin and Nash, 1998), but applied in several situations
important influences on the flow response transient as well as on the quality of in unguaged areas where the geometric properties of the catchment have been
water. The contact time with surface and subsurface material has direct control used to parameterize the physical processes (Yang et al., 1998; Yang et al., 2002).
on chemical and biochemical processes that determine the quality of water.
2.5.3 Geology
Topography also determines the type of vegetation that can adapt to the water-
shed landscape. In steep catchments, the likelihood of soil erosion and trans- The underlying geology of a watershed determines the mechanisms of sub-
port of the eroded sediments to the mainstream is quite high leading to a new surface flow of water. When the surface layers are highly porous, much of the
dimension in catchment hydrology. Sometimes, the topography of a catchment precipitation goes into the sub-soil by infiltration. In rocky, sandy, or gravel
is often represented by a few integrated parameters for purposes of modeling geologic formations, the infiltrated water in turn may percolate downward to
catchment processes. Examples of such parameters include the overland slope, the water table or to deeper aquifers to reappear with a time delay as base
the mainstream slope and the topographical index. The topographical index is flow.5 The chemical composition of geology also determines the quality of
a measure of the extent of flow accumulation at a specified point on the topo- ground water and subsequently surface water. In the process of the subterra-
graphical surface. It was first defined in the context of TOPMODEL (Bevan nean water flowing through geological formations, many types of minerals
and Kirkby, 1979) as topographical wetness index as: necessary for human health get dissolved. The water extracted from such
formations is commercially marketed as spring waters by many vendors.
α Geology also determines the movement and storage of water in aquifers.
λ = ln (2.6)
tan β Groundwater accounts for about 0.5% of earth’s fresh water and is a major
source for municipal and irrigation use. Depending upon the geology, aqui-
where λ is the topographical wetness index, α is the local upslope area drain- fers may be confined or unconfined. Confined aquifers are generally deep and
ing through a certain point per unit contour length, and tan β is the local are sandwiched between two impermeable layers of rock, whereas unconfined
slope in radians. It is used to describe the spatial distribution of soil moisture. aquifers are shallow and confined from below only. In confined aquifers,
Topographical index controls flow accumulation, soil moisture, evapotranspi- water is held under pressure, which in some cases can be high enough to flow
ration, chemical and biological composition in soil, etc., and is used in model- out to the ground surface as a jet if a hole (well) is drilled through the confin-
ing vegetation, spatial distribution of soil moisture, mapping soil organic ing layer, whereas in unconfined aquifers, the phreatic surface is at atmo-
matter, mapping slope failure susceptibility, and many more. A similar index, spheric pressure. Unconfined aquifers are prone to pollution from the surface,
known as downslope index has also been suggested to quantify topographic whereas confined aquifers are relatively free from surface pollution.
controls on hydrology (Hjerdt et al., 2004) as:
2.5.4 Soil Type
d
tan αd = (2.7) Soils can be classified in different ways. Engineers classify soils according to
Ld their engineering properties; soil scientists classify them according to their
where Ld is the horizontal distance to the point with an elevation d meters below morphology, behavior, and genesis. The widely used engineering classification
the elevation of the starting cell, following the steepest-direction flow path. is that used by the United Soil Classification System (USCS) which has three
categories: coarse-grained soils (e.g., sands and gravels); fine-grained soils
2.5.2 Geomorphology (e.g., silts and clays); and highly organic soils (peat). These categories may be
The natural shape of a terrain is determined by its geomorphology, which can further subdivided by taking into account properties such as color, water
be considered as the science aimed at understanding the processes that shape content, strength, etc., which are not specified in the coarse classification.
the earth’s crust over a range of time scales. The near-static shape of the earth Subclassifications include clay (fine particle size), silty clay, silty clay loam, silt
has been formed over millions of years by the physical and chemical processes loam, silt, clay loam, loam, sandy clay, sandy clay loam, sandy loam, loamy
taking place on or near the earth’s surface. The more dynamic shapes nested sand, and sand (coarse particle size) that are represented in a triangle. The soil
within the overall large static-scale shape take place over shorter time scales and science classification is more suited to land-use applications. US Department
are caused by the actions of wind, rain, snow, ice, climate, and human activities. of Agriculture classifies according the grain size in the form of a soil texture
There can also be abrupt changes caused by geologic activities such as earth- triangle. They may also be classified as residual soils or transported soils
quakes, volcanic eruptions, landslides, and tsunami. Changes in small-scale depending on their formation, which may be either by weathering rocks or
features such as formation of drainage channels, rills, gullies, and permanent transported by wind, water, or gravity, from a different location. In addition to
river channels take place over shorter time scales. The main processes that con- these classifications, there are also region specific classifications.
trol the short-term changes are weathering, erosion, and deposition. Sediment Soil type in a watershed determines the type of vegetation it can support.
production by weathering and erosion usually takes place in the higher alti- Soil is a non-renewable resource in the watershed, which requires careful
tudes, the transport in the middle altitudes, and the deposition in the lower management to support high-quality vegetation that includes agricultural
altitudes. For short-term hydrological analysis excluding sediment dynamics, a crops. Soils carry water, nutrients, and minerals necessary for plant growth,
catchment may be considered as geomorphologically in steady state. which is an indicator of a watershed health.
Fluvial geomorphology refers to the study of the formation of rivers and Soils and plants have a symbiotic relationship. Plants modify and develop
how they respond to human and climate induced changes in the watershed. It soils by creating pore spaces that increase the capacity to store more water,
is an important area of concern from the point of view of river management add organic matter when leaves die and decompose, protect the surface from
to reduce flood vulnerability and damages while maintaining a healthy erosion, while soils act as the foundation for all types of plants.
aquatic habitat. Channel changes may have implications for water supply, 2.5.5 Shape, Size, and Slope
navigation, aquatic habitat, and protection of property. Human-induced
changes include construction of dams and reservoirs, sand mining in the river The shape of the watershed contributes to the speed with which the precipi-
bed, while a common natural disturbance is a flood. The response to such tated water reaches the outlet. The size determines the quantity of runoff
changes includes bed and bank erosion, bed aggradations, and changes in the generated by the watershed. The widely used rational method for estimating
channel cross section. Degradation of the channel takes place downstream of the peak discharge relates it directly to the catchment area (Q = CIA, Q-peak
a dam due to reduction of sediment load, or due to increase in stream power discharge, I-rainfall intensity, A-catchment area, C-runoff coefficient). The
(product of discharge and slope) or decrease in sediment yield as a result of size, shape, and slope also determine the hydrograph6 shape. For example,
conservation measures. Aggradations can occur due to increase of upstream increase in the catchment size tends to increase the base length7 of the hydro-
sediment supply due to, for example, construction activities, increase of graph; shape of main stream and valleys tends to affect the velocity and hence
downstream bed level due to, for example, sea level rises, and due to basin
wide increase in sediment supply. When exposed to such changes, the sedi- 5
Baseflow is that part of streamflow which comes from deep subsurface flow and
ment equilibrium is disturbed, but over a period of time, rivers attain new delayed shallow subsurface flow. It is also sometimes referred to as dry weather flow.
6
A hydrograph is a graph that displays the flow rate as a function of time.
4
Time of concentration is the time taken by a water particle to travel from the furthest 7
Base length is the length along the time axis between the beginning of the hydro-
point in the catchment to its outlet. graph and the end of the hydrograph.
02_Singh_ch02_p2.1-2.8.indd 5 8/22/16 10:39 AM

the time to peak; in a large catchment, depending upon where the high-inten- which the millennium development goals were established, the Rio Summit
sity occurs, the base length of the hydrograph may increase or decrease. If the (Earth Summit), held in Rio de Janeiro, June 3–14, 1992, as an inter govern-
high-intensity rainfall occurs near the outlet, the time base will be short. In mental summit under the auspices of the United Nations Conference on
comparison, if the high-intensity rainfall occurs at the far end of the catchment, Environment and Development (UNCED) and the aptly named commemo-
the time base would be relatively longer. If the watershed hillslopes are steep, rative summit Rio+20 held in Rio de Janeiro, from June 20–22, 2012. The
the overland flow will be swift, whereas in relatively flat watersheds, more focus of Rio+20 Summit was to ensure sustainability of the Earth system
water will infiltrate into the subsoil and delay the runoff process. The same while providing water, food, and energy security to all inhabitants of the
applies to the slopes of the stream network. earth. The goal was to forge a global agreement by establishing a set of sus-
tainable development goals acceptable to all parties. The water policies and
water laws enacted in the recent times for better management in a few
2.6 RIVER BASIN selected countries are summarized in the following sections.
A river basin is the land area drained by a river and its tributaries. The basic
difference between a river basin and a watershed is that a river basin has a 2.7.1 Water Policy in Japan
well-defined network of rivers most of which are perennial, which a watershed In Japan, the National Government is regulating the management of rivers
may not have. For example, in arid zones, drainage in a watershed may take and associated water resources through the River Law. According to the law,
place via overland flow and/or ephemeral drainage channels, which become each river administrator (national or regional government) is in charge of
dry once the precipitation has ceased. In a river basin, the river network is comprehensive river management, which include flood management, river
formed by geomorphological actions over long periods but could be altered by works, maintenance of infrastructure (e.g., dams and dikes), authorization of
human actions. The flow in the main river and tributaries take place perenni- water (resources) rights, and conservation of the river environment (includ-
ally albeit at a reduced rate during dry weather. In ancient times, most civiliza- ing water quality, regulation of river area use, etc.). Ground water is controlled
tions started along river banks, as they provided the basic essentials of life, by local governments. Water supply is controlled by municipal governments
namely, water and food. Examples include the Indus Valley Civilization (also and the private sector. Large rivers (109 in total) come under the river admin-
known as Harappan culture) circa 6000 BC along the River Indus (India and istration while other rivers are managed by local governments.
Pakistan), Ancient Nile Valley Civilization along the River Nile in Egypt circa Japan had a river administrative system for over 100 years. The first mod-
5000 BC and the Yellow River Valley civilization in China circa 1500 BC, all of ern river administrative system started after the enactment of the “old river
which were considered as hydraulic empires. Significant hydraulic civilizations law” in 1896, designed primarily for flood control. The “new river law” was
also existed in ancient Somalia, Sri Lanka, Mesopotamia, Mexico, and Peru. enacted in 1964 for the purpose of flood control and water use. Under this
A river basin is a land area for human habitation as well as the home for law, integrated river system management and water-use regulations were
many other biotic species. In the modern times, it is also a place for many introduced. Subsequently, to meet the changing social and economic needs,
types of economic activities such as agriculture and aquaculture, industry, and the river law was amended in 1997 with the objective of establishing a com-
tourism. Vegetation forms an important and essential component of a river prehensive river administrative system for flood control, water use, and the
basin as a link that provides exchange of water from the land to the atmo- conservation of the environment. Water reuse is practiced in industry and for
sphere. The vegetation canopy intercepts the rain thereby reducing the kinetic recreational purposes. Desalination by reverse osmosis is done in a small way
energy of rainfall which in turn reduces soil erosion. Plant roots make the in Fukuoka and Okinawa. Interbasin water transfers also take place for opti-
underlying soil more porous thereby increasing the capacity to store more soil mum utilization. The water sector in Japan is administered under five minis-
water, and transport water and minerals from the soil to the body of the plant. tries: Ministry of Land, Infrastructure, and Transport; Ministry of Health,
Dead leaves from vegetation when decomposed provide nourishment to the Labour, and Welfare; Ministry of Agriculture, Forestry, and Fisheries;
soil. Before decomposition, they also provide some protection to the soil sur- Ministry of Economy, Trade, and Industry; and Ministry of the Environment.
face. Plants also contribute to reduction of greenhouse gases by their ability to In rare situations, where conflicts occur, they are settled in an amicable man-
absorb carbon dioxide and release oxygen during the photosynthesis process. ner in the traditional Japanese way. There is also an initiative by the Cabinet
The greatest impact on a river basin is made by human population. Office to strengthen the coordination of functions of related ministries. Japan
Humans have a higher demand for water from the river basin. They alter the has 14 water related legislative enactments of which River Law is one.
natural river basin by actions, such as deforestation, building dams, urbaniza-
tion and other infrastructural development, migration of people and the 2.7.2 Water Policy in the European Union
accompanying increase of waste generation, all of which give rise to negative There have been three water-management directives in the European Union
impacts on the natural health of the river basin. There is a danger that such (EU). The Urban Waste Water Treatment Directive (91/271/EEC) of May 21,
actions also increase the risk of flooding the lowland areas of a river basin. 1991 is concerned about discharges of municipal and some industrial wastewa-
Most human activities in river basins alter and accelerate the natural hydro- ters. The Drinking Water Directive (98/83/EC) of November 3, 1998 concerns
logical cycle thereby threatening the sustainability of many river basins in the potable water quality and the Water Framework Directive (2000/60/EC) of
world today. October 23, 2000 concerns water resources management. Recently, the water
environment in the EU is under great pressure from economic activities, urban
and demographic developments, and climate change and has put forward a
2.7 RIVER BASIN MANAGEMENT
“Blueprint to Safeguard Europe’s Water Resources,” which outlines actions that
Earth’s water resources are limited and the per capita share of water is con- concentrate on better implementation of current water legislation, integration
tinuously decreasing as a result of increasing population. Approximately of water policy objectives into other policies, and filling the gaps in particular
97.5% of earth’s water resources are contained in the oceans which is not as regards water quantity and efficiency with a time frame target set to 2050.
suitable for human consumption except via an expensive process of desalina-
tion. The fresh water contained in ice caps and glaciers, inland lakes, soil 2.7.3 Water Policy in the United States
water, ground water, atmospheric water, and river and stream storage consti- In United States, conflicts continue to arise in constructing a framework of
tute the remaining 2.5% of which only about 0.5% is accessible. Although policies due to the contrasting climates of the Eastern and Western United
renewable and abundant in nature, fresh water has spatial and temporal vari- States. The difficulty lies in a lack of understanding of integrating policies in
ability. This temporal and spatial variability and the changing life styles of both sides of the country, while still maintaining a fair amount of regional
human population lead to water stress8 and water scarcity. To ensure sustain- control. California state has been the first to attempt to integrate both types of
ability of the Earth system while providing water, food, and energy security to the traditional policies. Major water policies are under the federal government,
all inhabitants of the earth it is imperative to have fair and equitable water for example, the Clean Water Act (CWA) of 1948. Water policy, management,
management policies. These were forged into development goals during sev- and pollution usually tend to fall under the jurisdiction of local and state gov-
eral summits of world leaders. These include the U.N. Millennium Summit of ernments. Federal government gives grants to local and state governments to
world leaders held at the UN headquarters from September 6–8, 2000, during construct facilities for sewage treatment. In the Western United Sates, the
“Prior appropriation doctrine” in which “first in time, first in rights” principle
together with the relative powers of the two parties is practiced. Under this
8
Water stress index is the fraction of the total annual runoff available for human use.
doctrine, the more powerful the party who diverts water, the more rights he
The widely used indicators of water stress, water scarcity, and water shortage are the
Falkenmark indices (Falkenmark, 1989). Based on water usage in multiple countries, has. Landowners can use the rivers as they see fit, as long as they do not disturb
Falkenmark assigns the following threshold values: > 1700 m3 per capita per year, no the natural flow. In the Eastern United States, the “Riparian rights doctrine” in
stress; 1000–1700 m3 per capita per year, stress; 500–1000 m3 per capita per year, scarcity; which landowners whose property adjoins a body of water have the right to
< 500 m3 per capita per year, shortage. make reasonable use of it as it flows through or over their property, is practiced.
02_Singh_ch02_p2.1-2.8.indd 6 8/22/16 10:39 AM

LAND USE 2-7
2.7.4 Water Policy in India Bangladesh (54 with India and 3 with Myanmar) including the three mighty
India has more than 18% of the world’s population, but has only 4% of world’s rivers the Ganges, the Brahmaputra, and the Meghna.
renewable fresh water resources and only 2.4% of world’s land area. Indian The earliest national river authority may perhaps be the Tennessee Valley
National Water Policy has been enacted and adopted in 1987, reviewed and Authority (TVA), which is a federally owned corporation in the United States
updated in 2002 and 2012. A new Draft National Water Framework Bill was established in May 1933 to develop the Tennessee Valley which had been seri-
introduced in April 2013. It covers uses of water, adaptation to climate change, ously affected by the great depression. Similar national authorities or commis-
enhancing water available for use, demand management and water use, effi- sions include the Yangtze River Water Resources Commission, the Yellow
ciency, water pricing, conservation of river corridors, water bodies and infra- River Conservancy Commission, the Pearl River Water Resources
structure, project planning and implementation, management of floods and Commission, Huai River Water Resources Commission, the Hai River Water
droughts, water supply and sanitation, institutional arrangements, trans- Resources Commission, the Songliao River Water Resources Commission,
boundary rivers, database and information systems for sharing information, and the Taihu Basin Authority in China.
and research and training needs. 2.8 MAJOR RIVER BASINS IN THE WORLD
The ranking of a river can be based on the value of a wide range of ranking
2.7.5 Water Policy in China
parameters such as length, average or maximum discharge, annual volume of
China with a population more than 1.3 billion has an annual per capita fresh flow to ocean, catchment area, depth of flow, width, sediment yield, popula-
water resources availability of 2100 m3, which is about a third of the world tion in the river basin, and many other characteristics. However, in general,
average. However, due to the monsoon climate and complex geographical the size and importance of rivers are ranked in terms of their length or dis-
conditions, temporal and spatial distribution of water resources is highly charge. The ten largest rivers in the world by discharge at the mouth are
uneven resulting in difficulties in food production and economic develop- Amazon (Brazil), Congo (Congo), Yangtze (China), Brahmaputra
ment in some parts of the country. For example, north China has about 70% (Bangladesh), Ganges (India), Yenisei (Russia), Mississippi (USA), Orinoco
of farmland, but only 16% of water resources compared with south China, (Venezuela), Lena (Russia), and Parana (Argentina). In Asia, the major river
which has only about 30% of farmland but with 84% water resources. The basins include the Yangtze, the Brahmaputra, the Mekong, and the Indus; in
maximum annual precipitation in southeast China is over 2000 mm, and the Africa the Nile and the Congo; in Europe, the Danube and the Rhine; in
minimum annual precipitation in northwest China is less than 50 mm. South America, the Amazon, the Orinoco and the Parana; in North America,
Recently, China has embarked on implementing a strict management system the Mississippi, the Colorado, the Columbia, and the St Lawrence. Table 2.2
to optimally utilize her water resources in the face of conflicting demands in gives a list of the 10 longest rivers in the world. Depending of the criterion
the economic sector and the protection of eco-environment. used, the rankings in Table 2.2 may change. It is also to be noted that the data
China has a centralized system of water management. The Ministry of available sometimes differ from source to source.
Water Resources formulates general policies, and develops strategies for water
administration, and enacts related laws and regulations. The Ministry of Water 2.9 LAND USE
Resources has established seven river basin commissions, namely, Changjiang
(Yangtze River), Yellow River, Huai River, Hai River, Songliao River, Pearl Land use in a watershed governs several processes in the hydrological cycle as
River, and Lake Tai. The Water Law, enforced in July 1998 and revised in 2002, well as the productivity of the land. Classification of land use is not unique but,
is the most fundamental legislation concerning water administration in China. based on the vegetation type, nine traditional classes can be identified. They are
This law organizes activities concerning the general development, use, and forests, land put to nonagricultural uses, barren and uncultivable land, perma-
conservation of water resources, river development and flood-control works, nent pastures and other grazing lands, miscellaneous tree crops and groves not
etc., as well as coordination of socioeconomic activities and related agencies included in the net area sown, culturable waste, fallow land other than current
concerned with water. The Water Law covers general rules for the development fallows, current fallows, and net area sown. A much simpler classification
and utilization of water resources, the maintenance of water supplies, the river based on the services provided by the land considers four categories, namely,
basin and water use facilities, water resources management, water disputes and land underlying buildings and structures, land under cultivation, recreational
the supervision and inspection of law enforcement, legal responsibility, and land and associated surface water, and other land and associated surface water.
supplementary provisions such as flood control, prevention, and control of
2.9.1 Types of Land Use
water pollution.
The main function of land from the point of view of food production and secu-
2.7.6 River Commissions/River Authorities rity is agriculture. According to Food and Agricultural Organization (FAO)
(International and National) approximately 30% of the world land area is used for agriculture and another
30% for forests (http://www.fao.org/docrep/018/i3107e/i3107e04.pdf). The
Many trans-boundary rivers experience conflicts when it comes to sharing percentage of land area taken for residential purposes is relatively small although
the resources of the river. They may arise as a result of differences in the social the percentage of urban areas is on the increase as a result of migration of people
and economic levels of the riparian countries, or be due to unilateral actions for better opportunities. With the demand for increased food production to feed
taken usually by upstream countries. For example, if an upstream country the increasing population the percentage of land used for agriculture is likely to
decides to build a dam for hydropower generation or for other purposes, the increase in the future. Since land is a finite and nonrenewable resource, there is
natural flow of the river is disturbed and controlled which may adversely always competition among various users. For example, in many countries agri-
affect the downstream countries. International river commissions and/or cultural and forestland are being converted to urban areas thereby threatening
agreements have been established in regions with trans-boundary rivers to the sustainability of the environment and related ecosystems.
avoid or minimize such conflicts and to share and assist in the management
of the resources in an optimal and equitable manner. Table 2.2 Ten Longest Rivers of the World
Examples include the Mekong River Commission (MRC), established on
April 5, 1995 with the participation of Cambodia, Lao PDR, Thailand, and Length Area Average
Vietnam, to which China and Myanmar joined as “dialogue partners” in 1996; Ranking Name Country/region (km) (106 km2) discharge (m3/s)
Organization of the Amazon Cooperation Treaty (OACT), established on July 1 Nile Africa 6,648 3.35 5,100
3, 1978 with the participation of Brazil, Peru, Bolivia, Colombia, Ecuador,
Venezuela, Guyana, Suriname, and French Guiana; the Danube Commission, 2 Amazon South America 6,500 7.05 175,000
initially established in 1948 by seven riparian countries but subsequently 3 Yangtze China 6,280 1.80 30,440
expanded to include Austria, Bulgaria, Croatia, Germany, Moldova, Slovakia, 4 Mississippi- The United States 6,020 3.22 16,792
Romania, Russia, Ukraine and Serbia; the Central Commission for Navigation Missouri
on the Rhine (CCNR), formally constituted in 1815 with the participation of
Germany, Belgium, France, The Netherlands, and Switzerland; the International 5 Yenisei Russia 5,539 2.60 19,830
Sava River basin Commission (ISRBC), established on December 29, 2004 with 6 Yellow China 5,400 0.75 1,775
the participation of Bosnia and Herzegovina, Federal Republic of Yugoslavia, 7 Ob-Irtysh Russia 5,410 2.98 12,300
Serbia, Republic of Croatia and Republic of Slovenia; and the Mosel Commission
established on December 21, 1962 with the participation of France, Germany, 8 Congo Congo 4,700 3.48 41,300
and Luxembourg. The latter four commissions operate within the European 9 Amur Asia 4,440 1.86 8,600
Commission. Similar bilateral treaties have been signed by India and Bangladesh 10 Lena Russia 4,400 2.49 17,000
with the establishment of the Joint Rivers Commission, which came into being
on November 24, 1972 to share the waters of the 57 trans-boundary rivers of (Source: Water Conservancy Encyclopedia China, Vols. 1–4)
02_Singh_ch02_p2.1-2.8.indd 7 8/22/16 10:39 AM

2.9.2 Effect of Land Use on the Basin Scale Hydrological Cycle The authors acknowledge the sources of the information reproduced in the
Land use practices have major impacts on water (quantity and quality), soil, two Tables in this chapter.
nutrients, and other bio-species. Deforestation tends to increase runoff and
sediment generation by erosion. Forest covers act as carbon sinks thereby con- REFERENCES
tributing to the reduction of global warming. They also tend to make the
underlying subsoil more porous thereby allowing more infiltration to take place Bevan, K. J. and M. J. Kirkby, “A physically based, variable contributing area
which in turn reduces overland flow. Conversion of forest land to urban areas model of basin hydrology,” Hydrological Sciences Bulletin, 24: 43–69, 1979.
has negative impacts on the sustainability of the land-water ecosystem. Urban Falkenmark, M., “The massive water scarcity threatening Africa—why isn’t
areas also have low infiltration capacities thereby reducing the water available it being addressed,” Ambio, 18 (2): 112–118, 1989.
in the subsoil that supports plant growth. Inappropriate land use can alter the Gupta V. K., E. Waymire, and C. T. Wang, “A representation of an instanta-
basin scale hydrological cycle which in some situations may lead to irreversible neous unit hydrograph from geomorphology,” Water Resources Research, 16:
consequences. Land use affects evapotranspiration, initiation of surface runoff, 855–862, 1980.
and washout of nutrients from soil, among other hydrological processes. Hack, J. T., “Studies of longitudinal stream profiles in Virginia and
Changes in land use in a catchment may be caused by a number of factors, Maryland,” U.S. Geological Survey Professional Paper 294-B, U.S. Government
such as urbanization, agricultural development, land cover changes by afforesta- Printing Office, Washington, D.C., pp. 45–97, 1957.
tion and deforestation, and construction of water infrastructures, such as dams, Hjerdt, K. N., J. J. McDonnell, J. Seibert, and A. Rodhe, “A new topographic
reservoirs, and water import/export facilities. Such changes have direct influ- index to quantify downslope controls on local drainage,” Water Resources
ences on runoff generation both temporally and spatially. They can also affect the Research, 40 (w05602): 1–6, 2004.
partitioning of precipitation between evapotranspiration and runoff. For exam- Horton, R. E., “Erosional development of streams and their drainage
ple, it has been reported that an annual runoff reduction of more than 50% has basins: hydro-physical approach to quantitative morphology,” Geological
taken place in Loess Plateau (Zhang et al., 2008) and in the mountainous catch- Society of America Bulletin, 56 (3): 275–370, 1945.
ments in Haihe River (Xu et al., 2014) in China as a result of soil conservation Jayawardena, A. W. and J. K. White, “A finite element distributed catchment
measures such as terracing, afforestation, and construction of sediment traps. model, I—Analytical basis,” Journal of Hydrology, 34 (3–4): 269–286, 1977.
Jayawardena, A. W. and J. K. White, “A finite element distributed catchment
2.9.3 Land Use and Runoff model, II—Application to real catchments,” Journal of Hydrology, 42 (3–4):
Land use has a direct relationship with runoff production. In vegetated catch- 231–249, 1979.
ments, the rate and quantity of overland runoff production is much less than Jenson, S. K. and J. O. Domingue, “Extracting topographic structure from
those of bare catchments. Erosion and sediment generation in bare catchments digital elevation data for geographic information systems analysis,”
are higher leading to land degradation and accumulation of sediments down- Photogrammetric Engineering Remote Sensing, 54: 1593–1600, 1988.
stream. Urbanization which increases the impervious surface areas in catch- Kennedy, R. G., “The prevention of silting in irrigation canals,” Proceedings
ments leads to a decrease in infiltration which in turn increases surface runoff. of the Institution of Civil Engineers, London, 1895, Vol. 119, pp. 281–290.
An increase in impervious surface areas associated with urbanization also Lacey, G., “Stable channels in alluvium,” Proceedings of the Institution of
slows down evaporative cooling, allowing surface temperatures to rise to levels Civil Engineers, London, 1930, Vol. 229, pp. 259–384.
higher than in rural areas. Urbanization in flood plain areas increases the risk Lane, E. W., “Design of stable channels,” Transactions, ASCE, 20 (2776):
of flooding due to increased peak discharge and volume, and decreased time 1234–1279, 1955.
to peak. Urban watersheds, on average convert about 90% of the storm rainfall Lighthill, M. H. and Whitham, G. B., “On kinematic waves, I. Flood move-
to runoff (runoff coefficient ≈ 0.9), whereas nonurban forested watersheds the ment in long rivers,” Proceedings of the Royal Society, London, Series A, 1955,
conversion rate is about 25% (runoff coefficient ≈ 0.25). Catchment runoff also Vol. 229, pp. 281–316.
increases after deforestation and, as expected, decreases after afforestation. Mandelbrot, B. B., The Fractal Geometry of Nature, W. H. Freeman, New
Water infrastructure development such as construction of dams, reservoirs, York, pp. 188–189, 1983.
and diversions from/and to such structures can also affect the runoff. The Rodriguez-Iturbe, I. and J. B. Valdes, “The geomorphologic structure of the
effects of land use change on river flows are more evident in arid climates, hydrologic response,” Water Resources Research, 15: 1409–1420, 1979.
where the low flows are more sensitive to land use changes. Rosso, R., B. Bacchi, and P. La Barbera, “Fractal relation of mainstream
length to catchment area in river networks,” Water Resources Research, 27 (3):
2.9.4 Land Use and Water Quality 381–388, 1991.
Higher concentration of population resulting from urbanization also substan- Shamseldin A. Y. and J. E. Nash, “The geomorphological unit hydro-
tially degrades water quality, especially when there is no wastewater treatment. graph—a critical review,” Hydrology and Earth System Sciences, 2 (1): 1–8,
The resulting degradation of inland and coastal waters impairs water supplies, 1998.
causes oxygen depletion and fish kills, increases harmful algal blooms, and Shields, A., “Anwendung der ahnlichkeitsmechanik und der turbulenz-
contributes to the spread of waterborne diseases. It is important to note that forschung auf die geschiebebewegung,” Mitteilungen der Preussischen
land use changes are aimed at increasing economic productivity, but the price Versuchsanstalt fur Wasserbau und Schiffbau, No. 26, Germany, Berlin, 1936
that has to be paid is the impairment to ecosystems and biodiversity, with (English Translation by W.P. Ott and J.C. van Uchelen, California Institute of
potentially serious and unquantifiable costs. Intensive agriculture using various Technology, Pasadena, CA, 1936, p. 36).
types of fertilizes to increase productivity has increased food production but at Strahler, A. N., “Hypsometric (area-altitude) analysis of erosional topology,
the same time caused extensive environmental damage. Intensive agriculture Geological Society of America Bulletin 63 (11): 1117–1142, 1952.
increases erosion and sediment yield, and leaches nutrients and agricultural White, J. K. and A. W. Jayawardena, “Discussion of ‘A finite element
chemicals to groundwater, streams, and rivers. Agriculture has become the larg- approach to watershed hydrology’ by C. Taylor et al.,” Journal of Hydrology,
est source of excess nitrogen and phosphorus in waterways and coastal zones. 27: 357–358, 1975.
Increased fertilizer use has led to degradation of water quality in many regions. Woolhiser, D. A. and J. A. Ligget, “Unsteady one dimensional flow over a
plane; the rising hydrograph,” Water Resources Research, 3 (3): 753–771, 1967.
2.10 CONCLUSION
Xie, H., Fractals in rock mechanics, Geomechanics Research Series 1, A. A.
Balkema, Rotterdam, The Netherlands, 1993.
In this early chapter of the book, an attempt has been made to highlight and Xu, X., D. Yang, H. Yang, and H. Lei, “Attribution analysis based on the
describe the basic components of a watershed including their main character- Budyko hypothesis for detecting the dominant cause of runoff decline in
istics, basin scale hydrological processes and the characteristics that impact Haihe basin,” Journal of Hydrology, 510: 530–540, 2014.
them, river basins, and their management, including water policies in some Xu, Q. Water Conservancy Encyclopedia China [M]. China Waterpower
selected countries/regions, and land use and their relationship to other hydro- Press, Beijing, 2006.
logical processes. The Chapter is introductory in its presentation since much Yang, D., S. Herath, and K. Musiake, “Development of a geomorphology-
of the related details will appear in subsequent chapters. based hydrological model for large catchments.” Annual Journal of Hydraulic
Watersheds, or river basins, have been the cradles of civilization since Engineering, JSCE, 42: 169–174, 1998.
ancient times. What appears to be the challenge for human population is to Yang, D., S. Herath, and K. Musiake, “A Hillslope-based hydrological model
ensure the continuity and sustainability of the ecosystems in watersheds, using catchment area and width functions.” Hydrological Sciences Journal, 47
which in recent times have been threatened by human activities. A balanced (1): 49–65, 2002.
approach in which optimal utilization of the resources of the watershed, the Zhang, X., L. Zhang, J. Zhao, P. Rustomji, and P. Hairsine, “Responses of
river system, and land productivity are sought while at the same time ensur- streamflow to changes in climate and land use/cover in the Loess Plateau,
ing the sustainability of the entire system appears to be the way forward. China.” Water Resources Research, 44: W00A07, 2008.
02_Singh_ch02_p2.1-2.8.indd 8 8/22/16 10:39 AM

Chapter 3
Water Balance
BY
C. PRAKASH KHEDUN AND VIJAY P. SINGH
ABSTRACT turnover, time is defined as τ o = S / O , where S is the storage capacity of the

water body and O is the drainage rate. The residence time, which is the time
The water balance is an expression of the law of conservation of mass, formu-
that a water molecule remains in a reservoir, can also be defined. Since every
lated as the continuity equation. The formulation requires that the balance is
water molecule has a different residence time, the average residence time is
done for a system of specified geometry at a specified time scale. In hydrology, ∞
the system can be a watershed, a channel or a segment thereof, a lake, a plant most commonly used. It is defined as τr = ∫ τr ⋅ f (τr ) ⋅d τr , where f (·) is a
0
root zone, or an aquifer. The system can be extended to a region, continent, function and τ r is the residence time of one molecule (Mussy and Higy,
or even the globe, but it becomes difficult to accurately compute the different 2011). This continuous renewal allows a larger volume of water to be with-
components of the water balance for such a large system because of the lack drawn than is actually available on land at a point in time. For example, the
of data, especially at small time scales, and our less than complete under- total volume of water available for consumption is about 2000 km3, while the
standing of the interactions between components. The time scale, depending total volume of water withdrawn is 3800 km3/year—only 10% of the maxi-
on the spatial size of the system, can be hourly, daily, weekly, monthly, sea- mum available renewable freshwater resources [Fig. 3.2 (Oki and Kanae,
sonal, or yearly. The objective of this chapter is to present rudimentary 2006)]. Evaporation from the oceans represents the bulk of the moisture
aspects of water balance and its computation. in the atmosphere, while terrestrial evapotranspiration contributes only
about 20%. Further, transpiration contributes between 80 and 90% of the
3.1 INTRODUCTION total water flux from land, recycling 62,000 ± 8000 km3 of water per year
(Jasechko et al., 2013).
Computation of water balance of a hydrologic system entails computing the
individual components of the hydrologic cycle and their interactions. The role
of water in each process or component of the cycle is complex. As water 3.3 WATER ON THE EARTH
moves in the hydrological cycle, it not only interacts with other processes Over 70% of the Earth is covered with water. Unfortunately, most of this water
within the cycle, but also dynamically interacts with other components of the is in the oceans; about 97% of the world’s water is saltwater and is undrinkable
atmosphere, biosphere, cryosphere, lithosphere, and pedosphere. or unusable without some form of treatment to remove the salt. Desalination,
The amount of water in each state—ice, liquid, and vapor—varies over time. however, is an extremely expensive and energy intensive process. Thus, the
For example, on a seasonal scale, ice cover and rain-to-snow ratio changes with water in the ocean is inaccessible for most countries or communities around
fluctuations in temperature. Furthermore, changes, whether natural or anthro- the world. Only about 3% of the water on Earth is fresh. The freshwater
pogenic, in any one of the components of the hydrological cycle or the Earth’s reserves include: rivers and lakes, marshes, the vadoze zone, groundwater
system, affects the balance of water in the hydrological cycle. Although Chap. aquifers, glaciers and other permanently snow-covered regions, permafrost,
1 deals with the hydrologic cycle, it is appropriate here to revisit different com- biological entities, and the atmosphere (Table 3.1). Not all of this water is
ponents from the water balance perspective. After a brief description of the readily accessible or exploitable—68.7% is locked in ice caps and glaciers and
hydrological cycle, its components, and hydrologic fluxes, a discussion on the 30.1% is in fresh groundwater, leaving a very small portion (1.3%) in surface
amount of water on the Earth highlighting the spatial distribution of water is and other freshwater sources for human consumption.
given. Empirical equations for the computation of water budget is given next,
followed by a presentation of different water balance models, ranging from the 3.3.1 Earth’s Water Balance
simple bucket model to more complex land surface models. Finally, the influ- On the continental scale, the water balance of the major rivers and self-contained
ence of natural and anthropogenic effects on the water balance is discussed. hydrologic systems gives an estimate of renewable water resources available
for human consumption (Table 3.2). Water on Earth is not uniformly distrib-
3.2 HYDROLOGIC FLUXES uted. Asia, followed by South America and Africa, receives the largest volume
of precipitation, while Antarctica receives the least. Evaporation in Antarctica
The hydrologic cycle and its various components are shown in Fig. 3.1. The is nil, while on all the other continents, it is greater than half of the total vol-
main components of the hydrological cycle are precipitation, interception, ume of precipitation. The total evaporation from land as a whole is equal to
evapotranspiration (evaporation from free water surfaces and transpiration about 60% of the total precipitation.
from plants), depression storage, infiltration, surface runoff, interflow, and Discharge from rivers in Asia and South America combined is more than
groundwater flow. 50% of Earth’s total river discharge (Table 3.3). In Africa, one of the largest
The interchange of water between open water bodies and land via the continents by area, river flow is only 10% of the world’s total flow. This
atmosphere not only ensures the replenishment of water in lakes and riv- equates to a mean discharge of 4.8 l/s/km2—the least among all continents.
ers, but also prevents deterioration in water quality. The rate of renewal In contrast, the mean discharge for South America is 21 l/s/km2, while that
varies spatially and temporally, depending on the climate. This renewal, or for Asia, Europe, and North America is about 10 l/s/km2 for each.
3-1
03_Singh_ch03_p3.1-3.12.indd 1 8/22/16 11:21 AM

Precipitation
Evapotranspiration
Channel precipitation
Interception
Depression storage Surface runoff
Ground surface
Infiltration
Soil moisture Interflow
Percolation
Capillary rise
Groundwater storage Groundwater flow
Streamflow
Underground flow
into or out of the area
Figure 3.1 Components of the hydrologic cycle [Source: US EPA, 2015].
Figure 3.2 Global hydrological fluxes (1000 km3/year) and storages (1000 km3) [Source: Oki and Kanae, 2006].
2
03_Singh_ch03_p3.1-3.12.indd 2 8/22/16 11:21 AM

Water on THE Earth 3-3
Table 3.1 Water Reserves on Earth (Source: Shiklomanov, 1993)

Percentage of global reserves
Distribution area Volume Layer
(103 km2) (103 km3) (m) of total water of fresh water
World ocean 361,300 1,338,000 3,700 96.5 –

Ground water 134,800 23,400 174 1.7 –
Fresh water 10,530 78 0.76 30.1
Soil moisture 16.5 0.2 0.001 0.05
Glaciers and permanent snow cover 16,227 24,064.1 1,463 1.74 68.7
Antarctic 13,980 21,600 1,546 1.56 61.7
Greenland 1,802 2,340 1,298 0.17 6.68
Arctic islands 226 83.5 369 0.006 0.24
Mountainous regions 224 40.6 181 0.003 0.12
Ground ice/permafrost 21,000 300 14 0.022 0.86
Water reserves in lakes 2,0548.7 176.4 85.7 0.013 –
Fresh 1,236.4 91 73.6 0.007 0.26
Saline 822.3 85.4 103.8 0.006 –
Swamp water 2,682.6 11.47 4.28 0.0008 0.03
River flows 148,800 2.12 0.014 0.0002 0.006
Biological water 510,000 1.12 0.002 0.0001 0.003
Atmospheric water 510,000 12.9 0.025 0.001 0.04
Total water reserves 510,000 1,385,9844.61 2,718 100 –
Total fresh water reserves 148,800 35,029.21 235 2.53 100
Table 3.2 Water Balance at the Continental Scale (Source: Shiklomanov, 1993)
Precipitation Evaporation Runoff
Continent (mm) (km )
3
(mm) (km ) 3
(mm) (km3)
Europe 790 8,290 507 5,320 283 2,970

Asia 740 32,200 416 18,100 324 14,100
Africa 740 22,300 587 17,700 153 4,600
North America 756 18,300 418 10,100 339 8,180
South America 1,600 28,400 910 16,200 685 12,200
Australia and Oceania 791 7,080 511 4,570 280 2,510
Antarctica 165 2,310 0 0 165 2,310
Land as a whole 800 119,000 485 72,000 315 47,000
Areas of external runoff 924 110,000 529 63,000 395 47,000a
Areas of internal runoff 300 9,000 300 9,000 34 1,000b
a
Including underground water not drained by rivers.
b
Lost in the region through evaporation.
Table 3.3 River Runoff for Each Continent (Source: Shiklomanov, 1993)
Annual river runoff
Portion of total Area Specific discharge
Continent (mm) (km3) runoff (%) (103 km2) (l/s/km2)
Europe 306 3,210 7 10,500 9.7

Asiaa 332 14,410 31 43,475 10.5
Africab 151 4,570 10 30,120 4.8
North Americac 339 8,200 17 24,200 10.7
South America 661 11,760 25 17,800 21.0
Australiad 45 348 1 7,683 1.44
Oceania 1,610 2,040 4 1,267 51.1
Antarctica 160 2,230 5 13,977 5.1
Total land area 314 46,770 100 149,000 10.0
a
Asia includes Japan, the Philippines, and Indonesia.
b
Africa includes Madagascar.
c
North and Central America.
d
Australia includes Tasmania.
03_Singh_ch03_p3.1-3.12.indd 3 8/22/16 11:21 AM

3-4 Water Balance
3.3.2 Per Capita Water Availability

Total annual precipitation or river discharge gives a limited idea of the amount Precipitation Evaporation
of water available for human consumption, especially given that the world’s
population is not uniformly distributed. Per capita water availability depends
on both the amount of renewable water available and the total population Runoff
dependent on it. On a continental scale, Australia and Oceania, even though
have the lowest percentage runoff, have the highest per capita water availability. Dry soil
Asia, home to two-third of the world’s population, has the highest annual river
Field capacity
runoff by volume, but has the least per capita water availability (3920 m3/
person/year), followed by Europe (4230 m3/person/year), and Africa (5720 m3/
person/year) (Shiklomanov, 1998).
At the country scale, water availability depends on a number of factors, Water
depth Wet soil
including: surface area, climate regime, geographic and physical conditions,
geomorphology, and population. Two countries on the same continent may
have very different per capita water availability. Brazil and Peru, for example,
both located in South America have distinctly different renewable per capita Figure 3.3 Manabe’s simple bucket hydrologic model.
water availability: 28,124 and 54,023 m3/inhabitant/year, respectively; the latter
being one of the highest in the world (FAO, 2015). Iceland (515,152 m3/inhabit-
ant/year), Guyana (301,250 m3/inhabitant/year), and Suriname (183,673 m3/ bucket. When it rains, if precipitation is greater than evaporation, the mois-
inhabitant/year) have the highest total internal renewable water resources per ture capacity of the soil sample increases up to the point where it cannot hold
capita, mainly due to their small population (330,000, 800,000, and 539,000, any more water and any excess water thus runs off. This simplification is not
respectively). Canada, with a population of over 35 million, has a total internal totally accurate; it has been established that runoff occurs if the rate of pre-
renewable water resource per capita of 81,007 m3/inhabitant/year. On the other cipitation exceeds the rate of infiltration irrespective of the degree of soil
side of the spectrum, Kuwait (negligible), Bahrain (3 m3/inhabitant/year), and saturation (Loaiciga et al., 1996).
the United Arab Emirates (16.05 m3/inhabitant/year) have the lowest total Budyko (1956) used the simple bucket model to calculate Earth’s energy
internal renewable water resource per capita, mainly due to their geographic balance and Manabe (1969) was the first to include this hydrologic parame-
location. China and India, two of the most populous countries, have only 1986 terization scheme within a general circulation model (GCM). The simple
and 1155 m3/inhabitant/year, respectively. Of the 181 countries included in the bucket model ignores vegetation and other important hydrologic processes;
FAO’s AQUASTAT database, 39% have less than 1700 m3/inhabitant/year; 27% subsequent models have been developed to capture more complex and realis-
have less than 1000 m3/inhabitant/year; and 16% have less than 500 m3/inhabit- tic hydrologic processes.
ant/year, where 1700, 1000, and 500 m3/inhabitant/year are the United Nations
thresholds for water stressed, scarce, and absolute scarcity, respectively. The Bucket with a Bottom Hole Model
number of people living in each of these regions is increasing, hence putting Kobayashi et al. (2001) proposed the bucket with a bottom hole model
more pressure on the limited water resources available. (BBH); a simple bucket model with a hole at the bottom that allows both
downward and upward movement of water across the bottom surface. BBH
3.4 WATER BALANCE MODELING is an active surface soil layer of depth D (about 10 cm thick), where the water
content can go below the field capacity and an underlying soil moisture
Water balance is based on the principle of conservation of mass, where the reservoir. The change in W, the equivalent depth of water in the soil layer, is
change in storage ( ∆S ) is the difference between the inflow and outflow of given by:
water, expressed as:
∆S = P − R − E − G (3.1) ∆W = W (t + 1) − W (t ) = Pr (t ) − E(t ) − Gd (t ) − Rs (t ) (3.2)
where P is the precipitation, R is the runoff, E is the evaporation, and G is the where t is the time in days, Pr is the daily precipitation, E is the daily evapotrans-
deep groundwater storage. The water balance, or water budget, can be computed piration, Gd is the daily gravity drainage (positive) or capillary rise (negative),
for part of the hydrologic cycle (soil profile, terrestrial, or atmosphere compo- and Rs is the daily surface runoff. E = M ⋅ E p, where M = min [W /σWmax , 1],
nent), or over a defined spatial dimension (root zone, catchment area, continen- E p is the daily potential evapotranspiration (PET), Wmax is the total water-
tal, or global). The water budget can also be computed for any temporal scale. holding capacity (porosity × D ), and s is the resistance of ground cover to
3.4.1 Water Balance Models W − a 
evapotranspiration. Gd = exp   − c, where a is nearly equal to or less than
 b 
The hydrologic cycle is governed by both local processes and remote ocean-
atmosphere interactions. On the local scale the land surface exerts a non- WFD, the field capacity, b is the soil moisture recession parameter, and c is the
negligible influence on the atmospheric boundary layer through momentum, daily potential capillary rising. When W equals bucket capacity WBC = ηWmax,
energy, and water fluxes. This may occur over large spatial and temporal scales, where η (≤ 1) is the moisture retaining capacity of the soil, and Pr > E + Gd ,
and influences both the weather and the global climate system. This interac-
tion can be broadly classified into two types: that affecting the physical climate Rs = Pr − E − Gd (3.3)
system and that impacting the biogeochemical cycles. The physical system
involves the exchanges of radiation, sensible heat, latent heat, and momentum, runoff is still possible even if W < WBC, when Pr > E + Gd + (WBC − W )
and has an immediate impact on the wind regime, precipitation, skin tem-
perature, and soil moisture. The biogeochemical cycles involve the exchanges Rs = Pr − E − Gd − (WBC − W ) (3.4)
of gases, which affect the radiative transfer characteristics, and consequently,
the energy budget of the planet (Sellers, 1991). These processes affect the A modified version of the model (BBH-B) was later proposed to capture
movement of water, its interactions within the hydrological cycle, and its inter- bioprocesses. It is a two-layer model with the intent of modeling interception
actions with other components of the atmosphere, biosphere, cryosphere, losses and macropore flow in the root zone (Kobayashi et al., 2006).
lithosphere, and pedosphere. These processes are complex. Understanding and
modeling the role of water in these processes helps in the development of more Simple Water Balance Model
precise hydrological models. A large number of water balance models have Schaake et al. (1996) developed a “simple water balance” (SWB) model which
been reported in the literature (Singh, 1995; Singh and Frevert, 2002a, b, 2006; is intermediate between the one-parameter simple bucket model and the
Singh and Woolhiser, 2002). In this section, some of the different water bal- more complex hydrological models. The main aim of the model was to
ance models, ranging from the simple one-parameter model to more complex improve the representation of runoff, with respect to the simple bucket
ones that include most known hydrological processes, are presented for pur- model, without including every complex physical process involved. The
poses of illustrating the rudiments for computing water balance. model parameters include both a physical representation and a statistical
averaging of the processes. SWB can be applied to the catchment scale or to
Simple Bucket Model represent runoff in the coupled atmospheric-hydrological models (e.g., the
The “bucket” (or leaky bucket) hydrologic model is the simplest water bal- NOAH land surface model). It can help in refining our understanding of the
ance model and the fundamental building block of most rainfall-runoff level of complexity required in realistically modeling land surface processes
models (Fig. 3.3). In this model, a single-layer soil sample is simulated as a and to estimate model parameters.
03_Singh_ch03_p3.1-3.12.indd 4 8/22/16 11:21 AM

Water Balance Modeling 3-5
Thornthwaite’s Water Balance Model and Mather (1957). It was written in APL and later translated in other lan-
Thornthwaite (1931) produced climatic maps of North America and later that guages. McCabe and Markstrom (2007) developed a monthly water balance
of the Earth (Thornthwaite, 1933) using evaporation data to compute a mois- model, based on Thornthwaite’s method, driven by a graphical user interface.
ture index. He utilized two concepts: precipitation effectiveness and tempera- The model can be used for water balance assessment, research, or as a class-
ture efficiency, defined as the accumulated sum of monthly precipitation to room instruction tool. The main components of the model are: snow accu-
evaporation ratio and the accumulated sum of temperature to evaporation mulation, direct runoff, snowmelt, evapotranspiration and soil-moisture
ratio, respectively, in producing those maps. Later, Thornthwaite (1948) pro- storage, and runoff (Fig. 3.4). The inputs to the model are monthly tempera-
posed a revised climate classification scheme and introduced the term poten- ture (T, in °C), monthly total precipitation (P, in mm), and the latitude of the
tial evapotranspiration (PET) (Brutsaert, 2005). He recognized that the location of interest. The latter determines the day length, a variable that
moisture index should not be based on actual evaporation and transpiration determines PET. The hydrological partitioning and steps in the computa-
from the soil but on PET. Potential evaporation was derived empirically from tions are summarized as follows.
mean temperature. The method was further modified by Thornthwaite and Snow Accumulation The model first determines the amount of monthly
Mather (1955) to accommodate a wider range of soils and vegetation types. precipitation (P) that is rain ( Prain ) or snow (in mm). Mean monthly tempera-
Air temperature is assumed to be correlated with the integrated effects of ture (T) determines whether precipitation is snow, rain, or a mix of both.
net radiation and other controls of parameters affecting evapotranspiration. It When the mean monthly temperature is below Tsnow, all precipitation is con-
is used as an index of the energy available for evapotranspiration, shared sidered snow and if T is greater than Train, all precipitation is considered to be
between heating of the atmosphere and evapotranspiration (Calvo, 1986). rain. When T is between Tsnow and Train, the amount of precipitation that is
PET is computed over a 12-h (amount of daylight) and a 30-day month: snow is given by:
a
 10T   T −T 
PET = 1.6  a  (3.5) Psnow = P ×  rain  (3.7)
 I  Train − Tsnow 
where Ta is the mean monthly air temperature (°C), and I is the heat index for
the station, which is the sum of 12 monthly heat indices, given by: and Prain = P − Psnow . The value for Train has been empirically found to be 3.3°C,
12 1.514
while that for Tsnow varies with elevation. For elevations less than 1000 m,
T  Tsnow = −10°C and for elevations more than 1000 m, Tsnow = −1°C .
I = ∑ ai  (3.6)
i =1
5  Direct Runoff Direct runoff (DRO) is the fraction of rainfall that does
not infiltrate, for example, from impervious surfaces, and runoff from rain-
and a = 0.000000675 I 3 − 0.0000771I 2 + 0.01792 I + 0.49239 . The values for a fall excess after infiltration. The percentage of precipitation that becomes
and I can be obtained from tables in Thornthwaite and Mather (1957). The DRO has been empirically found to be about 5%. The amount of precipita-
information required to compute the water balance at a site includes: latitude, tion after DRO is Premain = Prain − DRO .
mean monthly air temperature, mean monthly precipitation, necessary con- Snow Melt The fraction of snow storage (snowstor) that melts in a month
version and computational tables, and information on water holding capacity (SMF) is given by:
of the depth of soil for which the balance is to be computed. Black (2007) notes  T − Tsnow 
that the latter method to compute water balance is useful for “description, SMF =   × meltmax (3.8)
classification, management, and research,” but given that the method is based Train − Tsnow 
on monthly data, precipitation falling at the end of the month may not be
allocated to the appropriate month. where the meltmax is the maximum melt rate set to 0.5. If SMF is greater than
Several computer programs have been written to compute water balance meltmax, then SMF = meltmax. The snow water equivalent of the amount of
using the Thornthwaite method. Black (1981) was probably the first to write snow (SM) that is melted in a month is given as SM = Psnow × SMF. The total
a computer program to implement the method described in Thornthwaite liquid water input to the soil ( Ptotal ) is the sum of SM and Premain.
Temperature (T) Precipitation (P)
Potential
evapotranspiration
(PET)
Snow (Psnow)
Rain (Prain)
Snow melt
Actual (SM)
Snow
evapotranspiration Direct runoff (DRO)
storage
(AET)
(snostor)
Surplus runoff (RO)
Soil-moisture storage capacity (STC)
Soil-moisture storage (ST)
Figure 3.4 Components of the Thornthwaite monthly water balance (Source: McCabe and Markstrom, 2007).
03_Singh_ch03_p3.1-3.12.indd 5 8/22/16 11:21 AM

3-6 Water Balance
Evapotranspiration and Soil-Moisture Storage Actual evapotranspiration semidistributed macroscale hydrologic model that can solve both the water
(AET) is a function of PET, Ptotal , soil-moisture storage (ST), and soil moisture and energy balance. It can be run in either the water balance mode or the
withdrawal (STW). Monthly PET is calculated through the Hamon equation water-and-energy balance mode; the former being less computationally
(Hamon, 1961) intensive. VIC was developed for incorporation in GCMs and has under-
gone several updates and modifications over the last 20 years. It has been
PETHamon = 13.97 × d × D 2 × Wt (3.9) used in studies ranging from water resources management, including the
effect of climate change, to land-atmosphere interaction.
where PETHamon is PET in mm, d is the number of days in a month, and D is VIC operates at the grid scale and subgrid variations are captured statisti-
4.95 × e0.062×T cally. The main features of VIC is its ability to incorporate subgrid heteroge-
the mean monthly hours daylight units in 12 h, and Wt = ,a
100 neity in land surface vegetation classes, subgrid variability in soil moisture
saturated water vapor density term, in g/m . When Ptotal for a month is less
3 storage capacity, drainage from the lower soil moisture zone as a nonlinear
than PET, AET is equal to Ptotal plus the amount of moisture that can be with- recession, and topography, thus enabling more realistic hydrology in moun-
drawn from storage in the soil. STW decreases linearly with decreasing ST, tainous regions by allowing for orographic precipitation and temperature
such that as the soil becomes drier, it becomes more difficult to remove water lapse rates. Within the grid-based VIC, vegetation characteristics [e.g., leaf
from the soil and less is available for AET: area index (LAI), albedo, minimum stomatal resistance, relative fraction of
roots in each soil layer, etc.] are assigned for each vegetation tile.
  ST   Evapotranspiration is obtained through the Penman–Monteith equation.
STW = STi−1 − abs ( Ptotal − PET ) ×  i−1   (3.10) Total AET include evaporation from the canopy layer, transpiration from each
  STC  
vegetation tile, and evaporation from bare soil, weighted by their respective
where STi−1 is the soil moisture storage for the previous month and STC is the surface cover area fractions. For each land cover type, there is one canopy
soil moisture storage capacity, estimated to be 150 mm. If Ptotal + STW < PET, layer. It intercepts rainfall according to the Biosphere-atmosphere transfer
then water deficit is calculated as PET − AET; if Ptotal > PET, then AET = PET scheme parameterization as a function of LAI [Dickinson et al., 1986]. There
and water in excess of PET replenishes ST. When ST > STC, the excess water can be multiple soil layers for each land cover type. Soil characteristics
is deemed surplus (S) and eventually runs off. (e.g., texture, hydraulic conductivity, etc.) are constant within each grid, even
Runoff Generation Runoff (RO) is generated from S at a specified rate of though vegetation characteristics may vary. In a three-layer representation,
0.5, which determines the fraction of surplus that becomes runoff for a month the two top layers are designed to represent the dynamic response of soil to the
and the excess is carried over to the next month. The total monthly runoff infiltrated rainfall; if the middle layer is wetter, diffusion is permitted from
RO total = DRO + RO. the latter to the upper layer. The third layer receives moisture from the middle
Steenhuis and Van Der Molen (1986) extended the Thornthwaite–Mather layer through gravity drainage following the Brooks–Corey relationship for
procedure to use daily input values and account for delay caused by percola- unsaturated hydraulic conductivity (Brooks and Corey, 1964). Runoff from
tion through the unsaturated zone in calculating recharge from soil moisture the bottom layer is modeled according to drainage description in the Arno
balance. A computer program to compute soil water balance with spatial and model (Franchini and Pacciani, 1991).
temporal variations in groundwater recharge in a geographic information For each time step, soil moisture distribution, infiltration, drainage between
system setting has been developed by Westenbroek et al. (2010). soil layers, surface runoff, and subsurface runoff are calculated for each land
cover tile. For each grid cell, total heat fluxes (latent heat, sensible heat, and
Soil and Water Assessment Tool ground heat) effective surface temperature, and total surface and subsurface
The Soil and Water Assessment Tool (SWAT; Arnold et al., 1998; http://swat runoffs are obtained by aggregating over the land cover tiles weighted by their
.tamu.edu) is a modeling effort of the USDA Agricultural Research Service. It fractional coverage. Each grid is modeled independently without any horizontal
has been in continuous development for over 30 years to meet the growing water flow and runoff is obtained for each grid cell. Runoff is not uniform
and diverse needs of the hydrological, environmental, and water resources across the cell. A stand-alone routing model collects and transport the grid-
management community. SWAT is a physically based continuous-time model based surface runoff and base flow toward the outlet of each grid. Water in the
that operates on a daily time step and is capable of predicting the impact of channel is no longer included in the water budget accounting procedure. Once
water management on sediment and other nonpoint source pollutions, and runoff exits a cell, a channel routing, based on the Saint–Venant equation,
agricultural yields in ungagged watersheds. The main components of the simulated drainage and discharge at the basin outlet. A detailed description of
model are: weather, hydrology, soil temperature and properties, plant growth, the model, including a historical overview, complete set of algorithms, calibra-
nutrients, pesticides, bacteria and pathogens, and land management. tion, and validation procedures, is available in Gao et al., (2009).
The hydrologic modeling is carried out at the watershed or subwatershed
scale. A basin is either subdivided into smaller subbasins or can be subdivided Noah
down to hydrologic response units (HRUs), characterized by their prevailing The community Noah [the acronym comes from the first letter of the follow-
land use, soil type, and management characteristics. The model requires daily ing four organizations: the National Centers for Environmental Protection,
precipitation, minimum and maximum temperature, solar radiation, relative the Oregon State University (Department of Atmospheric Sciences), the Air
humidity, and wind speed as inputs. The average air temperature determines Force, and the Hydrologic Research Lab—NWS (now Office of Hydrologic
if precipitation should be in the form of snow or not. Evapotranspiration can Development) land surface model (Chen et al., 1996; Koren et al., 1999; http://
be computed via either Penman–Monteith (Monteith, 1965), Priestly–Taylor www.ral.ucar.edu/research/land/technology/lsm.php) is a stand-alone one-
(Priestley and Taylor, 1972), or Hargreaves (Hargreaves et al., 1985) equa- dimensional column model. It can simulate soil moisture (both liquid and
tions. Hydrologic simulations and water balance at the HRU scale includes frozen), soil temperature, skin temperature, snowpack depth, snowpack water
canopy interception of precipitation, partitioning of precipitation, snowmelt, equivalent, canopy water content, and the water and energy flux terms of the
and irrigation water between surface runoff and infiltration, redistribution of surface water and energy balance (Mitchell, 2005).
water within the soil profile, evapotranspiration, lateral subsurface flow from The model has a snow layer and a canopy layer. The snow layer simulates
the soil profile, and return flow from shallow aquifers. A number of routines snow accumulation, sublimation, melting, and heat exchange at snow-
are available in SWAT to simulate crop yields, biomass outputs, nutrients and atmosphere and snow-soil interfaces. Precipitation is deemed snow if the
pesticide applications, and conservation and water-management practices. temperature of the lowest atmospheric layer is less than 0°C. The surface skin
Flow from all HRUs are aggregated to the subbasin level and routed using temperature is determined by applying a single linearized surface energy
either the variable-rate storage method or the Muskingum method. Sediment, balance equation (Mahrt and Ek, 1984). The ground heat flux is governed by
nutrient, pesticide, and bacterial loadings or concentrations are also aggre- the diffusion equation for soil temperature. The governing equations for the
gated at the HRU level and routed through channels, reservoirs, and other physical processes of the soil-vegetation-snowpack medium are solved
water bodies within the basin. through a finite difference spatial discretization and the Crank–Nicolson time
SWAT has been widely adopted across the world. Among others, it is being integration scheme. The total evaporation, in the absence of snow, is the sum
used in total maximum daily load (TMDL) analyses and to support conserva- of evaporation of precipitation that has been intercepted by the plant canopy,
tion efforts in the United States (U.S.). A complete review of the SWAT model transpiration from the roots and canopy of vegetation, and direct evaporation
and its applications is available in Gassman et al. (2007). from the topmost soil layer. Potential evaporation is calculated using
Penman-based energy balance and includes a stability-dependent aerody-
Variable Infiltration Capacity Model namic resistance after Mahrt and Ek (1984). The wet canopy evaporation
The variable infiltration capacity model (VIC: Liang et al., 1994; http:// follows Noilhan and Planton (1989) and Jacquemin and Noilhan (1990) and
www.hydro.washington.edu/Lettenmaier/Models/VIC/index.shtml) is a evaporation from bare soil follows Mahfouf and Noilhan (1991).
03_Singh_ch03_p3.1-3.12.indd 6 8/22/16 11:21 AM

Natural and Anthropogenic Effects on the Water Balance 3-7
Figure 3.5 Effect of ENSO around the world. [Source: National Oceanic and Atmospheric Administration (2012).]
The soil profile extends to a depth of 2 m, divided into four layers (0–0.1, The effect of ENSO is neither distributed uniformly across the globe, not even
0.1–0.4, 0.4–1, and 1–2 m) from the ground surface to the bottom. The root across a single continent, nor in time during its period of occurrence (Lyon,
zone is confined to the upper 1 m, while the lower layer acts as a reservoir 2004). As ENSO cycles through its different phases (El Niño, La Niña, and
with gravitational free-drainage at the bottom (Chen and Dudhia, 2001). The neutral), its impacts vary both spatially and temporally (Fig. 3.5). The two
volumetric soil moisture content is determined using the diffusive form of hydrological variables that are most affected during ENSO are precipitation
Richard’s equation, which is derived from Darcy’s law under the assumption and temperature. Some regions may experience above-average rainfall, lead-
of rigid, isotropic, homogeneous, and 1D vertical flow. Surface runoff is the ing to devastating floods, while others may endure extended periods of pre-
excess after infiltration (Schaake et al., 1996). cipitation deficit leading to excruciating droughts, the drying of biomass, and
Noah has been tested and validated in both coupled and uncoupled modes forest fires.
(Mitchell, 2005). A complete description of the model physics and computa- ENSO affects water balance mostly in the southeastern part of Africa and
tional algorithm is available in Chen and Dudhia (2001) and Grunmann Madagascar and the eastern part along the equator. Dry and warm conditions
(2005). are felt during El Niño from December through February in the southeastern
part, while the eastern part benefits from higher precipitation. The opposite
pattern is noted during La Niña. The influence of ENSO is visible in the sea-
3.5 NATURAL AND ANTHROPOGENIC EFFECTS ON THE
sonal vegetation dynamics as highlighted in the normalized difference vegeta-
WATER BALANCE
tion index (NDVI) (Philippon et al., 2014).
The hydrological cycle and each of the components of the water balance is The hydrology of Asia and Australia can be seriously affected during
sensitive to local changes happening within the watershed and to changes in ENSO. El Niño causes abnormally warm and dry conditions during December
the climate. These changes are noticeable at various spatial and temporal through February over the region extending from India to northern Australia.
scales. The most immediate changes are noticed when land use land cover is Dry conditions extend further south affecting hydrological conditions in the
affected. Natural variations in the climate and changes due to anthropogenic eastern part of Australia, leading to an increase in the risk of bush fires (Fuller
contribution of greenhouse gases and aerosols in the atmosphere are felt on and Murphy, 2006; Williams and Karoly, 1999). The Indian summer mon-
most hydro-meteorological variables and affect the water balance. soon, which lasts from June to September and accounts for up to 80% of the
rainfall in the region, is negatively correlated with ENSO—the region receives
3.5.1 Natural: Climate Variability less than average rainfall during El Niño events (Kirtman and Shukla, 2000).
The most common drivers of climate variability include large-scale circula- Delays in the onset of the monsoon and changes in the amount and pattern
tion patterns, sunspots, and volcanic eruptions. The former is the most of precipitation affects soil moisture and may lead to agricultural drought,
important variable and can have dramatic influence on the climate, weather, and has a strong impact on several major rivers [e.g., the Huang He (Yellow)
and other hydro-meteorological variables. The lag and duration between the and Yangtze Rivers in China, the transboundary Mekong, Ganges-
onset of these phenomena and their impacts locally can range from a few Brahmaputra-Megna, Indus, etc.] fed from the Himalayan glaciers (Thompson
months to years, and at times, even decades. A number of climate variability et al., 2000). Nearly 1.5 billion people in the most populous region in the
patterns have been identified. Some exert negligible influence on the hydro- world depends on those rivers for sustenance.
meteorological cycle, while others can drastically affect water balance. The Given the geographical distance separating the central Pacific Ocean and
extent and severity with which they influence local conditions is often a func- Europe, the influence of ENSO is hardly noticeable on the hydrology of the
tion of the geographical distance between the watershed of interest and the European regions (Fraedrich, 1990, 1994). El Niño (La Niña) episodes have
poles of the climate teleconnection patterns. been associated with enhanced cyclonic (anticyclonic) activities. El Niño has
The most important large-scale circulation pattern is the El Niño-Southern also been associated with more variable winter conditions, which affect snow-
Oscillation (ENSO). ENSO is a coupled ocean-atmosphere phenomenon fall, and ultimately streamflow in snow fed rivers. Correlation between ENSO
related to sea surface temperature anomalies in the central and eastern equa- and mass of glaciers in the European Alps is almost negligible at the yearly
torial Pacific and the associated sea-level pressure difference, the Southern timescale (Durand et al., 2009), but weak correlations may be detected on a
Oscillation (SO). ENSO has a recurrence pattern of 3–6 years and every event multidecadal scale (Efthymiadis et al., 2007).
normally lasts for around a year. El Niño events are often, but not always, ENSO alters the path of the jet stream, influencing weather and storm
followed by La Niña events, also referred to as the cold phase of ENSO. tracks on the North American continent. A dip in the position of the jet
03_Singh_ch03_p3.1-3.12.indd 7 8/22/16 11:21 AM

3-8 Water Balance
stream, during El Niño, results in warmer winters in western Canada and Changes in the vegetation type or crop cycle also affect the hydrology of a
southern Alaska and northeastern U.S., while the southern U.S. experiences watershed by altering evapotranspiration regime. The conversion of range-
cooler and wetter conditions. California experiences above average precipita- land to cropland, for example, has been found to enhance recharge (Scanlon
tion, which can often lead to flash floods (Mo and Higgins, 1998; Schonher et al., 2005). A decrease in evapotranspiration and increase in streamflow and
and Nicholson, 1989). Influx of excess moisture during El Niño also increases baseflow was observed when cropping systems changed from perennial to
snowfall in the Sierra Nevada mountain range (Kunkel and Angel, 1999). The annual (Schilling et al., 2008).
southern part of Mexico, on the other hand, experiences warmer and drier Infrastructure related to urbanization, such as the construction of roads,
summers. During La Niña, the dip in the jet stream shifts west of its normal form barriers, modify drainage patterns, and disrupt the movement of water
position toward the Central Pacific resulting in cooler winters in eastern within the basin and thus affect water balance. Other anthropogenic factors,
Canada, while the weather conditions in southern U.S. are warm and dry. such as water abstraction from surface and groundwater sources, diversions
Water availability in some sections of the Rio Grande basin, which extends and interbasin water transfers, and discharge of effluent, also affect the water-
from the Rockies in Colorado to the Gulf of Mexico, can increase by over shed scale water balance. Dams alter flow regimes and affect both the
300% and decrease by nearly 100% during the warm and cold phases of upstream and downstream hydrology of the watershed—marked impacts on
ENSO, respectively (Khedun et al., 2012). flow, sediment, aquatic and riparian ecology, and the environment are visible
In South America, El Niño has been associated with severe drought in especially downstream, almost immediately after the construction of a dam,
Mexico and most of Brazil, while Argentina, Paraguay, and Uruguay have and changes in the channel geomorphology becomes apparent in the long
experienced increased precipitation. The coastal areas of Ecuador, northern term (Lu and Siew, 2006; Mathias Kondolf and Batalla, 2005; Singer, 2007).
Peru, and southern Chile also benefit from above-average precipitation, while Around the world, between 4510 and 5330 km3 of water, equivalent to
snowfall in the mountainous Andes decreases, causing the glaciers to retreat. 11–13% of total annual river runoff (~41,000 km3/year), is impounded
During La Niña, however, the region can receive up to four times more pre- behind major dams (Gornitz, 2000).
cipitation (Haylock et al., 2006). Extreme precipitation events have also been
noted during ENSO (Grimm and Tedeschi, 2009). Groundwater Abstraction
ENSO is not the only climate variability pattern to influence water balance. The component of precipitation that drains through the soil matrix and does
Other climate teleconnection patterns, such as the Pacific Decadal Oscillation not evaporate or transpire back to the atmosphere through plants becomes
(PDO), the North Atlantic Oscillation, the Atlantic Multidecadal Oscillation, groundwater. This water is stored in aquifers. Aquifers are classified as uncon-
the Quasi Biennial Oscillation, the El Niño Modoki, etc., have been found to fined, confined, and fossil aquifers. Unconfined aquifers do not have a confin-
influence the hydrological cycle, either alone, or in conjunction with ENSO. ing upper layer and their upper boundary is the water table. They are often
In the U.S., for example, PDO has been found to modulate the effect of ENSO, shallow and are recharged by water percolating through the unsaturated zone
where El Niño (La Niña) during the positive (negative) phase of the PDO, directly above the aquifer and are thus prone to changes in climatic factors,
may lead to stronger climate responses, than when they are evolving in oppo- especially precipitation and temperature. Confined aquifers are found
site phases (Gershunov and Barnett, 1998; Khedun et al., 2014). between layers of impermeable material both above and below the aquifer.
They are saturated with water and are under pressure. These aquifers are
3.5.2 Anthropogenic Influences recharged by precipitation in the area where the aquifer crops out and has
Earth’s surface has been subject to considerable anthropogenic influences, similar characteristics as an unconfined aquifer. Fossil aquifers have ground-
especially over the last century. Modification of the natural state of the water- water that has remained trapped, as a result of geologic activities, in the rock
shed upsets local level water balance, and as these alterations expand, the formations for thousands or millions of years. Unlike confined and uncon-
country and continental scale hydrological patterns and water balance are fined aquifers, they are not replenished and are thus a nonrenewable resource.
affected. Changes in the climate as well affect the spatiotemporal hydrome- Examples of fossil aquifers include the Ogallala underlying the U.S. Great
teorological variables and thus the water balance. Plains and the Nubian Sandstone Aquifer, one of the largest known fossil
aquifers, located on the Eastern end of the Sahara Desert.
Land Use Land Cover Changes Surface and ground water balance are interlinked. Water balance models
The relationship between climate, land cover, and the hydrological cycle is can be used to estimate the contribution of groundwater to surface water or
complex. Climate determines vegetation type and density of cover, which, in vice versa. Groundwater and surface water interaction comprises natural
turn, influences the hydrology of the watershed—both the above- and sub- exchange fluxes with rivers, fluxes due to groundwater abstraction, aquifer
surface components of the hydrological cycle are affected by land cover char- recharge, and changes in evapotranspiration (Rassam et al., 2012).
acteristics. Vegetation cover intercepts and stores a fraction of precipitation; Groundwater abstraction can be from both replenishable and nonreplenish-
part of which eventually reaches the ground through stemflow and through- able aquifers; it is being mined to meet domestic, industrial, and agricultural
fall, while some evaporates back into the atmosphere. A higher proportion of uses. Globally, groundwater contributes 43% (545 km3/year) of the total
rainfall is intercepted during light- or short-duration events compared to high consumptive irrigation water use (Siebert et al., 2010) and has played a key
intensity or long duration ones. Other vegetation characteristics, such as leaf role in containing the food crisis. In several agricultural areas around the
area, leaf density, and orientation also determine interception storage. Part of world, groundwater has become the primary source of irrigation. Satellite
the water that reaches the ground and does percolate deep into groundwater measurements have revealed dramatic loss of groundwater resources in
storage is available for evaporation. Some evaporates directly from the soil northern India (Rodell et al., 2009), western U.S. (Famiglietti et al., 2011), and
surface, while others transpire through the plant cover. Evapotranspiration is elsewhere (Famiglietti, 2014). Groundwater depletion is even more apparent
influenced by plant type, growth stage and growth cycle, and evaporation during extended droughts (Long et al., 2013).
demand of the atmosphere, which is controlled by temperature, humidity, and Therefore, groundwater, especially fossil water, is an important component
wind speed. The complex nature of the interaction between precipitation with of the hydrological cycle and any water balance exercise. It can be deemed an
vegetation and its impact on local hydrology is extensively discussed in the input to the water balance equation, just like precipitation, and accounted as
literature (e.g., Crockford and Richardson, 2000; Llorens and Domingo, 2007; a flux of water vapor, through evaporation and evapotranspiration, into the
Martinez-Meza and Whitford, 1996). atmosphere. Recent studies (DeAngelis et al., 2010; Puma and Cook, 2010)
Disruption of the vegetation cover, for example, when forests are replaced have shown that irrigation can significantly increase downwind precipitation.
with agriculture, urbanization, etc., directly impacts the hydrology and water Groundwater also, subsequently, increases the volume and rate of streamflows
balance of the watershed. Infiltration capacity of the soil and evapotranspira- to the ocean and has been found to be partly responsible for sea-level rise
tion are reduced, and surface runoff and streamflow increase when forests are (Gornitz, 2000; Konikow, 2015; Pokhrel et al., 2012).
cleared, and the reverse trend is observed when cleared areas are reforested.
Urbanization further increases the area of impermeable surfaces, again inhib- Climate Change
iting the infiltration of precipitation and reduces evapotranspiration leading Earth’s climate is constantly changing. Evidence from tree rings and other
to greater runoff and streamflow. The magnitude of these changes depends on paleo-climatological records indicate that changes in the climate are also
the watershed characteristics, the prevailing climate, and spatial area impacted reflected in the hydrological system at both spatial and temporal scales (Gray
by land use land cover changes. Results from a study based in the upper et al., 2004; Gray and McCabe, 2010). Global warming, due to anthropogenic
Midwest United States showed that the conversion of forest to cropland emissions of greenhouse gases, a result of industrialization, is having a dis-
resulted in a decrease in evapotranspiration by 2.5% and an increase in runoff cernible influence on the physical and biological systems across the globe. An
by 20% and baseflow by 4%, while urbanization caused a decrease in evapo- increase in the amount of energy on Earth’s surface results in an intensifica-
transpiration and baseflow by 70 and 38%, respectively while surface runoff tion of the hydrological cycle, which disrupts every component of the water
increased by 1200% (Mishra et al., 2010). balance.
03_Singh_ch03_p3.1-3.12.indd 8 8/22/16 11:21 AM

REFERENCES 3-9
A warmer planet implies that the atmosphere has a higher water holding Brutsaert, W., Hydrology: An Introduction, Cambridge University Press,
capacity, as given by the Clausius–Clapeyron equation New York, 2005.
Budyko, M. I., The Heat Balance of the Earth’s Surface (translated from
des Lv (T ) es Teplovoi balans zemnoi poverkhnosti.Rep.), Weather Bureau, Washington,
= (3.11)
dT RvT 2 1956.
Calvo, J. C., “An evaluation of Thornthwaite’s water balance technique in
where es is the saturation vapor pressure, T is the temperature, Lv is the latent predicting stream runoff in Costa Rica,” Hydrological Sciences Journal, 31 (1):
heat of vaporization, and Rv is the gas constant for water vapor (Shelton, 51–60, 1986, doi:10.1080/02626668609491027.
2009). A higher atmospheric concentration of water vapor, a greenhouse gas, Chen, F. and J. Dudhia, “Coupling an advanced land surface–hydrology
results in the capture of more of the outgoing radiation from the Earth’s sur- model with the Penn State–NCAR MM5 modeling system. Part I: model
face, thus leading to even further warming. implementation and sensitivity,” Monthly Weather Review, 129 (4): 569–585,
Warmer temperatures enhances evapotranspiration, which results in lower 2001, doi:10.1175/1520-0493(2001)129<0569:CAALSH>2.0.CO;2.
soil moisture for plants and a decrease in groundwater recharge and reduced Chen, F., K. Mitchell, J. Schaake, Y. Xue, H-L. Pan, V. Koren, Q. Y. Duan,
baseflow, and therefore, less water in the streams, rivers, lakes, and wetlands. et al., “Modeling of land surface evaporation by four schemes and comparison
It also affects snow accumulation and enhances the melting of land-based with FIFE observations,” Journal of Geophysical Research, 101 (D3): 7251–7268,
glaciers resulting in changes in the timing and volume of runoff and stream- 1996, doi:10.1029/95jd02165.
flow in snow and glacier-fed catchments. On the other hand, an increase in Cisneros, B. E., T. Oki, N. W. Arnell, G. Benito, J. G. Cogley, P. Döll, T. Jiang,
evaporation from large water bodies such as lakes and oceans implies more et al., Freshwater resources, Climate Change 2014: Impacts, Adaptation, and
water in the atmosphere and hence an increase in global precipitation. Higher Vulnerability. Part A: Global and Sectoral Aspects. Contribution of Working
precipitation results in an increase in the soil moisture and recharge compo- Group II to the Fifth Assessment Report of the Intergovernmental Panel on
nent of the water balance equation, where the increase in precipitation is Climate Change, edited by C.B. Field, V.R. Barros, D.J. Dokken, K.J. Mach,
greater than that in evapotranspiration. This complex interaction between the M.D. Mastrandrea, T.E. Bilir, M. Chatterjee, K.L. Ebi, Y.O. Estrada, R.C.
climate and the hydrological cycle, and the redistribution of water in the water Genova, B. Girma, E.S. Kissel, A.N. Levy, S. MacCracken,, P.R. Mastrandrea,
balance equation, has been explored through GCMs. L.L. White, et al., Cambridge University Press, Cambridge, United Kingdom,
The Intergovernmental Panel on Climate Change (IPCC; Cisneros et al., and New York, 2014, pp. 229–269.
2014) states that as the climate changes, precipitation across the globe is pro- Crockford, R. H. and D. P. Richardson, “Partitioning of rainfall into
jected to increase. However, this increase will not be equally distributed throughfall, stemflow and interception: effect of forest type, ground cover and
across the planet. Watersheds located in high latitudes and equatorial Pacific climate,” Hydrological Processes, 14 (16–17): 2903-2920, 2011, doi:10.1002/
are likely to experience an increase in mean annual precipitation, while those 1099-1085(200011/12)14:16/17<2903::AID-HYP126>3.0.CO;2-6.
located in the mid-latitude and subtropical dry regions are likely to suffer a DeAngelis, A., F. Dominguez, Y. Fan, A. Robock, M. D. Kustu, and
decrease in mean precipitation. On the other hand, mean precipitation in D. Robinson, “Evidence of enhanced precipitation due to irrigation over the
catchments located in several mid-latitude wet regions is likely to increase. Great Plains of the United States,” Journal of Geophysical Research: Atmospheres,
Soil moisture in dry regions is likely to further decrease, hence increasing the 115 (D15): D15115, 2010, doi:10.1029/2010JD013892.
frequency of meteorological and agricultural droughts, leading to less surface Dickinson, R. E., A. Henderson-Sellers, P. J. Kennedy, and M. F. Wilson,
and groundwater. Biosphere-atmosphere Transfer Scheme (BATS) for the NCAR Community
Intensification of the hydrological cycle also implies changes in the fre- Climate Model, Document No. NCAR Technical Note NCAR/TN-275-+STR,
quency, intensity, spatial extent, duration, and timing of extreme precipitation 1986. doi: 10.5065/D6668B58.
events. Changes at both the lower tail (e.g., reduction in precipitation leading Durand, Y., G. Giraud, M. Laternser, P. Etchevers, L. Mérindol, and
to droughts) and the upper tail (e.g., high-intensity rainfall resulting in B. Lesaffre, “Reanalysis of 47 years of climate in the French Alps (1958–2005):
floods) of range of observed values can be expected. These changes may be in climatology and trends for snow cover,” Journal of Applied Meteorology and
the following three ways: (i) a shift in the mean which will result in less low- Climatology, 48 (12): 2487–2512, 2009, doi:10.1175/2009JAMC1810.1.
magnitude events and more high-magnitude events; (ii) an increase in stan- Efthymiadis, D., P. D. Jones, K. R. Briffa, R. Böhm, and M. Maugeri,
dard deviation and thus variability which equates with more low- and “Influence of large-scale atmospheric circulation on climate variability in the
high-magnitude events; and (iii) a change in the shape of the distribution, Greater Alpine Region of Europe,” Journal of Geophysical Resources, 112
where low magnitude events remain almost constant while high magnitude (D12): D12104, 2007, doi:10.1029/2006jd008021.
events increases (IPCC, 2012). Famiglietti, J. S., “The global groundwater crisis,” Nature Climate Change,
4 (11): 945–948, 2014, doi:10.1038/nclimate2425.
3.6 CONCLUSION
Famiglietti, J. S., M. Lo, S. L. Ho, J. Bethune, K. J. Anderson, T. H. Syed,
et al., “Satellites measure recent rates of groundwater depletion in California’s
Hydrological processes are complex. Each process plays an important role in Central Valley,” Geophysical Research Letters, 38 (3): L03403, 2011, doi:10.1029/
the water balance—both at the spatial and temporal scales. Natural factors, 2010GL046442.
such as climate variability patterns, and anthropogenic factors, such as land use FAO, AQUASTAT database, Food and Agriculture Organization of the
land cover and climate change, influence the hydrological cycle and hence United Nations (FAO), 2015. url: http://www.fao.org/nr/water/aquastat/
affect, to varying degrees, the water balance components. A wide range of main/index.stm Accessed on: 23 May 2015.
models, ranging from the one-parameter simple bucket model to more com- Fraedrich, K., “European grosswetter during the warm and cold extremes
plex ones, aimed at realistically representing each process, has been developed. of the El Niño/Southern Oscillation,” International Journal of Climatology,
These models help in our understanding of the role of water and its interac- 10 (1): 21–31, 1990, doi:10.1002/joc.3370100104.
tions within the hydrological cycle and with other physico-chemical and Fraedrich, K., “An ENSO impact on Europe?,” Tellus A, 46 (4): 541–552,
biological processes in other elemental cycles. Insights into the movement of 1994, doi:10.1034/j.1600-0870.1994.00015.x.
water and its vulnerabilities can help in the assessment and mitigation of the Franchini, M. and M. Pacciani, “Comparative analysis of several conceptual
impacts of natural and anthropogenic factors on water availability. rainfall-runoff models,” Journal of Hydrology, 122 (1–4): 161–219, 1991,
doi:http://dx.doi.org/10.1016/0022-1694(91)90178-K.
REFERENCES
Fuller, D. and K. Murphy, “The ENSO-Fire dynamic in insular southeast
Asia,” Climatic Change, 74 (4): 435–455, 2006, doi:10.1007/s10584-006-
Arnold, J. G., R. Srinivasan, R. S. Muttiah, and J. R. Williams, “Large area 0432-5.
hydrologic modeling and assessment. Part I: model development,” Journal of Gao, H., Q. Tang, X. Shi, C. Zhu, T. Bohn, F. Su, J. Sheffield, et al., Water
the American Water Resources Association, 34 (1), 73–89, 1998, doi:10.1111/ Budget Record from Variable Infiltration Capacity (VIC) Model Algorithm
j.1752-1688.1998.tb05961.x. Theoretical Basis Document, 2009.
Black, P. E., The Thornthwaite Water Budget in APL, SUNY College of Gassman, P. W., M. R. Reyes, C. H. Green, and J. G. Arnold, “The soil and
Environmental Science and Forestry, Syracuse, New York, 1981. water assessment tool: historical development, applications, and future
Black, P. E., “Revisiting the Thornthwaite and Mather water balance,” research directions,” Transations of the ASABE, 50 (4): 1211–1250, 2007.
Journal of the American Water Resources Association, 43 (6), 1604–1605, 2007, Gershunov, A. and T. P. Barnett, “ENSO influence on intraseasonal extreme
doi:10.1111/j.1752-1688.2007.00132.x. rainfall and temperature frequencies in the contiguous United States: obser-
Brooks, R. H. and A. T. Corey, Hydraulic Properties of Porous MediaRep., vations and model results,” Journal of Climate, 11 (7): 1575–1586, 1998,
Colorado State University, Fort Collins, Colorado, 1964. doi:10.1175/1520-0442(1998)011<1575:EIOIER>2.0.CO;2.
03_Singh_ch03_p3.1-3.12.indd 9 8/22/16 11:21 AM

3-10 Water Balance
Gornitz, V., Impoundment, groundwater mining, and other hydrologic Loaiciga, H. A., J. B. Valdes, R. Vogel, J. Garvey, and H. Schwarz, “Global
transformations: impacts on global sea level rise, Sea Level Rise: History and warming and the hydrologic cycle,” Journal of Hydrology, 174 (1–2): 83–127,
Consequences, edited by B. C. Douglas, M. S. Kearney, and S. P. Leatherman, 1996, doi:http://dx.doi.org/10.1016/0022-1694(95)02753-X.
Academic Press, San Diego, California, 2000, pp. 97–119. Long, D., B. R. Scanlon, L. Longuevergne, A. Y. Sun, D. N. Fernando, and
Gray, S. T., C. L. Fastie, S. T. Jackson, and J. L. Betancourt, “Tree-ring-based H. Save, “GRACE satellite monitoring of large depletion in water storage in
reconstruction of precipitation in the Bighorn Basin, Wyoming, since 1260 a.d.” response to the 2011 drought in Texas,” Geophysical Research Letters, 40 (13):
Journal of Climate, 17 (19): 3855–3865, 2004, doi:10.1175/1520-0442(2004) 3395–3401, 2013, doi:10.1002/grl.50655.
017<3855:TROPIT>2.0.CO;2. Lu, X. X. and R. Y. Siew, “Water discharge and sediment flux changes over
Gray, S. T. and G. J. McCabe, “A combined water balance and tree ring the past decades in the Lower Mekong River: possible impacts of the Chinese
approach to understanding the potential hydrologic effects of climate change dams,” Hydrology and Earth System Sciences, 10 (2): 181–195, 2006, doi:10.5194/
in the central Rocky Mountain region,” Water Resources Research, 46 (5): hess-10-181-2006.
W05513, 2010, doi:10.1029/2008WR007650. Lyon, B., “The strength of El Niño and the spatial extent of tropical
Grimm, A. M. and R. G. Tedeschi, “ENSO and extreme rainfall events in drought,” Geophysical Research Letter, 31 (21): L21204, 2004, doi:10.1029/
South America,” Journal of Climate, 22 (7): 1589–1609, 2009, doi:10.1175/ 2004gl020901.
2008JCLI2429.1. Mahfouf, J. F. and J. Noilhan, “Comparative study of various formulations of
Grunmann, P. J., Variational Data Assimilation of Soil Moisture Information, evaporations from bare soil using in situ data,” Journal of Applied Meteorology,
University of Maryland, College Park, 2005. 30 (9): 1354–1365, 1991, doi:10.1175/1520-0450(1991)030<1354:CSOVFO>
Hamon, W. R., “Estimating potential evapotranspiration,” Journal of the 2.0.CO;2.
Hydraulics Division, Proceedings of the American Society of Civil Engineers, 87: Mahrt, L. and M. Ek, “The influence of atmospheric stability on potential
107–120, 1961. evaporation,” Journal of Climate and Applied Meteorology, 23 (2): 222–234,
Hargreaves, G. L., G. H. Hargreaves, and J. P. Riley, “Agricultural benefits 1984, doi:10.1175/1520-0450(1984)023<0222:TIOASO>2.0.CO;2.
for Senegal river basin,” Journal of Irrigation and Drainage Engineering, 111 Manabe, S., “Climate and the ocean circulation. I. The atmospheric circula-
(2): 113–124, 1985, doi:10.1061/(asce)0733-9437(1985)111:2(113). tion and the hydrology of the Earth’s surface,” Monthly Weather Review, 97 (11):
Haylock, M. R., Peterson, T. C., Alves, L. M., Ambrizzi, T., Anunciação, Y. 739–774, 1969, doi:10.1175/1520-0493(1969)097<0739:CATOC>2.3.CO;2.
M. T., Baez, J, Barros, V. R., et al., “Trends in total and extreme South Martinez-Meza, E. and W. G. Whitford, “Stemflow, throughfall and chan-
American rainfall in 1960–2000 and links with sea surface temperature,” nelization of stemflow by roots in three Chihuahuan desert shrubs,” Journal
Journal of Climate, 19 (8): 1490–1512, 2006, doi:10.1175/JCLI3695.1. of Arid Environments, 32 (3): 271–287, 1996, doi:http://dx.doi.org/10.1006/
IPCC, Managing the risks of extreme events and disasters to advance cli- jare.1996.0023.
mate change adaptation. A Special Report of Working Groups I and II of the Mathias Kondolf, G. and R. J. Batalla, Hydrological effects of dams and
Intergovernmental Panel on Climate Change, Cambridge, UK, New York, water diversions on rivers of Mediterranean-climate regions: examples from
2012, p. 582. California, Developments in Earth Surface Processes, edited by G. Celso and J.
Jacquemin, B. and J. Noilhan, “Sensitivity study and validation of a land B. Ramon, Elsevier, 2005, Chap. 11, pp. 197–211, doi:http://dx.doi.
surface parameterization using the HAPEX-MOBILHY data set,” Boundary- org/10.1016/S0928-2025(05)80017-3.
Layer Meteorology, 52 (1): 93–134, 1990, doi:10.1007/bf00123180. McCabe, G. J. and S. L. Markstrom, A Monthly Water-Balance Model
Jasechko, S., Z. D. Sharp, J. J. Gibson, S. J. Birks, Y. Yi, and P. J. Fawcett, Driven By a Graphical User Interface, U.S. Department of Interior, U.S.
“Terrestrial water fluxes dominated by transpiration,” Nature, 496 (7445): Geological Survey, Reston, VA, USA, 2007, p. 6.
347–350, 2013, doi:10.1038/nature11983. http://www.nature.com/nature/ Mishra, V., K. A. Cherkauer, D. Niyogi, M. Lei, B. C. Pijanowski, D. K. Ray,
journal/v496/n7445/abs/nature11983.html (supplementary-information). L. C. Bowling, et al., “A regional scale assessment of land use/land cover and
Khedun, C. P., A. K. Mishra, J. D. Bolten, H. K. Beaudoing, R. A. Kaiser, J. climatic changes on water and energy cycle in the upper Midwest United
R. Giardino, and V. P. Singh, “Understanding changes in water availability in States,” International Journal of Climatology, 30 (13): 2025–2044, 2010,
the Rio Grande/Río Bravo del Norte basin under the influence of large-scale doi:10.1002/joc.2095.
circulation indices using the Noah land surface model,” Journal of Geophysical Mitchell, K. E., The Community Noah land-surface model. User’s Guide,
Research: Atmospheres, 117 (D5): D05104, 2012, doi:10.1029/2011JD016590. Public Release Version 2.7.1, 2005. url: http://www.ral.ucar.edu/research/land/
Khedun, C. P., A. K. Mishra, V. P. Singh, and J. R. Giardino, “A copula-based technology/lsm/noah/Noah_LSM_USERGUIDE_2.7.1.pdf.
precipitation forecasting model: Investigating the interdecadal modulation of Mo, K. C. and R. W. Higgins, “Tropical influences on California precipita-
ENSO’s impacts on monthly precipitation,” Water Resources Research, 50: tion,” Journal of Climate, 11 (3): 412–430, 1998, doi:10.1175/1520-0442(1998)
580–600, 2014, doi:10.1002/2013WR013763. 011<0412:TIOCP>2.0.CO;2.
Kirtman, B. P. and J. Shukla, “Influence of the Indian summer monsoon on Monteith, J. L., “Evaporation and the environment,” paper presented at 19th
ENSO,” Quarterly Journal of the Royal Meteorological Society, 126 (562): Symposium: Society of Experimental Biology: The State and Movement of Wate
213–239, 2000, doi:10.1002/qj.49712656211. in Living Organisms, Cambridge University Press, Swansea, UK, 1965.
Kobayashi, T., R. Iwanaga, J. I. Teshima, D. Ikegami, D. Yasutake, W. He, and Mussy, A. and C. Higy, Hydrology. A Science of Nature, Science Publishers,
H. Cho, “An improved bucket with a bottom hole model of soil hydrology setting Enfield, NH, USA, 2011.
in bioprocesses,” 1st International Conference BIOHYDROLOGY 2006, Impact of National Oceanic and Atmospheric Administration, National Weather
Biological Factors on Soil Hydrology, Prague, Czech Republic, 2006. Service, Weather Impacts of ENSO, Jetstream—Online School for Weather,
Kobayashi, T., S. Matsuda, H. Nagai, and J. I. Tesima, “A bucket with a bot- 2012. url: http://www.srh.noaa.gov/jetstream/tropics/enso_impacts.htm.
tom hole (BBH) model of soil hydrology,” Soil-Vegetation-Atmosphere Transfer Accessed on 27 July 2016.
Schemes and Large-Scale Hydrological Models. Sixth IAHS Scientific Assembly, Noilhan, J. and S. Planton, “A simple parameterization of land surface pro-
edited by A. J. Dolman, A. J. Hall, M. L. Kavvas, T. Oki, and J. W. Pomeroy, cesses for meteorological models,” Monthly Weather Review, 117 (3): 536–549,
IAHS Press, Maastricht, The Netherlands, 2001. 1989, doi:10.1175/1520-0493(1989)117<0536:ASPOLS>2.0.CO;2.
Konikow, L. F., “Long-term groundwater depletion in the United States,” Oki, T. and S. Kanae, “Global hydrological cycles and world water resources,”
Groundwater, 53 (1): 2–9, 2015, doi:10.1111/gwat.12306. Science, 313 (5790): 1068–1072, 2006, doi:10.1126/science.1128845.
Koren, V., J. Schaake, K. Mitchell, Q. Y. Duan, F. Chen, and J. M. Baker, “A Philippon, N., N. Martiny, P. Camberlin, M. T. Hoffman, and V. Gond,
parameterization of snowpack and frozen ground intended for NCEP weather “Timing and patterns of the ENSO signal in Africa over the last 30 years:
and climate models,” Journal of Geophysical Research, 104 (D16): 19569–19585, insights from normalized difference vegetation index data,” Journal of Climate,
1999, doi:10.1029/1999jd900232. 27 (7): 2509–2532, 2014, doi:10.1175/JCLI-D-13-00365.1.
Kunkel, K. E. and J. R. Angel, “Relationship of ENSO to snowfall and Pokhrel, Y. N., N. Hanasaki, P. J. F. Yeh, T. J. Yamada, S. Kanae, and T. Oki,
related cyclone activity in the contiguous United States,” Journal of Geophysical “Model estimates of sea-level change due to anthropogenic impacts on terrestrial
Research, 104 (D16): 19425–19434, 1999, doi:10.1029/1999jd900010. water storage,” Nature Geoscience, 5 (6): 389–392, 2012, doi:http://www.nature.
Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges, “A simple hydro- com/ngeo/journal/v5/n6/abs/ngeo1476.html (supplementary information).
logically based model of land surface water and energy fluxes for general Priestley, C. H. B. and R. J. Taylor, “On the assessment of surface heat flux
circulation models,” Journal of Geophysical Research: Atmospheres, 99 (D7): and evaporation using large-scale parameters,” Monthly Weather Review, 100
14415–14428, 1994, doi:10.1029/94JD00483. (2): 81–92, 1972, doi:10.1175/1520-0493(1972)100<0081:OTAOSH>2.3.CO;2.
Llorens, P. and F. Domingo, “Rainfall partitioning by vegetation under Puma, M. J. and B. I. Cook, “Effects of irrigation on global climate during
Mediterranean conditions. A review of studies in Europe,” Journal of Hydrology, the 20th century,” Journal of Geophysical Research: Atmospheres, 115 (D16):
335 (1–2): 37–54, 2007, doi:http://dx.doi.org/10.1016/j.jhydrol.2006.10.032. D16120, 2010, doi:10.1029/2010JD014122.
03_Singh_ch03_p3.1-3.12.indd 10 8/22/16 11:21 AM

REFERENCES 3-11
Rassam, D. W., I. Jolly, and T. Pickett, Guidelines for Modelling Groundwater- Singh, V. P., Compute Models of Watershed Hydrology. Water Resources
Surface Water Interactions in eWater Source: Towards Best Practice Model Publications, Highlands Ranch, CO, 1995.
Application, eWater Cooperative Research Centre, Bruce, Australia, 2012. Singh, V. P. and D. K. Frevert, Mathematical Models of Small Watershed
Rodell, M., I. Velicogna, and J. S. Famiglietti, “Satellite-based estimates of Hydrology and Applications. Water Resources Publications, Highlands Ranch,
groundwater depletion in India,” Nature, 460 (7258): 999–1002, 2009. CO, 2002a.
Scanlon, B. R., R. C. Reedy, D. A. Stonestrom, D. E. Prudic, and K. F. Singh, V. P. and D. K. Frevert, Mathematical Models of Large Watershed
Dennehy, “Impact of land use and land cover change on groundwater Hydrology. Water Resources Publications, Highlands Ranch, CO, 2002b.
recharge and quality in the southwestern US,” Global Change Biology, 11 (10): Singh, V. P. and D. A. Woolhiser, “Mathematical modeling of watereshed
1577–1593, 2005, doi:10.1111/j.1365-2486.2005.01026.x. hydrology,” Journal of Hydrologioc Engineering, 7 (4): 270–292, 2002.
Schaake, J. C., V. I. Koren, Q-Y. Duan, K. Mitchell, and F. Chen, “Simple Steenhuis, T. S. and W. H. Van Der Molen, “The Thornthwaite-Mather
water balance model for estimating runoff at different spatial and temporal procedure as a simple engineering method to predict recharge,” Journal of
scales,” Journal of Geophysical Research: Atmospheres, 101 (D3): 7461–7475, Hydrology, 84 (3–4): 221–229, 1986, doi:http://dx.doi.org/10.1016/0022-
1996, doi:10.1029/95JD02892. 1694(86)90124-1.
Schilling, K. E., M. K. Jha, Y-K. Zhang, P. W. Gassman, and C. F. Wolter, Thompson, L. G., T. Yao, E. Mosley-Thompson, M. E. Davis, K. A.
“Impact of land use and land cover change on the water balance of a large Henderson, and P.-N. Lin, “A high-resolution millennial record of the South
agricultural watershed: historical effects and future directions,” Water Asian monsoon from Himalayan ice cores,” Science, 289 (5486): 1916–1919,
Resources Research, 44 (7): W00A09, 2008, doi:10.1029/2007WR006644. 2000, doi:10.1126/science.289.5486.1916.
Schonher, T. and S. E. Nicholson, “The relationship between California Thornthwaite, C. W., “The climates of North America: according to a new
rainfall and ENSO events,” Journal of Climate, 2 (11): 1258–1269, 1989, classification,” Geographical Review, 21 (4): 633–655, 1931, doi:10.2307/209372.
doi:10.1175/1520-0442(1989)002<1258:TRBCRA>2.0.CO;2. Thornthwaite, C. W., “The climates of the Earth,” Geographical Review, 23
Sellers, P., Modeling and observing land-surface-atmosphere interactions (3): 433–440, 1933, doi:10.2307/209629.
on large scales, Land Surface-Atmosphere Interactions for Climate Modeling— Thornthwaite, C. W., “An approach toward a rational classification of cli-
Observations, Models and Analysis, edited by E. F. Wood, Kluwer Academic mate,” Geographical Review, 38 (1): 55–94, 1948.
Publishers, The Netherlands, 1991. Thornthwaite, C. W. and J. R. Mather, “The water balance,” Publications in
Shelton, M. L., Hydroclimatology: Perspectives and Applications, Cambridge Climatology, 8 (1): 86, 1955.
University Press, Cambridge, UK, 2009. Thornthwaite, C. W. and J. R. Mather, “Instructions and tables for comput-
Shiklomanov, I. A., World fresh water resources, Water in Crisis. A Guide ing potential evapotranspiration and the water balance,” Publications in
to the World’s Fresh Water Resources, edited by P. H. Gleick, Oxford University Climatology, 10 (3): 127, 1957.
Press, New York, 1993, pp. 13–24. US EPA. BASINS 4.1 (Better Assessment Science Integrating point & Non-point
Shiklomanov, I. A., World Water Resources. A New Appraisal and Assessment Sources) Modeling Framework - Tutorials and Training, Lecture 5 Hydrologic
for the 21st Century, United Nations Educational, Scientific and Cultural Processes. National Exposure Research Laboratory, RTP, North Carolina. url:
Organization, Paris, France, 1998. https://www.epa.gov/exposure-assessment-models/basins-tutorials-and-training.
Siebert, S., J. Burke, J. M. Faures, K. Frenken, J. Hoogeveen, P. Döll, and 2015.
F. T. Portmann, “Groundwater use for irrigation—a global inventory,” Westenbroek, S. M., V. A. Kelson, W. R. Dripps, R. J. Hunt, and K. R. Bradbury,
Hydrology and Earth System Sciences, 14 (10): 1863–1880, 2010, doi:10.5194/ SWB—A Modified Thornthwaite-Mather Soil-Water-Balance Code for Estimating
hess-14-1863-2010. Groundwater Recharge, U.S. Geological Survey, Reston, VA, USA, 2010, p. 60.
Singer, M. B., “The influence of major dams on hydrology through the Williams, A. A. J. and D. J. Karoly, “Extreme fire weather in Australia and
drainage network of the Sacramento River basin, California,” River Research the impact of the El Niño-Southern Oscillation,” Australian Meteorological
and Applications, 23 (1): 55–72, 2007, doi:10.1002/rra.968. Magazine, 48: 15–22, 1999.
03_Singh_ch03_p3.1-3.12.indd 11 8/22/16 11:21 AM

PART 2
DATA COLLECTION AND
PROCESSING
04_Singh_ch04_p4.1-4.8.indd 1 8/22/16 11:22 AM

04_Singh_ch04_p4.1-4.8.indd 2 8/22/16 11:22 AM
Chapter 4
Hydrometeors and Quantitative
Precipitation Estimation
BY
BEKKI WARD HARJO
ABSTRACT that it covers a limited area, so precipitation occurs between gaging sites. Thus
gages may not always sample rainfall maxima and minima, which are crucial
Hydrometeorological data are collected using radars, gages, satellites, and
to a good hydrometeorological analysis.
human observations. Quantitative precipitation estimates are developed from
multisensor analysis and are sometimes supplemented with climatological 4.2.2 Remotely Sensed Measurements
data. Automated algorithms can help identify and estimate hydrometeors, but
Ground-based radars can provide estimation of the phase, quantity, and ele-
human-performed quality control is still needed to ensure a high-quality final
vation of hydrometeors in the atmosphere. Satellites can provide visual
estimate.
images, thermal imagery, and platforms for radiometers to help determine the
quantity and phase of hydrometeors and also show the extent of snow and ice
4.1 INTRODUCTION after an event has occurred.
4.1.1 Precipitation Phases 4.2.3 Ground-Based Indirect Measurements
Hydrometeors are various forms of water in the atmosphere. Hydrometeors Indirect measurements of hydrometeorological data are used to re-create the
in liquid form include rain, dew, fog, drizzle, and cloud droplets. Solid hydro- historical estimates of rain and snow. Tree rings can be used as a proxy of
meteors include snow, hail, ice crystals, and graupel. Mixed-phase hydrome- annual precipitation in many locations. Striations in glaciers can also be used
teors include wet snow and water-covered hail. Most hydrometeors are large as a surrogate measurement of long-term annual precipitation.
enough to fall as precipitation, typically formed when ice crystals in the
clouds grow big enough to overcome updrafts in the clouds and begin falling
as snow. A much smaller percentage of precipitation forms as liquid drops 4.3 REMOTE SENSING OF PRECIPITATION
and grows by the process of collision and coalescence when ambient cloud- 4.3.1 Doppler Radars—Single Polarization
level temperatures are above freezing.
Real-time hydrometeorological data are collected with a network of Doppler
4.1.2 Impacts at Surface of Hydrometeor Types radars. These Doppler radars utilize radar pulses to estimate the size, type,
and Temperatures location, and movement of hydrometeors in the atmosphere. Doppler radar is
The impacts at the earth’s surface can be significantly different based on the preferred sensor for determining the timing and location of precipitation.
the temperature of the hydrometeor and the temperature of the air near the But as will be discussed later in this chapter, the amount of estimated pre-
earth’s surface, especially when the temperature is near freezing. A difference cipitation can at times be in considerable error. Doppler radars have different
of 1°C or 2°C in temperature of the air, the hydrometeor, or the ground operational modes, including clear air and precipitation modes.
surface can mean the difference between an ice storm (with ice buildup on Different radar wavelengths are used worldwide. Longer wavelengths are
trees and power lines) and a very cold rain. It is possible for rain to remain beneficial to obtaining adequate radar coverage in convective precipitation
supercooled liquid when it falls through the atmosphere through slightly and severe weather. Shorter wavelengths can provide uniform radar coverage
below freezing air temperatures, but once the droplets hit the earth or another for widespread precipitation patterns over a smaller range than radars with
surface, if the temperature of the surface is below freezing, the water droplets longer wavelengths due to attenuation issues.
will quickly freeze and cause a buildup of ice on the surface. The measurement that radar makes of the strength of the radar pulse after
returning from hitting hydrometeors or other objects is called reflectivity. The
reflectivity radar product (Z) is the most important single polarization radar
4.2 TYPES OF HYDROMETEOROLOGICAL DATA product conveying hydrometeorological information. Relationships between
the reflectivity and the rainfall rates have long been established. Areas of
4.2.1 Ground-Based Direct Measurements extremely high reflectivity indicate that heavy rainfall or hail is falling, while
Ground-based measurements provide a ground truth when analyzing radar- snow typically shows as bands of low reflectivity.
based precipitation data at specific points. Direct measurement instruments The power of returned radar pulses is used to calculate a reflectivity factor
include rain gages, snow pillows, and hail pads, which all measure the quantity (Z) using the Probert–Jones equation. The Z factor is then used in various
of hydrometeors once they have fallen to the ground. Snow pillows measure the equations (often referred to as Z–R power relationships) to calculate a rainfall
weight of the fallen snow and convert it to a liquid equivalent. Hail pads show rate, R. The strength of the return signal to the radar is determined by the
the approximate size and representative quantity of hail that fell at a location. quantity and size of hydrometeors within the radar beam. Different Z–R
Rain gages at the Earth’s surface can provide point measurements of the relationships can be used based on the meteorological conditions. The equa-
amount of precipitation that has fallen. However, rain gage density is such tion Z = 300 R1.4 is used as the default equation over much of the United States,
4-3
04_Singh_ch04_p4.1-4.8.indd 3 8/22/16 11:22 AM

4-4 Hydrometeors and Quantitative Precipitation Estimation
but the equation Z = 250 R1.2 is used for tropical rain processes. Large num- 4.3.4 Satellites
bers of Z–R relationships have been experimentally tested over the years and With little ability to obtain gage data or coverage from land-based radars,
additional research is still being conducted to determine appropriate Z–R satellite information is particularly valuable over the oceans and countries
relationship for different locations, climates, and atmospheric conditions. such as Mexico with limited surface-based sensors. Recent technological
There are automated algorithms that can estimate the precipitation based advances in satellites allow for the use of multiple radiometers using multiple
on the radar products alone or in combination with automated rain gages. algorithms to estimate the quantity of hydrometeors. Satellites estimate pre-
Cumulative rainfall products can also be automatically produced in a suite of cipitation based on thermal imagery and microwave radiometers. As the
radar products, typically on a 1-h, 3-h, or 24-h, or storm total basis. A single cloud height increases, the temperature of the hydrometeors in the clouds
relationship between the reflectivity and gage data is calculated and can be decreases. The temperature differential within the cloud formation can help
applied to the entire field of the radar, if the radar operator so chooses to survey precipitation formation and severe weather patterns. In general, colder
employ it. This is used to mitigate the shortcomings in radar precipitation cloud tops are correlated with heavier precipitation rates, and warmer cloud
estimates due to atmospheric condition or hydrometeor type. However, it is tops with lighter precipitation rates.
important to note that these unbiased, fully automated products may contain Early satellite programs focused on rainfall measurements on the tropics as
inaccurate precipitation rates because of radar artifacts or limitations in the these latitudes can be easier to obtain adequate coverage from orbit. Current
radar data, such as incomplete radar beam filling, blocked beams, overshooting satellite programs seek to expand the coverage over land. However, there are
beams, range degradation, melting layer, hail, or a nonstandard meteorological current limitations with satellite measurements at higher latitudes and over
atmosphere. land. Higher latitude measurements by geostationary satellites are hindered
4.3.2 Doppler Radars—Dual Polarization
by a high angle of the Earth’s surface from the satellite. Satellite observations
over land are affected by increased albedo in snow-covered areas and local-
Doppler radars with dual polarization send out radar pulses in both the ized phenomena such as mountains that alter cloud formation. Due to these
horizontal and vertical directions. This technology can help distinguish the limitations, satellite precipitation estimates are generally less skillful than
types of hydrometeors and other objects found in the radar beam by measur- those derived from radar and have larger areal displacement and quantitative
ing their length and width. Based on the reflectivity, differential reflectivity errors. However, satellite-derived precipitation estimates are useful in multi-
(the measurement of the strength and frequency shift of both the horizontal sensor precipitation algorithms, especially in areas lacking radar coverage or
and vertical radar pulse returns), velocity, correlation coefficient (the mea- ground-based rain gage observations.
surement of how similarly the horizontally and vertically polarized pulses are
behaving within a pulse volume), and specific differential phase (KDP) (the
4.4 HYDROMETEOROLOGICAL DATA PROCESSING
phase shift of both the horizontal and vertical radar pulse returns) products,
the object detected is classified into one of 10 different classifications: Hail, 4.4.1 Gridded Multisensor Precipitation Estimates
Graupel, Big Drops, Heavy Rain, Rain, Wet Snow, Dry Snow, Ice Crystals, In an effort to mitigate radar deficiencies, a gridded quantitative precipitation
Ground Clutter, and Biological. Being able to quickly identify and classify estimate, using multiple sensors, is performed. For a hydrometeorological
the types of hydrometeors makes dual-polarization radar a more valuable analysis of large areas, a mosaic of radar coverage is used in combination with
tool over the single-phase Doppler radar technology. The difference between rain gages, snow pillows, and other information. In areas without radar cover-
the horizontally polarized return power and the vertically polarized return age, climatological data are generally used to estimate the precipitation that
power helps define the shape of the hydrometeors. Figure 4.1 shows how fell, augmented by available gage data and satellite estimates. Radar precipita-
dual-polarization technology is able to differentiate between different types tion estimates are raised or lowered based on the relationship between the
of hydrometeors based upon the distinctive length and width ratios of the radar and gage readings. Both field and local biases are calculated and either
radar returns. can be applied.
There are several radar products available with dual polarization Doppler In order to minimize accumulative errors, the quality-control step of using
radar that convey hydrometeorological information. Of the dual polarization additional 24-h manual precipitation observations to calibrate the hourly data
fields now available, the KDP product is one of the most useful for determin- from the past day is performed. A daily postanalysis is performed to adjust
ing areas of heavy rain. The phase of the radar pulse shifts as it passes through the hourly precipitation estimates from the previous day. The postanalysis
hydrometeors and back again. The magnitude of this shift is a good indicator works by time distributing the 24-h accumulation into individual 1-h
of the size and density of the precipitation. As with the single polarization amounts based on the estimated precipitation expected to have fallen in each
radar, there are automated algorithms that can estimate the instantaneous individual hour. The multisensor optimal estimation procedure is then re-run
precipitation rate, or the cumulative precipitation based solely on radar for each of the individual hours making up the 24-h total, but including the
products, or using a multisensor analysis with automated rain gages. newly time distributed gage observations.
4.3.3 Phased Array Radars When multiple sensors are used to process hydrometeorological data, vari-
ous methodologies can be used to estimate the contribution of each input. A
Phased array radars are now being used to provide information about current grid is developed by optimally combining data from radars and gages. A
weather conditions, especially about storm structure during severe weather. mosaic of radar coverage is typically developed from lowest level available
In the future, it is expected that phased array radar will be able to better radar coverage. However, it is also possible to distance-weight the radar infor-
provide information about hydrometeors as well. Phased array radars are dif- mation or to utilize the average or maximum value of overlapping radar
ferent from the traditional Doppler radars in that the radar pulses can be coverage to obtain a more accurate estimate in some weather conditions. The
electronically steered to sample certain areas of the atmosphere, instead of the influence of individual gages is also generally distance weighted, unless
entire radar unit physically moving in a defined rotational pattern of differing there is local knowledge to the contrary. In some areas that tend to experience
preset elevation angles. more uniform precipitation patterns and have few gages, a single radar bias
correction is developed from the available gages compared to the radar field.
This relationship, known as a field bias, then is applied to the entire radar field
to obtain precipitation estimates. This radar bias is generally less reliable than
localized biases.
Human input to quantitative precipitation estimates continues to be impor-
tant even as technology advances. Unlike humans, algorithms cannot make
subjective decisions about current meteorological conditions or gage behavior
and have no familiarity with the region’s topography, meteorological pro-
cesses, and areas of insufficient radar coverage. An example of a 24-h gridded
multisensor quantitative precipitation estimate (QPE) incorporating a mosaic
of radar coverage, rain gages, snow pillows, precipitation climatology, human
observations, and other information is shown in Fig. 4.2. The QPE in the top
part of Fig. 4.2 is based only on a mosaic of raw radar average reflectivity. The
circular rings show the radar coverage available from each Doppler radar in
the area. The QPE in the bottom section of Fig. 4.2 includes the radar reflec-
tivity, observed hourly gage data and the biases calculated from these data,
daily human observations, precipitation climatology, and input from the
Figure 4.1 Dual polarization technology helps classify types of hydrometeors. hydrometeorological forecaster’s local knowledge.
04_Singh_ch04_p4.1-4.8.indd 4 8/22/16 11:22 AM

Hydrometeorological Data Quality Assurance and Control 4-5
Figure 4.2 24-Hour radar only and multisensor quantitative precipitation estimates.
4.5 HYDROMETEOROLOGICAL DATA QUALITY artifacts. There are a multitude of items in addition to hydrometeors
ASSURANCE AND CONTROL within a radar beam that can return signal back to the radar unit. And
It is important that hydrometeorological data go through appropriate data as discussed earlier, satellite hydrometeorological data also have limitations
quality control. While the spatial and temporal resolution of radar-based as they do not produce the same accuracy of information as the land-
products is superior to other sensors, there are several limitations of based Doppler radars used in conjunction with gages and climatological
Doppler radar data, ranging from physical limitations to artificial data information.
04_Singh_ch04_p4.1-4.8.indd 5 8/22/16 11:22 AM

4-6 Hydrometeors and Quantitative Precipitation Estimation
4.5.1 Nonprecipitation Radar Echoes 4.5.7 Cone of Silence

Anthropogenic items can appear on radar as ground clutter. While the major- A Doppler radar typically sends out radar beams from angles of 0.5° to 19.5°.
ity of Doppler radars have been built near airports or other open places, Because all the beams are transmitted radially and at preselected angles based
buildings and structures built near radars can block radar beams. Moving on the volumetric scan type chosen for operations, the radar is prevented
man-made items, such as wind turbines and traffic on elevated roadways, can from transmitting beams overhead. Because of the fixed angles of operability,
also return radar signals. Natural obstructions such as mountains can also there will be a cone-shaped area centered over the radar where the radar will
obstruct radar beams. Even far away trees can block radar beams in very flat not be able to see, although returns from hydrometers can still be obtained at
regions. There are algorithms in place within the processing of the radar data, low elevations near the radar.
which mitigate the effects of ground clutter and other nonmeteorological
targets. 4.5.8 Virga
Biological matter can be clearly seen on the radar at times. Birds can be Virga is falling precipitation that evaporates or sublimates before reaching the
picked up on the radar leaving their roosts in the morning and bats can be ground surface. Hydrometeors in virga produce radar returns, but the rain or
seen leaving theirs at night. Insects and small birds in gust fronts make snow does not reach the ground. Virga is common in regions with high tem-
the gust front show as a clearly delineated line across the radar image. peratures and low humidity.
Chaff is a radar countermeasure that consists of small strips of aluminum,
metalized glass fiber, or plastic fiber released into the atmosphere during 4.5.9 Incorrect Z–R Relationships and Algorithm Errors
military maneuvers. The chaff shows up on radars as a streak of high reflectiv- Different types of storms have different meteorology patterns that require
ity. Smoke plumes from fires and ash from volcanic eruptions show up on different Z-R relationships. Using an incorrect Z–R relationship can have a
radars as the smoke and ash are blown away by the wind. detrimental impact on precipitation estimates. Weather with low reflectivity,
such as light snow, often has radar returns too low to be considered hydrome-
4.5.2 Radar Artifacts teors by the algorithm, even though precipitation is falling and reaching the
Certain atmospheric conditions can cause the radar beam to be anomalously ground. The high reflectivity of hail can cause the radar algorithm to calculate
bent back into the ground causing significant ground clutter to show up in a much higher rainfall estimate because of the presence of hail.
radar data. In addition, spurious spikes and radar artifacts occur in Doppler
radar data frequently. A three-body scatter spike is an anomalous radar arti- 4.5.10 Data Errors
fact that can occur in the radar image down radial of areas with large or wet Sampling errors are common in rain and snow measurements. Rainfall gages
hail. Side lobe errors can occur when the returned power signal to the radar may introduce errors such as tipping buckets not keeping pace with heavy
unit is quite high. During the radar volume scans that occur during sunrise, a rainfall. If automated rain gages are not heated, they may report zero precipi-
spurious radar spike will occur to the east of the radar location toward the tation during freezing events. Automated gages may also introduce errors
rising sun. The same phenomenon occurs again at sunset with a spurious such as dew being measured as rainfall or by continuing to drip rain after the
radar spike occurring to the west toward the setting sun. These spikes occur rainfall events have ended because of a partially clogged opening. Completely
because the sun also produces radiation at the same wavelength of the radar. clogged gages report zero precipitation during events. Rain gages require
routine maintenance because they can be totally or partially clogged with bird
4.5.3 Beam Blockage nests, spider webs, leaves, or other debris.
The radar beam can be physically blocked by mountains, buildings, water Snow pillows can have measurement errors due to the mechanical nature
towers, or other items. It can also be blocked by intense precipitation, espe- of the gages. Because snow pillows measure the weight of the snow, changes
cially by significant quantities of large hail. There are times that radar beams in barometric pressure and temperature can cause the gage to measure erro-
are blocked by areas of high reflectivity in the atmosphere, such as severe neous positive or negative values.
storms. Storms with intense hail or rain can essentially block or attenuate The siting of gages to keep appropriate distances from trees and buildings
radar beams from traveling through the storm’s core or lowering the power or can prevent some gage errors. However, strong winds may prevent rain or
the return signal from returning back to the radar unit. snow from falling into rain gages or snow pillows even in open areas. Blowing
snow can also alter snowpack estimates.
4.5.4 Range Degradation and Earth Curvature The consistency of hydrometeorological data is dependent upon the train-
As a radar beam travels, the beam samples higher altitudes due to the Earth’s ing and skill of the human observer. Human observations can differ for a
curvature. Storm systems with low clouds, particularly winter weather events variety of reasons, such as differences in eyesight acuity or the introduction of
like snow, can be undersampled and not show up on radar when they are not numerical rounding into observations. Snowfall observations can be particu-
located near the radar. When radar beams travel farther than the operating larly inconsistent. Observations of snow are reported for inches of depth of
range of the radar, range folding can occur when the returns are confused fresh fallen snow or for older more densely packed snow. Snow depths also
with those returning from within the operating range. vary from place to place due to drifting. In order to mitigate differences due
Atmospheric conditions can also cause changes in the typical range of a to the consolidation of the older snow pack and localized drifting, human
Doppler radar. Subrefraction occurs when the beam bends less than the cur- observations of the water equivalent of the snow that fell over the past
vature of the Earth, resulting in a longer effective radar range, but possibly 24 hours are reported. However, differences occur between observers due to
underestimating cloud heights. Super refraction happens when the radar their methods and equipment when melting the snow to determine the snow
beam bends back toward the Earth’s surface more than the curvature of the water equivalent. Some of the liquid water will evaporate if the snow is melted
Earth. Low-level inversions where the temperature is cooler at the Earth’s too fast.
surface than at higher elevations in the atmosphere can produce super refrac-
tion, especially near dawn and dusk. Ducting is extreme super refraction
where the radar beam bends back toward the Earth’s surface instead of 4.6 QUANTITATIVE PRECIPITATION ESTIMATE, DATA USE,
ARCHIVING, AND ACCESSIBILITY
increasing with elevation. Super refraction tends to show an increase in radar
returns from ground clutter. QPE data are used in weather and climate forecasts, including flash flood
warnings, fire weather forecasting, and droughts. QPE data are crucial in
4.5.5 Nonuniform Beam Filling
impact-based weather forecasts, such as forecasted snow, ice, or freezing rains
In general operations, no radar beam is filled with the exact same size and during rush hour traffic. Some of the usages of QPE data in engineering
type of hydrometeor. Water droplets, hail, and snow exist in various sizes and include hydrologic and hydraulic modeling, flood studies, water resources
shapes. Only the average information of all the information in each area is planning, and development activity. There are multiple planning studies that
returned to the radar. If large wet hail is located in the same area as small use QPE data, such as precipitation recurrence intervals, climate stationarity
raindrops, the average reflectivity for the entire area will be returned back to studies, probable maximum precipitation (PMP), and probable maximum
the radar. This would raise the reflectivity of the entire area, showing up as a flood (PMF) studies. Uses of these planning studies include application in
heavier rain. building codes, development standards for public works, and coastal surge
studies. QPE information is used in reservoir planning and yield studies,
4.5.6 Bright Banding
while short-term QPE data are used in the operations of reservoirs, locks and
The elevation when the snow begins to melt, as it is falling, is detected as an dams, and hydropower facilities.
area of higher reflectivity and the radar algorithm is programmed to estimate Many facets of society use quantitative precipitation estimates, including
this as heavy rain. This melting level can produce a concentric band or circle farmers, ranchers, landscapers, outdoor enthusiasts, and many corporations.
of higher reflectivity known as the “bright band” in the radar products. Corporations use them to correlate sales to weather patterns. Kayakers, skiers,
04_Singh_ch04_p4.1-4.8.indd 6 8/22/16 11:22 AM

Quantitative Precipitation Estimate, Data Use, Archiving, and Accessibility 4-7
hunters, and others use them to plan their outdoor activities. Quantitative QPE data are not useful unless they are available. The period of available
precipitation estimates are used in drought monitoring and irrigation plan- data should be as long as possible as the length of the data archive will affect
ning to maximize crop yields. Estimations of the anticipated streamflow the usefulness of many studies and applications. It is also important that data
resulting from the melt of a mountain snowpack can be estimated in the early be archived in a format that is readily accessible to others, such as Standard
spring to assist farmers and reservoir controllers in calculating the quantity of Hydrometeorological Exchange Format (SHEF), network Common Data
water that will be available for irrigation later during the year from down- Form (netCDF), or WaterXML. As more GIS applications are developed, it is
stream reservoirs. also important that the QPE data be GIS compatible.
04_Singh_ch04_p4.1-4.8.indd 7 8/22/16 11:22 AM

Chapter 5
Streamflow Data
BY
ROBERT R. HOLMES, JR.
ABSTRACT
The importance of streamflow data to the world’s economy, environmental

health, and public safety continues to grow as the population increases. The
collection of streamflow data is often an involved and complicated process.
The quality of streamflow data hinges on such things as site selection, instru-
mentation selection, streamgage maintenance and quality assurance, proper
discharge measurement techniques, and the development and continued
verification of the streamflow rating. This chapter serves only as an overview
of the streamflow data collection process as proper treatment of consider-
ations, techniques, and quality assurance cannot be addressed adequately in
the space limitations of this chapter. Readers with the need for the detailed
information on the streamflow data collection process are referred to the
many references noted in this chapter.
5.1 STREAMFLOW
Streamflow is the volumetric discharge, expressed in volume per unit time

(typically cubic feet per second, ft3/s, or cubic meters per second, m3/s), that
occurs in a stream or channel and varies in time and space. Knowledge of the
quantity and temporal and spatial distribution of streamflow is crucial to the
science of hydrology, emergency management, and associated fields such as Figure 5.1 Streamgage mounted on the downstream pier of a bridge over the
water resources infrastructure design, water resources management, and Gasconade River at Rich Fountain, Missouri, USA, with a hydrographer making a
environmental protection. discharge measurement using an ADCP mounted on a tethered boat during a flood
in August, 2013.
5.2 TYPES OF STREAMFLOW DATA
Streamflow data encompass both water–surface elevation (stage) and volu- This chapter is restricted to discussing only those streamflow data derived
metric discharge (discharge), being collected at various intervals depending from simple stage-discharge or index-velocity ratings. The chapter on
on the need. Often streamflow data are collected through means of perma- streamflow ratings (see Chap. 6), as well as Rantz (1982), describes more
nent or semipermanent facilities installed on or near the stream, which are involved rating methods that utilize water–surface slope and rate of stage
termed streamgages (Fig. 5.1), although other names such as gauges, change for the ratings. Internationally, procedures for collecting stage and
streamflow monitoring stations, and river gages are used to refer to these discharge data can be found in selected publications by the International
installations. Organization for Standardization (ISO) and the World Meteorological
Continuous streamflow data are collected at a frequency sufficient to fully Organization (WMO). In many parts of the world and certainly in the United
characterize the streamflow hydrograph. Typical frequency of continuous States, the methods described in reports by the U.S. Geological Survey
streamflow data at streamgages ranges from 1 to 5 min for small streams (USGS) have been adopted.
(drainage area typically less than 50 km2), which are “flashy” (Fig. 5.2a) to The streamflow rating is developed by correlating discretely measured
60 min for large rivers (drainage area typically greater than 250,000 km2), discharge (discharge measurements) with concurrently collected stream data.
which rise and fall more gradually (Fig. 5.2b). In essence, the autonomously collected stream data act as a surrogate for
discharge by means of the streamflow rating.
5.3 STREAMGAGE OPERATION 5.3.1 Selection of a Streamgage Site
The purpose of a streamgage is to obtain a continuous record of discharge. Selection of the exact location of the streamgage along a reach should maxi-
Currently (2015), there is no widely-accepted, economically-viable method mize the quality of the streamflow ratings following existing guidelines
to directly determine discharge continuously, although research to develop (Rantz, 1982; Levesque and Oberg, 2012; World Meteorological Organization,
methods for this purpose is ongoing (Brakenridge et al., 2012; Fulton and 2010a; Sauer and Turnipseed, 2010; Carter and Davidian, 1968).
Ostrowski, 2008; Costa et al., 2006; Bjerklie et al., 2005). Rather, continuous Site selection involves three major factors. First and foremost, the
time series of discharge are typically computed by applying a streamflow rat- streamgage has to be located such that it can collect reliable and accurate
ing relation (streamflow rating) to autonomously collected stream data, such stream data. Second, the site must permit the making of discrete discharge
as stage, slope of water surface, rate of stage change, or stream index velocity. measurements (discharge measurements) in a safe and efficient manner.
5-1
05_Singh_ch05_p5.1-5.8.indd 1 8/22/16 11:23 AM

5-2 Streamflow Data
(a)
25
Boneyard data
20
Discharge (m3/s) 15
10
0
5/29/2008 5/31/2008 6/2/2008 5/29/2008 6/4/2008 6/6/2008 6/10/2008
0:00 0:00 0:00 0:00 0:00 0:00 0:00
Date and time
(b)
25,000
St. Louis data
20,000
Discharge (m3/s)
15,000
10,000
5000
0
1/31/2008 3/21/2008 5/10/2008 6/29/2008 8/18/2008
0:00 0:00 0:00 0:00 0:00
Date and time
Figure 5.2 Hydrograph for (a) Boneyard Creek near Urbana, Illinois, USA (drainage area 11.6 km2) and (b) Mississippi River at St. Louis, Missouri, USA (drainage area 1,810,000
km2) (Data from U.S. Geological Survey National Water Information System, http://waterdata.usgs.gov/nwis).
Third, the site must be well suited to the strategy chosen for developing and that the concurrency between the autonomously collected stream data and
maintaining the streamflow rating. the discharge measurement is well established. Otherwise, the discharge mea-
Reliability of stream data means that the data are complete and its delivery surements will require adjustment for time of travel (Rantz, 1982).
uninterrupted as near real-time data have proven crucial to efficient water The strategy for developing the streamflow rating has major implications
resources management. Ideally, the streamgage should be located in relatively for the location of the streamgage. A simple rating is used where the relation
calm pools or downstream from protective features to minimize the chances between stage and discharge is single-valued or univariate, meaning the rat-
of damage or failure of the structure, sensors, data logger, telemetry, and ing has only one value of discharge for a particular stage (Fig. 5.3a). For
associated accessories. Care should be taken to elevate moisture-sensitive simple ratings, the relation is controlled by features of the stream (Fig.5.3b).
equipment above the potential maximum flood height. These features are known as “controls” (see Chap. 6 on rating controls) and
Accuracy of stream data requires that the sensors be located so that the the accuracy and stability of the simple rating are directly related to the shape
sensor is reporting the true ambient value for that parameter. For example, in and stability, respectively, of the control. Locating the streamgage where the
the collection of stage, locations for sensor placement should be avoided controls are well-established is important, particularly the low-water (section)
where the water–surface elevation is artificially influenced by localized fea- control as that provides stability for the simple rating at low-water. With the
tures, such as bridge piers, inducing water–surface elevations different than exception of steep mountain streams, the streamgage should be located
the true ambient stream elevation. Additionally, accurate stage data depend upstream from the most stable low-water control that can be found during
on a stable mounting platform for the sensor apparatus. reconnaissance. Controls with cross-section shapes resembling a narrow
The location of the discharge measurements should provide hydraulic “V-shape” will promote increased sensitivity to small changes in stage and
conditions that enable an accurate and safe measurement across the spectrum result in improved accuracy of streamflow estimates. In some cases, it may be
of hydrologic conditions, including flood. The section where the discharge desirable to construct an artificial low-water control to increase stability and
measurement is made must be close enough in proximity to the streamgage ease the determination of a low-water rating.
05_Singh_ch05_p5.1-5.8.indd 2 8/22/16 11:23 AM

Streamgage Operation 5-3
(a) Rating curve
k
Overban
trol
con
el
Stage (m or ft)
nn 1
Cha <2
ol
n contr
Sectio 1
>2
Discharge (m3/s or ft3/s)
(b) Stream cross section
Overbank
Channel control
Section control
Streamgage section control
(c) Stream profile
Streamgage and associated

water-level sensor
Water surface for low Water surface for Water surface for high
flow when section medium flow when flow when channel control
control effective channel control involves both main
effective channel and overbank
Flow
age
amg
Stre ction
se
rol
cont
Figure 5.3 (a) Streamgage simple rating curve with effective controls identified along with the appropriate slope of the rating for those controls, (b) stream cross section view
showing water levels associated with controls, and (c) reach of river showing streamgage in relation to associated controls at various stages.
5.3.2 Methods of Stage Measurement

Where the channel slope is subject to backwater or highly unsteady flow A fundamental part of any streamgage is stage data, the collection of which
conditions, the water–surface slope also may need to be incorporated into the is governed by guidelines and standards to ensure quality data (Sauer and
rating. Backwater conditions often arise due to the presence of a downstream Turnipseed, 2010; World Meteorological Organization, 2010a). The stage of
dam or, perhaps, from variable inflow from a contributing stream. Passage of a stream or river is the height of the water–surface elevation above a par-
a significant flood wave along a low-slope channel may also induce slope ticular local datum, often known as the gage datum. The gage datum is often
effects. In some cases, an auxiliary gage must be established to measure the established such that a zero value is below the channel bed to ensure non-
slope along the reach between it and the primary streamgage, and appropriate negative values at all stages. Gage datum is often surveyed into known geo-
site and logistical evaluations for the auxiliary gage must be included in the detic datum to relate stage to known elevations of adjoining physical
site evaluation process. The fall between the two should exceed 0.5 ft, the features and enable comparison of stages between streamgages. Stage is
slope should be uniform along the entire reach, and the reach should be as far measured in feet or meters, often determined and reported to an accuracy
upstream from the backwater as possible. Unfortunately, the need for an aux- of the nearest one hundredth of a foot or one thousandth of a meter (Sauer
iliary station or for incorporation of other treatments for slope effects may not and Turnipseed, 2010).
be evident until several high-flow measurements documenting the passage of Stage can be determined either by nonrecording manual observation or
flood waves are obtained and plotted against concurrent stage to reveal the through automated sensors attached to recording data loggers. Manual
presence of a loop-rating. Rantz (1982) provides additional details. observation systems include vertical graduated plates (staff gages) attached
05_Singh_ch05_p5.1-5.8.indd 3 8/22/16 11:23 AM

5-4 Streamflow Data
to fixed mounts in the water or weighted tapes or wires in which the observa- the sum of the products of the area and mean velocity of each subsection
tion of stage is made by lowering the weight to the water surface from a (International Organization for Standardization, 2007a)
known elevation and subtracting the length of the wire or tape. The tradi- n
tional automated stage sensor over the years has been the float and counter-
Q = ∑(ai vi ) (5.1)
weight in a stilling well, but new technology has provided an array of options i =1
including submersible and nonsubmersible pressure transducers, optical
sensors, and radio detection and ranging (radar) sensors (Sauer and where Q is the total discharge, ai is the area of subsection i, and vi is the average
Turnipseed, 2010). Selection of the sensor must be determined based on velocity of subsection I. The midsection velocity–area method to determine
economics and site conditions. For example, a float and counterweight in a discharge (Fig. 5.4) is a commonly used cross section discretization method in
stilling well sensor provide excellent data; however, the combination of con- the United States (Turnipseed and Sauer, 2010). With the advent of acoustic
struction cost and sedimentation of the well may mean that a noncontact Doppler current profilers (ADCP) technology, efficient discretization of the
(sensor is mounted above the water) radar stage sensor may provide a better channel cross section into small subsections (both horizontal and vertical dis-
alternative when all factors are considered. Conversely, regular large magni- cretizations) has increased the number of data points collected in each mea-
tude winds or surface ice at a site might induce enough errors in stage measurement and reduced the measurement time. Thousands of velocity and depth
surement to influence the choice of a submersible pressure sensor over a observations (Fig. 5.5) are made as the ADCP travels approximately perpen-
radar stage sensor. dicular to streamflow (Mueller et al., 2009) with the ADCP mounted rigidly
Often the data loggers for automated stage sensor systems are attached to either inside a tethered boat or near the bow of a manned boat (Fig. 5.5).
telemetry, which transmit the data in near real-time to centralized computer Instruments to determine velocity include mechanical current meters
systems. Options for telemetry include satellite, telephone, and radio. (both vertical axis and horizontal axis-type meters), electromagnetic meters,
Telemetered data provide a number of advantages including timely detection acoustic Doppler velocity meters (ADVMs), and optical strobe velocity
and correction of streamgage problems, enhancement of flood warning and meters (Turnipseed and Sauer, 2010; Bureau of Reclamation, 2001).
forecasting, and operation of water resources projects. Mechanical meters have been used around for over 100 years. The predomi-
nant mechanical meter used in the United States today is the AA Price
5.3.3 Methods of Discharge Measurement mechanical current meter, which was first introduced in the nineteenth
Discharge measurements are made using a variety of methods following century (Frazier, 1974); however, the ADVMs have largely supplanted the use
standard protocols (Turnipseed and Sauer, 2010; World Meteorological of mechanical meters by many water resources agencies around the world.
Organization, 2010a; Herschy, 2008; Rantz, 1982; Benson and Dalrymple,
1967) and are grouped into two broad categories: direct and indirect meth- 5.3.4 Methods of Index Velocity Measurement
ods. Direct measurements of discharge use instruments, sensors, or devices Continuous measurement of velocity, termed index velocity, is used at select-
to “directly” measure the discharge on site when the streamflow of interest ed sites to derive streamflow by the index–velocity method (see Chap. 6).
is occurring. Indirect methods are used primarily to measure the peak dis- Four types of metering systems have typically been used to measure index
charge from a flood event after the flood has passed and the stage has velocity (Rantz, 1982): standard mechanical current meters, deflection
receded. Indirect methods utilize forensic surveys of both the remnant high meters, electromagnetic meters, and acoustic meters. Over time, the acoustic
water marks and the channel geometry in concert with the estimates of the velocity meter (AVM) and ADVM have become the predominant tools for
boundary roughness to estimate the peak discharge from the principles of index velocity methods (Levesque and Oberg, 2012). The AVM consists of
conservation of energy, momentum, and mass. A variety of indirect mea- two sensors positioned at an angle to the flow and utilize the acoustic travel
surement techniques exist that can be employed based on the stream time technique (Laenen, 1985). ADVMs, first introduced in 1997, consist of a
hydraulics and presence or absence of structures (Benson and Dalrymple, single unit utilizing the Doppler principal (Morlock et al., 2002).
1967; Matthai, 1967; Bodhaine, 1968; Dalrymple and Benson, 1967; In the United States, the relatively low-cost but robust ADVM has become
Hulsing, 1967). more predominant than the AVM (Levesque and Oberg, 2012). The ADVM
Most of the discharge measurements made at streamgages are direct mea- is mounted in an up-looking orientation to ensonify a vertical volume of
surements based on the velocity–area method, whereby the cross section of water, or it is mounted in a side-looking orientation (Fig. 5.6) to ensonify a
the stream is discretized into subsections (Fig. 5.4) and the total discharge is horizontal volume of water.
b1 b2 b3 b4 b5 b6 Water surface bn–1 bn
1 n
d2 dn–1
d3
2 d4
(n–1)
d5
d6
3
5 Streambed
6
Explanation
Boundary of subsection
1, 2, 3, .......... n Subsection number
b1, b2, b3 ..... bn Width of subsection
d1, d2, d3 ..... dn Depth of subsection
Figure 5.4 For the midsection velocity–area method, the total discharge is determined by the sum of the products of the area and mean velocity of each subsection.
05_Singh_ch05_p5.1-5.8.indd 4 8/22/16 11:23 AM

Streamgage Operation 5-5
ADCP mounted rigidly near

bow of a manned boat
20
Depth (m)
19
28
37
0 250 500 750 1000
Lateral distance (m)
0.004 1.051 2.098 3.145 4.192

Velocity magnitude (m/s)
Figure 5.5 Photo of data collection with associated data output from a manned-boat ADCP measurement for the Mississippi River below Cairo, Illinois, USA, on May 1, 2011.
Bridge pier
Horizontally ensonified volume
Vertically ensonified volume
Figure 5.6 Diagram of the ensonified volume from a horizontally- and vertically-mounted ADVM.
05_Singh_ch05_p5.1-5.8.indd 5 8/22/16 11:23 AM

5-6 Streamflow Data
5.4 QUALITY ASSURANCE OF STREAMGAGE DATA 5.5 DERIVED STREAMFLOW UNCERTAINTY
The quality of streamgage data is directly dependent on (1) regular field Uncertainty in computed (derived) streamflow is a product of the uncer-
inspections and maintenance of streamgages, (2) collection of quality dis- tainty from three primary sources: streamgage data (stage, index velocity, and
charge measurements through time and throughout the range of stage, and slope), discharge measurements, and streamflow rating. These three sources
(3) proper construction, application, and verification of streamflow ratings. have their many sources of uncertainty. For example, discharge measurement
High-quality stage data are obtained by adhering to standard techniques sources of uncertainty include the uncertainty in measurement of cross-
and quality assurance measures (Sauer and Turnipseed, 2010; World sectional area, uncertainty in measurement of water velocity, uncertainty in
Meteorological Organization, 2010a; Herschy, 2008). At each streamgage, an computation assumptions, and random and systematic errors (Turnipseed
independent nonrecording reference gage should be established, and its eleva- and Sauer, 2010). Each of these in turn arise from other sources of
tion should be related to the local gage datum. This datum should be verified uncertainty; for example, the uncertainty in cross sectional area arises from
by regular differential surveys and tied to multiple stable reference marks uncertainty in width and depth measurements. Clearly, uncertainty has a
away from the immediate vicinity of the streamgage (Kenney, 2010). Regular cascading effect.
field visits to the streamgage, augmented by intermittent high-streamflow Although much work has been done to quantify the uncertainty of dis-
visits, should be made to verify the recording gage against the reference gage charge measurements (Cohn et al., 2013; International Organization for
and ensure the working condition of the equipment. Corrections to the prin- Standardization, 2007b; Oberg and Mueller, 2007; Herschy, 1999; Pelletier,
ciple gage of as little as 0.003 m are made if the low-water control is a perma- 1989) and the uncertainty of streamflow ratings used to derive continuous
nent weir (Kennedy, 1983). streamflow data (Morlot and others, 2014; Schmidt, 2002; Clarke, 1999;
Accuracy of streamflow data begins with high-quality stage data. The fol- Masson and others, 1987), no quantitative standard for determination of
lowing are stage-accuracy standards of the USGS as given by Sauer and uncertainty of derived streamflow has yet to be agreed upon. The USGS
Turnipseed (2010): assigns a qualitative assessment of the general accuracy of the derived stream-
flow based on the amount of missing stage data, the quality of the stage data,
“..the overall accuracy of stage data ….is either 0.01 feet [3.05 mm] or 0.2 percent of
the hydrographer-assigned accuracy rating of discharge measurements, and
the effective stage, whichever is greater. For example, the required accuracy would be
0.06 ft [18.3 mm] at an effective stage of 30 ft [9.14 m], 0.02 ft [6 mm] at 10 ft the stability of the streamflow rating. An accuracy assignment of “excellent”
[3.05 m], and 0.01 ft [3.048 mm] at all effective stages less than 7.5 ft [2.29 m]. means that 95% of the daily mean values of streamflow are correct within 5%;
Effective stage is defined as the height of the water surface above the orifice, intake, “good” is within 10%; “fair” is with within 15%; and “poor” is less than “fair”
or other point of exposure of the sensor in the water body.” (Kennedy, 1983).
The International Organization for Standardization (2008) does not give a
standard for stage data, but rather outlines three performance classes of water 5.5.1 Streamflow Data Dissemination and Archival
level measurement devices based on their nominal uncertainty. Nominal Dissemination and archival of real-time and historic data are important
uncertainty of less than ±0.1%, less than ±0.3%, and less than ±1% of the for management, emergency response, and design for water and environ-
range of stage correspond to performance classes 1, 2, and 3, respectively. mental resources. Real-time streamflow data are often needed for timely
The discharge measurements are crucial to establishing the streamflow rat- operational decisions by water resources and emergency managers. Data
ing. Discharge measurements should be made both through time and across are often telemetered by way of satellite radio, telephone, or terrestrial
the range of stage using proper protocols and techniques (Turnipseed and radio to central locations to enable real-time decisions. These data are
Sauer, 2010; World Meteorological Organization, 2010a; International often concurrently broadcast and made available via the World Wide Web
Organization for Standardization, 2007a; Herschy, 2008). These techniques (Table 5.1).
include a thorough review of the measurement notes and associated electronic Historic data are important to enable contextual understanding of stream-
files and records to ensure that prevailing hydraulic conditions are consistent flow probabilities, rainfall-runoff characteristics, and stream hydraulics.
with the assumptions on which the measurement computations are based. These data are sometimes also available on the World Wide Web (Table 5.1).
Once the streamflow rating is established, continued discharge measure- Although certain countries of the world have collected streamflow data for
ments are important to either verify or adjust the streamflow rating. The decades, their data have not always been made widely available. Web portals
hydrographer’s judgment will determine the required density and frequency such as the World Hydrological Cycle Observing System (WHYCOS) and
of discharge measurements based on the stability of the streamflow rating Global Runoff Data Centre (GRDC) both operated by the World
(Kennedy, 1983, 1984). Proper development and maintenance of the stream- Meteorological Organization (Table 5.1) exist to promote free exchange of
flow rating using standard protocols and techniques (World Meteorological streamflow data around the world and particularly in those countries who do
Organization, 2010b; Kennedy, 1983) are crucial to the derivation of a con- not have centralized capability to serve their data. Value-added synthesis of
tinuous time series of discharge. Development of streamflow ratings is streamflow data such as USGS WaterWatch (Fig. 5.7) provides powerful map-
addressed in more detail in Chap. 6. based hydrologic situational awareness.
Table 5.1 Sources of Streamflow Data and Assessments

Region Agency/Organization Type of information/Data Data age Web location of streamflow data
Canada Environment Canada Water Office Stage and Discharge Real Time and Historic http://wateroffice.ec.gc.ca/index_e.html
United Kingdom Environment Agency Stage Real Time http://apps.environment-agency.gov.uk/
river-and-sea-levels/default.aspx
United States U.S. Army Corps of Engineers River Gages Network Stage Real Time and Historic http://rivergages.com
United States U.S. Bureau of Reclamation Water Operations Reservoir Releases and Real Time http://www.usbr.gov/main/water/
Volumes
United States U.S. Geological Survey National Water Stage and Discharge Real time and Historic http://water.usgs.gov
Information System
United States U.S. Geological Survey WaterWatch Assessment Maps Real Time and Historic http://waterwatch.usgs.gov/
World World Meteorological Organization GRDC Stage and Discharge Historic http://www.bafg.de/GRDC/EN/Home/
homepage_node.html
World World Meteorological Organization WHYCOS Stage and Discharge Real Time and Historic http://www.whycos.org/whycos/
05_Singh_ch05_p5.1-5.8.indd 6 8/22/16 11:23 AM

REFERENCES 5-7
Explanation—Percentile classes
<10 10–24 25–75 76–90 >90

Low High
Much below Below Normal Above Much above
normal normal normal normal
Figure 5.7 WaterWatch assessment of hydrologic condition across regions of the United States on May 16, 2014 (http://waterwatch.usgs.gov/).
REFERENCES Dalrymple, T. and M. A. Benson, Measurement of Peak Discharge by the

Slope-Area Method, U.S. Geological Survey Techniques of Water-Resources
Benson, M. A. and T. Dalrymple, General field and office procedures for
Investigations Book 3, Reston, Virginia, 1967, Chap. A2, p. 12.
indirect discharge measurements, U.S. Geological Survey Techniques of Water-
Frazier, A. H, Water Current Meters in the Smithsonian Studies in History
Resources Investigations Book 3, Reston, Virginia, 1967, Chap. A1, p. 12.
and Technology, No. 28, Smithsonian Institution Press, Washington, D.C.,
Bjerklie, D. M., D. Moller, L. C. Smith, and S. L. Dingman, “Estimating
1974, p. 95.
discharge in rivers using remotely sensed hydraulic information,” Journal of
Fulton, J. and J. Ostrowski, “Measuring real-time streamflow using emerg-
Hydrology, 309: 191–209, 2005.
ing technologies: radar, hydroacoustics, and the probability concept,” Journal
Bodhain, G. L., Measurement of peak discharge at culverts by indirect
of Hydrology, 357: 1–10, 2008.
methods, U.S. Geological Survey Techniques of Water-Resources Investigations
Herschy, R. W., Streamflow Measurement, 3rd ed., Taylor and Francis
Book 3, Reston, Virginia, 1968, Chap. A3, p. 60.
Group, United Kingdom, 2008, p. 507.
Brakenridge, G. R., S. Cohen, A. J. Kettner, T. De Groeve, S. V. Nghiem, J.
Herschy, R. W., Uncertainties in hydrometric measurements, Hydrometry,
P. M. Syvitski, and B. M. Fekerte, “Calibration of satellite measurements of
2nd ed. John Wiley, New York, 1999, pp. 355–370.
river discharge using a global hydrology model,” Journal of Hydrology, 475:
Hulsing, H., Measurement of peak discharge at dams by indirect method,
123–136, 2012.
U.S. Geological Survey Techniques of Water-Resources Investigations Book 3,
Bureau of Reclamation, Water Measurement Manual, Water Resources
Reston, Virginia, 1967, Chap. A5, p. 29.
Technical Publication, Washington, D.C., 2001, http://www.usbr.gov/pmts/
Kennedy, E. J. Discharge ratings at gaging stations, U.S. Geological Survey
hydraulics_lab/pubs/wmm/.
Techniques of Water-Resources Investigations Book 3, Reston, Virginia, 1984,
Carter, R. W. and J. Davidian, General procedure for gaging streams, U.S.
Chap. A10, p. 59.
Geological Survey Techniques of Water-Resources Investigations Book 3, Reston,
Kennedy, E. J., Computation of continuous records of streamflow, U.S.
Virginia, 1968, Chap. A6, p. 13.
Geological Survey Techniques of Water-Resources Investigations Book 3, Reston,
Clarke, R. T., “Uncertainty in the estimation of mean annual flood due to
rating-curve indefinition,” Journal of Hydrology, 222: 1–4, 1999.
Kenney, T. A., Levels at gaging stations, U.S.Geological Survey Techniques
Cohn, T. A., J. E. Kiang, and R. R. Mason, Jr., “Estimating discharge mea-
and Methods Book 3, Reston, Virginia, 2010, Chap. A19, p. 60.
surement uncertainty using the interpolated variance estimator,” ASCE Jour-
International Organization for Standardization, Hydrometry—Water Level
nal of Hydraulic Engineering, 139: 502–510, 2013.
Measuring Devices, ISO 4373:2008, Geneva, Switzerland, 2008, p. 20.
Costa, J. E., R. T. Cheng, F. P. Haeni, N. Melcher, K. R. Spicer, E. Hayes, W.
International Organization for Standardization, Hydrometry—Measurement
Plant, et al., “Use of radars to monitor stream discharge by noncontact meth-
of Liquid Flow in Open Channels Using Current Meters or Floats, ISO 748:2007,
ods,” Water Resources Research, 42: 1–14, 2006.
Geneva, Switzerland, 2007a, p. 58.
05_Singh_ch05_p5.1-5.8.indd 7 8/22/16 11:23 AM

5-8 Streamflow Data
International Organization for Standardization, Hydrometry—Velocity- Mueller, D. S., C. R. Wagner, M. S. Rehmel, K. A. Oberg, and F. Rainviulle,
Area Method Using Current Meters—Collection and Processing of Data for Measuring discharge with acoustic Doppler current profilers from a moving
Determination of Uncertainties in Flow Measurement, ISO 1088, Geneva, boat, U.S. Geological Survey Techniques and Methods Book 3, Reston, Virginia,
Switzerland, 2007b. 2009, Chap. A22, p. 95.
International Organization for Standardization, Measurement of Liquid Oberg, K. A. and D. S. Mueller, “Validation of streamflow measurements
Flow in Open Channel—Water Level Measuring Devices, ISO 4373–1979, made with acoustic Doppler current profilers,” ASCE Journal of Hydraulic
Geneva, Switzerland, 1979, p. 18. Engineering, 133: 1421–1432, 2007.
Laenen, A., Acoustic velocity meter systems, U.S. Geological Survey Tech- Pelletier, P. M., “Uncertainties in streamflow measurement under winter ice
niques of Water-Resources Investigations Book 3, Reston, Virginia, 1985, Chap. conditions, a case study: The Red River at Emerson, Manitoba, Canada,” Water
A17, p. 38. Resources Research, 25 (8): 1857–1868, 1989.
Levesque, V. A. and K. A. Oberg, Computing discharge using the index Rantz, S. E., Measurement and computation of streamflow, U.S. Geological
velocity method, U.S. Geological Survey Techniques and Methods Book Survey Water-Supply Paper 2175, 1982, p. 631.
3, Reston, Virginia, 2012, Chap. A23, p. 148. Sauer V. B. and D. P. Turnipseed, Stage measurement at gaging stations, U.
Masson, J., M. Ghio, C. Lallement, C. Parsy, and J. Philippe, Debitmetrie: S. Geological Survey Techniques and Methods Book 3, Reston, Virginia, 2010,
Precision des Stations de Jaugeage, La Houille Blanche, France, 1987, Chap. A7, p. 87.
pp. 333–338. Schmidt, A. R., Analysis of stage-discharge relations for open-channel
Matthai, H. F., Measurement of peak discharge at width contractions by flows and their associated uncertainties, Ph.D. Thesis, University of Illinois,
indirect methods, U.S. Geological Survey Techniques of Water-Resources Inves- Urbana-Champaign, 2002, p. 329.
tigations Book 3, Reston, Virginia, 1967, Chap. A4, p. 44. Turnipseed, D. P. and V. B. Sauer, Discharge measurements at gaging sta-
Morlock, S. E., H. T. Nguyen, and J. H. Ross, “Feasibility of acoustic Dop- tions, U.S. Geological Survey Techniques and Methods Book 3, Reston,
pler velocity meters for production of discharge records from U.S. Geological Virginia, 2010, Chap. A8, p. 87.
Survey streamflow-gaging stations,” U.S. Geological Survey Water Resources World Meteorological Organization, Field work, Manual on Stream Gauging,
Investigations Report 01-4157, Reston, Virginia, 2002, p. 56. World Meteorological Organization Number 1044-V1, 2010a, Vol. 1, p. 252.
Morlot, T., C. Perret, A. C. Favre, and J. Jalbert, “Dynamic rating curve World Meteorological Organization, Computation of discharge, Manual on
assessment for hydrometric stations and computation of the associated uncer- Stream Gauging, World Meteorological Organization Number 1044-V2,
tainties: quality and station management indicators,” Journal of Hydrology, 2010b, Vol. 2, p. 198.
517: pp. 173–186, 2014.
05_Singh_ch05_p5.1-5.8.indd 8 8/22/16 11:23 AM

Chapter 6
Streamflow Ratings
BY
ROBERT R. HOLMES, JR.
ABSTRACT 6.2 RATING CONTROLS

Autonomous direct determination of a continuous time series of streamflow For a simple rating, it is assumed that the relation between stage and discharge
is not economically feasible at present (2014). As such, surrogates are used to is controlled by a section (section control) or reach of channel (channel con-
derive a continuous time series of streamflow. The derivation process entails trol); both of these features are referred to as station control (Kennedy, 1984).
developing a streamflow rating, which can range from a simple, single-valued Corbett (1943) explains this as follows: “In order to have a definite and endur-
relation between stage and streamflow to a fully dynamic one-dimensional ing stage-discharge relation, the channel must have certain physical features
model based on hydraulics of the flow. capable of regulating or stabilizing flow past the gage to such an extent that
for a given stage of the water surface, the discharge will always be the same.”
6.1 INTRODUCTION A station control essentially “controls” the relation between the stage and
discharge and prevents hydraulic disturbances from translating past the con-
In the absence of discrete measurements of streamflow (discharge measure- trol (and affecting the stage-discharge relation).
ments), streamflow typically is determined through surrogate measures of Depending on the state of flow (subcritical or supercritical flow), the con-
other variables such as stage, water-surface slope, rate of water surface rise, or trol will either be downstream or upstream of the streamgage location. The
index velocity at a streamgage. Streamflow is derived from the surrogate vari- flow type is determined by the Froude number (F) as:
ables using what is termed a “rating.” The rating model is developed and cali-
brated utilizing discharge measurements, collected onsite by field staff using V Q
standard methods (Turnipseed and Sauer, 2010; International Standards F= = (6.1)
gD gA T
3
Organization, 2007).
Types of ratings used are broken into two broad categories: simple stage-
streamflow ratings (simple rating) and complex ratings. A simple rating where V is the mean velocity, g is the acceleration of gravity, D is the hydrau-
relates discharge to stage, assuming a unique univariate (single-value) relation lic depth, Q is the discharge, A is the cross-sectional area, and T is the top
between stage and discharge (Fig. 6.1). Simple ratings do not work well for width. Flow is subcritical when F is less than 1 and supercritical when F is
streamgages on low-gradient streams, streams with variable backwater greater than 1. The critical state of flow is when F equals 1.
(Fig. 6.2), streams with large amounts of channel or overbank storage, or For subcritical flow, the control will be downstream of the streamgage and
streams with highly unsteady flow. A complex rating relates discharge to stage that control will not allow any effects downstream of the control (due to
and other variables because of the lack of a unique, univariate relation changing streamflow or channel conditions) to translate upstream through
between stage and discharge. Complex rating methods range from simply the control. For supercritical flow (Froude number greater than 1), the control
adding a second independent variable to the process of computing stream- will be upstream from the streamgage and would not allow any effects
flow to computer models that solve conservation-of-momentum and conser- upstream of the control (due to changing streamflow or channel conditions)
vation-of-mass partial differential equations. to translate downstream through the control. Thus, depending on the state of
Traditionally, the simplest form of rating model (stage streamflow) is tried flow, the stage-discharge rating relation at the streamgage will neither be
first, with progressively more complex models developed if the simpler forms impacted by hydraulic changes that occur upstream (subcritical flow) or
fail to provide an accurate relation. downstream (supercritical flow) from the control, nor will effects translate in
Rating development and verification constitute an ongoing process the other direction through the control feature. For the remainder of this
throughout the life of a streamgage, requiring systematic analysis and peri- chapter, the discussion will focus on the more common condition of subcriti-
odic evaluation to refine (or reconstruct) the rating in response to changes in cal flow (control downstream).
stream hydraulics. This analysis and evaluation must be undergirded by a For section control, a single section of the channel is the controlling feature
thorough understanding of both the science of river hydraulics and the many and, strictly defined, is configured (e.g., by constriction, bump in channel
assumptions made for the various rating methods. Those sites producing the bottom, gravel/rock riffle, artificial weir/flume, or some combination) such
best rating-derived streamflow data are the result of a systematic process that that the flow passes through the critical-flow state. At the state of critical flow,
is put in place to regularly analyze available rating methods in light of new F equals 1 and from Eq. (6.1) the stage (which is directly related to depth of
data, keeping in mind that alterations and corrections made on the basis of flow) is only dependent on the discharge and the channel geometry. This
new data must accurately reflect the underlying stream hydraulics—and aspect is very useful for ratings, as it implies a unique relation between stage
therefore are rational and defensible. and discharge (assuming unchanging channel geometry). The equation for a
6-1
06_Singh_ch06_p6.1-6.14.indd 1 8/23/16 4:35 PM

6-2 Streamflow Ratings
U.S. Geological Survey streamgage 07015720, Bourbeuse River near High Gate, Missouri
102.10
82.10
62.10
52.10 Explanation
42.10 Simple rating
32.10
Discrete measurement of streamflow
22.10
12.10
10.10
8.10
7.10
Stage, in ft
6.10
5.10
4.10
3.10
2.90
2.70
2.60
2.50
2.40
2.30
2.20
0.00499995 0.01 0.1 1 10 100 1000 10,000 77,800
Streamflow, in ft3/s Data from U.S. Geological Survey, 2014
Figure 6.1 Example of a simple rating for the Bourbeuse River near High Gate, Missouri, USA (U.S. Geological Survey streamgage 07015720).
U.S. Geological Survey streamgage 05586100, Illinois River at Valley City, Illinois
28.00
26.00 Explanation
Simple rating
24.00 Discrete measurement of streamflow
22.00
20.00
18.00
16.00
Stage, in ft
14.00
12.00
10.00
8.00
6.00
4.00
2.00
0.60
2000 10,000 30,000 50,000 70,000 90,000 110,000 118,700

Figure 6.2 Example of a site for which the simple rating shown will not determine streamflow accurately, Illinois River at Valley City, Illinois, USA (U.S. Geological
Survey streamgage 05586100).
06_Singh_ch06_p6.1-6.14.indd 2 8/23/16 4:35 PM

Rating Controls 6-3
straight-line rating (plotted in logarithmic space) that characterizes flow over k 2 1

an artificial or natural section control can be expressed (Rantz, 1982) as: Q = AR 3 S0 2 (6.6)
n
Q = P(G − e )N (6.2) Equation (6.6) represents the assumption of steady, uniform flow and
undergirds the simple stage-discharge relation for channel control; Eq. (6.6)
where Q is streamflow, P and N are constants depending on the geometry and demonstrates that for a channel reach of unchanging geometry, roughness,
configuration of the natural section and its approach section, G is the stage, and bed slope, a unique relation between stage and streamflow occurs. For a
and e is the stage (or gage height) of zero flow (GZF). GZF corresponds to the rectangular channel, Eq. (6.6) becomes
stage when streamflow just ceases. The exponent N ranges from 1.5 for rect-
k 5 1
angular sharp-crested weirs to 3.3 for Columbus-type controls (Rantz, 1982). Q = BD 3 S0 2 (6.7)
For natural control sections, N is almost always equal to or greater than 2. The n
values of P are also highly variable, being influenced largely by the shape and where B is channel width and D is depth of flow.
configuration of the artificial or natural control section. Steady, uniform flow rarely exists in natural streams and rivers because of
For medium-to-high streamflow, the influence of the section control is lost the dynamic and spatial variability of open-channel flow (unsteadiness, con-
(submerged) and channel control is typically in effect. Channel control convective acceleration, and differential hydrostatic pressure forces). When no
sists of all the physical features of the downstream channel (such as geometry, abrupt changes occur in the flow depth (and thus velocity), this flow is called
bed slope, expansion/contraction, sinuosity, and bed/bank roughness) that gradually varied flow (Chow, 1959). As such, the controls on the stage-
determine the stage of the river for a particular discharge (Rantz, 1982). streamflow rating are complex. For channel control in simple ratings, the
Channel control can be conceptualized and formulated using the flow- assumption is made that the bed slope (S0) can be substituted for the friction
resistance relation for open channels; flow resistance in an open channel can slope (Sf) in Eq. (6.6). This assumption needs further examination.
be described by the Manning equation (Yen, 1992): When flow is predominantly one dimensional, the friction slope can be
expressed through a rearrangement of the one-dimensional unsteady equa-
k 2 1 tion of motion (Henderson, 1966) as:
V = R 3 S f 2 (6.3)
n ∂ y v ∂v 1 ∂v
S f = S0 − − − (6.8)
where the constant k is 1.49 for English units and 1 for SI units, n is the ∂ x g ∂ x g ∂t
Manning resistance coefficient, R is the hydraulic radius, and Sf is the friction where y is flow depth, x is distance in the longitudinal direction of flow, v is
slope. The hydraulic radius is defined as: the velocity in the longitudinal direction, and t is time. Eq. (6.8) readily shows
A the effects of differential hydrostatic pressure force (∂ y ⁄ ∂ x ), convective
R= (6.4)
P acceleration (∂v ⁄ ∂ x ), and unsteadiness (∂v ⁄ ∂t ) on the friction slope.
where P is the wetted perimeter. It should be noted that Chezy’s1 equation is Additionally, the conservation of mass in one-dimensional flow with no side-
equally appropriate for describing the flow resistance (French, 1985), because
channel inflow can be stated as (Henderson, 1966):
Manning’s equation has its roots in the Chezy equation (Yen, 1992).
Because the streamflow is the product of the mean velocity and the cross- ∂Q ∂y
+T = 0 (6.9)
sectional area, A, we can express. Eq. (6.2) in terms of streamflow as: ∂x ∂t
k 2 1 with all variables in Eq. (6.9) having been previously defined. The components
Q = AR 3 S f 2 (6.5) to control complexity in Eq. (6.8), caused by the nonuniformity and unsteadi-
n ness of the flow, combined with Eq. (6.9), produce a hysteresis effect in the
For steady, uniform flow the friction slope (Sf) is equal to the bed slope (S0) stage-discharge relation as shown in Fig. (6.3). This hysteresis is the so-called
(Chow, 1959) and Eq. (6.5) becomes loop rating (Kennedy, 1984), or loop-rating curve (Henderson, 1966).
988
986
984
982
980
Stage, in ft
978
976
974
972
970
968
0 1000 2000 3000 4000 5000 6000 7000
Streamflow, in ft3/s
Figure 6.3 Example of a loop rating based on an unsteady flow routed by the unsteady flow option of the one-dimensional hydraulic model HEC-RAS (U.S. Army Corps of
Engineers, 2010) through a nonuniform rectangular channel.
1
Chezy’s equation V = C RS f where C is Chezy’s roughness coefficient.
06_Singh_ch06_p6.1-6.14.indd 3 8/23/16 4:35 PM

Equations (6.8) and (6.9) conceptually capture all one-dimensional causes 6.3 SIMPLE RATINGS
of hysteresis, including those phenomena that are dominated by conservation
Ratings are developed by first plotting the discharge measurements. For many
of mass concepts as demonstrated in Fig. 6.4 (Kennedy, 1984) which shows
sites, a simple rating will suffice, with segments of the simple rating repre-
the loop effects when the hysteresis is tied to channel storage during the rise
sented by a power relation (Rantz, 1982) illustrated in Fig. 6.5 and shown as
and fall of a hydrograph or when streams with large floodplains transition
Eq. (6.2). Because the independent variable stage is plotted on the y-axis (this
into and out of overbank flows.
is standard practice), the slope term N is defined as the run over the rise
Although all natural streams have hysteresis (Fig. 6.3), a complex rating
instead of the traditional rise over the run.
is not required in all cases because sometimes the degree of hysteresis is
In the United States, particularly in the U.S. Geological Survey, simple rat-
small and well within the uncertainty of discharge measurements made by
ings are often developed using logarithmic plots to take advantage of the
the hydrographer. Unsteady flows caused by rainfall events move as a
ability to linearize Eq. (6.2) as follows:
“wave” down the river. In those cases where the streamflow can be accu-
Log Q = Log P + N Log(G − e ) (6.10)
rately determined from a simple rating, the wave is said to move as a “kine-
matic wave” (Henderson, 1966)—Sf is approximated by S0 and conditions The value of e usually corresponds to the GZF (for section control) or effec-
meet the steady-uniform flow assumption. If the wave does not fit the tive GZF (for channel control). The term e is a value that, when subtracted
kinematic-wave model, or is subject to other factors (such as backwater) from stage in the curvilinear relation defined by the stage-discharge measure-
that preclude use of a simple rating, Eqs. (6.8) and (6.9) conceptually cap- ments, will cause the logarithmic rating curve to plot as a straight line; a
ture the complexity of the flow dynamics and provide the basis for rating straight-line relation helps when ratings need to be extended (or interpolated)
development. in unmeasured ranges of stage. The value e is also called the scale offset. For
(a) Rating loops from channel storage

26
24
22 Constant-stage rating
Bankfull stage
20
18
16 Overbank rise
Stage, in ft
14
12 Within-banks rise
10
8
6
4
2
(b) Rating loops from return of overbank flow

26
Recession from high
24
overbank rise
22
Bankfull stage
20 Recession from minor
overbank rise
18
16
Stage, in ft
14 All rising stages or recessions

from within-banks rises
12
10
8
6
4
2
0 10 20 30 40 50
Streamflow, in thousands of ft3/s
Figure 6.4 Typical loop ratings from single-storm events (adapted from Kennedy, 1984).
06_Singh_ch06_p6.1-6.14.indd 4 8/23/16 4:35 PM

Simple Ratings 6-5
of the floodplain, greatly increasing the conveyance capacity of the stream.

2 b=
1.7
Thus, for each increment of stage rise, a greater amount of water is conveyed
opposed to streamflow confined within the banks of the main channel.
1 6.3.1 Curve Shaping and Rating Extension for Simple Ratings
Discharge measurements over a range of stage form the basis for construction
(G–e), in ft
1.5 P
of accurate rating curves. Given discharge measurements have associated
errors, and adequate discharge measurements are sometimes not available to
define the entire range of the rating, hydraulic theory helps guide accurate
construction of the complete rating curve.
Channel surveys that provide detailed cross-section geometry (horizontal
Equation of curve shown is Q = 125(G–e)1.7 positions and associated elevations) enable proper geometrical characteriza-
tion of the streamgage reach. Figure 6.6 shows surveyed cross sections of
1
100 200 300 400 three different types of channel cross-section/control configurations and
Q, in ft3/s associated shape of the simple rating. Figure 6.6a has no section control and
a channel cross section that has no major break in geometry. The lack of sec-
Figure 6.5 Theoretical simple rating plotted as a straight line on log-log grid tion control and uniform cross-section geometry result in a single-segment
(adapted from Kennedy, 1984). linear rating of constant slope. Figure 6.6b shows no section control and a
major break in the channel cross section, where the water exceeds the bank-
regular-shaped section controls, the value of e will be the actual GZF. For full stage and flows through the wide floodplain on both sides of the main
channel controls and irregular-shaped section controls, the value of e will be channel. Channel control exists at all stages up to the bankfull condition of
the effective GZF, which usually is greater than the actual GZF (Rantz, 1982). the main channel (shown by the lower segment of the rating), and then the
Although logarithmic plotting is predominant, there are benefits to plot- rating bends over to the right as the stage exceeds the bankfull condition.
ting the low-flow part of the rating in arithmetic rectangular coordinates. Knowledge of the bankfull stage allows for proper shaping of the rating—the
This allows for more detailed analysis at small changes of stage and enables rating bends to the right at the bankfull stage, as the water conveyance
zero discharge to be plotted (which cannot be done in logarithmic space). increases substantially above this point. Figure 6.6c shows an effective section
Simple ratings often have multiple linear segments, in which each segment control inside the main channel, with a wide floodplain existing on both sides
represents a different control feature. For example, the lower part of the rating once the stage exceeds bankfull conditions. This results in a three-segment
may have section control and the upper part have channel control; each seg- rating curve as shown to the right in Fig. 6.6c.
ment often will have a different effective GZF (e). Kennedy (1984) provides Additionally, for those sites that have section control at lower stages, deter-
much discussion about the use of scale offsets and their selection. mination of GZF provides an anchor point on the lower end of the rating.
Experience with rating characteristics has shown that the term b typically Determination of GZF is done in the field, either by direct survey (using
will have a value of less than 2 for channel control and greater than 2 for sec- precision surveying equipment) or by determining the maximum depth of
tion control (Kennedy, 1984). The lower value of b for channel control is to be flow over the section control and subtracting this value from the observed
expected when one contrasts the exponent on the depth of flow in Eq. (6.7) stage.
with the exponent on depth in Eq. (6.2). Figure 6.1 demonstrates the higher When an artificial section control is present, the stage-streamflow relation
slope for the lower stages where section control is in effect and the transition (for the section control portion of the rating) can be computed using a weir
to channel control above approximately 7 ft stage where the slope is less than equation. Calibration of an appropriate weir equation is done by utilizing
2. Above approximately 16 ft stage, the curve bends back to the right. This discharge measurements made during section control conditions.
bend back to the right is not the result of the return of the simple rating to Slope-conveyance and step-backwater analyses are two hydraulic methods
section control, but rather it stems from the stage rising above the elevation that can be used to provide quantitative estimates of the relation between
(a)
Rating shape
Channel shape
No flood plain 1
G–e
(channel control) <2
(b)
Bankfull
G–e
Bankfull stage 1 stage

Wide flood plain <2
(channel control)
(c)
G–e
1
Wide flood plain and >2 <2
1
high section control
Figure 6.6 Relation of rating-curve shape to cross-section geometries (adapted from Kennedy, 1984).
06_Singh_ch06_p6.1-6.14.indd 5 8/23/16 4:35 PM

stage and streamflow under channel control conditions. These methods are the rating to above bankfull conditions. This prohibition is driven mainly by
useful to supplement discharge measurements in shaping the rating curve, as the fact that the relation of friction slope to stage at one control condition may
well as extrapolating the rating under certain constraints. not be the same as that for a different control condition.
The slope-conveyance method is the less data intensive and less rigorous of Step-backwater studies are considered to be detailed hydraulic analyses, but
the two methods. The method gets its name by restating the Manning equa- even these should be undertaken using a calibrated step-backwater model
tion [Eq. (6.5)] as: (Dyhouse, 2003). Calibration would consist of adjusting the model parame-
ters (mainly roughness) so that the modeled water surface-profile matches
1 (within tolerances) the observed water surface profile for a known stream-
Q = KS f 2 (6.11)
flow.
6.3.2 Temporary Changes in Control and the Shifting-
where K is the conveyance and equals Control Method for Simple Ratings
Natural streamgage controls may change temporarily due to a variety of rea-
k 2
sons, including scour/fill of the stream channel, changes in aquatic vegetation,
K = AR 3 (6.12)
n and accumulation of drift/debris. The changes in the control are detected in
the plotting position of the discharge measurement on the simple rating. The
From detailed cross-section surveys (typically at a single representative need for discharge measurements to detect changes to the control increases
cross section), the values of area and hydraulic radius for various values of during floods. An example of control changes that would only be detected by
stage can be determined. From knowledge of the site conditions or from discharge measurements is shown in Fig. 6.8. In this figure, two direct stream-
reverse calculations using one or more discharge measurements, one can flow measurements made during flooding in 2008 (numbers 353 and 354, Fig.
estimate the roughness value, n. Conveyance can then be computed for any 6.8) on the Platte River near Kearney, Nebraska resulted in a more than +1-ft
value of stage (Fig. 6.7). (ft) correction of the rating curve at the upper end (above a stage of 5.9 ft) of
The relation of friction slope to stage is typically derived from discharge the rating. At a streamflow of 15,000 ft3/s, the stage on the rating curve
measurements by changed by +1.2 ft, from approximately 7.0 to 8.2.
2
If the control changes are determined to be long-lasting or permanent, a
 Q new rating curve is often drawn. However, many control changes are tempo-
S f =   (6.13)
K  rary in nature (lasting weeks to months) before changing again. The shifting-
control method (Kennedy, 1983) is used in these situations. The
An example friction slope relation is shown on the right side of Fig. 6.7. shifting-control method keeps the existing rating (base rating) but computes
The streamflow for corresponding stages is computed by combining the con- a shifted rating curve based on discharge measurements. The shifted rating is
veyance and friction slope through Eq. (6.7) to produce the rating extension put into effect for the period that the control condition has changed. In some
shown on the left side of Fig. 6.7. cases, a site may need multiple shifted ratings to fully describe the relation
The step-backwater method (Chow, 1959; Davidian, 1984; French, 1985) between stage and streamflow for the various periods when the control condi-
consists of a survey of the reach downstream of the streamgage (assuming tions are changing. Figure 6.9 is an example of a site that has a temporary
subcritical flow), whereby the geometry of a number of cross sections is shifted rating for varying low-flow conditions.
defined and the roughness is characterized. Using the concepts of conserva- Shifted ratings are common for low-flow conditions, as section controls are
tion of energy and conservation of mass, the reach-length water-surface pro- often the most sensitive to scour/fill and buildup of debris. A given condition
file is computed for various streamflow values. Because a downstream water often lasts only a short while before changing again. The shifted rating(s) are
surface must be assumed (or known), often the downstream cross section is activated and deactivated based on hydrologic judgment relative to the
set far enough downstream of the streamgage that any errors in computed hydraulic factors causing the changes in control. For example, erosion of the
water surface at the streamgage (caused by an assumed downstream water- low-water section control (causing a shift to the right of the simple rating)
surface elevation) are minimized. may have resulted from a high-water event; then on the next high-water
Both methods work well for rating development and shaping, particularly event, the low-water control aggrades back to its original height and configu-
in ranges of stage where discharge measurements are available to augment the ration. The shifted rating may be activated as the flow recedes from the first
methods. In those situations where ratings need to be extrapolated beyond high-water event, and then removed during the next high-water event, as the
the highest observed measurement, great care must be exercised. It is the relation between stage and streamflow revert back to the original (base)
policy of the U.S. Geological Survey not to extrapolate ratings to more than rating.
twice the highest discharge measured by discrete observation (Rantz, 1982)
unless detailed hydraulic analysis (such as step-backwater) is conducted. The 6.4 COMPLEX RATINGS
slope-conveyance method is not considered a “detailed hydraulic analysis,”
but because it utilizes discharge measurements and channel geometry, it can A variety of complex rating methods exists to deal with situations where
be utilized for extrapolation of ratings beyond twice the highest measure- simple ratings will not work. The choice of method is dependent on the data
ment. It is important to note, however, that this method should only be used available and the cause of the complexity. The choice often cannot be made
within the same control conditions as the highest discharge measurement. For prior to the collection and analysis of discharge measurements. More dis-
example, if the highest discharge measurement was made below bankfull charge measurements are required to characterize complex ratings than for
conditions, the slope-conveyance method should not be used to extrapolate simple ratings.
Simple rating Conveyance curve Friction-slope curve

1.49 Q 2
1/2 Sf =
Q = KSf K= AR2/3
n K
G–e
Rating interpolation
and extension
Q K S
Figure 6.7 Example of a simple rating shaped by the slope-conveyance method.
06_Singh_ch06_p6.1-6.14.indd 6 8/23/16 4:35 PM

Complex Ratings 6-7
10.90
1.2 feet at
15,000 ft3/s
354
353 24
259 207 22
206
236 23
209
5.90 210
37
208
355 36
211
340
235
Stage, in ft
341 338
350 343
356
Explanation
Rating no. 9.0 starting on 04/28/2008
Rating no. 8.10 used prior to 4/26/2008
338 Discharge measurements subsequent to

10/01/2006 plus all above 7,000 cubic feet
per second are plotted
355 2009 flood measurement
1.90
1000 5000 10,000 30,000
3 /s
Streamflow, in ft Data from U.S. Geological Survey, 2014
Figure 6.8 Changes in the rating curve for the Platte River near Kearney, Nebraska, USA (U.S. Geological Survey streamgage 06770200) (from Holmes et al., 2010).
While fully dynamic numerical models are sometimes used, many methods predicted increase in water-surface slope to the estimated celerity of the flood
to deal with complexity were developed in the years prior to modern com- wave (Rantz, 1982). These methods are still in use today (2014).
puter computational capabilities that now simplify the application of Eqs. Even though we can mathematically characterize the physics of rating
(6.8) and (6.9). For example, when variable backwater substantially impacts a hydrodynamics, other factors present in natural systems can add to the com-
site (effectively becoming a downstream control), two streamgages would be plexity, creating the need for additional analysis and consideration. Mobile-
installed some distance apart and the water-surface slope would be used as an bed channels, particularly sand channels, present challenges because of the
additional independent variable for the rating. Other rating methods treat the dramatic changes in cross-section geometry and concurrent changes in
looped rating as a simple storage phenomena (Kennedy, 1984) or relate the channel-bed roughness (form roughness) caused by the evolution of bedforms
22.00
Explanation
12.00 Data pertaining to water year* 2012
10.00 Most recent measurements
8.00 Highest measurements of the last 10 years
7.00 Measurements from the last 2 years
6.00 Base rating
Shifted rating
5.00
*Water year is the 12-month period from October 1
through September 30 of the following calendar
4.00 year. The water year is designated by the calendar
Stage, in ft
year in which it ends. For example, water year

2012 is the period from October 1, 2011, through
September 30, 2012.
3.00
2.80
2.60
2.50
2.40
2.30
2.20
2.11
0.5 1 4 10 40 100 400 1000 4000 10,000 40,000
Figure 6.9 Base rating curve and temporary shifted ratings.
06_Singh_ch06_p6.1-6.14.indd 7 8/23/16 4:35 PM

over the range of flows (Simons and Richardson, 1962). These unstable chan- Qm is now the computed streamflow for each increment in time that data are
nels require special consideration in rating analysis (Kennedy, 1984). collected.
Two general types of slope rating methods exist: (1) constant-fall ratings
6.5 SLOPE RATINGS and (2) variable-fall ratings, in which variable-fall ratings are further divided
into (a) limiting-fall and (b) normal-fall categories. Constant-fall ratings are
For sites with variable backwater, the friction slope (and water-surface slope) typically used when backwater is present at all times for all stages. Variable-
varies for a given stage. As evidenced from Eq. (6.5), varying friction slope fall ratings are preferable for those sites where backwater is intermittent.
means that multiple values of streamflow would be computed for the same The limiting-fall slope rating “limits” the streamflow to a maximum value
stage. The causes of this variable backwater are most often due to variable equal to the rated streamflow. In other words, the maximum value of dis-
stage at a downstream confluence or reservoir/lake (Rantz, 1982). Rating charge corresponding to a particular stage is limited to the value indicated by
complexity due to variable backwater can often be addressed by use of a stage- the simple rating, even if the measured fall is greater than rated fall. For
fall-streamflow rating, or more commonly referred to as a slope rating. Slope example, if the rated fall for a stage of 10 ft was 0.75 ft, but the measured fall
ratings are determined empirically for discrete observations of (1) streamflow, was 1.3 ft, the value of the fall ratio in Eq. (6.16) reverts to 1.0 instead of 1.73.
(2) stage at a base streamgage (the streamgage at which the streamflow data Therefore, Qm would be equal to Qr. Limited-fall ratings work best in situa-
are being computed, typically the upstream streamgage), and (3) the fall of the tions such as when a streamgage operates using a simple rating except during
water surface between the base streamgage and an auxiliary streamgage. high water, where a downstream dam creates backwater during the high
There should be no tributary inflow in the reach between the base and auxil- stages.
iary streamgages, the water-surface fall will ideally be a minimum of 0.5 ft, A normal-fall slope rating does not limit the discharge to any maximum
and each streamgage should be precisely synchronized and surveyed to the value from the rating curve. When the measured fall is greater than the rated
same vertical control. Adherence to these guidelines will improve the reli- fall, the value of the streamflow computed is allowed to be larger than the
ability of the slope rating (Kennedy, 1984); however, fall between streamgages rated streamflow. Normal-fall slope ratings are sometimes used where high-
of less than 0.5 ft can still provide satisfactory computation of the streamflow water measurements fail to indicate a limiting position for the discharge rat-
(Rantz, 1982). ing curve.
To develop a slope rating, numerous discharge measurements are made The development of a slope rating follows a similar (but not identical)
with concurrent stage and fall values recorded for each measurement. The iterative pattern, regardless of the type. Trial versions of the various compo-
following equation (Rantz, 1982) nents (simple rating, fall rating, and factor curve) are estimated through vari-
E
ous means utilizing the discharge measurement, followed by iterations to
Qm  fm  (6.14) improve the fit of the observations to the component relations. Figure 6.10
Qr =  f  shows the variable-fall rating for the Mississippi River at Thebes, Illinois (U.S.
r
Geological Survey streamgage 07022000). Detailed discussions of construct-
ing slope ratings can be found in Kennedy (1984).
is the major component of the slope rating method, where Qm (measured
streamflow) is the actual measured (or true) streamflow for the observed
stage (call it equal to HC), Qr (rated streamflow) is the streamflow taken from 6.6 RATE OF CHANGE OF STAGE METHODS
a simple base rating built for “normal” conditions when the stage equals HC, Rate of change of stage methods were developed as simplified methods to deal
fm (measured fall) is the actual fall associated with the measured streamflow with rating loops. A few of these methods include (1) storage method
at stage equal to HC, fr (rated fall) is the expected fall under “normal” condi- (Kennedy, 1984), (2) Jones method (Jones, 1915), (3) Lewis method (Corbett,
tions at stage equal to HC, and E is an exponent that has a value typically 1943), (4) Wiggins method (Rantz, 1982), and (5) Boyer method (Rantz,
f
ranging from 0.4 to 0.6. The term m is known as the fall ratio. Equation 1982; Kennedy, 1984). These methods utilize the development of a simple
fr rating, with an adjustment factor determined from a factor curve. Each
(6.14) has its basis in Eq. (6.5), whereby the friction slope is approximated by method employs a different factor curve, depending on the primary loop-
the water-surface slope. When the slopes for measured and normal condi- causing phenomena the method was developed to address. For example, the
tions are input, the ratio of the measured conditions and normal conditions storage method was developed for those situations where storage in the chan-
becomes nel is the main cause of the loop (Fig. 6.4a). The Boyer method is used to
km 2 1
address situations, where the loop effect cannot be addressed by simple stor-
Am Rm 3 S fm2
Qm nm age alone (Kennedy, 1984). The Jones, Lewis, and Wiggins methods are simi-
= (6.15)
Qr kr 2 1 lar to the Boyer, but not as widely used because of modifications present in
Ar Rr 3 S fr2
nr the Boyer method that make it easier to use (Rantz, 1982).
The Boyer method is the predominant rate of change of stage method in
Because the stages for the measured and normal condition are the same, area, the U.S. Geological Survey and is documented here. The Boyer equation is
roughness, hydraulic radius, and reach length are the same for the both mea-
sured and normal conditions. Equation (6.15) therefore reduces to Eq. (6.14).
Qm  J  (6.17)
From Eq. (6.15), one can see where the theoretical value for the exponent E in = 1+ 
Eq. (6.14) is equal to 0.5. Qr  USc 
The slope rating method assumes the friction slope is approximated by the
water-surface slope. From Eq. (6.8), one can see that this approximation works where Qm is measured streamflow, Qr is the rated streamflow, J is the rate of
change of stage, U is the celerity of the flood wave, and Sc is the friction slope
so long as: (a) the reach has very gradual changes in cross sectional geometry for steady-flow conditions measured at the same stage as Qm. In earlier meth-
  ods, such as the Jones method, the values of U and Sc were evaluated sepa-
resulting in negligible convective acceleration v ∂v ≅ 0 , and (b) the flow is
 g ∂ x  rately, but the Boyer method evaluates them as one. Jones (1915) has an
 1 ∂v  explanation of how Eq. (6.17) was developed based on the water-surface slope
not highly unsteady resulting in negligible local acceleration  ≅ 0 .
 g ∂t  differences between the slope when the flow is at steady state and when the
flood wave has arrived in the upstream end of the reach and just arriving
Slope ratings typically consist of three components: (1) a relation of stream- (with no change in stage yet) on the downstream end of the reach.
flow to stage for “normal” conditions (simple rating), (2) a relation of fall to To develop a Boyer method rating (Fig. 6.11), numerous discharge mea-
fm Q surements are made with concurrent stage and rate of change of stage values
stage (fall curve), and (3) a relation of to m Q (factor curve). The con-
f
r r recorded for each measurement. A simple rating is developed, whereby the
struction of the simple rating, fall curve, and factor curve is done by trial- first trial simple rating is drawn to be to the left of the positive rates of stage
and-error iteration utilizing the discharge measurements made across the change (rising limb of hydrograph) and to the right of the negative rates of
range of stages, streamflows, and fall conditions. stage change (falling limb). Equation (6.17) is then rearranged to compute
Once a slope rating is established, Eq. (6.14) can be rearranged to provide values of 1 ⁄ USc for each discharge measurement (Qm), associated value of Qr
a computation of the streamflow for the backwater condition from the trial simple rating that corresponds to the observed stage, and the
rate of change of stage (J) during the measurement. A 1 ⁄ USc factor curve is
E
f  then developed from the computed values of 1 ⁄ USc . Extra weight should be
Qm =  m  × Qr (6.16) given to those values of 1 ⁄ USc , where the rate of stage change is larger, with
 fr 
the maximum value of 1 ⁄ USc corresponding to just-above bankfull conditions.
06_Singh_ch06_p6.1-6.14.indd 8 8/23/16 4:35 PM

Rate of Change of Stage Methods 6-9
(a) Simple base rating

100
Explanation
Simple base rating
Discrete measurement of streamflow
Adjusted streamflow
Stage, in ft
10
1
10,000 100,000 1,000,000
Qr , in ft3/s Data from U.S. Geological Survey, 2014
(b) Fall curve (c) Factor curve

50 1
45 0.9
40 0.8
35 0.7
Stage, in ft
Qm / Qr
30 0.6
25 0.5
20 0.4
15 0.3
10 0.2
5 0.1
0 0
0 1 2 3 4 5 6 7 8 0 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1
Fr , in ft Fm / Fr
Data from U.S. Geological Survey, 2014 Data from U.S. Geological Survey, 2014
Figure 6.10 Variable-fall slope rating for the Mississippi River at Thebes, Illinois, USA (U.S. Geological Survey streamgage 07022000).
The value of 1 ⁄ USc goes to zero at two extremes on the graph—at the low end The values of Qadj are then plotted against the stage to further refine the
where section control is just submerged, and at the upper end where the simple rating following by recomputation of 1 ⁄ USc and a refinement of the
floodplain contains most of the flow (Kennedy, 1984). An iterative process is factor curve. Great care should be taken to make sure the factor curve bends
followed to further refine the simple rating and factor curve. In the second as gently as the data will allow and that 1 ⁄ USc does not become too large at
iteration, an adjusted streamflow, Qadj, is computed for each Qm from the fol- any one stage; if these steps are not taken, the value under the radical can
lowing rearrangement of Eq. (6.17), substituting Qadj for Qr. become negative during periods of rapidly falling stages, resulting in no solu-
Qm tion during those periods. The iterative process can stop when the percentage
Qadj = errors between the Qr from the simple rating and Qadj computed for that
 J  (6.18) iteration are minimized.
1+ 
 US  c
06_Singh_ch06_p6.1-6.14.indd 9 8/23/16 4:35 PM

100
90
80 Explanation
70
60 Discrete observation of streamflow and
measurement number—number
50 307 below (in parentheses) is dh/dt
(–.30)
40 Measured
278 Adjusted for changing stage effect

30 (–.08) by relation
Qm 1
= 1+ ×J
20 Q adj USc 20
278 314 242 112 39 243 206
146
(–.08)
197
15 235
(+.11) 277 127 15
(–.45) 192 (+.04)
Stage, in ft
319 289 (+.04)

(–.30) (+.09)
10 205 10
9 (–54) 208
8 (+1.33)
7
6
5 307
(–.30)
4
328
313 (+.04)
3
0.0
0.5
1.0
529 1
335
2 USc
315
1
10 100 1000 10,000 20,000
Figure 6.11 Example of the Boyer method applied at Lightning Creek near McCune, Kansas, USA (U. S. Geological Survey streamgage 07184000) (adapted from Kennedy, 1984).
6.7 DYNAMIC-FLOW MODEL METHODS developed such that it could readily be used to provide a derived stream-
Fully dynamic one-dimensional computational hydraulic models are used for flow based on observed stage (Schraffranek, 1989). BRANCH has been
those situations where the friction slope is variable for the same stage and the used successfully for the operation of streamgages (Bossong and Jeffcoat,
methods presented previously are insufficient to deal with the complexity. 1989; Bossong, 1989; Neely, 1989; Hampson, 1989). In streamflow-rating
These models utilize the one-dimensional conservation of momentum equa- mode, BRANCH requires cross-sectional geometry, reach length, channel
tion [Eq. (6.8)] and the one-dimensional conservation of mass equation [Eq. roughness, and time series of both the upstream and downstream water-
(6.9)]. surface elevations (with these two time series serving as the boundary
Equations 6.8 and 6.9 can have other forms, and are also known as the conditions for the model). The time series are readily generated by estab-
Saint-Venant equations, the one-dimensional unsteady shallow-water equa- lishing an auxiliary streamgage either upstream or downstream of the base
tions, the unsteady open-channel flow equations, or some other variant (Lee, streamgage.
1989). The applicability of these governing equations to accurately model the
streamflow dynamics requires that the fundamental underlying assumptions 6.8 INDEX-VELOCITY METHOD
of these equations are met, including (Cunge You et al., 1980):
1. Streamflow is one-dimensional with the water surface horizontal and the An increasingly common method to derive a continuous time series of
velocity uniform for the cross section. streamflow, particularly at streamgages with rating complexity, is the index-
2. The pressure is approximately hydrostatic because the flow is gradually velocity method (Levesque and Oberg, 2012). Streamflow can be computed at
varied (streamline curvature is small) with negligible vertical accelerations. a channel cross section as the product of the mean velocity and the cross-
3. The water is of constant density and is incompressible. sectional area.
4. The boundary friction and Reynolds stresses (turbulence) can be The cross-sectional area can be readily determined through bathymetric
accounted for by using a flow-resistance equation such as the Manning equa- and traditional land surveys. Using these data, the relation between stage and
tion [Eq. (6.5)]. cross-sectional area (Fig. 6.12a) can be determined (stage-area rating).
5. The channel bottom remains stable with time. The frequency of resurvey and verification of the stage-area rating depends
6. The channel slope is small. on the stability of the channel. The U.S. Geological Survey recommends that
Analytic solutions to these two partial differential equations [Eqs. (6.8) and the channel be resurveyed every year for the first 3 years, followed by a mini-
(6.9)] are only possible in simplified situations (Lai, 1982). As such, numerical mum of every 3 years so long as the channel remains stable (Levesque and
methods are often used to provide discrete approximations to these governing Oberg, 2012).
equations in the time and space domain. The numerical methods can be cat- Autonomous, continuous-direct determination of mean channel velocity is
egorized by their solution scheme as finite difference methods (Fread, 1974; very difficult and typically not cost efficient. However, continuous measure-
Cunge et al., 1980; Schaffranek et al., 1981; Franz and Melching, 1997; Barkau, ment of velocity (index velocity) of a certain portion of a river, stream, or
2001), finite-element methods (Smith and Cheng, 1979; Delong and canal can be made and used as an index to estimate the mean channel veloc-
Schoellhamer, 1989), and method of characteristics (Lai, 1967; Goldberg and ity (Levesque and Oberg, 2012). Discharge measurements and concurrent
Wylie, 1983). readings of the index velocity are made at a range of expected mean velocities.
Regardless of the numerical method utilized, initial and boundary condi- Once sufficient discharge measurements have been made, a relation to predict
tions (stage, streamflow, or some combination) are required. The BRANCH mean velocity can be developed (index-velocity rating). Mean velocity
model (Schraffranek et al., 1981) is a finite-difference model that was typically can be a function of the stream-wise component of the measurement
06_Singh_ch06_p6.1-6.14.indd 10 8/23/16 4:35 PM

Shifting-Control Method for Dealing with Rating Complexities 6-11
index velocity (Fig. 6.12b), but may also require additional independent or an alternative path must be found. Most often, rating complexity at these
variables such as stage. Accuracy of the index-velocity rating must be checked sites is dealt with by making numerous discharge measurements and either
periodically by making additional discrete observations of streamflow and (1) using a simple rating with the shifting-control method or (2) if frequent
adjusted accordingly. enough, the measurements are simply connected through linear interpolation
One marked advantage of the index-velocity method is that it can work to provide the streamflow for the desired time interval of interest (e.g.,
very well in bidirectional or reverse flow, both situations where other complex hourly estimates of streamflow).
rating methods either are unable or find it difficult to characterize these con- Figure 6.14a shows the loop presented during the historic 2011 flood on
ditions. Figure 6.13 shows an index-velocity rating for Trail Creek at the Missouri River at Bismarck, North Dakota (U.S. Geological Survey
Michigan City Harbor, Indiana. Because the water-surface gradient is small, streamgage 06342500). Using the shifting-control method, multiple shifted
reverse flows occur when winds blow upstream; however, the index-velocity ratings (Fig. 6.14b) were used during the 2011 flood to allow for accurate
method is able to compute a derived streamflow. determination of derived streamflow. Based on over 50 discharge measure-
ments in a 111-day period, a total of five individual shifted ratings were
6.9 SHIFTING-CONTROL METHOD FOR DEALING WITH applied during the flood to mimic the complexity of the rating. This approach
RATING COMPLEXITIES is used at a number of locations where other complex methods have not
At certain sites, none of the available complex-rating methods will provide an proven reliable, and/or the cost of the extra discharge measurements required
acceptable level of accuracy for the derived data. Concessions have to be made is less than the cost of collecting the necessary ancillary data to implement a
in those situations, and either the reported data quality must be downgraded, particular complex-rating method.
(a) Stage-area rating

30
25
20
Stage, in ft
15
10
0
0 500 1000 1500 2000 2500 3000
Cross-sectional Data from U.S. Geological Survey, 2014
area, in ft2
(b) Index-velocity rating

2.50
Explanation
Linear (discrete observations of streamflow)

2.00 Discrete observation of streamflow
Measured velocity, in ft/s
1.50
1.00
0.50
0.00
0 0.50 1.00 1.50 2.00 2.50
Index velocity, ft/s Data from U.S. Geological Survey, 2014
Figure 6.12 Stage-area and index-velocity rating for Iroquois River at Foresman, Indiana (U.S. Geological Survey streamgage 05524500).
06_Singh_ch06_p6.1-6.14.indd 11 8/23/16 4:35 PM

0.60
Explanation
0.40 Index-velocity rating

Discrete observation of streamflow
0.20
Mean channel velocity, in ft/s
–0.20
–0.40
–0.60
–0.80
–0.80 –0.60 –0.40 –0.20 0 0.20 0.40 0.60
Index velocity, in ft/s Data from U.S. Geological Survey, 2014
Figure 6.13 Index velocity rating for Trail Creek at Michigan City Harbor, Indiana (U.S. Geological Survey streamgage 04095380).
(a)
30
20 16
N o.
g
7-1-11 6-26
7-4-11
25 tin
19 Ra 6-20
7-14-11
6-25
6-19
6-15 6-17 6-18
7-22-11 6-15 6-14
18 6-12
7-27-11 6-5
6-13 6
6-4 6-8 .1
20 6-13 6-9 6-11 No
17 6-6
ng
6-7
R ati
120,000 130,000 140,000 150,000 6-3-11
5-27-11
Stage, in feet
5-25-11
15
6-2-11 EXPLANATION
6-1-11
2011 flood measurements
5-19-11
Rising
Steady state
Falling
1952 flood measurements

10
Pre-dam measurements
Post-dam measurements
7.1
30,000 40,000 50,000 60,000 80,000 100,000 140,000 200,000 240,000 300,000 400,000 500,000
Streamflow, in cubic feet per second Data from U.S. Geological Survey, 2014
Figure 6.14 Simple rating and shifted ratings for the Missouri River at Bismarck, North Dakota (U.S. Geological Survey streamgage 06342500) during the 2011 flood.
06_Singh_ch06_p6.1-6.14.indd 12 8/23/16 4:35 PM

REFERENCES 6-13
(b)
20
18
Explanation
Base rating in use
Stage, in ft
16 Shifted rating used on recession
First shifted rating in time
Other dotted lines—series of straight line

shifted ratings that were prorated in
time to simulate the loop on the rise
14
Discrete measure of streamflow on rise
Discrete measure of streamflow during

backwater from Heart River
Discrete measure of streamflow on fall
12
60,000 80,000 100,000 200,000

Streamflow, in ft3/s
Data from U.S. Geological Survey, 2014
Figure 6.14 (Continued).
6.10 UNCERTAINTY IN RATINGS Chow, V. T., Open-Channel Hydraulics, McGraw-Hill, New York, 1959,
p. 680.
Over the past few decades, awareness has increased regarding the importance
Clarke, R. T., “Uncertainty in the estimation of mean annual flood due to
and subsequent need for estimates of the uncertainty of streamflow data.
rating-curve indefinition,” Journal of Hydrology, 222: 1–4, 1999.
Much work has been done to derive methodologies both for the uncertainty
Cohn, T. A., J. E. Kiang, and R. R. Mason, Jr., “Estimating discharge mea-
of discharge measurements (Cohn et al., 2013; International Standards
surement uncertainty using the interpolated variance estimator,” ASCE Jour-
Organization, 2007b; Oberg and Mueller, 2007; Herschy, 1999; Pelletier, 1988)
nal of Hydraulic Engineering, 139: 502–510, 2013.
and the uncertainty of streamflow ratings used to derive continuous stream-
Corbett, D. M., Stream Gaging Procedure, U.S. Geological Survey Water-
flow data (Morlot et al., 2014; Schmidt, 2002; Clarke, 1999; Masson et al.,
Supply Paper 888, U.S. Government Printing Office, Washington, D.C., 1943,
1987). Unfortunately, this awareness has not translated into an agreed-upon,
p. 245.
systematic method to estimate the uncertainty of rating curves and the sub-
Cunge, J. A., F. M. Holly, Jr., and A. Verwey, Practical Aspects of Computa-
ject remains a matter of research (Le Coz, 2012).
tional River Hydraulics, Pitman Publishing, London, 1980, p. 420.
Rating curve uncertainty depends on a number of factors including: (1)
Davidian, J., Computation of water-surface profiles in open channels, U.S.
characteristics of hydraulic controls at the site, (2) stability and condition of
Geological Survey Techniques of Water Resources Investigations Book 3, Reston,
hydraulic controls at the site, (3) frequency and distribution of discharge
measurements (through time and range of streamflow), (4) accuracy of dis-
Delong, L. L. and D. H. Schoellhamer, “Computer Program HYDRAUX a
charge measurements, and (5) accuracy of stage observations. Current
model for simulating one-dimensional, unsteady, open-channel flow,” U.S.
research notwithstanding, the standard operational practice at present (2014)
Geological Survey Water Resources Investigations Report 88-4226, Reston,
in the United States by the U.S. Geological Survey is to categorize the final
Virginia, 1989, p. 57.
derived streamflow data as excellent, good, fair, or poor, based on the analyst’s
Dyhouse, G. R., J. Hatchett, and J. Benn, Floodplain Modeling Using HEC-
qualitative assessment of the aforementioned factors at the site for various
RAS, Bentley Institute Press, Exton, PA, 2003, p. 696.
data-collection periods (Rantz, 1982; Kennedy, 1983).
Fread, D. L., Numerical Properties of Implicit Four-Point Finite-Difference
REFERENCES Equations of Unsteady Flow, NOAA Technical Memorandum NWS HYRDO-
18, Office of Hydrology, Washington, D.C., 1974, p. 38.
Barkau, R. L., UNET one-dimensional unsteady flow through a full net- Franz, D. D. and C. S. Melching, “Full equations (FEQ) model for the solu-
work of open channels, U.S. Army Corps of Engineers Generalized Computer tion of the full,” Dynamic Equations of Motion for One-Dimensional Unsteady
Program User’s Manual, April 2001 Version 4.0, U.S. Army Corps of Engineers, Flow in Open Channels and Through Control Structures, U.S. Geological Sur-
Hydrologic Engineering Center, Davis, California, 2001, p. 400. vey Water Resources Investigations Report 96–4240, Reston, Virginia, 1997,
Bossong, C. R., “Application of the BRANCH and LTM models to the p. 258.
Coosa River, Alabama,” Proceedings of the Advanced Seminar on One-Dimen- French, R. H., Open-Channel Hydraulics, McGraw-Hill, New York, 1985,
sional Open-Channel Flow and Transport Modeling, U.S. Geological Survey p. 705.
Water Resources Investigations Report 89–4061, Reston, Virginia, 1989, Goldberg, D. E. and E. B. Wylie, “Characteristics method using time-line
pp. 27–29. interpolations,” American Society of Civil Engineers Journal of the Hydraulics
Bossong, C. R. and H. H. Jeffcoat, “Application of the BRANCH model to Division, 105 (HY5): 670–683, 1983.
determine flow in the Alabama River,” Proceedings of the Advanced Seminar Hampson, P. S., “One-dimensional flow modeling of the St. Johns River at
on One-Dimensional Open-Channel Flow and Transport Modeling, U.S. Geo- Jacksonville, Florida,” Proceedings of the Advanced Seminar on One-Dimensional
logical Survey Water Resources Investigations Report 89–4061, Reston, Open-Channel Flow and Transport Modeling, U.S. Geological Survey Water
Virginia, 1989, pp. 23–26. Resources Investigations Report 89–4061, Reston, Virginia, 1989, pp. 31–34.
06_Singh_ch06_p6.1-6.14.indd 13 8/23/16 4:35 PM

Henderson, F. M., Open Channel Flow, McMillan, New York, 1966, p. 522. quality and station management indicators,” Journal of Hydrology, 517: 173–
Herschy, R. W., Uncertainties in Hydrometric Measurements, Hydrometry, 186, 2014.
2nd ed., John Wiley, New York, 1999, pp. 355–370. Neely, B. L., “Flow determination of the Arkansas River at Little Rock,
Holmes, Jr., R. R., T. A. Koenig, and K. A. Karstensen, “Flooding in the Arkansas, using the BRANCH model,” Proceedings of the Advanced Seminar
United States Midwest, 2008,” U.S. Geological Survey Professional Paper 1775, on One-Dimensional Open-Channel Flow and Transport Modeling, U.S. Geo-
Reston, Virginia, 2010, p. 64. logical Survey Water Resources Investigations Report 89–4061, Reston,
International Standards Organization, Hydrometry—Measurement of Liq- Virginia, 1989, p. 30.
uid Flow in Open Channels Using Current-Meters or Floats, ISO 748, Geneva, Oberg, K. A. and D. S. Mueller, “Validation of streamflow measurements
Switzerland, 2007a. made with acoustic Doppler current profilers,” ASCE Journal of Hydraulic
International Standards Organization, Hydrometry—Velocity-Area Method Engineering, 133: 1421–1432, 2007.
Using Current-Meters—Collection and Processing of Data for Determination of Rantz, S. E., “Measurement and computation of streamflow,” U.S. Geologi-
Uncertainties in Flow Measurement, ISO 1088, Geneva, Switzerland, 2007b. cal Survey Water-Supply Paper 2175, Reston, Virginia, 1982, p. 631.
Jones, B. E., “A method of correcting river discharge for changing stage,” Schaffranek, R. W., R. A. Baltzer, and D. E. Goldberg, A model for simula-
U.S. Geological Survey Water-Supply Paper 375, U.S. Government Printing tion of flow in singular and interconnected channels, U.S. Geological Survey
Office, Washington, D.C., 1915, pp. 117–130. Techniques of Water Resources Investigations Book 7, Reston, Virginia, 1981,
Kennedy, E. J., “Computation of continuous records of streamflow,” U.S. Chap. C3, p. 110.
Geological Survey Techniques of Water-Resources Investigations Book 3, Reston, Schaffranek, R. W., “Evolution and status of the branch-network unsteady
Virginia, 1983, Chap. A13, p. 53, http://pubs.usgs.gov/twri/twri3-a13/ flow model,” Proceedings of the Advanced Seminar on One-Dimensional
Kennedy, E. J., “Discharge ratings at gaging stations,” U.S. Geological Survey Open-Channel Flow and Transport Modeling, U.S. Geological Survey Water
Techniques of Water-Resources Investigations Book 3, Reston, Virginia, 1984, Resources Investigations Report 89–4061, Reston, Virginia, 1989,
Chap. A10, p. 59 , http://pubs.usgs.gov/twri/twri3-a10/ pp. 14–16.
Lai, C., “Computation of transient flows in rivers and estuaries by the mulitiple- Schmidt, A. R., Analysis of stage-discharge relations for open-channel
reach method of characteristics,” U.S. Geological Survey Professional Paper 575-D, flows and their associated uncertainties, Ph.D. Thesis, University of Illinois,
U.S. Government Printing Office, Washington, D.C., 1967, pp. 273–280. Urbana-Champaign, 2002, p. 329.
Lai, C., “Numerical modeling of unsteady open-channel flow,” Advances in Simons, D. B. and E. V. Richardson, The effect of bed roughness on depth-
Hydroscience, Academic Press, 14: 161–332, 1982. discharge relations in alluvial channels, U.S. Geological Survey Water-Supply
Le Coz, J., “A literature review of methods for estimating the uncertainty Paper 1498-E, 1962, p. 26.
associated with stage-discharge ratings,” Technical Report of the WMO Initia- Smith, L. H. and R. T. Cheng, “A gelerkin finite element solution of the
tive on Assessment of the Performance of Flow Measurement Instruments and one-dimensional unsteady flow problems using a cubic hermite polynomial
Techniques, 2012, p. 21. basis function: American Society of Civil Engineers,” Proceedings of the 15th
Lee, J. K., “The one-dimensional equations of open-channel flow,” Proceed- International Conference on Coastal Engineering, Honolulu, Hawaii, 1976,
ings of the Advanced Seminar on One-Dimensional Open-Channel Flow and pp. 3358–3376.
Transport Modeling, U.S. Geological Survey Water Resources Investigations Turnipseed, D. P. and V. B. Sauer, Discharge measurements at gaging sta-
Report 89–4061, Reston, Virginia, 1989, pp. 6–13. tions, U.S. Geological Survey Techniques and Methods Book 3, Reston, Vir-
Levesque, V. A. and K. A. Oberg, “Computing discharge using the index ginia, 2010, Chap. A8, p. 87, http:/pubs.usgs.gov/tm/tm3-A8/
velocity method,” U.S. Geological Survey Techniques and Methods Book 3, U.S. Army Corps of Engineers, HEC-RAS River Analysis System User’s
Reston, Virginia, 2012, Chap. A23, p. 148. Manual, Version 4.1, Report CPD-68, p. 766, 2010.
Masson, J., M. Ghio, C. Lallement, C. Parsy, and J. Philippe, Debitmetrie: U.S. Geological Survey, 2014, USGS Water Data for the Nation: U.S. Geo-
Precision des Stations de Jaugeage, La Houille Blanche, France, 1987, logical Survey, National Water Information System, http://nwis.waterdata.
pp. 333–338. usgs.gov//nwis (accessed September 2014).
Morlot, T., C. Perret, A. C. Favre, and J. Jalbert, “Dynamic rating curve assess- Yen, B. C., Channel Flow Resistance: Centennial of Manning’s Formula,
ment for hydrometric stations and computation of the associated uncertainties: Water Resources Publications, Littleton, CO, 1982, p. 453.
06_Singh_ch06_p6.1-6.14.indd 14 8/23/16 4:35 PM

Chapter 7
Hydrologic Information
Systems
BY
JONATHAN L. GOODALL, VENKATESH MERWADE, ILYA ZASLAVSKY, DAVID G. TARBOTON,
JEFFERY S. HORSBURGH, DANIEL P. AMES, AND ALVA COUCH
ABSTRACT integrate, analyze, visualize, and archive these data. The volume and scope of
data types relevant to hydrology make the design and implementation of such
Hydrologic information systems (HIS) are software designed to address data
software systems an important research challenge. Also, the fact that hydro-
management and computational challenges in hydrology. HIS development is
logic datasets tend to be geographically distributed, meaning the data are
being driven by the need to manage a rapidly growing volume of available
collected, curated, and disseminated by autonomous groups, adds to the chal-
data for characterizing hydrologic systems. HIS software builds from geo-
lenge of gathering and integrating hydrologic data to support science.
graphic information system (GIS) principles and more general information
Like other scientific and engineering disciplines facing similar challenges,
and communication technology (ICT) principles to create data types and
a subdiscipline of hydrologic research has emerged called hydroinformatics
transmission strategies tailored for hydrologic data and models. A key chal-
that focuses on these informatics and computational challenges (Kumar et al.,
lenge addressed by HIS software is integrating distributed and heterogeneous
2005). One area of active research within hydroinformatics is the design and
data under the control of different federal, state, and local data providers.
development of hydrologic information systems (HIS). Through collabora-
Hydrology requires a broad set of data types including geospatial, multi-
tions among hydrologists, water resource engineers, computer scientists, and
dimensional array, and model input and output file formats. These data must
software engineers, new methods and software tools have been created over
be accessed, transformed, analyzed, and loaded into models to understand
the past decade to improve hydrologic data management through the adop-
large and complicated hydrologic systems. This chapter presents principles
tion of HIS software. There are activities across the globe to design and create
driving the design of HIS software, the Consortium of Universities for the
HIS software tools (e.g., D’Agostino et al., 2014; Stenson et al., 2011), and a
Advancement of Hydrologic Science, Inc. (CUAHSI) Hydrologic Information
complete review of these activities is beyond the scope of this chapter. Rather,
System (HIS) as an example HIS software, and current research to create next
we focus on the Consortium of Universities for the Advancement of
generation HIS software.
Hydrologic Science, Inc. (CUAHSI) HIS as an example of an HIS with more
than a decade of development effort and broad adoption.
Following this introduction is a brief background on data-management
7.1 INTRODUCTION principles relevant to designing and creating an HIS. Next, computing meth-
Continuing exponential growth in both quantity and accessibility of earth odologies for integrating components of an HIS are presented, focusing in
science data will have major impacts on both the way we conduct scientific particular on a service-oriented perspective that has received the most atten-
research and the resulting knowledge derived from such research for the tion within the community. Then, the CUAHSI HIS is introduced, including
foreseeable future. The growth of data available to scientists has led to discus- its three core software components for data publication, data discovery, and
sions of a new “fourth paradigm” of scientific progress, where breakthroughs data access. Finally, the chapter concludes with current research within the
are driven by computing methodologies that aim at exploring and manipulat- academic community aimed at designing and building next generation HIS
ing massive datasets (Hey et al., 2009). While there is a wealth of new datasets software.
available to support science, there remain important challenges, such as the
tedious process of integrating subsets of large, heterogeneous datasets col- 7.2 HYDROLOGIC DATA MANAGEMENT
lected and maintained by different organizations into a single data collection
to support specific studies (Halevy et al., 2006). Big data and data science HIS software systems target the entire data life cycle (Faundeen et al., 2014)
thrive not only on large data collections, but also large collections of smaller and seek to provide tools and techniques to assist in the movement of data
datasets that have high potential for reuse and sharing (Lynch, 2008; Mayer- through it. The data life cycle (Fig. 7.1) begins with planning for data collec-
Schönberger and Cukier, 2013). The approaches and software described tion, proceeds to acquiring data, and then to processing and analyzing data.
within this chapter address these data challenges for hydrologic science and Once data have been acquired, they must be preserved through an archival
engineering. system and published for use by the scientific community. These last two
Massive quantities of water-related data are generated daily through remote steps of data archival and publication have not been typical of most scientific
sensing and in situ observational networks (Overpeck et al., 2011). Together analyses; however, increased attention by research funding agencies and new
with geospatial data, such as soil maps, digital elevation models, land use means for rewarding data publication, similar to current academic award
maps, and hydrography data representing digitized versions of once map- structures for journal paper publication, seek to change this practice (Lawrence
based data, hydrologists have detailed information available to support a et al., 2011). Throughout this data life cycle, describing data using data mod-
variety of analyses. Software systems are required to organize, store, transmit, els, metadata, and documentation is critical. Maintaining this information
7-1
07_Singh_ch07_p7.1-7.10.indd 1 8/23/16 4:38 PM

7-2 Hydrologic Information Systems
Publish/
Plan Acquire Process Analyze Preserve
share
Describe (metadata, documentation)
Manage quality
Back up and secure
Figure 7.1 The data life cycle. [Source: Faundeen et al., 2014.]
and keeping track of the provenance of data, as it is transformed between Hierarchical Data Format, or a raster format (e.g., GeoTIFF). Observational
phases of the data life cycle is also important. Finally, there are needs for data are typically stored as text files (e.g., as comma separated values, or CSV)
curating data, for providing quality assurance and quality control, and for or extensible markup language (XML) files (e.g., as Water Markup Language,
developing procedures to backup, secure, and replicate data across geographic or WaterML). Geospatial data or a collection of observational data might be
sites to minimize the risk of data loss. organized within a RDBMS using a data model, such as the ArcHydro Data
Central to the design of any information system are data models. Well- Model (Maidment, 2002) or the CUAHSI HIS Observations Data Model
defined data models describe how to encode information in a clear and con- (ODM) (Horsburgh et al., 2008).
sistent way so that it can be easily processed and understood by both humans Many agencies that fund hydrologic research and data collection, including
and computers, as it flows through the data life cycle. A data model serves as the National Science Foundation, are now requiring formal data-management
a standardized means for organizing information. The aim of data model plans as part of funded research proposals. Data storage, management, pres-
design is to create a representation of the real world as a set of entities with ervation, and publication are all part of data management that can be facili-
attributes and interrelationships (Chen, 1976). Abstracting data into these tated through adoption of a standardized data-storage model and/or
structures and normalizing them into granular entities defined by a formal data-representation model. From the perspective of archiving and publishing
data model makes it easier to integrate data across data providers who imple- data, data-storage and transmission formats must be capable of encoding
ment the same data model. Without coordination, organizations may store both the data themselves and the context and metadata surrounding the data
data in ad hoc data models that are not easily integrated. In such cases, inte- (i.e., the information needed to interpret the data) (Janée et al., 2008). In most
grating data between two or more organizations that have their own methods cases, reuse of existing data is difficult, if not impossible, without this contex-
for data organization requires a mapping between data models and the poten- tual metadata. Because both the data and their context may be complex, a
tial for significant data loss. well-designed underlying and archival data model can facilitate both the
Metadata, the so-called “data about data,” are a critical component of the capture and representation of this information as well as guard against its loss
data modeling process, and defining metadata standards requires careful in the longer term as part of a preservation system. Part of managing data
planning to ensure adequate description of data without excessive detail includes being prepared to update the underlying storage model over time as
(Duval et al., 2002). Metadata are a component of data modeling in that they needs and applications change, which includes the need to recode existing
describe the attributes of entities. One use of metadata is to foster interoper- data to incorporate new metadata and representations. Data volumes are also
ability across scientific communities so that information collected in one an important consideration. Terabyte scale data cannot be efficiently repre-
community can be effectively understood within a second community. sented via data-storage models that are abstracted and reduced in real time.
Metadata standards have seen increasing adoption within water-related disci- Current research is investigating aggregative storage models that instead
plines, with Geographic Information Systems (GIS) being a prime example. combine data in real time.
Controlled vocabularies within metadata can be used to restrict the language
or terms that can be used within certain entity attributes. The purpose is to
7.3 SERVICE-ORIENTED ARCHITECTURES FOR
have precise, machine interpretable vocabularies within each discipline so INTEGRATING DISTRIBUTED HYDROLOGIC DATA
that machine-mediated translation of vocabularies across disciplines is AND MODELS
possible without human intervention and with minimum computational
overhead. System integration is an important component of an HIS in that it defines
A key challenge in creating a hydrologic data model is the variety of infor- how components are united into a single system. One method for system
mation needed for conducting hydrologic analyses. Observational data, maps, integration that has seen broad adoption within HIS and cyberinfrastructures
remote-sensing data, model outputs, and other types of data are all required in other scientific domains is through the use of a service-oriented architec-
for many hydrologic studies. These data have fundamentally different proper- ture (SOA). In SOA, data, models, analysis, and visualization tools are
ties, with different data formats, structures, and semantics. Most of the work exposed as functional units, or services, accessible to one or more software
in the field to date has focused on data modeling for observational hydro- client applications (Castronova et al., 2013; Goodall et al., 2008; Goodall et al.,
logic data, but current work seeks to broaden hydrologic data modeling to 2011). Services can then be integrated into larger applications and workflows
include other data classes (Tarboton et al., 2014). More work is needed to to perform tasks (Ludascher et al., 2003). The benefit of service-orientation is
identify a list of all data models required within the field, identify overlaps that it provides a loose coupling between system components, where all
and gaps in data model needs, and coordinate the synchronization of data exchanges occur through a defined application programming interface. The
across data models. This extends to metadata and especially controlled internal workings of each service are independent and, therefore, can be
vocabularies used for describing key hydrologic concepts as well as concepts designed and evolved by different organizations along different design paths.
in fields that interface with hydrology (Beran and Piasecki, 2009). Like component-based architectures, interfaces act as agreed upon contracts
Data models have multiple physical implementations, including opera- between services and clients (Argent, 2004).
tional databases, archival formats, formats for transmission of data over the Web services are a particular case of SOA where services are exposed on
Web, and formats for object-oriented computer programs. Standardization of the Internet. Web services have been developed for many applications related
data models means less translation, confusion, and misinterpretation of infor- to hydrology, including exposing observational data, geospatial data, and
mation as it moves between subcomponents of an HIS and through the data array-based data stored using NetCDF, and other data types. Search tools for
life cycle. Data can be stored in many different ways, ranging from binary or locating observational data based on keyword, geographic region, and time
text files, to structured markup files, to relational database-management sys- constraints are available. Client applications able to read and operate from
tems (RDBMS). The desirable file format for a particular dataset depends on Web services are able to effectively use data streaming directly over the
the class of data being stored, its intended use (e.g., operational use versus Internet, making for practical uses in GIS and related fields. For example, a
archival), and any needs to optimize compression for data transmission number of widely used Web services standards for both spatial and temporal
between distributed systems or for optimal data access. Many file formats are data has been emplaced by the Open Geospatial Consortium, including Web
commonly used in hydrology. Radar data, remote-sensing data, or model Mapping Service, Web Feature Service, Web Processing Service, and Water
simulation output data might be commonly stored within a binary file format, Markup Language version 2.0 (WaterML2) (Michaelis and Ames, 2008;
such as the Network Common Data Form (NetCDF) (Rew and Davis, 1990), Michaelis and Ames, 2009; Taylor et al., 2013). These and other standards for
07_Singh_ch07_p7.1-7.10.indd 2 8/23/16 4:38 PM

The CUAHSI Hydrologic Information System as an Example HIS 7-3
HIS Central
Data discovery and
integration platform
Se
rvi
ice
s ce Met
rv tion an ad
e
s tra ng d d ata
ata egis esti
a
d Da ta t sea
t a e r rv ta hem rch
Me rvic g ha ca
r ts e m
e
S tal o eta
Ca da
ta
Data services
HydroServer HydroDesktop
Water data services
Data publication Data synthesis and
Spatial data services
platform research platform
Figure 7.2 CUAHSI HIS Architecture. [Source: Zaslavsky and Maidment, 2011.]
service interfaces and data formats have become critical in SOA so that ser- ODM databases. HydroServer includes several software utilities for working
vices, metadata catalogs, and software clients can remain interoperable as new with ODM databases, including tools for loading data, visualization, and
resources are added to the system. This is analogous to the World Wide Web, quality control of time series data. Each ODM database is then attached to a
were standards, such as HTTP and HTML are crucial to the success of the Water One Flow (WOF) Web service that transmits data in WaterML format
system. and consists of a set of service methods (e.g., GetSites, GetSiteInfo,
With the use of Web services for exposing data, locating data distributed GetVariableInfo, and GetValues) that can be called from many different pro-
across different Web services and integrating them becomes a challenge. gramming languages and software environments for retrieving data from an
Metadata catalogs provide this service by indexing metadata across Web ser- ODM database over the Internet. These methods have been implemented
vices into a searchable collection that helps users find the appropriate Web within a single Web application that is easily installed and configured. The
service for the information of interest. Keeping catalogs in sync as informa- coupling of ODM and WOF ensures that all of the hydrologic time series data
tion within the different Web services changes can be challenging, especially published on HydroServers are available on the Internet in a standard,
with dynamic datasets such as real-time data. Systems differ in their method interoperable format that is platform and programming language agnostic. In
and frequency with which metadata is kept up to date. Metadata might be addition, registration of WOF services running on a HydroServer with the
harvested on a weekly basis using Web services for most datasets, but in other CUAHSI HIS Central (http://hiscentral.cuahsi.org/) cataloging service
cases, metadata might be harvested more or less frequently than weekly ensures that they can be discovered and more easily accessed by HIS client
depending on the frequency with which the data catalog is being updated. applications, like HydroDesktop (http://www.hydrodesktop.org) (Ames et al.,
2012), the WaterML R package (Kadlec et al., 2015), or the CUAHSI Water
Data Center data client (http://data.cuahsi.org).
7.4 THE CUAHSI HYDROLOGIC INFORMATION SYSTEM HydroServers are also capable of publishing geospatial datasets as data
AS AN EXAMPLE HIS services using Esri’s ArcGIS Server or other web mapping applications. A
The CUAHSI HIS serves as an example of an HIS to support research in the Capabilities database contains a listing of all services that have been published
hydrologic science and engineering community (Maidment et al., 2009). The on a HydroServer (both WOF and geospatial services) as well as appropriate
CUAHSI HIS is organized in three main components: HydroServer, HIS metadata to describe each service. A configuration tool enables HydroServer
Central, and HydroDesktop (Fig. 7.2) (Tarboton et al., 2009). HydroServer administrators to edit the contents of the Capabilities database. As an alterna-
provides a software stack for data publication, HIS Central provides a means tive to accessing the data via web services, a suite of Web applications can be
for searching across a network of HydroServer instances, and HydroDesktop installed that makes the data published on a HydroServer available via a web
serves as a client application within which users can search, download, ana- browser. These Web applications use the Capabilities database to automati-
lyze, and visualize data across a network of HydroServer instances. cally configure their contents and include: (1) an Internet map that combines
GIS datasets and time series published using ODM and WOF services for a
7.4.1 Data Publication Using HydroServer region, (2) the Time Series Analyst for visualizing observational data, (3) a
HydroServer, which is a software stack for publishing hydrologic observations HydroServer website that describes the regions and services published on the
data on the Internet (Conner et al., 2013; Horsburgh et al., 2010; Tarboton et HydroServer, and (4) the Capabilities Web service, which publishes the capa-
al., 2010; Tarboton et al., 2009), was designed to enable producers of hydro- bilities (i.e., a metadata description of the services that the HydroServer con-
logic data to join the distributed CUAHSI HIS network of published water tains) of the HydroServer in XML format. The latest versions of the CUAHSI
data Web services by establishing their own server and publishing their own HIS software tools are available at http://wdc.cuahsi.org.
data. HydroServer software makes data available via Web services on the
Internet, and, when services are registered with the central cataloging service 7.4.2 Metadata Integration Using HIS Central
called HIS Central, the data become discoverable through HIS client applica- The central catalog for CUAHSI HIS is hosted in the Microsoft Azure cloud
tions, like HydroDesktop (Ames et al., 2009; Ames et al., 2012). HydroServer and administrated by the CUAHSI Water Data Center, an NSF-supported
includes software for management of data on the server, as well as Web appli- facility dedicated to running and curating said catalog. On a weekly basis,
cations for presenting data to consumers. HydroServer also enables publica- WOF services are harvested, and the metadata for each service are included
tion of geospatial datasets for a study area or region and the linkage of in the catalog. The results of this harvesting are exposed by catalog Web ser-
geospatial datasets with point observational datasets that are hosted on the vices that are very closely related to WOF services, and include extra param-
same server. Several versions of the HydroServer software stack have been eters to determine the measured properties that are of interest as well as the
developed for different platforms and purposes. For example, the original geographic extent of interest. Thus, the catalog can be considered as the hub
version, based on Microsoft Internet Information Server, SQL Server, and of a federated data service including all data sources as subcatalogs. Responses
ArcGIS Server (Horsburgh et al., 2010) has been ported to the Mono frame- from catalog services list metadata for the WOF services that contain data, so
work (Couch et al., 2014) and has been converted to PHP and MySQL for use that accessing data requires additional queries to the data sources. Thus, the
on a Linux-based operating system (Conner et al., 2013; Kadlec and Ames, catalog contains only metadata. HIS catalog holdings are surprisingly diverse
2012). and heterogeneous, and include servers running all versions of HydroServer’s
Figure 7.3 shows an overview of the major components of HydroServer. WOF services, as well as custom WOF services written by others to interface
Observations made at fixed monitoring points are loaded into one or more their data sources with the HIS catalog and HIS in general. Data from the
07_Singh_ch07_p7.1-7.10.indd 3 8/23/16 4:38 PM

ODM databases and

HydroServer website Time series analyst
WaterOneFlow web services
Capabilities
WaterOneFlow WaterOneFlow WaterOneFlow database
configuration
tool
ODM ODM ODM
WaterOneFlow
services
table
HydroServer capabilities
Web service HydroServer map
Capabilities
database
ArcGIS server spatial data services
Spatial
services
table
Figure 7.3 General overview of HydroServer components.
United States Geologic Survey are transformed into WOF services by a cus- in the workflow is to document metadata in the catalog that are not recorded
tom HydroServer running at the CUAHSI Water Data Center. on HydroServers and not available from WOF service harvesting, including
Registering a WOF service in the catalog requires several steps, as docu- information on funding sources used to collect the data.
mented in Fig. 7.4. In the figure, arrows between boxes and data sources repre- A final piece of the catalog system is the CUAHSI-controlled vocabulary: a
sent information flow. First, the WOF service is registered in the catalog as a community moderated vocabulary for metadata that is utilized both within
data source and an initial metadata harvest is performed. The next step is to HydroServer and the catalog. This includes a controlled vocabulary for vari-
map “variables”—the WOF service name for a measured property—to “con- able names that are recognized by the catalog and automatically tagged with
cepts” that are the catalog’s canonical names for measured properties. While appropriate concepts in the CUAHSI concept ontology. Users can propose
“variables” are defined for each data source, “concepts” are defined only in the new variable names in the vocabulary and thus add these terms to the variable
catalog and serve as canonical and searchable representations of what might be names that are automatically mapped to concepts, thus saving a step in the
locally defined variable names. This is done in two ways, by automatically map- data-publication workflow.
ping variable names that are defined in the CUAHSI controlled vocabulary to
their matching concepts, and by allowing the user to manually map noncon- 7.4.3 Data Access, Visualization, and Analysis Using
forming variable names to concepts in the catalog. In this way, the catalog HydroDesktop
allows data providers to utilize locally meaningful names for variables, and also Web services for data search and discovery through the CUAHSI WDC and
define generally meaningful concepts that serve as search targets within the data download from HydroServers are generally only available to end users
catalog. The tagger is shown in Fig. 7.5, and includes a graphical representation through additional client software tools. The Web user interfaces to interact
of the CUAHSI “ontology”: a classification system for the catalog’s concepts with data via HydroServer are simple clients for interacting with Web services.
created to make choosing appropriate concepts easier. A final and optional step More complex clients can be developed by writing scripts in MatLab, R, Python,
Represent data in Tag noncanonical Enter non-WOFS

Upload to Register service in Request catalog
upload compatible variables with metadata (funding
HydroServer catalog harvest
form concepts sources)
Harvesting results
HydroServer
Source data HydroCatalog
(ODM)
WOFS Catalog services
User data client
Figure 7.4 Data discovery workflow for CUAHSI HIS.
07_Singh_ch07_p7.1-7.10.indd 4 8/23/16 4:38 PM

HydroShare as a Next-Generation HIS Software 7-5
Figure 7.5 The HydroTagger allows tagging variable names with concepts that represent canonical names of properties measured.
or another programming language to execute data searches on the WDC cata- being added to the map for all dataset sites having data that meet the search
log or to connect to a HydroServer and download data in a WaterML format. criteria, and (6) selecting one or more of the discovered sites and downloading
Such an approach allows for automation of complex and repetitive tasks that are data from these sites using a “Download” button. Once data are downloaded
dependent on these services and their data. For example, the WaterML R pack- within HydroDesktop, they are stored in a local database that closely mimics the
age supports data search and download from directly within the R statistical HydroServer ODM database and are available locally on the client computer for
computing software on any operating system (Kadlec et al., 2015). visualization through the HydroDesktop graphing tool, exploration using the
In addition to custom client scripts and tools that a user may create, tabular view, and exporting to external software for further analysis.
CUAHSI has developed a GIS-based desktop application called HydroDesktop Building on the extensibility model within HydroDesktop, several addi-
that serves as a complete example program for accessing and executing the tional functions and capabilities have been added, with more expected in the
fundamental HIS web services (Ames et al., 2012). HydroDesktop is a free future. For example, the United States Environmental Protection Agency
and open-source software application. This means that the installation pack- Watershed Delineation extension uses a Web service hosted by the United
age, the executables, and the entire source code itself is free for users to States Environmental Protection Agency to automatically identify stream
download, use, inspect, modify, and repurpose. The software is also highly networks and watershed boundaries that represent drainage areas upstream of
extensible, allowing for third-party engineers and programmers to add new a user-selected point. This tool follows the Web services-oriented paradigm
functions to a “ribbon-style” toolbar. With a fully exposed source code and of HydroDesktop and enables rapid identification of catchment areas. The
extension interface, users and developers can learn from the implementation HydroR extension provides users with integrated access to the R statistical
approach and develop new clients as the need may arise. software (Horsburgh and Reeder, 2014). The extension links R to the underly-
HydroDesktop is available for use on Microsoft Windows operating sys- ing HydroDesktop database and extracts time series for analysis within the R
tems. The software can be installed using an executable installer available for environment. The HydroModeler extension provides an implementation of
downloaded from http://www.hydrodesktop.org/. The installation package the OpenMI modeling interface standard (Gregersen et al., 2007; Moore and
creates a new start menu item called “CUAHSI HIS” and a submenu identify- Tindall, 2005), allowing modelers to link data from the HIS network with a
ing the current HydroDesktop release version. On start-up, HydroDesktop model or group of simulation models (Goodall et al., 2010). Models can be
offers the user several default templates from which to choose, including developed as plug-and-play components (Castronova and Goodall, 2010)
North America, Europe, Global, and others. These templates include with input and output data automatically transferred between models and the
prepackaged GIS datasets to enable spatial searches for data. Additionally, an HydroDesktop database for further visualization and analysis due to a map-
online basemap tool allows the user to choose between many common online ping between the OpenMI standard and the ODM.
basemaps (e.g., Bing Maps) for background display.
A typical workflow through HydroDesktop would include: (1) identifying a 7.5 HYDROSHARE AS A NEXT-GENERATION HIS
spatial search domain by drawing a box on the map or selecting a polygon from SOFTWARE
a GIS dataset, (2) identifying search terms of interest from a large list of com-
mon water quantity and quality variables, (3) choosing an observation time The CUAHSI HIS was designed for sharing time series data, but hydrolo-
range of interest, (4) selecting WOF services from a list of cataloged CUAHSI gists often require other classes of data for their analysis. Indeed a strength
HydroServers, (5) executing the search function, resulting in place-markers of HIS is the ability to generate knowledge by integrating information from
07_Singh_ch07_p7.1-7.10.indd 5 8/23/16 4:38 PM

Figure 7.6 (1) Data are searched by area, keyword, time range, and data source through the HydroDesktop user interface. (2) Results are shown in a GIS map.
(3) Metadata for hydrologic series are displayed. (4) Data are downloaded into a local database.
multiple sources. However, for this potential to be fully realized, a broader picks up from here, allowing the user to next post the results (data and/or
class of data than those addressed in the CUAHSI HIS needs to be accom- model) to HydroShare as resources, retaining provenance information on the
modated and the collaborative nature of the production of hydrologic original data source (4). HydroShare will also support direct entry of new
research products needs to be supported. HydroShare is a research project resources. Upon ingestion, background actions parse metadata and enable
being pursued to address these needs. HydroShare is an online collaborative analysis based on rules and policies. The user shares posted resources with
system being developed for open sharing of hydrologic data and models colleagues (5), designating who has permission to access the resources. A
(http://www.hydroshare.org). The goal of HydroShare is to enable scientists group collaborates on refining the analysis, model, or result. HydroShare
to easily discover and access hydrologic data and models, retrieve them to tracks provenance supporting reproducibility and transparency. After itera-
their desktop, or perform analyses in a distributed computing environment tion, the result is finalized and submitted for publication (6). At this point, the
that may include grid, cloud, or high-performance computing (HPC) model resources produced (data, model, workflow, and paper) are made immutable,
instances as necessary. The HydroShare development team is a collabora- access is opened, and permanent persistent identifiers (e.g., DOIs) are
tion of 10 institutions supported by NSF. assigned. The data may be moved to a permanent repository under the aus-
The vision behind the development of HydroShare is to enable more rapid pices of the CUAHSI Water Data Center (7).
advances in hydrologic understanding through collaborative data sharing, In HydroShare, the term “data” is interpreted in a broad sense to mean any
analysis, and modeling. Understanding will be advanced through the ability product or resource that is an object to be shared in HydroShare. This includes
to integrate information from multiple sources. Products (data, results, and data in a number of formats (time series, geographic features, geographic
models) can then be published as new resources that can be shared with col- grids, space time arrays, model programs, model instances, scripts, and collec-
laborators. HydroShare uses the Web to deliver its functionality in line with tions of resources). A structured approach is used for the representation of
software as a service and data as a service trends in cyberinfrastructure. each resource type using a formal resource data model (Tarboton et al., 2014).
Sharing and publication of data using the CUAHSI HIS is challenging because This approach discriminates HydroShare from some other resource sharing
it requires data publishers to establish or gain access to a HydroServer and for systems where objects with any described format are accepted. The structure
the recipient to use the HydroDesktop client software. Both require software of HydroShare resources enables their use in analysis and modeling functions
installation and have platform dependencies. HydroShare is Web-based, so that HydroShare can include a collection of value added analysis tools that
delivering its primary functionality through a Web browser client and web operate on these resources. These value-added tools are the incentive to
services that can be used on any computing platform. attract users to the system. They provide hydro-value-added functionality
HydroShare development is being driven by a collaborative data analysis over the functionality available in general purpose digital object sharing
and publication use case that extends the existing CUAHSI HIS data-sharing systems.
functionality into a dynamic, collaborative environment that leads to the The HydroShare architecture is comprised of a website and Web service
eventual archival publication of data and models (Fig. 7.8) (Tarboton et al., interface using the Representational State Transfer protocol to expose as
2014). At (1) data are observed and then loaded (2). In CUAHSI HIS, data are much functionality as possible via services accessible by client software. The
loaded into an ODM relational database on a HydroServer that publishes HydroShare website, Web service interface, and social media functions are
them using WOF services. Metadata are harvested by the HIS Central catalog being developed using the Django web application framework. The inte-
and support geographic and context-based data discovery. A HydroDesktop grated Rule-Oriented Data System is being used to manage federated data
client user (3) discovers, downloads, and analyzes the data, or uses them in a content and perform rule-based background actions on data and model
model. Steps 1–3 are supported by the existing CUAHSI HIS. HydroShare resources, including parsing to generate metadata catalog information and
07_Singh_ch07_p7.1-7.10.indd 6 8/23/16 4:38 PM

HydroShare as a Next-Generation HIS Software 7-7
Figure 7.7 (1) HydroDesktop displays data in tabular and (2) graphical form and includes a number of scientific graphing options.
the execution of models and workflows. HydroShare thus provides a new, publishing, sharing, and running the Soil Water Assessment Tool (SWAT), is
Web-based system for advancing model and data sharing. It provides access an example of such a system. SWATShare provides modelers with simplified
to a wider range of hydrologic data types, including models, using standards- access to XSEDE (http://xsede.org) and cloud resources for running SWAT
based data formats and interfaces. The standard representation of models models, including sensitivity analysis, autocalibration, and spatiotemporal
facilitates model sharing and discovery functionality as well as transmittal visualization of model results. Specifically, the SWATShare interface enables
and execution of models in HPC systems. searching of available SWAT models through a geographic view of water-
An important goal of HydroShare is to provide more hydrologists with sheds, uploading of SWAT models, editing of metadata information for
access to HPC resources to support modeling. One way HydroShare will do uploaded SWAT models, and running of uploaded SWAT models and output
this is by interoperating with existing software systems that provide HPC visualization. Such an infrastructure is expected to reduce the redundancy in
resources to hydrologists. SWATShare, a cyberinfrastructure designed for creating multiple SWAT models for the same area, and, at the same time,
Figure 7.8 Collaborative data analysis use case driving the development of the HydroShare system.
07_Singh_ch07_p7.1-7.10.indd 7 8/23/16 4:38 PM

Figure 7.9 Collaborative research and access to HPC through SWATShare.
encourage enhancement of existing models by incorporating modifications heterogeneous data collections so that they appear to an end user as a single,
from multiple users with diverse modeling needs. Similarly, by using tools, semantically and syntactically consistent data collection. Current work on the
such as SWATShare, modelers can execute multiple simulations, including CUAHSI HydroShare system seeks to advance on the CUAHSI HIS design by
model calibrations that can sometimes run for weeks to perform one model supporting a broad range of data classes through the adoption of a resource-
calibration, without consuming their local computing resources. centric data model approach. HydroShare will support different data classes,
In addition to increasing computational efficiency, SWATShare can enable each with unique metadata but sharing core common metadata terms derived
collaborative science more effectively and efficiently through its model shar- from already established metadata standards. The CUAHSI efforts represent
ing, publishing, and visualization capabilities. Figure 7.9 shows an example of only one of the many large-scale HIS efforts being actively developed across
collaborative research through SWATShare. In Fig. 7.9, Users A and B are the globe.
climate and watershed hydrologists, who are interested in studying the impact Ultimately, continued progress in the design of HIS will come from both
of climate change and agricultural-management practices on hydrology, increased demand for new tools and capabilities as well as growth within the
respectively. User C, who is interested in exploring SWAT model and its community of scholars with a primary focus in hydroinformatics research.
parameters, can use the models developed by Users A and B to get to the Hydroinformatics is a growing subdiscipline with research interests in
research directly without spending time to create the models for addressing addressing the data and computational challenges facing the hydrology and
his/her research questions. water resources communities. Hydroinformatics includes researchers focused
SWATShare will interoperate with HydroShare by using HydroShare to on many aspects of data-intensive hydrologic analysis from machine-learning
store model data with defined metadata for SWAT model instances. Users will techniques, to data mining, to the design of next-generation HIS software
be able to push model instances directly from SWATShare into HydroShare systems. Like other scientific disciplines facing similar data and computa-
for sharing model results with the broader community. Modelers will be able tional challenges, progress requires individual researchers with a primary
to push a model instance resource stored within HydroShare directly to focus and training at the interface of the domain science and computer and
SWATShare for visualization and execution. Both systems will benefit from information sciences. A new generation of researchers with such interdisci-
the provided interoperability by being able to focus on their own core func- plinary training is needed to create novel computational techniques and tools
tionality and use cases while still allowing the other system to complement that advance scientific understanding and sustainable management of water
and expand on its core functionality. HydroShare will be able to support shar- systems.
ing of other hydrologic model instances through standardization of model
metadata and ontologies (Elag and Goodall, 2013; Morsy et al., 2014).
REFERENCES
Ames, D. P., J. S. Horsburgh, J. L. Goodall, T. L. Whiteaker, D. G. Tarboton,

7.6 CONCLUSION
and D. R. Maidment, “Introducing the open source CUAHSI Hydrologic
HIS are integrated collections of software tools and hardware systems that Information System desktop application (HIS Desktop),” Proceedings of the
address steps along the data life cycle. From sensors in the field, to transmis- 18th World IMACS Congress and MODSIM09 International Congress on
sion systems, publication systems, integration approaches, analysis tools, Modelling and Simulation, Modelling and Simulation Society of Australia and
curation policies, and archiving systems, the goal is to more effectively use New Zealand and International Association for Mathematics and Computers
computers to automate many of the steps now done by scientists. The need in Simulation, MSSANZ, pp. 4353–4359, 2009.
for this automation is at least twofold. First is that the volume of data being Ames, D. P., J. S. Horsburgh, Y. Cao, J. Kadlec, T. L. Whiteaker, D. and
generated now through global modeling studies and sensing technologies Valentine, “HydroDesktop: web services-based software for hydrologic data
requires new tools for analysis. The current approach of downloading data discovery, download, visualization, and analysis,” Environmental Modelling
for analysis on personal computers simply will not scale as data volumes and Software, 37: 146–156, 2012.
continue to increase. The second need is to support data-intensive studies Argent, R. M., “An overview of model integration for environmental
so that they are efficient and reproducible through automation of tedious applications—components, frameworks and semantics,” Environmental
data gathering, translation, transformation, and other general processing Modelling and Software, 19 (3): 219–234, 2004.
tasks. Without these tools, scientific progress will be inhibited as data vol- Beran, B. and M. Piasecki, “Engineering new paths to water data,”
umes increase. Computers and Geosciences, 35 (4): 753–760, 2009.
The CUAHSI HIS provides an example of an HIS that focuses primarily on Castronova, A. M. and J. L. Goodall, “A generic approach for developing
the publication, access, and use of point-based observational time series process-level hydrologic modeling components,” Environmental Modelling
data. The system has proven successful in integrating autonomous and and Software, 25 (7): 819–825, 2010.
07_Singh_ch07_p7.1-7.10.indd 8 8/23/16 4:38 PM

REFERENCES 7-9
Castronova, A. M., J. L. Goodall, and M. M. Elag, “Models as web services Lawrence, B., C. Jones, B. Matthews, S. Pepler, and S. Callaghan, “Citation
using the open geospatial consortium (OGC) web processing service (WPS) and peer review of data: moving towards formal data publication,” Interna-
standard,” Environmental Modelling and Software, 41: 72–83, 2013. tional Journal of Digital Curation, 6: 4–37, 2011.
Chen, P. P.-S., “The entity-relationship model—toward a unified view of Ludascher, B., I. Altintas, and A. Gupta, “Compiling abstract scientific
data.” ACM Transactions on Database Systems (TODS), 1 (1): 9–36, 1976. workflows into web service workflows,” Proceedings of the 15th International
Conner, L. G., D. P. Ames, and R. A. Gill, “HydroServer Lite as an open Conference on Scientific and Statistical Database Management, IEEE,
source solution for archiving and sharing environmental data for indepen- pp. 251–254, Cambridge, Massachusetts, 2003.
dent university labs,” Ecological Informatics, 18: 171–177, 2013. Lynch, C. “Big data: how do your data grow?” Nature, 455 (7209): 28–29,
Conner, L. G., D. P. Ames, and R. A. Gill, “HydroServer Lite as an open 2008.
source solution for archiving and sharing environmental data for indepen- Maidment, D. R. Arc hydro: GIS for Water Resources, ESRI Press, Redlands,
dent university labs,” Ecological Informatics, 18: 171–177, 2013. CA, 2002.
Couch, A., R. Hooper, J. Pollak, and M. Seul, “Enabling water science at the Maidment, D. R., R. P. Hooper, D. G. Tarboton, and I. Zaslavsky, “Accessing
CUAHSI Water Data Center,” International Environmental Modelling and and sharing data using CUAHSI water data services,” Hydrology, Hydrogeology
Software Society (iEMSs) 7th International Congress on Environmental and Water Resources, edited by I. Cluckie, Y. Chen, V. Babovic, L. Konikow, A.
Modelling and Software, edited by D. P. Ames, N. W. T. Quinn, and A. E. Mynett, S. Demuth, and D. A. Savic, IAHS Publication no. 331, Proceedings
Rizzoli, IEMSS, San Diego, CA, 2014, p. 7. of Symposium JS4 held in Hyderabad, India, 2009, pp. 213–223.
D’Agostino, D., Clematis, A., Galizia, A., Quarati, A., Danovaro, E., Mayer-Schönberger, V., and K. Cukier, Big Data: A Revolution that Will
Roverelli, L., Zereik, et al., “The DRIHM project: a flexible approach to inte- Transform How We Live, Work, and Think, Houghton Mifflin Harcourt,
grate HPC, grid and cloud resources for hydro-meteorological research,” London, England, 2013.
Proceedings of the International Conference for High Performance Computing, Michaelis, C. D., and D. P. Ames, “Web Mapping Service (WMS), Web
Networking, Storage and Analysis, IEEE Press, pp. 536–546, New Orleans, Feature Service (WFS), and Web Processing Service (WPS),” Encyclopedia
Louisiana, 2014. of GIS, edited by S. Shekhar, H. Xiong, Springer, New York, 2008,
Duval, E., W. Hodgins, S. Sutton, and S. L. Weibel, “Metadata principles pp. 1259–1261.
and practicalities,” D-lib Magazine, 8 (4): 16, 2002. Michaelis, C. D. and D. P. Ames, “Evaluation and implementation of the
Elag, M. and J. L. Goodall, “An ontology for component-based models of OGC web processing service for use in client-side GIS,” Geoinformatica, 13:
water resource systems,” Water Resources Research, 49 (8): 5077–5091, 2013. 109–120, 2009.
Faundeen, J. L., T. E. Burley, J. A. Carlino, D. L. Govoni, H. S. Henkel, S. L. Moore, R. V. and C. I. Tindall, “An overview of the open modelling
Holl, V. B. Hutchison, et al., “The United States Geological Survey Science interface and environment (the OpenMI).” Environmental Science and Policy,
Data Lifecycle Model,” U.S. Geological Survey Open-File Report 2013–1265, 8 (3): 279–286, 2005.
p. 4, http://dx.doi.org/10.3133/ofr20131265, 2014. Morsy, M. M., J. L. Goodall, C. Bandaragoda, A. M. Castronova, and
Goodall, J. L., A. Castronova, M. Elag, and M. Ercan, “An integrated J. Greenberg, “Metadata for describing water models,” International
modeling environment within the CUAHSI Hydrologic Information System,” Environmental Modelling and Software Society (iEMSs) 7th International
Proceedings of the AGU Fall Meeting Abstracts, p. 1237, San Francisco, Congress on Environmental Modelling and Software, edited by D. P. Ames,
California, 2009. N. W. T. Quinn, and A. E. Rizzoli, IEMSS, San Diego, CA, 2014, p. 7.
Goodall, J. L., J. S. Horsburgh, T. L. Whiteaker, D. R. Maidment, and I. Overpeck, J. T., G. A. Meehl, S. Bony, and D. R. Easterling, “Climate data
Zaslavsky, “A first approach to web services for the National Water Information challenges in the 21st century,” Science, 331 (6018): 700–702, 2011.
System,” Environmental Modelling and Software, 23 (4): 404–411, 2008. Rew, R. and G. Davis, “NetCDF: an interface for scientific data access,”
Goodall, J. L., B. F. Robinson, and A. M. Castronova, “Modeling water Computer Graphics and Applications, IEEE, 10 (4): 76–82, 1990.
resource systems using a service-oriented computing paradigm,” Environmen- Stenson, M., P. Fitch, J. Vleeshouwer, A. Frost, Q. Bai, J. Lerat, B. Leighton,
tal Modelling and Software, 26 (5): 573–582. et al., “Operationalising the Australian Water Resources Assessment (AWRA)
Gregersen, J., P. Gijsbers, and S. Westen, “OpenMI: open modelling system,” Proceedings of the Water Information Research and Development
interface,” Journal of Hydroinformatics, 9 (3): 175–191, 2007. Alliance Science Symposium, Melbourne, Australia, 2011.
Halevy, A., A. Rajaraman, and J. Ordille, “Data integration: the teenage Tarboton, D., J. Horsburgh, K. Schreuders, D. Maidment, I. Zaslavsky, and D.
years,” Proceedings of the 32nd international conference on Very large data Valentine, “The HydroServer platform for sharing hydrologic data,” Proceedings
bases, VLDB Endowment, pp. 9–16, Seoul, Korea, 2005. of the AGU Fall Meeting Abstracts, San Francisco, California, 2010.
Hey, A. J. G., S. Tansley, and K. M. Tolle, The Fourth Paradigm: Data-Intensive Tarboton, D. G., J. S. Horsburgh, R. Idaszak, J. Heard, D. Valentine,
Scientific Discovery, Microsoft Research, Redmond, Washington, 2009. A. Couch, D. P. Ames, et al., “A resource centric approach for advancing
Horsburgh, J. S., and S. L. Reeder, “Data visualization and analysis within a collaboration through hydrologic data and model sharing,” 11th International
Hydrologic Information System: integrating with the R statistical computing Conference on Hydroinformatics HIC 2014, HIC 2014, New York, 2014,
environment,” Environmental Modelling and Software, 52: 51–61, 2004. p. 8.
Horsburgh, J. S., D. G. Tarboton, D. R. Maidment, and I. Zaslavsky, “A Tarboton, D. G., J. S. Horsburgh, D. R. Maidment, T. Whiteaker, I. Zaslavsky,
relational model for environmental and water resources data,” Water Resources M. Piasecki, J. L. Goodall, et al., “Development of a community hydrologic
Research, 44: 12, 2008. information system,” 18th World IMACS Congress and MODSIM09 Interna-
Horsburgh, J. S., D. G. Tarboton, K. A. Schreuders, D. R. Maidment, tional Congress on Modelling and Simulation, edited by R. S. Anderssen, R. D.
I. Zaslavsky, and D. Valentine, “HydroServer: a platform for publishing space- Braddock, and L. T. H. Newham, eds., MSSANZ, Cairns, Australia, 2009,
time hydrologic datasets,” Proceedings of the AWRA Specialty Conference on pp. 988–994.
GIS and Water Resources, Orlando, Florida, 2010. Tarboton, D. G., R. Idaszak, J. S. Horsburgh, J. Heard, D. P. Ames, J. L.
Janée, G., J. Mathena, and J. Frew, “A data model and architecture for long- Goodall, L. Band, “HydroShare: advancing collaboration through hydrologic
term preservation,” Proceedings of the 8th ACM/IEEE-CS Joint Conference on data and model sharing,” International Environmental Modelling and Software
Digital Libraries, ACM Press, New York, pp. 134–144, 2008. Society (iEMSs) 7th International Congress on Environmental Modelling and
Kadlec, J., and D. P. Ames, “Development of a lightweight hydroserver and Software, edited by D. P. Ames, N. W. T. Quinn, and A. E. Rizzoli, IEMSS, San
hydrologic data hosting website,” Proceedings of the AWRA Spring Specialty Diego, CA, 2014, p. 7.
Conference on GIS and Water Resources, New Orleans, Louisiana, 2012. Taylor, P., S. Cox, G. Walker, D. Valentine, and P. Sheahan, “WaterML2. 0:
Kadlec, J., B. StClair, D. P. Ames, and R. A. Gill, “WaterML R package for development of an open standard for hydrological time-series data exchange,”
managing ecological experiment data on a CUAHSI HydroServer,” Ecological Journal of Hydroinformatics, 16(2), 425–446, 2014.
Informatics 28: 19–29, 2015. Zaslavsky, I. and D. R. Maidment, “Service orientation in the design of a
Kumar, P., M. Folk, M. Markus, and J. C. Alameda, Hydroinformatics: Data community hydrologic information system.” Geoinformatics: Cyberinfrastruc-
Integrative Approaches in Computation, Analysis, and Modeling, CRC Press, ture for the Solid Earth Sciences, edited by G. R. Keller and C. Baru, Cambridge
Boca Raton, Florida, 2006. University Press, Cambridge, 2011, pp. 193–209.
07_Singh_ch07_p7.1-7.10.indd 9 8/23/16 4:38 PM

Chapter 8
Remote Sensing Techniques and
Data Assimilation for Hydrologic
Modeling
BY
ILIANA E. MLADENOVA, GREY NEARING, JOHN D. BOLTEN, AND VENKAT LAKSHMI
ABSTRACT algorithm development and the broad availability of satellite platforms that
measure the electromagnetic spectrum provide the necessary basis for the
Water is a vital natural resource. Society is in constant need of adequate esti-
retrieval of a large number of geophysical parameters, including digital eleva-
mates of available water supply and expected demand, and timely knowledge
tion models (DEMs), precipitation, surface temperature, evapotranspiration
of the potential occurrence of extreme hydrologic events such as floods and
(ET), soil moisture, runoff, ground water storage, land cover/use and various
droughts. Thus, the accurate monitoring and forecasting of the various com-
snow-, ice-, and vegetation-related characteristics. The most common sensor
ponents of the Earth’s hydrologic cycle is essential. This can be done by using
types that are currently used to infer the aforementioned variables are multi-
hydrologic models, remote sensing techniques or by utilizing a combined
spectral scanners that measure the spectral reflectance of the Earth’s surface
approach, which allows us to integrate the available satellite observations into
over a variety of narrow bands, thermal sensors that measure the emitted
the hydrologic models using various data assimilation techniques. This
radiation in the thermal portion of the electromagnetic spectrum and
chapter provides of overview of remote sensing and data assimilation tech-
microwave sensors that acquire observations in the longest wavelengths of the
niques and satellite systems used for hydrologic modeling, water resource
spectrum (1 mm to 1 m).
management, and water-related disaster monitoring and forecasting.
8.1 INTRODUCTION 8.3 REMOTE SENSING IN HYDROLOGICAL SCIENCES:

A HISTORICAL PERSPECTIVE
Our relationship with water is complex and delicate. We depend on regular
access to clean water for sustaining life, and our way of life. Accurately and Though there had been numerous aircraft- and ground-based surveys in mid
frequently characterizing various forms of water in the Earth’s hydrologic 1900s, the true start of global remote sensing in support of earth sciences and
cycle enables improved understanding of this relationship. Thus, society hydrology can be linked to the launch of systems such as Landsat (launched
requires reliable tools for assessing water availability, water demands, moni- in 1972) and the Advanced Very High Resolution Radiometer (AVHRR)
toring and preparing for extreme hydrological events as well as estimating (launched in 1978) (Holben, 1986; Loveland et al., 2000). Landsat and
human impacts on water quality and improving the management and main- AVHRR and the AVHRR’s successor MODIS (Moderate Resolution Imaging
tenance of ecosystems. This is done most efficiently by frequently monitoring Spectroradiometer), launched on two platforms, Aqua and Terra, in 1999 and
the water cycle and characterizing current and future behavior of the hydro- 2002, respectively, have been instrumental on quantifying the terrestrial veg-
logic system via stochastic or deterministic models. To this end, the monitor- etation and the surface temperature for the past 30 years. The advent of
ing, estimating, and forecasting of hydrologic state variables can be improved microwave remote sensing for water-related hydrological variables is
by assimilating observations from direct (i.e., ground) and remote (i.e., air associated with the launch of the Scanning Multichannel Microwave
and satellite)-based sensors able to infer or measure hydrologic state variables. Radiometer in 1978 (Njoku et al., 1980) and the Special Sensor Microwave
This chapter gives an introduction to notable remote-sensing platforms for Imager in 1987 (Hollinger et al., 1990). More recently, advances have been
soil moisture, surface temperature, vegetation, groundwater, and precipitation made to optimize the operational systems frequency-, resolution-, and cover-
vetted in hydrologic modeling applications, and an overview of relevant data age-wise and made the sensors more susceptible to the hydrological variable
assimilation concepts and methods for these applications. that is monitored. Soil moisture measurements, which were previously
derived using C- or X-band data are now available from two L-band missions,
Soil Moisture Ocean Salinity (SMOS) and Soil Moisture Active Passive
8.2 REMOTE SENSING THEORY
(SMAP), launched in 2009 and 2015, respectively. Building upon the legacy of
Remote sensing can be broadly defined as obtaining observations without the the Tropical Rainfall Measurement mission (TRMM), the Global Precipitation
sensor actually being in contact with the object being sensed. Each compo- Measurement (GPM) mission was launched in 2014, which provides global
nent of the hydrologic cycle has its distinct signal and responds to different coverage at 10 km resolution within 3 h. However, still many of the available
regions of the electromagnetic spectrum. By detecting the electromagnetic systems characterize with resolutions that are not suitable for fine-scale or
energy emitted or reflected from the Earth’s surface, remote sensing allows field-based studies. Therefore, great research efforts have been put toward the
the rapid and timely collection of data that can be used to provide hydrome- development of downscaling techniques (Anderson et al., 2004; Atkinson,
teorological information over large areas on a routine basis. Advances in 2013; Fang et al., 2013).
8-1
08_Singh_ch08_p8.1-8.6.indd 1 8/22/16 11:35 AM

8-2 Remote Sensing Techniques and Data Assimilation for Hydrologic Modeling
8.4 REMOTE SENSING: METHODS AND TECHNIQUES in regional and continental water storage (Schmidt et al., 2006), and have
8.4.1 Soil Moisture been widely used for estimating regional water storage variations and ground-
water withdrawal (Zaitchik et al., 2010; Swenson et al., 2003; Rodell, 2009),
Soil moisture is one of the smallest components of the hydrologic cycle; how- monitoring terrestrial water storage in snow-dominated basins (Forman et al.,
ever, it plays an important role in all of the terrestrial cycles. Due to their 2012), and glacier melting (Velicogna and Wahr, 2006).
strong sensitivity to the dielectric properties of the soil, soil moisture is most
often monitored using microwave observations acquired at frequency of 8.4.3 Surface Temperature and Vegetation Indices
10 GHz and lower (Ulaby et al., 1986).
MODIS
AMSR-E The MODIS provides a total of 44 data products related to the land, ocean,
The Advanced Scanning Microwave Radiometer (AMSR-E) was the first sat- and atmosphere including measurements of surface temperature (Ts),
ellite mission to incorporate soil moisture as an operational product (Njoku Normalized Difference Vegetation Index (NDVI), and Leaf Area Index (LAI).
et al., 2003; Shibata et al., 2003). AMSR-E was operational for almost a decade MODIS is onboard both the Terra and Aqua satellite platforms, which were
(2002–2011) and provided brightness temperature observations at six fre- launched on December 1999 and May 2002, respectively. Both platforms have
quencies ranging between 6.9 and 89 GHz twice a day (1:30 AM/PM). sun-synchronous polar orbits with equatorial crossing times of 10:30 AM
Multiple AMSR-E channels (i.e., 6, 10, and 18 GHz) were used in conjunction (Terra) and at 1:30 PM (Aqua). MODIS utilizes 36 spectral bands between
with a radiative transfer model to simultaneously retrieve surface soil mois- 0.405 and 14.385 μm. Its spatial resolution at nadir is 250 m (band 1~2),
ture, vegetation water content, and effective surface temperature (Njoku et al., 500 m (band 3~7), and 1000 m (band 8~36) (Justice et al., 1998, 2002). The
2003; Jackson et al., 2010). algorithms to derive MODIS NDVI, Ts, LAI, and Land Cover Maps are well
established and extensively evaluated: NDVI (Tucker, 1979; Myneni et al.,
SMOS 1995); LAI (Myneni et al., 2002); Land cover classification (DeFries et al.,
Soil moisture is best observed using brightness temperature data collected at 1998; Friedl et al., 2002) and surface temperature (Justice et al., 1998; Wan
~1.4 GHz (Wang and Choudhury, 1981). This consequently motivated the and Li, 1997). The NDVI and LAI are generally used as biweekly composites
launch of the first L-band-based satellite, the SMOS sensor by the European in order to minimize effects of cloud contamination.
Space Agency in November 2009. SMOS employs a Microwave Imaging GEOS
Radiometer using Aperture Synthesis, 2D interferometric L-band radiometer
operating at 1.4 GHz with 69 antenna receivers distributed on a Y-shaped Ts and several other products (e.g., radiance, rainfall rate, snow cover, etc.)
deployable antenna array and central hub. In addition to soil moisture, SMOS relevant to monitoring a number of hydrological processes are estimated
provides global estimates of ocean salinity (Kerr et al., 2001). SMOS operates using observations acquired from the Geostationary Operational
in a Sun-synchronous orbit with an equatorial overpass time of 6 AM and Environmental Satellite (GOES). GOES has been fundamental for weather
6 PM. The SMOS retrieval algorithm derives soil moisture at 43 km and it has forecasting since the early 1970s. It is in geosynchronous orbit and measures
been demonstrated to have accuracy of approximately 0.04 m3/m3 (Al Bitar et five wavelengths, including in the infrared and the visible range.
al., 2012; Pan et al., 2012). ET cannot be measured directly using a remote-sensing technique; how-
ever, approaches that combine satellite-based estimates and energy-based
AMSR2 modeling (Anderson et al., 1997) use remote-sensing estimates of surface and
air temperature, radiance, cloud cover, albedo, vegetation dynamics, and land
The extensive data record collected by AMSR-E is currently being extended
cover obtained from MODSI, GOES, and Landsat to estimate ET.
by the AMSR2 sensor onboard the GCOM-W1 (Global Climate Change
Mission) platform, which was launched in July 2012. AMSR2 provides almost 8.4.4 Precipitation
coverage of the Earth every 2 days, and technically, it is very similar to its
Precipitation is a key component of the hydrologic cycle and can be estimated
predecessor AMSR-E. It offers a few additional features, such as an extra
using data from various wavelengths. TRMM, GPM, AMSR-E, and AMSR2,
7.3 GHz channel for Radio Frequency Interference mitigation and improved
for example, utilize microwave frequencies, while the GOES and Meteosat
calibration using the hot-load.
utilize frequencies in the visible and infrared portion of the spectrum. The
SMAP1 later is very susceptible to the presence of clouds, enabling precipitation esti-
mation by relating the cloud characteristics to the rain rates (Adler and Negri ,
Launched in January 2015, the SMAP mission is an L-band-based mission 1988; Cheng et al., 1993). Cloud cover is transparent at microwave frequen-
that was developed by NASA in response to the National Research Council’s cies, thus in this case, rain rate is estimated directly from the energy emitted
Decadal Survey (Entekhabi et al., 2010). It provides global observations of soil from the water droplets (Huffman et al., 2007). Because there are several ways
moisture and freeze/thaw at unprecedented spatial resolution. By combining to estimate precipitation, the available remote sensing products differ in accu-
two microwave sensors, a synthetic aperture radar (centered at 1.41 GHz) and racy, spatial resolution, and sampling frequency. For example, IR-based prod-
a radiometer (centered at 1.26 GHz), SMAP addresses this shortcoming and ucts characterize with higher temporal resolution as compared to the
generate several different soil moisture products at three different resolutions, microwave-based only products. Often global satellite-based products are
3 km for the radar only product, 9 km for the combined radar-radiometer generated by merging observations from various systems and are calibrated
product, and 36 km for the radiometer only product. In addition to these using gauge-based data, which allows to overcome the limitations of the indi-
products, SMAP will provide an L4 surface and root zone soil moisture product, vidual remote sensing systems (Huffman et al., 2007; Joyce et al., 2011). Kidd
which will be generated by integrating the SMAP based retrievals into a land and Huffman (2011) offer an extensive overview of global precipitation mea-
surface model driven with meteorological forcing data. surements. Remote sensing of precipitation, although a well-established con-
cept, is quite difficult. Ongoing research aims to enhance the accuracy and
8.4.2 Groundwater resolution of the available products.
TRMM
GRACE
The Tropical Rainfall Measurement Mission Microwave Imager (TRMM
The Gravity Recovery and Climate Experiment (GRACE) is a joint partnership
TMI) was launched in 1997. TMI has an orbit that extends from 35oN to 35oS,
between the National Aeronautics and Space Administration (NASA) in the
limiting the coverage to the Tropics. The sensors on board the TMI have nine
United States and Deutsche Forschungsanstalt für Luft und Raumfahrt in
horizontally and vertically polarized channels with frequencies from 10.7
Germany. It is a dedicated dual-satellite mission, launched in March 2002 to
GHz to 85.5 GHz. The TMI completes an orbit every 91 min, making 15.7
make global, highly accurate measurements of Earth’s gravity field over a
orbits per day. Thus, the TMI overpasses any particular area at different local
spatial range ~400 km every month (Tapley et al., 2004). Small spatial and
times each day, providing one or more observations in the target area daily.
temporal gravitational variations can be detected by GRACE by measuring
The TRMM repeats its orbital cycle every 42 days, passing over the same spot
the slight alterations of the orbits of the twin spacecrafts. Small changes in the
at the same local time.
Earth’s gravity caused by gas, liquid, or solid (i.e., ice) mass transports are
reflected in the Earth’s gravity field and recorded by GRACE. The GRACE GPM
observations of the time-variable gravity fields are used to quantify variations The GPM mission, launched in 2014, is designed to produce a variety of
1
At the time of publication the SMAP active system was no longer operational. The precipitation products. It includes a dual-frequency Precipitation Radar
science team is working on and will provide an enhanced product using a different (DPR) and a passive-based GPM Microwave Imager (GMI). The DPR collects
methodology. coaligned observations at Ku-band and Ka-band, while the GMI includes
08_Singh_ch08_p8.1-8.6.indd 2 8/22/16 11:35 AM

Data Assimilation: Theory 8-3
13 microwave channels ranging between 10 and 189 GHz. The combination 8.5.2 Approximating the Full Bayesian Solution
of these two instruments and multiple observational frequencies allows to Bayes’ theorem is intractable for all but the simplest (e.g., linear Gaussian)
simultaneously estimating different rain rates and determine the size of the models, and the most common methods for approximating a solution to
precipitation particles. GPM covers the Earth from 65°S to 65°N and is Eq. (8.1) involve one or more of the following: (i) independence assumptions,
capable of providing precipitation estimates every 2–3h. The satellite flies at (ii) variational or sampling approaches, and (iii) parametric approximations.
an altitude of 407 km in a non–sun-synchronous orbit.
Filters versus Smoothers
8.4.5 LIDAR
The biggest practical issue is that the posterior distribution has dimension
Light Detection and Ranging (LIDAR) is a remote-sensing technique that equal to the product of the size of the domain by the number of states in the
measures distances (or ranges) to the Earth by sending pulses of light (Jensen, model by the number of time steps—this product is generally very large. To
2007). LIDAR systems are most often mounted on aircraft platforms and reduce the size of the inference problem, we can use the Markov property of
typically consist of a laser, a scanner, and a GPS receiver, which enables the Eq. (8.2) and treat the present state as independent of future observations so
collection of precise, high-resolution, three-dimensional information about that at each time step t the full solution is approximated by a series of time
the Earth’s shape (Campbell, 2002). The most significant application of step solutions like:
LIDAR in hydrologic is its use for the generation of DEM (Murphy
et al., 2007). p( zt | xt , ut , ht ) p( xt | xt −1 , ut , m, y1:t −1h1:t −1 )
p( x1:t | z1:t , m, h1:t ) = ∫ p(ut )dut
8.5 DATA ASSIMILATION: THEORY

∫ p(zt | xt , ut , ht ) p(xt | xt −1 , ut , m, y1:t −1h1:t −1 )dxt
(8.4)
Data assimilation is the process of combining information from models and
observations into estimates of the state of a dynamic system (Reichle, 2008). Equation (8.4) is applied sequentially to estimate a series of refined prob-
In this case, we are generally concerned with estimating mass and/or energy ability distributions over the modeled states at each step. The samples from
states of the hydrologic cycle from remote sensing data. The model is con- the posterior distribution at each time step are run through the model using
tinuous—in both time and space between remote-sensing observations, and Eq. (8.2) to obtain samples of the prior at the next time step. Equation (8.4) is
also allows us to extrapolate information in remote-sensing observations to called a filter, and Eq. (8.1) is called a smoother (McLaughlin, 2002). Using a
parts of the hydrologic system that are not directly observed. From a practical filter rather than a smoother reduces the dimension of the inference problem
perspective, the way this works is to use the remote-sensing observations to by a factor equal to the number of time steps.
constrain model simulations: data assimilation is fundamentally the applica-
tion of Bayes’ theorem to constrain probability distributions over model Variational versus Ensemble
simulations (Wikle and Berliner, 2007). So far, we have sampled the probability distributions on the right-hand side
The most general form of the data assimilation equation is of Eqs. (8.1) and (8.4), and we have used a filtering strategy to reduce the
complexity of the problem. We still need a way to approximate the integra-
p( z | x , u, h) p( x | u, m) tions over the state distribution and forcing data. Absent analytical solu-
p( x | z , m, h) = ∫ p(u)du (8.1) tions (which are generally unavailable), these integrals can be solved using
∫ p(z | x , u, h) p( x | u, m)dx Monte Carlo approximation. Alternatively, we can seek some particular
statistic of the posterior distribution that maximizes some criteria. The
where p(.) denotes a probability distribution, x is the time series of the former is typically done via Monte Carlo methods and the latter via a varia-
dynamic states simulated by the model (e.g., soil moisture, soil temperature, tional approach.
groundwater storage, plant water content, etc.) over the entire simulation The most common variational methods seek the maximum a posteriori
period and domain, z are all of the remote-sensing observation available in solution for conditioned Brownian motion with Gaussian observation error.
that domain, and u are the model forcing data or boundary conditions. In The filter version of this (often called 3D-Var referencing the three spatial
general, it is assumed that any forcing data contains errors, so it is necessary dimensions of distributed atmospheric models) minimizes the cost function:
to integrate over the uncertainty in these data. The dynamical model is
denoted through m, and the observation operator through h, which simulate
the relationship between the satellite observation and the system states simu- J ( xt ) = (m( x T −1  T −1
t −1 , ut ) − xt ) Pt (m( xt −1 , ut ) − xt ) + ( zt − ht ( x −t )t ) Rt ( zt − ht ( x −t )t )
lated by the model. The following subsections describe common methods for (8.5)
applying this equation to real problems, including how to approximate the
integrations and the various probability distributions. where Pt is the covariance of p( xt |xt −1ut , mh1:t ) and Rt is the covariance of εt
(assumed a priori). For linear models this minimizes at:
8.5.1 Estimating the Prior and Likelihood
t −1 , ut ) + St (Qt + Rt ) ( zt − ht ( x −t )t ) (8.6)
−1
The prior distribution over the state system is given by p( x | u, m) in Eq. (8.1). xt = min( J ( xt )) = m( x
This is obtained simply by running the model m with some recognition that xt
the model is not exactly correct (if hydrology models were as accurate as they
are precise, there would be no need for satellites). The most common way to where Qt is the covariance of the modeled observation (depends on Pt) and St
do this is to add some Gaussian error to the model states at each timestep, so is the covariance between xt and yt. More generally, Eq. (8.5) can be mini-
that the evolution of the modeled states is given by: mized using an iterative method with or without derivatives:
dJ
xt = m(ut , xt −1 ) + t (8.2) = ( yt − zt )T Rt−1St + ( m( x T −1
t −1 ) − x t ) Pt
(8.7)
dxt
By running the model several times (to create an ensemble of simulations)
and adding a different random error at each time step to each ensemble mem- The most common Monte Carlo method is the ensemble Kalman filter
ber, we obtain a sample of the distribution over the model states. (EnKF) (Houtekamer and Mitchell, 2001), which uses n independent and
Similarly, samples of the Bayesian likelihood distribution, denoted identically distributed (iid) samples of xt and yt to estimate St, Pt, and Qt
p( z | x , u, h) in Eq. (8.1), can be obtained by sampling a noisy observation required by Eqs. (8.6) and (8.7). This allows us to estimate the first two
process that may be time-dependent: moments of the posterior for a nonlinear dynamic model, but still requires
linear observation operators—pseudocode is given in supplement 8.1. The
yt = ht ( xt , ut ) + εt (8.3) EnKF can be generalized to a smoother, and pseudocode for a few simple
smoother approximations is given by Ravela and McLaughlin (2007).
This sampling process provides model-based estimates y of the actual
remote-sensing image data z. Samples y are taken to define a probability dis- Parametric versus Non-Parametric
tribution over the image data. All methods discussed earlier rely on parametric assumptions about model
It is important to note that even though it is typically Gaussian random and observation operators and their associated uncertainty distributions.
error that is added to both the model and observation operator, that the dis- These assumptions are inherent in the covariance terms St, Pt, and Qt. Such
tributions p( x | u, m) and p( z | x , u, h) are not necessarily Gaussian over time assumptions can be alleviated using a bootstrapping approach (Gordon et al.,
since the dynamical model m and observation operators ht may be nonlinear. 1993) called a sequential resampling filter or particle filter (van Leeuwen, 2010).
08_Singh_ch08_p8.1-8.6.indd 3 8/22/16 11:35 AM

8-4 Remote Sensing Techniques and Data Assimilation for Hydrologic Modeling
Samples of the posterior distribution at each timestep are obtained by system for the USDA Foreign Agricultural Service crop forecasting system
reweighting an iid Monte Carlo sample of the prior with probabilities propor- (Bolten et al., 2009), and (3) a system to provide qualitative crop condition
tional to: assessments for the stability of global food markets (Becker-Reshef, 2010).
p( xt ,i | u1:t ,i , m, h1:t , y1:t ) ∝ p( yt | xt ,i , ut ,i , ht ) (8.8)

REFERENCES
Pseudocode for a simple resampling filter is given in supplement 8.2.
Adler, R. F. and A. J. Negri, “A satellite infrared technique to estimate
8.5.3 Dealing with Bias tropical convective and stratiform rainfall,” Journal of Applied Meteorology
When we use certain parametric data assimilation techniques like the EnkF and Climatology, 27 (1): 30–51, 1988.
or Eq. (8.6), then bias in either m or h will result in bias in posterior state Al Bitar, A., D. Leroux, Y. H. Kerr, O. Merlin, P. Richaume, A. Sahoo, and
estimates. There are currently two ways to deal with this (Kumar et al., 2012): E. F. Wood, “Evaluation of the SMOS soil moisture products over continental
(i) mapping observations to the model climatology or (ii) calibrating the U.S. using the SCAN/SNOTEL network,” IEEE Transactions on Geoscience
model to the observation climatology. The land surface modeling community and Remote Sensing, 50 (5): 1572–1586, 2012.
has generally preferred the former in an effort to avoid changing the climatol- Anderson, M. C., J. M. Norman, G. R. Diak, W. P. Kustas, and J. R.
ogy of the lower boundary condition in weather forecasting models, whereas Mecikalski, “A two-source time-integrated model for estimating surface
the streamflow modeling community generally takes the latter approach. We fluxes using thermal infrared remote sensing,” Remote Sensing of the
briefly cover a few of these bias-correction methods. Environment, 60: 195–216, 1997.
Anderson, M. C., J. M. Norman, J. R. Mecikalski, R. D. Torn, W. P. Kustas,
CDF Matching and J. B. Basara, “A multiscale remote sensing model for disaggregating
If we believe that the majority of bias lies in the retrieval (i.e., with the observa- regional fluxes to micrometeorological scales,” Journal of Hydrometeorology,
tion operator h), we will force the cumulative density function (CDF) of the 5: 343–363, 2004.
model to match that of the observations. Empirical CDF with finite precision Atkinson, P. M., “Downscaling in remote sensing, International,” Journal of
Δy are constructed for individual dimensions of both modeled and observed y: Applied Each Observation and Geoinformation, 22: 106–114, 2013.
Bolten, J., W. Crow, X. Zhan, C. Reynolds, and T. Jackson, Assimilation of
1 T a satellite-based soil moisture product in a two-layer water balance model for
CDF* ( y min + α∆ y ) = ∑1
T t =1 yt ≤ ( ymin +α∆y )
(8.9)
a global crop production decision support system, Data Assimilation for
Atmospheric, Oceanic, and Hydrologic Applications, Springer-Verlag, Berlin,
where ymin is the minimum collectively among all retrievals (z) and model- Heidelberg, 2009.
predicted retrievals (y), is the histogram bin number, and CDF* is either the Becker-Reshef, I., C. Justice, M. Sullivan, E. Vermote, C. Tucker, A.
model or observation cumulative density function. Observations are then Anyamba, J. Small, et al., “Monitoring global croplands with coarse resolution
translated into the model climatology using a linear interpolation: Earth observations: The Global Agriculture Monitoring (GLAM) Project,”
 CDFobs ( zt ) − CDFobs ( y min + ∆ y (α * ))  Remote Sensing, 2 (6): 1589–1609, 2010.
zt* = y min + ∆ y α * + * * 
(8.10) Bulygina, N. and H. Gupta, “Correcting the mathematical structure of a
 CDFobs ( y min + ∆ y ( α + 1)) − CDFobs ( y min + ∆ y ( α ))  hydrological model via Bayesian data assimilation,” Water Resources Research,
47 (5): W05514, 2011.
α * = max(CDFmod ( y min + α∆ y ) ≤ CDFobs ( zt )) Campell, J., Introduction to Remote Sensing, 3rd ed., The Guilford Press,
α New York, 2001.
Cheng, M., R. Brown, and C. Collier, “Delineation of precipitation areas
An analytical solution considers only the first two moments of the cumula-
using Meteosat infrared and visible data in the region of the United Kingdom,”
tive density function:
Journal of Applied Meteorology and Climatology, 32: 884–898, 1998.
 T  2 DeFries, R. S., M. Hansen, J. R. G. Townshend, and R. Sohlberg, “Global
1 T  
1  1   ∑ i=1  yi − T ∑ i=1yi   land cover classification at 8 km spatial resolution: the use of training data
zt* = ∑ i=1yi +  zt − ∑ i=1zi  
T T
 (8.11) derived from Landsat imagery in decision tree classifiers,” International
T  T  T  1 T  
2
Journal of Remote Sensing, 19 (16): 3141–3168, 1998.
 ∑ i=1  i
z − ∑ i 
z
T i=1   Entekhabi, D., E. G. Njoku, P. E. O’Neill, K. H. Kellogg, W. T. Crow, W. N.
Edelstein, J. K. Entin, et al., “The Soil Moisture Passive Active (SMAP) mis-
Simultaneous State and Parameter Estimation sion,” IEEE Transactions on Geoscience and Remote Sensing, 98 (5): 704–716,
If we wish to attribute the majority of bias to the dynamic model, m, observa- 2010.
tions can be used to simultaneously estimate model states and parameters. Fang, B., V. Lakshmi, R. Bindlish, T. J. Jackson, M. Cosh, and J. Basara,
The robust way to do this is by expectation-maximization, and although this “Passive microwave soil moisture downscaling using vegetation index and
has been demonstrated in hydrology (Bulygina and Gupta, 2011; Wilkinson skin temperature,” Vadose Zone Journal, 12 (3), 2013, doi:10.2136/
et al., 2011) less rigorous methods are more common. Examples include vzj2013.05.0089.
simply calibrating the parameters and then performing data assimilation (e.g., Forman, B. A., R. H. Reichle, and M. Rodell, “Assimilation of terrestrial
Kumar et al., 2012), treating the parameters as dynamic and adding them to water storage from GRACE in a snow dominated basin,” Water Resources
the state vector to be estimated (called state augmentation) (e.g., Moradkhani Research, 48: W01507, 2012, doi:10.1029/2011WR011239.
et al., 2005), or performing data assimilation at each iteration of an optimiza- Friedl, M. A., D. K. Mclver, J. C. F. Hodges, X. Y. Zhang, D. Muchoney, A.
tion search (Vrugt et al., 2005). H. Strahler, C. E. Woodcock et al., “Global land cover mapping from MODIS:
algorithms and early results,” Remote Sensing of Environment, 83 (1–2):
287–302, 2002.
8.6 SUMMARY
Gordon, N. J., D. J. Salmond, and A. F. M. Smith, “Novel approach to non-
Remote sensing of the hydrologic cycle continues to evolve with the advance- linear non-Gaussian Bayesian state estimation,” IEE Proceedings F Radar and
ment of new sensor capabilities, innovative model integration and scaling Signal Processing, 140 (2): 107–113, 1993.
approaches, more robust antenna design and computer processing potential, Holben, B., “Characteristics of maximum-value composite site images from
and new additions to the constellation of satellite-based and aircraft-based temporal AVHRR data,” International Journal of Remote Sensing, 7 (11):
remote sensing platforms. This year marks the 55th anniversary of the 1417–1434, 1986.
first-Earth observing satellite, the NASA Television Infrared Observation Hollinger, J. P., J. L. Peirce, and G. A. Poe, “SSM/I instrument evaluation,”
Satellite (TIROS). Since that time, remote sensing and modeling of the hydro- IEEE Journal of Geoscience and Remote Sensing, 28 (5): 781–790, 1990.
logic cycle has quickly transitioned from ambitious, impracticable concepts to Houtekamer, P. L. and H. L. Mitchell, “A sequential ensemble Kalman filter
feasible and reliable applications regularly used to improve decision-making for atmospheric data assimilation,” Monthly Weather Review, 129 (1):
by policyholders and water managers. Many new and exciting applications of 123–137, 2001.
data assimilation systems integrating remote sensing observations are now Huffman, G. J., D. T. Bolvin, E. J. Nelkin, D. B. Wolff, R. F. Adler, G. Gu, Y.
operationally employed by several agencies and organizations including: (1) Hong, et al., “The TRMM multisatellite precipitation analysis (TMPA):
multiscale drought monitoring tools for the United States Drought Monitor Quasi-global, multiyear, combined-sensor precipitation estimates at fine
(USDM) (Rodell, 2012), (2) an operational soil moisture data assimilation scale,” Journal of Hydrometeorology, 8 (1): 38–55, 2007.
08_Singh_ch08_p8.1-8.6.indd 4 8/22/16 11:35 AM

References 8-5
Jackson, T. J., M. H. Cosh, R. Bindlish, P. J. Starks, D. Bosh, M. Seyfried, Pan, M., A. K. Sahoo, E. F. Wood, A. Al Bitar, D. Leroux, and Y. H. Kerr, “An
D. C. Goodrich, et al., “Validation of advanced microwave scanning radiom- initial assessment of SMOS derived soil moisture over the continental United
eter soil moisture products,” IEEE Transactions on Geoscience and Remote States,” Selected Topics in Applied Earth Observations and Remote Sensing,
Sensing, 48 (12): 4256–4272, 2010. 5 (5): 1448–1457, 2012, doi: 10.1109/JSTARS.2012.2194477.
Jensen, J., Remote Sensing of the Environment: An Earth Resource Perspective, Ravela, S. and D. McLaughlin, “Fast ensemble smoothing,” Ocean Dynamics,
2nd ed., Prentice Hall, Upper Saddle River, NJ, 2007. 57 (2): 123–134, 2007.
Joyce, R. J., P. Xie, and J. E. Janowiak, “Kalman filter–based CMORPH,” Reichle, R. H., “Data assimilation methods in the Earth sciences,” Advances
Journal of Hydrometeorology, 12: 1547–1563, 2011. in Water Resources, 31(11): 1411–1418, 2008.
Justice, C. O., E. Vermote, J. R. Townshend, R. Defries, D. P. Roy, D. K. Hall, Rodell, M., I. Velicogna, and J. S. Famiglietti, “Satellite-based estimates of
and V. V. Salomonson, “The Moderate Resolution Imaging Spectroradiometer groundwater depletion in India,” Nature, 460: 999–1002, 2009.
(MODIS): land remote sensing for global chance research,” IEEE Transactions Rodell, M., “Satellite gravimetry applied to drought monitoring,” Remote
on Geoscience and Remote Sensing, 36(4): 1228–1249, 1998. Sensing of Drought: Innovative Monitoring Approaches, edited by B. Wardlow,
Kalman, R. E., “A new approach to linear filtering and prediction prob- M. Anderson, and J. Verdin, CRC Press/Taylor and Francis, Boca Raton, 2012,
lems,” Transactions of the ASME–Journal of Basic Engineering, 82 (Series D): pp. 261–280.
35–45, 1960. Schmidt, R., P. Schwintzer, F. Flechtner, C. Reigber, A. Güntner, P. Döll, G.
Kidd, C. and G. Huffman, “Review Global precipitation measurements” Ramillien, et al., “J., GRACE observations of changes in continental water
Meteorological Applications, 18: 334–353, 2011. storage,” Global and Planetary Change, 50: 112–126, 2006.
Kerr, Y., P. Waldteufel, J-P. Wigneron, J. Martinuzzi, J. Font, and M. Berger, Shibata, A., K. Imaoka, and T. Koike, “AMSR/AMSR-E level 2 and 3 algo-
“Soil moisture retrieval from space: the Soil Moisture and Ocean Salinity rithm developments and data validation plans of NASDA,” IEEE Transactions
(SMOS) mission,” IEEE Transactions on Geoscience and Remote Sensing, 39 on Geoscience and Remote Sensing, 41 (2): 195–203, 2003.
(8): 1729–1735, 2001. Swenson, S., P. J. F. Yeh, J. Wahr, and J. Famiglietti, “A comparison of ter-
Kumar, S. V., R. H. Reichle, K. W. Harrison, C. D. Peters‐Lidard, S. restrial water storage variations from GRACE with in situ measurements
Yatheendradas, and J. A. Santanello, “A comparison of methods for a priori from Illinois,” Geophysical Research Letters, 33 (16): L16401, 2006, doi:
bias correction in soil moisture data assimilation,” Water Resources Research, 10.1029/2006gl026962.
48 (3): 2012. Tapley, B. D., S., Bettadpur, M. Watkins, and C. Reigber, “The gravity recov-
Loveland, T. R., B. C. Reed, J. F. Brown, D. O. Ohlen, Z. Zhu, L. Yang, and ery and climate experiment: mission overview and early results,” Geophysical
J. W. Merchant, “Development of global land cover characteristics database Research Letters, 31: L09607, 2004.
and IGBP discover from 1 km AVHRR data,” International Journal of Remote Tucker, C. J., “Red and photographic infrared linear combinations for
Sensing, 21(6–7): 1303–1330, 2000. monitoring vegetation,” Remote Sensing of Environment, 8 (2):127–150, 1979.
McLaughlin, D., “An integrated approach to hydrologic data assimilation: Ulaby F., R. K. Moore, and A. K. Fung (Eds.), Microwave Remote Sensing:
interpolation, smoothing, and filtering,” Advances in Water Resources, 25 Active and Passive: From Theory to Applications, Artech House, Norwood,
(8–12): 1275–1286, 2002. MA, 1986, Vol. III.
Metropolis, N., “The beginning of the Monte Carlo method,” Los Alamos Velicogna, I. and J. Wahr, “Acceleration of Greenland ice mass loss in spring
Science, 15 (584): 125–130, 1987. 2004,” Nature, 443: 329–331, 2006.
Moradkhani, H., S. Sorooshian, H. V. Gupta, and P. R. Houser, “Dual state– Vrugt, J. A., C. G. H. Diks, H. V. Gupta, W. Bouten, and J. M. Verstraten,
parameter estimation of hydrological models using ensemble Kalman filter,” “Improved treatment of uncertainty in hydrologic modeling: combining the
Advances in Water Resources, 28 (2): 135–147, 2005. strengths of global optimization and data assimilation,” Water Resources
Myneni, R. B., F. G. Hall, P. J. Sellers, and A. L. Marshak, “The interpreta- Research, 41 (1): W01017, 2005.
tion of spectral vegetation indexes,” IEEE Transactions on Geoscience and Wang, J. R. and B. J. Choudhury, “Remote sensing of soil moisture content
Remote Sensing, 33 (2): 481–486, 1995. over bare field at 1.4 GHz frequency,” Journal of Geophysical Research, 86 (C6):
Myneni, R. B., S. Hoffman, Y. Knyazikhin, J. L. Privette, J. Glassy, Y. Tian, 5277–5282, 1981.
Y. Wang et al., “Global products of vegetation leaf area and fraction absorbed Wan, Z. and Z-L. Li, “A physics-based algorithm for retrieving land-surface
PAR from year one of MODIS data,” Remote Sensing of Environment, 83 (1–2): emissivity and temperature from EOS/MODIS data,” IEEE Transactions on
214–231, 2002. Geoscience and Remote Sensing, 35 (4): 980–996, 1997.
Murphy, P. N. C., J. Ogilvie, F-R. Meng, and P. Arp, “Stream network mod- Wikle, C. K. and L. M. Berliner, “A Bayesian tutorial for data assimilation,”
eling using lidar and photogrammetric digital elevation models: a comparison Physica D-Nonlinear Phenomena, 230 (1–2): 1–16, 2007.
and field verification,” Hydrological Processes, 22 (12): 1747–1754, 2007. Wilkinson, R. D., M. Vrettas, D. Cornford, and J. E. Oakley, “Quantifying
Njoku, E., J. M. Stacey, and F. T. Barath, “The SEASAT scanning multichan- simulator discrepancy in discrete-time dynamical simulators,” Journal of
nel microwave radiometer (SMMR)-instrument description and perfor- Agricultural, Biological, and Environmental Statistics, 16 (4): 554–570, 2011.
mance,” IEEE Journal of Ocean Engineering, 5 (2): 100–115, 1980. Zaitchik, B. F., M. Rodell, and R. Reichle, “Assimilation of GRACE terres-
Njoku, E., T. J. Jackson, V. Lakshmi, T. K. Chan, and S. V. Nghiem, “Soil trial water storage data into a land surface model: results for the Mississippi
moisture retrieval from AMSR-E,” IEEE Transactions on Geoscience and river basin,” Journal of Hydrometeorology, 9: 535–548, 2008.
Remote Sensing, 41 (2): 215–229, 2003.
08_Singh_ch08_p8.1-8.6.indd 5 8/22/16 11:35 AM

Chapter 9
Geographic Information
Systems
BY
ROBERT E. GRIFFIN, JAMES F. CRUISE, WALTER L. ELLENBURG, MOHAMMAD Z. AL-HAMDAN, AND
CAMERON HANDYSIDE
ABSTRACT buy-in from national agencies (plus the enormous influx of data that the
satellite era garnered) created the foundation for sustained interest and
Geographic Information Systems (GIS) are commonly employed in a wide
applicability in the field of hydrology.
range of hydrologic studies. These systems can be used to store, query, and
analyze geographic data and current applications include the native integra-
tion and handling of data for hydrologic modeling. This chapter primarily 9.2 BASIC PRINCIPLES OF GIS
focuses on U.S. federal sources of model input data (including elevation, soils,
In its most basic form, a GIS consists of computer hardware, software, and
land cover land use, basins and hydrologic networks, and meteorological
geospatial data, and while software and hardware are constantly changing,
datasets), GIS applications in hydrology, and future directions in the field.
several key aspects to geospatial datasets remain universal. These concepts
Also covered in this chapter, and of particular concern for hydrologists, are
enable the analysis and display of geographic data and at their core include:
the ways in which a GIS handles multi-source data integration and the repre-
data models, coordinate systems, and the relational attributes of the dataset.
sentation of model inputs (precipitation, surface parameters, and watersheds
Data models are ways to store and display geographic datasets and include
for distributed modeling).
discrete object models, or vector data models, and field-based models, or
raster data models. The vector data model stores discrete spatial features
9.1 INTRODUCTION
(points, lines, and polygons) by storing the x,y,z nodes that form the feature
shapes as well as the attribute of the feature, an example being state adminis-
Geographic Information Systems (GIS) are computer systems for storing, trative boundaries provided by the U.S. Census Bureau through TIGER/Line
querying, analyzing, and displaying geospatial datasets. Although first used files. Unlike the vector model which is often used for discrete points in space,
for natural resource management, GIS integration is now common to most the raster model is more often used to represent continuous features, such as
hydrologic studies. GIS has undergone quite a dramatic change since the term elevation, temperature, precipitation, or satellite imagery. Raster data models
was first coined in the 1960s, in part due to the technological revolution of the store these continuous data in the form of contiguous and even grid cells or
past 30 years. Tasks that are relatively simple to conduct for humans, such as pixels, whose values represent the feature at that location. Raster data can vary
being able to tell which cities are closest to a floodplain from a map are some- in their cell values (the information content) as well as pixel size (the effective
what analytically complex for a synthetic environment and thus require a lot ground sample distance of the dataset; its “resolution”) and the pixel depth, or
of processing power. The limiting factor of GIS, like all models, is the inherent the radiometric resolution of the data.
complexity of real world systems, and thus it makes sense that the evolution A second basic principle in GIS concerns coordinate systems and data
of GIS coincides with the advent of advanced computing. projections, primarily important in enabling layers to be geographically
The field of cartography has existed for a long time, but the idea to aligned with one another. Often, datasets are gathered from a variety of
inference causality based on geographic location dates back to the sources, each with a different coordinate system or projection, and must be
nineteenth century with John Snow’s mapping of cholera deaths in London. put in the same reference system in order to perform subsequent analysis.
However, it was during the 1960s, brought on in part by the Cold War, that Coordinate systems are in essence location reference systems and can be
GIS as we know it began. The term was first coined by Roger Tomlinson of based on a three-dimensional (3D) ellipsoid (geographic coordinate system)
the Canada Land Inventory [Tomlinson, 1984]; the “first” system being used or flattened to a two-dimensional plane (projected coordinate system). While
to store, collate, and analyze data about land usage in Canada. The early era geographic coordinate systems are typically measured in degrees minutes
of GIS was led by a few individuals and organizations pursuing disparate seconds or decimal degrees, projected coordinate systems are measured in
goals, but things began to change once commercialization of the field began. ground units. One well-known data projection is the transverse Mercator
The intense commercial activity, spearheaded by Environmental Systems projection, a common derivative of which is the Universal Transverse
Research Institute (ESRI) provided an influx of spatial data-processing capa- Mercator, which represents a secant case of the projection system and divides
bilities available to both businesses and private users. At the same time, the Earth into contiguous hemispherical zones with linear distances typically
national agencies were beginning to adopt the technologies for a range of measured in meters. Datums, mathematical models of the Earth’s ellipsoidal
reasons including data management, increased data demand, and transpar- shape, are used together with coordinate systems to accurately determine the
ency purposes. The U.S. Census Bureau, as an example, began creating 3D coordinates of a point on the Earth’s surface.
Topologically Integrated Geographically Encoding and Referencing (TIGER) One function of a GIS is to enable a wide range of data layers to be prepro-
files in the 1980s, first publishing them in the 1990 decadal census. The cessed to matching coordinate systems and datums, thus allowing them to be
9-1
09_Singh_ch09_p9.1-9.8.indd 1 8/22/16 11:35 AM

9-2 Geographic Information Systems
analyzed in conjunction with one another. From the perspective of hydro-

logical analysis, this means that a NASA impervious surface product can be
efficiently and accurately overlaid on a terrain surface derived from a U.S.
Geological Survey (USGS) dataset as well as with administrative boundaries
and transportation networks from the U.S. Census Bureau and interpolated
rain gage data from the National Oceanic and Atmospheric Administration
(NOAA). Through integration in a GIS, disparately sourced datasets can be Digital terrain model
analyzed in multiple dimensions. Digital surface model
The third guiding principle in GIS involves the ability to create complex
queries on geographic features according to their related attribute informa-
tion. One might, for instance, want to select all watersheds over a certain size Figure 9.1 Comparison of DSM and DTM. Sources and Accuracies of Selected
and that have more than 50% of their land area in agriculture; a GIS enables Available Elevation Datasets (Current December 2014).
this. The ability to perform such queries is premised on the relational aspect
of linked geometric shapes (features) and their many attributes which can be others. The four primary ways that surface elevations are mapped include
stored as numbers, text, dates, and a range of other query-able formats. radar, lidar, stereoscopy, and direct topographic survey, the first three of
A GIS allows digital terrain to be analyzed in 3D, data attributes to be que- which produce 3D representations of the Earth’s surface in digital elevation
ried and combined, vectors to be interpolated, network connectivity to be models (DEMs), digital gridded datasets in which pixel values represent ele-
determined, and pattern analysis to be conducted, to name but a few of the vation. It must be noted that some elevation sources provide representations
analytical processes that can be performed. Since its inception in the 1960s, of the Earth’s surface as seen from remote-sensing platforms, surfaces which
GIS has been used to perform studies and process data in industries as diverse can include tree canopy and the built environment in addition to visible ter-
as public safety, natural resources, health, business, transportation, sustain- rain; these are referred to as DSMs or digital surface models and include those
able development, and disaster response. Much of what determines the use- DEMs based on aerial photogrammetry, stereoscopic interpretation, and
value of a GIS is not the analytical capabilities of the system, but rather the Interferometric Synthetic Aperture Radar datasets. Other methods such as
datasets that are available to the researcher. lidar and direct topographic survey more accurately represent the physical
terrain of the Earth’s surface; once processed, these are referred to as DTMs
9.3 DATA SOURCES AND CHARACTERISTICS or digital terrain models. The differences between these two methods of
recording elevation are illustrated in Fig. 9.1. Having been derived from his-
The past several decades have seen the proliferation of publicly available GIS- toric direct topographic survey (and more recently, Light Detection and
ready datasets in the United States, the developed world, and to some extent Ranging — lidar), the USGS National Elevation Dataset (NED) represents a
in developing countries. These data form the input elements for hydrologic DTM. Although not as accurate for hydrologic modeling, DSMs are more
models within GIS as well as stand-alone environments, and they can repre- readily available on a global scale through datasets, like SRTM, GDEM, and
sent relatively static layers, periodically updated information, and near real- GMTED2010, which are often at a coarser resolution as well.
time feeds. Thus, users can select the most appropriate datasets for their Table 9.1 presents common available elevation datasets at the national and
geographic and temporal areas of study. The wide range of datasets, the con- global scales for hydrologic analysis, including their horizontal resolutions
stantly changing landscape of data delivery and agency responsibility, and and vertical accuracies, key factors for determining usability in hydrologic
data standards hint at the importance of understanding what datasets to use models. The root-mean-square error (RMSEz) shown in Table 9.1 is a widely
and when. Here, we provide a list of hydrologically relevant GIS datasets, used measure of accuracy, representing the square root of the average of dif-
predominantly applicable to study in the United States though including ferences between DEM sources and high-precision ground sample points
global examples with the caveat that data quality, cost, and availability can (Maune et al., 2007). Assuming normal distribution, the vertical linear accuracy
change quite dramatically from country to country and a more thorough of any value at the 95% confidence level will be 1.96 times the RMSEz
globally comprehensive and site specific list would necessitate an entirely (Dewberry, 2012). Lidar datasets currently being acquired as part of the USGS
separate manual. 3DEP initiative can thus be considered accurate at 95% confidence to within
9.3.1 Model Input Data 20–40 cm vertically, while binned at 0.5—1 m horizontal grid cells.
The brief descriptions in this section are intended to provide some context for
data types and sources available for hydrologic analyses in a GIS, with Soil Characteristics
Table 9.2 outlining characteristics of select datasets. The authors acknowledge Data on soil composition is critical in hydrologic studies, with soil database
that this list is incomplete, focused mostly on applications in the United accuracy becoming increasingly recognized as an important consideration.
States, and relevant to the time period this was written in 2014. Specific data- GIS provides a means to define soil variability spatially across a watershed,
sets, sources, and user interfaces change so often that we attempt to cover the though long before GIS existed soil maps were being produced in areas
broad classes of data inputs while highlighting only a handful of specific around the world, for example, in the United States through the National
datasets themselves. Cooperative Soil Survey. Soil data are now widely available as digital GIS data
at high resolutions in some cases (e.g., SSURGO data in the United States) and
Elevation Datasets at somewhat lower spatial resolutions for larger global datasets (e.g., WISE
Elevation is a basic input into hydrologic models in that it determines deriva- and HWSD). Generally, this soil information takes the form of soil character-
tive terrain characteristics including slope, aspect, hydrologic networks, and istic databases and associated geometries, with some datasets more applicable
Table 9.1 Sources and Accuracies of Selected Available Elevation Datasets (Current December 2014)
Source data H. Resolution Source dates Source type V. Accuracy Typical applications
3DEP lidar 0.5–1m 2011–present DSM & DTM 10–20cm RMSEz* urban/rural watersheds
NED photogrammetry/aerial/digitized 3m, 10m, 30m 1925–present DTM 1.55m RMSEz** rural watersheds
contours/field survey/lidar/ifsar
data
SRTM ifsar data 30m, 90m 2000 DSM 4.01m RMSEz† river basins/continents
ASTER GDEM2 stereoscopy 30m 2009 DSM 8.68m RMSEz †
river basins/continents
GMTED2010 photogrammetry/aerial/digitized 250–1000m 2000–2011 DEM ~30m RMSEz†† continents/global
contours/field survey/lidar/ifsar
data
*[Dewberry, 2012; Sugarbaker et al., 2014]
**RMSE calculated for coterminous US from April 2013 [Gesch et al., 2014]
†
[Gesch et al., 2012]
††
RMSE ranges from 25–42m for 30 arcsec product, 29–32m for 15 arcsec product, and 26–30m for 7.5 arcsec product [Danielson and Gesch, 2011]
09_Singh_ch09_p9.1-9.8.indd 2 8/22/16 11:36 AM

Data Sources and Characteristics 9-3
to design studies (e.g., SSURGO) and others more relevant for planning stud- 9.3.2 Data Models and Repositories
ies in GIS. Table 9.2 lists the more common soil data sources for the Given the large amount of data sources, the usefulness of data distribution
United States and globally. services and models are evident. Many of the data sources mentioned earlier
can be found concurrently at the same repository, making the retrieval of data
Land Cover Land Use much easier. For example, NASA’s Earth Observing System Data and
One of the most important input datasets for watershed modeling is land Information System Distributed Active Archive Centers provide many web-
cover land use (LCLU). In hydrologic modeling, LCLU data are needed for based interface tools to search and subset the data of any its past or current
characterizing the impervious cover, assigning impervious values, and Earth-observing satellites and field-measurement campaigns. Perhaps most
calculating the curve number. Table 9.2 gives examples of remotely sensed useful for hydrologists, the National Map, a collaborative effort of many fed-
LCLU products that are used in hydrologic modeling in different areas eral agencies, spearheaded by the USGS, an effort which provides an easily
around the world, with spatial resolutions of these datasets ranging from queried collection of data ranging from elevation and hydrography to built
30 m to 1 km. structures and land cover.
In addition to the many data repositories, data models and tools are
Basins and Hydrologic Networks increasingly being developed for use specifically in a GIS. The EPA’s
These features can be calculated through terrain analysis natively within most Automated Geospatial Watershed Assessment tool collates the data needed
GIS software packages, though often these are available as predefined feature for a common hydrological model Soil and Water Assessment Tool (SWAT)
classes through government agencies. For example, the U.S. Department of and uses a GIS to aid in the automation of the spatial parameterization. The
Agriculture-Natural Resources Conservation Service (NRCS), the USGS, and Environmental Data Collector has been developed into an extension of ESRI’s
the Environmental Protection Agency (EPA) have collaborated to develop a ArcGIS platform that allows users to connect to and download environmental
standardized watershed boundary dataset, which delineates Hydrologic Unit data served by Thematic Real-Time Environmental Distributed Data Services
Code (HUC) boundaries in the United States. The accuracy of these derived directly into the GUI for display and analysis. Repositories and models like
basins, subbasins, and hydrologic networks are a function of available survey these will continue to play a specific role in the application of GIS in hydrol-
data and the horizontal and vertical resolutions of the input DEMs (see Table ogy, as data access and integration become more important for calibration
9.1). A list of example hydrologic and basin datasets available for the and validation of hydrologic models.
United States is provided in Table 9.2. 9.3.3 Data Integration and Handling
Meteorological Datasets As innovations in hydrologic modeling in GIS continue to progress, the vol-
ume of the data required and how they are handled and processed need to be
Weather data are available from numerous sources, including meteorological addressed. Though the increases in computer processing speed and advanced
stations, satellite products, and reanalysis datasets, and in a variety of formats data formats (e.g., Network Common Data Form) will most definitely aid in
and ranges in their content type and dates. In the United States, for instance, overcoming these challenges, the new paradigms of data collection and sharing
the National Oceanic and Atmospheric Administration’s (NOAA) National will play a large role as well.
Climatic Data Center provides free, public access to land-based weather mea- The data needed to fully represent the scenarios within more realistic GIS
surements, such as the Cooperative Observed Network data, which are com- models are becoming increasingly large and come from many sources. This is
piled basic daily measurements, made by over 10,000 volunteers. Satellite especially true in hydrology where the datasets can range from soil character-
measurements like those from the Geostationary Operational Environmental istics to climate records to land-use histories as discussed earlier. Data-
Satellite (GOES) program, are also used to define meteorological variables. integrated solutions are needed to address these potential bottle-necks. Data
GOES imagery is used to produce a variety of geophysical variables that integration is the process of combining heterogeneous data which reside in
define weather. As shown in Table 9.2, meteorological data, such as precipita- different sources or formats and/or resolutions and are made compatible with
tion, air temperature, humidity, solar radiation, and atmospheric pressure and each other. More information about spatial data integration schema and their
potential evaporation, can be also obtained from reanalysis datasets, such as applications can be found in Fotheringham and Rogerson (2013); Weng
those of the North American Land Data Assimilation System and Global (2010) specifically deals with the integration of remote-sensing data.
Land Data Assimilation System forcing datasets at 0.125° and 0.25° spatial In developing a data integration scheme, one must understand what the
resolutions and hourly and 3-h temporal resolutions, respectively (Cosgrove data represent, their reference frame, and accuracy. With respect to the scale
et al., 2003; Rodell et al., 2004). or measurement of the data, do the data represent a dichotomy or a presence/
absence value? Are they categorical, ranked, count, or continuous? Integrating
Ancillary Datasets data with differing measures and scales can be tedious. GIS provides the abil-
A wide range of GIS-ready ancillary datasets are currently available online for ity to draw geospatial inferences from the data, possibly creating new datasets
download and through map and feature services. Depending on the specific for integration.
application of the GIS for hydrology, the types of features available from pub- The second issue refers to the spatial and temporal aspects of the data. Are
licly available sources include transportation networks, wetlands, flood zones, the data represented in lines, points, polygons, or gridded raster formats?
administrative boundaries, and demographic data to name but a few exam- Transitions from vector to raster, and vice versa, are routine GIS applications
ples. Table 9.2 highlights some of the more commonly used ancillary datasets and can aid in the integration process. The issues of spatial reference and
for hydrology applications. coverage are compounded when integrating different spatial datasets in that
Table 9.2 Publicly Available Input Datasets for Hydrologic Models, Both for Design and Planning Purposes (Current as of December 2014)
Model input datasets Coverage Agency Time frame Format Linear resolution
Elevation Models
National Elevation Dataset (NED) US USGS 2014 Raster 3 m, 10 m, 30 m, 60 m
Shuttle Radar Topography Mission (SRTM) Global NASA 2000 Raster 30 m, 90 m
ASTER Global DEM (GDEMv2) Global NASA 2009/2011 Raster 30 m
EU-DEM Europe EU-Copernicus 2000 Raster 25 m
Lidar (3DEP) **
US USGS 2014 Vector <1 m
GMTED2010 Global USGS 2010 Raster 250 m, 500 m, 1 km
Canadian Digital Elevation Data L1 (CDED1v3) Canada NRC-CTI 1945–2010 Raster 10–100 m
Contour Maps (DLGs) US USGS 2014 Vector –
Photogrammetric interpretation** – – – Raster –
(Continued)
09_Singh_ch09_p9.1-9.8.indd 3 8/22/16 11:36 AM

Table 9.2 Publicly Available Input Datasets for Hydrologic Models, Both for Design and Planning Purposes (Current as of December 2014) (Continued)
Model input datasets Coverage Agency Time Frame Format Linear Resolution
Soil Characteristics
SSURGO/gSSURGO** US NRCS 1900–2005 Vector –
STATSGO US NRCS 1900–2005 Vector –
World Inventory of Soil Emission Profiles (WISE) Global ISRIC 1991–1996 Raster 60 km
Australian Soil Resource Information System (ASRIS) Australia CSIRO 1960s–2014 Raster/Vector 250 m
European Soil Database (ESDB) - SGDBE Europe EC-JRC Raster/Vector 1–10 km
Soil and Terrain Database Program (SOTER) Global FAO/UNEP/ISRIC 1986–2014
Harmonized World Soil Database (HWSD) Global IIASA-FAO 2012 Raster 1km
CONUS-Soil US PSU 1900–2005 Raster 1km
National Soil DataBase (NSDB) - Detailed Soil Survey Canada CanSIS 1970s–2010s Vector –
Land Cover Land Use (LCLU)
National Land Cover Database (NLCD) US USGS-MRLC 1992, 2001, 2006, Raster 30m
2011
Cropland Data Layer (CDL) US USDA-NASS 1997–2014 Raster 30m
MODIS Land Cover Product (MCD12Q1) Global NASA 2001–2014 Raster 500m
Coastal Change Analysis Program (C-CAP) US NOAA 1985, 1992, 1996, Raster 30m
2001, 2006, 2011
AVHRR Land Cover Global GLCF-UMD-NASA 1981–1994 Raster 1km
CORINE Land Cover Europe EA-EEA 1990, 2000, 2006, Raster/Vector 100m
2012
GlobCover Global ESA 2004–2006, 2009 Raster 300m
GLC2000 Global ESA 2000 Raster 1km
Orthoimagery** – – – Raster –
Basins/Networks
HydroSHEDS Global USGS-WWF 2000 Raster/Vector 90m
National Hydrography Dataset (NHD)** US USGS 2014 Vector –
(water bodies, stream network)
Watershed Boundary Dataset (WBD) US USGS 2014 Vector –
Meteorological Data
Radar (NEXRAD) US NOAA NCDC 1991–2014 Raster 2km, 4.7km
Local Radar **
– – – Raster –
Cooperative Observer Program (COOP)** US NWS 1890–2014 Vector –
Automated Surface Observing System (ASOS)** US NWS 1990s–2014 Vector –
Global Historical Climatology Network (GHCN) Global NOAA NCDC 1990s–2011 Raster/Text –
North American Land Data Assimilation System (NLDAS) US NASA ESDIS/NCEP 1979–2014 Raster 15km
Global Land Data Assimilation System (GLDAS) Global NASA ESDIS/NCEP 1949–2014 Raster 30km
Ancillary Datasets
Census Bureau (TIGER)—DC and ACS** US US Census Bureau 1990, 2000, 2010 Vector –
(admin boundaries, transportation networks, demographics)
National Flood Hazard Layer (NFHL)** US FEMA - MSC/NFIP 2014 Vector/WMS –
Dartmouth Atlas of Global Flood Hazard Global DFO 2000–2014 Raster 250m
National Transportation Atlas Database (NTAD) US US DOT 2008–2014 Vector –
National Highway Planning Network (NHPN) US US FHWA 2014 Vector –
National Wildlife Refuge System (NWRS) Trans. Assets US US FWS 2000, 2006–2007 Vector/WMS –
OpenStreetMap (OSM) Global OSM 2014 Vector –
Global Lakes and Wetlands Database (GLWD) Global WWF/CERS 2004 Raster/Vector 1km
National Wetlands Inventory (NWI) US US FWS 1977–2014 Vector/WMS –
**Datasets appropriate for design
09_Singh_ch09_p9.1-9.8.indd 4 8/22/16 11:36 AM

Representation of Model Inputs 9-5
the spatial reference of the source data must be known so accurate transfor-
mations may be applied as discussed at the beginning of this chapter. If the
coverage areas of two or more datasets are not aligned, certain inferences
about the data must be made to fully integrate the two. This problem is often
seen when the resolution of boundaries are different, where the boundaries of A4 4
one dataset appear smoothed out compared to the other. Interpolating or
planarizing boundary features are the most common approaches for dealing
with this issue; however, errors can arise when the data represented by the 1
polygonal area are assumed to be continuous when they are not. The varying
resolutions of raster datasets represent a further complication. Weighted A1
3
averaging and aggregation are typically used when scaling data up, in other
words “binning” data or effectively reducing the resolution of the dataset, A2 A3
common tasks that most GIS are natively programmed to perform. These 2
same queries can be asked of the temporal data. Do the data refer to specific
points in time or do they represent a period of time, or averages?
Aggregating data in GIS may be a simple task; however, the inverse, scaling
down, proves a much more complicated problem, for both spatial and tempo-
ral datasets. This occurs most often in climate datasets where the resolution
is too coarse for hydrological applications. Spatial downscaling methodolo-
gies can range from empirical or statistical methods which usually require Figure 9.2 Calculation and representation of a Thiessen Polygon surface.
some sort of regression to more sophisticated dynamical methods that incor-
porate higher resolution models (Fowler et al., 2007; Haylock et al., 2006; considered to be sufficiently accurate. The spatial and temporal resolution of
Al-Hamdan et al., 2013). this product may not always be conducive to the scale of the model needed;
As for temporal disaggregation, usually stochastic weather generators are therefore, disaggregation of the rainfall, both spatially and temporally may be
applied. These tasks expose a well-known short coming of GIS, its inherent necessary and may be accomplished using tools within the GIS (e.g., kriging),
static nature. Temporal and spatial disaggregation using stochastic or empiri- while temporal disaggregation may require the use of a weather generator
cal approaches are currently beyond the tasks of a desktop GIS. Over time, tool, or an open programming module within the GIS environment as
however, object-oriented programing languages have become more support- described in Sec. 9.3.3.
ed in GIS frameworks and currently provide users more control. Common
open-source languages such as Python (python.org) and R (r-project.org) are 9.4.2 Surface Parameters
most often used; these scripting languages allow complex tasks such as the The second important spatial coverage in the database is the topographic
disaggregation methodologies mentioned earlier to be incorporated directly distribution. Topographic data are displayed in some form of digital terrain
into the GIS workflow without having to perform them externally. In general, model either in raster or vector format. In the United States, some of the most
the incorporation of object-oriented programming languages enables GIS to commonly encountered data have been digital versions of the USGS 1:24,000
go beyond its current capabilities as essentially a database manager, to more (7.5’) topographic maps, mostly used in smaller scale hydrologic models
dynamic data modeling. (Table 9.2). The digital maps have a base spatial resolution of 30 m with a
contour interval of 10 ft (3.3 m) and a RMSEz vertical accuracy of one half of
9.4 REPRESENTATION OF MODEL INPUTS
a contour (Vieux, 2004). The recent advent and widespread adoption of lidar
A key aspect of hydrologic modeling that is facilitated by GIS is the spatial technology for topographic mapping has the potential to greatly facilitate
representation of the input data to the model. The spatial representation of hydrologic modeling; with lidar, it is possible to produce topographic cover-
the data forms the basis for the calculation of model parameters in a spatial ages at practically any desired resolution with tremendous vertical accuracy.
framework and can generally be divided into several categories: precipitation For example, the 3DEP initiative through the USGS has an ambitious goal of
data, surface parameters, and watershed representation for distributed full lidar coverage over the coterminous United States within the next decade,
modeling. providing freely available data with a horizontal resolution of 0.5–1 m and a
vertical accuracy of 10–20 cm RMSEz or a vertical linear error of 20–40 cm
9.4.1 Precipitation Data (see Table 9.1). Large-scale adoption of this technology has the benefit of both
First among the parameters are the climatic variables that are to be employed increasing the spatial (x,y,z) accuracy of available DEM data as well as creat-
by the model, including not only precipitation at a minimum, but also tem- ing DTM datasets that are current, thus making them applicable to hydro-
perature, solar radiation, relative humidity, and wind speed if evapotranspira- logic design studies. As noted by Sugarbaker et al. (2014), only 26% of the
tion is to be estimated. Sources and characteristics of meteorological data are USGS NED dataset represents terrain data collected after the year 2000; the
given in Table 9.2. The spatial distribution of the precipitation should be remainder of the coterminous U.S. NED data range from 1923 to 2000.
commensurate with the scale of the model and the modeling approach. A topographic surface may be generated from the DEM by using a TIN
For large-scale lumped parameter modeling, Thiessen polygons or Voronoi (Fig. 9.3), which can be produced from raster data (Vieux, 2004), but is more
diagrams, have historically been employed (Fig. 9.2). Thiessen polygons are often associated with data in vector format (Singh and Fiorentino, 1996). The
used to spatially distribute point rainfall from two or more locations over a process, known as tessellation (Singh and Fiorentino, 1996) produces a set of
basin or subbasin. This method assigns areas of influence to each rain gage nonoverlapping triangular facets of irregular sizes (Vieux, 2004). As
using perpendicular bisectors between gages. Then, the individual areas are described earlier, Delaunay triangulation is one tessellation method, while
weighted by the total basin area and the weights are applied to the precipita- more complex methods can handle breaks in the contours that might be
tion for each gage and then summed to obtain the “weighted average” pre- associated with sinks, wetlands, etc. (Gold, 1989). For simplicity, most TINs
cipitation for the storm for the basin. assume planar triangular facets. The vertices of the triangle edges (see
In deriving Thiessen Polygons for a basin, GIS most often employ Fig. 9.3 as follows) represent terrain features while the edges represent linear
Delaunay triangulation to produce Triangulated Irregular Network (TIN) features (Vieux, 2004), together creating a 3D surface. However, any eleva-
lines connecting each gage location to its two nearest neighbors to form tri- tion point from the whole surface can be easily interpolated from the TIN
angles as shown in Fig. 9.3 (Elangovan, 2006). Delaunay triangulation ensures (Singh and Fiorentino, 1996).
that no vertex of a polygon is within the circle transcribing the triangle con- A TIN has several advantages in that data storage is minimized and the
necting three adjacent data points, and that all angles are acute, a requirement variable sizes of the triangles can conform to irregular topographic surfaces
of the method. fairly well. Steep terrain will be represented by many smaller triangles, while
Another source of precipitation information that is rapidly becoming flat terrain may only take a few larger triangles as shown in Fig. 9.3. In
operational in hydrologic modeling in the United States is derived from radar general, if a gridded distributed hydrologic modeling strategy is to be
reflectivity data (see Table 9.2). A number of commercial vendors can followed, then allowing the data to remain in a gridded format (either the
distribute weather products to their customers in GIS format at spatial and original DEM or some downscaled grid) might be advisable, while if the data
temporal resolutions requested by the customer; however, the most common were in vector format, then a TIN surface would facilitate the extraction of the
source of radar-derived precipitation is from the NOAA NEXRAD. Currently, necessary hydrologic information.
radar data alone are not sufficiently accurate to support hydrologic modeling, For small-scale modeling, the standard source in the United States for soil
although the National Weather Service Stage IV multisensor data set is data is the Soil Survey Geographic (SSURGO) dataset identified in Table 9.2.
09_Singh_ch09_p9.1-9.8.indd 5 8/22/16 11:36 AM

Figure 9.3 A topographic surface represented by a TIN.
It originally consisted of polygons representing each soil series in the data, but mathematical transformation of basin x,y,z coordinates using external com-
recently a gridded data set has been produced by the NRCS at grid resolutions puter codes reading text files was employed to formulate the system, but since
of 10, 30, 90, or 100 m. For broader scale hydrologic work, the State Soil the development of GIS the cascading plane formulation has become easier.
Geographic (STATSGO) database might be appropriate. In STATSGO, similar With the advent of GIS, gridded models became popular, as seen in Fig. 9.5.
soil series from SSURGO were combined into larger soil associations. In doing The fundamental difference between the gridded and cascading plane formu-
this, some details of soil properties are lost, since all soils of similar taxonomy lations is that in the former, a uniform rectangular grid is overlain onto the
(silty clay loams for example) do not behave the same. Beyond STATSGO, a basin and the elevation of each grid is determined. There is no attempt to map
multilayer gridded soils database for the conterminous United States the grid pattern onto the basin topography, but the grids may be made small
(CONUS-Soil) was developed at Penn State University consisting of a 1-km- enough (if the native data allows) such that a fair representation of the topog-
gridded database with soil parameters given in 11 standard layers. Its major raphy results. The gridded formulation fits neatly into a solution technique
advantage is that the key soil parameters are given for each layer, whereas this where the governing equations (e.g., kinematic wave) are solved by numerical
is not the case with the two vector based soils coverages. In the latter cases, techniques where the solution grid conforms to the GIS grid. In this case, the
only the soil series or association is given and an attribute database, or a look- equations and boundary conditions remain in their original coordinates.
up table must be included that contains the necessary data. CONUS-Soil is The problems associated with the simple gridded formulation are evident
very useful for large-scale hydrologic modeling, but is not used much in in the above mentioned Fig. 9.5. Since natural watersheds are not made up of
smaller scale models where STATSGO or SSURGO are more appropriate. simple straight lines, the gridded pattern will not conform to basin
Land cover data are easily available over the United States from the National
Land Cover Database noted in Table 9.2. An accuracy analysis of the 2006
dataset (Wickham et al., 2013) concluded that the overall accuracy, as com-
pared to digital imagery of selected pixel samples, was 78%. This statistic is low
compared to those of other studies associated with specific land cover classifi-
cations reported by Cruise et al. (2010), which were more than 90%. Studies of
the effects of errors in land cover on hydrologic model simulations are sparse,
but an early study by Jackson et al. (1977) concluded that satellite based land
cover parameters might be suitable for planning and analysis models, but
would not be advisable for design purposes. Given the errors associated with
the more modern databases as cited earlier, that conclusion is likely still valid.
The Leaf Area Index (LAI) is a specific parameter associated with land Figure 9.4 Part of a basin represented by cascading planes.
cover, a proxy which can be calculated at relatively low cost and high fre-
quency. Although field-scale studies of LAI of specific vegetation species have
been reported (e.g., Jordan, 1969; Chen et al., 1997) LAI values are now most
often derived from remotely sensed data. NASA produces a near real time LAI
product of the contiguous United States derived from Moderate Resolution
Imaging Spectroradiometer (MODIS) satellite data at a 1-km spatial resolu-
tion, a scale compatible with large-scale modeling but problematic for small-
scale distributed modeling. It is more common for hydrologic models to
employ look-up tables of LAI for each land cover in their database with the
LAI in these tables being long term monthly averages for each land cover.
9.4.3 Watershed Representation for Distributed

Modeling
A watershed can be represented in a GIS for modeling in a number of ways.
In the beginning, the modeler must make a fundamental decision: use the
GIS to gain the most realistic representation of basin physical characteristics
in space, which will exert a profound impact on the modeling approach and
setup, or employ a simplified representation of the watershed that will lead to
a more straightforward model. One of the earliest catchment representations
for distributed modeling was the cascading planes configurations shown in
Fig 9.4 (Lane and Woolhiser, 1977). Various procedures have been developed
to map watershed surfaces and channels into rectilinear planes as outlined by
Lane and Woolhiser (1977) and Guo and Urbonas (2009). The central idea is
that the planes conform to the actual topography of the basin to the extent
possible with runoff routed from planes to their successive downslope planes
depending on slope, aspect, and relative roughness principles. Originally, Figure 9.5 A watershed represented using a gridded model.
09_Singh_ch09_p9.1-9.8.indd 6 8/22/16 11:36 AM

Current Status and Future Directions 9-7
boundaries and hydrographic feature and thus, unavoidable errors are intro- stream network extraction, ordering, and segmenting. In addition, GRASS
duced into the analysis. In order to avoid the errors associated with gridded features an interface with R statistics software (r-project.org) and is capable of
models, other formulations have been derived. For example, the TIN formula- performing Hortonian statistics. GRASS provides complete access to its inter-
tion discussed earlier can be used to discretize basin topography, boundaries, nal structure and algorithms and any module can be customized or improved,
and features better than is possible with simple rectangular schemes. It is many of which are used in other open source desktop GIS.
possible to discretize the TIN representation into a system of smaller finite
elements that can form the basis of the solution domain (Goodrich et al., 1991; 9.5.2 Hydrology Wrappers and Toolkits for GIS
Tisdale et al., 1998). These small elements will better represent the watershed Applications
features than grids or larger triangles. Finite element–based methods are GIS is frequently used in hydrology as an interface with existing distributed
more thoroughly discussed by Vieux (2004). and semidistributed models. In fact, GIS is probably the reason for the
popularity of distributed models as it often used as a static spatial database
9.5 MODEL/GIS INTERFACES manager. GIS is most often applied to the pre- and postprocessing of
hydrologic models; however, some hydrologic modeling software is fully
In the early years of GIS, modeling techniques and analysis were limited to integrated into a GIS with complete functionality. There are many hydrology
research specialists, where detailed text-based datasets were created and used models coupled with GIS and it is beyond the scope of this section to
as input files via a command prompt into complex (usually FORTRAN- describe in detail each one. Table 9.3 provides a summary of widely used
based) programs. Applications of GIS were usually only found in college existing hydrological models that have advanced their functionality by
campuses or research-oriented agencies. The introduction of the graphical integrating GIS at some level.
user interface (GUI) drastically changed the way GIS is used, as it did most
computer applications. With increased usability, many practicing hydrologists 9.6 CURRENT STATUS AND FUTURE DIRECTIONS
now employ GIS technology in all manner of hydrologic applications. There
are many desktop-ready GIS GUIs that exist in both open-source and com- Current and future developments of GIS in hydrology exist in the refinement
mercial or proprietary software and that have applicability in hydrology. To and advancement of existing techniques developed over the past few decades.
cover each in detail is not the purpose of this section; moreover, the list would In general, 4D (x,y,z, and t) modeling is the next major frontier in geospatial
not be static. Instead, this section describes GIS platforms that have specific analysis. Within GIS, it is important to distinguish 3D visualization and 3D
hydrologic functionality and that are commonly represented in research mapping. The mountain in Google Earth is just a visualized DEM underneath
literature. Existing hydrologic toolkits and models that are wrapped around a a folded/stretched 2D raster. Moving from 3D visualization to actual 3D maps
GIS for more functionality in both analysis and design are also presented. in GIS is a development on the horizon. The transition from simple points or
pixels to “voxels” (3D grid cube) has many challenges, not the least of which
9.5.1 Native GIS Functionality is the math of a seamless coordinate system that is projected upward from the
This chapter has outlined the range of ways GIS is used to aid in hydrological Earth’s surface.
analysis and design, identifying how functions, such as delineating a water- One of the biggest limitations to GIS applications in hydrology is its inher-
shed can be performed on any complete GIS platform. Given the rising influ- ent static nature. The inclusion of a temporal variable allows the analysis to
ence of GIS on the field of hydrology, certain platforms have move beyond abstraction and onto truly dynamic processes. In its most basic
hydrology-specific modules that make the analysis more straightforward. applications, spatial models incorporate a temporal component by time-
In the commercial field, ESRI desktop GIS is commonly used in both stamping layers. Distributions of geomorphic states are shown at different
public and private applications. With more than 40% of market share, ESRI’s times without explicit temporal relations among the layers (snapshot model,
ArcGIS is the most widely used commercial GIS software in the world and it LAN92). This simplified model lacks the ability to describe changes in space
specifically includes a module to aid in the analysis of hydrologic systems. though time. Given the complexity of including the time domain in a GIS
The ArcHydro module consists of a framework for database design and auto- [fundamentally a Database Management system (DBMS)], the snapshot
mated tools for common applications. By first assigning key identifying approach remains a widely used method (Gebbet & Pedesma, 2014), though
attributes, routine applications, such as point delineations, flow directions the incorporation of time in GIS has been near the top of the GIS research
and networks, and attribute-based tracing, are automated for ease of use. In agenda for more than 30 years (O’Sullivan, 2005).
addition, ArcHydro allows for the integration of time series data (e.g., water A second frontier in GIS concerns the source and accessibility of the data
quality and sediments), a common point of difficulty for many integrations. as well as the associated processing capabilities. Given the current state of the
Another common GIS platform in hydrological analysis is the open-source field in GIS focused on usability and increasing data and models, cloud stor-
Geographic Resources Analysis Support System (GRASS). GRASS is a general age and computing have dominated recent advancements, though a growing
purpose GIS that contains a number of modules for management, processing, platform in cloud computing is that of a modeling web service to produce
analysis, and visualization as well as modules for the hydrological sciences heterogeneous datasets. These consist of a set of conceptual model routines
(Jasiewicz et al., 2011). The hydrology-based modules streamline tasks such implemented on a common platform and linked to global and local datasets
as basin delineation and morphometric characterization, flow directions, and through web standards provided by the Open Geospatial Consortium (open-
Table 9.3 Commonly-Used Hydrologic Model Interfaces and Integration Schemes

Model Description/full name Platform Integration scheme Web reference
ArcSWAT Soil and Water Assessment Tool ArcGIS Built as an extension onto ArcGIS. http://swat.tamu.edu/software/arcswat/
ArcGIS serves as the SWAT GUI
BASINS Better Assessment Science Integrating point MapWindowGIS Exports shapefiles for further analysis http://water.epa.gov/scitech/datait/ mod-
& Non-point Sources with more advanced GIS els/basins/index.cfm
MODFLOW U.S. Geological Survey modular finite-differ- ArcGIS Geodatabase design framework for use in http://ahgw.aquaveo.com/Steady_State.
ence groundwater flow model ArcHydro pdf
HEC-GeoRAS Hydrologic Engineering Centers River ArcGIS GIS wrapper designed for pre and post http://www.hec.usace.army.mil/software/
Analysis System processing of model output hec-georas/
HEC-GeoHMS Hydrologic Engineering Center’s Hydrologic ArcGIS GIS wrapper designed for pre and post http://www.hec.usace.army.mil/software/
Modeling System processing of model output hec-geohms/
WEAP Water Evaluation And Planning system Native GIS that exports shapefiles for fur- http://www.weap21.org/
ther analysis with more advanced GIS
TOPMODEL TOPography based hydrological MODEL GRASS Fully integrated into GRASS as a stand- http://grass.osgeo.org/grass64/manuals/
alone module r.topmodel.html
SWMM Storm Water Management Model ArcGIS Proprietary software exist linking ArcGIS http://www.innovyze.com/products/
as a GUI infoswmm/
09_Singh_ch09_p9.1-9.8.indd 7 8/22/16 11:36 AM

geospatial.org). Web-enabled models are a promising approach to process Gold, C. Surface interpolation, spatial adjacency in GIS. Three Dimensional
large and distributed datasets. This technology holds a real future in GIS Applications In GIS, edited by J. Raper, Taylor and Francis, London, 1989,
applications in hydrology. With the ability to share data online, open source pp. 21–35.
platforms are blossoming. Additionally, this accessibility creates a stage for Goodrich, D. C., D. A. Woolhiser, and T. O. Keefer, “Kinematic routing
new data streams. The cost of data collection can be high so the ability to using finite elements on a triangular irregular network,” Water Resources
capitalize on the amount information that can be pulled from nontraditional Research, 27 (6): 995–1003, 1991.
sources will be integral. Guo, J. C. and B. Urbonas, “Conversion of natural watershed to kinematic
wave cascading plane,” Journal of Hydrologic Engineering, 14 (8): 839–846,
2009.
REFERENCES
Haylock, M. R., G. C. Cawley, C. Harpham, R. L. Wilby, and C. M. Goodess,
Al-Hamdan, M. Z., W. L. Crosson, M. G. Estes, Jr., S. M. Estes, D. A. “Downscaling heavy precipitation over the United Kingdom: a comparison of
Quattrochi, and D. Johnson, Downscaling MODIS land surface temperature dynamical and statistical methods and their future scenarios,” International
for urban public health applications. American Geophysical Union Fall Journal of Climatology, 26 (10): 1397–1415, 2006.
Meeting, December 9–13, 2013, San Francisco. Jackson, T. J., W. N. Fitch, and R. M. Ragan, “Test of Landsat-based urban
Chen, J. M., P. M. Rich, S. T. Gower, J. M. Norman, and S. Plummer, “Leaf hydrologic modeling,” Journal of the Water Resources Planning and
area index of boreal forests: theory, techniques, and measurements,” Journal Management Division, 103 (1): 141–158, 1977.
of Geophysical Research, 102 (D24): 29429–29443, 1997. Jasiewicz, J. and M. Metz, “A new GRASS GIS toolkit for Hortonian analy-
Cosgrove, B. A. “Real-time and retrospective forcing in the North sis of drainage networks,” Computers and Geosciences, 37 (8): 1162–1173,
American Land Data Assimilation System (NLDAS) project,” Journal of 2011.
Geophysical Research, 108 (D22): 8842, 2003. Jordan, C. F., “Derivation of leaf-area index from quality of light on the
Cruise, J. F., C. A. Laymon, and O. Z. Al-Hamdan, “Impact of 20 years of forest floor,” Ecology, 50: 663–666, 1969.
land-cover Change on the hydrology of streams in the southeastern Lane, L. J. and D. A. Woolhiser, “Simplifications of watershed geometry
United States,” Journal of the American Water Resources Association, 46 (6): affecting simulation of surface runoff,” Journal of Hydrology, 35 (1–2): 173–
1159–1170, 2010. 190, 1977.
Danielson, J., and D. Gesch, D. Global multi-resolution terrain elevation Maune, D., J. Maitra, and E. McKay, Accuracy standards and guidelines.
data 2010 (GMTED2010). Open-File Report 2011-1073, U.S. Geological Digital Elevation Model Technologies and Applications—the DEM User’s
Survey, South Dakota. Manual, 2nd ed., edited by D. F. Maune, American Society for Photogrammetry
Dewberry, Inc. National Enhanced Elevation Assessment—final Report and Remote Sensing, Bethesda, MD, 2007, pp. 65–97.
(revised March 29, 2012), Fairfax, VA, 2012. Retrieved from http://www.dew- O’Sullivan, D., “Geographical information science: time changes every-
berry.com/Consultants/GeospatialMapping/. Accessed December 01 2014. thing,” Progress in Human Geography, 29 (6): 749, 2005.
Elangovan, K. GIS: Fundamentals, Applications and Implementations. New Rodell, M., P. R. Houser, U. Jambor, J. Gottschalck, K. Mitchell, C-J. Meng,
India Publishing, 2006, Delhi. and D. Toll, “The global land data assimilation system,” Bulletin of the
Fotheringham, S. and P. Rogerson (eds.). Spatial Analysis and GIS. CRC American Meteorological Society, 85 (3): 381–394, 2004.
Press, 2013, Boca Raton. Singh, V. and M. Fiorentino, Geographical Information Systems in Hydrology,
Fowler, H. J., S. Blenkinsopa, and C. Tebaldi, “Linking climate change mod- Springer, 1996, Netherlands.
elling to impacts studies: recent advances in downscaling techniques for Sugarbaker, L. J., E. W. Constance, H. K. Heidemann, A. L. Jason, V. Lukas,
hydrological modelling,” International Journal of Climatology, 27: 1547–1578, D. L. Saghy, and J. M. Stoker, The 3D Elevation Program Initiative—A Call for
2007. Action, 2014. U.S. Geological Survey Circular 1399, South Dakota.
Gebbert, S., and E. Pebesma, “A temporal GIS for field based environmental Tisdale, T. S., P. D. Scarlatos, and J. M. Hamrick, “Streamline Upwind
modeling,” Environmental Modelling & Software, 53, 1–12, 2014. Finite-Element Method for Overland Flow,” Journal of Hydraulic Engineering,
Gesch, D. B., M. J. Oimoen, Z. Zhang, D. J. Meyer, and J. J. Danielson, 124 (4), 350–357, 1998.
“Validation of the ASTER Global Digital Elevation Model Version 2 over the Tomlinson, R. F., “Geographic Information Systems—a new frontier,” The
conterminous United States. Imaging a Sustainable Future, edited by M. Operational Geographer, 5 (1): 31–35, 1984.
Shortis and N. El-Sheimy, International Archives of Photogrammetry, Vieux, B. E., Distributed Hydrologic Modeling Using GIS, Springer Science
Remote Sensing and Spatial Information Sciences 39 (B1), 2012, pp. 281–286, and Business, 2004, Netherlands.
Netherlands. Weng, Q., Remote Sensing and GIS Integration: Theories, Methods, and
Gesch, D. B., M. J. Oimoen, and G. A. Evans, Accuracy assessment of the Applications, McGraw-Hill, 2010, New York.
U.S. Geological Survey National Elevation Dataset, and comparison with Wickham, J. D., S. V. Stehman, L. Gass, J. Dewitz, J. A. Fry, and T. G. Wade,
other large-area elevation datasets—SRTM and ASTER. U.S. Geological Survey “Accuracy assessment of NLCD 2006 land cover and impervious surface,”
Open-File Report 2014-1008, 2014, U.S. Geological Survey, South Dakota. Remote Sensing of Environment, 130: 294–304, 2013.
09_Singh_ch09_p9.1-9.8.indd 8 8/22/16 11:36 AM

Chapter 10
Design of Hydrologic Networks
BY
ASHOK K. MISHRA AND VIJAY P. SINGH
ABSTRACT It will be pertinent to revisit which the agencies-local, state, and federal-col-
lect what data in the United States. The National Weather Service of the U.S.
There are several methods that are used for designing hydrologic networks, Department of Commerce is responsible for the collection of hydrometeorol-
but all of these methods are based on empirical data already collected or to be gical data, including rainfall and evaporation. Of all the hydrologic data,
collected. Examples of these methods include correlation, information vari- rainfall data collection has the longest history. Snowfall data, on the other
ance, transfer function variance, network theory, and entropy theory. All hand, especially in the Western United States, are collected by Natural
these methods are mostly statistical in character and involve little hydrology. Resources Conservation Service of the U.S. Department of Agriculture.
This chapter provides a snapshot of these methods. Streamflow, groundwater, and sediment data as well as geomorphic data are
collected by the Geological Survey of the U.S. Department of the Interior,
referred as USGS. Collection of groundwater and sediment data started much
10.1 INTRODUCTION
later. Water quality data are collected by departments at the state governmen-
Hydrologic information collected over a basin constitutes the basic input for tal level that deal with environment and/or water resources. Groundwater
the design of various water resources projects, such as planning and design of data are also collected by these departments. USGS also collects water quality
reservoirs, drainage systems, flood protection works, parking lots, highway and groundwater data. Soil moisture data are collected on a limited basis by
embankments, arterial airfields, watershed protection works, water distribu- Agricultural Research Service of the U.S. Department of Agriculture. Data on
tion systems, and irrigation networks; as well as for the assessment of climate interception and forest hydrology are collected by Forest Service of the U.S.
change impact on water resources, determination of environmental flows, Department of Agriculture. In addition, there are a number of local agencies
and determination of eco-indices. Hydrologic data are useful for different or bodies that collect a variety of data on hydrology. Data collected by federal
stakeholders, for example, hydrologists, agronomists, climatologists, hydroge- agencies are freely available, but that may not be true for the data collected by
ologists, ecologists, and water resources managers. Water resources managers local or state agencies.
need hydrologic information for better reservoir operation and providing Water quality networks, groundwater networks, and sediment networks
information on flood forecasting and drought assessment. were established much later than were rainfall and streamflow. With growing
Data collection is considered as the first step in the development of sound emphasis on ecosystems health and integrity, environmental quality, and cli-
and sustainable environmental policy (CEAC, 1987), which also requires bet- mate change impact assessment, networks for the measurements of other
ter coordination of data collection activities to maximize the information that components of the hydrologic cycle are now receiving significant attention.
can be derived from the collected data. There is a continuous effort to manage Examples of such networks are the networks for measurement of soil mois-
continuous supply of reliable and timely data, and in this regard, World ture, land use, snow, urban pollution, water pollution, and air quality. With
Meteorological Organization (WMO) plays an important role in supporting remote sensing technologies becoming popular, spatial hydrologic data are
the National Hydrological Services in adopting modern data collection tech- now becoming available. In this chapter, the discussion will be focused on
niques and developing their data management capabilities. rainfall and streamflow networks.
During recent decades, there has been a decline in the amount of hydro-
logic data, especially river flow data, being collected in many regions of the
world. Most of these regions are located in Africa, Latin America, and Asia 10.3 NECESSITY OF HYDROLOGIC NETWORKS
(WMO, 1996). Possible reasons for the reduction in hydrologic networks are Hydrologic networks are necessary to provide information needed for fresh-
insufficient funding, inadequate institutional framework, and a lack of appre- water planning and management, and quantify climatic variability (e.g., spa-
ciation of the worth of long-term hydrologic data. Many studies have empha- tial and temporal distribution of rainfall) over the basin, and the networks
sized that current data collection networks are inadequate for providing the should, therefore, be able to capture the spatial variability of rainfall and
information required to understand and explain changes in natural ecosys- storm runoff for improving flood management. Hydrologic data are useful to
tems. investigate climatic patterns, such as trend analysis of mean annual minima of
river flows, cyclic patterns of streamflow corresponding to cyclic patterns of
rainfall. Lack of hydrologic information can lead to poor decisions on invest-
10.2 HYDROLOGIC NETWORKS
ment in water supply, hydropower, irrigation, and other projects, and the
Over the years, much emphasis in the hydrologic network design literature failure to provide timely warnings of drought and flood events to the popula-
has been on streamflow and precipitation networks. This is partly because tion. The consequences of poor hydrologic network design can include pay-
rainfall and streamflow measurements have a long history and have ubiqui- ing too much for a project, loss of property or lives, a shortfall of revenue or
tous applications in design of water resources projects and civil engineering water, or unforeseen adverse impacts on other users (e.g., through water log-
design works. Also, these networks are federal networks in the United States. ging, salinization, impacts on wetlands, lakes, floodplains, estuaries, etc.).
10-1
10_Singh_ch10_p10.1-10.6.indd 1 8/22/16 11:37 AM

10-2 Design of Hydrologic Networks
10.4 IMPACT OF HYDROLOGIC NETWORK DENSITY ON There is a large body of literature on the design of monitoring networks in
STREAMFLOW ESTIMATES surface hydrology (Moss, 1986; van der Made, 1986, 1987; Harmancioglu
Hydrologic network density influences the accuracy of streamflow simula- et al., 1999; Krstanovic and Singh, 1988, 1992a, b; Alfonso et al., 2010; Mishra
tion. Krajewski et al. (1991) found that temporal sampling density (of rain and Coulibaly, 2009, 2010, 2014; Li et al., 2012; Singh, 2013, 2015). Different
gages) affected hydrograph characteristics, such as peak magnitude, time to methods have been developed for network evaluation and design, including
peak, and runoff volume. For determining the effect of the number of rain (1) statistical methods; (2) network theory based method; (3) information
gages on streamflow accuracy, Georgakakos et al. (1995) considered a number theory based methods; (4) users survey method; (5) hybrid method; (6) phys-
of rain gauges (from 1 to 11) in streamflow simulation in two American river iographic components; and (7) sampling methods.
basins with an area of about 2000 km2. It was concluded that five rain gauges 10.6.1 Statistical methods
were adequate and concluded that 11 rain gauges were more than adequate to
represent mean areal precipitation over the catchment. The appropriate num- Statistical methods are amongst the oldest methods used for network design
ber and locations of rain gauges, which yield the best and worst hydrological (Wilm et al., 1939). These methods typically assume that the basin response,
modeling performance depends on the local geographical and climatic pat- land use, and climate remain the same over time. Commonly used statistical
terns (Dong et al., 2005). Also affecting the model performance is the combi- methods are (a) correlation, (b) variance reduction, (c) transfer function, and
nation of rain gauges (Anctil et al., 2006) and the spatial resolution of rainfall (d) transfer function variance.
input in model calibration and application (Bardossy and Das, 2006). The (a) Correlation: Cross-correlation between every two stations can be used
complexity in the identification of an optimal subnetwork is a challenge. This as a measure of information between them. If x and y denote records of two
discussion reveals that a hydrologic network is driven by multiple factors, stations, then the coefficient of correlation between them, denoted as rxy, can
which need to be considered in design. be expressed as:
Cov( x , y ) (10.1)
rxy =
10.5 DESIGN CONSIDERATIONS Sx S y

The methodology for design of hydrologic network should consider the
objectives of sampling, variables to be sampled, the locations of measurement in which Cov (x, y) is the covariance between x and y, Sx is the standard
stations, frequency of sampling, duration of sampling, use and users of data, deviation of x, and Sy is the standard deviation of y. It is logical to assume that
and socioeconomic considerations. Further, it is important to consider the the correlation between stations decreases with increasing distance. Often,
amount of information that each gaging station produces and a station with the relation between the correlation coefficient and distance is expressed in an
higher information should receive higher priority, but this information must exponential form as:
be balanced with site-specific uses and users of the information. A station
whose data is used by more and diverse users will be preferred to the station  d (10.2)
r (d ) = exp  − 
whose data is used by fewer and less diverse people.  λ
Current design practices have two modes. The first is space design that
determines the number of gaging stations and their locations. The second is where r is the correlation coefficient, d is the distance, and λ is parameter,
time design that determines the frequency, time interval, and duration of often called correlation length whose inverse yields the correlation decay rate.
sampling. It is possible to supplement the information in one mode by the Thus, a plot of the coefficient of correlation against distance between stations
information in the other mode, called space-time tradeoff, using appropri- and fixing a threshold value of the correlation coefficient permit the determi-
ate transfer methods or cross-correlation structure. It is important to nation of the location of gages and the number of gages.
remember that a network design must not be permanent but must be (b) Variance reduction methods: In general, statistical methods are essen-
amended in light of changing conditions and demands. Another important tially based on the minimization of variance. The information–variance
consideration is the cost-effectiveness of data collection. In other words, method involves space-time trade-off, that is, the lack of data in one domain
long-term benefits of data collection and the accomplishments of objectives (e.g., time) can be compensated for by extrapolation of similar data in
as a result must outweigh the cost of data collection. That is, the network another domain (e.g., space). A higher variance indicates a greater measure-
must yield the greatest benefit or most information for the budget allocated. ment error, implying more gages are needed in the basin and vice-versa
One does not need to be reminded that often it is the budget that controls (Silverman et al., 1981). The variance method can be applied to the spectral
the network design. density function to measure the distribution of the variance of data (say,
rainfall) over the range of its inherent frequencies. Variance can be
expressed as a function of correlation in time, correlation in space, length of
10.6 DESIGN OF HYDROLOGIC NETWORKS:
operation of the network, and geometry of the gauging array (Rodriguez-
METHODOLOGIES
Iturbe and Mejia, 1974).
One of the fundamental issues that must be incorporated in any kind of (c) State space stochastic models: Hydrologic data can be regarded as a
gauge network design is the estimation of streamflow/precipitation at manifestation of a multidimensional stochastic process, which can be sto-
ungauged sites from information at gauged sites, such that space–time vari- chastically modeled. Therefore, by combining the knowledge of such a pro-
ability of streamflow/rainfall is sampled optimally. In addition to spatial cess and multivariate estimation theory and optimization methods, one can
density, which is an integral part of network design, additional information determine the least-cost network (Bras and Rodriguez, 1985). This suggests
needs to be considered for hydrologic network design, such as current needs that data collection can be viewed as a statistical estimation problem. The
of users, variability of streamflow/rainfall patterns, maintenance costs, and stochastic model can include spatial variability of rainfall and has, therefore,
accessibility. WMO (1965) emphasized the importance of establishing an an advantage in network design in regions with strong climatic variations and
optimum network, and suggested that hydrometric stations should be land use changes.
divided into three categories: principal stations, secondary stations, and (d) Dimension reduction methods: These methods are used to reduce
special stations. Principal stations provide the basis for statistical studies multidimensional data sets to lower dimensions for analysis. Clustering
and should be continued, whereas secondary stations should operate for a approach is one such example, where the partitioning of data set is done into
limited number of years to establish a good correlation with principal sta- subsets that share some common traits—often proximity according to some
tions, and special stations are established for specific purposes and for defined distance measure. The objective is to identify groups of homogenous
limited periods of time. stations for optimal interpolation and to locate stations if needed in the
Criteria commonly used to identify representative stations are as follows: future. Commonly used methods are: principal component analysis and clus-
(1) Primary stations: Select the primary stations that provide information for tering methods. This approach can be used for network rationalization, that
multiple stakeholders and must be kept in the network. (2) Scenario analysis: is, eliminating redundant stations.
In addition to primary stations, select a combination of stations and rank dif- (e) Network analysis for regional information: The network analysis for
ferent scenarios based on their capacity to reduce overall modeling errors. regional information (NARI) method maximizes regional information
(3) Cost comparison: Evaluate and compare the cost associated with maintain- within a limited budget and time horizon. It was developed by simulating
ing and continuing different combinations of stations. (4) Optimal network: regional regressions of the logarithms of streamflow parameters against the
The combination of stations providing more information as well as minimum logarithms of drainage area. The underlying regression model will have a
error and minimum cost can be chosen as representative stations. The follow- finite model error that is homoscedastic with respect to drainage area.
ing section provides an overview of the historical and more recent develop- Tasker and Moss (1979) conducted a study for network of stream-gaging
ments in hydrometric network design. stations in northwest Arizona in order to identify a logical gauging strategy
10_Singh_ch10_p10.1-10.6.indd 2 8/22/16 11:37 AM

Design of hydrologic networks: Methodologies 10-3
for the area by including network operating costs for planning periods of 10, homogeneous regions having more than 70 gaging stations. It was found that
20, and 30 years. the number of stations needed depended on the time interval. If the time
(f) Network analysis using generalized least squares: Analogous to NARI, interval was 1 year then only one raingage was needed in each homogeneous
the network analysis using generalized least squares (NAUGLS) also maxi- region and other stations would reproduce more or less the same information
mizes regional information within a limited budget. However, the NAUGLS and would therefore be redundant. Mishra and Coulibaly (2014) applied the
method conveys more information than the NARI method. Parameters of the entropy method to identify redundant stations as well as most important sta-
NAUGLS method are estimated using the generalized least squares (GLS) tions based on three seasons for the Upper St. Lawrence basin in Canada.
estimator proposed by Stedinger and Tasker (1985). Tasker (1986) demon- Daily streamflow data for 25 stations (Table 10.1) were collected from
strated an application of NAUGLS by identifying an efficient gaging plan for Environment Canada hydrometric database (HYDAT CD-ROM Version
a specific operating budget. His study also provided insights into the amount 2.04, 2004) and the data were categorized into three seasons: April–July,
of regional information lost or gained by reducing or increasing the operating August–November, and December–March. They calculated mutual informa-
budget. tion (MI) based on the kernel density approach with the Gaussian kernel
(g) Spatial interpolation techniques: Geostatistical methods are spatial function. Stations with a high value of MI score would indicate a strong
interpolation techniques that have been applied to hydrologic network dependence between two variables. The stations rankings were derived based
design. Kriging methods are popular among spatial interpolation techniques, on the amount of information contents for different seasons (Table 10.1). It
which consist of two parts. First, analysis of spatial variation using vario- can be observed from the table that there is a variation in the station rankings
grams, which plot the variance of sample measurements as a function of dis- for different seasons, indicating that seasonal climate variability has an impact
tance is done. Second, an appropriate parametric model is then typically fitted on the information contents shared between stations. Based on station rank-
to calculate distance weights for interpolation. Once a suitable variogram ings, the most important stations were: 02AB006 and 02CF013. The station
model has been selected, weights are used to create a continuous surface for rankings for station 02AB006 were: 1 (April–July), 3 (August–November),
the entire basin, based on the variogram model and the values and location of and 1 (December–March), whereas for station 02CF013, they were 2 (April–
measured points. Using a simple and disjunctive kriging, Kassim and July), 1 (August–November), and 2 (December–March). The stations that
Kottegoda (1991) estimated the optimum locations of recording rain gauges shared more information had higher ranks 02CC005, 02CC010, 02CA006,
(Birmingham, the U.K.) for the determination of storm characteristics to be and 02CA003, with ranks varying between 21 and 25. The spatial interpola-
used in forecasting and design. tion methods can be used for constructing a contour map to identify critical
areas based on transinformation measures. The spatial pattern of mutual
information for the Upper St. Lawrence basin is shown in Fig. 10.1. Based on
10.6.2 Network Theory this contour map one can identify surplus and deficit locations of stations.
The network theory can be applied to determine such properties of a network This illustration can be considered as a preliminary step to identify critical
as clustering, topology, adjacency, and centrality. There are different measures areas for future streamflow network design. However, to optimize network
to represent these properties, such as clustering coefficient, degree distribu- design in each region, a finer spatial analysis is recommended for further
tion, average shortest length, degree centrality, and community structure. study. Policy makers can increase/decrease the number of stations based on
These properties permit to determine the gaging stations that are similar and the needs of the users based on a users survey approach and/or budgetary
those that are dissimilar. Thus, redundant stations can be identified and constraints.
eliminated. The theory of networks is described in a separate chapter in the
handbook and will not be repeated here. Sivakumar (2015) has described a
generic theory of networks in hydrology and Sivakumar and Woldemeskel Table 10.1 Location of Stations Selected in Upper St. Lawrence
(2015) have discussed a network-based analysis of spatial rainfall connec- Basin, Canada
tions.
Station rank
Station ID April–July August–November December–March
10.6.3 Entropy-Based Methods
02AB006 1 3 1
The main entropy concepts needed for hydrologic network design are mar-
ginal entropy, joint entropy, transinformation, and multivariate correlation. 02AB008 3 6 4
These entropy concepts and entropy theory are described in a separate chap- 02AB017 7 4 3
ter and will therefore not be repeated here. Hydrologic network design, using 02AC002 5 8 6
entropy-based methods, is based on the concept that the measurements or
data of a station contain a certain amount of information and this information 02AD010 8 7 12
can be quantified using the entropy concept. The information of measured 02BA003 16 18 16
data of any station is denoted by marginal entropy. The joint or collective
02BB003 13 16 20
information of two stations is then denoted by joint entropy. Likewise, the
information common between two stations is denoted by transinformation. 02BB004 6 5 11
The correlation coefficient, based on entropy, is called the informational cor- 02BC004 4 2 7
relation coefficient. The coefficient of correlation amongst many stations is
02BF001 12 9 13
called multivariate correlation.
It is logical to say that stations should have as little transinformation as 02BF004 11 13 10
possible, which means the stations should be as much independent of each 02CA002 14 15 15
other as possible. The stations with lower transinformation value mean they
02CA003 22 23 22
share less common information and can be treated as more independent. On
the other hand, stations with larger transinformation indicate that they are 02CA006 23 22 21
more dependent and likely duplicate the same information. In simple words, 02CB003 18 12 18
additional stations can be located at places where small transinformation
02CC005 25 25 24
values are obtained. Using these entropy concepts, it is possible to select a
station in a network that is most informative. One can then determine sta- 02CC008 10 10 5
tions that are most independent of each other. Hence, one can expand or 02CC010 24 24 25
reduce an existing network by identifying the zones with minimum hydro-
02CD001 19 14 19
logic information. For example, loss or gain of information at a specific loca-
tion can decide whether to add or remove a gage station. Recent years have 02CD006 15 17 17
witnessed a multitude of applications of these entropy concepts in hydrologic 02CE002 21 21 23
network design (Harmancioglu et al., 1999; Mishra and Coulibaly, 2010, 2014;
02CF007 20 19 14
Li et al., 2012; Alfosno, 2010; Alfonso et al., 2010; Singh, 2013, 2015).
Krstanovic and Singh (1988, 1992a, b) employed marginal and joint 02CF008 17 20 9
entropies and transinformation to evaluate the raingage network in the entire 02CF012 9 11 8
state of Louisiana for different time intervals, ranging from 1 h to 1 day to 1
week to 1 month to a year. The state is divided into 10 meteorologically 02CF013 2 1 2
10_Singh_ch10_p10.1-10.6.indd 3 8/22/16 11:37 AM

10-4 Design of Hydrologic Networks
April–July August–November December–March
Upper St. Lawrence basin

Figure 10.1 Contour maps of mutual information for April–July, August–November, and
December–March for Upper St. Lawrence basin (dark grey with higher mutual information denote
highly deficit and medium grey denotes excess with lower mutual information values).
10.6.4 Sampling Strategies of user needs was identified: (a) most users used historical data; (b) cooperating
Streamflow in a basin is impacted by the spatio-temporal distribution of pre- agencies most often used data for current-use needs; (c) most frequently used
cipitation, number of rain gauges, rainfall sampling interval, and discharge information is annual peak flow, followed by daily mean flow; (d) most com-
measurement interval. Therefore, it is important to evaluate the networks based mon data applications are hydrologic/hydraulic design; (e) 40% of the users
on different sampling options. Tsintikidis et al. (2002) investigated possible sites were interested in processed data, such as low-flow and water-quality statistics,
for the placement of additional gauges by developing a series of surrogate flood frequency, flood stages, and flood forecasts, that were the most valuable
operational networks from the dense network using an increasing number of additional information for their uses; and (f) 25% of the respondents expressed
gauges and quantifying the reduction of estimation error in each case. Due to an interest in using real-time data.
the placement of an additional station, there was a reduction in local variance
up to 70%, with substantial reduction of the estimation bias for a study area 10.7 CONCLUSION
located in the headwaters of the Middle and North Forks of the American River.
There is a large variety of methods for monitoring, evaluating, and designing
10.6.5 Hybrid Methods hydrologic networks. The networks to be installed must be based not only on
technical considerations but also socioeconomic considerations. There
To maximize the hydrologic information from a network, hybrid methods should be an interface between the data obtained from the networks and the
explore either the advantage of individual methods and/or the combination of fate of data upon collection. Issues pertaining to data storage, archiving, pro-
methods based on using the output from one method as an input to another cessing, and retrieval are equally important and must be dealt with equal
method. Consider the problem of a network design that consists of minimiz- importance. Further, there must be a greater appreciation of the worth of data
ing both the accuracy of areal mean estimation and the economic cost of data outside of the technical or professional communities and equally important is
collection. This problem can be solved by combining: (a) kriging for variance the appreciation of the value of the people engaged in data monitoring and
reduction, (b) optimization method for the cost reduction analysis, and (c) collection. This issue greatly affects the quality and reliability of data being
developing a hydrologic model for conducting different network scenarios as collected, especially in developing countries.
needed by stakeholders, for example, evaluating flow (high and low) predic-
tions.
Marcus et al. (2003) conducted a study to evaluate the streamgauging net- REFERENCES
work in the state of Illinois for quantifying the regional value of each station Alfonso, L., Optimization of Monitoring Networks for Water Distribution
through an assessment of the information transfer among gaging records for Systems, Ph.D. dissertation, Delft University of Technology, Delft, The Neth-
different (low, average, and high) flow conditions. The problem was addressed erlands, 2010.
by combinations of methods: (a) entropy-based information measures were Alfonso, L., R. Price, and A. Lobbrecht, “Information theory-based
derived to determine the regional value of each station within a region and approach for location of monitoring water levels gauges,” Water Resources
then stations were ranked based on the total net information. (b) The general- Research, 46: w03528, 2010, DOI: 10.1029/2009WR008101.
ized least square regression (GLS) method was used to compare the regional Anctil, F., N. Lauzon, V. Andréassian, L. Oudin, and C. Perrin, “Improve-
value of stations. (c) Both entropy and GLS methods were combined by ment of rainfall-runoff forecasts through mean areal rainfall optimization,”
including a function of the negative net information as a penalty function in Journal of Hydrology, 328: 717–725, 2006.
the GLS. Bardossy, A. and T. Das, “Influence of rainfall observation network on
model calibration and application,” Hydrology and Earth System Sciences, 3:
10.6.6 User Survey Method
3691–3726, 2006.
The user survey method is based on the stakeholder’s feedback whether to Bras, R. L. and I. Rodriguez-Iturbe, Random Functions and Hydrology,
continue/discontinue gauging stations, depending upon the type of data Addison-Wesley, Reading, MA, 1985.
requirement in the basin. One of the common approaches is to develop a set CEAC (Canadian Environmental Advisory Council), Canada and Sustain-
of questions, or questionnaire, by giving importance to a set of parameters, for able Development. Report of the World Commission on Environment and
example, (a) length and quality of record, and (b) usefulness of data for water Development and Its Implications for Canada, 8–12, 1987.
resources projects in the region. The survey questions provide greater impor- Chao Li., V. P. Singh, A. K. Mishra, “Entropy theory-based criterion for
tance to the parameters that are more important for the region. Since this hydrometric network evaluation and design: maximum information mini-
approach relies on personal judgment, it is important to take feedback from mum redundancy,” Water Resources Research, 48: W05521, 2012.
different sectors and users. The users can be chosen from a broad range of Dong, X., M. Dohmen-janssen, and M. J. Booij, “Appropriate spatial sam-
disciplines, for example, hydrologists, meteorologists, agronomists, as well as pling of rainfall for flow simulation,” Hydrological Sciences Journal, 50 (2):
planning commissions, and local, state, and federal agencies. This approach 279–298, 2005.
involves: (a) an assessment of individual stations in terms of the extent to Georgakakos, K. P., D. H. Bae, and D. R. Cayan, “Hydroclimatology of
which they contribute to the full set of priority considerations; (b) the weights continental watersheds: 1. Temporal analyses,” Water Resources Research, 31
that are combined for a particular station to evaluate its relative value or ben- (3): 655–675, 1995.
efits derived as well as costs to maintain those stations; (c) stations receiving Harmancioglu, N. B., O. Fistikoglu, S. D. Ozkul, V. P. Singh, and M. N.
lowest weights that are identified as possible candidates for discontinuance if Alpaslan, Water Quality Monitoring Network Design, Kluwer Academic Pub-
budgetary considerations required such an action; and (d) multiple combina- lishers, Boston, MA, 1999, p. 299.
tions of stations that, within the existing budget, are compared based on their Kassim, A. H. M. and N. T. Kottegoda, “Rainfall network design through com-
importance in the region in order to find the best combinations. parative kriging methods,” Hydrological Sciences Journal, 36 (3): 223–240, 1991.
For example, Singh et al. (1986) developed a questionnaire for 158 users of Krajewski, W. F., V. Lakshmi, K. P. Georgakakos, and S. C. Jain, “A Monte
streamflow data, including consulting firms, planning commissions, state and Carlo study of rainfall sampling effect on a distributed catchment model,”
federal agencies, public utilities, university professors, and others. A combination Water Resources Research, 27 (1): 119–128, 1991.
10_Singh_ch10_p10.1-10.6.indd 4 8/22/16 11:37 AM

REFERENCES 10-5
Krstanovic, P. F. and V. P. Singh, “Application of entropy theory to multi- Sivakuamr, B., “Networks: a generic theory for hydrology,” Stochastic Envi-
variate hydrologic analysis,” Technical Report WRR8 and WRR9, Department ronmental Research and Risk Analysis, 29: 761–771, 2015.
of Civil Engineering, Louisiana State University, Baton Rouge, LA, 1988. Sivakumar, B. and F. M. Woldemeskel, “A network-based analysis of spatial
Krstanovic, P. F. and V. P. Singh, “Evaluation of rainfall networks using rainfall connections,” Environmental Modeling Software, 69: 55–62: 2015 (in
entropy: 1. Theoretical development,” Water Resources Management, 6: 279– press).
293, 1992a. Stedinger, J. R. and G. D. Tasker, “Regional hydrologic analysis—ordinary
Krstanovic, P. F. and V. P. Singh, “Evaluation of rainfall networks using weighted and generalized least squares compared,” Water Resources Research,
entropy: 2. Application,” Water Resources Management, 6: 295–314, 1992a. 21 (9):1421–1432, 1985.
Markus, M., H. V. Knapp, and G. D. Tasker, “Entropy and generalized least Tasker, G. D. and M. E. Moss, “Analysis of Arizona flood data network for
square methods in assessment of the regional value of streamgages,” Journal regional information,” Water Resources Research, 15 (6): 1791–1796, 1979.
of Hydrology, 283: 107–121, 2003. Tasker, G. D., “Generating efficient gauging plans for regional information,”
Mishra, A. K. and P. Coulibaly, “Developments in hydrometric network Proceedings of the Budapest Symposium on Integrated Design of Hydrological
design: a review,” Reviews of Geophysics, 47: RG2001, 2009. Networks, edited by M. E. Moss, International Association of Hydrological
Mishra, A. K. and P. Coulibaly, “Variability in Canadian seasonal stream- Sciences (IAHS) Publication, 1986, Vol. 158, pp. 269–281.
flow information and its implication for hydrometric network design,” Jour- Tsintikidis, D., K. P. Georgakakos, J. A. Sperfslage, D. E. Smith, and T. M.
nal of Hydrologic Engineering, 19 (8): 05014003, 2014. Carpenter, “Precipitation uncertainty, rain gauge network design within
Mishra, A. K. and P. Coulibaly, “Hydrometric network evaluation for Cana- Folsom Lake watershed,” Journal of Hydrologic Engineering, 7 (2): 175–184,
dian watersheds,” Journal of Hydrology, 380 (3–4): 420–437, 2010. 2002.
Moss, M. E. (Ed.), Integrated Design of Hydrological Networks, IAHS Publi- Van der made, J. W., “Design aspects of hydrological networks,” Proceedings
cation No. 158, IAHS Press, Wallingford, England, 1986, p. 415. and Information No. 15, TNO Committee on Hydrological Research, The
Rodriguez-Iturbe, I. and J. M. Mejia, “The design of rainfall networks in Hague, The Netherlands, 1986, p. 192.
time and space,” Water Resources Management, 10 (4): 713–728, 1974. Van der made, J. W., Analysis of Some Criteria for Design and Operation of
Silverman, B. A., L. K. Rogers, and D. Dahl, “On the sampling variance of Surface Water Gauging Networks, Van Gorchum, Assen, The Netherlands,
raingage networks,” Journal of Applied Meteorology, 20: 1468–1478, 1981. 1987, p. 437.
Singh, K. P., G. S. Ramamurthy, and M. L. Terstriep, Illinois Streamgaging Wilm, H. A., A. Z. Nelson, and H. C. Story, “An analysis of precipitation
Network Program: Related Studies and Results. Illinois State Water Survey measurements on mountain watersheds,” Monthly Weather Review, 67 (6):
Miscellaneous Publication 94, Champaign, Illinois, 1986. 163–172, 1939.
Singh, V. P., Entropy Theory and Its Application in Environmental and Water World Meteorological Organization, Guide to Hydrometeorological Prac-
Engineering, Wiley-Blackwell, Sussex, 2013. tices, 1st ed., WMO No. 168, TP.82, Geneva, Switzerland, 1965.
Singh, V. P., Entropy Theory in Hydrologic Science and Engineering, World Meteorological Organization, The Adequacy of Hydrological Net-
McGraw-Hill Education, New York, 2015. works: A Global Assessment, WMO/TD 740, Geneva, 1996, p. 56.
10_Singh_ch10_p10.1-10.6.indd 5 8/22/16 11:37 AM

PART 3
METHODS
11_Singh_ch11_p11.1-11.6.indd 1 8/22/16 11:38 AM

11_Singh_ch11_p11.1-11.6.indd 2 8/22/16 11:38 AM
Chapter 11
Artificial Neural Networks
BY
GOKMEN TAYFUR AND VIJAY P. SINGH
ABSTRACT FFNN, briefly describes ANNs and their applications to modeling hydrologi-
cal processes and highlights their limitations and merits.
Artificial neural networks (ANNs) are employed for prediction, forecasting,
and simulation of hydrological systems. They are effective in handling
dynamic, nonlinear, and noisy data, especially when the underlying relation- 11.2 HISTORICAL DEVELOPMENT
ships are not fully understood. They are flexible, self-adaptive and less
assumption-dependent, and intrinsically general, and perform well in most McCulloch and Pits (1943) designed the first artificial neuron performing
applications where other models either tend to fail or become cumbersome. simple logical functions such as “A and B” or “A or B.” Hebb (1949) devel-
Nevertheless, there are some drawbacks and limitations of ANNs, such as oped the first learning law for a neuron. Farley and Clark (1954) and
overtraining, getting trapped in local minimum and insufficient capacity of Rochester et al. (1956) performed computer simulations of neurons.
extrapolation. Also, they require good quality and quantity of data, and they Neuroscience was influential in the development of artificial neurons and so
can not explicitly reveal the relationship between the physical problem and were the physiologists and engineers in the early years.
the network structure. This chapter, focusing on the commonly used feed In late 1950s and early 1960s, researchers developed a large class of ANNs
forward neural network (FFNN) approach, provides a short historical called perceptrons. The perceptron learning rule used an iterative weight
account of developments leading up to the present time, discusses relevant adjustment. Rosenblatt (1958) was the first one who developed the percep-
concepts and methods, limitations, drawbacks, and concludes with a sum- tron having three layers—input, hidden, and output. This system could learn
mary of the outstanding problems and directions for future work. to connect or associate a given input to a random output. Widrow and Hoff
(1960) developed a different system called ADALINE (Adaptive Linear
Element), which was an analogue electronic device made up of simple com-
11.1 INTRODUCTION
ponents. It employed the least-mean-squares learning rule.
Models in hydrology vary from simple black-box to complicated distributed The interest in artificial neural networks (ANNs) began to fade away in late
physics-based. Black-box models relate input to output without concentrating 1960s and early 1970s. Nevertheless, several researchers continued working
on the description of underlying physical processes. The use of time-series on developing neuro-morphically based computational methods for pattern
methods may be, on the other hand, complicated by nonstationarity and recognition. During that period, several paradigms, which continue to be
nonlinearity in the data, requiring experience and expertise from a modeler enhanced, were generated. Amari (1967) established a mathematical theory
(Imrie et al., 2000). The physics-based models that strive to account for for a learning basis (error-correction method) dealing with adaptive pattern
physical processes provide more accurate results at the expense of greater data classification. Klopf (1972) developed a basis for learning through artificial
requirements and use of sophisticated numerical techniques. Although these neurons based on a biological principle for neuronal learning called heterosta-
models are complicated with many parameters, the information that they are sis. Werbos (1974) developed the back-propagation learning method.
required to provide may be simple, such as mean annual groundwater Fukushima (1975) developed a step-wise trained multilayered neural net-
recharge rate over part of the basin or 100-year flood. Singh and Woolhiser work, called cognition, for interpretation of hand written characters. Carpenter
(2002) provided a historical perspective and challenges of hydrologic model- and Grossberg (1988) developed the Adaptive Resonance Theory networks
ing. based on biologically plausible models and influenced the exploration of
One of the recent developments in hydrologic modeling is the development resonating algorithms.
of soft computing techniques, such as artificial neural networks (ANNs), Progress during the 1970s and 1980s led to the renewed interest in ANNs.
Fuzzy logic (FL), and genetic algorithms (GAs), originally developed in the The first applications of ANN in hydrology started in the early 1990s to model
area of control engineering. The ANNs were developed in the early 1950s, the rainfall intensity fields, aquifer parameter values, streamflow, and rainfall-
FL in the 1960s, and the GAs in the 1970s and are now widely used in many runoff (French et al., 1992; Aziz and Wong, 1992; Kang et al., 1993; Halff et al.,
disciplines, including hydrology. These techniques are effective in handling 1993). ASCE (2000a) summarizes applications of ANNs to a wide range of
dynamic, nonlinear, and noisy data, especially when the underlying relation- hydrologic problems from the early 1990s to 2000. Applications of ANN in
ships are not fully understood (See and Openshaw, 1999). The reason for an hydrology have continued, encompassing sediment transport, dispersion,
increasing interest in these techniques is that they are flexible, self-adaptive evapotranspiration, seepage, snowmelt, river basin classification, infiltration,
and less assumption-dependent, and intrinsically general, and perform well in soil water retention, and event-based rainfall runoff. These and more applica-
most applications where other models either tend to fail or become cumber- tions can be found in Abrahart et al. (2004), Govindaraju and Rao (2010),
some (Nayak et al., 2005; Nourani et al., 2014). This chapter, focusing on Kingston et al. (2010), and Tayfur (2012).
11-3
11_Singh_ch11_p11.1-11.6.indd 3 8/22/16 11:38 AM

11-4 Artificial Neural Networks
11.3 ARTIFICIAL NEURAL NETWORKS where zi is the normalized value, xi is the corresponding value in the input set,
xmin is the minimum value of the input set, and xmax is the maximum value of
Artificial Neural Networks (ANNs) have been inspired from information
the input set.
processing of the biological nervous system, such as the brain. ANN is a mas-
There are several network training algorithms such as Radial Basis
sively parallel-distributed processor made up of simple processing units
Function, Conjugate Gradient, Cascade Correlation, and Generalized
(Fig. 11.1). It resembles human brain in two aspects: (1) it learns from the
Regression that are summarized in ASCE (2000b) and Tayfur (2012).
experiments and experience; and (2) it stores knowledge through synaptic
However, the commonly employed network training algorithm in hydrology
weights. In ANN: (1) information processing occurs in many neurons; (2)
is the back-propagation algorithm. The objective of the algorithm is to deter-
signals are passed between neurons through connection links; (3) each con-
mine optimal weights that would generate an output vector Y= (y1, y2, ..., yp)
nection link has an associated weight that represents its connection strength;
as close as possible to the target values of the output vector T = (t1, t2,...,tp)
and (4) each neuron typically applies an activation function to its net input to
with the selected accuracy. The optimal weights are found by minimizing a
determine its output signal.
predetermined error function (E) of the following form (ASCE, 2000b):
In hydrologic applications a three-layer feedforward ANN, as shown in Fig.
11.1, is generally considered. In the feedforward ANN, input quantities (xi)
are fed into the input layer neurons that, in turn, pass them on to the E=∑ ∑( yi - ti )2 (11.6)
P p
hidden layer neurons (zi) after multiplication by connection weights (vij). A
hidden layer neuron adds up the weighted input received from each input where yi is the component of an ANN output vector Y; ti is the component of
neuron (xivij) and associates it with a bias (bj). The result (netj) is then passed a target output vector T; p is the number of output neurons; and P is the
on through a nonlinear transfer function, such as sigmoid or hyperbolic tan- number of training patterns.
gent function, to produce an output. The net information is expressed as In the back propagation algorithm, the effect of input is first passed for-
(Dawson and Wilby, 1998): ward through the network to reach the output layer. After the error is com-
puted, it is then propagated back toward the input layer with modified
netj = ∑ xivij − bj (11.1) weights. The gradient-descent method, along with the chain rule of differen-
tiation, is employed to modify the network weights as (Imrie et al., 2000;
The activation functions are bounded, monotonically increasing and have ASCE, 2000b):
continuous derivatives. Equations (12.2) and (12.3) that stand for the sigmoid
and tangent hyperbolic activations functions can, respectively, be defined as ∂ E(wij ) (11.7)
follows (Imrie et al., 2000):
wijnew = wijold − η
∂wij
1
f(net j ) = − net (11.2) where η is the learning rate. For Fig. 11.1, the derivative term in Eq. (11.7) can
1+ e j
be derived using the chain rule and expressed by Eqs. (11.8) and (11.9) for the
2 sigmoid and tangent hyperbolic functions, respectively, as (Tayfur, 2012):
f (net j ) = −2net − 1
(11.3) p q
1+ e j ∂E

∂wij
= ∑ ∑(tkj − z kj )(−1) f (wij )[1 − f (wij )]y j (11.8)
k =1 j =1
11.4 ANN TRAINING AND TESTING
p q
Before initiating network training, random numbers are first assigned to con- ∂E
nection weights. There are no strict rules, although the range for the random
∂wij
= ∑ ∑(tkj − z kj )(−1)[1 − f (wij )][1 + f (wij )]y j (11.9)
k =1 j =1
values is generally set between –3 and +3. The practical rule is to set the range
in (–3/n; 3/n), where n is the number of neurons in the input layer. Next, Equations (11.8) and (11.9) are used when the connection weights between
input–output data used for the network training are normalized, depending hidden and output layers are updated. Equation (11.7) for the sigmoid and
upon the chosen activation function. Generally, data are compressed into [0.1, hyperbolic tangent functions, when updating the connection weights between
0.9] for the sigmoid function and into [–0.9, 0.9] for the tangent hyperbolic input and hidden layers, can be expressed by Eqs. (11.10) and (11.11), respec-
function, respectively, as (Tayfur, 2012): tively, as follows (Tayfur, 2012):
0.8 ( xi − x min ) (11.4) p q

zi = 0.1 + vijnew = vijold − η∑ ∑(tkj − z kj )(−1) f (wij )[1 − f (wij )]y j [1 − y j ]wij xi (11.10)
x max − x min k =1 j =1
p q
vijnew = vijold − η∑ ∑(tkj − z kj )(−1)[1 + f (wij )][1 − f (wij )][1 + y j ][1 − y j ]wij xi (11.11)
1.8 ( xi − x min ) 
zi =   − 0.9 (11.5) k =1 j =1
 x max − x min  Upon training where input–output pairs are fed into the network and
model tries to minimize the error function by doing iterations. During train-
ing, as the number of iterations increase, the error is supposed to decrease as
shown in Fig. 11.2. In order to see whether the training of a network is suc-
Bias Bias cessfully accomplished, one can feed the network with the same training input
node node pairs to obtain model outputs. In this procedure, no more target outputs are
fed into the network. The model outputs can then be plotted against mea-
bl bk sured actual data to visualize the model performance. One can further fit a
vl1 wk1 regression line to the scattered data and also obtain the equation of the fitting
line and the correlation coefficient (R2). If slope of the fitting line (m) is close
v11 w11 to one and the intercept (b) is close to zero and R2 is high, then training can
x1 z1 y1
be said to be successful.
Once training is successfully accomplished, one can then confidently apply
x2 z2 the ANN to the test data. In this case, the user can enter any test data into the
y2 network to produce model output. In order to evaluate the performance, the
produced output is first de-normalized. For example, if Eq. (11.5) was to be
employed for the compaction of data before training, then Eq. (11.12) can be
vnm wmp used to decompress the data (Tayfur, 2012):
xn zm yp
[1.8x min + (z
Input Input Hidden Output Output yi = i + 0.9)(xmax − x min )] (11.12)
vector layer layer layer vector 1.8
Figure 11.1 Schematic representation of a three-layer artificial neural network. where yi stands for de-normalized value.
11_Singh_ch11_p11.1-11.6.indd 4 8/22/16 11:38 AM

Shortcomings in ANN Modeling 11-5
E
E
Error for
cross-validation set
Stopping stage
Error for training set

Epoch
Figure 11.2 Desired error variations with respect to number of iterations.
Epoch
Figure 11.4 Schematic representation for trend of error during overtraining.
Several ways can be used to assess the performance of the model: (1) plot- training set, but poorly for the test data. Cross-validation is employed to pre-
ting of measured data versus model output data using a line graph to visually vent overtraining. For this purpose, the original data set is portioned into
see the trend; (2) plotting of measured data versus model output data using a three groups: training set, cross-validation set, and testing set. The cross-val-
scatter plot and fitting a regression line and obtaining R2; (3) computing error idation set is used as a check for the network to terminate the iterations, when
measures, such as mean absolute error, root mean square error, and mean the error for the training set continues to decrease while the error for the
absolute relative error; and (4) computing the coefficient of efficiency (E) that cross-training set begins to increase (Fig. 11.4). If the available data set is too
accounts for model errors in estimating mean or variance of observed data small for portioning, the simplest way to prevent overtraining is to stop train-
sets. Although E ranges from minus infinity (poor model) to 1.0 (perfect ing when the error does not decrease significantly.
model) [see Eq. (11.13)], its value is supposed to be close to one for perfect For satisfactory input–output mapping, the quality and quantity of data is
modeling (Moramarco et al., 2005). crucial. Tayfur and Guldal (2006) pointed out that cross-correlation can shed
light on the number of input variables and the quality of data. Furthermore,
N
they showed that training and testing data sets should have comparable values
∑ (Omi − Opi )2
E = 1.0 − iN=1 (11.13) for the mean and range parameters. The quantity of data is also important.
They pointed out that at least 90 days of time series data are required for
∑ (Omi − Om )2 satisfactory prediction of TSS from precipitation.
i =1
During iterations, it is possible to get trapped at a local minimum rather
where N is number of observations, Om is the measured data, Op is the than the global minimum. At global minimum, the optimal connection
predicted data, O m is the mean of measured data, O p is the mean of the strengths are obtained resulting in desired solution. To overcome local mini-
predicted data. (5) Computing the discrepancy ratio (DR) for discrete data mum, one or more of the followings can be employed: (1) learning parameter
predictions, such as predicting dispersion coefficient for different rivers, from (δ) can be dynamically adjusted. (2) Network can be initialized with different
which an accuracy of a model can be obtained (Tayfur, 2012). DR is expressed set of initial weights. (3) Different network structures can be employed, such
by Eq. (11.14) and if DR = 0.0, then there is an exact-prediction. Otherwise, as double hidden layer. (4) Momentum factor (α), shown by Eq. (11.15), can
there is either an over-prediction (DR > 0.0) or under-prediction (DR < 0.0) be used (ASCE, 2000b):
(Seo and Cheong, 1998).
∂E
 Op  ∆vij (n ) = −δ + α ∆vij (n − 1) (11.15)
DR = log  (11.14) ∂vij
 Om 
The accuracy of a model can be found by number of DR values between where ∆vij (n ) and ∆vij (n − 1) are the weight increments between nodes i
–0.3 and 0.3, relative to the total number of data values (Seo and Cheong, and j during the nth and (n–1)th pass (or epoch), respectively; δ is learning
1998). For better accuracy, a narrower range, such as 0.18 and –0.18, can be rate; and α is momentum factor.
set for the DR values. Although ANN has a very successful interpolation capability, it lacks the
extrapolation ability for the cases for which it is not trained (Tayfur et al.,
11.5 DRAWBACKS OF ANN TECHNIQUE 2007). Imrie et al. (2000) developed a modified cascade-correlation training
algorithm to improve extrapolation capability of ANNs. Giustolisi and
The ANN technique has a number of limitations, including: (1) over-training, Laucelli (2005) tried an avoiding overfitting technique to increase the gener-
(2) quantity and quality of data, (3) local minimum, and (4) extrapolation. alization performance of ANN.
Overtraining is the case where the network parameters are too fined-tuned to
the training data set such that the network starts trying to fit the noise com-
11.6 SHORTCOMINGS IN ANN MODELING
ponent of the data set thus memorizing individual examples rather than the
trend (Fig. 11.3). When this happens, the network performs very well over the There are several minor drawbacks in ANN modeling and these are as fol-
lows:
1. The selection of the number of hidden layers and neurons in the hidden
Predicted
layers is quite arbitrary. A single hidden layer is most often sufficient in

hydrological applications. Nevertheless, when there are an excess number of
input variables with more than a single output variable, then two hidden lay-
ers can be employed. The number of neurons in the hidden layer can be found
Trend by the trial and error procedure. The number of neurons in the hidden layer,
at least at the beginning when forming the network, can be decided as (Tayfur,
2012):
Overtrain
N h ≤ 2N i + 1 (11.16)
where Nh is the number of neurons in the hidden layer, and Ni is the num-
Measured ber of neurons in the input layer.
2. The data set needs to be sufficiently large so that a user can group the
Figure 11.3 Schematic representation for overtraining. data separately for training and testing. Another set may be desirable for
11_Singh_ch11_p11.1-11.6.indd 5 8/22/16 11:38 AM

11-6 Artificial Neural Networks
cross-validation in order to monitor over-training. There is no rule on what Giustolisi, O. and D. Laucelli, “Improving generalization of artificial neural
the minimum number of data sets would be. networks in rainfall-runoff modeling,” Hydrological Sciences Journal, 50 (3):
3. The assignment of values for the learning rate and momentum factor is 439–457, 2005.
based on the experience of the user. These parameters can have values in the Halff, A. H., H. M. Halff, and M. Azmoodeh, “Predicting runoff from rain-
[0, 1] range. Assigning very small values within this range may cause a slug- fall using neural networks,” Journal of Hydraulic Engineering, 119 (7): 760–
gish convergence, while large values can destabilize the network training. 765, 1993.
Tayfur (2012) has pointed that larger values can be assigned when the number Hebb, G., The Organization of Behavior, Wiley, New York, 1949.
of neurons in the hidden layer is fewer. Imrie, C. E., S. Durucan, and A. Korre, “River flow prediction using artifi-
4. The connection weights of ANN do not correspond to any system cial neural networks: generalization beyond the calibration range,” Journal of
parameters. Unlike individual weights, the relative input contributions do Hydrology, 233: 138–153, 2000.
have physical meaning and therefore the use of such measures would facilitate Kang, K. W., J. H. Kim, J. C. Park, and K. J. Han, “Evaluation of hydrological
the assessment of a trained ANN and enable the constraints to be applied to forecasting system based on neural network model.” Proceedings of the 25th
the objective function to direct the search for the optimum weight vector into IAHR Congress, 257–264, Delft, The Netherlands, 1993.
physically feasible regions of the search space. Kingston, G., H. Maier, and M. Lambert, Bayesian Artificial Neural Net-
works: With Applications in Water Resources Engineering, VDM, Verlag, 2010,
ISBN-13: 3639223248.
11.7 FUTURE DIRECTION
Klopf, A. H., Brain Function and Adaptive Systems—A Heterostatic Theory,
Despite many major and minor shortcomings, ANN an easy, user-friendly, Air Force Cambridge Research Laboratories, Research Report, Hanscom Air
data-driven, and practical modeling tool, has significant potential in hydrol- Force Base, Maryland, 1972.
ogy. Exploring a physical concept through a fully black-box model of ANN McCulloch, W. S. and W. A. Pitts, “A logical calculus of the ideas immanent
has been criticized by many researchers. However, there have been some in nervous activity,” Bulletin of Mathematical Biophysics, 5: 115–133, 1943.
efforts to mitigate this problem in recent years. Zhang and Govindaraju Moramarco, M., S. Barbetta, F. Melone, and V. P. Singh, “Relating local
(2003) and Nourani et al. (2012) developed ANNs that can explicitly account stage and remote discharge with significant lateral inflow,” Journal of Hydro-
within its architecture for the geomorphologic characteristics of a watershed. logic Engineering, 10 (1): 58–69, 2005.
Chen et al. (2014) recently suggested using copula-entropy method to iden- Nayak, P. C., K. P. Sudheer, D. M. Rangan, and K. S. Ramasastri, “Short-
tify input variables for an ANN model. Substantial research can be diverted to term flood forecasting with a neurofuzzy model,” Water Resources Research,
make ANN physically based and thus overcome its major drawbacks. 41, W04004, doi: 10.1029/2004WR003562, 2005.
Rochester, N., J. H. Holland, L. H. Haibt, and W. L. Duda, “Tests on a cell
assembly theory of the action of the brain, using a large digital computer,” IRE
REFERENCES
Transactions on Information Theory, IT-2: 80–93, 1956.
Abrahart, R., P. E. Kneale, and L. M. See (Eds.), Neural Network for Hydro- Nourani, V., O. Kalantari, and A. H. Baghanam, “Two semi-distributed
logical Modelling, CRC Press, Taylor & Francis Group, 2004, ISBN-13: ANN-based models for estimation of suspended sediment load,” Journal of
9058096197, London, UK. Hydrologic Engineering, 17 (12): 1368–1380, 2012.
Amari, S. “Theory of adaptive pattern classifiers,” IEEE Transactions, Nourani, V., A. H. Baghanam, J. Adamowski, and O. Kisi, “Applications of
EC-16: 299–307, 1967. hybrid wavelet-artificial intelligence models in hydrology. A review,” Journal
ASCE, “Artificial neural networks in hydrology. II: Hydrologic applica- of Hydrology, 514 (6): 358–377, 2014.
tions,” Journal of Hydrologic Engineering, 5 (2): 124–137, 2000a. Rosenblatt, F. The Perceptron: A Probabilistic Model for Information Storage
ASCE, “Artificial neural networks in hydrology. I: Preliminary concepts,” and Organization in the Brain, Cornell Aeronautical Laboratory, New York,
Journal of Hydrologic Engineering, 5 (2): 115–123, 2000b. Psychological Review, Vol. 65, No. 6, 1958, pp. 386–408.
Aziz, A. R. R. and K. F. V. Wong, “Neural network approach to the deter- See, L. and S. Openshaw, “Applying soft computing approaches to river
mination aquifer parameters,” Ground Water, 30 (2): 164–166, 1992. level forecasting,” Hydrological Sciences Journal, 44 (5): 763–778, 1999.
Carpenter, G. A. and S. Grossberg, “The art of adaptive pattern recognition Seo, I. W. and T. S. Cheong, “Predicting longitudinal dispersion coefficient
by a self-organizing neural network,” Computer, 21: 77–88, 1988. in natural streams,” Journal of Hydraulic Engineering, ASCE, 124 (1): 25–32,
Chen, L., L. Ye, V. P. Singh, J. Zhou, and S. Guo, “Determination of input 1998.
for artificial neural networks for flood forecasting using copula entropy Singh, V. P. and D. A. Woolhiser, “Mathematical modeling of watershed
method,” Journal of Hydrologic Engineering, 19 (11), 10.1061/(ASCE)HE.1943- hydrology,” Journal of Hydrologic Engineering, ASCE, 7 (4): 270–292, 2002.
5584.0000932, 04014021. 2014. Tayfur, G. and V. Guldal, “Artificial neural networks for estimating daily
Dawson, C. W. and R. Wilby, “An artificial neural network approach to total suspended sediment in natural streams,” Nordic Hydrology, 37 (1):
rainfall-runoff modeling,” Hydrological Sciences Journal, 43 (1): 47–66, 1998. 69–79, 2006.
Farley, B. G. and W. A. Clark, “Simulation of self-organizing systems by Tayfur, G., T. Moramarco, and V. P. Singh, “Predicting and forecasting flow
digital computer,” IRE Transactions on Information Theory, IT-4: 76–84, 1954. discharge at sites receiving significant lateral inflow,” Hydrological Processes,
French, M. N., W. F. Krajewski, and R. R. Cuykendall, “Rainfall forecasting 21: 1848–1859, 2007.
in space and time using a neural network,” Journal of Hydrology, 137: 1–31, Tayfur, G. Soft Computing in Water Resources Engineering—Artificial Neural
1992. Networks, Fuzzy Logic, Genetic Algorithm. WIT Press, Southampton, 2012.
Fukushima, K., “Cognitron: a self-organizing multilayered neural net- Werbos, P., Beyond regression: new tools for prediction and analysis in behav-
work,” Biological Cybernetics 20: 121–136, 1975. ioral sciences, PhD Dissertation, Harvard University, Cambridge, MA, 1974.
Govindaraju, R. S. and A. R. Rao (Eds.), Artificial Neural Networks in Widrow, B. and M. E. Hoff, “Adaptive switching circuits,” Proceedings of the
Hydrology, Springer, 2010, ISBN-13: 978-9048154210. Netherlands. Wescon Convention Record, Part 4, 1960, pp. 96–140.
11_Singh_ch11_p11.1-11.6.indd 6 8/22/16 11:38 AM

Chapter 12
Fuzzy Logic
BY
ISTVAN BOGARDI
ABSTRACT
1
Membership
Fuzzy logic is able to encode imprecision, vagueness, and uncertainties com- Excellent quality
mon in applied hydrology by using the concept of the membership function
that defines partial membership in a set. Two main domains of fuzzy logic are
considered: function of fuzzy numbers and fuzzy rule-based modeling
(FRM). Methodologies of fuzzy number functions consider the vertex meth-
od, the extension principle and fuzzy numerical simulation. Fuzzy set applica-
tion examples cover water supply risk, surface and groundwater flow and
transport, regression analysis, geostatistics, classification, and decision mak-
ing. FRM calculates output of a process from a given input using fuzzy logic,
specifically fuzzy rules. FRM shows much potential in cases when a causal
Poor quality
relationship is well established but difficult to calculate under real-life condi-
tions, when data are scarce and imprecise, or when a given input vector has
several contradictory responses which may be true to varying degrees. It can 0
23 31
involve qualitative information and expert’s experience. Application examples Springflow (m3/min)
of FRM cover teleconnection, the hydrological cycle, rainfall-runoff, stream-
flow, water quality, reservoir operation, and classification. A supplement to Figure 12.1 Membership function of a quality karstic thermal spring (Bogardi et al.,
the chapter contains basic definitions and operations, methodological details, 1983).
and a step-by-step FRM application example.
12.1 FUZZY LOGIC BASICS

subjective description of a person’s uncertainty.” Clearly, aleatoric uncertain-
ties may be preferably encoded by probabilities while epistemic uncertainties
Mining operation in a karstic region may harm quality thermal springs. by fuzzy sets. This approach bridges a historical gap between probability
However, there is no consensus about the definition of “quality” represented theory and fuzzy set theory (Ross et al., 2009). The simplest method of con-
by the mean annual flow. Experts agree on a lower limit below which a spring sidering imprecision is to perform an interval analysis. An uncertain param-
is out of the quality set. Also, an upper limit is accepted indicating excellent eter can take on any value within such an interval. With more information on
quality. Between these limits spring quality gradually increases. How is it pos- the uncertain parameter, the interval model can be sharpened; that is, we
sible to define this set of quality spring? Fuzzy sets or fuzzy logic (Zadeh, determine the possibility that the parameter can take on certain value(s)
1965) is capable of handling situations where membership in a set cannot be within the interval. To utilize additional information on an uncertain param-
defined on a yes/no basis; in other words, the boundaries of sets are vague. eter, multiintervals can be defined. As Fig. 12.2 shows, the first bit of addi-
The membership function is a central concept in fuzzy set theory. The value tional information results in another interval estimate [c, d] inside the first
of the membership function of an element in an ordinary set is either 0 or 1; interval [a, b]. The possibility, or degree of membership, of this second inter-
that is, it is not or it is a member of the set. However, an element may be a val [c, d] corresponds (in this example) to 0.6. Thus, the two intervals [a, b]
member of a fuzzy set to some degree, so its membership value can be and [c, d] represent the uncertain parameter with the two ‘‘levels of member-
between 0 and 1 (Fig. 12.1). Formal definition of the membership function as ship” 0.6 and 1. Accordingly, fuzzy logic can be considered as an extension of
well as basic operations including union and intersection on fuzzy sets are interval analysis. A fuzzy subset M is called a fuzzy number if M is a normal
given in the Supplement. convex fuzzy subset of the set of real numbers. The membership value of a
Applied hydrology commonly involves a system of concepts, principles, real number reflects the ‘‘likeliness” of the occurrence of that number; the
and methods for dealing with modes of reasoning that are approximate rather level sets (intervals in this case) reflect different sets of numbers with a given
than exact. In other words, hydrology is hampered by uncertainties caused by minimum likeliness. The simplest type of fuzzy number is triangular, that is,
nature (e.g., climate), limited data, and imprecise modeling. Probability and linear on either side of the peak. Figure 12.2 gives an example of a triangular
statistics, as the traditional tool to deal with such situation, has limitations as fuzzy number (TFN). The arithmetic of intervals (Supplement) can be used
Bruno de Finetti (1974) proclaims, ‘‘Probability does not exist; it is a to add, subtract, multiply, or divide fuzzy numbers.
12-1
12_Singh_ch12_p12.1-12.6.indd 1 8/22/16 11:39 AM

12-2 Fuzzy Logic
m(x) m
1 1
D (–) S
D S
0.6 c d
0
a = 0.6 level set –3 0 3 5 6 9 11 15
Shortage : D (–) S > 0
D (–) S = [d1–s2, d2–s1] = [3–15, 6–9, 11–5] = [–12, –3, 6] · 105 (m3/year)
Width Figure 12.3 Difference between demand and supply.
0
a=4 5 b=7
Transmissivity (10–2 m2/s) are also uncertain. Probabilistic risk analysis commonly uses frequency-based
Figure 12.2 Membership function of transmissivity “more than 4, less than 7,” about
statistical information; however, this is rarely available in climate change situ-
5·10–2 m2/s. ations. As an alternative, nonprobabilistic risk analysis is used that is rela-
tively simple, practical, and applicable with available data/information.
Uncertainty in the three elements—supply, demand, and consequences—is
The assessment of membership function has been a great challenge of fuzzy expressed by TFNs with a minimum, a maximum, and a plausible value
logic applications. Like in statistics (e.g., selection and fitting distributions), (CC-WaterS, 2012). A numerical example illustrates the methodology:
there is lack of consensus on the preferred assessment methodology. Two Demand: (3,6,11)· 105 m3/year, and Supply: (5,9,15)· 105 m3/year. Water short-
main groups of approaches are described in details with examples by Ross age (demand—supply) is calculated as the difference between two fuzzy
(2010): (1) Most common is manual mapping of knowledge elicited from numbers (Fig. 12.3).
experts by intuition, inference or, rank ordering. For example, simply agree- If shortage is D (–) S ≥ 0, it is not a fuzzy number in this example since its
ing on a lower and upper limit and a plausible value, a TFN can be obtained. max ≠ 1. Normalize to obtain shortage as a fuzzy number (Fig. 12.4),
(2) Automatic generation of membership functions using empirical data with Shortage: (0,0,6) ·105 m3/year. In other cases, if the shortage D (–) S may have
the help of neural networks, genetic algorithms, or entropy minimization a max = 1, then normalization is not necessary.
principle. These methods are promising but need considerable data set and Water supply risk is obtained by multiplying shortage and consequences (as
computational effort. specific economic loss) in Fig. 12.5. Shortage: [0,0,6] ·105 m3/year; Specific
From the early application of fuzzy logic to hydrology (Bogardi et al., Loss: [1,2,4] · $/102 m3;
1983), a large amount of research has been pursued, and, at present, fuzzy Risk: [0,0,6] · [1,2,4] = [0,0,24]· 103 $/year.
logic has become a practical tool in hydrologic analysis and water resources Water quality insufficiencies (Ganoulis, 2008) are also accounted for in a
decision making (Ross, 2010). In this chapter, two main areas of hydrological similar matter, and multicriterion decision making is performed in a fuzzy
applications describe the methodology and provide case examples. logic framework to obtain a composite risk as a fuzzy number (CC-WaterS,
m
12.2 FUNCTION OF FUZZY NUMBERS
1.0
12.2.1 Methodology
Three common methods for computing functions y = f (x1, x2,…,xj) of j fuzzy
variables include (1) discretize membership values with level sets, this is the Shortage as a fuzzy number: [0, 0, 6]
vertex method (Dong and Shah, 1987); (2) directly calculate membership
function value µy for various x values, using the extension principle (Zadeh, D (–) S 0.5
1965); (3) fuzzy numerical simulation (Chanas and Nowakowski, 1988). The
Nonnormal fuzzy set of shortage
simple and most practical vertex method is based on the α-level set concept
(Fig. 12.2). Since each α-level set corresponds to an interval number, we can
use interval number operations according to the given function in order to
obtain the interval number of the function at that level. To repeat this opera- 0.0
tion with several α-levels, the membership function can be constructed –3 0 6
(Supplement). According to the extension principle, given membership func- Figure 12.4 Shortage as a fuzzy number.
tions µj(xj), the membership function µy can be calculated by the optimization
problem (Supplement):
µ(y) = sup{min(µ1(x1), …, µj(xj)} if y = f (x1, x2, …, xj) m
1.0
= 0 otherwise
Fuzzy numerical simulation generates possible values of membership func-
tions and these realizations are used with the given function (Supplement).
Specific
12.2.2 Application Examples 0.5 loss
Water supply risk
Water Supply Risk Analysis under Climate
Change Shortage as a
Risk, in general, includes the combination of the uncertain adverse events and fuzzy number
the consequence of these events. Here, adverse events correspond to water 0.0
shortages and water quality insufficiencies. Due to the inherently uncertain
climate change estimation and future socioeconomic conditions, all elements 0 12 4 6 24
of risk (water supply, water demands, and consequences of water shortages) Figure 12.5 Calculation of water supply risk.
12_Singh_ch12_p12.1-12.6.indd 2 8/22/16 11:39 AM

Fuzzy Rule-Based Modeling (Fuzzy Inference) 12-3
2012). Finally, risk-management actions are ranked for waterworks under the related to the distance of that point to the cluster centers, and range from 0
pressure of climate change. (out of a cluster) to 1 (full member of a cluster). Hard and fuzzy c-means
methods use similar objective functions to minimize inner distances, but
Groundwater Flow and Transport FCM also adds the respective membership values as kind of weights. In an
interesting applications, typhoon tracks over the western North Pacific have
Numerical simulation of groundwater flow under steady state and transient been identified both by hard c-means and FCM (Kim et al., 2011). In this
conditions involves the prediction of hydraulic head in space and time. In the case, the results by the FCM are more useful in isolating influences of the
modeling process, parameter uncertainty is one of the major barriers to reli- physical phenomena such as El Niño and La Niña.
able prediction of groundwater system response. Stochastic modeling tech-
niques have been used extensively for consideration of uncertainty that stems
from few measurement data in the presence of spatial variability. Alternatively, Fuzzy Decision Making
groundwater flow and transport modeling can be performed by the help of Objectives, constraints, and parameters of water resources decision models
fuzzy-set theory. Finite difference method may be used to approximate the are often imprecisely known, and therefore amenable to fuzzy-logic formula-
governing equation, in which aquifer parameters such as transmissivity are tion. In case of selection among alternatives, the objectives are calculated as
considered to be fuzzy numbers. If variables in the system of equations are fuzzy numbers. Several methods are available for ranking objectives as fuzzy
fuzzy numbers instead of crisp values, then the dependent variable (e.g., numbers (Ross, 2010), but mostly their centroids are used for ranking
hydraulic head) is also fuzzy. Using the vertex method or fuzzy numerical (Supplement). Both single- and multiobjective rankings have been widely
simulation, a nonlinear optimization algorithm can solve the linear system of applied in practice (Vujetic and Simonovic, 2011; CC-WaterS, 2012). Linear
interval equations (Dou et al., 1997). With the vertex method at each α-level programming tools can be used with fuzzy linear objectives and/or con-
cut, the hull of the solution of hydraulic heads is searched to find the end- straints, leading to compromise solutions (Chang and Ayyub, 2001). Classical
points of the head intervals which bound the possible head values for each stochastic dynamic programming of reservoir operation can be enhanced by
node in the finite difference grid. considering multiple objectives (e.g., power generation, flood control, irriga-
tion) formulated as fuzzy sets. Optimal release policy is calculated by maxi-
Surface Water mizing the minimum expected membership values of the fuzzy objectives
(Chandramouli and Nanduri, 2011).
Well-known uncertainties have been described by fuzzy sets, including, Monte-Carlo methods can also be used for solving fuzzy optimization
among others, river discharge and stage (Shrestha and Simonovic, 2010), problems (Buckley and Jowers, 2008).
runoff coefficient (Sen and Altunkaynak, 2005), unit hydrograph (Bardossy et
al., 2006), the effect of precipitation uncertainty (Maskey, 2004), or stream
pollution (Mpimpas et al., 2008).
12.3 FUZZY RULE-BASED MODELING
(FUZZY INFERENCE)
Fuzzy Regression
12.3.1 Methodology of Fuzzy Rule-Based Modeling
Fuzzy regression may be used whenever a relationship between variables is
imprecise and/or data are inaccurate and/or sample size is insufficient as is Fuzzy rule-based modeling (FRM) calculates output of a process or system
the case in many hydrological applications (Bardossy et al., 1990). Fuzzy from a given input using fuzzy logic, specifically fuzzy rules. FRM shows
regression expresses regression parameters as fuzzy numbers that leads to the much potential in cases when a causal relationship is well established but dif-
dependent variable, y also as fuzzy number. The purpose is to find a regres- ficult to calculate under real-life conditions, when data are scarce and impre-
sion model that minimizes a measure of vagueness of y, for example the area cise, or when a given input vector has several contradictory responses which
under the membership function. This measure of vagueness replaces the least may be true to varying degrees. It can involve qualitative information and
squares criterion of statistical regression. The minimization of vagueness expert’s experience. These features are often present in applied hydrology. In
leads to a linear programming problem in linear regression and a nonlinear direct contrast to neural networks, which take a training set and use black,
problem in nonlinear regression. As an example, transmissivity and storage impenetrable models, FRM is highly visible and capable of utilizing expert’s
coefficient for an aquifer using imprecise pump-test data can be calculated experience. The primary difference between fuzzy control and FRM is that
with a modified fuzzy least-squares regression method (Mathon et al., 2008). fuzzy control applies feedback, whereas FRM simulates processes usually
This allows for analysis of the uncertainty that comes from a lack of knowl- without a feedback mechanism. The methodology of FRM to model hydro-
edge about aquifer heterogeneity. logical processes consists of the following steps (Bardossy and Duckstein,
1995) (more details can be found in the Supplement):
1. Selection and representation of input (premises) and output (responses)
Fuzzy Geostatistics variables by fuzzy numbers. As an example consider the teleconnection
Often in applied geostatistics (see Chap. 20) limited amount of exact measure- between atmospheric circulation patterns, CP, and the ENSO phenomena
ment data at exact locations is available. On the other hand, different type and (represented by the Southern Oscillation Index, SOI) as premises and a com-
quality of additional inexact data, regional hydrogeological, or expert infor- mon drought index, PDSI as response. The entire range of possible premise
mation may be available. Fuzzy set methods can handle inexact data and and response values is divided into several overlapping classes each forming
locations as well as the combination of exact and inexact data (Bardossy et al., a fuzzy number, namely for monthly CP occurrence: very rare, rare,
1988; Diamond, 1989). Kriged values and the estimation variance are medium, frequent, and very frequent (Fig. 12.6a), for SOI: strong El Niño,
expressed as fuzzy numbers. As an example, 20 direct permeability measure- weak El Niño, Neutral, weak La Niña, and strong La Niña phases (Fig. 12.6b)
ments and 20 inexact data stemming from fuzzy regression between geoelec- and for PDSI: 18 classes from extremely dry to extremely wet conditions
trical measurement and permeability have been used for obtaining a (Fig. 12.6c).
permeability contour map. Fuzzy geostatistics is a useful tool for the estima- 2. Definition of the training and validation data sets. The entire data set of
tion of hydrogeological parameters required for groundwater modeling premises and response is split into two parts: a training set and a validation
(Piotrowski et al.,1996), or for designing soil-sampling grids when precise set. The training set will be used to learn the fuzzy rules and the validation set
information on spatial variability is unavailable (Lark, 2000). to validate these rules, namely, how correctly they can estimate the observed
response.
3. Learning of the fuzzy rules using available measurement information,
Fuzzy Classification (Clustering) expert knowledge, limited small-scale process modeling, or the combinations
C-means clustering is a common technique to group data sets that contain thereof. A typical fuzzy rule with K fuzzy premises and one fuzzy response
similar information into relatively homogeneous clusters by minimizing can be written as follows:
inner distances between data points and cluster centers. It is hard (crisp) IF (premise 1 is Ai1,1) AND (premise 2 is Ai2,2) AND … AND (premise K is
partitions because each data point must be assigned to a single cluster. In AiK,K) THEN (response is Bj).
hydrology, this requirement may be too restrictive since a data point may 4. Application of the fuzzy rules using the validation data set. The response
pertain to more than one cluster. For instance, to classify watersheds in real for each data point in the validation set is estimated by using the rules of the
world, many watersheds may pertain to several clusters with different system constructed in the calibration phase.
degrees (Basu and Srinivas, 2014). The widely used fuzzy c-means method 5. Evaluate the fuzzy rule-based model in terms of how well it reproduces
(FCM) assigns fuzzy membership values to the each data point which is the statistical properties of observed drought index in the validation set.
12_Singh_ch12_p12.1-12.6.indd 3 8/22/16 11:39 AM

12-4 Fuzzy Logic
(a) (b)
1 1
m m
Strong La Niña
Medium
Rare
Neutral
Strong EI Niño
Weak La Niña
Weak EI Niño
Very frequent
Very rare
Frequent
0 0
0 ¼·max ½·max ¾·max max –3 –2 –1 0 1 2 3
Monthly relative frequency of a given CP class SOI values
(c)
1
m
Dry 8
Dry 7
Dry 6
Dry 5
Dry 4
Dry 3
Dry 2
Dry 1
Extreme wet
Normal
Wet 1
Wet 2
Wet 3
Wet 4
Wet 5
Wet 6
Wet 7
Wet 8
Extreme dry
0
–9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9
Drought index values
Figure 12.6 Fuzzy numbers defined on premises: (a) CP, (b) SOI, and response, (c) Drought Index.
12.3.2 Application Examples of Fuzzy Rule-Based with measured temperature and ET. The results were compared to the calibra-
Modeling tion data set and calculation with the Blaney-Criddle method, and FRM
Teleconnection showed high efficiency in predicting ET. FRM was applied to the description
of infiltration and unsaturated flow (Bardossy, 1996). The fuzzy rules describ-
The relationship between CP types and ENSO as premises and local monthly
ing vertical water movement were formulated with the help of relative soil
precipitation in Arizona under climate change, as response, was analyzed by
moisture contents. Numerical solution of the Richards equation provided a
Galambosi et al. (2009). Monthly frequency distributions of CP types gener-
synthetic learning set, and the results were compared to measurement data.
ated from GCM data under 2 × CO2 and different ENSO scenarios were used
Although, rule-based models are less accurate than process modeling, in real-
to estimate monthly precipitation under climate change. FRM performed
life problems pure analytical modeling may give a false impression of accu-
better than regression analysis, probably due to its nonlinear approximation
racy. The gain in simplicity, computational speed, and flexibility may
capability.
compensate for the possible loss in accuracy.
Elements of the Hydrological Cycle
Rainfall-Runoff
Temperature as premise was considered to estimate potential evapotranspira-
tion (ET) as response in Greece (Tzimopoulos et al., 2008). FRM was trained Among many FRM to simulate event-based runoff, Tayfur and Singh (2006)
employed rainfall and infiltration rate, and time as premises to estimate
event-based discharge. The fuzzy rules were learned and validated using
measured data. The results compared to neural net and physically based
1 models indicated FRM advantages, namely transparency and consideration of
uncertainties in the fuzzy inference procedure through the IF-THEN state-
Station 7 ments.
Relative frequency
0.75
(SW-Hungary)
Streamflow
0.5 Lagged discharges on three upstream river branches served premises to fore-
cast flood hydrograph on a downstream river (Zhang et al., 2012). Historical
Estimated PDSI
0.25 data provided the learning and validation data set. The accuracy of FRM was
Observed PDSI better than the Muskingum–Cunge scheme. Using hydraulic flood routing
FRM can replace the momentum equation but retain the continuity equation
0 (Gopakumar and Mujumdar, 2009). The training set originated from
–8 –6 –4 –2 0 2 4 6 8
PDSI values observed discharge-area relationships. For heterogeneous sub-reaches, opti-
1
mized fuzzy rules were obtained by a neuro-fuzzy algorithm (Nurnberger
et al., 1999).
Relative frequency
Station 12
0.75 Water Quality
(SE-Hungary)
FRM and neural net (ANN) was used to estimate river sediment concentra-
0.5 tion from measured lagged discharge and stage data (Lohani et al., 2007).
Both FRM and ANN were found to be considerably better than the conven-
Estimated PDSI
0.25 tional rating curve method. However, FRM resulted in more accurate esti-
Observed PDSI mates than ANN for almost all river stages.
0 Reservoir Operation Control
–8 –6 –4 –2 0 2 4 6 8
PDSI values
Operation policy of a multipurpose reservoir can be obtained by FRM (Mehta
and Jain, 2009). Premises include reservoir level, inflow, demand, and flood
Figure 12.7 Empirical cumulative frequency distributions of the observed and storage to calculate release as response. Measurement data of premises and
estimated PDSI time series in summer (station 7: Pécs, and station 12: Szeged). actual releases were used to define the learning and validation sets. Both
12_Singh_ch12_p12.1-12.6.indd 4 8/22/16 11:39 AM

REFERENCES 12-5
Mamdani-type FRM and adaptive neuro-fuzzy interactive system (Ross, Ganoulis, J., Engineering Risk Analysis of Water Pollution: Probabilities and
2010) offered good results, but FRM proved more users friendly. Fuzzy Sets, John Wiley & Sons, 2008, p. 317. Somerset, NJ
Gopakumar, R. and P. P. Mujumdar, “A fuzzy dynamic flood routing model
FRM Classification for natural channels,” Hydrological Processes, 23: 1753–1767, 2009.
Four chemical indicators were used as premises to classify groundwater qual- Kim, H., J. Kim, C. Ho, and P. Chu, “Pattern classification of typhoon tracks
ity for irrigation purposes (Priya, 2013). The fuzzy rules allowed certain flex- using the fuzzy c-means clustering method,” Journal of Climate, 24: 488–508,
ibility because slightly different indicators may correspond to a given class. 2011.
The model was validated using ground water quality data and the results Lark, L. M., “Designing sampling grids from imprecise information on soil
improved a widely accepted standard classification system. FRM was applied variability, an approach based on the fuzzy kriging variance,” Geoderma,
to classify daily atmospheric circulation patterns (CP) for estimating coastal- 98: 35–59, 2000.
wave properties (Pringle et al., 2014). Rules were defined corresponding to Lohani, A. K., N. K. Goel, and K. K. S. Bhatia, “Deriving stage-discharge-
the geographical location of pressure anomalies. Degree of fulfillment (see sediment concentration relationships using fuzzy logic,” Hydrological Sciences
Supplement) were computed for each rule, and the rule with the highest DOF Journal, 52 (4): 793–807, 2007.
value was assigned to the CP class for that day. Maskey, S., V. Guinot, and R. K. Price, “Treatment of precipitation
uncertainty in rainfall runoff modeling: a fuzzy set approach,” Water
Resources Research, 27: 889–898, 2004.
ACKNOWLEDGMENT
Mathon, B. R., M. M. Ozbek, and G. F. Pinder, “Transmissivity and storage
This chapter is dedicated to the memory of my colleague and friend, Lucien coefficient estimation by coupling the Cooper–Jacob method and modified
Duckstein who pioneered the introduction and application of fuzzy logic in fuzzy least-squares regression,” Journal of Hydrology, 353 (3–4): 267–274,
hydrology. 2008.
Mehta, R. and S. K. Jain, “Optimal operation of a multi-purpose reservoir
using neuro-fuzzy technique,” Water Resources Management, 23: 509–529,
REFERENCES
2009.
Bardossy, A., “The use of fuzzy rules for the description of elements of the Mpimpas, H., P. Anagnostopulos, and J. Ganoulis, “Uncertainty of model
hydrological cycle,” Ecological Modelling, 85: 59–65, 1996. parameters in stream pollution using fuzzy arithmetic,” Journal of
Bardossy, A., I. Bogardi, and W. E. Kelly, “Imprecise (Fuzzy) information in Hydroinformatics, 10 (3): 189–200, 2008.
geostatistics,” Mathematical Geology, 1 (4): 287–311, 1988. Nurnberger, A., D. Nauck, and R. Kruse, “Neuro-fuzzy control based on the
Bárdossy, A., I. Bogardi, and L. Duckstein, “Fuzzy regression in hydrology,” NEFCON-model: recent developments,” Soft Computing, 2 (4): 168–182,
Water Resources Research, 26 (7): 1497–1508, 1990. 1999.
Bardossy, A. and L. Duckstein, Fuzzy Rule-Based Modelling with Piotrowski, J.A., F. Bartels, A. Salski, and G. Schmidt, Estimation of
Applications to Geophysical, Biological and Engineering Sciences, CRC Press, hydrogeological parameters for groundwater modelling with fuzzy geostatis-
Boca Raton, FL, 1995. tics: closer to nature?, Calibration and Reliability in Groundwater Modelling,
Bardossy, A., G. Mascellani, and M. Franchini, “Fuzzy unit hydrograph,” IAHS, 1996, no. 237, pp. 511–520. Wallingford, UK
Water Resources Research, 42, W02401, doi:10.1029/2004WR003751, 2006. Pringle, J., D. D. Stretch, and A. Bárdossy, “Automated classification of the
Basu, B. and V. V. Srinivas, “Regional flood frequency analysis using kernel- atmospheric circulation patterns that drive regional wave climates,” Nat.
based fuzzy clustering approach,” Water Resources Research, 50: 3295–3316, Hazards Earth Syst. Sci., 14: 2145–2155, 2014.
2014. Priya, K. L., “A fuzzy logic approach for irrigation water quality assessment:
Bogardi, I., L. Duckstein, and A. Bardossy, “Regional management of an a Case Study of Karunya Watershed, India,” Journal of Hydrogeology and
aquifer under fuzzy environmental objectives,” Water Resources Research, 19 Hydrologic Engineering, 2:1. doi:10.4172/2325-9647.1000104 2 (1), 2013.
(6): 1394–1402, 1983. Ross, J. L., M. M. Ozbek, and G. F. Pinder, “Aleatoric and epistemic uncer-
Buckley, J. J. and L. J. Jowers, Monte Carlo Methods in Fuzzy Optimization, tainty in groundwater flow and transport simulation,” Water Resources
Studies in Fuzziness and Soft Computing, Springer-Verlag, Berlin Heidelberg, Research,: 2009.45, W00B15, doi:10.1029/2007WR006799, 2009
2008, p. 222. Ross, T. J., Fuzzy Logic with Engineering Applications, 3rd ed., John Wiley &
Chang, Y. O. and B. M. Ayyub, “Fuzzy regression methods—a comparative Sons, 2010. Somerset, NJ.
assessment,” Fuzzy Sets and Systems, 119: 187–203, 2001. Sen, Z. and A. Altunkaynak, “A comparative fuzzy logic approach to runoff
CC-WaterS, Climate change and impacts on water supply, EU Transnational coefficient and runoff estimation,” Hydrological Processes, 20 (9): 1993–2009,
Program, 2012, http://www.ccwaters.eu/ 2006.
Chanas, S. and M. Nowakowski, “Single value simulation of fuzzy vari- Shrestha, R. R. and S. P. Simonovic, “Fuzzy sets theory based methodology
ables,” Fuzzy Sets and Systems, 25: 43–57, 1988. for the analysis of measurement uncertainties in river discharge and stage,”
Chandramouli, S. and U. V. Nanduri, “Comparison of stochastic and fuzzy Canadian Journal of Civil Engineering, 37 (3): 429–440, 2010.
dynamic programming models for the operation of a multipurpose reservoir,” Tayfur, G. and V. P. Singh, ANN and fuzzy logic models for simulating
Water and Environment Journal, 25: 547–554, 2011. event-based rainfall-runoff, Journal of Hydraulic Engineering, 132(12), Dec.
de Finetti, B., Theory of Probability, John Wiley & Sons, New York, 1974. 2006
Diamond, P. “Fuzzy kriging” Fuzzy Sets and Systems 33(3), 315–332, 1989. Tzimopoulos, C., L. Mpallas, and G. Papaevangelou, “Estimation of evapo-
Dong, W. M. and H. C. Shah, “Vertex method for computing functions of transpiration using fuzzy systems and comparison with the Blaney-Criddle
fuzzy variables,” Fuzzy Sets and Systems, 21 (1): 65–78, 1987. method,” Journal of Environmental Science and Technology, 1: 181–186, 2008.
Dou, C., W. Woldt, I. Bogardi, and M. Dahab, “Steady state groundwater Vujetic, D. and S. P. Simonovic, Water Resources Decision Making Under
flow simulation with imprecise parameters,” Water Resources Research, 31 (1), Uncertainty, Report 073, University of Western Ontario, April, 2011, p. 148.
2709–2719, 1995. London, Ontario, Canada,
Galambosi, A., E. C. Ozelkan, and L. Duckstein, “The impact of ENSO and Zhang, C., Y. Wang, L. Zhang, and H. Zhou, “A fuzzy inference method
macrocirculation patterns on precipitation under climate change,” Environ. based on association rule analysis with application to river flood forecasting,”
Geol., 58: 929–935, 2009. Water Science and Technology, 66 (10): 2090–2098, 2012.
Zadeh, L., “Fuzzy sets, ” Information and Computation, 8: 338–353, 1965.
12_Singh_ch12_p12.1-12.6.indd 5 8/22/16 11:39 AM

Chapter 13
Evolutionary Computing:
Genetic Algorithms
BY
AKIRA KAWAMURA AND TAREK MERABTENE
ABSTRACT by Storn and Price (1997), among others. All these techniques are now col-
lectively considered to be under the umbrella of EC, and the algorithms
Evolutionary computing (EC) techniques in hydrology and water resources
involved are termed evolutionary algorithms (EAs), because their natural
emerged in the early 1990s as procedures to alleviate some of the complexities
evolution biological basis is essentially the same (Ghosh and Tsutsui, 2003).
encountered in modeling and optimizing hydrologic systems and processes.
At present, there exist many different variants of EC techniques (Maier et al.,
Essentially, among many EC applications in hydrologic sciences, genetic algo-
2014).
rithms (GAs) have been the most commonly used methods for its simplicity,
The general scheme that forms the common basis for all EC techniques is
robustness, and practicability in implementation. This Chapter presents a
an iterative and stochastic process that operates on a population of individuals.
short history of EC techniques, their fundamentals and applications in
Each individual represents a potential solution to the problem being solved.
hydrology through an extensive elaboration around GAs as an example. GAs
Every individual is evaluated by means of a “fitness function,” a measure of its
applications to optimization problems in hydrology cover numerous fields
goodness for the problem at hand, and the higher value the better. Based on
such as hydrologic-data mining, water resources management, reservoir
the fitness values the fitter individuals are searched in the considered search
operations, groundwater, water quality, and rainfall-runoff modeling, among
space. The search process is basically driven by the three fundamental steps
others. Although the implementation procedure of GA is adequately similar
of selection, reproduction, and replacement (Ghosh and Tsutsui, 2003; Eiben
to any EC application, its specific elements, such as choice of the objective
and Smith, 2007). During the selection step, fitter individuals have higher
function, the genotype coding, selection strategy, and the crossover tech-
chance of being selected than less-fit ones. During the reproduction step,
nique, etc., must be identified according to the properties of the target hydro-
“recombination” and/or “mutation” is applied to the selected individuals in
logic problems to be solved. To bring hands-on application to water engineers,
the population to yield a new population. During the replacement step, indi-
practitioners and scientists, the GA application in the field of conceptual
viduals of the original population are substituted by the newly created indi-
rainfall-runoff modeling is elaborated. The four attached supplements to this
viduals. The whole process is repeated until a termination criterion is
chapter give additional illumination and step-by-step implementation, from
reached.
coding to result evaluation, of GA application to the optimization of the Tank
Model parameters as an example.
13.2 EVOLUTIONARY COMPUTING IN HYDROLOGY: AN
OVERVIEW
13.1 INTRODUCTION
Many hydrologic systems and processes are highly complex, nonlinear, time-
Evolutionary computing (EC) refers to the study of some guided random varying, and spatially distributed incorporating chaotic nature and random-
search techniques, developed based on the concepts of genetics and natural ness. With improved data acquisition and processing capabilities in the field,
evolution (Box, 1957) inspired from the Darwinian principles of survival of recently hydrologic sciences are supported by immense quantities of data.
the fittest. The idea of applying Darwinian principles to automated problem Data-driven EC techniques in hydrology and water resources emerged in the
solving dates back to 1948 when Turing (1948) proposed the idea of devising early 1990s to take full advantage of the information contained in such enor-
machine learning (i.e., the “Turing Test”) using “genetical or evolutionary mous data, and there has been a significant growth since then (Babovic and
search.” Effective implementations of EC techniques for design search and Rao, 2010). In recent years, many different EC techniques are being applied in
optimization, however, started in the 1960s when Bremermann (1962) carried hydrologic sciences, and they have shown to be very useful in a broad range
out computer experiments on “optimization through evolution and recombi- of hydrologic contexts due to their various advantages, which are most likely
nation.” During the era, three core EC approaches independently emerged responsible for their widespread adaptability (Awad, 2003; Keane, 2010;
(Eiben and Smith, 2007): (1) “evolutionary programming” by Fogel et al. Maier et al., 2014). Essentially, any EC technique is applied in order to solve
(1966), (2) “genetic algorithms” by Holland (1973), and (3) “evolution strate- problems which are generally considered to be in the category of so-called
gies” by Rechenberg (1973). Thereafter, in the 1990s other forms of EC were “optimization.” It should be noted, however, that EC techniques do not ensure
proposed, such as “genetic programming” by Koza (1990), “ant colony opti- an optimal solution, and that they may need parameter tuning for good per-
mization” by Colorni et al. (1991), SCE-UA (shuffled complex evolution formance. In the following, we focus and elaborate on “genetic algorithms,”
— University of Arizona) method by Duan et al. (1992), “particle swarm which have been the most commonly used method among many EC applica-
optimization” by Kennedy and Eberhart (1995), and “differential evolution” tions in hydrologic sciences (Nicklow et al., 2010).
13-1
13_Singh_ch13_p13.1-13.4.indd 1 8/22/16 11:41 AM

13-2 Evolutionary Computing: Genetic Algorithms
13.3 GENETIC ALGORITHMS random selection has proved to be more efficient to maintain population
13.3.1 Overview diversity (Mitchell, 1998).
The “roulette wheel sampling” (RWS) was the first and most popular
Following Goldberg (1989), the robustness and advantage of Genetic method to execute the fitness-proportionate selection (Goldberg, 1989). The
Algorithms (GAs) rise from their possibility of (1) coding parameters flexibly, RWS promotes the fitter individuals into the mating pool and ensures popula-
(2) conducting the search from a population point of view, (3) waiving major tion diversity by giving all individuals a relative chance to be selected for
mathematical restrictions on the objective function, such as continuity and reproduction. As the GA generation progresses, however, RWS does not guar-
differentiability, and (4) progressively improving the solution by stochastic, antee the survival of the fittest chromosome. Furthermore, the bias of the
not deterministic, rules. method toward the fittest parents runs a risk of premature convergence of the
The implementation of GAs has been extensively presented in the special- algorithm (Baker, 1985). To overcome the disadvantages of RWS, different
ized literature (e.g., Goldberg, 1989; Mitchell, 1998). In essence, GAs take selection methods were proposed, namely “tournament selection,” “stochastic
place in two spaces referred to as genotype and phenotype. Genotype is the universal sampling,” “linear rank selection,” “exponential rank selection,”
search space represented by a randomly generated population of “coded” “truncation selection,” “Boltzmann tournament selection,” and so on (Goldberg,
chromosomes (i.e., individuals). Phenotype is the solution space representing 1989; Mitchell, 1998). Decision on the suitable selection procedure has a sig-
the “decoded” variables of the problem at hand. Each chromosome in the nificant impact on narrowing the search space toward the most promising
genotype space carries several genes made up of a sequence of alleles (e.g., a region of global optimum, improving the convergence rate, and reducing the
0 or 1 in binary coding). Each gene is mapped over a fixed string length of number of generations to reach the optimal solution (Jebari and Madiafi, 2013).
alleles to represent one variable. The length of the chromosome is equal to the
total length of all genes congregated. During each GA evolution, called gen- 13.3.4 GA Operators
eration, chromosomes are rated for their fitness by the objective function. GA operators are the key to implement reproduction mechanisms by recom-
Fitter individuals have a greater probability to be selected as “parents” in the bination of parents into a new generation of offspring. Important reproduc-
“mating pool.” In order to progressively generate better “offspring” through- tion operators include crossover (i.e., recombination), mutation, inversion/
out the generations, a proficient selection routine must be conceived while permutation, elitism, and fitness scaling.
making usage of the “Schema Theorem” or the “Building Block Hypothesis” Crossover is the procedure of exchanging one or more substrings between
(Goldberg, 1989). During selection and reproduction processes, the popula- pairs of chromosomes. It is considered by Holland (1973) as the most impor-
tion diversity is maintained using genetic operators, namely crossover, muta- tant of all GA operators. Crossover or recombination occurs between two
tion, elitism, elimination, reordering, and reinsertion. selected chromosomes with some specific probability. All strings in the popu-
13.3.2 Parameters Representation and Coding lation have the chance of undergoing crossover, but not all will participate in
the process during a single generation. This is because the probability of
GAs require primarily all parameters to be represented and optimally esti- crossover is less than unity. A number of crossover techniques have been
mated (i.e., potential solutions) by a specific “genetic code” using some finite proposed. Holland’s initial analysis of one-point crossover has been extended
alphabet. The binary encoding is the most common coding (Mitchell, 1998), to two-point, n-point, and uniform crossover (De Jong, 1975). Examples of
even though any other coding (e.g., octal, hexadecimal, ternary, integer, real- crossover are depicted in Supplement 13.2.
valued, and permutation coding) suitable for the problem can be used Mutation randomly changes the allele values of some locations in the chro-
(Kumar, 2013). mosome. It is applied to each string in the new population assuming a low
Under the binary coding, potential solutions are represented as a finite- mutation probability. A common implementation procedure of mutation
length string of 0s and 1s. The length of the gene (i.e., substring) reflects the consists of flipping the bit at a randomly chosen locus. This allows the algo-
level of accuracy and/or the range values of the parameter. If the solution is a rithm to reintroduce genetic diversity, expand search space, and avoid getting
vector, as it is for most applications in hydrology, the string is created by trapped in local optima. It is worth noting that “evolutionary strategies” stress
concatenating the binary coding of vector coordinates into a single string (i.e., the role of mutation as the core operator for reproduction (Spears, 1993).
a chromosome made of multiple genes). The relationship between the real Inversion/permutation procedure consists of reversing the order of adja-
value of a parameter and its binary string is achieved by coding and decoding cent sections of the chromosome, thus rearranging the order in which genes
processes. The coding of a variable R, whose domain is [Rmin , Rmax ] with a are arrayed. Though inversion/permutation can potentially bring high-fitness
desired precision ε, requires a string of at least ( Rmax − Rmin ) bits length. The individuals together so that they are less subject to later disruption due to
ε crossover (Koza, 1990), its practical efficiency is still to be documented.
string length L is estimated to meet the following criteria: Elitism, first introduced by De Jong (1975), forces the GAs to retain some
of the best individuals at each generation to pass to the next generation with-
2 L−1 <
( Rmax − Rmin ) ≤ 2L − 1 (13.1) out being modified. With elitism process, the diversity of GA population is
ε well kept while the best individual is protected from being lost or destroyed
by crossover or mutation. Nevertheless, while elitism is proven to signifi-
The decoding of the binary string into its real value R is given by: cantly improve the GAs’ performance by protecting good individuals, it is also
proven that excessive selection of elites increases the number of duplicates
Rmax − Rmin L
R = Rmin +
2 L − 1 i=1
( )
∑ bi ∗ 2i−1 (13.2) within the new generation to a point where it negatively impacts the perfor-
mance (Mitchell, 1998).
Fitness scaling can greatly improve genetic diversity among individuals
where bi is the bit value at position i (bi = 0 or 1). An example of coding is shown toward the end of GA run by overcoming the risk of premature convergence,
in Supplement 13.1. created by the bias of the “fitness proportionate selection” like RWS (Goldberg,
13.3.3 Fitness Function, Selection, and Reproduction 1989). Among all the proposed fitness scaling methodologies, linear scaling is
still largely used for its simplicity and efficiency to amplify the distinction
In natural sciences, the fitness of a chromosome determines its aptitude to between good and very good chromosomes (Kreinovich et al., 1993).
survive in its environment and reproduce offspring. In GAs, the fitness func-
tion is derived from the objective function and used in successive genetic
operations. It is used to characterize individuals’ performance in the problem 13.3.5 Termination Criteria
domain (i.e., fitness landscape), leading to a robust search less likely to get Termination criteria specify the conditions to stop the GA run. Some of the
trapped in regions of local optima. The objective function refers to the func- commonly used conditions for termination are based on (1) generation num-
tion to be optimized. In hydrology, optimization problems usually imply the ber: the user-specified maximum number of evolutions (i.e., generations)
minimization of a single or multiple objective functions. The fitness is there- have been reached, which is commonly used; (2) evolution time: the elapsed
fore inversely proportional to the objective function (see Sec. 13.4.2). evolution time exceeds the user-specified maximum evolution time; (3) con-
Selection is carried out stochastically through a fitness-based process vergence of fitness values: the fitness is considered as converged or, for a given
known as “fitness-proportionate selection” (Holland, 1973), where each indi- period of time, the fitness improvement remains under an acceptable thresh-
vidual in the population has a chance to be selected as a “parent” into the old; (4) population convergence: the average fitness across the current popu-
“mating pool” for reproduction according to an expected number of times. lation is less than a user-specified percentage; and (5) gene convergence: a
Reproduction is the process of creating offspring from two selected parents by user-specified average value of a gene across all of the chromosomes in the
applying GA operators. The selection of pairs can be simply executed by current population is less than a user-specified percentage. Since GAs employ
selecting each consecutive pairs in the mating pool. However, a probabilistic stochastic processes to search the fitness landscape, there are no guarantees to
13_Singh_ch13_p13.1-13.4.indd 2 8/22/16 11:41 AM

Conclusion and Future Directions 13-3
reach an optimum. Therefore, the use of one or a combination of different With respect to the population size, Mitchell (1998) suggests a minimum
termination criteria is prevailing. of 25 to 50 individuals per population, and practical applications of CRMMs
recommend a population size of 100 individuals (Wang, 1991;
13.4 GA APPLICATIONS IN HYDROLOGY
Tanakamamu, 1993).
Selection of a suitable objective function is essential to the success of any
There exist a wide range of GA applications to optimization problems in EC technique in hydrology. In application to CRRMs, the key is to use an
hydrology, which cover hydrologic-data mining, water resources manage- adequate indicator to compare observed and simulated runoff for peak flows
ment, reservoir operations, groundwater modeling, and automatic calibration as well as low flows (Duan et al., 1992). If a single objective function is judged
of distributed and conceptual models for rainfall-runoff analysis, among oth- to be inefficient, then multiple objective functions must be sought (Madsen,
ers (refer to Nicklow et al., 2010; Babovic and Rao, 2010; Yoo and Kim, 2014). 2000). Objective functions generally used in CRRMs include the simple least
Hydrologic systems and processes are highly interdependent in nature, and square (SLS), the root mean square error (RMSE), the heteroscedastic maxi-
similarly involve a large number of parameters to be optimally estimated. mum likelihood estimator (HMLE) (Sugawara, 1979; Duan et al., 1992), and
Although the GA implementation procedure is adequately similar to any the Nash–Sutcliffe efficiency (NSE) (Nash and Sutcliffe, 1970) that is also
hydrologic application, its specific elements such as choice of the objective commonly used to measure the model’s efficiency.
function, the genotype coding, selection strategy, and the crossover tech- As a powerful selection mechanism, the RWS with probability fitness
nique, etc., are somewhat different according to the properties of the target evaluation function has been successfully tested for CRRMs (Wang, 1991;
hydrologic problems to be solved. In this section, GA applications mainly in Franchini and Galeati, 1997). The method implements a probabilistic selec-
the field of conceptual rainfall-runoff models (CRRMs) are elaborated by the tion based on the fitness ranking of the individuals prior to selection. Another
way of examples. more powerful method, judged to be applicable to hydrologic optimization
problems, is known as the “Boltzmann tournament selection” inspired from
13.4.1 Conceptual Rainfall-Runoff Models “simulated annealing” (Goldberg, 1989).
If one-point and/or two-point crossover are tested and proven not to be
Among all the aforementioned hydrologic processes, optimization of CRRMs
efficient for the model structure (because of the large genotype), then alterna-
is a central focus for many hydrologists owing to their widespread use and
tively the uniform crossover should be explored as it has been reported to
crucial importance for all aspects of watershed management. CRRMs are
perform very well for CRRMs (Seibert, 2000).
commonly used as a means to predict catchment runoff, especially in case
With regards to calibration of CRRMs, a number of studies have been
that the field observation data are not sufficient (in quality and quantity) to
conducted that compared GA with other EC techniques (e.g., Duan et al.,
apply distributed rainfall-runoff models. Many CRRMs commonly contain a
1992; Gan and Biftu, 1996; Cooper et al., 1997; Thyer et al., 1999). These stud-
set of interconnected storages in cascade in their model structure to account
ies demonstrated that hybrid EC methods such as the SCE-UA method are
for the different subflows contributing to the runoff hydrograph. Their hydro-
superior search algorithms. The SCE-UA method has been widely applied for
logic responses are conceptually represented by a set of unknown parameters,
calibration of various CRRMs, including the SAC-SMA model (e.g.,
which are not necessarily measurable or have physical meaning, but need to
Sorooshian et al., 1993; Gan and Biftu, 1996), the Tank model (Tanakamaru
be evaluated and optimized through calibration. As a result, optimal param-
and Burges, 1996; Cooper et al., 1997), the Xinanjiang model (Gan and Biftu,
eter estimation is a crucial problem for CRRMs.
1996), and so on.
In general, CRRMs involve a large number of parameters to be optimally
estimated depending on the model structure. For example, the “Tank Model”
(Sugawara, 1961) has 16 parameters in its typical four-stage version. The 13.5 CONCLUSION AND FUTURE DIRECTIONS
illustration and formulation of the four-stage Tank Model and an example of
In light of the increasing availability of hydrologic data as well as the growing
GA implementation for the Tank Model are provided in Supplements 13.3
complexity of hydrologic models, data-driven EC techniques are playing an
and 13.4, respectively. Other examples of popular CRMMs (Singh, 1995) are
increasingly important role in providing promising solutions to a wide range
the Stanford Watershed Model (Crawford and Linsley, 1966), the SAC-SMA
of problems in hydrologic sciences. At present, many different EC techniques
(SACramento Soil Moisture Accounting) model (Burnash et al., 1973), the
exist and new variants are continuously being proposed, claiming their nov-
HBV (Hydrologiska Byråns Vattenbalansavdelning) model (Bergström,
elty and practical efficacy. The EC techniques are undoubtedly attractive, but
1976), and the Xinanjiang model (Zhao et al., 1980). In general, the aforemen-
they require problem-domain expertise and certain precautions to bring out
tioned models can be structured to involve from 6 to more than 20 parameters
their full potential (Babovic and Rao, 2010). The main research challenge for
to be evaluated through manual or automatic calibration. Manual parameter
applying present EC techniques to the hydrologic problems is to shift research
calibration of CRMMs has proven to be cumbersome and time consuming,
efforts from the development and performance assessment for relatively
and only efficient if maintained by an experienced and knowledgeable expert
simple case studies to efficient and accurate real-world applications, espe-
hydrologist. As a result, nowadays, automatic calibration has been more fre-
cially for decision support in complex, and uncertain application contexts
quently adopted (Yilmaz et al., 2010). It is worth mentioning, however, that no
(Maier et al., 2014). EC applications in hydrology are generally still in their
automatic calibration can guarantee the uniqueness of the obtained parameter
infancy and there is a need to direct research efforts into the associated
sets due to the problem of so-called “equifinality.” Generally, the subspace
domains.
associated with each parameter contains a great number of local optima that
are not necessarily located in the vicinity of the global optimum (Sorooshian
and Gupta, 1983; Merabtene et al., 1997). To date, several GA schemes have REFERENCES
been successfully applied to the optimal calibration of various CRRMs, which
Awad, H., A. Kawamura, K. Jinno, and Y. Kuno, “Evolutionary computing
include the Tank Model (Tanakamaru, 1993), the Xinanjiang model (Wang,
techniques for optimal pressure regulation in water distribution networks,”
1991), NAM model (Madsen, 2000), and so on.
Annual Journal of Hydraulic Engineering, JSCE, 47: 865–870, 2003.
Babovic, R. and R. Rao, “Evolutionary computing in hydrology,” Advances
13.4.2 Remarks in GA Implementation for CRRMs in Data-Based Approaches for Hydrologic Modeling and Forecasting, edited by
The feasible search range of model parameters for any CRRM must be set B. Sivakumar and R. Berndtsson, World Scientific Publishing, Singapore,
based on specialized literature. In general, the lower and upper bounds for Chap. 7, 2010, pp. 347–369.
each parameter can be obtained by using sensitivity analysis (Duan et al., Baker, J. E., “Adaptive selection methods for genetic algorithms,” Proceedings
1992), and/or manual calibration (Merabtene et al., 1997). If the range of the of the 1st International Conference on Genetic Algorithms, edited by J. J.
parameters cannot be properly predetermined, then the feasible domains Grefenstette, Lawrence Erlbaum Associates, Hillsdale, NJ, 1985, pp. 101–111.
must be set based on the problem’s operational constraints. Merabtene et al. Bergström, S., “Development and application of a conceptual runoff-
(2002) proposed to delineate the practically feasible parameter domains in model for Scandinavian catchments,” Reports RHO7, SMHI, Norrköping,
their water resources management application. Sweden, 1976.
When creating the genotype space using binary coding, each parameter is Box, G. E. P., “Evolutionary operation: a method of increasing industrial
to be discretized into a gene of length L bits according to Eq. (13.1). For the productivity,” Journal of Applied Statistics, 6 (2): 81–101, 1957.
sake of simplicity, the parameters are usually mapped using the same number Bremermann, H. J., “Optimization through evolution and recombination,”
of alleles. Following Wang (1991), L = 7 bits is a suitable gene length to con- Self-Organizing Systems, edited by M. C. Yovitts et al., Spartan Books,
form to the “Schemata Theorem.” Note that other coding systems have also Washington, D.C., 1962, pp. 93–106.
been successfully used for CRRMs, such as integer coding (Muleta and Burnash, R. J. C, R. L. Ferral, R. A. McGuire, and R. A. McGuire, A
Nicklow, 2005) and real coding (Franchini and Galeati, 1997; Seibert, 2000). Generalized Streamflow Simulation System Conceptual Modeling for Digital
13_Singh_ch13_p13.1-13.4.indd 3 8/22/16 11:41 AM

13-4 Evolutionary Computing: Genetic Algorithms
Computers, the Joint Federal-State River Forecast Center, U.S. Department of Mitchell, M., An Introduction to Genetic Algorithm, The MIT Press,
Commerce, National Weather Service, State of California, Department of Cambridge, MA, 1998.
Water Resources, Sacramento, CA, 1973. Muleta, M. K. and J. W. Nicklow, “Decision support for watershed
Colorni, A., M. Dorigo, and V. Maniezzo, “Distributed optimization by ant management using evolutionary algorithms,” Journal of Water Resources
colonies,” Proceedings ECAL91—European Conference on Artificial Life, Planning and Management, 131 (1): 35–44, 2005.
Elsevier, Paris, France, 1991, pp. 134–142. Nash, J. E. and J. V. Sutcliffe, “River flow forecasting through conceptual
Cooper, V. A., V. T. V. Nguyen, and J. A. Nicell, “Evaluation of global models. Part I—A discussion of principles,” Journal of Hydrology, 10 (3):
optimization methods for conceptual rainfall–runoff model calibration,” 282–290, 1970.
Water Science and Technology, 36 (5): 53–60, 1997. Nicklow, J., P. Reed, D. Savic, T. Dessalegne, L. Harrell, A. Chan-Hilton, M.
Crawford, N. H. and R. K. Linsley, “Digital simulation in hydrology: Karamouz, et al., “State of the art for genetic algorithms and beyond in water
Stanford watershed model IV,” Technical Report, No. 39, Department of Civil resources planning and management,” Journal of Water Resources Planning
Engineering, Stanford University, 1966. and Management, 136 (4): 412–432, 2010.
De Jong, K. A., “An analysis of the behavior of a class of genetic adaptive Rechenberg, I., Evolutionsstrategie: Optimierung Technischer Systeme Nach
systems,” Doctoral Dissertation, University of Michigan, Ann Arbor, MI, 1975. Prinzipien der Biologischen Evolution, Frommann-Holzboog, Stuttgart, 1973.
Duan, Q., S. Sorooshian, and V. K. Gupta, “Effective and efficient global Seibert, J., “Multi-criteria calibration of a conceptual runoff model using a
optimization for conceptual rainfall-runoff models,” Water Resources GA,” Hydrology and Earth System Sciences, 4 (2): 215–224, 2000.
Research, 28 (4): 1015–1031, 1992. Singh, V. P. (Ed.) Computer Models of Watershed Hydrology, Water
Eiben, A. E. and J. E. Smith, “Introduction to evolutionary computing,” Resources Publications, Highlands Ranch, CO, 1995.
Natural Computing Series, Springer, Berlin Heidelberg New York, 2007, pp. Sorooshian, S. and V. K. Gupta, “Automatic calibration of conceptual
1–14, corrected 2nd printing. rainfall-runoff models: the question of parameter observability and
Fogel, L. J., A. J. Owens, and M. J. Walsh, Artificial Intelligence Through uniqueness,” Water Resources Research, 19 (1): 260–268, 1983.
Simulated Evolution. John Wiley & Sons, New York, NY, 1966. Sorooshian, S., Q. Duan, and V. K. Gupta, “Calibration of rainfall-runoff
Franchini, M. and G. Galeati, “Comparing several genetic algorithm models: application of global optimization to the Sacramento soil moisture
schemes for the calibration of conceptual rainfall-runoff models,” Hydrological accounting model,” Water Resources Research, 29 (4): 1185–1194, 1993.
Sciences Journal, 42 (3): 357–379, 1997. Spears, W. M., “Crossover or mutation?,” Foundations of Genetic Algorithms,
Gan, T. Y. and G. F. Biftu, “Automatic calibration of conceptual rainfall– edited by L. D. Whitley, Morgan Kaufmann, San Mateo, CA, Vol. 2, 1993,
runoff models: optimization algorithms, catchment conditions, and model pp. 221–237.
structure,” Water Resources Research, 32 (12): 3513–3524, 1996. Storn, R. and K. Price, “Differential evolution—A simple and efficient
Ghosh, A. and S. Tsutsui, (Ed.) Advances in Evolutionary Computing: heuristic for global optimization over continuous spaces,” Journal of Global
Theory and Applications, Springer-Verlag Berlin Heidelberg New York, 2003. Optimization, 11: 341–359, 1997.
Goldberg D. E., Genetic Algorithms in Search, Optimization and Machine Sugawara, M., “On the analysis of runoff structure about several Japanese
Learning, Addison-Wesley, Longman, Boston, MA, 1989. rivers,” Japanese Journal of Geophysics, 2 (4): 1–76, 1961.
Holland, J. H., “Genetic algorithms and the optimal allocation of trials,” Sugawara, M., “Automatic calibration of the Tank Model,” Hydrological
SIAM Journal on Computing, 2 (2): 88–105, 1973. Sciences Bulletin, 24 (3): 375–388, 1979.
Jebari, K. and M. Madiafi, “Selection methods for genetic algorithms,” Tanakamaru, H., “Parameter identification of the tank model with the
International Journal of Emerging Sciences, 3 (4): 333–344, 2013. genetic algorithm,” Annuals of Disaster Prevention Research Institute, 36 (B-2):
Keane, A., “An introduction to evolutionary computing in design search and 231–239, 1993.
optimisation,” Theoretical Aspects of Evolutionary Computing, edited by L. Kallel, Tanakamaru, H. and S. J. Burges, “Application of global optimization to
B. Naudts, and A. Rogers, Springer-Verlag Berlin Heidelberg, 2010, pp. 1–11. parameter estimation of the Tank model,” Proceedings International Conference
Kennedy, J. and R. Eberhart, “Particle swarm optimization,” Proceedings of on Water Resources and Environmental Research, Kyoto University, Kyoto,
IEEE International Conference on Neural Networks IV, IEEE Service Center, 1996, pp. 39–46.
Perth, Australia, 1995, pp. 1942–1948. Turing, A. M., “Intelligent machinery, a heretical theory,” The Turing Test:
Koza, J. R. “Genetic programming: a paradigm for genetically breeding Verbal Behavior as the Hallmarks of Intelligence, edited by S. M. Shieber, the
populations of computer programs to solve problems,” Computer Science MIT Press, Cambridge, MA, 1948, pp. 105–110.
Department Technical Report, Stanford University, Stanford, CA, 1990, STAN- Thyer, M., G. Kuczera, and B. C. Bates, “Probabilistic optimization for
CS-90-1314. conceptual rainfall–runoff models: a comparison of the shuffled complex
Kreinovich, V., C. Quintana, and O. Fuentes, “Genetic algorithms: what evolution and simulated annealing algorithms,” Water Resources Research, 35
fitness scaling is optimal?,” Cybernetics and Systems: An International Journal, (3): 767–773, 1999.
24 (1): 9–26, 1993. Yilmaz, K. K., J. A. Vrugt, H. V. Gupta, and S. Sorooshian, “Model
Kumar, A., “Encoding schemes in genetic algorithm,” International Journal calibration in watershed hydrology,” Advances in Data-Based Approaches for
of Advanced Research, IT Engineer, 2 (3): 1–7, 2013. Hydrologic Modeling and Forecasting, edited by B. Sivakumar and R.
Madsen, H., “Automatic calibration of a conceptual rainfall-runoff model Berndtsson, World Scientific Publishing, Singapore, Chap. 3, 2010, pp.
using multiple objectives,” Journal of Hydrology, 235: 276–288, 2000. 53–105.
Maier, H. R., Z. Kapelan, J. Kasprzyk, J. Kollat, L. S. Matott, M. C. Cunha, Yoo, D. G. and J. H. Kim, “Meta-heuristic algorithms as tools for hydro-
G. C. Dandy, et al., “Evolutionary algorithms and other metaheuristics in logical Science,” Geoscience Letters, 1 (4): 1–7, 2014.
water resources: current status, research challenges and future directions,” Wang, Q. J., “The genetic algorithm and its application to calibrating
Environmental Modelling and Software, 62: 271–299, 2014. conceptual rainfall-runoff models,” Water Resources Research, 27 (9):
Merabtene, T., K. Jinno, and A. Kawamura, “Interactive user interface for 2467–2471, 1991.
rainfall-runoff analysis by tank model,” Memoirs of the Faculty of Engineering, Zhao, R. J., Y. L. Zhuang, L. R. Fang, X. R. Liu, and Q. S. Zhang, “The
Kyushu University, 57 (3): 107–120, 1997. Xinanjiang model,” Proceedings of the Oxford Symposium on Hydrological
Merabtene, T., A. Kawamura, K. Jinno, and J. Olsson, “Risk assessment for Forecasting, IAHS Publication No. 129, IAHS Press, Wallingford, UK, 1980,
optimal drought management of integrated water resources system using pp. 351–356.
genetic algorithm,” Hydrological Processes, 16 (11): 2189–2208, 2002.
13_Singh_ch13_p13.1-13.4.indd 4 8/22/16 11:41 AM

Chapter 14
Relevance Vector Machine
BY
SHIVAM TRIPATHI AND RAO S. GOVINDARAJU
ABSTRACT 14.2 BACKGROUND
Recent advances in statistical learning theory have yielded tools that are Tipping (2001) developed RVM as a Bayesian alternative to support vector
improving our capabilities for analyzing large and complex datasets. Among machine (SVM; Vapnik, 1998), which in the late 1990s was setting new
such tools, relevance vector machines (RVMs) are finding increasing applica- records for prediction performances in many areas, including hydrology
tions in hydrology because of their excellent generalization properties, and (Tripathi et al., 2006). Like SVM, the RVM relies on kernel trick to approxi-
their probabilistic interpretation that yields prediction uncertainty. RVMs mate nonlinear functions. However, while SVM achieves good generalization
combine the strengths of kernel-based methods and Bayesian theory to estab- performance by implementing the structural risk-minimization principle that
lish relationships between a set of input vectors and a desired output. This attempts to minimize an upper bound on the generalization error by striking
chapter briefly reviews the motivation behind RVMs and their different for- a right balance between the training error and the capacity of machine (i.e.,
mulations for regression and classification problems, discusses subtleties the ability of machine to learn any training set without error), RVM achieves
involved in their implementation, and surveys their applications in hydrology. the same by implementing Bayesian Ockham’s Razor. The following sections
The chapter concludes by drawing connections between RVMs and other briefly present the motivation and background of the kernel trick and sparse
statistical models used in hydrology and presents directions for their possible Bayesian learning in the context of RVM.
extensions.
14.2.1 Kernel Trick
Many linear models have an equivalent dual representation in which the input
14.1 INTRODUCTION vectors x and x ′ appear in the form of inner product x T x ′. (In this chapter,
A large fraction of hydrological data tends to be high dimensional (i.e., exhib- matrices are represented by uppercase bold italic letters. All vectors are column
its large variability over a wide range of space and time scales) and severely vectors and are represented by lowercase bold italic letters, and normal font italic
underconstrained (sparse coverage over the input space), and is often inter- letters represent scalars.) Such models can be converted into nonlinear models
spersed with spurious data that confound analyses. For hydrologic applica- by mapping the input vector into a nonlinear feature space φ( x ) [e.g., if x is a
tions, the challenges in data modeling are: (a) achieving good generalization two-dimensional vector, i.e., x = ( x1 , x 2 )T , φ( x ) = ( x12 , 2 x1x 2 , x 22 )T is a
performance in high-dimensional data where the curse of dimensionality possible quadratic mapping into a three-dimensional feature space], and in
stipulates exponential increase in the amount of data for good predictions, (b) their dual representation, the input vector appears in the form of φ( x )T φ( x ′ ).
representing and effectively combining the available physical knowledge of The kernel trick (or kernel substitution) is to replace the inner product by a
hydrological systems into data-learning algorithms, (c) characterizing and kernel function K( x , x ′ ) = φ( x )T φ( x ′ ) that operates in the original space
quantifying uncertainty in predictions, and finally (d) making learning algo- rather than the feature space. Thus, the inner product can be computed with-
rithms computationally efficient to handle voluminous data. out explicitly mapping the data into the feature space, which is often compu-
Relevance vector machine (RVM) offers a novel approach for addressing tationally cheaper [for the quadratic mapping exemplified earlier, the kernel
these challenges. It has several appealing attributes (listed as follows) and, function is given by K( x , x ′ ) = ( x T x ′ )2 = ( x1x1′ + x 2 x 2′ )2 = φ( x )T φ( x ′ )]. The
consequently, has found several applications in hydrology. kernel function can be constructed either by explicitly mapping φ( x ) and
1. RVM provides a probabilistic framework yielding both model and pre- then finding the corresponding kernel function or by directly specifying the
diction errors, which not only quantify the inherent uncertainty in predic- function. A necessary condition for K( x , x′ ) to be a valid kernel is that
tions but also separate uncertainties due to noise in the training data and the Gram matrix K , whose (i , j )th element is K( xi , x j ), is a positive
extrapolation during prediction. semidefinite matrix (Bishop, 2006).
2. RVM includes a Bayesian construct that induces regularization through
14.2.2 Sparse Bayesian Analysis
automatic relevance determination (ARD), yielding more robust models that
are not overly complex. The kernel trick provides a simple and elegant way of capturing nonlinear
3. RVM design involves relatively little subjectivity, compared to other relationships between input vectors and corresponding outputs using linear
common data-driven or machine-learning approaches. algorithms. However, most of the kernel methods from the late 1990s did not
This chapter reviews the background and motivation of RVM, presents its provide important measures such as predictive distribution of the evidence as
mathematical formulation, discusses subtleties involved in its implementa- they were not developed with a probabilistic foundation. Thus, the important
tion, and surveys its applications in hydrology. The focus is on RVM regres- issue of model uncertainty remained unsolved. A possible way to overcome
sion, which is more common in hydrological applications than RVM this problem is to exploit versatility of kernel methods in conjunction with
classification; however, the suggestions made are equally applicable to classi- Bayesian learning. Based on this idea, Tipping (2001) developed RVMs that
fication problems. Further, the chapter investigates reasons for sparsity in are kernel-based methods formulated under a Bayesian construct. RVMs
RVM results, draws connections between RVM and other statistical models provide sparse solution to regression and classification tasks by implementing
used in hydrology, and comments on its limitations and possible extensions. Bayesian ARD (MacKay, 1996) in the transformed feature space.
14-1
14_Singh_ch14_p14.1-14.8.indd 1 8/22/16 11:42 AM

14-2 Relevance Vector Machine
14.3 MATHEMATICAL FORMULATION 14.3.2 Inference

14.3.1 Model specification After assigning priors and defining the likelihood function [Eq. (14.3)], the
posterior distribution of the model parameters, p(w , α , σ 2ε | y ) , is estimated
Consider a finite training sample of N patterns {( xi , yi ), i = 1,…, N }. The ith
by applying Bayes’ rule. For RVM, the posterior distribution does not have a
input pattern xi denotes a vector of m-independent variables (i.e.,
closed form and hence has to be approximated by one of the following three
xi = [xi1 ,…, xim ] ∈ℜm and X = [x1 ,…, x N ]T ), and yi ∈ℜ ( y = [ y1 ,…, y N ]T ) is
methods:
the corresponding dependent variable.
Further, let the regression relationship be defined as: 1. Type II maximum likelihood: This approximation was used by Tipping
N
(2001) in the first RVM implementation. In this approximation, the posterior
yi = f ( xi , X ; w ) + εi = ∑ w j K( xi , x j ) + εi (14.1) distribution is decomposed into two components as:
j =0
( ) (
p w , α , σ 2ε | y = p w | y , α , σ 2ε p α , σ 2ε | y (14.5) )( )
where w = [w0 , w1 ,…, w N ]T is a weight vector, εi is an error term assumed to
have Gaussian distribution with mean zero and variance σ 2ε , and K( xi , x j ) The first term on the right-hand side of Eq. (14.5) is the posterior probability
is a kernel function. There are several possibilities for the choice of kernel of the weight w given by σ 2ε and α, and is normally distributed,
function, including linear, algebraic polynomials, trigonometric polynomials,
p(w | α , σ 2ε , y ) = N (µ w , ∑ w2 ), where mean and covariance are µ w = σ −ε 2 ∑w ΦT y
radial basis functions (RBFs), and sigmoid functions. The RBF kernel which
is more commonly used in hydrological application is defined as: and ∑w = (σ −2 T −1
ε Φ Φ + A ) , with A = diag(α 0 , α1 ,, α N ) and Φ N × N +1 is the
design matrix with elements Φij = K( xi , x j ), i = 1,, N ; j = 0,, N . The
 1 if j = 0 second term on the right-hand side of Eq. (14.5) is the posterior probability
 of σ 2ε and α. The estimation of this probability is analytically intractable and
K( xi , x j ) = 
(
 exp − xi − x j

2
σ 2kernel ) otherwise
(14.2)
is approximated by delta function at its mode, that is,
where σ 2kernel is the width of the kernel function. The training of an RVM
( ) ( )
p α , σ 2ε | y ~ δ α max , σ ε2 max (14.6)
model involves estimation of parameters w , σ 2ε , and σ 2kernel . The parameter
σ 2kernel is often assumed to be known. In Sec. 14.4.2, we show that σ 2kernel is where σ 2ε max and α max are values for which p(α , σ 2ε | y ) reaches its maximum
one of the most sensitive parameters controlling the generalization perfor- value. It turns out that maximizing p(α , σ 2ε | y ) is equivalent to maximizing
mance of RVM, and suggest few heuristic techniques for estimating it. The the marginal likelihood p( y | α , σ 2ε ) given by
likelihood of the dataset for this model [Eq. (14.1)] can be written as:
N
 1 2 ( ) ( )
p y | α , σ 2ε = ∫ p y | w , α , σ ε2 p ( w | α )d w
( )
p y w , σ 2ε = ∏
1 
exp − 2 [ yi − f ( xi , X ; w )]  (14.3)
{ }
2 σ 1
 
( )
2
2πσ ε −1 −12
i =1  ε
= ( 2π )
exp − y T σ 2ε I + ΦA −1Φ T y
−N 2
σ 2ε I + ΦA −1ΦT
2
In RVMs, learning is achieved by specifying a prior distribution for model (14.7)
parameters and estimating a posterior distribution for parameters by using
the likelihood function [Eq. (14.3)]. An ARD prior of the following form is Closed-form solutions of σ 2ε max and α max are not available. Tipping (2001)
assigned to each weight vector proposed an iterative estimation method based on a type of ARD (MacKay,
1996). After convergence, the learning algorithm provides hyperparameter
 1
p w j α j = N  0, 
 αj
( ) (14.4)
posterior mode ( σ 2ε max and α max ) and posterior distributions of weights con-
ditioned on σ 2ε max and α max . During the process of learning, some values of
α j may approach infinity, and so the corresponding weights w j tend to be
delta functions centered at zero, and are thus automatically omitted from
where α j are called hyperparameters. The choice of a zero-mean Gaussian Eq. (14.1). Only the remaining patterns corresponding to nonzero weights are
prior for weights expresses a preference for smaller weights, and hence for a deemed relevant for function approximation [hence the term, relevance vector
smoother model [Eq. (14.1)]. To complete the model specification, the machine (RVM)].
remaining parameters, namely hyperparameters, α j , and variance of the noise
term, σ 2ε , are assigned uniform uninformative priors over a logarithmic scale. 2. Variational Bayes: The aim here is to approximate the posterior distribu-
Figure 14.1a shows the schematic of RVM using directed acyclic graph (DAG). tion p(w , α , σ 2ε | y ) by q(w , α , σ 2ε | y ) such that q(w , α , σ 2ε | y ) is similar to
p(w , α , σ 2ε | y ) , and at the same time has a closed form. This method is based
on the principles of calculus of variations and is therefore referred to as varia-
tional inference or variational Bayes method. In the machine learning litera-
a W s2 a W s2 ture, it is also known as ensemble method (MacKay, 2005). A complete
Bayesian treatment of RVM through variational inference was proposed by
Bishop and Tipping (2000) in which the approximating distribution is
assumed to factorize as:
m yn m yn Bn
( ) ( )
p w , α , σ 2ε | y ≈ q w , α , σ 2ε | y = qw ( w )qα ( α )qτ ( τ ) (14.8)
zn
where the precision τ = σ −2ε . In variational RVM, the prior distribution of w
is same as that of original RVM; however, α and τ are assigned flat gamma
Xn Xn X̃ n priors resulting in a closed form for q(w , α , σ 2ε | y ). Sparsity in this formula-
tion is achieved by pruning weights that are very close to zero.
n = 1,...N n = 1,...N
3. Markov Chain Monte Carlo: The third alternative for approximating the
posterior distribution is stochastic approximation by using sampling algo-
(a) RVM (b) VNRVM rithms. Markov Chain Monte Carlo (MCMC) methods, developed in the
Figure 14.1 RVM and VNRVM representation using DAGs. A circle or a node physics literature, are the most popular sampling methods used in Bayesian
denotes a random variable, a shaded circle represents observed variable, a nonshaded inference because they are efficient in sampling high-dimensional spaces,
circle represents a model parameter or hidden variable, solid elliptical dot represents a easy to design, and applicable to large class of distributions (Bishop, 2006).
deterministic parameter, and an arrow denotes a distribution of the variable on its head These methods sample by constructing a Markov chain whose stationary or
conditioned on the variable on its tail, for example, an arrow connecting x to y
equilibrium distribution is the required posterior distribution. Juncai et al.
represents p( y | x ). A plate marked n = 1,, N denotes that there are N nodes of the
type of which only one example is given inside the plate. In VNRVM, the symbol x n (2010) provided an MCMC implementation of RVM for studying biological
denotes nth input vector with ζ n measurement error that follows Gaussian distribution sequence data. Like variational noisy relevance vector machine (VNRVM—to
with zero mean and variance Bn−1. be described later), proper prior distribution is used and sparsity is achieved
14_Singh_ch14_p14.1-14.8.indd 2 8/22/16 11:43 AM

Application of RVM 14-3
by pruning weights that have relatively smaller magnitudes. The MCMC 1000
implementation for RVM is computationally very expensive and can only be
used for relatively small datasets.
800
14.3.3 Prediction
After having determined the posterior distribution of model parameters, the
predictive distribution of the dependent variable y ∗ for a new set of indepen- 600
dent variables x ∗ is obtained by integrating out the model parameters
400
( ) ( )( )
p y ∗ | y = ∫∫∫ p y ∗ | y , w , σ 2ε max p w , α , σ 2ε | y dw dα dσ 2ε (14.9)
If type II maximum likelihood approach is used for parameter estimation, the 200
predictive distribution is Gaussian with mean and variance given by
µ y∗ = µ w K( x * , X ) and σ 2y∗ = σ ε2 max + K( x * , X )T Σ w K( x * , X ) , respectively. For
variational RVM, the predictive distribution does not have a closed form, but 0
it can be approximated by a Gaussian distribution if the precision is assumed 0 200 400 600 800 1000
to be tightly concentrated around its mean value. For MCMC RVM, the Figure 14.2 Locations of electrical conductivity measurements obtained from Zhang
integration in Eq. (14.9) can be performed numerically by using the samples (1990). The squares show the measurement locations, while the rhombus shows the
generated during the estimation of posterior distribution. location of outlier in the dataset.
14.3.4 RVM for Classification

In a classification problem, the dependent variable ti is a binary variable
ti ∈{0,1} . The RVM framework can be extended to the classification problem parameter, σ kernel (Schölkopf and Smola, 2002). The kernel function can be
by applying a sigmoid function to the regression model [Eq. (14.1)] as: interpreted as a nonlinear transformation that maps the input space to a
higher-dimensional feature space. According to Cover’s theorem (Cover,
1 1965), a linear function can be formulated in the higher-dimensional feature
Ω( f ) = (14.10) space to seek a nonlinear relationship between inputs and outputs in the
1 + e− f original input space. Thus, the problem of choosing a kernel is equivalent to
finding an appropriate form of data representation for learning. Although
where, to keep the notation uncluttered, f ( xi , X ; w ) is represented by f . The there is no available general solution to this problem, many empirical and
likelihood of the dataset for this classification model is given by theoretical studies provide insights for specific applications (Schölkopf and
Smola, 2002; Cherkassky and Mulier, 1998). Tripathi and Govindaraju (2007)
N
suggested several techniques for practical selection of kernel parameters for
p ( t w ) = ∏ Ω( f )  1 − Ω( f ) 
ti 1−ti
(14.11)
i =1
hydrological applications.
Tipping (2001) proposed a method for estimating kernel parameters of
RVM by maximizing the likelihood function [Eq. (14.3)]. This can be
where t = [t1 ,…,t N ]. The parameters of the classification model are same as
achieved by taking the derivative of log-likelihood with respect to kernel
those for the regression model [Eq. (14.1)], except that there is no noise vari-
parameter σ kernel , and performing gradient-based local optimization [see
ance term σ 2ε . Model parameters are estimated as before, by first assigning
Tipping, 2001, p. 235 for details]. If the kernel parameter σ kernel becomes a
prior distributions and then estimating their posterior distributions using the
free variable to be determined during the learning process, the RVM [given
likelihood function [Eq. (14.11)] and applying Bayes’ rule. Unlike the regres-
by Eq. (14.1)] will become nonlinear in the unknown parameters. Estimating
sion model, the posterior distribution of w does not have a closed form and
parameters in such a context will need a rigorous nonlinear optimization
has to be approximated. The other aspects of the classification model formu-
algorithm as the likelihood surface can have multiple local maxima
lation remain the same as the regression model. For a new input vector, the
(Quinonero-Candela, 2004). In fact, Tipping (2001) cautions that maximizing
RVM model makes a probabilistic prediction for the category of the depen-
the likelihood is not guaranteed to yield optimum value of kernel parameter
dent variable. Additional details of the RVM model for classification can be
σ kernel in terms of model accuracy.
found in Bishop (2006, p. 353) and Tipping (2001, p. 218).
Figure 14.3 shows the value of rescaled log-likelihood and number of sta-
tistically relevant vectors (RVs) identified by the RVM model, as a function of
14.4 APPLICATION OF RVM
σ kernel , developed for the approximation of electrical conductivity values at
the Marana site, and Fig. 14.4 illustrates the electrical conductivity surface
The computer implementation of RVM is straightforward; however, its suc-
cessful and efficient application requires careful consideration of RVM’s
characteristics. This section discusses a few of those characteristics by using
the following two example datasets (Tripathi and Govindaraju, 2007). 200 200
Rescaled like hood
Number of relevant vectors
1. Marana data: Electrical conductivity measurements (surrogate for
hydrologic conductivity) at 129 locations over a 1000 m by 1000 m agricul-
tural field near Marana, Arizona (Fig. 14.2; Zhang, 1990).
150
2. Sinc data: A synthetic data generated from sinc function,
sinc( x ) = sin( x ) x , which has been a popular choice for studying kernel
methods (Tipping, 2001).
150 100
14.4.1 Presence of Outliers
A preliminary data analysis to identify outliers should be performed because
like other nonlinear statistical models, RVM results are sensitive to the pres-
ence of outliers. If outliers are detected, either they should be removed or 50
robust variants of RVM (Yang et al., 2007; Mitra et al., 2010) should be used.
The electrical conductivity dataset has one measurement that was unusually
large in comparison to other measurements and hence removed from further
analysis (Fig. 14.2). 100 0
100 101 102 103 104 105
14.4.2 Selection of Kernel Function
Figure 14.3 The value of rescaled log likelihood and number of vectors deemed rele-
It is widely acknowledged that a key factor that determines the generalization vant by the RVM model developed for the approximation of electrical conductivity val-
capabilities of kernel methods, including RVM, is the choice of kernel width ues at the Marana site with varying kernel width σ kernel .
14_Singh_ch14_p14.1-14.8.indd 3 8/22/16 11:43 AM

Electrical conductivity (mmho/cm)

1.6
1.4
1.2
1
0.8
1000
0.6
900
0.4 800
0.2 700
900 600
800
700 500 x (m)
600 400
y (m) 500
400 300
300
200 200
100
Figure 14.4 Performance of the RVM method for the approximation of electrical conductivity values at the Marana
site with kernel width σ kernel = 20 as obtained from maximizing the likelihood function. Out of a total of 128 vectors,
112 vectors (shown as solid circles) were considered statistically relevant by the RVM model. Clearly, RVM overfits
the data.
generated by RVM for the value of σ kernel corresponding to maximum likeli- 14.4.5 Choice of Inference Method
hood. The generated surface is very complex with many sharp peaks and The original RVM uses type II maximum likelihood approach that approxi-
valleys. This is an example of the RVM model overfitting the training data. mates posterior distributions by delta functions at their modes. This assump-
Further, it may be noted that 112 of 128 vectors were chosen as being relevant tion is justified when the number of data points ( N ) is very large because, in
in this RVM application. Thus, a complex model has been adopted with no that case, the posterior distribution is tightly concentrated around its mean.
advantage of sparsity suggesting that estimation of kernel parameter by However, in typical hydrologic applications, N tends to be small and this
maximizing the likelihood function alone may be undesirable. Figure 14.5a assumption may not hold. If N is small and computational resources are not
shows the fitted surface by RVM using the value of σ kernel obtained by follow- constrained, MCMC method is recommended. If computational resources are
ing the procedure given in Tripathi and Govindaraju (2007). The RVM pro- constrained, variational Bayes is recommended.
vides a reasonably smooth surface with only 11 relevant vectors. Figure 14.5b
shows the corresponding prediction variance surface that is quite revealing. 14.4.6 Relevant Vectors
The high variance estimates are associated with locations where the separa- The number of RVs identified by RVM is an indicator of complexity of the
tion between observed values and model prediction is large, as one would function—more the number of RVs more complex is the function learned by
expect intuitively. Similarly, the model variances tend to be larger for loca- RVM. Figure 14.4 shows a very complex electrical conductivity surface for
tions that have little or no data support. Marana site generated by RVM that selected 112 of 128 data points as RVs,
while Figure 14.5 shows a relatively smooth surface generated by RVM that
14.4.3 Extrapolation deemed only 11 of 128 data points as relevant. We recommend that number
The original RVM by Tipping (2001) using type II maximum likelihood of relevant vectors selected by RVM should be analyzed and reported. While
approach is notorious because its prediction becomes increasingly uncertain RVs have statistical significance in deciding the complexity of function
when extrapolating outside the domain of training data. Rasmussen and approximated by RVM, assigning physical significance to them requires cau-
Quinonero-Candela (2005) proposed an augmentation approach, which treats tion. Our experiments suggest that small perturbations in data do not sig-
test data point as a new relevant vector, to solve this problem, and termed it as nificantly change the number of RV, but the data points being selected as
RVM healing. Figure 14.6 compares the predictive uncertainty estimated from relevant may change. Nevertheless, RVs represent prototypical examples from
standard and augmented RVM. Evidently, the predictive standard deviation of which useful information about the physical process being modeled can be
standard RVM is smaller in extrapolation than in interpolation. Hence, when gleaned (Khalil et al., 2006).
extrapolation is inevitable, we recommend using augmented RVM.
14.4.4 Engaging Observational Uncertainty in RVM 14.5 EXAMPLES FROM HYDROLOGY AND FUTURE SCOPE
In many hydrological applications, the input vector x has large measure RVMs have been successfully applied in many water resources problems.
ment uncertainties; however, the standard implementation of RVM ignores Some of the earlier applications included geospatial interpolation
these, assuming inputs as deterministic. Quinonero-Candela (2004) modified (Govindaraju, 2005), groundwater quality modeling (Khalil et al., 2005a),
the RVM algorithm to engage input uncertainties during prediction phase real-time management of reservoir releases (Khalil et al., 2005b), modeling of
and demonstrated its advantages for multistep forecast. Tripathi and chaotic hydrologic time series (Khalil et al., 2006), and studying the impact of
Govndaraju (2012) proposed VNRVM to engage uncertainties in input data climate change on regional hydrology (Ghosh and Mujumdar, 2008). It has
during both learning and prediction phases. Figure 14.1b shows the sche- also found applications in precipitation nowcasting (Barillec and Cornford,
matic of VNRVM and Fig. 14.7 shows the performance of VNRVM in learn- 2009), sediment transportation (Dogan et al., 2009), and forecasting of hydro-
ing sinc function when the input data have known uncertainties. It is evident climatic variables (Tripathi and Govindaraju, 2012; Samui and Dixon, 2012;
that the function learned by the standard RVM is overly smooth whereas that Joshi et al., 2013; Srivastava et al., 2013; Bachour et al., 2014; Bai et al., 2014).
by VNRVM is closer to the true function. Results from our other experiments RVMs provide a promising alternative to many statistical hydrologic prob-
show that engaging uncertainties in input data provides not only better pre- lems because of their excellent generalization properties. In addition, they
diction but also a more reliable estimate of predictive uncertainty; hence we have the added advantage of probabilistic interpretation that yields prediction
recommend that uncertainties in input data should be used in developing uncertainty. However, several avenues should be explored to make this tech-
RVM models. nique better suited to a wide variety of hydrologic data. In particular,
14_Singh_ch14_p14.1-14.8.indd 4 8/22/16 11:43 AM

Miscellaneous Topics 14-5
(a)
0.6
Electrical conductivity (mmho/cm)

0.4
0.2
0.1
0.8
0.6
0.4
0.2
800 1000
900
600 800
700
400 600
y (m) 500
400
200 300 x (m)
100 200
(b)
0.052
0.06
Variance of conductivity
0.048
(mmho2/cm2)
0.046
0.044
0.042
0.04
0.038
0.036
800 1000
900
600 800
700
400 600
500
y (m)
200 300 400 x (m)
200
100
Figure 14.5 Performance of the RVM method for the approximation of electrical conductivity values at the Marana site
with kernel width σ kernel = 180 . The top panel shows the RVM-fitted surface (mesh) along with measurements in circles.
Out of a total of 128 vectors, 11 vectors (shown as solid circles) were considered statistically relevant by the RVM model.
The bottom panel shows the prediction variance surface of the RVM estimates.
incorporating and representing physics of the hydrologic system into RVM 14.6.2 Connection with Other Models
learning, modifying the RVM formulation to handle flexible priors, and mak- RVMs share common features with many statistical models. They are
ing the algorithm robust against outliers will make RVMs more amenable to designed to emulate sparseness of SVM under Bayesian framework.
hydrologic problems. Quinonero-Candela (2004) has shown that RVM is a special case of a
Gaussian process, commonly referred to as Kriging in hydrological literature,
14.6 MISCELLANEOUS TOPICS with the covariance governed by the kernel function. RVM can be casted as a
DAG and treated as sparse probabilistic graphical models. Further, it is related
14.6.1 Computational Complexity to classic Bayesian linear regression in feature space with ARD priors. These
Training of RVM using type II maximum likelihood requires inversion of linkages of RVM with other models provide insights into functioning of RVM
matrix of size N × N , where N is the number of data points, that requires that can be used to further improve and adapt RVM for hydrological
O( N 3 ) computations. Bishop (2006, p. 352) has proposed a sequential learn- applications.
ing algorithm for RVM that scales like O( M 3 ) where M is the number of
relevant vectors, which is typically much smaller than N . Variational and 14.6.3 Computer Implementation
MCMC inferences of RVM are computationally much more expensive and The algorithm of RVM based on type II maximum likelihood approach can
are recommended when N is relatively small. be readily programmed. Software packages for computer implementation of
14_Singh_ch14_p14.1-14.8.indd 5 8/22/16 11:43 AM

Figure 14.6 RVM approximation of sinc function along with ±1 predictive standard deviation by standard
RVM and augmented RVM. The prediction beyond the training data range (–10,10) represents extrapolation
by RVM.
Figure 14.7 Learning sinc function from hundred samples of noisy data (plus symbol). A standard RVM
algorithm (RVM) learns an oversmooth function (dotted line with cross marker), whereas the VNRVM
estimate (dotted line) closely matches the true function.
RVMs are also freely available to help users test RVM models for their prob- Bai, Y., P. Wang, C. Li, J. Xie, and Y. Wang, “A multi-scale relevance vector
lems. Some popular implementation tools for RVMs are: (a) SparseBayes regression approach for daily urban water demand forecasting,” Journal of
software developed by Tipping and Faul (2003) for Matlab, (b) smooth RVM Hydrology, 517: 236–245, 2014, doi:10.1016/j.jhydrol.2014.05.033.
for Matlab by Schmolck and Everson (2007), (c) kernlab for R by Karatzoglou Barillec, R. and D. Cornford, “Data assimilation for precipitation nowcast-
et al. (2004), and (d) Dlib-ml for C++ by King (2009). ing using Bayesian inference,” Advances in Water Resources, 32 (7):
1050–1065, 2009, doi:10.1016/jadvwatres.2008.09.004.
Bishop, C. M., Pattern Recognition and Machine Learning, Information
REFERENCES
Science and Statistics, 1st ed., Springer, New York, 2006.
Bachour, R., W. Walker, A. Ticlavilca, M. McKee, and I. Maslova, “Estima- Bishop, C. M. and M. E. Tipping, “Variational relevance vector machines,”
tion of spatially distributed evapotranspiration using remote sensing and a Proceedings of the 16th Conference on Uncertainty in Artificial Intelligence,
relevance vector machine,” Journal of Irrigation and Drainage Engineering, 140 edited by C. Boutilier and M. Goldszmidt, Morgan Kaufmann, San Francisco,
(8): 04014029, 2014, doi:10.1061/(ASCE)IR.1943-4774.0000754. CA, 2000, pp. 46–53.
14_Singh_ch14_p14.1-14.8.indd 6 8/22/16 11:43 AM

REFERENCES 14-7
Cherkassky, V. and F. Mulier, Learning from Data: Concepts, Theory, and Vision and Pattern Recognition (CVPR), San Francisco, CA, USA, 2010,
Methods, Wiley Interscience, New York, 1998. pp. 1887–1894, doi:10.1109/CVPR.2010.5539861.
Cover, T. M., “Geometrical and statistical properties of systems of linear Quinonero-Candela, J., Learning with uncertainty—Gaussian processes
inequalities with applications in pattern recognition,” IEEE Transactions on and relevance vector machines, Ph.D. dissertation, Technical University of
Electronic Computers, EC-14 (3): 326–334, 1965. Denmark, Kongens Lyngby, 2004.
Dogan, E., S. Tripathi, D. A. Lyn, and R. S. Govindaraju, “From flumes to Rasmussen, C. E. and J. Quinonero-Candela, “Healing the relevance vector
rivers: can sediment transport in natural alluvial channels be predicted from machine through augmentation,” Proceedings of the 22nd International
observations at the laboratory scale?,” Water Resources Research, 45: W08433, Conference on Machine Learning, Bonn, Germany, 2005, pp. 689–696. ACM
2009, doi:10.1029/2008WR007637. Press, New York, NY, USA.
Ghosh, S. and P. P. Mujumdar, “Statistical downscaling of GCM simulations Samui, P. and B. Dixon, “Application of support vector machine and rele-
to streamflow using relevance vector machine,” Advances in Water Resources, vance vector machine to determine evaporative losses in reservoirs,”
31 (1): 132–146, 2008. Hydrological Processes, 26 (9): 1361–1369, 2012, doi:10.1002/hyp.8278.
Govindaraju, R. S., “Bayesian learning and relevance vector machines for Schmolck, A. and R. Everson, “Smooth relevance vector machine: a
hydrologic applications,” Second Indian International Conference on Artificial smoothness prior extension of the RVM,” Machine Learning, 68 (2): 107–135,
Intelligence, Pune, India, December 20–22, 2005, pp. 1078–1093. 2007.
Joshi, D., A. St-Hilaire, A. Daigle, and T. B. M. J. Ouarda, “Databased com- Schölkopf, B. and A. J. Smola, Learning with kernels: support vector
parison of sparse Bayesian learning and multiple linear regression for statisti- machines, regularization, optimization, and beyond, Adaptive Computation
cal downscaling of low flow indices,” Journal of Hydrology, 488: 136–149, and Machine Learning, 1st ed., MIT Press, Cambridge, MA, 2202.
2013, doi:10.1016/j.jhydrol.2013.02.040. Srivastava, P. K., D. Han, M. R. Ramirez, and T. Islam, “Machine learning
Juncai, H., W. Fengbi, M. Huanzhang, and Z. Mingtian, “A Markov chain techniques for downscaling SMOS satellite soil moisture using MODIS land
Monte Carlo sampling relevance vector machine model for recognizing tran- surface temperature for hydrological application,” Water Resources Manage-
scription start sites,” Proceedings of the International Conference on Artificial ment, 27 (8): 3127–3144, 2013.
Intelligence and Computational Intelligence, Sanya, China, 2010, pp. 185–188, Tipping, M. E., “Sparse Bayesian learning and the relevance vector
doi:10.1109/AICI.2010.277. machine,” Journal of Machine Learning Research, 1 (3): 211–244, 2001,
Karatzoglou, A., A. Smola, K. Hornik, and A. Zeileis, “Kernlab—an S4 doi:10.1162/15324430152748236.
package for kernel methods in R,” Journal of Statistical Software, 11 (9): 1–20, Tipping, M. E. and A. Faul, “Fast marginal likelihood maximization for
2004. sparse Bayesian models,” Proceedings of the Ninth International Workshop on
Khalil, A., M. N. Almasri, M. McKee, and J. J. Kaluarachchi, “Applicability Artificial Intelligence and Statistics, edited by C. M. Bishop and B. J. Frey, Key
of statistical learning algorithms in groundwater quality modeling,” Water West, FL, 2003. Society for Artificial Intelligence and Statistics, New Jersey,
Resources Research, 41 (5): W05010, 2005a, doi:10.1029/2004WR003608. USA.
Khalil, A., M. McKee, M. Kemblowski, and T. Asefa, “Sparse Bayesian Tripathi, S. and R. S. Govindaraju, “On selection of kernel parameters in
learning machine for real-time management of reservoir releases,” Water relevance vector machines for hydrologic applications,” Stochastic Environ-
Resources Research, 41 (11): W11401, 2005b, doi:10.1029/2004WR003891. mental Research and Risk Assessment, 21 (6): 747–764, 2007, doi:10.1007/
Khalil, A. F., M. McKee, M. Kemblowski, T. Asefa, and L. Bastidas, “Multi- s00477-006-0087-9.
objective analysis of chaotic dynamic systems with sparse learning machines,” Tripathi, S. and R. S. Govindaraju, “Appraisal of statistical predictability
Advances in Water Resources, 29 (1): 72–88, 2006, doi:10.1016/j.advwa- under uncertain inputs: SST to rainfall,” ASCE Journal of Hydrologic Engineer-
tres.2005.05.011. ing, 16 (6): 970–983, 2011.
King, D. E., “Dlib-ml: a machine learning toolkit,” Journal of Machine Tripathi, S., V. V. Srinivas, and R. S. Nanjundiah, “Downscaling of precipi-
Learning Research, 10, 1755–1758, 2009. tation for climate change scenarios: a support vector machine approach,”
MacKay, D. J. C., “Bayesian methods for backpropagation networks,” Journal of Hydrology, 330 (3–4): 621–640, 2006.
Models of Neural Networks III, edited by E. Domany, J. L. van Hemmen, and Vapnik, V. N., Statistical Learning Theory, John Wiley & Sons, New York,
K. Schulten, Springer, New York, 1996, pp. 211–254. 1998.
MacKay, D. J. C., Information Theory, Inference and Learning Algorithms, Yang, B., Z. Zhang, and Z. Sun, “Robust relevance vector regression with
Version 7.2, Cambridge University Press, 2005, ebook: http://www.inference. trimmed likelihood function,” IEEE Signal Processing Letters, 14 (10):
phy.cam.ac.uk/mackay/itila/book.html. Accessed on February 09, 2016. 746–749, 2007.
Mitra, K., A. Veeraraghavan, and R. Chellappa, “Robust RVM regression Zhang, R., Soil variability and geostatistical applications, Ph.D. disserta-
using sparse outlier model,” Proceedings of the IEEE Conference on Computer tion, The University of Arizona, Tucson, 1990.
14_Singh_ch14_p14.1-14.8.indd 7 8/22/16 11:43 AM

Chapter 15
Harmonic Analysis and Wavelets
BY
D. LABAT
ABSTRACT In order to provide a temporal localization of the frequency content of a

signal, it is necessary to introduce orthogonal bases that do not correspond to
This chapter emphasizes the different aspects of the implementations of the an infinite interval but are restricted on a temporal interval and equal to zero
continuous and orthogonal wavelet analyses in hydrology. These new tech- outside this interval.
niques appear as particularly well suited for applications in hydrology and The first option proposed consists of the windowed Fourier transform or
hydrogeology. Most of the signals encountered in hydrology are characterized Gabor methods (Gabor, 1946), a time-frequency transform that allows for
by intermittent or transient features that lead to nonstationary data series, temporal discrimination of the frequency but with an arbitrary function of
such as temperature, rainfall, or river flow time series (Labat, 2005; Sang, constant width in terms of both time and frequency. The windowed Fourier
2013). When time series appear as ergodic and stationary, Fourier analyses transform is then defined by:
lead to interesting results especially in term of power law statistical behavior
on high frequency bands. However, classical Fourier algorithms only deliver
 t −t 
limited information since they do not provide any temporal localization of the y (t1 , t 2 , ω ) = exp(i 2πωt ) ⋅Φ  1 2 
frequency or periodicity. For example, when dealing with long term period-  σ  (15.3)
icities, it is important to decipher a continuous oscillation on the entire time +∞
interval from an intermittent process or to decipher a signal with a time- ⇒ SG(ω , t1 − t 2 ) = ∫ x (t ) y * (t1 , t 2 , ω )dt
−∞
varying periodicity (such a chirp for example) from two oscillations. The
main advantage of harmonic and wavelet analysis is that the window size is
not fixed but varies with the frequency following the famous Heisenberg ( )
t −t
Where Φ 1 σ 2 is the windowed Fourier function different from zero only
around t2 on a time domain [–w/2:w/2]. A commonly used function is a
incertitude theorem.
Gaussian of width s defined by:
The basic objective of harmonic and wavelet analyses is to achieve a com-
plete time-scale representation of localized and transient phenomena occur-
ring at different time scales.  t −t   (t − t )2 
Φ  1 2  = exp  −i ⋅ σ 2 1 2  (15.4)
 σ   2 
15.1 INTRODUCTION
The short-time (windowed) Fourier and Gabor transforms as still limited
Analyzing a spatial or temporal signal basically consists of looking for simi- by the nonadaptation of the size of the window to the scale of the process.
larity between the signal and well-known mathematical functions usually Therefore, we encourage to use wavelet transforms in order to overcome this
underlined by a clear physical meaning. We will restrict here our approach to issue.
temporal signals function of time t. In mathematics, the degree of similarity
between two square continuous-time integrable signals x(t) and y(t) can be
indicated using the classical scalar product IXY defined by: 15.2 THE CONTINUOUS WAVELET TRANSFORM
The wavelet transform constitutes another type of projection that constitutes

+∞
I XY = ∫ x (t ) y (t )dt (15.1) a more valuable analysis when dealing with signals encountered in the Earth
−∞
sciences as first demonstrated by Grossman and Morlet (1984).
If, I(x,y) is equal to zero, the two functions are said to be orthogonal meaning The first step consists of changing the projection basis in order to provide
that there are no similarity between them. If, I(x, y) is equal to one, the two a time-scale dependent window of projection that overcomes the issue of the
functions are similar. constant width-interval present in classical time frequency projections. The
From this point of view, the Fourier transform can be viewed as a projec- classical orthonormal complex Fourier basis {exp(iwnt), n∈Z}, where wn is a
tion of the temporal domain into the frequency domain (w) using the set of discrete frequency and (wn = 2pn/T) is replaced by a two-parameter basis
complex functions: {ya,t(t), (a,t) ∈(R*+ × R)}, which allows for a time-scale discrimination of
processes. Therefore, the coefficients of the wavelet transform of a square-
+∞ integrable continuous-time signal x(t) are defined by the linear integral
y (t , ω ) = exp(i 2πωt ) ⇒ S(ω ) = ∫ x (t ) y * (t , ω )dt (15.2) operator (corresponding to the classical L2 inner product):
−∞
where ∗ stands for the complex conjugate. +∞ 1  t − τ

The scalar product noted S(w) is the spectral amplitude of the signal x(t) at C x (a , τ ) = ∫ x (t )ψ ∗a ,τ (t )dt with ψ a ,τ (t ) = ψ  (15.5)
−∞ a  a 
the frequency w and allows to determine if a specific periodicity is present or
not in the signal. The reader will find comprehensive reviews of spectral where ∗ corresponds to the complex conjugate. The function y(t), which can be
analysis in Walker (1997), Papoulis (1964), Priestley (1981), and Tary et al. real or complex, plays the role of a convolution-kernel, and is called a wavelet.
(2014) among others. The parameter a can be interpreted as a dilation (a > 1) or contraction (a < 1)
15-1
15_Singh_ch15_p15.1-15.8.indd 1 8/22/16 11:45 AM

15-2 Harmonic Analysis and Wavelets
(a)
1000
800
Annual runoff m3/s

600
400
200
0
1820 1840 1860 1880 1900 1920 1940 1960 1980 2000
Years
(b)
4
8
Period (years)
16
32
1820 1840 1860 1880 1900 1920 1940 1960 1980 2000
Years
(c) x 104
3.5
Global wavelet spectrum
3
2.5
2
1.5
1
0.5
0
128 64 32 16 8 4
Period (years)
Figure 15.1 Morlet wavelet analysis of the mean annual runoff of the Vannern Gota river. From top
to bottom: (a) temporal fluctuations of annual runoff, (b) two-dimensional Morlet wavelet spectrum, (c)
global Morlet wavelet spectrum. In the third plot, the dashed line correspond to the 90% statistical con-
fidence interval from red noise.
factor of the wavelet function y(t), corresponding to different scales of observa- The wavelet function must fulfill some strict mathematical conditions (Heil
tion. The parameter t can be interpreted as a temporal translation or shift of the and Walnut, 1989; Jawerth and Sweldens, 1994). For example, the time-scale
function y(t), which allows for a study of the signal x(t) locally around the time localization property requires that wavelet functions be characterized, in both
t (Fig. 15.1). A frequency interpretation of Eq. 15.5 is also of interest. Effectively, time and frequency domains, by compact supports or, at least, by a suffi-
using the Parseval theorem, the wavelet coefficients of a continuous time-signal ciently fast decay. Mathematically, these properties can be formalized by two
x(t) can also be obtained via the relation: conditions called “admissibility conditions.” As mentioned earlier, one of
these conditions is a rapid decrease of y(t) around the origin of time, imply-
1 +∞ 1 +∞ ing that:
C x (a , τ ) =
2π ∫−∞
ˆ (aω )e iωτ dω =
xˆ(ω ) aψ
2π ∫−∞
xˆ(ω )Ψ a ,τ (ω )dω (15.6)
+∞
∫−∞ ψ(t )dt = 0 (15.8)
This formulation indicates that the wavelet coefficients correspond to the
filtering of x̂ [Fourier transform of x(t)] by a set of band-pass filters Ya,t In addition to a normalization of the wavelet function (the wavelet function
related to the Fourier transform of the wavelet function y(t) [denoted by must satisfy ψ = 1, where ⋅ is the classical L2-norm), more restrictive
ψˆ (ω )] via the relation: conditions must be imposed on the wavelet function in order to obtain a
reconstruction formula. More precisely, the continuous-time signal x(t) can
ˆ (aω )e iωτ (15.7)
Ψ a ,τ (ω ) = a ψ be retrieved from its wavelet decomposition coefficients Cx(a,t) only if the
wavelet function has N vanishing moments, that is,
From this point of view, the parameter a can be physically interpreted as a
dilation or contraction of the filter (corresponding to different nonoverlapping +∞ k
frequency intervals), and the parameter t can be interpreted as a phase shift.

∫−∞ t ψ (t )dt = 0, k = 1,, N − 1 (15.9)
15_Singh_ch15_p15.1-15.8.indd 2 8/22/16 11:45 AM

SIGNAL ENERGY REPARTITION IN THE WAVELET FRAME 15-3
the condition (15.11), is then said to have regularity of order N. These condi- In such an orthonormal basis, all wavelets y are orthonormal to both their
tions are called “regularity conditions.” There is a very large choice in wavelet translates and their dilates. Therefore, it is possible to construct a complete
functions. The simplest wavelet function is the Haar function, which is used basis under which all square integrable signals x(i) can be expanded as a
for wavelet spectrum estimation in this chapter (presented later). It is linear combination of translates and dilates of orthonormal wavelets:
defined by:
+∞ +∞
1 for 0 < x < 1 / 2 x (i ) = ∑ ∑ C xj ,k (i )ψ j ,k (i ) (15.15)

 j =0 k =−∞
ψ ( x ) = −1 for 1 / 2 ≤ x < 1 (15.10)
0 elsewhere Equation (15.15) is the discrete-time formulation of signal synthesis; it can

also be interpreted as successive approximations (in a mean square sense) of
Then, several other wavelets such as Daubechies or Symlets (Daubechies, the analyzed discrete signal x(i) by the sequence {xn(i)}n=1,..,N defined by:
1990, 1992) are also of practical use when dealing with hydrological signals.
Conditioned by these restrictions, there exists a reconstruction formula, n−1 +∞
which allows the synthesis of the continuous time signal x(t) through its xn (i ) = ∑ ∑ C xj ,k (i )ψ j ,k (i ) (15.16)
j =0 k =−∞
wavelet coefficients CX(a,t):
1 +∞ +∞ dadτ This constitutes the conceptual basis of multiresolution analysis.

x (t ) =
Kψ ∫0 ∫−∞ Cx (a, τ)ψ a ,τ (t ) a2
(15.11) Multiresolution wavelet analysis (Mallat, 1989) allows the decomposition of
a function or signal in a progression of successive “approximations” and
“smooths,” corresponding to different scales j. The difference between the
Moreover, another property of the continuous wavelet transform is its ability actual signal and its approximation of order n is called the “residual.”
to conserve energy between time domain and time-scale domain, that is, Intuitively, the approximation is relatively smooth and the detail, being com-
posed of high frequency components, is rougher. Note that the detail corre-
+∞ 1 +∞ +∞ 2 dadτ sponds to the difference between two successive levels of approximation j
∫−∞ x ∫0 ∫−∞ Cx (a, τ)
2
(t )dt = (15.12)
Kψ a2 and j + 1.
To sum up, the multiresolution analysis allows an orthogonal decompo-
One disadvantage of these nonorthogonal wavelets is the variance repartition sition of a sampled hydrologic signal in terms of approximations and details
of the signal in the continuous wavelet framework is characterized by a of increasing order of resolution. Because of the orthogonality of the decom-
redundancy of information among the wavelet coefficients. This redundancy position, it should be noted that the multiresolution induces a more efficient
is mainly due to two properties. First, the Morlet wavelet and all other and easier interpretation of energy distribution on the different scales of the
mother wavelet used in continuous wavelet transform do not strictly have a decomposition.
fully compact support although it decays to zero. Second, redundancy also
comes from continuous shift in time and scale of the wavelet during the con- 15.4 SIGNAL ENERGY REPARTITION IN THE
tinuous transform. Then, the correlation between coefficients appears as WAVELET FRAME
intrinsic to the wavelet-kernel and not to the analyzed signal. The wavelet
The repartition of energy in the signal allows for the determination of the
coefficient interpretation is then strongly dependent on the chosen projection
scales, which concentrate the essential dynamics of a signal. Therefore, a
basis, which must be adapted to the specific problem at hand. For example,
determination of the temporal variations of the distribution of energy across
the Mexican Hat wavelet allows for a good temporal resolution but a poor
the scales is an important application of the wavelet transform.
frequency resolution, whereas the Morlet wavelet allows for a good frequency
Previously, the Fourier transform allowed for the determination of the
resolution although with a less precise temporal resolution (Labat, 2000).
distribution of energy across frequencies w using the Power spectrum density
PXX(w) defined for a signal x(t) by:
15.3 DISCRETE TIME WAVELET TRANSFORM AND
2
MULTIRESOLUTION ANALYSIS 1 T
T ∫0
PXX (ω ) = x (t )2 exp(−i 2πωt )dt (15.17)
The continuous wavelet analysis is essentially a signal analysis tool but
because of the redundancy issue, it cannot be considered as an energy-
repartition tool. Therefore, discrete time wavelet transform is now presented. The companion concept of wavelet energy was defined first by Hudgins
For practical applications, the scientist or engineer does not have at his or her (1993) and then by Brunet and Collineau (1995). Other results concerning
disposal a continuous-time signal process but rather only a discrete-time wavelet variance estimators and applications can be found in Percival (1995)
signal, which is denoted herein by x(i). Then, the time-scale domain needs to or Percival and Mofjeld (1997). This concept is based on the conservation of
be discretized and one can choose the general form ({a0-j, k.t0.a0-j}, (k,j) ∈ Z2), energy between the time domain and the time-scale domain, that is,
where a0 (a0 > 1) and t0 are constants.
The discretized version of Eq. (15.5) defines the discrete wavelet transform +∞ 1 +∞ +∞ 2 dadτ
∫−∞ x ∫0 ∫−∞ Cx (a, τ)
2
(t )dt = (15.18)
coefficients, as follows: Kψ a2
+∞
Cx ( j , k ) = ∫ x (t )ψ ∗ j ,k (t )dt with ψ j ,k (t ) = a0 j 2 ψ (a0 jt − kτ0 ) (15.13) Therefore, the total signal variance can be distributed in the wavelet domain
−∞ and the wavelet spectrum WX(a,t) of a continuous-time signal x(t), which can
be defined, by analogy with the Fourier analysis, as the modulus of its wavelet
where ∗ corresponds to the complex conjugate. The integer j is the scale factor coefficients (Liu, 1995):
(analogous to the parameter a) and the integer kt0 is the translation factor
(analogous to the parameter t). As in the continuous-time wavelet transform,
2
WX (a, τ ) = C X (a, τ )C ∗ X (a, τ ) = C X (a, τ ) (15.19)
the reconstruction formula can be applied with appropriate choice of the
coefficients a0 and t0. This allows for representation of a discrete-time signal
This wavelet spectrum can also be averaged in time or in scale, but at the cost
process in terms of series expansion and leads ultimately to the concept of
of a loss of information (Torrence and Compo, 1998). On the one hand, time
multiresolution analysis.
averaging gives the variance signal distribution between the different scales,
Once the time-scale domain has been discretized, some particular values
referred to as the global averaged wavelet power spectrum. On the other
for a0 and t0, corresponding to an octave representation (Daubechies, 1990,
hand, scale averaging allows the temporal identification of a particular com-
1992) of the scales, are chosen, namely: a0 = 2 and t0 = 1. The continuous grid
ponent (i.e., the temporal variability at a given scale or range of scales) of the
is then replaced by a discrete dyadic grid of the form: ({2-j, k.2j}, (k,j)∈Z2). The
signal.
orthogonal discrete wavelet transform coefficients Cj,k are then defined by the
The determination of the characteristic periods of oscillation is achieved
convolution product:
using the global wavelet spectrum. Statistical confidence limits of the peaks
+∞ identified in the wavelet power spectrum have been discussed by Torrence
C xj ,k = ∫ x (t )ψ ∗ j ,k (t )dt with ψ j ,k (t ) = 2 j 2 ψ (2 j t − k ) (15.14) and Compo (1998), Ge (2007), Liu et al. (2007), and Schaefli et al. (2007). Two
−∞
15_Singh_ch15_p15.1-15.8.indd 3 8/22/16 11:45 AM

limits are classically considered in order to restrict analysis to statistically However, real-part or imaginary-part wavelet analysis does not provide a
significant-only fluctuations: clearly separate information. As in the Fourier analysis, only modulus and
• Only periodicity different from red noise is taken into consideration and arguments are considered here as interpretative. This conclusion is confirmed
95% and 90% confidence intervals are generally depicted in the figures, by Mizuno-Matsumoto et al. (2001, 2002), who observed, in a preliminary
following Torrence and Campo (1998) calculations. analysis, no significant differences between the definitions. Wavelet cross-
• Edge effects [and then limitations of the integration time interval in correlation is thus defined as:
Eq. (15.19)], lead to the presence of a “cone of influence.” This corre-
sponds to the region where variance cannot be estimated properly
2 2
because of edge effects. That leads to an under-estimation of variance as RWC XY (a, τ ) + IWC XY (a, τ )
a result of zero-padding of the original time series (Torrence and WR XY (a, τ ) = (15.25)
Compo, 1998). Due to the edge effects leading to the presence of a cone WC XX (a,0) ⋅ WCYY (a,0)
of influence, if T is the total length of the series, only periods smaller
than T/(2√2) are considered as significant. where RWCXY and IWCXY are, respectively, the real and the imaginary part of
According to Meneveau (1991), the concept of wavelet variance can be also the cross-wavelet correlation function.
defined by using the orthogonal multiresolution approach. The total energy In conclusion, the introduction of wavelet cross-correlation allows for a
(variance) TX(m) captured by the detail Dm at scale m of a signal x(i) of length rapid identification of the degree of correlation between processes at a given
2M can be estimated as: scale and the determination of the delay between these processes.
2 M −m
1
∑ ( Dmx , j )
2 15.6 WAVELET CROSS SPECTRUM AND COHERENCE
Tx (m) = (15.20)
2 M −m j =1 By analogy with the Fourier cross-spectrum, Liu (1995) first defines the wave-
let cross-spectrum WXY(a,t) between two signals x(t) and y(t) by:
Note that these concepts were first implemented to study intermittent turbu-
lent signal above the forest canopy (Katul et al., 1994; Turner et al., 1994;
WXY (a, τ ) = C X (a, τ )C ∗Y (a, τ ) (15.26)
Katul and Parlange, 1995). One of the main advantages of the discrete
orthogonal wavelet transform is its ability to concentrate the energy of the
signal in a limited number of coefficients. where CX(a,t) and C∗Y(a,t) are, respectively, the wavelet coefficient of the
continuous-time signal x(t) and the conjugate of the wavelet coefficient of
y(t). The averaging techniques applied to the wavelet spectrum are also used
15.5 WAVELET ANALYSIS OF THE TIME–SCALE here for expressing the cross-covariance of the signals x and y and its distribu-
RELATIONSHIP BETWEEN TWO SIGNALS tion through scales. However, Maraun and Kurths (2004), among others,
In this section, we focus on the introduction of wavelet indicators to investi- demonstrated that wavelet cross-spectrum appears to be nonsuitable to inter-
gate the relationship between two given signals (Labat et al., 2000, 2002), for pret interrelation between two processes (ENSO and NAO fluctuations in
example, rainfall rates/discharges or discharge/climate circulation indices. their example) and recommend the introduction and use of wavelet coher-
Classical cross-correlation makes it possible to quantify the degree of simi- ence analysis. Effectively, cross wavelet analyses only indicate the presence of
larity between two signals, denoted as. The cross-correlation RXY between two common periodicities between two signals but no causal relationship is put in
stochastic processes x(t) and y(t) is defined as: evidence.
The notion of coherence in signal processing consists, from a general point
of view, of a measure of the correlation between two signals or between two
C XY (t1 − t 2 ) representations of these signals. To overcome the problems inherent to non-
RXY (t1 − t 2 ) = with CXY(t1 – t2)=E[x(t – t1)y(t – t2)] (15.21)
C XX (0) ⋅ CYY (0) stationary signals, it has been proposed to introduce the wavelet coherence
(Torrence and Webster, 1999; Lachaux et al., 2002). The main issue is that an
application of a similar formula to Eq. (15.25) leads to a coherence always
that indicates that cross-correlation analysis provides a quantitative measure equal to 1, as pointed out by Labat et al. (1999a) and Labat (2005). Attempts
of the relatedness of two signals, shifting in time with respect to each other. It that have been made to avoid this problem essentially differ on two points: the
allows for the identification of common components occurring at the same fundamental definition of wavelet coherence and the estimation of the wave-
delay. However, if the signals include nonstationary components, the cross- let cross-spectrum.
correlation becomes invalid. The same conclusion arises if both signals are Torrence and Webster (1999) propose to determine wavelet coherence
characterized by highly variable processes occurring over a wide range of using a smoothed estimate of the wavelet spectrum and to define a smooth
scales. The continuous wavelet cross-correlation is introduced, which over- wavelet spectrum and cross spectrum, denoted as SWXX(a,t) and SWXY(a,t),
comes the limitations of classical correlations. respectively:
Considering two correlated processes x(t) and y(t), the wavelet cross-
correlation function, denoted as WCXY(a,u), for a given scale a and a given
t +δ /2
time delay u can be defined as: SWXX (a, τ ) = ∫ *
WXX (a, τ )WXX (a, τ )dadτ (15.27)
t −δ /2
WC XY (a,u) = E[WXX (a, τ ) ⋅WYY (a, τ + u)] (15.22) t +δ /2

SWXY (a, τ ) = ∫ *
WXX (a, τ )WYY (a, τ )dadτ (15.28)
t −δ /2
At this point, some authors (see later) recommend distinguishing the real part
noted as RWCXY (a,t) from the imaginary part IWCXY (a,t) of the wavelet The scalar d represents the size of the two-dimensional filter (Lachaux et al.,
cross-correlation, arguing that the real part alone is sufficient to quantify the 2002). It constitutes an important parameter of the wavelet coherence and
strength of correlation between the two signals at a given scale a. The wavelet must be adequately determined for an acceptable estimation of the wavelet
cross-correlation is then defined by: coherence. We therefore recommend to the reader to properly apply the
Torrence and Webster (1999) indications. The wavelet coherence can then be
RWC XY (a,u) defined by analogy with Fourier coherence as:
WR XY (a,u) = (15.23)
RWC XX (a,0) ⋅ RWCYY (a,0)
SWXY (a, τ )
WC(a, τ ) = (15.29)
Sello and Bellazzini (2000) propose to consider only the real part of the wavelet  SWXX (a, τ ) ⋅ SWYY (a, τ ) 
transform and define a wavelet local correlation coefficient WLCC(a,u) based  
*
on the wavelet cross-spectrum, WXY (a,u) = WXX (a,u) ⋅WYY (a,u), as:
Schwarz inequality still ensures that WC takes a value between 0 and 1.
Statistical confidence limits of wavelet coherency based on red-noise back-
Real(WXY (a,u))
WLCC(a,u) = (15.24) ground are discussed in Grinsted et al. (2004), Maraun and Kurths (2004),
WXX (a,u) ⋅ WYY (a,u) and Schaefli et al. (2007).
15_Singh_ch15_p15.1-15.8.indd 4 8/22/16 11:45 AM

REFERENCES 15-5
15.7 APPLICATIONS OF WAVELET TRANSFORMS IN neural networks (ANNs) coupled to wavelet models, the so called hybrid
HYDROLOGY AND EARTH SCIENCES wavelet-ANN models and derivatives.
Wavelet analysis has become more and more popular in the Earth sciences,
especially in order to identify coherent structures in time series. In fact, there 15.8 PERSPECTIVES
is a good chance that wavelets will be used by scientists from all different
fields dealing with time series that clearly exhibit transient processes. This In conclusion, wavelet transforms appear as a powerful tool to identify transient
popularity is possible largely because of the existence of free wavelet programs processes in hydrological time series whereas multiresolution analyses for now
(especially in MATLAB) made available through the Web [see, for example, appear as restricted to isolate a given component corresponding to a given scale.
www.amara.com/current/wavelet.html, atoc.colorado.edu/research/wavelets, Wavelet cross-correlation and wavelet coherence are also now being extensively
or Uvi wave toolbox (Gonzales et al., 1996) for univariate analyses, and used to provide more detailed information about the time–scale relationship
www.pol.ac.uk/home/research/waveletcoherency/ for wavelet bivariate analy- between two signals. In these ways, wavelet analysis provides a better investigating
ses). Table 15.1 provides a brief review of wavelet applications in hydrology. tool than classical Fourier analysis, and thus deserves far more attention.
In hydrology, wavelets can, at the fundamental level, be used to detect tran- However, it must be kept in mind that wavelets cannot be considered as the
sient or periodic processes in relevant time series. But as mentioned earlier, “Grail,” since the interpretation of the results from the wavelet analysis must
continuous wavelet transform is also well suited for identification of interan- always rely on physical processes. Wavelet research in hydrology should be
nual variability in hydrologic and meteorologic time series. Some applications dedicated to the development of wavelet-based models coupled to neural
of wavelets in hydrology also concern much longer periods, ranging from a networks or Volterra expansion taking into account the intrinsic multiscale
century to several million years. Multiresolution concepts can also be used as and nonlinear nature of many physical relationships. Wavelets will become a
a stochastic tool to generate synthetic time series and recent developments in more complete and better simulation tool for the study of multiscale proper-
wavelet hydrologic modeling are also based on the introduction of artificial ties of hydrological and climatological processes.
Table 1.1 Studies on the Applications of Wavelets in Hydrology

Climatology Baliunas et al. (1997), Benner (1999), Higuchi et al. (1999), Lucero and Rodriguez (1999), Compagnucci et al. (2000), Janicot and Sultan
(2000), Jury and Melice (2000), Kulkarni (2000), Saco and Kumar (2000), Jung et al. (2002), Lucero and Rozas (2002), Tyson et al. (2002), Min
et al. (2003), Zheng et al. (2003), Anctil and Coulibaly (2004), Coulibaly and Burn (2004), Coulibaly (2006), Partal and Küçük (2006), Su et al.
(2008), Sujino et al. (2004), Franco Villoria et al. (2012), Labat et al. (2012), Carey et al. (2013), de Souza et al. (2010), Dong et al. (2014),
Dieppois et al. (2013), Bourrel et al. (2014), Zhang et al. (2014), Küçük et al. (2009), Li et al. (2009), Gobena et al. (2009), Liu et al. (2014),
Briciu and Mihaila (2014), Xu et al. (2014), Restrepo et al. (2014), Chevalier et al. (2013), Hierro et al. (2013), Assani et al. (2008), Masséi et al.
(2011), Nalley et al. (2012), Xu et al. (2013), Zhu and Meng (2010), Li et al. (2012), Rivera et al. (2013)
Precipitation fields and Kumar and Foufoula-Georgiou (1993), Takeuchi et al. (1994), Turner et al. (1994), Katul and Parlange (1995), Kumar (1996), Venugopal and
micrometeorology Foufoula-Georgiou (1996), Katul et al. (1998, 2001), Szilagyi et al. (1999), Furon et al. (2008)
Hydrology Dohan and Whitfield (1997), Fraedrich et al. (1997), Whitfield and Dohan (1997), Smith et al. (1998), Jury and Melice (2000), Labat et al.
(2000, 2001, 2004, 2005), Gaucherel (2002), Lafrenière and Sharp (2003), Lark et al. (2003), Labat (2005, 2006, 2008), Schaefli et al. (2007),
Kang and Lin (2007), Kuss et al. (2014), Slimani et al. (2009), Dong et al. (2014), Zhang et al. (2013), Lauzon et al. (2004), Parent et al. (2006)
Hydrologic Anctil and Tapé (2004), Bayazit and Aksoy (2001), Bayazit et al. (2001), Chou and Wang (2002), Aksoy et al. (2004a, b), Sujono et al. (2004),
modeling Cannas et al. (2006), Chou (2007), Lane (2007), Adamkowsi (2008a, 2008b), Kisi (2008, 2009), Partal and Cigizoglu (2008), Unal (2004), Kim
and Waldes (2003), Sahay and Sehgal (2014), Shoaib et al. (2014), Nourani et al. (2009a, 2009b, 2014a, 2014b), Goyal (2013, 2014), Sehgal et al.
(2014), Sahay and Srivastava (2014), Tiwari and Chatterjee (2010), Wang et al. (2009), Hu et al. (2008), Salerno and Tartari (2009), Rao and
Krishna (2009), Coulibaly and Baldwin (2008), Xu et al. (2014), Pramanik et al. (2011), Danandeh et al. (2013), Karthikeyan and Nagesh
Dumar (2013), Yu et al. (2013), Rathinasamy et al. (2013), Sang et al. (2009)
REFERENCES Briciu, A. E. and D. Mihăilă, “Wavelet analysis of some rivers in SE Europe

and selected climate indices,” Environmental Monitoring and Assessment,
Adamkowski, J. F., “River flow forecasting using wavelet and cross-wavelet
186 (10): 6263–6286. doi: 10.1007/s10661-014-3853-z.
transform models,” Hydrological Processes, 22 (25): 4877–4891, 2008a.
Cannas B., A. Fanni, L. See, and G. Sias, “Data preprocessing for river flow
Adamowski, J. F., “Development of a short-term river flood forecasting
forecasting using neural networks: wavelet transforms and data partitioning,”
method for snowmelt driven floods based on wavelet and cross-wavelet
Physics and Chemistry of the Earth, 31: 1164–1171, 2006.
analysis,” Journal of Hydrology, 353, 247–266, 2008b.
Carey, S. K., D. Tetzlaff, J. Buttle, H. Laudon, J. McDonnell., K. McGuire, J.
Aksoy H., T. Akar, and N. E. Unal, “Wavelet analysis for modeling
Seibert, et al. “Use of color maps and wavelet coherence to discern seasonal
suspended sediment discharge,” Nordic Hydrology, 35 (2): 165–174, 2004.
and interannual climate influences on streamflow variability in northern
Aksoy, H, Z. F. Toprak, A. Aytek, and N. E. Unal, “Stochastic generation of
catchments,” Water Resources Research, 49 (10): 6194–6207, 2013.
hourly mean wind speed data,” Renewable Energy, 29: 2111–2131, 2004.
Chevalier, L., B. Laignel, N. Massei, S. Munier, M. Becker, I. Turki, A.
Anctil, F. and P. Coulibaly, “Wavelet analysis of the interannual variability
Coyneld, et al. “Hydrological variability of major French rivers over the past
in Southern Quebec streamflow,” Journal of Climate, 17 (1): 163–173, 2004.
decades, assessed from gauging stations and GRACE observations,”
Anctil, F. and D. G. Tapé, “An exploration of ANN rainfall-runoff forecasting
Hydrological Sciences Journal, 59(10): 1844–1855.
combined with wavelet decomposition,” Journal of Environmental Engineering
Chou, C-M., “Applying multi-resolution analysis to differential hydrological
and Science, 3: S121–S128, 2004.
grey models with dual series,” Journal of Hydrology, 332: 174–186, 2007.
Assani, A. A., F. Lajoie, M. E. Vadnais, and G. Beauchamp, “Analyse de
Chou, C-M. and R-Y. Wang, “On-line estimation of unit hydrographs using
l’influence de l’oscillation Arctique sur la variabilité interannuelle des
the wavelet based LMS algorithm,” Hydrological Sciences Journal, 47 (5):
précipitations dans le bassin versant de la rivière Saint-François (Québec,
721–738, 2002.
Canada) au moyen de la méthode des corrélations canoniques,” Revue des
Compagnucci, R. H., S. A. Blanco, M. A. Figliola, and P. M. Jacovkis,
sciences de l’eau/Journal of Water Science, 21 (1): 19–33, 2008.
“Variability in subtropical Andean Argentinean Atuel river: a wavelet
Bayazit, M. and H. Aksoy, “Using wavelets for data generation,” Journal of
approach,” Environmetrics, 11: 251–269, 2000.
Applied Statistics, 28 (2): 157–166, 2001.
Coulibaly, P., “Spatial and temporal variability of Canadian seasonal pre-
Bayazit M., B. Önöz, and H. Aksoy, “Nonparametric streamflow simulation
cipitation (1900–2000),” Advances in Water Resources, 29: 1846–1865, 2006.
by wavelet or Fourier analysis,” Hydrological Sciences, 46 (4): 623–634, 2001.
Coulibaly, P. and C. K. Baldwin, “Dynamic neural networks for
Benner, T., “Central England temperatures: long term variability and tele-
nonstationary hydrological time series modeling,” Practical Hydroinformatics,
connections,” International Journal of Climatology, 19: 391–403, 1999.
Springer, Berlin, Heidelberg, 2008, pp. 71–85.
Bourrel, L., P. Rau, B. Dewitte, D. Labat, W. Lavado, A. Coutaud, A. Vera,
Coulibaly, P. and D. H. Burn, “Wavelet analysis of variability in annual
et al. “Low‐frequency modulation and trend of the relationship between
Canadian streamflows,” Water Resource Research, 40 (3): W03105, 2006, doi:
ENSO and precipitation along the Northern to Center Peruvian Pacific coast,”
10.1029 /2003WR002667.
Hydrological Processes, Volume 29, Issue 6, pages 1252–1266, 2014.
15_Singh_ch15_p15.1-15.8.indd 5 8/22/16 11:45 AM

Danandeh Mehr, A., E. Kahya, and E. Olyaie, “Streamflow prediction using Jawerth, B. and W. Sweldens, “An overview of wavelet based multiresolution
linear genetic programming in comparison with a neuro-wavelet technique,” analyses,” SIAM Review, 36 (3): 377–412, 1994.
Journal of Hydrology, 505: 240–249, 2013. Jung, H-S., Y. Choi, and G-H. Lim, “Recent trends in temperature and
Daubechies I., “The wavelet transform, time-frequency localization and sig- precipitation over South Korea,” International Journal of Climatology, 22:
nal analysis,” IEEE Transactions on Information Theory, 36 (5): 961–1005, 1990. 1327–1337, 2002.
Daubechies, I., Ten lectures on wavelets, CSBM—NSF Series Application Jury, M. R. and J. L. Melice, “Analysis of Durban rainfall and Nile river flow
Mathematics, No. 61, SIAM, 1992, p. 357, doi : http://dx.doi.org/10.1137/ 1871–1999,” Theoretical and Applied Climatology, 67: 161–169, 2000.
1.9781611970104. Kang S. and H. Lin, “Wavelet analysis of hydrological and water quality
de Sousa, A. M., N. M. Castro, F. A. Canales, J. A. Louzada, M. I. Vitorino, signals in an agricultural watershed,” Journal of Hydrology, 338: 1–14,
and E. B. de Souza, “Multiscale variability of the evapotranspiration in eastern 2007.
Amazonia,” Atmospheric Science Letters, 11 (3): 192–198, 2010. Karthikeyan, L. and D. Nagesh Kumar, “Predictability of nonstationary
Dieppois, B., A. Durand, M. Fournier, and N. Massei, “Links between time series using wavelet and EMD based ARMA models,” Journal of
multidecadal and interdecadal climatic oscillations in the North Atlantic and Hydrology, 502: 103–119, 2013.
regional climate variability of northern France and England since the 17th Katul, G. G. and M. B. Parlange, “Analysis of land surface heat fluxes using
century,” Journal of Geophysical Research: Atmospheres, 118 (10): 4359–4372, the orthonormal wavelet approach,” Water Resources Research, 31 (11):
2013. 2743–2749, 1995.
Dohan, K. and P. H. Whitfield, “Identification and characterization of Katul, G. G., M. B. Parlange, and C. R. Chu, “Intermittency, local isotropy,
water quality transients using wavelet analysis. I Wavelet analysis and non-gaussian statistics in atmospheric surface layer turbulence,” Physics
methodology,” Water Science and Technology, 36 (5): 325–335, 1997. of Fluids, 6 (7): 2480–2492, 1994.
Dong, L., J. Shimada, M. Kagabu, and C. Fu, “Teleconnection and climatic Katul, G., C-T. Lai, K. Schäfer, B. Vidakovic, J. Albertson, D. Ellsworth, and
oscillation in aquifer water level in Kumamoto plain, Japan,” Hydrological R. Oren, Multiscale analysis of vegetation surface fluxes: from second to years.
Processes, 29 (7): 1687–1703, 2014. Advances in Water Resources, 24: 1119–1132, 2001.
Finazzi, F., R. Haggarty, C. Miller, M. Scott, and A. Fassò, “A comparison of Katul, G. G., J. Schieldge, C-I. Hsieh, and B. Vidakovic, “Skin temperature
clustering approaches for the study of the temporal coherence of multiple perturbations induced by surface layer turbulence above a grass surface,”
time series,” Stochastic Environmental Research and Risk Assessment, 1–13, Water Resource Research, 34 (5): 1265–1274, 1998.
2014. Kim, T. W. and J. B. Valdes, “A nonlinear model for drought forecasting
Fraedrich, K., J. Jiang, F-W. Gerstengarbe, and P. C. Werner, “Multiscale based on conjunction of wavelet transforms and neural networks,” Journal of
detection of abrupt climate changes: application to Nile river flood levels,” Hydrologic Engineering, 8: 319–328, 2003.
International Journal of Climatology, 17: 1301–1315, 1997. Kisi, Ö., “Stream flow forecasting using neuro-wavelet technique,”
Franco Villoria, M., E. M. Scott, T. Hoey, and D. Fischbacher-Smith, Hydrological Processes, 22: 4142–4152, 2008.
“Temporal investigation of flow variability in Scottish rivers using wavelet Kişi, Ö., “Wavelet regression model as an alternative to neural networks for
analysis,” Journal of Environmental Statistics, 3 (6): 2012. monthly streamflow forecasting,” Hydrological Processes, 23 (25): 3583–3597,
Furon, A. C., C. Wagner-riddle, C. R. Smith, and J. S. Warland, “Wavelet 2009.
analysis of wintertime and spring thaw CO2 and N2O fluxes from agricultural Kisi, Ö. and Cimen, M. “A wavelet-support vector machine conjunction
fields,” Agricultural and Forest Meteorology, 148: 1305–1317, 2008. model for monthly streamflow forecasting,” Journal of Hydrology, 399 (1):
Gabor, D., “Theory of communication,” Journal of the Institute of Electrical. 132–140, 2011.
Engineers (London), 93 (III): 429–457, 1946. Küçük, M., E. Kahya, T. M. Cengiz, and M. Karaca, “North Atlantic
Gardner, W. A., “A unifying view of coherence in signal processing,” Signal oscillation influences on Turkish lake levels,” Hydrological Processes, 23 (6):
Processing, 29: 113–140, 1992. 893–906, 2009.
Gaucherel, C., “Use of wavelet transform for temporal characterisation of Kulkarni, J. R., “Wavelet analysis of the association between the southern
remote watersheds,” Journal of Hydrology, 269 (3–4): 101–121, 2002. oscillation and the Indian summer monsoon,” International Journal of
Ge, Z., “Significance tests for the wavelet power and the wavelet power Climatology, 20: 89–104, 2000.
spectrum,” Annales Geophysicae, 25: 2259–2269, 2007. Kumar, P., “Role of coherent structure in the stochastic dynamic variability
Gobena, A. K. and T. Y. Gan, “The role of Pacific climate on low-frequency of precipitation,” Journal of Geophysical Research, 101 (26): 393–404, 1996.
hydroclimatic variability and predictability in southern Alberta, Canada,” Kumar P. and E. Foufoula-Georgiou, “A multicomponent decomposition of
Journal of Hydrometeorology, 10 (6): 1465–1478, 2009. spatial rainfall fields. Segregation of large and small scale features using
Gonzales S. S., P. N. Gonzales, and G. S. Garcia, Uvi-Wave Wavelet Toolbox wavelet transform,” Water Resources Research, 29 (8): 2515–2532, 1993.
for Use with Matlab, University of Vigo, Department of Communication Kuss, A. J. M. and J. J. Gurdak, “Groundwater level response in US Principal
Technology, Spain, 1996, p. 196. Aquifers to ENSO, NAO, PDO, and AMO,” Journal of Hydrology, 519 (B):
Goyal, M. K., “Monthly rainfall prediction using wavelet regression and 1939–1952, 2014.
neural network: an analysis of 1901–2002 data, Assam, India,” Theoretical and Labat, D., “Recent advances in wavelet analyses: Part 1. A review of
Applied Climatology, 1181): 25–34, 2013. concepts,” Journal of Hydrology, 314: 275–288, 2005.
Goyal, M. K., “Modeling of sediment yield prediction using M5 Model Tree Labat, D., “Oscillations in land surface hydrological cycle,” Earth and
Algorithm and wavelet regression,” Water Resources Management, 28 (7): Planetary Science Letters, 242: 143–154, 2006.
1991–2003, 2014. Labat, D., “Wavelet analysis of annual discharge records of the world’s
Grinsted, A., J. C. Moore, and S. Jevrejeva, “Application of the cross wavelet largest rivers,” Advances in Water Resources, 31: 109–117, 2008.
transform and wavelet coherence to geophysical time series,” Nonlinear Labat, D., R. Ababou, and A. Mangin, “Analyse en ondelettes en Hydrologie
Processes in Geophysics, 11: 561–566, 2004. karstique: 1ère Partie: analyse univariée de pluies et débits de sources
Grosmann, A. and J. Morlet, “Decomposition of Hardy functions into karstiques,” Comptes Rendus de l’Académie des Sciences de Paris, 329 (série
square integrable wavelets of constant shape,” SIAM Journal on Mathematical IIa): 873–879, 1999a.
Analysis, 15(4): 723–736, 1984. Labat, D., R. Ababou, and A. Mangin, “Analyse en ondelettes en Hydrologie
Heil, C. E. and D. F. Walnut, “Continuous and discrete wavelet transforms,” karstique: 2ème Partie: analyse croisée pluies-débits,” Comptes Rendus de
SIAM Review, 31 (4): 628–666, 1989. l’Académie des Sciences de Paris, 329 (série IIa): 881–887, 1999b.
Hierro, R., P. Llamedo, A. Torre, and P. Alexander, “Oscillation modes of Labat, D., R. Ababou, and A. Mangin, “Rainfall-runoff relations for karstic
humidity over the Amazon basin derived from GPS RO profiles,” Journal of springs: Part II: Continuous wavelet and discrete orthogonal multiresolution
Geophysical Research: Atmospheres, 118 (23): 13–121, 2013. analyses,” Journal of Hydrology, 238: 149–178, 2000.
Higuchi, K., J. Huang, and A. Shabbar, “A wavelet characterization of the Labat, D., R. Ababou, and A. Mangin, “Introduction of wavelet analyses to
North Atlantic Oscillation variation and its relationship to the North Atlantic rainfall/runoffs relationships for a karstic basin: the case of Licq-Atherey
sea surface temperature,” International Journal of Climatology, 19: 1119–1129, karstic system (France),” Ground Water, 39 (4): 605–615, 2001.
1999. Labat, D., R. Ababou, and A. Mangin, “Analyse multirésolution croisée de
Hu, C., Y. Hao, T. C. J. Yeh, B. Pang, and Z. Wu, “Simulation of spring flows pluies et débits de sources karstiques,” Comptes Rendus de l’Académie des
from a karst aquifer with an artificial neural network,” Hydrological Processes, Sciences de Paris, Géosciences, 334: 551–556, 2002.
22 (5): 596–604, 2008. Labat, D., J. Ronchail, J. Callède, J. L. Guyot, E. de Oliveira, and W.
Janicot, S. and B. Sultan, “Intra-seasonal modulation of convection in the Guimarães, “Wavelet analysis of Amazon hydrological regime variability,”
West African monsoon,” Geophysical Research Letters, 28 (2): 523–526, 2000. Geophysical Research Letter, 31: L02501, 2004.
15_Singh_ch15_p15.1-15.8.indd 6 8/22/16 11:45 AM

REFERENCES 15-7
Labat, D., J. Ronchail, and J-L. Guyot, “Recent advances in wavelet analyses: Nourani, V., M. T. Alami, and M. H. Aminfar, “A combined neural-wavelet
Part 2—Amazon, Parana, Orinoco and Congo discharges time scale model for prediction of Ligvanchai watershed precipitation,” Engineering
variability,” Journal of Hydrology, 314: 289–311, 2005. Applications of Artificial Intelligence, 22 (3): 466–472, 2009.
Labat, D., J. C. Espinoza, J. Ronchail, G. Cochonneau, E. de Oliveira, J. C. Nourani, V., A. H. Baghanam, A. Y. Rahimi, and F. H. Nejad, “Evaluation
Doudou, and J. L. Guyot, “Fluctuations in the monthly discharge of Guyana of wavelet-based de-noising approach in hydrological models linked to
Shield rivers, related to Pacific and Atlantic climate variability,” Hydrological artificial neural networks, Computational Intelligence Techniques in Earth and
Sciences Journal, 57 (6): 1081–1091, 2012. Environmental Sciences, Springer, The Netherlands, 2014, pp. 209–241.
Lachaux, J. P., A. Lutz, D. Rudrauf, D. Cosmelli, J. M. Le Van QuyenMartin- Nourani, V., A. Hosseini Baghanam, J. Adamowski, and O. Kisi,
erie, and F. Varela, “Estimating the time-course of coherence between single “Applications of hybrid wavelet–artificial intelligence models in hydrology: a
trial brain signals: an introduction to wavelet coherence,” Neurophysiology review,” Journal of Hydrology, 514: 358–377, 2014.
Clinic, 32: 157–174, 2002. Nourani, V., Komasi, M., & Mano, A, “A multivariate ANN-wavelet
Lafrenieres, M. and M. Sharp, “Wavelet analysis of inter-annual variability approach for rainfall–runoff modeling,” Water Resources Management, 23
in the runoff regimes of glacial and nival stream catchments, Bow Lake, (14): 2877–2894, 2009.
Alberta,” Hydrological Processes, 17 (6): 1093–1118, 2003. Papoulis, A., Probability, Random Variables and Stochastic Processes,
Lane S. N., “Assessment of rainfall–runoff models based upon wavelet McGraw Hill, New York, 1964, p. 588.
analysis,” Hydrological Processes, 21: 586–607, 2007. Parent, A. C., F. Anctil, and L. Parent, “Characterization of temporal
Lark, R. M., S. R. Kaffka, and D. L. Corwin, “Multiresolution analysis of soil variability in near-surface soil moisture at scales from 1 h to 2 weeks,” Journal
data on electrical conductivity of soil using wavelets,” Journal of Hydrology, of Hydrology, 325: 56–66, 2006.
272, 276–290, 2003. Partal, T. and M. Küçük, “Long-term trend analysis using discrete wavelet
Lauzon, N, F. Anctil, and J. Petrinovic, “Characterization of soil moisture components of annual precipitations measurements in Marmara region
conditions at temporal scales from a few days to annual,” Hydrological (Turkey),” Physics and Chemistry of the Earth, 31: 1189–1200, 2006.
Processes, 18: 3235–3254, 2004, doi: 10.1002/hyp.5656. Partal, T. and H. K. Cigizoglu, “Estimation and forecasting of daily sus-
Li, C. H., Z. F. Yang, G. H. Huang, and Y. P. Li, “Identification of pended sediment data using wavelet-neural networks,” Journal of Hydrology,
relationship between sunspots and natural runoff in the Yellow River based 358 (3–4): 317–331, doi:10.1016/j.jhydrol.2008.06.013
on discrete wavelet analysis,” Expert Systems with Applications, 36 (2): 3309– Percival, D., “On estimation of wavelet variance,” Biometrika, 82 (3): 619–631,
3318, 2009. 1995.
Li, D., W. Wang, S. Hu, and Y. Li, “Characteristics of annual runoff variation Percival, D. and H. O. Mofjeld, “Analysis of subtidal coastal sea level
in major rivers of China,” Hydrological Processes, 26 (19): 2866–2877, 2012. fluctuations using wavelets,” Journal of the American Statistical Association, 92
Liu, P. C., Wavelet spectrum analysis and ocean wind waves, edited by E. (493): 868–880, 1997.
Foufoula-Georgiou and P. Kumar, Wavelets in Geophysics, Academic Press, Pramanik, N., R. Panda, and A. Singh, “Daily river flow forecasting using
New York, 1995, pp. 151–166. wavelet ANN hybrid models,” Journal of Hydroinformatics, 13 (1): 49–63, 2011.
Liu, Y., X. S. Liang, and R. H. Weisberg, “Rectification of the bias in the Priestley, M. B., Spectral Analysis and Time Series, Academic Press, San
wavelet spectrum,” Journal of Atmospheric and Oceanic Technology, 24 (12), Diego, CA, 1981, p. 890.
2093–2102, 2007. Rao, Y. S. and B. Krishna, Modelling Hydrological Time Series Data Using
Liu, B., C. Xiao, and X. Liang, “Evaluation of spatial and temporal Wavelet Neural Network Analysis, IAHS Publication, Wallingford, UK, Vol.
characteristics of precipitation variations in Jilin Province, Northeast China,” 333, 2009, p. 101.
Theoretical and Applied Climatology, 122 (1): 129–142, 2014. Rathinasamy, M., J. Adamowski, and R. Khosa, “Multiscale streamflow
Lucero, O. A. and D. Rozas, “Characteristics of aggregation of daily rainfall forecasting using a new Bayesian Model Average based ensemble multi-
in a middle-latitudes region during climate variability in annual rainfall wavelet Volterra nonlinear method,” Journal of Hydrology, 507: 186–200,
amount,” Atmospheric Research, 61: 35–48, 2002. 2013.
Lucero, O. A. and N. C. Rodriguez, “Spatial organization of decadal and Restrepo, J. C., J. C. Ortíz, J. Pierini, K. Schrottke, M. Maza, L. Otero, and J.
bidecadal rainfall fluctuations in southern North America and Southern Aguirre, “Freshwater discharge into the Caribbean Sea from the rivers of
South America,” Atmospheric Research, 57: 219–246, 2001. Northwestern South America (Colombia): magnitude, variability and recent
Lucero, O. A. and N. C. Rodriquez, “Relationship between interdecadal changes,” Journal of Hydrology, 509: 266–281, 2014.
fluctuations in annual rainfall amount and annual rainfall trend in a Rivera, J. A., O. C. Penalba, and M. L. Betolli, “Inter-annual and
southern mid-latitudes region of Argentina,” Atmospheric Research, 52: inter-decadal variability of dry days in Argentina,” International Journal of
177–193, 1999. Climatology, 33 (4): 834–842, 2013.
Mallat, S., “A theory for multiresolution signal decomposition: the wavelet Saco, P. and P. Kumar, “Coherent modes in multiscale variability of stream-
representation,” IEEE Transactions on Pattern Analysis and Machine flow over the United States,” Water Resources Research, 36 (4): 1049–1067,
Intelligency, 11: 674–693, 1989. 2000.
Maruan, D. and J. Kurths, “Cross wavelet analysis: significance testing and Sahay, R. R. and V. Sehgal, “Wavelet-ANFIS models for forecasting
pitfalls,” Nonlinear Processes in Geophysics, 11: 505–514, 2004. monsoon flows: case study for the Gandak River (India),” Water Resources, 41
Massei, N., B. Laignel, E. Rosero, A. Motelay‐massei, J. Deloffre, Z. L. Yang, (5) : 574–582, 2014.
and A. Rossi, “A wavelet approach to the short‐term to pluri‐decennal Sahay, R. R. and A. Srivastava, “Predicting monsoon floods in rivers
variability of streamflow in the Mississippi river basin from 1934 to 1998,” embedding wavelet transform, genetic algorithm and neural network,” Water
International Journal of Climatology, 31 (1): 31–43, 2011. Resources Management, 28 (2) : 301–317, 2014.
Meneveau, C., “Analysis of turbulence in the orthonormal wavelet Salerno, F. and G. Tartari, “A coupled approach of surface hydrological
representation,” Journal of Fluid Mechanics, 232: 469–520, 1991. modelling and Wavelet Analysis for understanding the baseflow components
Min, S-K., W-T. Kwon, E-H. Park, and Y. Choi, “Spatial and temporal of river discharge in karst environments,” Journal of Hydrology, 376 (1): 295–
comparisons of droughts over Korea with East Asia,” International Journal of 306, 2009.
Climatology, 23: 223–233, 2003. Sang, Y. F., “A review on the applications of wavelet transform in hydrology
Mizuno-Matsumoto, Y., G. K. Motamedi, W. R. S. Webber, and R. P. Lesser, time series analysis,” Atmospheric Research, 122: 8–15, 2013.
“Wavelet cross-correlation analysis can help predict whether bursts of pulse Sang, Y. F., D. Wang, and J. C. Wu, Study on the WCC method for time
stimulation will terminate after discharges,” Clinical Neurophysiology, 113: series data analysis, Sixth International Conference on Fuzzy Systems and
33–42, 2002. Knowledge Discovery, 2009. FSKD’09, Los Alamitos, CA, IEEE, Vol. 3, 2009,
Mizuno-Matsumoto, Y., T. Yoshimine, Y. Nii, A. Kato, M. Taniguchi, J. K. pp. 593–597.
Lee, T. S. Ko, et al., “Landau-Kleffner Syndrome: localization of epileptogenic Schaefli, B., D. Maraun, and M. Holschneider, “What drives high flow
lesion using wavelet cross—correlation analysis,” Epilepsy and Behaviour, 2: events in the Swiss Alps? Recent developments in wavelet spectral analysis
288–294, 2001. and their application to hydrology,” Advances in Water Resources, 30: 2511–
Nalley, D., J. Adamowski, and B. Khalil, “Using discrete wavelet transforms 2525, 2007.
to analyze trends in streamflow and precipitation in Quebec and Ontario Sehgal, V., R. R. Sahay, and C. Chatterjee, “Effect of utilization of discrete
(1954–2008),” Journal of Hydrology, 475: 204–228, 2012. wavelet components on flood forecasting performance of wavelet based
Niu, J., “Precipitation in the Pearl River basin, South China: scaling, ANFIS models,” Water Resources Management, 28 (6): 1733–1749, 2014.
regional patterns, and influence of large-scale climate anomalies,” Stochastic Sen, A. K. and K. Fang, “Multiscale variability and atmospheric teleconnec-
Environmental Research and Risk Assessment, 27 (5): 1253–1268, 2013. tions of the precipitation time series reconstructed from tree ring records in
15_Singh_ch15_p15.1-15.8.indd 7 8/22/16 11:45 AM

the xiaolong mountain, central China,” Journal of Coupled Systems and Unal, N. E., H. Aksoy, and T. Akar, “Annual and monthly rainfall data
Multiscale Dynamics, 2 (1): 9–14, 2014. generation schemes,” Stochastic Environmental Research and Risk Assessment,
Shoaib, M., A. Y. Shamseldin, and B. W. Melville, “Comparative study of 18 (4): 245–257, 2004.
different wavelet based neural network models for rainfall–runoff modeling,” Venugopal, V. and E. Foufoula-Georgiou, “Energy decomposition of
Journal of Hydrology, 515: 47–58, 2014. rainfall in the time-frequency-scale domain using wavelet packets,” Journal of
Slimani, S., N. Massei, J. Mesquita, D. Valdés, M. Fournier, B. Laignel, and Hydrology, 187: 3–27, 1996.
J. P. Dupont, “Combined climatic and geological forcings on the spatio- Walker, J. S., “Fourier analysis and wavelet analysis,” Notices of The
temporal variability of piezometric levels in the chalk aquifer of Upper American Mathematical Society, 44 (6): 658–670, 1997.
Normandy (France) at pluridecennal scale,” Hydrogeology Journal, 17 (8): Wang, W., J. Jin, and Y. Li, “Prediction of inflow at three gorges dam in
1823–1832, 2009. Yangtze River with wavelet network model,” Water Resources Management, 23
Smith, L. C., D. L. Turcotte, and B. L. Isacks, “Stream flow characterization (13): 2791–2803, 2009.
and feature detection using a discrete wavelet transform,” Hydrological Whitfield, P. H. and K. Dohan, “Identification and characterization of
Processes, 12: 233–249, 1998. water quality transients using wavelet analysis. II Application to electronic
Su, B., M. Gemmer, and T. Jiang, “Spatial and temporal variation of extreme water quality data,” Water Science and Technology, 36 (5): 337–348, 1997.
precipitation over the Yangtze River Basin,” Quaternary International, 186: Xu, J., Y. Chen, and W. Li, “The nonlinear hydro-climatic process: a case
22–31, 2008. study of the Tarim Headwaters, NW China, Water Resources Research in
Sujono J., S. Shikasho, and K. Hiramatsu, “A comparison of techniques for Northwest China, Springer, The Netherlands, 2014, pp. 289–310.
hydrograph recession analysis,” Hydrological Processes, 18: 403–413, 2004, doi: Xu, J., Y. Chen, W. Li, Q. Nie, Y. Hong, and Y. Yang, “The nonlinear hydro-
10Ð1002/hyp.1247. climatic process in the Yarkand River, northwestern China,” Stochastic
Szilagyi J., M. B. Parlange, G. G. Katul, and J. D. Albertson, “An objective Environmental Research and Risk Assessment, 27 (2): 389–399, 2013.
method for determining principal time scales of coherent eddy structures using Xu, J., Y. Chen, W. Li, Q. Nie, C. Song, and C. Wei, “Integrating wavelet
orthonormal wavelets,” Advances in Water Resources, 22 (6): 561–566, 1999. analysis and BPANN to simulate the annual runoff with regional climate
Takeuchi, N., K. Narita, and Y. Goto, “Wavelet analysis of meteorological change: a case Study of Yarkand River, Northwest China,” Water Resources
variables under winter thunderclouds over the Japan sea,” Journal of Management, 28 (9): 2523–2537, 2014.
Geophysical Research, 99 (D5): 10751–70757, 1994. Yu, S. P., J. S. Yang, and G. M. Liu, “A novel discussion on two long-term
Tary, J. B., R. H. Herrera, J. Han, and M. van der Baan, “Spectral forecast mechanisms for hydro-meteorological signals using hybrid wavelet-
estimation—what is new? What is next?,” Reviews of Geophysics, 52(4): 723– NN model,” Journal of Hydrology, 497: 189–197, 2013.
749, 2014, doi:10.1002/2014RG000461. Zhang, D., H. Hong, Q. Zhang, and X. Li, “Attribution of the changes in
Tiwari, M. K. and C. Chatterjee, “Development of an accurate and reliable annual streamflow in the Yangtze River Basin over the past 146 years,”
hourly flood forecasting model using wavelet–bootstrap–ANN (WBANN) Theoretical and Applied Climatology, 119(1): 323–332, 2015.
hybrid approach,” Journal of Hydrology, 394 (3): 458–470, 2010. Zhang, Y. Y., D. Y. Zhong, and B. S. Wu, “Multiple temporal scale
Torrence, C. and G. P. Compo, “A practical guide to wavelet analysis,” relationships of bankfull discharge with streamflow and sediment transport in
Bulletin of the American Meteorological Society, 79: 61–78, 1998. the Yellow River in China,” International Journal of Sediment Research, 28 (4):
Torrence, C. and P. J. Webster, “Interdecadal changes in the ENSO- 496–510, 2013.
Monsoon system,” Journal of Climate, 12: 2679–2690, 1999. Zheng, D., X. Ding, Y. Zhou, and Y. Chen, “Earth rotation and ENSO
Turner, B. J., M. Y. Leclerc, M. Gauthier, K. E. Moore, and D. R. Fitzjarrald, events: combined excitation of interannual LOD variations by multiscale
“Identification of turbulence structures above a forest canopy using a wavelet atmospheric oscillations,” Global and Planetary Change, 36: 89–97, 2003.
transform,” Journal of Geophysical Research, 99 (D1): 1919–1926, 1994. Zhu, L. and J. Meng, “Study on rainfall variations in the middle part of
Tyson, P. D., G. R. J. Cooper, and T. S. McCarthy, “Millenial to multi- Inner Mongolia, China during the past 43 years,” Environmental Earth
decadal variability in the climate of Southern Africa,” International Journal of Sciences, 60 (8): 1661–1671, 2010.
Climatology, 22: 1105–1117, 2002.
15_Singh_ch15_p15.1-15.8.indd 8 8/22/16 11:45 AM

Chapter 16
Outlier Analysis and Infilling of
Missing Records in Hydrologic
Data
BY
UMED S. PANU AND WAIWAH NG
ABSTRACT required, and mainly governed by the data characteristics in terms of their
occurrence in space and time.
Hydrologic data analyses invariably suffer from challenges arising from the
This chapter deals with the basic definitions and presents discussion neces-
existence of outliers and missing records in datasets. In absence of adequate
sary for the understanding of the concepts of outliers and missing records, as
understanding and/or interpretation of such datasets usually lead to biased
well as their handling methods in hydrologic analysis. The definitions, types,
results in hydrologic analyses. Fundamental definitions and germane discus-
causes, identification/treatment methods of outliers as well as the infilling
sion necessary for understanding the concepts of outliers and missing records
procedures of missing records are presented. Specific topics including how to
as well as their handling methods in hydrologic analyses are offered. The sig-
improve existing methods and approaches, and future research directions are
nificance of outliers and missing records in hydrologic datasets is espoused in
discussed in subsequent sections.
view of analysis, forecasting, and synthesis of hydrologic data. Specific topics
including suggestions for improvement of existing methods, and future
research directions pertaining to handling of outliers and missing records are 16.1.1 Characteristics of Hydrologic Data
discoursed. This chapter endeavors to provide succinct yet fundamental back- Measurement and reporting of hydrologic data on various hydrologic pro-
ground information concerning outliers and missing records in hydrologic cesses can be classified into four types (Yevjevich, 1972): (1) data recorded in
datasets. time; (2) data recorded in space; (3) data recorded in time and space; and (4)
data concurrently recorded in time and/or space on two or more variables.
Most of historically collected data on hydrologic processes, specifically on
16.1 INTRODUCTION
streamflows, belong to the first type. Streamflows and other hydrologic phe-
In geophysical sciences, such as hydrology, observations are regularly record- nomena are observed in discrete time and thus time becomes an indexing
ed to capture changes in historical processes under a specific space interval or variable. The resulting temporal data can be referred to as univariate time
time period. Hydrologic data may contain extreme observations (outliers) series with single variable, or multivariate time series with multiple variables.
due to subtle changes experienced by the process of a system, or due to the Hereafter, the temporal hydrologic data is referred to as hydrologic data.
impact of external factors. Such outliers are indeed true observations. In Hydrologic data, such as streamflow and precipitation records, typically
contrast, outliers in dataset can be caused by instrumental errors (e.g., faulty comprises of data without negative values. The existence of lower bound of
equipment, malfunction of instrument, and improper exposure of instru- zero and occurrences of high outliers cause positive skewness in the distribu-
ment), or human errors (e.g., misreading the instrument, wrong entry of data tion of hydrologic data. On the other hand, rivers or streams originating in
in records/computer, and error in computation). In hydrology, the concern lakes and swamps cause negative skewness because of prolonged periods of
for outliers commonly arises in context of regional homogeneity assessment low flows. The lower limit of zero and presence of skewness invariably make
and frequency-analysis studies. Without a careful investigation of possible the hydrologic data to be non-normally distributed and their distributions
causes and/or overlooking the importance of outliers, practitioners may with heavy tails (Klemes, 2000). In addition, hydrologic data may be regarded
incorrectly include or omit outliers in their analysis. Consequently, the over/ composed of deterministic and stochastic components. The deterministic
underestimation of the impact of outliers may lead to over conservative or component may be comprised of cyclicity and/or some kind of monotonic
under design in water resources systems. trend. The stochastic component could be pure random or may display some
At times, the datasets may be riddled with missing records. Missing records persistence.
are common to hydrologic datasets for reasons such as loss of records, mal- The existing techniques in hydrologic literature do not explicitly consider
function of instruments, temporary termination of the data collection at a groups of similar data values and are rather based on the consideration of
specific location or period because of relatively less interest to the survey or individual data values where both cyclicity and trend effects are removed by
data-collection agencies, or lack of financial, human, or apparatus resources applying prespecified, sometimes arbitrary, procedures. In contrast, tech-
to continue measurements. Under such different circumstances, some prac- niques, such as pattern recognition, explicitly consider similar values in the
tical and rational approaches are required to handle missing records. The form of groups (or patterns) where cyclicity is considered as an integral part
presence of missing records in a dataset presents unique challenges to the of the pattern itself and hence no need arises for its isolation. In a seminal
analyst because these scenarios lead to less-reliable estimates of parameters, paper, Unny et al. (1981) respecified a “deterministic trend” as a “known
which bias the results of analysis. Methods for handling of missing records are trend,” which in any case, must be removed before analysis.
16-1
16_Singh_ch16_p16.1-16.8.indd 1 8/22/16 11:47 AM

16-2 Outlier Analysis and Infilling of Missing Records in Hydrologic Data
16.1.2 Outliers and Missing Records and four types of outliers caused by their nature of occurrence: incorrect data
Their Significance entry, computer coding failure, data contamination, and true observation(s).
Outliers in a dataset in essence are undesirable (or unwanted) values similar Knowledge on the causes of outliers can help in the determination of a
to weed plants in a cultivated agricultural field. For example, maize plants, suitable treatment. Confirmed erroneous observation(s) can simply be
although physically exist in a wheat field (i.e., statistically considered as a part removed. As for the treatment of true extreme observation(s), the user needs
of the population), are indeed considered by the farmer as undesirable or to perform statistical outlier tests to assess whether or not the extremal
unwanted (i.e., weed plants). In other words, outlier (e.g., maize plants) may observation(s) likely belong to the same population. The user will then decide
be subjectively determined based on the interest of the user. Outliers (i.e., the treatment of the identified outlier(s) as to be “removed,” “retained,” or
extreme high and/or low observations) being true observations on the phe- “revised” (Grubbs, 1969). Usually, such a decision-making process may
nomena and/or hydrologic process and yet may be considered as observations involve investigation of outliers using other variables, risk-assessment consid-
which do not conform to the remainder of data. A key question for an analyst erations, and the experience of the user.
is whether an observation is an outlier or not, and should such an observation
be discarded from the dataset simply because it appears to be inconsistent 16.2.2 Approaches for Outlier Analysis
with rest of the data. In the design and management of water resources sys-
Outlier analysis can be viewed as comprising of two components, namely, the
tems, it is important to be aware of anomalies (i.e., outliers) in data as they can
outlier detection and the outlier treatment. The statistical purpose of identify-
induce bias in the estimation of parameters, and may consequently lead to
ing suspicious observation(s) is to prevent data contamination, to recognize
improper water resource management policies or infrastructure design.
mixed groups, to avoid distortion of the presumed underlying probability
When the interior data of a distribution are in error, these data are called as
distribution, and to enhance the quality of hydrologic data.
inliers. Inliers can arise due to systematic or processing error. Statistical analysis
Outlier detection and their treatment constitute a broad field encompass-
can be seriously affected by inliers because inliers are difficult to distinguish
ing a wide array of applications. Outlier analysis in different data fields typi-
from good data values. Sometimes, they are even difficult to find and correct.
cally requires specific types of techniques such as in time series data, temporal
Relevant detail pertaining to methods of handling inliers in the context of flood
outlier analysis examines anomalies in the behavior of the data across time.
frequency analysis is found elsewhere (Singh and Nakashima, 1981).
The data type clearly governs the selection of the method/technique to be
In addition to outliers, the existence of missing records is also one of the
used for outlier analysis. A key aspect of any outlier detection technique is the
common problems observed in hydrologic data. This problem can affect the
nature of hydrologic data.
fitting of statistical models and thus, leads to increment of uncertainty in
Outlier analysis in temporal data may be categorized (Gupta et al., 2014) in
model outputs (Little and Rubin, 1987). Assessment of incomplete datasets
a wide variety of ways representing different facets of the analysis. Outlier
invariably raises the level of complexity in statistical analyses. The varying
analysis in various aspects of hydrologic data can also be visualized to dis-
nature of missing observations under different circumstances requires ade-
cover rare and interesting occurrences similar to traditional techniques, but
quate, practical, and rational treatment because improper handling of missing
the hydrologic data is likely to offer new challenges.
observations may significantly distort the results of analyses.
Outliers can be detected in a variety of ways such as through concentration
of residuals, Studentized t-test distribution, cluster analysis (Mahalonobis
16.2 CONCEPTS AND METHODS FOR distance, Mds), etc.; however, there exists some commonalties among the
OUTLIER ANALYSIS various competing outlier detection methods. The process of outlier detection
can be divided into four basic processing units. (1) An observation by itself is
In some fields, outlier detection is also known as event detection, novelty
natural and boundless that cannot be discriminated as outlier without suitable
detection, anomaly detection, deviant discovery, change point detection, fault
specification of a defining boundary. (2) Adjustment of a dataset may be car-
detection, intrusion detection, or misuse detection. Presently, majority of
ried out to enhance the tractability of a dataset. (3) A proper presumed type
studies are devoted to three types of outliers, namely: contextual outliers (i.e.,
of distribution or pattern can isolate a target group from other groups in
anomalous in specific contextual attributes, e.g., longitude/latitude for the
the dataset. (4) Depending upon the interest of the user, the data excluded
spatial data, and time for temporal data), point outliers (i.e., individually
outside the boundary can be considered as outliers. The manner in which the
anomalous), and collective outliers (i.e., a collection of anomalies).
sample dataset is arranged and bounded will directly affect the outcome.
Without clearly specifying the scope and purpose of an outlier analysis, the
16.2.1 DEFINITION AND TYPES OF OUTLIERS results of such an analysis will vary. The treatment of outlier(s) is highly
dependent on the interest of the user.
An outlier can broadly be conceptualized (Aggarwal, 2013) as comprising of
a regular component, a noise component, and an anomaly component as
depicted as follows. 16.2.3 Methods for Outliers Detection
In streamflow analyses, outliers are referred to extreme high/low flows in
historical records. Such extremal observations can highly deviate from the
Strong or weak outlier
rest of the observations (Hu, 1987). Outliers in a dataset may bias the esti-
mates of missing observations and simply discarding outliers could lead to
Regular data Noise Inconsistency underestimation of the true missing observations. Thus, the selection and use
of an appropriate outlier detection approach is not only critical but also cru-
cial in the design and operation of water resource systems.
Grade of OUTLYINGNESS increases from left to right Classical outlier detection is confined to pure statistical considerations
with disregard to the nature of the extremal observation(s) and to the need of
Figure 16.1 A Pictorial representation of conceptual aspects of an outlier. users. However, two major categories of outlier detection analysis are based
on the premises of accommodation (i.e., outlier inclusion with/without
revision) and discordance (i.e., outlier exclusion).
Despite an intuitive understanding on the meaning of outlier by many In a particular application, it is important to evaluate the available options
analysts, there exist varied interpretations. Barnett and Lewis (1978) stated of the selected method to meet the detection sensitivity for a specific case. For
that an outlier is “an observation” (or subset of observations), which appears example, in flood frequency analyses, a conservative outlier detection
to be inconsistent with the remainder of that set of data. Similarly but more approach is suggested because omissions of outlier(s) can result in serious
precisely, Hawkins (1980) defined an outlier as “an observation which devi- consequences due to under or over estimation of the design flow. This can be
ates so much from other observations as to arouse suspicions that it was achieved by using more sensitive outlier detection methods.
generated by a different mechanism.” Alternatively, Beckman and Cook Technically, the selection of a suitable outlier detection method involves
(1983) defined an outlier as “a collective to refer to either a contaminant or a different considerations, such as parametric versus nonparametric, sensitive
discordant observation.” In turn, Iglewicz and Hoaglin (1993) provided inter- versus insensitive, one-tail or two-tails, time and efficiency, type of
pretation of a discordant observation as “any observation that appears sur- application, number of outliers, number of variables, sample size, selection
prising or discrepant to the analyst” and likewise a contaminant observation of probability distribution, determination of the level of significance, outlier
as “any observation that is not a realization from the target distribution.” treatment, and risk assessment. Least square regression is an especially
An outlier is generally recognized either as an error or a true observation. useful method for multivariate outlier detection. Autoregressive Integrated
The terms such as “mixing” and “slippage” are also used to explain the outlier Moving Average (ARIMA) is useful for detection of outlier(s) in time series
phenomena (Hawkins, 1980). Tabachnick and Fidell (1996) even visualized datasets.
16_Singh_ch16_p16.1-16.8.indd 2 8/22/16 11:47 AM

Concepts and Methods for Handling Missing Records 16-3
The U.S. Water Resources Council (US-WRC) method is one of the most sampled is assumed to be normal. For this test procedure, consider a sample of
commonly used hydrologic techniques in North America for outlier detection size n, where the sample is arranged with the suspect value in front and followed
and outlier deletion (McCuen, 2004). This method is based on the principles by its nearest neighbor and then remainder values are arranged in ascending (or
of hypothesis testing with an underlying assumption of normal distribution, descending) order. The order is determined based on whether the suspect value
but applied to the log Pearson Type III (LP3) distributed data (Grubbs and is the largest or the smallest. The null hypothesis that the outlier belongs to the
Beck, 1972). Studies on the hydrologic outlier analysis by Kottegoda (1984) sample is rejected if the test statistic value exceeds the tabulated critical value
compared different statistical outlier methods in maximum flow data series, (Kanji, 1993). The advantage of the Dixon method is that it does not consider
whereas Singh (1980), Hu (1987), Spencer and McCuen (1996), and Panu et the mean and standard deviation, which in some cases cannot be accurately
al. (2000) primarily examined the application of the US-WRS method in obtained. This method is good for a small sample size, but its constraint is that
hydrologic data series. Pegram (1997) examined detection and identification the sample size should be greater than 3 and less than 25. For sample sizes larger
of extremal observation(s) in rainfall data and Singh (1980) in groundwater than 25, the Dixon method may not be appropriate.
data series. • The US-WRC method: The US-WRC method is the most popular
Other than US-WRC method, the US-EPA (U.S. Environment Protection method used for flood frequency analysis in North America. The method is
Agency) also employs an outlier analysis approach from the Scout multivari- based on Grubb’s test and normality assumption. The test procedure for a
ate analysis package (Singh, 1993), which includes the nonparametric outlier sample size (n) is to obtain the value of deviate (Ko) at 10% significance
approach (e.g., Mds) for extremal data detection. For flood frequency analysis level from the Table for Outlier Test Deviates (Ko) [Appendix 4 of Bulletin
in Canada, the Consolidated Frequency Analysis (CFA) package (Pilon et al., 17B (U. S. Interagency Advisory Committee, 1982)]. Compute the mean (Ỹ)
1985) is utilized for the outlier testing process. and standard deviation (Sy) of the logarithms of observations including the
In water resources engineering, the nonstatistical considerations (e.g., the extremal observation(s) being considered, but excluding zero-flood obser-
intent of outlier analysis and the interest of the user) are usually more critical vations, peak flows below the gage base, and outliers (extremal observa-
than the statistical procedures (outlier detection). To cope with the needs of tions) previously detected and identified. Compute the value of the detection
different users, statisticians have developed numerous methods for outlier(s) criterion for high outliers (Yoh) as Yoh = Ỹ + Ko Sy. Compare the logarithm
detection as briefly described below. of the extremal observation(s) being considered (Yh) with the criterion
• Box-plot: The box-plot serves in revealing the center of the data, their (Yoh); if Yh > Yoh then the extremal observation can be considered as a high
spread and distribution, and the presence of outliers. A box-plot displays the outlier (McCuen, 2004).
three quartiles as well as the minimum, the median, and the maximum of the • Consolidated Frequency Analysis method: In the CFA analysis package
data. Quantiles are values taken at intervals from the inverse of the cumulative (Pilon et al., 1985), the identification of outliers is conducted based on the
distribution function of a random variable. It includes a rectangular box that methods suggested by Grubbs and Beck (1972). Only high outliers are consid-
shows the interquartile range with the lower edge at the first quartile, and the ered for evaluation because of their importance in flood frequency analysis.
upper edge at the third quartile. A line is drawn through the box at the second
quartile (which is the 50th percentile or the median). A line at either end of 16.2.4 Methods for Outliers Treatment
the box extends to the extreme values which are observations between 0 and Often, a hydrologic data analyst is faced with a crucial question whether an
1.5 times that of the interquartile range from the edge of the box. Observa- observation is an outlier or not, and should such an observation be discarded
tions lying between 1.5 and 3 times the interquartile range from the edge of from the dataset simply because it appears to be inconsistent with rest of the
the box are called outliers. Observations that are beyond three times the hydrologic data. The treatment of outliers is still under debate because of the
interquartile range from the edge of the box are called extreme outliers user specific requirements and the varied nature of hydrological data sets. The
(Montgomery and Runger, 1994). Further, the box-plot method is simple to principal argument of handling outliers has been presented by Rousseeuw and
use and easy interpret. The disadvantage of the method is its reliance on the Leroy (1987), who proposed the use of robust regression methods, such as the
interquartile range, which may not have a strong relationship to the nature of least median of squares to cope with outliers in the dataset. On the other hand,
the dataset. Dixon (1951) has proposed the outlier removal approach. Irrespectively, there
• Mahalonobis Distance: The Mds of a point is a measure of the distance are suggestions that whenever outlier(s) is eliminated from the results, one
of the point from the mean point of a cluster. It is used to find outliers in should at least report outlier(s) along with results of the statistical analysis.
multivariate data and, unlike an absolute distance measurement, it takes Another remedial suggestion pertaining to outlier(s) treatment is to enlarge the
into account the “shape” of the cluster. The Mds method is very sensitive to sample size to aid in better assessing the nature of extremal observation(s).
intervariable changes in the sample dataset. The Scout is a multivariate Moreover, an adequate decision on a specific treatment should be based on the
analysis package of the US-EPA (Singh, 1993) and can be used for computa- results of analysis obtained from cases of the “with” and “without” outlier(s) in
tion of the Mds. One of the major functions of this package is outlier(s) the dataset. In some cases, the removal or preservation of outliers in the dataset
detection and utilizes both the classical and the robust approaches to outlier does not affect the design outcome. The International Organization for
analysis. Standardization also recommends that outlier(s) should not be excluded purely
• Chauvenet’s criterion: This criterion, based on the normal distribution, based on statistical evidence, and further investigations for discrepancies are
advises the rejection of an extreme observation if the probability of occur- required. However, it is not always possible to know the exact cause of such
rence of such a deviant observation from the mean of the n measurements is discrepancies in individual data values.
less than ½ n. Obviously, the choice of ½ n is arbitrary, but is reasonable and Basically, three approaches for the outlier treatment are available for consid-
can be validated (Taylor, 1984). Nevertheless, some analysts disagree with eration, such as outlier exclusion, outlier revision, and outlier inclusion. The
such a statement. Analysts at the U.S. National Bureau of Standards (Natrella, exclusion of outliers is particularly selected when the data are suspect to contain
1963) point out that such a criterion easily rejects when the sample size, n, is erroneous observations. In some cases, outliers can be excluded because the
too small. observation exceeds the limit of measuring equipment. However, most of the
• Extreme studentized: The extreme studentized method is similar to the time, the nature of observed outliers is still uncertain. Frequently, engineering
Chauvenet’s criterion. Instead of using normal error integral distribution, it practitioners tend to arbitrarily adopt the treatment of excluding outliers that
uses t-distribution. And it also assumes that extreme observations in either can enhance the fitness to a specific model (e.g., linear regression and normal
direction are considered rejectable. Population mean is unknown and stan- distribution). Some scientists or engineers may suggest a conservative treat-
dard deviation is assumed to be the sample standard deviation. The test pro- ment concept that data should not be excluded unless a reasonable doubt does
cedure is to: (1) choose α (the probability [or risk]), a user is willing to accept exist. In that case, modification can be an approach to reduce the impact of
in rejecting an observation that really belongs into the group. (2) Look up the extreme data and to facilitate the analysis. However, the degree of modification
q1 − α (n, ∞) value from the Standard Studentized Deviate Table with n being is hard to justify. The inclusion approach is to adopt only those statistical meth-
the number of observations in the sample, and v being the degrees of free- ods that are robust enough to manage the extreme data without being unduly
dom. (3) Compute w = q1 − α * σ and (4) if (Xn − X1) > w, reject the observation biased by individual outliers. Such an approach is getting acceptance in hydro-
that is a suspect; otherwise, retain it (Natrella, 1963). logical analyses (e.g., L-moment and nonparametric methods).
The use of the t-distribution, instead of the normal distribution, makes it
more realistic in reflecting the use of the sample distribution rather than
16.3 CONCEPTS AND METHODS FOR HANDLING
population distribution. This is especially true for cases where the sample size
MISSING RECORDS
is less than 30.
• Dixon method: The objective of the Dixon method (Dixon, 1951) is to Conceptual understanding of a missing record satisfying the varying nature
investigate the significance of the difference between a suspicious extremal of its interpretations by various users/investigators is challenging because the
observation and other observations in the sample. The population that is being significance and complexity of a missing record varies according to the
16_Singh_ch16_p16.1-16.8.indd 3 8/22/16 11:47 AM

time-scale or space-scale usage of a data series. In hydrologic data, missing cases of highly correlated data with small gaps. However, in cases where esti-
records have a common occurrence and can have a significant impact on the mation of missing records exceeds the allowable sequential correlation regime
conclusions that can be inferred from the data. The existence of missing (or the dataset is seemingly random-like), then traditional estimation
records is not uncommon whenever and wherever a hydrologic time series approaches may encounter difficulties for infilling missing records. Thus,
(or a set of observations) is recorded. Missing records in such time series overreliance on temporal correlations in the estimation process may not be
represent a great challenge for statistical analysis as a complete dataset is appropriate.
normally required. Understanding the occurrence and nature of missing For infilling of missing records in daily precipitation data, the probability
records, and statistical properties of the observed data are essential for the distribution and spatial dependence among nearby sites are normally more
infilling of missing records. relevant than the temporal dependence of precipitation series. It is because
the temporal dependence normally suffers from the adverse effect of tempo-
16.3.1 Definition and Types of Missing Records ral intermittence, which generally is more apparent than the adverse impact
induced by the spatial intermittence. Therefore, the missing records of one
A missing observation or missing data (or record) in a particular spatial or site can be estimated through the correlative relationships expressed in
temporal location for a variable signifies the nonexistence of a value either in parameters of a multivariate normal distribution model. The missing records
numerical or any other format. can also be infilled or estimated through a conditional sampling obtained
The categorization of missing records into three basic types by Little and from the distribution model.
Rubin (1987) is well accepted in the literature, namely: (1) Missing Completely Most streamflow infilling and/or of records extension procedures are based
at random (MCAR): this is the highest level of randomness. It occurs when on ordinary least-squares (OLS) regression equations that are used to estimate
the probability of an instance having a missing value for an attribute does not individual missing values at stations with short-term records. OLS regression,
depend on either the known values or the missing values. At this level of unfortunately, results in extended records with smaller variances than the his-
randomness, any missing records treatment method can be applied without torical records. Techniques other than OLS regression that can be used to esti-
risk of introducing bias in the database. (2) Missing at random (MAR): the mate missing values and that tend to preserve the variance of the historical
probability of an instance having a missing value for an attribute depends on records include regression plus noise (Matalas and Jacobs, 1964) and two
the known values and not on the value of the missing observation itself. That alternatives to regression described by Hirsch (1982), which are referred to as
is, the missing observation exists due to some external influences, which can MOVE-l and MOVE-2 (Maintenance of Variance Extension, Types 1 and 2).
be explained by other measured variables in the study. The pattern of missing Alley and Burns (1983) developed a procedure capable of selecting any of the
records is considered to be traceable or predictable. (3) Non-ignorable (NI): above mentioned procedures for infilling of missing streamflow records, which
when the probability of an instance having a missing value for an attribute is routinely utilized by the United States Geological Survey (Parrett et al., 1989
depends on the value of that attribute, then the pattern of missing observa- and Vining et al., 1996.) The procedure selects the best base station from sev-
tions is nonrandom and is not predictable from other variables in the data- eral base stations and also has an option that allows for consideration of both
base. Missing observations can only be explained by the very variable(s) for cyclic (monthly) and noncyclic (annual) extension equations for each individu-
which the observation is missing. al prediction. With the cyclic (monthly) equation, missing data are estimated
In fact, the MCAR assumption is not easy to meet. Most of the time, MAR from streamflow data for the selected month; with the noncyclic (annual) equa-
is the only assumption being met in practical cases. Even such a widely tion, missing data are estimated from all concurrent streamflow data for the
accepted assumption may not always be true because missing values on a period of record. Specifically, the missing data for a particular month at a site of
number of variables and the assessment of dependence on other variables interest are estimated using both the cyclic (monthly) equation and the noncyc-
may be considerably complex. The purpose of understanding the cause and/ lic (annual) equation, and a standard error of prediction is computed for both
or nature of the missing records is to facilitate the selection of a suitable estimates. The equation that has the smallest standard error of prediction then
method and ensure compliance of the selected method with the underlying is used to estimate the missing data. Similar methods including the kriging
assumptions. interpolation method have been utilized in Canada, among others by Ontario
Ministry of Natural Resources (2008) for addressing the issues of missing
16.3.2 Approaches for Handling Missing Records records in climatic data. Some of the most commonly used methods for hand-
There exist three primary approaches for handling missing records such as: ling missing records are presented as follows.
(1) Omitting Missing Records: Procedures based on complete records are
under this category. When some variables are not recorded, such an approach 16.3.3 Methods for Handling of Missing Records
simply discards incomplete records and only analyzes the complete records. There exist numerous methods for handling of missing records which may vary
This is a relatively easy approach to carry out and deemed satisfactory with from ad hoc methods to involving linear or nonlinear regression approach.
small amounts of missing records. However, it is not usually efficient because Such methods can be broadly grouped into two categories. The first category
it is not able to fully utilize the available information and thus can lead to involves methods based on assumptions of normal probability distribution or
serious biases. (2) Infilling Missing Records: The missing records are routinely temporal dependence in a univeriate time series. The methods in the second
infilled and the resultant completed data are analyzed by standard methods. category are developed based on assumptions of spatial dependence among
(3) Accommodating Missing Records: Based on a predefined model, a broad observations at multiple sites using multivariate time series. Some of the most
class of parameters estimation procedures is developed using the dataset commonly employed methods are briefly described as follows.
containing missing records (without infilling). This approach is flexible and
avoids the use of ad hoc methods. Methods Based on the Assumption of Normal
In hydrology, many studies concerning handling of missing records in Probability Distribution or Temporal
streamflow or precipitation data rely on the association between predictor Dependence (Category-I)
and predictant variables. Basically, missing records can be estimated based on The following methods are commonly used for infilling of missing records in
assumptions related to probability distribution, spatial dependence, temporal datasets exhibiting strong temporal dependence (e.g., daily streamflows data),
dependence, or other forms of dependence. For instance, missing records can or datasets complying the underlying assumption of normal distribution (e.g.,
be estimated using the neighboring observations within a time series itself annual maximum peak flow data). These methods are applicable to different
(Elshorbagy et al., 2000b) or the spatial observations from nearby sites (Ng et observations of hydrologic variables as long as the nature and characteristics
al., 2009). For infilling purposes, it is normally preferable to use the neighbor- of variables obeying the underlying assumption. Additional detail on these
ing observations exhibiting higher correlation with missing records. classical statistical methods can be found elsewhere (Little and Rubin, 1987).
For infilling missing records in daily streamflow data, the temporal depen- • List-wise or case-wise data-deletion method: This method omits the
dence normally plays a significant role compared to the probability distribu- entire missing record from the analysis. Such an operation may result in
tion of the dataset. It is because the behavior of streamflow evolves reduction of the sample size for analysis, but such an approach is generally
continuously in a time series and the strength of embedding autocorrelation acceptable if the dataset is large enough and the number of missing records is
is more apparent. Therefore, the missing records of one site can be estimated relatively small.
through correlation relationships of the univariate time series itself (i.e., • Pair-wise data-deletion method: This method computes the statistics
through an autocorrelation model) or from corresponding data of a nearby based on the available pair-wise data, such as that used in cases of bivariate
stream(s), or of a gauging station located either upstream or downstream on correlation or covariance. Similar concerns as described for the list-wise dele-
the same river. tion are also applicable in this method.
A typical approach for infilling missing records may involve an estimation • Mean substitution method: This method substitutes a mean, mode, or
based on linear regression models, which may perform reasonably well in median value computed from available cases to infill missing records on the
16_Singh_ch16_p16.1-16.8.indd 4 8/22/16 11:47 AM

Discussion and Concluding Remarks on methods for outliers and infilling of missing records 16-5
remaining cases. This method is simple, but may underestimate the variance • Kriging technique: Kriging is a spatial interpolation technique that
of the selected variable. Also, in a trendy time series, this method may miss a arrives at interpolated surfaces with minimum error. The technique is used
number of valuable coherent signals around the missing records. regularly in geotechnical analysis and also utilized in hydrologic analysis,
• Last value/observation carried forward method: This is a simple method among other by Yates et al. (1986), Dingman et al. (1988), and Seo et al.
as its name implies, but may only be reasonable when the data set is consistent (1990). The advantage of kriging over OLS techniques is that the estimated
and stable. values at observation sites are equal to the actual measurement (Gebhardt et
• Hot deck imputation method: This method identifies the case that is al., 1988). In kriging technique, the precipitation at a site is considered as a
most similar to that with a missing record and substitutes the most similar function of a predictable trend and a random component. Kriging may treat
known case value for the missing record. the trend as a constant (simple kriging), or may describe the trend as a poly-
• Linear model: An equation is developed based on the available data for nomial (universal kriging). A difference in variance of measurements
a given variable. An improvement of this method may involve uncertainty in between rain gages is related to the distance between rain gages, called a
the imputed record. The model can be static or dynamic (e.g., ARIMA, expo- semivariogram. The direction (of the distance) is assumed not to affect rain
nential moving average, OLS model, etc.). This method is acceptable in cases gage values.
of trendy series or in the existence of a significantly clear temporal pattern. The above mentioned categories are a typical classification of different
However, with cases of noisy data, infilling may be difficult and a nonlinear missing records handling methods/techniques. Some of the methods stated
model may be required. earlier can fall into more than one category. The selection of a method should
• Nonlinear model: Mathematical models where the exponent is not equal still be determined in view of the nature of missing records and the type of
to 1 [e.g., artificial neural networks (ANNs), etc.] are used. data structure. For example, the selection of a method may remain inconclu-
• Expectation Maximization (EM) algorithm: The EM algorithm is an sive for estimating missing daily precipitation due to varied correlations arising
iterative procedure with two discrete steps. In the expectation (E) step, one from considerations of neighboring observations, the period and the length
computes the expected values of the log-likelihood of the available data. In of the dataset with missing records, the distance between sites, and the impact
the maximization (M) step, one substitutes the expected values for the miss- of temporal/spatial intermittence.
ing records obtained from the E step, and then maximizes the likelihood
function as if no data were missing when obtaining the new parametric
estimates. The procedure iterates through these two steps until convergence 16.4 DISCUSSION AND CONCLUDING REMARKS
occurs. ON METHODS FOR OUTLIERS AND INFILLING OF
• Data-segmentation method: This method groups data into segments MISSING RECORDS
based on assigned common patterns or criteria and then substitutes the There exists numerous methods for the consideration; however, the selec-
missing records value using statistical information (e.g., group mean) of the tion of a proper method to handle the problem of outliers or missing
segment. records would be a challenge as many criteria may have to be considered.
• Multiple imputation approach: Instead of the single imputation For instance, the selection of an infilling method depends on criteria, such
approach, the multiple imputations approach is gaining increasing attention as the record length (observed and missing), the presumed type of missing
and acceptance because the results of such analyses are further extendable to data, the structure of the observed data (any spatial and temporal depen-
other statistical analyses. One advantage of the multiple imputations approach dence established), and the intent of the use (i.e., effectiveness, efficiency,
is its ability to generate more than one dataset for imputation. Another advan- robustness, accuracy, precision, and parsimoniousness). As this chapter
tage is that it does not underestimate the variability of the infilled records, intends to serve as a comprehensive material for the target readers of the
which is a common problem with the single imputation approach. handbook, a broad discussion relevant to foregoing considerations is pre-
• k-Nearest Neighbor method: It is one of the simplest methods to infill sented in this section; which explores the relevant topics related to outliers
missing records through selecting k number of nearest neighbors (i.e., candi- and missing records along with the outstanding problems concerning the
dates) for the predictor. One of the randomly drawn candidates becomes the existing methods.
observation to infill missing record(s). The nearest neighbor can be selected
either geometrically or by taking sites with highest correlation to the target 16.4.1 Outlier Detection and Treatment
site. The existing methods for outlier detection and treatment as described earlier
perform reasonably in handling outlier problems. However, these methods do
Methods Based on the Assumption of Spatial have limitations in their application because either the dataset does not meet
Dependence (Category-II) the underlying assumptions of the method or the method itself has inherent
The following methods are commonly used for infilling of missing records in problems, such as conceptualization, mathematical or computational short-
datasets exhibiting strong spatial dependence (e.g., precipitation data at mul- comings. To overcome such shortcomings, many investigators endeavor to
tiple sites within a homogenous region). Additional detail on the methods can explore the use of alternative methods.
be found elsewhere (ASCE, 1996; Teegavarapu, 2012). In the case of univariate hydrologic data, outlier detection and treatment
• Thiessen method: The Thiessen method (1911) is for approximating the have received reasonable attention. However, with multivariate hydrologic
distribution area around precipitation gages for the purpose of distributing data, this has not been the case, despite the fact that multivariate analysis is
average precipitation depths over an area. A Thiessen polygon network is well established in statistics. Recently, a pioneering attempt in this direction
manually constructed using perpendicular bisectors to lines between gages has been reported by Chebana and Ouarda (2011). It was found that outliers
and carefully removing overlapping bisectors until an even spatial distribu- have a serious effect on the modeling of hydrologic data in multivariate
tion is obtained. Computerized versions of this method are available (Diskin, settings.
1969; Shih and Hamrick, 1975; Croley and Hartmann, 1985). The identification of outliers has always been a difficult task to undertake
• Isohyetal method: The Isohyetal method is still a useful and fairly accu- due to the existence of a few consistent identification procedures and guide-
rate technique for determining the spatial distribution of precipitation lines for application to water resources as well as many other domains related
(France, 1985). An isohyetal map is constructed by plotting the stations and to the environment and natural resources. There is a need to identify proce-
their appropriate precipitation values on a topographic map of the area. A dures for outliers especially due to the availability of increasing amount of
user judgement decides on the placement of isohyetal contours affected by data in a variety of forms.
topography within the area. Estimated average precipitation is computed by
16.4.2 Infilling of Missing Records
the weighted average for the subject area. Hamlin (1983) demonstrated that
the number of stations and the skill of the analyst make a difference in the The commonly used techniques for the estimation of missing data in hydrol-
resolution of isohyetal maps. ogy, such as streamflows, are based on regression analysis, time series analy-
• Linear and multiple regression methods: These methods offer a straight- sis, ANNs, and interpolation techniques. Despite progress in the development
forward and useful technique in extrapolating precipitation records of an area of methods/techniques for infilling of missing records, some issues remain
with incomplete records. Regression relationships between two or more largely unsettled. Numerous techniques of data infilling can be primarily
stations are developed with the existing data. Their relationships are then classified into two classes: statistical methods and computer intelligence
used to infill missing records. Salas (1993) provided an excellent discussion methods. Regardless of the common usage of statistical methods in handling
on criteria for improving estimators of parameters in details. missing records, there has been a slow, but steady shift toward computational
• Polynomial interpolation method: This method (Tabios and Salas, 1985) intelligence methods even though these methods will take time to attain a
utilizes a polynomial function fitted to the stations. A least squares statistical higher accuracy. Without detailed knowledge of the processes under investi-
treatment resolves the weighting of a station based on coordinates of the station. gation, these types of techniques develop functional relationships between
16_Singh_ch16_p16.1-16.8.indd 5 8/22/16 11:47 AM

input and output data by finding an optimum set of network parameters Ng et al., 2007; and others). Another nonparametric approach based on the
through the training process. ANNs have become one of the most promising Generalized Regression Neural Network (Ng et al., 2007) has also been used
techniques for infilling of missing records (Panu et al., 2000). Invariably, the for infilling of missing records with reasonable success.
ANN-based techniques have been shown to be superior compared to the • Grouping approach: Proper grouping may help generalize relationships
traditional statistical methods for infilling missing records. within grouped elements and also help to specialize relationships among
grouped elements. Grouping may not be able to specialize the accuracy within
16.4.3 Limitations of Existing Methods for Outliers groups, but the generalization of the infilling process may help develop more
and Infilling of Missing Records robust mechanisms for handling data sets containing outliers.
A review of limitations of various methods utilized in handling of outliers and In general, grouping appears everywhere. Some group formations may
missing records is summarized succinctly in general terms as follows. simply be arbitrary and hence not justifiable either hydrologically or
• Univariate analysis: In many instances, it is understood that more than one physically. Other group formations may be physically valid, such as grouping
independent variable is correlated with the dependent variable and consequently into four seasons. Therefore, rational and systematic data organization (e.g.,
effect and/or contribute to resulting value(s) of the dependent variable. Univari- clustering, partition, and data segmentation) is necessary for meaningful and
ate analysis may be simple to implement, but it may underestimate the complex- successful data infilling analyses. The concept of Grouping has been pre-
ity of a system through the omission of impacts of all contributing factors. sented by Panu et al. (1978). Some studies in relation to this concept have also
• Linear regression approach: Linear regression approaches used for the been adopted for the simulation of precipitation approach using the diction-
outlier analysis or handling missing records have proven to be a very simple ary formulations (Ng and Panu, 2010).
but valuable tool for modeling and thus are widely used in all disci- For example, one of the grouping approaches utilizes the concepts of chaos
plines. However, such an approach may be relatively inflexible and restrictive theory in the development of infilling technique for the missing records in
in modeling data with high variability both in time and space. streamflow data (Elshorbagy et al., 2002; Ng et al., 2007; and others). Recently,
• Parametric approach: Many of the methods related to outlier detection the utility and benefits of considering groups in hydrologic datasets for infill-
and infilling of missing records are based on the parametric approach. This ing of missing records have been specified by Harvey et al. (2012).
approach does not effectively deal with the variability in the data, and hence • Accommodation analysis: Accommodation analysis refers to methods of
the assumption corresponding to a parametric approach is subjectively handling outliers or missing records that require neither detection of outliers
selected requiring justification and validity. nor infilling of individual missing records. The methods under this category
• Single-valued approach: The single-valued approaches ignore informa- are considered robust enough to handle problematic data directly without the
tion contained in and among groups of data. A few published reports (Panu need of a single or specific address to the problematic observation or
et al., 1978; Unny et al., 1981) are available in hydrologic literature that discuss sequence of observations. For example, some studies utilize L-moments to
and describe the use of concepts of extracting information from data groups. handle data with outliers without detection. Also, users can directly estimate
• Ad hoc analysis: Ad hoc analysis may be able to address a single and parameters of a distribution using likelihood estimation without the need for
specific problem, but its usefulness may be limited because it lacks diversity infilling the missing records.
in application. A general and robust analysis method is therefore required. • Implication of climate change in outlier analysis or infilling of missing
records: Global climate change directly affects regional precipitation patterns
16.4.4 Handling Outliers or Infilling Missing that subsequently change streamflows, and there is a good correspondence
Records First? between local precipitation/streamflow patterns and atmospheric circulation
Which operation should come first, the outlier analysis or handling missing (Stehlik et al., 2002). It is envisioned that additional considerations relative to
records? The answer to this dilemma is still inconclusive because the result of outlier analysis or handling of missing records are required due to global
each operation may or may not affect the operation of the other. However, four climate change. For example, since more extreme events are envisioned to
potential scenarios may occur: (1) the results of these two operations (i.e., out- occur in the future due to climate change, some outlier detection methods
lier detection and infilling of missing records) do not affect each other, proceed- may become oversensitive in identifying outliers, if more frequent extreme
ing with either of the operations first is not a concern; (2) the existence of events are anticipated to occur in the future. For infilling of missing data,
missing records does not affect the outlier analysis, but not the other way multiple imputation techniques may become more attractive due to their
around, the outlier analysis may proceed first; (3) the outlier does not affect the increased ability to handle the higher uncertainty anticipated to occur in
handling of missing records, but not the other way around, the missing records hydrologic data under global climate change. In addition, the development of
may proceed first; and (4) the results of these two operations may affect each suitable techniques involving the downscaling approach of the variable of
other, thus performing whichever operation first may bias the result of analysis. interest at specific time scale is likely to be advantageous in infilling of miss-
One of the methods to manage the problem is through alternatively conducting ing records.
each operation iteratively until a stable result is obtained. • Data mining, machine learning, and computer intensive computations:
With the advancement of computing technology over the past decades, hydro-
logical analyses have also gone through a significant change. Improved comput-
16.5 FUTURE RESEARCH DIRECTIONS ing hardware and data collection techniques have produced a large data base
Toward addressing some of the outstanding problems mentioned earlier in available for analysis. Data mining plays an important role in using and inter-
Sec. 16.4.3, several potential future research directions are envisioned. Since preting such data. Many machine learning algorithms have also been devel-
the subject matter related to missing data and outliers are rather diverse, the oped; for example, ANN and Genetic Algorithm techniques are gaining interest
future research directions are presented through a broad discussion of rele- among many researchers, scientists, and practitioners in handling missing
vant issues and concerns. records and outlier-related problems. As computing becomes faster and more
• Multivariate analysis: Multivariate techniques allow researchers to inves- efficient, some studies that would not have been feasible earlier, have now
tigate relationships between variables in an overarching way and to quantify become possible due to the availability of high capacity operational systems.
the relationship between variables. They can control association between
variables by using cross tabulation, partial correlation and multiple regres- REFERENCES
sions, and introduce other variables (e.g., output of a radar or a regional cli-
mate model) to determine the links between the independent and dependent Aggarwal, C. C., Outlier Analysis, Springer, New York, 2013.
variables or to specify the conditions under which the association occurs. Alley, W. M. and A. W. Burns, “Mixed-station extension of monthly
This will enhance our understanding and interpretations of the relationship streamflow records: American Society of Civil Engineers,” Journal of Hydrau-
among data and would allow for more powerful and robust tests of signifi- lic Engineering, 109 (10): 1272–1284, 1983.
cance compared to univariate analysis. ASCE Task Committee on Hydrology Handbook of Management Group D,
• Use of nonlinear regression and non-normal distribution: To address the Hydrology Handbook (ASCE Manuals on Engineering Practice No. 28, 2nd
need for flexibility in analysis, the nonlinear regression approach may be ed.), ASCE, New York, NY, USA, 1996.
required. The advantages of using a nonlinear approach utilizing ANN for Bardossy, A. and G. G. Pegram, “Infilling missing hydrological data-meth-
handling missing streamflow records have been demonstrated among others ods and consequences,” AGU Fall Meeting Abstracts, San Francisco, Califor-
by Panu et al. (2000) and Khalil et al. (2001). Also, Bardossy and Pegram nia, USA , 2013, p. 1.
(2013) proposed a method for estimating missing records using a non-normal Barnett, V. and T. Lewis, Outliers in Statistical Data, John Wiley & Sons,
distribution based on copulas. New York, 1978.
• Nonparametric approach: This approach is gaining interest through the Beckman, R. J. and R. D. Cook, “Outliers with discussion,” Technometrics,
k-Nearest neighbor for infilling of missing records (Elshorbagy et al., 2000; 25(2): 161–163, 1983.
16_Singh_ch16_p16.1-16.8.indd 6 8/22/16 11:47 AM

REFERENCES 16-7
Box, G. E. P. and G. M. Jenkin, Time Series Analysis: Forecasting and Control Ng, W. W., U. S. Panu, and W. C. Lennox, “Comparative studies in problems
(Revised edition in 1986), Campbell, San Francisco, CA, 1976. of missing extreme daily streamflow records,” Journal of Hydrologic Engineer-
Chebana, F. and T. B. M. J. Ouarda, “Depth-based multivariate descriptive ing, 14 (1): 91–100, 2009.
statistics with hydrological applications,” Journal of Geophysical Research, 116: Ng, W. W. and U. S. Panu, “Comparisons of traditional and novel stochastic
D10120, doi: 10.1029/2010JD015338, 2011. models for the generation of daily precipitation occurrences,” Journal of
Croley, T. E. II and H. C. Hartmann, “Resolving Thiessen polygon,” Journal Hydrology, 380 (1): 222–236, 2010.
of Hydrology, 76(1): 363–379, 1985. Ontario Ministry of Natural Resources, Climate Data Gap Filling Project,
Dingman, S. L., D. M. Seely-Reynolds, and R. C. Reynolds, “Application of Government of Ontario, Toronto, 2008, p. 37.
kriging to estimating mean annual precipitation in a region of orographic Panu, U. S., T. E. Unny, and R. K. Ragade, “A feature prediction model in
influence,” JAWRA, 24(2): 329–339, 1988. synthetic hydrology based on concepts of pattern recognition,” Water
Diskin, M. H., “Thiessen coefficients by a Monte Carlo procedure,” Journal Resources Research, 14 (2): 335–344, 1978.
of Hydrology, 8(3): 323–335, 1969. Panu, U. S., M. Khalil, and A. Elshorbagy, “Streamflow data infilling
Dixon, W. J., “Ratios involving extreme values,” Annals of Mathematical techniques based on concepts of groups and neural networks,” Artificial
Statistics, 22(1): 68–78, 1951. Neural Networks in Hydrology, Springer, The Netherlands, 2000, Vol. 36, pp.
Elshorbagy, A. A., U. S. Panu, and S. P. Simonovic, “Group-based estima- 235–258.
tion of missing hydrological data: I. Approach and general methodology,” Parrett, C., D. R. Johnson, and J. A. Hull, “Estimates of monthly streamflow
Hydrological Sciences Journal, 45(6): 849–866, 2000. characteristics at selected sites in the upper Missouri river basin, Montana:
Elshorbagy, A. A., U. S. Panu, and S. P. Simonovic, “Group-based estima- base period water year 1937–1986,” U.S. Geological Survey, Water-Resources
tion of missing hydrological data: II. Application to streamflows,” Hydrologi- Investigations Report 89–4082, 1989, p. 109, USGS, Reston, Virginia, USA.
cal Sciences Journal, 45(6): 867–880, 2000a. Pegram, G., “Patching rainfall data using regression methods. 3. Grouping,
Elshorbagy, A. A., S. P. Simonovic, and U. S. Panu, “Estimation of missing patching and outlier detection,” Journal of Hydrology, 198 (1–4): 319–334,
streamflow data using principles of chaos theory,” Journal of Hydrology, 1997.
255(1): 123–133, 2002b. Pilon, P. J., R. Condie, and K. D. Harvey, Consolidated Frequency Analysis
Gandin, L. S., “Objective analysis of meteorological fields (Ed. Hardin, R.),” Package—CFA—User Manual. Water Resources Branch, Inland Waters Direc-
Jerusalem: Israel Program for Scientific Translations, 242, 1965. Reproduced by torate, Environment Canada, Ottawa, Ontario, Canada, 1985.
the Royal Meteorological Society, UK, Quarterly Journal, 92(393): 447, 1966. Rousseeuw, P. J. and A. Leroy, Robust Regression and Outlier Detection,
Gebhardt, K., C. Bohn, and N. Gordon, “Estimating daily precipitation in Wiley, New York, 1987.
the northern Great Basin,” Saval Ranch Research and Evaluation Project Final Salas, J. D., “Analysis and modeling of hydrologic time series,” Handbook of
Report, Bureau of Land Management, Reno, NV, 1988, 4 (25), pp. 4–25. Hydrology, edited by D. R. Maidment, McGraw-Hill, New York, 1993, pp.
Grubbs, F. E., “Procedures for detecting outlying observations in samples,” 19.1–19.72.
Technometrics, 11: 1–21, 1969. Seo, D. J., W. F. Krajewski, and D. S. Bowles, “Stochastic interpolation of
Grubbs, F. E. and G. Beck, “Extension of sample sizes and percentage points rainfall data from rain gages and radar using cokriging: 1. Design of experi-
for significance tests of outlying observations,” Technometrics, 14(4): 847–854, ments,” Water Resources Research, 26 (3): 469–477, 1990.
1972. Shih, S. F. and R. L. Hamrick, “Modified Monte Carlo technique to compute
Gupta, M., J. Gao, C. Aggarwal, and J. Han, “Outlier detection for temporal Thiessen coefficients,” Journal of Hydrology, 27 (3–4): 339–356, 1975.
data: a survey,” IEEE Trans. on Knowledge and Data Engineering, 25(1): 1–20, Singh, A., “Omnibus robust procedures for assessment of multivariate
2014. normality and detection of multivariate outliers,” Multivariate Environmental
Hamlin, M. J., “The significance of rainfall in the study of hydrological Statistics, North-Holland, Amsterdam, 1993.
processes at basin scale,” Journal of Hydrology, 65(1–3): 73–94, 1983. Singh, D. P., Flood frequency modeling and outliers. Organic Geochemistry,
Hawkins, D. M., Identification of Outliers, Chapman and Hall, London, ASCE, New York. 1980.
1980. Singh, K. P. and M. Nakashima, A new methodology for flood frequency
Harvey, C. L., H. Dixon, and J. Hannaford, “An appraisal of the perfor- analysis with objective detection and modification of outliers/inliers, Illinois
mance of data-infilling methods for application to daily mean river flow Institute of Natural Resources, State Water Survey Division, SWS Report #
records in the UK,” Hydrology Research, 43(5): 618–636, 2012. 272, 1981, Urbana-Champaign, Illinois, USA.
Hirsch, R. M., “A comparison of four streamflow record extension tech- Spencer, C. S. and R. McCuen, “Detection of outliers in Pearson type III
niques,” Water Resources Research, 18(4): 1081–1088, 1982. data,” Journal of Hydrologic Engineering, 1: 2–10, 1996.
Hu, S., “Problems with outlier test methods in flood frequency analysis,” Stehlik, I., J. J. Schneller, and K. Bachmann, “Immigration and in situ gla-
Journal of Hydrology, 96(1): 375–383, 1987. cial survival of the low-alpine Erinus alpinus (Scrophulariaceae),” Biological
Iglewicz, B. and D. Hoaglin, “How to detect and handle outliers,” American Journal of the Linnean Society, 77: 87–103, 2002.
for Quality, Milwaukee, WA, 1993, Vol. 16. Tabachnick, B. G. and L. S. Fidell, Using Multivariate Statistics, Harper Col-
Kanji, G. K., 100 Statistical Tests, SAGE, London, 1993. lins College Publishers, New York, 1996.
Khalil, M., U. S. Panu, and W. C. Lennox, “Groups and neural networks Tabios, G. Q. and J. D. Salas, “A comparative analysis of techniques for
based streamflow data infilling procedures,” Journal of Hydrology, 241 (3): spatial interpolation of precipitation,” Water Resources Bulletin, 21(3):
153–176, 2001. 365–380, 1985.
Klemeš, V., “Tall tales about tails of hydrological distributions,” Journal of Taylor, S. J., “Estimating the variances of autocorrelations calculated from
Hydrologic Engineering, 5 (3): 227–231, 2000. financial time series,” Applied Statistics, 33 (3): 300–308, 1984.
Kottegoda, N. T., “Investigation of outliers in annual maximum flow series,” Teegavarapu, R. S. V., Floods in a Changing Climate: Extreme Precipitation,”
Journal of Hydrology, 72 (1–2): 105–137, 1984. Cambridge University Press, Cambridge, UK, 2012.
Little, R. J. A. and D. B. Rubin, Statistical Analysis with Missing Data, Wiley, Thiessen, A. H., “Precipitation for large areas,” Monthly Weather Review, 39
New York, 1987. (7): 1082–1084, 1911.
Matalas, N. C. and B. A. Jacobs, “A correlation procedure for augmenting U.S. Interagency Advisory Committee on Water Data, Guidelines for Deter-
hydrologic data, U.S. Geological Survey Professional Paper 434-E, 1964, p. 7, mining Flood Flow Frequency, Bulletin 17-B of the Hydrology Subcommittee:
USGS, Reston , Virginia, USA. Reston, Virginia, U.S. Geological Survey, Office of Water Data Coordination,
McCuen, R. H., Hydrologic Analysis and Design, 3rd ed., Prentice Hall, 1982, p. 183.
Upper Saddle River, New Jersey, 2004. Unny, T. E., U. S. Panu, C. D. McInnes, and A. K. C. Wong, “Pattern analy-
Montgomery, D. C. and G. C. Runger, Applied Statistics and Probability for sis and synthesis of time dependent hydrologic data,” Advances in Hydrosci-
Engineers, Wiley, New York, 1994. ences, Academic Press, New York, Vol. 12, 1981, pp. 222–244.
Muthén, B., D. Kaplan, and M. Hollis, “On structural equation modeling Vining, K. C., D. R. Johnson, and C. Parrett, “Synthesis of monthly natural
with data that are not missing completely at random,” Psychometrika, 52 (3): flows for selected sites in the Musselshell River Basin, Montana, base period
431–462, 1987. 1929–1989,” U.S. Geological Survey, Water-Resources Investigations Report
Natrella, M. G., Experimental Statistics, National Bureau of Standards 96–4094, 1996, p. 43, USGS, Reston, Virginia, USA.
Handbook 91, U.S. Government Printing Office, Washington, D.C., 1963. World Meteorological Organization (WMO), Report No. 3: The Beaufort
Ng, W. W., U. S. Panu, and W. C. Lennox, “Chaos based analytical tech- scale of wind force, Reports on Marine Science Affairs, WMO, Geneva, 1970.
niques for daily extreme hydrological observations,” Journal of Hydrology, 342 Yevjevich, V., Probability and Statistics in Hydrology, Water Resources Pub-
(1): 17–41, 2007. lication, Fort Collins, CO, 1972.
16_Singh_ch16_p16.1-16.8.indd 7 8/22/16 11:47 AM

Chapter 17
Linear and Nonlinear Regression
BY
KUMER PIAL DAS
ABSTRACT 17.1.2 Assumptions of Linear Regression
Regression analysis is one of the most commonly used statistical methods of Let us take a look at the following assumptions of linear regression:
analyzing hydrologic data. This analysis is used to detect a relation between • A linear relationship exists between dependent and independent vari-
the values of two or more variables, of which at least one of them is indepen- ables. There are a number of ways to check whether a linear relationship
dent in nature. The objective is to test whether there is any statistically sig- exists between two variables. However, creating a scatterplot is the most
nificant relation exist among dependent variable and independent variable(s). convenient way of checking linearity. Steps of creating a scatterplot using
For example, regression analysis can be used for detecting relations between Microsoft Excel is straightforward and we can create a linear regression line
crop growth and water table depth or runoff and rainfall data. This chapter (trend line) as well. When the chart window is highlighted, a Chart menu
describes, with example, simple linear regression model, bivariate regression appears. From the Chart menu, we can add a trend line to the chart. If
model, multiple regression model, and nonlinear regression model. It also the relationship displayed in your scatterplot is not linear, an execution of
provides example related to hydrology and the steps of using MS Excel for linear regression is not recommended. Alternative approaches are: nonlinear
analyzing hydrologic data. regression or polynomial regression.
• Both dependent and independent variables should be measured at the
continuous level (i.e., they are either interval or ratio variables).
17.1 LINEAR AND NONLINEAR REGRESSION
• Statistical independence of the errors means that all errors are indepen-
Regression is perhaps the most popular form of statistical analysis. It can be dent of one another.
used for two purposes: explanation and prediction. The use of regression • Homoscedasticity of the errors can be interpreted as variances of the
analysis (both linear and nonlinear) in hydrology is obvious and evident errors constant over the range of x (e.g., in the case of time series data, it
because in numerous occasions, we wish to investigate how the changes in would be constant over time), versus the predictions and any independent
one or more variable(s) affect each other. These variables may be linked by an variable. If the variance changes over a range of x, which is known as het-
exact straight line relationship or by a nonlinear relationship. This chapter eroscedastic, we need to transform the response variable and possibly the
briefly discusses these relationships that are used in hydrology. predictor variables to achieve a linear and homoscedastic relationship.
• Approximate normality of the error distribution. In other words, we
17.1.1 Simple Linear Regression Model assume that εi ∼ N (0,σ 2 ). Two common methods of checking this assump-
In simple linear regression model, there is only one predictor variable and the tion are: (a) a histogram (with a superimposed normal curve) or (b) a normal
regression function is linear. The model is concerned with relationships of the form: P–P plot.
Response variable = Model function + Random error It should be noted that the most common model used in hydrology is based
on the aforementioned assumptions of a simple linear regression model.
In mathematical terms we write, yi = β0 + β1xi + εi ,
i = 1, 2,…, n (17.1) 17.1.3 Correlation Coefficient
where Correlation measures the dependability of the relationship, it is a measure of
1. yi and xi are the values of the response or dependent variable (y) and how well one variable can predict the other and it is important to know the
predictor or independent variable (x) in the ith trial, respectively. correlation coefficient before we start fitting the regression model. The
2. β0 and β1 are the unknown parameters and usually known as regres- Pearson (or product moment) correlation coefficient is denoted by r and it is
sion coefficients. In particular, β0 is the y-intercept (the point where the a measure of linear association. The value of r does not depend on which of
regression line crosses the y-axis)
and β1 is the slope of the line (the change the two variables under study is labeled x and y . It is also independent of the
in outcome variable y with a unit change in the predictor variable x ). units in which x and y are measured. The value of r is always in between -1
3. Finally, it is assumed that εi is a random error term which is indepen- and +1 [r = 1 can happen if and only if all (xi , yi ) pairs lie on a straight line
dent of xi . Moreover, E(εi ) = 0, and V (εi) = σ 2 . The standard deviation of εi with positive slope and r = −1 is the case when this happens with negative
is constant over the whole range of variation of xi , this property is called slope]. Datasets that follow some nonlinear monotonic function will exactly
“homoscedasticity.” Also, E(εiε j ) = 0 (for all i , j ; i ≠ j , i , j = 1, 2,..., n) which states have r < 1. The square of the sample correlation coefficient provides the value
that the random errors are uncorrelated. of the coefficient of determination (r 2 ) that would result from fitting the
Since the model (17.1) is linear in regard to the parameters β0 and β1 , it is simple linear regression model. The correlation coefficient r can form the
called a simple linear regression model. Unless a model is specifically called basis of a statistical test of independence where the null hypothesis is that yi ’s
nonlinear, it can be considered linear in its parameters, and the word linear is are not dependent from the xi ’s and are identically and independently distrib-
usually omitted. The value of the highest power of a predictor variable in the
r n−2
model is called the order of the model. Note that the model (17.1) is linear in uted normal random variables. The test statistics t is defined as = . The
the independent variable x too and that is why it is also called a first-order 1− r 2
null hypothesis is rejected if t > tα ,n−2.
linear model. 2
17-1
17_Singh_ch17_p17.1-17.10.indd 1 8/22/16 12:03 PM

17-2 Linear and Nonlinear Regression
17.1.4 Fitting the Regression Line to be in standard form because we get b0 and b1 only on the left side of the
The n pairs of measurements (xi , yi ), i = 1, 2,…, n can be plotted as points in following normal equations:
the Cartesian plane. This will result in a scatter diagram. A generic scatter plot
is displayed in Fig. 17.1. The parameters β0 and β1 are somewhat in the nature n (b0 ) + nX (b1 ) = nY , and nX (b0 ) + ∑ Xi2 (b1 ) = ∑ XiYi (17.3)
of population parameters which can never be known exactly and must be
estimated by their corresponding statistics denoted by b0 and b1. One widely By solving the two equations in Eq. (17.3) for b0 and b1, we get
used method known as ordinary least squares (OLS) is a statistical technique
which attempts to find the function which most closely approximates the data b0 = Y − b1 X
(a “best fit”). In other words, it is an approach to fitting a model to the
observed data. In technical terms, the OLS is used to fit a straight line through where
a set of data points, so that the sum of the squared vertical distances (called
∑ i=1XiYi − nXY = ∑ i=1( Xi − X )(Yi − Y ) = SSXY
n n
residuals) from the actual data points is minimized.
From Eq. (17.1) the sum of squares of deviation from the true line, also b1 = (17.4)
∑ i=1Xi2 − nX ∑ i=1( Xi − X )2
n 2 n
SSXX
known as the error sum of squares, can be written as:
n n
Q ( β0 , β1 ) = SSε = ∑εi 2 = ∑( yi − β0 − β1xi )2 (17.2) Here, SSXY is called the corrected sums of cross product and SSXX is called the
i =1 i =1 corrected sum of squares.
Given the estimators b0 and b1, the fitted model is
The method begins with assumed values for b0 and b1 and suppose that the
relation between the independent and dependent variable is given by
yˆ = b0 + b1x = y − b1x + b1x = y + b1 ( x − x ) (17.5)
y = b0 + b1xi; some b0 ’s and b1’s will be better fits than others. Let ŷi = b0 + b1xi
be the value of the y estimated by the regression equation when x has the
value xi ; then the difference between the actual value (yi ) and estimated value The aforementioned estimation procedure is called regressing y upon x and
( yî ) is called the residual or the error denoted by ei (Fig. 17.1). Mathematically, Eq. (17.5) is called the fitted regression line or regression equation. Since equal-
ei = yi − yî = yi − b0 − b1xi . As expected, different b0 ’s and b1’s will cause each ei ity of all xi is not a possibility, it can be claimed that there will be a unique
to have a different value. solution with respect to (b0 , b1 ).
The OLS method requires that we choose our estimates b0 and b1, so that
when substituted for β0 and β1 in Eq. (17.2), we get the least possible value of Example 1 (Rainfall and Runoff)
SSε . We can determine b0 and b1 by using calculus. More specifically, we dif-
ferentiate Eq. (17.2) twice, first with respect to β0 and then with respect to β1 Consider a data set of rainfall and runoff information obtained from Pee Dee
and set those results equal to zero. In other words, River, South Carolina, USA. A sample of annual data are shown as follows. We
will construct a scatter plot for these data and find the fitted regression line
∂Q(β0 , β1 ) ∂Q(β0 , β1 ) for them.
= 0 and = 0
∂β 0 bb
∂β1 bb
0 1 0 1
Table 17.1 Eighteen Observations of Rainfall and Runoff Data
These two partial derivatives produce the following two equations:
n Rainfall Runoff Rainfall Runoff Rainfall Runoff
−2∑( yi − b0 − b1xi ) = 0 Year (X) (in) (Y) (cfs) Year (X) (in) (Y) (cfs) Year (X) (in) (Y) (cfs)
i =1 1966 57.81 668 1972 46.76 915 1978 36.56 776
n
1967 50.89 692 1973 54.46 988 1979 56.49 1170
−2∑ xi ( yi − b0 − b1xi ) = 0 1968 45.14 760 1974 46.82 624 1980 37.3 1160
i=1
1969 28.28 792 1975 54.12 1740 1981 54.43 510
1970 57.99 505 1976 53.35 541 1982 65.2 1360
After some algebraic development, we get the following set of equations
1971 73.06 2010 1977 46.5 770 1983 55.35 984
popularly known as normal equations for estimating β0 and β1 . They are said
(xn, yn)
(xn, y^n)
(xi, yi) y^ = b0 + b1x

b0 = y + b1x
ei = yi – yî ∑(x – x–)(yi – y–)
y b1 = i
(xi, yî) ∑(xi – x–)2
(x3, y3)
(x1, y^1)
(x1, y1)
x1 x2 x3 xi xn
x
Figure 17.1 Scatter plot of n pairs of measurements ( xi , yi ), i = 1, 2,…, n .
17_Singh_ch17_p17.1-17.10.indd 2 8/22/16 12:03 PM

Linear and Nonlinear Regression 17-3
17.1.5 Creating a Scatter Diagram with Excel The left term of the aforementioned equation is the total variation in the
A scatter diagram of the aforementioned dataset is displayed in Fig. 17.2. response Y , the first term in the right side is the variation in mean response,
and the second term is the residual value. In other words, the aforementioned
Scatterplot [runoff against rainfall (1966–1983)]
equation quantifies the basic regression line concepts, Data = Fit + Residual.
2500 If we square both sides of this and sum from i = 1,2,..., n, we obtain
n n n
2000
∑( yi − y )2 = ∑( yî − y )2 + ∑( yi − yî )2 (17.7)
i=1 i=1 i=1
n
It can be shown that the cross-product term, CPT = 2∑( yî − y )( yi − yî )
Runoff (cfs)
1500
i=1
becomes zero. The aforementioned equation may also be written as
1000 SST = SSM + SSE , where, as before, SS stands for sum of squares and T, M, and
E are notations for total, model, and error (residual), respectively. For simple
500 linear regression, the mean square model (MSM ) is defined as MSM =
n
SSM ∑i=1( yî − y )
2
= , where degrees of freedom, DFM = 1, because in the

0 DFM 1
0 10 20 30 40 50 60 70 80
Rainfall (in) model, there is one explanatory variable, x . Similarly, the corresponding mean
n
SSE ∑i=1( yi − yî )
2
Figure 17.2 Scatter plot for rainfall and runoff data (Example 1).
square error (MSE ) is defined as MSE = = which is nothing
DFE n−2
but the estimate of the variance about the population regression line (σ 2 ).
The data described in Example 1 will be used to find the fitted regression line. Note that, each sum of squares is associated with a number known as its DF
18
which indicates how many independent pieces of information involving the n
From the data, we have, n = 18, x = 51.14, y = 942.5, Sxx = ∑( xi − x )2 = 104.09, independent numbers y1 , y 2 ,..., yn are needed to compile the sum of squares.
18 i=1
n
and Sxy = ∑( xi − x )( yi − y ) = 1797.96.
i =1
For example, the SS about the mean, ∑( yi − y )2 needs (n −1) independent
SXY 1797.96 i=1
Thus, from Eq. (17.4) b1 =
= = 17.27 and b0 = y − b1 x = 942.5- pieces and so the sum of squares has (n − 1) DF. So, corresponding to Eq. (6),
SXX 104.09
we can show the split of DF as
(17.27)*51.14 = 59.31 which provides the fitted regression line as:
yˆ = 59.31 + 17.27 x n − 1 = 1 + (n − 2)

We can tabulate for each of the 18 values xi , at which a yi observation is avail- ANOVA calculations are displayed in an analysis of variance table, which
able, the fitted value yî , and the residual yi − yî as in Table 17.2. has the format for simple linear regression according to Table 17.3.
As expected from the definition of the least squares, the residuals sum to
zero. However, in practice, the sum may not be exactly zero due to rounding.
Table 17.3 ANOVA Table for Simple Linear Regression
17.1.6 Analysis of Variance
Analysis of variance (ANOVA) consists of calculations that provide informa- Sum of
tion about levels of variability within a regression model. It also forms a basis Source DF squares Mean square F
for tests of significance. Consider the following identity:
n
n 2
yi − y = ( yî − y ) + ( yi − yî ) (17.6) Model 1 ∑( yî − y )2 MSM = = MSM /MSE
i=1 DFM 1
n
SSE ∑i=1( yi − yî )
n 2
Table 17.2 Observations, Fitted Values, Residuals, and Sum of Squares

Error n−2 ∑( yi − yî )2 MSE =
DFE
=
n−2
i=1
(Example 1)
Obs n
SST
no. yi ŷi yi − yî ( yi − yî )2 ŷi − y ( yî − y )2 Total n −1 ∑( yi − y )2 DFT
i=1
1 668 1,058 -390 152,100 115 13,225
2 692 938 -246 60,516 -4 16
3 760 839 -79 6,241 -104 10,816
4 792 548 244 59,536 -395 156,025 The F column provides a statistic for testing the alternative hypothesis
5 505 1,061 -556 309,136 118 13,924 that β1 ≠ 0 against the null hypothesis that β1 = 0. We compare the ratio
6 2,010 1,321 689 474,721 379 143,641 F = MSM / MSE with the 100(1 − α )% point of the tabulated F (1, n − 2) distribu-
7 915 867 48 2,304 -76 5,776 tion in order to determine whether β1 can be considered nonzero on the basis
8 988 1,000 -12 144 57 3,249 of the data. So, for a 0.05-level test, we would reject H 0 if F exceeds the 95th
9 624 868 -244 59,536 -75 5,625
percentile of F (1, n − 2). In other words, when the numerator (MSM ) is large
relative to the denominator (MSE), then the ratio is large and there is evidence
10 1,740 994 746 556,516 51 2,601
against the null hypothesis.
11 541 981 -440 193,600 38 1,444
12 770 862 -92 8,464 -80 6,400 17.1.7 Inferences Related to Linear Regression
13 776 691 85 7,225 -252 63,504
Linear regression attempts to model the relationship between two variables by
14 1,170 1,035 135 18,225 92 8,464
fitting a linear equation to observed data. Thus, often we want to make infer-
15 1,160 703 457 208,849 -239 57,121
ences about β1 , the slope of the population regression model. For example, we
16 510 999 -489 239,121 57 3,249 may test:
17 1,360 1,185 175 30,625 243 59,049
18 984 1,015 -31 961 73 5,329 H 0 : β1 = 0 (X and Y are not linearly related)
Sum 2,387,820 559,458
H1 : β1 ≠ 0 (X and Y are linearly related)
17_Singh_ch17_p17.1-17.10.indd 3 8/22/16 12:03 PM

βˆ1 19% of the variation. This dataset indicates that the correlation between rain-
The test statistic for the aforementioned hypothesis is defined as t = fall and runoff may depend on other characteristics. It is our understanding
s(βˆ1 ) that there are a few other characteristics in this example that may influence
which follows the t distribution with (n − 2) DF. Alternatively, we can test the relationship between rainfall and runoff.
whether x and y are positively linearly related (β1 > 0) or x and y are nega-
tively linearly related (β1 < 0). Even though we have discussed two different
tests (a t-test and an F-test) for the test of H 0 : β1 = 0 versus H1 : β1 ≠ 0, we can 17.2 MEASURES FOR GOODNESS OF FIT
Sxy
show that these two tests are equivalent due to the mathematical fact that By using the definition of correlation coefficient (r = ), the regression
F (1, v ) = {t (v )}2 . equation can also be rewritten from Eq. (17.5) as: σ x .σ y
For example, using the data mentioned in Example 1, we can test whether
rainfall is linearly related to runoff by testing the aforementioned hypothesis. σy
yˆ − y = r (x − x ) (17.8)
We have used the following steps in MS Excel 2013 to analyze the data in σx
Example 1:
1. Go to the Data tab and click Data Analysis (if you do not have Data By squaring the aforementioned expression and taking the expected value
Analysis button, you may need to load the Analysis ToolPak Add-In). of the squares, the error variance can be written as:
2. Select Regression and click OK. σ ε2 = σ y2 (1 − r 2 )
3. In the resulting dialogue box, select the Y Range (A1: A 20). This is
the predictor variable.
17.2.1 Inference on Regression Coefficients
4. Similarly, select the X Range (B1:B20). These are the explanatory
variables. In order to place confidence intervals on β0 and β1 and to test hypothesis
5. In the same dialogue box, click the Confidence Level box and enter concerning them, it is essential to know the variance of b0 and b1 which will be
the desired level of confidence in the corresponding cell to the right. denoted by σ b20 and σ b21 , respectively. Considering the normality assumption
6. Select an Output Range. of random error [εi ~ N (0,σ 2 )], for all i independent, the inference for the
7. Click Residuals. slope (β1 ) can be obtained using Eq. (17.4):
8. Click OK. n n
The resulting output obtained from Excel has been displayed in Table 17.4. ∑i=1(xi − x )( yi − y ) = ∑i=1(xi − x ) yi = n k y
For example, using the data mentioned in Example 1, we can test whether b1 = n n ∑i i
rainfall is linearly related to runoff by testing the aforementioned hypothesis. ∑i=1(xi − x )2 ∑i=1(xi − x )2 i=1
We have used the following steps in MS Excel 2013 to analyze the data in
Example 1:
1. Go to the Data tab and click Data Analysis (if you do not have Data ( xi − x )
where ki = n and yi ~ N (β0 + β1xi ,σ 2 ).
Analysis button, you may need to load the Analysis ToolPak Add-In) ∑ j=1(x j − x )2
2. Select Regression and click OK.
3. In the resulting dialogue box, select the Y Range (A1: A20). This is the  
 σ2  σ2
predictor variable. b
It follows that 1 ~ N β
 1 , n
2
 . Thus, σ b1 = n which
4. Similarly, select the X Range (B1:B20). These are the explanatory  ∑ ( x j − x )2  ∑ j=1(x j − x )2
 j=1 
variables.
5. In the same dialogue box, click the Confidence Level box and enter s
indicates that SE (b1 ) = .
the desired level of confidence in the corresponding cell to the right. n
6. Select an Output Range. ∑ j=1(x j − x )2
7. Click Residuals.
8. Click OK. b1
The resulting output obtained from Excel has been displayed in Table 17.4. If the model is correct, then the quantity is distributed as a t distri-
SE(b1 )
b1 − β1
Table 17.4 ANOVA Table for Example 1 (Rainfall and Runoff Data) bution with (n − 2) DF. In other words, ~ tn−2 . The confidence limits of
SE(b1 )
Summary output β1 can be estimated from
Regression statistics Lβ1 = b1 − tn−2 (1 − α 2) SE (b1 )

Multiple R 0.435591
R square 0.18973985 U β1 = b1 + tn−2 (1 − α 2) SE (b1 )
Adjusted R square 0.139098
Standard error 386.25 Test of the hypothesis concerning β1 can be made by noting that
Observations 18
b1 − β1
~ tn−2 . Thus, H 0 : β1 = 0 versus H1 : β1 ≠ 0 is tested by computing the test
ANOVA SE(b1 )
b
Source df SS MS F Significance F statistic t = 1 under H 0 . H 0 is rejected if t > tn−2 (1 − α 2). A typical
SE(b1 )
Regression 1 558,996.8 558,996.8 3.74 0.070
Residual 16 2,387,126 149,195.4
P-value graph is displayed in Fig. 17.3.
Total 17 2,946,123
The inference for the intercept (β0 ) can be obtained in a similar fashion. In
this situation, we start with the following equation:
Lower Upper
Coefficient SE t-stat P value 95% 95% n n n
yi
b0 = y − b1x = ∑ − ∑ ki xyi = ∑ci yi
Intercept 59.19 465.33 0.1272 0.900367 -927.26 1045.65
i=1
n i=1 i=1
X variable 1 17.27 8.92 1.93 0.07 -1.64 36.19
1 (x − x )
Thus, we estimate with 95% confidence that the runoff in Pee Dee River, where ci = − n i x and yi ~ N (β0 + β1xi ,σ 2 ) which leads us to
n ∑ ( x j − x )2
South Carolina, USA for each additional increase of 1% rainfall will increase j=1
by an amount somewhere between -1.64 and 36.19 cfs. Note that the experi-
mental range of predictor variable X is varied from 28.28 and 73.06 in. Our   
estimate is best when the value of rainfall is within the experimental range. In  2

general, more rainfall should produce more runoff. But the lack of signifi- b0 ~ N  β0 , σ 2  1 + n x 
n 2
cance and apparent nonlinearity of the hydrologic response of the basin can 
  ∑ j=1 j
( x − x ) 
be observed from a low R 2 and higher P value. The model is explaining about
17_Singh_ch17_p17.1-17.10.indd 4 8/22/16 12:03 PM

Multiple Linear Regression 17-5
represents a hyperplane in the (k + 1)-dimensional space of ( x1 , x 2 ,…, x k , y ).

0.30 The regression coefficient (β j ). represents the expected rate of change in the
response variable as x j changes when the remaining predictor variables
x1 , x 2 ,…, x j −1 , x j+1 ,…, x k are kept fixed. Since β j is the expected change in y as
x j changes by one unit while all other independent variables remain fixed, the
0.25 tn–2 regression coefficients β ’s are also referred to as partial regression coefficients.
17.3.2 Estimation of Regression Coefficients

0.20 The OLS method that we have used to fit simple linear regression can be
Density
extended to fit the multiple linear regression model. The least-squares

method proceeds by minimizing the sum of squared deviations of the
0.15 observed yi (i = 1, 2,…, n) from true means obtained from Eq. (17.9) E(yi) =
β0 + β1xi1 + β 2 xi 2 + ..... + β k xik , that is by minimizing
0.10 n
Q(β0 , β1 ,…, β k ) = ∑( yi − β0 − β1xi1 − β 2 xi 2 − ..... − β k xik )2 (17.10)
i =1
0.05
over choices of (β0 , β1 ,…, β k ). As before we use calculus to minimize the
aforementioned Eq. (17.9).
–3 –2 –1 0 1 2 3 We take the partial derivatives of Eq. (17.10) with respect to β0 , β1 ,…, and
β k and equating them to zero. We will have the following (k + 1) equations:
tobs
 ∂Q n
P value = Area in the two tails 
 ∂β 0
(
= −2 ∑ yi − βˆ0 − βˆ1xi1 − βˆ2 xi 2 − ..... − βˆk xik = 0
i=1
)
Figure 17.3 A typical P-value graph.

n
 
 ∂Q
(
ˆ ˆ 2 ˆ ˆ
 ∂β = −2 ∑ yi xi1 − β0 xi1 − β1xi1 − β 2 xi1xi 2 − ..... − β k xi1xik = 0
1 i= 1
)
1 x2  
Thus, σ b20 = σ 2  + n , which indicates that SE(b 0) =  ……………
n ∑ ( x j − x )2   n
 
j=1 
 ∂β
k i=1
(
 ∂Q = −2 ∑ y x − βˆ x − βˆ x x − βˆ x x − ..... − βˆ x 2 = 0
i ik 0 ik 1 i1 ik 2 ik i 2 k ik )
  
s 1 + x2 . 
n n 2
 ∑ (
j=1 j
x − x )  (17.11)
The basis for confidence interval and hypothesis tests can be defined as The aforementioned set of equations can also be converted into standard
b0 − β0 form, referred to as normal equations, as follows:
~ tn−2 . The confidence limits of β1 can be estimated from
SE(b0 )
n n n n
Lβ0 = b0 − tn−2 (1 − α 2) SE (b0 ) nβˆ0 + βˆ1 ∑xi1 + βˆ2 ∑xi 2 +…+ βˆk ∑xik = ∑yi
i=1 i=1 i=1 i=1
n n n n n
U β0 = b0 + tn−2 (1 − α 2) SE (b0 ) βˆ0 ∑xi1 + βˆ1 ∑xi12 + βˆ2 ∑xi1xi 2 +…+ βˆk ∑xi1xik = ∑xi1 yi (17.12)
i=1 i=1 i=1 i=1 i=1
Thus, H 0 : β0 = 0 versus H1 : β0 ≠ 0 is tested by computing the test statistic
b0 ……………
t= under H 0 . H 0 is rejected if t > tn−2 (1 − α 2). The significance of the
SE(b0 ) n n n n n
overall regression equation can be evaluated by testing the above hypothesis βˆ0 ∑xik + βˆ1 ∑xi1xik + βˆ2 ∑xik xi 2 +…+ βˆk ∑xik 2 = ∑xik yi
(H 0 : β0 = 0 ) too. If the hypothesis is accepted, then ŷ = y or the regression i=1 i=1 i=1 i=1 i=1
line does not explain a significant amount of the variation in y. In this situa-
tion, we recommend using y as an estimator of y regardless of the value of x .
Solving the above system of (k + 1) equations for the (k + 1) unknowns
17.3 MULTIPLE LINEAR REGRESSION
βˆ0 , βˆ1 ,…, and β̂ k , we obtain the least square estimators for β0 , β1 ,…, and β k .
17.3.1 Introduction
Quite often we wish to model the response variable as a function of several
other quantities in the same equation because it is reasonable to have a 17.3.3 An Example Related to Hydrology
response variable being dependent on more than one quantity. For example, Example 2
the peak rate of runoff from watersheds in a given area may be dependent Consider a similar example discussed in the simple regression model, where
on the rainfall, watershed area, and slope of the mainstream. Multiple monthly runoff is likely to be dependent on the rainfall in the same month
regression is used to model situations like these where the objective is to and in the previous months. Consider the following set of hypothetical data.
build a model that relates a dependent variable to more than one indepen- To analyze these data using Excel, we proceed exactly as we did in the
dent variable. Multiple regression analysis is often of practical value because simple linear regression example. The only change here is to include more
it can yield important insights into data dynamics. The general multiple than one column in the Input X Range. Note, however, that the regressors
linear regression model is an extension of model discussed in Sec. 17.1: must to be in contiguous columns. Moreover, it should be noted that Excel
restricts the number of regressors. Excel calculates the values of β0 , β1 , …, β k
yi = β0 + β1x1 + β 2 x 2 + ,..., + β k x k + ε , (17.9) to make the predictions from Eq. (17.9) as accurate (in the sense of minimiz-
ing the sum of squared errors) as possible.
where, y is a dependent variable, x1 , x 2 ,..., x k are independent variables and
β1 , β 2 ,..., β k are unknown parameters. This model is linear in the parameters β j . 17.3.4 ANOVA for Multiple Regression
As before, ε is a random error incurred when observing y at ( x1 , x 2 ,…, x k ). The ANOVA calculations for multiple regression are nearly identical to the
Also, E(ε ) = 0 and V (ε ) = σ 2 . The multiple linear regression model (17.9) calculations for simple linear regression, except that the DF are different,
17_Singh_ch17_p17.1-17.10.indd 5 8/22/16 12:04 PM

Table 17.5 Fifteen Observations of Rainfall and Runoff Data significant which is difficult to justify physically. But again note that this is a
trial dataset. In a situation like this, stepwise regression, which is a semiauto-
Runoff Rainfall in Rainfall in Rainfall in mated process of building a model by successively adding or removing vari-
month (i) month (i) month (i–1) month (i–2) ables based solely on the t-statistic of their estimated coefficients can be used.
(y) (x1) (x2) (x3) The regression coefficients table is very informative. The column
Month (i) (mm) (mm) (mm) (mm) Coefficients gives the least squares estimates of β j and the column SE gives
1 948 680 452 562 the standard errors of the least square estimates b j of β j . The column t-stat
2 940 662 504 247 gives the computed t-statistic H 0 : β j = 0 against H a : β j ≠ 0 . As expected, the
3 940 588 422 258 Column P value gives the P value for test of H 0 : β j = 0 versus H a : β j ≠ 0 . For
4 945 557 528 323 this example, this equals the P{ t > t − stat}, where t is a t-distributed random
5 941 584 547 228 variable with 12 DF and t-stat is the computed value of the t-statistic given in
6 957 852 515 525 the previous column. It should also be noted that this P value is for a two-
7 935 527 512 260 sided test, and it must be adjusted in case of a one sided test. Columns Lower
8 952 574 545 615 95% and Upper 95% provide a 95% confidence interval for β j . In summary,
9 943 689 325 315 the regression output provides us a fitted model for the aforementioned
10 935 514 495 254 example as follows:
11 945 591 542 654
y = 912.5 + 0.03x1 + 0.003x 2 + 0.023x 3
12 941 459 410 412
13 935 312 478 528
14 950 600 548 605 17.4 NONLINEAR REGRESSION
15 945 745 565 445 17.4.1 Introduction
The linear least squares method fits a straight line or a flat plane to a set of
since the number of explanatory variables is more than one. In the multiple data points. However, it is possible that the true relationship we want to model
regression case, the F column provides a statistic for testing the alternative is curved rather than flat. In this situation, we use nonlinear regression where
hypothesis that at least one of the parameters, β j ≠ 0 for j =1, 2,..., k against the we discuss models which are nonlinear in the parameters. There are many
examples in hydrology where nonlinear regression is used. For example, daily
MSM
null hypothesis that β1 = β 2 = ... = β k = 0. We compare the ratio F = with evaporation can be computed by nonlinear regression with temperature, solar
MSE radiation, wind velocity, relative humidity, and sunshine hours. In this section,
the 100(1 − α )% point of the tabulated F (k , n − k − 1) distribution. When the we present some commonly used nonlinear regression models and discuss
numerator (MSM ) is large relative to the denominator (MSE), then the ratio is point and interval estimation for unknown parameters. We also illustrate how
large and there is evidence against the null hypothesis. It must be noted that the nonlinear regression functions can be transformed to linear regression model.
F-test in multiple regression does not indicate which of the parameters β j for
17.4.2 Some Commonly Used Families of Nonlinear
j =1, 2,..., k is not equal to zero, rather it indicates only that at least one of the
Regression in Hydrology
parameters is linearly related to the response variable.
SS While a linear equation has one basic form, nonlinear equations can take
The ratio M is an important quantity and is denoted by R 2 . It is known many different forms and that is why nonlinear regression covers many dif-
SST
ferent forms. A few popular nonlinear equation forms are:
as the squared multiple correlation coefficient.
1. An exponential growth (decay) equation is defined as Yi = β0e β1xi + εi ,
As before, a measure of quality of the regression is the F statistic. Formally,
where β1 is the growth (decay) rate and εi is the error term. The equation states
this F statistic tests the following null and alternative hypothesis.
that the response variable is growing (or shrinking) at a constant rate of β1 .
H0 : b1 = 0, b2 = 0, . . . , bk = 0
2. A constant-elasticity equation can be defined as Yi = β0 xiβ1 + εi . This is
H1 : At least one of β1 , β 2 ,..., β k is not zero.
primarily used for demand curves equation, where x is some continuous vari-
It must be noted that β0 is not involved in this test. The fraction of the
able that is always larger than 0, β0 determines the scale, and β1 is the elastic-
variation in Y explained by the regression is denoted by R 2 and is defined as
ity of y with respect to x. As before, εi is the error term and has a mean of 1
SS
R 2 = Model . and is always larger than 0. The equation states that the elasticity of y with
SSTotal respect to x is constant, β1 .
Understanding and interpreting the regression output is very important. R2 3. Sigmoidal curves (S-shaped curves) arise in various applications,
is of greatest interest of all. R 2 = 0.796569375 means that 79.7% of the varia- including engineering, signal detection theory, and physiological or pharma-
tion of runoff ( yi ) around its mean ( y ) is explained by the regressors x1 , x 2 , cological studies. Various types of growth data often correspond to sigmoidal
and x 3 . The standard error (SE) refers to the estimated standard deviation of curve. For example, the weekly water/air temperature relationship can be
the error term. It is sometimes called the SE of the regression. The column described by a continuous S-shaped curve. Some of the nonlinear regression
labeled F gives the overall F-test of H 0 : β1 = 0, β 2 = 0,..., β k = 0 versus H1 : At functions that have been used in such situations include the Gompertz model,
least one of β1 , β 2 ,..., β k is not zero. The column labeled Significance F has the Logistic regression model, Morgan–Mercer–Flodin model, Weibull-type functions,
associated P-value. Since 0.000403 < 0.05, we reject the null hypothesis at − ( β1 +β2 xi )
and Richard’s model. The Gompertz model is defined as: Yi = β0e −e + εi.
significance level 0.05. It should be noted that Excel can be used to calculate
As it can be seen from the mathematical definition, the Gompertz model is a
the P value as well. For example, FINV(0.000403,3,11) = 14.36.
three-parameter model with the parameters β1 , β 2 , and β3 . The logistic regres-
The hypothetical dataset provides a good illustration with high R 2 and
β0
significant regression. However, the dataset also indicates that β 2 is not sion model has three parameters too and can be defined as: Yi = + εi .
1 + e −( β1 +β2 xi )
Table 17.6 ANOVA Table for Multiple Linear Regression
Source DF Sum of squares Mean square F
n
n 2
Model k ∑( yî − y )2 MSM =

DFM
=
1
MSM /MSE
i=1
n
SSE ∑ i=1(Yi − Yi )
n ˆ 2
Error n − k −1 ∑( yi − yî )2 MSE =
DFE
=
n − k −1
i =1
n
SST
Total n −1 ∑( yi − y )2
i=1 DFT
17_Singh_ch17_p17.1-17.10.indd 6 8/22/16 12:04 PM

Nonlinear Regression 17-7
Table 17.7 ANOVA Table for Example 2 (Rainfall and Runoff Data)
Summary output
Regression statistics
Multiple R 0.892507353
R square 0.796569375
Adjusted R square 0.741088295
Standard error 3.268703567
Observations 15
ANOVA
Source df SS MS F Significance F
Regression 3 460.20 153.40 14.36 0.000403
Residual 11 117.53 10.68
Total 14 577.73
Coefficient SE t-stat P value Lower 95% Upper 95%
Intercept 912.5 7.53 121.1 1.52 × 10-18 895.91 929.08
X variable 1 0.03 0.00 4.64 0.00 0.02 0.05
X variable 2 0.003 0.01 0.27 0.79 -0.03 0.03
X variable 3 0.023 0.00 4.03 0.0019 0.01 0.03
This logistic regression model can be used to flood susceptible mapping and transformations of the response variable, the predictor variables, the parame-
risk area delineation (Pradhan, 2009). The Weibull model can be defined as: ters, or any combination of these such that the transformed function is linear in
 X β2
{− i  }
the unknown parameters. In addition, transformations are applied to correct
Yi = β0e  β1 
+ εi . Finally, the Richard’s model has the following definition problems of non-normality or unequal variances. As expected, transformations
need to apply to the original data prior to performing regression and it must be
β0 assumed that the error in the transformed equation has the desired properties
Yi = + εi .
[1 + e −( β1+β2 xi ) ]β3 (normal distribution with mean 0). Many transformation techniques are avail-
able; however, only a relatively small number of them are regularly used:
4. The segmented polynomial model has some use too. It can be defined
(a) Usually transformations are chosen from the “power family” of trans-
as:
formations, where each value is replaced by x p , with p being an integer or
 half-integer. For example, p = −0.5 indicates a reciprocal square root transfor-
β + β x + β 2 xi2 + εi , xi ≤ α
Yi =  0 1 i mation and the data values must all be positive.
 γ 0 + γ1xi + εi , xi > α (b) Log transformation is another common approach. Exponential
Numerous other useful families of nonlinear regression functions exist. We growth or decay equation discussed earlier can be transformed by logarithmic
will consider intrinsically the following form of the nonlinear models: transformations. Antilogarithm technique is also used.
(c) Some other transformation techniques include: polynomial regres-
Yi = f ( xi′ ; β ) + εi (17.13) sion and Box-Cox transformations. Polynomial regression model can be
approximated by the forward selection procedure or backward elimination
where f ( xi′ ; β ) is a nonlinear function relating E(Y ) to the predictor variables,
approach. Box-Cox technique allows one to find an optimum transformation
xi . Moreover, xi is a k ×1 vector of independent variables, β is a p ×1 vector of
of the response variable using maximum likelihood methods.
parameters, and εi′ s are identically and independently distributed random
variables with mean 0 and variance σ 2 . 17.4.5 Example 3 (A Trial Dataset)
To illustrate the fitting and analysis of a nonlinear regression model, we will
17.4.3 Assumptions
use a simple example, where the model has only two parameters and the
The assumptions for nonlinear regression is almost identical to that for linear sample size is reasonably small. Say we would like to develop a regression
regression with an exception that in this case, the regression function is a model for predicting annual maximum series (Y), measured in gallon of a
nonlinear function of the unknown parameters. desired variable based on geographical distances (X), measured in ft.
Observed annual maximum series and geographical distances expected to
17.4.4 Transforming to Linearity follow an exponential decay model (Kjeldsen and Jones, 2009). Consider the
Sometimes, it is possible to transform a nonlinear regression function Yi to a following dataset (Table 17.8) to explore the relationship between X and Y.
linear function. Appropriate transformation techniques have to be used, so that Our objective is to (1) describing this relationship mathematically and (2)
Table 17.8 Thirty-Six Hypothetical Data of Distances and Annual Maxima
Annual Annual Annual

Distances maximum Distances maximum Distances maximum
Catchment (i) (X) (ft) (Y) (gallon) Catchment (i) (X) (ft) (Y) (gallon) Catchment (i) (X) (ft) (Y) (gallon)
1 10 1325 13 300 468 25 821 124
2 14 1256 14 300 1250 26 842 114
3 15 1254 15 330 425 27 900 78
4 20 1200 16 380 389 28 924 69
5 35 1025 17 450 400 29 980 60
6 50 800 18 520 354 30 1021 52
7 70 758 19 530 300 31 1054 45
8 100 854 20 600 258 32 1124 40
9 20 1525 21 650 241 33 1230 41
10 120 652 22 700 201 34 1254 38
11 180 650 23 752 25 35 1289 40
12 250 556 24 760 156 36 1321 35
17_Singh_ch17_p17.1-17.10.indd 7 8/22/16 12:04 PM

Scatterplot of Example 3 Table 17.9 ANOVA Table for Example-3 (Distances vs. Annual Maxima)
1800
Summary output
1600
Regression statistics
1400
Annual maximum
1200 Multiple R 0.959133908

R square 0.919937854
1000 Adjusted R 0.917583085
800 square
Standard error 0.377835964
600 Observations 36
400
ANOVA
200
Significance
0 Source df SS MS F F
0 200 400 600 800 1000 1200 1400
Regression 1 55.77 55.77 390.674 0
Distances
Residual 34 4.85 0.14
Figure 17.4 Scatter plot of distances and annual maxima (Example 3). Total 35 60.63
Lower Upper
Coefficient SE t-stat P-value 95% 95%
creating a model from this dataset. A scatter plot of the data is shown in Intercept 7.10 0.10 69.12 0 6.89 7.31
Fig. 17.4. As it can observed from the scatterplot, the relationship between the X variable 1 -0.0029 0.00 -19.76 0 0.00 0.00
predictor variable and the response variable is exponential in this example
and should follow a two-parameter nonlinear exponential model
(Yi = β0e β1xi + εi ) discussed in (1). Since a model is available, our next step
17.4.6 Parameter Estimation
would be to estimate the regression parameters β0 and β1 .
The aforementioned scatterplot clearly indicates a curvilinear trend to Method of least squares is widely used in nonlinear regression too. Geometric
the data, to be more specific the above plot can be fitted to an exponential illustration and linear approximation can also be used to estimate parameters
function. It is possible to linearize the relationship between X and Y by in nonlinear regression model.
using the log transformation of the data on the y-axis. In other words, we The least square estimates of β defined in Eq. (17.13) is denoted by β̂ and
take the base of the natural logarithms, log e ( yi ). The log-transformed data is the set of parameters that minimizes the sum of squared residuals (errors):
are displayed in Fig. 17.5. A linearized trend can be observed from this log- n
{ ( )}
2
transformed scatterplot, which indicates that a simple linear regression SSE (βˆ ) = ∑ (Yi − f xi′ ; βˆ (17.14)
model can be used to describe the relationship between distances and i=1
annual maximum.
SSE
Mean squared error (MSE) is defined as MSE = and it is an unbiased
(n − p)
estimate of σ 2 . As in linear regression, our objective is to obtain the normal
Scatterplot of log-transformed data (Example 3)
8 equations. To obtain those equations, we consider the partial derivatives of
7 SSE (βˆ ) with respect to each β̂ j , and set them equal to zero. This gives us a
Log annual maximum
system of p equations. Each normal equation is given by

6
5 ∂SSE (βˆ )
n  ˆ 
4
∂βˆ
{
= −2∑ (Yi − f ( xi′ ; βˆ ) }∂ f ∂(xβˆ′; β ) = 0
i (17.15)
j i=1  j 
3
2
These partial derivatives are functions of the parameters and an explicit
1
solution for β̂ cannot be obtained. For example, consider the exponential
0 ∂Y
0 200 400 600 800 1000 1200 1400 growth model discussed earlier: Yi = β0e β1 xi + εi. The derivatives are: i = e β1xi
Distances ∂Yi ∂β 0
β1 xi
and = β0e ( xi ). From Eq. (17.14), the normal equations are:
Figure 17.5 Scatter plot of log-transformed data for distances and annual maxima ∂ β1
(Example 3). n n
∑{(Yi − βˆ0eβ x }eβ x  = 0, and ∑{(Yi − βˆ0eβ x }βˆ0eβ x (xi ) = 0. . Even for
ˆ ˆ ˆ ˆ
1 i 1 i 1 i 1 i
i=1 i=1
We follow the steps in linear regression examples (Examples 1 and 2) this straightforward example, there is no analytic solution for β̂ . The estima-
to model the relationship. We have the following information: n = 36 , tion of parameters usually requires the use of iterative methods (e.g.,
Gauss–Newton algorithm and Levenberg–Marquardt algorithm) and most of
X = 553.2, and Y = 5.5 [note that this is the mean of transformed Y = log e ( yi )].
n
the commonly used statistical software packages provide routines for calcu-
β=
∑i=1( Xi − X )(Yi − Y ) = −0.0029, α = Y − Xβ = 5.5 − (0.0029) ∗ 553.2 = 7.10.
lating βˆ1 , βˆ2 ,…, βˆ p.
n
∑i=1( Xi − X )2
17.4.7 Example 4 (Weekly Water/Air Temperature)
So, Ŷ = α + β X = 7.10-0.0029X. So, our regression equation can be written It has been documented on many occasions [(e.g., in (Maidment, 1993)] that
as log e (Yˆ ) = 7.10-0.0029X which can be written as Yˆ = e7.10e−0.0029 X or the relationship between weekly water and air temperature can be well
explained by a continuous S-shaped curve. Mohseni et al. (Gupta and
Yˆ = 1214.38e−0.0029 X .
Guttman, 2013) proposed the following equation to estimate stream
We can calculate the ANOVA table (Table 17.9) and resulting signifi-
temperature, Ts :
cance with the following quantities: ∑ x =19916 , ∑ y = 197.91, ∑(x − x )2 =
α −µ
6,629,868, ∑( y − y ) 2
= 60.63, ∑( x − x )( y − y ) = −19229, ∑( y − yˆ )2 = 4.85, Ts = µ + (17.16)
1 + e y (β − Ts )
and ∑ ( yˆ − y )2 = 55.75.
17_Singh_ch17_p17.1-17.10.indd 8 8/22/16 12:04 PM

REFERENCES 17-9
where, µ is a parameter used to generate nonzero minimum temperature, α is REFERENCES

the estimated maximum stream temperature, γ is a measure of the steepest
Gupta, B. C. and I. Guttman, Statistics and Probability with Applications for
slope of the function, and β is the air temperature at the point of inflection.
Engineers and Scientists, Wiley, Hoboken, NJ, 2013.
The slope (γ ) is a function of the slope tanθ at the point of inflection and can
Haan, C. T., Statistical Methods in Hydrology, The Iowa State University
4 tanθ Press, Ames, IO, 1977.
be estimated as γ = . The logistic function Eq. (17.16) will produce an
α −µ Kjeldsen, T. and A. J. David, “An exploratory analysis of error components
in hydrological regression modeling,” Water Resources Research, 45: 1–13,
S-shaped curve and can be explained using the nonlinear technique. The
2009.
least square method discussed earlier will be used to estimate the four
Maidment, D. R., Handbook of Hydrology, McGraw-Hill, New York, 1993.
parameters (α , β ,γ , and µ ). Similar argument discussed in Eq. (17.14) will
Mohseni, O., G. H. Stefan, and R. E. Troy, “A nonlinear regression model
lead us to the following:
for weekly steam temperatures,” Water Resources Research, 34 (10):
2685–2692, 1998.
∑εi2 = ∑{Tobs − µ − 1 + eγ (β −T ) }
n n 2
α −µ Pradhan, B., “Flood susceptible mapping and risk area delineation using
(17.17)
i=1 i=1
s
logistic regression, GIS and remote sensing,” Journal of Spatial Hydrology,
9 (2): 1–18, 2009.
Note that, α and µ are the linear coefficients and thus can be estimated
using the normal equations directly. However, the other two parameters
(β and γ ) must be estimated using a numerical approximation method
(e.g., Gauss–Newton algorithm).
17_Singh_ch17_p17.1-17.10.indd 9 8/22/16 12:04 PM

Chapter 18
Time Series Analysis and Models
BY
ÓLI G. B. SVEINSSON AND JOSE D. SALAS
ABSTRACT 18.2 PROPERTIES OF HYDROLOGICAL TIME SERIES
Analysis and modeling of time series of hydroclimatic processes have been Hydroclimatic series are generally defined at discrete time intervals, such as
widely used for planning and management of water resources systems. Time annual and seasonal (e.g., monthly, weekly, daily, and hourly). Seasonal data
series models are capable of modeling complex dependence structure across are usually affected by the annual hydrologic cycle (e.g., monthly data show a
multiple sites and multiple time scales, and may be useful for generating syn- 12-month periodic structure). In addition, hourly data may exhibit distinct
thetic time series, forecasting, evaluating operational rules for water resources diurnal cycle due to Earth’s rotation. Other cycles due to sunspot activity or
systems, sizing reservoirs, detecting trends and shifts, filling in missing data, large-scale atmospheric forcing, such as El-Niño-Southern Oscillation may
and extending short records. This chapter includes a wide range of models and also affect the variability of hydroclimatic data. Time series are said to be
modeling techniques for representing hydrological processes, such as precipi- stationary if the statistical properties of the time series do not change with
tation and streamflows, in continuous and discrete time scales. The emphasis time, that is, the probability distribution of the process is the same at all times.
for continuous models has been around the concept of point processes while Conversely, if any statistical property depends on time then the process is
that for discrete time has been based on short memory and long memory nonstationary with regards to that statistical property. For example, monthly
models such as AR, ARMA, and FARMA including their discrete counter- hydrologic time series, which generally have periodic mean and periodic
parts as well as periodic and product models. In addition, univariate and covariance are nonstationary.
multivariate shifting mean (SM) models are included so as to represent com- Statistical analysis of hydroclimatic time series generally includes estimat-
plex shifting mechanisms of the underlying hydroclimatic processes that may ing basic sample statistics, such as mean, variance, skewness, autocorrelation
arise from the effect of low frequency components of the atmospheric and function (ACF), partial autocorrelation function (PACF), and spectrum.
oceanic systems. Furthermore, the chapter discusses disaggregation and non- Other statistics, such as storage-drought- and flood-related statistics, may be
parametric approaches and some basic issues pertaining to uncertainty and of particular interest for analyses of streamflow series (Salas, 1993; Sveinsson,
nonstationarity. Time series analysis and modeling of hydrologic data can be 2014; Yevjevich, 1972). Basic unbiased sample statistics of time series yt of
used for evaluation of impacts and risks that can arise from changes due to length N, such as the mean, variance, and skewness coefficient are calculated,
natural variability, human intervention, and climate change in hydrologic respectively, as:
time series. N
N ∑ ( yt − y )3
1 N 1 N
18.1 INTRODUCTION y = ∑ yt ; s =
2
∑ ( yt − y ) ; g = ( N −t =1)(
2 1 (18.1)
N t =1 N − 1 t =1 N − 2)s 3
The field of stochastic hydrology has been developed in the early 1950s and
1960s by several hydrologists, such as Hurst, Matalas, Thomas, Fiering, and
Yevjevich, who developed the basic concepts, models, and applications where y is a measure of location, s is a measure of spread, and g is a measure
thereof and inspired the work and contributions by many others in the 1970s, of shape of the distributional properties of the data. Hydrologic series are
1980s, and 1990s, which are included in books and chapters such as Bras and generally autocorrelated although the series of maximum annual values may
Rodriguez-Iturbe (1985), Hipel and McLeod (1994), Loucks et al. (1981), be uncorrelated. The lag-h autocorrelation function ACF(h) is ρh = Ch /C0,
Salas et al. (1980), and Salas (1993). Many of the statistical procedures and where Ch is the lag-h autocovariance function ACVF(h),
time-series models are available in common software packages, such as Ch = E[( yt +h − µ )( yt − µ )]. The corresponding sample estimate rh is simply the
Matlab, R, SAS, SPSS, specialized software for classical time series analysis cross-correlation between consecutive values of the series at lag-h, that is,
(e.g., Brockwell and Davis, 2002), and time-series analysis of hydrologic data,
such as SPIGOT (Grygier and Stedinger, 1990) and SAMS (Salas et al., 2006; ch 1 N −h
Sveinsson et al., 2009). rh = , with ch = ∑ ( yt +h − y )( yt − y ) (18.2)
c0 N t =1
Time-series models may be used for generating (simulating) and forecast-
ing hydrologic processes such as streamflows that may occur in the future.
Synthetically generated series enable estimating the probability distribution of where ch is the lag-h sample autocovariance function and c0 is the maximum
key decision parameters, such as reservoir storage, planning and testing oper- likelihood estimate of the sample variance. The autocorrelation of annual
ating rules, estimating future power output of hydroelectric systems, and streamflow series usually arises from storage effects of the soil, groundwater,
evaluating the performance of an irrigation system. Forecasting data may be and lake storage. The correlogram is a plot of rh versus h, where a fast-decaying
used for operation and management of water resources systems. This chapter correlogram is an indication of a short memory process, whereas a slow-
includes a variety of models and methods developed in literature. Some of decaying correlogram may suggest a long-memory process. The PACF is also
them have been used for modeling complex water systems, such as the a measure of serial dependence and is obtained by repeatedly fitting an
Colorado River, the Great Lakes and St. Lawrence River, and the Nile autoregressive model of order p to the time series with PACF(p) being equiv-
River systems. alent to the autoregressive parameter fp in the fitted AR(p) model.
18-1
18_Singh_ch18_p18.1-18.12.indd 1 8/22/16 12:06 PM

18-2 Time Series Analysis and Models
Approximate confidence intervals for both the ACF and PACF of a white correlated residuals may suggest that a higher order model is needed.
noise process at the α significance level are ± z1–α/2/N0.5 (Brockwell and Davis, Furthermore, synthetically simulated series from the model of the same
2002), where z1–α/2 is the 1–α/2 quantile of the standard N(0,1) distribution, length as the historical series should be capable of approximately reproducing
for example, approximate 95% confidence intervals are given by ±1.96/N0.5. the historical statistical properties of the original time series.
Both ACF and PACF are used to select the order of the autoregressive-moving A number of alternative time-series models are presented and discussed in
average [ARMA] (p, q) model (Sec. 18.5.1). For multiple time series, the the following sections. They include univariate models, such as ARMA and
sample lag-h cross-correlation coefficient across sites i and j is rhij = chij / c0ii c0jj , periodic ARMA (PARMA) (periodic ARMA), fractional ARMA and PARMA
N −h for modeling long-memory series, gamma autoregressive (GAR) models for
1
where chij = ∑ ( yt(+i )h − y (i ) )( yt( j ) − y ( j ) ) , h = …, –2, –1, 0, 1, 2, …., is the lag-h
N t =1
gamma distributed data as well as periodic gamma autoregressive (PGAR),
shifting mean (SM) models capable of modeling processes exhibiting shifts or
cross-covariance estimator and c0ii is the maximum likelihood estimator of the jumps in the mean level, Markov Chain and discrete ARMA and PARMA for
variance at site i. modeling discrete processes, and product models for modeling intermittent
Similarly, for seasonal time series yv,t, where ν = 1,...,N is the year, t = 1,...,w processes. Multivariate versions of the ARMA, PARMA, and SM models are
is the season, and w is the number of seasons per year, the unbiased seasonal also included. In addition, disaggregation models for modeling jointly the
sample statistics are calculated for each season t in the same way as in annual and seasonal process of both single and multiple time series are pre-
Eq. (18.1) earlier. Likewise, the sample lag-h ACF of a seasonal time series is sented as well as nonparametric models as an alternative to parametric mod-
the correlation between yv,t and yv,t–h, that is, els. Furthermore, the use of exogenous parameters in models for short-term
forecasting can improve forecasting performance over models based on past
ch ,τ 1 N observations of the forecasted process. Some of these models are briefly
rh ,τ =
c0,τ c0,τ −h
, with ch ,τ = ∑ ( yν ,τ − yτ )( yν ,τ −h − yτ −h ) (18.3)
N ν =1
reviewed.
Note that the lag-1 season-to-season correlation coefficient r1,1 represents 18.4 MODELING OF CONTINUOUS TIME PROCESSES
the correlation between season 1 of the current year with season w of the
previous year. If the lag-h season-to-season correlations are markedly differ- Alternative models are available for modeling hydrological processes at vari-
ent throughout the year, for example, different dependence structure during ous time scales such as annual, monthly (seasonal), and daily. In principle, if
the spring freshet and during the winter, then it is an indication that a model one can develop a model that can adequately represent the underlying process
with periodic parameters should be used to preserve the season-to-season in continuous time, then one should be able to derive the model at any level
correlations. Note that for preserving both annual and seasonal statistics of of aggregation such as a day or month. For illustration, we will consider mod-
the sample data, disaggregation models must be applied (Sec. 18.8). eling rainfall in which rainfall occurrence in continuous time follows a
Spectral analysis of the time series may be used to detect cyclic compo- Poisson process (Le Cam, 1961). Let the number of storms N(t) arriving to a
nents. The spectral density function can be estimated in various ways by both given point in a time interval (0,t) be Poisson distributed with parameter lt
parametric and nonparametric methods (e.g., Brockwell and Davis, 1991, (l = storm arrival rate). Thus, if n storms arrive in the interval (0,t) at times
2002; Jenkins and Watts, 1968; Percival and Walden, 1993; Salas et al., 1980) t1,...,tn, the number of storms in any time interval T is also Poisson with
and most software packages today include one or more methods for estima- parameter lT. Assume further that the rainfall amount R associated with a
tion of the spectrum. Similarly, for slowly decaying correlograms (long- storm arrival is white noise, that is, rainfall amounts r1,...,rn are associated with
memory processes), the Hurst slope (Koutsoyiannis, 2002; Salas, 1993) should storms occurring at t1,...,tn, and N(t) and R are independent. Such rainfall
be estimated. generating process has been called Poisson white noise (PWN). The cumula-
tive rainfall in the interval (0,t), denoted Z (t ), is a compound Poisson process
and the cumulative rainfall over successive nonoverlapping time intervals T,
18.3 TIME-SERIES MODELING
is given by Yi = Z (iT ) − Z (iT − T ), i = 1,2,... . The statistical properties of Yi,
A variety of models are available for modeling univariate and multivariate assuming that Z(t) is generated by a PWN model, has been widely studied
series of annual and seasonal time scales. The selected model aims at repro- (e.g., Cadavid et al., 1992; Eagleson, 1978). For example, its ACF is equal to
ducing important statistics of the process under consideration. Most para- zero for all lags greater than zero, which contradicts actual observations. In
metric time series models assume that the underlying process is normally PWN, the rainfall is assumed to occur instantaneously with zero duration,
distributed and stationary in the mean and variance. Depending on the time which is unrealistic. Instead, the Poisson rectangular pulse (PRP) model
series under consideration, the first step prior to model fitting may involve (Rodriguez-Iturbe et al., 1984) assumes that rainfall occurs with a finite ran-
the following analysis (Brockwell and Davis, 2002; Salas, 1993; Sveinsson, dom duration D and a random intensity I. The pair of intensities and dura-
2014): tions (i1,d1),...,(in,dn) correspond to n storms occurring at times t1,...,tn in the
interval (0,t). The storms may overlap; consequently, the process Yi becomes
Test for Test for Is seasonality present autocorrelated. Although the PRP model is better conceptualized than PWN,
normality trends in the mean and it is still limited when applied to rainfall data.
variance? Alternative models based on the concept of clusters have been suggested
Transform to Remove trends
(Le Cam, 1961; Neyman and Scott, 1958). The Neyman–Scott cluster process
normal if if necessary Remove if necessary by
can be thought as a two-level mechanism for rainfall generation, storm gen-
necessary seasonal standardization eration governed by a Poisson process, and associated with each storm, a
number of precipitation bursts that are Poisson (or geometric) distributed
The foregoing choices whether or not to transform to normal, remove trends, occurs. Then, if an instantaneous random depth R describes the rainfall burst,
and seasonality, depend on the properties of the selected model as we will see the resulting process is known as Neyman–Scott white noise, while if the rain-
later. fall burst is a rectangular pulse the process is known as Neyman–Scott rectan-
Generally, the typical steps involved in time series modeling are: (1) model gular pulse. Extensive developments, applications, and reviews on such
identification, (2) parameter estimation, and (3) model testing and validation. models have been made (e.g., Burlando and Rosso, 1991; Cadavid et al., 1992;
The sample correlation of the time series (ACF and PACF) is often used to get Cowpertwait and O’Connell, 1997; Entekhabi et al., 1989; Foufoula-Georgiou
an idea of the type of model that may be fitted to the given series, and the and Guttorp, 1986, 1987; Kavvas and Delleur, 1981; Obeysekera et al., 1987;
model with the minimum residual variance is often selected as the best. This Rodriguez-Iturbe et al., 1984, 1987a, and 1988). In addition, temporal rainfall
does not penalize for the number of parameters and a common practice is models based on Cox processes (e.g., Smith and Karr, 1983), renewal pro-
using the Akaike Information Criterion (AIC) and the Schwarz Information cesses (e.g., Buishand, 1977; Foufoula-Georgiou and Lettenmaier, 1987),
Criterion (SIC) for selecting the best model which penalize for the number of Barlett–Lewis processes (e.g., Gyasi-Agyei and Willgoose, 1997; Rodriguez-
parameters used in the model (Brockwell and Davis, 2002; Shumway and Iturbe et al., 1987a), and space-time multidimensional rainfall models have
Stoffer, 2000). been developed (e.g., Islam et al., 1988; Krajewski and Smith, 1989; Smith and
All the sample information on lack of fit is contained in the residuals. In Krajewski, 1987; Wilks, 1998).
most time-series models, the residuals are assumed to be normally distributed
with mean zero. Thus, the frequency distribution of the residuals should
18.5 UNIVARIATE MODELING
resemble as normally distributed. In addition, the residuals should be uncor-
related and also independent of the explanatory variables used in the model. Univariate time series models are used for modeling hydroclimatic processes,
Non-normal residuals may indicate lack of transformation of the data, while such as streamflow at a single site. There are various types of models available
18_Singh_ch18_p18.1-18.12.indd 2 8/22/16 12:06 PM

Univariate Modeling 18-3
depending on the characteristics of the underlying time series. In this section, where et is an independent noise term, f is the autoregression coefficient, and
we give an overview of commonly used univariate models, some of them have Zt is a three-parameter gamma distributed variable with marginal density
been specifically developed for streamflow, while some others for precipita- function given by:
tion. In general, models intended for shorter time scales, such as hourly or
daily, are more complex than models for larger time scales. α β ( z − λ )β −1 exp [−α ( z − λ )]
fZ (z ) = (18.7)
18.5.1 ARMA and FARMA Models Γ(β )
The stationary ARMA models (Bras and Rodriguez-Iturbe, 1985; Brockwell
and Davis, 2002; Box et al., 1994; Hipel and McLeod, 1994; Salas, 1993) is in which l, a, and b are the location, scale, and shape parameters, respec-
likely the most commonly used for modeling streamflow and other hydrocli- tively. The noise term, εt, can be obtained by (Lawrence, 1982):
matic series. Physical justification of ARMA models for rainfall-runoff pro-
cess (Sec. 18.11) have been suggested (e.g., Salas and Smith, 1981). The  η=0
 if M = 0 (18.8)
ARMA(p,q) model of autoregressive order p and moving average order q is ε = λ (1 − φ ) + η , where  Uj
 η = ∑ j =1Y j φ
M
defined as: if M >0

p q
Yt = µY + ∑ φi (Yt −i −µY ) + ε t − ∑θ j ε t − j (18.4) where M is Poisson distributed with mean – b ln(f), Uj, j =1,2,.... are inde-
i =1 j =1 pendent identically distributed (iid) random variables with uniform (0,1)
distribution, and, Yj, j = 1,2,.... are iid exponential random variables with
where Yt is normally distributed with mean µY and variance s2(Y), εt is an mean (1/a).
independent normally distributed noise with mean zero and variance The GAR(1) model has four parameters, {f, l, a, b}. It may be shown that
σ 2 (ε ) and independent of past observations, and {f1,…,fp} and {q1,…, qq} the population moments of Zt are functions of the model parameters as:
are the autoregressive and moving average parameters, respectively. Simpler
models such as AR(1) Yt = µY + φ1 (Yt −1 − µY ) + ε t and ARMA(1,1) β β 2
µ=λ+ , σ2 = 2, γ = , ρ1 = φ (18.9)
Yt = µY + φ1 (Yt −1 − µY ) + ε t − θ1ε t −1 are often adequate, but higher order α α β
ARMA models may be necessary depending on the particular case. The char-
acteristics of ACF and PACF of ARMA(p,q) models are summarized in where m, s2, g, and r1 are the mean, variance, skewness coefficient, and the
Table 18.1. ARMA models are flexible and can accommodate features of alter- lag-1 autocorrelation coefficient, respectively. Estimation of the GAR(1)
native models, such as fractional Gaussian noise (Mandelbrot and Wallis, model parameters can be made by the MOM using the results obtained by
1969), broken line (Mejia et al., 1972), and SM (Sveinsson et al., 2003). For Matalas (1967), Wallis and O’Connell (1972), and the simulation studies con-
details on parameter estimation of ARMA models [e.g., method of moments ducted by Fernandez and Salas (1990). For more details on the estimation
(MOM), least squares, and maximum likelihood], the interested reader may method and applications thereof, the interested reader is referred to Fernandez
refer to Box et al. (1994) and Brockwell and Davis (2002). and Salas (1990).
Fractional ARMA models, that is, FARMA(p, d, q), are direct generaliza-
tion of ARMA and ARIMA models (Brockwell and Davis, 1991; Hipel and 18.5.3 Shifting Mean Models
McLeod, 1994; Hosking, 1981). FARMA models are stationary with fractional The shifting mean (SM) model (Sveinsson and Salas, 2006; Sveinsson et al.,
differencing (1–B)d = 1 – dB – 0.5 d(1 – d) B2 – … , where B is the backward 2003, 2005, 2009) is characterized by sudden shifts or jumps in the mean and
shift operator BnYt = Yt–n. Note that –0.5 < d < 0.5 is applied instead of an can be used to model high and low frequency shifts of hydroclimatic pro-
integer d as is used for the nonstationary ARIMA model. Then, the fractionally cesses (Sveinsson et al., 2003). The underlying process is assumed to be char-
differenced time series is fitted by an ARMA model as: acterized by multiple stationary states, which only differ from each other by
having different means that vary around the long-term mean. Thus, the pro-
q
cess is stationary and autocorrelated, where autocorrelation arises from the
φ ( B)(1 − B)d (Yt − µY ) = ε t − ∑θ j ε t − j (18.5)
sudden shifting pattern in the mean. The SM model is defined as:
j =1
t
where f(B) = 1 – f1B – f2B2 – … – fpBp is the autoregressive operator. Also Xt = Yt + Zt , where Zt = ∑ Mi I( Si −1 ,Si ] (t )
note that due to the fractional differencing operator, the FARMA model has i =1 (18.10)
an infinite autoregressive part. An interesting feature of FARMA process is its
long memory, where the ACF converges to zero with increasing lag-h at a where {Xt} is a sequence of random variables representing the hydrological
much slower rate than that for the traditional ARMA process. The FARMA process of interest; {Yt} is a sequence of iid Normal (mY, σ Y2 ) variables that are
model has been used for modeling monthly and daily flows (Montanari et al., independent of the mean level sequence {Zt}, {Mt} is iid Normal
2
1997, 2000), and the FARMA(0, d, 0) is similar to fractional Gaussian noise, (0, σ M = σ Z2 ), and I is the indicator function with Si = N1 + N 2 ++ N i , S0 = 0 ,
which has been used in studies of the long-term dependence in hydrocli- in which Ni is the length the process remains in mean level i (a discrete sta-
matic series (Kousoyiannis, 2002). An overview of fitting FARMA models is tionary delayed renewal sequence with Ni iid positive geometric with param-
given by Brockwell and Davis (1991) and Hipel and McLeod (1994). Unlike eter p). Thus, the average length of each state of the process is 1/p. SM models
ARIMA models, where d is fixed prior to estimating the parameters of the have been used for modeling the Pacific Decadal Oscillation (PDO) index
differenced series, all parameters in the FARMA model must be estimated and the annual flows of the Niger River (Sveinsson et al., 2003), the quarter-
simultaneously. Note that for practical applications 0 < d < 0.5. monthly annual maximum outflows of Lake Ontario, and the 7-day annual
low flows of the Paraná River, Argentina (Sveinsson et al., 2005), where Yt was
18.5.2 GAR Model
modeled by skewed distributions. SM model with AR(1) persistence
The GAR(1) model is similar to the AR(1) model except that the underlying (Sveinsson, 2002; Sveinsson and Salas, 2006), and Bayesian version of the SM
process is gamma distributed instead of normal. The GAR(1) model can be model has been proposed (Fortin et al., 2004), and it has been shown that the
expressed as (Lawrence and Lewis, 1981): SM models belongs to the class of hidden Markov-chain models (Fortin et al.,
2004; Thyer and Kuczera, 2000). Parameters are estimated by the MOM as:
Zt = φ Zt −1 + ε t (18.6)
ρˆ ( X ) ρˆ ( X )
pˆ = 1 − 2 ; µˆY = µˆ X ; σˆ M
2
= σˆ X2 1 ; σˆ Y2 = σˆ X2 − σˆ M
2
(18.11)
ρˆ1 ( X ) (1 − pˆ )
Table 18.1 Properties of the ACF and PACF for ARMA(p,q) Models where parameters are feasible if ρˆ1 ( X ) > ρˆ 2 ( X ) > ρˆ12 ( X ) . To reduce the effects
AR(1) AR(p) MA(q) ARMA(p,q)
of sample variability of the estimated ACF and avoid unfeasibility of the SM
model parameters, Sveinsson et al. (2005, 2009) suggested fitting the exact
ACF Decays Tails off Zero for Tails off σ 2 (1 − p)h
geometrically lag > q form of the model ACF ρh ( X ) = M2 2
, h = 1,2, to the sample ACF
σ Y +σ M
2
PACF Zero for Zero for Tails off Tails off using least squares, that is, rh = a bh with a = σ M / σ X2 and b = 1 – p, where
lag > 1 lag > p 0 < {a, b} < 1.
18_Singh_ch18_p18.1-18.12.indd 3 8/22/16 12:06 PM

18.5.4 Markov-Chain Models Likewise, the DARMA(1,1) model may be expressed as (Jacobs and Lewis,
We have seen in Sec. 18.4 that the models and properties of cumulative pro- 1978):
cesses, such as rainfall over successive nonoverlapping time periods, that is,
discrete time rainfall, can be derived from continuous time models. Xi = U iYi + (1 − U i )Zi−1 , i = 1, 2, (18.14)
However, one may formulate models directly at discrete time scales, such as
in which Ui is an iid Bernoulli (0,1) process with parameter P[U i = 1] = β ,
hourly and daily. In these cases, the theory of Markov chains has been
0 ≤ β ≤ 1, so that Xi = Yi with probability β , and Xi = Zi−1 with probability
widely used for modeling not only rainfall, but also other hydrologic pro-
1 − β , Yi is another iid Bernoulli variable with parameter p1 (as aforemen-
cesses such as streamflow, soil moisture, and water storage in reservoirs (e.g.,
tioned), Zi–1 is a DAR(1) process with parameters l and p1 as in Eq. (18.12),
Buishand, 1977; Chang et al., 1984; Feyerherm and Bark, 1964; Katz,
and the variables {Xi}, {Zi}, and {Yi} are stationary and have the same probabil-
1977a, b; Katz and Parlange, 1995; Richardson and Wright, 1984; Roldan and
ity distribution p0 and p1. The ACF of Xi is ρk ( X ) = cλ k−1 , k ≥1, where
Woolhiser, 1982).
c = (1 − β )(λ + β − 2λβ ). It is a slower decaying correlogram (longer memory)
Consider that a discrete time process X(t) = j, j=1,...,r (where j denotes a
than that of the DAR(1). The run length distributions, estimation procedures,
state and r is the number of states) is represented by a first-order Markov
and applications of the binary DAR(1) and DARMA(1,1) processes can be
chain (a simple Markov chain). For instance, in modeling daily rainfall, one
found in literature (e.g., Buishand, 1978; Chang et al., 1984; and Chung and
may consider only two states, j=1 for a dry day (no rain) and j=2 for a wet day.
Salas, 2000).
A simple Markov chain is defined by a square transition probability matrix
P(t), with elements pij(t)=P[X(t)=j|X(t–1)=i] for all i,j pairs. The marginal 18.5.6 Product Models
distribution of the chain being at any state j at time t is denoted by
The AR, ARMA, FARMA, SM, and GAR models discussed in the previous
qj(t)=P[X(t)=j], j=1,...,r, and qj(0) is the distribution of the initial states. The
sections are useful for modeling hydrologic processes, such as sreamflow in
Markov chain is homogeneous or stationary if P(t) does not depend on time
perennial rivers, but they are impractical for intermittent streamflows in
and in this case, the notations P, pij, and qj are used. Common estimation
ephemeral streams and precipitation. Intermittent processes can be modeled
methods include the MOM and maximum likelihood and estimation of the
as the product (Jacobs and Lewis, 1978):
n-step transition probability pij(n ) is available in literature (e.g., Guttorp, 1995;
Wilks, 1995). To test whether a simple Markov chain is an adequate model for Yt = Xt Zt (18.15)
the process under consideration, one can check some of the underlying
assumptions of the model and whether it is able to reproduce some relevant where Yt is a nonnegative intermittent hydrologic variable, Xt is a discrete
properties of the rainfall process. For example, one can compare the n-step autocorrelated variable, Zt is a positive-valued continuous autocorrelated
transition probability pij(n ) with that obtained from the data, pîj(n ). Likewise, the variable, and Xt and Zt are assumed to be mutually uncorrelated. For example,
AIC has been helpful in selecting the order of Markov chains (e.g., Katz, the variable Xt may be represented by a dependent (1,0) Bernoulli process and
1981). Although in some cases, simple Markov chains may be adequate for Zt by an AR(1). Then the product Yt is intermittent and autoregressive.
representing the occurrence of rainfall, often more complex models may be Product models have been applied for modeling short-term rainfall (e.g.,
necessary. For instance, in modeling daily rainfall throughout the year, the Buishand, 1977; Chang et al., 1984) and intermittent streamflows (Chebaane
parameters of the Markov chain may vary with time and the estimates of the et al., 1995; Salas and Chebaane, 1990). Further elaboration on product mod-
transition probabilities pij can be fitted with trigonometric series to smooth els considering periodic processes will be illustrated in Sec. 18.6.5.
out sample variations (e.g., Roldan and Woolhiser, 1982). Higher order
Markov chains may be necessary as well. Chin (1977) analyzed daily precipi-
tation records for many stations across the United States and concluded that 18.6 UNIVARIATE PERIODIC MODELING
for the majority of stations Second- and third-order models were preferred For seasonal hydrologic time series (e.g., monthly series) seasonal statistics
for the winter months, while a first-order model was better for the summer such as the mean and variance may be reproduced by a stationary model, such
months. In addition, for daily rainfall, Roldan and Woolhiser (1982) used as ARMA by removing seasonality in the underlying seasonal series, that is,
maximum likelihood for estimating Fourier series coefficients for alternating subtracting the seasonal mean and dividing by the seasonal standard devia-
renewal processes and Markov chains while Katz and Parlange (1995) and tion. This procedure assumes that season-to-season correlations are the same
Martinez and Salas (2004) used mixed models with periodic Markov chains throughout the year. However, hydrologic time series, such as monthly
for hourly rainfall. streamflows, generally exhibit different month-to-month correlations.
Periodic models have been suggested in literature for modeling such periodic
18.5.5 DAR and DARMA Models dependence structure (e.g., Hipel and McLeod, 1994; Loucks et al., 1981;
Markov chains have been widely used for modeling hydrologic time series Salas, 1993; Salas et al., 1980). Note that despite preserving the seasonal cova-
that are temporally correlated. However, in many cases geophysical data may riance, annual statistics may not be preserved. Sections 18.6.3 and 18.8 later
not be Markovian. Chang et al. (1984) showed that simple Markov chains do further refer to this issue.
not properly describe some sequences of daily rainfall, and alternative models 18.6.1 PARMA and PFARMA Models
such as DARMA (discrete autoregressive moving average) were needed.
DARMA models are the discrete counterparts of ARMA models and they are PARMA models have been suggested in literature for modeling periodic
more suitable for modeling persistence characteristics with longer memory dependence structure. For w seasons, a PARMA(p, q) model consists of w
than simple Markov chains. DARMA(p,q) models (p = autoregressive order individual ARMA(p, q) models, where the dependence is across seasons
and q = moving average order) were originally developed by Jacobs and Lewis instead of years. A PARMA(p, q) model may be expressed as (e.g., Hipel and
(1978). Low-order DARMA models have been used for modeling precipita- McLeod, 1994; Salas, 1993)
tion and streamflow (e.g., Buishand, 1978; Chang et al., 1984; Chebaane et al.,
1992, 1995; Chung and Salas, 2000; Cancelliere and Salas, 2010). Some rele- p q
vant definitions and properties regarding DAR(1) and DARMA(1,1) models Yν ,τ = µτ + ∑ φi ,τ (Yν ,τ −i − µτ −i ) + εν ,τ − ∑θ j ,τ εν ,τ − j (18.16)
are briefly reviewed as follows. i =1 j =1
The DAR(1) model may be defined as (Jacobs and Lewis, 1978):
Xi = Vi Xi−1 + (1 − Vi )Yi , i = 1, 2,, (18.12) where Yν ,τ represents the hydrologic process for year n and season t and µτ
and σ τ2 (Y ) are the seasonal mean and seasonal variance, respectively. The
where Vi is an independent Bernoulli (0,1) process with parameter noise term εn,t is uncorrelated normally distributed with mean zero and vari-
P[Vi = 1] = λ , 0 ≤ λ ≤ 1 and Yi is another independent Bernoulli (0,1) process ance σ τ2 (ε ), {f1,t,…,fp,t} are periodic autoregressive parameters, and {q1,t,…,
with P[Yi = 1] = π 1 . Model (18.12) implies that Xi = Xi−1 with probability λ and qq,t} are periodic moving average parameters. In many practical applications,
Xi =Yi with probability 1 − λ . Its ACF is given by ρk ( X ) = λ k , k ≥1. It may be low-order models, such as PAR(1), PAR(2), and PARMA(1,1) may be ade-
shown that the DAR(1) process is a first-order Markov chain with one-step quate, although residuals should always be tested to ensure adequate model
transition probabilities given by: fit. For details on parameter estimation and applications, the interested reader
is referred to the references mentioned earlier. Fractional PARMA, that is,
 λ + (1 − λ )π j , if i = j FPARMA(p,d,q) (Hipel and McLeod, 1994), is developed in a similar way as
 FARMA (Sec. 18.5.1). For each season of the year, a separate FARMA model
p(i , j ) = P( X k+1 = j | X k = i ) =  , i , j = 0, 1 (18.13)
 (1 − λ )π j , if i ≠ j is fitted as for the PARMA model. Hui and Li (1988) derived maximum
likelihood estimation for the special case FPARMA(p,d,0).
18_Singh_ch18_p18.1-18.12.indd 4 8/22/16 12:06 PM

Univariate Periodic Modeling 18-5
18.6.2 PGAR Model

the seasonal mean. We will fit an ARMA(P,Q)ω model to Yt as Φ( Bω )Yt =Θ( Bω )ξt
Stochastic modeling of skewed hydrologic series has always been a challenge
in which xt is correlated normal with mean zero and variance σ 2 (ξ ) ,
because most models available in literature assume that the underlying pro-
cess is normally distributed, thereby requiring data transformations to nor- Φ( Bω ) = 1 − (Φ1B1ω + ...+ Φ P B Pω ), and Θ( Bω ) = 1 − (Θ1B1ω + ...+ ΘQ BQω ) . For
mality prior to applying the models. While this procedure has been a standard instance, for P=2 and Q=0, the model is written as Yt = Φ1 Yt −ω + Φ 2Yt −2ω + ξt ,
practice, it is always useful having an alternative where no transformation to which is like an ARMA(2,0), but defined over blocks of ω seasons (years).
normality is necessary. Section 18.5.2 presented the GAR model, which is Recall that t is time in seasons. In addition, the residual xt is fitted as
stationary autoregressive with a gamma marginal distribution. While such a φ ( B)ξt =θ ( B)ε t . Then, combining the two aforementioned models above gives
model is useful for some hydrologic series such as annual streamflow, its
application to processes such as weekly rainfall is impractical. An alternative Φ( Bω )φ ( B)Yt =Θ( Bω )θ ( B) ε t (18.18)
model for periodic hydrological processes at the seasonal time scales (e.g.,
monthly) is the PGAR (Fernandez and Salas, 1986). which is the stationary multiplicative ARMA(p,q)x(P,Q)ω process.
Let the underlying periodic process denoted by Qn, t be autoregressive hav- The drawback with the foregoing multiplicative model for seasonal pro-
ing a gamma marginal distribution with parameter set (at, bt, lt), respec- cesses such as streamflows is that there is no provision for reproducing peri-
tively, the scale, shape, and location parameters. Then, Zν ,τ = Qν ,τ − λτ is odic covariance structure. A model that incorporates such periodic features is
gamma with two parameters at and bt . There are two types of periodic the multiplicative PARMA (XPARMA) which may be written as (Salas and
gamma models, an additive model Zν ,τ = φτ Zν ,τ −1 + εν ,τ , where Zn, t is Abdelmohsen, 1992):
defined for season t (t =1,…, w) of year n (n =1,…,N) in which w is the num-
ber of seasons per year and N is the number of years considered, Zν ,0 = Zν −1,ω , Φτ ( Bω )φτ ( B)Yν ,τ = Θτ ( Bω )θτ ( B) εν ,τ (18.19)
ft is a periodic autoregressive parameter, and en, t is a random noise that has
where Yn,t = periodic series after removing the seasonal mean, en,t = normal
a particular distribution such that Zn, t is gamma. The model requires that bt
increases from one season to the next or remains constant throughout the uncorrelated with mean zero and variance σ τ2 (ε ), Ft (Bw) = 1–(F1,t B1w +...+
year. The second type of periodic gamma model is a product Zν ,τ = Zνθτ,τ −1 Wν ,τ , Fp,t BPw), and Qt (Bw) = 1–(Q1,t B1w +...+ QQ,t BQw). For instance, the PARMA(1,1)
where qt is a periodic autoregressive exponent and Wn, t is a random noise, x(1,1)ω is a multiplicative process which is written as Yn,t = F1,t Yn–1,t +
which also has a particular distribution such that Zn, t is gamma. The product φ1,τ Yν ,τ −1 − Φ1,τ φ1,τ Yν −1,τ −1 + εν ,τ − Θ1,τ εν −1,τ − θ1,τ εν ,τ −1 + Θ1,τ θ1,τ εν −1,τ −1.
model requires that the shape parameter bt decreases from one season to the Monthly flows of the Nile River at Aswan (1871–1989) were used for com-
next or remains constant. In each case, details of the noises en, t and Wn, t can paring the PAR(1), PARMA(1,1), and XPARMA(1,1)x(1,1)12 models. The
be found in Fernandez and Salas (1986). model parameters were estimated by least squares based on the logarithmic
In order to accommodate a varying shape parameter bt across the year, a transformed data. For each model, samples of the same length as the historical
mixture model was suggested as: record were generated which were used for comparing the monthly and
annual statistics versus those obtained from the historical record. The Nile
Zν ,τ = φτ Zν ,τ −1 + Zνθτ,τ −1 Eν ,τ (18.17) River flows show significant year-to-year dependence for monthly flows, sig-
nificant month-to-month correlations as well as significant dependence struc-
ture for annual flows. The results showed that generally all models reproduced
in which Eν ,τ = εν ,τ if βτ ≥ βτ −1 or Eν ,τ = Wν ,τ if βτ < βτ −1 and the parameters the month-to-month correlations quite well. However, the yearly correlations
 α  β 
1/2 for each month and the annual correlations were not reproduced by the
ft and ft are given by φτ = ρ1,τ  τ   τ  if βτ ≥ βτ −1 otherwise ft = 0 PAR(1) and PARMA(1,1) models. On the other hand, the correlations derived
 α τ −1   βτ −1  from the XPARMA(1,1)x(1,1)12 model resembled closely those from the his-
 β 1/2
and θτ = ρ1,τ  τ  if bt < bt–1, otherwise θτ = 0, where ρ1,τ is the correla- torical record. Likewise the storage- and drought-related statistics were better
 βτ −1  reproduced by the multiplicative model (Salas and Abdelmohsen, 1992).
tion between Zn, t and Zn, t–1. For further details on the PGAR model, the
18.6.4 PDAR and PDARMA Models
interested reader is referred to Fernandez and Salas (1986). The model has
been tested extensively and its performance compares favorably respect to the Just as PAR and PARMA models are useful for modeling continuous valued
traditional approach based on normal transformation and using normal- hydrological process at the seasonal time scale, PDAR and PDARMA models
based models. However, the PGAR model is limited to a lag-1 correlation and are useful for discrete valued processes, such as the occurrence of hourly
is only available for single sites. Applications of the PGAR model to weekly rainfall or the occurrence of seasonal streamflows in dry regions. For illustra-
and monthly streamflows are available (e.g., Fernandez and Salas, 1986; tion, we assume modeling the occurrence of hourly rainfall in places where
there is evidence of the daily cycle.
Chebaane et al., 1992).
The occurrence of hourly rainfall may be represented by Xν ,τ = 1 if rainfall
18.6.3 XPARMA Model occurs during the hour t of day v and Xν ,τ = 0 otherwise. Such binary-
dependent variable may be expressed mathematically by a periodic discrete
A desirable objective in simulating hydrological processes is the preservation
autoregressive model of order one, PDAR(1), as:
of statistics at multiple time scales (e.g., monthly and annual). For example,
for generating seasonal flows two approaches have been followed in literature. Xν ,τ = Vν ,τ Xν ,τ −1 + (1 − Vν ,τ )Wν ,τ ; ν =1,2,...; τ =1,2,...,ω (18.20)
The first has been generating seasonal flows directly using a periodic model,
while the second has been generating annual flows in a first step, then obtain such that Xν ,0 = Xν −1,ω , where ω is the number of hours. In addition, Vν ,τ
the seasonal flows by temporal disaggregation (Sec. 18.8). Low-order PARMA and Wν ,τ are independent Bernoulli with P[Wν ,τ = 1] = π 1,τ , 0 ≤ π 1,τ ≤ 1,
models generally do not capture the annual flow properties when annual P[Wν ,τ = 0] = π 0,τ = 1 − π 1,τ , and P[Vν ,τ = 1] = λτ , 0 ≤ λτ ≤ 1. Therefore, Xν ,τ =
flows are obtained from seasonal flows (Obeysekera and Salas, 1986). While Xν ,τ −1 with probability λτ and Xν ,τ = Wν ,τ with probability 1 – λτ .
disaggregation approach has been useful from the practical standpoint, one Furthermore, it may be shown that PDAR(1) is equivalent to a periodic simple
may not be satisfied from the theoretical standpoint, since it is only logical Markov chain with transition probability matrix
that if a model that is applied directly to seasonal flows fails to reproduce the
statistics at the aggregated level (annual), it must be because the model was 0 1
not right in the first place. The XPARMA model below offers an alternative.
0  λτ + (1 − λτ )π 0,τ (1 − λτ )π 1,τ 
The stationary ARMA(p,q) process may be expressed as φ ( B)Yt = θ ( B)ε t in  
which Yt is stationary with mean zero, et = normal noise with mean zero Pτ =  
 (1 − λ )π  (18.21)
and variance σ 2 (ε ), φ ( B) = 1 − (φ1B1 + ... + φ p B p ) and θ ( B) = 1 − (θ1B1 + ...+ θ q Bq ) 1 τ 0,τ λτ + (1 − λτ )π 1,τ
 
with B j Zt = Zt − j . Likewise, the PARMA(p,q) process may be expressed as
φτ ( B)Yν ,τ =θτ ( B)εν ,τ where Yν ,τ = periodic correlated process with mean A number of properties that are useful for parameter estimation, simulation,
zero, εν ,τ = normal noise with mean zero and variance σ τ2 (ε ), ν is the year and forecasting can be determined (Chebaane et al., 1992, 1995) such as: (a)
the expected value E[ Xν ,τ ] = µτ ( X ), that is, the probability of rainfall during
and t is the season, φτ ( B) = 1 − (φ1,τ B1 + ... + φ p ,τ B p ), and qt = (B) = 1– the hour t, P[ Xν ,τ = 1], τ =1,...,ω ; (b) the variance and the lag-k autocovari-
(θ1,τ B1 + ... + θ q ,τ Bq ) with B j Zν ,τ = Zν ,τ − j for j ≤ t, otherwise B j zν ,τ = zν −1,ω +τ − j . ance; (c) the n-step periodic transition probability matrix Pτ(n ); and (d) the
Box et al. (1994) suggested a multiplicative model that could be useful for probabilities of the durations of dry and wet periods (i.e., the sequence of
seasonal series. Let Yt represent the underlying seasonal series after removing nonrainy and rainy hours, respectively).
18_Singh_ch18_p18.1-18.12.indd 5 8/22/16 12:06 PM

Katz and Parlange (1995) suggested that in order to improve the estimation
which µτ ( Z ) and σ τ2 ( Z ) were estimated from the conditional hourly rainfall
of upscale statistics, for example, daily statistics, one could improve the mod-
data and used the moment Eqs. (18.25a) and (18.25b) to estimate the param-
eling of the occurrence process Xν ,τ . For example, low-order periodic
eters µτ ( N ) and σ τ2 ( N ). The autoregressive parameter ft was determined by
DARMA (PDARMA) models, such as the PDARMA(1,1) model, can be for-
the MOM. Clearly using an autocorrelated non-Gaussian model directly in
mulated (Chebaane et al., 1992) as:
terms of the underlying precipitation variable Z would be desirable. Two
Xν ,τ = Uν ,τ Wν ,τ + (1 − Uν ,τ ) Zν ,τ −1 (18.22a) alternatives that assume that Z is gamma distributed are the GAR model and
the PGAR model as described in Secs. 18.5.2 and 18.6.2.
Zν ,τ = Vν ,τ Zν ,τ −1 + (1 − Vν ,τ )Wν ,τ (18.22b)
18.7 MULTIVARIATE MODELING
where Uν ,τ and Vν ,τ are independent Bernoulli processes with probabilities
Multivariate models are used for modeling temporal and spatial dependence
P[Uν ,τ = 1] = β1,τ , 0 ≤ βτ ≤ 1 and P[Vν ,τ = 1] = λτ , 0 ≤ λτ ≤ 1, and Wν ,τ as
structure at multiple lags. Analysis and modeling of multiple time series is
defined earlier.
more involved than for a single time series. The basic principles are similar
The parameters of PDAR and PDARMA models can be estimated by the
except that vector and matrix notations are used for model definition, proper-
MOM and maximum likelihood (Chebaane et al., 1992; Katz and Parlange,
ties, and parameter estimation. Full multivariate models generally have many
1995). Applications of the PDAR(1) model for modeling the occurrence of
parameters, but simpler contemporaneous models in which the parameter
hourly rainfall have been made (e.g., Katz and Parlange, 1995; Martinez and
matrices are diagonal are practical. Both models can preserve complex tem-
Salas, 2004). Also Salas and Chebaane (1990) and Chebaane et al. (1992) used
poral dependence structure for each site, but only lag-0 correlations across
the PDAR(1) model for modeling the occurrence of monthly streamflows in
sites are preserved in contemporaneous models, which is usually considered
arid regions.
adequate for hydroclimatic series, while higher lagged correlations can be
18.6.5 Product Periodic Models preserved in full multivariate models.
In Sec 18.5.6, we have considered product models for stationary processes. In 18.7.1 Multivariate ARMA and PARMA Models
this section, we consider periodic (seasonal) processes, such as hourly rainfall
Properties of multivariate ARMA and PARMA models have been widely
in some regions and monthly streamflows in dry regions. For the ease of
studied (e.g., Bernardo and Smith, 1994; Box et al., 1994; Brockwell and
explanation, let Yν ,τ be a positive intermittent variable that represents hourly
Davis, 1991, 2002; Hipel and McLeod, 1994; Shumway and Stoffer, 2000).
precipitation (or any time interval smaller than a day such as 4 h) in which
Simpler contemporaneous ARMA and PARMA models are often used for
n = day, t = time interval (h) within the day. The variable Yν ,τ can be
hydroclimatic data at multiple sites (Hipel and McLeod, 1994; Salas, 1993;
expressed as the product:
Salas et al., 1980). The full multivariate ARMA (MARMA) and the contem-
Yν ,τ = Xν ,τ Zν ,τ ; ν = 1,2,...; τ =1,2,...,ω (18.23) poraneous ARMA (CARMA) models for n sites are expressed as:
where w = number of time intervals in the day (e.g., w = 24 for hourly pre- p q
cipitation), Xν ,τ is a binary (0,1) periodic-dependent variable that represents Yt = ∑ Φi Yt −i + ε t − ∑ Θ j ε t − j (18.26)
the occurrence process, and Zν ,τ is a periodic-dependent continuous vari- i =1 j =1
able representing the precipitation amount. In addition, Xν ,τ and Zν ,τ are
where Yt is a n × 1 column vector of normally distributed zero mean variables
assumed to be mutually independent. Alternatively, Eq. (18.23) can be
(i.e., it is assumed that the mean of the data have been subtracted prior to
expressed as Yν ,τ = Zν ,τ if Xν ,τ = 1 and Yν ,τ = 0 if Xν ,τ = 0 . The model is
modeling). The CARMA model is more parsimonious than MAR. For
assumed to be valid for a given season or month of the year. Modeling of the
example, the number of parameters for MAR(2) and CARMA(2,0) models
discrete periodic process Xν ,τ has been described in Sec. 18.6.4. Therefore,
are 2n2 + 0.5(n2 + n) and 2n + 0.5(n2 + n), respectively. Similarly, the full
the remaining part of this section addresses briefly modeling of the continu-
multivariate MPARMA and contemporaneous PARMA with periodic param-
ous variable Zν ,τ . The variable Zν ,τ represents the amount of rainfall during
eters, that is, CPARMA, can be expressed as:
the rainy hours, that is, Zν ,τ > 0 if Xν ,τ = 1. The variable Zν ,τ is generally
skewed and autocorrelated, and may be modeled in several ways depending p q
on the data at hand. For example, depending whether or not the hour-to-hour Yν ,τ = ∑ Φi ,τ Yν ,τ −i + εν ,τ − ∑ Θ j ,τ εν ,τ − j (18.27)
correlations vary with the hour a model with constant or periodic parameters i =1 j =1
will be selected. In addition, short-term precipitation may be highly skewed. where Yν ,τ is a n × 1 column vector of normally distributed zero mean vari-
For example, the skewness coefficient of July daily rainfall for Denver Airport ables for year n and season t. The autoregressive and moving average param-
is about 5 based on 1949–1990 record. Thus, one will need to include this eter matrices are full n × n matrices for the MARMA and MPARMA models,
characteristic in the model. Two possibilities are: Gaussian models that but they are diagonal for the CARMA and CPARMA models. The n × 1 noise
require a previous step of transformation and non-Gaussian models. They are vector {εt} is iid MVN (0,G) and vector {εν ,τ } ~ iid MVN ( 0 , Gτ ) for the sta-
further described as follows. tionary and periodic models, respectively. For all models, G and Gt are full
Several alternatives are available for modeling Zν ,τ , such as low-order n × n matrices, enabling the preservation of the lag-0 correlations across sites.
ARMA and PARMA models after transformation. For example, suppose a In practice low-order models, such as MAR(1), MAR(2), and MARMA(1,1),
power transformation is used to transform Z into a normal variable N as and their periodic counterparts (Hipel and McLeod, 1994; Matalas, 1967;
Nν ,τ = [Zν ,τ ]cτ , where ct is an exponent that may vary with the hour, and a Salas and Pegram, 1979; Salas et al., 1980) have been used for modeling
PAR(1) model is fitted to represent the variable N as: streamflows at multiple sites. The moment equations of the MAR(p) model
p p
σ τ (N ) are C 0 = G + ∑Φi CTi and C h = ∑Φi C h−i , h ≥ 1, where C h = E[Yt YtT−h ] is the
Nν ,τ = µτ ( N ) + φτ  Nν ,τ −1 − µτ −1 ( N ) + σ τ ( N ) εν ,τ (18.24)
σ τ −1 ( N )  i =1 i =1
lag-h population cross-covariance matrix. For the MAR(1) model, the param-
in which e is a standard normal random variate. The parameters of this model eters are estimated by Φˆ = c c −1 and G ˆ = c − c c −1c T , where c is the lag-h
1 1 0 0 1 0 1 h
are µτ ( N ) , σ τ2 ( N ) , and φτ . Using the power transformation one could find a sample cross-covariance matrix. Similarly, for the MPAR(1) model, the
ˆ = c c −1 ˆ −1 T
relationship between the moments of N and Z. For illustration if ct =1/2 one parameters are estimated by Φ 1,τ 1,τ 0,τ −1 and Gτ = c 0,τ − c 1,τ c 0,τ −1c 1,τ .
can find that Since the CARMA and CPARMA models have diagonal parameter matrices,
they can be decoupled into univariate ARMA and PARMA models and the
µτ ( Z ) = µτ2 ( N ) + σ τ2 ( N ) (18.25a) parameters estimated independently for each site using the corresponding
univariate model-estimation procedures. This allows for identification of the
σ τ2 ( Z ) = 2σ τ2 ( N )[σ τ2 ( N ) + 2 µτ2 ( N )] (18.25b) best univariate model for each site. After having estimated the diagonal
parameter matrices what remains is estimation of the noise variance-covariance
Katz and Parlange (1995) used the PAR(1) model for July hourly rainfall of matrix G or Gt. In either case, the process is the same and it is illustrated here
Denver Airport except that the transformation parameter was assumed to be for the CARMA model, which requires that the individual ARMA models are
a constant value, that is, ct = 1/8 and the autoregression coefficient ft was also causal. Causality implies that CARMA(p, q) can be written as an infinite mov-
assumed a constant value throughout the day. The parameters µτ ( N ) and ∞ ∞
σ τ2 ( N ) were estimated from the transformed data and used the moment ing average model, for example, Yt = ∑ Ψ j ε t − j with C 0 = ∑ Ψ j GΨ Tj , where
p
relationships as a way of verifying the estimates of µτ ( Z ) and σ τ2 ( Z ) . j =0 j =0
Ψ 0 = I and Ψ j = −Θ j + ∑Φi Ψ Tj −i . Since for contemporaneous models, Ψ j
Alternatively, Martinez and Salas (2004) applied an estimation procedure in
i =1
18_Singh_ch18_p18.1-18.12.indd 6 8/22/16 12:06 PM

Nonparametric Models 18-7
are diagonal then the ith row and jth column element of G is estimated from as for the Great Lakes system in North America (Fagherazzi et al., 2005) or
∞
Gˆ ij = c0ij ∑ ψkiiψkjj . the Nile River system in Africa (Salas et al., 1995). Software packages such as
k =0
SAMS (Salas et al., 2006; Sveinsson et al., 2009) and SPIGOT (Grygier and
18.7.2 Multivariate SM Model Stedinger, 1990) have been developed for applications of alternative disag-
Multiple time series of different variables and different hydrologic regimes gregation modeling techniques.
may require mixing of models. For example, shifts in time series of one hydro- In general, annual flow series at key sites or regions can be disaggregated in
logic variable may not be present in the series of another variable because of stages, for preserving the temporal (annual and seasonal) and spatial statis-
differences in their geographic locations. Sveinsson (2002) and Sveinsson and tics. The disaggregation models introduced by Valencia and Schaake (1973)
Salas (2006) developed a multivariate contemporaneous shifting mean (CSM) and Mejia and Rousselle (1976) are often used in practice. The Mejia–
model for simulating the annual net basin supplies of the Great Lakes. Since Rousselle model for spatial disaggregation of annual or seasonal data from n
not all of the Great Lakes data showed shifting patterns, the CSM model was key sites to m subsites can be respectively represented as:
mixed with a CARMA model for simulating jointly the entire Great Lakes
Yν = A Xν + B εν + C Yν −1 (18.29)
system. The CSM-CARMA model is capable of preserving the spatial lag-0
correlations between different sites and other important statistical character-
istics of the individual sites. In addition, for sites modeled by the CSM model Yν ,τ = Aτ Xν ,τ + Bτ εν ,τ + Cτ Yν ,τ −1 (18.30)
certain dependence structure in both time and space is preserved through the where X is a vector of flows at key sites, which is modeled by say an ARMA
mean level process Zt. The CARMA portion of the model is capable of pre- or CARMA model, Y is a vector of flows at subsites, and A, B, and C are full
serving more complex time dependence structure for each individual site m × n, m × m, and m × m parameter matrices, respectively. The independent
depending on the AR and MA order of the model. In Sveinsson and Salas noise vector εv is iid MVN(0, G = BBT), and εv,t is iid MVN (0, Gt = BtBtT),
(2006), the Yt component in the CSM model was also contemporaneous where B (or Bt) is the Cholesky decomposition of G (or Gt). This model
AR(1). The mixed CSM-CARMA model for n sites can be written as: preserves the lag-1 correlation coefficient in space and time between all sub-
sites through the matrices C and Ct . The parameters are estimated by the
 X (1)   Y (1)   (1)  MOM (Hipel and McLeod, 1994; Salas, 1993; Sveinsson et al., 2009). The
 t   t   Zt  model can be repeatedly applied for further spatial disaggregation from sub-
        sites to subsequent sites as is common when modeling complex streamflow
 (n1 )   (n1 )    
 Xt   Yt   Zt(n1 ) 
networks. The disaggregation aforementioned schemes were initially pro-
 (n +1)  =  (n1 +1)  +  
(18.28) posed for temporal disaggregation of annual flows into seasonal flows, requir-
 Xt   Yt
1
  0  ing many parameters for temporal disaggregation. Alternatively, condensed
         models (Lane) (Lane, 1979, 1981) and contemporaneous models (Spigot)
     
 Xt
(n )
Y
  t
(n )
  0 (Grygier and Stedinger, 1988, 1990) with periodic parameters for temporal
disaggregation have been suggested reducing the number of parameters
required significantly (Sveinsson, 2014). The Lane and Spigot models for
where n = n1+ n2, the first n1 sites follow a CSM model and the remaining n2 disaggregating annual Yν to seasonal Yν ,τ data (year ν, season t) for n sites
sites a CARMA(p,q) model. For the CSM model, the mean level sequence can be respectively represented as:
t
Zt = ∑ Mi I (Si −1 ,Si ] (t ) is defined in the same way as for the univariate SM Yν ,τ = Aτ Yν + Bτ εν ,τ + Cτ Yν ,τ −1 (18.31)
i =1
model (Sec. 18.5.3) and {Yt } ~ iid MVN (µ Y , C Y (0)) . If the shifts in the means
Yν ,τ = Aτ Yν + Bτ εν ,τ + Cτ Yν ,τ −1 + Dτ εν ,τ (18.32)
are caused by changes in natural processes, such as low-frequency variability
of the climate system and the mean levels are correlated in space at lag-0, then
{Mi } ~ iid MVN[0 , C M (0)] and a necessary and sufficient condition for {Zt} to where Lane’s model in Eq. (18.31) has full n × n periodic parameter matrices
be stationary in the covariance C Z (h) = (1 − p )h C M (0), h = 0,1, , is that and Spigot model in Eq. (18.32) has diagonal contemporaneous periodic
N1 , N 2 , is a common sequence for all sites, that is, the time series of the matrices except for matrix B which is full. Both models preserve the lag-1
hydrologic variables within a geographic region would all exhibit shifts at season-to-season correlations between all subsites through the matrix C and
the same time periods (note that if {Mi} are assumed uncorrelated in space the the lag-0 correlations in space through matrix B. These models can be repeat-
stationarity condition where N1 , N 2 , is a common sequence is not neces- edly applied in a stage-wise manner down to finer time scales (Santos and
sary). The parameter-estimation procedure for the CSM model is similar as Salas, 1992), further reducing the number of parameters in each disaggrega-
for other contemporaneous models. That is, first the CSM is uncoupled into tion step. Lane’s model requires adjustments of seasonal values to ensure
univariate SM models and the parameters at each site estimated using the additivity up to the aggregated level while Spigot model includes a transfor-
univariate SM model procedures (Sec. 18.5.3). Then the common parameter mation dependent term [last term in Eq. (18.32)] ensuring approximate
of the positive geometric distribution, p, for all sites is taken as the average additivity of the model in the original domain.
of the at site sample estimates [or as a weighted average if the sample For annual data, only spatial disaggregation models are needed, while for
records are of different lengths as in Sveinsson and Salas (2006)]. seasonal data both spatial and temporal disaggregation models are used. The
Subsequently, the parameters of each univariate SM model are reestimated foregoing models can be applied in different order for seasonal data, for
using the common p̂. Lastly, the nondiagonal elements of CM(0) and CY(0) example, “Spatial-Temporal Disaggregation” first disaggregating annual data
ˆ (0) = (1 − pˆ )−1 C
ˆ (1) and made symmetric by replacing in space from key sites to subsites to subsequent sites, etc., and then applying
are estimated from C M X temporal disaggregation for different groups of sites, or “Temporal-Spatial
ij
the off diagonal elements cˆM (0) and cˆMji (0) with their respective averages, Disaggregation” first disaggregating key sites annual data down to seasonal
then C ˆ (0) = Cˆ (0) − C
ˆ (0) . and then disaggregating the seasonal data in space from key sites to subsites,
Y X M
Estimation of the CARMA part of the model (if n2>0) is made by fitting the etc. When using disaggregation models for data generation for data requiring
best ARMA model for each site using the parameter estimation procedure for normalizing transformation or when various elements of the parameter
the multivariate CARMA model (Sec. 18.7.1). For estimation of the variance- matrices are simplified in condensed or contemporaneous models, then sum-
covariance matrix of the noise (G) of the CARMA component Yt, the proce- mability may be lost and adjustments may be needed to ensure additivity
dure of the CARMA model is applied, where each of the elements of Yt constraints in the original domain (Grygier and Stedinger, 1988; Sveinsson
corresponding to the CSM component is considered as an ARMA(0,0) model. et al., 2009).
The upper left n1 × n1 part of the n × n estimated G matrix is replaced by the
estimate of CY(0). 18.9 NONPARAMETRIC MODELS
Nonparametric models are an attractive alternative to parametric models.

18.8 DISAGGREGATION MODELS
They do not require the data to be transformed to normal and they can cap-
Disaggregation models are capable of reproducing statistics at different aggre- ture features, such as bimodality and nonlinear dependence structure that is
gation levels, such as annual and seasonal. Disaggregation in space and time difficult to capture with traditional parametric models (Lall, 1995), although
(downscaling) is important for modeling complex hydroclimatic processes, it has also been suggested for hydrologic data to log-transform the data prior
whether it is downscaling precipitation to fine time scales or disaggregating to kernel density estimation (Rajagopalan et al., 1997, 2010) or use skewed
flows and precipitation in a complex stream network system (Hipel and gamma kernels (Salas and Lee, 2010). The simplest nonparametric models are
McLeod, 1994; Santos and Salas, 1992; Salas, 1993; Sveinsson et al., 2009) such pure resampling or bootstrapping models, such as the simple bootstrap based
18_Singh_ch18_p18.1-18.12.indd 7 8/22/16 12:06 PM

on resampling with replacement of the historical time series. To account for is then directly proportional to the cross-correlation coefficients at different
autocorrelation block bootstrapping can be used (Vogel and Shallcross, 1996). lags of the residual series. In practice, the selected model(s) should be physi-
Nonparametric models based on conditional kernel density estimation have cally consistent and be included among those with the lowest AIC and SIC.
been used for monthly streamflow simulation and disaggregation of annual to For details on alternative estimation procedures, refer to Box et al. (1994),
monthly (Lee et al., 2010; Sharma et al., 1997; Tarboton et al., 1998). The Brockwell and Davis (2002), and Hipel and McLeod (1994).
k-nearest neighbor resampling models have been used for resampling month- Uncertainty of model parameters is a well-known issue, but not always
ly streamflows, disaggregating annual to monthly flows, and simulating taken into account to estimate predictive distribution or confidence limits of
rainfall and other climatic variables (Lall and Sharma, 1996; Lee and Ouarda, decision variables or forecasts. For some models, such as ARMA and
2011; Prairie et al., 2006, 2007, and 2008; Rajagopalan and Lall, 1999; Salas PARMA, the sample distributions of the parameters are defined (e.g.,
and Lee, 2010; Yakowitz and Karlsson, 1987). Brockwell and Davis, 2002; Hipel and McLeod, 1994), Also, since most para-
metric modeling schemes are linear in the parameters and assume that the
data are normally distributed, the inclusion of parameter uncertainty can be
18.10 STOCHASTIC SIMULATION, FORECASTING, AND achieved by adopting the theory of parameter inferences from the classical
UNCERTAINTY univariate and multivariate regression model (Johnson and Wichern, 2002).
Stochastic simulation and forecasting and methods to deal with parameter In addition, Bayesian modeling (Bernardo and Smith, 1994) offers a frame-
uncertainty have been discussed in literature (e.g., Lettenmaier and Wood, work to take into account the uncertainty in the model parameters. When
1993; Salas, 1993; Valdés et al., 2002; Sveinsson, 2014; Salas et al., 2014). Time- multiple time series are synthetically generated from a parametric model for
series models enable generating (simulating) equally likely synthetic series analysis of the distribution of certain design related statistics, it can be argued
that may be used for planning and management of water resources systems. that parameter uncertainty is being somewhat implicitly embedded in the
Random generators of univariate and multivariate standard normal variables generation process, since for each synthetic series the parameters could be
are available in many software packages. Exact generation procedures are reestimated and they certainly would be different than the original parameter
available for simple models (e.g., Salas and Abdelmohsen, 1993), while set. For complex models, repeated one-step ahead prediction can be used for
approximations are generally used for complex correlated models that depend estimating the prediction interval. Furthermore, if non-normality of the
on past observations and residuals’ initial assumptions. In practice, a short residuals is of concern bootstrapping of the residuals with replacement and
warm-up period is often used to remove the effects of initial assumptions, reestimation of the model parameters can be used as an alternative for esti-
which is equivalent to generating a longer time series that is needed and dis- mating the prediction interval (Sveinsson et al., 2008a).
carding the values of the warm-up period. The length and the number of the
simulated series may vary. For comparing model and historical statistics, the 18.11 CONCEPTUAL STOCHASTIC MODELING
length of the simulated series should be the same as the length of the historical
record, while for operational and planning studies the length of the simulated In previous sections, we presented stochastic models, such as ARMA and
series should match the planning horizon. The number of generated series FARMA that may be useful for simulating hydrological processes, such as
(e.g., 1000) will provide the information to estimate the probability distribution streamflows. However, these models are also used in several fields other than
or the uncertainty of the decision parameter. hydrology. In this section, we briefly illustrate a conceptual framework which
Forecasting can be done based on the models presented in this chapter may justify applying the foregoing models for streamflow processes. For this
simply by starting simulations at the end of the historical records conditioned purpose, we follow the concepts outlined in Fiering (1967) and Salas and
on the most recent historical observations with all noises initialized to zero. Smith (1981), where the modeling framework is considered at the annual
Adaptive forecasting or Kalman filtering techniques (Lettenmaier and Wood, time scale. The key input in the simple basin system is precipitation. Some
1993) can be used allowing the model parameters to be updated at each time studies in literature suggest that precipitation is white or near white noise, say
step prior to making the next step-ahead forecast. Forecast performance of short memory Markovian (e.g., Blender et al., 2006; Potter, 1979); while oth-
different model alternatives is often measured in terms of mean-absolute- ers indicate that in some regions precipitation exhibits long-term variability
forecast error and root-mean-square-forecast error (Lettenmaier and Wood, and shifting patterns similar to those observed in some atmospheric/oceanic
1993). Forecasting models including exogenous information usually perform indices (e.g., Enfield et al., 2001) and Koutsoyiannis and Langousis (2011)
better than models based only on past observations of the underlying process. suggested that precipitation has long memory. In the following conceptual
For example, for streamflow forecasting the exogenous variables may be basin watershed both short- and long-memory precipitation are assumed.
precipitation, climatic indices such as El Niño-Southern Oscillations (ENSO) Let us define a basin, where Xt is average annual precipitation, Gt is the
and PDO or other measures of atmospheric and oceanic circulation influenc- groundwater storage, and Qt is the annual streamflow (t is in years). The fol-
ing the local climate (e.g., Sveinsson et al., 2008b; Salas et al., 2011). If there lowing assumptions are made (Fig. 18.1): annual infiltrated precipitation,
are many exogenous variables available alternative models could be fitted and It = aXt ; annual evapotranspiration, Et = bXt ; and annual surface runoff
the resulting forecasts combined by averaging or weighting (Sveinsson et al., Rt = (1 − a − b)Xt , in which 0<a<1 and 0<b<1 are coefficients. In addition,
2008a). Furthermore, the parameter space may be reduced using principal baseflow contribution to the stream is assumed proportional to groundwater
components on the exogenous variables. Popular models including exoge- storage, that is, c Gt–1, and the groundwater flow passing downstream is g Gt–1,
nous variables are transfer function noise (TFN) models and simpler alterna- where c and g are coefficients such that 0<c<1, 0<g<1, and 0<c+g<1. Then
tives, such as ARMAX and PARMAX, which are extensions of the ARMA and using the storage equation for groundwater and denoting φ = 1 − c − g , it may
PARMA models (Box et al., 1994; Brockwell and Davis, 2002; Hipel and be shown Gt = φ Gt −1 + a Xt . Streamflow is made up of surface runoff and
McLeod, 1994). A TFN model and ARMAX model with one exogenous vari- groundwater contribution. It follows, Qt = (1 − a − b)Xt + cGt −1 . Combining the
able can be presented respectively as: foregoing equations, the model for annual streamflow becomes (Salas and
Smith, 1981)
ω ( B) b θ ( B) Qt = φ Qt −1 + eXt − θ Xt −1 (18.35)
Yt − µY = B ( Xt − µ X ) + εt (18.33)
δ ( B) φ ( B)
ω ( B) b θ ( B)
Yt − µY = B ( Xt − µ X ) + εt (18.34)
φ ( B) φ ( B)
Xt
with Yt being the dependent variable, Xt being the exogenous variable leading Et = b Xt
the dependent variable by lag-b, ω(B), δ(B), θ(B), and f(B) are polynomials in
B and εt is the white noise term. When Yt is weekly autocorrelated, it is not
uncommon to have the order p = 0 and q = 0 for the autoregressive operator It = a Xt
f(B) and the moving average operator θ(B), respectively, and thus the only (1 – a – b)Xt
input being the exogenous variable. The main difference between the two Qt
models relates to the estimation procedure, although both models can be Gt Gt
expressed as traditional regression models. In the ARMAX model, typically c Gt–1
both series are prewhitened using appropriate ARMA models and then
causal relationships between the residual series are analyzed at different lags. g Gt–1
In the TFN model, the input series Xt is prewhitened and the output series Yt
is then filtered with the same model. The impulse response function ω(B)/δ(B) Figure. 18.1 Conceptual stochastic model (after Salas and Smith, 1981).
18_Singh_ch18_p18.1-18.12.indd 8 8/22/16 12:07 PM

REFERENCES 18-9
where e = 1 − a − b, and θ = φ e − ac . The models for Gt and Qt can be written underlying process to make it stationary prior to fitting an appropriate sta-
as: tionary model. Synthetic time series generated from the fitted model are then
corrected into the original domain by reapplying the historical or projected
φ ( B)Gt = a Xt (18.36) anthropogenic changes creating, for example, several sets of time series that
are only valid for certain periods in the future. Such synthetic time series may
φ ( B)Qt = θ ( B) Xt (18.37) be used for estimating design events or adaptation strategies to deal with the
expected future climatic change. Hydropower companies relying on concep-
tual rainfall-runoff models for estimation of historical records are already
where φ ( B) = 1 − φ B and θ ( B) = e − θ B. One can make additional assump- applying similar methods to correct historical flows into the future taking into
tions on the precipitation input Xt. First, we consider the cases of no- account trends in both temperature and precipitation due to climate change
memory or short-memory precipitation. It may be shown that (i) Gt is an (e.g., Sveinsson et al., 2012).
AR(1) model and Qt is an ARMA(1,1) if Xt is white noise normal, and (ii) Gt While the various concepts and models included in this chapter are useful
is AR(2) model and Qt ARMA(2,1) if Xt is AR(1). A similar modeling frame- for stationary and nonstationary conditions as stated earlier, the actual devel-
work has been extended to seasonal time scales leading to PAR and PARMA opment and application in practice may be limited due to lack of appropriate
models (Salas and Obeysekera, 1992). data. Experience shows that often significant shifts in the streamflow regime
Suppose now that precipitation input has long memory (long-term persis- (e.g., in the Nile River system) has been wrongly attributed to human inter-
tence), for example, assume that Xt is a FARMA(0,d,0) process. Certainly that vention until longer data and advances in understanding of the underlying
will bring “long memory” into our conceptual models of Gt and Qt . Thus, we climate showed the misconception. Thus, one must be careful in attributing
assume that the cause of changes. Conversely and upward trend in the flow regime may be
(1 − B)d Xt = εt or Xt = (1 − B)−d εt (18.38) quickly attributed to the so-called “climate change,” while the real cause may
be the effect of significant land use changes on the basin.
It follows that the model for Gt becomes
φ ( B)Gt = a (1 − B)− d ε t or φ ( B)(1 − B)d Gt = a ε t (18.39) REFERENCES

which is a FARMA(1,d,0) process. Likewise, the model for streamflow Qt Bernardo, J. M. and M. F. A. Smith, Bayesian Theory, Wiley, Chichester,
becomes UK, 1994.
φ ( B)(1 − B)d Qt = θ ( B)ε t (18.40) Blender, R., K. Fraedrich, and B. Hunt, “Millennial climate variability:
GCM-simulation and Greenland ice cores,” Geophysical Research Letters, 33
that is, a FARMA(1,d,1). (4), L04710, doi:10.1029/2005GL024919, 2006.
Bras, R. L. and I. Rodriguez-Iturbe, Random Functions and Hydrology,
Addison-Wesley, Menlo Park, CA, 1985.
18.12 FINAL REMARKS AND FUTURE CHALLENGES
Box, G. E. P., G. M. Jenkins, and G. C. Reinsel, Time Series Analysis: Fore-
Traditionally stochastic models of hydrologic series have been used for a casting and Control, 3rd ed., Upper Saddle River, Prentice Hall, New Jersey,
variety of purposes, such as evaluating operational rules of water resources 1994.
systems, sizing of reservoirs, filling in missing data and extending records, Brockwell, P. J. and R. A. Davis, Introduction to time series and forecasting,
detection of trends and shifts, and forecasting hydrologic events that may Springer Texts in Statistics, 2nd ed., Springer-Verlag, New York, 2002.
occur in the future. Hydroclimatic data show different types of dependence Brockwell, P. J. and R. A. Davis, Time Series: Theory and Methods, Springer-
structure due to the effect of storage, annual hydrologic cycle, and persistence Verlag, New York, 1991.
and cycles in atmospheric circulation patterns. Time-series models are capa- Buishand, T. A., “Some remarks on the use of daily rainfall models,” Journal
ble of modeling such complex temporal and spatial dependence structure. of Hydrology, 36: 295–308, 1977.
They are useful for modeling stationary and nonstationary hydroclimatic data Buishand, R. A., The Binary DARMA(1,1) Process as a Model for Wet-Dry
that may exhibit trends, shifts, and oscillating features with short or Sequence, Technique Note 78-01, Department of Mathematics, Agricultural
long memory. University, Wageningen, The Netherlands, 1978.
Challenges facing time-series analysis and forecasting in today’s world are Burlando, P. and R. Rosso, “Comment on ‘Parameter estimation and sensi-
in large part due to human intervention and climate variability, which tivity analysis for the modified Bartlett-Lewis rectangular pulses model of
may cause significant changes in the distributional properties of the process rainfall by Islam et al., J. Geophys. Res., vol. 95, no. D3, 1990, p. 2093–2100,’”
under consideration (e.g., flow and precipitation). Human intervention, such Journal of Geophysical Research, 96 (no. D5): 9391–9395, 1991.
as construction of dams, weirs, and diversions for water regulation, may cause Cadavid, L. G., J. D. Salas, and D. C. Boes, “Disaggregation of short-term
changes in the flow regime, sediment transport, and water quality. In addi- precipitation records,” Water Resources Papers, Colorado State University, Fort
tion, land use changes, such as urbanization, deforestation, and large-scale Collins, CO, Vol. 106, 1992.
conversion of land for irrigation, may cause changes in the streamflow pattern Cancelliere, A. and J. D. Salas, “Drought probabilities and return period for
(e.g., Salas and Pielke, 2003). In practice, some of the changes resulting from annual streamflows series,” Journal of Hydrology, 391: 77–89, 2010.
human intervention are quite clear, requiring flow records to be naturalized Chang, T. J., M. L. Kavvas, and J. W. Delleur, “Daily precipitation modeling
(e.g., effects of dams, weirs, and diversions need to be removed from the by discrete autoregressive moving average processes,” Water Resources
records) prior to fitting a model and generating synthetic time series. Research, 20 (5): 565–580, 1984.
However, even if the cause of some apparent changes in say streamflow data, Chebaane, M., J. D. Salas, and D. C. Boes, “Modeling of monthly intermit-
may be evident, naturalization of measured flow data may not be a simple task tent streamflow processes,” Water Resources Papers 105, Colorado State
depending on the complexity of the system at hand. Often errors accumulate University, Fort Collins, CO, 1992.
and propagate and may lead to significant changes in the data and corre- Chebaane, M., J. D. Salas, and D. C. Boes, “Product periodic autoregressive
sponding statistics. processes for modeling intermittent monthly streamflows,” Water Resources
Climate change is dominated by human influences (Karl and Trenberth, Research, 31 (6): 1513–1518, 1995.
2003; Salas et al., 2014; Weaver, 2003), and to-date there does not appear to be Chin, E. H., “Modeling daily precipitation occurrence process with Markoc
any study attributing the global warming observed over the last 50 years to chains,” Water Resources Research, 13 (6): 949–956, 1977.
any known natural causes. Knowledge of large-scale forcing factors, such as Chung, C. and J. D. Salas, “Drought occurrence probabilities of dependent
ENSO, PDO, and AMO along with global climate variability, can aid in hydrologic processes,” Journal of Hydrologic Engineering (ASCE), 5 (3): 259–
understanding the variability of hydrological data, such as streamflows and 268, 2000.
evaluate impacts and risks under changing conditions (e.g., Jain and Lall, Cowpertwait, P. S. P. and P. E. O’Connell, “A regionalized Neyman-Scott
2001; Sveinsson et al., 2003, 2005; Salas et al., 2014). Changing climate due to model of rainfall with convective and stratiform cells,” Hydrology and Earth
anthropogenic impacts may lead to changing distributional properties of the System Sciences, 1: 71–80, 1997.
process under consideration, such as trends in the mean (location), variance Eagleson, P., “Climate, soil and vegetation. 2. The distribution of annual
(spread), and skewness (shape) (Sveinsson, 2014), calling for sophisticated precipitation derived from observed storm sequences,” Water Resources
risk analyses of possible climate change in the future. The time-series analysis Research, 14: 713–721, 1978.
methods presented in this chapter can be used for modeling changing distri- Enfield, D. B., A. M. Mestas-Nunez, and P. J. Trimble, “The Atlantic multi-
butional properties over time, where either time varying parameters are used decadal oscillation and its relation to rainfall and river flows in the Continen-
or historical deterministic anthropogenic changes are removed from the tal U.S.,” Geophysical Research Letters, 28 (10): 2077–2080, 2001.
18_Singh_ch18_p18.1-18.12.indd 9 8/22/16 12:07 PM

Entekhabi, D., I. Rodriguez-Iturbe, and P. S. Eagleson, “Probabilistic repre- Krajewski, W. F. and J. A. Smith, “Sampling properties of parameter
sentation of the temporal rainfall process by a modified Neyman-Scott rect- estimators for a storm field rainfall model,” Water Resources Research, 25 (no.
angular pulse model: parameter estimation validation,” Water Resources 9): 2067–2075, 1989.
Research, 25 (no. 2): 295–302, 1989. Lall, U., “Recent advances in nonparametric function estimation: hydraulic
Fagherazzi, L., R. Guay, D. Sparks, J. Salas, and O. G. B. Sveinsson, Stochas- applications,” U.S. National Report to International Union of Geodesy and
tic Modeling and Simulation of the Great Lakes – St. Lawrence River System, Geophysics, 1991–1994, Reviews of Geophysics, 33: 1093–1102, 1995.
Report prepared for the Lakes Ontario – St. Lawrence River Study of the Lall, U. and A. Sharma, “A nearest neighbor bootstrap for resampling
International Joint Commission, Ottawa and Washington, 2011. hydrologic time series,” Water Resources Research, 32 (3): 679–693, 1996.
Fernandez, B. and J. D. Salas, “Periodic gamma autoregressive processes for Lane, W. L., “Corrected parameter estimates for disaggregation schemes,”
operational hydrology,” Water Resources Research, 22 (10): 1385–1396, 1986. International Symposium on Rainfall Runoff Modeling, Mississippi State
Fernandez, B. and J. D. Salas, “Gamma-autoregressive models for stream- University, MS, 1981.
flow simulation,” Journal of Hydrologic Engineering (ASCE), 116 (11): 1403– Lane, W. L., Applied Stochastic Techniques (Last Computer Package), User
1414, 1990. Manual, Division of Planning Technical Services, U.S. Bureau of Reclamation,
Feyerherm, A. M. and L. D. Bark, “Statistical methods for persistent pre- Denver, CO, 1979.
cipitation patterns,” Journal of Applied Meteorology (now Journal of Climate Lawrance, A. J., “The innovation distribution of a gamma distributed
and Applied Meteorology), 4: 320–328, 1964. autoregressive process,” Scandanivian Journal of Statistics, 9: 234–236, 1982.
Fiering, M. B., Streamflow Synthesis, Harvard University Press, Cambridge, Lawrance, A. J. and P. A. W. Lewis, “A new autoregressive time series model
MA, 1967, p. 135. in exponential variables (NEAR(1)),” Advances in Applied Probability, 13 (4):
Fortin, V., L. Perreault, and J. D. Salas, “Retrospective analysis and forecast- 826–845, 1981.
ing of streamflows using a shifting level model,” Journal of Hydrology, 296 Le Cam, L. A. “A stochastic description of precipitation,” Proceedings of the
(1–4): 135–163, 2004. IV Berkeley Symposium on Mathematics, Statistics & Probability, edited by
Foufoula-Georgiou, E. and P. Guttorp, “Compatibility of continuous rain- J. Newman, University of California Press, Berkeley, CA, 1961, pp. 165–186.
fall occurrence models with discrete rainfall observations,” Water Resources Lee, T. S., J. D. Salas, and J. Prairie, “An enhanced nonparametric stream-
Research, 22: 1316–1322, 1986. flow disaggregation modeling and genetic algorithm,” Water Resources
Foufoula-Georgiou, E. and D. P. Lettenmaier, “A Markov renewal model of Research, 46 (W08545): 1–14, 2010.
rainfall occurrences,” Water Resources Research, 23 (no. 5): 875–884, 1987. Lee, T. and T. B. M. J. Ouarda, “Identification of model order and number
Foufoula-Georgiou, E. and P. Guttorp, “Assessment of a class of Neyman- of neighbors for k-nearest neighbor resampling,” Journal of Hydrology, 404:
Scott models for temporal rainfall,” Journal of Geophysical Research, 92 136–145, 2011.
(no. D8): 9679–9682, 1987. Lettenmaier, D. P. and E. F. Wood, Hydrologic Forecasting, Handbook of
Gyasi-Agyei, Y. and G. R. Willgoose, “A hybrid model for point rainfall Hydrology, edited by D. R. Maidment, McGraw-Hill, New York, Chap. 26, 1993.
modeling,” Water Resources Research, 33 (7): 1699–1706, 1997. Loucks, D. P., J. R. Stedinger, and D. A. Haith, Water Resources Systems
Grygier, J. C., and J. R. Stedinger, “Condensed disaggregation procedures Planning and Analysis, Prentice-Hall, Englewood Cliffs, New Jersey, 1981.
and conservation corrections for stochastic hydrology,” Water Resources Mandelbrot, B. B. and J. R. Wallis, “Computer experiments with fractional
Research, 24 (10): 1574–1584, 1988. Gaussian noises. Part 1, Averages and variances,” Water Resources Research, 5
Grygier, J. C., and J. R. Stedinger, SPIGOT, A Synthetic Streamflow Genera- (1): 228–241, 1969.
tion Software Package, technical description, version 2.5, School of Civil and Martinez, A. and J. D. Salas, “Modelamiento estocastico de lluvias horarias,”
Environmental Engineering, Cornell University, Ithaca, New York, 1990. Ingeneria del Agua, 11 (1): 29–39, 2004.
Guttorp, P., Stochastic Modeling of Scientific Data, Chapman Hall, London, Matalas, N. C., “Mathematical assessment of synthetic hydrology,” Water
1995, p. 372. Resources Research, 3 (4): 937–945, 1967.
Hipel, K. and A. I. McLeod, Time Series Modeling of Water Resources and Mejia, J. M., I. Rodriguez-Iturbe, and D. R. Dawdy, “Streamflow simulation:
Environmental Systems, Elsevier, Amsterdam, 1994. 2. The broken line process as a potential model for hydrologic simulation,”
Hosking, J. R. M., “Fractional differencing,” Biometrika, 68: 165–176, 1981. Water Resources Research, 8 (4): 931–941, 1972.
Hui, Y. V. and W. K. Li, On Fractional Differenced Periodic Processes, Tech- Mejia, J. M. and J. Rousselle, “Disaggregation models in hydrology revis-
nical Report, Chinese University of Hong Kong, Hong Kong, 1988. ited,” Water Resources Research, 12 (3): 185–186, 1976.
Islam, S., R. L. Bras, and I. Rodriguez-Itrube, “Multidimensional modeling Montanari, A., R. Rosso, and M. S. Taqqu, “Fractionally differenced
of cumulative rainfall: parameter estimation and model adequacy through a ARIMA models applied to hydrologic time series: identification, estimation
continuum of scales,” Water Resources Research 24 (7): 985–992, 1988. and simulation,” Water Resources Research, 33 (5): 1035–1044, 1997.
Jacobs, P. A., and P. A. W. Lewis, “Discrete time series generated by mix- Montanari, A., R. Rosso, and M. S. Taqqu, “A seasonal fractional ARIMA
tures I: correlational and runs properties,” Journal of Royal Statistical Society, model applied to the Nile River monthly flows at Aswan,” Water Resources
B40: 94–105, 1978. Research, 36: 1249–1259, 2000.
Jain, S. and U. Lall, Floods in a changing climate: does the past represent Neyman, J. and E. L. Scott, “Statistical approach to problems of cosmology,”
the future? Water Resources Research, 37 (12): 3193–3205, 2001. Journal of Royal Statistics Society, Series B, 20 (no. 1): 1–43, 1958.
Jenkins, G. M. and D. G. Watts, Spectral Analysis and Its Applications, Obeysekera, J. T. B. and J. D. Salas, “Modeling of aggregated hydrologic
1st ed., Holden-Day, New York, 1968. time series,” Journal of Hydrology, 86: 197–219, 1986.
Johnson, R. A. and D. W. Wichern, Applied Multivariate Statistical Analysis, Obeysekera, J. T. B., G. Tabios, and J. D. Salas, “On parameter estimation of
5th ed., Prentice-Hall, New York, 2002. temporal rainfall models,” Water Resources Research, 23 (no. 10): 1837–1850, 1987.
Karl, T. R. and K. E. Trenberth, “Modern global climate change,” Science, Percival, D. B. and A. T. Walden, Spectral Analysis for Physical Applications:
302: 1719–1723, 2003. Multitaper and Conventional Univariate Techniques, Cambridge University
Katz, R. W., “Precipitation as a chain-dependent process,” Journal of Press, Cambridge, 1993.
Applied Meteorology, 16: 671–676, 1977a. Potter, K. W., “Annual precipitation in Northwest United States: long memory,
Katz, R. W., “An application of chain-dependent processes to meteorology, short memory or no memory,” Water Resources Research, 15 (2): 340–346, 1979.
Journal of Applied Meteorology, 14: 598–603, 1977b. Prairie, J., K. Nowak, B. Rajagopalan, U. Lall, and T. Fulp, “A stochastic
Katz, R. W., “On some criteria for estimating the order of a Markov chain,” nonparametric approach for streamflow generation combining observational
Technometrics, 23 (3): 243–249, 1981. and paleoreconstructed data,” Water Resources Research, 44 (6): W06423,
Katz, R. W. and M. B. Parlange, “Generalizations of chain-dependent pro- 2008, doi: 10.1029/2007WR006684.
cesses: application to hourly precipitation,” Water Resources Research, 31: Prairie, J. R., B. Rajagopalan, T. J. Fulp, and E. A. Zagona, “Modified K-NN
1331–1341, 1995. model for stochastic streamflow simulation,” Journal of Hydrology Engineer-
Kavvas, M. L. and J. W. Delleur, “A stochastic cluster model of daily rainfall ing, 11 (4): 371–378, 2006.
sequences,” Water Resources Research, 17 (no. 4): 1151–1160, 1981. Prairie, J., B. Rajagopalan, U. Lall, and T. Fulp, “A stochastic nonparametric
Koutsoyiannis, D., “The Hurst phenomenon and fractional Gaussian noise technique for space-time disaggregation of streamflows,” Water Resources
made easy,” Hydrological Sciences Journal, 47 (4): 573–595, 2002. Research, 43 (3): W03432, 2007, doi:10.1029/2005WR004721.
Koutsoyiannis, D. and A. Langousis, Precipitation, Treatise on Water Sci- Rajagopalan, B. and U. Lall, “A k-nearest-neighbor simulator for daily pre-
ence, edited by P. Wilderer and S. Uhlenbrook, Vol. 2, pp. 27–78, Academic cipitation and other weather variables,” Water Resources Research, 35 (10):
Press, Oxford, 2011. 3089–3101, 1999.
18_Singh_ch18_p18.1-18.12.indd 10 8/22/16 12:07 PM

REFERENCES 18-11
Rajagopalan, B., U. Lall, and D. G. Tarboton, “Evaluation of kernel density Sharma, A., D. G. Tarboton, and U. Lall, “Streamflow simulation: a non-
estimation methods for daily precipitation resampling,” Journal of Stochastic parametric approach,” Water Resources Research, 33 (2): 291–308, 1997.
Hydrology and Hydraulics, 11 (6), 523–547, 1997. Shumway, R. H. and D. S. Stoffer, Time Series Analysis and Its Applications,
Rajagopalan, B., J. D. Salas, and U. Lall, Stochastic methods for modeling 1st ed., Springer, New York, 2000.
precipitation and streamflow, Advances in Data-Based Approaches for Hydro- Smith, J. A. and A. F. Karr, “A point process model of summer season rain-
logic Modeling and Forecasting, edited by B. Sivakumar and R. Berndtsson, fall occurrences,” Water Resources Research, 19 (no. 1): 95–103, 1983.
World Scientific, Singapore, 2010. Smith, J. A. and W. F. Krajewski, “Statistical modeling of space-time rainfall
Richardson, C. W. and D. A. Wright, WGEN: A Model for Generating Daily using radar and rain gage observations,” Water Resources Research, 23
Weather Variables, USDA-ARS, ARS-8, August, 1984, p. 83. (no. 10): 1893–1900, 1987.
Rodriguez-Iturbe, I., V. K. Gupta, and E. Waymire, “Scale considerations in Sveinsson, O. G. B., Modeling of Stationary and Non-Stationary Hydrologic
the modeling of temporal rainfall,” Water Resources Research, 20 (no. 11): Processes, Ph.D. dissertation, Colorado State University, Fort Collins, CO,
1611–1619, 1984. 2002, p. 206.
Rodriguez-Iturbe, I., D. R. Cox, and V. Isham, “Some models for rainfall Sveinsson, O. G. B., Time Series Analysis of Hydrologic Data. In: Handbook
based on stochastic point processes,” Proceeding of the Royal Society of of Engineering Hydrology: Modeling Climate Change and Variability, edited by
London, Series A, 410: 269–288, 1987. S. Eslamian, CRC Press, Taylor & Francis Group, Chap. 27, 2014.
Rodríguez-Iturbe, I., D. R. Cox, and V. Isham, “A point process model for Sveinsson, Ó. G. B., U. Lall, V. Fortin, L. Perrault, J. Gaudet, S. Zebiak, and
rainfall: further developments,” Proceedings of the Royal Society of London A, Y. Kushnir, “Forecasting spring reservoir inflows in Churchill falls basin in
417: 283–298, 1988. Québec, Canada,” Journal of Hydrologic Engineering (ASCE), 13 (6): 426–437,
Rodriguez-Iturbe, I., B. Febres de Power, and J. B. Valdes, “Rectangular 2008a.
pulses point process models for rainfall: analysis of empirical data,” Journal of Sveinsson, Ó. G. B., U. Lall, J. Gaudet, S. Zebiak, Y. Kushnir, and V. Fortin,
Geophysical Research, 92 (no. D8), 9645–9656, 1987. “Analysis of climatic states and atmospheric circulation patterns that influ-
Roldan, J. and D. A. Woolhiser, “Stochastic daily precipitation models: 1. A ence Quebéc spring streamflows,” Journal of Hydrologic Engineering (ASCE),
comparison of occurrence processes,” Water Resources Research, 18 (5): 13 (6): 411–425, 2008.
1451–1459, 1982. Sveinsson, O. G. B., T. S. Lee, J. D. Salas, W. L. Lane, and D. K. Frevert,
Salas, J. D., Analysis and modeling of hydrologic time series, Handbook of Stochastic Analysis, Modeling and Simulation (SAMS): Version 2009 User’s
Hydrology, edited by D. R. Maidment, McGraw-Hill, New York, Chap. 19, Manual, Department of Civil and Environmental Engineering, Colorado
1993. State University, Fort Collins, CO, 2009.
Salas, J. D. and M. W. Abdelmohsen, “Determining streamflow drought Sveinsson, O. G. B., Ú. Linnet, and E. B. Elíasson, “Hydropower in Iceland.
statistics by stochastic simulation,” U.S.-PRC Bilateral Symposium on Droughts Impacts and adaptation in future climate,” Climate Change and Energy Sys-
and Arid Region Hydrology, Tucson, AZ, September 16–20, 1992, pp. tems: Impacts, Risks and Adaptation in the Nordic and Baltic Countries, edited
227–233. by T. H. Thorsteinsson and H. Björnsson, Nordic Council of Ministers,
Salas, J. D. and M. Abdelmohsen, “Initialization for generating single site TemaNord 2011:502, Chap. 10, 2012.
and multisite low order PARMA processes,” Water Resources Research, 29 (6): Sveinsson, O. G. B. and J. D. Salas, “Multivariate shifting mean plus persis-
1771–1776, 1993. tence model for simulating the Great Lakes net basin supplies,” Proceedings of
Salas, J. D. and G. G. S. Pegram, “A seasonal multivariate multilag autore- the 25th AGU Hydrology Days, Colorado State University, Fort Collins, CO,
gressive model in hydrology,” Modeling Hydrologic Processes, edited by H. J. 2006, pp. 173–184.
Morel-Seytoux, J. D. Salas, T. G. Sanders, and R. E. Smith, Water Resources Sveinsson, O. G. B., J. D. Salas, and D. C. Boes, “Prediction of extreme
Publications, Fort Collins, CO, June 1977, pp. 125–145. events in hydrologic process that exhibit sudden shifting pattern,” Journal of
Salas, J. D., D. C. Boes, G. C. S. Pegram, and V. Yevjevich, “The Hurst phe- Hydrologic Engineering (ASCE), 10 (4): 315–325, 2005.
nomenon as a preasymptotic behavior,” Journal of Hydrology, 44: 1–15, 1979. Sveinsson, O. G. B., J. D. Salas, D. C. Boes, and R. A. Pielke, Sr., “Modeling
Salas, J. D. and M. Chebaane, “Stochastic modeling of monthly flows in the dynamics of long term variability of hydroclimatic processes,” Journal of
streams of arid regions,” Proceedings of the International Symposium on Hydrometeorology, 4: 489–505, 2003.
HY&IR Division ASCE, San Diego, CA, 1990, pp. 749–755. Tarboton, D. G., A. Sharma, and U. Lall, “Disaggregation procedures for
Salas, J. D., J. W. Delleur, V. Yevjevich, and W. L. Lane, Applied Modeling of stochastic hydrology based on nonparametric density estimation,” Water
Hydrologic Time Series, Water Resources Publications, Littleton, CO, 1980. Resources Research, 34 (1): 107–119, 1998.
Salas J. D., C. J. Fu, and B. Rajagopalan, “Long range forecasting of Thyer, M. and G. Kuczera, “Modelling long term persistence in hydro-
Colorado streamflows based on hydrologic, atmospheric, and oceanic data,” climatic time series using a hidden state Markov model,” Water Resources
Journal of Hydrologic Engineering (ASCE), 16 (6): 508–520, 2011. Research, 36 (11): 3301–3310, 2000.
Salas, J. D., R. S. Govindaraju, M. Anderson, M. Arabi, W. Suarez, W. Valdés, J. B., P. Burlando, and J. D. Salas, “Stochastic forecasting of precipi-
Lavado, and T. R. Green, “Introduction to hydrology,” Handbook of Environ- tation and streamflow processes,” Handbook of Weather, Climate, and Water,
mental Engineering: Modern Water Resources Engineering, edited by L. K. edited by T. D. Potter and B. Colman, John Wiley & Sons, New York, 2002,
Wang and C. T. Yang, Springer, New York, Vol. 1, Chap. 1, 2014. pp. 642–665.
Salas, J. D. and T. Lee, “Non-parametric simulation of single site seasonal Valencia, D. and J. C. Schaake, Jr., “Disaggregation processes in stochastic
streamflows,” Journal of Hydrologic Engineering, 15 (4): 284–296, 2010. hydrology,” Water Resources Research, 9 (3): 580–585, 1973.
Salas, J. D. and J. Obeysekera, “Conceptual basis of seasonal streamflow Vogel, R. M. and A. L. Shallcross, “The moving blocks bootstrap versus
time series models,” Journal of Hydrologic Engineering (ASCE), 118 (8): parametric time series models,” Water Resources Research, 32 (6): 1875–1882,
1186–1194, 1992. 1996.
Salas, J. D. and R. A. Pielke, Sr., “Stochastic characteristics and modeling of Wallis, J. R. and P. E. O’Connell, “Small sample estimation of r1,” Water
hydroclimatic processes,” Handbook of Weather, Climate, and Water, edited by Resources Research 8 (3): 707–712, 1972.
T. D. Potter and B. Colman, John Wiley & Sons, Hoboken, NJ, Chap. 32, 2003, Weaver, A. J., “The science of climate change,” Geoscience Canada, 30 (3):
pp. 587–605. 91–109, 2003.
Salas, J. D. and R. A. Smith, “Physical bases of stochastic models of annual Wilks, D. S., Statistical Methods in the Atmospheric Sciences: An Introduc-
flows,” Water Resources Research, 17: 428–430, 1981. tion, Academic, New York, Vol. 59, 1995.
Salas, J. D., N. Saada, and C. H. Chung, “Stochastic modeling and simula- Wilks, D. S., “Multisite generalization of a daily stochastic precipitation
tion of the Nile river system monthly flows,” Technical Report No. 5, Comput- generation model,” Journal of Hydrology, 210 (1–4): 178–191, 1998.
ing Hydrology Laboratory, Engineering Research Center, Colorado State Yakowitz, S. and M. Karlsson, “Nearest-neighbor methods with application
University, Fort Collins, CO, 1995. to rainfall-runoff prediction,” Stochastic Hydrology, edited by J. B. Mcneil and
Salas, J. D., O. G. B. Sveinsson, W. L. Lane, and D. K. Frevert, “Stochastic G. J. Humphries, Reidel, Hingham, MA, 1987, pp. 149–160.
streamflow simulation using SAMS-2003,” Journal of Irrigation and Drainage Yevjevich, V., Stochastic Processes in Hydrology, Water Resources Publica-
Engineering, 132 (2): 112–122, 2006. tions, Littleton, CO, 1972.
Santos, E. G. and J. D. Salas, “Stepwise disaggregation scheme for synthetic
hydrology,” Journal of Hydraulic Engineering (ASCE), 118 (5): 765–784, 1992.
18_Singh_ch18_p18.1-18.12.indd 11 8/22/16 12:07 PM

Chapter 19
Statistical Detection of
Nonstationarity: Issues and
Needs
BY
RICHARD H. MCCUEN
ABSTRACT change can occur over all or just part of the record length. A change in the
mean can be abrupt or gradual. Even if the mean does not change, a nonsta-
Books have been written on the subject of this chapter, so the subject matter
tionary series can result from a change in the variance; this usually implies
of this chapter tries to look at the issue from a different perspective.
that a factor responsible for the sampling variation has changed. Since the
Specifically, the objectives are: (1) to provide an overview of exploratory
type of change is not always evident from a graphical analysis, statistical tools
methods of analyzing time series potentially influenced by change; (2) to
are often applied to detect the significance of apparent change. If a change is
discuss problems relevant to many standard hypothesis tests used to analyze
apparent from either a graphical or a statistical analysis, a physical reason for
hydrologic time series; (3) to provide an overview of problems related to
the change should be identified before accepting the change as being an accu-
modeling of data that has been influenced by nonstationary causing factors.
rate reflection of the system.
In addition to graphical analyses, methods such as split filtering, spectrum
Models are used to make forecasts or predictions. Nonsystematic variation
analysis, and differencing are discussed. Error variation is usually considered
confounds the modeling of hydrologic time series. When a hydrologic system
constant, but nonstationary error variation is often the norm. Given the
undergoes change, the nonstationarity adds another level of complexity to the
importance in breakpoints to successfully detecting trends, assessing the dis-
modeling. Some models originate from theoretical considerations while oth-
tribution of breakpoints is discussed. The age-old issue of decision making
ers are entirely empirically based, but most current modeling efforts integrate
using statistical levels of significance is also discussed.
theory and the results of analyses of measured data. Theory alone is not
19.1 INTRODUCTION
generally adequate by itself, as little is known about the variations of the many
variables that are subject to change. While empirical analyses may provide
Change is inevitable! Measured data can reflect change, but it is not always accurate regenerations of the measured data, the resulting empirical models
easy to accurately detect and quantify an expected change via analysis of may not provide accurate estimates when they are extrapolated to conditions
measured data. Even where the physical change introduces systematic change outside the range of measurements. This is especially true where the pro-
into measured data, the systematic effect is often obscured by the variation cesses being modeled involve systematic trends about which theory provides
that is referred to as sampling variation. While some of the sampling variation little help.
may be random in nature, some of it may relate to other variables, such as Changes to hydrologic processes may influence values of measured hydro-
antecedent moisture content that were not measured. In brief, time series logic data. If the processes change gradually, such as a slow increase in urban-
analysis involves two steps: detect the change and then model the change. ization over a large portion of the record length, then the measured data will
Where forecasts and uncertainty assessments are needed, it is necessary to likely show a gradually varying trend. Change can be abrupt, such as when
model both the systematic and nonsystematic components of the processes. 25 acres of a 50-acre forested watershed is clear-cut in a short period of time.
Measured values of past occurrences are used in both the detection and mod- In such cases, a spike may be evident in the flood record. Both the gradual
eling of the change. Statistically separating the systematic variation from the trend and the abrupt spike represent nonstationarity. Thus, the first task in
total variation is often difficult, and failure to select an accurate model of the analyzing a time series of a hydrologic variable is to separate the systematic
systematic behavior can even prevent the detection of change. variation from the sampling variation.
Analysis of time-dependent processes is not unique to hydrologic engineer- The topic of the detection and modeling of nonstationarity could be
ing. Those who deal with electronic measurements have been attempting to addressed in a series of books, as it is an important topic that has been part of
accurately separate the signal from the noise for a longer time span than those many professional publications. The goal of this chapter is very narrow; spe-
who deal with hydrologic temporal sequences. Economists have long been cifically I will present the analysis of nonstationarity from the perspective of
decomposing economic data to identify trends especially to characterize the exploratory analysis followed by statistical detection, which itself is explor-
bull and bear market cycles. Poll takers try to identify cultural trends using atory. Three specific objectives will be briefly covered:
polling results that are subject to considerable sampling variation. Many of the • To discuss exploratory methods for analyzing hydrologic data that one
methods used today by hydrologists were first employed in areas such as these. suspects is subject to change that has resulted in a nonstationary time series.
A stationary time or space series is one where the statistical characteristics • To identify problems with standard statistical hypothesis tests that are
do not change over the length of the record. These characteristics include the commonly used in the exploration of measured hydrologic data.
mean, variance, and skewness. If a series were separated into many parts, then • To explore issues that confound the detection and modeling of nonsta-
the individual parts should have the same characteristics, at least within sam- tionarity, especially issues related to the nonsystematic variation that is very
pling variation. Change in a time series can be gradual or abrupt. A gradual prevalent in hydrologic data.
19-1
19_Singh_ch19_p19.1-19.6.indd 1 8/22/16 12:07 PM

19-2 Statistical Detection of Nonstationarity: Issues and Needs
19.2 EXPLORATORY METHODS FOR DETECTION exists. While it is designed principally to detect periodicities, a temporal trend
OF CHANGE can appear as a low frequency spike. Although the significance of a particular
The general purpose of exploratory data analysis is to gather sufficient under- spike on a power spectrum is difficult to assess, the results can still be useful
standing of the expected change so that experimental hypotheses about in exploratory analyses. As with filtering, several power spectral analyses may
change expected in the measured data can be developed. Detection analyses need to be made using different intervals. Interpretation of the data requires
should focus on the functional nature of the nonstationary period, the experience. Sample size is especially important in power spectrum analyses.
expected variation of any breakpoint, and the relative magnitude of the sam- The method does not provide a mathematical model of the trend.
pling variation. The investigative approach used in assessing the significance Differencing is another exploratory method that can be useful in making a
of each of these factors will influence the success of the overall exploratory preliminary assessment of a series that is potentially nonstationary. This pro-
effort. cedure creates a new series by taking the difference between adjacent values
When a time series is suspected of including the effect of change, the data and using the differences as a new series. If the new series consists of values
should initially be graphically analyzed. One intent of a graphical study is to that are statistically identical and independent, then the original series could
assess the functional nature of the trend. Additionally, if part of a time series be modeled with a linear function. If the differenced series does not suggest a
is believed to be the result of stationary conditions prior to the time of change, linear trend, then a second-difference series should be computed. If this dif-
then the location of the breakpoint will need to be assessed. The methods of ferencing produced a constant series, then the resulting series would suggest
analysis selected should depend on the expected type of change, for example, that the original series had a nonlinear trend. If all of the differenced series
abrupt versus gradual, linear versus nonlinear, and partial versus complete lack a trend, then the original series was likely stationary. Excessive random-
record change. The accuracy with which a trend function reflects the varia- ness in the original series will minimize the effectiveness of differencing. A
tion during the period of change depends, in part, on the correctness of the statistical test, such as the runs test or Kendall’s tau test, can be used to iden-
function selected. However, because the inherent sampling variation in tify when a differenced series is random.
hydrologic data is usually considerable, the effectiveness of most graphical A systematic, comprehensive approach to exploratory analysis can improve
analyses in suggesting the true nature of the nonstationarity is often minimal. the efficiency of an analysis and simultaneously increase the chance of devel-
Breakpoints in time series that involve change are especially difficult to accu- oping a representative model of the change. While the focus is usually to
rately detect using graphical analyses. Graphical analyses of time-dependent understand the systematic variation while suppressing the sampling variation,
measurements are usually more beneficial in detecting episodic or abrupt hydrologic data are often dominated by nonsystematic variation.
changes. Understanding this source of variation is critical to the accurate identification
Numerous methods for exploratory analysis beyond graphical analysis of nonstationarity. If the change to the physical processes is expected to cause
have been proposed and used. Past research in this area has led to the devel- change to the characteristics of the sampling variation, not just the systematic
opment of various methods for data manipulation. Several of these methods variation, then understanding the change to the nonsystematic variation is
for data manipulation are discussed in the following paragraphs. important to detect the true nature of the nonstationarity.
After initial analyses using graphical methods, filtering methods are gener-
ally the next method applied in time series analyses. Moving average filtering 19.3 STATISTICAL EXPLORATION OF NONSTATIONARITY
is the most common smoothing technique, but split filtering is often used
where seasonal variations are expected. Moving average filtering replaces Statistical tests are, in a sense, exploratory in that a test does not directly yield
each measured value in a time series with the average of a series of adjacent a model of the nonstationary processes. Instead, a statistical test seeks to
values. Since the most effective length of a filter is generally unknown, sev- determine whether or not a sequence is random in some characteristic. A test
eral different filter lengths are usually applied and individually interpreted. It compares the time series of interest to the conditions of theoretical random-
is important to recognize that filtering can remove variation associated with ness, for example, a change did not occur. The theory of a test recognizes that
the systematic processes, not just the variation associated with the nonsys- incorrect conclusions can occur, that is, in most analyses type I and type II
tematic processes. errors usually have nonzero probabilities of occurrence. While decisions can
Split filtering is a special case of the moving average filtering method. With legitimately be made using statistical tests, any decision should be considered
split filtering, the time series is separated into short duration sequences. Each as exploratory in the sense that it recognizes that errors could be significant
measured value in a sequence is replaced by the mean of the sequence. The from a hydrologic perspective.
filter length for a sequence does not have to be of constant duration. The Hydrologic data are subject to many different conditions of change, which
length of a season is commonly used with hydrologic data. While this filtering includes measurements thought to have: (1) an abrupt change, such as urban
method involves qualitative assessment of the smoothed series, it will gener- development over a short period of time; (2) gradual change over the entire
ally eliminate more of the sampling variation than moving average filtering; record length; or (3) a record that includes a portion that reflects stationary
however, it can be less effective for periodic functions where the filter length conditions and a portion for nonstationary conditions. Every statistical test
is not properly selected. Like the moving average filtering method, results and its critical values have sensitivities that differ depending on the types of
from several analyses based on different filter lengths should be evaluated. change. As such, a table of critical values for a specified statistical test may not
Where extreme events, such as outliers, are present in a time series, filtering be useable for a change in conditions that differ from the conditions for which
methods may not be sufficient to minimize the confounding effect of the the tabled values were developed. For example, the alternative hypothesis for
extreme events. Extreme events increase the total variation of a time series. the runs test is that the values are not random. Yet, the test has been used to
Therefore, extreme events in the stationary portion of the time series can test for both gradual and abrupt changes. The three conditions previously
reduce the likelihood of detecting a trend. Conversely, extreme events in the identified do not reflect the same fundamental theory, so different sets of
nonstationary portion of a time series can increase the likelihood of conclud- critical values would be appropriate even when the same test statistic is used.
ing that the trend is significant. For a time series that includes several extreme This fact applies to all tests, not just the runs test. It is, therefore, always
events, the series may be transformed using logarithms, with the magnitude important to verify that tabled critical values are appropriate for the condi-
of the extreme events experiencing the greatest change due to transformation. tions of change being investigated.
Filtering methods applied to the logarithms can be more effective in identify- The ability of a specific statistical test to identify the existence of change
ing the nature of an inherent trend. Ultimately, it is necessary to properly deal depends on the nature of the change and the extent of the sampling variation.
with the extreme events, which may involve censoring. Hydrologic change can be gradual, that is, occur over a long period of time,
Auto- and cross-correlation analyses can be useful for detecting trends. or abrupt, that is, occur over a short part of the record. Some tests are sensitive
Autocorrelograms that decrease at a low rate from the lag-zero point suggest to the specific nature of the change and even the location within the time
that the time series may be structured about a temporal trend. However, the series where change occurred. Tests that are sensitive to abrupt change can
time interval on which an autocorrelogram is computed can influence the lead to incorrect decisions if they are applied to data characterized by gradual
extent of the autocorrelation for small lags. The statistical significance of an change. Tests intended to detect gradual change may be sensitive to the loca-
autocorrelation coefficient should be identified, but since critical values of tion of a breakpoint. Knowing about such sensitivities can reduce the chance
correlation coefficients are based on an assumption of independence, the of applying the wrong test or using inappropriate critical values.
significance of a computed autocorrelation coefficient may not be interpreted The existence of change suggested by the result of a hypothesis test depends
in the same way as one based on independent events. Critical values for test- on the criterion used to indicate statistical significance. Too frequently, deci-
ing a Pearson correlation coefficient will not be appropriate for testing auto- sion making with a statistical test is based on an arbitrary level of significance
correlation coefficients. or a subjectively assessed rejection probability. This possibility dictates that
Power spectrum analysis, which is primarily used to detect periodic varia- the implications of the two types of statistical errors be clearly assessed and
tions, is another exploratory method that can suggest whether or not a trend used in justifying a decision. Some have argued that the levels of significance
19_Singh_ch19_p19.1-19.6.indd 2 8/22/16 12:07 PM

Effect of Nonconstant Error Variation 19-3
and rejection probabilities are arbitrary decision criteria. Computed rejection A simple example can be used to illustrate the generation of the two data
probabilities of statistical tests can be sensitive to the form of the change. series. The trend for each series will be the linear model:
Critical values of a test statistic are often used in making decisions with
statistical tests. Every statistical test is sensitive to the various factors that were yt = 0.8 + 0.1 t (19.2)
used in developing the critical values. For example, many parametric tests are where t is time (1 to 8). The random variation added to each sample point will
based on the assumption of the normal distribution. Any sensitivity, such as be standard normal deviates. In the first series, the normal deviates are ran-
to normality, should be considered in selecting a test to assess the significance domly sequenced. In the second series, the same set of normal variates is
of expected change in hydrologic data. Even distribution-free tests are sensi- used, but the individual values are sequenced according to their absolute
tive to general characteristics of measured data. When a test is applied to values. The following two series illustrate the two data generation scenarios:
measurements that reflect one type of change and the type of change is differ-
ent from the change in which the critical values are based, then decisions Trend 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6
made with the critical values may be incorrect. Random –0.8 0.4 0.7 –0.3 –0.6 0.2 0.9 –0.5
The runs test for sequential independence of values in a time series trans-
forms the continuously distributed values into a binary series (e.g., + vs. – or Series 1 0.1 1.4 1.8 0.9 0.7 1.6 2.4 1.1
0 vs. 1). Values above a criterion such as the median are assigned one symbol, Trend 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6
while those below the criterion are assigned to the other symbol. A run is
defined as a sequence of one or more values of the same symbol. If a series is Random 0.2 –0.3 0.4 –0.5 –0.6 0.7 –0.8 0.9
really randomly distributed, then the series will have a moderate number of
Series 2 1.1 0.7 1.5 0.7 0.7 2.1 0.7 2.5
runs. A time series is considered nonstationary if the series is represented as
having an unexpectedly large or small number of runs. While the runs test is Both series have a mean of 1.25. The standard deviation of series 1 is 0.711,
not statistically powerful, it is insensitive to outliers and the probability distri- while that of series 2 is 0.715. The Spearman–Conley correlation coefficient
bution of the original time series. was computed for both of the above series and tested for statistical signifi-
The Kendall test for trend uses essentially the same hypotheses as the runs cance. The first series was not found to be significant, with a Spearman–
test but is more powerful. The test is intended to detect monotonically Conley coefficient of 0.26, which has a rejection probability much greater
increasing or decreasing trends. The test compares each value in the time than 10%. This suggests that the values are uncorrelated in time, that is, they
series to ever other value in the time series, with the comparisons checking are stationary. The second series has a correlation coefficient of 0.45, with a
for change in a specific direction—increasing or decreasing. The Kendall test rejection probability of about 5.5%, which suggests that the series has a trend
is sensitive to either gradual or abrupt change, but the sensitivity differs with and is, therefore, statistically detected as nonstationary.
each type of change. The overall intent of the simulation was to investigate the influence that the
The Pearson and Spearman correlation coefficients can be used to detect random variation can have on the ability of a test to detect a trend. In generat-
autocorrelations in a series. While the Pearson test assumes the data are nor- ing the data sets, the characteristics of the trend and the random variation
mally distributed, the Spearman test is nonparametric, that is, distribution were systematically varied in order to show different effects, specifically the
independent. These tests are often misused with time series because the user effect of sample size, degree of scatter in the nonsystematic variation, and the
uses time (i.e., 1, 2, 3, …) as the second variable. Both tests are based on the scatter in the trend relative to the scatter of the nonsystematic variation.
assumption of two independent variables, and the time sequence is not a The results are summarized in Table 19.1.
random variable. For a univariate time series, the Spearman–Conley test was The first analysis assessed whether or not a temporally increasing variation
designed to test for serial independence. in the time series measurements would influence the ability of a statistical test
to detect change. Set A of Table 19.1 is the result of change of the random
variation, with standard deviations of the random component of 0.3, 0.7, and
19.4 EFFECT OF NONCONSTANT ERROR VARIATION
1.0. A sample size of 20 and the standard deviation of the trend were held
Some types of change introduce a trend into the principal variable. For constant for all samples. With a constant St of 0.623, the relative variations
example, urbanization causes peak discharge rates to increase. Change may would be 0.48, 1.11, and 1.59. The Kendall tau test was applied to each of the
also cause the secondary, unmeasured variables to change. For example, 10,000 generated series. The mean tau statistic decreased as the variation of
urbanization also causes change to soil moisture processes, baseflow rates, the random component increased even though the same systematic trend was
flow patterns, loss rates, and channel characteristics. These secondary changes used. This implies that the increased random scatter made it more likely that
can change the variances of the peaks and increase the sampling variation the trend could not be detected. This result was expected, but it is the differ-
reflected in the variation of the principal variable. Specifically, as the total ences in the mean values of tau that are noteworthy. The differences between
imperviousness increases, variations of soil moisture and vegetation can cause the mean tau statistics between the constant (Tc) and increasing random (Tr)
variations in the peak discharge rates that go beyond the change in the mean components differed by about 0.08, which would reflect significant variation
discharge rates due to the amount of imperviousness. Imagine a trend in peak in rejection probabilities. For a sample size of 20, the critical values of tau for
discharge rates that increases linearly with time. The change in the other 10 and 5% differ by 0.053, so the difference of 0.08 is not an insignificant dif-
variables can cause greater variation in the peak discharge rates, as the ference. In summary, if the Kendall test is applied to a process where a sec-
increasing trend progresses. Thus, the scatter about the trend line is increas- ondary variable introduces a nonconstant variation into measured values of a
ingly greater as time advances than the scatter prior to the breakpoint. time series, then the tau statistic will be inflated, which suggests that an even
Does this secondary effect of change in the variance of the variable influ- stronger trend in the data is present and likely due to the principal causative
ence the ability of a statistical test to detect the trend? For example, urbaniza- variable.
tion may increase both the magnitude of the discharges, as represented by a
linear trend, and the variance of the measured values. This would appear that
TABLE 19.1 Kendall Tau Statistics for Constant (Tc) and
the sampling variation is increasing as the mean discharge increases. To inves-
Increasing Random (Tr) Variation as a Function of Sample Size (n)
tigate this possibility, a simulation experiment was designed. Data sequences and the Variation of the Trend (St) and the Random (Sr) Variation
were generated for two conditions: constant variation and increasing vari-
ance. A linear trend was assumed to reflect the effect of change in the princi- Test n St Sr Tc Tr
pal variable y: A1 20 0.623 0.297 0.754 0.834
yt = a + bt (19.1) A2 20 0.623 0.692 0.493 0.613
where t is time (integer from 1 to n). A set of n random normal deviates were A3 20 0.623 0.989 0.376 0.435
generated to represent the sampling variation. To form the first time series,
B1 10 0.603 0.682 0.534 0.642
the deviates were randomly added to each value of the linear trend. A second
time series was generated by ordering the same random normal deviates B2 30 0.607 0.695 0.478 0.605
according to their absolute magnitude and adding the signed values to the
B3 50 0.595 0.697 0.467 0.599
sample points of the trend. Thus, the two time series included the same trend
values and the same deviates, but with the second set, the deviates were C1 30 0.304 0.695 0.273 0.380
ordered by absolute value magnitude. Thus, the two series would have very C2 30 0.607 0.695 0.478 0.605
similar total variations in spite of the systematically increasing variation of the
second series. Ten thousand series were generated for both scenarios. C3 30 0.911 0.695 0.612 0.735
19_Singh_ch19_p19.1-19.6.indd 3 8/22/16 12:07 PM

Sample size is usually a factor that influences statistical decisions. Set B of TABLE 19.2 The Effect of Filtering of Kendall’s Tau Statistic
Table 19.1 shows the effect of sample size on the ability to detect a trend using (T) for the Linear (L) and Exponential (e) and for Selected
Kendall’s test. Ten thousand simulations were made using the same variation Sample Sizes (n) and Values of the Standard Deviation of the
in the trend (~0.6) and in the random variation (~0.7) for samples of 10, 30, Trend Values (St) and the Standard Deviation of the Random
and 50. As the sample size increased, the magnitude of the trend decreased, Numbers (Sr)
so the relative variation would not change. Again, the sampling variation was n Sr TLr TLs Ter Tes
either random or ordered by magnitude. Neither of the mean tau statistics
St
was overly sensitive to the sample size, although both mean tau values
decreased with increasing sample size. This inverse trend was expected, as it 10 2.0 0.77 0.93 0.76 0.92
is known that critical values of tau decrease with increasing sample size.
1.0 0.55 0.75 0.54 0.74
Again, the mean values of tau for the case of randomly distributed sampling
variation were less than those where the random variation increased as the 0.5 0.32 0.48 0.32 0.47
trend line increased, which again shows that the hypothesis test of a trend can
be influenced by variation of secondary variables. 50 2.0 0.72 0.84 0.58 0.69
As the level of urbanization increases, the ability to detect a trend in peak
discharge rates should increase. Increasing the variation of the peaks should 1.0 0.52 0.69 0.42 0.55
have about the same effect for different trend magnitudes. Set C of Table 19.1
0.5 0.31 0.47 0.26 0.38
shows the effect of increased variation of the trend (St) relative to variation of
the random component (Sr), that is sampling variation. As the magnitude of
the trend increased, the mean tau significantly increased, which indicated a
greater likelihood that the trend is significant, more significant than it actu- added to each ordinate of the trend was normally distributed and varied from
ally is. The differences between the mean values of tau remained essentially 50 to 200% of the variation in the trend. For each scenario, the simulation
the same as the trend increased, with a difference of about 0.11 between Tc included 10,000 samples of size n, where n varied from 10 to 50.
and Tr. With each sample size of n, the Kendall tau test was conducted to test the
In summary, hypothesis tests, such as Kendall’s test, often ignore the effect alternative hypothesis that the sample included a trend. The mean value of tau
of secondary variables. The tests assume that the process causing the variation was determined for each combination of the three factors: sample size, rela-
in the time series is independent of other factors. The results of these analyses tive error variation (St/Sr), and the a priori smoothing. The results are given
show that a second factor that causes a nonconstant error variation in the in Table 19.2 and yield several interesting observations.
measured data can significantly influence the rejection probability of a statis- First, filtering of the time series caused the values of tau to increase. On
tical test. Multivariate statistical tests may be more effective in detecting average, filtering of the linear model increased the value of tau by about 0.16,
change in hydrologic data that is dominated by sampling variation. which reflects a very significant change in the rejection probability, that is, the
p-value. The increased tau means that the trend was more evident following
19.5 EFFECT OF A PRIORI FILTERING OF TIME SERIES the smoothing and that the filtering had a greater effect on the nonsystematic
variation than on the systematic variation. This was true for both the linear
Filtering or smoothing of a time series is intended to reduce the random and exponential models.
variation without reducing the systematic variation. When filtering is under- Second, as the relative variation of the trend component increased, tau
taken, the total variation of both the raw series and the smoothed series increased, as expected. The increase was very significant, with marginally
should be computed and compared. While filtering will reduce the total significant values of tau for St/Sr = 0.5 to highly significant values for St/Sr > 1.
variation, the hope is that the smoothing will only reduce noise, not variation This result implies that the Kendall test statistic depends heavily on the rela-
associated with the signal. In some cases, filtering can cause a greater reduc- tive variation.
tion in the systematic variation than in the nonsystematic variation. For Third and most importantly, the functional nature of the trend can be a
example, if a time series included a periodic trend (e.g., annual cycle) and the significant factor in the outcome of the Kendall test. When comparing the
smoothing interval was set equal to the period, then the smoothing would mean values of tau for the two functions of Eq. 19.3, the linear trend is more
eliminate the systematic variation associated with the cycle. Filtering should easily detected than is the exponential trend. This is especially true for the
not be applied without an understanding of its potential impact. larger sample size. For a sample size of 10, the values of tau were nearly iden-
The results of filtering are often interpreted as part of a graphical assess- tical, as the early portion of the exponential model more nearly approximates
ment, usually completed prior to performing a statistical test on the time the corresponding portion of the linear model. As the length of the time series
series. The question then arises: Can filtering influence the decision of a increased, the curvature of the exponential model has a greater influence on
hypothesis test? Should the test be made on the filtered series or the nonfil- the value of tau. For the larger sample, the early portion of the exponential
tered series? If filtering reduces the variation associated with a trend inherent function is relatively low sloped, which results in smaller values of tau. The
to the time series, then the likelihood of acceptance of the null hypothesis of linear model has similar total variation but a steeper early section that results
a lack of a trend will increase, which would unnecessarily increase conclu- in higher values of tau. This result clearly shows that the rejection probability
sions and actions related to the lack of a trend. Conversely, if the filtering for a hypothesis test depends on the trend model used to represent change.
reduces the variation associated with the nonsystematic variation more than
the variation associated with the systematic change, then the test is more
likely to detect an effect when it is applied to the smoothed series. Knowing 19.6 DISTRIBUTION OF A BREAKPOINT
the potential effect of filtering can be important to assessing the results of Many hydrologic time series include both stationary and nonstationary sec-
statistical hypothesis testing. tions. In order to model both parts, the breakpoint must be identified.
A series of simulation analyses were undertaken to assess the effect of the Assume that a nonstationary section with an increasingly linear slope follows
filtering of measured data prior to performing a hypothesis test for trend a stationary section, with sampling variation being significant in each section.
detection. Various factors would influence the extent of the effect; this Of course, both the magnitude of the trend and the time of the breakpoint are
includes the sample size, the model structure, and the relative error variation, not known, that is, they are both values of random variables. If the breakpoint
which was defined here as the ratio of the standard deviation of the trend is not identified accurately, then any function fitted to the assumed nonsta-
values (St) to the standard deviation of random variation (Sr). Thus, time tionary section, which is based on an inaccurate breakpoint, will also likely be
series were generated by adding systematically varying values of a trend and inaccurate. The fact that a breakpoint is a random variable indicates that an
non-systematic variation; this ensured that the time series included a trend exact value should not be expected. A breakpoint has a sampling distribution
and the null hypothesis of no trend should be rejected. Two model structures and any assumed value is subject to sampling variation.
were used to represent the systematic component: The ability of a hypothesis test to detect a trend where a breakpoint exists
zero-intercept linear form: y = bt (19.3a) depends on several factors, including: (1) the magnitude of the sampling
variation in both the stationary and nonstationary sections of the time series;
and (2) the relative steepness of the trend of the nonstationary section; (3) the
relative lengths of the two sections as well as the total series length; and (4)
exponential form: y = c(edt –1) (19.3b)
the functional nature of the trend. Since each of these factors is associated
in which t is time, and b, c, and d are empirical coefficients. Both of these with randomness, then the breakpoint will be a random variable and have a
functions pass through the origin, that is, y = 0 at t = 0. The random variation distribution function, which depends on the previously stated factors. Unless
19_Singh_ch19_p19.1-19.6.indd 4 8/22/16 12:07 PM

Conclusion 19-5
the location of a breakpoint is known exactly from other data, the variation of hydrologic data. Second, critical values for most tests may not apply to situa-
its probability distribution reflects it uncertainty. Therefore, knowing or not tions that involve hydrologic change. This relates to the discordance between
knowing the sampling distribution of a breakpoint could have an impact on hypothesis statements about hydrologic change and the conditions on which
decision making. the development of the statistical tests were based. Third, statistical analysis
To show the effect of the relative variation in the two sections on the dis- of hydrologic data generally assume a temporally univariate condition, which
tribution of a breakpoint, samples were generated with the first half of each inflates the nonsystematic variation and makes it difficult to detect change.
series being stationary and the second half being nonstationary based on a Fourth, in the case of hydrologic data that involves both stationary and non-
linear trend (Eq. 20.1). The sampling variation of the values of the stationary stationary sections, the sampling distribution of the breakpoint is not always
portion about the mean was assumed constant over time. The variation of recognized as a source of uncertainty. Given issues such as these, it seems
the nonstationary section included both the variation associated with the reasonable to ask whether currently available statistical methods are adequate
trend and that of the sampling variation. It would be easier to detect the loca- to justify a conclusion of nonstationarity of hydrologic data.
tion of the breakpoint if the trend portion had a significant slope; then the One intent of the analyses discussed herein was to assess issues related to
relative variation between the two sections would be more evident. To quan- the statistical detection of hydrologic change. A wide variety of statistical tests
tify the relative variation between the two sections, the mean ratio Rs of the have been developed to test empirical data for change. While these tests are
standard deviations of the stationary part to the nonstationary section (SS/Sn) useful, a reliable decision criterion has not gained wide acceptance. The pit-
was used. falls of using the level of significance as a criterion are well known. For
A series of simulations, each involving 10,000 samples, generated time series example, the use of a seemingly arbitrary level of significance such as 5% is not
composed of both stationary and nonstationary functions. A linear function adequate for a number of reasons. More importantly, the implications of the
was used to generate individual values for the nonstationary trend in each type II error are more decisive than those of the type I error, so the level of
generated series. The slope of the linear trend was changed to vary the variance significance needs to be replaced with a more objective and relevant criterion.
of the individual values that represented the nonstationary section. Random Analyses presented herein have shown that the functional structure of the
variation was added to each sample point including the stationary part of each trend can be an important factor in the detection of change. When the non-
section. The standard deviations of both sections were computed, with the stationary processes are assumed to follow one functional form while another
ratio (Rs) of the standard deviation of stationary section to the standard devia- form is more theoretically justified, then the statistical detection of a trend
tion of the nonstationary section used to reflect the level of nonstationarity. Rs may be suppressed. While this issue needs more investigation, it does suggest
would be small when the trend of the nonstationary section is dominant. For that the decision criterion is not just a function of the sample size and level of
each sample, the Kendall tau test was conducted sequentially starting with the significance. The distribution of the test statistic appears to be a function of
first three points, then the first four, the first five, etc. For each of the 10,000 the model structure. This can complicate both the detection of change and the
samples, the point at which the minimum tau occurred was noted; this time acceptance of a broadly applied decision criterion.
would indicate the most likely breakpoint. As the ratio of the two standard The power of a test is an important criterion when selecting a statistical
deviations decreased, as would occur as the slope steepened, the likelihood of test. The power of a test varies with factors such as the sample size, the extent
the breakpoint going beyond the end of the stationary section would also of the sampling variation, and the nature of the change. Some tests are sensi-
decrease. This reflects a relative rise of the slope of the nonstationary section, tive to abrupt change while others are sensitive to gradual change. Generally,
such that the actual breakpoint is easier to detect. Thus, a smaller portion of the power of a test increases with increasing sample size and decreases with
the computed values of tau occurs in the nonstationary section. If Sn is large, increasing sampling variation. When selecting a test, issues such as these
then a trend exists, and Ss/Sn is small. If a trend exists, then large values of should be investigated in order to ensure that the rest is appropriate for detec-
Kendall’s tau are expected because of the nonstationary section. The minimum tion of the specific type of change. Guidelines such as the following reflect
tau will then appear in the section associated with the stationary section. Thus, knowledge that should be gathered when selecting a test:
the probability density function of the minimum tau would have a relatively • Spearman’s test is more powerful for detecting a gradual trend than for
small proportion in the section associated with the stationary series. detecting an abrupt change.
Table 19.3 summarizes the results of the simulations. Each of the 10,000 • Spearman’s test is not correct when time or space is the second variable;
generated series had a record length of 20, with the breakpoint between times in such cases, the Spearman–Conley test should be used.
of 10 and 11. For low standard deviation ratios, that is, less than 0.1, the break- • As the sample size of a trend increases, the power of the runs test
point was always located in the stationary section because the nonstationary increases.
trend was relatively large. For larger ratios of the standard deviations, sample • The runs test is more powerful when the abrupt change is near the center
values of the breakpoint occurred over the entire record length, including as of the record length than when the change is near either end of the series.
far as the end of the nonstationary section, which essentially suggests that the • The Mann–Whitney test is more powerful for an abrupt change than for
location of a breakpoint cannot be accurately identified. This occurs because a series with a gradual change, especially when the abrupt change is near the
the trend is rather shallow. The proportion of tau values in the nonstationary center of the record length.
section is an indication of the probability distribution of the breakpoint. • The Kendall test is most powerful when used with data characterized by
a gradually varying trend.
Similar information is available for other tests.
19.7 CONCLUSION
Numerous methods of trend detection have been proposed; this includes
Statistical detection of change has some limitations. First, the issue of the both function fitting algorithms and statistical hypothesis tests. The former
decision criterion, that is the level of significance, needs to be investigated. provides a model while the latter only infers whether or not a trend exists.
The use of standard values such as 5% will likely not be appropriate with Both approaches focus on the systematic variation. Given that the sampling
variation imbedded in hydrologic time series is often very significant to the
point of masking the presence of a systematic trend, new approaches that
TABLE 19.3 Assessment of Breakpoint Location as a focus on the sampling variation should be developed. Part of the sampling
Function of the Standard Deviation Ratio (Rs) = Ratio of the variation may be due to variation associated with unmeasured variables, so
Standard Deviations of the Stationary and Nonstationary
function-fitting methods could be used to statistically hold such variation
Sections; Pn = Proportion of Tau Values in Cells 11 to 20, that
is the Nonstationary Section; C = Largest cell Number for a constant. Statistical methods such as multivariate part and partial correlation
Minimum Tau and regression exist and could form the basis for new methods of modeling
multivariate induced change. For example, part regression analysis could be
Rs Pn C used to remove the effect of antecedent moisture conditions from a series of
0.008 0.0000 10 measured peak discharge rates, which would then enable the trend due to a
causative factor such as urbanization to be more easily detected and modeled.
0.032 0.0000 10
Hydrologic time series are frequently dominated by variation that appears
0.080 0.0000 10 to be random. This supposedly random variation makes it difficult to identify
a trend. Is all of the variation really random or is variation related to factors
0.128 0.0012 11
that have not been controlled or measured? For example, time series of peak
0.361 0.0405 14 discharge rates are characterized by considerable scatter. Antecedent moisture
0.699 0.0974 18
condition (AMC) is likely a causative factor of some of the variation, but since
AMC statistics and measurements on AMC are generally not available, varia-
1.000 0.2364 20 tion in the peak discharge rates that is associated with AMC is lumped in with
19_Singh_ch19_p19.1-19.6.indd 5 8/22/16 12:07 PM

the true sources of random variation. If progress is to be made in the model- would improve. The rationality of the coefficients of the multifunction trend
ing of nonstationary processes, it seems that metrics to represent these sec- should be considered as they reflect on the effects that will be attributed to the
ondary factors should be developed so that measurements can be made. This individual components.
would help control the variation that is falsely assumed to be sampling varia- All of this discussion leads to the conclusion that there is both bad news
tion. This approach may require the development of new methods of multi- and good news. The bad news can be summarized as follows:
variate detection of trends. • Unaccounted for variation in measured hydrologic data is very signifi-
Trends are currently represented using a mathematical function that cant and prevents exploratory assessments of hydrologic data from providing
depends only on time of occurrence as the predictor variable. If measure- convincing statistical evidence of change.
ments on a causative factor, such as AMC or temperature, were available, then • Critical values for many traditional statistical tests were not designed for
the relation used to represent the trend could be made a function of time and the assessment of the types of change inherent to measured hydrologic data.
the other variable. This may treat the variable and time as interactive or inde- • Current knowledge of the functional forms that are appropriate for mod-
pendent, which would depend on the existence of a physical relation. The eling hydrologic change is deficient, and the arbitrary or incorrect selection
significance of the external variable would need to be checked. If the variable of a functional form can introduce uncertainty into statistical results of
is an important part of the sampling variation, then the relation should physical change.
account for some of the sampling variation and then the coefficients related The good news would then follow that these problematic issues provide the
to trend would be more physically realistic and the goodness-of-fit statistics opportunity for research.
19_Singh_ch19_p19.1-19.6.indd 6 8/22/16 12:07 PM

Chapter 20
Spatial Analysis and
Geostatistical Methods
BY
DIPANKAR DWIVEDI, BAPTISTE DAFFLON, BHAVNA ARORA,
HARUKO M. WAINWRIGHT, AND STEFAN FINSTERLE
ABSTRACT unconstrained approaches used for describing subsurface heterogeneity,

While Section 20.2 and Section 20.3 focus on data-driven approaches, Section
Quantifying and mapping spatial and temporal patterns in soil and landscape
4 describes relevant inverse modeling and parameter estimation for con-
properties are crucial to modeling hydrological and biogeochemical fluxes
straining heterogeneous field simulations. The focus of this chapter is on
reliably. The field of spatial analysis and geostatistical methods has seen
providing an overview of the different techniques available for analyzing
tremendous growth over the past two decades, especially with the advent of
heterogeneity in the subsurface. Consequently, the discussion of the different
distributed hydrologic modeling, development of data mining and assimila-
techniques is not comprehensive but selective in the sense that the most
tion approaches, and increase in computational power. Of course, numerous
important techniques are broadly described.
challenges are still associated with characterizing and representing heteroge-
neity given that data are sparse, heterogeneous, complex, and convoluted. In
such a scenario, traditional methods of data analysis and integration into 20.2 DATA TYPES AND METHODS
physical models may not be adequate. On the other hand, spatial analysis and Time-series data include values that are typically collected by continuous
geostatistical methods provide an excellent basis to handle such data and measurement over time. For instance, a pressure transducer measures the
further study data mining and data model integration in hydrology with groundwater level continuously. Space-series data (spatial data) include
increasing complexity in data (e.g., big data). Accordingly, we attempt to point-scale (or pointwise) attributes (e.g., source of the data, date, salinity,
provide readers a modest background into various data types and methods of temperature) collected at spatial or georeferenced locations. Generally,
data mining and spatial analysis. We also describe traditional and advanced directional (e.g., east of), topological (e.g., overlap), or other spatial relations
geostatistical methods to explore structural information in heterogeneous (e.g., distance, adjacent to) inherent in spatial data are absent in nonspatial
fields and to generate heterogeneous models with or without data constraints. data (Shekhar et al., 2001). The distinctive characteristics that separate spatial
We believe that this chapter will provide the reader with a sense of the vast data from nonspatial data is a point of debate (Anselin, 1999; Pfeiffer, 1996).
diversity of methods used in spatial analysis and geostatistics. In general, spatial data can be subdivided into two major categories: continu-
ous or discrete (Shekhar et al., 2003). The Earth science field typically deals
20.1 INTRODUCTION with spatially continuous measurements. For example, soil moisture values
are often collected to understand water and energy exchanges at the land
Spatial patterns are of fundamental importance to many disciplines of the surface–atmosphere interface. Soil moisture data are collected across multiple
Earth sciences such as geology, geomorphology, and hydrology. Historically, scales (e.g., point, field, satellite footprint-scale), yet soil moisture data gener-
there has been more emphasis on observing and interpreting temporal ally need to be inferred at locations or scales that have not been directly
patterns (e.g., streamflow), particularly in catchment hydrology. Until sampled. Additionally, some attributes such as spatial attributes may or may
recently, not many studies have addressed the challenges of spatial heteroge- not vary in time. For example, soil moisture content changes in time, whereas
neity, particularly at the large scale. However, there has been a growing permeability can be regarded as constant over time (time scale of years). Data
interest in representing and understanding spatial patterns in hydrology that include both spatial and temporal attributes are categorized as spatiotem-
through the advancement of spatially distributed hydrological models. poral data. Nitrate time-series data collected at various locations are a
Despite having models that are capable of interpreting spatial responses, there examples of spatiotemporal data. Spatiotemporal data have implicit or explicit
are numerous challenges associated with characterizing and representing dependencies or patterns among closely or sometimes remotely spaced data
spatial heterogeneity in these hydrological models. For instance, some points in time and space as well. Deciphering such patterns in the data is the
challenges may include how to (a) describe, characterize, and quantify spatial broader goal of spatiotemporal analysis. Spatiotemporal analysis can be done
heterogeneity; (b) identify interlinkages between processes and patterns; through both data-driven and model-driven approaches that will be explored
(c) represent spatial heterogeneity in models across spatial scales. While all of in this chapter.
these issues are open research questions, we hope to give readers some
background of geospatial methods and insights into hydrological behavior. In
20.3 SPATIAL ANALYSIS
this chapter, we will discuss spatial analysis and geostatistical methods, with
emphasis on selected standard topics to showcase theory and applications in In the Earth sciences, spatial analysis is concerned with making inferences
hydrologic science and technology. about processes or patterns affecting heterogeneous geological media using
The chapter is organized as follows: Section 20.2 describes various data their topological attributes, geometric features, or geographical properties.
types in Earth sciences, especially in hydrology and associated methods for Spatial analysis is a field that has seen growth through several historical
data analysis, Section 20.3 introduces the concept of spatial analysis and data disciplines. One of the earliest applications of spatial analysis has been docu-
mining techniques, Section 20.4 provides a description of constrained and mented by Krige (1952) who observed an improvement in his estimates of ore
20-1
20_Singh_ch20_p20.1-20.10.indd 1 8/26/16 4:37 PM

20-2 Spatial Analysis and Geostatistical Methods
grades in mining blocks when taking into account the grades in neighboring variance is represented in a minimum number of dimensions. These multi-
blocks. This application was partly inspired the work of Matheron on the variate statistical techniques are useful for dimension reduction (Jolliffe, 2002).
fundamental theory of random processes and his theory of regionalized The transformed attributes have a descriptive power that is more easily
variables (Matheron, 1965). Spatial analysis is successfully applied in cartogra- ordered than the original attributes of the data. SVD is closely related to PCA;
phy for map visualization and data maps (Getis, 2008). Other examples of however, SVD is more general. SVD gives two sets of basis vectors—one along
spatial pattern analysis include confirming the theory of Gondwanaland the column of the data matrix (the length of the data) and the other along the
which states that all continents formed one land mass (Shekhar and Chawla, rows of the data matrix (dimensions) of the data. By contrast, PCA gives basis
2003a), and describing spatial patterns in stream networks (Ganio et al., 2005). vectors only along the rows of the data matrix. EOF is also similar to PCA, but
One major challenge with the large array of techniques available for spatial more general. EOF can be used to analyze both spatial and temporal patterns.
analysis is that the value of a specific technique depends on the objective and In other words, EOF analysis is analogous to the geographically weighted
the nature of the data, and thus choosing the adequate technique is not always PCAs (Monahan et al., 2010). These techniques are popular in hydrology for
trivial (Pfeiffer, 1996; Dale and Fortin, 2010). This section describes a few identifying major factors and patterns in the hydrologic data such as water
fundamental issues related to the handling of spatial data. The other challeng- quality (e.g., Helena et al., 2000; Dwivedi et al., 2013).
ing issue in spatial analysis is the modifiable areal unit problem (MAUP).
MAUP is primarily concerned with the fact that the boundaries of the areal Wavelet Analysis
units are arbitrary, modifiable and user-dependent, and the resulting analysis Wavelet analysis is a technique that facilitates multigranularity decomposition
is inconsistent. MAUP was first recognized by Openshaw and Taylor (1979), of both spatial and temporal datasets through the use of a data window that
as spatial bias was introduced in their analysis of correlation coefficients, can be stretched and translated with a flexible resolution in both space
which changed values for different levels of spatial aggregation (Wong, 1981). (or time) and frequency domains (Kumar and Foufoula-Georgiou, 1997;
Another major factor influencing spatial data analysis is the scale at which Martínez et al., 2011). This space-frequency localization property of wavelets
observations are made versus the scale at which spatial variation is present can be very useful as it allows the Earth science data to be examined at a detail
(i.e., process scale). Other considerations are that spatial relationships can matched to its scale (Lau and Weng, 1995; Kumar and Foufoula-Georgiou,
exist at different levels of granularity. 1997; Arora et al., 2013). Figure 20.1 illustrates the use of wavelets in charac-
With the rapid advancement of analytical cartography, geographical infor- terizing spatial and temporal variability in near-surface soil moisture data.
mation system (GIS), and remote sensing techniques, steps are needed to avoid The local wavelet spectrum of soil moisture clearly reveals power features
combining incompatible datasets. In the Earth sciences, emphasis should be related to the precipitation events on the 1–48 hour scales.
placed on error propagation analysis and the avoidance of the ecological
fallacy (ecological vs. individual inference), that is, errors stemming from 20.3.2 Inferring Structural Information
of Heterogeneous Model or Data
inferring the outcome at individual units based on the results of aggregation
(Portnov et al., 2007). Sampling designs should also account for spatial One clear distinction that separates spatial statistics from classical statistics is
heterogeneity or incorporate additional information (such as soft information). that spatial data tend to be highly correlated while statistical analyses can
work with the assumption of independent data (Anselin, 1999). This spatial
property is evident from the first law of geography which states that “every-
20.3.1 Data Mining Techniques
thing is related to everything else, but near things are more related than distant
Earth science data that represent high spatial, temporal, or spectral resolu- things” (Tobler, 1979). In spatial statistics, this property is called spatial auto-
tion, and dimensions—from in-situ measurements, remote sensing, LiDAR correlation and is quantified using measures such as Moran’s I and Geary’s C
data collection missions, etc.—are increasingly large and complex because of (Shekhar et al., 2010; Cressie, 2015). There are local indicators of spatial
advancements in technology and instrumentation. Data reduction techniques autocorrelation as well, which may be more suited for Earth science
are required to (a) represent data more tractably, (b) filter redundancy from applications.
oversampling, and (c) apply computationally expensive algorithms for inter- Describing or removing the autocorrelation structure is one of the most
preting patterns or generating value-added data products. In general, spatial important and challenging aspect of analyzing spatial data (Legendre, 1993;
data mining is a relatively young field that is concerned with extracting inter- Griffith and Peres-Neto, 2006). One of the earliest manifestations of removing
esting, potentially useful and nontrivial patterns from large spatial datasets the impact of spatial dependence was reported by Fisher (Fisher, 1937), who
(Shekhar and Chawla, 2003c). Some examples include feature selection (e.g., was primarily concerned with estimating crop response to agronomic prac-
cluster analysis, classification) and axis rotation [e.g., principal component tices and different varieties in the field.
analysis (PCA), singular value decomposition (SVD), and empirical orthogo-
Empirical Variograms
nal function (EOF) analysis], and transformation (e.g., wavelets).
In geostatistics, the property value (such as permeability, porosity) at a given
Cluster Analysis location u is defined as a random variable Z(u). The underlying assumption
Cluster analysis is an unsupervised pattern recognition technique that reveals is that the values at two locations are more likely to be similar and more
natural structure of a data set without making a priori assumptions about the correlated when the two locations are closer. Such spatial correlation
data, in order to apportion a set of observations into subsets or clusters based structure is defined as a function of the lag distance h by the spatial covari-
on their nearness or similarity (Ashley and Lloyd, 1978; Vega et al., 1998). ance or correlation C(h). Alternatively, we may define the measure of
Spatial clustering is defined as the process of grouping a set of similar spatial difference, referred to as semivariogram, since the two locations are more
objects into meaningful subclasses (i.e., clusters) (Han and Kamber, 2006). A likely to have different values when they are farther apart. The variogram
well-known historical example of spatial clustering includes the mapping of g(h) is defined as:
cholera, which had become an epidemic in Soho, London, in 1854. Spatial
mapping by Dr. John Snow, an epidemiologist, helped locate a water pump as 2 γ (h) = E{[Z (u + h) − Z (u)]2 } (20.1)
the source of the disease (Shekhar and Chawla, 2003a). where Z(u) and Z(u + h) are two random variables separated by distance h.
Spatial clustering algorithms can be sub-divided into six categories: partition- We assume the field to be stationary, that is, the mean and covariance struc-
ing, hierarchical, grid-based, model-based, constraint-based, and density-based ture do not change over space and E[Z(u)] = E[Z(u − h)]. The variogram is
clustering algorithms (Berkhin, 2006). The choice of the clustering algorithm is linked to the covariance C(h) by
based on the type of data, the expected outcomes, and the trade-offs between
quality and speed of clustering (Han et al., 2001). For example, spatial clustering γ(h) = 2[C(h) − C(0)] (20.2)
can be applied to point data with applications in disease surveillance or crime where C(0) is the data variance. When anisotropy is suspected, a variogram
analysis, and to polygon data with applications in census surveys or landscape needs to be defined by a distance in each of the principle and orthogonal
ecology. Applications of clustering in the Earth sciences include the detection of directions. In real applications, the variogram often does not become zero as
seismic faults by classifying the records of earthquakes, or the generating h approaches zero, which implies that there is variability at a single location.
thematic maps in GIS by clustering feature vectors (Ester et al., 2000). This is called the nugget effect, which is attributed to several factors, such as
measurement errors and small-scale variability.
PCA, SVD, and EOF In practice, the empirical (or experimental) variogram is computed based
PCA, EOF, or SVD are techniques based on an orthogonal linear transforma- on the data by taking the difference of each pair according to Eq. (19.1).
tion that rotates the data to a new coordinate system so that most of the Because it is not possible to find the same distance for more than one pair,
20_Singh_ch20_p20.1-20.10.indd 2 8/26/16 4:37 PM

Heterogeneous field estimation and simulation 20-3
1–48 h
Aug 1st Aug 1st
July 15th 0.75

July 15th
1.5
Date
Date
3
July 1st July 1st
June 15th June 15th

90 60 30 0 4 2 0 4 2 0
(a) Distance (m) (b) Av.var soil moisture (%) (c) Av.var rain (mm2)
Figure 20.1 (a) Distance–time wavelet analysis of soil moisture for a predefined time window of 1–48 h. The shaded contours are at normalized
variance of 0.75, 1.5, and 3; the bold line is the 95% confidence level using a lag-1 autocorrelation. (b) Scale-averaged wavelet power of soil mois-
ture. The bold line is computed from data series; the thin line is computed from PCA. (c) Scale-averaged wavelet power of precipitation. The dashed
line is the 95% confidence level using a lag-1 autocorrelation. [Reproduced from Parent, A.-C., Anctil, F., & Parent, L.-C. (2006). Characterization of
temporal variability in near-surface soil moisture at scales from 1 h to 2 weeks. Journal of Hydrology, 325, 56-66, Fig. 3, with permission from Elsevier.]
bins are created for all distances in the interval [h – Dh/2, h + Dh/2], and the variogram models. In addition, various assumptions have to be evaluated
square mean is taken in each bin according to Eq. (19.1). There are several such as the presence of anisotropy and stationarity.
issues in computing such an empirical variogram. The variogram approach is,
for example, known to perform the best when the overall distri- 20.4 HETEROGENEOUS FIELD ESTIMATION
bution of Z(u) is Gaussian. Several transformation methods, such as normal AND SIMULATION
transform and the Box–Cox transform, are available to address variables that
Nonspatial attributes (e.g., elevation, temperature, soil moisture) are repre-
are non-Gaussian. In addition, exploratory data analysis is required for
sented in space by surfaces or slices at a given z. Surfaces can be visually
removal of obvious trends in the data, choosing the principle and orthogonal
displayed as points, contours, triangulated irregular networks grids, maps,
directions, and defining distance bin sizes and angle tolerance to compute the
and images. Typically, a surface is a spatial data structure that defines space
variogram for each lag and direction. There are many software packages
(e.g., topographic features) or attributes associated with space (e.g., soil mois-
available, such as gstat (Pebesma and Gräler, 2015) and geoR (Diggle and
ture) as an array of cells that are arranged for locations x, y in space. Because
Ribeiro Jr, 2007) in R, GSLIB (Deutsch and Journel, 1998), and SGeMS (Remy
it is impractical to obtain measurements for each raster cell, various methods
et al., 2009).
of spatial interpolation are used to derive the surface of nonspatial attributes
or properties from sample points. Several interpolation methods are based on
Variogram Model the principle of spatial autocorrelation and level and strength of interdepen-
Although the empirical variogram provides the structural information of dence. Surface interpolation methods can be broadly categorized
subsurface heterogeneity, it is often too noisy or erratic to use for interpola- as deterministic and geostatistical. Deterministic methods include inverse
tion or geostatistical simulations. The common practice is to define a con- distance weighting, Barnes interpolation, bilinear interpolation, bicubic inter-
tinuous function of a parametric variogram model and to determine its polation, Delaunay triangulation, natural neighbor, and spline interpolation.
parameters through fitting this model to the empirical variogram. There are These methods are simpler and based on direct measurements. However,
several variogram (or covariance) functions commonly used in subsurface geostatistical methods, such as kriging and co-kriging, are more advanced and
science, including spherical, exponential, Gaussian and Matern functions. sophisticated techniques for surface interpolation and surface modeling since
The chosen function and their parameters (such as scale parameters) control they include the data-defined heterogeneity structure, such as spatial correla-
the spatial pattern and variability. Equations for variogram models are given, tions and anisotropy.
for example, in Deutsch and Journel (1998), Rubin (2003), and Diggle and The most common use or goal of geostatistics is to map a spatially hetero-
Ribeiro Jr. (2007). An appropriate variogram model can be selected by com- geneous property over a spatial domain based on sparse measurements. In the
paring the residual of fitting, or information criteria such as Bayesian or deterministic sense, this procedure is called interpolation between measured
Akaike information criteria (Diggle and Ribeiro, 2007). locations. In statistical terms, it is equivalent to estimate the property values at
There are many software packages available for fitting the variogram mod- unsampled locations conditioned on sparse observations. Such a map or image
els (e.g., geoStat, geoR, GSLIB). In addition, recent advances allow us to (in 2D or 3D)—having a property value at each pixel—is commonly called a
estimate the variogram parameters (including an anisotropy ratio and trans- field. A random variable is defined at each pixel to represent the uncertain
formation) directly from the data without creating an empirical variogram property value, and the entire map is defined as a random vector, called
(Diggle and Ribeiro, 2007). In any case, it is recommended to inspect and random field.
compare the fitted variogram to the empirical variogram, since such fitting is The number and density of data locations are generally dependent on vari-
often inaccurate in the presence of outliers or other problems in the data. ous factors, including measurement device specificities and field accessibility.
Real-world applications often pose a challenge in variogram modeling and There are many challenges to distribute sampling points for capturing the
fitting. For example, the domain of interest might have multiple scales of heterogeneity representative of a studied region and estimating the property
heterogeneity (Neuman, 1994; Deutsch and Journel, 1998), having short- field accurately. Several geostatistics-based methodologies have been pro-
range and long-range structures or zonal anisotropy. One remedy for such posed for designing sampling schemes and facilitating spatial mapping
cases is to have a nested variogram model, which is the sum of different (Nowak et al., 2003; Diggle and Ribeiro, 2007).
20_Singh_ch20_p20.1-20.10.indd 3 8/26/16 4:37 PM

(a) (b) (c)

100 100 100 3
80 80 80 2
1
60 60 60
0
40 40 40
–1
20 20 20 –2
0 0 0 –3
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
(d) (e) (f)

100 100 100 3
80 80 80 2
1
60 60 60
0
40 40 40
–1
20 20 20
–2
0 0 0 –3
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Figure 20.2 Synthetic random field with different covariance functions and scales: (a–c) the scale parameter is 10 and (d–f) the scale parameter is 30. (a) and
(d) are based on an exponential, (b) and (e) on a Gaussian, and (c) and (f) on a spherical variogram model.
20.4.1 Kriging and Co-kriging where s∗2 is the overall variance of the field. The kriging variance represents
Kriging is a traditional interpolation technique that has forged the early the uncertainty in the kriged estimate so that the regions far from point
acceptance of geostatistics (Krige, 1952; Matheron, 1965) and the key measurements have higher kriging variance than their vicinity.
component of variogram-based estimation. While we refer to several Various kriging-based techniques have been developed to handle more
textbooks for details (Kitanidis, 1997; Rubin, 2003; Diggle and Ribeiro, 2007; complex datasets and field heterogeneity with specific characteristics. While
Remy et al., 2009), here we explain the key concept behind Kriging required the reader is referred to geostatistical textbooks (Deutsch and Journel, 1998),
for general understanding of the various other estimation approaches. this section lists several of them along with their characteristics:
In Kriging, the property value at an unsampled location is described by a - Simple kriging refers to the case of kriging with a stationary mean, variance,
linear combination of other measurement values: and variogram-covariance function over the domain of interest.
n - Ordinary kriging implies the re-estimation of the mean value m∗ from the
Y ∗ (u) − m(u) = ∑ λα (u)[z (uα ) − m(uα )] (20.3) data at each location based on kriging neighborhood, and thus corresponds
α=1 to simple kriging where the constant mean value is replaced by the location-
where Y∗(u) is the estimate at the unsampled location u, m(u) is the estimated dependent mean value.
mean, la (a = 1…n) are weights applied to the n data values, z(ua ) (a = 1…n) - Kriging with a trend model, also referred as universal kriging, assumes that
are the n data values, and m(uα) are the mean estimates at the data locations. the mean follows a prescribed trend model, such as a polynomial function.
The weights la(u) are determined to ensure unbiasedness, that is, E[Y(u) The trend is removed before performing simple kriging. Because a trend is
− Y∗(u)] = 0, and also minimum variance of the error Var(Y − Y∗). The first simulated, universal kriging is unstable if extrapolated outside the data
weights are computed by the linear equation: spatial coverage.
n - Nonlinear kriging implies that some transform has been applied on the
∑ λβ (u) C(uβ − uα ) = C(u − uα ) (20.4) sample dataset.
- Co-kriging is the extension of kriging to two or more types of variables
α=1
where ub − ua presents the distance between the data point a and b (Kitanidis and Vomvoris, 1983). It involves the empirical or theoretically
(a, b = 1,…,n). This system of n equations with n unknowns is solved to derived correlations between the different types of measurements (for exam-
estimate the weights. The kriging estimate is the linear unbiased minimum- ple, hydraulic head observations for permeability estimation).
variance estimator. The left-hand side of Eq. (20.4) is related to redundancy - Indicator kriging is a variant of kriging in which the property of interest is a
in the data, while the right-hand side is related to the closeness of the data to categorical variable (i.e., indicator) rather than a continuous variable. It is
the location being estimated. While kriging can be compared to other often used to characterize geological units or sedimentary facies.
weighted average estimators (such as inverse distance interpolation), kriging 20.4.2 Heterogeneous Field Generation
involves the use of a variogram that enforces the spatial correlation structure
derived from the datasets, which allows more complex and realistic Kriging provides the best-unbiased linear estimate at each pixel and interpo-
estimation (for example, accounting for anisotropy) than inverse distance lates between point measurements. The kriged field, however, is known to
interpolation. smooth out extreme values, such as high permeability zones (e.g., Rubin,
In addition, the kriging variance s∗2(u) can be defined at a given location by: 2003). When we simulate a hydrological process based on such a smoothed
parameter field, the results are not often realistic. Mathematically, it is known
n that kriged fields are not adequate unless the subsequent process model (e.g.,
σ *2 (u) = σ 2 − ∑ λβ (u) C(uβ − uα ). (20.5) a flow model) is perfectly linear and the interest is only on the mean and vari-
α=1 ance behavior of outcomes (Diggle and Ribeiro, 2007). Having a nonlinear
20_Singh_ch20_p20.1-20.10.indd 4 8/26/16 4:37 PM

Heterogeneous field estimation and simulation 20-5
2 2
0 0
–2 –2
(a)
2 2
0 0
–2 –2
(b)
2 2
0 0
–2 –2
(c)
1 Figure 20.4 Two equally likely randomly generated fields in the same parameter as
in Figure 20.3 the spectral method was used.
0.5
spatial range in which the data points and adjacent pixels are considered. Also,
0 many variations of the sequential Gaussian simulation are available, depend-
(d) ing on the data and stationarity characteristics (Deutsch and Journel, 1992;
Remy et al., 2009; Pyrcz and Deutsch, 2015).
Figure 20.3 Synthetic example of simple kriging interpolation: (a) true (or reference) The Cholesky decomposition method creates the covariance matrix among
field, (b) sparse measurements, (c) interpolated field, and (d) kriging variance. all the pixels and decomposes the matrix into a lower and upper triangular
matrix. By generating non-correlated Gaussian variables at all pixels, and
multiplying them with the triangular matrix, we can generate the spatially
process model, it is necessary to perform a stochastic simulation to (1) gener- correlated Gaussian field. With built-in functions in software packages, this is
ate equally possible fields given data and spatial correlation structure, (2) usually the fastest and simplest way to generate an unconditional Gaussian
simulate hydrological processes on each field, and (3) analyze the ensemble of field, although creating a large covariance matrix of size O(n2) often causes a
those outputs (e.g., Maxwell et al., 1999). We call a field unconditional when memory problem and limits the size of the field (n is the number of pixels in
it is generated given only the structural parameters (i.e., mean and covari- the field). To create a conditional field, the difference between the uncondi-
ance), as opposed to a conditional field when the data values are honored at tional field and observed values is kriged (Rubin, 2003).
the measurement locations (Rubin, 2003). The measurements can be, for The spectral method is based on the spectral decomposition of the covari-
example, permeability data derived from a core-permeameter test or slug test. ance matrix using the fast Fourier transform (FFT). This method requires
Recent research focuses on generating heterogeneous fields while involving computing the eigenvalues of circulant matrices in O(n log2n) operations
hydrological or other process models such that the conditional data corre- while only storing O(n) elements of the matrix (Dietrich and Newsam, 1996;
spond to the output of those process models as a function of the parameter Nowak et al., 2003). The conditional simulation can be done by kriging
field y (i.e., the generated heterogeneous field property); z = f (y) (e.g., Chen (Rubin, 2003) or as proposed by Dietrich and Newsam (1996). With built-in
et al., 2004, 2006). For example, solute concentrations at various times and functions of FFT in standard software, the spectral method is often much
locations, or arrival times of geophysical signals are used as data observations faster than the sequential method, and can handle larger fields than the
(z) based on which geostatistical parameters of the property field y are esti- Cholesky method. The limitation of the spectral method is that it only works
mated using inverse modeling. Such process models can be complex and for a uniform grid.
nonlinear, involving, for example, a numerical hydrological model for simu-
lating flow and transport, or a geophysical model for simulating electromag- Facies-Based Simulation
netic wave propagation. The number of observations is often smaller than the Instead of simulating property values (such as permeability), the facies-based
number of variables in the field so that the inverse problem is ill-posed and approach generates categorical (or indicator) variables. A facies is a geological
the solution has uncertainty. To alleviate this problem, the stochastic inverse unit that has a unique distribution of permeability or geochemical properties
modeling methods aim to simulate equally plausible fields y given the data z (Ritzi et al., 1995; Sassen et al., 2012). Distributing such facies is often consid-
(e.g., Chen et al., 2004, 2006). A number of methods have been proposed for ered adequate to capture the spatial heterogeneity at the scale relevant to
this purpose such as simulated annealing and Markov-Chain Monte-Carlo many hydrogeological applications (Carle, 2000). There are two widely used
methods (explained more in detail in Secs. 4.2.4 and 4.2.5). simulation approaches: (1) variogram-based sequential indicator simulation,
and (2) transition-probability approach.
Gaussian Field Simulation Sequential indicator simulation is similar to sequential Gaussian simula-
The most common approach is the Gaussian random field generation, assum- tion, except that it is based on an indicator (rather than real-valued)
ing that the field follows a multivariate Gaussian distribution with known mean variogram and two-point statistics (Deutsch and Journel, 1992). The indica-
and covariance. Gaussian random fields can be generated using (1) sequential tor at each pixel is sequentially generated from the conditional probability
Gaussian simulation, (2) Cholesky decomposition, and (3) the spectral method. given the observed values and adjacent pixels within the correlation range.
Sequential Gaussian simulation is probably the most common approach in GSLIB is often used to perform sequential indicator simulations (Deutsch and
hydrogeology, as it is implemented in the widely distributed software package Journel, 1992).
GSLIB (Deutsch and Journel, 1992). In this approach, each pixel is sampled The transition-probability approach was developed later as an alternative
sequentially in a random order. The value at each pixel is randomly generated to the variogram-based approach, which often performs poorly when there
given the conditional mean and variance (calculated using kriging) based on are more than three facies (Carle and Fogg, 1996). It defines the conditional
the correlation between this pixel and data values and previously sampled probability of an indicator at each pixel directly given the adjacent pixels and
pixels. To limit the computational expense, the user needs to determine the data. It is more flexible than the variogram-based approach to integrate
20_Singh_ch20_p20.1-20.10.indd 5 8/26/16 4:37 PM

geological knowledge and spatial patterns. The transition-probability Markov Chain Monte-Carlo
approach is implemented in the software T-PROGS (Carle, 1999). In the Bayesian method, the goal is to sample the field y from the posterior
distribution conditioned on the data z p(y|z). The Markov-chain Monte-
Multipoint Geostatistics
Carlo (MCMC) approach is known to be the most efficient way to sample a
Simulating curvilinear features is possible through data conditioning, nonsta- high-dimensional random vector (i.e., a field) conditioned on data (Andrieu
tionary statistic models, or the use of facies modeling. However, in some cases et al., 2003; Gamerman and Lopes, 2006). MCMC has been used in Bayesian
a promising alternative is the use of multipoint statistics. The multipoint estimation of hydrological properties (e.g., Michalak, 2003; Chen et al., 2004,
model of spatial continuity is generally a training image, an arbitrary or 2006; Vrugt et al., 2005, 2006; Arora et al., 2012). The advantage of MCMC is
exhaustive representation of the heterogeneity at the scale of the investigated that it represents the entire multivariate posterior distribution conditioned
field. Characterizing and reproducing curvilinear features (such as sinuous various types and scales of datasets (e.g., Wikle et al., 2001), which enables us
fluvial channels) require many point pairs and thus data. The question arises, to sample random fields for uncertainty quantification studies.
however, if a (somewhat subjective) explicit image-based model is more appro- In the estimation, the posterior distribution is first inverted using Bayes’
priate than a (poorly informing) implicit two-point statistic variogram model. theorem, as p(y|z) µ p(z|y)p(y), where the likelihood p(z|y) is computed from
After the construction of a training image and the inference of probability the process model z = f(x) , and the prior p(y) is defined based on the known
density functions at each location (analogous to inferring a variogram information of the field. Although there have been many sampling schemes
model), the sequential simulation will for each randomly selected location in proposed within MCMC, the most common ones are (1) Gibbs sampling for
the simulated field scan the relevant training image for a facies configuration the parameters with analytical distributions (e.g., normal and multivariate
that replicates the data (and already simulated locations). A Monte Carlo normal), and (2) Metropolis–Hasting (MH) sampling for other distribution.
approach is used to generate a random value from the probability function Gibbs sampling is much more efficient but only possible when the process
and to choose the corresponding facies. The sequential simulation is per- model f is approximately linear or linearized. The MH sampling generates the
formed by visiting all locations in a random order. Various methods are being proposal values of the field (or each pixel) using unconditional/conditional
developed to improve run time and pattern reproduction, including, for simulation, and accepts/rejects depending on the acceptance probability
example, the use of a direct sampling approach that uses multivariate training defined by the posterior distribution.
images and enables additional complexity and dimensionality in the training
image objects (Mariethoz et al., 2010). The reader is referred to textbooks and Estimating Geostatistical Parameters Using
review papers for more details (Hu and Chugunova, 2008; Mariethoz and a Multiphysics Joint Inversion Approach
Caers, 2014). The previous sections describe geostatistical approaches in which a spatially
In hydrology, multipoint geostatistics has been used to investigate the effect correlated attribute field is first characterized based on available data of the
of complex small-scale sedimentary heterogeneity on the short-term migra- same (or related) attribute, and then these attributes are estimated (through
tion of a contaminant plume in a cross-bedded aquifer (Huysmans and interpolation or geostatistical simulation) at unsampled locations. Once a
Dassargues, 2009). Multipoint geostatistics has also been combined with heterogeneous property field is generated, it is used in a process model to
geologic-process models for conditional simulation of 3-D heterogeneity calculate a certain system response (e.g., contaminant migration times) either
(Michael et al., 2010) and to produce maps of geobody geometry and within a deterministic or stochastic framework. If geostatistics is used in the
thickness. Feyen and Caers (2006) used multipoint statistics to quantify the context of predictive process modeling, it is important to capture the relevant
uncertainty of groundwater flow and transport model predictions for various aspects of heterogeneity, that is, those that are most influential on the system
levels of hydrogeological uncertainty in hydrofacies structure and the intrafa- response of interest. For example, depending on the objective of a numerical
cies variability of the hydraulic properties. simulation, it may be more important to capture the heterogeneity or z of
conducting (e.g., high-permeability) rather than resisting (e.g., low-
Simulated Annealing permeability) features. An indiscriminate application of geostatistical concepts
Simulated annealing (SA) iteratively perturbs a target parameter field may thus lead to a description of spatial heterogeneity that is suboptimal for the
(random field or simulation) until it matches (to a level considered acceptable) intended purpose.
the data (involving a process model f or not), structural constraints (e.g., a One way to overcome this pitfall is to integrate the process simulation into
variogram model) or any type of constraint that can be expressed in an objec- the geostatistical estimation. Rather than estimating, for example, the nugget
tive function. In practice, all perturbations in the generated target parameter effect, correlation length, and sill value of a variogram model from measured
field that lower the global objective function are unconditionally accepted, permeability data, these geostatistical parameters may be determined by
whereas those that do not are rejected or occasionally accepted according to matching simulated to observed tracer breakthrough data through the use of
a Boltzmann-type probability distribution controlled by a temperature inverse modeling. In this way, the geostatistical field is tailored to the flow and
parameter T. The higher the value of T, the more likely an unfavorable per- transport processes of interest. Note that measured permeability data can
turbation will be accepted. Throughout the SA process, T is gradually lowered readily be integrated into this framework.
such that the algorithm has a chance to reach a global, optimal “energy state.” Estimating geostatistical parameters based on hydrological data alone often
Once the objective function is deemed small enough, the SA procedure is yields an ill-posed inverse problem. It is thus essential to include as many
terminated. The procedure is run multiple times to assess uncertainty. complementary data as possible into the analysis. Geophysical observations
Advantages of using the SA approach for data integration are that it is not (e.g., electromagnetic and acoustic data) are often collected in survey con-
limited to simple Gaussian statistics, and it is able to incorporate any figurations that yield sufficient spatial coverage suitable to infer heteroge-
constraint on the output realizations that can be expressed in the form of a neous properties. In this framework, the geophysical response to the
multicomponent objective function. This comes with the caveat that compu- subsurface structure and changes in the system state are matched jointly with
tational efficiency and convergence decreases with constraint complexity, and hydrogeological observations, such as pressure, flow rate, and concentration
that the choice of the weights attributed to the various components can be measurements. Such a joint hydrogeophysical inversion is a promising
subjective. For more detailed information on the use of SA for subsurface data approach to obtain spatially heterogeneous property fields that are relevant to
integration, we refer the reader to geostatistics textbooks and papers (Deutsch flow and transport problems (Kowalsky et al., 2004, 2005; Lehikoinen et al.,
and Journel, 1992; Deutsch and Cockerham, 1994; Deutsch and Wen, 2000; 2010; Doetsch et al., 2013).
Parks et al., 2000; Kelkar and Perez, 2002). To concurrently estimate geostatistical parameters along with hydrogeo-
Simulated annealing has been applied for a large range of hydrological logical properties and petrophysical coefficients, the pilot-point method
problems. For example, Deutsch and Wen (2000) used this approach to recon- [originally proposed by d e Marsily (1978)] is often employed. In this method,
struct high-resolution images from sparse data. They showed that SA was attribute values are estimated at select locations, which are used as conditioning
reproducing small-scale features particularly well while matching the vario- points for geostatistical interpolation or simulation. By changing a pilot-point
grams at small lags. Day-Lewis et al. (2000) used this approach to generate value, the attribute field in the vicinity of the point changes, leading to a
three-dimensional realizations of fracture networks, conditioned on borehole change in the simulated system response that can be used to calibrate the
data, inferred hydraulic connections between packer-isolated borehole inter- model. The geostatistical constraint thus acts as a model-reduction method
vals, and an indicator variogram model of spatial variability that yield maps that enables the characterization of site-specific heterogeneity without the
of fracture geometry and the probability distribution of specific fracture need to estimate parameters for each pixel of the model.
zones. Dafflon and Barrash (2012) used SA to integrate ground-penetrating The method is demonstrated here using a synthetic field experiment in
radar tomography data, borehole porosity log data, and spatial variability which water is released from a pond into a highly heterogeneous soil exhibit-
constraints to generate high-resolution simulations of hydraulic parameter ing a spatially correlated, anisotropic, inclined permeability structure
fields needed to perform realistic flow and transport simulations. (see Fig. 20.5a). Two monitoring boreholes are equipped with neutron probes
20_Singh_ch20_p20.1-20.10.indd 6 8/26/16 4:37 PM

REFERENCES 20-7
log (k) Sliq.

X X X X
–11.50 0.95
–0.5 X X –11.75 –0.5 X X 0.90
–12.00 0.85
–12.25 0.80
X X –12.50
X X 0.75
–1.0 –12.75 –1.0 0.70
X X –13.00 X X 0.65
Depth [m]
Depth [m]
–13.25 0.60
X X –13.50 X X 0.55
–1.5 –13.75 –1.5 0.50
X X –14.00 X X 0.45
–14.25 0.40
–2.0 X X –14.50 –2.0 X X 0.35
0.30
0.25
0.20
–2.5 –2.5 0.15
1 2 3 1 2 3
X [m] X [m]
(a) (b)
log (k) σlog(k)

X X X X
–11.50 1.3
–0.5 X X –11.75 –0.5 X X 1.2
–12.00 1.1
–12.25 1.0
X X –12.50
X X 0.9
–1.0 –12.75 –1.0 0.8
X X –13.00 X X 0.7
Depth [m]
–13.25 Depth [m] 0.6

X X –13.50 X X 0.5
–1.5 –13.75 –1.5 0.4
X X –14.00 X X 0.3
–14.25
–2.0 X X –14.50 –2.0 X X
–2.5 –2.5
1 2 3 1 2 3
X [m] X [m]
(c) (d)
Figure 20.5 (a) Spatially correlated, anisotropic, inclined permeability structure used for simulating synthetic infiltration experiment. Also shown are the locations of GPR anten-
nas (squares: transmitting; circles: receiving) and position of pilot points (crosses). (b) Liquid saturation distribution after one day of water release, locations of neutron probes in
boreholes (squares), and GPR straight-ray paths used for inversion. (c) Median permeability field estimated by joint hydrogeophysical inversion using 20 realizations. (d) Standard
deviation of estimated permeabilities.
for water content measurements; the same boreholes are used for ground techniques are valuable in analyzing complex systems with multiple interacting
penetrating radar (GPR) surveys (see Fig. 20.5b). The joint inversion is for- components that render uncertainty in system response. Because hydrogeo-
mulated as a nonlinear weighted least-squares problem in which the infiltra- logical phenomena are remarkably complex, and the description of the true
tion rates, water contents, and GPR arrival times are matched by iteratively system (e.g., aquifer, reservoir) is unknown, geostatistical techniques provide
updating the parameters of a spherical variogram model, permeability modi- a stochastic framework to address the uncertainty. Also, the availability of
fiers at 14 pilot points, porosity, parameters of the water retention and relative commercial software [e.g., MATLAB, ArcGIS (Geographic Information
permeability curves, as well as petrophysical parameters used to calculate the System)] and open source software (e.g., Quantum GIS, R, and Python librar-
dielectric constant of the partially saturated soil. To account for unidentifiable ies for spatial analysis) has facilitated the use of spatial analysis and geostatis-
small-scale heterogeneity, multiple joint inversions were performed with tics in multidisciplinary fields. Next-generation research challenges and
different seed numbers. The median of the resulting permeability fields is promising directions for advances in spatial analysis and geostatistics could be
shown in Fig. 20.5c. The salient features of the soil structure are well identi- in socio-hydrologic fields that can help us understand and address growing
fied between the two boreholes, whereas the averaging process results in an food–water–energy demands.
unstructured effective permeability in the regions where no geophysical data
are available. This is also reflected by the standard deviation shown in ACKNOWLEDGMENTS
Fig. 20.5d. More details can be found in Finsterle and Kowalsky, (2008).
The approach presented in this section provides the means to characterize We acknowledge partial support from the Subsurface Scientific Focus Area at
select aspects of the soil structure that are relevant to the system behavior at a Lawrence Berkeley National Laboratory funded by the U.S. Department of
given site. Geostatistics is used in this approach as a model-reduction Energy, Office of Science, Office of Biological and Environmental Research
technique within a multiphysics joint inversion framework. under Award Number DE-AC02-05CH11231.
20.5 SUMMARY REFERENCES
Since the inception and use of geostatistical methods for the appraisal of ore Andrieu, C., N. De Freitas, A. Doucet, and M. I. Jordan, “An introduction
reserves in mining by pioneers like Georges Matheron and Danie G. Krige, to MCMC for machine learning,” Machine Learning, 50 (1–2): 5–43, 2003,
spatial analysis and geostatistical methods have been used in several other doi:10.1023/A:1020281327116.
fields, such as hydrogeology, hydrology, geochemistry, geography, forestry, Anselin, L., “Interactive techniques and exploratory spatial data analysis,”
soil science, agriculture, and social sciences. Spatial analysis and geostatistics Geography Information System Prince Technology Management Application,
are a set of numerical techniques to characterize spatial attributes. These 253–266, 1999.
20_Singh_ch20_p20.1-20.10.indd 7 8/26/16 4:37 PM

Arora, B., B. P. Mohanty, and J. T. McGuire, “Uncertainty in dual Ganio, L. M., C. E. Torgersen, and R. E. Gresswell, “A geostatistical
permeability model parameters for structured soils,” Water Resources approach for describing spatial pattern in stream networks,” Frontiers in
Research, 48: 2012, doi:10.1029/2011WR010500. Ecology and the Environment, Vol. 3, 2005, pp. 138–144.
Arora, B., B. P. Mohanty, J. T. McGuire, and I. M. Cozzarelli, “Temporal Getis, A., “A history of the concept of spatial autocorrelation: a geographer’s
dynamics of biogeochemical processes at the Norman Landfill site,” Water perspective, Geographical Analysis, 40 (3), 297–309, 2008, doi:10.1111/
Resources Research, 49: 6909–6926, 2013, doi:10.1002/wrcr.20484. j.1538-4632.2008.00727.x.
Ashley, R. P. and J. W. Lloyd, An example of the use of factor analysis and Griffith, D. A. and P. R. Peres-Neto, “Spatial modeling in ecology: the flex-
cluster analysis in groundwater chemistry interpretation, Journal of Hydrology, ibility of eigenfunction spatial analyses,” Ecology, 87 (10): 2603–2613, 2006,
39 (3–4), 355–364, 1978, doi:10.1016/0022-1694(78)90011-2. doi:10.1890/0012-9658(2006)87[2603:SMIETF]2.0.CO;2.
Berkhin, P., “A survey of clustering data mining techniques,” Grouping Han, J. and M. Kamber, Data Mining: Concepts and Techniques, Elsevier,
Multidimensional Data, Springer Berlin Heidelberg, Netherlands, pp. 25–71, 2006. Waltham, 2006.
Carle, S. F., T-PROGS: Transition probability geostatistical software, Han, J., M. Kamber, and A. K. H. Tung, “Spatial clustering methods in data
University of California, Davis, CA, 1999. mining: A survey,” Geographic Data Mining and Knowledge Discovery, Taylor
Carle, S. F. and G. E. “Fogg, transition probability-based indicator geostatis- and Francis, London and New York, pp. 30–50, 2001.
tics,” Mathematical Geology, 28 (4): 453–476, 1996, doi:10.1007/BF02083656. Helena, B., R. Pardo, M. Vega, E. Barrado, J. M. Fernandez, and
Chen, J., S. Hubbard, Y. Rubin, C. Murray, E. Roden, and E. Majer, L. Fernandez, “Temporal evolution of groundwater composition in an alluvial
“Geochemical characterization using geophysical data and Markov Chain aquifer (Pisuerga River, Spain) by principal component analysis,” Water
Monte Carlo methods: a case study at the South Oyster bacterial transport site Research, 34 (3): 807–816, 2000, doi:10.1016/S0043-1354(99)00225-0.
in Virginia,” Water Resources Research, 40 (12), 2004, doi:10.1029/2003WR002883. Hu, L. Y. and T. Chugunova, “Multiple-point geostatistics for modeling
Chen, J., S. Hubbard, J. Peterson, K. Williams, M. Fienen, P. Jardine, and subsurface heterogeneity: a comprehensive review,” Water Resources Research,
D. Watson, “Development of a joint hydrogeophysical inversion approach and 44 (11): 1–14, 2008, doi:10.1029/2008WR006993.
application to a contaminated fractured aquifer,” Water Resources Research, Huysmans, M. and A. Dassargues, “Application of multiple-point geostatis-
42: 1–13, 2006, doi:10.1029/2005WR004694. tics on modelling groundwater flow and transport in a cross-bedded aquifer
Cressie, N., Statistics for Spatial Data, John Wiley & Sons, New York, 2015. (Belgium),” Hydrogeology Journal, 17 (8): 1901–1911, 2009, doi:10.1007
Dafflon, B. and W. Barrash, “Three-dimensional stochastic estimation of /s10040-009-0495-2.
porosity distribution: benefits of using ground-penetrating radar velocity Jolliffe, I. T., Principal Component Analysis, John Wiley & Sons, Ltd.,
tomograms in simulated-annealing-based or Bayesian sequential simulation New York, 2002.
approaches,” Water Resources Research, 48, 2012, doi:W0555310.1029/ Kelkar, M. and G. Perez, Applied Geostatistics for Reservoir Characterization,
2011wr010916. Society of Petroleum engineers, Richardson, TX, 2002.
Dale, M. R. T. and M-J. Fortin, “From graphs to spatial graphs,” Annual Kitanidis, P. K., Introduction to Geostatistics: Applications in Hydrogeology,
Review of Ecology, Evolution, and Systematics, 41 (1): 21–38, doi:10.1146 Cambridge University Press, Cambridge, UK, 1997.
/annurev-ecolsys-102209-144718, 2010. Kitanidis, P. K. and E. G. Vomvoris, “A geostatistical approach to the
Day-Lewis, F. D., P. A. Hsieh, and S. M. Gorelick, “Identifying fracture-zone inverse problem in groundwater modeling (steady state) and one-dimensional
geometry using simulated annealing and hydraulic-connection data,” Water simulations,” Water Resources Research, 19 (3): 677, 1983, doi:10.1029/
Resources Research, 36 (7): 1707–1721, 2000. WR019i003p00677.
de Marsily, G., De l’identification des systemes hydrogeologiques, PhD Kowalsky, M. B., S. Finsterle, and Y. Rubin, “Estimating flow parameter
thesis, Paris VI, France, 1978. distributions using ground-penetrating radar and hydrological measure-
Deutsch, C. V. and P. W. Cockerham, “Practical considerations in the appli- ments during transient flow in the vadose zone,” Advances in Water Resources,
cation of simulated annealing to stochastic simulation,” Mathematical 27: 583–599, 2004, doi:10.1016/j.advwatres.2004.03.003.
Geology, 26 (1): 67–82, 1994. Kowalsky, M. B., S. Finsterle, J. Peterson, S. Hubbard, Y. Rubin, E. Majer,
Deutsch, C. V. and A. G. Journel, GSLIB: Geostatistical Software Library and A. Ward, and G. Gee, “Estimation of field-scale soil hydraulic and dielectric
User’s Guide, Oxford University Press, New York, p. 340, 1992. parameters through joint inversion of GPR and hydrological data,” Water
Deutsch, C. V. and X. H. Wen, “Integrating large-scale soft data by Resources Research, 41 (11): 1–19, 2005, doi:10.1029/2005WR004237.
simulated annealing and probability constraints,” Mathematical Geology, 32 Krige, D. G., “A statistical approach to some basic mine valuation problems
(1): 49–67, 2000. on the witwatersrand,” Journal of Chemical, Metallurgical, and Mining Society
Dietrich, C. R. and G. N. Newsam, “A fast and exact method for multidi- of South Africa, 52(6), 201–215, 1952, doi:10.2307/3006914.
mensional Gaussian stochastic simulations: extension to realizations Kumar, P. and E. Foufoula-Georgiou, “Wavelet analysis for geophysical
conditioned on direct and indirect measurements,” Water Resources Research, applications,” Reviews of Geophysics, 35 (97): 385, 1997, doi:10.1029/97RG00427.
32 (6): 1643–1652, 1996, doi:10.1029/93WR01070. Lau, K.-M. and H. Weng, “Climate signal detection using wavelet transform:
Diggle, P. and P. J. Ribeiro, Model-Based Geostatistics, Springer Science & how to make a time series sing,” Bulletin of the American Meteorological
Business Media, New York, 2007. Society, 76 (12): 2391–2402, 1995, doi:10.1175/1520-0477(1995)076<2391:
Doetsch, J., M. B. Kowalsky, C. Doughty, S. Finsterle, J. B. Ajo-Franklin, CSDUWT>2.0.CO;2.
C. R. Carrigan, X. Yang, S. D. Hovorka, et al., “Constraining CO2 simulations Legendre, P., “Spatial autocorrelation: trouble or new paradigm?,” Ecology,
by coupled modeling and inversion of electrical resistance and gas composi- 74 (6): 1659–1673, 1993, doi:10.2307/1939924.
tion data,” International Journal of Greenhouse Gas Control, 18: 510–522, Lehikoinen, A., J. M. J. Huttunen, S. Finsterle, M. B. Kowalsky, and J. P.
2013, doi:10.1016/j.ijggc.2013.04.011. Kaipio, “Dynamic inversion for hydrological process monitoring with electri-
Dwivedi, D., B. P. Mohanty, and B. J. Lesikar, “Estimating Escherichia coli cal resistance tomography under model uncertainties,” Water Resources
loads in streams based on various physical, chemical, and biological factors,” Research, 46: 1–14, 2010, doi:10.1029/2009WR008470.
Water Resources Research, 49: 2896–2906, 2013, doi:10.1002/wrcr.20265. Mariethoz, G. and J. Caers, Multiple-Point Geostatistics: Stochastic Modeling
Ester, M., A. Frommelt, H-P. Kriegel, and J. Sander, “Spatial data mining: with Training Images, Willey-Blackwell, West Sussex, UK, 2014.
database primitives, algorithms and efficient DBMS support,” Data Mining Mariethoz, G., P. Renard, and J. Straubhaar, “The direct sampling method
and Knowledge Discovery, 4 (2–3): 193–216, 2000, doi:10.1023/ to perform multiple-point geostatistical simulations,” Water Resources
A:1009843930701. Research, 46 (11), 1–14, 2010, doi:10.1029/2008WR007621.
Feyen, L. and J. Caers, Quantifying geological uncertainty for flow and Martínez, B., M. A. Gilabert, F. J. García-Haro, A. Faye, and J. Meliá,
transport modeling in multimodal heterogeneous formations, Advances in “Characterizing land condition variability in Ferlo, Senegal (2001-2009) using
Water Resources, 29 (6): 912–929, 2006, doi:10.1016/j.advwatres. multi-temporal 1-km Apparent Green Cover (AGC) SPOT Vegetation data,”
2005.08.002. Global and Planetary Change, 76 (3–4): 152–165, 2011, doi:10.1016/j.
Finsterle, S. and M. B. Kowalsky, “Joint hydrological-geophysical inversion gloplacha.2011.01.001.
for soil structure identification,” Vadose Zone Journal, 7 (1): 287–293, 2008, Matheron, G., Les variables Régionalisées et leur Estimation, Masson, Paris,
doi:10.2136/vzj2006.0078. 1965.
Fisher, R. A., An Agricultural Experiment in Randomised Blocks. The Design Maxwell, R. M., W. E. Kastenberg, and Y. Rubin, “A methodology to inte-
of Experiments, Oliver and Boyd, Edinburgh, U.K., 1937. grate site characterization information into groundwater-driven health risk
Gamerman, D. and H. F. Lopes, Markov Chain Monte Carlo: Stochastic assessment,” Water Resources Research, 35 (9): 2841, 1999, doi:10.1029/
Simulation for Bayesian Inference, 2nd ed., CRC Press, New York, 2006. 1999WR900103.
20_Singh_ch20_p20.1-20.10.indd 8 8/26/16 4:37 PM

REFERENCES 20-9
Michael, H. A., H. Li, A. Boucher, T. Sun, J. Caers, and S. M. Gorelick, Rubin, Y., Applied Stochastic Hydrogeology, Oxford University Press,
“Combining geologic-process models and geostatistics for conditional simu- New York, 2003.
lation of 3-D subsurface heterogeneity,” Water Resources Research, 46 (5), Sassen, D. S., S. S. Hubbard, S. A. Bea, J. Chen, N. Spycher, and M. E.
2010, doi:10.1029/2009WR008414. Denham, “Reactive facies: an approach for parameterizing field-scale reactive
Michalak, A. M., “A method for enforcing parameter nonnegativity in transport models using geophysical methods,” Water Resources Research, 48:
Bayesian inverse problems with an application to contaminant source identi- 1–20, 2012, doi:10.1029/2011WR011047.
fication,” Water Resources Research, 39 (2): 1–14, 2003, doi:10.1029/ Shekhar, S. and S. Chawla, “Introduction to spatial data mining,” Spatial
2002WR001480. Databases: A Tour, Prentice Hall, New Jersey, 2003b, pp. 237–284.
Monahan, A. H., J. C. Fyfe, M. H. P. Ambaum, D. B. Stephenson, and G. R. Shekhar, S. and S. Chawla, Spatial Databases: A Tour, Prentice Hall,
North, “Empirical Orthogonal Functions: The Medium is the Message,” New Jersey, 2003c.
Journal of Climate, 22 (24), pp. 6501–6514, 2010, ISSN 1520-0442. Shekhar, S., Y. Huang, W. Wu, C. T. Lu, and S. Chawla, “Data mining for
Neuman, B. C., Scale in distributed systems, Readings in Distributed scientific and engineering applications,” Massive Computing, edited by R. L.
Computing Systems, Computer Society Press, 1994, pp. 463–489. Grossman, C. Kamath, P. Kegelmeyer, V. Kumar, and R. R. Namburu,
Nowak, W., S. Tenkleve, and O. A. Cirpka, “Efficient computation of linear- Springer US, Boston, MA, 2001.
ized cross-covariance and auto-covariance matrices of interdependent Shekhar, S., P. Zhang, and Y. Huang, “Trends in spatial data mining,”
quantities,” Mathematical Geology, 35 (1): 53–66, 2003, doi:10.1023/ Science, 7: 357–379, 2003, doi:10.1007/0-387-25465-X_39.
A:1022365112368. Shekhar, S., P. Zhang, and Y. Huang, “Spatial data mining,” Data Mining
Openshaw, S. and P. Taylor, A million or so correlation coefficients: three and Knowledge Discovery Handbook, edited by O. Maimon and L. Rokach,
experiments on the modifiable areal unit problem. Statistical Applications in Springer, New York, 2010.
the Spatial Sciences, edited by Wrigley N, Pion, London, UK, 1979. Tobler, W., “Cellular geography,” Philosophy in Geography, edited by S. Gale
Parks, K. P., L. R. Bentley, and A. S. Crowe, Capturing geological realism in and G. Olsson, Reidel, Dordrecht, The Netherlands, 1979.
stochastic simulations of rock systems with Markov statistics and simulated Vega, M., R. Pardo, E. Barrado, and L. Debán, “Assessment of seasonal and
annealing, Journal of Sedimentary Research, 70 (4): 803–813, 2000. polluting effects on the quality of river water by exploratory data analysis,”
Pebesma, E. and B. Gräler, Introduction to Spatio-Temporal Variography, Water Research, 32 (12): 3581–3592, 1998, doi:10.1016/S0043-1354
CRAN, University of Munster, 2015. (98)00138-9.
Pfeiffer, D., Issues Related to Handling of Spatial Data, Massey University, Vrugt, J. A., C. G. H. Diks, H. V. Gupta, W. Bouten, and J. M. Verstraten,
Palmerston North, June 1996, pp. 23–28. “Improved treatment of uncertainty in hydrologic modeling: Combining the
Portnov, B. A., J. Dubnov, and M. Barchana, “On ecological fallacy, assessment strengths of global optimization and data assimilation,” Water Resources
errors stemming from misguided variable selection, and the effect of aggregation Research, 41: 1–17, 2005, doi:10.1029/2004WR003059.
on the outcome of epidemiological study,” Journal of Exposure Science and Vrugt, J. A., H. V. Gupta, S. C. Dekker, S. Sorooshian, T. Wagener, and
Environmental Epidemiology, 17 (1): 106–21, 2007, doi:10.1038/sj.jes.7500533. W. Bouten, “Application of stochastic parameter optimization to the
Pyrcz, M. J. and C. V. Deutsch, “Geostatistical reservoir modeling,” Sacramento Soil Moisture Accounting model,” Journal of Hydrology, 325
Mathematical Geosciences, 47 (4): 497–499, 2015, doi:10.1007/s11004- (1–4), 288–307, 2006, doi:10.1016/j.jhydrol.2005.10.041.
015-9588-8. Wikle, C. K., R. F. Milliff, D. Nychka, and L. M. Berliner, “Spatiotemporal
Remy, N., A. Boucher, and J. Wu, Applied Geostatistics with SGeMS: A User’s hierarchical Bayesian modeling: tropical ocean surface winds,” Journal of
Guide, Cambridge University Press, Cambridge, UK, 2009. American Statistical Association, 96 (454): 382–397, 2001, doi:10.1198/
Ritzi, R. W., D. F. Dominic, N. R. Brown, K. W. Kausch, P. J. McAlenney, 016214501753168109.
and M. J. Basial, “Hydrofacies distribution and correlation in the Miami Wong, D. W. S., “The modifiable areal unit problem,” Quantitative
Valley Aquifer System,” Water Resources Research, 31 (12): 3271–3281, 1995, Geography: A British View, 38 (1934): 60–70, 1981, doi:10.2307/2277827.
doi:10.1029/95WR02564.
20_Singh_ch20_p20.1-20.10.indd 9 8/26/16 4:37 PM

Chapter 21
Frequency Distributions
BY
VIJAY P. SINGH AND LAN ZHANG
ABSTRACT that takes on values xi, i = 1, 2,…, N, with probabilities pi, i = 1, 2,…, N, where
N is the number of values.
The use of frequency distributions is now ubiquitous in hydrology. There is a
multitude of distributions that are used, depending on the type of data or Table 21.1 Commonly Applied Discrete Probability Distributions
variable. Most of these distributions are asymmetrical and can be classified as in Hydrology
discrete or continuous. The continuous distributions belong to one or the Distribution p.m.f. Parameters
other type of distribution families, for example, Pearson family. Some distri-
butions are general and encompass several distributions as special cases, and exp(−axi )
some distributions can be obtained from other distributions by using appro- Maxwell–Boltzmann pi = (21.1) a>0
∑ i =1 exp(−axi )
N
priate transformations. This chapter briefly discusses those frequency

distributions that are used in hydrology and their properties.  N  k N −k
Binomial P( K = k ) =   p (1 − p) (21.2) N , p
 k 
21.1 INTRODUCTION
Geometric pm = p(1 − p)m−1 (21.3) p
The use of frequency distributions can be traced to the design of hydraulic
structures, such as dams, spillways, levees, culverts, urban drainage, and
 N −1 
arterial runways requiring flood frequency analysis that started around 1914 Negative Binomial P(n) =  n −1 k
 (1 − p) p (21.4) k > 1, p
in the United States (Fuller, 1914; Hazen, 1914). Many distributions have since  k −1 
been used all over the world and no one distribution is universally accepted.
For example, Water Resources Council (1967) recommended the use of the an exp (−a)
Poisson Pn = (21.5) a > 0;
log-Pearson type III distribution that has now become a standard in the n!
United States. In China, the Pearson type III distribution is a standard, n = 0,1,2,…
whereas in Europe the generalized extreme value distribution is the preferred
distribution. Besides flood frequency analysis, frequency distributions are now 21.2.1 Maxwell–Boltzmann (M–B) Distribution
commonly employed in hydrometeorology, river engineering, and all branches
of hydrology, including surface water, groundwater, subsurface water, water Expressed in Eq. (21.1), parameter a can be determined from the knowledge
quality, and snow, ice, and glacier. Indeed there is hardly any area in hydrology of the mean of X. Fiorentino et al. (1993) employed this distribution to repre-
that does not employ a frequency distribution of one kind or another. sent the probability distribution of elevations of links in a river basin. The
Barring a few exceptions, most frequency distributions in hydrology are M–B distribution may be considered as one of the discrete equivalents of the
asymmetrical, which arise for two reasons. In the first case, the data may exponential distribution.
come from a mixture of two or more homogeneous populations and may be 21.2.2 Binomial Distribution
heterogeneous. An example is floods caused by a mixture of snowmelt and
rainfall. In this case, the flood data will be heterogeneous. In the second case, The binomial distribution can be expressed as Eq. (21.2), where parameters
data are homogeneous but the tendency of deviation to either side of the N, k, and p represent the total number of trials (Bernoulli trials taking on
mean is not equal. This is the case most frequently dealt with in hydrology. either success or failure), the total number of successes, and the probability of
When doing frequency analysis, sometimes we deal with two or three vari- success for any trial.
ables that are not independent of each other. Then, bivariate or trivariate This distribution has been used to determine the number of floods
distributions are needed. However, this chapter discusses only univariate occurring in a given number of years (USACE, 1993).
frequency distributions that have found significant application in hydrology. 21.2.3 Geometric Distribution
Univariate distributions can be classified into two groups: (1) discrete and (2)
continuous. Singh (2013) discussed many of these distributions. Most of the This distribution describes the number of trials before the next success occurs
distributions used in hydrology are continuous, although the discrete and can be given as Eq. (21.3), where parameter p defines the probability of
distributions have also been applied. The parameter estimation for the distri- success on any trial, and parameter pm denotes the probability of success on the
butions is outside the scope of this chapter, which will be discussed in the mth trial with the failure occurring for the previous m − 1 trials. This distribu-
chapter that follows. In the following sections, the discrete distributions will tion has been used to determine the number of days before the occurrence of
be discussed first. the next rain. The geometric distribution is a discrete analogue of the expo-
nential distribution.
21.2 DISCRETE FREQUENCY DISTRIBUTIONS 21.2.4 Negative Binomial Distribution
There are many discrete distributions but only those frequently used in hydrol- The negative binomial distribution describes the number of trials N needed to
ogy are discussed here, as listed in Table 21.1. Let there be a random variable X obtain exactly k successes and can be given as Eq. (21.4). Here the number of
21-1
21_Singh_ch21_p21.1-21.12.indd 1 8/22/16 12:22 PM

21-2 Frequency Distributions
successes is fixed and the number of trials varies. Equation 21.4 suggests that 21.3.2 Burr Family
k - 1 successes in the preceding n - 1 trails have occurred already, and describes Burr (1942) suggested a differential equation, given as Eq. (21.7), for describ-
the probability of k - 1 successes in n - 1 trials. This distribution is a reversal ing the CDF where x is in the range for which the solution is sought. It may
of the binomial distribution and has been used to determine the probability of be noted that Eq. (21.7) is similar to Eq. (21.6), if Eq. (21.7) is recast with PDF
rain occurring on the nth day if a given number of rainy days have occurred as: dy /dx = y (m − y ) g ( x , y ); y = f ( x ) such that the recasted expression is the
during n - 1 days. Pearson system of distribution.
Depending on the expression of g(x, y), Burr (1942) provided 20 different
21.2.5 Poisson Distribution
types of distributions. Tadikamalla (1980) showed that the Lomax, compound
The Poisson distribution, given by Eq. (21.5), describes the probability of the Weibull, Weibull, logistic, and log-logistic distributions are special cases of
number of successes in a certain length of time or space where the probability the Burr type XII distribution. He also showed that two-parameter kappa
of success per unit of time or space is defined. Its parameter a is equal to the family of distributions, used by Mielke (1973) for analyzing precipitation data,
average number of successes over the length of time or space. This distribu- and three-parameter kappa distributions, used by Mielke and Johnson (1974),
tion is a discrete equivalent of the gamma distribution. It is frequently used in are special cases of the Burr type III distribution.
rainfall modeling. Dooge (1972) used it in the development of a generalized
unit hydrograph theory. 21.3.3 Halphen Family
Halphen (1941) proposed a family of frequency distributions, now called
Halphen distributions (Perreault et al., 1999), that offer sufficient flexibility to
21.3 CLASSIFICATION OF CONTINUOUS FREQUENCY
DISTRIBUTIONS
fit a variety of hydrologic data. There are two restrictions imposed on the
distributions: (i) the lower bound is zero, and (ii) the distribution parameters
Continuous frequency distributions can be grouped into two classes: (1) need to be estimated using sufficient statistics. The Halphen family of distribu-
exponential and (2) power. Examples of the exponential class are negative tions includes harmonic, type A, and type B distributions. Dvorak et al. (1988)
exponential, normal, lognormal, extreme value, gamma, Pearson type III, showed that these three distributions can be obtained by solving the differen-
Weibull, logistic, and beta. Pareto, Wakeby, Burr, and kappa distributions are tial equation given as Eq. (21.8). The harmonic distribution is also known as
examples of the power class. These distributions can also be classified based the hyperbola distribution, and the Halphen type A distribution is recognized
on the number of parameters they have, such as one-parameter (e.g., as a re-parameterized form of the generalized inverse Gaussian distribution.
exponential), two-parameter (e.g., gamma), three-parameter (e.g., Pearson),
four-parameter (e.g., kappa), and five-parameter (e.g., Wakeby) distributions. 21.3.4 Johnson Family
Most of the continuous distributions belong to one or the other distribution Johnson (1949) proposed three systems of frequency distributions, called SL,
family, such as Pearson (1895), Edgeworth (1898), Charlier (1905), Halphen SB, and SU that, together with the normal distribution, encompass a variety of
(1941, 1955), Johnson (1949), Burr (1942), and asymptotic extreme value shapes encountered in practice. With the aid of a simple transformation, these
(e.g., Gumbel, 1958; Fisher-Tippett, 1928) families. Four of the more popular distributions can be transformed to a normal distribution, which is a very
families are discussed here with the general equations listed in Table 21.2. desirable feature. Let X be a random variable transformed to normality by
z = f (x), where z is a unit normal variable. Johnson (1942) introduced
21.3.1 Pearson Family Eq. (21.9) in which a, b are the shape parameters, c is the location parameter,
Pearson (1895) developed a family of distributions that result from a solution d is the scale parameter, and f is a nondecreasing function of x. Depending on
of the differential equation given as Eq. (21.6). In this equation, f ( x ) = dF ( x )/dx the choice of f, an infinite system of frequency distributions can be generated.
is probability density function (PDF) and F(x) is the cumulative probability One can also use y = (x − c)/d and express z in terms of y, since the distribu-
distribution function (CDF) of random variable X; and a, c0, c1, and c2 are the tion of x will be the same as that of y, because x = c + dy. Systems SB and SL
parameters. By solving Eq. (21.6), Pearson (1985) derived five types of fre- were generated by using the log-transformation given as Eq. (21.10) in which
quency distributions, depending on skewness and range: (1) type I: limited z is a unit normal variable. SB is for bounded systems and SL for unbounded
range in both directions and skewness; (2) type II: limited range and symme- systems. Likewise, system SU was constructed using the transformation given
try; (3) type III: limited range in one direction only and skewness; (4) type IV: as Eq. (21.11). The Johnson system is kind of all encompassing and can
limited range in both directions and skewness; and (5) type V: unlimited range generate virtually any distribution.
in both directions and symmetry, that is, the normal distribution. These five
types of Pearson distributions encompass a large variety of distribution
21.4 CONTINUOUS FREQUENCY DISTRIBUTIONS
shapes. Of these, the Pearson type III (log-Pearson type III) distribution is
used most commonly in hydrology, particularly for flood frequency analysis A large number of frequency distributions are used in hydrology. Many of
(e.g., Bobée, 1975). these distributions are discussed in Kite (1977), Johnson et al. (1994), Singh
Table 21.2 Four Popular Probability Distribution Families

Family Equation Parameters
1 d f (x )
Pearson = −(a + x )/ (c0 + c1 x + c2 x 2 ) (21.6) a, c0 , c1 , c2
f (x ) d x
dy
Burr = y(1 − y ) g ( x , y ), y = F ( x ); g ( x , y ) > 0 if y ∈[0,1] (21.7)
dx
1 d y a0 + a1 x + a2 x 2
Halphen = , y = f ( x ) (21.8) a0 , a1 , a2 , q
y dx xq
Johnson  x −c
z = a + bf  (21.9) a, b, d > 0; c ∈R
 b 
 x −c 
Johnson SL & SB z = a + b log  (21.10)
 c + d − x 
2
 x −c x −c  x −c
Johnson SU z = a + b sinh −1  = a + b log( +  + 1) (21.11)
 d  d  d 
21_Singh_ch21_p21.1-21.12.indd 2 8/22/16 12:22 PM

Continuous Frequency Distributions 21-3
(1998, 2013), and Rao and Hamed (2000). Let X , f ( x ), F ( x ) be the random The normal distribution is a popular distribution, because it is simple and
variable, its probability density function (PDF), and its cumulative distribu- extensively tabulated. Further, if a random variable is not normally distrib-
tion function (CDF), respectively. Table 21.3 presents most of these uted, it can be transformed to follow a normal distribution. The ability to
distributions. transform any variable to a normally distributed variable was the motivation
underlying the development of the Johnson family of distributions. Hazen
21.4.1 Normal Distribution (1914) was one of the first to employ the normal distribution through
The PDF and CDF of the normal distribution are expressed as Eqs. (21.12a)– probability paper to analyze hydrological data. The normal distribution has
(21.12b) respectively, in which parameters m, σ are the population mean and also been fitted to annual rainfall and runoff data (Markovic, 1965). The nor-
population standard deviation of random variable X. The CDF of the normal mal distribution arises when the causative factors for a variable are additive.
distribution may also be represented as the standardized normal distribution The central limit theorem makes the normal distribution very useful. The
by normalizing X as: z = ( x − µ )/ σ , where Z is known as the standard normal central limit theorem states the mean of the independent identical distributed
variate. The PDF of Z is an even function so that f (z) = −f (z). Thus Eq. (21.12a) (IID) random variables, regardless of the distribution, is approximately
with µ = 0, σ = 1 can be approximated numerically (Abramowitz and Stegun, normally distributed.
1965) as:
21.4.2 Inverse Gaussian or Wald Distributions
f ( z ) = (a0 + a2 z 2 + a4 z 4 + a6 z 6 + a8 z 8 + a10 z 10 )−1 , 0 ≤ z < ∞ , where: Wald (1947) derived a class of distributions, now referred to as Wald distribu-
tions or inverse Gaussian distributions. Johnson et al. (1994) gave a full
a0 = 2.5052367, a2 = 1.2831204, a4 = 0.2264718, a6 = 0.1306469, account of this class. The two-parameter inverse Gaussian distribution has
been expressed in several different forms. The usual form is given as Eqs.
a8 = −0.0202490, a10 = 0.0039132. The error in the approximation (21.13a) and (21.13b) for its PDF and CDF, respectively, in which m3 /l and
is less than 2.3 × 10−4. m /l are the variance and square of the coefficient of variation for random
Correspondingly, the CDF of Z can be approximated numerically as: variable X, respectively.
If l = m2, the two-parameter Inverse Gaussian distribution reduces to the
1 one-parameter Inverse Gaussian distribution. Additionally, adding the loca-
F ( z ) = 1 − f ( z ) a1u + a2u 2 + a3u3 + a4u 4 + a5u5  , u = , p = 0.2316419 tion parameter a : a ∈R , we have the three-parameter Inverse Gaussian distri-
1 + pz
bution given as Eqs. (21.14a) and (21.14b). Strupczewski et al. (2003) found
where, a1 = 0.319381530, a2 = 0.356563782, a3 = 1.781477937, the solution of a linear diffusion wave equation for flow routing in open
a4 = 1.821255978, and a5 = 1.330274429. The error in this approximation is channels subject to impulse input as the inverse Gaussian distribution. Over
less than 7.5 × 10−8. part of the parameter space this distribution is approximately normal.
Table 21.3 Commonly Applied Continuous Probability Distributions in Hydrology

Distribution p.d.f. Parameters c.d.f.
1  ( x − µ )2  x 1  (t − µ )2   x −µ
Normal f (x ) = exp  − (21.12a) µ ∈R F(x ) = ∫ exp  − dt = Φ 
1
 2σ 2  −∞
1
 2σ 2   σ 
σ (2π ) 2 σ (2π ) 2
σ >0 Φ(): standard Normal distribution function (c.d.f.) (21.12b)
1
 λ 2  λ ( x − µ )2   λ  x   2λ   λ  x  
Inverse Gaussian: f (x ) =  3  exp  −
(21.13a) µ > 0; λ > 0 F(x ) = Φ   − 1  + exp  µ  Φ  − x  µ + 1  (21.13b)
Two-Parameter  2π x   2 µ 2 x   x  µ   
(Wald) x > 0
1
 λ 2  λ ( x − µ − a)2   λ  x −a   2λ   λ  x −a 
Inverse Gaussian f (x ) =  exp  − (21.14a) µ > 0; λ > 0 F(x ) = Φ  − 1  + exp   Φ  − + 1  (21.14b)
Three-Parameter  2π ( x − a)3   2 µ 2 ( x − a) 

 x −a  µ   µ  x − a  µ 
x >a a ∈R

1  (ln x − µ )2   ln x − µ 
Log-Normal (LN) f (x ) = exp  − (21.15a) µ ∈R F(x ) = Φ  (21.15b)
1
 2σ 2   σ 
xσ (2π ) 2
x > 0
σ >0 σ x = exp(2µ + σ 2 )(exp(σ 2 ) −1) (21.15d)
2
 σ 
µx = exp  µ +  (21.15c)
 2 
1 (ln( x − a) − µ )2  ln( x − a) − µ 
Three-Parameter f (x ) = exp(− ) (21.16a) a, µ , ∈R F(x ) = Φ  (21.16b)
1
2σ 2  σ 
Log-Normal (LN3) ( x − a) σ (2π ) 2
x − a > 0 σ >0
1  x  x
Exponential f ( x ) = exp  −  , x > 0 (21.17a) a>0 F ( x ) = 1 − exp  −  (21.17b)
a  a  a
1−b
 1

1  b( x − c )  b  b( x − c )  b    b( x − c )  1 
Generalized f (x ) =  1 −  exp  −  1 −  (21.18a) a > 0; F ( x ) = exp  −  1 −   (21.18b)
b
Extreme Value a a  

 a  
   a  

(GEV) b( x − c )
1− >0 b, c ∈R
a
Extreme Value
1  (x − c)  (x − c)    x −c
Type I (EV I) f ( x ) = exp  − − exp  −  (21.19a) a>0 F ( x ) = exp  − exp  −  (21.19b)
a  a  a     a 

x −c ≥0 c ∈R
(Continued)
21_Singh_ch21_p21.1-21.12.indd 3 8/22/16 12:22 PM

Table 21.3 Commonly Applied Continuous Probability Distributions in Hydrology (Continued)


1  ln x − c  ln x − c     ln x − c  
Log-EV I f (x ) = exp  − − exp  −  (21.20a) a>0 F ( x ) = exp  − exp  −  (21.20b)
ax  a  a     a  
ln x − c ≥ 0 c ∈R
− β −1
β  x −c   x − c  −β    x − c  −β 
Extreme Value f (x ) =   exp  −    (21.21a) a, β > 0; F ( x ) = exp  −    (21.21b)
Type II (EVII) a a    a     a  
x >c
β −1
β  x −c   x −c β    x −cβ 
Extreme Value f (x ) = −  exp  −  −   (21.22a) a > 0; F ( x ) = exp  −  −   (21.22b)
Type III (EVIII) a a    a     a  
x ≤c c ∈R , β > 0
a −1
  x a   x a
a x
Weibull f (x ) =   exp  −    , x > 0 (21.23a) a, b > 0 F ( x ) = exp  −    (21.23b)
b b   b    b 
b −1 x
1  x  x 1 x
Γ(b) ∫0
a b −1
Gamma f (x ) =   exp  −  , x > 0 (21.24a) a, b > 0 F(x ) = y exp(− y )dy ; y = (21.24b)
aΓ (b )  a   a a
b x c
Γ  ,   
c  x
b −1
  x  c
 c  a 
Generalized Gamma f ( x ) =   exp  −    (21.25a) a, b, c > 0 F(x ) = 1 − (21.25b)
 b  a    a   b
aΓ   Γ 
 c  c
Γ(,): incomplete Gamma function.

b −1 x −c
1  x −c  x −c 1 x −c
Γ(b) ∫0
f (x ) = exp  − a, b > 0 a y b −1 exp(− y )dy ; y =
Pearson Type III    , x ∈[c , +∞ ) (21.26a) F(x ) = (21.26b)
aΓ(b)  a   a  a
c ∈R
b −1 ln x − c
1  ln x − c   ln x − c  1 ln x − c
Γ(b) ∫0
Log-Pearson f (x ) =   exp  −  (21.27a) a, b > 0 F(x ) = a y b−1 exp(− y )dy ; y = (21.27b)
Type III axΓ(b)  a   a  a
x ∈[exp(c ), +∞ ) c ∈R
Γ(a + b) a−1 Γ(a + b) x a−1

Γ(a)Γ(b) ∫0
Beta f (x ) = x (1 − x )b−1, x ∈(0, 1) (21.28a) a, b > 0 F(x ) = x (1 − x )b−1 dx (21.28b)
Γ(a)Γ(b)

 x −b
exp  − 
 a  1
Logistic f (x ) = ; x ∈(−∞ , +∞ ) (21.29a) b ∈R , a > 0 F(x ) = (21.29b)
  x −b
2
 x −b
a  1 + exp  −  1 + exp  − 
  a    a 
1 1
Three-Parameter f (x ) = 2 (21.30a) a > 0, F(x ) = 1 (21.30b)
1 1
Log-Logistic  −  +1  b( x − c ) 
−
 b( x − c )  b   b( x − c )  b 1+ 1+
b
a1+ 1+   1 +   a 

  a    a  b ≥ 1,

x ≥c c ∈R
a−1 exp(−(1 − b) y ) 1
Generalized Logistic f ( x ) = (21.31a) a > 0, F(x ) = , b ≠ 0
(1 + exp(− y ))2 1
 b( x − c ) b
1+ 1− 
 a 
  b( x − c )   ( x − c )(1 − b) 
 −b −1 ln  1 −  , b ≠ 0 exp  −
  a  x  a  1
y= b, c ∈R F(x ) = ∫ 2 dx , b = 0 (21.31b)
 x − c , b = 0
−∞
 x −c  2
a  exp  −  + 1
 a   a  

−b
f ( x ) = bab x −b−1 , x > a (21.32a) a, b > 0  x (21.32b)
Two-Parameter F(x ) = 1 −  
 a
Pareto
 1
−1
 1
 1  bx  b bx    bx   b
  1 −  , x > 0, < 1, b ≠ 0  1 −   a   , b ≠ 0
1 −  
Generalized f (x ) =  a a a (21.33a) a > 0, b ∈R F(x ) =  (21.33b)
Two-Parameter  1 x   x
exp(− ), b = 0 1 − exp − , b = 0
Pareto 
 a a 
  a 
21_Singh_ch21_p21.1-21.12.indd 4 8/22/16 12:22 PM

Classification of Continuous Frequency Distributions 21-5


 1
−1 a>0  1
 1  b( x − c )  b   b( x − c )  b
 1−  , b ≠ 0  1 − 1 −
Generalized Three- f (x ) =  a a  (21.34a) b, c ∈R F(x ) =  a  (21.34b)
Parameter Pareto  1  x −c   x −c
 exp  −  , b = 0  1 − exp  − a 
 a  a  
n
Γ x  
1  x  −1
n  2
2
Chi-Square f (x ) = n exp  −  x 2 , x > 0 (21.35a) n : degree of F(x ) = (21.35b)
 2 n
 n
22 Γ    freedom Γ  
 2  2
x  n
 n   −1 n x 
Γ x   = ∫ 2 t  2  exp(−t )dt = γ  ,  : lower incomplete Gamma function.
 2 0 2 2
2
y n−1  y2   n  n
Chi f ( y) = exp  −  , y = x , X ~ χ 2 (n) (21.36a) n : degree of F ( y ) = Γy 2   / Γ   (21.36b)
n
n  2  2  2
22 Γ  
−1
freedom 2
 
2
2 −1
1   x − a  1 1  x − a
Cauchy f (x ) = 1 +    (21.37a) a ∈R , b > 0 F ( x ) = + arctan  (21.37b)
πb  b   2 π  b 

Burr Type XII f ( x ) = kcx c −1 (1 + x c )−( k +1) , x ≥ 0 (21.38a) k, c > 0 F ( x ) = 1 − (1 + x c )− k (21.38b)
 j  j 
Γ  k −  Γ  + 1
 c  c 
M j = E( X j ) = k (21.38c)
Γ(k + 1)
 c −1 c − ( k +1)  c −k
 ckx  1 + x   1− 1+ x 
 a   a    a 
 
Burr Type XII f (x ) =  − ( k +1)
; x ≥ 0 (21.39a) a, k , c > 0 F(x ) =  −k
(21.39b)
(Three-Parameter)  ckx c −1   x  c     x c
 c  1 +  a    1 −  1 +   
 a     a 
k
 xc 
Burr Type III f ( x ) = kcx − c −1 (1 + x − c )− c −1 , x > 0 (21.40a) k, c > 0 F(x ) =  (21.40b)
 1 + x c 
 j  j
Γ  k +  Γ 1− 
j  c  c
M j = E(Y ) = k (21.40c)
Γ(k + 1)

Burr Type II f ( x ) = k exp(− x )(1 + exp(− x ))−( k +1) , x ≥ 0 (21.41a) k>0 F ( x ) = (1 + exp(− x ))− k (21.41b)
− ( k +1) −k
k x  x
Lomax f (x ) =  1 +  , x ≥ 0 (21.42a) k>0 F(x ) = 1 −  1 +  (21.42b)
a a  a
2a
  x 2 x 
a
 x
 ( a − 1)  b  + a  b  a
   b 
Kappa f (x ) = 1 , x ≥ 0 (21.43a) a, b > 0 F(x ) = 1 (21.43b)
+1
(Two-Parameter)  a
x a   x a a
x  a +     a +  b  
  
b   
ac
x
a2c   ac ac −1
 b
, x ≥ 0  x   x 
Kappa f (x ) = (21.44a) a, b, c > 0 F(x ) =    a +    (21.44b)
 x 
ac 2  b   b 
(Three-Parameter)
x  a +   
  b 
1   x m 
Halphen-A f A (x ) = x n−1 exp  −a  +   , x > 0 (21.45) m > 0, a > 0
2mn Kn (2a)   m x 
Kn () : modified Bessel function of second kind n ∈R
of other n.
  x  2 ax 
1
Halphen-B fB (x ) = x 2n−1 exp  −   +  (21.46) m > 0, n > 0
m2n efn (a)   m m
x > 0 ; efn () : exponential factorial function a ∈R
(Continued)
21_Singh_ch21_p21.1-21.12.indd 5 8/22/16 12:22 PM


  m  am  2
1
Halphen-IB f IB ( x ) = x −2n−1 exp  −   + (21.47) m > 0, n > 0
m−2n efn (a)   x x 
x >0 a ∈R
1
(1 − (1 − m)ax )1−m
Tsallis f (x ) = 1
, x > 0 (21.48a) a>0
∞
∫0 (1 − (1 − m)at )1−m dt
1
m-exponential expm ( x ) = (1 − (1 − m)x )1−m (21.48b)
−1
 ∞

f ( x ) = expm (−ax )  ∫ expm (−ax )dx  (21.48c)
 0 
−1
∞ 
Escort f ( x ) = f m ( x )  ∫ f m (t )dt  (21.49) m: Tsallis
0  entropy index
1
−1−
Kousoyiannis f ( x ) = (1 + k(a0 + a1 x c1 )) k x c2 −1 (21.50)
Wakeby x = −a(1 − F )b + c(1 − F )− d + a0 (21.51) a, b, c , d > 0;

a0 ∈R
m
Sinepower f ( x ) = mnx n−1 cos( x n )(sin( x n ))m−1 , x ≥ 0 (21.52a) m, n > 0 F ( x ) =  sin( x n ) (21.52b)
y − y min
x=
ym − y min
1 − F0
Extended Sinepower f ( x ) = mny n−1 cos( y n )(sin( y n ))m (21.53a) m, n > 0 F ( x ) = F0 + (1 − F0 )(sin( y n ))m (21.53b)
zm − z min
z − z min z −z x
x= , xm = m min , y = F0 : CDF at z = z min
z max − z min z max − z min xm
General distribution f ( x ) = (1 − F0 )abx a−1 (1 − x a )b−1 , x ∈[0,1] (21.54a) a, b ≥ 0 F ( x ) = F0 + (1 − F0 )(1 − (1 − x a ))b (21.54b)
for double bounded z − z min
variable x= F0 : CDF at z = z min
z max − z min
a
Boughton K = A+ , a = 1.009 A2.04 (21.55)
 T 
ln ln   −A
 T − 1
T
: recurrence interval determined from the plotting position
formula;
K: frequency factor for the selected recurrence interval as:
− −
log Q = q+ KS , q = mean(log q)
S = std(log q)
−k
f ( x|b) = b exp(−bx ) x ≥ 0, b > 0 : Exponential  x
Compound (21.56a) a, k > 0 F(x ) = 1 −  1 +  (21.56b)
 a
distribution akb k −1
Exponential- f (b) = exp(−ab), a, k > 0 : Gamma
Gamma Γ( k )
k x ( )
− k +1
∞
f ( x ) = ∫ f ( x|b) f (b)d b =  1 + 
0 a  a 
−n
 xa 
Compound Weibull f ( x | b) = abx a−1 exp(−bx a ), x ≥ 0, a, b > 0 (Weibull) (21.57a) a, c , n > 0 F(x ) = 1 −  1 +  (21.57b)
(Weibull-Gamma)  c 
cnbn−1
f (b) = exp(−cb), c > 0, n > 0;(Gamma)
Γ(n)
∞ ancn x a−1
f ( x ) = ∫ f ( x|b) f (b)d b = ,x ≥0
0 ( x a + c )n+1
−1
 xa 
Weibull-Exponential f ( x|b) = abx a−1 exp(−bx a ), x ≥ 0, a, b > 0 (21.58a) a, c > 0 F(x ) = 1 −  1 +  (21.58b)
 c 
f (b) = c exp(−cb), c > 0
∞ acx a−1
f ( x ) = ∫ f ( x|b) f (b)d b =
0 ( x a + c )2
− ( a +b ) −1
c  x
ac −1
  x c B y (a , b )   x  −c 
Generalized Beta of f (x ) =    1 +  d   (21.59a) a, b, c , d > 0 F(x ) = ; y =  1 +    (21.59b)
second kind dB(a, b)  d    B(a, b)   d 
B(,): complete Beta Function. By (,): incomplete Beta function
21_Singh_ch21_p21.1-21.12.indd 6 8/22/16 12:22 PM

Continuous Frequency Distributions 21-7
21.4.3 Log-Normal Distributions 21.4.8 Extreme Value Type III (EVIII) Distribution
The log-normal probability distribution arises when the causative factors are The EV III distribution has an upper bound. The EV III distribution belongs
multiplicative. Since Chow (1951, 1954) proposed a log-probability law in to the Weibull family also known as inverse Weibull distribution. The EV III
hydrology, this distribution has been widely used for frequency analysis of distribution may be obtained if the shape parameter b > 0. Let b = 1/β , its PDF
hydrological data. Aitchison and Brown (1957) gave a full account of the and CDF are given as Eqs. (21.22a)–(21.22b). In EV III distribution, parameters
log-normal distribution. Let Y = log X, 0 < x < ∞. If Y follows a normal distri- a, β , c are the scale, shape, and location parameters, respectively. The EV III
bution, X follows a log-normal distribution. The two-parameter PDF and distribution has been used for frequency analysis of low flow extremes.
CDF of X are given as Eqs. (21.15a)–(21.15b), where m and σ are the mean Figure 21.1 shows the GEV density function plots for different values of
and standard deviation of Y. The parameters m and σ can be related to the shape parameter b to represent the EVI (Gumbel), EVII, and EVIII distributions.
mean and standard deviation of original random variable X given as Eqs.
(21.15c)–(21.15d). If z = (ln x − µ )/σ , the variable Z follows the standardized
normal distribution. GEV density function for a = 2, c = 3
The three-parameter log-normal distribution is frequently used in fre- 0.35
quency analysis of extreme floods, seasonal flow volumes, flow duration EVI
curves, rainfall intensity-duration, soil water retention, pore radius, and pore EVII, b = −1.25
capillary pressure (e.g., Stedinger, 1980). The PDF and CDF of the three- 0.3
EVIII, b = 1.25
parameter log-normal distribution are given as Eqs. (21.16a)–(21.16b).
Again, the standardized variable Z following the standard normal distribution 0.25
is obtained as z = [ln(x − a) − my]/σy .
21.4.4 Exponential Distribution 0.2
The exponential distribution is applied to model the distribution of the time
pdf
between independent events with the constant average rate. It has been used
for the frequency analysis of rainfall depth, intensity, and duration, time 0.15
interval between flood events, and reliability analysis. The PDF and the CDF
of the exponential distribution are given as Eqs. (21.17a)–(21.17b) in which 0.1
parameter a is the mean. The exponential distribution is related to the
Poisson process wherein the distribution of time between recurrences is
exponentially distributed. 0.05
21.4.5 Generalized Extreme Value Distribution

0
The generalized extreme value (GEV) distribution, introduced by Jenkinson −10 −5 0 5 10
(1955) and recommended by Natural Environment Research Council (1975) of x
Great Britain, is widely used for flood frequency analysis, particularly in west-
Figure 21.1 Plot of the GEV probability density function for different shape
ern Europe. Its PDF and CDF are given as Eqs. (21.18a)–(21.18b) in which
parameters.
parameters a, b , c are the scale, shape, and location parameters, respectively.
The range of X is controlled by shape parameter, b; for b > 0, it is bounded as
−∞ < x ≤ c + (a/b); for b < 0, it bounded as c + (a/b) ≤ x < ∞. Depending on the 21.4.9 Weibull Distribution
value of shape parameter b, the GEV distribution specializes into the extreme
The Weibull distribution, derived from the power transformation of the expo-
value type I (or Gumbel) distribution for b = 0, the extreme value type II (or
nential variable, is used for frequency analysis of low flow and precipitation
Fréchet) distribution for b < 0, the extreme value type III distribution for b > 0,
characteristics, hazard and survival functions, and risk and reliability analysis.
the reverse Raleigh distribution for b = 2, and the reverse exponential distribu-
This distribution is a reverse GEV distribution. Murthy et al. (2004) have
tion for b = 1. EV I, EV II, and EV III distributions are further discussed in what
described a whole family of Weibull distributions. Its PDF and CDF are given
follows.
as Eqs. (21.23a)–(21.23b), where a, b are the shape and scale parameters,
21.4.6 Extreme Value Type I (EVI) and respectively.
Log-EVI Distributions
21.4.10 Gamma Distribution
The extreme value type I (EVI) distribution, also known as the Gumbel distri-
bution (Gumbel, 1941, 1942), is one of the most popular distributions for fre- The two-parameter gamma distribution is used for synthesis of instantaneous
quency analysis of extremes of floods, droughts, and rainfall and meteorological or finite-period unit hydrographs and flood frequency analysis. Its PDF and
data. By limiting the shape parameter b → 0, the GEV distribution converges to CDF are given as Eqs. (21.24a)–(21.24b), where a, b are the scale and shape
EV I (or Gumbel) distribution. Its PDF and the CDF are given as Eqs. (21.19a)– parameters, respectively. Following Abramowitz and Stegun (1958), the
(21.19b), respectively. In the EV I distribution, a is the scale parameter and c is gamma distribution reduces to the chi-square distribution, if y = x / a, that is,
the location parameter. Additionally, the EV I distribution has a constant skew- F ( y ) = F ( χ 2|v ) : χ 2 = 2 y , v = 2b (degree of freedom). For v > 30 the gamma
ness coefficient (about 1.14) and constant kurtosis (about 5.4). If the return distribution tends to approach a normal distribution with zero mean and unit
x −c variance, where the variate u is defined through the following transformation:
period T is defined as T = 1/(1 − F) and let y = then the reduced variate χ2 2 9υ
a u =[( )1 3 + − 1]( )1 2 .
  T − 1  υ 9υ 2
yT can be defined as yT = − ln  − ln  . .
  T  
21.4.11 Generalized Gamma Distribution
Let Y = lnX. If Y has an EV I distribution then X has a log-EV I distribution
Using the Boltzmann–Gibbs–Shannon (BGS) entropy, Papalexiou and
with PDF and CDF given as Eqs. (21.20a)–(21.20b). The log-EV I distribution
Koutsoyiannis (2012) derived the three-parameter generalized gamma distri-
has been used for frequency analysis of instantaneous flood peaks (Reich,
bution and tested it on 11,519 daily rainfall records from across the globe. The
1970). The greatest attraction of EVI is that it is expressed in closed form.
PDF and CDF of this distribution can be expressed as Eqs. (21.25a)–(21.25b),
21.4.7 Extreme Value Type II (EVII) Distribution respectively, where a is the scale parameter, and b, c are the shape parameters.
The generalized gamma distribution is a flexible distribution that includes
The EVII distribution, known as Fréchet distribution, may be obtained if the
exponential, two-parameter gamma, Weibull, chi-squared, and other distri-
shape parameter b < 0 in GEV distribution. The EV II distribution has also
butions as special cases. Figure 21.2 plots the generalized gamma distribution
1
been applied to the frequency analysis for the extreme values. Let b = − < 0. and its special cases.
β
Its PDF and CDF are given as Eqs. (21.21a)–(21.21b). In EVII distribution, a, 21.4.12 Pearson Type III and Log-Pearson
c, and b are the scale, location, and shape parameters, respectively. In hydrol- Type III Distributions
ogy, the EVII distribution may be applied to model the annual extremes, for The Pearson type III (PIII) distribution is popular in flood and extreme rain-
example, extreme rainfall (Koutsoyianni, 2004). fall frequency analysis and can be considered as a three-parameter gamma
21_Singh_ch21_p21.1-21.12.indd 7 8/22/16 12:22 PM

Generalized Gamma and its special distributions The three-parameter log-logistic distribution (LLD) has been used for fre-
0.35 quency analysis of precipitation and streamflow data. Its PDF and CDF can
Generalized Gamma: a = 4, b = 3, c = 2 be written as Eqs. (21.30a)–(21.30b) where a, b, c are the scale, shape, and
Gamma: a = 4, b = 3 location parameters.
0.3 Exponential: a = 4
Weibull: a = 4, b = 3 21.4.15 Generalized Logistic distribution
0.25 The generalized logistic (GLO) distribution has been recommended as the
standard for flood frequency analysis in the United Kingdom (IH, 1999;
Kjeldsen and Jones, 2004). The GLO distribution is a three-parameter distri-
0.2 bution proposed by Hosking and Wallis (1997). Its PDF and CDF can be writ-
ten as Eqs. (21.31a)–(21.31b) in which a, b , c are the scale, shape, and location
pdf
parameters, respectively. The range of random variable X is dependent on

0.15
a a
these parameters as: x ∈( −∞ , c +  for b > 0, x ∈ c + , +∞ ) for b < 0, and
b  b
0.1 x ∈(−∞ , +∞ ) for b = 0. The event with T-year return period can be expressed
a
in the closed-form as: xT = c + (1 − (T − 1)− b ).
0.05 b
21.4.16 Pareto Distributions
0 The Pareto distribution, introduced by Pickands (1975), is a heavy tailed

0 5 10 15 20 distribution and is applied to model peak over threshold flood values. A full
x account of the Pareto distributions was given by Arnold (1983). Three forms
of the Pareto distribution have been used: (1) two-parameter, (2) generalized
Figure 21.2 Plots of generalized gamma probability density function and its
special cases. two-parameter, and (3) three-parameter generalized. The PDF and CDF of
two-parameter Pareto distribution can be expressed as Eqs. (21.32a)–(21.32b),
where a and b are the scale and shape parameters, respectively. This is a spe-
distribution. Its PDF and CDF can be expressed as Eqs. (21.26a)–(21.26b),
cial form of the Pearson type VI distribution. The two-parameter generalized
where a, b, and c are the scale, shape, and location parameters, respectively.
Pareto distribution has a PDF and CDF given as Eqs. (21.33a)–(21.33b),
The log-Pearson type III (LPIII) distribution is one of the most frequently
where a, b are the scale and shape parameters, respectively. If the shape
used distributions for frequency analyses of hydrologic extremes and was
parameter b = 0, this distribution reduces to the exponential distribution. The
recommended by Water Resources Council (1967) of the United States. If
three-parameter generalized Pareto distribution has its PDF and CDF
Y = lnX follows the PIII distribution, X follows the LPIII distribution with
expressed as Eqs. (21.34a)–(21.34b), where a, b , c are the scale, shape, and
PDF and CDF expressed as Eqs. (21.27a)–(21.27b), where a, b, and c are the
location parameters, respectively. Again, if b = 0, this distribution reduces to
scale, shape, and location parameters, respectively. Figure 21.3 shows an
the exponential distribution with a location parameter.
example of the density functions of the PIII and LPIII distributions.
21.4.13 Beta Distribution 21.4.17 Chi-Square Distributions
The beta distribution is used in reliability and safety analysis of civil engineer- For a Chi-square random variable with n degrees of freedom, the PDF and the
ing systems where variables or parameters are generally bounded and skewed CDF can be expressed as Eqs. (21.35a)–(21.35b). Lancaster (1969) gave a full
random quantities. This distribution is a Pearson type I distribution. Its PDF account of this distribution. The standard Chi-square distribution with
and CDF can be expressed as Eqs. (21.28a)–(21.28b), where a, b are the shape n degrees of freedom is the same as the gamma distribution with shape
parameters. Singh (1996) found the solution of kinematic converging and parameter n/2 and scale parameter 2. The mode of the distribution is at 0 if
diverging overland flows in terms of the beta distribution. n ≤ 2, and at n − 2 for n > 2. If y = x , where X has a chi-square distribution
with n degrees of freedom, Y has a chi-distribution with n degrees of freedom
21.4.14 Logistic and Log-Logistic Distributions with PDF and CDF as Eqs. (21.36a)–(21.36b). It may be interesting to note
The logistic distribution has been applied to frequency analysis of floods. Its that Eq. (21.36) becomes the half normal distribution for n = 1, the Raleigh
PDF and CDF can be expressed as Eqs. (21.29a)–(21.29b), where a is the scale distribution for n = 2, and the Maxwell–Boltzmann distribution in continu-
parameter, and b is the location parameter. If Y = [ln(X − b)/a)]/a has a logis- ous domain for n = 3. Originally the MB distribution was derived as a discrete
tic distribution then X has a log-logistic distribution. The log-logistic distribution. The mode of Chi-square distribution occurs at 0 if n ≤ 1 and at
distribution (LLD) is similar in shape as the log-normal distribution and has n − 1 if n > 1.
been used for frequency analysis of survival data and precipitation data. 21.4.18 Cauchy Distribution
The Cauchy distribution has PDF and CDF given as Eqs. (21.37a)–(21.37b) in
Pearson type distribution which a and b are the location and scale parameters, respectively. Cauchy dis-
0.04 tribution has no mean, variance, or higher moments. It is interesting to note
PIII(4,6,2) that the distribution is symmetrical about x = a, has median as a, and has upper
LPIII(2,1.2,2.07)
0.035 and lower quartiles as a ± b. The PDF exhibits points of inflexion at a ± b 3 . At
these points the values of CDF are 0.273 and 0.727, whereas the corresponding
0.03 values are 0.159 and 0.841 for the normal distribution. The Cauchy distribution
has heavier tails than the Gumbel distribution and can be used for frequency
0.025 analysis of extremes. Nathan and McMahon (1990) state that although the data
may fit it, it is unsuitable for low flow analysis.
pdf
0.02 21.4.19 Burr Distributions

Three distributions of the Burr family (Burr, 1942) that have been more
0.015 widely used in hydrology and hydrometeorology are types II, III, and XII,
because their forms and their inverses are expressed in simple closed forms.
0.01 Tadikamalla (1980) showed that several distributions are special cases of the
Burr type XII distribution.
0.005 21.4.20 Burr Type XII Distribution
The PDF and CDF of this distribution can be expressed as Eqs. (21.38a)–
0
20 40 60 80 100 120 140 160 180 200 (21.38b) in which k , c are the shape parameters. The PDF is unimodal at
x x = [(c − 1)/(kc + 1)1 c ], provided c >1, and L-shaped if c ≤ 1. The Burr type XII
is a two-parameter distribution but can be extended to (i) a three-parameter
Figure 21.3 Plots of PT III and LPT III probability density functions.
21_Singh_ch21_p21.1-21.12.indd 8 8/22/16 12:22 PM

Conclusion 21-9
distribution by incorporating the scale parameter, and (ii) a four-parameter would lead to several exponential and power distributions, including beta
distribution by incorporating both the location and scale parameters. A scale prime, gamma, Weibull, and Pareto.
parameter a can be incorporated in Eq. (21.38) in two ways given as Eqs.
(21.39a)–(21.39b). To express the PDF and CDF for the four-parameter Burr 21.4.29 Wakeby Distribution
XII distribution, we can simply substitute x in Eq. (22.39) as x − m, given Houghton (1978) introduced a five-parameter distribution, called Wakeby,
µ ∈R , x ≥ µ . It is interesting that the j-th power moment of X following the for modeling flood flows. In inverse form, the Wakeby distribution is
two-parameter Burr type XII distribution can be computed using Eq. (21.38c). expressed as Eq. (21.51) in which F is the uniform variate (0, 1), a, b, c, and d
are the shape parameters, and a0 is a location parameter. This distribution has
21.4.21 Burr Type III Distribution found significant use for modeling hydrologic extremes (Griffiths, 1989;
Applying the transformation Y = 1/X to the Burr Type XII distribution, the Rahman et al., 2015).
Burr type III distribution can be expressed as Eqs. (21.40a)–(21.40b) with the
j-th power moment of Y given as Eq. (21.40c). Tadikamalla (1980) showed that 21.4.30 Sinepower Distributions
as k → ∞, the distribution function reduces to F(x) = exp(−x −c) which can be Kumaraswamy (1976, 1978, 1980) introduced sinpower distributions for mod-
considered as the distribution of a reciprocal Weibull variate. In other words, eling bounded data. Let Y be a random variable with lower bound ymin and
if X has the Weibull distribution, Y = 1/X has the aforementioned reciprocal mode ymode. Using a transformation, the sinepower distribution has PDF and
Weibull distribution. CDF as Eqs. (21.52a)–(21.52b). It may be noted that x is expressed in radians,
and (xmin ) = 0 when x = xmin = 0; f (x) = 0 when x = 0 and x = xmax, since
21.4.22 Burr Type II Distribution
xnmax = p /2 and cos(xn) = 0. Kumaraswamy (1976) showed that n = ln(p/2)/ ln(xmax),
If X = c log y, Eqs. (21.40a) and (21.40b) lead to the PDF and CDF for the Burr
type II distribution which can be expressed as Eqs. (21.41a)–(21.41b).
( n
and m = sec 2 x mode)−
( )
 n − 1 tan x mode 
n
n  . Kumaraswamy (1978) extended the

n x mode 
21.4.23 Lomax Distribution  
For analyzing business data, Lomax (1954) employed a distribution whose sinepower distribution by using a transformation, whose PDF and CDF are
PDF and CDF can be written as Eqs. (21.42a)–(21.42b) in which a is a scale expressed as Eqs. (21.53a)–(21.53b). Kumaraswamy (1980) presented a more
parameter and k is a shape parameter. The Lomax distribution is a special case general distribution for double bounded variables with PDF and CDF as
of the three-parameter Burr type XII distribution with c = 1. Lomax (1954) Eqs. (21.54a)–(21.54b) using a transformation.
f (x ) k
defined the failure rate function as Z = = . 21.4.31 Boughton Distribution
1 − F ( x ) a( x + a)
21.4.24 Kappa Distributions Boughton (1980) empirically proposed a three-parameter distribution
expressed by Eq. (21.55) for modeling flood flows from 78 watersheds from
For analyzing precipitation data, Mielke (1973) used the two-parameter kappa eastern Australia. He also showed a relation between frequency factor and
distribution with PDF and CDF, respectively, as Eqs. (21.43a)–(21.43b) where recurrence interval.
a, b are the shape and scale parameters, respectively. The three-parameter
kappa distribution can be written for its PDF and CDF as Eqs. (21.44a)– 21.4.32 Compound Distributions
(21.44b) where a, c are the shape parameters and b is the scale parameter. Some aforementioned distributions are examples of compound distributions.
Hosking (1994) discussed the four-parameter kappa distribution that is a Dubey (1970) has shown that the Lomax distribution is analogous to the
reprameterization of the three-parameter kappa distribution with an addi- exponential-gamma distribution given as Eqs. (21.56a)–(21.56b) which is the
tional location parameter. Singh and Deng (2003) estimated its parameters same as Eqs. (21.42a)–(21.42b) for the Lomax distribution.
using entropy. It specializes into the three parameter kappa distribution plus
Pareto, GEV, Gumbel, logistic, and uniform distributions, and is related to the 21.4.33 Compound Weibull Distribution
Burr type III and type XII distributions. The compound Weibull distribution is obtained by considering the scale
21.4.25 Halphen Distributions parameter in the Weibull distribution as gamma distributed (Dubey, 1966)
expressed as Eqs. (21.57a)–(21.57b) which is the same as the Burr type XII
Perreault et al. (1999a, b) described the Halphen distribution types A, B, and distribution.
IB (Halphen, 1941, 1955). Following their work, these three distributions have
PDFs as Eqs. (21.45)–(21.47). The gamma and inverse gamma distributions 21.4.34 Weibull Exponential Distribution
can be shown to be limiting cases of the Halphen distributions. These distribu- The Weibull-exponential distribution can be obtained by considering the
tions can be applied to a variety of hydrologic variables. Perreault et al. (1999a, scale parameter b in Eq. (21.58a) as exponentially distributed, that is, n = 1.
b) discussed mathematical properties for these distributions. El Adlouni et al. Then, the Weibull exponential distribution can be expressed as Eqs. (21.58a)–
(2010) compared the performances of Halphen and GEV distributions for (21.58b) which is the same as the Burr type XII distribution.
flood frequency analysis. They showed Halphen distributions are good candi-
dates to model the extreme value variables. 21.4.35 Generalized Beta of the Second
Kind Distribution
21.4.26 Tsallis Distribution
Papalexiou and Koutsoyiannis (2012) derived a four-parameter beta of sec-
Drainage basin characteristics, such as width function, exhibit power law or ond kind (GB2) distribution with the use of the BGS entropy, and applied it
multifractal scaling. For frequency analysis of such characteristics, the Tsallis to 11,519 daily rainfall records from across the globe. The PDF and CDF of
distribution that is derived from the Tsallis entropy may be appropriate this distribution can be expressed as Eqs. (21.59a)–(21.59b), where d is the
(Keylock, 2005). The Tsallis distribution can be expressed as Eq. (21.48a) in scale parameter and a, b, c are the shape parameters. The GB2 is a flexible
which m > 0 is the entropy index, and a is a parameter. Equation (21.48a) is distribution that nests Beta Prime, lognormal, Weibull, gamma, generalized
related to the exponential distribution. Using the definition of m-exponential gamma, log-logistic, Burr types III and XII, Lomax, and other distributions.
given as Eq. (21.48b), Eq. (21.48a) can be expressed as Eq. (21.48c). For m → 1, Figure 21.4 plots the generalized beta distribution of second kind and some
Eq. (21.48) will converge to the exponential distribution. of its special case.
21.4.27 Escort Distribution
The distribution is given as Eq. (21.49). Keylock (2005) used this distribution 21.5 CONCLUSION
for analyzing the recurrence interval of extreme floods. A wide variety of frequency distributions, ranging from one-parameter to
five-parameter, are available for hydrological analyses. Some of the distribu-
21.4.28 Kousoyiannis Distribution
tions are quite general and lead to other distributions as special cases.
Using Tsallis entropy, Koutsoyiannis (2005a, b) proposed a four-parameter Literature survey shows that for most hydrological frequency analyses two- or
distribution that can be expressed as Eq. (21.50), where a is a scale parameter, three- or exceptionally four-parameter distributions suffice. All of these dis-
and c1, c2, and k are shape parameters. Parameter a0 is determined by substi- tributions are derived without reference to hydrology. It would be desirable to
tuting Eq. (21.50) into the total probability equation so is not considered as a investigate if some of the distributions could be derived using hydrological
parameter. Here, k = (1 − m)/m where m is the Tsallis entropy index. With laws and then estimate their parameters using measurable hydrologic charac-
appropriate transformations, Koutsoyiannis (2005a) showed that Eq. (21.50) teristics. This might further help unify some of these distributions.
21_Singh_ch21_p21.1-21.12.indd 9 8/22/16 12:22 PM

Generalized Beta distribution of second kind Deutsch, C. V. and P. W. Cockerham, “Practical considerations in the appli-
1 cation of simulated annealing to stochastic simulation,” Mathematical
Generalized B2 (0.5,4,2,2) Geology, 26 (1): 67–82, 1994.
0.9 Generalized Gamma (b to inf, d = b*b1/a) Deutsch, C. V. and A. G. Journel, GSLIB: Geostatistical Software Library and
Beta 2 (5,4,1,1)(c = d = 1) User’s Guide, New York, 1992.
0.8 BR12 (1,4,0.8,2)(a = 1) Deutsch, C. V. and X. H. Wen, “Integrating large-scale soft data by simu-
BR3 (3,1,0.1,2)(b = 1) lated annealing and probability constraints,” Mathematical Geology, 32 (1):
0.7 Log−Logistic (1,1,1,3)(a = b = c = 1) 49–67, 2000.
Dietrich, C. R. and G. N. Newsam, “A fast and exact method for multidi-
0.6
mensional Gaussian stochastic simulations: extension to realizations
conditioned on direct and indirect measurements,” Water Resources Research,
pdf
0.5
32 (6): 1643–1652, 1996, doi:10.1029/93WR01070.
0.4
Diggle, P. and P. J. Ribeiro, Model-Based Geostatistics, Springer Science &
Business Media, New York, 2007.
0.3 Diggle, P. J. and P. J. Ribeiro, Jr., “Geostatistical design,” Model-Based
Geostatistics, 199–212, 2007, doi:10.1007/978-0-387-48536-28.
0.2 Doetsch, J., M. B. Kowalsky, C. Doughty, S. Finsterle, J. B. Ajo-Franklin, C.
R. Carrigan, X. Yang, S. D. Hovorka, et al., “Constraining CO2 simulations by
0.1 coupled modeling and inversion of electrical resistance and gas composition
data,” International Journal of Greenhouse Gas Control, 18: 510–522, 2013,
0 doi:10.1016/j.ijggc.2013.04.011.
0 2 4 6 8 10 Dwivedi, D., B. P. Mohanty, and B. J. Lesikar, “Estimating Escherichia coli
x loads in streams based on various physical, chemical, and biological factors,”
Figure 21.4 Plots of generalized beta distribution of second kind and some of its Water Resources Research, 49: 2896–2906, 2013, doi:10.1002/wrcr.20265.
special cases. Ester, M., A. Frommelt, H-P. Kriegel, and J. Sander, “Spatial data mining:
database primitives, algorithms and efficient DBMS support,” Data Mining
and Knowledge Discovery, 4 (2–3): 193–216, 2000, doi:10.1023/
REFERENCES A:1009843930701.
Andrieu, C., N. De Freitas, A. Doucet, and M. I. Jordan, “An introduction Feyen, L. and J. Caers, Quantifying geological uncertainty for flow and
to MCMC for machine learning,” Machine Learning, 50 (1–2): 5–43, 2003, transport modeling in multi-modal heterogeneous formations, Advances in
doi:10.1023/A:1020281327116. Water Resources, 29 (6): 912–929, 2006, doi:10.1016/j.advwatres.2005.08.002.
Anselin, L., “Interactive techniques and exploratory spatial data analysis,” Finsterle, S. and M. B. Kowalsky, “Joint hydrological-geophysical inversion
Geography Information Systems: Principles, Techniques Management and for soil structure identification,” Vadose Zone Journal, 7 (1): 287–293, 2008,
Applications, eds., P. Longley, M. Goodchild, D. Maguire, and D. Rhind. doi:10.2136/vzj2006.0078.
Cambridge: Geoinformation, Int., 253–266, 1999. Fisher, R. A., An Agricultural Experiment in Randomised Blocks. The Design
Arora, B., B. P. Mohanty, and J. T. McGuire, “Uncertainty in dual permea- of Experiments, Oliver and Boyd, Edinburgh, U.K., 1937.
bility model parameters for structured soils,” Water Resources Research, 48: Gamerman, D. and H. F. Lopes, Markov Chain Monte Carlo: Stochastic
2012, doi:10.1029/2011WR010500. Simulation for Bayesian Inference, 2nd ed., CRC Press, New York, 2006.
Arora, B., B. P. Mohanty, J. T. McGuire, and I. M. Cozzarelli, “Temporal Ganio, L. M., C. E. Torgersen, and R. E. Gresswell, “A geostatistical
dynamics of biogeochemical processes at the Norman Landfill site,” Water approach for describing spatial pattern in stream networks,” Frontiers in
Resources Research, 49: 6909–6926, 2013, doi:10.1002/wrcr.20484. Ecology and the Environment, Vol. 3, 2005, pp. 138–144.
Ashley, R. P. and J. W. Lloyd, “An example of the use of factor analysis and Getis, A., “A history of the concept of spatial autocorrelation: a geographer’s
cluster analysis in groundwater chemistry interpretation,” Journal of perspective,” Geographical Analalysis, 40 (3), 297–309, 2008, doi:10.1111/
Hydrology, 39 (3–4), 355–364, 1978, doi:10.1016/0022-1694(78)90011-2. j.1538-4632.2008.00727.x.
Berkhin, P., “A survey of clustering data mining techniques,” Grouping Griffith, D. A. and P. R. Peres-Neto, “Spatial modeling in ecology: the flex-
Multidimensinal Data, 25–71, 2006, doi:10.1007/3-540-28349-82. ibility of eigenfunction spatial analyses,” Ecology, 87 (10): 2603–2613, 2006,
Carle, S. F., T-PROGS: Transition probability geostatistical software, 1999. doi:10.1890/0012-9658(2006)87[2603:SMIETF]2.0.CO;2.
Carle, S. F. and G. E. Fogg, “Transition probability-based indicator geosta- Han, J. and M. Kamber, Data Mining: Concepts and Techniques, 2nd ed.,
tistics,” Mathematical Geology, 28 (4): 453–476, 1996, doi:10.1007/ Elsevier Inc., Oxford, 2006.
BF02083656. Han, J., M. Kamber, and A. K. H. Tung, “Spatial clustering methods in data
Chen, J., S. Hubbard, Y. Rubin, C. Murray, E. Roden, and E. Majer, mining: a survey,” Geographic Data Mining and Knowledge Discovery,
“Geochemical characterization using geophysical data and Markov Chain pp. 188–217, 2001.
Monte Carlo methods: a case study at the South Oyster bacterial transport site Helena, B., R. Pardo, M. Vega, E. Barrado, J. M. Fernandez, and L.
in Virginia,” Water Resources Research, 40 (12), 2004, doi:10.1029/ Fernandez, “Temporal evolution of groundwater composition in an alluvial
2003WR002883. aquifer (Pisuerga River, Spain) by principal component analysis,” Water
Chen, J., S. Hubbard, J. Peterson, K. Williams, M. Fienen, P. Jardine, and D. Research, 34 (3): 807–816, 2000, doi:10.1016/S0043-1354(99)00225-0.
Watson, “Development of a joint hydrogeophysical inversion approach and Hu, L. Y. and T. Chugunova, “Multiple-point geostatistics for modeling
application to a contaminated fractured aquifer,” Water Resources Research, subsurface heterogeneity: a comprehensive review,” Water Resources Research,
42: 1–13, 2006, doi:10.1029/2005WR004694. 44 (11): 1–14, 2008, doi:10.1029/2008WR006993.
Cressie, N., Statistics for Spatial Data, John Wiley & Sons, New York, 2015. Huysmans, M. and A. Dassargues, “Application of multiple-point geostatis-
Dafflon, B. and W. Barrash, “Three-dimensional stochastic estimation of tics on modelling groundwater flow and transport in a cross-bedded aquifer
porosity distribution: benefits of using ground-penetrating radar velocity (Belgium),” Hydrogeology Journal, 17 (8): 1901–1911, 2009, doi:10.1007/
tomograms in simulated-annealing-based or Bayesian sequential simulation s10040-009-0495-2.
approaches,” Water Resources Research, 48, 2012, doi:W0555310.1029/ Jolliffe, I. T., Principal Component Analysis, 2nd ed., Springer Verlag, New
2011wr010916. York, 2002.
Dale, M. R. T. and M-J. Fortin, “From graphs to spatial graphs,” Annual Kelkar, M. and G. Perez, Applied Geostatistics for Reservoir Characterization,
Review of Ecology, Evolution, and Systematics, 41 (1): 21–38, doi:10.1146/ Society of Petroleum engineers, Richardson, TX, 2002.
annurev-ecolsys-102209-144718, 2010. Kitanidis, P. K., Introduction to Geostatistics: Applications in Hydrogeology,
Day-Lewis, F. D., P. A. Hsieh, and S. M. Gorelick, “Identifying fracture-zone Cambridge University Press, Cambridge, 1997.
geometry using simulated annealing and hydraulic-connection data,” Water Kitanidis, P. K. and E. G. Vomvoris, “A geostatistical approach to the
Resources Research, 36 (7): 1707–1721, 2000. inverse problem in groundwater modeling (steady state) and one-dimensional
de Marsily, G., De l’identification des systemes hydrogeologiques, PhD simulations,” Water Resources Research, 19 (3): 677, 1983, doi:10.1029/
thesis, Paris VI, France, 1978. WR019i003p00677.
21_Singh_ch21_p21.1-21.12.indd 10 8/22/16 12:22 PM

REFERENCES 21-11
Kowalsky, M. B., S. Finsterle, and Y. Rubin, “Estimating flow parameter Parks, K. P., L. R. Bentley, and A. S. “Crowe, Capturing geological realism
distributions using ground-penetrating radar and hydrological measure- in stochastic simulations of rock systems with Markov statistics and simulated
ments during transient flow in the vadose zone,” Advances in Water Resources, annealing,” Journal of Sedimentary Research, 70 (4): 803–813, 2000.
27: 583–599, 2004, doi:10.1016/j.advwatres.2004.03.003. Pebesma, E. and B. Gräler, Introduction to Spatio-Temporal Variography,
Kowalsky, M. B., S. Finsterle, J. Peterson, S. Hubbard, Y. Rubin, E. Majer, A. 2015.
Ward, and G. Gee, Estimation of field-scale soil hydraulic and dielectric Pfeiffer, D., Issues Related to Handling of Spatial Data, Massey University,
parameters through joint inversion of GPR and hydrological data,” Water Palmerston North, June, 1996, pp. 23–28.
Resources Research, 41 (11): 1–19, 2005, doi:10.1029/2005WR004237. Portnov, B. A., J. Dubnov, and M. Barchana, “On ecological fallacy, assess-
Krige, D. G., “A statistical approach to some basic mine valuation problems ment errors stemming from misguided variable selection, and the effect of
on the witwatersrand,” Journal of Chemical, Metallurgical, and Mining Society aggregation on the outcome of epidemiological study,” Journal of Exposure
of South Africa, 52 (6), 119–139, 1952, doi:10.2307/3006914. Science and Environmental Epidemiology, 17 (1): 106–21, 2007, doi:10.1038/
Kumar, P. and E. Foufoula-Georgiou, “Wavelet analysis for geophysical sj.jes.7500533.
applications,” Reviews of Geophysics, 35 (97): 385, 1997, doi:10.1029 Pyrcz, M. J. and C. V Deutsch, “Geostatistical reservoir modeling,”
/97RG00427. Mathematical Geosciences, 47 (4): 497–499, 2015, doi:10.1007/s11004-015-
Lau, K-M. and H. Weng, “Climate signal detection using wavelet trans- 9588-8.
form: how to make a time series sing,” Bulletin of the American Meteorological Remy, N., A. Boucher, and J. Wu, Applied Geostatistics with SGeMS: A User’s
Society, 76 (12): 2391–2402, 1995, doi:10.1175/1520-0477(1995)076<2391:CSD Guide, Cambridge University Press, Cambridge, UK, 2009.
UWT>2.0.CO;2. Ritzi, R. W., D. F. Dominic, N. R. Brown, K. W. Kausch, P. J. McAlenney,
Legendre, P., “Spatial autocorrelation: trouble or new paradigm?,” Ecology, and M. J. Basial, “Hydrofacies distribution and correlation in the Miami
74 (6): 1659–1673, 1993, doi:10.2307/1939924. Valley Aquifer System,” Water Resources Research, 31 (12): 3271–3281, 1995,
Lehikoinen, A., J. M. J. Huttunen, S. Finsterle, M. B. Kowalsky, and J. P. doi:10.1029/95WR02564.
Kaipio, “Dynamic inversion for hydrological process monitoring with electri- Rubin, Y., Applied Stochastic Hydrogeology, Oxford University Press,
cal resistance tomography under model uncertainties,” Water Resources Oxford, 2003.
Research, 46: 1–14, 2010, doi:10.1029/2009WR008470. Sassen, D. S., S. S. Hubbard, S. A. Bea, J. Chen, N. Spycher, and M. E.
Mariethoz, G. and J. Caers, Multiple-Point Geostatistics: Stochastic Modeling Denham, “Reactive facies: an approach for parameterizing field-scale reactive
with Training Images, Wiley-Blackwell, Chichester, UK, 2014. transport models using geophysical methods,” Water Resources Research, 48:
Mariethoz, G., P. Renard, and J. Straubhaar, “The direct sampling method 1–20, 2012, doi:10.1029/2011WR011047.
to perform multiple-point geostatistical simulations,” Water Resources Shekhar, S. and S. Chawla, “Introduction to spatial data mining,” Spatial
Research, 46 (11), 1–14, 2010, doi:10.1029/2008WR007621. Databases: A Tour, Prentice Hall, New Jersey, 2003a.
Martínez, B., M. A. Gilabert, F. J. García-Haro, A. Faye, and J. Meliá, Shekhar, S. and S. Chawla, “Introduction to spatial data mining,” Spatial
“Characterizing land condition variability in Ferlo, Senegal (2001-2009) using Databases: A Tour, Prentice Hall, New Jersey, 2003b, pp. 237–284.
multi-temporal 1-km Apparent Green Cover (AGC) SPOT Vegetation data,” Shekhar, S. and S. Chawla, Spatial Databases: A Tour, Prentice Hall, New
Global and Planetary Change, 76 (3–4): 152–165, 2011, doi:10.1016/j.glopla- Jersey, 2003c.
cha.2011.01.001. Shekhar, S., Y. Huang, W. Wu, C. T. Lu, and S. Chawla, “Data mining for
Matheron, G., Les variables Régionalisées et leur Estimation, Masson, Paris, scientific and engineering applications,” Massive Computing, edited by R. L.
1965. Grossman, C. Kamath, P. Kegelmeyer, V. Kumar, and R. R. Namburu, Springer
Maxwell, R. M., W. E. Kastenberg, and Y. Rubin, “A methodology to inte- US, Boston, MA, 2001.
grate site characterization information into groundwater-driven health risk Shekhar, S., P. Zhang, and Y. Huang, “Trends in spatial data mining,”
assessment,” Water Resources Research, 35 (9): 2841, 1999, doi:10.1029/ Science, 7: 357–379, 2003, doi:10.1007/0-387-25465-X_39.
1999WR900103. Shekhar, S., P. Zhang, and Y. Huang, “Spatial data mining,” Data Mining
Michael, H. A., H. Li, A. Boucher, T. Sun, J. Caers, and S. M. Gorelick, and Knowledge Discovery Handbook, edited by O. Maimon and L. Rokach,
“Combining geologic-process models and geostatistics for conditional simu- Springer, New York, 2010.
lation of 3-D subsurface heterogeneity,” Water Resources Research, 46 (5), Tobler, W., “Cellular geography,” Philosophy in Geography, edited by S. Gale
2010, doi:10.1029/2009WR008414. and G. Olsson, Reidel, Dordrecht, The Netherlands, 1979.
Michalak, A. M., “A method for enforcing parameter nonnegativity in Vega, M., R. Pardo, E. Barrado, and L. Debán, “Assessment of seasonal and
Bayesian inverse problems with an application to contaminant source identi- polluting effects on the quality of river water by exploratory data analysis,”
fication,” Water Resources Research, 39 (2): 1–14, 2003, doi:10.1029/ Water Research, 32 (12): 3581–3592, 1998, doi:10.1016/S0043-1354(98)00138-9.
2002WR001480. Vrugt, J. A., C. G. H. Diks, H. V. Gupta, W. Bouten, and J. M. Verstraten,
Monahan, A. H., J. C. Fyfe, M. H. P. Ambaum, D. B. Stephenson, and G. R. “Improved treatment of uncertainty in hydrologic modeling: Combining the
North, “Empirical Orthogonal Functions: The Medium is the Message,” strengths of global optimization and data assimilation,” Water Resources
Journal of Climate, 22 (24), 6501–6514, 2010. Research, 41: 1–17, 2005, doi:10.1029/2004WR003059.
Neuman, B. C., “Scale in distributed systems,” Readings in Distributed Vrugt, J. A., H. V. Gupta, S. C. Dekker, S. Sorooshian, T. Wagener, and W.
Computing Systems, IEEE Computer Society Press Los Alamitos, CA, 1994, Bouten, “Application of stochastic parameter optimization to the Sacramento
pp. 463–489. Soil Moisture Accounting model,” Journal of Hydrology, 325 (1–4), 288–307,
Nowak, W., S. Tenkleve, and O. A. Cirpka, “Efficient computation of linear- 2006, doi:10.1016/j.jhydrol.2005.10.041.
ized cross-covariance and auto-covariance matrices of interdependent Wikle, C. K., R. F. Milliff, D. Nychka, and L. M. Berliner, “Spatiotemporal
quantities,” Mathematical Geology, 35 (1): 53–66, 2003, doi:10.1023/ hierarchical Bayesian modeling: tropical ocean surface winds,” Journal of
A:1022365112368. American Statistical Association, 96 (454): 382–397, 2001,
Openshaw, S. and P. Taylor, “A million or so correlation coefficients: three doi:10.1198/016214501753168109.
experiments on the modifiable areal unit problem,” Statistical Applications in Wong, D. W. S., “The modifiable areal unit problem,” Quantitative
the Spatial Sciences, edited by Wrigley N, Pion, London, UK, 1979. Geography: A British View, 38 (1934): 60–70, 1981, doi:10.2307/2277827.
21_Singh_ch21_p21.1-21.12.indd 11 8/22/16 12:22 PM

Chapter 22
Calibration, Parameter
Estimation, Uncertainty, Data
Assimilation, Sensitivity
Analysis, and Validation
BY
EZIO TODINI AND DANIELA BIONDI
ABSTRACT allows either to estimate the parameters value if no assumptions on their

nature are made or to adjust their value in order to compensate for the
This chapter discusses the nature of the different sources of uncertainty and
unavoidable simplifications and errors in model and data, if the parameters
their effect on the parameters estimates. Techniques aimed at estimating the
are assumed to have physical meaning.
most appropriate parameter values are also introduced, from the simple trial
More recently, under the Bayesian influence, calibration has been extended
and error and the least squares approaches to more sophisticated Bayesian
from the original scope of finding a specific “optimal” value for the parame-
techniques involving assumptions on the prior distributions and the defini-
ters to the derivation of a parameter density (the “posterior” parameter den-
tion of appropriate likelihood functions.
sity) incorporating all the uncertainties descending from the model
Models are simplified representations of complex reality, observations are
assumptions, the required simplifications, and the data errors. Under this
affected by direct measurement or indirect estimation error, while scaling
approach, parameter estimation is then a way for assessing and reducing
from the infinitesimal to coarser finite scales introduces representation
model prediction uncertainty.
errors, etc. Therefore, model reproductions of natural phenomena are hardly
One important issue is to bear in mind the difference between calibrating
perfect and parameters are thus introduced as the knobs, which allow tuning
a model and estimating the “true” values of parameters (if they exist).
models thus improving their adequacy to observed realizations.
Calibrating, means tuning the parameter values in order to match the model
Chapter 22 discusses parameters nature, which can vary from purely sto-
output with observations, while estimating the parameters aiming at finding
chastic and uncertain to deterministic, linked to physically meaningful quan-
their true, albeit unknown value, means finding values that not necessarily
tities, but still distorted by the presence of model and measurement errors. In
will improve the model performances, due to the not necessarily correctness
all cases the presence of uncertainties and errors in models, data, initial and
of the model, the eventual nonlinear links between parameters and model
boundary conditions requires the use of appropriate techniques to estimate
predictions and measurement errors.
the expected value of parameters conditional to the chosen models and the
Therefore, an unbiased optimal parameter estimator will not necessarily
observations (deterministic estimation) or, in a Bayesian perspective, the
lead to a calibrated model.
posterior density of parameters (probabilistic estimation).
Calibration of parameters is sometimes confused with identification, which
Calibration and data assimilation techniques are introduced as means to
falls beyond the scope of this chapter. While calibration refers exclusively to
adjust, update or improve parameter estimated values. Assessment of simula-
parameters, identification also involves adjusting model structure. Frequently,
tion uncertainty, namely the uncertainty of the model output for any given
data-driven models are based on the linear/nonlinear combination of simple
known observation, is also presented together with the sensitivity analysis of
elementary structures (such as the linear channel and linear reservoir in Nash
models on the errors of different nature.
cascade (Nash, 1958, 1960), the elementary nonlinear reservoir in the Tank
model (Sugawara, 1967, 1995) and the nonlinear generic building block in the
22.1 INTRODUCTION hydrological models proposed by Fenicia et al. (2011) or the “perceptrons”
(Rosemblatt, 1958; Minsky and Papert, 1969) in the Artificial Neural Networks
Hydrological models are simplified representations of reality, combining models. Identification, which also involves calibration for any selected hypo-
prior knowledge assumptions to information extracted from observations. thetic structure, is then needed to find a parsimonious combination of the
From the very simple hydrological models, with no prior assumptions made, elementary models producing the most likely fit with observations.
to the more complex ones based on physically meaningful structural assump-
tions, they all require calibration.
22.2 PARAMETER UNCERTAINTY
Models are a combination of structure, describing the links among state
variables, and parameters. Parameters, according to the modeler conceptual- This section aims at clarifying the concept of uncertainty in parameter esti-
ization, have different meanings, ranging from simple model “tuning” knobs mates, which depends on several factors including the prior assumptions on
to physically meaningful quantities. the nature of the model (regression type, linear/nonlinear, structural,
Whichever the meaning, calibration is needed in order to match the model physically based, etc.), of parameters (undefined as in regression models or
output (simulation/prediction) to the observations. Traditional calibration physically meaningful as, for instance, the roughness coefficient in hydraulic
22-1
22_Singh_ch22_p22.1-22.20.indd 1 8/22/16 1:31 PM

22-2 Calibration, Parameter Estimation, Uncertainty, Data Assimilation, Sensitivity Analysis, and Validation
models), as well as the errors in observations and the estimation approach Ot + ∆t = Ot e−k∆t + (1 − e−k∆t )I + e t + ∆t (22.3)
[moments, least squares, maximum likelihood(ML), multicriteria, etc.] used
which can be transformed into a simple linear regression model by setting
in the parameter calibration phase.
a = e−k∆t:
22.2.1 The Nature of Models and Parameters
Ot + ∆t = I + a (Ot − I ) + e t + ∆t (22.4)
The nature of a model depends largely on the assumptions made by the mod-
eler on the basis of his prior knowledge. It is evident that in order to preserve the causative nature of Eq. (22.1), the
A wide range of modeling possibilities exist from the very simple linear value of a must be greater than zero and smaller than 1, because
regression models, only requiring a weak hypothesis on the probability distri- 0 ≤ e − k∆t ≤ 1; ∀k > 0, ∆t ≥ 0. Nonetheless, when measurement errors are large,
bution of errors, to the structural physically based models, where prior and due to the linearity assumption on the link between outflow and stored
hypotheses are made on the links and interactions among state variables and volume, the linear regression estimated value for a may even result negative
parameters. (α < 0) or larger than 1 (α > 1), because the range of existence of α in the
To better understand, let us start from the simple linear regression model, linear regression, if not explicitly imposed, lies between −∞ and +∞, thus loos-
which can be viewed as the estimation of the correlation coefficient among ing the physical (at least in terms of mass balance) nature of our model.
standardized Normal variables. When applying the linear regression model, Therefore, a way to impose the physicality on the parameter is to perform a
no additional prior assumptions are needed on causative relations between “constrained” linear regression, by introducing a number of constraints rep-
predictand, which is the variable of interest, and the predictor, the explana- resenting our prior assumptions in the estimation phase. In the case of the
tory variable used to infer its value. example, it is sufficient to add the constraint 0 < α < 1 to the minimization of
It is well known in statistics that correlation does not imply cause and the sum of squared errors (essentially the simple linear regression), by solving
effect. The most well-known example commonly used to show this concept is the following quadratic programming problem:
the storks and babies correlation estimated by Dr. Gustav Fischer on annual N 2
minα ∑ ti =1 [(Oti +∆t − I ) − α (Oti − I )] (22.5)
counts made 1930–1936 in Oldenburg Germany (Fischer, 1936, 1938, 1940),  s .t . 0 < α < 1
reproduced in Fig. 22.1, which resulted in R = 0.94. 
where N is the number of observed data.
80 Following the same concept, when estimating the Unit Hydrograph ordi-
nates, Natale and Todini (1976a, b) showed that physically meaningful
constraints, aimed at imposing mass balance between inflow and outflow (the
75 sum of all the ordinates of the Unit Hydrograph, must equal 1) and that water
flows downstream (all the ordinates must be non-negative), reduce the range
of existence parameters, thus leading to sharper parameter estimates.
70 Needless to say that introducing constraints in the estimation requires a
certain care otherwise when the imposed constraints are incorrect, the
estimation variance will still be reduced at the cost of increasingly higher biases.
Babies
65 The aforementioned concepts can be broaden by introducing in the estima-

tion phase all the available prior knowledge, either in the form of prior prob-
ability densities assumptions, as advocated by Bayesian Inference, or by
60 introducing physically meaningful constraints in order to obtain sharper
y = 0.1506x + 36.231 parameter estimates.
R2 = 0.88374 The case of the Muskingum model can be used to illustrate this point.
55 The Muskingum routing model was developed by McCarthy (1938, 1940)
as a linear model with constant parameters, which fixed value had to be esti-
mated from input and output observations. Similarly to the linear reservoir
50 case presented earlier, the model was physically based in terms of mass bal-
130 150 170 190 210 230 250
ance, while approximated in terms of the momentum balance. Cunge (1969)
Storks found a way of linking the Muskingum original parameters to physical quan-
Figure 22.1 Regression between annual number of babies born and annual number tities (namely, discharge, water stage, celerity, surface width, friction slope,
of storks seen in Oldenburg in the years 1930–1936. time, and space integration steps) and, at the same time, to better represent
the momentum balance equation leading to a parabolic flood routing model.
In order to estimate the previously mentioned physical quantities, the
In the example, the number of babies born, the predictand, is well pre-
Muskingu–Cunge model requires more information than the original one,
dicted on the basis of the number of storks, the predictor, although no cause
such as the cross section geometry and the friction coefficient, but, although
and effect can be implied.
calibration of the friction coefficient (the unique physically meaningful
Causative relations (namely, cause and effect, the basic assumption in most
parameter) is still needed, the model performs better and where calibration is
of hydrological and hydraulic models), always descend from the axiomatic
not available can still be run from literature-based friction coefficient values.
hypothesis of a continuum (a continuum of mass, energy, relation, etc.).
Summarizing, the physical nature and the structure of a model depends on
To clarify this point, let us apply the linear regression model to a more
the amount of information the modeler has introduced in terms of prior
physically based (and therefore, causative) example of a linear reservoir.
knowledge of the phenomena under study. Similarly, with the increasing
In this case, we “a priori” assume a continuum of mass reflected into the
physical meaning of model equations, also parameters may be assumed to
mass balance equation:
retain physical meaning, which allows to derive first guess estimates from
dV available information and to reduce their domain of existence, potentially
= I − O (22.1) leading to sharper parameter or parameter probability densities estimates.
dt
in which the change in time of the storage V is linked to I, the inflow, and O, 22.2.2 Sources of Uncertainty
the outflow. It is evident from Eq. (22.1) that our prior assumption on a con- There are several sources of uncertainty affecting the parameter estimates.
tinuum, makes the inflow causative on the outflow. The first source of uncertainty is due to the model itself because a model is
Moreover, let us assume the outflow O to be proportional to the stored always a schematization and a simplification of a more complex reality. The
volume, namely, O = kV, through a parameter k > 0, which is a first-order used schematizations and simplifications never allow reproducing reality at
approximation of a hydraulic correct momentum-balance equation. 100% and the discrepancy will resort into differences (eM ) between observed
If I is constant in time, the solution to Eq. (22.1), an ordinary differential and modeled quantities, affecting the estimated parameter values.
equation, is A second source of uncertainty is due to the input (e I ) and output (e O )
observation errors. Sometimes, these errors may be negligible, but most of the
Ot + ∆t = Ot e−k∆t + (1 − e−k∆t)I (22.2) times, they are large strongly influencing the parameter estimates. Given the
If one wants to estimate parameter k from observations, measurement and tendency of estimating parameters on the basis of deviations of model output
model-schematization errors must also be taken into account, which can be from the observations, input errors and output errors may have substantially
done by adding an error term et + ∆t to the previous equation, leading to: different effects on parameter estimates.
22_Singh_ch22_p22.1-22.20.indd 2 8/22/16 1:31 PM

Parameter estimation 22-3
Errors in initial and boundary conditions assumptions (ε IC ; ε BC ) also affect In its simplest formulation, LS allows estimating the parameters of a simple
many structural models, and in particular, the nonlinear chaotic models (such or a complex model by minimizing the sum of squares of deviations between
as most of the meteorological and climatological models) resorting into large observed and model predicted values.
uncertainties affecting the determination of the parameter values. The Ordinary Least Squares
In addition to all these sources of uncertainty converging into the
parameter-estimation process, the final parameter estimate values will be Given a vector of observations Y ≡ [ y1 , y 2 , …, yn ] (water levels, discharges,
unavoidably affected by the selected estimation procedure (ε ES ). etc.) with n, the number of observations, using a model M ( X , ϑ ), a function
of known and/or observed quantities X and of a set of unknown (or at least
22.2.3 Parameter Estimation: A Way for Assessing and uncertain) parameter values ϑ, it is possible to simulate the observations by
Reducing Uncertainty producing a corresponding vector of forecasts, namely, Y ≡ [ y , y , …, y ].
1 2 n
As stated in the introduction, model parameters can either be viewed as  is a product of the model M ( X ,ϑ ), it will be different
Given that Y for any
unknown quantities of interest to be determined from observations, or as
k 
given realization ϑ of the set of parameters, namely, Y(ϑ ). The least-square
k
tuning knobs essentially aimed at calibrating the model. In both cases, given approach allows to estimate the model parameters by minimizing the follow-
that we do not know their true value (if it exists) parameters must be viewed  (ϑ), namely:
ing function of the residuals d(ϑ) = Y − Y
{ }
as uncertain variables strongly affected by all the errors described earlier.
ϑ LS = minϑ J[ε (ϑ )] = minϑ J[Y − Y (ϑ )] = min 1 ∑n [ y − y (ϑ )]2 (22.7)
ϑ i =1 i
ϑ {ε M , ε I , ε O , ε IC , ε BC , ε ES } (22.6) n i
If the scope is to find the true albeit unknown value of parameters, the In principle, to just estimate a set of parameters, one does not need to make
estimation must aim at deriving unbiased minimum variance estimates of the assumptions when using the LS approach. Nonetheless, if one wants to be able
parameters, starting from assumptions on the probability distribution of of evaluating the properties of the estimator or drawing confidence limits
errors and, possibly, on the description of a prior probability density of the around the predicted values, a number of assumptions on the nature of the
parameter values. residuals are required. The classical assumption, which makes the least
On the contrary, if the scope is to calibrate a model, the estimation must squares identical to the simple linear regression, is that d are assumed to be
aim at deriving unbiased minimum variance estimates of the model output, weakly stationary (namely, stationary up to the second-order moment), inde-
and not necessarily of the parameters still starting from assumptions on the pendent and Normally distributed random variables. Under this assumption,
probability distribution of errors and, following a Bayesian approach, on the the least squares approach, called the ordinary least squares (OLS) approach,
description of a prior probability density of the parameter values. is derived as the minimum of the trace of the covariance assuming a multi-
Finally, given that the scope for building and using models is to improve variate homoscedastic Normal distribution
available knowledge allowing taking better decisions, a third, more rewarding
objective of calibration, known in the literature as Bayesian Inference, is to

ϑ
{ }
ϑ OLS = min trace E ( ε − µ ε )( ε − µ ε )T = min trace Σ
ϑ
(22.8)
find the full posterior density of parameters, not limiting the process to the
estimation of their expected value, or sometimes their modal one. In this { }
with le the mean of d and ¬ =E ( ε − µ ε )( ε − µ ε )T = σ ε2I, its covariance
matrix, a diagonal matrix with constant variances under homoscedasticity
approach, regardless to their potential physical nature, parameters are consid-
assumptions, which leads to:
ered as the final receptors of all the uncertainty and the scope is to improve
the description of their prior probability density using the observations. The (ε − µε ) (ε − µε ) T
resulting sharper posterior probability density is then used, by means of the ϑ OLS = min trace Σ = min = min σ ε2 (22.9)
ϑ ϑ n ϑ
model, to compute predictive uncertainty as well as expected benefits and
costs affecting decisions, as described in Chap. 26. The Generalized Least Squares
Frequently, the assumptions of weak stationarity and independence do not
22.3 PARAMETER ESTIMATION
hold, which implies the need for extending OLS to the heteroscedastic case,
This section describes the standard techniques commonly used for estimating where ¬ is a full symmetrical matrix. In theory, minimizing the trace of the
parameters of hydrological models, such as the maximum likelihood, the least correlation matrix, can do it:
squares, and the more recent multicriteria approaches. All these approaches
( ε − µ ε )T Σ ε−1 ( ε − µ ε )
aim at finding either an expected value or a modal value for the parameters.
In addition, Bayesian approaches will be described, which aim at inferring the ϑ
{
min trace E ( ε − µ ε )Σ ε−1 ( ε − µ ε )T = min
ϑ
} n
(22.10)
full posterior probability density of parameters to be used in the assessment In practice, this approach, which is called the generalized least squares
of predictive uncertainty and in the estimation of the expected utility in deci- (GLS) approach, cannot be directly solved because, while trace Σ ε = nσ ε2 can
sion making (see Chap. 26). No mention is made of other approaches, such as be estimated under the assumption of ergodicity, this is not the case for the
moments and cumulants (Dooge, 1973), etc., which were used in the pre- and full matrix. To solve the GLS problem, special approaches have been devised,
early computer era to estimate parameters of linear hydrological models, such such as the feasible generalized least squares (FGLS). In the first stage, one
as the Muskingum or the Nash cascade. Today, moments, cumulants, proba- solves the OLS problem. Using the OLS-estimated parameters, ϑ , one can
bility weighted moments, and L-moments are essentially used to directly produce a set of residuals ε (ϑ ) = Y − Y
(ϑ ) from which one can iteratively
estimate parameters of probability distribution functions. obtain an approximated estimate of ¬.
The main properties required in an estimator are: consistency, unbiased- Please note that the GLS approach corresponds to maximizing the natural
ness, efficiency, robustness, and minimum variance. logarithm of probability of d under the assumption of a multi-Normal distri-
• Consistency: An estimator is consistent if the estimator converges to the bution.
true value almost surely as the number of observations approaches infinity.
• Unbiasedness: An estimator is unbiased if its expected value is equal to 1
− ( ε − µ ε )T Σε−1 ( ε − µ ε )
e 2
the true value. ϑ GLS = max log = min( ε − µ ε )T Σ ε−1 ( ε − µ ε ) (22.11)
• Efficiency: An estimator is efficient if it produces the smallest error ϑ (2π )n/2 || Σ ε ||1/2 ϑ
covariance matrix among all unbiased estimators, it is also regarded apart from constants terms that are not relevant in maximization.
optimally using the information in the measurements. A well-known
efficiency criterion is the Cramér–Rao (Rao, 1945; Cramér, 1946) lower The Constrained Least Squares
bound. As described in Sec. 22.2.1, there are several cases for which, unacceptable
• Robustness: An estimator is robust if it is insensitive to the gross mea- parameter estimates can result, as for instance, when the numerosity of the
surement errors and the uncertainties of the model. sample is rather small and at the same time, the variance of errors is large. It
• Minimal variance: Variance reduction is the central issue of various is therefore necessary to reduce the domain of existence of parameter values
Monte Carlo approximation methods, most improvement techniques are by introducing constraints describing, in the estimation phase, physically
variance-reduction oriented. meaningful ranges or regularity conditions.
An example of the constrained linear systems (CLS) approach is given for
22.3.1 The Least-Squares Estimator the estimation of unit hydrograph ordinates, which may result unstable, oscil-
The introduction of the least squares (LS) approach is generally attributed to lating, negative, and generally not preserving the mass balance condition
Carl Friederich Gauss (1809), who used it in the determination of planetary (namely, that their sum must equal 1), when OLS or GLS approaches are
orbits. taken. Several researchers dealt with the problem, as for instance, Eagleson
22_Singh_ch22_p22.1-22.20.indd 3 8/22/16 1:31 PM

et al. (1966) who introduced non-negativity constraints by means of linear Under a set of regularity conditions (Newey and McFadden, 1994) and
programming. Few years later, Natale and Todini (1976a,b), followed by Bree provided that the Likelihood function is a continuous function twice differ-
(1978) and Bruen and Dooge (1984), introduced non-negativity constraints, entiable in the domain of existence of the observations, which, in the case of
mass balance constraints, and regularity conditions in the estimation of unit hydrological models, as previously mentioned, are the residuals d(ϑ), MLE
hydrograph ordinates. A quadratic programming approach was effectively shows interesting properties.
used to solve the constrained least squares problem and the reduction of It can be proven that, under the aforementioned conditions, the MLE is a
estimation variance was shown both analytically and numerically by Natale consistent, efficient, and asymptotically Normal estimator. Consistency
and Todini (1976a, b). means that the estimated value will tend to the true value when the numeros-
As a concluding remark, linear or nonlinear constrained minimization ity of the sample tends to infinity; efficiency is the property for which the
approaches should be used whenever the numerosity of the available data covariance of the MLE, which can be estimated by using Fisher Information
sample is insufficient to allow for reliable parameter estimates and/or the vari- or Fisher Information Matrix in the multivariate parameter case (Lehmann
ance of the difference between observed and computed values is quite large. and Casella, 1998), reaches the Cramér–Rao lower bound (Rao, 1945;
Cramér, 1946) when the sample size tends to infinity; asymptotic normality
22.3.2 The Maximum Likelihood Estimator implies that the MLE tends to the Gaussian distribution with mean the true
The maximum likelihood estimator (MLE) was introduced by Fisher (1922) value and covariance matrix equal to the inverse of the Fisher Information
as an alternative approach to the method of least squares. MLE allows esti- Matrix (Lehmann and Casella, 1998).
mating parameters of a probability distribution given a sample of observations.
22.3.3 The Multicriteria Approaches
Given a sample of n-independent observations ( x1 , x 2 , …, xn ), extracted
from a population Ω x with probability density function f x ( x ϑ ), where ϑ are As discussed in the previous sections dealing with least square or ML
the relevant parameters, the probability density of the sample can be com- approaches, parameter estimation is essentially based on the minimization or
puted as: maximization of a convex function of departures of the modeled values from
n
the observed ones, with the objective of best reproducing observations using
f ( x1 , x 2 , …, xn ϑ ) = f x ( x1 ϑ ) × f x ( x1 ϑ ) ×… × f x ( xn ϑ ) = ∏ i =1 f x ( xi ϑ ) (22.12) the model. Sometimes, when dealing with lumped or semidistributed hydro-
Equation (22.12) shows that the probability of observing a specific sample logical models, the objective of parameter estimation requires reflecting other
of length n can be estimated conditionally on the “knowledge” of the parame- issues, such as the timing of the peak flow; improved performances in repro-
ters ϑ, namely, on the assumption that ϑ values are known. The Likelihood ducing low, medium, or high flows; better preservation of mass balance; etc.
function was introduced by looking at f ( x1 , x 2 , …, xn ϑ ) the other way round, (Gupta et al., 1998). This has motivated multicriteria calibration approaches
namely, by interpreting Eq. (22.12) as the likelihood of observing a specific in which multiple sets of observations and/or multiple evaluation criteria are
value for the parameters once the value of the sample is known. In other words: employed (Gupta et al., 1998; Legates and McCabe, 1999; Madsen, 2000; Yapo
et al., 1998). Moreover, in the case of fully distributed physically based mod-
(ϑ ; x1 , x 2 , …, xn ) = f ( x1 , x 2 , …, xn ϑ ) (22.13) els, measurements are not limited to the flow at the gauged sections; several
additional state variables measured in space and time are also available
The MLE of the parameter values is then obtained by maximizing the like- (Refsgaard, 1997), such as snowpack elevation and water equivalent, soil
lihood of observing them conditionally upon the observed sample. moisture, groundwater table elevation, areas at saturation, etc. Assimilation of
ϑ MLE = max ϑ (ϑ ; x1 , x 2 , …, xn ) (22.14) all these additional data may be appropriate to better reflect the internal
physical nature of the state variables. Accordingly, the new calibration criteria
Please note that formally, under a Bayesian inferential approach, Eq. (22.13) are either introduced into the objective function assigning them individual
is not fully correct. A proper relation should in fact be derived from the Bayes weights and aggregated into a single measure to be maximized (or mini-
theorem as: mized) using single-objective optimization procedure or compromised by
searching for Pareto-optimal, namely boundary nondominated, solutions
f ( x1 , x 2 , …, xn ϑ ) f (ϑ )
'(ϑ ; x1 , x 2 , …, xn ) = f (ϑ x1 , x 2 , …, xn ) = (22.15) (Pareto, 1896–1897, 1906) through evolutionary multiobjective optimization
f ( x1 , x 2 , …, xn ) (EMO) algorithms. Application of the single-objective optimization can be
In Eq. (22.15), while the denominator is independent from the parameters found in Madsen (2000), who used the aggregation (or scalarization)
and can be neglected in view of the maximization, this does not apply to the approach together with the SCE-UA single-objective optimization algorithm
density of parameters f (ϑ), which unfortunately is unknown. In practice, the developed by Duan et al. (1992). Several applications of Pareto-optimal com-
validity of Eq. (22.13) is introduced by assuming f (ϑ) as a uniform probability promise solutions can be found in the literature using the NSGA-II
density (in other words, we assume a “noninformative prior” on the parame- (Nondominated Sorted Genetic Algorithm-II due to Deb et al., 2002), the
ters), therefore, independent from the value of ϑ, and the Likelihood can be Multiobjective Complex Evolution (MOCOM-UA due to Yapo et al., 1998),
viewed as proportional to the density of the parameters given the sample: the Multiobjective Shuffled Complex Evolution Metropolis (MOSCEM) algo-
rithm (Vrugt et al., 2003), and the AMALGAM (A Multi-ALgorithm
(ϑ ; x1 , x 2 , …, xn ) ∝ f (ϑ x1 , x 2 , …, xn ) (22.16) Genetically Adaptive Multiobjective) method (Vrugt and Robinson, 2007). In
Given the multiplicative nature of the Likelihood, it is common practice to the latter case, a set of Pareto-optimal solutions is initially found, while the
minimize its natural logarithm (the Log-Likelihood): most satisfactory solution is successively selected either subjectively or by
maximizing an appropriate (subjective) utility function, not necessarily an
ϑ MLE = max ϑ log L(ϑ ; x1 , x 2 , …, xn ) = max Σni=1 log f x ( xi ϑ ) (22.17) economic benefit/cost function, but rather a function expressing the decision
ϑ maker’s perception of benefits, costs, or risk (see Chap. 26 for a more detailed
Please note that, when the MLE is not directly applied to the estimation of definition).
parameters of a probability distribution, but rather to estimate the parameters Several examples of application of multicriteria approaches in calibration
of a model, such as a hydrological or a hydraulic model, one must convert the and validation of hydrological models can be found in the literature (Duan et
observations into variables for which a probability distribution can be al., 1992; Seibert, 2000; Beldring, 2002; Boyle et al., 2003; Vrugt et al., 2003),
assumed. while an extensive list of multicriteria and Pareto-optimal applications can be
For instance, one can proceed at defining a vector of residuals found in Efstratiadis and Koutsoyiannis (2010).
 (ϑ ), namely, the differences between the observed quantities of
ε (ϑ ) = Y − Y
interest and the modeled ones, similarly to what described in Sec. 22.3.1 for 22.3.4 The Bayesian Inferential Approaches
the LS estimator. In order to introduce the Bayesian Inferential approaches, it is worthwhile
If the observations are not independent, the Likelihood can still be defined quoting Draper and Krnjajić (2013):
pending the availability of an expression for the multivariate joint density of
the sample. “In the Bayesian modeling paradigm for inference, prediction and decision-making, there
are three fundamental ingredients: ϑ, something unknown or only partially known to You
For instance, if the sample is distributed according to a multi-Normal den-
(a generic person wishing to reason sensibly in the face of uncertainty; D, an information
sity, then the MLE becomes (data) source that You judge to be relevant to decreasing Your uncertainty about ϑ; and B”
1
− ( ε − µ ε )T Σε−1 ( ε − µ ε )
(please note that B may correspond to a hydrological model and/or to) “a set of proposi-
e tions (true-false statements) summarizing Your background assumptions and judgments
ϑ MLE = max log 2 n/2 = min( ε − µ ε )T Σ ε−1 ( ε − µ ε ) (22.18)
ϑ (2π ) || Σ ε ||1/2 ϑ about relevant aspects of ϑ (e.g., that ϑ ≥ 0 if ϑ represents the mean remission time for a
specified set of patients with a given disease) and D (e.g., that the data set arose as the result
which, as can be noticed, coincides with the GLS estimator. of a randomized controlled trial with the following design: ...).
22_Singh_ch22_p22.1-22.20.indd 4 8/22/16 1:31 PM

Data assimilation 22-5
In this framework the physicist R. T. Cox (1946, 1961) (also see Jaynes, 2003) began with a Decision Theory (Berger, 1985; Bernardo and Smith, 1994; De Groot, 2004),
conditional plausibility operator p(A\B) acting on propositions A and B and calibration and as discussed earlier by Draper and Krnjajić (2013) decisions should be
results for Bayesian model specification developed from this a full probability calculus in taken on the basis of an expected utility E{U ( y )}, which can be a combination
which, for example, propositions such as ϑ ≥ t for real-valued ϑ bring the usual machinery
of expected benefits and expected losses. U(y) can be estimated only if the
of cumulative distribution functions and densities to bear on the process of uncertainty
quantification. probability density function of the conditioning variable y is known. For
instance, y can be the water level to be compared to a threshold value above
Cox proved that—under a reasonable set of axioms involving internal logical consistency which losses are no more null. Since both in planning and management y is
and representation of degrees of plausibility by real numbers Your uncertainty quantifica- generally unknown, the expected value of the utility function can be defined as:
tion, if You wish to be rational, should be based (for inference and prediction) on the con- +∞
ditional probability distributions p(ϑ|B) and p(D|ϑ, B); for decision-making, it had E {U ( y )} = ∫ U ( y ) f ( y )d y (22.20)
0
previously been shown by Ramsey (1926) that the only other relevant ingredients are a set
A of possible actions a and a utility function U(a, ϑ*), expressing (in real-valued terms) the where f (y) is the probability density expressing our incomplete knowledge (in
gain that would result if You chose action a and the unknown ϑ actually took the value ϑ*. other words, our uncertainty) on the value that will actually occur.
In this formulation p(ϑ|B)—usually called Your prior distribution, although temporal In order to assess the value of y, we make a model providing an estimate for
considerations need not arise in specifying it—quantifies all of Your information about ϑ y, namely, ŷ, which hopefully will be less uncertain than y, although still not
external to D, and p(D|ϑ, B)—typically referred to as Your sampling distribution, even fully known. Following the Bayesian approach ŷ can be expressed as a function
though it is not necessary in this approach to consider other data sets that might have been of the parameters, namely, ŷ (ϑ) and, as described in Sec. 22.3.4 a posterior
observed but were not—quantifies Your predictive uncertainty about D given ϑ, before D density for the parameters can be assessed.
has arrived. Note that Cox’s use of the phrase logical consistency has nothing to do with the
repeated-sampling idea of asymptotic consistency ….”
Moreover, following what is described in Chap. 26, the conditional density
of y given the model predicted value ŷ, namely, f (y|ŷ (ϑ)) can be derived using
“It then follows from Cox’s and Ramsey’s results in this framework that the three basic sta-
one of the available predictive uncertainty postprocessors (Chap. 26, Sec. 26.6)
tistical activities of inference, prediction and decision-making are each governed by a single and used to estimate an expected predictive density:
equation (familiar in Bayesian work), as follows:
E[ f { y | y (ϑ )}] = ∫ f [ y | y (ϑ ) ] f "(ϑ | D , B)d ϑ (22.21)
• ( inference) p(ϑ D , B) = c p(ϑ B) p(D ϑ , B), in which c is a positive normal- Ωϑ
izing constant and p(ϑ D , B)—usually called Your posterior distribution, with Ωϑ the domain of existence of the parameters and the explicit dependence
although again temporal considerations are not central to the formula- on D and B has been omitted for the sake of clarity. Note that the integral
tion—quantifies the totality of Your information about ϑ, both internal appearing in Eq. (22.21) is usually evaluated via Monte Carlo sampling (see
and external to D; Chap. 23), to overcome the analytical complexity.
• (prediction) p(D D , B) = ∫ p(D ϑ , D , B) p(ϑ D , B)dϑ , where D* is Using the expected predictive density of Eq. (22.21), one can then reliably
∗ ∗
Θ estimate the expected value of the utility function E{U ( y )} to be used within
future data” (please note with reference to Chap. 27: not the model fore- the decision-making framework:
casted quantity but rather the real unknown one) “and Q is the space over
+∞
which You acknowledge Your uncertainty about ϑ. Often D provides no E{U ( y )} = ∫ U ( y )E [ f { y | y (ϑ )}]d y (22.22)
0
additional information about D* if ϑ is known, in which case this equation
simplifies to p(D ∗ D , B) = ∫ p(D ∗ ϑ , B) p(ϑ D , B)dϑ and involves only The alternative approach where an optimal set of parameters ϑ* is estimated
Θ instead of the full posterior parameter density, namely:
previously-defined quantities; and
• (decision) The optimal action a* is given by +∞
E{U ( y )} = ∫ U ( y ) f [ y | y (ϑ ∗ )]d y (22.23)
0
a∗ = (a∗ D , B) = arg max a∈A E(ϑ D ,B ) {U (a,ϑ )} = arg max a∈A ∫ U (a,ϑ ) p(ϑ D , B)dϑ leads to a different result, because, unavoidably:
Θ
(22.19)
in other words, a* is the action that maximizes expected utility, where the expectation is E[ f { y | y(ϑ )}] = ∫ f [ y | y (ϑ )] f ′′(ϑ | D , B)d ϑ ≠ f [ y | y (ϑ )] (22.24)
Ωϑ
over Your posterior uncertainty about ϑ given D.” This is what hydrologists tend to do in practice, and, as shown by Liu et al.
The described Bayesian approach shows that if one wants to improve the (2005), due to the integral nature of Eq. (22.23), it can often be used in deci-
effectiveness of prediction as well as the reliability of decisions, it is not suffi- sion-making, particularly when using structural and physics-based models,
cient to estimate a specific “optimal value” ϑ* for the parameters in the infer- provided one can show that the difference between the two expressions is not
ential problem, but it is rather necessary to assess the full parameter large, namely,
probability distribution (or density) in the form of the “posterior” probability
density of parameters p(ϑ D , B) given a set of propositions B (usually your E[ f { y | y (ϑ )}] = ∫ f [ y | y (ϑ )] f ′′(ϑ D , B)d ϑ ≈ f [ y | y (ϑ ∗ )] (22.25)
Ωϑ
model and/or physical constraints) and a set of observations D. The “poste-
rior” probability density can then be used to estimate the expected predictive 22.3.6 Choice of the Most Appropriate Parameter
density as well as the expected utility function value to be optimized (see Estimation Approach
Chap. 26). It is rather difficult to provide guidelines on which is the most appropriate
22.3.5 Parameter Estimation vs. Posterior Parameter
parameter estimation tool to be used. This will depend at the same time on
Densities Estimation the availability of information on the quantitative description of the phenom-
enon to be represented (structural equations, time-independent and time-
Why one should spend larger amount of computational efforts in order to dependent data describing the spatial domain, the initial and boundary
estimate the full posterior density of parameters f ′′(ϑ D , B) (with B a model conditions, the forcing input functions, etc.) as well as the quality of results
and/or a set of propositions and D the set of observations used to assess the one would like to obtain.
posterior density), instead of estimating a single “optimal” set ϑ* is a major Simple approaches with few available input-output data should restrict
aspects to be clarified. To do so, let us consider that even in the case where themselves to OLS approaches. If long series of input-output data are avail-
perfect values for ϑ exist, due to a correct physical interpretation of the phe- able, one can apply GLS and ML approaches.
nomenon under study, all the errors floating around (see Sec. 22.2) will Multicriteria approaches can be used with more complex structural models
unavoidably distort their estimated value ϑ . This is the classical nature of the when the approximations in the internal model links can distort parameter
“inverse problem,” where even small perturbations in the observations can estimates.
generate large errors in the estimated parameter values. Nonetheless, whenever important decisions are linked to the modeling
Moreover, not necessarily perfect values for ϑ exist, due to the inaccuracies exercise, it is essential that one resorts to the Bayesian approaches in order to
in the description of phenomena or the coarseness of the model used, which derive a representation of the full density of parameters and not only an
will make the optimal parameter values vary randomly or according to spe- expected or a modal value.
cific situations.
In other words, if not the parameters, at least the parameter estimates must
22.4 DATA ASSIMILATION
be viewed as random variables, which best description is uniquely provided
by the full probability density f ′′(ϑ D , B). A part of the challenge for every modeling/forecasting problem is the combi-
Bearing in mind that the most common scope for using models, both in the nation of the knowledge embodied in a model and of the information that can
planning and in the management phases, is to improve decisions, following be gained from independent observations to find the best representation of
22_Singh_ch22_p22.1-22.20.indd 5 8/22/16 1:31 PM

the dynamic behavior of a system. Both model predictions and observations In most of DA techniques, to render the problem mathematically treatable,
are imperfect and affected by uncertainty, but an appropriate merging has these equations are linearized introducing the tangent linear operators Mt and
proved to synergistically provide more accurate results, that cannot be Ht:
obtained when used individually (McLaughlin, 2002; Liu and Gupta, 2007;
Reichle, 2008; Liu et al., 2012). Charney et al. (1969) first suggested combin- x t = Mt x t −1 + w t (22.28)
ing current and past data in a meteorological model, using the model’s equa-
tions to provide continuity in time and dynamic coupling amongst the z t = Ht x t + v t (22.29)
meteorological fields. From this approach, several techniques, known as data Moreover, assumptions on the statistical properties of the errors terms are
assimilation (DA), have developed in meteorology, hydrology, and other also made: many approximations consider terms wt, and vt as zero-mean
fields. white noise (namely, uncorrelated in time) sequences with Gaussian probabil-
In the broader sense, DA is the process with which the information content ity distribution, characterized by covariance matrices Qt and Rt respectively,
of observations (data) is transferred into other data records, into model input and null cross covariances.
state variables, parameter estimates, or model output. DA is also highly
related to calibration, but, when dealing with models, it is generally applied to 22.4.2 Variational Data Assimilation
already “calibrated” models in order to update state estimates or to adjust the The aim of variational data assimilation (VDA) techniques is to find the best
model performance in accordance to the latest measured quantities. fit, in the least-squares sense, between selected output quantities of the model
For instance, one may refer to DA when infilling data record gaps or when and their observed equivalent, by minimizing an objective function or cost
setting up the input state variables of a model, such as the snow cover or the function J. The cost function J quantifies an aggregate of errors (assumed
soil moisture content, based on recent point or spatial measurements. DA is independents and additive at different time), each being weighted by the
also directly used to update model output to match the latest observations or corresponding error covariance matrix. Considering a time-independent
to improve model parameter estimates. problem, the objective function J can be written as:
In the case of record gap infilling, the DA problem consists in using the
information content of the surrounding measurement records to estimate the J ( x ) = 1 / 2( x − x b )T B −1 ( x − x b ) + 1 / 2(z − z )T R −1 (z − z ) (22.30)
expected value of the missing observations, possibly using approaches leading where xb refers to a “background” estimate of the state vector, for example, the
to minimum variance estimates. DA techniques were firstly developed and result from a previous assimilation analysis, characterized by a background
used operationally in meteorology (Daley, 1991) and oceanography (Bennet, error covariance matrix B. In practice, xb usually corresponds to the “first
1992). The application of DA in hydrology is more recent, although in dra- guess,” that is, the point used to initiate the minimization procedure.
matic growth under the motivation in understanding and possibly reducing VDA can thus be considered simply as an optimization problem, where the
the uncertainty involved in hydrologic modeling, but mostly because of the state vector x is “calibrated,” that is, the variable with respect to which the cost
increasing availability of satellite observation at higher spatiotemporal resolu- function is minimized. When the vector of model-predicted observations ẑ is
tions. Hydrological applications range from addressing the assimilation of calculated through a linear observation operator H, J is a quadratic functional
traditional in situ observations, such as stage, discharge, rainfall, soil moisture, of x with a unique minimum and provides a solution which is equivalent to the
and snowpack measurements (e.g., Todini et al., 1976; Todini and Wallis, BLUE (best linear unbiased estimator) result (see Sec. 22.4.3).
1978; Kitanidis and Bras,1980a, b; Walker and Houser, 2001; Young, 2002; Seo Depending on the spatial dimension of the state variable, VDA methods
et al., 2003, 2009; Vrugt et al., 2005; Clark et al., 2008) to incorporating can be one-dimensional (1D-Var), two-dimensional (2D-Var), and three-
remotely sensed retrievals of various quantities, such as soil moisture, land dimensional (3D-Var). Introducing the time dimension into the assimilation
surface temperature, snow water equivalent, and/or snow cover area or extent, scheme provides the four-dimensional variational assimilation (4D-Var, 3D
surface water level (Entekhabi et al., 1994; Houser et al., 1998; Castelli, 1999; in space + 1D in time). Figure 22.2 shows a schematic representation of
Galantowicz et al., 1999; Walker et al., 2001; Boni et al., 2001; Brocca et al., 3D-Var and 4D-Var.
2010; Reichle et al., 2010; Rodell and Houser, 2004; Giustarini et al., 2011; De The optimization objective function can be reformulated to handle obser-
Lannoy et al., 2012, among others). vations at several different times within a prescribed assimilation window, as
Although, most DA applications were meant for updating model dynamic follows:
states, stochastic DA techniques have been also developed and applied to
joint-state-parameter estimation (Todini, 1978a, b; Thiemann et al., 2001; J ( x ) = 1 / 2( x − x b )T B −1 ( x − x b ) + 1 / 2 Σ N (z − z t )T R −1 (z − z t ) (22.31)
o o o o o t =0 t t t
Moradkhani et al., 2005a, b; Labarre et al., 2006), by coupling state DA with
where the subscript refers to the time, varying from the initial time to N the
parameter calibration techniques (Vrugt et al., 2005) either using a technique
number of observation time points within the assimilation interval. The
known as “state augmentation,” that is, essentially extending state vector with
4D-Var is a generalization of 3D-Var, where the control variable is the state
the model parameters (e.g., Gelb, 1974) or by adding a second Kalman Filter
vector xo at the beginning of each assimilation window that is calibrated to
(Todini, 1978a, b; Labarre et al., 2006) or an Ensemble Kalman Filter
allow also for the best fit between the model trajectory and the observations
(Moradkhani et al., 2005a) in the space of parameters.
over the assimilation interval. In this sense, the 4D-Var is a dynamic observer
Anyway, the focus of DA is largely on model state updating, referring to the
method, well suited for smoothing problems.
problem of assimilating observations that can be distributed in space and in
It is worth mentioning that in the 4D-Var, the model equation is imposed
time.
as a strong constraint in the minimization of the objective function, that is,
There are two basic approaches to DA, variational and sequential tech-
the sequence of model states xt must be a solution of the model equation,
niques (Bouttier and Courtier, 1999), which essentially differ in their numer-
which is considered to be perfect. A straightforward and elegant way to incor-
ical cost, their optimality, and their suitability for real-time data assimilation.
porate the model constraint in the cost function is to introduce Lagrangian
22.4.1 State-Space Formulation
multipliers.
In application involving high-dimensional and nonlinear models, mini-
To illustrate the basic concept of data assimilation, let us consider a generic, mizing the objective function in Eq. (22.31) is not analytically feasible, and
typically nonlinear, dynamic state-space formulation of a stochastic model, advanced numerical (i.e., variational) methods are required. The solution is
that predicts the system state vector xt at time t as a function of the system sought iteratively, by performing several evaluations of the cost function and
state vector estimate at previous time steps (for the sake of clarity, the forcing, of its gradient ∇J in order to approach the minimum using suitable descent
or control, vector ut, and model parameters ϑ are not included). methods (steepest descent, conjugate gradient, or quasi-Newton method).
x t = M( x t −1 ) + w t (22.26) Nevertheless, to obtain a treatable minimization problem, especially because
of the calculation of the second term of J, simplifications and approximations
where M() is a linear or nonlinear model operator and wt a model error vec- are needed, like linearizing the state and observation equations. With these
tor. Bearing in mind that there is not always an immediate relationship premises, the adjoint model technique is typically used as it permits efficient
between the state variables and the observations, the following equation pro- calculation of the gradient of the cost function and is proved to be large-scale
vides a general map that relates states to observations zt: efficient (e.g., Huang and Yang, 1996).
z t = H( x t ) + v t (22.27) KF and 4D-Var are valid and produce equivalent results at the end of the

interval if model operator M and observation operator H are linear (namely,
where H() is a linear or nonlinear observation operator, and vt the observation M and H), whereas the covariance propagation in 4D-Var is implicit and only
error. applies within the assimilation interval.
22_Singh_ch22_p22.1-22.20.indd 6 8/22/16 1:31 PM

Adding another term to the objective function in Eq. (22.31) that takes into Statistical Correction
account errors in model formulation, yields the so-called weak-constraint Statistical correction can be considered as a derivative of the direct insertion
4D-Var: this method is numerically very costly and barely used. Examples of technique. The approach assumes that the model pattern is correct but con-
application of VDA techniques in hydrology are given by Castelli et al. (1999), tains a nonuniform bias: the predicted states are thus scaled and shifted to
Reichle et al. (2001), and Seo et al. (2003, 2009). match mean and variance of the observations.
State
State
Xb Xb0
X X0
Observation
Background Updated prediction
Analysis Previous prediction
Time Assimilation window Time

Figure 22.2 Schematic representation of different assimilation approaches. 3D-Var and other sequential DA techniques (left);
4D-Var (right).
22.4.3 Sequential Data Assimilation Successive Correction

The sequential techniques (or direct observer) approach the DA problem by The method consists of successively adjusting a background estimate for an
sequentially updating the model “background” states, the prior knowledge individual model state by nearby observations. The analysis at each time step
about the system states, using the additional information brought by a new is found through a series of iterations that use weight W to smooth observa-
observation z. The improved estimate of the state vector xa, which is called the tions into the model states (Bratseth, 1986) by modifying the states at all grid
“analysis,” is obtained through a correction applied to the background states, points within a user-defined influence radius for each observation. The
which depends on the difference between the observations and the model Cressman function is one of the most commonly used weighting system to
predicted observations (known as the innovation) multiplied by a weighting account for the distance between the model grid point and the observation.
factor Kt: At each time step t, the approach is usually applied consecutively to each
observation s as follows:
x ta = x bt + K t (z t − zˆ t ) (22.32)
x ts+1 = x ts + Wts (z t − z t )
The subscript t refers to the time of the update.
The fundamental difference between the various assimilation methods is where z t is evaluated through the observation operator considering the state
the choice of the gain matrix Kt, which represents the relative uncertainty estimate at the sth iteration and the analysis x ta is the value obtained at the end
in the observations and in model estimates, and is a number between 0 and 1 of the iterations.
in the scalar situation. With the aim to look for an unbiased estimate of x ta
with minimal variance, that is, a BLUE, and assuming a linear observation Optimal Interpolation
operator, that can be achieved choosing a proper Kt from the solution to the The method, also referred to as statistical interpolation, approximates the
variational optimization problem posed in Eq. (22.30), that is, such that objec- “optimal” solution of Eq. (22.32) through an algebraic simplification of the
tive function J is a minimum. This has been shown analytically (see, for computation of the gain K in Eq. (22.33). The analysis equation can be
instance, Jazwinski, 1970; Gelb, 1974) to produce: regarded as a list of scalar equation, one per model state in the vector x, and
the analysis increment for each of them is given by the corresponding line of
K t = PtbHtT (Ht PtbHtT + Rt )−1 (22.33)
K times the innovation vector. The fundamental hypothesis in optimal inter-
where Ptbis the background error covariance matrix of x bt . It can also be shown polation is that for each model variable only a few observations are important
that choosing the optimal least squares gain, the analysis error covariance Pta in determining the analysis increment, so that background and observation
matrix simplifies into error covariance submatrices restricted to selected observations are consid-
ered in the gain matrix. The algorithm also requires the background error
Pta = (I − K t Ht )Ptb (22.34) covariance matrix to be specified, that is often approximated with a fixed
where I is the identity matrix. Please note that Eqs. (22.32)–(22.34) are similar structure for all time steps and relies on the design of empirical autocorrela-
to the basic equations of the KF approach to data assimilation developed in tion functions given only by distance (Lorenc, 1981).
the 1960s for optimal systems control, which will be detailed in Sec. 22.4.4
Nudging
together with its variants.
Other sequential methods can be assimilated to Eq. (22.32) by choosing a Nudging (or Newtonian relaxation) has its roots in control theory and funda-
proper gain (Walker and Houser, 2005), such as direct insertion, statistical mentally consists of adding to the model equation a feedback term to gradu-
correction, successive correction, optimal interpolation/statistical interpola- ally “nudge” the model state toward the observations. The correction term is
tion, and nudging. While these approaches, briefly described as follows, are assumed proportional to the difference between the observation and the
computationally efficient and easy to implement, the updates do not always equivalent quantity computed by the model, multiplied by a gain matrix
account for observation uncertainty and do not propagate the state covari- accounting for the uncertainties of both modeled and observed variables. The
ance matrix explicitly. Houser et al. (1998) provide an example of application model solution can be nudged toward either individual observations; in this
of several of these alternative assimilation approaches. case, it is referred to as observation nudging, or gridded analysis obtained from
observation using other assimilation schemes (Stauffer and Seaman,1990),
Direct Insertion leading to the so-called analysis nudging.
It is one of the earliest and most simplistic approaches to data assimilation. As Nudging differs from the previously introduced sequential DA techniques
the name suggests, the forecast model states are directly replaced with the in the fact that through the numerical model the time dimension is included
observations, whenever one is available, by assuming explicitly that the model and thus it can be considered as a continuous form of four-dimensional data
has no useful information and that the observations are perfect. assimilation methods, that is, those that use information from more than one
22_Singh_ch22_p22.1-22.20.indd 7 8/22/16 1:31 PM

time like 4D-Var. Nudging allows the model to adjust gradually in a realistic With respect to the Wiener filter, the advantage of the KF lays in the fact
manner over multiple time steps to changes introduced by DA. that optimality (in terms of unbiasedness and minimum variance) is recur-
sively imposed at each step in time as opposed to the batch form required by
22.4.4 The Kalman Filter the Wiener filter. This also allows to locally linearizing in time non linear
The typical sequential data assimilation technique is based on the KF dynamic processes to produce what is known as the extended Kalman filter
(Kalman, 1960; Kalman and Bucy, 1961). Given its importance, this section (EKF) (Jazwinski, 1970).
provides the basic elements of the KF, while a more extensive derivation can The basic problem in applying the KF is that the optimality conditions
be found in Jazwinski (1970) or Gelb (1974). only hold when the state transition matrix Φt ,t −1, together with the model
Before introducing the KF, it is worthwhile explaining the domain of applica- and measurement error statistics ( w t , Qt , v t , Rt ) are fully known. Mehra
tion of a “filter.” A filter is essentially a data assimilation tool. In systems engi- (1970) provided a solution to the problem of estimating the unknown error
neering, there is a clear distinction between “filter,” “predictor,” and “smoother.” statistics by imposing the time independence of the innovation, a condition
A filter aims at estimating the value of a quantity at time t0 conditionally to associated with the KF optimality. The estimation of the state transition
available observations up to time t0; a predictor aims at estimating the value of matrix parameters (generally known as the “hyperparameters”) showed to
a quantity at time t > t0 conditionally to available observations up to time t0; a be a far more complex problem. Several approaches can be found in the lit-
smoother aims at estimating a quantity at time t < t0 conditionally to available erature for solving the nonlinear estimation problem originated by the
observations up to time t0. simultaneous estimation of both state and parameter values, ranging from
The KF (Kalman, 1960; Kalman and Bucy, 1961) is the recursive extension developing the KF in the parameter space (Mayne, 1965; Sage and Husa,
of the Wiener Filter (Wiener, 1949) to linear, or locally linearized in time, 1969) to the use of the EKF on state vector enlarged with the parameters
stationary as well as nonstationary processes. The original formulation of the (Jazwinski, 1970) or from the use of ML Cooper and Wood (1982a, b)
KF is in continuous time, but in hydrological applications it has always been (Gupta and Mehra, 1974; Wood and O’Connell, 1985) to the method of
applied in its discretized form. moments (Wojcik, 1993) and to full Bayesian approaches (Mantovan et al.,
In discrete form, the KF is build around two vectors: the state vector xt , 1999). Following the instrumental variables (IV) approach (Kendall and
which contains the n variables that uniquely characterizes the system under Stuart, 1967; Young, 1974; Young and Whitehead 1977), Todini (1978a, b)
study and the measurement vector zt , which contains the m available mea- realized that the posterior state estimate x t t is the best possible IV, being
surements. totally independent from measurement noise and a minimum variance esti-
In its simplest discrete formulation, two equations describe the system. The mator of the true unknown state due to optimality of the KF. Accordingly, he
first one represents the system dynamics through the evolution in time of the developed the mutually interactive state parameter (MISP) estimation tech-
state vector, while the second one represents the measurement system by nique by using two mutually conditional KFs: one in the space of the state
describing which of the state variables are being directly or indirectly mea- conditional to the previous step parameter estimates and one in the space of
sured. the parameters conditional to the previous and the last updated state esti-
mates (Todini, 1978a, b). MISP, was recently found by Mantovan et al. (1999)
x t = Φt ,t −1 x t −1 + Γ t w t known as the “model” or “system” equation (22.35) very close to ML, but much less demanding in computer time, while a full
and Bayesian approach, requiring the use of the Gibbs sampler (Gelfand and
Smith, 1990; Carter and Khon, 1994), to produce the posterior distributions,
z t = Ht x t + v t known as the “measurement” equation (22.36) had to be abandoned due to its exaggerated computer time requirements.
where x t [1, n] is the state vector, namely, the vector containing all the n suf- MISP was also recognized as the most efficient unbiased estimator in other
ficient state variables able to fully represent the system’s dynamics; Φt ,t −1 [n, n] fields, such as the automatic speech recognition (Labarre et al., 2006).
is the state transition matrix transforming the state vector from its state at Several examples of application of KF, in hydraulic and hydrological appli-
time (t − 1) to its state at time t, which may vary at each step in time; cations can be found in Chiu (1978), while specific applications to rainfall
w t [1, p ≤ n ] is an unknown random Gaussian time independent process, runoff models can be found in Hino (1973), Szollosy-Nagy (1976) for on-line
with mean w t and covariance matrix Qt , used to represent the model error, estimation of linear transfer function models, in Todini and Wallis (1978) for
while matrix Γt [n, p] is an appropriate matrix to relate the dimensions; on-line state and parameter estimation of a threshold-type multiple input
z t [1, m ≤ n] is the measurement vector, namely, the vector containing the m single output ARX model and in Georgakakos (1986a, b) for the on-line
observations and matrix Ht [n, m] is an appropriate matrix to relate the update of the Sacramento model parameters using an EKF.
dimensions; vt [1, m] is the measurement error, represented as a random Before concluding this section, it is worthwhile pointing out that the KF,
Gaussian time independent process with mean v t and covariance matrix Rt, being a “filter” is not a “predictor.” Therefore, it is an excellent tool for assim-
which is also assumed as being independent from the model error wt . ilating observations and calibrating state and parameter values, but rapidly
For the sake of simplicity, following the original derivation by Kalman becomes very poor when used for “forecasting” at future times, since observa-
(1960) the “control” term has been omitted from the model equation, while a tions are no more available. Therefore, the KF cannot be directly used for
measurement error term has been added. The KF aims at finding x t t the assessing what will be defined as “forecasting uncertainty,” unless the problem
unbiased minimum variance estimate of the unknown state xt together with is reversed and the model forecast, which is the only available information at
its error covariance matrix Pt t conditional to the knowledge of an unbiased a future times, is now taken as a “pseudomeasurement” of a future unknown
priori state estimate x t t −1 and its covariance matrix Pt t −1 (which fully represent realization.
the stochastic process due to the hypothesis of Gaussian errors) and the latest
noise-corrupted measurements zt together with its measurement error statis- The Extended Kalman Filter
tics (again mean and covariance are sufficient due to the Gaussian hypothe- The KF provides optimal estimates of the state for linear systems when both
sis). This estimate is obtained using the following equations, a neat derivation the model and the measurement systems are affected by independent
of which can be found in Gelb (1974). Gaussian noise. When dealing with nonlinear dynamical systems, application
At each step in time, the estimates of the state and of the error covariance of the KF implies local linearization Smith et al. (1962), McElhoe (1966) are
are extrapolated from the previous step: among the first ones who studied techniques to apply the standard linear KF
x t t −1 = Φ x
t ,t −1 + Γ w t state extrapolation
t −1 t −1 t (22.37) methodology to a linearization of the true nonlinear system. This approach,
called EKF, is suboptimal and can easily lead to divergence. The EKF became
Pt t −1 = Φt ,t −1 Pt −1 t −1 ΦtT,t −1 + Γ t Qt Γ tT error covariance extrapolation (22.38) the algorithm of choice in numerous nonlinear estimation, and machine
learning applications, including estimating the state of a nonlinear dynamic
Then the following quantities are estimated: system as well estimating parameters for nonlinear system identification.
υt = z t − Ht x t t −1 − v t known as the “Innovation” (22.39) To overcome the EKF limitations, new developments have appeared which
aim at overcoming the need for linearizing nonlinear models and the need for
the Gaussian assumption of model and measurement errors. New approaches
K t = Pt t −1HTt (Ht Pt t −1 HTt + Rt )−1 known as the “Kalman Gain” (22.40) appeared such as the ensemble Kalman filter (EnKF) (Evensen, 1994, 2003)
Finally the a priori estimates can be updated to include the latest measurements: and its simplified version known as the unscented Kalman filter (UKF) (Julier
and Uhlmann , 1997; Wan and van der Merwe, 2001) and finally the particle
x t t = x t t −1 + K t υt state update (22.41) filter (PF) (Gordon et al., 1993), which is the closest to a full Bayesian
approach, together with its simplified version, the unscented particle filter
Pt t = (I − K t Ht ) Pt t −1 error covariance update (22.42) (UPF) (van der Merwe et al., 2000).
22_Singh_ch22_p22.1-22.20.indd 8 8/22/16 1:31 PM

Interesting overviews of all these development can be found in Chen (2003) probability distributions. Nonetheless, there are several disadvantages that
and Feng et al. (2011). limit the use of the PF.
The large number of samples, much larger than that required by the EnKF,
The Ensemble Kalman Filter limit the PF to low dimensionality problems. In addition in the PF, there is a
The EnKF, introduced by Evensen (1994, 2003), is an interesting approach danger of degeneration of posterior densities, which requires frequent careful
aimed at overcoming the linearization problem, which produces suboptimal resamplings.
estimates when dealing with complex nonlinear models. EnKF is presently An example of application of the PF in hydrology can be found in
used for assimilating data in large nonlinear ocean and atmospheric models Moradkhani et al. (2005).
as well as in hydrological models.
The simple and innovative approach of the EnKF is to modify the state and Unscented Particle Filter
covariance extrapolation [Eqs. (22.37) and (22.38)]. In van der Merwe et al. (2000), the unscented transformation was used to
In the EnKF, one starts by extracting an ensemble of m possible states at approximate the appropriate “proposal distribution” of the PF, which resulted
time t0 , namely x it0 t0, with ∀ i = 1, …, m, from a prior distribution, usually a in the so-called UPF, allowing for a noticeable reduction of computational
Gaussian distribution with mean µ x t and covariance matrix Px t . efforts. The UPF couples the PF with the UKF (Mohammadi and Asif, 2011;
0 0
Zhao, 2014). In the UPF, the optimal proposal distribution function is approx-
The extrapolation of each individual member is then performed using the
imated as a Gaussian distribution whose statistics (mean and error covariance
model M instead of the state transition matrix, as in the standard KF:
matrix) are computed using the UKF. Presently, applications of the UPF and
x it t −1 = M(x it −1 t −1 ) (22.43) its variants are mostly limited to object tracking (Rui and Chen, 2001).
The state and the error covariance are then directly estimated from the 22.4.5 Choice of the Most Appropriate DA Approach
extrapolated members: The identification of the most suitable assimilation technique for a given vari-
able is still an open issue in DA for hydrological applications. A number of
x t t −1 = µ x t t −1 + Γ t w t state extrapolation (22.44)
questions and issues concerning hydrological DA, and particularly land surface
DA, are open areas of research mainly related to the development: of bias cor-
Pt t −1 = Px t t −1 + Γ t Qt Γ tT error covariance extrapolation (22.45) rection schemes (prior to data assimilation) in order to reduce systematic bias
1 m 1 m between observations and model outputs; of techniques facing the spatial reso-
where µ x t t −1 = ∑x it t −1 and Px t t −1 = ∑ i =1 ( x it t −1 − µ x t t −1 )2, while the other lution problem; of improved representation of model and observational errors;
m i =1 m
of data assimilation algorithms that can efficiently account for the typical
KF equations remain unchanged. nonlinear nature of hydrological models; of optimization of data assimilation
After the first step from t0 to t0 +1, there is no more need of generating a computational efficiency for use in large operational hydrological applications;
new ensemble because it is possible to update each individual member of the of multivariate hydrological assimilation methods to exploit observations from
ensemble: multiple sensors with complementary information (Houser et al., 2010).
The key factors in determining the choice of the most appropriate assimila-
x it t = x it t −1 + K t (z t − Ht x it t −1 − v t ) (22.46)
tion techniques depend on the suitability for the prescribed application, on
EnKF has been extensively used in several domains due to its simplicity. the computational feasibility, and on the optimality of the solution.
The only requirement is to generate a sufficiently numerous sample to reliably Filtering problems, like real-time assimilation, where the data is ingested
allow to estimate the mean and the covariance matrix (generally ≤ 100). sequentially as it becomes available, can be solved using sequential approaches
Also, the EnKF needs hyper parameters, which in this case are the model or with variational methods if a new smoothing problem is defined at each
M parameters, to be estimated. In that respect, an interesting extension of the measurement time. This is usually the framework for forecasting problems,
MISP approach to EnKF was recently introduced by Mordakhani et al. (2005). where the observations up to the current time are used to update the initial
conditions. Smoothing techniques are suited for analyzing historic data
The Unscented Kalman Filter (reanalysis), so that having observations for an entire study interval, they
The UKF (Julier and Uhlmann, 1997; Wan and van der Merwe, 2001) is a allows to include information also from later observations to improve the
variant of the EnKF, aimed at reducing the numerosity of ensemble members state estimation at a given time.
to be generated, which may prove useful when the model simulation cost is Sequential DA methods that considers approximations to the optimal filter
very high. equations or alternative solving methodologies, like direct insertion or meth-
Instead of estimating the extrapolated mean and covariance matrix from a ods based on the old Creissman analysis scheme, are computationally effi-
numerous randomly chosen ensemble, as in the EnKF, a small but representa- cient and relatively easy to implement, but the updates often account for
tive “deterministic” sample is used in UKF to calculate mean and covariance observational accuracy in an empirical way, and information on estimation
terms. Essentially, ( 2n + 1), sigma points (n being the state vector dimension), uncertainty is limited. Furthermore, the common assumption that the obser-
are chosen based on a square-root decomposition of the prior covariance. vations are more accurate than model forecasts, can lead to discontinuities in
These sigma points are propagated through the true nonlinearity, without the time series of the updated state that can cause possible model shocks.
approximation, and then a weighted mean and covariance are estimated. Moreover, these approaches are ineffective in data sparse regions (Houser et
UKF has the advantage of reducing the number of required simulations for al., 2010). On the other hand, methods based on probabilistic interpretation
estimating the extrapolated mean and covariance matrix, but, as opposed to in defining error statistics, such as the variational and other sequential meth-
the EnKF suffers from the asymmetry of the error probability distribution ods, although under different simplifying assumptions are able in theory to
and should be used for symmetrical distributions if not Gaussian. represent the uncertainty in the background, in the observations and in the
resulting analysis.
The Particle Filter In this perspective, operational 4D-Var method, which has been widely
The PF introduced by Gordon et al. (1993) was developed on the basis of used in weather forecast (traditionally viewed as an initial value problem; by
Smith and Gelfand (1992) sampling-resampling approach and falls in the contrast, typical land surface models are dissipative in nature), assimilates
category of the sequential Monte Carlo (SMC) algorithms aimed at correctly uncertain data into a deterministic model under the assumption that the
describing the posterior density of the state vector given the model and the model is perfect (strong constraint). In weak-constraint 4D-Var, a term in the
measurements. The required posterior density function is represented by a set penalty function involving model errors is included. Through the 4D-Var,
of random samples with associated weights, which are then used in the esti- being a smoothing algorithm, there is a larger potential for dynamically based
mation phase. In practice, a sample of the state is generated from a prior dis- balanced analyses and performs satisfactorily when data are sparse.
tribution and the prediction phase, similar to that of the EnKF, implies Nevertheless, three major drawbacks of 4D-Var must be mentioned. First, its
routing this sample through the model. The update phase of the PF implies numerical cost is high compared to approximate versions of the KF or
resampling and weights estimation (Smith and Gelfand, 1992), for instance, ensemble-based methods (e.g., EnKF and PF). Second, its formalism does not
using the principle of the “important sampling” (Doucet, 1998). allow to determine the analysis error directly. Finally, the definition of the
As the number of samples becomes very large, the SMC representation adjoint model, is time consuming and may be difficult for a system (e.g.,
approaches the usual functional description of the posterior density function, hydrological models) which exhibits nonlinearities and threshold processes
and the SMC filter approaches the optimal Bayesian estimate. (Lahoz and Schneider, 2014).
The advantages of the PF lie in the fact it allows estimating the full poste- The classic KF provides the optimal state estimate for linear systems and
rior probability density in cases of nonlinear systems with non-Gaussian error provides information on estimation uncertainty through a dynamic updating
22_Singh_ch22_p22.1-22.20.indd 9 8/22/16 1:31 PM

of the forecast (background) error covariance through time: on the other Differential Analysis
hand, it is of limited use in hydrological applications where physical models Differential analysis evaluates local sensitivity at the nominal value of the
are often nonlinear. For nonlinear dynamics the EKF can be used, but neither parameter space by taking the partial derivative of y with respect to an input
the optimality of the analysis nor the equivalence with 4D-Var holds, and if factor xi , while all parameters are held constant to their nominal value. The
the nonlinearities are strong, it is prone to become unstable. The computa- derivative values are themselves the metrics of sensitivity:
tional cost (larger than that of 4D-Var even with small models) of propagating
error information result in making both KF and EKF impractical for high ∂y
dimensional systems. Sj = (22.47)
∂ xi x nom
Ensemble based-data assimilation methods, like EnKF and PF approaches
and their variants, although can be computationally demanding depending where the subscript x nom indicates that the derivative is taken at the nomi-
on the size of the analyzed ensemble, are becoming attractive for near real- nal value in the space of the input. Derivatives are usually normalized by the
time applications as they do not require the derivation of a tangent linear reference value at which the derivative is calculated, thus removing the effects
operator or adjoint equations, and offer great flexibility with respect to the of units and allowing the comparison of sensitivity indices related to param-
form of model error that can be included. eters with different units of measure. Inherent to this calculation are the
assumptions that the higher order partials are negligible and there is no cor-
relation between input parameters. The calculation of model derivatives is a
22.5 SENSITIVITY ANALYSIS delicate task. A straightforward and frequently used way for approximating
the sensitivities is realized by applying finite difference schemes, described in
Sensitivity analysis (SA) is the study of how the variation in the output of a literature as the brute force or the indirect method (as it does not require
model can be apportioned, qualitatively or quantitatively, to different sources manipulation of the underlying model equations and it is applied in a black-
of variation (Saltelli et al., 2000), including model parameters, forcing, initial box manner); other methods for calculating the local sensitivities are the
conditions or boundary conditions. direct method, Greens function method, and the polynomial approximation
An extensive analysis of the relation between model input factors and out- method. A powerful and promising strategy for reliable and accurate deriva-
put variables represents an essential screening tool that can be helpful for tives evaluation is the implementation of algorithmic differentiation (Hwang
manifold purposes: identifying potential deficiencies in model structure; et al., 1987; Griewank, 2000), which is a set of techniques to numerically
providing guidance for model simplification; identifying model inputs having evaluate the derivative of a function specified by a computer program.
a significant influence on model responses; finding critical regions in the
space of inputs. A related practice is uncertainty analysis, which has a greater 22.5.2 Screening Techniques
focus on assessing how the data errors propagate through the model and Screening techniques are SA methods that also make use of local sensitivity
affect the output (see Chap. 26). measures to compute parameter sensitivities, but calculated at a larger num-
In this section, although often referring to a generic “input” or “factor,” the ber of points in the parameter space. One-at-a-time (OAT) designs (Daniel,
focus is mainly on analyzing the sensitivities of model outputs to model 1973) and Morris’ design (Morris, 1991), briefly described in the following,
parameters: in this perspective, one of the main goal of SA, is to rank, quali- are among the most popular screening methods. Other screening SA methods
tatively or quantitatively, the various parameters based on their relative are the factorial, the fractional factorial and the iterated fractional factorial
importance in influencing the model response, with the aim of providing designs, and sequential bifurcation (Box and Draper, 1987).
information to achieve an appropriate and parsimonious model structure as
well as the minimum number of parameters to facilitate calibration. Referring One at a Time
to the degree to which a parameter affects model output, the terms Conceptually, one of the simplest and most common approaches is that of
“influential,” “important,” “effective,” or “correlated” are almost interchange- perturbing repeatedly model inputs OAT, holding others at their nominal
ably used in literature, while it should be pointed out that only the model values (Daniel, 1973). A sensitivity ranking can be obtained quantifying the
responses could be properly defined as “sensitive” or “insensitive” with difference (or the percentage change) in the model output due to the change
respect to the parameter variation. in the input variable. Each input factor, for instance, can be individually per-
Any SA approach can be considered as having two components: (1) a strat- turbed by a factor of its standard deviation (Downing et al., 1985) or varied
egy for sampling the model parameters space and (2) a numerical or graphical from its minimum to its maximum value.
measure to quantify the impacts of sampled parameters on the model output
of interest. The implementation of these two components varies widely and Morris’ Design
several studies have reviewed and classified existing SA methods (Iman and The experimental design proposed by Morris (1991) is composed of individu-
Helton, 1988; Hamby, 1994; Melching, 1995; Frey and Patil, 2002; Cacuci, ally randomized OAT experiments, in which the impact of changing one
2003; Oakley and O’Hagan, 2004; Saltelli et al., 2000, 2004, 2008). Many tech- factor at a time is evaluated in turn, exploring several regions of the parameter
niques have been applied also to hydrological models (e.g., Carpenter et al., space. The method relies on a local sensitivity measure, called elementary
2001; Sieber and Uhlenbrook, 2005; van Griensven et al., 2006; Pappenberger effect of xi (EE i), due to a perturbation ∆, and expressed as:
et al., 2006; Tang et al., 2007) even if the increasing complexity of distributed
models poses several challenges in terms of their computational demands and f ( x1 ,…, xi + ∆ ,…, xn ) − f ( x1 ,…, xn )
EE i ( x1 ,…, xn , ∆ ) = (22.48)
their potential overparameterization. ∆
Throughout this section, the generalized model functional form is repre- The sensitivity measures proposed by Morris for each factor xi are the
sented by y = f ( x ), where it has been assumed a single model output y, and mean and the standard deviations of the elementary effects computed at dif-
x is the vector of independent variables xi with i=1,…n potentially including ferent randomly sampled points in a parameter space normalized to the unit
model parameters, forcing, initial conditions, etc. It is worth noting that SA n-dimensional hypercube.
can be accomplished considering either a particular model output y or a 22.5.3 Global Sensitivity Analysis
generic error function describing the degree of model fit to the observations.
SA techniques can be classified as local, screening and global methodolo- Global SA methods attempt to explore the full parameter space within pre-
gies, depending on how the parameters are perturbed. A synthetic description defined feasible parameter ranges and distributions, and to address the effect
of the main approaches is provided in the following section. on model output resulting from considering all input parameters simultane-
ously and from their interactions. The available approaches to perform global
SA, can be grouped, not without overlapping, into random sampling meth-
22.5.1 Local Sensitivity Analysis
ods, variance-based methods and response surface techniques.
Local methods focus on the study of model response behavior for small
changes in input factors, generally considered individually, around a specified Random Sampling Methods
value in the parameter space, called “nominal value,” which can be taken from These approaches consist in sampling at random the vector of input param-
the literature, or assumed as the calibrated or the mean value of the parame- eters and propagating their effect through the model to explore the relative
ters. The major drawbacks of these methods are their inability to fully explore mapping of the output functions. The parameters are sampled from a prior
the input space and to account for parameter interactions, making them parameters distribution using one of the several available strategies, such as
prone to underestimating actual model sensitivities (Tang et al., 2007). random Monte Carlo sampling, importance sampling or Latin hypercube
Moreover, the results are function of selection of the nominal value. sampling (McKay et al., 1979).
22_Singh_ch22_p22.1-22.20.indd 10 8/22/16 1:32 PM

Validation techniques 22-11
The sensitivity measures inferred from sampling SA methods can be evalu- Vij = V [E ( y | xi , xj)] − Vi − Vj (22.50)
ated in several ways: scatter plots, regression and correlation-based methods,
and sensitivity tests involving segmented input distributions. ……………
Scatter Plots The Vi terms are called first-order terms, or main effect of xi , and they
Parameter sensitivity can be assessed qualitatively by plotting the output vari- represent the reduction of V that would be expected if the true value of xi is
able against individual input variables. Scatter plots are useful for quick visual known: it is worth noting that the Vi estimate is a more general sensitivity
assessment of the influence of individual inputs on model response: the pat- measure than a regression coefficient, as it also works for nonlinear models.
tern of the plot can directly reveal linear or monotonic relations, the existence In the same way, the second-order terms Vij accounts for the interaction effect
of thresholds, or other complex dependencies. Other widely used graphical between the parameters xi and xj.
tools to analyze relations between inputs and outputs are cobweb plots. In this framework, the sensitivity of the output variance to each individual
parameter or parameter combinations is then assessed based on the Sobol’s
Regression- and Correlation-Based Methods sensitivity indices:
The regression-based SA relies on determining a multiple regression between Vi
the model output and the model parameters as explanatory variables. In some first order
Si =
V
way, this relationship is used to replace a highly complex model with a simpli-
V ji
fied “response surface.” Regression analysis typically involves fitting a linear second order S ji = (22.51)
relationship between inputs and the output which in its simplest form corre- V
sponds to the first order polynomials. …………….
The fitted regression coefficients (RCs), representing the expected change
in the output y per unit change in xi when all the remaining independent V~i
variables are constant, provide an absolute sensitivity measure of model total order STi = 1 −
V
parameters. A standardization process is often necessary to remove the effects
of units to allow comparisons and rankings. The standardized regression where Si denotes the sensitivity that results from the main effect of param-
coefficients (SRCs) are obtained by the ratio of parameter and output stan- eter xi ; the second order sensitivity indices, Sji , define the sensitivity that
dard deviations. If the inputs are independent, the SRCs give the fractional results from the interaction of parameters xj and xi , and so on for the higher
contribution of each factor to the variance of y. orders till a total number of 2n−1 of indices. In Eq. 22.51 V~i represents the
Whenever a nonlinear but monotonic relationship between input and out- variance explained by all the factors except xi . The total order index STi
put exists, a rank transformation (Iman and Conover, 1979) should be applied measures the contribution of xi to the output variance with respect to the
to get a linear relationship: rank regression coefficient and standardized rank single variable and to every interaction with the other input variables. The
regression coefficient are the correspondent qualitative sensitivity measures. computation of Sobol’s indices is based on the computation of the variances
A drawback of rank-based sensitivity measures is that the properties of the of conditional expectations, which is very time consuming. Estimation
model y can be distorted by the rank transformation (Saltelli and Sobol’, 1995). procedures for the variance-based sensitivity indices are the Sobol’s method
Correlation coefficients can also be used as a measure of sensitivity. The (Sobol’, 1993), the Fourier amplitude sensitivity test (FAST) method (Cukier
Pearson correlation coefficient, or the corresponding Spearman correlation et al., 1978; Saltelli et al., 1999), and correlation ratios (McKay, 1995).
coefficient based upon rank-transformed values, can be used to characterize
the degree of linear relationship between the output values and sampled val- Response Surface Technique
ues of individual inputs. A possible way to account for correlations among The response surface method (Box and Wilson, 1951; Myers and Montgomery,
other input variables relies in the calculation of partial correlation coefficients 1995), also known as meta modeling or surrogate modeling, is based on
or partial rank correlation coefficient. building a “response surface,” that is, a simplified model that approximately
Sensitivity Tests Involving Segmented Input represents the behavior of the model response as a relatively simple function
Distributions (Monte Carlo Filtering) of the chosen model parameters, and is often applied also in the calibration
process. The surface is then used to replace the original model for subsequent
Alternative SA techniques involve some form of dividing or segmenting analysis (Monte Carlo, etc.). A typical approach to response surface develop-
input parameters into two or more empirical distributions based on an asso- ment is to use regression methods. Among the other response surface models,
ciated partitioning of the output distribution according to some criterion the polynomial chaos expansions are quite effective for the SA (Wiener,
(Hamby, 1994). Either graphical methods (e.g., marginal cumulative distri- 1938). The response surface method is generally complex and computation-
bution function plots) or statistical tests, such as Kolmogorov–Smirnov and ally demanding, therefore, it is generally used in later stages of the analysis,
Cramer–von Mises t tests, test and Mann–Whitney test, F test and squared- investigating a limited number of factors.
ranks tests, are used to compare the input distributions, or some of their
characteristics (e.g., means or variances), created by segmentation. If the 22.5.4 Choice of the Most Appropriate SA Approach
distribution of a certain parameter, xi , is proved to be different in the sub- Although tools based on OAT approaches are widely used in investigations aim-
samples, xi is considered to be an influential parameter. Another kind of ing at SA, it has been recognized that their use is unsuitable unless the analyzed
approach involving segmented input distribution, often applied in hydrology, model is proved to be linear. Global SA methods, based on MC schemes, in spite
is the generalized SA, or regional SA (RSA), by Spear and Hornberger (1980), of their higher computational cost, have been identified as a best practice when
which inspired the generalized likelihood uncertainty estimation procedure dealing with nonlinear models, ensuring thorough exploration of the multidi-
by Beven and Binley (1992). The RSA approach partitions the MC-generated mensional space of input uncertainty, while capture possible interactions
parameter sets into two groups “behavioral/non-behavioral,” according to among variables. Moreover, by allowing for the characterization of the proper-
model performance or behavior. ties of posterior distribution, these approaches can be implemented as part of
Variance-Based Methods further analyses aimed at model calibration and uncertainty quantification. The
use of quantitative model-free measures of sensitivity, such as variance-based
Variance-based methods are a particular form of sampling-based SA tech- measures, is also recommended in problem related to factors prioritization.
niques allowing a full exploration of the input space, accounting for interactions
and nonlinear responses. The approach considers that each parameter is
responsible for a fraction of the output variance and takes this fraction as the 22.6 VALIDATION TECHNIQUES
reference sensitivity measure. For independent input factors, the total variance Although wide variations of the adopted terminology exist in the literature
of a model output y can be decomposed into a series of contributions that (Anderson and Bates, 2001; Matott et al., 2009) the term validation, well
result from individual parameters as well as parameter interactions: known in hydrology and environmental modeling, is commonly used to
V(y) = ∑i Vi + ∑i < j Vij + ∑i < j < k Vijk + … + Vi2…n (22.49) indicate a procedure aimed at assessing the performance of models, proce-
dures, and decisions in different settings, such as the adequate representation
where the various terms Vi , Vij , Vijk ,…, Vi2…n are the conditional variances, of reality in the scientific context or the robustness of decision making based
obtained by averaging over the possible values of the fixed parameter or set of on simulation or on forecasting.
parameters: In the scientific context, the term validation may include the representation
Vi = V [E ( y | xi )] of a single physical process or the overall behavior of several interacting
processes.
22_Singh_ch22_p22.1-22.20.indd 11 8/22/16 1:32 PM

Therefore, the choice of the proper model testing system and validation observations should be used respectively for calibration and validation and
criteria should be guided by several factors, mainly depending on the nature results compared. The model should be judged acceptable only if the two
of the simulated variables and the model purpose. Before going into details, it results are similar with admissible errors.
is necessary to outline some important premises. This procedure coincides with the so-called split sample test in the classic
Referring to utilization of hydrological models, it is emphasized the dis- hierarchical validation scheme proposed by Klemeš (1986), as well as with the
tinction between: (i) simulation models intended as exploratory or research first level of the theoretical scheme of Gupta et al. (2008).
tools, that is, whose purpose is mainly focused on the understanding and Further tests have been proposed by Klemeš to evaluate model transfer-
correct representation of hydrological processes; (ii) simulation models devel- ability in space and time, that is, under (e.g., climatic) conditions different
oped for operational purposes, typically for water-management decisions (in from those corresponding to the available flow record. Particularly valuable
particular, for planning, design, etc.) whose objective is to improve the reli- when dealing with simulation of a streamflow record in ungauged basins, is
ability and robustness of decisions; and (iii) forecasting models that are the proxy-basin test, which evaluates the geographical transposability of the
primarily addressed to the estimation of the variable of interest at some model to similar basins: the model should be calibrated on a basin and vali-
specified future time, or its predictive probability density in view of increasing dated on another basin and vice versa. If the two validation results are accept-
the reliability and robustness of decisions aimed at reducing flood risk or issu- able and similar, the model has a level of credibility with regard to its ability
ing flood alerts under the uncertainty of future events (see Chap. 26). to simulate the streamflow in a third (ungauged) basin adequately.
In the latter two cases, the aim is to extend into the future (uncondition- The current Sec. 22.6.2 introduces the techniques useful for validating
ally—simulation—or conditionally to the latest measurements—forecasting), performance of model deterministic simulations, while the techniques used
the available hydrological knowledge and in particular variables on which for assessing the accuracy of hydrological predictions, and in particular
decision should be based. In this perspective, the criteria for model perfor- probabilistic predictions, are detailed in Sec. 22.6.3.
mance assessment must be user-oriented, or operational (Klemeš, 1986), that
is, they usually but not necessarily, only measure the correctness of the esti- Numerical Techniques
mates of hydrological variables generated by the model as well as the perfor- The performance metrics (or indexes) provide a quantitative and aggregate
mance in terms of correct decisions. In general, this type of assessment does estimate of model reliability and are generally expressed as a function of the
not guarantee, that the model is correct from a scientific point of view or simulation errors. Some metrics have a statistical foundation, as the likeli-
sometimes that the model is a very good model, provided it allows taking the hood functions, the Akaike Information Criterion, the Bayesian Information
correct decisions at the right time. Criterion and the Kashyap Information Criterion. The last three statistic cri-
Moreover, taking inspiration from applied meteorology, and the pioneer teria account for the mathematical complexity of the model by including the
work by Murphy (1993) in the field of forecast verification, a not negligible number of model parameters in the metric computation.
part of the “goodness” of a forecast is its value, defined as the degree to which A wide number of metrics can be derived from the general expression (van
the forecast helps a decision maker to realize some incremental economic der Molen and Pintér, 1993):
and/or other benefit. 1/b
A broader concept of validation should be applied to the former case: the 1 N τ
aim is at comprehensively evaluating the consistency, and the coherence with F =  ∑ y s ,t − yo ,t  τ ≥ 1, b ≥ 1 (22.52)
 N i =1 
real world, of the model thought as an ISO (input-state-output) system.
Model performance can be addressed by means of qualitative and quantita- or from the analogous relation function based on the relative deviations,
tive criteria. The former essentially rely on the graphical comparison between τ 1/b
observed and simulated data, whereas the latter are based on numerical per- 1 N y −y 
formance metrics. Both approaches are fundamental tools to be used in F =  ∑ s ,t o ,t
 yo ,t ≠ 0, τ ≥ 1, b ≥ 1 (22.53)
 N i =1 yo ,t 
complementary fashion, since they are able to capture distinct aspects of  
model performance. where F is the performance metric, N is the number of observations, while ys,t
and yo,t are the simulated and observed values at time t, respectively.
22.6.1 Performance Validation of Exploratory Research Tools In particular, the metrics related to Eq. (22.53) are dimensionless and, thus,
provide a more balanced evaluation of model performance over the entire
It has been argued that when verifying the model performance by simply
study period. Metrics derived from both expressions do not have an upper
comparing outputs and observations, the validation assessment lends only a
boundary, while a null value indicates a perfect fit.
rather modest level of credibility limited to interpolation within the range of
According to the values assumed by τ and b parameters, the two expres-
the database and confined to the task for which the model has actually been
sions provide different metrics, some of which are listed in Table 22.1. For
tested (Klemeš, 1986). In other words, these tests do not allow to assess if the
higher τ , the metric is more sensitive to large differences between simulated
model structure and parameterization are consistent with the physics of the
and observed values. To assess the quality of the model fit, other indexes, for
simulated processes (Oreskes, 2003), that is, the hydrological soundness of
example, the Janus coefficient (Power, 1993), compare the model errors in the
the model. The concept of a broader “scientific” or “internal” validation origi-
validation and the calibration period. Other commonly adopted goodness-of-
nates from these considerations: synthesizing, the scientific validation has the
fit metrics are based on regression operations between simulated and
goal of verifying that right outputs are produced for the right reason (Kirchner
observed data (Legates and McCabe, 1999): for example, the coefficient of
et al., 2006; Aumann, 2007).
determination R2, the index of agreement d (Wilmott et al., 1985), and the
The idea that a model performs well only if it returns satisfactory results,
coefficient of efficiency Nash–Sutcliffe Efficiency coefficient (NSE) intro-
provides a limited view of the practical utility and scientific value of a model,
duced by Nash and Sutcliffe (1970), which is by far the most utilized index in
and does not provide meaningful basis for extrapolation.
hydrological applications.
The scientific validation may include the usual validation techniques
Due to its large popularity, it is worthy to focus on Nash–Sutcliffe coeffi-
applied in performance validation of simulation models and extend them to
cient, whose main characteristics are as follows: (i) it measures the departure
model input and state variables.
from unity of the ratio between the mean squared error and the variance of
In this framework, the quantification of the different sources of uncertainty
the observations; (ii) it varies between −∞ and 1; and (iii) a null value is
(e.g., observations, process, parameterization, and model structure) is crucial
obtained when the simulation is identically equal to the mean value of the
(Todini, 2007).
observed series.
In this perspective also, providing examples of poor model performance is
The diagnostic properties of the Nash–Sutcliffe efficiency have been
very useful to highlight model weaknesses and to guide further model
recently investigated in detail by Gupta et al. (2009), who eventually proposed
improvements.
an innovative index, called kling–gupta efficiency, that in a multiobjective
Although the obvious costs of this type of validation hinders its application
perspective is expressed as an explicit function of three components: correla-
in most of practical applications, it may have a broader proactive impact on
tion, variability error, and bias error.
modeling choices guiding monitoring campaigns, specific investigations or
Other frequently used indexes include those based on the rank correlation
simply a wiser choice of the modeling tools.
criteria, such as the Spearman and Kendall coefficients. The reader is referred
to Jachner and van den Boogaart (2007) and Dawson et al. (2007) for a
22.6.2 Performance Validation of Simulation Models detailed survey on performance metrics.
The typical approach adopted to evaluate model performance requires the Many criticisms have been addressed to traditional lumped metrics for
comparison between simulated outputs on a set of observations that were not their lack of diagnostic power or inability to capture differences between dif-
used for model calibration. When possible, two nonoverlapping periods of ferent model or parameter sets leading to ambiguous situations characterized
22_Singh_ch22_p22.1-22.20.indd 12 8/22/16 1:32 PM

Validation techniques 22-13
Table 22.1 Numerical metrics used to evaluate model performances

Performance metric Expression
N
1
Mean absolute error F1 =
N
∑| ys ,t − yo,t |
t =1
N
1
∑| ys ,t − yo,t |
2
Mean squared error F2 =
N t =1
12
1 N
2
Root mean squared error F3 = 
N
∑| ys ,t − yo,t | 
t =1
1
Minimax objective function F4 = max | y s ,t − yo ,t |
N
N
1 y s ,t − yo ,t
Average absolute percentage error F5 = 100
N
∑ yo ,t
t =1
N 2
1 y s ,t − yo ,t
Mean squared relative error F6 = 100
N
∑ yo ,t
t =1
2
 ∑tN=1 ( y s ,t − y s )( yo ,t − yo ) 
Coefficient of determination (R2) F7 = 
N 2 12 N 2 12
  ∑t =1 ( y s ,t − y s )   ∑t =1 ( yo ,t − yo )  
∑tN=1 ( y s ,t − y s )2
Index of agreement (d) F8 = 1 −
∑tN=1 (| y s ,t − y s | + | yo ,t − yo |)2
∑tN=1 ( y s ,t − yo ,t )2
Nash-Sutcliffe efficiency coefficient F9 = 1 −
∑tN=1 ( yo ,t − yo )2
by equifinality. As a result, more powerful evaluation tool like multi-objective forecasting. These techniques are both applied to forecasts issued in hindcast
criteria that combines different (weighted) performance metrics into one mode as well as to real-time operational forecasts. The term hindcast refers to
overall objective function (e.g., Madsen, 2000) have been proposed. the case of a model forecast where the model is used forced by known input,
such as when simulating forecasting using past data with the scope of assessing
Graphical Techniques potential benefits for using predictions in decision making.
Graphical techniques allow a subjective and qualitative validation. Despite the Model forecasts can be either deterministic or probabilistic. In the tradi-
plethora of existing goodness-of-fit metrics, visual inspection still represents tional deterministic approach, a unique “best” output is produced. More
a fundamental step in model validation as it allows the study of temporal recently, a number of techniques, such as ensemble forecasting and statistical
dynamics of model performance and facilitate, the identification of patterns post-processing approaches, have been proposed to account for different
in error occurrence. In most cases, they are based on a graphic comparison of sources of uncertainty and produce probabilistic forecasts.
simulated and measured time series (see Fig. 22.3 left). This kind of plots can In the aforementioned essay by Murphy (1993), three types of “goodness”
be difficult to read, especially when the observation period is long. Scatterplots of a forecast are distinguished:
of simulated versus observed variables are more easily interpretable and pro- Consistency: The degree to which the forecast corresponds to the forecaster’s
vide an objective reference given by the 1:1 line of perfect fit (see Fig. 22.3 best judgment about the situation, based upon his/her knowledge base
center). Other common graphical representations are residual plots, and the Quality: The degree to which the forecast corresponds to what actually
comparison of simulated and observed data through box-plot or quantile– happened
quantile Q–Q plots (see Fig. 22.3 right). Value: The degree to which the forecast helps the decision maker to realize
some incremental economic and/or other benefit
22.6.3 Performance Validation of Forecasting Systems Traditionally, forecast verification has emphasized quality, but it is impor-
Forecast verification techniques, mainly developed in the atmospheric sciences, tant to note that quality is not necessarily the same as forecast value. A
have only recently begun to be used for evaluating forecasts in hydrological forecast has high quality if it predicts the observed conditions well according
modeling, especially in view of the increasing interest in probabilistic flood to some objective or subjective criteria. It has value if it helps the user to make
30 30 12
Observed
Sim. discharge (m3/s)
25
Simulated
Discharge (m3/s)
Redisual (m3/s)
20 20 6
15
0
10 10
5
–6
0 0
1 Dec 2002 1 Feb 2003 0 10 20 30 1 Dec 2002 1 Feb 2003
Observed discharge (m3/s)
Figure 22.3 Graphical methods used to evaluate deterministic model forecast performances (Biondi et al., 2012). Observed and simulated time
series (left), scatter plot (center), residual plot (right).
22_Singh_ch22_p22.1-22.20.indd 13 8/22/16 1:32 PM

a better decision. When the cost of a missed event is high, the deliberate Practically SS = 0, indicates no improvement over the reference forecast; a
overforecasting of a rare event may be justified, even though a large number positive value forecast being evaluated is superior to the reference one. This
of false alarms may also result. relative measure can be easily constructed using almost all the measures of
An important discrimination when verifying forecasting models regards accuracy detailed in the previous section.
the form of the variable under analysis that can be a continuous variable or a
categorical (discrete-valued) variable, that is, a variable that can take one and Probabilistic Forecast
only one of a finite set of possible values. In the following, a brief description While in meteorological applications Joliffe and Stephenson’s (2003) book
of some of the verification tools available in the different situations is carried offers an informative description of current verification practices, in hydro-
out, separating the case of deterministic forecast from the probabilistic one. logical applications a generalized consensus has not yet been reached on how
to assess probabilistic forecasts.
In case of probabilistic forecast for binary (yes/no) events, a commonly
Deterministic Forecast
adopted verification tool is the Brier score, which takes the form of the more
Deterministic (but also a probabilistic) hydrological forecasts can be trans- general ranked probability score (RPS) when it is intended to be applicable to
formed into a categorical “yes, an event will happen,” or “no, the event will not multicategory forecasts. The relative operating characteristic (ROC) is,
happen” forecasts according to some critical value threshold (e.g., the water instead, a graphical tool that illustrates the performance of the forecast as
level exceeds or not a given threshold value). a binary classifier system by varying its discrimination probability threshold
Taking inspiration from applied meteorology, the contingency table, which (Mason, 1982). The curve is produced by plotting the hit rate versus the false
shows the frequency of yes and no forecasts and occurrences, together with alarm rate at various threshold settings. The area under the ROC curve is
the frequency of their combinations (hit, miss, false alarm, and correct nega- frequently used as a score. Another useful graphical device is the reliability
tive), is a common way to analyze what types of errors are being made. A large diagram that plots the conditional distribution of observations (binary pre-
variety of categorical scores can be computed from the elements of the con- dictands) given the forecast probability, where the range of forecast probabil-
tingency table to describe particular aspects of forecasts performance: accu- ities is divided into K bins. The sample size in each bin is often included as a
racy, bias score, hit rate or probability of detection, false alarm rate, false histogram or values beside the data points.
alarm ratio, threat score or critical success index, Gilbert skill score, Pierce When dealing with continuous predictands, standard goodness-of-fit metrics
skill score, Heidke skill score, among the others (see Wilks, 1995, for a used for deterministic forecast do not allow a complete and fair evaluation of
detailed review). It is worth noting that particular attention has to be paid in the forecast performance and appropriate diagnostic approaches are required.
interpreting the results of those scores that are heavily influenced by the most Probabilistic verification techniques, mainly developed in the atmospheric
common category, usually “no event,” when applied to the case of rare events. sciences, have only recently begun to be used, together with specifically
Many of the previous indexes (e.g., Accuracy, Heidke skill score, and Pierce designed tools, for evaluating probabilistic forecasts in hydrological modeling.
skill score) can be extended for verifying multicategory forecasts: they also In his essay for a distribution-based approach, Murphy (1993), distinguishes
start with a contingency table showing the frequency of forecasts and obser- nine attributes that contribute to the fullest description of the multifaceted
vations in the various bins, where a perfect forecast system would have values nature of the probabilistic forecast quality: association, accuracy, skill, bias,
of nonzero elements only along the diagonal, and values of 0 for all entries off sharpness, uncertainty reliability, resolution, and discrimination (Table 22.2),
the diagonal. However, any multicategory forecast verification can be easily while Gneiting et al. (2003) only concentrate on “calibration” and “sharpness.”
converted to a series of binary-type verifications. Each of these attributes carries fundamental information about the forecast
When dealing with deterministic forecasts of a continuous-valued variable, performance: the first three attributes characterize the joint distribution of the
which, until present, has been the most typical forecast in hydrology, many forecasts and observations; attributes four to six define properties of the mar-
goodness-of-fit metrics, for example, the mean error, mean absolute error, ginal distribution of forecasts and observations, or are concerned with their
root mean square error, mean squared error, similar to the ones mentioned for relationship; the last four attributes concern the conditional distributions of the
the performance validation of simulation models (Sec. 22.6.2), can be used. forecasts given the observations, and the conditional distributions of the obser-
vations given the forecasts. Some recent techniques have been specifically
Skill Scores designed to quantify their weight in the process of verifying hydrologic proba-
bilistic forecasts; contributions in this field include the work of Welles (2005),
Forecast skills are usually referred to the relative accuracy of a set of forecasts
Laio and Tamea (2007), Engeland et al. (2010), Bartholmes et al. (2009), Bogner
with respect to some set of standard control or reference forecasts. Common
and Pappenberger (2011), Franz and Hogue (2011), and Thyer et al. (2009).
choices for the reference are climatological average values of the predictand,
With the aim of verifying if the forecast is correct under a statistical view-
persistence forecasts, forecasts from other models or random forecasts
point, Laio and Tamea (2007) proposed the use of the probability integral
(Wilks, 1995). In this perspective, forecast skill is usually presented as a skill
transform, evaluating the cumulative distribution function of the predictions
score (SS) interpreted as a percentage improvement over the reference fore-
in correspondence to the observed value, and of a Q–Q plot. In an ideal situ-
cast:
ation, the predictive cumulative probability values should be independently
and uniformly distributed on the interval [0,1]. The approach also provides a
score forecast − scorereference
SS = (22.54) diagnostic tool to investigate the causes behind deviations from uniformity
score perfect forecast − scorereference (Figs. 22.4 and 22.5).
Table 22.2 List of Attributes for Various Aspects of Probabilistic Forecast Quality (Adapted from Murphy, 1993)
Attribute Definition
Association Strength of the linear relationship between individual pairs of forecasts and observations
Accuracy Average correspondence between individual pairs of forecasts and observations
Skill Relative accuracy of forecasts of interest over some reference forecasts
Bias Correspondence between mean forecast and mean observed value
Sharpness Variability of the forecasts
Uncertainty Variability of the observations
Reliability Characterize the conditional distribution of the observations given the forecasts
Ability of the system to sort the observations into groups of different frequency distributions (difference between conditional mean observation
Resolution
and unconditional mean observation averaged over all forecasts)
Discrimination 1 Characterize the conditional distribution of the forecasts given the observations
Ability of the system to provide different forecasts for those occasions having different observed values of the predictands (difference between con-
Discrimination 2
ditional mean forecast and unconditional mean forecast averaged over all observations). It is the converse of resolution.
22_Singh_ch22_p22.1-22.20.indd 14 8/22/16 1:32 PM

REFERENCES 22-15
alarm is given, no preventive action is done: if the event is not observed, no

Ri /n loss occurs, but if the event does occurs, a loss L is expected. In particular, the
decision-maker aims at taking decisions that will minimize the expected
1 expenditure, which is the sum of the cost and the loss terms.
Over The (relative) value score, the same concept used for SS has been intro-
prediction duced with the aim of assessing the relative improvement in economic value
of a forecast system with respect to a reference forecast or to a predefined
Narrow strategy (e.g., always taking actions or never taking actions). Available scores
forecast are often expressed as a function of a cost-loss ratio, that is, the cost of taking
action expressed as a fraction of potential loss incurred if the protective action
is not taken and adverse event occurs (Richardson, 2000). Because the cost/
loss ratio can be different for different forecast users, the value score is gener-
ally plotted with respect to various cost-loss situations; in case of probabilistic
Under forecasts, it is necessary to compute relative value curves for the entire range
prediction of probabilities, then select the optimal values (the upper envelope of the
relative value curves) to represent the value of the probabilistic forecast sys-
tem. The approach was basically developed for discrete variables (such as
Large alert/no alert) while recently Laio and Tamea (2007) proposed a modification
forecast of the mentioned approach to deal with the evaluation of probabilistic
0 forecasts of continuous-valued variables.
0 1 zi The evaluation of the economical value of a forecast, mostly used in
Figure 22.4 Possible outcomes of a probability plot of the predictive cumulative
meteorology and recently extended to hydrological applications, is useful
probability values zi corresponding to the observations, versus their corresponding when there is no specific application. In hydrological applications to reser-
empirical frequency distribution function, represented as the rank order divided by the voirs, where a long historical record of previous management is available, it is
sample size Ri / n or by the Weibull plotting position. Deviations from the bisector line, more convenient to use combined prediction management in hindcast mode
which correspond to a reliable forecast, denote interpretable deficiencies related to the and to compare the obtained economical result to the historical management
spread or the prediction bias of the probabilistic forecast. From Laio and Tamea, 2007. performances, similarly to what described in Chap. 26, Sec. 26.8.3. The
Observed data that fall inside the band (%)
100 100
Observed data falling inside the band (%)
90 90
80 80
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Uncertainty band around the mean (%) Uncertainty band around the mean (%)
Figure 22.5 Probability plot: theoretical and empirical cumulated distribution of residuals sampled at 10% increments in calibration (left) and verification (right) periods.
From Coccia and Todini, 2011.
Figure 22.5 shows an example of successful assessment of the predictive benefits of the combined prediction/decision can be better understood by the
distribution on the Baron Fork river in the United States (Coccia and Todini, water resources managers.
2011). As can be observed, both in calibration and verification periods the
empirical cumulated distribution show limited deviations from the theoretical
45° line. REFERENCES
Among numerical SS, it is worth mentioning the continuous ranked prob- Anderson M. G. and P. D. Bates, Model Validation: Perspectives in
ability score (CRPS), which is a further generalization of the RPS to continuous Hydrological Science, John Wiley & Sons, 2001.
predictands: CRPS encompasses information about reliability, resolution, and Aumann, C. A., “A methodology for developing simulation models of com-
uncertainty within a single scalar measure of forecast quality (Hersbach, 2000). plex systems,” Ecological Modelling, 202: 385–396, 2007, doi:10.1016/ j.ecol-
model.2006.11.005.
Economical Values of Forecasts Bartholmes, J. C., J. Thielen, M. H. Ramos, and S. Gentilini, “The European
A common used approach of meteorologists for evaluating forecasts requires Flood Alert System EFAS—Part 2: statistical skill assessment of probabilistic
the definition of a cost-loss function to determine the potential “value” of a and deterministic operational forecasts,” Hydrology and Earth System Sciences,
forecast and to obtain information to be used within a decision making 13 (2):141–153, 2009.
framework. One way of quantifying the potential value of a forecast and more Beldring, S., “Multi-criteria validation of a precipitation–runoff model,”
appropriately of a forecasting system is to use the concept of potential eco- Journal of Hydrology, 257: 189–211, 2002.
nomic value, estimated using simple cost/loss models. This approach has Bennett, A. F., Inverse Methods in Physical Oceanography, Cambridge
been originally proposed for the evaluation of probabilistic forecasts of dis- University Press, Cambridge, 1992, p. 346.
crete-valued variables (Epstein, 1969; Murphy, 1969). Berger, J. O., Statistical Decision Theory and Bayesian Analysis, 2nd ed.,
In this cost/loss model, taking actions, on the basis of a forecast on the New York: Springer-Verlag,1985, ISBN:0-387-96098-8, MR 0804611.
occurrence of a future event, is associated with a cost C, which is implied by Bernardo, J. M. and A. F. M. Smith, Bayesian Theory, John Wiley & Sons,
an alarm issued by the user to protect his environment. By contrast, when no Chichester,1994.
22_Singh_ch22_p22.1-22.20.indd 15 8/22/16 1:32 PM

Beven, K. J. and A. M. Binley, “The future of distributed models: model Cukier, R. I., H. B. Levine, and K. E. Shuler, “Nonlinear sensitivity analysis
calibration and uncertainty prediction,” Hydrological Processes, 6: 279–298, of multiparameter model systems,” Journal of Computational Physics, 26: l–42,
1992. 1978.
Biondi, D., G. Freni, V. Iacobellis, G. Mascaro, and A. Montanari, Cunge, J. A., “On the subject of a flood propagation computation method
“Validation of hydrological models: conceptual basis, methodological (Muskingum method), Delft, The Netherlands,” Journal of Hydrological
approaches and a proposal for a code of practice,” Physics and Chemistry of the Research, 7 (2): 205–230, 1969.
Earth, 42–44: 70–76, 2012. Daley, R., Atmospheric Data Analysis, Cambridge University Press, New
Bogner, K. and F. Pappenberger, “Multiscale error analysis, correction, and York, 1991.
predictive uncertainty estimation in a flood forecasting system,” Water Daniel, C., “One-at-a-time-plans,” Journal of the American Statistical
Resources Research, 47: W07524, 2011,http://dx.doi.org/10.1029/ 2010WR00 Association, 68: 353–360, 1973.
9137. Dawson, C. W., R. J. Abrahart, and L. M. See, “HydroTest: a web-based
Boni G., D. Entekhabi, and F. Castelli, “Land data assimilation with satellite toolbox of evaluation metrics for the standardized assessment of hydrological
measurements for the estimation of surface energy balance components and forecasts,” Environmental Modelling & Software, 22: 1034–1052, 2007.
surface control on evaporation,” Water Resources Research, 37 (6): 1713–172, Deb, K., A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multi-
2001. objective genetic algorithm: NSGA-II,” IEEE Transactions on Evolutionary
Bouttier, F. and P. Courtier, Data Assimilation Concepts and Methods, Computation, 6 (2): 182–197, 2002.
ECMWF Training Notes, Available online at: http://www.ecmwf.int. De Lannoy, G. J. M., R. H. Reichle, K. R. Arsenault, P. R. Houser, S. Kumar,
Box, G. E. P. and N. R. Draper, Empirical Model Building and Response N. E. C. Verhoest, and V. R. N. Pauwels, “Multiscale assimilation of Advanced
Surfaces, John Wiley & Sons, New York, 1987. Microwave Scanning Radiometer EOS snow water equivalent and Moderate
Box, G. E. P. and K. B. Wilson, “On the experimental attainment of opti- Resolution Imaging Spectroradiometer snow cover fraction observations in
mum conditions (with discussion),” Journal of the Royal Statistical Society northern Colorado,” Water Resources Research, 48: W01522, 2012,
Series B, 13 (1): 1–45, 1951. doi:10.1029/2011WR010588.
Boyle, D. P., H. V. Gupta, and S. Sorooshian, Multicriteria calibration of De Groot, M., Optimal Statistical Decisions, Wiley Classics Library (origi-
hydrologic models, Calibration of Watershed Models, edited by Q. Duan, H. V. nally published in 1970), 2004, ISBN:0-471-68029-X.
Gupta, S. Sorooshian, A. N. Rousseau, and R. Turcotte, American Geophysical Dooge, J. C. I., Linear Theory of Hydrologic Systems, USDA Technical
Union, Washington, D. C., 2003, doi:10.1002/9781118665671.ch14. Bulletin 1468, US Department of Agriculture, Washington, D. C., 1973.
Bratseth, A. M., “Statistical interpolation by means of successive correc- Doucet, A., On sequential Monte Carlo methods for Bayesian filtering,
tions,” Tellus, 38A: 439–447, 1986. Department of Engineering, University of Cambridge, Technical Report
Bree, T., “The stability of parameter estimation in the general linear model,” CUED/F-INFENG/TR.310, 1998, p. 26.
Journal of Hydrology, 37: 47–66, 1978. Downing, D. J., R. H. Gardner, and E. O. Hoffman, “An examination of
Brocca, L., F. Melone, T. Moramarco, W. Wagner, V. Naeimi, Z. Bartalis, and response-surface methodologies for uncertainty analysis in assessment mod-
S. Hasenauer, “Improving runoff prediction through the assimilation of the els,” Technometrics, 27: 151–163, 1985.
ASCAT soil moisture product,” Hydrology and Earth System Sciences, 14: Draper, D. and M. Krnjajic, Calibration Results for Bayesian Model
1881–1893, 2010, doi:10.5194/hess-14-1881-2. Specification, Technical Report, Department of Applied Mathematics and
Bruen, M. and J. C. I. Dooge, “An efficient and robust method for estimat- Statistics, University of California, Santa Cruz, 2013, https://users.soe.ucsc.
ing unit hydrograph ordinates,” Journal of Hydrology, 70: 1–24, 1984. edu/~draper/draper-krnjajic-2013-draft.pdf.
Cacuci, D. G., Sensitivity & Uncertainty Analysis, Volume 1: Theory, Duan, Q., S. Sorooshian, and V. Gupta, “Effective and efficient global opti-
Chapman & Hall, 2003. mization for conceptual rainfall–runoff models,” Water Resources Research, 28
Carpenter, T. M., K. P. Georgakakos, and J. A. Sperfslagea, “On the para- (4): 1015–1031, 1992.
metric and NEXRAD-radar sensitivities of a distributed hydrologic model Eagleson, P. S., R. R. Mejia, and F. March, “Computation of optimum realiz-
suitable for operational use,” Journal of Hydrology, 253: 169–193, 2001. able unit hydrographs,” Water Resources Research, AGU2 (4): 755–764, 1966.
Carter, C. K. and R. Khon, “On Gibbs sampling for state-space models,” Efstratiadis, A. and D. Koutsoyiannis, “One decade of multi-objective cali-
Biometrika, 81: 641–653, 1994. bration approaches in hydrological modelling: a review,” Hydrological Sciences
Castelli, F., D. Entekhabi, and E. Caporali, “Estimation of surface heat flux Journal, 55 (1): 58–78, 2010.
and an index of soil moisture using adjoint-state surface energy balance,” Engeland, K., B. Renard, I. Steinsland, and S. Kolberg, “Evaluation of statis-
Water Resources Research, 35 (10): 3115–3125, 1999. tical models for forecast errors from the HBV model,” Journal of Hydrology,
Charney, J. G., M. Halem, and R. Jastrow, “Use of incomplete historical data 384: 142–155, 2010.
to infer the present state of the atmosphere,” Journal of Atmospheric Sciences, Entekhabi, D., H. Nakamura, and E. G. Njoku, “Solving the inverse prob-
26: 1160–1163, 1969. lem for soil moisture temperature profiles by sequential assimilation of mul-
Chen, Z., “Bayesian filtering: from Kalman filters to particle filters, and tifrequency remotely sensed observations,” IEEE Transactions on Geoscience
beyond, 2003, CiteSeerX: 10.1.1.107.7415. and Remote Sensing, 32 (2): 438–448, 1994.
Chiu, C-L. (ed.), Application of Kalman Filter to Hydrology, Hydraulics and Evensen, G., “Sequential data assimilation with a non-linear quasi-geo-
Water Resources, University of Pittsburgh, PA, 1978. strophic model using Monte Carlo methods to forecast error statistics,”
Clark, M. P., D. E. Rupp, R. A. Woods, X. Zheng, R. P. Ibbitt, A. G. Slater, Journal of Geophysical Research, 99 (C5): 10143–10162, 1994.
J. Schmidt, M. J. and Uddstrom, “Hydrological data assimilation with the Evensen, G., “The Ensemble Kalman Filter: theoretical formulation and
ensemble Kalman filter: use of streamflow observations to update states in a practical implementation,” Ocean Dynamics, 53: 343–367, 2003.
distributed hydrological model, Advances in Water Resources, 31: 1309–1324, Feng, Z., X. Wen-fang, and L. Xi, “Overview of nonlinear Bayesian filtering
2008, doi:10.1016/j.advwatres.2008.06.005. algorithm,” Procedia Engineering, 15: 489–495, 2011.
Coccia, G. and E. Todini, “Recent developments in predictive uncertainty Fenicia, F., D. Kavetski, and H. H. G. Savenije, “Elements of a flexible
assessment based on the Model Conditional Processor approach,” Hydrology approach for conceptual hydrological modeling: 1. Motivation and theoretical
and Earth System Sciences, 15: 3253–3274, 2011, doi:10.5194/hess-15-3253- development, Water Resources Research, 47: W11510, 2011, doi:10.1029/2010
2011. WR010174.
Cooper, D. M. and E. F. Wood, “Identification of multivariate time series Fischer, G., Ornithologische Monatsberichte, Berlin, Vol. 44, Issue 2, 1936.
and multivariate input-output models,” Water Resources Research, 18 (4): Fischer, G., Statistiches Jahrbuch Deutscher Gemeinden, Jena, 1932–1938,
937–946, 1982a. pp. 27–33.
Cooper, D. M. and E. F. Wood, “Identification of multivariate time series Fischer, G., Ornithologische Monatsberichte, Berlin, Vol. 48, Issue 1, 1940.
and multivariate input-output models: application to rainfall-runoff pro- Fisher, R. A.,“On the mathematical foundations of theoretical statistics,”
cesses,” Water Resources Research, 18 (5): 1352–1364, 1982b. Philosophical Transactions on Royal Society of London Series A, 222: 309–368,
Cox, R. T., “Probability, frequency and reasonable expectation,” American 1922.
Journal of Physics, 14: 1–13, 1946. Franz, K. J. and T. S. Hogue, “Evaluating uncertainty estimates in hydro-
Cox, R. T., The Algebra of Probable Inference, Johns Hopkins University logic models: borrowing measures from the forecast verification community,”
Press, BL, 1961. Hydrology and Earth System Sciences,15: 3367–3382, 2011.
Cramér, H., Mathematical Methods of Statistics, Princeton University Press, Frey, H. C. and S. R. Patil, “Identification and review of sensitivity analysis
Princeton, NJ, 1946, ISBN:0-691-08004-6,OCLC 185436716. methods,” Risk Analysis, 22 (3): 553–578, 2002.
22_Singh_ch22_p22.1-22.20.indd 16 8/22/16 1:32 PM

REFERENCES 22-17
Galantowicz, J., D. Entekhabi, and E. Njoku, “Tests of sequential data Jazwinski, A. H., Stochastic Processes and Filtering Theory, Academic Press,
assimilation for retrieving profile soil moisture and temperature from 1970.
observed L-band radiobrightness,” IEEE Transactions on Geoscience and Jolliffe, I. T. and D. B. Stephenson, Forecast Verification: A Practitioner’s
Remote Sensing, 37 (4): 1860–1870, 1999. Guide in Atmospheric Science, John Wiley & Sons, Chichester, 2003, ISBN
Gauss C. F., Theoria Motus Corporum Coelestium in sectionibus conicis 0-471-49759-2.
solem ambientium (Theorie der Bewegung der Himmelskörper, die Sonne in Julier, S. and J. K. Uhlmann, “A new extension of the Kalman filter to non-
Kegelschnitten umkreisen), Theory of the Motion of Heavenly Bodies Moving linear systems,” Proceedings of the AeroSense: The 11th International
about the Sun in Conic Sections, English translation by C. H. Davis, Dover, Symposium on Aerospace/Defence Sensing, Simulation and Controls, 1997.
New York, reprinted 1963. Kalman, R. E., “A new approach to linear filtering and prediction prob-
Gelb, A. (Ed.), Applied Optimal Estimation, MIT Press, Cambridge, MA, lems,” Transaction of the ASME—Journal of Basic Engineering, 82 (Series D):
1974, p. 374. 35–45, 1960.
Gelfand, A. E. and A. F. M. Smith, “Sampling based approaches to calculat- Kalman, R. E. and R. S. Bucy, “New results in linear filtering and prediction
ing marginal densities,” Journal of the American Statistical Association, 85: theory,” Transaction of the ASME—Journal of Basic Engineering, 83: 95–108,
389–409, 1990. 1961.
Georgakakos, K. P., “A generalized stochastic hydrometeorological model Kendall, M. G. and A. Stuart, The Advanced Theory of Statistics, 2nd ed.,
for flood and flash-flood forecasting, 1. Formulation,” Water Resources Charles Griffin and Co., London, Vol. II, 1967.
Research, 22 (13): 2083–2095, 1986a. Kirchner, J. W., “Getting the right answers for the right reasons: linking
Georgakakos, K. P., “A generalized stochastic hydrometeorological model measurements, analyses, and models to advance the science of hydrology,”
for flood and flash-flood forecasting, 2. Case studies,” Water Resources Water Resources Research, 42: W03S04, 2006, doi:10.1029/2005WR004362.
Research, 22 (13): 2096–2106, 1986b. Kitanidis, P. K. and R. L. Bras, “Real-time forecasting with a conceptual
Giustarini, L., P. Matgen, R. Hostache, A. Montanari, D. Plaza, V. R. N. hydrologic model: 1. Analysis of uncertainty,” Water Resources Research, 16
Pauwels, G. J. M. De Lannoy, et al., “Assimilating SAR-derived water level data (6): 1025–1033, 1980a.
into a hydraulic model: a case study,” Hydrology and Earth System Sciences, 15: Kitanidis, P. K. and R. L. Bras, “Real time forecasting with a conceptual
2349–2365, 2011. hydrologic model: 2. Applications and results,” Water Resources Research, 16
Gneiting, T., F. Balabdaoui, and A. Westveld, “Verifying probabilistic fore- (6):1034–1044, 1980b.
casts: calibration and sharpness,” Proceedings of the Workshop on Ensemble Klemeš, V., “Operational testing of hydrological simulation models,”
Weather Forecasting in the Short to Medium Range, Val-Morin, QC, Canada, Hydrological Sciences Journal, 31 (1): 13–24, 1986.
www.cdc.noaa.gov/people/tom.hamill/ef_workshop_2003_ schedule.html. Labarre, D., E. Grivel, Y. Berthoumieu, E. Todini, and M. Najim,
Gordon, N. J., D. J. Salmond, and A. F. M. Smith, “Novel approach to non- “Consistent estimation of autoregressive parameters from noisy observa-
linear/non-Gaussian Bayesian state estimation,” Proceedings of the Institute of tions based on two interacting Kalman filters,” Signal Processing, 86 (10):
the Electrical Engineering, F, 140 (2): 107–113, 1993. 2863–2876, 2006.
Griewank, A., Evaluating derivatives. Principles and techniques of algorith- Lahoz, W.A. and P. Schneider, “Data assimilation: making sense of Earth
mic differentiation, Frontiers in Applied Mathematics, SIAM, Philadelphia, observation,” Frontiers in Environmental Science, 2: 1–28, 2014, doi:10.3389/
2000. fenvs.2014.00016.
Gupta, N. K. and R. K. Mehra, “Computational aspects of maximum likeli- Laio, F. and S. Tamea, “Verification tools for probabilistic forecasts of con-
hood estimation and reduction of sensitivity function calculations,” IEEE tinuous hydrological variables,” Hydrology and Earth System Sciences, 11 (4):
Transactions on Automatic Control, 19: 774–783, 1974. 1267–1277, 2007.
Gupta, H. V., S. Sorooshian, and P. O. Yapo, “Toward improved calibration Legates, D. R. and G. J. McCabe, “Evaluating the use of ‘goodness-of-fit’
of hydrological models: multiple and noncommensurable measures of infor- measures in hydrologic and hydroclimatic model validation,” Water Resources
mation,” Water Resources Research, 34: 751–763, 1998. Research, 35 (1): 233–241, 1999.
Gupta H. V., T. Wagener, and Y. Liu, “Reconciling theory with observations: Lehmann, E. L. and G. Casella, Theory of Point Estimation, 2nd ed.,
elements of a diagnostic approach to model evaluation,” Hydrological Springer, 1998, ISBN:0-387-98502-6.
Processes, 22: 3802–3813, 2008. Liu, Y. and H. V. Gupta, “Uncertainty in hydrologic modeling: toward an
Gupta, H. V., H. Kling, K. K. Yilmaz, and F. G. Martinez, “Decomposition integrated data assimilation framework,” Water Resources Research, 43:
of the mean square error and NSE performance criteria: implications for W07401, 2007, doi:10.1029/2006WR005756.
improving hydrological modelling,” Journal of Hydrology, 377: 80–91, 2009. Liu, Y., A. H. Weerts, M. Clark, H-J. Hendricks Franssen, S. Kumar,
Hamby, D. M., “A review of techniques for parameter sensitivity analysis of H. Moradkhani, D-J. Seo, et al., “Advancing data assimilation in operational
environmental models,” Environmental Monitoring and Assessment, 32 (2): hydrologic forecasting: progresses, challenges, and emerging opportunities,”
135–154, 1994. Hydrology and Earth System Sciences,16: 3863–3887, 2012.
Hersbach, H., “Decomposition of the continuous ranked probability score Liu, Z., M. V. L. Martina, and E. Todini, “Flood forecasting using a fully
for ensemble prediction systems,” Weather and Forecasting, 15: 559–569, 2000. distributed model: application of the TOPKAPI model to the Upper Xixian
Hino, M., “On-line prediction of hydrologic systems,” Proceedings of the XV Catchment,” Hydrology and Earth System Sciences, 9: 347–364, 2005.
Conference IAHR, Istanbul, 1973, pp. 121–129. Lorenc, A., “A global three-dimensional multivariate statistical interpola-
Houser, P. R., W. J. Shuttleworth, J. S. Famiglietti, H. V. Gupta, K. H. Syed, tion scheme,” Monthly Weather Review, 109: 701–721, 1981.
and D. C. Goodrich, “Integration of soil moisture remote sensing and hydro- Madsen, H., “Automatic calibration of a conceptual rainfall–runoff model
logic modeling using data assimilation,” Water Resources Research, 34: 3405– using multiple objectives,” Journal of Hydrology, 235: 276–288, 2000.
3420, 1998. Mantovan, P., A. Pastore, and S. Tonellato, Recursive estimation of system
Houser, P. R., G. J. M. DeLannoy, and J. P. Walker, Landsurface data parameter in environmental time series, Classification and Data Analysis,
assimilation, Data Assimilation: Making Sense of Observations, edited by W.A. edited by M. Vichi and O. Optiz, Springer, 1999, pp. 311–318.
Lahoz, B. Khattatov, and R. Ménard, Springer, Berlin, 2010, pp. 549–597. Mason, I., “A model for assessment of weather forecasts,” Australian
Huang, X-Y. and X. Yang, Variational data assimilation with the Lorenz Meteorological Magazine, 30: 291–303, 1982.
model: high-resolution limited area model (HIRLAM), Technical Report 26, Matott, L. S., J. E. Babendreier, and S. T. Purucker, “Evaluating uncertainty
Danish Meteorol. Inst., Copenhagen, 1996, p. 42. in integrated environmental models: a review of concepts and tools,” Water
Hwang, D., D. W. Byun, and M. T. Odman, “An automatic differentiation Resources Research, 45: 2009, doi:10.1029/2008WR007301.
technique for sensitivity analysis of numerical advection schemes in air qual- McCarthy, G. T., The Unit Hydrograph and Flood Routing, Unpublished
ity models,” Atmospheric Environment, 31 (6): 879–888, 1997. manuscript presented at a conference of the North Atlantic Division, U.S.
Iman, R. L. and W. J. Conover, “The use of the rank transformation in Army, Corps of Engineers, June 24, 1938.
regression,” Technometrics, 21: 499–509, 1979. McCarthy, G. T., Flood routing, Flood Control, The Engineer School, Fort
Iman, R. L. and J. C. Helton, “An investigation of uncertainty and sensitivity Belvoir, VA, Chap. V, 1940, pp. 127–177.
analysis techniques for computer models,” Risk Analysis , 8 (1): 71–90, 1988. McElhoe, B. A., “An assessment of the navigation and course corrections
Jachner, S. and K. G. van den Boogaart, “Statistical methods for the qualita- for a manned flyby of Mars or Venus, IEEE Transactions on Aerospace and
tive assessment of dynamic models with time delay (R Package qualV),”Journal Electronic Systems, 613–623, 1966.
of Statistical Software, 22 (8): 30, 2007. McKay, M. D., Evaluating Prediction Uncertainty, Technical Report
Jaynes, E. T., Probability Theory: The Logic of Science, Cambridge University NUREG/CR-6311, US Nuclear Regulatory Commission and Los Alamos
Press, Cambridge, 2003. National Labortory, Los Alamos, NM, 1995.
22_Singh_ch22_p22.1-22.20.indd 17 8/22/16 1:32 PM

McKay, M. D., R. J. Beckman, and W. J. Conover, “A comparison of three variational data assimilation approach,” Water Resources Research, 37: 2353–
methods for selecting values of input variables in the analysis of output from 2364, 2001.
a computer code,” Technometrics, 21: 239–245, 1979. Reichle, R. H., “Data assimilation methods in the Earth sciences,” Advances
McLaughlin, D., “An integrated approach to hydrologic data assimilation: in Water Resources, 31: 1411–1418, 2008, doi:10.1016/j.advwatres.2008.01.001.
interpolation, smoothing, and filtering,” Advances in Water Resources, 25: Reichle, R. H., S. V. Kumar, S. P. P. Mahanama, R. D. Koster, and Q. Liu,
1275–1286, 2002. “Assimilation of satellite-derived skin temperature observations into land
Mehra, R. K., “On the Identification of variances and adaptive Kalman fil- surface models,” Journal of Hydrometeorology, 11: 1103–1122, 2010,
tering.” IEEE Transactions on Automatic Control, 15: 175–184, 1970. doi:10.1175/2010JHM1262.1.
Melching, C. S., Reliability estimation, Computer Models of Watershed Richardson, D. S., “Skill and relative economic value of the ECMWF
Hydrology, edited by V. P. Singh, Water Resources Publication, Highlands ensemble prediction system,” Quarterly Journal of the Royal Meteorological
Ranch, CO, 1995, pp. 69–118. Society, 126: 649–667, 2000.
Minsky, M. L. and S. S. Papert, Perceptrons: An Introduction to Computational Rodell, M. and P. R. Houser, “Updating a land surface model with MODIS-
Geometry, MIT Press, Cambridge, MA, 1969. derived snow cover,” Journal of Hydrometeorology, 5: 1064–1075, 2004.
Mohammadi, A. and A. Asif, Consensus-based distributed unscented Rosenblatt, F., “The perceptron: aprobabilistic model for information stor-
Kalman filter, Proceedings IEEE Statistical Signal Processing Workshop (SSP), age and organization in the brain,” Cornell Aeronautical Laboratory,
978-1-4577-0568-7/11: 2011, pp. 237–240, http://www.eecs.yorku.ca/~asif/ Psychological Review, 65(6): 386–408,1958, doi:10.1037/h0042519.
publications/p237-ssp.pdf. Roulin, E. and S. Vannitsem, “Skill of medium-range hydrological ensem-
Moradkhani, H., S. Sorooshian, H. V. Gupta, and P. Houser, “Dual state– ble predictions,” Journal of Hydrometeorology, 6: 729–744, 2005.
parameter estimation of hydrologic models using ensemble Kalman filter,” Roulin, E.,“Skill and relative economic value of medium-range hydrological
Advances in Water Resources, 28: 135–147, 2005a. ensemble predictions,” Hydrology and Earth System Sciences, 11: 725–737, 2007.
Moradkhani, H., K-L. Hsu, H. V. Gupta, and S. Sorooshian, “Uncertainty Rui Y. and Y. Chen, “Better proposal distributions: object tracking using
assessment of hydrologic model states and parameters: sequential data unscented particle filter,” Proceedings of the IEEE International Conference on
assimilation using the particle filter,” Water Resources Research, 41: W05012, Computer Vision and Pattern Recognition, 2001, pp. 786–793, http://research.
2005b, doi:10.1029/2004WR003604. microsoft.com/en-us/um/people/yongrui/ps/cvpr01a.pdf.
Morris, M. D., “Factorial sampling plans for preliminary computational Sage, A. P. and G. W. Husa, “Adaptive filtering with unknown prior statis-
experiments,” Technometrics, 33: 161–174, 1991. tics,” Proceedings of the 1969 Joint Automatic Control Conference, 1969, pp.
Murphy, A. H., “Measures of the utility of probabilistic predictions in cost- 760–796.
loss ratio decision situation in which knowledge of the cost-loss ratio is Saltelli, A. and I. M. Sobol’, “About the use of rank transformation in sensi-
incomplete,” Journal of Applied Meteorology, 8: 863–873, 1969. tivity analysis of model output,” Reliability Engineering & System Safety, 50
Murphy, A. H., “What is a good forecast? An essay on nature of goodness (3): 225–239, 1995.
in weather forecasting,” Weather Forecasting, 5 (8): 281–293, 1993. Saltelli, A., S. Tarantola, and K. P-S. Chan, “A quantitative model indepen-
Myers, R. H. and D. C. Montgomery, Response Surface Methodology: Process dent method for global sensitivity analysis of model output,” Technometrics,
and Product in Optimization Using Designed Experiments, John Wiley & Sons, 41 (1): 39–56, 1999.
New York, 1995. Saltelli, A., K. Chan, and M. Scott (Eds.), Sensitivity Analysis, John Wiley &
Nash, J. E., The form of the instantaneous unit hydrograph, IUGG General Sons, Chichester, 2000.
Assembly of Toronto, IAHS Publication, Vol. III,1958, pp.114–121. Saltelli, A., S. Tarantola, F. Campolongo, and M. Ratto, Sensitivity Analysis
Nash, J.E.,“A unit hydrograph study with particular reference to British in Practice: A Guide to Assessing Scientific Models, Wiley, Hoboken, NJ, 2004.
catchments,” Proceedings of the Institution of Civil Engineers, 17: 249–282, Saltelli, A., M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D. Gatelli, M.
1960. Saisana, and S. Tarantola, Global Sensitivity Analysis: The Primer, John Wiley
Nash, J. E. and J. V. Sutcliffe, “River flow forecasting through conceptual & Sons, 2008.
models,” Journal of Hydrology, 10: 282–290, 1970. Seibert, J., “Multi-criteria calibration of a conceptual runoff model using a
Natale, L. and E. Todini, “A stable estimation for linear models:1. Theoretical genetic algorithm,” Hydrology and Earth System Sciences, 4 (2): 215–224, 2000.
development and Monte-Carlo experiments,” Water Resources Research, 12 Seibert, J. and J. J. McDonnell, “On the dialog between experimentalist and
(4): 667–671, 1976a. modeler in catchment hydrology: use of soft data for multicriteria model
Natale, L. and E. Todini, “A stable estimator for linear models: 2. Real world calibration,” Water Resources Research, 38: 2002, doi:10.1029/2001WR000978.
hydrologic applications,” Water Resources Research, 12(4): 672–675, 1976b. Seo, D. J., V. Koren, and N. Cajina, “Real-time variational assimilation of
Newey, W. K. and D. McFadden, Large sample estimation and hypothesis hydrologic and hydrometeorological data into operational hydrologic fore-
testing, Handbook of Econometrics, edited by R. Engle, and D. McFadden, casting,” Journal of Hydrometeorology, 4 (3): 627–641, 2003.
Elsevier Science, Chap. 35, Vol. 4, pp. 2111–2245, 1994, ISBN 0-444-88766-0. Seo, D-J., L. Cajina, R. Corby, and T. Howieson, “Automatic state updating
Oakley, J. and A. O′Hagan, “Probabilistic sensitivity analysis of complex for operational streamflow forecasting via variational data assimilation,”
models: a Bayesian approach,” Journal of the Royal Statistical Society, Series B, Journal of Hydrology, 367: 255–275, 2009, doi:10.1016/j.jhydrol.2009.01.019.
66: 751–769, 2004. Sieber, A. and S. Uhlenbrook, “Sensitivity analyses of a distributed catch-
Oreskes, N., The role of quantitative models in science, The Role of Models: ment model to verify the model structure,” Journal of Hydrology, 310: 216–
Ecosystem Science, edited by C. D. Canham, et al., Princeton University Press, 235, 2005.
Princeton, 2003. Smith, A. F. M. and A. E. Gelfand, “Bayesian resampling without tears: a
Pappenberger, F., I. Iorgulescu, and K. Beven, “Sensitivity analysis based on sampling-resampling perspective,” The American Statistician, 46 (2): 84–88,
regional splits and regression trees (SARS-RT),” Environmental Modelling 1992.
Software, 21: 976–990, 2006. Smith, G. L., S. F. Schmidt, and L. A. McGee, Application of statistical filter
Pareto, V., Cours d’économie politique professé à l’Université de Lausanne (in theory to the optimal estimation of position and velocity on board a circumlunar
French), Lausanne, Rouge, Vol. 2, 1896–1897. vehicle, NASA Technical Report R-135, USA, 1962.
Pareto, V., Manuale d’Economia Politica (in Italian), Società Editrice Sobol, I. M., “Sensitivity analysis for nonlinear mathematical models,”
Libraria, Milano, 1906. Mathematical Modeling & Computational Experiment, 1: 407–414, 1993.
Power, M., “The predictive validation of ecological and environmental Spear, R. C. and G. M. Hornberger, “Eutrophication in peel inlet, II.
models,” Ecological Modelling, 68: 33–50, 1993. Identification of critical uncertainty via generalised sensitivity analysis,”
Ramsey, F. P., Truth and probability, The Foundations of Mathematics and Water Research, 14: 43–49, 1980.
Other Logical Essays, edited by R. B. Braithwaite, Kegan Paul, London, 1926, Stauffer, D. R. and N. L. Seaman, “Use of four-dimensional data assimila-
pp. 156–198. tion in a limited-area mesoscale model, Part I: Experiments with synoptic-
Rao, C. R., “Information and the accuracy attainable in the estimation of scale data,” Monthly Weather Review, 118: 1250–1277, 1990.
statistical parameters,” Bulletin of the Calcutta Mathematical Society, 37: Sugawara, M., The flood forecasting by a series storage type model:
81–89, 1945. International Symposium on Floods and their Computation, IAHS Publ. 85,
Refsgaard, J. C., “Parameterisation, calibration and validation of distributed IAHS Press, Wallingford, 1967, pp. 1–6.
hydrological models,” Journal of Hydrology, 198: 69–97, 1997. Sugawara, M., Chapter 6: Tank model, Computer Models of Watershed
Reichle, R.,D. Entekhabi, and D. McLaughlin, “Downscaling of radio Hydrology, edited by V. P. Singh, Water Resources Publications, Littleton, CO,
brightness measurements for soil moisture estimation: a four-dimensional 1995.
22_Singh_ch22_p22.1-22.20.indd 18 8/22/16 1:32 PM

REFERENCES 22-19
Tang Y., P. Reed, T. Wagener, and K. van Werkhoven, “Comparing sensitiv- strengths of global optimization and data assimilation,” Water Resources
ity analysis methods to advance lumped watershed model identification and Research, 41: W01017, 2005, doi:10.1029/2004WR003059.
evaluation,” Hydrology and Earth System Sciences, 11 (2): 793–817, 2007, Vrugt, J. A. and B. A. Robinson, “Improved evolutionary search from geneti-
doi:10.5194/hess-11-793-2007. cally adaptive multi-method search,” Proceedings of the National Academy of
Thiemann, T., M. Trosset, H. Gupta, and S. Sorooshian, “Bayesian recursive Sciences USA, 104 (3): 708–711, 2007, doi:10.1073/pnas.0610471104.
parameter estimation for hydrologic models,” Water Resources Research, 37 Walker, J. P., G. R. Willgoose, and J. D. Kalma, “One-dimensional soil mois-
(10): 2521–2535, 2001. ture profile retrieval by assimilation of near-surface observations: a comparison
Thyer, M., B. Renard, D. Kavetski, G. Kuczera, S. Franks, and S. Srikanthan, of retrieval algorithms,” Advances in Water Resources, 24 (6): 631–650, 2001.
“Critical evaluation of parameter consistency and predictive uncertainty in Walker, J. P. and P. R. Houser, “A methodology for initializing soil moisture
hydrological modelling: a case study using Bayesian total error analysis,” in a global climate model: assimilation of near surface soil moisture observa-
Water Resources Research, 45: W00B14, 2009, http://dx.doi.org/10.1029/200 tions,” Journal of Geophysical Research, 106 (D11): 11761–11774, 2001.
8WR006825. Walker, J. P. and P. R. Houser, Hydrologic data assimilation, Advances in
Todini, E., A. Szollosi-Nagy, and E. F. Wood, Adaptive state-parameter Water Science Methodologies, edited by A. Aswathanarayana, A.A. Balkema,
estimation algorithm for real time hydrologic forecasting: a case study, The Netherlands, 2005, p. 230.
Proceedings of IIASA/WMO Workshop on the Recent Developments in Real Wan, E. and R. van der Merwe, The unscented Kalman filter, Kalman
Time Forecasting/Control of Water Resources Systems, IIASA, Laxenburg, Filtering and Neural Networks, Chap. 7, edited by S. Haykin, John Wiley &
Austria, 1976. Sons, New York, 2001.
Todini, E., Mutually interactive state/parameter estimation (MISP) in Welles, E., Verification of River Stage Forecasts, PhD Dissertation, University
hydrological applications, Identification and Control in Environmental of Arizona, 2005, p. 157, http://www.nws.noaa.gov/oh/rfcdev/projects/
Systems, edited by G. C. Vansteenkiste, Elsevier, New York, 1978a. EdwinWellesV6.pdf.
Todini, E., Mutually interactive state/parameter estimation (MISP), Wiener, N., “The homogeneous chaos,” American Journal of Mathematics,
Application of Kalman Filter to Hydrology, Hydraulics and Water Resources, 60: 897–936, 1938.
edited by C-L. Chiu, University of Pittsburgh, Pittsburgh, PA, 1978b, pp. Wiener, N., Extrapolation, Interpolation, and Smoothing of Stationary Time
135–150. Series, John Wiley & Sons, New York, 1949, ISBN:0-262-73005-7.
Todini, E. and J. R. Wallis, A real time rainfall-runoff model for an online Wilks, D. S., Statistical methods in the atmospheric sciences, International
flood warning system, Application of Kalman Filter to Hydrology, Hydraulics Geophysics Series, Academic Press, ISBN-10:0127519653; ISBN-13:978-
and Water Resources, edited by C-L. Chiu, University of Pittsburgh, Pittsburgh, 0127519654, Vol. 59: 1995, p. 464.
PA, 1978, pp. 355–368. Wilmott, C. J., S. G. Ackleson, R. E. Davis, J. J. Feddema, K. M. Klink, D. R.
Todini, E., “Hydrological catchment modelling: past, present and future,” Legates, J. O’Donnell, and M. C. Rowe, “Statistics for the evaluation and com-
Hydrology and Earth System Sciences, 11: 468–482, 2007. parison of models,” Journal of Geophysical Research, 90: 8995–9005, 1985.
van der Merwe, R., J. F. G. de Freitas, A. Doucet, and E. Wan, The Wojcik, P. J., On-line estimation of signal and noise parameters and the
unscented particle filter, Technical Report CUED/F-INFENG/TR 380, adaptive Kalman filtering, Apporximate Kalman Filtering, edited by G. Chen,
Cambridge University Engineering Department, Cambridge, 2000. World Scientific Publishing, Singapore, 1993.
van der Molen, D. T. and J. Pinter, “Environmental model calibration under Wood, E. F. and P. E. O’Connell, Real Time Forecasting, Chapter 15, edited
different specifications: an application to the model SED,” Ecological by M. G. Anderson, and T. P. Burt, John Wiley & Sons, 1985.
Modelling, 68: 1–19, 1993. Yapo, P. O., H. V. Gupta, and S. Sorooshian, “Multi-objective global optimi-
van Griensven, A., T. Meixner, S. Grunwald, T. Bishop, M. Diluzio, and zation for hydrologic models,” Journal of Hydrology, 204: 83–97, 1998.
R. Srinivasan, “A global sensitivity analysis tool for the parameters of multi- Young, P. C., “Recursive approach to time series analysis,” Bulletin of the
variable catchment models,” Journal of Hydrology, 324 (1–4): 10–23, 2006. Institute of Mathematics and Its Applications, 12 (5/6): 209–224, 1974.
Vrugt, J. A., H. V. Gupta, L. A. Bastidas, W. Bouten, and S. Sorooshian, Young, P. C. and P. Whitehead, “A recursive approach to time-series analy-
“Effective and efficient algorithm for multi-objective optimization of hydro- sis for multivariable systems,” International Journal of Control, 25 (3): 457–
logic models,” Water Resources Research, 39 (8): 1214, 2003, 482, 1977.
doi:10.1029/2002WR001746. Zhao, Y., “An improved unscented particle filter with global sampling strat-
Vrugt, J. A., C. G. H. Diks, H. V. Gupta, W. Bouten, and J. M. Verstraten, egy,” Journal of Computational Engineering, 2014: 1–6, 2014, http://dx.doi.
“Improved treatment of uncertainty in hydrologic modeling: Combining the org/10.1155/2014/175820.
22_Singh_ch22_p22.1-22.20.indd 19 8/22/16 1:32 PM

Chapter 23
Bayesian Methods
BY
GEORGE KUCZERA, DMITRI KAVETSKI, BEN RENARD, AND MARK THYER
ABSTRACT assumed, at different times, to respond differently to identical inputs). In

addition, in statistical analysis, it is common to assume that imperfections in
The Bayesian framework provides a systematic approach to making inferences
the model structure and errors in the observed data (including initial and
in applications where uncertainty can be represented by probability distribu-
boundary conditions) can be described using probability distributions. When
tions. Its relevance and richness derives from its close alignment with the sci-
applying Bayesian inference, as with most other statistical techniques, the
entific method where hypotheses are formulated, tested and refined. This
hydrologist must formulate an explicit hypothesis of how inherent random-
chapter describes and illustrates the three core steps of the Bayesian approach.
ness, in combination with model and data errors, affects system dynamics and
It then provides an introduction to the computational tools needed to make an
the production of observable quantities.
inference, followed by an overview of diagnostic techniques to scrutinize
An integral part of the model hypothesis is the specification of parameters,
assumptions. It concludes with a brief review of applications in hydrology.
that is, quantities that need to be assigned a value to fully describe how the
system functions. For example, the baseflow recession may depend on the
23.1 INTRODUCTION
fraction of aquifer storage that discharges into the stream in a day—the frac-
The Bayesian framework provides a systematic approach to making infer- tion would typically be treated as a parameter.
ences (or drawing conclusions) in applications where the uncertainties in the
observed data and the uncertainty in the quantities of interest can be (approx- 23.2.1.1 Example: The Regression
imately) represented using probability distributions (Gelman et al., 2004). model hypothesis
Bayesian methods enjoy wide application in hydrology. This chapter provides The classical regression model has seen wide application in hydrology and
an overview of the Bayesian approach with emphasis on core ideas and offers a simple but insightful illustration of the Bayesian framework. Consider
insights. It starts by describing the main concepts that define the Bayesian a typical modeling scenario where the interest is to infer the system parameters
approach using two simple linear regression models. It then provides an θ describing the deterministic relationship between an observed system output
introduction to the computational tools needed to make an inference, fol- yt and one or more predictor variables xt . Predictor variables are defined as
lowed by an overview of diagnostic techniques to scrutinize assumptions. It observed variables needed by the model to produce a simulated output. In the
concludes by briefly reviewing applications in hydrology. context of time series analysis, the index t generally denotes time, in spatial
analysis it denotes a spatial location, and so forth. For example, in rainfall-
23.2 THE BAYESIAN INFERENCE FRAMEWORK runoff modeling applications, yt is the observed streamflow, and xt might
comprise rainfall and potential evapotranspiration.
The Bayesian framework derives its name from Thomas Bayes’ work (1763)
The key hypothesis underlying the (univariate) classical regression model
on the so-called “inverse probability problem,” and defines a procedure for
is that yt is a random realization from a Gaussian distribution given by
updating what is known about a quantity of interest as new information
becomes available. While this represents the mathematical core of the yt ← N ( f ( xt , θ ), σ ε2 ) (23.1a)
Bayesian framework, its relevance and richness comes from the fact that it is
closely aligned with the scientific method where hypotheses are formulated, or, equivalently,
tested, and refined. In this section the Bayesian framework is presented as a
three-step process, as articulated by Gelman et al. (2004). Two linear regres- ε t ← N (0, σ ε2 ), where ε t = yt − f ( xt , θ ) (23.1b)
sion examples illustrate its application and demonstrate the importance of
understanding the information content of data. The term f ( xt , θ ), called the regression function, is a function of the
observed predictors xt and parameters θ . As seen in Eq. (23.1), it defines the
23.2.1 Step 1: Formulating the Model Hypothesis expected value of the output yt when the regression error εt has an expected
Bayesian inference starts with the formulation of a hypothesis H that value of zero.
describes how the system of interest produces quantities that are observable. The uncertainty (or inherent randomness) in the relationship between the
For example, when the system of interest is a catchment, the hypothesis may predictors and output is assumed to be fully described by a Gaussian error εt
represent the runoff-generation mechanism in response to rainfall events. In with zero mean and constant variance σ ε2 . Further, it is assumed that the
formulating this hypothesis, the hydrologist will generally draw upon the cur- errors (εt )t =1,...,n are statistically independent. These assumptions, along with
rent paradigm describing the physics of the system and may also propose the assumed form of the regression function, represent the model hypothesis
additional assumptions and/or enhancements to that paradigm. that is subjected to scrutiny in the third step of the Bayesian framework.
A key element of a statistical hypothesis is the acknowledgement that the Importantly, it is stressed that the unknown parameters subject to inference
system—including the environmental system, its model, and the observation comprise both the regression function parameters θ as well as the regression
systems—is not deterministic. The loss of determinism may correspond to error parameters γ (here, γ = σ ε ). The generic notation Θ = (θ , γ ) is used to
the natural system itself being nondeterministic (e.g., if the catchment is denote the full list of inferred quantities.
23-1
23_Singh_ch23_p23.1-23.10.indd 1 8/22/16 1:33 PM

23-2 Bayesian Methods
23.2.2 Step 2: From Prior to Posterior Combining the prior and likelihood using Eq. (23.3) and applying a series of
The second step of the Bayesian approach involves a learning process in algebraic manipulations leads to the posterior pdf of θ1, which itself is a
which observations of system behavior denoted by the data D are used to Gaussian distribution,
update our knowledge about the parameters Θ of the model hypothesis H. In
θ1 | σ ε2 , y , x ~ N (θˆ1 , σ θ2 ) (23.6a)
the Bayesian approach, knowledge (or its opposite, uncertainty) about a
parameter is represented by a probability distribution—the tighter the prob- where θˆ1 and σ θ2 are, respectively, the posterior mean and variance of θ1:
ability distribution, the less uncertainty there is about the parameter.
The learning process is formally encapsulated by Bayes equation  n  n
p(D | Θ, H) p(Θ | H)
 ∑ x t yt  ∑ x t yt
θ 
p(Θ | D , H) = (23.2) θˆ1 = σ θ2  o2 + t =1 2  → t =1n as σ o → ∞ (23.6b)
p(D | H) σo σε
  ∑ xt2
  t =1
where p(Θ | H) is the prior probability density function (pdf), which describes
the uncertainty about Θ prior to acquiring the data D, and p(Θ | D , H) is the σ ε2
posterior pdf, which describes the uncertainty about Θ after processing the σ θ2 = n
(23.6c)
σ2
information in D. ∑ xt2 + σ ε2
The term p(D | Θ, H) is often referred to as the “likelihood function,” and t =1 o
can be thought of as the “probability” of observing the data D for a given value
of Θ. More precisely, unless D is a discrete variable, p(D | Θ, H) is the probabil- To illustrate what can be learned from this posterior, consider a synthetic
ity density of D given Θ. data set comprising 20 observations generated using θ1 = 0.3 and σ ε = 25.
Note that, in parameter estimation contexts, the term p(D | Θ, H) is treated Figure 23.1a presents a scatter plot of the predictor and output data, along
as a function of Θ (even though it is defined as the probability density of D). with the regression line drawn using θˆ1, while Fig. 23.1b presents a plot of the
For this reason, the likelihood function is often written as L(Θ; D , H), espe- posterior pdf of θ1 assuming no useful prior knowledge (i.e., the prior stan-
cially in the frequentist literature. dard deviation σ o was set to a large value). This is referred to as “vague” or
“uninformative” prior information.
The denominator p(D | H) = ∫ p(D | Θ, H) p(Θ | H)d Θ , often referred to as
Θ
the “marginal likelihood” of the data D, ensures the posterior density inte-
grates to 1. In many cases, direct computation of p(D | H) is prohibitive. In 50
such cases, noting that p(D | H) does not depend on Θ, the posterior pdf can
be expressed up to a constant of proportionality 250 40
Posterior pdf
p(Θ | D , H) ∝ p(D | Θ, H) p(Θ | H) (23.3) 30
Y 150
For conciseness, it is usual practice to omit the model hypothesis H from the 20
list of conditioning variables. Here the conditioning on H is made explicit to
stress that the prior and posterior distributions only make sense in the context 50 10
of hypothesis H, which itself is likely to be subject to uncertainty and is scru- 0 0
tinized in the third step (Sec. 23.4).
0 200 400 600 800 0.26 0.28 0.30 0.32
Posterior distributions involving three or more parameters can be difficult
to examine graphically. Analytical examination and summary of the posterior (a) X (b) q1
also becomes difficult or impossible except for special cases (e.g., the linear
Figure 23.1 One-parameter regression function example: (a) Scatter plot of predictor
Gaussian regression models presented in Secs. 23.2.2.1 and 23.2.2.2). and output data; (b) plot of posterior pdf of q1 assuming no prior information.
Fortunately, the recent development of approximation techniques, such as
Markov chain Monte Carlo methods, has provided the tools to work with
complicated posteriors and has played a major role in the rapid growth in
Bayesian applications (Kuczera and Parent, 1998; Bates and Campbell, 2001; The posterior mean θˆ1 defines the slope of the solid line in Fig. 23.1a. It can
Vrugt et al., 2008; Cressie et al., 2009; Renard et al., 2011; and many others). be shown that this is the line of “best fit” in the sense the sum of the squared
errors is minimized (when the prior is sufficiently vague). Perusal of Fig. 23.1a
23.2.2.1 Example: A One-Parameter Linear suggests considerable uncertainty in drawing the line of best fit. This is con-
Regression Function firmed by the posterior of θ1 in Fig. 23.1b, which shows that the slope of the
The mechanics of Bayesian updating are illustrated using two simple exam- regression line is most likely to lie somewhere between 0.27 and 0.32.
ples which are amenable to analytical treatment yet offer rich insights. In the absence of useful prior information (i.e., σ o is large), the posterior
n
The first example is the one-parameter linear regression function
variance of θ1 reduces to σ ε2 ∑ xt2. In this case, when the predictor values lie
f ( xt , θ ) = θ1xt (23.4) t =1
close to 0, the denominator term is small, leading to large uncertainty in the
For example, this relationship could be used as a possible (highly simplified!) slope parameter. A corollary of this behavior is that uncertainty in θ1 can be
description of the dependence of annual catchment runoff on annual catch- reduced by seeking observations of outputs corresponding to large predictor
ment precipitation. values—a geometrically intuitive finding relevant to experimental design.
For simplicity, assume the error variance σ ε2 is known. Suppose our prior Another approach to reduce the uncertainty in θ1 is to use an informative
uncertainty about θ1 is described by a Gaussian distribution N (θ o , σ o2 ) with prior with prior variance σ o2 << σ ε2 . Assuming such a prior is available, the
mean θ o and variance σ o2. In our example, where θ1 is the runoff coefficient, posterior mean given by Eq. (23.6b) can be expressed as a weighted average of
n n
the hydrologist may already have some prior knowledge of its possible values.
For instance, she might choose θ o and σ o such that most of the prior mass for the prior mean and the least-squares estimate of the slope ∑ xt yt ∑ xt2
t =1 t =1
θ1 is between 0 and 1; knowing that the catchment is located in a semiarid which is solely based on the data,
region, she might favor values of θ1 closer to 0. More generally, hydrologists
n
frequently use prior information derived from some form of regionalization
or information transfer. For example, see Viglione et al. (2013) for applica- ∑ x t yt σ ε2
tions in flood frequency analysis. θˆ1 = wθ o + (1 − w ) t =1n with w = n
(23.7)
If the data D consists of n observations of the predictor variables
x T = ( x1 , , xn ) and outputs y T = ( y1 , , yn ) then the assumption of inde-
∑ xt2 σ ε2 + σ o2 ∑ xt2
t =1 t =1
pendent Gaussian errors leads to the likelihood function (or sampling distri-
bution of the observed outputs) It can be seen that the weighting of the prior versus the data is determined by
a ratio involving the prior variance, the variance of regression errors and the
n
1  ( y − θ x )2  sum of squared predictor values. In particular, a tighter prior on θ1 (smaller
p( y | θ1 , σ ε2 , x ) = ∏ exp − t 12 t  (23.5) σ o) leads to a weight closer to 1, hence favoring the information contained in
t =1 2πσ ε2  2σ ε  the prior over the information brought by the observed data.
23_Singh_ch23_p23.1-23.10.indd 2 8/22/16 1:33 PM

The Bayesian Inference Framework 23-3
More generally, Bayesian inference merges prior information and the infor- useful information about the slope. However, there are an infinite number of
mation in the data, with the relative weighting being dependent on the precision combinations of θ1 and θ 2 that add up to any particular value of the slope of
of each source of information. Ultimately, if one source of information is much the solid line.
more precise than others, it dominates the inference making the contribution of In the absence of useful prior knowledge, nonidentifiability leads to ill-
other sources inconsequential. posed problems. Hydrological modeling is replete with models that are poorly
identifiable. Examples include rainfall-runoff models with multiple runoff
23.2.2.2 Example 2: A Nonidentifiable pathways that can mimic each other or are not activated by the rainfall, and
Two-Parameter Linear Regression Function distributed models where different parameters are assigned to each subcatch-
In the Bayesian framework, the posterior summarizes our knowledge about the ment or cell. The parameters of such models calibrated to observed runoff at
parameters given the available information (including the model hypothesis). a single location are likely to be nonidentifiable. Often this problem can be
Comparing the prior and posterior allows one to gauge the information con- remedied by fixing a subset of the parameters, or by applying similarity rules.
tent of the data D, the key idea being that “informative” data should signifi- From a Bayesian perspective this represents the use of (possibly subjective)
cantly reduce prior uncertainty. This is an important consideration in informative prior knowledge.
hydrological modeling, where limited data availability can impose significant
constraints on what can be learned. In this section, the linear regression 23.2.2.3 Predictive Distributions
example from Sec. 23.2.2.1 is modified to illustrate this issue and highlight the Once the posterior has been inferred, it is possible to make inferences on any
importance of understanding the posterior. quantity of interest that depends on Θ using the posterior predictive distribu-
Suppose the regression function is f ( xt , θ ) = θ1xt + θ 2 xt . Though this tion. For example, the posterior predictive distribution of the observed output
appears to be similar to the regression function in the first example, the y f given a new predictor value x f is
inference is fundamentally different—the information in the data D is now
used to make an inference about two parameters, θ1 and θ 2. Most readers
will intuitively question such an inference, noting in particular the mutu-
p( y f | x f , D , H) = ∫ Θ p( y f | x f , Θ, H) p(Θ | D , H)d Θ (23.9)
ally compensatory behavior of θ1 and θ 2. Indeed, this highly simplified For the standard models discussed earlier, Eq. (23.9) can be written more
example is instructive for understanding identifiability issues in more explicitly as
complex models.
It can be shown that the posterior variance of θ1 is
p( y f | x f , D , H) = ∫θ ,σ ε
p( y f | f ( x f , θ ), σ ε , H) p(θ , σ ε | D , H)dθ dσ ε (23.10)
 n 
 ∑ xt2   2
where p( y f | f ( x f , θ ), σ ε , H) is the pdf of a Gaussian distribution with mean
2  2 t =1
σ o1 1 + σ o 2 2 
  σ 2 + σ ε as σ 2 → ∞ given by f ( x f , θ ) and standard deviation given by σ ε , and p(θ , σ ε | D , H) is
o2 o1
 σε   n
the posterior pdf of parameters (θ , σ ε ).
Var(θ1 | D ) = n
→ ∑ xt2 (23.8) In essence the predictive distribution uses the posterior of Θ to weight the

∑t  2
x 2 t =1
distribution of y f for each possible value of Θ. In the case of the two linear
 σ o1 as σ o22 → ∞
(
1 + σ o21 + σ o22 ) t =1
σ ε2
regression models, analytical expressions for the predictive distribution can
be derived under the assumptions of Gaussian prior knowledge and constant-
variance Gaussian errors (Press, 1989). In other cases, numerical approxi-
mations can be employed (e.g., see Gelfand and Smith, 1990).
where σ o21 and σ o22 are the prior variances of θ1 and θ 2, respectively. A similar
For the linear regression function θ1xt , under the assumptions of Sec. 23.2.2.1,
expression can be derived for the posterior variance of θ 2.
the posterior predictive distribution of y f is also Gaussian,
Figure 23.2 plots the posterior standard deviation of θ1 as a function of the
prior standard deviations σ o1 and σ o 2 for the same data as used in Fig. 23.1.
It can be seen that as the prior standard deviations increase, the posterior
(
y f ~ N µyf , σ yf2 ) (23.11a)
standard deviation of θ1 also increases. Indeed, in the absence of useful prior
information on θ1 and θ 2 (i.e., for σ o21 → ∞, σ o22 → ∞ ), the data provide no µ yf = θˆ1x f (23.11b)
useful information on the inference of θ1. This outcome, referred to as non-
identifiability, has a strong geometric interpretation. Referring to Fig. 23.1a,  x 2f 
the slope of the solid line is θ1 + θ 2. As seen in Fig. 23.1b, the data provide σ yf2 = σ ε2 1 + n 2
 (23.11c)
 2 σε 
 ∑ t σ2 
x +
 t =1 o 
1.0
If using uninformative priors (i.e., σ o2 → ∞ ), the variance of the posterior
0.7
predictive distribution reduces to:
0.8 0.6  x2 
lim σ 2
= σ 2
1 + n f  (23.12)
Prior standard deviation for θ2
yf ε
σ o2 →∞  2
 ∑ xt 
0.5  t =1 
0.6
0.4 Consider the predictive variance σ yf2 in Eq. (23.12). The first term, σ ε2 , arises
from the inherent and irreducible uncertainty expressed by the regression
error model. The second term, given by the ratio in the brackets, represents
0.4 0.3 the contribution of uncertainty in the slope parameter θ1 to predictive uncer-
tainty in y f . Unlike the first term, which is fixed and does not depend on x f ,
the second term increases as x f increases in magnitude and decreases as the
0.2 number of observations increases.
0.2 Using the data from Fig. 23.1a, Fig. 23.3 shows the 90% predictive proba-
0.1 bility intervals due to uncertainty in the slope parameter θ1 and compares
them to the predictive intervals due to total uncertainty. It can be seen that,
for predictor values x f outside the range of observations, the uncertainty in
0.0 the slope parameter dominates the total uncertainty. In other words, as the
0.2 0.4 0.6 0.8 1.0 extrapolation beyond the range of observed predictor values increases, the
Prior standard deviation for θ1 effects of uncertainty in the slope are amplified to dominate the inherent
uncertainty in the predictor–output relationship. In contrast, for predictor
Figure 23.2 Posterior standard deviation of q1 for the two-parameter regression values lying within the range of observed data, the inherent uncertainty
function as a function of prior standard deviations so1 and so2. dominates.
23_Singh_ch23_p23.1-23.10.indd 3 8/22/16 1:33 PM

500 namely ∫ z(θ ) p(θ | D )dθ . If the loss function is asymmetric implying the
500
θ
consequences of under- and overdesign are different, the best estimate may
not be the posterior mean of z (θ )—see for example Botto et al. (2014). Note
300
300
Y
Y
that this analysis holds for the case where z is a deterministic function of q
and would need to be generalized in the case where z is a stochastic function
dependent on the entire vector Θ.
0 100
0 100
Decision-making could be viewed as a fourth step in the application of the
Bayesian framework. However, as the specification of the loss function is gen-
0 500 1000 1500 2000 0 500 1000 1500 2000 erally determined by broader societal considerations, it is not falsifiable or
X X verifiable using the empirical scientific method. In this respect, it is “external”
(a) (b) to the Bayesian framework and scientific analysis.
Figure 23.3 Predictive distribution for one-parameter linear regression function 23.2.5 Relationship with Other Probabilistic
fitted to the data in Fig. 23.1. Dashed lines correspond to 90% probability intervals: Inference Frameworks
(a) Uncertainty in slope parameter only; (b) total uncertainty (solid) compared with
It is worth noting that the Bayesian framework is not the only inference
parametric uncertainty (dashed). The gray lines are explained in Sec. 23.3.3.
approach based on probability theory. However, there are differences in the
interpretation of probability. In the frequentist paradigm, probability refers
For the two-parameter linear regression function θ1xt + θ 2 xt , again making strictly to the relative frequency of an event in a large number of trials,
the assumptions from Sec. 23.2.2.2, the posterior predictive distribution of y f is whereas the Bayesian paradigm extends the use of probability to encompass
“subjective degrees of confidence or belief ” in the odds of an event taking
( )
y f ~ N µyf , σ yf2 (23.13a) place. Hence in Bayesian inference, knowledge about inferred model param-
eters can be represented by a probability distribution (i.e., the parameters are
 n σ2  treated as random variables), while in frequentist inference, the parameters
 ∑ xt yt + 2 ε 2 (θ o1 + θ o 2 )  are treated as fixed (albeit unknown) quantities. Despite these fundamental
 σ o1 + σ o2 
µ yf =  t =1 n  x f (23.13b) differences, both approaches often produce similar results particularly in
σ ε2
 ∑ xt + σ 2 + σ 2
2
 data-rich applications. The interested reader is referred to Berger and Berry
(1988), Efron (1986), and others for a broader discussion.
 t =1 o1 o2 
 x 2f  23.3 COMPUTATIONAL METHODS

σ yf2 = σ ε2 1 + n 2
 (23.13c)
 σ  The posterior distribution describes a surface in parameter space. Figure 23.4
 ∑ xt + σ 2 + σ 2 
2 ε
illustrates such a surface in the case of two parameters where the contours rep-
 t =1 o1 o2 
resent lines of constant posterior density. Although in many practical cases the
computation of the posterior using Eq. (23.3) is numerically straightforward,
The second term in the predictive variance is the contribution of uncertainty communicating the information contained in the posterior can be challenging
in the slope θ1 + θ 2 and grows with x f . In the absence of any useful prior infor- when more than two parameters are involved. In some applications, the poste-
mation on the parameters (i.e., for σ o21 → ∞, σ o22 → ∞), the posterior predic- rior can be shown to correspond to (or approximate) well-understood distribu-
tive variance reduces to that in Eq. (23.12), tions—e.g., see the classic text by Box and Tiao (1973). However, in many
applications, the posterior is not amenable to such analysis, particularly when
 x 2f  many parameters are involved. Modern Bayesian methods use optimization
lim σ yf2 = σ ε2 1 + n  (23.14) approaches (Sec. 23.3.1) to search for a good point estimate of the parameters
σ o21 →∞ ,  2
σ o22 →∞  ∑ x t  (typically the parameter set that maximizes the posterior) and sampling tech-
 t =1  niques such as Markov chain Monte Carlo methods (Sec. 23.3.2) to approximate
the posterior by a number of parameter samples that summarize its “shape” and
despite the fact that the posterior variance on the individual parameters is help modelers draw conclusions about the quantities of interest.
infinite [see Eq. (23.8)].
These two examples highlight some of the insights that can be obtained
from the posterior. The first highlights that extrapolation of a model beyond
observed predictor data may produce an explosion in predictive uncertainty Θ2
because the extrapolation amplifies the propagation of parameter uncer-
tainty into output uncertainty. The second shows that high parameter
uncertainty need not imply high predictive uncertainty provided the predic-
tions are not extrapolated beyond the observed data.
23.2.3 Step 3: Scrutiny of the Model Hypothesis
The inference described in the second step only makes sense in the context of
the hypothesis H. The question then arises, is the hypothesis H credible? Box
and Draper (1987, p. 424) have observed that although all models are
“wrong,” the purpose of a model is to provide a useful approximation. The
objective of the third step in the Bayesian framework is to ask whether the Θ1
model hypothesis is credible. Can the hypothesis be reconciled with available Unnormalized posterior pdf
evidence to a degree deemed fit for purpose? This is the subject of Sec. 23.4. 0 large
23.2.4 From Predictions to Making Decisions

Figure 23.4 Illustration of a bivariate posterior pdf (with denser gray colors repre-
In the Bayesian approach, decision-making is guided by assessing the conse- senting higher probability density regions).
quences of under- or over-estimation of a variable of interest z. One formal
approach to do this is to specify a loss function L[zˆ, z ] that describes the loss or 23.3.1 Using Point-Estimates from the Posterior
consequence when an estimate ẑ does not correctly represent the “true” value z Distribution: Optimization
(DeGroot, 1970). In the special case where z is a deterministic function of q and
The simplest approach for making use of a posterior distribution is to find its
keeping in mind that the “true” value of q is uncertain, one way to select the
“mode” or “optimum,” i.e., the parameters for which the posterior pdf is
estimate ẑ is to minimize the expected loss, defined as ∫ L[zˆ , z (θ )] p(θ | D )dθ . maximized. This is closely related to the traditional idea of calibration, which
θ
The estimate depends on the choice of loss function and the information seeks to optimize an a priori selected objective function. In Bayesian inference,
available to infer q. If the loss function is symmetric and quadratic, that is, the objective function is the posterior distribution itself—as discussed previ-
L[zˆ, z (θ )]∝ [zˆ − z (θ )]2 , the optimal estimate is the posterior mean of z (θ ), ously, it is derived by building probability models of the system of interest.
23_Singh_ch23_p23.1-23.10.indd 4 8/22/16 1:33 PM

Computational Methods 23-5
Many optimization techniques have been described in the literature. Θ2

Traditional methods include gradient-based optimizers, in particular, Candidate sample (Θ(*), Θ(*))
1 2
Newton-type methods (Nocedal and Wright, 1999). These methods are very
efficient for smooth objective functions, and mature software exists in most
programming languages. The most popular methods of this class are Quasi– Current sample (Θ(i–1) , Θ(i–1) )
1 2
Newton methods for general objective functions and Gauss–Newton methods
for least-squares-type objective functions (Nocedal and Wright, 1999). A
limitation of Newton-type methods is their poor performance when the Jump
objective function is not smooth, because in this case the gradient values distribution
needed for the algorithm are either difficult to estimate or may not even exist
(e.g., at the points of discontinuity). The limitations of traditional Newton-
type methods can be alleviated using optimization methods that work solely Jump! Θ1
with function values, so that gradient values are not needed, and indeed even
their existence is not required for the algorithm to function. The Nelder– Unnormalized posterior pdf
0 large
Mead simplex method (Nelder and Mead, 1965) is one of the best-known
nonsmooth optimization methods. Nonsmooth methods are particularly
attractive when combined with evolutionary strategies, which randomize the Figure 23.5 Illustration of an MCMC step for a bivariate posterior pdf. A candidate
search in order to allow the optimizer to escape local optima and improve the sample is generated from an easy-to-sample jump distribution; the candidate is then
chances of finding the global optimum. In hydrology, the shuffled complex accepted or rejected depending on the ratio of probability densities at the proposed
versus old parameter values.
evolution method has become very popular due to its robust performance for
nonsmooth and irregular-shaped objective functions (Duan et al., 1992). Box 23.1 The Metropolis algorithm using a Gaussian jump distribution.
Other well-known optimizers popular in hydrological applications include
the dynamic-dimensioned search (Tolson and Shoemaker, 2007), the 0. Choose a starting point Θ(0) and a jump covariance matrix Σ J .
AMALGAM method (Vrugt et al., 2007), and others. 1. Repeat for i =1: N sam, where Nsam is the number of samples to be generated
It is worth noting that, in many cases, nonsmooth objective functions are
a. Sample a candidate vector Θ(*) from a Gaussian distribution centered
not an inherent property of the model, but are spurious artifacts caused by
on the current sample Θ(i−1) and with covariance matrix Σ J :
poor numerical implementation of these models (Kavetski and Clark, 2010).
Θ(∗) ← N (Θ(i−1) , Σ J )
In such cases, it may be preferable to improve the model implementation to
b. Compute the ratio between the new and the previous posterior pdf
take advantage of powerful gradient-based optimization methods while also
improving model behavior (Hill et al., 2015). p(Θ(∗) | D )
τ=
An obvious limitation of optimization is that it provides little indication of p(Θ(i−1) | D )
the uncertainty in the optimized parameters. This can often be remedied by c. Accept the candidate vector with probability min(τ , 1), that is, set
carrying out sensitivity analysis around the optimal parameter values. Θ(i ) = Θ(∗); otherwise reject the candidate vector, that is, set Θ(i ) = Θ(i−1)
Assuming the posterior distribution of the parameters is multivariate
Gaussian, it is also possible to formally derive approximations to the posterior
covariance matrix, by differentiating twice the log-posterior distribution at the generating a vector Θ(∗) from a multivariate Gaussian jump distribution. The
−1 new point is then accepted with a probability equal to the ratio of the “new”
 ∂2 log p(Θ | D ) 
mode, Cov(Θ | D ) = −  , where the derivatives can be either posterior density evaluated at the candidate point Θ(∗) and the “old” posterior
 ∂Θ2  density evaluated at point Θ(i−1). This means that “uphill” jumps are always
evaluated analytically or approximated using finite differences (Gelman et al., accepted, while “downhill” jumps are accepted at least occasionally (the lower
2004). This approximation is broadly analogous to the computation of the the posterior density of the proposed point, the lower the chance of accepting
Fischer information in frequentist methods, except that the Bayesian poste- a jump there). Of great practical significance is the fact that this algorithm
rior is used in lieu of the likelihood. In many cases, this approximation can only requires the posterior in the form of Eq. (23.3)—the constant of propor-
provide useful estimates of uncertainty; however, its accuracy can be poor if tionality is not required as it cancels out in the ratio of the posteriors.
the posterior departs substantially from a Gaussian shape. These limitations The Metropolis algorithm can be generalized to allow for an arbitrary jump
motivate the more expensive, but also more powerful Markov Chain Monte distribution—this is known as the Metropolis–Hastings algorithm. The only
Carlo methods, which are described next. difference is that the acceptance ratio is not simply equal to the posterior ratio,
but now encompasses a term corresponding to a ratio of the jump densities.
23.3.2 Sampling from the Posterior Distribution: 23.3.2.3 Examples: Revisiting the Two Linear
MCMC Samplers
Regression Models
One way to communicate the uncertainty about the parameters is to randomly
The Metropolis sampler is illustrated for the one-parameter linear regression
sample many parameter values from the posterior distribution, and then use
problem yt ← N (θ1xt , σ ε2 ) presented in Sec. 23.2.2.1.
these values to explore the properties of the posterior distribution (variability,
Consider first the case where σ ε is known. The only unknown quantity is the
shape, dependence). The parameter samples can then be used to generate a set
regression parameter θ1, for which a vague prior N (0, 100002 ) is assigned. The
of predictions that will naturally quantify predictive uncertainty.
Metropolis sampler is then applied to the unnormalized posterior (using the R
23.3.2.1 General Principles of package MCMC). The variance of the Gaussian jump distribution is set to 0.12.
MCMC Sampling Figure 23.6a shows the direct output of the MCMC sampler: 50,000 values of
the unknown parameter θ1 have been sampled from the posterior distribution,
Markov Chain Monte Carlo (MCMC) samplers provide a general method for
(approximately) drawing random samples from any probability distribution.
10 20 30 40 50
0.26 0.28 0.30 0.32
In Bayesian applications, MCMC methods are used to sample from posterior

Normalized count
distributions. The main idea in MCMC sampling is to explore the distribution

of interest (referred to as the “target” distribution) by a series of successive
jumps. This is illustrated in Fig. 23.5. Starting from the current parameter set
θ1
Θ(i−1), a jump is proposed to a new parameter set Θ(*). The jump is obtained by
sampling from a “jump” (or “proposal”) distribution, typically selected as a
distribution that is easy to sample from. The core idea in MCMC sampling is
to probabilistically accept or reject the proposed parameter set Θ(*) in such a
0
way that the resulting process generates realizations that converge to the target
0 10000 30000 50000 0.26 0.28 0.30 0.32
distribution (which, in Bayesian applications, is the posterior distribution).
Iteration θ1
23.3.2.2 Metropolis and Metropolis–Hastings (a) (b)
Samplers
Figure 23.6 MCMC sampling for the one-parameter linear regression function with
The algorithm in Box 23.1 describes the Metropolis algorithm. Given a cur- known error standard deviation se : (a) Trace of 50,000 samples of q1; (b) Normalized
rent point Θ(i−1) in the parameter space, the jump to another point is made by histogram of the 50,000 samples of q1 along with the exact pdf derived in Eq. (23.6).
23_Singh_ch23_p23.1-23.10.indd 5 8/22/16 1:33 PM

60 50 0.10
Normalized count
Normalized count
50 40 0.08
30 0.06
40
σε
20 0.04
30
10 0.02
20
0 0.00
0.26 0.28 0.30 0.32 0.26 0.28 0.30 0.32 0.34 15 20 25 30 35 40 45
θ1 θ1 σε
(a) (b) (c)
Figure 23.7 MCMC sampling for the one-parameter linear regression model with unknown error standard deviation se : (a) Scatterplot of 50,000 samples
of (q1, se); (b) normalized histogram of the 50,000 samples of q1; (c) normalized histogram of the 50,000 samples of se .
as illustrated in Fig. 23.6b, where the normalized histogram closely corresponds of the sampler. For general guidelines, the reader is referred to Gelman et al.
to the exact posterior shown in Fig. 23.1b. These 50,000 values can be used to (2004) and Robert and Casella (2004); for examples of specific samplers, see
quantify any property of the posterior distribution (e.g., its mean, variance, or Haario et al. (2001), Thyer et al. (2002), Renard et al. (2006), Vrugt et al. (2009),
higher-order moments) and to represent uncertainty using probability intervals. Kuczera et al. (2010), Hoffman and Gelman (2014), and others.
Now consider the case where σ ε is unknown. There are two unknown quan-
tities Θ = (θ1 , σ ε ), so the posterior distribution is two-dimensional. The same 23.3.3 Using MCMC Samples for Predictive
Uncertainty Quantification
Gaussian prior for θ1 is used, and a wide uniform prior U[0,10000] for σ ε. For
the Gaussian jump covariance a diagonal matrix is used with diagonal terms An appealing property of MCMC sampling is that it considerably simplifies
equal to 0.12. Figure 23.7 shows the result of MCMC sampling. Panel (a) is a uncertainty analysis for predictions. Indeed, since MCMC samples by defini-
scatterplot of 50,000 values of (θ1 , σ ε ), which are realizations from the bivari- tion represent the uncertainty in all sampled quantities, predictive uncer-
ate posterior distribution of (θ1 , σ ε ). It is seen that θ1 and σ ε are uncorrelated. tainty can be evaluated using simple propagation.
The two other panels show normalized histograms for each parameter, repre- Consider for instance MCMC sampling from the one-parameter linear
senting marginal posterior distributions. For instance, one can observe that the regression model, yielding a set of values (θ1(i ) , σ ε(i ) )i=1:N sim. Each value θ1(i )
marginal posterior distribution of σ ε is asymmetric (skewed). defines a possible regression line, and all these lines represent the parametric
Our attention now turns to the nonidentifiable two-parameter linear uncertainty, that is, the uncertainty due to the imperfect knowledge of the
regression model yt ← N (θ1xt + θ 2 xt , σ ε2 ). The first panel of Fig. 23.8 shows slope of the regression (gray lines in Fig. 23.3a). This information can be
that the 50,000 values of θ1 and θ 2 generated by the MCMC sampler do not further summarized by computing, for a given predictor value x f , a 90%
seem to converge. The second panel demonstrates the very strong negative probability interval from the samples (θ1(i )x f )i=1:N sim. Repeating this procedure
correlation between the parameters. This correlation is due to the fact that the for an ensemble of values of the predictor yields the 90% probability interval
data only provide useful information to identify the slope θ1 + θ 2, and any shown as dashed lines in Fig. 23.3.
combination of θ1 and θ 2 that adds up to the same value yields the same pos- It is apparent from Fig. 23.3a that the parametric uncertainty interval is
terior pdf value. It is instructive to examine the behavior of the 50,000 values much too narrow when compared to the spread of observed data, especially
of θ1 + θ 2 computed from the MCMC samples. The MCMC samples of θ1 + θ 2 for predictor values close to zero. This is because parametric uncertainty
are perfectly well-behaved (last two panels of Fig. 23.8), which reflects the fact represents only one part of the total predictive uncertainty. In addition,
that the slope of the regression can indeed be identified from the data even it is necessary to account for the inherent uncertainty in the relationship
though its individual components θ1 and θ 2 cannot. between y and x—which is exactly what the (probabilistic) regression model
(i.e., the regression function and the regression error) seeks to represent.
23.3.2.4 Practical Aspects: Monitoring Hence, to derive the total predictive uncertainty, a random output has to be
Convergence and Efficient MCMC Sampling
sampled from the full model defining the hypothesis H for each MCMC
The appeal of MCMC samplers lies in their generic ability to sample from sample (θ1(i ) , σ ε(i ) ). For a given predictor x f , this yields a set of values
virtually any distribution (but see later for important caveats). They have been
successfully applied to many complex distributions, including high-dimen- { y (i ) ← N[θ1(i )x f , (σ ε(i ) )2 ]}i=1:N sim
. Figure 23.3b shows the corresponding values
sional ones. However, like any numerical method, MCMC samplers require as gray lines, along with a 90% predictive interval (dashed lines).
care to avoid misleading conclusions, and they may require considerable tun-
ing and/or parameter transformations to work efficiently (or feasibly). In
23.4 DIAGNOSTICS TO SCRUTINIZE
particular, MCMC convergence proofs are asymptotic with respect to the MODEL ASSUMPTIONS
number of samples and as such do not guarantee that this limit can be
approached in a feasible (or even finite) amount of time. Therefore, two issues An integral part of the Bayesian framework (and any scientific endeavor) is
of practical importance are monitoring convergence to avoid poor approxi- scrutiny of the assumptions of the hypothesis H using a range of diagnostics.
mations and selecting “good” jump distributions to maximize the efficiency The model predictions and parameter estimates are conditional on the
0.24 0.26 0.28 0.30 0.32 0.34
50
4
40
Normalized count
2
30
θ1 + θ2
2
θ2
0
θ
20
1
−2
θ1
10
0
θ2
−4
–1
0 10000 30000 50000 –4 –3 –2 –1 0 1 0 10000 30000 50000 0.26 0.28 0.30 0.32 0.34
Iteration θ1 Iteration θ1 + θ2
(a) (b) (c) (d)

Figure 23.8 MCMC sampling for the two-parameter linear regression function with slope q1 + q2: (a) traces of 50,000 samples of q1 and q2; (b) scatterplot of 50,000 samples of (q1, q2);
(c) traces of 50,000 samples of q1 + q2; (d) normalized histogram of the 50,000 samples of q1 + q2.
23_Singh_ch23_p23.1-23.10.indd 6 8/22/16 1:34 PM

Applications in hydrology 23-7
assumptions of the hypothesis. Consequently, unsupported assumptions may Fig. 23.9 shows that the SLS model produces constant precision for all flows,
lead to misleading predictive distributions. In this section the key elements of while the WLS model produces higher precision for low flows than for high
a range of diagnostics are outlined with a focus on probabilistic predictions flows. Ideally, predictive distributions should have high reliability and high
and residual errors. precision, though this is often difficult to achieve in practice.
23.4.1 Diagnostics to Evaluate Probabilistic Predictions 23.4.2 Diagnostics to Evaluate Residual Error
Model Assumptions
In the context of Bayesian inference, model predictions take the form of
probabilistic predictive distributions (see Fig. 23.9a, b for examples; and If the predictive QQ plot identifies unreliable probabilistic predictions, more
Chap. 2, Sec. 2.7 of this book). A predictive distribution is considered to be probing diagnostics are needed to identify which assumptions need to be
statistically reliable when it adequately captures the distributional properties of refined. A common assumption in regression models is that the residual
the observed data (i.e., when the observations can be viewed as samples from errors, defined as the difference between observed and predicted model
the predictive distribution). From a methodological point of view, this requires outputs, are independent and identically distributed and follow a Gaussian
a diagnostic approach that compares the observed output with the entire pre- distribution with mean zero and constant standard deviation. A range of
dictive distribution of the model output. This is a more stringent test than statistical diagnostics is available to evaluate these assumptions including
standard goodness-of-fit assessments (e.g., using the Nash–Sutcliffe statistic or QQ plots to evaluate the Gaussian error assumption, plots of residuals
the root mean square error), which are limited to comparing a single time against the model predictors and/or time to test the assumption of zero mean
series of simulated outputs with a single time series of observed outputs. A and constant standard deviation, and autocorrelation functions of the
generic approach for checking reliability is the predictive quantile-quantile residuals to evaluate the assumptions of independence. The reader is referred
(QQ) plot, used for the verification of probabilistic forecasts in hydrology and to Kuczera (1983) and Thyer et al. (2009) for a description of the available
meteorology (Gneiting et al., 2007; Laio and Tamea, 2007; Thyer et al., 2009). diagnostics.
An appealing feature of the predictive QQ plot is that departures from sta-
tistical reliability can be directly interpreted, as seen in Fig. 23.9c. Figure 23.9d 23.5 APPLICATIONS IN HYDROLOGY
shows an example, which compares two predictive distributions for the GR4J
model fitted to the French Broad River catchment (Evin et al., 2014). The dis- Bayesian inference has seen considerable application in hydrological model-
tributions are obtained using two different residual error models: (a) Gaussian ing. This section describes selected applications in flood frequency analysis
errors with constant variance, referred to as “standard least squares” (SLS) and and rainfall-runoff modeling to convey important insights offered by the
(b) Gaussian errors with variance increasing with model predictions, referred Bayesian approach.
to as “weighted least squares” (WLS) (note that the naming of these methods
reflects their origins in “least squares” optimization). In this example—as well 23.5.1 Flood Frequency Analysis
as in our general experience in rainfall-runoff model calibration—the WLS Bayesian applications in flood frequency analysis have focused on the pooling
error model provides a more reliable predictive distribution because it of information from different sources to reduce the uncertainty in extreme
accounts for the heteroscedasticity of residual errors. flood quantiles (e.g., see Stedinger and Cohn, 1986; Kuczera, 1999; Reis et al.,
Another important attribute of the predictive distribution is its precision, 2005; Micevski and Kuczera, 2009; Viglione et al., 2013, and others). The dif-
which describes the “width” of the predictive distribution. The example in ferent sources of information include: time series of gauged annual maximum
25 Observed 25 Observed
Predicted Predicted
20 50% probability 20 50% probability
Runoff (mm/day)
Runoff (mm/day)
90% probability 90% probability
15 15
10 10
5 5
0 0
2400 2410 2420 2430 2440 2450 2400 2410 2420 2430 2440 2450
(a) Predictive distribution using SLS model (b) Predictive distribution using WLS model
1 1.0
Under-prediction SLS
WLS
Quantile of observed p-value
0.8 Over-estimated
Quantile of observed p-value
0.8
predictive
uncertainty
0.6 0.6
0.4 0.4
0.2 Over-prediction
0.2
Under-estimated predictive
uncertainty
0 0.0
0 0.2 0.4 0.6 0.8 1
0.2 0.4 0.6 0.8
0.0 1.0
Theoretical quantile of U[0,1]
Theoretical quantile of U[0,1]
(c) Interpretation of predictive QQ plot (d) Predictive QQ plot for SLS and WLS
Figure 23.9 Predictive distribution for GR4J model calibration to French Broad River (see Evin et al., 2014 for details): (a) using SLS residual error model; (b) using WLS resid-
ual error model; (c) interpretation of predictive QQ plot; and (d) predictive QQ plot for both SLS and WLS.
23_Singh_ch23_p23.1-23.10.indd 7 8/22/16 1:34 PM

Box–Cox transformation). Autocorrelation can be represented by expressing

errors at time t as a function of errors at preceding time steps, in the simplest
2000 case, using autoregressive models.
A range of residual error models have been proposed to describe residual
Discharge (x 1000 ML/day)
errors that are heteroscedastic, exhibit significant autocorrelation, and possibly
500 time and state-dependent bias. These include mixture modeling approaches
(Schaefli et al., 2007), the skewed exponential power distribution approach
200 (Schoups and Vrugt, 2010), autoregressive bias error models (Del Giudice et al.,
2013), and others. A key conclusion from these studies is that the choice of
residual error model impacts significantly on model parameters and predic-
50 tions. Residual model development remains an ongoing area of research,
including in operational contexts [e.g., see Evin et al. (2013, 2014), Li et al.
20 (2015), and others].
10
23.5.3 Decompositional Treatment of Uncertainty
5 The aggregated approaches described in Sec. 23.5.2 are often sufficient in
operational applications where the main objective is to predict streamflow.
2 However, a logical question of research and operational significance is to
Gf G10 U U10 Uf Gf G10 U U10 Uf attempt to disaggregate, or decompose, the total predictive uncertainty into its
Figure 23.10 Posterior distributions of 10- and 100-year flood quantiles showing constituent sources. For example, how much predictive uncertainty is due to
median and 90% probability limits obtained using different information: Gf = 54-year the effect of rainfall errors, how much predictive uncertainty is due to the
gauged record, G10 = 10-year gauged record; U = regional distribution; U10 = G10 + U; effects of model error? These questions can be investigated using “decomposi-
Uf = U + Gf (taken from Micevski and Kuczera, 2009). tional” Bayesian approaches.
The key difference between “aggregated” versus “decompositional”
approaches is that the latter require separate probability models for each
discharges, number of exceedances above a discharge threshold, rating curve source of uncertainty that the modeler wishes to disaggregate. For example,
error and regionalized flood estimates. By way of illustration, Fig. 23.10 shows early applications of the Bayesian Total Error Analysis (BATEA) framework
the posterior distribution for 10- and 100-year flood quantiles for different used separate error models to describe rainfall errors and residual errors
combinations of information. While the 10-year gauged record produces very (Kavetski et al., 2006). Subsequently, the BATEA framework was extended to
wide 90% limits, pooling this information with regional flood information include descriptions of model structural error (Kuczera et al., 2006), rating
produces 90% limits comparable to using the full 54-year gauged record. curve error (Thyer et al., 2009), etc. However, as more and more sources of
error are described separately, the problems of non-identifiability and ill-
23.5.2 Aggregated Treatment of Uncertainty in posedness become more and more pronounced. Similar to the two-parame-
Rainfall-Runoff Models ter linear regression example where parameters θ1 and θ 2 interact with one
Traditional applications of Bayesian analysis in rainfall-runoff models have another, parameters describing rainfall errors can interact strongly with
closely followed the regression template described in Sec. 23.2.1. In particular, parameters describing model structural errors. This was directly demonstrated
all sources of uncertainty—whether data errors, model structural errors, or by Renard et al. (2010).
catchment stochasticity—are lumped into a single residual error term. Avoiding ill-posedness in decompositional approaches requires additional
Broadly speaking, this approach can be described as an “aggregated” treat- information. This can take the form of prior knowledge in Bayes equation. For
ment of uncertainty. The challenge then is to provide a statistically reliable example, in a later application of BATEA, Renard et al. (2011) used the geosta-
description of this residual error term. Analysis of typical residual errors tistical technique of conditional simulation to estimate the uncertainty in catch-
shows that they exhibit two important characteristics: (i) heteroscedasticity: ment rainfall time series computed from a limited number of raingauges. A
larger streamflows are usually subject to larger errors, and (ii) autocorrelation: direct rating curve error analysis was carried out using stage-discharge data,
the errors are serially correlated in time, for example, large positive errors are again allowing an informed error model to be developed. The only error term
usually followed by a sequence of large positive errors [e.g., Sorooshian and for which prior information was not available was for the parameters describing
Dracup (1980) and others]. model structural errors in the rainfall-runoff model itself. However, as shown
These characteristics invalidate the assumptions of standard regression, where in Renard et al. (2011), the presence of informative priors on the rainfall and
the errors are assumed to be independent random variables with zero mean and runoff observation errors constrained the inference of structural errors, yield-
a constant variance. Therefore, the residual error model must be revised to ing a well-posed inference. Figure 23.11 illustrates the insights gained using this
accommodate heteroscedasticity and autocorrelation. Heteroscedasticity can be strategy showing that structural errors dominate the total uncertainty.
represented either directly (e.g., by parameterizing the standard deviation of Decompositional approaches are of clear interest to hydrologists as they
errors as a function of streamflow) or indirectly (e.g., using a logarithmic or offer the promise of detailed quantitative insight into the relative significance
Observed Observed
Total 90% predictive interval Total 90% predictive interval
Structural errors 90% predictive interval Rainfall errors 90% predictive interval
18 18
16 16
14 14
Runoff (mm)
Runoff (mm)
12 12
10 10
8 8
6 6
4 4
2 2
0 0
550 600 650 700 550 600 650 700
(a) Time (day) (b) Time (day)
Figure 23.11 Partial probability limits reflecting the estimated effects of rainfall and model structural errors on
streamflow predictions (taken from Renard et al., 2009).
23_Singh_ch23_p23.1-23.10.indd 8 8/22/16 1:34 PM

REFERENCES 23-9
of different sources of error. However, at least two factors preclude widespread Evin, G., M. Thyer, D. Kavetski, D. McInerney, and G. Kuczera, “Comparison
application of decompositional approaches: (i) the formulation of informative of joint versus postprocessor approaches for hydrological uncertainty estima-
priors on data errors requires significant effort, and (ii) the computational cost tion accounting for error autocorrelation and heteroscedasticity,” Water
of decompositional methods is typically much higher than for aggregated Resources Research, 50: 2350–2375, 2014.
methods. It is stressed that, like any mathematical modeling approach, a Gelfand, A. E. and A. F. M. Smith, “Sampling-based approaches to calculat-
decompositional method is only as good as its underlying assumptions. Hence, ing marginal densities,” Journal of the American Statistical Association, 85:
applications where a decompositional method is applied with poorly con- 398–409, 1990.
structed data error models are unlikely to produce meaningful uncertainty Gelman, A., J. B. Carlin, H. S. Stern, and D. B. Rubin, Bayesian data analysis,
decompositions. Uncertainty decomposition is an active research topic, full of Texts in Statistical Science, Chapman & Hall/CRC, Boca Raton, 2004, p. 696.
exciting promises and unresolved challenges (Kuczera et al., 2006; Vrugt et al., Gneiting, T., F. Balabdaoui, and A. E. Raftery, “Probabilistic forecasts,
2008; Reichert and Mieleitner, 2009; Renard et al., 2011; and others). calibration and sharpness,” Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 69: 243–268, 2007.
Haario, H., E. Saksman, and J. Tamminen, An adaptive Metropolis algo-
23.6 CONCLUSION
rithm. Bernoulli, 7 (2): 223–242, 2001.
The Bayesian framework offers a systematic approach to inference in cases Hill, M. C., D. Kavetski, M. Clark, M. Ye, M. Arabi, D. Lu, L. Foglia, et al.,
where probability models are well suited to describing the data, which is often “Practical use of computationally frugal model analysis methods,”
the case in hydrologic applications. The framework involves an iterative three- Groundwater, 2015, doi: 10.1111/gwat.12330.
step process involving hypothesis (or model) building, inference using Bayes Hoffman, M. D. and A. Gelman, “The No-U-Turn sampler: adaptively set-
equation and scrutiny of assumptions, most likely leading to a revision of the ting path lengths in Hamiltonian Monte Carlo,” Journal of Machine Learning
hypothesis. While the second step can be regarded as formulaic, the first and Research, 15: 1593–1623, 2014.
third steps rely on the skill of the hydrologist. The quality of the inference Kavetski, D., G. Kuczera, and S. W. Franks, “Bayesian analysis of input
ultimately relies on how competently all three steps have been applied. uncertainty in hydrological modeling: 1. Theory,” Water Resources Research,
Bayes equation can be thought of as representing a learning process through 42 (3): W03407, 2006.
which current knowledge is revised as new information becomes available. Kavetski, D. and M. P. Clark, “Ancient numerical daemons of conceptual
Information can come from multiple sources, which can be data-based or hydrological modeling: 2. Impact of time stepping schemes on model analysis
subjective. In Bayesian inference, lack of knowledge in quantities of interest and prediction,” Water Resources Research, 46: W10511, 2010.
such as model parameters is described using probability distributions, and the Kuczera, G., “Improved parameter inference in catchment models: 1.
updating of prior knowledge into posterior knowledge is accomplished using Evaluating parameter uncertainty,” Water Resources Research, 19 (5): 1151–1162,
the likelihood function. This procedure requires the hydrologist to develop a 1983.
hypothesis of how deterministic and random processes within the system Kuczera, G. and E. Parent, “Monte Carlo assessment of parameter uncer-
produce the observations. This can be a challenging task, demanding not only tainty in conceptual catchment models: the Metropolis algorithm,” Journal of
knowledge of the physics of the hydrologic system and its observation systems Hydrology, 211 (1–4): 69–85, 1998.
but also an understanding of random processes. However, in our view, the Kuczera, G., “Comprehensive at-site flood frequency analysis using Monte
payoff is a deeper understanding and representation of the mix of determinis- Carlo Bayesian inference,” Water Resources Research, 35 (5): 1551–1558, 1999.
tic and random processes that drive hydrologic systems. Kuczera, G., D. Kavetski, S. W. Franks, and M. T. Thyer, “Towards a
Bayesian total error analysis of conceptual rainfall-runoff models: character-
ising model error using storm-dependent parameters,” Journal of Hydrology,
ACKNOWLEDGMENTS
331 (1–2): 161–177, 2006.
We thank Hoshin Gupta, Ezio Todini, and Alberto Viglione for insightful Kuczera, G., D. Kavetski, B. Renard, and M. Thyer, “A limited-memory
comments and suggestions made as part of the review of this book chapter. acceleration strategy for MCMC sampling in hierarchical Bayesian calibration
of hydrological models,” Water Resources Research, 46:1–6, 2010.
Laio, F. and S. Tamea, “Verification tools for probabilistic forecasts of con-
REFERENCES
tinuous hydrological variables,” Hydrology and Earth System Sciences, 11 (4):
Bates, B. and E. Campbell, “A Markov chain Monte Carlo scheme for 1267–1277, 2007.
parameter estimation and inference in conceptual rainfall-runoff modeling,” Li, M., Q. J. Wang, J. C. Bennett, and D. E. Robertson, “A strategy to over-
Water Resources Research, 37 (4): 937–947, 2001. come adverse effects of autoregressive updating of streamflow forecasts,”
Bayes, T., “An essay towards solving a problem in the doctrine of chances,” Hydrology and Earth System Sciences, 19 (1): 1–15, 2015, doi: 10.5194/
Philosophical Transactions of the Royal Society (London), 53: 370–418, 1763 hess-19-1-2015.
(reprinted in Biometrika, 1958, 45, 293–315). Micevski, T. and G. Kuczera, “Combining site and regional flood informa-
Berger, J. O. and D. A. Berry, “Statistical analysis and the illusion of objec- tion using a Bayesian Monte Carlo approach,” Water Resources Research,
tivity,” American Scientist, 76 (2): 159–165, 1988. 45:1–11, 2009, doi:10.1029/2008WR007173.
Botto, A., D. Ganora, F. Laio, and P. Claps, “Uncertainty compliant design Nelder, J. A. and R. Mead, “A simplex method for function minimization,”
flood estimation,” Water Resources Research, 50 (5): 4242–4253, 2014. Computing Journal, 7: 308–313, 1965.
Box, G. E. P. and N. R. Draper, Empirical Model-Building and Response Nocedal, J. and S. J. Wright, Numerical Optimization, Springer
Surfaces, Wiley, New York, 1987. Science+Business Media, New York, 2006.
Box, G. E. P. and G. C. Tiao, Bayesian Inference in Statistical Analysis, Press, S. J., Bayesian Statistics: Principles, Models and Applications, John
Addison-Wesley, Reading, MA, 1973. Wiley & Sons, New York, 1989.
Cressie, N., C. A. Calder, J. S. Clark, J. M. ver Hoef, and C. K. Wikle, Reichert, P. and J. Mieleitner, “Analyzing input and structural uncertainty
“Accounting for uncertainty in ecological analysis: the strengths and limitations of nonlinear dynamic models with stochastic, time-dependent parameters,”
of hierarchical statistical modeling,” Ecological Applications, 19 (3): 553–570, Water Resources Research, 45:1–19, p. 2009.
2009. Reis Jr., D. S., J. R. Stedinger, and E. S. Martins “Bayesian generalized least
DeGroot, M. H., Optimal Statistical Decisions, McGraw-Hill, New York, squares regression with applications to log Pearson type 3 regional skew esti-
1970. mation,” Water Resources Research, 41 (10):1–14, 2005.
Del Giudice, D., M. Honti, A. Scheidegger, C. Albert, P. Reichert, and J. Renard, B., V. Garreta, and M. Lang, “An application of Bayesian analysis
Rieckermann, “Improving uncertainty estimation in urban hydrological and MCMC methods to the estimation of a regional trend in annual maxima,”
modeling by statistically describing bias,” Hydrology and Earth System Water Resources Research, 42 (12):1–17, 2006.
Sciences, 17 (10): 4209–4225, 2013. Renard, B., E. Leblois, G. Kuczera, D. Kavetski, M. Thyer, and S. W. Franks,
Duan, Q., S. Sorooshian, and V. Gupta, “Effective and efficient global opti- “Characterizing errors affecting areal rainfall estimates: application to uncer-
mization for conceptual rainfall-runoff models,” Water Resources Research, tainty quantification and decomposition in hydrologic modelling,” 32nd
28: 1015–1031, 1992. Hydrology and Water Resources Symposium, Engineers Australia, Canberra,
Efron, B., “Why isn’t everyone a Bayesian?” The American Statistician, Newcastle, Australia, 2009.
40 (1):1–5, 1986. Renard, B., D. Kavetski, M. Thyer, G. Kuczera, and S. W. Franks,
Evin, G., D. Kavetski, M. Thyer, and G. Kuczera, “Pitfalls and improvements “Understanding predictive uncertainty in hydrologic modeling: le challenge
in the joint inference of heteroscedasticity and autocorrelation in hydrological of identifying input and structural errors,” Water Resources Research, 46:
model calibration,” Water Resources Research, 49 (7): 4518–4524, 2013. W05521, 2010, doi: 05510.01029/02009WR008328.
23_Singh_ch23_p23.1-23.10.indd 9 8/22/16 1:34 PM

Renard, B., D. Kavetski, E. T. Leblois, M. Thyer, G. Kuczera, and S. W. Franks, Thyer, M., B. Renard, D. Kavetski, G. Kuczera, S. W. Franks, and S. Srikanthan,
“Toward a reliable decomposition of predictive uncertainty in hydrological “Critical evaluation of parameter consistency and predictive uncertainty in
modeling: characterizing rainfall errors using conditional simulation,” Water hydrological modeling: a case study using Bayesian total error analysis,” Water
Resources Research, 47 (11): W11516, 2011. Resources Research, 45: W00B14, 2009, doi:10.1029/2008WR006825.
Robert, C. P. and G. Casella, Monte Carlo statistical methods. Springer Tolson, B. A. and C. A. Shoemaker, “Dynamically dimensioned search
Texts in Statistics. Springer, New York, 2004, p. 650. algorithm for computationally efficient watershed model calibration,” Water
Schaefli, B., D. B. Talamba, and A. Musy, “Quantifying hydrological model- Resources Research, 43: W01413, 2007.
ing errors through a mixture of normal distributions,” Journal of Hydrology, Viglione, A., R. Merz, J. L. Salinas, and G. Blöschl, “Flood frequency
332 (3–4): 303–315, 2007. hydrology: 3. A Bayesian analysis,” Water Resources Research, 49:675–692,
Schoups, G. and J. A. Vrugt, “A formal likelihood function for parameter 2013, doi: 10.1029/2011WR010782.
and predictive inference of hydrologic models with correlated, heterosce- Vrugt, J. A. and B. A. Robinson, “Improved evolutionary optimization from
dastic and non-Gaussian errors,” Water Resources Research, 46: W10531, genetically adaptive multimethod search,” Proceedings of the National
2010. Academy of Sciences of the United States of America, 104 (no. 3): 708–711,
Sorooshian, S. and J. A. Dracup, “Stochastic parameter estimation proce- 2007, doi: 10.1073/pnas.0610471104.
dures for hydrologic rainfall-runoff models: correlated and heteroscedastic Vrugt, J. A., C. J. F. ter Braak, M. P. Clark, J. M. Hyman, and B. A. Robinson,
error cases,” Water Resources Research, 16 (2): 430–442, 1980. “Treatment of input uncertainty in hydrologic modeling: doing hydrology
Stedinger, J. R. and T. A. Cohn, “Flood frequency analysis with histori- backward with Markov chain Monte Carlo simulation,” Water Resources
cal and paleoflood information,” Water Resources Research, 22: 785–793, Research, 44: W00B09, 2008, doi:10.1029/2007WR006720, 002008.
1986. Vrugt, J. A., C. J. F. ter Braak, C. G. H. Diks, B. A. Robinson, J. M. Hyman,
Thyer, M., G. Kuczera, and Q. J. Wang, “Quantifying parameter uncertainty and D. Higdon, “Accelerating Markov chain Monte Carlo simulation by differ-
in stochastic models using the Box-Cox transformation,” Journal of Hydrology, ential evolution with self-adaptive randomized subspace sampling,” International
265 (1–4): 246–257, 2002. Journal of Nonlinear Sciences and Numerical Simulation, 10 (3): 273–290, 2009.
23_Singh_ch23_p23.1-23.10.indd 10 8/22/16 1:34 PM

Chapter 24
Optimization Approaches for
Integrated Water Resources
Management
BY
ANDRÉ DOZIER, MAZDAK ARABI, JOHN W. LABADIE,
AND DARRELL FONTANE
ABSTRACT Simulation benefits from being intuitive and descriptive of the system it rep-
resents, but finding good solutions through simulation proves difficult in
Meeting competing demands for water with adequate quality to support
large, multidimensional, and highly constrained decision spaces. Numerical
populations, economies, and ecosystems proves difficult due to uncertainties
optimization techniques efficiently explore these spaces for more beneficial
in climate, land use, and demographics. Increasingly, sustainability and equi-
solutions (Hashimoto, 1982).
tability factors are included as additional criteria with which to assess perfor-
Differing and sometimes conflicting stakeholder perspectives about water
mance of water management solutions, requiring stakeholder input and
introduce multiple objectives, or criteria, with which to judge system perfor-
environmental assessment. Optimization methods provide systematic
mance (Reed, 2008). Thus, for several decades, managers and scientists
approaches to evaluate tradeoffs in meeting integrated water resources
throughout the globe have searched for more integrated, sustainable, and
management (IWRM) goals. This chapter analyzes trends in optimization
equitable approaches to natural resource management, leading to wide-
techniques as research moves from single-objective technical solutions to
spread adoption of an IWRM paradigm (White, 1998). This paradigm shift
multiobjective stakeholder-driven solutions. Linear programming (LP) and
led to the development of management criteria within hydrographic, socio-
dynamic programming (DP) were the preferred optimization techniques
economic, environmental, and political-administrative contexts for quanti-
until the turn of the twenty first century, when heuristic or evolutionary
fying sustainability and equitability (van der Zaag et al., 2002; Loucks and
methods became widely popular. Simulation-based optimization and decom-
Gladwell, 1999). Such integrated water management imposes multiple objec-
position techniques provide the means to improve system representation
tives on traditional engineering approaches, rendering technical approaches
while performing optimization. A growing number of researchers have recog-
potentially less efficient in a pragmatic sense (i.e., physically, financially, or
nized that technical solutions do not perform well in addressing sustainability
politically) in favor of more equitable and sustainable management schemes
and equitability, and thus have proposed role-playing games and serious
(Giordano and Shah, 2014).
online video games as stakeholder-driven approaches. Future research should
The goal of this chapter is to portray the trends in optimization methods
focus on increasing the efficiency of multiobjective analysis and simulation-
developed to address the challenges introduced by IWRM. The objectives of
based optimization, as well as further exploration of methods to harvest
this work are to
stakeholder solutions within optimization.
1. Define optimization in the context of integrated water management
2. Outline trends in optimization methods in water-related subdisciplines
24.1 INTRODUCTION
3. Discuss challenges and gaps to present a roadmap for future research
Previous literature and textbooks already provide general formulations and
Extreme weather-related events, climate change impacts, aging water infra- algorithms for solving linear, convex, nonconvex, nonsmooth, discrete, sto-
structure, and degradation of freshwater resources present increasingly chal- chastic, and dynamic optimization problems (Loucks and van Beek, 2005). To
lenging dilemmas for water resource managers throughout the globe avoid redundancy, this chapter instead highlights trends in optimization
(Trenberth et al., 2003; Bouwer, 2000; Schwarzenbach, 2006). Growing popu- methodologies as the water management community moves toward an
lation and uncertain climate narrow the window of acceptable water system IWRM paradigm.
operations because of increased competition for water resources, uneven
distributions of water in space and time, and encroachment on riverine-
ecology landscapes due to development (Oki and Kanae, 2006; Vörösmarty 24.1.1 Optimization in the Context of Integrated
et al., 2000; Gourbesville, 2008). Legal restrictions and financial limitations Water Resources Management
further constrain operations and management of freshwater resources Although this chapter lies within a hydrology handbook, scientific investiga-
(Neuman, 1998; Saleth and Dinar, 2000). Geophysical processes, including tions within hydrology rarely utilize optimization methods except in calibra-
hydrological processes, operate as a holistic system regardless of whether tion of hydrologic models (covered in Chap. 63) and briefly in the study and
human-based management schemes view them as such (Reid et al., 2010). modeling of optimality in ecohydrological systems (Eagleson and Tellers,
Detailed simulation models of water systems often aid decision-making by 1982). For this reason, we focus on water resources management applications
assessing the effects of solution strategies formulated by system managers in and problem perspectives. IWRM, more thoroughly discussed in Chap. 140,
a very detailed and methodological manner (Rani and Moreira, 2009). refers to a process that “promotes the coordinated development and
24-1
24_Singh_ch24_p24.1-24.8.indd 1 8/22/16 1:38 PM

24-2 Optimization Approaches for Integrated Water Resources Management
management of water, land, and related resources, in order to maximize the others have utilized optimization techniques that directly incorporate
resultant economic and social welfare in an equitable manner without com- stakeholder values as weighted objectives in multiobjective optimization
promising the sustainability of vital ecosystems” (Agarwal et al., 2000). or preferences in multi-criteria decision analysis (MCDA) (Voinov and
Although some question the IWRM paradigm (Giordano and Shah, 2014), its Bousquet, 2010). Analysis of literature indicates an emerging trend in the
widespread adoption has set guiding principles that require a change from search for integrated, sustainable, and equitable water management:
traditional engineering to integrated analysis of water projects (Bouwer, 2000; gamification.
Gourbesville, 2008). Table 24.1 lists previous literature surveys that contextu-
alize optimization problems and algorithms used within various water 24.2 TRENDS
resource subsystems. Optimization in the context of IWRM integrates and
synthesizes these various subproblems so as to address decision-making more Previous literature reviews cover more specific topics within water management
holistically (Cai, 2008; Kelly et al., 2013). Thus, a special focus is placed on (see Table 24.1), and a few papers broadly review water management as a whole
integrative and multiobjective optimization techniques that combine tradi- (McKinney et al., 1999; Singh, 2012; Loucks et al., 2005). We therefore briefly
tional engineering approaches with economic, social, or environmental pro- analyze trends in optimization methods used in each field. Emphasis on the
cesses, impacts, and tradeoffs. development of methodologies over time highlights the rapidly growing popular-
Although every optimization problem embeds some kind of simulation ity of multiobjective analysis and the use of evolutionary computation techniques
or prescriptive model of the system under consideration, detailed and real- to solve both single and multiobjective problem formulations. Figure 24.1 displays
istic simulation models have rarely appeared within the historical context of a water-related literature trend analysis (see appendix) clearly indicating that
optimization. This has hindered adoption of these detailed models by stake- evolutionary algorithms have surpassed every other form of optimization in
holders and operations engineers (Labadie, 2004; Yeh, 1985). Highly simpli- water-related topics since the early 2000s. Multiobjective analysis has also been
fied system representation within optimization models leads to large rapidly climbing to a prominent position in water system operations research.
uncertainty in the impacts and tradeoffs of resultant management policies. Most multiobjective analyses consist of techniques constrained to multiple objec-
As computing power and speed continue to improve, however, more com- tives such as epsilon-constraint, the weighting method, MCDA, and analytical
plex system descriptions have entered into optimization methodologies. hierarchy process, while about one-third of multiobjective analyses stem from
Simulation-based optimization techniques such as reinforcement learning multiobjective evolutionary algorithms. These trends show a growing desire
(RL) and evolutionary computation offer promising capabilities to incorpo- within water-related research fields to improve realistic system representation,
rate more realistic models within optimization (Rani and Moreira, 2009; search nonsmooth objective functions, and incorporate more sustainable and
Lee and Labadie, 2007; Rieker and Labadie, 2012; Belaineh et al., 1999; equitable solutions. Selected subdisciplines are discussed in the following subsec-
Safavi et al., 2009). tions, followed by elaboration on techniques useful for interdisciplinary synthesis
Optimization in a traditional sense provides technical solutions through and incorporation of stakeholder input.
the use of computer-based numerical techniques. However, a number of
researchers and practitioners within the water resource management com- Table 24.1 Literature surveys of optimization methods used for operations
munity have come to believe that effective, sustainable, and equitable water research and management of various water resources systems
management requires stakeholder participation (Kelly et al., 2013; Seppälä,
2002; Voinov and Bousquet, 2010; Steins and Edwards, 1999; Barreteau et System Surveys
al., 2007). The public tends to adopt stakeholder-driven, bottom-up solu-
tions more easily than technical solutions within a democratic setting General overviews Loucks and van Beek, 2005; McKinney et al., 1999;
Singh, 2012
(Voinov and Bousquet, 2010; Korfmacher, 2001; Gaddis et al., 2010),
Reservoirs Labadie, 2004; Yeh, 1985; Wurbs, 1993;
whereas technical, top-down solutions in other political settings may be
Wardlaw and Sharif, 1999
more efficient at achieving practical outcomes (Giordano and Shah, 2014).
Irrigation networks Singh, 2013
Empirical evidence indicates that appropriate collective action effectively
Groundwater Wagner, 1995; Gorelick, 1983; Yeh, 1992
manages a multiple-use common-pool resource, which includes most water
Distribution network design Kang and Lansey, 2011; Walski, 1995
resource projects (Steins and Edwards, 1999; Ostrom, 1990). Today, many
Floodplain management Kumar et al., 2010
water management practitioners have taken to elicitation of stakeholder
Conjunctive use Singh, 2014
ideas by performing participatory modeling tasks with stakeholders, while
Total (7059) Total (7059)

104 104
Evolutionary (3282) Single objective (5445)
Nonlinear programming (1374) Multiobjective (1652)
Linear programming (1296)
Cumulative publications
Dynamic programming (1097)

103 Multiobjective-specific (958) 103
Decomposition (293)
102 102
101 101
100 100
1950 1960 1970 1980 1990 2000 2010 2020 1950 1960 1970 1980 1990 2000 2010 2020
Year
Figure 24.1 Literature trends on a semi-log scale from 1950 to 2013 showing the number of articles that mention various optimization
methods over time (left panel), and the number of articles that are identified to be either single-objective or multiobjective analyses
(right panel). “Multiobjective-Specific” optimization methods consist of those specific to multiobjective analysis such as epsilon-con-
straint, compromise programming, and multi-criteria decision analysis, but do not include those methods that can also be used for sin-
gle-objective analysis such as multiobjective genetic algorithms. Matching articles are constrained to those with a match for water-related
topics in the title, abstract, or keywords sections of publication metadata. Metadata is provided by Web of Science®. More information on
how this figure is produced is found in the appendix.
24_Singh_ch24_p24.1-24.8.indd 2 8/22/16 1:38 PM

Trends 24-3
24.2.1 Reservoir Operations and Management 24.2.2 Irrigation Systems

Reservoir operations problems represent classic testing grounds for new Irrigation systems herein may refer either to operations of conveyance sys-
optimization techniques due to their large degree of nonlinearity, multiob- tems or single farm crop selection and patterns. Common objectives for
jective purposes, and significant economic impact. Common objectives of operating irrigation systems include maximization of farm profit or water
reservoir operations problems include maximizing revenue from sale of productivity, and minimization of waterlogging, groundwater depletion, or
hydropower or water supply, minimizing shortages, maximizing energy transpiration (Singh, 2013; Ali and Talukder, 2008). Since the 1970s, LP and
production, and minimizing vulnerability and damage due to floods, DP in addition to stochastic variants of these techniques have dominated
among many others (Wurbs, 1993). Since the 1960s, many researchers have irrigation system operations research, where NLP remains less popular
utilized LP in addition to its stochastic variants such as chance-constrained (Singh, 2013; Kipkorir, 2001). When linked with design or economic objec-
linear programming (CCLP) and stochastic linear programming (with tives, Lagrangian multiplier methods aid solution of nonlinear terms and
decomposition) to perform reservoir system optimization (Labadie, 2004; relaxation of nonlinear constraints (González-Cebollada and Macarulla,
Yeh, 1985). Network flow optimization solves a subset of linear optimiza- 2012). Since the late 1990s, researchers have rapidly introduced GAs as the
tion problems and remains a popular simulation and optimization approach dominant heuristic technique (Singh, 2013).
in practice due to its speed and graphical representation of system con-
straints (Labadie, 2004; Wurbs, 1993).
Techniques such as interior point (IP), quadratic programming (QP), 24.2.3 Groundwater
sequential linear (and sequential quadratic) programming (SLP and SQP), Groundwater management can take many different forms, but in general it
Frank-Wolfe (or conditional gradient) method, generalized reduced gra- aims to improve groundwater quantity, quality, or allocation (Wagner,
dient (GRG), and conjugate gradient method (CGM) comprise some of 1995). Objective functions include minimizing pumping costs, maximizing
the mathematical, gradient-based techniques used in reservoir optimiza- profit, minimizing storage requirements or stream depletion, minimizing
tion to solve nonlinear problems, commonly referred to as nonlinear water deficit in scarce locations, and maximizing well production or
programming (NLP) techniques (Labadie, 2004; Yeh, 1985; Wurbs, 1993; hydraulic head, among many others (Gorelick, 1983). Since the late 1960s
Dai and Labadie, 2001). Some methods such as CGM do not directly and early 1970s, LP has been the most commonly used method to optimize
handle constraints, and therefore do not generally apply to a large number groundwater management, and methods such as the embedding method or
of problems. Integer programming methods such as branch-and-bound, response matrices were introduced to spatially and temporally approximate
cutting plane, and branch-and-cut help to address introduction of discrete groundwater response to management decisions within an LP formulation
variables in addition to nonlinear constraints and objective function (Gorelick, 1983; Yeh, 1992). Other early methods included mixed integer
terms (Labadie, 2004; Higgins et al., 2011). Optimization of problems programming techniques, QP, OCT, and DP (Singh, 2012; Gorelick, 1983;
with nonlinear constraints can be performed using Lagrange multipliers Yeh, 1992). As in other water subdisciplines, evolutionary optimization
and duality theory, a methodology also known as the method of multipli- methods such as GA, SA, and PSO applied to groundwater management
ers (MOM) or the Augmented Lagrangian method (Labadie, 2004; problems rapidly replaced the prominent optimization methods in the late
Bertsekas, 1996). 1990s to early 2000s (Singh, 2012; Wagner, 1995). Many simulation-based
Researchers have preferred to use DP, optimal control theory (OCT), and optimization techniques have been explored to ensure proper solution of
RL due to the sequential, nonlinear, nonconvex, nonsmooth, and stochastic groundwater flow equations (Wagner, 1995; Gorelick, 1983). These simula-
nature of many reservoir system management and operational problems tion techniques present mechanisms for integrating surface water into
(Labadie, 2004; Lee and Labadie, 2007; Rieker and Labadie, 2012). After the groundwater management (Singh, 2012).
discovery of DP by Richard Bellman in the mid-1950s, DP quickly rose in
the 1960s as a technique to solve reservoir operations problems, outpacing
application of LP techniques even though LP was formulated first (Cottle et 24.2.4 Water Distribution Networks
al., 2007; Dreyfus, 2002). To alleviate the so-called “curse of dimensionality” Water distribution system design cannot easily be obtained through use of
when considering multireservoir systems, researchers in the late 1960s and traditional mathematical-programming techniques due to the discrete
early 1970s began to develop DP variants that enhance solution speed of nature of the decision variables and cost functions, uncertain demand,
multidimensional problems such as Incremental DP (IDP) and DP and complex energy function (Taher and Labadie, 1996). Thus, design
Successive Approximations (DPSA) (Labadie, 2004; Yeh, 1985). Additionally, techniques tend toward direct search or simulation-based optimization.
functional approximation methods such as orthogonal polynomials, cubic Objective functions usually include minimizing total cost (both construc-
splines, and artificial neural networks (ANNs) can be used to speed objec- tion and operation of pipes, pumps, etc.) with constraints regarding
tive function evaluation (Tejada-Guibert et al., 1995; Bertsekas and pressure and flow criteria (Kang and Lansey, 2011; Walski, 1995).
Tsitsiklis, 1995). Researchers have historically utilized hybrid variants of mathematical and
Since the advent of the personal computer in addition to rapidly increasing heuristic techniques, trading off accurate system representation with
processing speeds in the early 1990s, literature in reservoir operations computational tractability (Cunha and Sousa, 1999). To improve system
research has been highlighting a widespread, growing interest in evolutionary representation, researchers have more recently employed fully heuristic
programming techniques, also referred to as heuristic techniques. programming techniques such as GAs, PSO, ACO, SA, TS, SCE, harmony
Evolutionary computation can handle nonsmooth, nonconvex, multimodal, search (HS), immune algorithms (IAs), memetic algorithms (MAs), and
and stochastic problems, but typically requires many more objective function shuffled frog-leaping algorithms (SFLA), among others (Cunha and
evaluations, converges more slowly, cannot guarantee convergence on an Sousa, 1999; di Pierro, 2009). MCDA methods also help to support deci-
optimum, and is often unsuitable for dynamic optimization problems due to sion making around design and planning for water distribution networks
the large search space [10]. Heuristic techniques used for reservoir operations (Scholten et al., 2014).
problems include genetic algorithms (GAs), particle swarm optimization
(PSO), ant colony optimization (ACO), honey-bee mating optimization
(HBMO), simulated annealing (SA), tabu search (TS), and shuffled complex 24.2.5 Floodplain Management
evolution (SCE) (Rani and Moreira, 2009; Labadie, 2004; Wardlaw and Sharif, The application of optimization for floodplain management represents a rela-
1999; Baltar and Fontane, 2008). Hybrids of these heuristic techniques with tively small field of study, perhaps due to its excessively large data and model-
mathematical techniques have served to address computational inefficiency ing requirements for land use change, or perhaps because of the small number
and highly constrained systems. of floodplain managers. Flood planning aims at minimizing expected value of
Methods such as epsilon-constraint, the weighting method, goal program- flood damages through various permanent and emergency actions consisting
ming, compromise programming, multiobjective GAs (MOGAs), and other of both structural and nonstructural flood mitigation strategies (Lund, 2002).
variants of heuristic techniques search multiobjective problems for a Very few researchers have applied LP, NLP, mixed-integer programming, or
Pareto optimal solution (Rani and Moreira, 2009; Labadie, 2004). MCDA DP to the field of floodplain management (Lund, 2002). Starting in 1997 with
techniques optimize over a finite number of alternatives using stakeholder the optimization of layout and sizing of detention systems by Yeh and
input to assign preferences or otherwise rank the alternatives based on Labadie, GAs have been used to optimize floodplain decisions in both single-
multiple criteria (i.e., objectives) (Labadie, 2004; Voinov and Bousquet, and multiple-objective forms (Woodward, 2013; Yeh, 1997). Expert systems
2010). Similar trends are observed in other subdisciplines of water and MCDA techniques have also been applied to flood management (Kumar
resources management. et al., 2010).
24_Singh_ch24_p24.1-24.8.indd 3 8/22/16 1:38 PM

24.2.6 Water Supply Planning and Conjunctive Use Simulation-based optimization techniques require a methodology to repre-
Water supply planning and conjunctive use of surface water and groundwa- sent operating rules given current system state. Tables, linear or polynomial
ter resources aim at diversifying water sources to improve reliability of rules, expert systems, regression models, fuzzy rule-based systems, ANNs, and
water supply and water quality (Singh, 2012). Optimization in the context hybrid approaches provide mechanisms to map system state to operating rule
of conjunctive use has evolved from determining irrigation water source decisions (Rani and Moreira, 2009; Labadie, 2004). Because no structure is
decisions in the 1960s to integrating other aspects of water resources man- assumed a priori, optimization methods that utilize fuzzy rules or ANNs
agement such as demand-side management, groundwater remediation, and often outperform those using simple operating rule structures such as linear
water ownership transfers (Pulido-Velázquez et al., 2006). Surface water and operating rules (Wan et al., 2006).
groundwater interactions complicate conjunctive use modeling, and have 24.2.8 Incorporating Stakeholder Values and Human
thus consumed a large portion of research in the field (Safavi et al., 2009). Knowledge in Optimization
The earliest formal optimization techniques used for conjunctive water sup-
ply planning consisted of DP and LP, with LP and its stochastic variants As mentioned earlier, properly implemented integrated water management
emerging as the dominate approach (Singh, 2012). Due to discrete variables requires stakeholder input, yet most of the optimization techniques dis-
and model integration requirements of water supply problems, decomposi- cussed so far are purely numerical mechanisms guided by water managers,
tion, integer programming, and mixed-integer programming techniques researchers, and engineers. MCDA, Bayesian belief networks (BBN), com-
have dominated the nonlinear mathematical approaches to conjunctive use promise programming, and the analytic hierarchy process (AHP) incorpo-
optimization (Watkins and McKinney, 1998; Cai et al., 2001). Since the rate stakeholder values as preferences associated with the various criteria,
2000s, evolutionary methods such as GAs and MOGAs have quickly entered and thus guide technical solutions towards stakeholder values (Voinov and
literature as simulation-based solutions to optimize conjunctive use prob- Bousquet, 2010). PROMETHEE and ELECTRE III apply fuzzy sets within
lems (Safavi et al., 2009; Peralta et al., 2014; Singh, 2014). To improve hydro- MCDA to represent uncertainties that can exist in representing stakeholder
logic system representation without significantly increasing computation preferences (Labadie, 2004). MCDA methods have been successfully
time, ANNs and other linear and nonlinear regression techniques often act applied within reservoir operations and water distribution system design
as a surrogate for more sophisticated, physicallybased approaches. For (Labadie, 2004; Scholten et al., 2014). Other methods such as expert or
example, stream-aquifer interactions often require excessive computations knowledge-based systems directly take past human experience to build a
through a finite difference model such as MODFLOW, but not when repre- model of the decision-making process that finds solutions to a previously
sented within an ANN (Triana et al., 2010). recognized problem (Liao, 2005).
Stakeholder participation has progressed from building decision support
24.2.7 Integrated Analysis of Multipurpose Water systems through the use of simulation and optimization methodologies
Resource Systems (Voinov and Bousquet, 2010), to being the decision support system itself (Le
As discussed previously, IWRM requires accurate representations of water sys- Bars and Grusse, 2008; Lankford et al., 2004). A growing community has
tems in order to manage water sustainably and equitably across economic, attempted to optimize allocation of resources in a more equitable and sustain-
sociopolitical, and environmental (ESE) spheres (Cai, 2008; McKinney et al., able manner by utilizing experimental games, also referred to as simulation
1999; Cai et al., 2003). Optimization applications for floodplain and water qual- games or role-playing games (Barreteau et al., 2007; Valkering et al., 2012).
ity management integrate social, environmental, and legal systems due to their Since the 1930s and 1940s, the field of experimental economics has utilized
proximity to land use change and management (Lund, 2002; Tong and Chen, experimental games to study human behavior and economic values
2002). Land use change, ecosystem water use, and societal interactions with and (Kahneman and Smith, 2002; Neumann and Morgenstern, 1944). Economists
perspectives on water will increasingly play larger roles in simulation and opti- now view these lab experiments as a common tool to study behavior, perform
mization as research progresses to provide more representative modeling of conflict resolution, test policies, and develop new solutions for natural
water systems. resource management (Barreteau et al., 2007; Ehmke and Shogren, 2008).
Hierarchical optimization employs mature optimization techniques to One role-playing game showed that human adaptation tended toward non-
solve subproblems of an overarching master optimization problem, the dual structural flood control measures (Valkering et al., 2012).
of which is often solved with subgradient or evolutionary techniques Several physical board games, computer simulation games, and hybrids have
(Labadie, 2004; Cai et al., 2001). Dantzig-Wolfe decomposition has benefited also been utilized in localized contexts to solve water management issues
irrigation ditch planning by linking higher-level water allocation optimiza- throughout the globe (Le Bars and Grusse, 2008; Lankford et al., 2004; Dray et
tion with lower-level cropping-pattern optimization (Paudyal and Das Gupta, al., 2006; Rebolledo-Mendez et al., 2009). Experience from one study suggests
1990). Among other decomposition techniques, Lagrangian relaxation (LR), that a computer-assisted board game simplifies system representation enough
or the MOM, remains perhaps one of the most popular techniques used for to aid stakeholder understanding, expression of opinions, and generation of
optimization and model integration within hydrothermal coordination stud- solutions so as to consequently reduce collective disagreement (Dray et al.,
ies (Dozier, 2012), as well as in conjunctive surface-water and groundwater 2006). Serious online video gaming, or games with a purpose, extends economic
management (Gorelick, 1983). Hierarchical techniques have thus demon- experiments by crowdsourcing the search for integrated water management
strated utility in integrating across disciplines. solutions, in addition to the more important immediate and decentralized
OCT offers a very similar technique to LR that allows optimization of a analysis of the impacts and tradeoffs associated with those solutions (Le Bars
multistage, sequential decision problem simultaneously constrained by and Grusse, 2008). Games also provide a mechanism to arouse awareness over
potentially complex system dynamic equations through use of the Hamiltonian specific water issues (Rebolledo-Mendez et al., 2009; Rizzoli et al., 2014). When
(Labadie, 2004). OCT has been utilized to link economics optimization with applied to water resources management, video gaming simultaneously improves
natural resource management, mainly forests and fisheries, since 1969, where system representation through simulation and optimizes management strate-
recent research focuses on fine-tuning spatial and temporal resolution gies that are driven by stakeholders, a difficult, if not impossible, task for tradi-
(Dorfman, 1969; Sanchirico and Wilen, 2005). Unlike OCT, DP is an easily tional optimization methods (Harou et al., 2009).
discretized and constrained form of dynamic optimization, and has thus been
highly utilized within water management subdisciplines to perform inte- 24.3 CHALLENGES AND RESEARCH GAPS
grated hydroeconomic optimization (Harou et al., 2009). With the notable
exception of RL (Lee and Labadie, 2007; Rieker and Labadie, 2012), most DP Researchers recognize a need to integrate across scales of analysis and across
techniques require addition of multiple artificial state variables to incorporate disciplines within optimization to improve water system representation. For
processes beyond a simple Markov decision process, where the state is trans- example, integration of transmission and distribution scales in water distribu-
ferred through more than one discretized timestep (Shim et al., 2002). tion network design improves the search for cost-effective solutions (Kang
Addition of state variables significantly increases computation time, but and Lansey, 2011). Detailed simulation models offer the most straightforward
methods such as IDP and DPSA help to alleviate the large computation time capability to synthesize across subdisciplines in addition to being more attrac-
associated with large multidimensional problems (Labadie, 2004). Some tive and intuitive to water managers and other practitioners than less-
authors have adapted the differential dynamic programming (DDP) technique detailed, prescriptive optimization models (Labadie, 2004). Utilization of
to incorporate more detailed simulation (Culver and Shoemaker, 1993). simulation models within hierarchical optimization techniques through use
Simulation-based optimization techniques like RL, although less mature than of Lagrange multipliers, Hamiltonian functions, and simulation-based
traditional optimization techniques, provide a straightforward and intuitive dynamic optimization methods serves to synthesize across various scales
platform for integrating system representation across various disciplines (Pulido-Velázquez et al., 2006). Heuristic or evolutionary computation meth-
(Harou et al., 2009). ods also offer direct search capabilities by running a simulation model and
24_Singh_ch24_p24.1-24.8.indd 4 8/22/16 1:38 PM

REFERENCES 24-5
evaluating the objective function from model outputs, as is often performed 24.6 APPENDIX: LITERATURE TREND ANALYSIS
in automating model calibration.
A broad and automated analysis of literature from 1950 to 2013 provided
Difficulties with simulation-based optimization approaches arise due to
helpful insight into evolving trends within various subdisciplines of IWRM.
computational limitations. Within water resource management literature,
Figure 24.1 shows the number of publications regarding categorized meth-
high performance computation through massive parallelization, cloud com-
odologies in addition to a comparison of single objective and multiobjective
puting, and distributed data stores remain underutilized even though compu-
analyses. We assembled a list of more than 50 optimization methods catego-
tational benefits are clear. Promising and particularly useful techniques to
rized by LP (2 methods), nonlinear (mathematical) programming (17 meth-
hasten simulations are functional approximation methods and incremental or
ods), decomposition techniques (3 methods), DP (6 methods), evolutionary
successive approximation approaches (Labadie, 2004; Safavi et al., 2009;
computation (17 methods), and multiobjective analysis (7 methods). In
Bertsekas and Tsitsiklis, 1995; Johnson and Rogers, 2000). Future research
addition to mathematical, gradient-based methods, NLP techniques
will focus on reducing the computation time, improving convergence charac-
included discrete programming methods such as integer programming,
teristics, and generalizing software implementations of simulation-based
mixed-integer programming, branch-and-bound, and cutting plane meth-
optimization methods.
ods. The top four NLP techniques returned in the search for water resource
Technical optimization can produce top-down, pragmatic solutions to
management-related topics consist of conjugate-gradient method (147),
specific problems, but may not incorporate equitable or sustainable criteria
mixed-integer programming (106), QP (105), and integer programming
with meaningful stakeholder input. Thus, a variety of methods including
(100 records). We searched the Web of Science® database for each method
MCDA, BBN, AHP, expert systems, role-playing games, and serious online
within literature topics, which includes any occurrence of a word within the
video games have applied stakeholder values and even stakeholder-driven
title, abstract, or keywords. Searches were constrained to a specific water-
ideas within the searching mechanisms for solutions to various water resource
related discipline by adding a keyword to each query, such as “river,” “irriga-
management issues (Voinov and Bousquet, 2010; Barreteau et al., 2007; Le
tion system,” etc.
Bars and Grusse, 2008). Although they remain far from being mature and
The number of articles containing a multiobjective analysis was estimated
stable fields of study and practice in water resources management, role-play-
by placing an additional constraint on the query: AND (“multiobjective” OR
ing games and serious online video games offer a promising mechanism to
“multi-objective” OR “multiple objective*”). Single-objective analyses were
incorporate stakeholder-driven solutions directly within the optimization
queried using the NOT operator instead: NOT “multiobjective” NOT “multi-
technique. Such stakeholder-driven optimization results in large oscillations
objective” NOT “multiple objective*”. The total number of articles that con-
and adjustments in the search for optimal water management strategies due
tain multiobjective analyses is 1635 from 1950 to 2013, 958 of which also
to uncertainties, trial and error, inaccuracies, misperceptions, and insufficient
contain descriptions of multiobjective-specific methods (MCDA, goal pro-
knowledge of the systems (Le Bars and Grusse, 2008; Sterman, 1989; Diehl
gramming, compromise programming, AHP, etc.), and 580 of which also
and Sterman, 1995; Cronin et al., 2009). These issues acknowledge the formi-
discuss evolutionary methods. Figure 24.1 displays the total number of arti-
dable challenges faced in applying stakeholder-driven approaches, but also
cles found within Web of Science® for each category.
present exciting new research to enhance human-computer interactions for
Data included in the trend analysis are derived from the Web of Science®
crowdsourcing water management solutions.
prepared by Thomson Reuters®, Inc. (Thomson®), Philadelphia, Pennsylvania,
USA: © Copyright Thomson Reuters® 2015. All rights reserved. Data before
24.4 CONCLUSION the 1960s or 1970s is of lower quality and sometimes does not contain article
abstracts. Conference proceedings were incorporated into the metadata col-
IWRM served as a context for the goal of this chapter to analyze trends in
lection only after 1990, which may explain the large increase in all areas of
optimization techniques. Motivations for focusing on IWRM stem from a
research at that time.
growing interest in more integrative and holistic modeling solutions that
incorporate multiple, nontechnical objectives within economic, social, and
environmental spheres. Simulation modeling offers the best platform on
REFERENCES
which to build accurate models of physical water systems. Over the past 60
years, operations research literature in water management clearly indicates a Agarwal, A., M. S. de los Angeles, R. Bhatia, I. Chéret, S. Davila-Poblete, M.
trend toward more accurate system representation within optimization mod- Falkenmark, F. Gonzalez-Villarreal, et al., “Integrated water resources man-
els through use of evolutionary computation and other forms of simulation- agement,” Global Water Partnership, Swedish International Development
based optimization. Due to their capability to break large optimization Agency, Stockholm, Sweden, Vol. 4, 2000.
problems into subproblems of a vastly simpler nature, decomposition meth- Ali, M. H. and M. S. U. Talukder, “Increasing water productivity in crop
ods have become commonplace in hydrothermal coordination studies, and production—a synthesis,” Agricultural Water Management, 95 (11): 1201–
are of growing interest in conjunctive use and water quality management 1213, 2008.
studies. Due to the excessive computation time of heuristic and simulation- Baltar, A. M. and D. G. Fontane, “Use of multiobjective particle swarm
based techniques, parallel implementations of simulation models and optimi- optimization in water resources management,” Journal of Water Resources
zation methods may provide future research directions, in addition to Planning Management, 134 (3): 257–265, 2008.
improving algorithm searching capabilities, search space reduction, and Barreteau, O., C. Le Page, and P. Perez, “Simulation and gaming in natural
functional approximation methods. resource management,” Simulation & Gaming, 38 (2): 181–184, 2007.
IWRM imposes economic, social, and environmental criteria on a tradi- Belaineh, G., R. Peralta, and T. Hughes, “Simulation/optimization model-
tionally engineering-based field of study. Multiobjective analyses of water ing for water resources management,” Journal of Water Resources Planning
management strategies now often incorporate sustainability and equitability and Management, 125 (3): 154–161, 1999.
metrics across such criteria. Due to the difficulty in establishing suitable Bertsekas, D. P., Constrained Optimization and Lagrange Multiplier
metrics, practitioners have attempted to incorporate stakeholder values in Methods, Athena Scientific, Belmont, MA, 1996.
settling on predetermined solutions through MCDA and other similar Bertsekas, D. P. and J. N. Tsitsiklis, “Neuro-dynamic programming: an
techniques. Others have recognized the need to involve stakeholders in overview,” Proceedings of the 34th IEEE Conference on Decision and Control,
identifying the solutions, and have held workshops in which stakeholders New Orleans, LA, Vol. 1, 1995, pp. 560–564.
directly participated in modeling exercises and role-playing games. An Bouwer, H., “Integrated water management: emerging issues and chal-
emerging crowdsourcing practice, serious online video gaming, also prom- lenges,” Agricultural Water Management, 45 (3): 217–228, 2000.
ises to incorporate stakeholder-driven solutions and analyses into challeng- Cai, X., “Implementation of holistic water resources-economic optimization
ing integrated water management problems. models for river basin management–reflective experiences,” Environmental
Modelling & Software, 23 (1): 2–18, 2008.
Cai, X., D. C. McKinney, and L. S. Lasdon, “Integrated hydrologic-agro-
24.5 ACKNOWLEDGMENTS
nomic-economic model for river basin management,” Journal of Water
This work is partially supported by a grant from the Agriculture and Food Resources Planning and Management, 129 (1): 4, 2003.
Research Initiative of the USDA National Institute of Food and Agriculture Cai, X., D. C. McKinney, L. S. Lasdon, and D. W. Watkins, “Solving large
(NIFA), grant number #2012-67003-19904, and partially by National Science nonconvex water resources management models using generalized Benders
Foundation IGERT Grant No. DGE-0966346: I-WATER: Integrated Water, decomposition,” Operations Research, 49 (2): 235–245, 2001.
Atmosphere, Ecosystem Education, and Research Program at Colorado State Cottle, R., E. Johnson, and R. Wets, “George B. Dantzig (1914–2005),”
University. Notices of the AMS, 54 (3): 344–362, 2007.
24_Singh_ch24_p24.1-24.8.indd 5 8/22/16 1:38 PM

Cronin, M. A., C. Gonzalez, and J. D. Sterman, “Why don’t well-educated Lee, J-H. and J. W. Labadie, “Stochastic optimization of multireservoir sys-
adults understand accumulation? A challenge to researchers, educators, and tems via reinforcement learning,” Water Resources Research, 43 (11): 1–16, 2007.
citizens,” Organizational Behavior and Human Decision Processes, 108 (1): Loucks, D. P. and E. van Beek, Water Resources Systems Planning and
116–130, 2009. Management: An Introduction to Methods, Models and Applications, United
Culver, T. B. and C. A. Shoemaker, “Optimal control for groundwater reme- Nations Educational, Scientific and Cultural Organization (UNESCO), Paris,
diation by differential dynamic programming with quasi-Newton approxima- France, 2005.
tions,” Water Resources Research, 29 (4): 823–831, 1993. Loucks, D. P., E. van Beek, J. R. Stedinger, J. P. M. Dijkman, and M. T.
Cunha, M. da C. and J. Sousa, “Water distribution network design optimi- Villars, Water resources planning and management: an overview, Water
zation: simulated annealing approach,” Journal of Water Resources Planning Resources Systems Planning and Management: An Introduction to Methods,
and Management, 125 (4): 215–221, 1999. Models and Applications, United Nations Educational, Scientific and Cultural
Dai, T. and J. W. Labadie, “River basin network model for integrated water Organization (UNESCO), Paris, France, 2005, pp. 2–37.
quantity/quality management,” Journal of Water Resources Planning and Loucks, D. P. and J. S. Gladwell, Sustainability Criteria for Water Resource
Management, 127 (5): 295–305, 2001. Systems, Cambridge University Press, Cambridge, United Kingdom, 1999.
Diehl, E. and J. D. Sterman, “Effects of feedback complexity on dynamic Kang, D. and K. Lansey, “Revisiting optimal water-distribution system
decision making,” Organizational Behavior and Human Decision Processes, design: issues and a heuristic hierarchical approach,” Journal of Water
62 (2): 198–215, 1995. Resources Planning and Management, 138 (3): 208–217, 2011.
di Pierro, F., S-T. Khu, D. Savić, and L. Berardi, “Efficient multi-objective Kipkorir, E. C., D. Raes, and J. W. Labadie, “Optimal allocation of short-term
optimal design of water distribution networks on a budget of simulations irrigation supply,” Irrigation and Drainage Systems, 15 (3): 247–267, 2001.
using hybrid algorithms,” Environmental Modelling & Software, 24 (2): Korfmacher, K. S., “The politics of participation in watershed modeling,”
202–213, 2009. Environmental Management, 27 (2): 161–176, 2001.
Dorfman, R., “An economic interpretation of optimal control theory,” Labadie, J. W., “Optimal operation of multireservoir systems: state-of-the-
American Economic Review, 59 (5): 817–831, 1969. art review,” Journal of Water Resources Planning and Management, 130 (2): 93,
Dozier, A. Q., Integrated water and power modeling framework for renewable 2004.
energy integration, M.S. Thesis, Department of Civil Engineering, Colorado Lankford, B., C. Sokile, D. Yawson, and H. Levite, “The river basin game: a
State University, Fort Collins, CO, 2012. water dialogue tool,” International Water Management Institute, Colombo, Sri
Dray, A., P. Perez, N. Jones, C. Le Page, P. D’Aquino, I. White, and T. Lanka, 75, 2004.
Auatabu, “The AtollGame experience: from knowledge engineering to a com- Liao S., “Expert system methodologies and applications—a decade review
puter assisted role playing game,” Journal of Artificial Societies and Social from 1995 to 2004,” Expert Systems with Applications, 28 (1): 93–103, 2005.
Simulation, 9 (1): 1–12, 2006. Le Bars, M. and P. Le Grusse, “Use of a decision support system and a
Dreyfus, S., “Richard Bellman on the birth of dynamic programming,” simulation game to help collective decision-making in water management,”
Operations Research, 50 (1): 48–51, 2002. Computers and Electronics in Agriculture, 62 (2): 182–189, 2008.
Eagleson, P. S. and T. E. Tellers, “Ecological optimality in water-limited Lund, J. R., “Floodplain planning with risk-based optimization,” Journal of
natural soil-vegetation systems 2. Tests and applications,” Water Resources Water Resources Planning and Management, 128 (3): 202–207, 2002.
Research, 18 (2): 341–354, 1982. McKinney, D. C., X. Cai, M. W. Rosegrant, C. Ringler, and C. A. Scott,
Ehmke, M. D. and J. F. Shogren, “Experimental methods for environment “Modeling water resources management at the basin level: review and future
and development economics,” Environment and Development Economics, directions,” System-Wide Initiative on Water Management, International
14 (04): 419, 2008. Water Management Institute, Colombo, Sri Lanka, Vol. 6, 1999.
Gaddis, E. J. B., H. H. Falk, C. Ginger, and A. A. Voinov, “Effectiveness of a Neuman, J. C., “Beneficial use, waste, and forfeiture: the inefficient search
participatory modeling effort to identify and advance community water for efficiency in western water use,” Environmental Law, 28: 919–996, 1998.
resource goals in St. Albans, Vermont,” Environmental Modelling & Software, Oki, T. and S. Kanae, “Global hydrological cycles and world water
25 (11): 1428–1438, 2010. resources,” Science, 313 (5790): 1068–1072, 2006.
Giordano, M. and T. Shah, “From IWRM back to integrated water Ostrom, E., Governing the Commons: The Evolution of Institutions for
resources management,” International Journal of Water Resources Development, Collective Action, Cambridge University Press, Cambridge, UK, 1990.
30 (3): 364–376, 2014. Paudyal, G. N. and A. Das Gupta, “Irrigation planning by multilevel opti-
González-Cebollada, C. and B. Macarulla, “Comparative analysis of design mization,” Journal of Irrigation and Drainage Engineering, 116 (2): 273–291,
methods of pressurized irrigation networks,” Irrigation and Drainage, 61 (1): 1990.
1–9, 2012. Peralta, R. C., A. Forghani, and H. Fayad, “Multiobjective genetic algorithm
Gorelick, S. M., “A review of distributed parameter groundwater manage- conjunctive use optimization for production, cost, and energy with dynamic
ment modeling methods,” Water Resources Research, 19 (2): 305–319, 1983. return flow,” Journal of Hydrology, 511: 776–785, 2014.
Gourbesville, P., “Challenges for integrated water resources management,” Pulido-Velázquez, M., J. Andreu, and A. Sahuquillo, “Economic optimi-
Physics and Chemistry of the Earth, Parts A/B/C, 33 (5): 284–289, 2008. zation of conjunctive use of surface water and groundwater at the basin
Harou, J. J., M. Pulido-Velazquez, D. E. Rosenberg, J. Medellín-Azuara, J. R. scale,” Journal of Water Resources Planning and Management, 132 (6): 454–
Lund, and R. E. Howitt, “Hydro-economic models: concepts, design, applica- 467, 2006.
tions, and future prospects,” Journal of Hydrology, 375 (3–4): 627–643, 2009. Rani, D. and M. M. Moreira, “Simulation–optimization modeling: a survey
Hashimoto, T., J. R. Stedinger, and D. P. Loucks, “Reliability, resiliency, and and potential application in reservoir systems operation,” Water Resources
vulnerability criteria for water resource system performance evaluation,” Management, 24 (6): 1107–1138, 2009.
Water Resources Research, 18 (1): 14–20, 1982. Rebolledo-Mendez, G., K. Avramides, S. de Freitas, and K. Memarzia,
Higgins, A. J., B. A. Bryan, I. C. Overton, K. Holland, R. E. Lester, D. King, “Societal impact of a serious game on raising public awareness: the case of
M. Nolan, et al., “Integrated modelling of cost-effective siting and operation FloodSim,” in Proceedings of the 2009 ACM SIGGRAPH Symposium on Video
of flow-control infrastructure for river ecosystem conservation,” Water Games, New Orleans, LA, 2009, pp. 15–22.
Resources Research, 47 (5): W05519, 2011. Reed, M. S., “Stakeholder participation for environmental management: a
Johnson, V. and L. Rogers, “Accuracy of neural network approximators in literature review,” Biological Conservation, 141 (10): 2417–2431, 2008.
simulation-optimization,” Journal of Water Resources Planning and Reid, W. V., D. Chen, L. Goldfarb, H. Hackmann, Y. T. Lee, K. Mokhele, E.
Management, 126 (2): 48–56, 2000. Ostrom, et al., “Earth system science for global sustainability: grand chal-
Kelly, R. A., A. J. Jakeman, O. Barreteau, M. E. Borsuk, S. ElSawah, S. H. lenges,” Science, 330 (6006): 916–917, 2010.
Hamilton, H. J. Henriksen, et al., “Selecting among five common modelling Rieker, J. D. and J. W. Labadie, “An intelligent agent for optimal river-reser-
approaches for integrated environmental assessment and management,” voir system management,” Water Resources Research, 48 (9): W09550, 2012.
Environmental Modelling & Software, 47: 159–181, 2013. Rizzoli, A. E., A. Castelletti, A. Cominolab, P. Fraternali, A. D. dos Santos,
Kumar, D. N., K. Srinivasa Raju, and F. Baliarsingh, “Modeling for flood B. Storni, R. Wissmann-Alves, et al., “The SmartH2O project and the role of
control and management,” Natural and Anthropogenic Disasters: Vulnerability, social computing in promoting efficient residential water use: a first analysis,”
Preparedness and Mitigation, edited by M. Jha, Springer, The Netherlands, in 7th International Congress on Environmental Modelling and Software, San
2010, pp. 147–168. Diego, CA, 2014, pp. 1559–1567.
Kahneman, D. and V. Smith, “Foundations of behavioral and experimental Safavi, H. R., F. Darzi, and M. A. Mariño, “Simulation-optimization model-
economics,” Advanced Information on the Prize in Economic Sciences, pp. ing of conjunctive use of surface water and groundwater,” Water Resources
1–25, 2002. Management, 24 (10): 1965–1988, 2009.
24_Singh_ch24_p24.1-24.8.indd 6 8/22/16 1:38 PM

REFERENCES 24-7
Saleth, R. M. and A. Dinar, “Institutional changes in global water sector: van der Zaag, P., I. M. Seyam, and H. H. G. Savenije, “Towards measurable
trends, patterns, and implications,” Water Policy, 2: 175–199, 2000. criteria for the equitable sharing of international water resources,” Water
Sanchirico, J. N. and J. E. Wilen, “Optimal spatial management of renew- Policy, 4 (1): 19–32, 2002.
able resources: matching policy scope to ecosystem scale,” Journal of Voinov, A. A. and F. Bousquet, “Modelling with stakeholders,” Environmental
Environmental Economics and Management, 50 (1): 23–46, 2005. Modelling & Software, 25 (11): 1268–1281, 2010.
Schwarzenbach, R. P., B. I. Escher, K. Fenner, T. B. Hofstetter, C. A. von Neumann, J. and O. Morgenstern, Theory of Games and Economic
Johnson, U. von Gunten, and B. Wehrli, “The challenge of micropollutants in Behavior, Princeton, New Jersey, Princeton University Press, 1944.
aquatic systems,” Science, 313 (5790): 1072–1077, 2006. Vörösmarty, C. J., P. Green, J. Salisbury, and R. B. Lammers, “Global water
Scholten, L., A. Scheidegger, P. Reichert, M. Mauer, and J. Lienert, resources: vulnerability from climate change and population growth,” Science,
“Strategic rehabilitation planning of piped water networks using multi-crite- 289 (5477): 284–289, 2000.
ria decision analysis,” Water Research, 49: 124–143, 2014. Wagner, B. J., “Recent advances in simulation-optimization groundwater
Seppälä, O. T., “Effective water and sanitation policy reform implementa- management modeling,” Reviews of Geophysics, 33 (no. S2): 1021–1028, 1995.
tion: need for systemic approach and stakeholder participation,” Water Policy, Walski, T. M., “Optimization and pipe-sizing decisions,” Journal of Water
4: 367–388, 2002. Resources Planning and Management, 121 (4): 340–343, 1995.
Shim, K., D. Fontane, and J. W. Labadie, “Spatial decision support system Wan, Y., J. W. Labadie, K. Konyha, and T. Conboy, “Optimization of fre-
for integrated river basin flood control,” Journal of Water Resources Planning quency distribution of storm-water discharges for coastal ecosystem restora-
and Management, 128 (3): 190–201, 2002. tion,” Journal of Water Resources Planning and Management, 132 (5): 320–329,
Singh, A., “An overview of the optimization modelling applications,” 2006.
Journal of Hydrology, 466–467: 167–182, 2012. Wardlaw, R. and M. Sharif, “Evaluation of genetic algorithms for optimal
Singh, A., “Irrigation planning and management through optimization reservoir system operation,” Journal of Water Resources Planning and
modelling,” Water Resources Management, 28 (1): 1–14, 2013. Management, 125 (1): 25–33, 1999.
Singh, A., “Simulation–optimization modeling for conjunctive water use Watkins, D. W. and D. C. McKinney, “Decomposition methods for water
management,” Agricultural Water Management, 141: 23–29, 2014. resources optimization models with fixed costs,” Advances in Water Resources,
Steins, N. A. and V. M. Edwards, “Platforms for collective action in multiple-use 21 (4): 283–295, 1998.
common-pool resources,” Agriculture and Human Values, 16 (3): 241–255, 1999. White, G. F., “Reflections on the 50-year international search for integrated
Sterman, J., “Misperceptions of feedback in dynamic decision making,” water management,” Water Policy, 1 (1): 21–27, 1998.
Organizational Behavior and Human Decision Processes, 43: 301–335, 1989. Woodward, M., B. Gouldby, Z. Kapelan, and D. Hames, “Multiobjective
Taher, S. and J. W. Labadie, “Optimal design of water-distribution networks with optimization for improved management of flood risk,” Journal of Water
GIS,” Journal of Water Resources Planning Management, 122 (4): 301–311, 1996. Resources Planning and Management, 140 (2): 201–215, 2013.
Tejada-Guibert, J. A., S. A. Johnson, and J. R. Stedinger, “The value of Wurbs, R. A., “Reservoir-system simulation and optimization models,”
hydrologic information in stochastic dynamic programming models of a Journal of Water Resources Planning and Management, 119 (4): 455–472,
multireservoir system,” Water Resources Research, 31 (10): 2571–2579, 1995. 1993.
Triana, E., J. W. Labadie, T. K. Gates, and C. W. Anderson, “Neural network Yeh, C. and J. W. Labadie, “Multiobjective watershed-level planning of
approach to stream-aquifer modeling for improved river basin management,” storm water detention systems,” Journal of Water Resources Planning and
Journal of Hydrology, 391 (3–4): 235–247, 2010. Management, 123 (6): 336–343, 1997.
Trenberth, K. E., A. Dai, R. M. Rasmussen, and D. B. Parsons, “The chang- Yeh, W. W-G., “Reservoir management and operations models: a state-of-
ing character of precipitation,” Bulletin of the American Meteorological Society, the-art review,” Water Resources Research, 21 (12): 1797–1818, 1985.
84 (9): 1205–1217, 2003. Yeh, W. W-G., “Systems analysis in ground-water planning and manage-
Tong, S. T. Y. and W. Chen, “Modeling the relationship between land use ment,” Journal of Water Resources Planning and Management, 118 (3): 224–
and surface water quality,” Journal of Environmental Management, 66 (4): 237, 1992.
377–393, 2002.
Valkering, P., R. van der Brugge, A. Offermans, M. Haasnoot, and H.
Vreugdenhil, “A perspective-based simulation game to explore future
pathways of a water-society system under climate change,” Simulation &
Gaming, 44 (2–3): 366–390, 2012.
24_Singh_ch24_p24.1-24.8.indd 7 8/22/16 1:38 PM

Chapter 25
Nonparametric Methods
BY
UPMANU LALL AND BALAJI RAJAGOPALAN
ABSTRACT to simulate system performance for samples drawn from this distribution.
Likewise, the hydrologist may be interested in estimating the probability dis-
This chapter discusses two types of very broadly applicable methods for
tribution of the output variable(s) from the model given uncertain inputs and
hydrologic data analysis. The first is nonparametric function estimation, and
parameters. The traditional tools available are fitting the data to a specified
refers to a data driven approach for approximating target regression or prob-
probability distribution (e.g., the Normal, Generalized Extreme Value, or
ability density functions for use with hydrologic models. The second pertains
Gamma distribution). As with the earlier example, there may be many cases in
to nonparametric hypothesis testing and inference with a broad set of applica-
which the data may exhibit a structure (e.g., bimodality) that is not captured
tions ranging from trend identification to clustering to testing whether or not
well by these simple models. This may become more apparent as the sampling
two sets of data come from the same population.
frequencies for hydrologic data increase, and the total volume of data or the
sampling of the variable’s state space increases. In this case as well, the hydrol-
25.1 INTRODUCTION ogist may be interested in exploring methods that are flexible in fitting the
target function, while making some basic assumptions as to the continuity,
Hydrological modeling often refers to a set of conservation equations that differentiability, and other general properties of these functions. This recogni-
include terms for appropriate physical, chemical, and biological processes. tion has led to a rapid development and application of a wide range of methods
Sometimes, parameterizations of these processes in terms of effective of nonparametric function estimation. An early review of these developments
parameters and equations relating state variables are available. For instance, for hydrologic applications was offered by Lall (1995). Basic ideas have been
Darcy’s Law relates porous media-flow velocities to the spatial gradient of applied for the estimation of probability densities, regression functions, spatial
hydraulic potential, through the parameter hydraulic conductivity. The field interpolation and simulation, time series simulation and forecasting,
parameter, hydraulic conductivity may vary in space. In this case, given space trend analysis, and point process representation. Since these are the building
and time observations of hydraulic head, one can try to do inference on blocks for a wide range of hydroclimatic analyses, it is not surprising that
the spatial variation of the hydraulic conductivity, within the framework of the applications to precipitation, floods, streamflow, climate change, groundwater,
associated conservation equations. The hydrologist may also be interested in water quality, agro-hydro-ecology, and soil processes have been reported. The
developing contours of hydraulic head, which would necessitate interpolation number of methods proposed and used have also increased dramatically.
of the observed values in space and time. Such an interpolation could be done In addition to function estimation, hydrologists are often concerned with a
using the physics-based model, as part of a strategy to infer hydraulic variety of hypothesis testing and inference problems. A typical example would
conductivities, or be pursued directly with the data. In the first case, typically, be a hypothesis test to see whether two samples of data on a hydrologic process
a numerical discretization of the study domain is considered, and the hydraulic could have come from the same underlying population. Such a test may be
conductivity and hydraulic heads associated with each “grid” are estimated useful to assess the effect of a specific treatment, or land use or climate change.
such that the error of prediction of the observations is minimized. This leads One approach to such tests was to assume that the samples (after appropriate
to a large, complex, inverse problem to be solved, where the number of transformation) were drawn from a known probability distribution (e.g.,
parameters are typically much larger than the number of observations. To Normal), and to then use the sampling properties of that distribution to assess
make such a problem tractable, the modeler may assume a simple spatial whether or not the parameters of that distribution as estimated from the two
function (e.g., a polynomial in space) which has a small number of parameters samples are different from each other at a prespecified probability. In line with
to be estimated as a descriptor of the unknown hydraulic conductivity field, or the earlier discussion, one can question whether the assigned probability model
consider a regularization strategy (Sun, 2013), where the effective dimension is indeed adequate, and this has led to a corresponding direction in nonpara-
of the hydraulic conductivity and/or hydraulic head fields is reduced by metric methods, where much weaker assumptions are made as to the underly-
assumptions as to the continuity and differentiability of the parameter of ing probability model. An example of such nonparametric methods, in the
interest, and a trade-off between the spatial variation of the values of the target modern context of computational statistics is the bootstrap (Bradley Efron,
variable and the ability to replicate the observations is considered. There is an 1982, 2000; Noreen, 1989), which considers resampling the data from the
implicit recognition, that a simple polynomial function is unlikely to adequately underlying empirical probability distribution with or without a direct specifica-
represent the variation of the parameter of interest to the hydrologist, and that tion of the underlying probability function.
a more flexible inference strategy that does not impose such a model is needed.
Yet, the flexible strategy needs to make some general assumptions about the
25.2 DEFINITIONS
nature of variation of the parameter to constrain the system identification.
Similar considerations emerge in other estimation and inference problems A parametric model is one in which the functional form for the model is fully
faced by the hydrologist. The hydrologic model may need a specification of the specified by an explicit equation with a fixed number of parameters (e.g.,
external forcing to the model, for example, the space-time distribution of pre- location, scale, shape, correlation, or tail thickness), and the inference of these
cipitation. This forcing may be available from a numerical weather prediction parameters is assumed to then fully describe the underlying process.
or climate model, or need to be developed from irregularly sampled space- A nonparametric model does not assume a fixed functional form. Rather,
time precipitation data. Even if a point or lumped model is considered, it may it seeks to develop estimates of that functional form or test hypotheses under
be necessary to estimate the probability distribution of the external forcing and modest assumptions as to the behavior of that function. Such a model still has
25-1
25_Singh_ch25_p25.1-25.8.indd 1 8/22/16 1:41 PM

25-2 Nonparametric Methods
parameters that need to be specified, but the number of these parameters 2001; Wand, 2001; Yakowitz, 1989; and Szidarovszky, 1985). A univariate
needed may change (increase) as the amount of data available increases, and nonparametric regression estimator can be written as:
depending on the nature or degree of variation in the data. This definition n
covers the methods in the class of nonparametric density estimation and E[ y | x ] = ∑yi g ( x , xi )K (ui ) (25.4)
regression (Ruppert, 2001; Scott, 2015; Silverman, 1986), examples of which i=1
are provided in the next section. They also cover nonparametric hierarchical where E[.] is the expected value operator, and u and K(.) are defined as before.
Bayesian models (Quintana and Muller, 2004; Teh and Jordan, 2010), which K(.) need not be a probability density function, and g(.,.) may be a specific
allow the use of latent variables to form mixtures of parametric distributions parametric function (e.g., an mth order polynomial, with m preselected). The
that lead to quite general approximation properties. separation of g(.) and K(.) makes it clear that K(.) is a function that allows
A second type of nonparametric method in statistics is one that uses local estimation of a target function g(.). Splines and local regression or con-
distribution-free methods (Hollander, Wolfe, and Chicken, 2013; Maritz, ditional density estimation are easily expressed in this form. Since both g(.)
1995). These include methods based on order statistics or ranks, which do not and K(.) are functions of just x and xi, one can actually skip this distinction
require any assumption as to an underlying probability model for the data in and just describe the weight function as K(.), covering both aspects.
question. Such methods tend to be robust to the misspecification of the With these, very general definitions one can provide a classification for some
underlying probability model. of the common nonparametric function estimation methods in Table 25.1.
Nonparametric methods are much more flexible in representing the underly- From Table 25.1, it is clear that there is a large class of distinct nonparamet-
ing data processes than parametric methods. The price paid is that if the para- ric function estimation methods that share the common feature, that the data
metric model were indeed correct, then the nonparametric method, which plays a large role in determining the target function, and that while they differ
implicitly considers model uncertainty, will be less efficient than the parametric in the details of the estimation procedure, in all cases, they seek to estimate
model. Where a theoretical probability model is justified by the data, the associ- the target function through the application of a weight function to the obser-
ated parametric model will have a much smaller uncertainty associated with it vations. Combinations of these methods, for example wavelets and local
than the nonparametric approach, especially as the sample sizes increase. regression, or splines and Gaussian processes are also reported in the litera-
However, when there is uncertainty as to the correct model, and the data do not ture. While the basic ideas associated with many of these methods were pre-
correspond to the typical parametric models considered, the nonparametric sented 50–100 years ago, access to large-scale computation has facilitated the
methods lead to a lower bias of estimation, and more robust conclusions, espe- rapid development and application of these methods over the last 20 years,
cially as the sample sizes increase. They are also much more in the spirit of and they are now central to the Machine Learning and Computationally
“letting the data speak”, that is, exploratory analyses. Intensive Statistics Literature.
25.3.2 Nonparametric Inference
25.3 METHODS
Nonparametric inference refers to methods for statistical hypothesis tests
Density estimation and regression function estimation are the building blocks which make no or very few assumptions as to the particular probability
of models in time-series analysis, analysis of extremes, spatial interpolation, distribution that generated the data. There are many such tests (Conover,
simulation of random fields, and clustering or classification. The general 1999), and the ones most commonly used in hydrology are discussed in
structure of these building blocks is discussed first. Distribution-free methods Helsel and Hirsch (1992). These are briefly summarized as follows:
are used primarily for hypothesis testing as to trends or comparisons of • Kendall’s tau (Kendall, 1938): It measures statistical dependence between
multiple groups of data. An overview of these methods is provided next. two variables, through an examination of the association of their ranks. It is also
used as a test for trend, when the computation is performed between a hydro-
25.3.1 Nonparametric Function Estimation logic variable and time. Adaptation of Kendall’s tau for the case where the data
may have serial correlation is provided by Hamed and Ramachandra Rao (1998).
The goal of nonparametric density (intensity) and regression estimation is to
• Mutual Information (Moon, Rajagopalan, and Lall, 1995): It measures
approximate a wide class of target functions under a minimum of assumptions
dependence across two variables based on joint probability distribution of the
as to the nature of that function. Many of these methods can be considered in
variables.
the framework of kernel estimators, which strive to smooth a target function of
• Sen Slope (Helsel and Hirsch, 1992; Hirsch, Slack, and Smith, 1982): It is
the observed data through the application of a kernel or weight function. The
a robust estimator of the linear trend that is often estimated in conjunction
probability density function f(x) of a multivariate variable x can be estimated as:
with the Kendall’s tau.
n • Mann–Whitney U or Wilcoxon rank sum test (Mann and Whitney,
f ( x ) = ∑K (|| x − xi ||) (25.1) 1947): It tests whether two samples are drawn from the same distribution.
i =1 • Spearman’s rank correlation coefficient (Helsel and Hirsch, 1992): It
measures monotonic statistical dependence between two variables using their
where K(u) refers to a kernel function, ||u|| is a distance measure on the argu- ranks and the possible permutations of the ranks.
ment u, x is a point at which the density is desired, xi is d dimensional obser- • Kolmogorov–Smirnov test (Smirnov, 1948): It tests whether a sample is
vation vector for the ith observation, and n is the number of observations. drawn from a given distribution, or whether two samples are drawn from the
More generally, for any function g(x) of the data (e.g., the intensity func- same distribution. An extension for the multivariate case is provided by
tion), an estimator of its variation with x can be formed as: Fasano and Franceschini (1987).
n • Bootstrap (Efron and Tibshirani, 1993; Efron, 1982, 2000): It is a very
f ( x ) = ∑g ( xi )K (|| x − xi ||) (25.2) general way of drawing samples from the underlying distribution of the
i=1 data to test a variety of hypotheses, and to estimate confidence limits on
statistics computed on a finite sample. The most typical case is to draw a
Two examples can now be provided—for univariate kernel density estima- sample with replacement, from the original observations of the same
tion and for nonparametric regression. A univariate kernel density estimator length as the original sample. For the multidimensional case, a vector of
with a Gaussian kernel can be written as: the variables is drawn preserving the dependence across variables, unless
n −u 2 a hypothesis test of the dependence across the variables is of interest in
K (ui ) x − xi 1 which case one sample is drawn preserving, and by randomizing, the
f (x ) = ∑ ; ui = ; K (u) = e 2 (25.3)
i=1
nh( x , xi ) h( x , xi ) 2π variables for each draw. There are several cases or examples of bootstrap
methods:
where h(x, xi) is a bandwidth that can be a function of either or both the • Block bootstrap: The data are split into nonoverlapping blocks and
point of estimate x, and the location of observations xi or both, or it can be a observations are sampled from them.
constant h that determines the degree of smoothing of the target density func- • Moving block bootstrap (Kunsch, 1989): Sequential blocks of length b
tion. The kernel function and the bandwidth are the two parameters of the are formed and then resampled as blocks, in case serial dependence of length
method. Any valid probability density function can be used as a kernel. less than b is expected.
The multivariate extension (Scott, 2015) of (25.3) is particularly useful, • Smoothed bootstrap: Samples are drawn from a nonparametric density
since it can provide a basis for conditional density estimation, that is, f(y|x) = estimate (e.g., kernel density) fit to the data.
f(y,x)/f(x). One can draw samples from the conditional distribution, as moti- • Markovian bootstrap (Lall and Sharma, 1996): A time-series structure is
vated in (Lall and Sharma, 1996; Sharma, Tarboton, and Lall, 1997) or develop considered for the data, and a conditional sampling strategy using the condi-
regression equations as a conditional expectation (Eubank, 1988; Ruppert, tional density given the current prediction vector is used.
25_Singh_ch25_p25.1-25.8.indd 2 8/22/16 1:41 PM

Applications 25-3
Table 25.1 Examples of Some Nonparametric Function Estimation Methods and Their Attributes
Method Focus and parameters Notes

Histogram Density estimation, bin width Sensitivity to origin and bin width, largely replaced by kernel density
estimators
Kernel density and regression Linear convolution of the data using a specified kernel function; Conceptual basis for many methods. Kernel choice based on assumptions of
estimators optimize bandwidth smoothness and edge effects, bandwidth is optimized using best fit criteria.
References (Scott, 2015; Silverman, 1986)
k-nearest neighbor methods Generalization of kernel estimators to consider data adaptive Can often lead to better automatic performance than fixed bandwidth ker-
bandwidth based on the data density in the neighborhood of the nel estimators especially for regression. The k-nn density function will not
point of estimation-parameterized using the distance to the kth integrate to 1. References (Mack and Rosenblatt, 1979; Scott, 2015)
nearest neighbor
Local regression and likelihood Generalization of kernel density and regression estimators where Lowess (Cleveland and Devlin, 1988) and Locfit (see, Lall, Moon, Kwon,
estimators the target function is estimated using a higher order kernel or and Bosworth, 2006; Wand, 2001) are the most popular methods in this
polynomial locally, often with bandwidth chosen by the distance class (Loader, 2006). They provide diagnostics and fitting tools comparable
to k-nearest neighbors to those from most parametric methods
Spline methods Piecewise Polynomials with continuity and differentiability con- For density estimation, see Kooperberg and Stone (1991); for spline
straints at “knots” or specified breakpoints. Can be shown to regression, see Bosworth and Lall (1995), Craven and Wahba (1978),
have an equivalent high order kernel. The number and location Friedman (1991), Gu (2013), Ruppert (2001), Wahba (1990)
of knots translates into an effective variable bandwidth.
Kriging Parametric function assumed for spatial correlation of the Kriging was introduced for spatial estimation. Reference: (Cressie, 2015)
(and closely related dependent variable. The shape of the function determines the The Gaussian Process models (can be seen as finite mixture models) are
Gaussian process models) kernel, and its “range” parameter determines the bandwidth. closely related and can be applied for regression. Can include a parametric
Equivalence to splines under certain conditions. trend function. Models with splines as the trend function and a spatial
correlation kernel also exist. References (Neal, 1998; Shi and Wang, 2008)
Wavelets and wavelet regression Wavelets specify a kernel, but consider multiple bandwidths to The predictor and the predictands can be wavelet transformed and then a
capture different scales of variation of the target function. predictive model can be built using the appropriate scales for each variable.
Wavelet analysis is typically used to identify scales at which References (Donoho and Johnstone, 1995; Hollander et al., 2013; Oh and
there is a high variance or “signal.” Thresholding is used to Lee, 2005; Torrence and Compo, 1998)
determine a cutoff scale or bandwidth for regression.
Artificial neural networks The kernel function is identified through a multi-layer convolu- A very large literature with diverse developments and applications
(ANNs) tion of data-driven weights (Chap. 11). ANNs can be considered nonparametric if you allow the
number of hidden layers to be large Reference (Haykin, 2008)
Support vector machines Solve for a kernel function on the data such that the nonlinear A constrained optimization problem needs to be solved, and where the
function is approximated with a bounded error, and with a mini- kernel function is specified by the data, the computational cost can be high.
mum norm or total size of all the weights. Structural properties References (Daoud and Turabieh, 2013; Smola and Schölkopf, 2004;
of the kernel function are assumed. Suykens and Vandewalle, 1999; Vapnik, 2013)
Nonparametric Bayesian Typically finite mixture models for either density estimation or The data is used to determine the complexity of the model, given a
models regression, defined through a Dirichlet process, and associated noninformative prior over a potentially overspecified model.
prior probabilities across the candidate classes. References (Gershman and Blei, 2011; Müller and Quintana, 2004;
Teh and Jordan, 2009)
Cluster analysis Several methods used to identify distinct groups in a data set, k-means is the most commonly used and identified groups based on
i.e., mixtures in a density function. Methods include k-means maximizing distance between each groups centroid. Hierarchical
or k-medioids, k-d trees, hierarchical clustering, kernel density clustering has k-means as an end-point, but tries to partition the data in
estimation to identify multimodality, and Support Vector a tree so that each subcluster can be identified by a parent. K-d trees
Machines. partition the data into a tree as but use the variance along each
coordinate as a criterion to partition the tree along the median of the
coordinate in any partition with the largest spread. Silhouette methods
are useful for identifying the number of clusters in both k-means and
hierarchical clustering. References (Hartigan, 2012; Kaufman and
Rousseeuw, 2009; Lall and Bosworth, 1994; Minnotte, 1997;
Rousseeuw, 1987)
• Importance sampling (Tokdar and Kass, 2010): The data are resampled ties of the data are appropriately resampled while allowing randomness in
with probabilities specified by weights assigned to each original observation. the construction of the sequences. However, if the null hypothesis is that
• Bootstrap aggregating-bagging (Breiman, 1996): It is a model averaging there is no serial correlation, then the samples should be drawn randomly
approach to improve regression estimation by combining estimates from and the distribution of the bootstrap sample statistics should be compared to
randomly drawn samples from the original data. the statistics from the original sample.
• Multivariate smoothed bootstrap (Lall, Devineni, and Kaheil, 2015): The
multivariate data are sampled preserving cross-variable dependence using a 25.3.3 Software Implementation
combination of univariate nonparametric density estimates, and a smoothing The R statistical package (https://cran.r-project.org/) is an open-source
of the empirical multivariate copula. software that runs on multiple computing platforms, where all the methods
The bootstrap is the newest of these methods and has seen extensive indicated in this section are described with examples and implemented.
development and application since its introduction. If computational
resources are available, it provides the most general approach to nonpara- 25.4 APPLICATIONS
metric inference, since hypotheses about any function of the data can be Nonparametric methods have seen widespread application in hydrology.
explored. A key to the success of the bootstrap is that the samples drawn Broad categories of these applications are indicated as follows, with some
need to have similar properties to the original sample with respect to the representative references.
confidence intervals of interest or the hypothesis to be tested. This means
that if the original data have serial correlation, and one is interested in a
confidence interval of a statistic of the dataset, the bootstrap samples need to 25.4.1 Trend Identification
reflect a similar pattern. This could be achieved perhaps by importance Recognizing that hydrologic systems undergo changes overtime and space
sampling or by designing a block bootstrap so that the time sequence proper- due to variations in climate and regional changes in land use and/or
25_Singh_ch25_p25.1-25.8.indd 3 8/22/16 1:41 PM

regulation or diversion of flows, or pumping, the identification of the For hydroclimatic forecasting, nonparametric methods have been used
magnitude of trends in the mean, variance, and extremes of a time series or extensively to estimate conditional distributions given appropriate predictors
a collection of time series of hydrologic variables is necessary. If a (Abarbanel and Lall, 1996; Araghinejad, Burn, and Karamouz, 2006; Asefa,
statistically significant trend is identified, then the attribution of that trend Kemblowski, Lall, and Urroz, 2005; Bracken, Rajagopalan, and Prairie, 2010;
to putative causal factors is attempted. A possible strategy can be as follows: Coulibaly and Baldwin, 2005; Grantz, Rajagopalan, Clark, and Zagona, 2005;
• For each series, use Kendall’s tau and/or Spearman’s rank correlation Kalra and Ahmad, 2009; Kalra, Miller, Lamb, Ahmad, and Piechota, 2013;
versus time and the Sen Slope to assess whether there is a monotonic trend in Karlsson and Yakowitz, 1987; Kember, Flower, and Holubeshen, 1993; Lall,
the series. If multiple series are being analyzed, a field significance test to Moon, Kwon, and Bosworth, 2006; Lall et al., 1996; Moon, Lall, and Kwon,
establish whether the number of series where significant trends are found, 2007; Porporato and Ridolfi, 2001; Regonda, Rajagopalan, Lall, Clark, and
could be due to chance should be tested. For this, use a bootstrap to sample Moon, 2005; Sharma, Luk, Cordery, and Lall, 1999, 2000; Sharma, 2000; Smith,
each data vector (across all variables subjected to a trend test) with replace- Day, and Kane, 1992; Smith, 1991; Souza Filho and Lall, 2003). (Adamowski
ment and establish the probability that the number of series identified with and Feluch, 1991; Asefa, Kemblowski, McKee, and Khalil, 2006; Khan and
trends at a significance level, p, could have occurred by chance. An example Coulibaly, 2006; Kisi and Cimen, 2011; Li, Kwon, Sun, Lall, and Kao, 2009;
of this process is presented by (Douglas, Vogel, and Kroll, 2000; Krishnamurthy, Msiza, Nelwamondo, and Marwala, 2008; Wei, 2012.)
Lall, and Kwon, 2009; Yue, Pilon, and Phinney, 2003).
• Use wavelets to investigate nonmonotonic trends that may have a narrow 25.4.5 Stochastic Hydrology
band structure (Cioffi, Lall, Rus, and Krishnamurthy, 2015; Jain and Lall, Stochastic hydrologic models are described in Chap. 128. There have been
2001). Complement with LOWESS or LOCFIT to examine the nature of the significant developments of nonparametric methods in this area, with the
trend. Where, trends in extremes are of interest, local quantile modeling application of k-nearest neighbor and kernel density estimation methods,
(Chernozhukov, 2005; Koenker, Ng, and Portnoy, 1994; Sankarasubramanian moving block bootstrap methods and hidden Markov models being notable.
and Lall, 2003; Yu, Lu, and Stander, 2003) or modeling of a threshold exceed- Key references are Ghosh and Mujumdar (2007), Harrold, Sharma, and
ance process (Jain and Lall, 2000) may be of use. Sheather (2003), Khalil, Kwon, Lall, and Kaheil (2010), Kwon, Lall, and
• Consider clustering (Sun, Lall, Merz, and Dung, 2015; Westmacott and Obeysekera (2008), Kwon, Lall, and Khalil (2007), Mehrotra and Sharma
Burn, 1997) of the multiple series investigated to identify homogeneous (2005, 2006a, 2006b, 2007), Nowak, Prairie, Rajagopalan, and Lall (2010),
groups of sites. For each group, consider causal modeling using nonparamet- Prairie, Nowak, Rajagopalan, Lall, and Fulp (2008), Prairie, Rajagopalan, Lall,
ric regression on selected covariates for attribution of trends. and Fulp (2007), J. R. Prairie, Rajagopalan, Fulp, and Zagona (2006),
Rajagopalan, Lall, Tarboton, and Bowles (1997), Sharma and Lall (1999),
25.4.2 Spatial Interpolation and Simulation Sharma and O’Neill (2002), Sharma, Tarboton, and Lall (1997), Srinivas and
The typical problem addressed is that the hydrologic observations are Srinivasan (2005).
obtained at irregularly spaced locations, and either a grid average or a point
value at an unsampled location is of interest. Kriging is the most commonly 25.5 DISCUSSION
used method with applications to groundwater levels, rainfall, and streamflow
interpolation. However, other nonparametric regression methods (Coulibaly, Nonparametric methods are flexible, but given their data-adaptive nature
Bobée, Anctil, and Aravena, n.d.; Haylock et al., 2008; Hofierka, Parajka, tend to have lower bias but higher variance than a corresponding parametric
Mitasova, and Mitas, 2002; Hutchinson, 1995; Rajagopalan and Lall, 1998) alternative for function estimation. This statement assumes that the correct
have also been applied for the purpose. parametric alternative is known with certainty. Accounting for uncertainty in
Spatial simulation of precipitation or groundwater fields was also done the choice of a parametric model is equivalent to considering a diffuse
originally in the Kriging framework (Delhomme, 1978, 1979), but with addi- Bayesian prior over a set of candidate models. Loosely, this is conceptually
tional assumptions as to the probability distribution (e.g., Gaussian) associ- similar to what nonparametric Bayesian methods, that consider finite
ated with the target variable. One can also consider a spatial Markov or mixtures of candidate parametric models, or nonaparametric density or
Poisson process with nonhomogeneous intensity or transition intensity and regression models that consider a smoothing process over a high dimensional
estimate the space or time varying probabilities nonparametrically (Ali and set also try to do. Given computational resources, and a n-fold cross validation
Lall, 1998; Burton, Fowler, Kilsby, and O’Connell, 2010; Cressie, 2015; Diggle, design to test the performance of a candidate model, one can reliably choose
2013; Heikkinen and Arjas, 1998; Rajagopalan, Lall, and Tarboton, 1996). an appropriate nonparametric model, or converge to a parametric model.
Where time-series data are available, for example, for monthly or annual However, such testing of model performance needs to be done judiciously and
precipitation, simulations of the variables at multiple sites based on a non- thoroughly to avoid overfitting and poor subsequent predictive performance.
parametric approach, with empirical copulas and using a bootstrap frame- Typical measures of model fit, such as the generalized cross validation, the
work are also feasible (Lall, Devineni, and Kaheil, 2015). Akaike Information Criteria, and Mallows Cp tend to estimate the variance of
the estimate accounting for the effective number of parameters, but under the
25.4.3 Probability Density and Quantile/Return assumption that each candidate model is approximately unbiased. This may
Period Estimation not always be true, and hence formal cross validation needs to be performed
Some of the earliest applications (Adamowski and Feluch, 1990; Adamowski, to test model performance. Properly constructed bootstrap samples are often
1985; Apipattanavis, Rajagopalan, and Lall, 2010; Bardsley, 1989; Faucher, quite effective in such a setting as well.
Rasmussen, and Bobée, 2001; Moon, Lall, and Bosworth, 1993; Schuster and In all cases, it is worth emphasizing that the investigator needs to be clear on
Yakowitz, 1985) of nonparametric density estimation and regression methods the modeling objectives, and assumptions before choosing nonparametric
in hydrology were made for flood-frequency analysis. The motivation for versus parametric methods. In our view, nonparametric models should be used
much of this work was that the parametric estimators in vogue did not as exploratory tools to assess the performance of alternatives to specific para-
address mixtures of processes that may lead to extreme floods, and that the metric models, especially for relatively simple estimation settings. Often, the
upper tail behavior associated with large floods may not correspond to the standard parametric choices, especially in the presence of adequate data, are
data fit to the bulk of the distribution. These ideas were extended to consider surpassed by a nonparametric choice. This may be in terms of performance on
droughts, nonstationarity, and covariates (Durocher, Chebana, and Ouarda, a particular metric of fit, or more importantly in terms of the flexibility of
2015; López and Francés, 2013; Sadri and Burn, 2012; Sankarasubramanian specifying a model that captures the details of a physical process of interest very
and Lall, 2003; Santhosh and Srinivas, 2013; Yoo, Kwon, Kim, and Ahn, 2012). effectively. The k-nearest neighbor methods are particularly useful in this
regard since they allow an intuitive and effective design in many cases.
25.4.4 Hydroclimatic Forecasts and Nonlinear However, the k-nearest neighbor methods are subject to the choice of a distance
Dynamical Systems metric, and may be ineffective for high-dimensional predictors unless some
The recognition that hydroclimatic systems may potentially be modeled as attempt at dimension reduction is made, either directly exploiting the depen-
nonlinear dynamical systems led to a natural application of nonparametric dence structure of the predictors using Principal or independent component
methods (Abarbanel and Lall, 1996; Coulibaly and Baldwin, 2005; Lall, analysis, or through a linear projection of the dependent variable on the predic-
Sangoyomi, and Abarbanel, 1996; Porporato and Ridolfi, 2001) for the tors as illustrated in Souza Filho and Lall (2003) to assess appropriate weights.
recovery of the associated nonlinear map associated with the system. The The most interesting aspect of nonparametric methods is the ease with which
same recognition motivated the development of related methods (Lall and a complex modeling setting can be decomposed and addressed by a combina-
Sharma, 1996; Rajagopalan, Lall, and Tarboton, 1993; Sharma et al., 1997) for tion of different methods, for example, wavelets to decompose the long- and
nonlinear time series analysis of hydrologic systems that are described in the short-memory components of time series, and then a k-nearest neighbor or
chapter on stochastic hydrology. support vector or local polynomial modeling or moving block bootstrap
25_Singh_ch25_p25.1-25.8.indd 4 8/22/16 1:41 PM

REFERENCES 5
modeling of the dynamics of each component. Nonparametric Bayesian meth- iah/online/?IsisScript=iah/iah.xis&src=google&base=REPIDISCA&lang=p&

ods and Bayesian version of Gaussian process models are the emerging areas nextAction=lnk&exprSearch=12295&indexSearch=ID.
that promise a bridge between parametric and nonparametric methods for Craven, P. and W. Wahba, “Smoothing noisy data with spline functions,”
application. There are also emerging advances in learning methods for artificial Numerische Mathematik, 31 (4): 377–403, 1978, http://doi.org/10.1007/
neural network and support-vector machines. In this paper, we pointed out the BF01404567.
unifying framework under which most of these methods can be viewed. Their Cressie, N., Statistics for Spatial Data, John Wiley & Sons, New York, 2015.
effectiveness does vary, given the differences in implementation. The hydrolo- Retrieved from https://books.google.com/books?hl=en&lr=&id=MzN_
gist is well advised to consider this as a family of methods whose similarity and BwAAQBAJ&pgis=1.
difference need to be understood in the context of each application, while rec- Al Daoud, E. and H. Turabieh, “New empirical nonparametric kernels for
ognizing that asymptotically they have very similar behavior. support vector machine classification,” Applied Soft Computing, 13 (4):
1759–1765, 2013, http://doi.org/10.1016/j.asoc.2013.01.010.
Delhomme, J. P., “Kriging in the hydrosciences,” Advances in Water
REFERENCES
Resources, 1 (5): 251–266, 1978, http://doi.org/10.1016/0309-1708(78)
Abarbanel, H. D. I. and U. Lall, “Nonlinear dynamics of the Great Salt Lake: 90039-8.
system identification and prediction,” Climate Dynamics, 12 (4): 287–297, Delhomme, J. P., “Spatial variability and uncertainty in groundwater flow
1996, http://doi.org/10.1007/BF00219502. parameters: a geostatistical approach,” Water Resources Research, 15 (2):
Adamowski, K., “Nonparametric Kernel estimation of flood frequencies,” 269–280, 1979, http://doi.org/10.1029/WR015i002p00269.
Water Resources Research, 21 (11): 1585–1590, 1985, http://doi.org/10.1029/ Diggle, P. J., Statistical Analysis of Spatial and Spatio-temporal Point
WR021i011p01585. Patterns, 3rd ed., CRC Press, Boca Raton, 2013. Retrieved from https://books.
Adamowski, K. and W. Feluch, “Nonparametric flood‐frequency analysis google.com/books?hl=en&lr=&id=XR8bAAAAQBAJ&pgis=1.
with historical information,” Journal of Hydraulic Engineering, 116 (8), 1990. Donoho, D. and I. Johnstone, “Adapting to unknown smoothness via wave-
Retrieved from http://ascelibrary.org/doi/abs/10.1061/(ASCE)0733-9429 let shrinkage,” Journal of the American Statistical Association, 90 (432):
(1990)116:8(1035). 1200–1224, 1995, http://doi.org/10.2307/2291512.
Adamowski, K. and W. Feluch, “Application of nonparametric regression to Douglas, E. M., R. M. Vogel, and C. N. Kroll, “Trends in floods and low
groundwater level prediction” Canadian Journal of Civil Engineering, 18 (4): flows in the United States: impact of spatial correlation,” Journal of
600–606, 1991, http://doi.org/10.1139/l91-073. Hydrology, 240 (1–2): 90–105, 2000, http://doi.org/10.1016/S0022-
Ali, A. I. and U. Lall, “A continuous parameter homogeneous semi-Markov 1694(00)00336-X.
model for stratigraphic analyses from Borehole data,” Log Analyst, 39: 26–37, Durocher, M., F. Chebana, and T. B. M. J. Ouarda, “A nonlinear approach
1998. to regional flood frequency analysis using projection pursuit regression,”
Apipattanavis, S., B. Rajagopalan, and U. Lall, “Local polynomial-based Journal of Hydrometeorology, 16 (4) : 1561–1574, 2015. Retrieved from http://
flood frequency estimator for mixed population,” Journal of Hydrologic journals.ametsoc.org/doi/abs/10.1175/JHM-D-14-0227.1.
Engineering, 15 (9): 680–691, 2010, http://doi.org/10.1061/(ASCE)HE. Efron, B., The Jackknife, the Bootstrap and Other Resampling Plans, CBMS-
1943-5584.0000242. NSF Series. Philadelphia: SIAM 1982, http://doi.org/10.1137/1.9781611970319.
Araghinejad, S., D. H. Burn, and M. Karamouz, “Long-lead probabilistic Efron, B., “The bootstrap and modern statistics,” Journal of the American
forecasting of streamflow using ocean-atmospheric and hydrological predic- Statistical Association, 95 (452): 1293–1296, 2000, http://doi.org/10.1080/
tors,” Water Resources Research, 42 (3): 2006, http://doi.org/10.1029/ 01621459.2000.10474333.
2004WR003853. Efron, B. and R. J. Tibshirani, An Introduction to the Bootstrap, Refrigeration
Asefa, T., M. Kemblowski, U. Lall, and G. Urroz, “Support vector machines And Air Conditioning, 1993, Vol. 57, http://doi.org/10.1111/1467-9639.00050.
for nonlinear state space reconstruction: application to the Great Salt Lake Eubank, R., Spline Smoothing and Nonparametric Regression, Marcel-
time series,” Water Resources Research, 41 (12): 2005, http://doi. Dekker, New York, 1988.
org/10.1029/2004WR003785. Fasano, G. and A. Franceschini, “A multidimensional version of the
Asefa, T., M. Kemblowski, M. McKee, and A. Khalil, “Multi-time scale Kolmogorov-Smirnov test,” Monthly Notices of the Royal Astronomical Society,
stream flow predictions: the support vector machines approach,” Journal of 225 (1): 155–170, 1987, http://doi.org/10.1093/mnras/225.1.155.
Hydrology, 318 (1–4): 7–16, 2006, http://doi.org/10.1016/j.jhydrol.2005.06.001. Faucher, D., P. F. Rasmussen, and B. Bobée, “A distribution function based
Bardsley, W. E., “Using historical data in nonparametric flood estimation,” bandwidth selection method for kernel quantile estimation,” Journal of
Journal of Hydrology, 108: 249–255, 1989, http://doi.org/10.1016/0022- Hydrology, 250 (1–4): 1–11, 2001, http://doi.org/10.1016/S0022-
1694(89)90286-2. 1694(01)00359-6.
Bosworth, K. W. and U. Lall, “LOWLAD: a locally weightedL 1 smoothing Friedman, J. H., “Multivariate adaptive regression splines,” The Annals of
spline algorithm with cross validated choice of smoothing parameters,” Statistics, 19 (1): 1–67, 1991, retrieved from http://www.jstor.org/
Numerical Algorithms, 9 (1): 85–106, 1995. stable/2241837.
Bracken, C., B. Rajagopalan, and J. Prairie, “A multisite seasonal ensemble Gershman, S. J. and D. M. Blei, “A tutorial on Bayesian nonparametric
streamflow forecasting technique,” Water Resources Research, 46 (3): 2010, models,” Journal of Mathematical Psychology, 56:1–12, 2012. http://doi.
http://doi.org/10.1029/2009WR007965. org/10.1016/j.jmp.2011.08.004.
Breiman, L., “Bagging predictors,” Machine Learning, 24 (2): 123–140, Ghosh, S. and P. P. Mujumdar, “Nonparametric methods for modeling
1996, http://doi.org/10.1007/BF00058655. GCM and scenario uncertainty in drought assessment,” Water Resources
Burton, A., H. J. Fowler, C. G.Kilsby, and P. E. O’Connell, “A stochastic Research, 43 (7): 2007 http://doi.org/10.1029/2006WR005351.
model for the spatial-temporal simulation of nonhomogeneous rainfall Grantz, K., B. Rajagopalan, M. Clark, and E. Zagona, “A technique for
occurrence and amounts,” Water Resources Research, 46 (11): 2010, http://doi. incorporating large-scale climate information in basin-scale ensemble
org/10.1029/2009WR008884. streamflow forecasts,” Water Resources Research, 41 (10): 2005, http://doi.
Chernozhukov, V., “Extremal quantile regression,” Annals of Statistics, org/10.1029/2004WR003467.
33 (2): 806–839, 2005, http://doi.org/10.1214/009053604000001165. Gu, C., Smoothing Spline ANOVA Models, Springer Science & Business
Cioffi, F., U. Lall, E. Rus, and C. K. B. Krishnamurthy, “Space-time structure Media, New York, 2013. Retrieved from https://books.google.com/books?hl=
of extreme precipitation in Europe over the last century,” International Journal en&lr=&id=5VxGAAAAQBAJ&pgis=1.
of Climatology, 35 (8): 1749–1760, 2015, http://doi.org/10.1002/joc.4116. Hamed, K. H. and A. R. Rao, “A modified Mann-Kendall trend test for
Cleveland, W. S. and S. J. Devlin, “Locally weighted regression: an approach autocorrelated data,” Journal of Hydrology, 204 (1–4): 182–196, 1998, http://
to regression analysis by local fitting,” Journal of the American Statistical doi.org/10.1016/S0022-1694(97)00125-X.
Association, 83 (403): 596–610, 1988, http://doi.org/10.2307/2289282. Harrold, T. I., A. Sharma, and S. J. Sheather, “A nonparametric model for
Conover, W. J., “Practical Nonparametric Statistics,” The Statistician, Vol. 22, stochastic generation of daily rainfall occurrence,” Water Resources Research,
1999, http://doi.org/10.2307/2986830. 39 (10): 2003, http://doi.org/10.1029/2003WR002182.
Coulibaly, P. and C. K. Baldwin, “Nonstationary hydrological time series Hartigan, J. A., “Direct Clustering of a Data Matrix,” Journal of the
forecasting using nonlinear dynamic methods,” Journal of Hydrology, 307 (1–4): American Statistical Association, 2012. Retrieved from http://amstat.tandfon-
164–174, 2005, http://doi.org/10.1016/j.jhydrol.2004.10.008. line.com/doi/abs/10.1080/01621459.1972.10481214.
Coulibaly, P., B. Bobée, F. Anctil, and R. Aravena, “Artificial neural network Haykin, S., Neural Networks and Learning Machines, Pearson Prentice Hall,
modeling of water table depth fluctuations,” Water Resources Research, 37 (4): New Jersey, Vol. 3, 2008. Retrieved from http://www.amazon.com/
885–896, 2001, retrieved from http://bases.bireme.br/cgibin/wxislind.exe/ dp/0131471392.
25_Singh_ch25_p25.1-25.8.indd 5 8/22/16 1:41 PM

Haylock, M. R., N. Hofstra, A. M. G. Klein Tank, E. J. Klok, P. D. Jones, and Lall, U. and K. W. Bosworth, “Multivariate Kernel estimation of functions
M. New, “A European daily high-resolution gridded data set of surface tem- of space and time hydrologic data,” Stochastic and Statistical Methods in
perature and precipitation for 1950–2006,” Journal of Geophysical Research, Hydrology and Environmental Engineering, edited by K. W. Hipel, A. I.
113 (D20): D20119, 2008, http://doi.org/10.1029/2008JD010201. McLeod, U. S. Panu, and V. P. Singh, Vol. 10/3, Springer, Dordrecht, The
Heikkinen, J. and E. Arjas, “Non-parametric Bayesian estimation of a spa- Netherlands, pp. 301–315. http://doi.org/10.1007/978-94-017-3083-9_22.
tial Poisson intensity,” Scandinavian Journal of Statistics, 25(3): 435–450, 1998, Lall, U., N. Devineni, and Y. Kaheil, “An empirical, nonparametric simula-
http://doi.org/10.1111/1467-9469.00114. tor for multivariate random variables with differing marginal densities and
Helsel, D. R. and R. M. Hirsch, Statistical Methods in Water Resources, nonlinear dependence with hydroclimatic applications,” Risk Analysis : An
Elsevier, New York, 1992. Retrieved from https://books.google.com/books?hl Official Publication of the Society for Risk Analysis, 2015, http://doi.
=en&lr=&id=jao4o5X1pvgC&pgis=1. org/10.1111/risa.12432.
Hirsch, R. M., J. R. Slack, and R. A. Smith, “Techniques of trend analysis for Lall, U., Y-I. Moon, H-H. Kwon, and K. Bosworth, “Locally weighted poly-
monthly water quality data,” Water Resources Research, 18 (1): 107–121, 1982, nomial regression: parameter choice and application to forecasts of the Great
http://doi.org/10.1029/WR018i001p00107. Salt Lake,” Water Resources Research, 42 (5): 2006, http://doi.
Hofierka, J., J. Parajka, H. Mitasova, and L. Mitas, “Multivariate interpola- org/10.1029/2004WR003782.
tion of precipitation using regularized spline with tension,” Transactions in Lall, U., T. Sangoyomi, and H. D. I. Abarbanel, “Nonlinear dynamics of the
GIS, 6 (2): 135–150, 2002, http://doi.org/10.1111/1467-9671.00101. Great Salt Lake: nonparametric short-term forecasting,” Water Resources
Hollander, M., D. A. Wolfe, and E. Chicken, Nonparametric Statistical Research, 32 (4): 975–985, 1996, http://doi.org/10.1029/95WR03402.
Methods, John Wiley & Sons, New York, 2013, retrieved from https://books. Lall, U. and A. Sharma, “A nearest neighbor bootstrap for resampling
google.com/books?id=-V7jAQAAQBAJ&pgis=1. hydrologic time series,” Water Resources Research, 32 (3): 679–693, 1996.
Hutchinson, M. F., “Interpolating mean rainfall using thin plate smoothing Lee, T. and J. Salas, “Copula-based stochastic simulation of hydrological
splines,” International Journal of Geographical Information Systems, 9 (4): data applied to Nile River flows,” Hydrology Research, 42 (4): 318–330, 2011.
385–403, 1995, http://doi.org/10.1080/02693799508902045. Retrieved from http://www.iwaponline.com/nh/042/nh0420318.htm.
Jain, S. and U. Lall, “Magnitude and timing of annual maximum floods: Li, P-H., H-H. Kwon, L. Sun, U. Lall, and J-J. Kao, “A modified support
trends and large-scale climatic associations for the Blacksmith Fork River, vector machine based prediction model on streamflow at the Shihmen
Utah,” Water Resources Research, 36 (12): 3641–3651, 2000. Reservoir Taiwan,” International Journal of Climatology, 30 (8): 1256–1268,
Jain, S. and U. Lall, “Floods in a changing climate: Does the past represent 2009, http://doi.org/10.1002/joc.1954.
the future?,” Water Resources Research, 37 (12): 3193–3205, 2001. Loader, C. (2006). Local regression and likelihood. Springer Science &
Kalra, A. and S. Ahmad, “Using oceanic-atmospheric oscillations for long Business Media.
lead time streamflow forecasting,” Water Resources Research, 45 (3): 2009, López, J. and F. Francés, “Non-stationary flood frequency analysis in con-
http://doi.org/10.1029/2008WR006855. tinental Spanish rivers, using climate and reservoir indices as external covari-
Kalra, A., W. P. Miller, K.W. Lamb, S. Ahmad, and T. Piechota, “Using large- ates,” Hydrology and Earth System Sciences Discussions, 10 (3): 3103–3142,
scale climatic patterns for improving long lead time streamflow forecasts for 2013, http://doi.org/10.5194/hessd-10-3103-2013.
Gunnison and San Juan River Basins,” Hydrological Processes, 27 (11): 1543–1559, Mack, Y. P. and M. Rosenblatt, “Multivariate k-nearest neighbor density
2013, http://doi.org/10.1002/hyp.9236. estimates,” Journal of Multivariate Analysis, 9 (1): 1–15, 1979, http://doi.
Karlsson, M. and S. Yakowitz, “Nearest-neighbor methods for nonparamet- org/10.1016/0047-259X(79)90065-4.
ric rainfall-runoff forecasting,” Water Resources Research, 23 (7): 1300–1308, Mann, H. B. and D. R. Whitney, “On a test of whether one of two random
1987, http://doi.org/10.1029/WR023i007p01300. variables is stochastically larger than the other,” The Annals of Mathematical
Kaufman, L. and P. J. Rousseeuw, Finding Groups in Data: An Introduction Statistics, 18 (1): 50–60, 1947, http://doi.org/10.1214/aoms/1177730491.
to Cluster Analysis, John Wiley & Sons, New York, 2009. Retrieved from Maritz, J. S., Distribution-Free Statistical Methods, 2nd ed., CRC Press, Boca
https://books.google.com/books?hl=en&lr=&id=YeFQHiikNo0C&pgis=1. raton, 1995. Retrieved from https://books.google.com/books?hl=en&lr=&id=
Kember, G., A. C. Flower, and J. Holubeshen, “Forecasting river flow using 1Nrzy56BGPwC&pgis=1.
nonlinear dynamics,” Stochastic Hydrology and Hydraulics, 7 (3): 205–212, Mehrotra, R. and A. Sharma, “A nonparametric nonhomogeneous hidden
1993, http://doi.org/10.1007/BF01585599. Markov model for downscaling of multisite daily rainfall occurrences,”
Kendall, M. G., “A new measure of rank correlation,” Biometrika, 30 (1–2): Journal of Geophysical Research, 110 (D16): D16108, 2005, http://doi.
81–93, 1938, http://doi.org/10.1093/biomet/30.1-2.81. org/10.1029/2004JD005677.
Khalil, A. F., H. H. Kwon, U. Lall, and Y. H. Kaheil, “Predictive downscaling Mehrotra, R. and A. Sharma, “A nonparametric stochastic downscaling
based on non-homogeneous hidden Markov models,” Hydrological Sciences framework for daily rainfall at multiple locations,” Journal of Geophysical
Journal, 55 (3): 333–350, 2010, http://doi.org/10.1080/02626661003780342. Research, 111 (D15): D15101, 2006a, http://doi.org/10.1029/2005JD006637.
Khan, M. S. and P. Coulibaly, “Application of support vector machine in Mehrotra, R. and A. Sharma, “Conditional resampling of hydrologic time
lake water level prediction,” Journal of Hydrologic Engineering, 11 (3): 199–205, series using multiple predictor variables: A K-nearest neighbour approach,”
2006, http://doi.org/10.1061/(ASCE)1084-0699(2006)11:3(199). Advances in Water Resources, 29 (7): 987–999, 2006b, http://doi.org/10.1016/j.
Kisi, O. and M. Cimen, “A wavelet-support vector machine conjunction advwatres.2005.08.007.
model for monthly streamflow forecasting,” Journal of Hydrology, 399 (1–2): Mehrotra, R. and A. Sharma, “A semi-parametric model for stochastic gen-
132–140, 2011, http://doi.org/10.1016/j.jhydrol.2010.12.041. eration of multi-site daily rainfall exhibiting low-frequency variability,” Journal
Koenker, R., P. Ng, and S. Portnoy, “Quantile smoothing splines,” of Hydrology, 335 (1–2): 180–193, 2007, http://doi.org/10.1016/j.jhydrol.2006.
Biometrika, 81 (4): 673–680, 1994, http://doi.org/10.1093/biomet/81.4.673. 11.011.
Kooperberg, C. and C. J. Stone, “A study of logspline density estimation,” Minnotte, M. C., “Nonparametric testing of the existence of modes,” Annals
Computational Statistics & Data Analysis, 12 (3): 327–347, 1991, http://doi. of Statistics, 25 (4): 1646–1660, 1997, http://doi.org/10.1214/aos/1031594735.
org/10.1016/0167-9473(91)90115-I. Moon, Y. I., B. Rajagopalan, and U. Lall, “Estimation of mutual information
Krishnamurthy, C. K. B., U. Lall, and H. H. Kwon, “Changing frequency using kernel density estimators,” Physical Review E, 52 (3): 2318–2321, 1995.
and intensity of rainfall extremes over India from 1951 to 2003,” Journal of Moon, Y-I., U. Lall, and K. Bosworth, “A comparison of tail probability
Climate, 22 (18): 4737–4746, 2009. estimators for flood frequency analysis,” Journal of Hydrology, 151 (2–4):
Kunsch, H. R., “The Jackknife and the bootstrap for general stationary 343–363, 1993, http://doi.org/10.1016/0022-1694(93)90242-2.
observations,” The Annals of Statistics, 17 (3): 1217–1241, 1989, http://doi. Moon, Y-I., U. Lall, and H-H. Kwon, “Non-parametric short-term forecasts
org/10.1214/aos/1176347265. of the Great Salt Lake using atmospheric indices,” International Journal of
Kwon, H-H., U. Lall, and A. F. Khalil, “Stochastic simulation model for Climatology, 28 (3): 361–370, 2007, http://doi.org/10.1002/joc.1533.
nonstationary time series using an autoregressive wavelet decomposition: Msiza, I. S., F. V. Nelwamondo, and T. Marwala, “Water demand prediction
applications to rainfall and temperature,” Water Resources Research, 43 (5): using artificial neural networks and support vector regression,” Journal of
2007, http://doi.org/10.1029/2006WR005258. Computers, 3 (11): 1–8, 2008, http://doi.org/10.4304/jcp.3.11.1-8.
Kwon, H-H., U. Lall, and J. Obeysekera, “Simulation of daily rainfall sce- Müller, P. and F. A. Quintana, “Nonparametric Bayesian data analysis,”
narios with interannual and multidecadal climate cycles for South Florida,” Statistical Science, 19 (1): 95–110, 2004, http://doi.org/10.1214/
Stochastic Environmental Research and Risk Assessment, 23 (7), 879–896, 088342304000000017.
http://doi.org/10.1007/s00477-008-0270-2. Neal, R. M., “Regression and classification using Gaussian process priors,”
Lall, U., “Recent advances in nonparametric function estimation: hydraulic Bayesian Statistics, 6: 475–501, 1998.
applications (95RG00343),” US National Report to International Union of Noreen, E. W., Computer-Intensive Methods for Testing Hypotheses: An
Geodesy and Geophysics 1991–1994, 1: 1093, 1995. Introduction, Wiley-Interscience, New York, 1989.
25_Singh_ch25_p25.1-25.8.indd 6 8/22/16 1:41 PM

REFERENCES 25-7
Nowak, K., J. Prairie, B. Rajagopalan, and U. Lall, “A nonparametric sto- Sharma, A. and R. O’Neill, “A nonparametric approach for representing
chastic approach for multisite disaggregation of annual to daily streamflow,” interannual dependence in monthly streamflow sequences,” Water Resources
Water Resources Research, 46 (8): 2010, http://doi.org/10.1029/2009WR008530. Research, 38 (7): 5-1–5-10, 2002, http://doi.org/10.1029/2001WR000953.
Oh, H. S. and T. C. M. Lee, “Hybrid local polynomial wavelet shrinkage: Sharma, A., D. G. Tarboton, and U. Lall, “Streamflow simulation: a non-
wavelet regression with automatic boundary adjustment,” Computational parametric approach,” Water Resources Research, 33 (2): 291–308, 1997.
Statistics and Data Analysis, 48 (4): 809–819, 2005, http://doi.org/10.1016/j. Sharma, A., D. G. Tarboton, and U. Lall, “Streamflow simulation: a non-
csda.2004.04.002. parametric approach,” Water Resources Research, 33 (2): 291–308, 1997,
Porporato, A. and L. Ridolfi, “Multivariate nonlinear prediction of river http://doi.org/10.1029/96WR02839.
flows,” Journal of Hydrology, 248 (1–4): 109–122, 2001, http://doi.org/10.1016/ Shi, J. Q. and B. Wang, “Curve prediction and clustering with mixtures of
S0022-1694(01)00395-X. Gaussian process functional regression models,” Statistics and Computing, 18
Prairie, J., K. Nowak, B. Rajagopalan, U. Lall, and T. Fulp, “A stochastic (3): 267–283, 2008, http://doi.org/10.1007/s11222-008-9055-1.
nonparametric approach for streamflow generation combining observational Silverman, B. W., Density Estimation for Statistics and Data Analysis,
and paleoreconstructed data,” Water Resources Research, 44 (6): 2008, http:// Chapman and Hall, London, 1986. Retrieved from http://www.loc.gov/catdir/
doi.org/10.1029/2007WR006684. enhancements/fy0738/85021347-d.html.
Prairie, J. R., B. Rajagopalan, T. J. Fulp, and E. A. Zagona, “Modified K-NN Smirnov, N., “Table for estimating the goodness of fit of empirical distribu-
model for stochastic streamflow simulation,” Journal of Hydrologic Engineering, tions,” The Annals of Mathematical Statistics, 19 (2): 279–281, 1948, http://doi.
11 (4): 371–378, 2006, http://doi.org/10.1061/(ASCE)1084-0699(2006)11:4(371). org/10.1214/aoms/1177730256.
Prairie, J., B. Rajagopalan, U. Lall, and T. Fulp, “A stochastic nonparametric Smith, J. A., “Long-range streamflow forcasting using nonparametric
technique for space-time disaggregation of streamflows,” Water Resources regression,” Journal of the American Water Resources Association, 27 (1):
Research, 43 (3): 2007, http://doi.org/10.1029/2005WR004721. 39–46, 1991, http://doi.org/10.1111/j.1752-1688.1991.tb03111.x
Quintana, F. A. and P. Müller, “Nonparametric Bayesian data analysis,” Smith, J. A., G. N. Day, and M. D. Kane, “Nonparametric framework for
Statistical Science, 19 (1): 95–110, 2004, http://doi.org/10.1214/ long‐range streamflow forecasting,” Journal of Water Resources Planning and
088342304000000017. Management, 118 (1): 82–92, 1992, http://doi.org/10.1061/(ASCE)0733-
Rajagopalan, B., U. Lall, and D. G. Tarboton, “Simulation of daily precipita- 9496(1992)118:1(82).
tion from a nonparametric renewal model,” Utah State University, CO, 1993. Smola, A. J. and B. Schölkopf, “A tutorial on support vector regression,”
Rajagopalan, B., U. Lall, and D. G. Tarboton, “Nonhomogeneous Markov Statistics and Computing, 14 (3): 199–222, 2004, http://doi.
model for daily precipitation,” Journal of Hydrologic Engineering, 1 (1): 33–39, org/10.1023/B:STCO.0000035301.49549.88.
1996. Souza Filho, F. A. and U. Lall, “Seasonal to interannual ensemble stream-
Rajagopalan, B., U. Lall, D. G. Tarboton, and D. S. Bowles, “Multivariate flow forecasts for Ceara, Brazil: applications of a multivariate, semiparametric
nonparametric resampling scheme for generation of daily weather variables,” algorithm,” Water Resources Research, 39 (11): 2003, http://doi.
Stochastic Hydrology and Hydraulics, 11 (1): 65–93, 1997, http://doi. org/10.1029/2002WR001373.
org/10.1007/BF02428426. Souza Filho, F. A. and U. Lall, “Seasonal to interannual ensemble stream-
Rajagopalan, B. and U. Lall, “Locally weighted polynomial estimation of flow forecasts for Ceara, Brazil: applications of a multivariate, semiparametric
spatial precipitation,” Journal of Geographic Information and Decision Analysis, algorithm,” Water Resources Research, 39 (11): 1307, 2003.
2 (2): 44–51, 1998. Srinivas, V. V. and K. Srinivasan, “Hybrid moving block bootstrap for sto-
Regonda, S., B. Rajagopalan, U. Lall, M. Clark, and Y-I. Moon, “Local poly- chastic simulation of multi-site multi-season streamflows,” Journal of
nomial method for ensemble forecast of time series,” Nonlinear Processes in Hydrology, 302 (1–4): 307–330, 2005, http://doi.org/10.1016/j.jhydrol
Geophysics, 12 (3): 397–406, 2005, http://doi.org/10.5194/npg-12-397-2005. .2004.07.011.
Rousseeuw, P. J., “Silhouettes: a graphical aid to the interpretation and vali- Sun, N-Z., “Inverse Problems in Groundwater Modeling,” Springer Science
dation of cluster analysis,” Journal of Computational and Applied Mathematics, & Business Media, New York, 2013, retrieved from https://books.google.com/
20: 53–65, 1987, http://doi.org/10.1016/0377-0427(87)90125-7. books?hl=en&lr=&id=R7znCAAAQBAJ&pgis=1.
Ruppert, D., “Nonparametric regression and spline smoothing,” Journal Sun, X., U. Lall, B. Merz, and N. V. Dung, “Hierarchical Bayesian clustering
of the American Statistical Association, 2001, http://doi.org/10.1198/ for nonstationary flood frequency analysis: application to trends of annual
016214501753382417. maximum flow in Germany,” Water Resources Research, 2015, http://doi.
Sadri, S. and D. H. Burn, “Nonparametric methods for drought severity org/10.1002/2015WR017117.
estimation at ungauged sites,” Water Resources Research, 48 (12): 2012, http:// Suykens, J. A. K. and J. Vandewalle, “Least squares support vector machine
doi.org/10.1029/2011WR011323. classifiers,” Neural Processing Letters, 9 (3): 293–300, 1999, http://doi.
Sankarasubramanian, A. and U. Lall, “Flood quantiles in a changing cli- org/10.1023/A:1018628609742.
mate: seasonal forecasts and causal relations,” Water Resources Research, 39 Teh, Y. W. and M. I. Jordan, “Hierarchical Bayesian nonparametric models
(5): 1134, 2003. with applications,” Bayesian Nonparametrics, 2009, pp. 1–48, http://doi.
Santhosh, D. and V. V. Srinivas, “Bivariate frequency analysis of floods org/10.1017/CBO9780511802478.
using a diffusion based kernel density estimator,” Water Resources Research, Teh, Y. W. and M. I. Jordan, “Hierarchical Bayesian nonparametric models
49 (12): 8328–8343, 2013, http://doi.org/10.1002/2011WR010777. with applications,” Bayesian Nonparametrics: Principles and Practice, 2010,
Schuster, E. and S. Yakowitz, “Parametric/nonparametric mixture density pp. 158–207.
estimation with application to flood-frequency analysis,” Journal of the Tokdar, S. T. and R. E. Kass, “Importance sampling: a review,” Wiley
American Water Resources Association, 21 (5): 797–804, 1985, http://doi. Interdisciplinary Reviews: Computational Statistics, 2010, http://doi.
org/10.1111/j.1752-1688.1985.tb00173.x. org/10.1002/wics.56.
Scott, D. W., Multivariate Density Estimation: Theory, Practice, and Torrence, C. and G. P. Compo, “A practical guide to wavelet analysis,”
Visualization, Wiley, New York, 2015. Retrieved from https://books.google. Bulletin of the American Meteorological Society, 79 (1): 61–78, 1998, http://doi.
com/books?hl=en&lr=&id=XZ03BwAAQBAJ&pgis=1. org/10.1175/1520-0477(1998)079<0061:APGTWA>2.0.CO;2.
Sharma, A., “Seasonal to interannual rainfall probabilistic forecasts for Vapnik, V., The Nature of Statistical Learning Theory, Springer Science &
improved water supply management: Part 3—A nonparametric probabilistic Business Media, New York, 2013, retrieved from https://books.google.com/
forecast model,” Journal of Hydrology, 239 (1–4): 249–258, 2000, http://doi. books?hl=en&lr=&id=EqgACAAAQBAJ&pgis=1.
org/10.1016/S0022-1694(00)00348-6. Vogel, R. M. and A. L. Shallcross, “The moving block bootstrap versus
Sharma, A. and U. Lall, “A nonparametric approach for daily rainfall simu- parametric time series models,” Water Resources Research, 32 (6): 1875–1882,
lation,” Mathematics and Computers in Simulation, 48 (4–6): 361–371, 1999, 1996, retrieved from http://cat.inist.fr/?aModele=afficheN&cpsidt=
http://doi.org/10.1016/S0378-4754(99)00016-6. 10945420.
Sharma, A., K. C. Luk, I. Cordery, and U. Lall, “A comparison of model Wahba, G., “Spline models for observational data,” SIAM, 1990. Retrieved from
predictor identification strategies for long-term rainfall forecasting,” https://books.google.com/books?hl=en&lr=&id=BS6fGsebFNkC&pgis=1.
Hydrologic Modeling: Proceedings of the International Conference on Water, Wand, M. P., “Local regression and likelihood,” Journal of the American
Environment, Ecology, Socio-Economics, and Health Engineering (WEESHE): Statistical Association, 96 (453): 339–355, 2001, http://doi.org/10.1198/
October 18–21, 1999, Seoul National University, Seoul, Korea, 1999, p. 50. jasa.2001.s373.
Sharma, A., K. C. Luk, I. Cordery, and U. Lall, “Seasonal to interannual Wei, C-C., “Wavelet kernel support vector machines forecasting tech-
rainfall probabilistic forecasts for improved water supply management: niques: case study on water-level predictions during typhoons,” Expert
Part 2—Predictor identification of quarterly rainfall using ocean-atmosphere Systems with Applications, 39 (5): 5189–5199, 2012, http://doi.org/10.1016/j.
information,” Journal of Hydrology, 239 (1): 240–248, 2000. eswa.2011.11.020.
25_Singh_ch25_p25.1-25.8.indd 7 8/22/16 1:41 PM

Westmacott, J. R. and D. H. Burn, “Climate change effects on the hydro- Yoo, J., H-H. Kwon, T-W. Kim, and J-H. Ahn, “Drought frequency analysis
logic regime within the Churchill-Nelson River Basin,” Journal of Hydrology, using cluster analysis and bivariate probability distribution,” Journal of
202 (1–4): 263–279, 1997, http://doi.org/10.1016/S0022-1694(97)00073-5. Hydrology, 420–421: 102–111, 2012, http://doi.org/10.1016/j.jhy-
Yakowitz, S. J., “Nonparametric density estimation, prediction, and regres- drol.2011.11.046.
sion for Markov sequences,” Journal of the American Statistical Association, 80: Yu, K., Z. Lu, and J. Stander, “Quantile regression: applications and current
215–221, 1989. Retrieved from http://amstat.tandfonline.com/doi/abs/10.108 research areas,” Journal of the Royal Statistical Society: Series D (The
0/01621459.1985.10477164. Statistician), 52 (3): 331–350, 2003, http://doi.org/10.1111/1467-9884.00363.
Yakowitz, S. J. and F. Szidarovszky, “A comparison of kriging with nonpara- Yue, S., P. Pilon, and B. Phinney, “Canadian streamflow trend detection:
metric regression methods,” Journal of Multivariate Analysis, 16 (1): 21–53, impacts of serial and cross-correlation,” Hydrological Sciences Journal, 48 (1):
1985, http://doi.org/10.1016/0047-259X(85)90050-8. 51–63, 2003, http://doi.org/10.1623/hysj.48.1.51.43478.
25_Singh_ch25_p25.1-25.8.indd 8 8/22/16 1:41 PM

Chapter 26
Predictive Uncertainty
Assessment and
Decision Making
BY
EZIO TODINI
ABSTRACT 26.2 FORECASTING IN HYDROLOGY
This chapter introduces the problem of decision-making under uncertainty Prediction in hydrology has been widely used as a planning tool since the ‘50s
and how this uncertainty can be reduced using information provided by pre- while, from the ‘70s, forecasting has become an operational support for real-
dictive models. After defining prediction and forecasting, the concept of time operational reservoir management as well as flood alert and flood risk
“predictive” uncertainty is introduced highlighting its differences with “simu- reduction.
lation” uncertainty. Starting from the late ‘70s and early ‘80s a wide number of real-time flood
It also describes the more commonly used techniques for assessing the forecasting systems were developed and installed all over the world, but until
predictive densities and the reasons that have undermined until recently their the late ‘90s, predictions and/or forecasts where essentially “deterministic.”
operational use. A number of open issues are also discussed such as account- Hydrological and hydraulic models were used to determine either design or
ing for heteroscedasticity in the derivation of the predictive densities, the possible future discharges and water levels. Model output values were implic-
ways of accounting for parameter uncertainty as well as for input uncertainty, itly assumed as “real” quantities and directly used for design (as for instance
usually introduced in the form of ensembles. the well-known “design discharge”) or, in the case of real-time forecasting, as
The chapter concludes by discussing via a number of examples, the appro- the future “real” water level to be taken as the basis for operational decisions,
priate use of predictive uncertainty in decision-making and the problem of such as issuing an alert or opening a sluice.
correctly communicating to the decision makers and stakeholders the At the same time meteorologists from the coarse meso-scale models started
assessed uncertainty measures in the form of best available knowledge. developing limited area models which improved rainfall forecast performed
better than the simplistic “inertial” model (tomorrow’s weather will equal
26.1 INTRODUCTION today’s). The large uncertainty present in these forecasts induced meteorolo-
gists to introduce, in the ‘90s, the concept of “ensemble forecasts” aimed at
As can be found in Wikipedia (http://en.wikipedia.org/wiki/Prediction), providing a measure of the so-called “predictive uncertainty” of meteorologi-
“a prediction (Latin præ-, “before,” and dicere, “to say”) … is a statement cal forecasts.
about the way things will happen in the future, often, but not necessarily, Also, hydrologists started considering the problem of uncertainty in the
based on experience or knowledge. While there is much overlap between forecasts along two lines of thought. The first one, following meteorologists,
prediction and forecast, a prediction may be a statement that some outcome was based on the “ensembles,” leading recently to the establishment of the
is expected, while a forecast is more specific,” based on appropriate computa- worldwide research initiative Hydrological Ensemble Prediction Experiment
tions “and may cover a range of possible outcomes.” (HEPEX—http://hepex.irstea.fr). The second one, following a more formal
A prediction is most often centered around the idea of an outcome based statistical analysis, led to the development of models postprocessors based on
on deductive logic or beliefs rather than the approach. a number of alternative approaches. Today, the concept of uncertainty is more
Forecasting (http://en.wikipedia.org/wiki/Forecasting) “is the process of and more considered in hydrological forecasting not only in research but also
making statements about events whose actual outcomes (typically) have not in practice (Cloke and Pappenberger, 2009; Pappenberger et al., 2013).
yet been observed, as for instance, the estimation of some variable of interest at Both lines of thought will be discussed in this chapter in the light of a
some specified future date. Prediction is a similar, but more general term. Both proper use of the measure of uncertainty aimed at improving the reliability
might refer to formal statistical methods employing time series, cross-sectional and robustness of decisions.
or longitudinal data, or alternatively to less formal judgmental methods.” A
forecast is a means to validating a prediction based on an analysis of varying
26.3 MOTIVATIONS FOR CONVERTING DETERMINISTIC
factors and patterns, showing you how you got to this future logic or belief. TO STOCHASTIC PREDICTION
Usage of these terms can differ between areas of application; for example,
in hydrology, the terms “forecast” and “forecasting” are sometimes reserved Models are simplified representations of reality usually developed to (1) allow
for estimates of values at certain specific future times, while the term “predic- finding rationale and coherence among several sources of error affected data;
tion” is used for more general estimates, such as the number of times floods (2) better understand the ongoing real processes; and finally (3) predict
will occur over a long period. beyond the range of observations. Being simplified, sometimes overly simpli-
The term “prediction” in hydrology is also commonly used instead of “hind- fied and coarse, representations of reality, it is evident that even using “per-
cast prediction,” when a model is used forced by known input, for instance fect” error-free data, it is unlikely that their output will coincide with the
when simulating forecasting using past data. observed values.
26-1
26_Singh_ch26_p26.1-26.16.indd 1 8/22/16 1:44 PM

26-2 Predictive Uncertainty Assessment and Decision Making
Moreover, errors unavoidably affect point measurements and become days, etc.). In this case one or more models (meteorological, hydrological,
larger when extrapolating at larger scales, as in the case of point raingauge hydraulic, etc.) can be used to obtain a forecast, which allows reestimating
precipitation measurements, which are usually extrapolated to be representa- a sharper predictive density as a measure of predictive uncertainty (PU).
tive of areal averages over several square kilometers.
Also, wrong assumptions on initial and boundary conditions, required by 26.4.1 Role of Predictive Uncertainty in
Decision-Making
the more physically oriented models (as for instance the state of the atmo-
sphere at the beginning of the forecasting run in meteorological models; the In order to fully understand and appreciate the role of PU in decision-making,
initial soil moisture content in catchment rainfall runoff models; the it is necessary to realize that what causes the damages (or the benefits) is the
upstream and lateral inflow in a river for which a routing model is being actual future occurrence of water level, discharge or water volume, and not its
developed, etc.), are generally unknown or not fully known. prediction generated by a forecasting model. For instance, damages will occur
Using the available uncertain information in the calibration phase, aimed when the actual water level yt (and certainly not if its predicted value yˆt t0 at an
at estimating model parameter values, all these errors, in combination with earlier time t0) will overtop the dyke level y ∗ (Todini, 2008). Therefore a utility/
the adopted estimation technique, generate uncertain (frequently also biased damage function at any future time t > t0 must be expressed as a function yt ,
and inefficient) parameter estimates, which in turn affect model perfor- the actual unknown level that will occur at time t:
mances as well as model-based forecasts.
 ∗
Recently another source of uncertainty has been recognized to strongly  0 ∀yt ≤ y
affect hydrological forecasts. Hydrological catchment models, calibrated U ( yt ) =  (26.1)
using observed precipitation, are then operationally run using highly uncer-  g ( yt − y ∗ ) ∀yt > y ∗
tain meteorological forecasts, thus increasing the level of indetermination.
In the past, decision makers were given “deterministic” predictions and where g (∗) represents a generic function relating the cost of damages and
forecasts without a measure of uncertainty. Nonetheless, decision makers losses to the future, albeit unknown water stage yt . Please note that the utility
have always been aware of the fact that the predicted values were not “perfect,” function U ( yt ) may not necessarily represent actual damages, but rather the
in the sense that they would not necessarily match the actual future occur- decision maker’s assessment of the situation, based on his risk propensity
rence, and they used their experience to partly compensate, in a subjective (Nicholson et al., 2005).
way, this nonperfect behavior. For instance, during the design phase, they According to Decision Theory (Berger, 1985; Bernardo and Smith, 1994;
would raise the dam or the dyke level to increase safety or try to analyze De Groot, 2004), decisions should be taken on the basis of the expected utility
worst-case scenarios. Nonetheless, specifically in the case of real time fore- E{U ( yt )}, which can be a combination of expected benefits and expected
casts, the predicted quantities were treated as “real” quantities to be compared losses. E{U ( yt )} can be estimated only if the probability density function of
to thresholds triggering decisions, leading to a large number of wrong deci- the conditioning variable yt is known. For instance, yt can be the water level
sions (see Sec. 26.8 for an example). to be compared to a threshold value above which losses are no more null.
This lack of robustness descends from the fact that, in the planning phase, Since both in planning and management yt is generally unknown, the
operating rules are derived as a function of assumed “known” quantities. For expected value of the utility function can be defined as:
instance an alert is issued if the “known” water level is above a threshold level. +∞
Therefore the triggering mechanism is conditional on the “knowledge” of the E{U ( yt )} = ∫ U ( yt ) f ( yt )dyt (26.2)
future level, which is not true in operational settings since the true future 0
water level is unknown; only its virtual representation, the model forecast, is
known. In other words, the future value is not known with certainty, as the where f ( yt ) is the probability density expressing our incomplete knowledge
operating rules imply. (in other words our uncertainty) on the future value that will occur. This
Unfortunately, for many years this problem was not set into the correct density, which can be estimated from historical data, is generally too broad
perspective. Ensemble predictions and uncertainty postprocessors were because it lacks the conditionality on the current events. To improve reliabil-
mostly developed to provide a “measure” of uncertainty to warn the decision ity of decisions, it is essential to update this historical probability distribution
makers not to accept the forecasts at face value. Decision makers interpreted function by more realistically combining prior knowledge to one or more
this more as a lack of predictive capability rather than as an extra piece of hydrological models forecasts summarizing all the available information (like
information forecasters originally intended to provide. the rain forecast, the catchment geomorphology, the state of the river at
The result was that flood and water resources managers mostly rejected the moment of the forecast, etc.). This allows to provide a more informative
density f ( yt yˆt(1)t0 , yˆt(2) ˆ (m )
t0 , … , yt t0 ) expressing the predictive uncertainty on the
these approaches, because they were now told that forecasts were uncertain,
and, at the same time, they were not provided with any specific tool allowing value of a future predictand yt given one or more model forecasts
them to make appropriate use of this information. yˆt(1t0) , yˆt(2t0) , … , yˆt(mt0 ) issued at time t0.
It was only at the turn of the last century that forecasting uncertainty was When dealing with forecasting, the interest lies in the assessment of PU for
more clearly defined (Krzysztofowicz, 1999) and viewed as an appropriate prediction times t larger than t0. Nonetheless, in planning mode the same
way of communicating the best possible information arising from one or concept can be extended to simulation mode when t ≤ t0.
more forecasting models. This information could then be integrated into the Summarizing, for a decision maker to take rational decisions, is therefore
decision-making process using the tools provided by decision theory (Berger, necessary first of all to define his risk propensity (e.g., the decision maker
1985; Bernardo and Smith, 1994; De Groot, 2004). could be risk-prone when the level of damage is low and risk-averse when the
level of damage is high), by choosing the appropriate utility function; then
reach a decision by minimizing the expected value of risk on the basis of a
26.4 PREDICTIVE UNCERTAINTY
predictive density function conditional upon all the information he/she can
According to Decision Theory, decisions, under uncertainty of future events, gather and in particular on the available model forecasts.
should be taken by trading off between what one has to pay to implement a
specific decision and the expected benefit. For instance, the cost for raising a 26.4.2 Definition of Predictive Uncertainty
dyke to prevent flooding, which is a real cost once the decision is taken, must In clarifying to hydrologists the meaning of predictive uncertainty,
be confronted with the “expected benefit,” namely the expected reduction in Krzysztofowicz (1999) points out that “Rational decision making (for flood
damages due to the change in the probability of flooding given the higher warning, navigation, or reservoir systems) requires that the total uncertainty
dyke. The need for computing the “expected benefit” descends from the fact about a hydrologic predictand (such as river stage, discharge, or runoff
that the benefit is not a “certain” quantity since it depends on the unknown volume) be quantified in terms of a probability distribution, conditional on
future occurrence. all available information and knowledge” and that “Hydrologic knowledge is
Therefore, in order to compute the expected benefit, it is necessary to assess typically embodied in a deterministic catchment model.”
the probability density of the future occurrence (usually called the predictive These statements underline two aspects usually not fully and clearly under-
density); this can be done by using all available information, and, whenever stood. The first is that the objective of forecasting is the description of the
available, model forecasts. uncertainty of the actual future values of water stage, discharge, runoff
For instance, one could use the “climatological” probability density, namely volume, etc., rather than the uncertainty of predictions generated by the
the one estimated from historical records, but this density is generally quite hydrological forecasting models. The second one is that this uncertainty,
“widespread,” in the sense that it allows for an ample spectrum of possibilities. generally expressed in terms of a probability density (or probability distribu-
Usually, in operation, better information is available, particularly when tion) function, is “conditional” upon the hydrological forecasting model pre-
the expected event is not too far ahead in time (1 hour, few hours, 1 or more diction, which is now seen, similarly to the measurement of an imprecise
26_Singh_ch26_p26.1-26.16.indd 2 8/22/16 1:44 PM

Predictive Uncertainty 26-3
instrument, as the available, although uncertain, knowledge of the future. In

other words, the forecasting model prediction is now a way to reduce uncer-
tainty in the decision-making process and not the provider of deterministic yt fyt||y^*t|to (yt|y^*t|to)
(and therefore “certain”) future levels, flows, etc.
PU can be defined as the uncertainty of a future realization of a predictand
(the future quantity of interest) remaining conditional on all available infor-
mation, which is usually, but not necessarily, embedded in one or more model
forecasts. Several measures of PU are available, such as probability densities,
fuzzy sets, green sets, etc. In hydrological applications, a measure of PU is
Observed values
usually provided in terms of the conditional probability density of a predic-
tand given one or more than one predictors (for instance one or more model
forecasts).
Please note that in the case of PU, the uncertain quantity is not the model
forecast, which is known at the time of the forecast, but rather the future
realization, namely a dynamic future water level, discharge, water volume,
etc., yet to be observed.
Unfortunately, PU is usually confused with simulation uncertainty (SU),
which is alternatively defined as the uncertainty of a model forecast given
the knowledge of what has actually occurred. Although in the past it was
referred to as “emulation” or “validation” uncertainty, the term “simulation”
was here preferred and introduced to characterize this type of uncertainty,
because SU can only be estimated until the present time when both input
and output observations are available. When dealing with real-time applica-
tions this implies running the models in simulation or in hindcast mode. SU y^*t|to yt|to
cannot be estimated in prediction mode because observations are not avail- Model forecast
able and what will happen is still unknown. As can be understood from this
definition, in the case of SU the model forecast is the uncertain quantity, Figure 26.2 Predictive uncertainty (PU). Once the joint probability density between
while the actual value of what has happened is known, which implies that observed and predicted values is built, knowing model prediction yˆt∗t0, one can assess
SU can only be used on past observations and not in predictive mode. On using historical data, the uncertainty on what has actually happened yt .
the contrary PU can be estimated both in simulation (or hindcast) and in
prediction mode.
To better understand the concepts and the definition of PU, following The first one:
Todini (2008) let us introduce the concept of the joint-probability distribution
of the real quantity of interest, the predictand (namely the discharge or the
( ) f yt∗ , yˆt t0 (
f yt∗ , yˆt t0 ) (26.3)
water level at a specific cross section, etc.) yt, and the corresponding model (
f yˆt t0 yt∗ = ) =
( )
∫ ( )dyˆ
+∞
prediction yˆt t0 . Note that, without loss of generality, only one model is f y , ˆ
y ∗
t t t0 t t0
f yt∗
0
considered here for the sake of clarity.
If one plots observed values yt over the corresponding predictions yˆt t0, is a measure of the simulation uncertainty, describing the spread of the fore-
unless yt = yˆt t0, a cloud of points will be observed (as in Figs. 26.1 and 26.2) casted values yˆt t0 around a specific observed value yt∗, as graphically repre-
which gives a visual impression of the spread as well as of the quality of the sented in Fig. 26.1.
prediction. A two-dimensional probability density function can be fitted to The second one:
these points, which will correspond to the joint probability density f ( yt , yˆt t0 ).
( )f yt , yˆt∗t0 (
f yt∗ , yˆt t0 ) (26.4)
Knowing the joint probability density, it is then possible to derive two condi-
tional densities as the joint divided by the marginal probability.
(
f yt yˆt∗t0 = ) ∫ f ( y , y ) dy
+∞
ˆ ∗
=
( )
f yˆt∗t0
0 t t t0 t
is a measure of the predictive uncertainty, describing the spread of the future

unknown values yt around a specific predicted value yˆt∗t0, as graphically rep-
f^y (y^t|to|yt*) resented in Fig. 26.2.
yt t|to|yt* If several predictive models are available, this definition of PU can be easily
extended. In this case, assessing predictive uncertainty, means finding the
following conditional probability density:
yt* (
f yt yˆt(1)t0 , yˆt(2) ˆ (m )
t0 ,…, yt t0 (26.5) )
Observed values
where yˆt(1)t0 , yˆt(2)

t0 , … , yˆt(mt0 )
correspond to the predictions of m different models
(with m ≥ 1) concurring at reducing the uncertainty on the future realization
of yt .
Quoting Todini (2011), evaluating SU is fundamental in order to revise and
improve a given model: the aim in this case is to reduce SU in order to
improve model structure, model parameterization, and parameter estimation.
Therefore, when dealing with SU, it is necessary to identify and describe all
the components of uncertainty due to (1) the model structure; (2) the param-
eter estimates; (3) the observation measurements; (4) the initial conditions;
and (5) the boundary conditions or to use a Bayesian approach, which implies
the derivation of a posterior density of parameters used to marginalize the
effect of the uncertainty of individual components.
Although statistically correct, this does not necessary apply when dealing
with PU, owing to the structural nature of models typically used in meteoro-
y^t|to logical, hydrological, and hydraulic forecasts. The reason is the following.
Model forecast PU may be assessed using incorrect items such as (1) a given model structure;
Figure 26.1 Simulation uncertainty (SU). Once the joint probability density between (2) a given set of parameter estimates; (3) a given set observation measure-
observed and predicted values is built, knowing what has happened, which implies that ments; (4) given initial conditions; and (5) given boundary conditions.
the actual value yt∗ has been observed, one can assess the lack of accuracy (uncertainty) Nonetheless, the predictive density derived from using those incorrect issues
of the model prediction yˆt t0. can be correctly used in prediction provided that there is full “consistency”
26_Singh_ch26_p26.1-26.16.indd 3 8/22/16 1:44 PM

between the model and input data used in assessment and the model and The conversion is performed by transforming the original data into their
input data used in prediction. probability and from the probability into the corresponding values from a
It is certainly true that estimating the posterior density of parameters and standard Normal (0,1) distribution.
then using it to marginalize the model, parameters, input and initial, and
boundary conditions uncertainty leads to a more dense PU measure, but, as The Parametric Approach
shown by Liu et al. (2005), the difference is minimal while the computational Regardless to the probability distribution of the original data, it is possible to
cost can be extremely heavy and not compatible with operational prediction convert nonGaussian distributed data into the Normal space by probability
and decision. matching.
As can be seen from the previous equations, the assessment of PU aims at If Fy ( y | ϑ y ) = Prob{ y ≤ y} (with ϑ y a set of parameters characterizing the
providing the decision maker with a measure of the probability of any future probability distribution to be estimated) represents the probability distribution
(real) value of the predictand conditional upon all the knowledge and informa- of y, the conversion of variable y into the Normal space is performed by match-
tion, available up to the present, he was able to acquire through a learning infer- ing Fy ( y |ϑ y ) to the probability of a variable h, which probability distribution
ential process (and usually condense in one or more than one model predictions). function N (η ) = Prob{η ≤ η}, is the Standard (0,1) Normal distribution.
In practice, after estimating the parameter values ϑ̂ from the observations,
26.4.3 Difference of Predictive Uncertainty from h is obtained as the value of a Standard Normal variable, with probability
Sensitivity Analysis
P = Fy ( y|ϑ y ), namely:
While PU aims at assessing the uncertainty of a future “real” occurrence of a
quantity of interest, albeit unknown at the time of forecast, sensitivity analysis
(SA) aims at assessing the uncertainty of model forecasts descending from
( ( )) (26.6)
η = N −1 ( P ) = N −1 Fy y ϑˆ y
−1
imperfect knowledge of inputs, parameters, initial and boundary conditions, with N the inverse of the Standard (0,1) Normal distribution.
etc., and is somehow related to SU. In order to apply the post processors in the Gaussian space, both vectors
SA is generally performed by randomly perturbing model inputs, parame- y (observations) and ŷ (predictions) are converted into the corresponding
ters, initial and boundary conditions, etc., in order to assess their effect on the Standard (0,1) Normal variables h and η̂ using the appropriate marginal dis-
model output. Similarly to the assessment of SU, the scope of SA is to analyze tributions Fy ( y | ϑˆ y ) and Fyˆ ( yˆ | ϑˆ yˆ ) after estimating the required parameter
in detail the effects of the different sources of uncertainty, on the model fore- sets ϑ̂ y and ϑ̂ ŷ .
casts possibly aimed at reducing them. Alternative parametric transformations into the Normal space, which do
While SA stops at the assessment of the sensitivity of model performances not require fitting probability distributions, can also be found in the litera-
as a function of the different sources of uncertainty, SU needs a step further, ture. For example, in order to make the observations and the predictions
namely the marginalization of all these uncertainties, to provide an expected look approximately Normal one can apply the Box-Cox transformation (Box
uncertainty measure. A formally correct assessment of SU requires: and Cox, 1964) or the more recent log-sinh transformation introduced by
1. Assuming a prior joint density of all the sources of uncertainty (inputs, Wang et al., (2012).
parameters, initial and boundary conditions, etc.) expressing the prior knowl-
edge. In this phase the different densities could be assumed as independent; The Nonparametric Approach
2. Define a Likelihood function linking the model forecasts to the Frequently a convenient nonparametric approach, the Normal Quantile
observations; Transform (NQT) (Van der Waerden, 1952, 1953a, b) is used (Kelly and
3. Assess the posterior joint density of all the sources of uncertainty Krzysztofowicz, 1997; Bogner et al., 2012), to convert non-Gaussian distrib-
through a Bayesian inferential process; uted data into the Normal space.
4. Compute the expected SU by marginalizing the effects of inputs, param- Given a sample of n observations sorted in ascending order, the probability
eters, initial and boundary conditions, etc., uncertainty on the basis of the density f {Fy ( y k n )} of the cumulative probability of occurrence Fy ( y k n ) of the
assessed posterior joint density. kth value y k n observed (with 1 ≤ k ≤ n the rank order of y k n within the sample),
With a Bayesian perspective, by assuming the parameters to be the receptors can be estimated, in nonparametric form according to the rank order regard-
of all the uncertainty, the above procedure could be simplified by estimating the less of the actual value of y.
posterior parameter probability density to be used in the marginalization. Based on theoretical considerations (Wilks, 1948), the probability density
Nonetheless, the assessment of the SU due to the single source of error, is valu- f {Fy ( y k n )} is a Beta(α , β ) with parameters α = k and β = n − k + 1, namely:
able: it could lead to a better understanding of the SU (and therefore also of the
PU, being model and predictions related through the joint probability density of n!
Eqs. 26.3 and 26.4), and can help to reduce it in a meaningful way, for example f {Fy ( y k n )} = Beta(k , n − k + 1) = Fyk−1 (1 − Fy )n−k (26.7)
(k − 1)!(n − k )!
by prioritizing improvements by reducing the different sources of errors.
Similarly, one could use SA as the basis for a formally correct assessment of whose expected value (Gupta and Nadarajah, 2004) corresponds to the clas-
the expected PU by marginalizing the effects of inputs, parameters, initial and sical “Weibull plotting position:”
boundary conditions, etc., uncertainty on the basis of the assessed posterior
joint density. But this formally correct procedure, requiring extremely large
computational efforts incompatible with real time forecasting, may not be
{ ( )} = n k+1 (26.8)
E Fy y k n
necessary due to the “structural” nature of the models used in hydrological The NQT approach uses this probability to create a one-to-one link with the
predictions, which induces high correlation among the resulting perturbed Standard Normal (0, 1) variate having the same probability. In other words:
predictions, with little additional information content (Liu et al., 2005).
 k 
ηk n = N −1   (26.9)
26.5 TECHNIQUES AIMED AT ASSESSING  n +1 
PREDICTIVE UNCERTAINTY
 k 
In the last decades, several techniques were developed to assess PU. Some of these where the symbol N −1   represents the value of a N (0, 1) distributed
 n+ 1 
techniques have gained a certain degree of confidence and are currently opera- k
tionally used. Although most of these techniques can directly handle, or can be variable h with probability P(η ≤ ηk n ) = .
n +1
extended, multilocation and multitime predictions, for the sake of clarity they will The two vectors yi and ŷi are first ranked in ascending order and respec-
be presented here for the case of single time and single location prediction. i
tively transformed into their expected rank probabilities E { P ( y ≤ yi )} =
n +1
26.5.1 Conversion into the Normal Space i
and E { P ( yˆ ≤ yî )} = according to their ranking order i ∈ 1, 2, …, n. The
A number of the PU assessment techniques described below such as the n +1
Bayesian Forecasting System and the Model Conditional Processor, for rea- NQT transformation is successively applied by computing two new variables
sons of mathematical tractability, require that observations and forecasts are ηi and η̂i as the standard Normal Distribution deviates corresponding to the
preventively converted into the Normal space, while other approaches, such i i
same probabilities P(η ≤ ηi ) = and P(ηˆ ≤ ηî ) = . In such a way, the
as the Quantile Regression (QR), the Bayesian Model Averaging (BMA), and n +1 n +1
the Multiple Regression approaches, frequently improve their performances. marginal distributions of the transformed variables ηi and η̂ i will correspond
Conversion into the Normal space is also a way to handle certain types of to Standard Normal N (0, 1) distributions. Note that the product moment cor-
heteroscedasticity that can be reduced or disappear in the transformed space. relation in time (Edwards, 1976) of each variable together with the product
26_Singh_ch26_p26.1-26.16.indd 4 8/22/16 1:44 PM

Techniques aimed at assessing predictive uncertainty 26-5
moment cross correlation among variables will be obviously lost in this non- Step 3: Defining a Likelihood Function
linear transformation. Nonetheless, the Spearman rank correlation (Lehmann In Krzysztofowicz and Kelly (2000), the Likelihood derivation is based upon
and D’Abrera, 1998), which can be viewed as an approximation of the previ- a regression between the Normal space image of the observed and modeled
ous ones, will be fully preserved. variables h and η̂ :
Although not usually acknowledged, all the approaches transforming the ηˆt0 +k∆t = akηt0 +k∆t + bkηt0 + ck + ξt0 +k∆t , where ξt0 +k∆t is a N (0, σ k2 ) independent
problem in the Normal space by means of the NQT, require additional process, with ak , bk , ck , and σ k2 estimated through a linear regression in the
models to represent the tails of the distributions for probability quantiles Normal space.
smaller than 1/(n + 1) or larger than n/(n + 1). In general this is not a complex This allows to derive the Likelihood function as:
problem, but some care should be placed in the tail models to avoid unjusti-
fied extremely large uncertainty bands for the largest observed values
(Coccia and Todini, 2011; Bogner et al., 2012).
( ) (
Γ k ηˆt0 +k∆t ηt0 +k∆t , ηt0 = N akηt0 +k∆t + bkηt0 + ck , σ k2 (26.11) )
26.5.2 Bayesian Forecasting System which can be written, in terms of the probability density, as:
Krzysztofowicz (1999) introduced the Hydrological Uncertainty Processor (
1 ηˆt0 + k∆t −akηt0 + k∆t +bkηt0 +ck )
2
−
(HUP), which together with the Input Uncertainty Processor (IUP), gave rise 2 σ k2
e
to the Bayesian Forecasting System (BFS) with the aim of estimating PU from
a set of historical observations and a hydrological model prediction (Kelly and
(
γ k ηˆt0 +k∆t ηt0 +k∆t ,ηt0 = ) 2π σ k
(26.12)
Krzysztofowicz, 2000; Krzysztofowicz, 2002; Krzysztofowicz and Maranzano,

2004). The BFS was developed around the idea of converting both observa- Step 4: Deriving the Posterior
tions and model predictions into a Normal space to allow for the derivation Density Function
of the joint and the predictive conditional distribution using a mathematically
treatable multivariate distribution. In Krzysztofowicz earlier works this trans- Following the Bayes theorem, the definition of the posterior density is:
formation was performed by means of the NQT (Van der Waerden, 1952,
(
γ k ηˆt0 +k∆t , ηt0 +k∆t ηt0 )
1953a, b; Kelly and Krzysztofowicz, 1997; Bogner et al., 2012), a convenient (
ϕ k ηt0 +k∆t ηˆt0 +k∆t , ηt0 = )
nonparametric approach, but can also be done using appropriate parametric
distributions. In practice, after converting the observations and the model
(
γ k ηˆt0 +k∆t ηt0 )
forecasts available for the historical period into the Normal space, the BFS
combines the prior predictive uncertainty (in this case derived using an =
( ) (
γ k ηˆt0 +k∆t ηt0 +k∆t ,ηt0 ϕ k0 ηt0 +k∆t ηt0 ) (26.13)
∫ −∞ γ k (ηˆt +k∆t ηt +k∆t ,ηt )ϕ k0 (ηt +k∆t ηt ) dηt +k∆t
+∞
autoregressive model) with a Likelihood function in order to obtain the pos-
terior density of the predictand conditional to the model forecasts. From the 0 0 0 0 0 0
Normal space this conditional density is finally reconverted into the real
space in order to provide the predictive probability density. Given that both the prior density (Eq. 26.10) and the Likelihood function
The introduction of BFS generated a positive impact on the hydrological (Eq. 26.12) are normal-linear, Krzysztofowicz (1999), Krzysztofowicz and
community, in fact Krzysztofowicz (1999) was the first in hydrological fore- Herr (2001) apply the theory of conjugate families of distributions (De Groot,
casting who clarified the concept of predictive uncertainty. Nonetheless, BFS 2004) to derive the posterior density, which becomes:
has two major limitations. The first one relates to the fact that only one model
( )
2
1 ηt0 + k∆t − Akηˆ t0 + k∆t + Bkηt0 +Ck
at a time can be used in BFS, which is hardly extendable to multimodel fore- −
2 Tk2
e
casts. Moreover the used prior Lag-1 Markov process model, which is essen-
tially used for mere reasons of mathematical tractability in deriving the BFS,
(
ϕ k ηt0 +k∆t ηˆ t0 +k∆t , ηt0 = ) 2π Tk
(26.14)
frequently tends to be inadequate in representing the predictand particularly
at longer time steps, like in the case of a flood routing problem where the with
model is adequate for representing the recession but not necessarily the rising
limb of the flood wave. The Markovian structure assumed for the family of ak (1 − ρ 2 k ) ρ kσ k2 − akbk (1 − ρ 2 k ) −a c (1 − ρ 2 k )
Ak = ; Bk = ; Ck = 2 k k 2 k ; and
prior densities is only a prototype. For instance, Reggiani and Weerts (2008) ak (1 − ρ 2 k ) + σ k2
2 2 2k
ak (1 − ρ ) + σ k 2
ak (1 − ρ ) + σ k2
in the development of the operational flood-forecasting system of the river
(1 − ρ 2 k )σ k2
Rhine, removed some limitations of the prior density in the original BFS, by Tk2 = .
including upstream observations in the prior distribution estimation. ak2 (1 − ρ 2 k ) + σ k2
The basic ideas of BFS are here briefly summarized.
Step 5: From the Normal Space to the
Step 1: Observations and Forecasts are Real World
Converted from the Real World to the
Normal Space In order to convert the results into the real world and obtain the predictive
probability f k ( yt0 +k∆t | yˆt0 +k∆t , yt0 ), one must sample the predictive density of
The first step in Krzysztofowicz (1999) approach is the conversion into a Eq. (26.14) in the Normal space and reconvert the obtained quantiles into the
Normal space of both the n observations y and the corresponding model real space by a reverse process with respect to the one described in Step 1.
predicted values ŷ into their Normal N(0.1) h and η̂ , either by means of a This is due to the fact that the transformation is highly nonlinear. Note for
parametric or a nonparametric approach as explained in Sec. 26.5.1. instance that the mean value in the Normal space does not correspond to the
Step 2: Setting up the Prior Assumption mean value in the real world, it corresponds to the median (50% probability).
Although widely used (Reggiani and Weerts, 2008; Reggiani et al., 2009), the
The application of the Bayes theorem requires, first of all, the assumption on a performances of BFS are strongly affected by the quality of the a priori model
prior predictive probability. Krzysztofowicz assumes a Lag-1 Markov process defined in Step 2; this is generally expected when a relatively good hydrologic or
for h, the image of the observations in the Normal space. The model can be hydraulic forecasting model is available and, at the same time, the forecasting
formulated as ηt0 +∆t = ρηt0 + εt0 +∆t , with εt0 +∆t a Normal N (0, 1 − ρ 2 ) indepen- lead time is large enough to degrade the performances of the lag-1 Markov pro-
dent process, where the parameter ρ = E{ηtηt −∆t }, the Lag-1 Markov coefficient, cess assumed to develop the a priori predictive density (Todini, 2013). The theo-
is estimated from the observations. In order to issue a forecast k ∆t in advance, retical background of BFS does not prevent using more complex prior models
this model can be extended by successive applications of the Lag-1 model, lead- (Reggiani and Weerts, 2008), but the chosen ones must allow mathematical
k
tractability in the analytical derivation of the posterior density as in Step 4.
ing to the following predictive model ηt0 +k∆t = ρ kηt0 + ∑ ρ k− jεt0 + j∆t . This model
j =1 26.5.3 Quantile Regression
gives rise to a prior predictive probability Φ0k (ηt0 +k∆t t0 ) = N ρ kηt0 , 1 − ρ 2 k ,( ) The previously described technique (BFS) can handle a certain type of het-
which can be written, in terms of the probability density, as: eroscedasticity in the real space, but implies homoscedasticity of the error
variance in the Normal space, which is assumed to be independent from the
( k
1 ηt0 + k∆t − ρ ηt0 )
2
magnitude of the observed or forecasted values. In real cases this assumption
−
2 1− ρ 2 k may lead to a lack of accuracy, especially at reproducing high flows, because the
e
( )
ϕ k0 ηt0 +k∆t ηt0 =
2π (1 − ρ 2 k )
(26.10) NQT tends to increase the variance of the lower values. Moreover, the number
of observed and computed low and medium flows is much larger than that of
26_Singh_ch26_p26.1-26.16.indd 5 8/22/16 1:44 PM

high flows with the consequence of a higher weight in the determination of the The problem is correctly formulated and allows each quantile of the PU to
regression or the correlation coefficients used by the different approaches. As a be computed, but it requires the estimation of at least two parameters per
consequence, the estimation of high flows in the Normal Space will be affected quantile (in the linear case) and the number of parameters to be estimated
by a distortion in the mean as well as an overestimation of the variance, which may become quite large. Moreover, QR does not always improve by assuming
will inevitably increase when returning into the real space. homoscedasticity: this depends on the actual distribution of the errors.
Recently, in order to overcome this problem, the QR was introduced by Although disagreed with Koenker, Weerts et al. (2011) suggest that better
Koenker (2005) and applied by several authors as for instance Weerts et al. results could be obtained by transforming the observations and predictions by
(2011).  means of NQT before applying the QR. This is done in Figs. 26.3 and 26.4,
The QR approach tries to represent the error heteroscedasticity identifying which show two situations where the use of QR leads to opposite results.
a linear (or non-linear) variation of the quantiles of the predictive density as Figure 26.4 is an optimal situation for using QR because the variation of
a function of the model forecast. The τ th sample quantile is computed by error variance is linearly decreasing with the magnitude of the forecasts and
identifying the parameters with a sort of constrained type regression, solved the resulting quantiles well represent the real distribution of the data. On the
by means of Linear Programming. contrary, in Fig. 26.3 it is not possible to identify a linear variation of the error
The quantile regression estimates a set of parameters βτ for any given variance and the use of QR does not provide improved assessments of PU,
quantile probability t, with 0 ≤ τ ≤ 1. particularly for high forecast values. As can be seen in Fig. 26.3, the predictive
The quantile regression (Koenker, 2005) is defined as: uncertainty estimated with QR tends to be unrealistically small (zero at the
n
point of the crossing quantile lines) for large forecast values. This points out
min βτ ∈ℜ ∑ ρτ ( yi − ξτ ( yî , βτ )) (26.15) another problem of QR, which is the possible crossing of quantile lines due to
i =1
the independent estimation of the single quantiles.
where ρτ ( yi − ξ ( yî , βτ )) is a function of both t and yi − ξτ ( yî , βτ ) and is 26.5.4 Bayesian Model Averaging
defined as:
Introduced by Raftery (1993), Bayesian Model Averaging (BMA) has gained

 τ ( yi − ξτ ( yî , βτ )) ∀yi ≥ ξτ ( yî , βτ )
a certain popularity in the latest years. The scope of BMA is to estimate a
ρτ ( yi − ξ ( yî , βτ )) = 
 (τ − 1) ( yi − ξτ ( yî , βτ )) ∀yi < ξτ ( yî , βτ )
(
multimodel conditional predictive density f yt yˆt(1)t0 , yˆt(2) ˆ (m ) ˆ( j)
t0 , …, yt t0 , where yt t0 )
 is the jth model prediction and m the number of models, starting from the
(26.16) ( )
single model predictive densities fi yt yˆt(it)0 . While the single-model densities
( )
fi yt yˆt(it)0 follow the definition given for the PU, namely the conditional
The solution to this problem is then found using Linear Programming probability density defined as the ratio between the joint and the marginal
(Koenker, 2005). densities as per Eq. (26.4), the BMA multimodel predictive density is derived
Please note that these quantile regressions are performed in the original as a Bayesian mixture of densities:
untransformed space by regressing the observations over the model forecasts.
In the case of linear quantile bounds, ξτ ( yî , βτ ) = aτ + bτ yî the problem can
be set as follows:
(
f yt yˆt(1)t0 , yˆt(2) ) { (
ˆ (m ) ) (
ˆ (1) ˆ (2)) ˆ (m )
t0 ,…, yt t0 = E f1 yt yt t0 , f 2 yt yt t0 ,…, fm yt yt t0 ( )}
n (26.19)
min βτ ∈ℜ ∑ ρτ  yi − (aτ + bτ yî ) (26.17)
i =1 which results into an approximation of the formal multimodel conditional
with density:
) (f ( yˆ , yˆ ,…, yˆ ) )
(1) ( 2) (m)
f y , yˆ , yˆ ,…, yˆ


 τ  yi − (aτ + bτ yî ) ∀yi ≥ aτ + bτ yî
ρτ [ yi − (aτ + bτ yî )] = 
(
f yt yˆt(1)t0 , yˆt(2) ) (
ˆ (m ) ˆ (1) ˆ (2) ˆ (m)
t0 ,…, yt t0 ≅ f yt yt t0 , yt t0 ,…, yt t0 =
t
(1)
t t0 t t0
( 2)
t t0
(m)
t t0 t t0 t t0
 (τ − 1)  yi − (aτ + bτ yî ) ∀yi < aτ + bτ yî

(26.20)
(26.18) Differently from the BFS assumptions, in BMA all the models (including
the eventual AR prior model) are similarly considered as alternative models.
As can be seen this formulation implies that a couple of aτ , bτ values must be Raftery et al. (2005) developed the approach on the assumption that the pre-
estimated using Linear Programming, in order to represent the upper and dictand as well as the model forecasts were approximately Normally distrib-
lower quantile lines shown for instance in Fig. 26.3. uted, while Vrugt and Robinson (2007) relaxed this hypothesis and showed
2
Transformed observed data
5%
0
–4 –3 –2 –1 0 1 2 3 4 50%
95%
–1
–2
–3
–4
Transformed forecasted data
Figure 26.3 An optimal situation for using the QR in Normal space.
26_Singh_ch26_p26.1-26.16.indd 6 8/22/16 1:44 PM

Techniques aimed at assessing predictive uncertainty 26-7

1
5%
0
–4 –3 –2 –1 0 1 2 3 4 50%
–1 95%
–2
–3
–4
Figure 26.4 Poor results are obtained using QR in Normal space in the situation represented here, which, by the way, is
quite common in hydrological applications.
how to apply the BMA to Log-normal and Gamma distributed variables. As opposed to BFS, BMA can be easily extended to multiprediction prob-
In practice the Bayesian Inference problem, namely the need for estimating lems, such as for instance to derive multisite predictive densities relevant to
posterior densities, is overcome in the BMA by estimating a number of several water stages along the same river, thus benefitting from their spatial
weights via a constrained optimization problem. dependence.
 T m 26.5.5 Multiple Regression Approaches
 max log  = max ∑ log∑w f y yˆ ( j )
 wj
wj
t =1 j =1
j j t t t0 ( ) The Model Output Statistics Approach
 (26.21) Multiple Linear Regression approaches, also known as Model Output
m
 Statistics (MOS) were used in meteorological applications (Glahn and Lowry,
 s .t .∑ w j = 1 1972; Wilks, 1995) as an uncertainty post processor.
 j =1
In the MOS approach, the expected conditional mean and conditional vari-
ance are computed using a linear combination of models prediction:
(
where w j are the weights to be estimated, f j yt yˆt( tj0) is the conditional prob- ) yt = a0 + a1 yˆt(1)t0 + a2 yˆt(2) ˆ (m )
t0 + + am yt t0 + et (26.25)
ability of yt , the predictand at site s and time t, given yˆt( tj0), model j unbiased
prediction for yt, with T the total number of observation time intervals and m
with et as residual error, which variance is minimized via multiple-linear
the number of used models.
regression.
Once the weights w j have been estimated, BMA allows estimating the
This leads to:
expected predictive density of the predictand conditional to several model
predictions.
m
{
E yt yˆt(1)t0 , yˆt(2) }
ˆ (m ) ˆ ˆ ˆ (1) ˆ ˆ (2) ˆ ˆ (m )
t0 ,…, yt t0 = a0 + a1 yt t0 + a2 yt t0 + + a3 yt t0 (26.26)
( )
t0 ,…, yt t0 ≅ ∑w j f j yt yt t0 (26.22)
f yt yˆt(1)t0 , yˆt(2) ˆ (m ) ˆ( j) ( )
j =1 {
Var yt yˆt(1)t0 , yˆt(2) }
ˆ (m ) = Var{eˆt } (26.27)
t0 ,…, yt t0
From Eq. 26.22 one can then estimate the BMA expected prediction mean as:
m where eˆt = yt − a0 + a1 yˆt(1)t0 + a2 yˆt(2) ˆ (m )
t0 + + am yt t0 .
{
E yt yˆt(1)t0 , yˆt(2) ˆ (m ) }
t0 ,…, yt t0 = ∑ w j E yt yt t0
ˆ( j)
j =1
{ } (26.23) The multiple-linear regression implies that the distribution of residuals et
is the zero mean Normal distribution N (0,Var{e}), which is hardly encoun-
while an approximation of the variance results: tered in real world cases. This implies that in general the MOS approach
allows estimating the expected mean and variance, but not the full predictive
m
{ }
t0 ,…, yt t0 = ∑w j Var yt yt t0
Var yt yˆt(1)t0 , yˆt(2) ˆ (m ) ˆ( j) { } distribution, which becomes:
( )
j =1
N aˆ0 + aˆ1 yˆt(1)t0 + aˆ2 yˆt(2) ˆ ˆ (m ) ˆ
t0 + + a3 yt t0 , Var{e} (26.28)
2
 m m 
+∑w j  yˆt( tj0) − ∑w j E yt yˆt( tj0)

{ }  (26.24)


where Var{eˆ} = Var{eˆt }, which is assumed constant in time.
j =1 j =1 Please note that if observations and predictions are converted into the
Normal space prior to its application, MOS becomes similar to the MCP (see
In hydrological implementation of BMA, the probabilistic structure of data, Sec. 26.5.5) and provides a coherent estimator of the predictive distribution.
which are assumed to be approximately Normally distributed in the original
development, is inappropriate to represent highly skewed quantities such as The Ensemble Model Output
water discharges or water levels in rivers, as pointed out by Duan et al. (2007) Statistics Approach
and Vrugt and Robinson (2007). Therefore, a probabilistic transformation of Gneiting et al. (2005) introduced a variant of MOS for handling Ensembles,
data (such as NQT or Box-Cox transformations) is required in order to which they called EMOS (Ensemble Model Output Statistics). In EMOS,
improve the performances of the BMA approach. Gneiting et al. (2005) take into account the spread of the ensemble to avoid
Moreover, if the single model conditional distributions used in BMA are underdispersion by assuming that the Var{et } may vary in time as a function
not calibrated, yˆt t0 is not guaranteed to be an unbiased prediction of yt . In of the ensemble spread variance St2 , namely
which case a bias correction by means of a linear regression need to be applied
to yˆt t0 to get an unbiased prediction (Raftery et al., 2005). Var{et } = c + dSt2 (26.29)
26_Singh_ch26_p26.1-26.16.indd 7 8/22/16 1:44 PM

The EMOS approach uses a heteroschedastic regression leading to the follow- Due to the NQT, the marginal distributions are all N (0, 1); therefore, the
ing predictive density: variance-covariance matrices appearing in Eq. (26.33) coincide with the cor-
relation matrices:
(
N aˆ0 + aˆ1 yˆt(1)t0 + aˆ2 yˆt(2) ˆ ˆ (m ) 2
t0 + + a3 yt t0 , c + dSt (26.30) )
with conditional mean: ¬ηη = 1; ¬ηηˆm =  ρηgˆ 1 ρηηˆ 2  ρηηˆm 
 
{
E yt yˆt(1)t0 , yˆt(2) }
ˆ (m ) ˆ ˆ ˆ (1) ˆ ˆ (2) ˆ ˆ (m )
t0 ,…, yt t0 = a0 + a1 yt t0 + a2 yt t0 + + a3 yt t0 (26.31)  1 ρηˆ 1ηˆ2 ρηˆ 1ηˆm 
 
and conditional variance  ρˆ ˆ 
η 2η 1
=  (26.34)
{ }
(2) ¬ gˆmgˆm
Var yt yˆt(1)t0 , yˆ ,…, yˆt(mt0 ) ˆ 2 (26.32)
= cˆ + dS  ρηˆm−1ηˆm 
t
t t0  
 ρηˆmηˆ 1 ρηˆmηˆm−1 1 
In recent applications EMOS was extended to GEV (Scheuerer, 2014; Lerch  
and Thorarinsdottir, 2013) distributed variables, but in general, as in the case of
MOS, although the conditional mean and variance are acceptable, the estimated The joint observations/forecasts probability density is the multi-Normal
predictive distribution is hardly representative of the actual one, unless one density:
converts the observations and the forecasts into the Normal space as proposed
1 T
by Gneiting et al. (2005), similarly to what is done by MCP (see Sec. 26.5.5). − (η ,gˆm )− lη ,gˆ  ¬ g−ˆ1 gˆ (η ,gˆm ) − lη ,gˆ 
e 2 m m m m
Similarly to BMA, all the regression approaches can be easily extended to f (η , gˆm ) = (26.35)
handle multiprediction problems. (2π )(m+1) ¬η ,gˆm
26.5.6 The Model Conditional Processor Following the definition, the conditional predictive density at a specific
The analysis of the previously described approaches, together with the conve- time t can be analytically obtained by exploiting the properties of the Normal
nient properties of the multivariate Normal distribution, generated the idea of Distributions (Mardia et al., 1979) by dividing the joint by the marginal density.
generalizing the use of the NQT to derive what will be called the Model
(η − lη )2
Conditional Processor (MCP) (Todini, 2008). Based on the properties of the − t gˆm ,t
2 σ η2 t gˆm ,t
Multi-Normal distribution (Mardia et al., 1979), the MCP allows to directly f (ηt , gˆm ,t ) e
assessing the density of the predictand conditional on all the model forecasts f (ηt gˆm ,t ) = +∞
= (26.36)
through a multiple regression in the Normal space. Similar to what is done in ∫ −∞ f (ηt , gˆm ,t ) f (ηt )dηt 2πσ η t gˆm ,t
BFS, the derivation of the predictive density is performed after converting
observations y and model forecasts ŷ into the Normal space, which can be where
performed using either the NQT or by fitting parametric marginal distribu-  −1 ˆ
tions to both observations and forecasts as described in Sec. 29.5.1.  µηt gˆm,t = ¬ηgˆm ¬ gˆmgˆm gm ,t
Accordingly, the observations y and their forecasts yˆmT = [ yˆ (1) , yˆ (2) ,…, yˆ (m ) ]  2 −1
(26.37)
 σ ηt gˆm,t = 1 − ¬ηgˆm ¬ gˆmgˆm ¬ηTgˆm
are converted into the corresponding h and ηˆmT = [ηˆ(1) , ηˆ(2) ,…, ηˆ(m ) ], all char- 
acterized by Normal N (0, 1) marginal distributions.
As in BFS, observations and forecasts are assumed to be linearly related in More recently, Zhao et al. (2011) introduced the General Linear Model
the multi-normal Space, leading to the joint probability distribution Post-Processor (GLMPP), which uses the same conceptual MCP approach
N (µ η ,gˆm , Ση ,gˆm ) with and practically leads to the same MCP equations. Also Wang and Robertson
(2011), after converting seasonal flows in the Normal space, used a multisite
  0  probabilistic processor based on an approach quite similar to MCP.
   Similar to BFS, the basic MCP implies homoscedasticity in the Normal
 µ η ,gˆm =    space. In order to account for heteroscedasticity, the MCP approach was thus
  0 
(26.33) extended from the Normal to the Truncated Normal distributions (Coccia

  ¬  and Todini, 2011).
ηη ¬ ηgˆm  This approach allows decomposing the range of observations, according to the
 ¬ =
 η ,gˆm  ¬ T ˆ ¬  need, into two or three separate domains, for each of which the appropriate pre-
  ηgm ηˆmgˆm 
dictive Truncated Normal densities can be analytically developed (see Fig. 26.5).
4.5
3.5
2.5
1.5
0.5 5% quantile
–4.5 –3.5 –2.5 –1.5 –0.5 0.5 1.5 2.5 3.5 4.5
–0.5 50% quantile
–1.5 95% quantile
–2.5
–3.5
–4.5
Figure 26.5 Two different Truncated Normals are used to account for heteroscedasticity of data in the
Normal space giving rise to two predictive densities, one valid below and one above a given threshold to
be identified.
26_Singh_ch26_p26.1-26.16.indd 8 8/22/16 1:44 PM

Major reasons undermining the operational use of predictive uncertainty 26-9
26.5.7 Other Approaches a function of the determinant of the covariance matrix S[m, m], of the fore-
Non-Gaussian Copulas approaches (Sklar, 1959; Nelsen, 1999; Aas et al., cast ensemble, or of the predictive density, as an extension of the univariate
2009) have been recently introduced in climatological applications (Schölzel, standard deviation.
C. and Friederichs, 2008) and are used to set up uncertainty post processors A forecast is “calibrated” if predicted events declared to have probability p,
for forecasted wind and precipitation (Schefzik et al., 2013), and, in hydrol- actually occur p times on average (Gneiting et al., 2008).
ogy, for forecasted discharges (Madadgar et al., 2012). Schefzik et al. (2013) When the predictive density is represented by discrete ensemble forecasts
use a combination of BMA and empirical copulas where BMA is applied to or by discrete quantiles (as in QR), calibration assessment is usually per-
derive the predictive distribution while the Empirical Copula approach ECC formed via the verification rank histogram or Talagrand diagram (Anderson,
is applied to consider multivariate dependencies in temporal, spatial, and 1996; Hamill and Colucci, 1997; Talagrand et al., 1997; Hamill, 2001). The
spatio-temporal weather trajectories. This approach is similar to the Schaake Talgrand histogram is a plot of the histogram of the rank of the verifying
Shuffle (Clark et al., 2004). observation when pooled within the ordered ensemble values. Alternatively,
Kalman Filters (KFs) (Kalman, 1960) have also been used in order to assess the Kolmogorov–Smirnov two sample tests can also be used.
predictive uncertainty (Diomede et al., 2008), although in a peculiar way. Another verification technique is the probability integral transform (PIT)
Following the definition of “filters,” KFs can only be effectively applied up to (Dawid, 1984; Diebold et al., 1998; Gneiting et al., 2007) which is similar to
time t0, namely until measurements are available. In order to use KFs for the verification rank histogram.
prediction, when t > t0 , the model(s) output forecast(s) must be regarded as When the full continuous predictive density is derived (as in BFS, BMA,
error corrupted measurement(s) of any future realization (Todini, 2006) and, and MCP), the calibration assessment can be performed using reliability
similar to what is done in the BFS (Krzysztofowicz, 1999), a prior assumption diagrams (Wilks, 1995), mostly used in meteorological applications (Hamill
on the future realization is also necessary. and Colucci, 1997; Bremnes, 2003) and recently proposed in hydrological
Alternative approaches for determining predictive uncertainty have been forecasting by Laio and Tamea (2007). Chi-square or Kolmogorov–Smirnov,
proposed by Bogner and Pappemberger (2011); by Regonda et al. (2013), one sample tests can also be used.
namely the Hydrologic Model Output Statistics (HMOS); and by Brown and It is also worthwhile mentioning the Continuous Ranked Probability Score
Seo (2013), who introduced a nonparametric approach for the postprocessing (CRPS) (Matheson and Winkler, 1976; Hersbach, 2000) as a single scalar
of hydrological ensemble forecasts, which is similar to indicator co-kriging. measure that encompasses information about different attributes about fore-
Other approaches have also been studied but, although theoretically devel- cast quality (reliability, resolution, and uncertainty) and has been used in
oped, have not been yet used in practical applications. Among these, one can several recent hydrological applications related to predictive uncertainty
recall Montanari and Brath (2004) suggesting the regression of model errors assessment (e.g., Regonda et al., 2013; Engeland and Steinsland, 2014).
over the model forecast; Solomatine and Shrestha (2009) proposing the
“UNcertainty Estimation based on local Errors and Clustering” approach; 26.7 MAJOR REASONS UNDERMINING THE
Alvisi et al. (2010) introducing the use of “Grey Neural Networks.” OPERATIONAL USE OF PREDICTIVE UNCERTAINTY
26.5.8 Using Meteorological Predictions with the In the case of hydrological predictions, only rarely the additional information
Uncertainty Processors provided by the predictive density has been fully accepted and used within the
Meteorological models quantitative precipitation forecasts are one of the decision-making process. In most cases it has been merely used to attach a
major sources of errors in real-time flood forecasting. It is evident that if the measure of reliability to an alternative form of “deterministic” forecast based
uncertainty processor was developed by using measured precipitation as on the mean (mean forecast, mean of the ensemble, etc.).
input to the hydrological models, then “consistency” between the model and This suboptimal use of the uncertainty information was caused by several
input data used in assessment, and the model and input data used in predic- reasons, such as (1) unclear and mostly wrong definition of predictive uncer-
tion is not guaranteed (see Sec. 26.4.2). This implies the need for further tainty and its use; (2) unclear role and use of parameters uncertainty often
acknowledging, accounting for, and marginalizing the uncertainty embedded confused with predictive uncertainty; (3) unclear role and use of ensembles in
in the meteorological precipitation forecasts, as for instance correctly done by the assessment of predictive uncertainty; (4) unclear understanding and defini-
Kelly and Krzysztofowicz (2000) through the PUP component in the BFS. tion of the prediction-communication-decision chain; and (5) unclear under-
Nonetheless, it is worthwhile mentioning that in hydrological forecasting standing of the mechanism and the advantages for using predictive uncertainty
practice, PU is assessed in hindcasting mode over past available records, by in the decision-making process.
using either observed rainfall or past meteorological precipitation forecasts. All these reasons are discussed as follows.
Therefore, one can directly use the uncertain meteorological precipitation 26.7.1 Unclear and Mostly Wrong Definition of
forecasts as input to the hydrological model in the development of the uncer- Predictive Uncertainty and Its Use
tainty processor. This guarantees the required “consistency” between the
model and input data used in assessment and the model and input data used Another major problem that undermined the proper use of predictive
in prediction, and no further uncertainty assessment is necessary. uncertainty in the last two decades is the fact that predictive uncertainty has
been usually confused with SU or with the spread deriving from parameter-
sensitivity analyses. There are many cases in the literature where the predic-
26.6 VERIFICATION OF THE ESTIMATED tive density is essentially based on the model residuals analysis. This leads to
PREDICTIVE DENSITY various incorrect results.
As mentioned earlier, probabilistic predictions are characterized by the vari- In order to clarify this point, let us assume, for this demonstration without
ance of the predictive density (or the spread of quantiles), a property gener- loss of generality, that the predictand (the observation) y and the predictor (for
ally referred to as the “sharpness” of the forecast, which is a property of the example a model forecast) ŷ are both Normally distributed and the residual
prediction only. are homoscedastic.
Bearing in mind that the final goal of probabilistic forecasting is decision- The definition given in Sec. 26.4.3:
making, one of the objectives is to maximize the sharpness of the forecast
f ( y , yˆ )
PDFs conditional to calibration (Gneiting et al., 2003), calibration aims at f ( y yˆ ) = (26.39)
reaching consistency between the forecasted densities and what can be veri- f ( yˆ )
fied from actual observations.
A forecast technique is “sharp” if the predictive densities are denser on under the assumptions of normality and homoscedasticity leads to the follow-
average than predictive densities (or prediction intervals are shorter on aver- ing predictive density:
age than prediction intervals) derived from a less sharp approach. The denser
the predictive densities are, the sharper the forecast, and the sharper the bet- ( y −µ y yˆ )2
−
ter subject to calibration. 2σ y2 yˆ
e
Several measures of scatter and spread have been discussed in the statistical
literature as for example by Bickel and Lehmann (1979), Oja (1983), and
{ }
n µ y yˆ ; σ y2 yˆ =
2π σ y yˆ
(26.40)
Shaked and Shanthikumar (1994). Gneiting et al. (2008) as a measure of
sharpness for the multivariate prediction case suggest the use of the following where
statistics:
ρ yyˆσ y
Sm = 2m ¬ (26.38) µ y yˆ = µ y +
σ yˆ
(
( yˆ − µ yˆ ) ; σ y2 yˆ = 1 − ρ yy
2
)
2
ˆ σ y (26.41)
26_Singh_ch26_p26.1-26.16.indd 9 8/22/16 1:44 PM

with conditions, etc., and techniques, based on Bayesian inference, are used to
assess parameters’ posterior density.
Covar { yyˆ} Parameters’ posterior density should then be used to marginalize their effect,
µ y = E { y } ; µ yˆ = E { yˆ} ; σ y2 = Var { y } ; σ y2ˆ = Var { yˆ} ; ρ yyˆ =
σ yσ yˆ by computing an “expected” predictive density, but never be interpreted (as
unfortunately frequently happens) as a measure of the predictive density.
(26.42) Unfortunately, several SA approaches have been erroneously interpreted as
It can easily be shown that the simulation density: a mean to provide a measure of predictive uncertainty. A typical case relates
to the Generalized Likelihood Uncertainty Estimation (GLUE) (Beven and
( yˆ −µ yˆ y )
2
−
Binley, 1992), a widely used technique, which applies a Bayesian inferential
2σ y2ˆ y approach based on “less formal Likelihoods.” Apart from criticisms on the use
f ( yˆ , y ) e
f ( yˆ y ) = = n{µ yˆ y ;σ y2ˆ y } = (26.43) of the less formal likelihoods (Mantovan and Todini, 2006; Mantovan et al.,
f ( y) 2π σ yˆ y 2007), GLUE assumes that a measure of the cumulated predictive density is
given by the cumulated normalized likelihood. This leads to the GLUE pre-
where dictive probability (Beven and Freer, 2001):
ρ yyˆσ yˆ
µ yˆ y = µ yˆ +
σy
2
(
( y − µ y ); σ y2ˆ y = 1 − ρ yy 2
ˆ σ yˆ (26.44) )
p
Prob{ yˆt ≤ y } = ∑ {ϑ i yˆt (ϑ i ) ≤ y }
p
∑ {ϑ i yˆt (ϑ i )} (26.47)
i =1 i =1
differs from the formal definition of predictive density, since µ yˆ y ≠ µ y yˆ and
σ y2ˆ y ≠ σ y2 yˆ and may highly distort the estimation of expected damages or In Eq. (26.47), ϑ i represents one of the p parameter sets generated via Monte Carlo
benefits as discussed in Sec. 26.4.2, particularly due to the incorrectly esti- sampling, while {ϑ i yˆt (ϑ i )} the parameters likelihood. In practice, Eq. (26.47)
mated bias. represents the cumulated posterior parameter probability Prob"{ϑ i }, which in
Another erroneous approach, widely found in the literature, is the direct use GLUE is proportional to the Likelihood, namely:
of the distribution of residuals, which leads to the following predictive density: p
( y − yˆ − µε )2
Prob"{ϑ i } = {ϑ i yˆt (ϑ i )} ∑{ϑ i yˆt (ϑ i )} (26.48)
− i =1
2σ ε2
e
{
n yˆ + µε ; σ ε2 = } 2π σ ε
(26.45) due to the choice of a noninformative uniform prior Prob'{ϑ i }.
As can be noticed, Eq. (26.47) results do not comply to the formal defini-
where tion of predictive probability given in Sec. 26.4.2, while a more rigorous way
to proceed, when assuming uncertain parameters, would be (1) to derive the
µε = µ y − µ yˆ ; σ ε2 = Var { y − yˆ} = σ y2 + σ y2ˆ − 2 ρ yyˆσ yσ yˆ (26.46) posterior parameter density and (2) to marginalize the uncertainty using the
posterior parameters density:
Not only the density of Eq. (26.45) doesn’t follow the definition, but it can
p
also be easily proven that apart from the case when ρ yyˆ =1, which implies the
Prob{ yt ≤ yt∗ yˆt } = ∑Prob{ yt ≤ yt∗ yˆt (ϑ i )} Prob"{ϑ i } (26.49)
perfect forecast, it is always less sharp than the predictive density descending
i =1
from the definition. This can be easily shown for σ y2 = σ y2ˆ , in which case
σ y2 yˆ σ ε2 Equation (26.49) is written in discrete form due to the large but not infinite
2
2
= (1 − ρ yy
ˆ)≤ = 2(1 − ρ yyˆ ), as plotted in Fig. 26.6. In the special case of number p of generated parameter sets. Using the GLUE posterior parameter
σy σ y2 probability of Eq. (26.48) leads to:
σ y2 yˆ σ 2
ρ yyˆ =1, σ ε2 = 0 from Eq. 26.43 and 2 = ε2 = 0. p p
σy σy
Prob{ yt ≤ yt∗ yˆt } = ∑Prob{ yt ≤ yt∗ yˆt (ϑ i )} {ϑ i yˆt (ϑ i )} ∑ {ϑ i yˆt (ϑ i )}
i =1 i =1
4 (26.50)
3.5 an expression that highly differs from that of Eq. (26.47) where the condi-
tional probability term Prob{ yt ≤ yt∗ yˆt (ϑ i )} does not even appear.
3 To understand the rationale of Eq. (26.50), one must realize that each gen-
2(1–r yy^)
2.5 erated parameters set produces a different predictive probability conditional
on the specific parameter set. Some of these predictive probabilities corre-
2 sponding to more appropriate parameter values will be sharper, while others
corresponding to less accurate parameter values will appear rather flat.
1.5 Therefore, to obtain an “expected predictive distribution,” one must marginal-
1 ize the uncertainty embedded in the parameters by integrating, with respect
1–r 2yy^ to parameters, the product of the predictive probability function conditional
0.5 on the parameters times the posterior parameters’ distribution, the most
informative given the available information.
0 To better clarify the reasons for avoiding the use of GLUE as a measure of
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 predictive uncertainty, let us look at Fig. 26.7 displaying the results of a simu-
r yy^ lation where parameter uncertainty is assessed via 10,000 Monte Carlo
parameter sets realizations. The central part of this figure shows the results
σ y2 yˆ σ ε2 derived from Eqs. (26.47) and (26.50), while the two external graphs enlarge
Figure 26.6 Plot of (curved line) and
(straight line) for the special case
σ 2
y σ y2 the period ranging from 280 to 340 in order to improve readability. Both
2 2
σ y = σ yˆ . The proper definition of predictive uncertainty always leads to a sharper prob- graphs represent the observations as thick solid lines. The upper one shows
abilistic forecast, apart when ρ yyˆ = 1, namely the perfect (deterministic) forecast. the expected value of the observations (assumed unknown) conditional upon
the model forecast (thin solid line) and the 5–95% uncertainty band esti-
mated according to Eq. (26.50) (gray shaded band). The lower one, following
26.7.2 Unclear Role and Use of Parameters the GLUE approach, shows the median of the forecasts (thin solid line) and
Uncertainty Often Confused with Predictive Uncertainty the 5–95% uncertainty band estimated according to Eq. (26.47) (gray shaded
Frequently parameters-induced uncertainty has been confused with predic- band). It can be observed that while the shaded band of the upper graph
tive uncertainty. encompasses most observations this does not happen for the lower one.
In particular widely spread SA techniques have led to the wrong estimation Moreover, if one considers the period around the ordinate 310 it is possible
of predictive uncertainty. to understand that the very small spread of the GLUE uncertainty band is due
In a Bayesian perspective, parameters for a given model can be interpreted to the insensitivity of the model forecasts to the parameters’ variation, but is
as the receptors of all the uncertainty embedded in the forecasts, including not a representation of the probability of occurrence of the observed value,
uncertainty in observations, in parameterization, in initial and boundary taken as unknown, which in this case, would be predicted practically zero.
26_Singh_ch26_p26.1-26.16.indd 10 8/22/16 1:45 PM

Major reasons undermining the operational use of predictive uncertainty 26-11
1000
500
0
280 290 300 310 320 330 340
1500
Q [m3 s–1]
1000
500
0
0 50 100 150 200 250 300 350 400 450 500
(c)
1500
Q [m3 s–1]
1000
500
0
0 50 100 150 200 250 300 350 400 450 500
1000
500
0
280 290 300 310 320 330 340
Figure 26.7 Comparison between a standard Bayesian approach of Eq. (26.50) (upper graph) and the GLUE
approach of Eq. (26.47) (lower graph).
26.7.3 Unclear Role and Use of Ensembles in the model parameters have been constrained to be equal within each exchange-
Assessment of Predictive Uncertainty able group to account for exchangeability.
In recent applications, hydrological probabilistic forecasts have been based Exchangeability can also be exploited in the MCP approach (Todini, 2008;
either on uncertainty processors (Krzysztofowic, 1999; Todini, 2004; Todini, Coccia and Todini, 2011), by ranking the ensemble predictions in ascending
2008) or on ensembles, following meteorological traditional approaches and order at each time step and using, as predictors, the quantiles (Fig. 26.8b)
the establishment of the HEPEX program (http://hepex.irstea.fr). instead of the actual ensemble member (Fig. 26.8a).
Unfortunately, the direct use of ensembles as a measure of the predictive This allows using Eq. (26.37) with m predictors, where m is the number of
density is an incorrect practice, because the ensemble measures the spread of quantiles, because the ith quantile will remain consistently the ith quantile at
the forecast instead of the conditional probability of the future outcome con- each time step.
ditional on the forecast, as descends from the definition of predictive uncer- Alternatively without assuming exchangeability, MCP, essentially based on
tainty. Meteorological forecast ensembles, and therefore also hydrological a multiple regression in the Normal space, can be extended to use ensembles
forecasts ensembles, which results from the meteorological forecast ensem- to represent the local (in time) smaller or larger conditional predictive uncer-
bles as input, are typically biased and underdispersive (Hamill and Colucci, tainty as a function of the ensemble spread. This is done by modifying the
1997; Eckel and Walters, 1998) and therefore a bad direct estimator of the classical linear regression equations (Eq. 26.37), implying perfectly observed
predictive uncertainty. predictors, to alternative regression equations similar to the Kalman filter
Only few approaches are reported in the literature, which correctly use the ones, allowing for uncertain predictors
ensemble to estimate an expected conditional predictive density (Reggiani et
al., 2009), similarly to what is done when several predictive models are avail-  −1
able as in the BMA (Raftery et al., 2005) or MCP (Todini, 2008; Coccia and
Todini, 2011) approaches.
 l


η gˆ
t m ,t
= ¬ ηt
ˆ
gm ,t




¬ ˆ
gm ,t
ˆ
gm ,t
+ (1  ˆ
m 
R t  )
gm ,t − lgˆm
A major problem, limiting the correct use of ensembles, is in fact the dif-  (26.51)
−1
ficulty of defining the time dependence of the ensemble members, due to the  2  1  T
lack of a consistent ranking: in other words, when dealing with multiple  σ lηt gˆm ,t = 1 − ¬ ηt gˆm ,t  ¬ gˆm ,t gˆm ,t + Rt  ¬ ηt gˆm ,t
 m
models, the ith model remains the ith model regardless to the time of forecast,
while this does not happen when dealing with ensemble members, since there
is no definition for the ith member of an ensemble. with ĝm ,t the ensemble mean at time t, Rt the covariance matrix representing
To handle meteorological ensembles (Reggiani et al., 2009), use all the the spread of the ensemble at time t, and lĝm the mean over time of the
ensembles at the same time, while BMA and EMOS have been extended using ensemble means.” As opposed to Eq. (26.37), where the prediction variance is
the concept of exchangeable members (Fraley et al., 2010). Exchangeable constant in the Normal space, Eq. (26.51) allows to account for the predictive
members lack individually distinguishable physical features. BMA and EMOS uncertainty of both the ensemble mean and the ensemble spread.
26_Singh_ch26_p26.1-26.16.indd 11 8/22/16 1:45 PM

6500 6500
6000 6000
5500 5500
5000 5000
4500 4500
4000 4000
3500 3500
3000 3000
2500 2500
2000 2000
0 20 40 60 80 100 120 0 20 40 60 80 100 120
(a) (b)
Figure 26.8 The original 16 ensemble members displayed in (a) are resampled at each time step to provide the 16 ranked quantiles displayed in (b) and then used as predictors.
26.7.4 Unclear Understanding and Definition of the must be more effectively interpreted as a measure of “predictive knowledge”
Prediction-Communication-Decision Chain (Fig. 26.9). And it is the measure of predictive knowledge, by incorporating it
As discussed in Sec. 26.2, meteorologists and hydrologists provide decision in the decision-making process, that allows taking educated decisions.
makers with a measure of uncertainty more as “lack of knowledge” rather When dealing with measurable quantities, such as for instance water vol-
than “best knowledge,” which is commonly viewed as a “measure of caution” umes, water levels, etc., one of the most appropriate ways of describing our
at accepting the forecasts at face value rather than as an essential piece of understanding (knowledge) on their potential occurrence is by using a prob-
information aimed at “improving” decisions. Moreover, scientists, modelers, ability density function. While the total ignorance on a quantity can be math-
and forecasters are not fully aware of the decision-making procedures used in ematically represented by a uniform density lying over an extremely wide
operation while, at the same time, operational decision makers do not under- range (e.g., from 0 to +∞) and the perfect knowledge can be described by a
stand how a measure of uncertainty can be fully integrated in the decision- Dirac delta function (Dirac, 1958, pp. 58–61) over the exact value, any inter-
making process. mediate situation of knowledge can be mathematically represented by an
The basic problem is probably semantic, in the sense that what is commu- appropriate probability density function, which can be both interpreted as a
nicated as “predictive uncertainty” should in reality be more precisely termed measure of uncertainty or as a measure of knowledge.
“predictive knowledge” (Todini, 2014). If correctly assessed, “predictive knowledge” is the maximum level of infor-
As in the classical case of the half-empty/half-full glass, one must realize mation a forecaster can provide to the decision maker, who can then exploit
that what is generally referred-to as a measure of “predictive uncertainty” it to derive the most likely correct decision. Consequently, to maximize the
Total Predictive No
uncertainty uncertainty uncertainty
No Predictive Perfect
knowledge knowledge knowledge
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
–0.05
–4 –3 –2 –1 0 1 2 3 4
Figure 26.9 The representation of predictive uncertainty and predictive knowledge.
26_Singh_ch26_p26.1-26.16.indd 12 8/22/16 1:45 PM

Examples of proper use of predictive uncertainty to improve decisions 26-13
7"
6"
5" HT = 4.75m
H[m] 4"
3"
2"
1"
0"
1" 501" 1001" 1501" 2001" 2501" 3001" 3501" 4001" 4501" 5001" 5501" 6001"
Time [days]
Figure 26.10 The flood warning example. Comparison of observed discharge (black solid line), model forecasts
(dashed line), and expected water level conditional to the model forecast (grey line) compared to the threshold value
HT = 4.75 m (gray horizontal line).
likelihood of taking the correct decisions, it is fundamental to (1) correctly Comparison of the number of right/wrong decisions based on the model
assess a realistic measure of the available knowledge on the future events and forecast taken as deterministic (6 right–11 wrong) with the corresponding
(2) make full use of this measure of knowledge. decisions taken on the basis of the expected value conditional on the model
forecast (13 right–4 wrong) already reveals the increased robustness of the
26.7.5 Unclear Understanding of the Mechanism and probabilistic forecast over the deterministic one.
of the Advantages for Using Predictive Uncertainty in Moreover, larger benefits from probabilistic forecasts could descend from
the Decision-Making Process the full use of the probabilistic density forecast (Coccia and Todini, 2011),
Until present, very little effort has been placed into effectively demonstrating the such as issuing the warning:
real operational advantages when using probabilistic forecasts within the frame 1. On the basis of a probabilistic threshold, namely by issuing the alert
of the decision-making process. These advantages may be summarized as: when the probability of overtopping the dyke is higher than a given value
1. Increased robustness of decisions by reducing the number of missed and 2. When the expected value of an utility function expressing the risk pro-
false positives pensity of the decision maker is lower (or higher) than a given value
2. Increased expected benefits or reduced expected losses, descending from 3. When the expected value of an utility function expressing the risk pro-
the decisions pensity of the decision maker reaches a maximum (or a minimum)
These benefits must be demonstrated to the decision makers by:
26.8.2 A Theoretical Example: Management of an
1. Involving them from the start in the definition and in the refinement of Earth-Fill Dam Reservoir
the utility functions (see Sec. 26.4.1) according to their risk propensity and
requirements This extremely simplified problem is presented to enlighten the benefits for
2. Providing them with the suggested “optimal” expected decision, as a using the full measure of predictive knowledge instead of merely the model
result of the Bayesian decision process, rather than the probabilistic forecast forecast. The earth-fill dam reservoir in Fig. 26.11 is operated for irrigation
they do not fully grasp and energy production purposes. Operating rules have been established on a
3. Simulating on a long historical record the results of decisions taken on long-term basis, from which the average economic value L1 of water stored in
the basis of the Bayesian decision approach, which exploits the probabilistic the reservoir during the filling period is established to be:
forecasts, to estimate the expected values L1 ( X ) = c1 X (26.52)
4. Comparing the simulation results with the historical ones, generally
based on empirical rules or driven by deterministic forecasts where X[m ] is the volume of water and c1 [ € or $/m3 ] the cost per unit volume
3
Since improvements when using the probabilistic approach are so large, of water stored.
decision makers will inevitably increase their willingness to understand and
possibly use the proposed approaches. Volume
26.8 EXAMPLES OF PROPER USE OF PREDICTIVE Expected
UNCERTAINTY TO IMPROVE DECISIONS volume Release
This section includes theoretical and real world examples, demonstrating the
benefits of using the full predictive density as opposed to deterministic fore-
casts in decision-making. Damages
Current
26.8.1 A Theoretical Example: Issuing Flood Warnings volume
The following simple example is used to illustrate the increased robustness of
flood warnings when using probabilistic flood forecasts. At a specific cross
section, flood warnings are issued when the water stage Ht > HT = 4.75 m. In
order to gain operational time, a flood forecasting model is available.
Figure 26.10 compares several observed flood events Ht (black solid line),
the corresponding model forecasts Ĥt (dashed line) and the expected water
level conditional to the model forecast E{Ht | Hˆ t } (gray line) to the threshold
level HT (gray horizontal line). Figure 26.11 The simple schematic representation of the reservoir management problem.
26_Singh_ch26_p26.1-26.16.indd 13 8/22/16 1:45 PM

A flood wave has been forecasted which will raise the level of the reservoir After a series of algebraic manipulations, the following interesting result
from the current level to a future value, substantially higher than the reservoir can be obtained:
top (Fig. 26.11), which must be prevented for the safety of the dam. Losses
will anyway occur if the water volume becomes larger than the maximum  c 
Vx = µ V Vˆ − Vmax + σ V Vˆ N −1 1 − 1  (26.58)
reservoir content and the economical losses can be expressed as L2, a damage  c2 
function of the exceeding volume (V − Vmax ) (assumed to be linear for simpli-
fying the calculation as well as the reasoning without loss of generality).  c 
where N −1 1 − 1  represents the value of a Standard-Normal variable with
 c2 
 L (V ) = 0 V ≤ V c
2 max probability 1 − 1 .
 (26.53) c2
 L2 (V ) = c2 (V − Vmax ) V > Vmax The analysis of this result reveals that releasing Vx = µ V Vˆ − Vmax is best only
in two cases. Either when the unit losses for releasing water c1 = c2 /2 (which
where (V − Vmax )[m3 ] is the volume of water in excess, Vmax the maximum is not so, since the interesting case is when c1  c2), or in the case of a perfect
content in the reservoir, and c 2 [ € or $/m3 ] the unit cost of damages per unit model forecast (namely, V = µ V Vˆ = Vˆ and σ V Vˆ = 0) which again leads to
of water volume in excess. It is also reasonable to assume that c1  c2, to Vx = Vˆ − Vmax. In all the other cases of interest, one must release more than just
account for the consequences of flooding. Vx = µ V Vˆ − Vmax . The reason why is very simple. Not being entirely certain on
Let us first assume that a volume forecast V̂ is available and it is our inten- what may occur, one must use the precaution principle. The larger the uncer-
tion to use this quantity for establishing how much water should be preven- tainty is, the more cautious he must be and preventively release a larger
tively released in order to minimize the overall losses. In order to find the amount of water than what might be assumed optimal under the hypothetical
optimal release Vx one must then minimize the following function expressing assumption of perfect knowledge of the future.
the total losses:
26.8.3 A Real World Case: Operational Management of
minVx L(Vx ) = L1 (Vx ) + L2 (Vˆ ) = c1Vx + c2 (Vˆ − Vmax − Vx ) (26.54) the Lake Como
Lake Como in northern Italy is mainly used for irrigation and hydroelectrical
The minimum (being c1  c2 ) is obviously reached when Vx = Vˆ − Vmax , power production. Its outlet is controlled through a dam at Olginate and
namely when the stored water volume equals the maximum reservoir content. management of the lake has to cope with the necessity of saving water to
If V̂ could be a “perfect” forecast, namely the future value V would be iden- satisfy agricultural and hydroelectrical demands and, at the same time, must
tical to the predicted value V̂ (which implies assuming a Dirac delta over V̂ as guarantee safety against flooding, a risk increased in the past 40 years.
the measure of our predictive knowledge) then this rule represents the best The management problem is complicated by the relatively small reservoir
possible choice. But, V̂ cannot be perfect and it is unlikely that it will equal the control volume (246.5 mm3, roughly 1–20 of the yearly inflow volume com-
actual occurrence of the future volume V. bined with the reduced hydraulic capacity of the downstream gates that
In order to simplify the problem without losing generality, let us assume that allow releasing 900–1000 m3s–1 whereas the inflow reaches 1800–2000 m3s–1,
the predictive density, expressing our perception of the probability of occur- which leads to rapid rising and filling of the lake (3–5 days) in the case of
rence of a future value once we are aware of the model forecast, can be repre- large flood events.
sented by a Gaussian probability density with the following characteristics: The two basic requirements of management were: optimization of water
resources availability as long-term objective and flood risk management as a
(V −µV Vˆ )
2
− short-term objective. To cope with the long-terms objective, the operating
2 σ V2 Vˆ rules were established on a 10-day basis using a Stochastic Dynamic
e
f (V Vˆ ) = (26.55) Programming (SDP) algorithm (Alarcon and Marks, 1979; Todini, 1999)
2π σ V2 Vˆ based on transition matrices representing the probability of the inflow during
the ith 10-days period, conditional on inflow observed during the previous
period. In order to cope with floods, here considered as a short-term tempo-
where µ V Vˆ = E{V | Vˆ } is the expected conditional mean of the future value rary disturbance, a second short-term objective, namely flood crest subsid-
given the model forecast and σ V2 Vˆ = Var {V Vˆ } the variance of the departures ence, was also considered.
of the true unknown value from the expected conditional mean. As men- The shorter-term rule was then derived on the basis of daily inflow fore-
tioned earlier, if correctly assessed, this predictive density is the best represen- casts generated by a simple Nearest Neighbour (NN) approach (Yakovitz,
tation of our knowledge on the future occurrence. The conditional variance 1987). A conditional predictive density was estimated and used in the
σ V2 Vˆ is indicative of the quality of our knowledge. In the case of a utopic decision-making process for the estimation of the expected losses to be
“perfect” predictive model σ V2 Vˆ would tend to zero with the probability den- minimized (Todini, 1999).
sity degenerating into a Dirac delta over the value V = V. ˆ
The Como lake manager collaborated in the definition of a utility function
The effect of preventively releasing a volume of water Vx from the reservoir representing the discrete stages of increasing damages, which are summarized
will modify this conditional density as follows: in Fig. 26.12.
−
(
V − µ −V  2
 V Vˆ x 
)
2 σ V2 Vˆ
e
(
f V Vˆ ,Vx =) 2π σ V2 Vˆ
(26.56) 6
Penalty function [euros × 106]
5
While the variance will not be affected by the released volume Vx , the mean
of the predictive density, representing the future final volume stored in the 4
reservoir, will become µ V Vˆ − Vx , namely the expected future volume E{V Vˆ }
conditional on the model forecast minus the released volume Vx . The variance 3
of the predictive density is unchanged because no additional information
(knowledge) on the future event has been provided. 2
The optimal release can be found by minimizing the expected utility losses
(see Eq. 26.2) with respect to Vx , as: 1
0
( )
2
V − µ −V 

−
V Vˆ x 
 –120 –80 –40 0 40 80 120 160 200 240
+∞ 2 σ V2 Vˆ
e Lake level [cm]
minVx E{L(Vx )} = c1Vx + c2 ∫ (V − Vmax ) 2π σ V2 Vˆ
dV (26.57)
Figure 26.12 The utility function defined for the Como lake management problem.
Vmax
A high penalty for lake levels below –40 cm at Malgrate gauge and a step-wise increas-
ing quadratic function to represent the effects of high levels: at 120 cm the water starts
where the expectation is only taken over the second term, since Vx , is not flooding the main square; at 140 cm traffic must be stopped; at 173 the flood assumes a
uncertain: it is a fixed quantity, once the decision is taken. legal status for damages refund.
26_Singh_ch26_p26.1-26.16.indd 14 8/22/16 1:45 PM

REFERENCES 26-15
The resulting operating rule was simulated over a period of 15 years Dirac, P. A. M., The principles of quantum mechanics, 4th ed., International
(1981–1995) and compared to the recorded results deriving from the actual Series of Monographs on Physics 27, Oxford University Press, New York, 1958,
lake management. The obtained improvement was outstanding: p. 328, ISBN 978-0-19-852011-5.
1. The frequency of the city of Como flooding was reduced by over 30% Duan, Q. Y., N. K. Ajami, X. G. Gao, and S. Sorooshian, “Multi-model
2. The water deficit was reduced on the average of 110 mm3/year (12%) ensemble hydrologic prediction using Bayesian model averaging,” Advances
3. The electricity production was increased on average by 3% in Water Resources, 30 (5): 1371–1386, 2007.
A decision support system (DSS) was operationally installed in 1997 and Eckel, F. A. and M. K. Walters, “Calibrated probabilistic quantitative pre-
constantly used by the lake manager since. A recent evaluation of the Lake cipitation forecasts based on the MRF ensemble,” Weather Forecasting, 13:
Como management during a period of drought years 2000–2006 revealed 1132–1147, 1998.
that, although the results were extremely close, the real management outper- Edwards, A. L., “The correlation coefficient,” An Introduction to Linear
formed the simulated one. This was essentially due to the availability of Regression and Correlation, W. H. Freeman, San Francisco, CA, 1976, pp. 33–46.
external information, such as meteorological forecasts, satellite imagery of Engeland, K. and I. Steinsland, “Probabilistic postprocessing models for
cloud cover, etc., which additionally supported the lake manager in taking the flow forecasts for a system of catchments and several lead times,” Water
final decision. Resources Research, 50: 182–197, 2014, doi:10.1002/2012WR012757.
Fraley, C., A. E. Raftery, and T. Gneiting, “Calibrating multimodel forecast
ensembles with exchangeable and missing members using Bayesian model
REFERENCES
averaging,” Monthly Weather Review, 138 (1): 190–202, 2010.
Aas, K., C. Czado, A. Frigessi, and H. Bakken, “Pair-copula constructions of Glahn, H. R. and D. A. Lowry, “The use of model output statistics (MOS) in
multiple dependence,” Insurance Mathematics & Economics 44 (2): 182–198, 2009. objective weather forecasting,” Journal of Applied Meteorology, 11: 1203–1211, 1972.
Alarcon, L. and D. H. Marks, A Stochastic Dynamic Programming Model for Gneiting, T., F. Balabdaoui, and A. Westveld, “Verifying probabilistic forecasts:
the Operation of the High Aswan Dam, Rep. TR. 246, Ralph M. Pearson calibration and sharpness,” Proceeding of the Workshop on Ensemble Weather
Laboratory for Water Resources and Hydrodynamics, Department of Civil Forecasting in the Short to Medium Range, Val-Morin, QC, Canada, 2003, www
Engineering, Massachusetts Institute of Technology, Cambridge, MA, 1979. .cdc.noaa.gov/people/tom.hamill/ef_workshop_2003_ schedule.html.
Alvisi S., A. Bernini, and M. Franchini, “Grey neural networks for estimating Gneiting, T., A. E. Raftery, A. H. III Westveld, and T. Goldman, “Calibrated
the uncertainty band of a real time forecasting model,” Proceedings of the BHS probabilistic forecasting using ensemble model output statistics and mini-
Third International Symposium on Role of Hydrology in Managing Consequences mum CRPS estimation,” Monthly Weather Review, 133 (5): 1098–1118, 2005.
of a Changing Global Environment,” Newcastle upon Tyne, July, 2010. Gneiting, T., F. Balabdoui, and A. E. Raftery, “Probabilistic forecasts, cali-
Anderson, J. L., “A method for producing and evaluating probabilistic fore- bration and sharpness,” Journal of the Royal Statistical Society Series B:
casts from ensemble model integrations,” Journal of Climate, 9: 1518–1525, Statistical Methodology, 69: 243–268, 2007.
1996. Gneiting, T., L. I. Stanberry, E. P. Grimit, L. Held, and N. A. Johnson,
Berger, J. O., Statistical Decision Theory and Bayesian Analysis, 2nd ed., Assessing Probabilistic Forecasts of Multivariate Quantities, with an Application
Springer-Verlag, New York, 1985, ISBN 0-387-96098-8. MR 0804611. to Ensemble Predictions of Surface Winds, Technical Report no. 537, Department
Bernardo, J. M. and A. F. M. Smith, Bayesian Theory. John Wiley & Sons, of Statistics, University of Washington, 2008.
Chichester, UK, 1994, ISBN 0-471-92416-4. MR 1274699. Hamill, T. M. and S. J. Colucci, “Verification of Eta-RSM short-range
Beven, K. J. and A. M. Binley, “The future of distributed models: model ensemble forecasts,” Monthly Weather Review, 125: 1312–1327, 1997.
calibration and uncertainty prediction,” Hydrological Processes, 6: 279–298, Hamill, T. M., “Interpretation of rank histograms for verifying ensemble
1992. forecasts,” Monthly Weather Review, 129: 550–560, 2001.
Beven, K. J. and J. Freer, “Equifinalilty, data assimilation, and uncertainty Hersbach, H., “Decomposition of the continuous ranked probability score
estimation in mechanistic modeling of complex environmental systems using for ensemble prediction systems,” Weather Forecast, 15, 559–570, 2000.
the GLUE methodology,” Journal of Hydrology, 249: 11–29, 2001. Kalman, R. E., “A new approach to linear filtering and prediction prob-
Bickel, P. J. and E. J. Lehmann, “Descriptive statistics for nonparametric lems,” Journal of Basic Engineering, 82 (1): 35–45, 1960, doi:10.1115/1.3662552.
models IV. Spread,” Contributions to Statistics, edited by J. Jureckova, Kelly, K. S. and R. Krzysztofowicz, “A bivariate meta-Gaussian density for
Academia, Prague, 1979, pp. 33–40. use in hydrology,” Stochastic Hydrology and Hydraulics, 11: 17–31, 1997.
Bogner, K. and F. Pappenberger, “Multiscale error analysis, correction, and Kelly, K. S. and R. Krzysztofowicz, “Precipitation uncertainty processor
predictive uncertainty estimation in a flood forecasting system,” Water for probabilistic river stage forecasting,” Water Resources Research, 36 (9):
Resources Research, 47: 07524, 2011, doi:10.1029/2010WR009137. 2643–2653, 2000
Bogner, K., F. Pappenberger, and H. L. Cloke, “Technical note: the normal Koenker, R., Quantile Regression, Econometric Society Monographs,
quantile transformation and its application in a flood forecasting system,” Cambridge University Press, New York, 2005.
Hydrology and Earth System Sciences, 16: 1085–1094, 2012. Krzysztofowicz, R., “Bayesian theory of probabilistic forecasting via deter-
Box, G. E. P. and D. R. Cox, “An analysis of transformations,” Journal of the ministic hydrologic model,” Water Resources Research, 35: 2739–2750, 1999.
Royal Statistical Society: Series B, 26: 296–311, 1964. Krzysztofowicz, R. and K. S. Kelly, “Hydrologic uncertainty processor for
Bremnes, J. B., “Probabilistic forecasts of precipitation in terms of quantiles probabilistic river stage forecasting,” Water Resources Research, 36 (11):
using NWP model output,” Monthly Weather Review, 132: 338–347, 2003. 3265–3277, 2000.
Brown, J. D. and D.-J. Seo, “A nonparametric post-processor for bias- Krzysztofowicz, R. and H. D. Herr, “Hydrologic uncertainty processor for
correction of hydrometeorological and hydrologic ensemble forecasts,” probabilistic river stage forecasting: precipitation-dependent model,” Journal
Journal of Hydrometeorology, 11: 642–665, 2010. of Hydrology, 249: 46–68, 2001.
Clark, M., S. Gangopadhyay, L. Hay, B. Rajagopalan, and R. Wilby, “The Krzysztofowicz, R., “Bayesian system for probabilistic river stage forecast-
Schaake shuffle: a method for reconstructing space-time variability in fore- ing,” Journal of Hydrology, 268: 16–40, 2002.
casted precipitation and temperature fields,” Journal of Hydrometeorology, 5 Krzysztofowicz, R. and C. Maranzano, “Hydrologic uncertainty processor for
(1): 243–262, 2004. probabilistic stage transition forecasting,” Journal of Hydrology, 293: 57–73, 2004.
Cloke, H. L. and F. Pappenberger, “Ensemble flood forecasting: a review,” Laio, F. and S. Tamea, “Verification tools for probabilistic forecasts of con-
Journal of Hydrology 375 (3–4): 613–626, 2009. tinuous hydrological variables,” Hydrology and Earth System Sciences, 11:
Coccia, G. and E. Todini, “Recent developments in predictive uncertainty 1267–1277, 2007, http://dx.doi.org/10.5194/hess-11-1267-2007.
assessment based on the model conditional processor approach,” Hydrology Lehmann, E. L. and H. J. M. D’Abrera, Nonparametrics: Statistical Methods
and Earth System Sciences, 15 (10): 3253–3274, 2011. Based on Ranks, Prentice-Hall (rev. ed.), Englewood Cliffs, New Jersey, 1998.
Dawid, A. P., “Statistical theory: the prequential approach (with discus- Lerch, S. and T. L. Thorarinsdottir, “Comparison of non-homogeneous
sion),” Journal of the Royal Statistical Society, 174 (A): 278–292, 1984. regression models for probabilistic wind speed forecasting,” Tellus A, 65:
De Groot, M. Optimal Statistical Decisions, Wiley Classics Library 21206, 2013, http://dx.doi.org/10.3402/tellusa.v65i0.21206.
(Originally published 1970), 2004, ISBN 0-471-68029-X. Liu, Z., M. V. L. Martina, and E. Todini, “Flood forecasting using a fully
Diebold, F. X., T. A. Gunther, and A. S. Tay, “Evaluating density forecasts distributed model: application of the TOPKAPI model to the upper Xixian
with applications to financial risk management,” International Economic catchment,” Hydrology and Earth System Sciences, 9: 347–364, 2005.
Review, 39 (4): 863–883, 1998. Madadgar, S., H. Moradkhani, and D. Garen, “Towards improved reliability
Diomede, T., F. Nerozzi, T. Paccagnella, and E. Todini, “The use of meteo- and reduced uncertainty of hydrologic ensemble forecasts using multivariate
rological analogues to account for LAM QPF uncertainty,” Hydrology and post-processing,” Hydrological Processes, 15: 1133–1155, 2012, doi: 10.1002/
Earth System Sciences, 12: 141–157, 2008. hyp.9562.
26_Singh_ch26_p26.1-26.16.indd 15 8/22/16 1:45 PM

Mantovan, P. and E. Todini, “Hydrological forecasting uncertainty assess- Solomatine, D. P. and D. L. Shrestha, “A novel method to estimate model
ment: incoherence of the GLUE methodology,” Journal of Hydrology, 330: uncertainty using machine learning techniques,” Water Resources Research,
368–381, 2006. 45: W00B11, 2009, doi:10.1029/2008WR006839.
Mantovan, P., E. Todini, and M. L. V. Martina, “Reply to comment by Keith Talagrand, O., R. Vautard, and B. Strauss, “Evaluation of probabilistic pre-
Beven, Paul Smith and Jim Freer on ‘Hydrological forecasting uncertainty diction systems,” Proceedings of a Workshop Held at ECMWF on Predictability,
assessment: incoherence of the GLUE methodology,’” Journal of Hydrology, 338: 20–22 October 1997, European Centre for Medium-Range Weather Forecasts,
319–324, 2007. Reading, 1997, pp. 1–25.
Mardia, K. V., J. T. Kent, and J. M. Bibby, Multivariate Analysis. Probability Todini, E., “Using phase-space modelling for inferring forecasting uncer-
and Mathematical Statistics, Academic Press, London, 1979. tainty in non-linear stochastic decision schemes,” Journal of Hydroinformatics,
Matheson, J. E. and R. L. Winkler, “Scoring rule for continuous probability 01.2: 75–82, 1999.
distribution,” Management Science, 22: 1087–1095, 1976. Todini E., “Role and treatment of uncertainty in real-time flood forecasting,”
Montanari A. and A. Brath, “A stochastic approach for assessing the uncer- Hydrological Processes, 18 (14): 2743–2746, 2004.
tainty of rainfall-runoff simulations,” Water Resources Research, 40, W01106, Todini, E., “Rainfall-runoff models for real-time forecasting,” Encyclopedia
2004, doi: 10.1029/2003WR002540. of Hydrological Sciences, Part 11, Paper 123, John Wiley & Sons, Chichester,
Nelsen, R. B., An Introduction to Copulas, Springer, New York, 1999. UK, 2006, Online ISBN: 9780470848944, doi: 10.1002/0470848944.
Nicholson, N., E. Soane, M. Fenton-O’Creevy, and P. Willman, “Personality Todini, E., “A model conditional processor to assess predictive uncertainty
and domain specific risk taking,” Journal of Risk Research, 8 (2):157–176, 2005, in flood forecasting,” International Journal of River Basin Management, 6 (2):
ISSN 1366-9877. 123–137, 2008.
Oja H., “Descriptive statistics for multivariate distributions,” Statistics & Todini, E., “History and perspectives of hydrological catchment modeling,”
Probability Letters, 1: 327–332, 1983. Hydrology Research, 42 (2–3): 73–85, 2011.
Pappenberger, F., L. Stephens, S. J. van Andel, J. S. Verkade, M. H. Ramos, L. Todini, E., “From HUP to MCP: analogies and extended performances,”
Alfieri, J. D. Brown, et al., “Operational HEPS systems around the globe,” HEPEX- Journal of Hydrology, 477: 33–42, 2013.
Blog, 2013, http://hepex.irstea.fr/operational-heps-systems-around-the-globe/ Todini, E., “The role of predictive uncertainty in the operational manage-
Raftery, A. E., “Bayesian model selection in structural equation models,” ment of reservoirs,” Proceedings of ICWRS2014 Evolving Water Resources
Testing Structural Equation Models, edited by K. A. Bollen and J. S. Long, Sage, Systems: Understanding, Predicting and Managing Water–Society Interactions,
Newbury Park, CA, 1993, pp. 163–180. IAHS Publication no. 36X, Bologna, Italy, 2014.
Raftery, A. E., T. Gneiting, F. Balabdaoui, and M. Polakowski, “Using Van der Waerden, B. L., “Order tests for two-sample problem and their
Bayesian model averaging to calibrate forecast ensembles,” Monthly Weather power I,” Indagationes Mathematicae, 14: 453–458, 1952.
Review, 133 (5): 1155–1174, 2005. Van der Waerden, B. L., “Order tests for two-sample problem and their
Reggiani, P. and A. Weerts, “A Bayesian approach to decision-making power II,” Indagationes Mathematicae, 15: 303–310, 1953a.
under uncertainty: an application to real-time forecasting in the river Rhine,” Van der Waerden, B. L., “Order tests for two-sample problem and their
Journal of Hydrology, 356: 56–59, 2008, doi:10.1016/j.jhydrol.2008.03.027. power III,” Indagationes Mathematicae, 15: 311–316, 1953b.
Reggiani, P., M. Renner, A. Weerts, and P. Van Gelder, “Uncertainty assess- Vrugt, J. A. and B. A. Robinson, “Treatment of uncertainty using ensemble
ment via Bayesian revision of ensemble streamflow predictions in the opera- methods: Comparison of sequential data assimilation and Bayesian model
tional river Rhine forecasting system,” Water Resources Research, 45: W02428, averaging,” Water Resources Research, 43: W01411, 2007, doi:10.1029/
2009, doi:10.1029/2007WR006758. 2005WR004838.
Regonda, S. K., D-J. Seo, B. Lawrence, J. D. Brown, and J. Demargne, Wang, Q. J. and D. E. Robertson, “Multisite probabilistic forecasting of
“Short-term ensemble streamflow forecasting using operationally-produced seasonal flows for streams with zero value occurrences,” Water Resources
single-valued streamflow forecasts – a Hydrologic Model Output Statistics Research, 47: W02546, 2011, doi:10.1029/2010WR009333.
(HMOS) approach,” Journal of Hydrology, 497: 80–96, 2013, doi:10.1016/ Wang, Q. J., D. L. Shrestha, D. E. Robertson, and P. Pokhrel, “A log-sinh
j.jhydrol.2013.05.028. transformation for data normalization and variance stabilization,” Water
Shaked, M. and J. G. Shanthikumar, Stochastic Orders and Their Applications, Resources Research, 48: W05514, 2012 doi:10.1029/2011WR010973.
Academic Press, Boston, 1994. Weerts, A. H., H. C. Winsemius, and J. S. Verkade, “Estimation of predic-
Schefzik, R., T. L. Thorarinsdottir, and T. Gneiting, “Uncertainty quantifi- tive hydrological uncertainty using quantile regression: examples from the
cation in complex simulation models using ensemble copula coupling,” National Flood Forecasting System (England and Wales),” Hydrology and
Statistical Science, Statistical Science, 28 (4): 616–640, 2013. Earth System Sciences, 15 (1): 255–265, 2011.
Sklar, A., “Fonctions de reparttion à n dimensions et leurs marges (in Wilks, S. S., “Order statistics,” Bulletin of the American Mathematical
French),” Publications de l’Institut de statistique de l’Université de Paris, 8: Society, 54 (1–1): 6–50, 1948.
220–231, 1959. Wilks, D. S., Statistical Methods in the Atmospheric Sciences: An Introduction,
Scheuerer, M., “Probabilistic quantitative precipitation forecasting using Academic Press, Boston, 1995, p.467.
Ensemble Model Output Statistics,” Quarterly Journal of the Royal Meteorological Yakowitz, S., “Nearest neighbour methods for time series analysis,” Journal
Society, 140 (680): 1086–1096, 2014. of Time Series Analysis, 8 (2): 235–247, 1987.
Schölzel, C. and P. Friederichs, “Multivariate non-normally distributed Zhao, L., Q. Duan, J. Schaake, A. Ye, and J. Xia, “A hydro-logic post-processor
random variables in climate research – Introduction to copulas,” Nonlinear for ensemble streamflow predictions,” Advances in Geoscience, 29: 51–59,
Processes Geophysics, 15: 761–772, 2008. 2011.
26_Singh_ch26_p26.1-26.16.indd 16 8/22/16 1:45 PM

Chapter 27
Risk-Reliability Analysis
BY
YEOU-KOUNG TUNG AND LARRY W. MAYS
ABSTRACT 27.2 MEASURES OF RELIABILITY
Hydrosystems are generally designed with a certain stipulated capacity and In hydrosystems engineering design and analysis, loads usually arise from
are placed in natural environments to serve their intended functions and natural events (e.g., floods and storms) which occur randomly in time and in
purposes. The systems are often subject to the occurrences of various loads space. The conventional practice for measuring the reliability of a hydrosys-
and stresses (e.g., floods, rains, waves) which tend to hinder their perfor- tem infrastructure is the return period or recurrence interval, which is the
mance. The capacity of a hydrosystem and the imposed external loads, more long-term average (or expected) time period between two successive failure-
often than not, are random and subject to uncertainty. Therefore, evaluation causing events. An equivalent term is the mean-time-to-failure used in time-
of the reliability of a hydrosystem requires a study between the interaction of to-failure analysis described in Sec. 27.5. The main disadvantage in the
load and resistance. This chapter provides a summary of basic concepts and conventional use of return period is that reliability is measured only in terms
issues in reliability analysis. Several methods that are useful for reliability of expected time of occurrence of loads without considering their interactions
assessment and risk-based design of hydrosystems are described. with the resistance (Melchers, 1999). The conventional interpretation of
return period can be generalized as the average time period to failure of the
27.1 INTRODUCTION
system when all uncertainties affecting load and resistance are considered.
That is, the return period can be calculated as the reciprocal of the failure
Practically, all hydrosystems are designed to serve in natural environments probability computed by Eq. (27.2).
that are subject to various external stresses. The resistance, strength, capacity, Two other types of reliability measures that consider the relative magni-
or supply of a hydrosystem is its ability to accomplish the intended mission tudes of resistance and anticipated load (called design load) are frequently
satisfactorily without failure when subjected to external stresses or loads. used in engineering practice. One is the safety margin (SM) defined as the
Loads, stresses, or demands tend to cause failure of the system. When the difference between the resistance and the anticipated load, that is,
strength of the system is exceeded by the stress, failure occurs. The resistance
of a hydrosystem can be the flow carrying capacity, while the load could be SM = R - L (27.4)
the magnitude of flow passing through the system. The resistance of the sys- The other measure is called the safety factor (SF), a ratio of resistance to load
tem and the imposed loads, more often than not, are random and subject to defined as:
uncertainty. The evaluation of system reliability is to study the interaction of
load and resistance. SF = R/L (27.5)
Considering the randomness of resistance (R) and load (L), the reliability Several types of SFs are summarized in Table 27.1 and their applications to
of a hydrosystem is defined as the probability that the system capacity (i.e., the engineering systems are discussed by Yen (1979).
resistance) exceeds or equals the load, that is, the probability the system is
safe. Mathematically, the reliability of a hydrosystem can be expressed as:
ps = Pr( L ≤ R ) (27.1) Table 27.1 Different Types of SFs (after Yen, 1979)
Type of SFs Definition 1 Definition 2
in which Pr(•) represents the probability. On the other hand, the early defini-
tion of risk is the probability of the load exceeding the resistance Pr( L > R ), Preassigned Assigned number
which indicates the failure probability of the system. The relationship Central µR µL mR and mL are the true mean values of resistance and
between reliability (ps) and failure probability (pf) is load
pf = Pr( L > R ) = 1 − ps (27.2) Mean R L R and L are the mean values of resistance and load
estimated from the available data
Today, the term “risk” in risk analysis is more often defined as the product Characteristic Ro Lo Ro and Lo are the specified resistance and load
of the amount that may be lost and the probability of losing it which can be
Partial l/g = NL/NR pf = Pr(L > g R) = Pr(NL L > NR R)
expressed as:
Risk = Failure probability × Failure consequences (27.3)
Therefore, the determination of risk requires the evaluation of failure prob- From Eq. (27.1), the reliability of a system can be expressed in terms of the
ability as well as the consequences associated with the failure which some- SM as:
times indicates the system vulnerability. ps = Pr( R − L ≥ 0) = Pr(SM ≥ 0) (27.6)
Equations (27.1) and (27.2) are referred to as a static reliability model
which does not consider the time dependence of the resistances and/or loads. Using the safety margin for computing reliability requires knowing the prob-
They are generally applied for the system performance subject to a single ability distribution of SM. If that is the case, the reliability can be obtained by
worst load event. ps = 1 - FSM(0) in which FSM(•) is the cumulative distribution function (CDF)
27-1
27_Singh_ch27_p27.1-27.10.indd 1 8/22/16 1:47 PM

27-2 Risk-Reliability Analysis
of the random safety margin SM. Under some special cases, the distribution xk
of SM can be easily assessed without mathematical manipulations.
Although exact solutions for reliability computation are desirable, it might
not be practical because the exact probability density function (PDF) of the
Q(x) < 0
SM may not be easily derived. In such circumstances, the use of the normal
approximation based on the mean and variance of the SM might be a viable Failure region
alternative. Ang (1973) indicated that, provided ps < 0.99, reliability is not
greatly influenced by the choice of distribution for L and R and the assump-
tion of a normal distribution for SM is quite satisfactory. However, for a very
low failure probability (e.g., pf < 10–3), the shape of the tails of the distribution
then become very critical in which case accurate assessment of the distribu-
tion of SM or a direct integration procedure should be used to evaluate the
failure probability.
Q(x) > 0
The SF as defined in Eq. (27.5) is the ratio of resistance to load, R/L. The
Safe region
reliability then can be written as Pr(SF ≥ 1). Similar to the SM, reliability com-
putations using the safety factor requires knowing the PDF of SF. The simplest
case is that when both load L and resistance R have lognormal distributions.
The logarithmic transform of SF leads to the difference of ln(R) and ln(L)
which are both normally distributed. The reliability computation can then
Limit-state
proceed as for the SM case
function
R 
ps = Pr  ≥ 1 = Pr(SF ≥ 1) = Pr[ln(SF) ≥ 0] = Pr[ln( R ) − ln( L ) ≥ 0] (27.7)
L  xj
0
After some algebraic manipulations, Eq. (27.7) can also be expressed in terms
of the statistical properties of L and R directly as (Chow et al., 1988)
  2   Figure 27.1 System states defined by performance (limit-state) function.

 − ln  µ R 1 + Ω L  
2
  µ L 1 + Ω R  
p f = 1− Φ   (27.8) in which FQ(⋅) is the CDF of the performance function Q; and Q′ is the stan-
 ln (1 + Ω2L )(1 + Ω2R )  dardized performance function defined as Q′ = (Q - mQ)/sQ with mQ and sQ
being the mean and standard deviation of the performance function, respec-
where ΩL and ΩR are the coefficients of variation of load and resistance, tively; and b = mQ/sQ is frequently called the reliability index. The reliability
respectively. index was first introduced by Cornell (1969) and later formalized by Ang and
Cornell (1974).
27.3 PERFORMANCE FUNCTION AND RELIABILITY INDEX
The expression of reliability, ps, for some commonly used distributions of
Q(X) can be found in Table 27.1 of Tung et al. (2006). Referring to Fig. 27.2,
In hydrosystem engineering problems, the resistance and load defining the it can be seen that bsQ is the distance of the mean of the performance func-
performance function are frequently functions of several stochastic basic tion away from the critical point Q = 0 beyond which the system failure would
variables, that is, L = g(XL) = g(X1,X2,...,Xm) and R = h(XR) = h(Xm+1,Xm+2,...,XK), occur (i.e., Q < 0). In practice, the normal distribution is commonly used for
where XL and XR are vectors of stochastic basic variables defining the load Q(X), in which case the reliability can be simply computed as:
function, g(XL), and the resistance function, h(XR), respectively. Reliability
can be expressed in terms of a performance function which describes the state ps = 1 − Φ(−β ) = Φ(β ) (27.14)
of system performance. The system performance Q(XL, XR) can be described
in terms of the load L = g(XL) and the resistance R = h(XR) as one of the fol- For practically all probability distributions used in the reliability analysis, the
lowing forms: value of the reliability, ps, increases with the reliability index, b.

 R − L = h( X R ) − g ( X L ) = SM(X ) (27.9)
 fQ(q)
Q( X ) = Q ( X R , X L ) =  ( R /L ) − 1 = [h( X R ) g ( X L )] − 1 = SF(X ) − 1 (27.10)
 Safe region
 ln( R /L ) = ln[h( X R )] − ln[ g ( X L )] = ln[SF(X )] (27.11) Failure

region
where X = (XL, XR) is a vector containing a total list of uncertain parameters
used in defining the system performance function. Equation (27.9) is identi- bsQ
cal to the SM, while Eqs. (27.10) and (27.11) are based on the SF representa-
tion. Therefore, the parameter space defining the performance function, as 0 mQ q
shown in Fig. 27.1, can be divided into different regions:
Figure 27.2 Probability distribution of performance function Q.
>0, system performance is satisfactory (or safe region)

Q(x ) =0, limit-state surface (or failure surface)
 27.4 DIRECT INTEGRATION METHOD
<0, system performance is unsatisfactory (or failure region)
The computation of failure probability via Eq. (27.2), under the condition of
known joint PDF of the load and resistance, fL,R(l,r), can be expressed as:
In terms of the performance function, the failure probability is the proba-
bility that the performance function value is less than zero, that is, ∞ ∞ 
p f = Pr[Q( X ) < 0] (27.12)
pf = ∫  ∫ f L ,R (l , r ) dl dr (27.15a)
0 r 
Assuming that the probability distribution of the performance variable, fQ(q), or

is known or derived, the failure probability can be expressed as: ∞ l 
pf = ∫  ∫ f L ,R (l , r ) dr  dl (27.15b)
p f = Pr(Q < 0) = FQ (0) = FQ′ (−µQ /σ Q ) = FQ′ (−β ) (27.13) 0 0 
27_Singh_ch27_p27.1-27.10.indd 2 8/22/16 1:47 PM

First-Order Second-Moment Reliability Methods 27-3
If the load L and resistance R are statistically independent, Eqs. (27.15a) 1×100
and (27.15b) for failure probability can be written respectively as:
Failure probability, P(R<L)

∞ ∞  ∞
pf = ∫ f R (r )  ∫ f L (l ) dl dr = ∫ f R (r )[1 − FL (r )]dr = 1 − ER[FL ( R)] (27.16a) 1×10–1

0  r  0
or 1×10–2
∞ l  ∞
pf = ∫ f L (l )  ∫ f R (r )d r  dl =
 0 
∫ f L (l )FR (l ) dl = EL[FR ( L )] (27.16b) COV(R) = 0.0 COV(R) = 0.05 COV(R) = 0.1
0 0 1×10–3
COV(R) = 0.3 COV(R) = 0.5
in which fL(•) and fR(•) are the marginal PDF’s of load, L, and resistance, R,
respectively; FL(•) and FR(•) are, respectively, the CDFs of the load, L, and
1×10–4
resistance, R; and E[•] is the expectation operator. This type of reliability
0 50 100 150 200
computation is called load-resistance interference, which can be schemati-
cally shown in Fig. 27.3. The computation steps for the determination of Design hydrologic return period (years)
failure probability using Eq. (27.16b) are shown in Fig. 27.3. The method of Figure 27.4 Effect of resistance uncertainty on failure probability under COV(L) = 0.1.
direct integration for failure probability computation using Eqs. (27.15a),
(27.15b), (27.16a), and (27.16b) in general is analytically tractable only for a
very few special combinations of probability distributions. In general, numer- flood (load) and random flow carrying capacity (resistance) are independent
ical integrations are implemented. log-normal random variables. The figure clearly shows that by considering
only inherent randomness of hydrologic load [the bottom curve correspond-
ing to the coefficient of variation (COV), COV(R) = 0] the annual failure
probability is significantly underestimated as the uncertainty of resistance
COV(R) increases. In reality, the inherent natural randomness of hydrologic
fR(r) and fL(l)fL(l)
fR(r)
processes is only one of the many uncertainties in hydrosystem design and
management. Figure 27.4 clearly demonstrates the deficiency of the conven-
tional frequency analysis approach in reliability assessment of hydrosystems.
Cornell (1972) said it well that “It is important in engineering applications that
we avoid the tendency to model only those probabilistic aspects that we think we
know how to analyze. It is far better to have an approximate model of the whole
problem than an exact model of only a portion of it.”
l, r
27.5 FIRST-ORDER SECOND-MOMENT RELIABILITY METHODS
Evaluation of the reliability or failure probability of a hydrosystem, as

FR(l) and fL(l)fL(l)
shown in Eqs. (27.13) and (27.14), requires the estimation (or derivation) of
FR(l)
the probability distribution function of the performance function and its
statistical properties, such as the mean and standard deviation. In real-life
hydrosystems, the performance function Q(X) [e.g., Eqs. (27.9)–(27.11)] is
often too complex to analytically derive its statistical features. Therefore,
approximated approaches are used to estimate the statistical properties of
the performance function. This section briefly summarizes two practical
first-order second-moment (FOSM) methods that can be used for reliability
l evaluation of hydrosystem engineering problems. Both methods are based
on the Taylor series expansion of the performance function. The mean-
value FOSM (MFOSM) takes the expansion point at the means of stochastic
Area =
basic variables, whereas the advanced FOSM (AFOSM) takes an expansion
Failure probability point located on the limit-state function, Q(X) = 0.
FR(l) × fL(l)
The Taylor series expansion of the performance function Q(X) with respect
to a selected point of stochastic basic variables X = xo in the parameter space
can be expressed as:
K 
∂Q 
Q( X ) = qo + ∑   ( X k − x ko )
k=1  ∂ X k x
o
(27.18)
K K 
l 1 ∂2 Q 
+ ∑∑  ( Xi − x io) ( X j − x jo) + ε
Figure 27.3 Graphical illustration of computation steps for failure probability by 2 i=1 j=1  ∂ Xi ∂ X j 
xo
Eq. (27.16b).
in which K is the total number of stochastic basic variables; and e represents
In conventional infrastructural reliability analysis for hydrosystems, the the higher-order terms. The first-order partial derivative terms are called the
inherent randomness of geophysical events (e.g., floods, rainstorms, etc.) is sensitivity coefficients, each representing the rate of change of the perfor-
often the only uncertainty considered. The uncertainties associated with the mance function value Q with respect to the unit change of the corresponding
resistance of the hydraulic flow carrying capacity are largely ignored. Under variable at the expansion point xo.
such circumstances, the aforementioned mathematical expressions for failure Dropping the higher-order terms represented by e, Eq. (27.18) is a second-
probability then can be reduced to order approximation of the performance function Q(X). Further truncating
the second-order terms from Eq. (27.18) leads to the first-order approxima-
pf = P [L > r*] (27.17) tion of Q(X) as:
K 
in which the resistance R = r* is the designated value of resistance—a deter- ∂Q 
Q( X ) ≈ q o + ∑   ( X k − x ko ) (27.19)
ministic quantity. By considering inherent randomness of annual maximum  ∂ X k x
k=1
floods, the annual failure probability, pf, for a hydraulic structure designed o
with a capacity to accommodate a T-year flood, that is, r* = lT, is 1/T. or in a matrix form as:
Figure 27.4 shows the effect of hydraulic uncertainty on the overall
annual failure probability under the assumption that both annual maximum Q(X) = wo + sot . (X - xo) (27.20)
27_Singh_ch27_p27.1-27.10.indd 3 8/22/16 1:47 PM

where so = ∇xQ(xo) is the column vector of sensitivity coefficients with each ele- xk
ment representing ∂ Q/∂ Xk evaluated at X = xo. The mean and variance of Q(X)
obtained by the first-order approximation can be expressed, respectively, as:
K 
Q(x) < 0
∂Q  x*
E [Q ] ≈ q o + ∑   (µ k − x ko ) (27.21)
Failure region
k=1  ∂ X k x Q(x) > 0
o
Safe region
and
  K K  mk
∂Q   ∂Q 
K 
∂Q 
Var[Q] ≈ Var q o + ∑   ( X k − x ko ) = ∑∑     Cov( X k ,X j )
 k=1 
∂ X k x o 
 k=1 j=1  X k x o  ∂ X j x
∂
o
xj
(27.22) 0 mj
Q(x) = 0
in which Cov(Xk, Xj) represents the covariance of random variables Xk and Xj. Limit-state surface
In matrix forms, Eqs. (27.21) and (27.22) can be expressed as:
Figure 27.5 Expansion point for the AFOSM method.
E[Q] ≈ qo + sot . ( lx - xo) (27.23)
and x* = (xL*, xR*) for the first-order Taylor series is located on the failure surface
Var[Q] ≈ sot Cx so (27.24) defined by the limit-state equation, Q(x*) = 0.
As can be seen in Fig. 27.5, the number of possible such expansion points,
x*, satisfying Q(x*) = 0 is infinite. At each expansion point on the limit-state
in which lx and Cx are the column vector of the means and covariance matrix
surface, the corresponding estimated values of the mean, variance, and
of the stochastic basic variables X, respectively.
reliability index of the performance function can be obtained, according to
27.5.1 Mean-Value First-Order Second-Moment Eqs. (27.23), (27.24), and (27.27), by
Reliability Method
µQ∗ ≈ s ∗t ( µ x − x ∗ ) (27.29)
The mean-value first-order second-moment (MFOSM) method estimates the
mean (mQ) and variance (σ Q2 ) of the performance variable Q by Eqs. (27.21)
and (27.22) [or by (27.23) and (27.24)], respectively, in which the expansion
point is chosen at the means of the stochastic variables, that is, xo= lx. The σ Q2 ∗ ≈ s ∗t C x s ∗ (27.30)
resulting mean and variance of random Q can be estimated as:
µ Q , MFOSM ≈ Q(µ x ) (27.25) b* = mQ*/sQ* (27.31)
and in which quantities in Eqs. (27.29)–(27.31) with subscript “*” signify that they
K K −1 K are dependent on the location of the expansion point x*. Therefore, the value
σ Q2 , MFOSM ≈ ∑ sk2 σ k2 +2∑ ∑ ρkj (skσ k )(s jσ j ) (27.26) of estimated reliability would change with the expansion point x* on the
k=1 k =1 j =k+1 limit-state surface.
From the design viewpoint, one is more concerned with the combination of
where ρkj is the linear correlation coefficient between basic variables Xk and Xj. stochastic basic variables that would yield the lowest reliability or highest
Hence, the reliability index b is computed as: failure probability. The point on the failure surface associated with the lowest
b MFOSM = mQ, MFOSM/sQ, MFOSM (27.27) reliability is the one having the shortest distance to the point, where the means
of the stochastic basic variables are located. This point is called the design
point (Hasofer and Lind, 1974) or the most probable failure point (Shinozuka,
And, by Eq. (27.13), the reliability can be computed as: 1983). The task of the AFOSM method is to identify such a design point by
solving the following optimization model:
ps = Pr(Q ≥ 0) = 1 - FQ′ (-bMFOSM) (27.28)
s t∗ ( µ x − x ∗ )
Minimize β∗ = (27.32a)
(st∗ C x s∗ )
0.5
in which FQ(•) is the CDF of the performance variable Q; and Q′ is the stan-
dardized performance variable. Adopting normal distribution for Q the reli-
ability can be simply computed by Eq. (27.14).
Although the MFOSM method is simple and straightforward in use, it, Subject to Q(x*) = 0 (27.32b)
however, possesses some weaknesses which include: (1) inappropriateness of
the expansion point; (2) inability to handle distributions with a large skew Several iterative algorithms have been developed for determining the design
coefficient; (3) generally poor estimations of the mean and variance of nonpoint. The AFOSM method also can account for correlated, non-normal
linear functions; (4) sensitivity of the computed failure probability to the random variables in the performance function. Detailed descriptions of these
formulation of the performance function; and (5) limited ability to use avail- algorithms can be found elsewhere (Tung et al., 2006).
able probabilistic information (Ang and Tang, 1984; Yen et al., 1986; Tung et al., The AFOSM method has been applied to various hydrosystem engineer-
2006). Nevertheless, the MFOSM method provides good estimation for reli- ing problems, including storm sewers (Melching and Yen, 1986), dams
ability if the performance function Q(X) is relatively linear and/or the vari- (Cheng et al., 1982, 1993), sea dikes and barriers (Vrijling 1987, 1993), free-
ability of the involved stochastic variables is small. To reduce the effect of board design (Cheng et al., 1986), bridge scour (Yen and Melching, 1991),
nonlinearity, one way is to include the second-order terms in a Taylor series rainfall-runoff modeling (Melching et al., 1990; Melching, 1992), groundwa-
expansion as shown in Eq. (27.18). This would increase the burden of analysis ter pollutant transport (Sitar et al., 1987; Jang et al., 1990), open-channel
by having to compute the second-order partial derivatives and higher-order design (Easa, 1992), sediment transport (Bechtler and Maurer, 1992), back-
statistical moments, which may not be so easily and reliably obtained for water computations (Cesare, 1991; Singh and Melching, 1993), and water-
practical problems. quality modeling (Tung, 1990; Melching and Anmangandla, 1992; Melching
and Yoon, 1996; Han et al., 2001).
27.5.2 Advanced First-Order Second-Moment Although the AFOSM reliability method can overcome many of the defi-
Reliability Method ciencies of the MFOSM method, its ability and accuracy are still limited when
The main thrust of the advanced first-order second-moment (AFOSM) the performance function is highly nonlinear and the required accuracy is high.
method is to improve the deficiencies associated with the MFOSM method, In such circumstances, the second-order reliability methods are required. The
while keeping the simplicity of the first-order approximation. Referring to second-order methods could produce more accurate reliability estimation than
Fig. 27.5, the difference in the AFOSM method is that the expansion point the first-order methods at the expense of more complicated computations.
27_Singh_ch27_p27.1-27.10.indd 4 8/22/16 1:47 PM

Time-to-Failure Analysis 27-5
27.6 TIME-DEPENDENT (DYNAMIC) RELIABILITY MODELS 27.7.1 Failure Density Function, Failure Rate, and
Mean Time to Failure
In some situations, reliability evaluation of a hydrosystems engineering and
management scheme is made with respect to a specified time framework. For Using the failure density function, the reliability of a system or a component
example, one might be interested in the overflow probability of an urban within a time interval [0, t] can be expressed as:
storm water detention basin in the summer season when convective thun- ∞
derstorms prevail. Loads to most hydrosystems are caused by the repeated ps (t ) = ∫ fT (t )dt (27.38)
occurrences of hydrological events (e.g., floods, storms, and droughts) which t
are random by nature. Dynamic or time-dependent reliability analysis consid-
ers repeated applications of load and also can consider the change in the which represents the probability that the system experiences no failure within
distribution of resistance with time. The objective of the reliability computa- [0, t]. The failure probability, based on Eq. (27.41), then can be expressed as:
tions for the dynamic models is to determine the system reliability over a t
specified time interval in which the number of occurrences of loads is also a
random variable.
p f (t ) = 1 − ps (t ) = ∫ fT (t )dt (27.39)
0
Repeated loads on a hydrosystem are characterized by the time each load
is applied and the behavior of time intervals between the applications of Schematically, the reliability and failure probability are shown in Fig. 27.6.
load. From a reliability theory viewpoint, the uncertainty about the load
and resistance variables may be classified into three categories: determinis-
tic, random-fixed, and random-independent (Kapur and Lamberson, 1977). fT (t)
For the deterministic category, the loads assume values that are exactly
known a priori. For the random-fixed case, the randomness of loads varies
in time in a known manner. For the random-independent case, the load is
not only random but the successive values assumed by the load are statisti-
cally independent.
Consider a hydrosystem with a capacity R = r subject to n repeated loads ps(t)
pf(t)
L1, L2,…, Ln. When the system capacity, r, is fixed, the system reliability after
Time
n loads, ps(n, r) is 0 t
ps (n, r ) = Pr [( L1 ≤ r ) ∩ ( L2 ≤ r ) ∩ ... ∩ ( Ln ≤ r )] = Pr [( Lmax ≤ r )] (27.33) Figure 27.6 Schematic diagram of reliability and unreliability for time-to-failure
analysis.
in which Lmax = max{L1, L2,…, Ln}, the maximum load, is also a random vari-
able. If all random loads are independent and each has its own distribution, The probability of failure of a hydrosystem in some time interval (t, t + Δt)
Eq. (27.33) is reduced to can be expressed in terms of the reliability as:
n t +∆t ∞ ∞
ps (n, r ) = FLi (r ) (27.34)
∫ fT (t ) dt = ∫ fT (t ) dt − ∫ fT (t ) dt = ps (t ) − ps (t + ∆t )
t t t +∆t
Reliability under the fixed number of loads n on a system with a random
resistance can be expressed as: which defines the difference in reliability at time t, ps(t), and at time t + Δt,
∞
ps(t + Δt). The rate at which a failure occurs in the time interval (t, t + Δt) is
called the failure rate which is the conditional probability that a failure per
ps (n) = ∫ ps (n, r ) f R (r )dr (27.35) unit time occurs in the time interval, given that no failure has occurred prior
0
to time t.
Since the loads onto many hydrosystems are related to hydrological events, A hazard function, the instantaneous failure rate, is the limit of the failure
the occurrence of the number of loads, in general, is uncertain. The reliability rate as Δt replace with 0, that is,
of the system under random load in the specified time interval [0, t] can be
expressed as:  p (t ) − ps (t + ∆t )  1  d [ ps (t )]  fT (t )
m(t ) = lim  s = − = (27.40)
∞ ∆t →0  ∆t ps (t )  ps (t )  dt  ps (t )
ps (t ) = ∑ π (t | n) ps (n) (27.36)
n=0
The hazard function indicates the change in the failure rate over the life of
in which π (t | n) is the probability of n loads occurring in the time interval a hydrosystem or its components. The quantity m(t)dt is the probability that
[0, t]. The Poisson distribution, which has been found to be an appropriate a component fails during the time interval (t, t + dt).
model for the number of occurrences of hydrological events (Todorovic and The practical usefulness of the hazard function m(t) is that the failure den-
Yevjevich, 1969; Zelenhasic, 1970; Rousselle, 1972; Fogel and Duckstein, sity function fT(t) and reliability function ps(t) can be derived directly from it.
1982), can be used. From Eq. (27.40), the reliability function ps(t) can be expressed in terms of the
By combining the poisson distribution with Eq. (27.36), the time-dependent hazard function m(t) by integrating both sides of d[ps(t)]/ps(t) = -m(t)dt. The
reliability for the random-independent load and random resistance can be result then is
expressed as:
 t 
∞ −ν t n ∞ ∞ ps (t ) = exp − ∫ m(t ) dt  (27.41)
e (ν t )  0 
ps (t ) = ∑ ∫ f R (r ) [ FL (r )]n dr = ∫ f R (r ) e−νt [1−F (r )] dr (27.37)
L
n=0
n! 0 0
Substituting Eq. (27.41) into Eq. (27.40), the failure density function, fT(t),
where ν is the mean rate of occurrence of the load in [0, t] which can be esti- can be expressed, in terms of hazard function m(t), as (Kapur and Lamberson,
mated from historical data. 1977):
 t 
27.7 TIME-TO-FAILURE ANALYSIS
fT (t ) = m(t ) exp − ∫ m(τ ) dτ  (27.42)
Instead of considering detailed interactions of resistance and load over time,  0 
a system or its components can be treated as a black box and their perfor-
mance observed over time. This reduces the reliability analysis to a one- For an exponential distribution f T (t ) = λ e− λt , the corresponding reliability
dimensional problem involving only time as the random variable. In such
cases, time-to-failure (T) of a system or component of the system is the ran- f (t )
function is ps (t ) = e− λt . Then the hazard function is m(t ) = T = λ which is
dom variable with the PDF, fT(t), called the failure density function. Time- ps (t )
to-failure analysis is particularly suitable for assessing the reliability of a constant. The relationships of fT(t), ps(t), and m(t) for other types of failure
systems involving components which are repairable. density functions, can be found elsewhere (Henley and Kumamoto, 1981).
27_Singh_ch27_p27.1-27.10.indd 5 8/22/16 1:47 PM

The mean time to failure (MTTF) is the expected value of the time to failure A (t)
stated mathematically as:
∞
MTTF = ∫ τ fT (τ ) dτ (27.43) 1.0
0 Repairable
component
which has a unit of time (e.g., minutes, hours, etc.).
27.7.2 Repair Density Function, Repair Rate, and
Mean Time to Repair
A (∞)
Similar to the failure density function, the repair density function, g(t), Nonrepairable
describes the random characteristics of the time required to repair a failed component
component when failure occurs at time zero. The probability of repair, G(t),
is the probability that the component repair is completed before time t, given
that the component failed at t = 0. In other words, G(t) is the CDF of the 0.0 Time, t
repair time having the repair density function g(t).
The repair rate, r(t), similar to the failure rate through the hazard func- Figure 27.7 Availability for repairable and nonrepairable components (or systems).
tion, is the probability that the component is repaired per unit time given
that the component failed at time zero and is still not repaired at time t. The
quantity r(t)dt is the probability that a component is repaired during the Using exponential failure and repair density functions, the resulting failure
time interval (t, t + dt) given that the component is in failed state at time t. rate and repair rate, according to the definitions given in Eqs. (27.40) and
Similar to Eq. (27.43) the relation between repair density function, g(t), (27.44), are constants equal to their respective parameters. For a constant
repair rate, r(t), and repair probability, G(t), is failure rate and a constant repair rate, the analytical solutions of the entire
process can be derived. Henley and Kumamoto (1981) used Laplace trans-
g T (t )
r (t ) = (27.44) forms to derive the unavailability and availability as:
[1 − G(t )]
λ [
which can be derived in the same manner or analogous to the failure rate and U (t ) = 1 − e( λ +η )t ] (27.50a)
λ +η
hazard function. Given a repair rate, r(t), the repair density function, g(t), and
the repair probability, G(t), are determined, respectively, as: η λ
A(t ) = 1 − U (t ) = + [e( λ +η )t ] (27.50b)
λ +η λ +η
 t 
g (t ) = r (t ) exp − ∫ r (τ ) dτ  (27.45) The steady state or stationary unavailability, U(∞), and the stationary avail-
 0  ability, A(∞), when t → ∞, are
 t  MTTR λ
U (∞) = = (27.51a)
G(t ) = 1 − exp − ∫ r (τ ) dτ  (27.46) MTTF + MTTR λ + η
 0 
The mean time to repair (MTTR) is the expected value of the time to repair a
MTTF η
A(∞) = = (27.51b)
failed component, defined as: MTTF + MTTR λ + η
∞
27.8 MONTE CARLO SIMULATION
MTTR = ∫ tg (t ) dt (27.47)
0 Due to the complexity of physical systems and mathematical functions
encountered in real-life hydrosystems design and management problems,
which has a unit of time. derivation of the exact solution for the probabilistic characteristics of the
The mean time between failure (MTBF) is the expected value of time system response is difficult, if not impossible. In such cases, Monte Carlo
between two consecutive failures. For a repairable component, the MTBF is simulation (MCS) is a viable tool to provide numerical estimations of the
the sum of MTTF and MTTR, that is, stochastic features of the system response.
MTBF = MTTF + MTTR (27.48) MCS is a numerical procedure to reproduce random variables that pre-
serves the specified distributional properties. In MCS, the system response of
The mean time between repair (MTBR) is equal to MTBF. interest is repeatedly measured under various system parameter input sets
generated from the known or assumed probabilistic laws. MCS offers a practi-
27.7.3 Availability and Unavailability
cal approach to the reliability analysis because the random behavior of the
The reliability of a component is a measure of the probability that the compo- system response can be probabilistically duplicated.
nent would be continuously functional without interruption through the Two major concerns in the practical applications of MCS in uncertainty
entire period [0, t]. This measure is appropriate if a component is nonrepair- and reliability analyses are: (1) the requirement of a large amount of computa-
able and has to be discarded when the component fails. In case that the com- tions for generating random variates, and (2) the presence of correlation
ponents in the system are repairable, a measure that has a broader meaning among stochastic basic parameters. However, as computing power increases,
than that of the reliability is needed. the concern with the computational cost diminishes, and MCS is becoming
The availability, A(t), of a component is the probability that the component more practical and viable for uncertainty and unreliability analyses. Beck
is in operating condition at time t, given that the component was as good as (1985) notes that “when the computing power is available, there can, in gen-
new at time zero. The reliability generally differs from the availability because eral, be no strong argument against the use of Monte Carlo simulation.”
reliability requires the continuation of the operational state over the specified As noted previously, the accuracy of the model output statistics and probabil-
interval [0, t]. Subcomponents contribute to the availability, but not to the ity distribution (e.g., probability that a specified safety level will be exceeded)
reliability, if the subcomponent that failed before time t is repaired and is then obtained from MCS is a function of the number of simulations performed. For
operational at time t. As shown in Fig. 27.7, the availability is always larger models or problems with a large number of stochastic basic variables and for
than or equal to the reliability, that is, A(t) ≥ ps(t). For a repairable component, which low probabilities (< 0.1) are of interest, tens of thousands of simulations
A(t) > ps(t), while for a nonrepairable component, A(t) = ps(t). may be required. Rules for determining the number of simulations required for
The unavailability, U(t), at time t is the probability that a component is in convergence are not available, and, thus, replication of the MCS runs for a given
the failed state at time t, given that it started in the operational state at time number of simulations is the only way to check convergence (Melching, 1995).
zero. Availability and unavailability of the component are complimentary
events, therefore, 27.8.1 Univariate Random Number Generation
A(t) + U(t) = 1 (27.49) Generation of a univariate random number from a given probability distribu-
tion function can be made by three types of methods: (1) CDF-inverse method,
Knowing A(t) ≥ ps(t), it can be concluded that U(t) ≤ pf (t). (2) acceptance-rejection method, and (3) variable transformation method.
27_Singh_ch27_p27.1-27.10.indd 6 8/22/16 1:47 PM

Monte Carlo Simulation 27-7
The CDF-inverse method is based on the relationship of u = Fx(xu) in which 27.8.2 Multivariate Random Number Generation
u is the probability and xu is the corresponding quantile. For the majority of It is not uncommon for hydrosystem engineering problems to involve multiple
continuous probability distributions applied in hydrosystems engineering and random variables that are correlated and statistically dependent. For example,
analysis, Fx(x) is a strictly increasing function of x. Hence, random variates many data show that the peak discharge and volume of a runoff hydrograph
having a known CDF, Fx(x), can be generated by xu = Fx−1 (u) with u being a are positively correlated. To simulate systems involving correlated random
standard uniform random variate defined over the unit interval [0,1], denoted variables, generated random variates must preserve the probabilistic character-
by ~U(0,1). To efficiently apply the CDF-inverse method for generating ran- istics of the variables and the correlation structure among them. Although
dom numbers, an explicit expression between X and U is essential so that X multivariate random number generation is an extension of the univariate case,
can be obtained analytically from the generated U as shown in Table 27.2. mathematical difficulty and complexity associated with multivariate problems
The basic idea of the acceptance-rejection method is to replace the original increase rapidly as the dimension of the problem gets larger. Compared with
fx(x) by an appropriate PDF, hx(x), from which random variates can be easily generating univariate random variates, multivariate random variate generation
and efficiently produced. The generated random variate from hx(x), then, is is much more restricted to fewer joint distributions, such as multivariate nor-
subject to testing before it is accepted as one from the original fx(x). This mal, multivariate log-normal, and multivariate gamma (Ronning, 1977;
approach for generating random numbers has become widely used. Johnson, 1987; Parrish, 1990). Nevertheless, the algorithms for generating
The variable transformation method generates a random variate of interest univariate random variates serve as the foundation for many multivariate
based on its known relationship with other random variables whose variates Monte Carlo algorithms.
can easily be produced. For example, generating Chi-square random variates The idea of the CDF-inverse method can be applied when the joint CDF of
with n degrees of freedom can be achieved by utilizing the fact that the sum of correlated random variables, Fx(x) = F1,2,...,K(x1, x2, ..., xK), is available and can
n squared independent standard normal random variables gives a Chi-square be decomposed into
random variable with n degrees of freedom. One could generate Chi-square
random variates by first producing n standard normal random variates, then
squaring them, and finally adding them together. Fx(x) = F1(x1) × F2(x2|x1) × ⋅⋅⋅ × FK(xK|x1, x2, ..., xK–1) (27.52)
Table 27.2 List of Distributions Whose CDF Inverse Are Analytically Expressible
Distribution Fx(x) = x = Fx-1(u)
Exponential 1-exp(-b x), x > 0 -b ln(1-F)

Uniform (x-a)/(b-a) a+(b-a)F
Gumbel exp{-exp[-(x-x)/b)]} x-b ln[-ln(F)]
Weibull 1-exp{-[(x-x)/b]a} x+b[-ln(1-F)]1/a
Pareto 1-x-a (1-F)-(1/a)
Wakeby Not explicitly defined x+(a/b)[1-(1-F)b]-(g /d )[1-(1-F)-d]
Kappa {1-h[1-a(x-x)/b]1/a}1/h x+(b/a){1-[(1-Fh)/h]a}
Burr 1-(1+xa)-b [(1-F)-1/b - 1]1/a
Cauchy 0.5+tan (x)/p
-1
tan[p (F-0.5)]
Rayleigh 1- exp[-(x-x)2/2b2] x+{-2b2 ln(1-F)}1/2
Generalized Lambda Not explicitly defined x+aFb-g (1-F)d
Generalized extreme value exp[-exp(-y)] x+b{1-[-ln(F)]a}/a, a ≠ 0;
where y = -a-1 ln{1-a(x-x)/b}, a ≠ 0; x-b ln[-ln(F)], a=0
= (x-x)/b, a=0
Generalized logistic 1/[1+exp(-y)] x+b{1-[(1-F)/F]a}/a, a ≠ 0;
where y = -a–1ln{1-a(x-x)/b}, a ≠ 0; x-b ln[(1-F)/F], a=0
= (x-x)/b, a=0
Generalized pareto 1-exp(–y) x+b[1-(1-F)a]/a, a ≠ 0;
where y = -a–1 ln{1-a(x-x)/b}, a ≠ 0; x-b ln[1-F], a=0
= (x-x)/b, a=0
To generate univariate random numbers from those distributions not in which F1(x1) and Fk(xk|x1, x2, ..., xk–1) are, respectively, the marginal CDF
efficiently amenable to the CDF-inverse method or variable transformation and the conditional CDF of random variables X1 and Xk. Based on Eq. (27.55),
method (e.g., normal, log-normal, and gamma distributions), the accep- the algorithm using the CDF-inverse method to generate n sets of K multi-
tance-rejection method is used or other efficient algorithms are developed. variate random variates from a specified joint distribution is as follows
For example, the Box–Muller algorithm (Box and Muller, 1958) and the (Rosenblatt, 1952):
Marsagalia–Bray algorithm (Marsagalia and Bray, 1964) were developed
for generating the standard normal variate Z~N(0,1). After a standard x1 = F1−1 (u1 )
normal random variate Z is generated, the corresponding normal variate
x 2 = F2−1 (u2 x1 )
X~N(mx,sx) with mean mx and standard deviation sx can be obtained as (27.53)
X = µ x + σ x Z . Log-normal variates can be generated by utilizing the fact •••
that the log-transform of a log-normal variable becomes normally distrib-
x K = FK−1 (uK x1 , x 2 ,...., x K −1 )
uted variable.
There are several books focusing on generating univariate random num-
bers (Rubinstein, 1981; Dagpunar, 1988; Gould and Tobochnik, 1988; Law In general, the availability of joint multivariate CDFs is restrictive. Many
and Kelton, 1991). To facilitate the implementation of MCS, computer sub- applications requiring generating multivariate random variables assume multi-
routines in different languages are available (Press et al., 1989, 1992, 2002; variate normal under which efficient algorithms have been developed. Consider
IMSL, 1980). In addition, several spreadsheet-based computer software (e.g., a random vector X =(X1, X2, ..., XK)t having a multivariate normal distribution
Microsoft Excel, @Risk, and Crystal Ball) contain statistical functions allow- with a mean vector mx and covariance matrix Cx, denoted as X ~ N(mx,Cx).
ing the generation of random variates of various distributions. To generate high dimensional multivariate normal random variates with
27_Singh_ch27_p27.1-27.10.indd 7 8/22/16 1:47 PM

specified mx and Cx, the CDF-inverse algorithm previously described might

not be efficient. Two efficient algorithms, namely, the square-root method
and the spectrum decomposition method, for generating multivariate normal
random variates are commonly used and both algorithms are based on Q(x) < 0,
orthogonal transformation of the covariance matrix Cx or correlation matrix failure region
Rx (Tung et al., 2006). The result of the transformation is a vector of indepen-
dent normal variables, which can easily be generated by the algorithms
described earlier.
In most hydrosystem engineering problems, random variables often are
statistically and physically dependent with a mixture of different marginal
distributions of which the corresponding joint PDF or CDF is difficult to
establish. As a practical alternative, to properly replicate such behavior, the Q(x) = 0
MCS method can be developed to preserve the correlation relationships
among the stochastic variables and their marginal distributions.
To generate correlated random variables with a mixture of marginal distri-
butions, a methodology adopting a bivariate distribution model was first
suggested by Li and Hammond (1975). The practicality of the approach was
advanced by Der Kiureghian and Liu (1985) who based their method on the
Nataf bivariate distribution model (Nataf, 1962). Chang et al. (1994) utilized
this set of formulas, which transforms the correlation coefficient of a pair of
non-normal random variables to its equivalent correlation coefficient in the
Q(x) > 0,
bivariate standard normal space, for multivariate simulation. Other practical
alternatives, such as the polynomial normal transformation (Vale and Safe region
Maurelli, 1983; Chen and Tung, 2003), can serve the same purpose. Through
a proper normal transformation, the multivariate MCS can be performed in a
correlated standard normal space in which efficient algorithms, such as those Figure 27.8 Determination of the failure probability by MSC with simple random
previously described, are available. sampling.
27.8.3 Monte Carlo Integration

In reliability analysis, computations of system and/or component reliability that “variance reduction cannot be obtained from nothing; it is merely a way
and other related quantities (e.g., mean time to failure) involve integration of not wasting information.” Therefore, for a problem that is not known at the
operations as shown in Eq. (27.43). For more complex problems, such as load- initial stage of the study, pilot simulations can be performed for the purpose
resistance interference, the reliability is determined as: of gaining useful insight about the problem. The insight, then, can later be
incorporated into the variance-reduction techniques for a more efficient
ps = Pr [Q( X ) ≥ 0] = ∫ f x ( x ) d x (27.54)
simulation study. Therefore, most of the variance-reduction techniques
Q ( x )≥0
require additional effort on the part of analysts. Several variance-reduction
where Q(X) is the performance function involving correlated stochastic basic techniques that are commonly used are listed as follows.
variables X. As can be seen, computation of reliability by Eq. (27.54) involves The importance sampling technique concentrates the distribution of sam-
multidimensional integration. pling points in the part of the domain that is most “important” for the task,
For cases of integration in one or two dimensions, such as Eq. (27.54), rather than spreading them out evenly (Marshall, 1956).
where the integrands are well behaved (e.g., no discontinuity), conventional The antithetic-variates technique (Hammersley and Morton, 1956) achieves
numerical integration methods, such as the trapezoidal approximation or the variance-reduction goal by attempting to generate random variates that
Simpson’s rule are efficient and accurate. Gould and Tobochnik (1988) show would induce a negative correlation for the quantity of interest between sepa-
that the accuracy of conventional numerical integration schemes decreases rate simulation runs. A simple approach to generate negatively correlated
rapidly as the dimensionality of integration increases. For multiple integrals, uniform random variates with minimal computation is to let U2 = 1 - U1
such as Eq. (27.54), the Monte Carlo method becomes a more suitable because Cov(U, 1 - U) = -1/12. As any random quantity in MCS is related to
numerical technique for integration. the uniform random number, the average of the random quantity of interest
For low-dimensional integration, simple Monte Carlo integration tech- corresponding to U and 1 - U will have reduced variance than any of the
niques, such as the hit-and-miss method and the sample-mean Monte Carlo individual ones.
integration would be sufficient with the latter method more accurate. For Correlated-sampling techniques are especially effective for variance reduc-
reliability computation involving a multidimensional integral as in Eq. (27.54) tion when the primary objective of the simulation study is to evaluate small
the directional simulation, based on the idea of conditional probability, is a changes in the system performance or to compare the difference in system
simple procedure to improve the efficiency of the Monte Carlo solution for performances between two specific designs (Rubinstein, 1981; Ang and Tang,
equation (27.54). Bjerager (1988) demonstrates the directional simulation 1984). In this problem, variance reduction can be achieved if the random
through several examples and shows that the CoV of estimated reliability p̂s samples of the two quantities have a positive correlation. One easy way to
for a given sample size depends on the shape of the failure surface and the obtain positively correlated samples is to use the same sequence of uniform
value of the unknown reliability. For non-spherical failure surfaces, the CoV random variates from U(0,1) in both simulations.
increases as the dimensionality of the problem K increases. The stratified sampling technique is a well-established area in statistical
sampling (Cochran, 1966). Variance reduction by the stratified sampling
27.8.4 Variance-Reduction Techniques technique is achieved by taking more samples in important subregions.
MCS is a sampling procedure, results obtained from the procedure inevita- The Latin hypercube sampling (LHS) technique is a special method under
bly involve sampling errors, which decrease as the sample size increases the umbrella of stratified sampling which selects random samples of each
(Fig. 27.8). Increasing the sample size to achieve a higher precision generally random variable over its range in a stratified manner. The LHS technique
means an increase in computer time for generating random variates and divides the plausible range of each random variable into M (M ≥ K in prac-
data processing. Variance-reduction techniques aim at obtaining high accu- tice) equal-probability intervals. Within each interval, a single random variate
racy for the MCS results without having to substantially increase the sample is generated resulting in M random variates for each random variable. McKay
size. Hence, variance-reduction techniques enhance the statistical efficiency (1988) suggested that the use of twice the number of involved random vari-
of the MCS. When applied properly, variance-reduction techniques some- ables for sample size (M ≥ 2K) would be sufficient to yield accurate estimation
times can make the difference between an impossible expensive simulation of the statistics model output. Iman and Helton (1988) indicated that a choice
study and a feasible, useful one. of M equal to 4/3K usually gives satisfactory results. In hydrosystems engi-
Variance-reduction techniques attempt to reduce the error associated with neering, the LHS technique has been widely applied to sediment transport
the MCS results by utilizing known information about the problem at hand. (Yeh and Tung, 1993; Chang et al., 1993), water quality modeling (Jaffe and
Naturally, such an objective cannot be attained if the analyst is completely Ferrara, 1984; Melching and Bauwens, 2001; Sohrabi et al., 2003; Manache
ignorant about the problem. On the other extreme, the error is zero if the and Melching, 2004), rainfall-runoff modeling (Melching, 1995; Yu et al.,
analyst has complete knowledge about the problem. Rubinstein (1981) states 2001; Christiaens and Feyen, 2002).
27_Singh_ch27_p27.1-27.10.indd 8 8/22/16 1:47 PM

References 27-9
REFERENCES Iman, R. L. and J. C. Helton, “A investigation of uncertainty and sensitivity

analysis techniques for computer models,” Risk Analysis, 8 (1): 71–90, 1988.
Ang, A. H-S. “Structural risk analysis and reliability-based design,” Journal
IMSL (International Mathematical and Statistical Library), IMSL Inc.,
of the Structural Engineering Division, ASCE, 99 (ST9): 1981–1910, 1973.
Houston, 1980.
Ang, A. H-S. and C. A. Cornell, “Reliability bases of structural safety and
Jaffe, P. R. and R. A. Ferrara, “Modeling sediment and water column inter-
design,” Journal of the Structural Engineering Division, ASCE, 100 (9): 1755–1769,
actions for hydrophobic pollutants, parameter discrimination and model
1974.
response to input uncertainty,” Water Research, 18 (9): 1169–1174, 1984.
Ang, A. H-S. and W. H. Tang, Probability Concepts in Engineering and Design,
Jang, Y. S., N. Sitar, and A. Der Kiureghian, “Reliability approach to proba-
Volume II: Decision, Risk, and Reliability, John Wiley & Sons, New York, 1984.
bilistic modelling of contaminant transport,” Report No. UCB/GT-90/3,
Bechtler, W. and M. Maurer, “Reliability theory applied to sediment trans-
Department of Civil Engineering, University of California, Berkeley, CA,
port formulae,” Proceedings of the 5th International Symposium on River
1990.
Sedimentation, Karlsruhe, Germany, 1992, pp. 311–317.
Johnson, M. E., Multivariate Statistical Simulation, John Wiley & Sons, New
Beck, M. B., “Water quality management: a review of the development and
York, 1987.
application of mathematical models,” Lecture Notes in Engineering 11, edited
Kapur, K. C. and L. R. Lamberson, Reliability in Engineering Designs, John
by C. A. Brebbia and S. A. Orszag, Springer-Verlag, New York, 1985.
Wiley & Sons, New York, 1977.
Bjerager, P., “Probability integration by directional simulation,” Journal of
Law, A. M. and W. D. Kelton, Simulation Modeling and Analysis, McGraw-
Engineering Mechanics, ASCE, 114 (8): 1285–1301, 1988.
Hill, New York, 1991.
Box, G. E. P. and M. E. Muller, “A note on generation of random normal
Li, S. T. and J. L. Hammond, “Generation of pseudorandom numbers with
deviates,” Annals of Mathematical Statistics, 29: 610–611, 1958.
specified univariate distributions and covariance matrix,” IEEE Transactions
Cesare, M. A., First-order analysis of open-channel flow, Journal of
on Systems, Man, and Cybernetics, 5: 557–561, 1975.
Hydraulic Engineering, ASCE, 117 (2): 242–247, 1991.
Manache, G. and C. S. Melching, “Sensitivity analysis of a water-quality
Chang, C. H., J. C. Yang, and Y. K. Tung, “Sensitivity and uncertainty
model using Latin hypercube sampling,” Journal of Water Resources Planning
analyses of a sediment transport model: a global approach,” Journal of
and Management, ASCE, 130 (3): 232–242, 2004.
Stochastic Hydrology and Hydraulics, 7 (4): 299–314, 1993.
Marsagalia, G. and T. A. Bray, “A convenient method for generating normal
Chang, C. H., Y. K. Tung, and J. C. Yang, “Monte Carlo simulation for cor-
variables,” SIAM Review, 6: 260–264, 1964.
related variables with marginal distributions,” Journal of Hydraulic Engineering,
Marshall, A.W., “The use of multi-stage sampling schemes in Monte Carlo
ASCE, 120 (2): 313–331, 1994.
computations,” Symposium on Monte Carlo Methods, edited by M. A. Meyer,
Chen, X. Y. and Y. K. Tung, “Investigation of polynomial normal transfor-
John Wiley & Sons, New York, 1956, pp. 123–140.
mation,” Journal of Structural Safety, 25: 423–445, 2003.
McKay, M. D., “Chapter 4: Sensitivity and uncertainty analysis using a sta-
Cheng, S. T., B. C. Yen, and W. H. Tang, “Overtopping risk for an existing
tistical sample of input values,” Uncertainty Analysis, edited by Y. Ronen, CRC
dam,” Hydraulic Engineering Series No. 37, Department of Civil Engineering,
Press, Boca Raton, FL, 1988, pp. 145–186.
University of Illinois, Urbana, IL, 1982.
Melchers, R. E., Structural Reliability: Analysis and Prediction, 2nd ed.,
Cheng, S. T., B. C. Yen, and W. H. Tang, “Wind induced overtopping risk of
John Wiley & Sons, New York, 1999.
dams,” Stochastic and Risk Analysis in Hydraulic Engineering, edited by B. C.
Melching, C. S., “An improved first-order reliability approach for assessing
Yen, Water Resources Publications, Littleton, CO, 1986, pp. 48–58.
uncertainties in hydrologic modeling,” Journal of Hydrology, 132: 157–177,
Cheng, S. T., B. C. Yen, and W. H. Tang, “Stochastic risk modeling of dam
1992.
overtopping,” Stochastic and Risk Analyses in Hydraulic Design, edited by
Melching, C. S., “Reliability estimation,” Computer Models of Watershed
B. C. Yen and Y. K. Tung, American Society of Civil Engineers, New York,
Hydrology, edited by V. P. Singh, Water Resources Publications, Highlands
1993, pp. 123–132.
Ranch, CO, Chap. 3, 1995, pp. 69–118.
Christiaens, K. and J. Feyen, “Use of sensitivity and uncertainty measures
Melching, C. S. and S. Anmangandla, “Improved first-order uncertainty
in distributed hydrological modeling with an application to the MIKE SHE
method for water quality modeling,” Journal of Environmental Engineering,
model,” Water Resources Research, 38 (9): 1169, 2002, doi:10.1029/2001
ASCE, 118 (5): 791–805, 1992.
WR000478.
Melching, C. S. and W. Bauwens, “Uncertainty in coupled non-point source
Chow, V. T., D. R. Maidment, and L. W. Mays, Applied Hydrology, McGraw-
and stream water-quality models,” Journal of Water Resources Planning and
Hill, New York, 1988.
Management, ASCE, 127 (6): 403–413, 2002.
Cochran, W., Sampling Techniques, 2nd ed., John Wiley & Sons, New York,
1966. Melching, C. S. and B. C. Yen, “Slope influence on storm sewer risk,”
Cornell, C. A., “A probability-based structural code,” Journal of American Stochastic and Risk Analysis in Hydraulic Engineering, edited by B. C. Yen,
Concrete Institute, 66 (12): 974–985, 1969. Water Resources Publications, Littleton, CO, 1986, pp. 66–78.
Cornell, C. A., “First-order analysis of model and parameter uncertainties,” Melching, C. S. and C. G. Yoon, “Key sources of uncertainty in QUALE
Proceedings of the International Symposium on Uncertainties in Hydrologic and model of Passaic River,” Journal of Water Resources Planning and Management,
Water Resources Systems, Tucson, AZ, Vol. 2, 1972, pp. 1245–1272. ASCE, 122 (2): 105–113, 1996.
Dagpunar, J., Principles of Random Variates Generation, Oxford University Melching, C. S., H. G. Wenzel, Jr., and B. C. Yen, “A reliability estimation in
Press, New York, 1988. modeling watershed runoff with uncertainties,” Water Resources Research, 26
Der Kiureghian, A. and P. L. Liu, “Structural reliability under incomplete (10): 2275–2286, 1990.
probability information,” Journal of Engineering Mechanics, ASCE, 112 (1): Nataf, A., “Détermination des distributions de probabilités dont les marges
85–104, 1985. sont données,” Computes Rendus de l’Académie des Sciences, Paris, 255: 42–43,
Easa, S. M., “Probabilistic design of open drainage channels,” Journal of 1962.
Irrigation and Drainage Engineering, ASCE, 118 (6): 868–881, 1992. Parrish, R. S., “Generating random deviates from multivariate Pearson
Fogel, M. and L. Duckstein, “Stochastic precipitation modeling for evaluat- distributions,” Computational Statistics & Data Analysis, 9: 283–295, 1990.
ing non-point source pollution,” Statistical Analysis of Rainfall and Runoff, Press, W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, Numerical
edited by V. P. Singh, Water Resources Publication, Littleton, CO, 1982. Recipes in Pascal: The Art of Scientific Computing, Cambridge University
Gould, H. and J. Tobochnik, An Introduction to Computer Simulation Press, New York, 1989.
Methods: Applications to Physical Systems, Part 2, Addison-Wesley, Reading, Press, W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, Numerical
MA, 1988. Recipes in FORTRAN: The Art of Scientific Computing, Cambridge University
Hammersley, J. M. and K. W. Morton, “A new Monte Carlo technique Press, New York, 1992.
antithetic-variates,” Proceedings of Cambridge Physics Society, 52: 449–474, Press, W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, Numerical
1956. Recipes in C++: The Art of Scientific Computing, Cambridge University Press,
Han, K-Y., S-H. Kim, and D-H. Bae, “Stochastic water quality analysis New York, 2002.
using reliability method,” Journal of the American Water Resources Association, Ronning, G., “A simple scheme for generating multivariate gamma distri-
37 (3): 695–708, 2001. butions with non-negative covariance matrix,” Technometrics, 19 (2): 179–183,
Hasofer, A. M. and N. C. Lind, “Exact and invariant second-moment code 1977.
format,” Journal of the Engineering Mechanics Division, ASCE, 100 (1): 111–121, Rosenblatt, M., “Remarks on a multivariate transformation,” Annals of
1974. Mathematical Statistics, 23: 470–472, 1952.
Henley, E. J. and H. Kumamoto, Reliability Engineering and Risk Assessment, Rousselle, J., “On some problems of flood analysis,” Ph.D. Dissertation,
Prentice-Hall, Englewood Cliffs, NJ, 1981. Colorado State University, Fort Collins, CO, 1970.
27_Singh_ch27_p27.1-27.10.indd 9 8/22/16 1:47 PM

Rubinstein, R. Y., Simulation and The Monte Carlo Method, John Wiley & Virjling, J. K., “Probabilistic design of water retaining structures,” Engineering
Sons, New York, 1981. Reliability and Risk in Water Resources, edited by L. Duckstein and E. J. Plate,
Shinozuka, M., “Basic analysis of structural safety,” Journal of Structural Martinus Nijhoff, Dordrecht, The Netherlands, 1987, pp. 115–134.
Engineering, ASCE, 109 (3): 721–740, 1983. Vrijling, J. K., “Development in probabilistic design of flood defenses in the
Sitar, N., J. D. Cawlfield, and A. Der Kiureghian, “First-order reliability Netherlands,” Reliability and Uncertainty Analyses in Hydraulic Design, edited
approach to stochastic analysis of subsurface flow and contaminant transport,” by B. C. Yen and Y. K. Tung, American Society of Civil Engineers, New York,
Water Resources Research, 23 (5): 794–804, 1987. 1993, pp. 133–178.
Singh, S. and C. S. Melching, “Importance of hydraulic model uncertainty Yeh, K. C. and T. K. Tung, “Uncertainty and sensitivity of a pit migration
in flood-stage estimation,” Proceedings of the 1993 ASCE National Conference model,” Journal of Hydraulic Engineering, ASCE, 119 (2): 262–281, 1993.
on Hydraulic Engineering, Hydraulic Engineering ’93, edited by H-W. Shen, Yen, B. C., “Safety factor in hydrologic and hydraulic engineering design,”
S-T. Su, and F. Wen, Vol. 2, 1993, pp. 1939–1944. Reliability in Water Resources Management, edited by E. A. McBean, K. W. Hipel,
Sohrabi, T. M., A. Shirmohammadi, T. W. Chu, H. Montas, and A. P. and T. E. Unny, Water Resources Publications, Littleton, CO, 1979, pp. 389–407.
Nejadhashemi, “Uncertainty analysis of hydrologic and water quality predic- Yen, B. C., S. T. Cheng, and C. S. Melching, “First-order reliability analysis,”
tions for a small watershed using SWAT2000,” Environmental Forensics, 4 (4): Stochastic and Risk Analysis in Hydraulic Engineering, edited by B. C. Yen,
229–238, 2003. Water Resources Publications, Littleton, CO, 1986, pp. 1–34.
Todorovic, P. and V. Yevjevich, “Stochastic process of precipitation,” Yen, B. C. and C. S. Melching, “Reliability analysis methods for sediment
Hydrology Paper No. 35, Colorado State University, Fort Collis, CO, 1969. problems,” Proceedings of the 5th Federal Interagency Sediment Conference,
Tung, Y. K., “Evaluating the probability of violating dissolved oxygen stan- Las Vegas, Vol. 2, 1991, pp. 9.1–9.8.
dard,” Ecological Modelling, 51: 193–204, 1990. Yu, P. S., T. C. Yang, and S. J. Chen, “Comparison of uncertainty analysis
Tung, Y. K., B. C. Yen, and C. S. Melching, Hydrosystems Engineering Risk methods for a distributed rainfall-runoff model,” Journal of Hydrology, 244:
and Reliability Analysis, McGraw-Hill, New York, 2006. 43–59, 2001.
Vale, C. D. and V. A. Maurelli, “Simulating multivariate nonnormal distri- Zelenhasic, E., “Theoretical probability distribution for flood peaks,”
butions,” Psychometrika, 48 (3): 465–471, 1983. Hydrology Paper No. 42, Colorado State University, Fort Collins, CO, 1970.
27_Singh_ch27_p27.1-27.10.indd 10 8/22/16 1:47 PM

Chapter 28
Scaling and Fractals
BY
DANIELE VENEZIANO AND CHIARA LEPORE
ABSTRACT n
maps such that Fi (S ) ∩ Fj (S ) is almost empty and S = Φ(S ) =  i=1 Fi (S ). Recall
This chapter reviews concepts and methods of scale invariance with emphasis that F is a contraction map if it reduces the distance between any two distinct
on hydrologic applications. It covers the scale-invariant properties known as points. If the maps Fi contract S isotropically by the same factor r > 1, S is a
self-similarity and multifractality; methods to generate scale-invariant sets, fractal set with fractal (Hausdorff or box) dimension D = ln(n)/ ln(r ).
functions and measures; basic properties of selected classes of scale-invariant Hutchinson (1981) proved that an ss set S is uniquely defined by the transfor-
processes; ways to infer scale invariance from data; and frequently encoun- mations Φ and is the attractor of the sequence of sets obtained by recursively
tered deviations from exact scale invariance. applying Φ as S j = Φ(S j −1 ) starting from any closed bounded set S0.
Extensions to random sets (e.g., Falconer, 1986; Mauldin and Williams,
28.1 INTRODUCTION 1986; Graf, 1987; Zahle, 1988) essentially require invariance of probability
distributions under scaling. Many sets satisfy this property under contraction
Like many other scientific and technical fields, hydrology has greatly benefited or dilation with respect to the origin. If one further requires the set to be
from the exploitation of scale invariance, the property that an object like a set, stationary, the ss property should hold relative to any point of the embedding
a function, or a measure reproduces itself (in some sense) under scale-change space, but any such set must be either empty or equal to the entire space. A
transformations. In hydrology, scale-invariance concepts first appeared in the third and more interesting possibility is to require scale invariance relative to
1940s in the work of Robert Horton and subsequently of Arthur Strahler, any point of the random set (Zahle, 1988).
Ronald Shreve, John Hack, and Harold Hurst (Horton, 1945; Hurst, 1951; Most methods to generate self-similar sets use recursive procedures that at
Hack, 1957; Strahler, 1957; Shreve, 1966). The symmetries and laws for channel each step add detail or points to the set. For example, one may generate self-
networks and river flows discovered in these works were later recognized to be similar curves G in Rd using Hutchinson’s (1981) approach, starting from a
manifestations of scale invariance. Renewed impetus to scale-based models piece-wise linear curve G0 that connects n + 1 given points A1, A2, …., An+1.
came from Lucien Le Cam’s seminal paper on multiscale pulse representations Then one recursively applies a set of contractive transformations Φ = ( F1 ,...., Fn )
of rainfall (Le Cam, 1961) and from the systematization of fractal geometry and where Fi maps the segment A1An+1 into AiAi+1 to obtain increasingly detailed
the discovery of multifractal processes by Benoit Mandelbrot and others (e.g., representations (Barnsley, 1988; Tricot, 1995). Other recursive constructions
Mandelbrot, 1974, 1983; Frisch and Parisi, 1985; Schertzer and Lovejoy, 1987). include the random midpoint displacement method that generates Brownian
In the physical world where variability is ubiquitous, scale invariance rarely motion (Fournier et al., 1982) and replacement procedures to construct self-
manifests itself as a deterministic property whereby an object is made of exact similar rooted trees, which for example have been used as models of river
scaled replicas of itself. That is why in this review we focus on random objects networks (Mandelbrot and Vicsek, 1989). In application contexts, it is often
whose probability laws remain the same under scale-change transformations. more enlightening if one can find construction procedures rooted in the
Section 28.2 states the mathematical conditions of scale invariance for sets physics of a phenomenon. For example, rock fractures often form sequen-
and functions and describes mechanisms through which this property may tially in time, so that the joints at level j depend on the geometry of the
arise. Section 28.3 gives key properties of two important classes of scaling preexisting joints up to level j – 1 in addition to the stress field and local
random functions. Inference methods for stationary multifractal measures strength of the rock (Lee et al., 1990). Similarly, a river network may initiate
are briefly discussed in Sec. 28.4, while Sec. 28.5 delves with often-observed at the outlet and propagate and bifurcate upstream depending on the local
violations of scale invariance. This is an abbreviated version of a more erodibility of the soil, the local flow, and the slope of the terrain (Willgoose
extensive review by Veneziano and Langousis (2010). et al., 1991; Rodríguez-Iturbe and Rinaldo, 1997).
28.2 SCALE-INVARIANT SETS AND FUNCTIONS 28.2.2 Self-Similar Processes

Many terms are used in the literature to describe the scale invariance of ran-
The intuitive notion of scale invariance is that, when appropriately magnified, dom functions: self-similarity, self-affinity, mono-fractality, multifractality,
any part of an object is similar to the whole. By differently specifying what dynamic scaling, generalized scale invariance, etc. Here we distinguish only
“similar” mean and under what scaling operation the object is invariant, one between self-similarity (ss) and multifractality (mf), based on whether the
can generate a wide range of scale-invariance conditions. Here we focus on scale-change transformation under which a random process is invariant is
basic forms of scale invariance for random sets and functions and mention deterministic (ss) or stochastic (mf). So defined, self-similarity encompasses
frequently encountered variants, but do not aim at being exhaustive. In any self-affinity and monofractality, whereas dynamic scaling and generalized
given application, one should be clear on the specific scaling property that is scale invariance are specific forms of multifractality for random fields in
assumed or detected from data. spaces of dimension d ≥ 2.
28.2.1 Self-Similar Sets A random process X(t), t > 0, is said to be self-similar if, for some real H
and any r > 0,
For deterministic sets, a precise notion of scale invariance (called self-similar-
ity or ss) was first proposed by Hutchinson (1981). In this definition, a set S is d
self-similar if there exists a finite set Φ = ( F1 ,..., Fn ) of contraction similarity X (t ) = r − H X (rt ), t > 0 (28.1)
28-1
28_Singh_ch28_p28.1-28.6.indd 1 8/22/16 1:48 PM

28-2 Scaling and Fractals
d
where = denotes equality in distribution. The exponent H is called the self- normal, Levy stable, gamma, Student’s t, Poisson, and compound Pois-
similarity index and X(t) is said to be H-ss. One can readily extend the ss son distributions among others.
property to processes on the entire real line, random fields in spaces of 2. The distribution of Aro for any admissible scaling factor ro ≠ 1 completely
dimension d > 1, and random measures. One may also weaken Eq. (28.1) by defines the distribution of Ar for any other admissible r.
applying it to the increments of the process rather than the process itself. 3. If Eq. (28.5) holds under both contraction and dilation (for all r > 0), Ar
Except for trivial cases such as X(t) = constant, ss processes are nonstation- must be deterministic and equal r −H for some H. It follows that ss pro-
ary. For example, X(t) vanishes at 0 if H > 0 and vanishes at infinity if H < 0. cesses form a separating class between processes that are mf under
However, H-ss processes may have stationary increments. We refer to this contraction and under dilation.
important class as H-sssi processes. 4. Multifractal processes under contraction and under dilation are in one-
Next we describe three general ways to construct self-similar processes: trans- to-one correspondence (Veneziano, 1999). Hence, any property and
formation of stationary processes, limits of renormalized processes, and addi- construction method for one class has a counterpart in the other class.
tion of scaled processes. Lamperti (1962) showed that a stationary process, X(t), Since for most applications one is interested in processes that are multifrac-
t > 0, is H-ss if and only if there exists a stationary process Z(t), –∞ < t ≤ ∞, tal under contraction, when we use mf without specification, we refer to
such that multifractal scaling under contraction. Multifractality under dilation is
encountered in the Fourier representation of stationary mf processes
X (t ) = t H Z (ln t ), t > 0 (28.2) (Veneziano, 1999) and has been proposed in certain areas of hydrology, for
Notice that in this construction the scaling exponent H is controlled by the example to model how the maximum annual river flood depends on basin
deterministic factor tH, whereas the random process Z(lnt) is responsible for the area (Gupta et al., 1994).
differences among distinct H-ss processes X(t). The smoothest-possible H-ss In analogy to ss processes, general ways to construct mf processes include
processes, obtained by setting Z(t) to a deterministic or random constant C, are transformation of stationary processes, limits of renormalized processes, and
the power-law functions X(t)=CtH. Based on Eq. (28.2), one can verify that a the addition of scaled processes (Veneziano, 1999). Stationary multifractal
process X(t) is H-ss by checking that the transformed process Z(t)=e–HtX(et) is measures are generally constructed as products of scaled processes (multipli-
stationary. cative cascades). A brief description of these construction methods is given
Lamperti (1962) gave a second characterization of ss processes, as renor- below, mainly by pointing out the salient differences with similar construc-
malization limits: X(t) is H-ss if and only if there exists a process Y(t) and a tions of ss processes.
positive function a(u) such that, as u → ∞, Lamperti’s (1962) stationarity characterization of ss processes in Eq. (28.2)
can be extended to mf processes. The main difference is that the deterministic
a(ru)/ a(u) → r − H function tH that confers the H-ss property is replaced with a random function
d (28.3) that confers the mf property. Also Lamperti’s (1962) characterization of ss
a(u)Y (ut )→ X (t ) processes as renormalization limits, Eq. (28.3), can be extended to multifrac-
tal processes. The extension replaces the scaling function a(u) with a random
The construction methods given above have mainly theoretical interest.
process A(u) with certain limiting properties. Finally, the mf extension of the
Most H-ss processes are physically and numerically generated by adding inde-
weighted-sum characterization in Eq. (28.4) replaces the deterministic
pendent scaled copies of a given random process, as (Voss, 1988)
weights ro− jH with certain random weights Ar j . Like in the ss case, the discrete
o
∞ construction can be made continuous.
X (t ) = ∑ ro− jH Z j (rojt ) (28.4) The construction of stationary multifractal measures deserves special atten-
j =−∞ tion. These measures have been used extensively in hydrology, for example, to
where ro > 1 is a constant scale-change factor and the Zj (t) are independent represent the hydraulic conductivity of geologic materials (Veneziano and
copies of some random process Z(t) such that Z(0) = 0 if H > 0 and Z(–∞) = 0 Essiam, 2003) and the distribution of rainfall in space and/or time (e.g., Gupta
if H < 0. If Z(t) has stationary increments, then X(t) is H-sssi. A special case of and Waymire, 1993; Lovejoy and Schertzer, 1995; Over and Gupta, 1996; Deidda
Eq. (28.4) produces processes called Fractal Sums of Pulses. These processes et al., 2004). While these measures can be obtained through the methods men-
were first developed by Halford (1968) and later rediscovered and applied to tioned earlier, a construction procedure based on the product of scaled indepen-
rainfall by Lovejoy and Mandelbrot (1985). dent copies of a non-negative random function is simpler and more broadly
The process X(t) in Eq. (28.4) is H-ss for scaling by discrete factors r = rok used. Even simpler is the construction of discrete multiplicative cascades, which
where k is an integer. Continuous-summation extensions produce continuous- are multifractal measures with limited stationarity and discrete scaling proper-
scaling H-ss processes. The fractionally integrated a-stable measures of ties (e.g., Gupta and Waymire, 1993; Lovejoy and Schertzer, 1995).
Samorodnitsky and Taqqu (1994) are examples of such continuous constructions. The canonical construction of a discrete multifractal cascade considers a
random measure with mean value 1 inside the unit d-dimensional cube Sd.
28.2.3 Multifractal Processes One starts at level 0 with a uniform unit measure density inside the cubic tile
Multifractality is a more general form of scale invariance than self-similarity. Ω01 = Sd . At subsequent levels n = 1, 2, …, each tile at level n – 1 is partitioned
The property is variously defined, most often with reference to nonlinear into md cubic tiles, where m > 1 is an integer called the multiplicity of the
scaling of the moments with the moment order or to the fractal dimension of cascade. The measure density inside each tile Ωni (i = 1,..., mnd ) is obtained by
sets of points where a process displays various degrees of singularity. Here we multiplying the measure density in its parent tile at level n – 1 by a random
prefer a more fundamental definition based on scale invariance, from which variable Wni with unit mean value. The variables Wni are independent copies
nonlinear scaling of the moments and other properties follow. of a non-negative random variable W, called the generator of the cascade.
Certain classes of multifractal processes, including stationary multifractal This construction can be “densified” to obtain continuous-scaling station-
processes, exist only as random measures X(dt) or equivalently as generalized ary measures. Among such measures are those obtained by filtering and
∞ exponentiating a-stable (Gaussian or Levy-stable) measures (Schertzer and
processes X (h) = ∫ h(h)X (d t ) where h(t) is a test function (Yaglom, 1987). Lovejoy, 1987). These are sometimes called “universal” mf processes based on
−∞
For this reason, here we define multifractality for such generalized processes. the argument that the attraction of sums of iid variables to Levy-stable distri-
Ordinary processes X(t) are included because they can be viewed as measure butions should make them ubiquitous in nature. While this argument has
∞
densities such that X(dt)=X(t)dt and X (h) = ∫ h(t )X (t )d t. been refuted, universal multifractal processes retain significant theoretical
−∞
Let X(h) be a generalized random process on the real line, defined for test and practical interest, as they include multifractal generalizations of frac-
functions h in some class Q, and let hr (t) = rh(rt), 0 < r < ∞, be a scaled ver- tional Brownian and Levy motions.
sion of h(t). Notice that r > 1 produces contraction in time and 0 < r < 1
produces dilation. We say that X(h) is multifractal under contraction in Q if 28.3 SOME PROPERTIES OF H-SSSI PROCESSES AND
there exists a set of non-negative random variables A r such that MF CASCADES
d Here we mention basic properties of two classes of processes, which in addi-

X (hr ) = Ar X (h), for any r ≥ 1, hŒQ (28.5) tion to scaling possess stationarity properties: self-similar processes with
stationary increments (H-sssi processes) and stationary multifractal cascades.
If Eq. (28.5) holds for 0 < r ≤ 1, then X(h) is multifractal under dilation.
Important properties of mf processes include the following: 28.3.1 H-sssi Processes
1. log(Ar) must have infinitely divisible (id) distribution [on id distribu- The properties listed below concern the second moments, spectral density, and
tions, see (Feller, 1966)]. Infinitely divisible distributions include the fractal dimension of graphs of H-sssi processes. For a more detailed account,
28_Singh_ch28_p28.1-28.6.indd 2 8/22/16 1:48 PM

SOME PROPERTIES OF H-SSSI PROCESSES AND MF CASCADES 28-3
see Samorodnitsky and Taqqu (1994). We start by noting that for X(t) an H-sssi results, which involve the function K (q ) = logm E[W q ] (m is the multiplicity of
process with H > 0, the existence of moments of certain orders has implications the cascade) and its derivative K '(q ):
on the possible range of H, and vice-versa. Specifically, (a) if E[| X (t )|q ] < ∞ for • Z is nondegenerate if and only if K '(1) < 1.
some q < 1, then H < 1/q; (b) if E[| X (t )|] < ∞, then H ≤ 1; and (c) if 0 < H < 1, • E[Z q ], q > 1, is finite if and only if K (q ) < (q − 1).
then E[X(t)] = 0. Since processes with finite second moments satisfy
E[| X (t )|] < ∞, it follows from property (b) that for them H ≤ 1. These are also the conditions for nondegeneracy and finiteness of the
q-moment of ε n. The order q∗ > 1 above which the moments of Z and ε n
Second Moments diverge is found from the condition K (q∗ ) = (q∗ − 1). For an extension of these
Under the condition E[|X(t)2]< ∞, one can derive the second moments of X(t) results, see Guivarc’h (1990).
and its increments ξ (k ) = X (k + 1) − X (k ). For X(t) one finds Scaling of the dressed moments of order q < q* follows immediately from
Eq. (28.10), which gives
1 2H 2H

2
{ }
t + s − (t − s )2 H E[ X (1)2 ] (28.6)
E[ X (t )X (s )] = E[(ε n+1 )q ] = m K ( q ) E[(ε n )q ] (28.11)
The Gaussian process with zero mean and second moment function in Eq. (28.6) One can further show that if the spectral density of the cascade S(ω ) exists,
is known as fractional Brownian motion (f Bm) and is the only Gaussian H-sssi then
process.
Since the increments ξ (k ) = X (k + 1) − X (k ) are stationary, their second S(ω ) ∝| ω |−1+ K (2) (28.12)
moment function B(n) = E[ξ (s )ξ (s + n)] depends only on the lag n. This func- Since K(2) > 0, the exponent in Eq. (28.12) is larger than −1. Comparison
tion has the form with Eq. (28.8) shows that stationary multifractal measures have a slower-
1 decaying spectral density than H-sssi processes. This slower power-law decay
B(n) =
2
{ }
(n + 1)2 H − 2n 2 H + (n − 1)2 H E[ X (1)2 ] (28.7) of S(ω ) is often taken as first evidence of multifractality.
Consequences of Eq. (28.7) are: (a) if 0 < H ≤ 0.5, the increments of X(t) Distributions of Z and ε n
have negative correlation and X(t) has weak dependence; (b) if H = 0.5, the The distribution of the dressed densities ε n depends on the distributions of W
increments of X(t) are uncorrelated; and (c) if 0.5 < H ≤ 1, the increments of and Z. In turn, the distribution of Z depends on the distribution of W and the
X(t) have positive correlation and X(t) has long-range dependence [on weak multiplicity m. The latter dependence is embodied in the fundamental identity
and long-range dependence, see Beran (1994)].
d 1 m
Spectral Density Z= ∑Wi Zi (28.13)
m i =1
For H > 0, the process X(t) is nonstationary. However, if one relaxes the condi-
tion X(0) = 0, one may regard X(t) as a stationary process with infinite vari- where the variables Wi and Zi are independent copies of W and Z, respectively.
ance and a well-defined spectral density S(w) of the type Based on Eq. (28.13), the distribution of Z can be found iteratively as follows
(Veneziano and Furcolo, 2003): one starts by setting Z(0) = 1 and then updates
S(ω ) ∝ ω − (1+ 2 H ) (28.8) the distribution of Z( j ) at steps j = 1, 2,… using
Equation (28.8) holds for all H-sssi processes with finite second moments, 1 m
including fractional Brownian motion. The power-law form of S(w) in Z( j ) = ∑Wi Z( j−1)i (28.14)
m i =1
Eq. (28.8) is often used to validate self-similarity and estimate the index H. d
However, power-law spectral densities originate also under the condition of where the variables Z( j −1)i are independent copies of Z( j−1). As j → ∞, Z( j ) → Z .
multifractality (see later). Distinguishing between self-similarity and multi- An important feature of Z and ε n is that when the order of moment diver- ∗
fractality requires consideration of moments of several different orders. gence q* is finite the distributions have algebraic upper tails P[Z > z ] ~ z − q and
− q∗
P[ε n > ε ] ~ ε . However, for ε n the power-law behavior effectively starts at
Fractal Dimension of Graphs values of e that increase rapidly as the resolution level n increases and at high
d
Using the property that X(t) is statistically equivalent to X '(t ) = r − H X (rt ), resolutions the power-law tail may not be observable, even in large samples. In
one obtains that the Hausdorff and box fractal dimensions of the graph of the high-resolution limit n → ∞, the power-law tail becomes unimportant and
X(t) are both given by the distribution of ε n is attracted to one of three types: lognormal, log-Levy, or
a third type with a lognormal or log-Levy distribution of ε n+ = (ε n | ε n > 0)
D = 2 − H (28.9) (Veneziano and Furcolo, 2003).
For example, graphs of Brownian motion (for which H = 0.5) have fractal
dimension of 1.5. Extremes
Many applications require knowledge of the high quantiles of en or the distri-
28.3.2 Stationary mf Measures bution of the maximum of en over all the tiles of a cascade at level n. Several
The multifractal property in Eq. (28.5) has implications on the moments, such results can be derived, in approximation, from a property of the random
distribution, and other properties of X(h). In what follows, we focus on dis- variables Ar in Eq. (28.5). Specifically, let g be a given positive number. It fol-
crete multiplicative cascades on the line, but results are readily extended to lows from Cramer’s large deviation theorem (Cramér, 1938; Stroock, 1999)
cascades in spaces of higher dimension. that as r → ∞
A fundamental quantity in a discrete cascade is the average measure den-
sity inside a generic tile at level n. One should distinguish between two such P[ Ar > r γ ] ~ r − C (γ ) (28.15)
average densities: the “bare” density ε n ;b when the cascade construction is where ~ denotes equality up to a factor that varies slowly with r at infinity and
terminated at level n, and the “dressed” density ε n for the fully developed C(g ) is the Legendre transform of K(q),
cascade. These densities satisfy
C(γ ) = max{qγ − K (q )} (28.16)
ε n ;b = W1W2 ⋅⋅⋅Wn q
(28.10) Since the bare densities en ; b have the same distribution as Ar for r = mn,
ε n = ε n ;b Z Eq. (28.16) has direct implications on the upper tail of those densities when n is
where the variables Wi are independent with the distribution of the cascade high.
generator W. Z is the so-called dressing factor and is independent of ε n ;b. What happens for the dressed densities en = en ; b Z is less obvious because of
the effect of Z. The problem has been studied by Schertzer and Lovejoy (1987)
Existence, Moment Scaling, and and Veneziano (2002), where the following is proven. Define a “dressed
Spectral Density moment scaling function” K d (q ) as
The bare densities ε n ;b always exist and, as one can deduce from Eq. (28.10),
their moments are the nth powers of the corresponding moments of W. log E[ε rq ]  K (q ), q < q∗
Matters are more complicated for the dressed densities, because ε n is non- K d (q ) = lim = (28.17)
r →∞ log r ∞ , q ≥ q∗
degenerate (has nonzero probability of being finite non-zero) if and only if Z is
nondegenerate, and the qth moment of ε n exists if and only if the qth moment where q* is the order of moment divergence and let Cd (γ ) be the Legendre
of Z exists. Kahane and Peyriere (1976) proved the following fundamental transform of K d (q ),
28_Singh_ch28_p28.1-28.6.indd 3 8/22/16 1:48 PM

C(γ ), γ ≤γ ∗ The standard method to estimate K(q), say for a stationary multifractal
Cd (γ ) = max{qγ − K d (q )} =  (28.18) measure in the unit interval, is to partition that interval into sub-intervals of
∗ ∗ ∗
q
C(γ ) + q (γ − γ ), γ >γ ∗ length 2− n, where n = 1, 2, … is the resolution level, and for each n calculate
∗ ∗
with C(γ ) in Eq. (28.16) and γ = K ′(q ) the slope of K(q) at q*. Then, as the empirical moments of en, the average measure density at level n. After these
n → ∞, initial operations, the function K(q) is estimated in two steps (Harris et al.,
1997). In the first step, one calculates the empirical moments µˆnq = 〈ε nq 〉 and for
P[ε n > mnγ ] ~ m− nCd (γ ) (28.19) each q finds Kˆ (q ) as the slope of the linear regression of log 2 ( µˆnq ) against n over
Note that for γ > γ , Eqs. (28.18) and (28.19) give P[ε n > mnγ ] ~ m− nq γ ,
∗ ∗
a range [nmin , nmax ] of resolution levels. We refer to Kˆ (q ) as a nonparametric
which by putting ε = mnγ becomes P[ε n > ε ] ~ ε − q . This shows that when the
∗
estimator of the moment scaling function. In the second step one fits a para-
order of moment divergence q is finite, the dressed densities have an alge-
* metric model Kˆ par (q ) to Kˆ (q ) over a range [qmin , qmax ] of moment orders.
braic upper tail, as previously noted. As the resolution level nmax → ∞, Kˆ (q ) becomes linear with positive slope
Equation (28.19) is fundamental to the estimation of the extremes of sta- g1 for q > q1 where g1 and q1 are illustrated in Fig. 28.1 (Ossiander and
tionary multifractal measures. Since there are mn cascade tiles at resolution Waymire, 2000, 2002). In practice nmax is finite and Kˆ (q ) displays gradual
level n, the product mn P[ε n > ε ] is the expected number of exceedances of straightening as q increases, with an asymptotic slope γˆ1 that varies from
level e by en in a single cascade realization. One may take the reciprocal of this sample to sample. Using numerical simulation, Harris et al. (1997) studied the
quantity, bias and variance of Kˆ (q ) for different q and nmax .
The bias and variance of the parametric estimator Kˆ par (q ) depend in addi-
1 tion on the range of moment orders used in the second step. Following the
Tn (ε ) = (28.20)
mn P[ε n > ε ] approach of Harris et al. (1997), Furcolo and Veneziano (2008) assessed the
performance of both nonparametric and parametric estimators and suggested
as the return period of e expressed in units of cascade realizations. Solving
alternatives to reduce their mean square error.
Eq. (28.20) for e gives the return period values e(n, T). Using Eqs. (28.19) and
(28.20), one can show that (Hubert et al., 1998; Veneziano and Furcolo, 2002) 28.4.2 Estimation of Universal Parameters
mnγ 1 T 1/q1 , n → ∞ , T finite Stationary universal multifractal measures are widely used to model geo-
ε (n,T ) ~  n 1/q* (28.21) physical fields including atmospheric turbulence and temperature, cloud
m T , n finite, T → ∞ reflectivity, topographic surface, ocean surface radiance, rainfall and river
flow (eg., Schertzer and Lovejoy, 1987; Lovejoy and Schertzer, 1990; Lavallee
where g1 is such that C(γ 1 ) = 1 and q1 = C '(γ 1 ) is the moment order associated
et al., 1991; Schmitt et al., 1992; Olsson et al., 1993; Schmitt et al., 1993; Tessier
with g1; see Fig. 28.1. This result has been used to explain the approximate
et al., 1993; Deidda et al., 2004, 2006; Lilley et al., 2006; Lovejoy and Schertzer,
scaling of the intensity-duration-frequency curves with return period T and
2006; De Montera et al., 2009; Serinaldi, 2010; Sun and Barros, 2010;
temporal resolution mn.
Veneziano and Lepore, 2012). The amplitude scaling factors Ar have (maxi-
mally negatively skewed) log-Levy distribution and in the stationary case the
moment scaling function is given by
 C1
 (qα − q ), for α ≠ 1
K (q) K (q ) =  α − 1 (28.22)
C1q ln(q ), for α = 1
K (q)
(q–1) where 0 < a ≤ 2 is the index of stability of the Levy distribution and C1 is a
positive constant that controls the dispersion of ln(Ae ) (Schertzer and
Lovejoy, 1987). For a = 2, ln(Ae ) has normal distribution with mean value
m = −C1 and variance σ 2 = 2C1 (this corresponds to the sub-class of lognor-
mal multifractal measures), whereas for a → 0 the distribution of Ae develops
a mass e−C1 at eC1 and a mass (1 − e−C1) at zero (this limit corresponds to so-
1 called beta multifractal measures). A popular way to estimate a and C1 is the
double trace moment (DTM) method of Lavallée et al. (1991), which involves
(0, 0) q1 q* q certain sequences of exponentiation, normalization, and averaging opera-
tions. Veneziano and Furcolo (1999) observed that the original DTM method
g1 uses scaling relations that hold only in approximation and produces biased
1 estimates of the parameters (especially C1). They also suggested an unbiased
variant of the method.
–1
28.5 PROCESSES WITH LIMITED SCALE INVARIANCE
Figure 28.1 Illustration of the moment-scaling function K(q) and the parameters q1, Many natural phenomena display systematic deviations from scale invariance
q*, and g1 in Eq. (28.21).
or are scale invariant in some limited sense. In some cases, the mechanism
that generates a process is scale invariant, but the process itself is not. For
example, this is the case when a multiplicative cascade construction is termi-
28.4 INFERENCE OF SCALING FOR STATIONARY
nated at a finite resolution. Similarly, processes may have different scaling
MULTIFRACTAL MEASURES
properties in different resolution ranges, meaning that the generating mecha-
Methods to estimate the scaling parameters from data depend on the type of nism has distinct scale-invariance properties in different ranges, whereas the
scale invariance considered. Here we focus on stationary multifractal measures process itself is not scaling. In yet other cases, a process may become scale-
X(dt), whose scaling is characterized by the random factor Ar in Eq. (28.5) or invariant after a simple transformation. One example is exponentiated frac-
equivalently by the moment scaling function K (q ) = log r ( E[ Arq ]). First, we tional Brownian motion. Exponentiation destroys scale invariance, but the
consider the general problem of K(q) estimation and then remark on special- fact that the log of the process is scale invariant may still be of interest. In what
ized methods for universal multifractal measures. follows, we describe four different deviations from multifractal scaling that
have been noted in rainfall, turbulence, and hydraulic conductivity.
28.4.1 Estimation of K(q)
Since the factors Ar are not directly accessible, one typically finds K(q) from Bounded Cascades
the scaling of moments of variables X (h), such as average densities and wave- The generator W of a stationary multifractal cascade has the same distribution
let coefficients. The range of finite (and hence usable) moments depends on at all levels n, whereas bounded cascade models (Menabde, 1998; Menabde
which specific variable is analyzed and never exceeds the range of finite and Sivapalan, 2000) allow the distribution to depend on n. The variation with
moments of Ar . For example, if Ar has lognormal distribution then all its n is typically assumed to be systematic. For example, in a bounded variant of
moments are finite, but the average densities and wavelet coefficients have the lognormal cascade one often assumes that ln(Wn ) has normal distribution
finite moments only over a finite range of q. with mean value −0.5c nσ 2 and variance c nσ 2 , where σ 2 > 0 and 0 < c < 1 are
28_Singh_ch28_p28.1-28.6.indd 4 8/22/16 1:48 PM

References 28-5
given constants. The result is essentially exponentiated fractional Brownian In many cases nature displays departures from exact scale invariance. As
motion exp{BH (t )} with H = −0.5logm (c ) and m the multiplicity of the cascade. scale-invariant modeling and analysis are becoming well established, the next
While the process exp{BH (t )} does not scale, in the small-scale limit its incre- frontier should include a more systematic understanding and treatment of
ments have the same H-ss property as those of BH(t). In hydrology, bounded deviations from scale invariance. This would lead to models that, while in
cascades have been proposed as models of rainfall (Menabde et al., 1997; general more complicated than those based on self-similarity and multifrac-
Menabde and Sivapalan, 2000) and exponentiated fBm has been used to repre- tality, have wider applicability and better represent observed phenomena.
sent the hydraulic conductivity of natural aquifers (Molz and Bowman, 1993).
Extended Self-Similarity ACKNOWLEDGMENT
Extended self-similarity (ESS) is a departure from multifractality originally The authors gratefully acknowledge the useful comments by an anonymous
introduced to better describe turbulent velocity fields (e.g., Benzi et al., 1993). reviewer.
If ∆v(τ ) is the longitudinal velocity increment at distance t in the mean flow
direction, then a standard multifractal assumption is that for any r > 1 a ran- REFERENCES
dom variable Ar exists such that
Barnsley, M. F., Fractals Everywhere, 2nd ed., Academic Press, Boston,
d
∆v(τ ) = Ar ∆v(rτ ) 1988, pp. 205–207.
(28.23) Benzi, R., S. Ciliberto, R. Tripiccione, C. Baudet, F. Massaioli, and S. Succi,
⇒ E[| ∆v(τ )|q ] ∝ τ − K (q ) “Extended self-similarity in turbulent flows,” Physical Review E, 48, R29–R32,
1993.
where K (q ) = log r E[ Arq ]. Equation (28.23) holds reasonably well at high
Beran, J., Statistics for Long-Memory Processes, Vol. 61, CRC Press, Boca
Reynolds numbers, but breaks down at small scales where viscosity becomes
Raton, FL, 1994.
significant. To extend validity to the latter regime, Benzi et al. (1993) suggested
Cramér, H., “Sur un nouveau théoreme-limite de la théorie des probabili-
the weaker condition
ties,” Actualités scientifiques et industrielles, 736: 5–23, 1938.
E[| ∆v(τ )|q1 ] ∝ E[| ∆v(τ )|q2 ]K (q1 )/ K (q2 ) (28.24) De Montera, L., L. Barthès, C. Mallet, and P. Golé, “The effect of rain-no
rain intermittency on the estimation of the universal multifractals model
One can show (Veneziano and Langousis, 2010) that there are similarities parameters,” Journal of Hydrometeorology, 10, 493–506, 2009.
between the ESS property in Eq. (28.24) and the bounded cascades men- Deidda, R., M. G. Badas, and E. Piga, “Space-time scaling in high-intensity
tioned earlier. Tropical Ocean Global Atmosphere Coupled Ocean-Atmosphere Response
Experiment (TOGA-COARE) storms,” Water Resources Research, 40, 2004.
Dependence of the Generator Wn on the State doi:10.1029/2003WR002574.
at Level n Deidda, R., M. G. Badas, and E. Piga, “Space–time multifractality of
In studies of rainfall, Olsson (1998), Guntner et al. (2001), Veneziano et al. remotely sensed rainfall fields,” Journal of Hydrology, 322: 2–13, 2006.
(2006), and others found it necessary to generalize the cascade construction Falconer, K. J., “Random fractals,” Mathematical Proceedings of the
by making the distribution of Wn depend not only on n but also on the bare Cambridge, 100: 559–582, 1986.
rainfall intensity in the parent tile at level n − 1. Note that in this extension, Wn Feller, W., An Introduction to Probability Theory and Its Applications, Vol. 2,
depends on the bare rainfall intensity at level n − 1 (dependence on the dressed John Wiley & Sons, NewYork, NJ, 1996.
intensity at level n − 1 is present also in standard multifractal cascades). Fournier, A., D. Fussell, and L. Carpenter, “Computer rendering of stochas-
tic models,” Communications of the ACM, 25: 371–384, 1982.
Segmented Multifractality Fraedrich, K. and C. Larnder, “Scaling regimes of composite rainfall time-
A special case of dependence of W on the scale level n is when the cascade gen- series,” Tellus A, 45A: 289–298, 1993.
erator W is distributed like Wi for ni < n ≤ ni +1, where 1 = n1 < n2 < . . . < ns = ∞. Frisch, U., and G. Parisi. “Fully developed turbulence and intermittency,”
We refer to this property as segmented multifractality. Measures with seg- Turbulence and predictability in geophysical fluid dynamics and climate
mented multifractality are actually not multifractal at any scale, but if a range dynamics, edited by Ghil, M., R. Benzi and G. Parisi, North-Holland
[ni , ni+1 ] is wide, the measure displays approximate multifractality inside that Publishing Co., Amsterdamn, 1985, pp. 71–88.
range. For example, if it exists, the spectral density S(w) displays approximate Furcolo, P. and D. Veneziano, “Improved moment-scaling estimation for
power-law behavior in the frequency range associated with [ni , ni+1 ]. The stationary multifractal measures,” Symposium on New Statistical Tools in
existence of distinct scaling regimes in rainfall time series has long been Hydrology, Capri, Italy, October 13–14, 2008.
recognized; see for example Fraedrich and Larnder (1993). Graf, S., “Statistically self-similar fractals,” Probability Theory and Related
Fields, 74, 357–392, 1987.
Guivarc’h, Y., “Sur une extension de la notion de loi semi-stable,” Annales
28.6 CONCLUSION
de l’I.H.P. Probabilités et Statistiques, 26 (2): 261–285, 1990.
This review of scale invariant methods in hydrology is necessarily concise and Guntner, A., J. Olsson, A. Calver, and B. Gannon, “Cascade-based disag-
the reference list is definitely incomplete. Indeed, the literature on fractal/ gregation of continuous rainfall time series: the influence of climate,”
multifractal applications, even restricted to hydrology, is vast and growing at Hydrology and Earth System Sciences, 5: 145–164, 2001.
a rapid pace. However, the evolution of the underpinning concepts and meth- Gupta, V. K. and E. C. Waymire, “A statistical analysis of mesoscale rain-
ods, which are emphasized here, is more gradual. This should hopefully make fall as a random cascade,” Journal of Applied Meteorology, 32: 251–267,
the present review relevant for years to come. 1993.
Our aim has been to present a unified picture of scale invariance, emphasiz- Gupta, V. K., O. J. Mesa, and D. R. Dawdy, “Multiscaling theory of flood
ing the conceptually simple (yet far reaching) difference between self-similarity peaks: regional quantile analysis,” Water Resources Research, 30: 3405–3421,
and multifractality and the commonality of how self-similar and multifractal 1994.
processes can be generated. In our own work, we have found it beneficial Hack, J. T., Studies of longitudinal river profiles in Virginia and Maryland,
to think of multifractality as a scale invariance condition, as presented US Geological Survey Professional Paper, 1957, p. 294.
here. Halford, D., “A general mechanical model for |f|αspectral density random
There are areas in which further understanding and development are noise with special reference to flicker noise 1/|f|,” Proceedings of the IEEE, 56:
needed. One is the robust testing and inference of scale invariance: one too 251–258, 1968.
often concludes that an object is scale invariant from the observed approxi- Harris, D., A. Seed, M. Menabde, and G. Austin, “Factors affecting multi-
mate scaling of the moments. This may be deceptive as there are many scaling analysis of rainfall time series,” Nonlinear Processes in Geophhysics, 4:
nonscaling processes whose moments approximately scale. The problem is 137–156, 1997.
especially critical for the extremes of a process, which are controlled by Horton, R. E., “Erosional development of streams and their drainage
high-order moments that are difficult to estimate from data. It would be basins—hydrophysical approach to quantitative morphology,” Geological
desirable to go beyond the use of moments and develop inference techniques Society of American Bulletin, 56: 275–370, 1945.
that use marginal distributions at various scales. At a minimum, after a Hubert, P., H. Bendjoudi, D. Schertzer, S. Lovejoy, A multifractal explanation
scale-invariant model has been fitted to data, one should verify that the for rainfall intensity-duration-frequency curves, Proceedings of the Workshop
model accurately reproduces the observed marginal distributions at multiple “Heavy Rains and Flash Floods,” Publ., 2049, 21–28, Natl. Res. Counc.,
scales. Washington, D.C., 1998.
28_Singh_ch28_p28.1-28.6.indd 5 8/22/16 1:48 PM

Hurst, H. E., “Long-term storage capacity of reservoirs,” The American Rodríguez-Iturbe, I. and A. Rinaldo, Fractal River Basins: Chance and Self-
Society of Civil Engineers, 116: 770–799, 1951. organization, Cambridge University Press, Cambridge, UK, 1997.
Hutchinson, J. E., “Fractals and self similarity,” The Indiana University Samoradnitsky, G. and M. S. Taqqu, Stable Non-Gaussian Random Processes:
Mathematics Journal, 30: 713–747, 1981. Stochastic Models with Infinite Variance, CRC Press, Boca Raton, FL, 1994.
Kahane, J. P. and J. Peyriere, Sur certaines martingales de Benoit Mandelbrot, Schertzer, D. and S. Lovejoy, “Physical modeling and analysis of rain and
Advances in Mathematics, 22: 131–145, 1976. clouds by anisotropic scaling multiplicative processes,” Journal of Geophysical
Lamperti, J., “Semi-stable stochastic processes,” The American Mathematical Research: Atmospheres, 92: 9693–9714, 1987.
Society, 104: 62–78, 1962. Schmitt, F., D. Lavallee, D. Schertzer, and S. Lovejoy, “Empirical determina-
Lavallée,D., S. Lovejoy, D. Schertzer. Universal multifractal theory and tion of universal multifractal exponents in turbulent velocity-fields,” Physical
observations of land and ocean surfaces, and of clouds, SPIE proceedings 1558, Review Letters, 68: 305–308, 1992.
San Diego, pp. 60–75, 1991. Schmitt, F., D. Schertzer, S. Lovejoy, and Y. Brunet, “Estimation of universal
Lavallée, D., D. Schertzer, and S. Lovejoy, On the determination of the multifractal indices for atmospheric turbulent velocity fields,” Fractals-
codimension function, Non-Linear Variability in Geophysics, Springer, Complex Geometry Patterns and Scaling in Nature and Society, 1: 568–575,
Netherlands, 1991, pp. 99–109. 1993.
Le Cam, L., “A stochastic description of precipitation,” Proceedings of the Serinaldi, F., “Multifractality, imperfect scaling and hydrological properties
Fourth Berkeley Symposium on Mathematical Statistics and Probability, of rainfall time series simulated by continuous universal multifractal and
University of California Press, Berkeley, CA, 1961, pp. 165–186. discrete random cascade models,” Nonlinear Processes in Geophysics, 17:
Lee, J. S., D. Veneziano, and H. H. Einstein, Hierarchical fracture trace 697–714, 2010.
model, Rock Mechanics: Contributions and Challenges, Balkema, Rotterdam, Shreve, R. L., “Statistical law of stream numbers,” Journal of Geology, 74:
1990, pp. 261–268. 17–37, 1966.
Lilley, M., S. Lovejoy, N. Desaulniers-Soucy, and D. Schertzer, “Multifractal Strahler, A. N., “Quantitative analysis of watershed geomorphology,”
large number of drops limit in rain,” Journal of Hydrology, 328: 20–37, 2006. Transactions of the American Geophysical Union, 38: 913–920, 1957.
Lovejoy, S. and B. B. Mandelbrot, “Fractal properties of rain, and a fractal Stroock, D. W., Probability Theory: An Analytic View, Cambridge University
model,” Tellus A, 37: 209–232, 1985. Press, Cambridge, UK, 1999.
Lovejoy, S. and D. Schertzer, “Fractals, raindrops and resolution depen- Sun, X. and A. P. Barros, “An evaluation of the statistics of rainfall extremes
dence of rain measurements,” Journal of Applied Meteorology, 29: 1167–1170, in rain gauge observations, and satellite-based and reanalysis products using
1990. universal multifractals,” Journal of Hydrometeorology, 11: 388–404, 2010.
Lovejoy, S. and D. Schertzer, “Multifractals and rain,” In New Uncertainty Tessier, Y., S. Lovejoy, and D. Schertzer, “Universal multifractals—theory
Concepts in Hydrology and Water Resources, edited by Z. W. Kundzewicz, 61–103, and observations for rain and clouds,” Journal of Applied Meteorology, 32:
1995. Retrieved from http://www.cambridge.org/US/academic/subjects/ 223–250, 1993.
earth-and-environmental-science/hydrology-hydrogeology-and-water- Tricot, C., Curves and Fractal Dimension, Springer-Verlag, New York, NY,
resources/new-uncertainty-concepts-hydrology-and-water-resources 1995.
Lovejoy, S. and D. Schertzer, “Multifractals, cloud radiances and rain,” Veneziano, D., “Basic properties and characterization of stochastically self-
Journal of Hydrology, 322: 59–88, 2006. similar processes in Rd,” Fractals, 7: 59–78, 1999.
Mandelbrot, B. B., “Intermittent turbulence in self-similar cascades; diver- Veneziano, D., “Large deviations of multifractal measures,” Fractals-
gence of high moments and dimension of the carrier,” Journal of Fluid Complex Geometry Patterns and Scaling in Nature and Society, 10: 117–129,
Mechanics, 62: 331–358, 1974. 2002.
Mandelbrot, B. B., The Fractal Geometry of Nature (updated and aug- Veneziano, D. and P. Furcolo, “A modified double trace moment method of
mented edition), W. H. Freeman, New York, NY, 1983, p. 468. multifractal analysis,” Fractals-Complex Geometry Patterns and Scaling in
Mandelbrot, B. B. and T. Vicsek, “Directed recursion models for fractal Nature and Society, 7: 181–195, 1999.
growth,” Journal of Physics A: Mathematical and General, 22: L377, 1989. Veneziano, D. and P. Furcolo, “Multifractality of rainfall and scaling of
Mauldin, R. D. and S. C. Williams, “On the hausdorff dimension of some intensity-duration-frequency curves,” Water Resources Research, 38 (12):
graphs,” The American Mathematical Society, 298: 793–803, 1986. 1306, 2002. doi:10.1029/2001WR000372
Menabde, M., “Bounded lognormal cascades as quasi-multiaffine random Veneziano, D. and A. K. Essiam, “Flow through porous media with multi-
processes,” Nonlinear Processes in Geophysics, 5: 63–67, 1998. fractal hydraulic conductivity,” Water Resources Research, 39: 1166, 2003.
Menabde, M. and M. Sivapalan, “Modeling of rainfall time series and doi:10.1029/2001WR001018, 6
extremes using bounded random cascades and Levy-stable distributions,” Veneziano, D. and P. Furcolo, “Marginal distribution of stationary multi-
Water Resources Research, 36: 3293–3300, 2000. fractal measures and their Haar wavelet coefficients,” Fractals-Complex
Menabde, M., D. Harris, A. Seed, G. Austin, and D. Stow, “Multiscaling Geometry Patterns and Scaling in Nature and Society, 11: 253–270, 2003.
properties of rainfall and bounded random cascades,” Water Resources Veneziano, D. and A. Langousis, Scaling and fractals in hydrology,
Research, 33: 2823–2830, 1997. Advances in Data-Based Approaches for Hydrologic Modeling and Forecasting,
Molz, F. J. and G. K. Bowman, “A fractal-based stochastic interpolation edited by B. Sivakumar and R. Berndtsson, World Scientific, 2010. Retrieved
scheme in subsurface hydrology,” Water Resources Research, 29: 3769–3774, from http://www.worldscientific.com/doi/abs/10.1142/9789814307987_0004
1993. Veneziano, D. and C. Lepore, “The scaling of temporal rainfall,” Water
Olsson, J., “Evaluation of a scaling cascade model for temporal rainfall Resources Research, 48: 2012. doi:10.1029/2012WR012105
disaggregation,” Hydrology and Earth System Sciences, 2: 19–30, 1998. Veneziano, D., P. Furcolo, and V. Iacobellis, “Imperfect scaling of time and
Olsson, J., J. Niemczynowicz, and R. Berndtsson, “Fractal analysis of high- space-time rainfall,” Journal of Hydrology, 322: 105–119, 2006.
resolution rainfall time-series,” Journal of Geophysical Research: Atmospheres, Voss, R. F., Fractals in nature: from characterization to simulation, The
98: 23265–23274, 1993. Science of Fractal Images, Springer-Verlag, New York, NY, 1988, pp. 21–70.
Ossiander, M. and E. C. Waymire, “Statistical estimation for multiplicative Willgoose, G., R. L. Bras, and I. Rodriguez-Iturbe, “A coupled channel
cascades,” Annals of Statistics, 28 (6): 1533–1560, 2000. network growth and hillslope evolution model: 1. Theory,” Water Resources
Ossiander, M. and E. C. Waymire, “On estimation theory for multiplicative Research, 27: 1671–1684, 1991.
cascades,” Sankhyā: The Indian Journal of Statistics, Series A, 64 (2): 323–343, Yaglom, A. M., Correlation Theory of Stationary and Related Random
2002. Functions, Springer, New York, NY, 1987.
Over, T. M. and V. K. Gupta, “A space-time theory of mesoscale rainfall Zahle, U., “Self-similar random measures .1. Notion, carrying hausdorff
using random cascades,” Journal of Geophysical Research: Atmospheres, 101: dimension, and hyberbolic distribution,” Probability Theory and Related
26319–26331, 1996. Fields, 80: 79–100, 1988.
28_Singh_ch28_p28.1-28.6.indd 6 8/22/16 1:48 PM

Chapter 29
Nonlinear Dynamics
and Chaos
BY
BELLIE SIVAKUMAR
ABSTRACT However, significant developments in computational power and major

advances in measurement technology and mathematical concepts over the
The inherent nonlinear nature of hydrologic systems has been known for
past three decades have facilitated formulation of nonlinear approaches as
several decades. Nonlinearity in hydrologic systems is evident in many differ-
viable alternatives for complex hydrologic systems. This subsequently led to
ent ways and at almost all spatial and temporal scales. “Chaos” is a special
applications of nonlinear approaches to study the nonlinear and related prop-
form of nonlinearity and generally refers to situations where simple nonlinear
erties of hydrologic systems. These applications started to emerge during
deterministic systems with sensitive dependence on initial conditions leading
1980s–1990s, and have skyrocketed since then. The nonlinear approaches
to complex and random-looking behavior. Since the fundamental properties
that are popular in hydrology include, among others, nonlinear stochastic
of chaotic systems (i.e., nonlinear interdependence, hidden determinism and
methods, data-based mechanistic models, artificial neural networks, support
order, and sensitivity to initial conditions) are highly relevant in hydrology,
vector machines, evolutionary computing, fuzzy logic, wavelets, entropy-
numerous studies, since the 1980s, have applied the concepts of chaos theory
based techniques, and chaos theory (e.g., Rodriguez-Iturbe et al., 1989; Young
to study hydrologic systems. The application areas and problems include
and Beven, 1994; Kumar and Foufoula-Georgiou, 1997; Singh, 1997; ASCE
rainfall, river flow, rainfall-runoff, lake volume, sediment transport, ground-
Task Committee, 2000; Sivakumar, 2000; Dibike et al., 2001; Kavvas, 2003;
water flow and contaminant transport, modeling, prediction, noise reduction,
Gupta et al., 2007; Şen, 2009); see Sivakumar and Berndtsson (2010) and
scaling, disaggregation, missing data estimation, reconstruction of system
Jayawardena (2014) for further details on the applications of many of these
equations, parameter estimation, catchment classification, and others. This
nonlinear approaches in hydrology. This chapter focuses on nonlinear deter-
chapter provides an overview of chaos theory and its applications in hydrol-
ministic dynamics, popularly known as chaos theory.
ogy. It includes: (1) a brief account of the history of the development of chaos
In the nonlinear science literature, the term “chaos” is normally used to
theory; (2) a description of some of the most important methods for chaos
refer to situations where complex and random-looking behavior arises from
identification and prediction; (3) a discussion on the potential issues in the
simple nonlinear deterministic systems with sensitive dependence on initial
applications of chaos theory methods to hydrologic (and other real) systems;
conditions (Lorenz, 1963); the converse also applies. The three fundamental
and (4) a brief review of the applications of chaos theory in hydrology.
properties inherent in this definition: (1) nonlinear interdependence;
29.1 INTRODUCTION
(2) hidden determinism and order; and (3) sensitivity to initial conditions are
highly relevant in hydrologic systems and processes. For example: (1) nonlinear
The inherent nonlinear nature of hydrologic systems has been known for sev- interactions and interdependence are dominant among the components and
eral decades now (e.g., Minshall, 1960; Izzard, 1966; Amorocho, 1967; Dooge, mechanisms in the hydrologic cycle; (2) determinism and order are prevalent
1967; Singh, 1979). Nonlinearity in hydrologic systems is evident in many in daily temperature, annual river flow, and many other hydrologic processes
different ways and at almost all spatial and temporal scales. The hydrologic at one or more scales; and (3) runoff is highly sensitive to the time at which
cycle itself is an example of a system exhibiting nonlinear behavior, with almost rainfall occurs and solute transport in surface and sub-surface waters is highly
all of the individual components exhibiting nonlinear behavior as well. The sensitive to the time at which contaminants are released. The first property
climatic inputs and landscape characteristics are changing in a highly nonlin- represents the “general” nature of hydrologic phenomena, whereas the second
ear fashion, and so are the outputs, often in unknown ways. The rainfall-runoff and third represent their “deterministic” and “stochastic” natures, respectively.
process is nonlinear, almost regardless of the basin area, land use, rainfall Further, despite their complexity and random-looking behavior, hydrologic
intensity, and other influencing factors; see Singh (1988) for a comprehensive phenomena may also be governed by only a very few degrees of freedom (e.g.,
review of earlier black-box and conceptual models of nonlinear rainfall-runoff runoff in a well-developed urban catchment depends essentially on rainfall),
process. another fundamental idea of chaos theory (Sivakumar, 2004a).
Although the nonlinear nature of hydrologic systems had been recognized The finding that “complex and random-looking” behavior is not necessarily
at least as early as the 1960s, much of the early research during 1960s–1980s, the outcomes of complex systems but can also be from simple nonlinear deter-
including development and application of time series models, essentially ministic systems with sensitivity to initial conditions has far reaching implica-
resorted to linear (deterministic and stochastic) approaches (e.g., Thomas tions in hydrologic modeling, since most outputs from such systems (e.g., time
and Fiering, 1962; Crawford and Linsley, 1966; Freeze and Harlan, 1969; series of rainfall, river flow, water quality) are typically “complex and random-
Yevjevich, 1972; Valencia and Schaake, 1973; Fleming, 1975; Salas and Smith, looking.” One crucial implication of this finding is the need, first of all, to
1981; Bras and Rodriguez-Iturbe, 1985), which continue to be prevalent in identify the dynamic nature of the given system towards selecting an appropri-
hydrology. At least two factors contributed to this situation: (1) the assump- ate modeling approach, as opposed to simply resorting to a particular
tion that linear approaches are generally sufficient in view of simplicity and approach based on certain preconceived notion. In view of this, and particu-
relative to other uncertainties; and (2) lack of computational power to develop larly facilitated by the development of new nonlinear time series methods,
the (perhaps more complex) nonlinear mathematical models. applications of nonlinear dynamics and chaos concepts in hydrology started to
29-1
29_Singh_ch29_p29.1-29.12.indd 1 8/22/16 1:51 PM

29-2 Nonlinear Dynamics and Chaos
emerge in the 1980s (e.g., Hense, 1987; Rodriguez-Iturbe et al., 1989), and have The positive outcomes from these laboratory experiments encouraged
seen an enormous growth since then (e.g., Wilcox et al., 1991; Berndtsson search for chaos outside the “controlled” space—in nature. To this end,
et al., 1994; Jayawardena and Lai, 1994; Puente and Obregon, 1996; Porporato advances in computational power and measurement technology facilitated
and Ridolfi, 1997; Sivakumar et al., 1999a, 2001a, 2001b, 2014; Lambrakis et al., development, in the 1980s, of a new set of mathematical techniques for chaos
2000; Sivakumar, 2002, 2003; Regonda et al., 2004; Dodov and Foufoula- identification and prediction based on time series, with concepts of data
Georgiou, 2005; Hossain and Sivakumar, 2006; Dhanya and Nagesh Kumar, reconstruction, dimensionality, entropy, and predictability (e.g., Packard
2010, 2011; Kyoung et al., 2011; Sivakumar and Singh, 2012). Comprehensive et al., 1980; Takens, 1981; Grassberger and Procaccia, 1983a, b; Wolf et al.,
accounts of such applications can be found in Sivakumar (2000, 2004a, 2009). 1985; Farmer and Sidorowich, 1987). These techniques have, since then, been
The application areas and problems include rainfall, river flow, rainfall-runoff, employed for identification and prediction of chaos in many real systems,
lake volume, sediment transport, groundwater contaminant transport, including atmospheric, biologic, ecologic, economic, engineering, environ-
modeling, prediction, noise reduction, scaling, disaggregation, missing data mental, financial, political, and social systems.
estimation, reconstruction of system equations, parameter estimation, and
catchment classification, among others.
The outcomes of such studies are encouraging, as they generally reveal the 29.3 CHAOS CONCEPTS AND IDENTIFICATION METHODS
adequacy of simpler models to represent hydrologic systems and the possibil- In the context of nonlinear dynamics and chaos, a wide variety of methods
ity of more accurate short-term predictions. Despite this, there continue to have been developed for reconstruction of a time series, estimation of the
be criticisms and skepticisms on such studies and outcomes, essentially on invariants for chaos identification, and prediction. These methods include
the basis of potential limitations in the applications of chaos methods for phase space reconstruction, correlation dimension method, Kolmogorov
hydrologic time series, such as issues associated with data size, data noise, entropy method, Lyapunov exponent method, false nearest neighbor algo-
zeros, and others (e.g., Schertzer et al., 2002; Koutsoyiannis, 2006). While rithm, nonlinear local approximation prediction method, close returns plot,
such criticisms and skepticisms cannot be dismissed altogether and indeed surrogate data method (for detection of nonlinearity), and 0-1 test, among
need careful consideration, there are also concerns about potential “false- others. Some of the more popular methods are briefly described here. To put
alarms” arising from such criticisms; see Sivakumar et al. (2002a) and the methods in a proper context, a simple demonstration of their utility for
Sivakumar (2005) for details. Amid all of this, what is clear is that chaos identification of chaos (or distinguishing between stochastic and chaotic time
studies and their outcomes certainly provide different perspectives and new series) is also presented. Due to space constraints, however, results for only
avenues to study hydrologic systems. In fact, arguments as to the potential of two of the methods are presented.
chaos theory to serve as a bridge between our traditional and dominant
deterministic and stochastic theories have also been put forward (Sivakumar, 29.3.1 Phase-Space Reconstruction
2004a, 2011c).
This chapter aims to provide an overview of chaos theory and its applica- Phase space is a useful concept for representing the evolution of a dynamic
tions in hydrology. The rest of the chapter is organized as follows. Section 29.2 system (e.g., Packard et al., 1980). It is essentially a graph, whose coordinates
presents briefly the history of the development of chaos theory. Section 29.3 represent the variables necessary to completely describe the state of the sys-
describes some of the most important methods for chaos identification and tem at any moment. The trajectories of the phase space diagram describe the
prediction, especially those that have found applications in hydrology. evolution of the system from some initial state and, hence, represent the his-
Section 29.4 discusses some of the potential issues in the applications of chaos tory of the system. The “region of attraction” of these trajectories in the phase
methods to real systems. Section 29.5 reviews the applications of chaos theory space provides at least important qualitative information on the nature of the
in hydrology. Section 29.6 offers some final thoughts toward the future. system dynamic properties.
Phase space can be reconstructed based on a single-variable data series or
multivariable data series (in the latter case, it is called state space). The idea
29.2 CHAOS THEORY: A BRIEF HISTORY behind this concept is that a nonlinear system is characterized by self-
The roots of chaos theory date back to the studies, in about 1900, of Henri interactions, and that data of a single (or multi) variable should carry
Poincaré on the problem of the motion of three objects in mutual gravita- sufficient information about the entire, and often large-dimensional, system.
tional attraction: the “three-body problem.” Poincaré found that there can be Many methods are available for phase space reconstruction from an available
orbits that are non-periodic, and yet not forever increasing nor approaching time series, including the Whitney’s embedding theorem, Takens’ delay
a fixed point. However, chaos theory remained in the background for almost embedding theorem, fractal delay embedding prevalence theorem, and
half a century. The invention of high-speed computers in the 1950s changed filtered delay embedding (such as singular value decomposition or singular
this situation, as computers allowed better experimentation with equations, spectrum analysis or principal component analysis); see Sauer et al. (1991) for
especially the process of repeated iteration of mathematical formulas to study a comprehensive account. However, the Takens’ delay embedding theorem is
nonlinear dynamic systems. the most widely used one in chaos identification studies, and is described
Such experiments led to Edward Lorenz’s discovery, in 1963, of chaotic here.
motion on a “strange attractor” (Lorenz, 1963). Lorenz studied a simplified Given a single variable series, Xi, where i = 1, 2, ..., N, a multi-dimensional
model of convection rolls in the atmosphere to obtain insight into the unpre- phase space can be reconstructed according to the Takens’ delay embedding
dictability of the weather. He found that the solutions to his equations never theorem as follows (Takens, 1981):
settled down to equilibrium or to a periodic state, but instead continued to
oscillate in an irregular, aperiodic fashion. Furthermore, when the simula- Yj = (Xj, Xj+t , Xj+2t ,..., Xj+(m–1) t ) (29.1)
tions were started from two slightly different initial conditions, the resulting
where j = 1, 2,..., N – 1(m – 1)t ; m is the dimension of the vector Yj, called
behaviors became totally different. The implication was that the system was
embedding dimension; and t is an appropriate delay time. A correct phase
inherently unpredictable, as tiny errors in measuring the current state of the
space reconstruction in a dimension m generally allows interpretation of the
atmosphere could amplify rapidly. However, Lorenz also showed that there
system dynamics (if the variable chosen to represent the system is appropri-
was structure (in the chaos), as the solutions to his equations, when plotted in
ate) in the form of an m-dimensional map fT, given by
three dimensions, fell onto a butterfly-shaped set of points.
The 1970s witnessed the main developments in chaos theory. Ruelle and Yj+T = fT (Yj) (29.2)
Takens (1971) proposed a new theory for the onset of turbulence in fluids,
based on abstract consideration about “strange attractors.” May (1976) found where Yj and Yj+T are vectors of dimension m, describing the state of the system
examples of chaos in iterated mappings arising in population biology, and at times j (current state) and j + T (future state), respectively. With Eq. (29.2), the
stressed on the pedagogical importance of studying simple nonlinear systems, task is basically to find an appropriate expression for fT (e.g., FT) to predict the
to counterbalance the often misleading linear intuition fostered by traditional future.
education. Study of other simple nonlinear models, such as the Henon map The utility of phase space diagram for system identification is demonstrated
(Henon, 1976) and the Rössler system (Rössler, 1976), also revealed the hid- here through its reconstruction for two artificially generated series (Fig. 29.1a
den beauty of chaos. Feigenbaum (1978) discovered that there are certain and b) that look very much alike (both “complex” and “random”) but neverthe-
universal laws governing the transition from regular to chaotic behavior, and less are the outcomes of systems (equations) possessing significantly different
established a link between chaos and phase transitions. The study of chaos dynamic characteristics. The first set (Fig. 29.1a) is the outcome of a pseudo
then moved to the laboratory, with ingenious experiments set up and chaotic random number generation function
behavior studied in fluids, mechanical oscillators, and semiconductors (e.g.,
Swinney and Gollub, 1978; Linsay, 1981). Xi = rand( ) (29.3)
29_Singh_ch29_p29.1-29.12.indd 2 8/22/16 1:51 PM

CHAOS CONCEPTS AND IDENTIFICATION METHODS 29-3
which yields independent and identically distributed numbers (between 0 Many algorithms have been formulated for the estimation of the correla-
and 1). The second set (Fig. 29.1b) is the outcome of a fully deterministic tion dimension of a time series, but the Grassberger-Procaccia algorithm
two-dimensional map (Henon, 1976): (Grassberger and Procaccia, 1983a) has been the most popular. The algorithm
uses the concept of phase space reconstruction for representing the dynamics
Xi+1 = a – Xi2 + bYi; Yi+1 = X (29.4) of the system from an available single-variable time series (Eq. 29.1). For an
which yields irregular solutions for many choices of a and b, but for a = 1.4 m-dimensional phase space, the correlation function, C(r), is given by
and b = 0.3, a typical sequence of Xi is chaotic.
2
Figure 29.1c and d presents the phase space plots for these two series. These C(r ) = lim
N →∞ N ( N − 1)
∑ H (r − | Yi − Yj |) (29.5)
diagrams correspond to reconstruction in two dimensions (m = 2) with delay i, j
(1≤i < j ≤ N )
time t = 1, that is the projection of the attractor on the plane {Xi, Xi+1}. For the
first set, the points (of trajectories) are scattered all over the phase space (i.e., where H is the Heaviside step function, with H(u) = 1 for u > 0, and H(u) = 0 for
absence of an attractor), a clear indication of a “complex” and “random” nature u ≤ 0, where u = r – ||Yi – Yj||, r is the vector norm (radius of sphere) centered
of the underlying dynamics and potentially of a high-dimensional (and pos- on Yi or Yj. If the time series is characterized by an attractor, then C(r) and r are
sibly random) system. On the other hand, the projection for the second set related according to
yields a very clear attractor (in a well-defined region), indicating a “simple” and
“deterministic” (yet non-repeating) nature of the underlying dynamics and C(r ) ≈ α r ν (29.6)
r →0
potentially of a low-dimensional (and possibly chaotic) system. N →∞
29.3.2 Correlation Dimension Method where a is a constant and n is the correlation exponent or the slope of the Log
The dimension of a time series is, in a way, a representation of the number of C(r) versus Log r plot. The slope is generally estimated by a least square fit of
dominant variables present in the evolution of the corresponding dynamic a straight line over a certain range of r (scaling regime) or through estimation
system. Correlation dimension is a measure of the extent to which a data of local slopes between r values.
point affects the position of the other points of the attractor in phase space. The distinction between low-dimensional (and perhaps deterministic) and
The correlation dimension method uses the correlation integral (or function) high-dimensional (and perhaps stochastic) systems can be made using the n
for determining the dimension of the attractor and, hence, for distinguishing versus m plot. If n saturates after a certain m and the saturation value is low,
between low-dimensional and high-dimensional systems. then the system is generally considered to exhibit low-dimensional deterministic
1.00
1.35
0.80 0.81
0.60
Value, X
Value, X
0.27
0.40 –0.27
0.20 –0.81
0.00 –1.35
0 200 400 600 800 1000 0 200 400 600 800 1000
(a) Time (b) Time
1.00
1.35
0.80 0.81
0.60 0.27
Xi+1
Xi+1
0.40 –0.27
0.20 –0.81
0.00 –1.35
0.00 0.20 0.40 0.60 0.80 1.00 –1.35 –0.81 –0.27 0.27 0.81 1.35
(c) Xi (d) Xi
10.0 1.4
Correlation exponent
Correlation exponent
8.0
1.2
6.0
1.0
4.0
0.8
2.0
0.0 0.6
0 2 4 6 8 10 0 2 4 6 8 10
(e) Embedding dimension (f) Embedding dimension
Figure 29.1 Random versus chaotic data: (a) and (b) time series; (c) and (d) phase space; and (e) and ( f ) correlation dimension.
29_Singh_ch29_p29.1-29.12.indd 3 8/22/16 1:51 PM

dynamics. The saturation value of n is defined as the correlation dimension (d) the two points Y and Z evolve until their distance, L¢1, is greater than some
of the attractor, and the nearest integer above this value is generally an indica- arbitrarily small e. The Y value at k becomes Y2 and a new nearest neighbor,
tion of the number of variables dominantly governing the dynamics. On the Z2, is selected. This procedure continues until the fiducial trajectory reaches
other hand, if n increases without bound with increase in m, the system under the end of the time series. The replacement of the old point by its substitute
investigation is generally considered to exhibit high-dimensional stochastic point and the replacement of the error direction by a new directional vector
behavior. constitutes a renormalization of errors along the trajectory. The largest
To demonstrate the utility of the dimension concept, Fig. 29.1e presents the Lyapunov exponent, λ1, is then given by
correlation dimension results for the first set (Fig. 29.1a), whereas those for
the second set (Fig. 29.1b) are shown in Fig. 29.1f. In each case, embedding 1
M
L′j
λ1 = ∑ log 2 (29.9)
dimensions from 1 to 10 are used for phase space reconstruction. It is clear N∆t j =1 Lj
that the first set is the outcome of an infinite-dimensional system (i.e., absence
of saturation in dimension), whereas the second set is the outcome of a low- where M is the number of replacement steps (where some arbitrarily small e
dimensional system (with a correlation dimension value of 1.22). is exceeded) and N is the total number of timesteps that the fiducial trajectory
progressed.
29.3.3 Kolmogorov Entropy Method
The Kolmogorov entropy, or K entropy, is the mean rate of information cre- 29.3.5 False Nearest Neighbor Algorithm
ated by the system. It is important in characterizing the average predictability The false nearest neighbor (FNN) algorithm (Kennel et al., 1992) provides
of a system of which it represents the sum of the positive Lyapunov exponents information on the minimum embedding dimension of the phase space
(see Sec. 29.3.4 for details of Lyapunov exponent). The Kolmogorov entropy required for representing the system dynamics. It examines, in dimension m,
quantifies the average amount of new information on the system dynamics the nearest neighbor YjNN of every vector Yj, as it behaves in dimension m + 1.
brought by the measurement of a new value of the time series. In this sense, it If the vector YjNN is a true neighbor of Yj, then it comes to the neighborhood
measures the rate of information produced by the system. The value of K = 0 of Yj through dynamical origins. On the other hand, if the vector YjNN moves
for periodic or quasi-periodic (i.e., completely predictable) time series; K = ∞ far away from vector Yj as the dimension is increased, then it is declared a
for white noise (i.e., unpredictable by definition); and 0 < K < ∞ for chaotic “false nearest neighbor” as it arrived in the neighborhood of Yj in dimension
system. m by projection from a distant part of the attractor. When the percentage of
It is important to note that it is very difficult to directly estimate K entropy these false nearest neighbors drops to zero, the geometric structure of the
for a time series representing a system with greater than two degrees of free- attractor has been unfolded and the orbits of the system are now distinct and
dom. Therefore, in the study of chaos, the only alternative is to estimate an do not cross (or overlap).
approximation of the K entropy. This approximation is oftentimes the lower A key step in the false nearest neighbor algorithm is to determine how to
bound of the K entropy, which is the K2 entropy. Grassberger and Procaccia decide upon increasing the embedding dimension that a nearest neighbor is
(1983b) proposed the first algorithm for estimation of Kolmogorov entropy in false. Two criteria are generally used (e.g., Sangoyomi et al., 1996):
the context of chaos analysis. Their algorithm uses the correlation sum or
correlation integral, and the K2 entropy is given as follows: • If Rm+1( j) ≥ 2RA, the jth vector has a false nearest neighbor, where Rm+1(j)
is the distance to the nearest neighbor of the jth vector (i.e., YjNN) in an
1  embedding of dimension (m + 1), and RA is the standard deviation of the
K 2 (m) = lim  {log[Cm (r )] − log[Cm+1 (r )]} (29.7) time series Xi, i = 1, 2, …, N.
r →0  τ 
and • If [Rm+1( j) – Rm( j)] > eRm( j), the jth vector has a false nearest neighbor,
where e is a threshold factor (generally between 10 and 50), and the dis-
K 2 = lim[K 2 (m)] (29.8) tance Rm+1( j) is computed to the same neighbor that was identified with
m→∞
embedding m, but with the (m + 1)th coordinate (i.e., Xj-mt appended to
where t is the delay time, Cm(r) is the value of C(r) when the embedding
the jth vector and to its nearest neighbor with embedding m).
dimension is m, Cm+1(r) is the value of C(r) when embedding dimension of
phase space is m + 1. Again, the choice of t and m is key to calculation of The first criterion is needed because with a finite dataset, such as hydrologic
entropy. In practice, we need to consider dimension of embedding phase series, under repeated embedding, one may stretch out the points such that
space as well as t which has better simulating effect. The K2 entropy and they are far apart and yet cannot move any farther apart upon increasing the
Kolmogorov entropy are thought to have the same qualitative behavior. For dimension. The second criterion checks whether the nearest neighbors have
chaotic systems with noise, a variation to the above, known as the modified moved far apart on increasing the dimension. The appropriate threshold e is
correlation entropy, has also been proposed (Jayawardena et al., 2010). generally selected through experimentation.
29.3.4 Lyapunov Exponent Method 29.3.6 Nonlinear Local Approximation
Prediction Method
Lyapunov exponents are the average exponential rates of divergence (expan-
sion) or convergence (contraction) of nearby orbits in the phase space. Since An important purpose of identification of chaos in a time series is to attempt
nearby orbits correspond to nearly identical states, exponential orbital diver- more reliable short-term predictions than those possible with other methods
gence means that systems whose initial differences that may not be possible (e.g., stochastic methods), since the deterministic nature of chaotic systems
to resolve will soon behave quite differently; in other words, predictive ability allows reliable short-term predictions, although their sensitive dependence
is rapidly lost. Any system containing at least one positive Lyapunov exponent on initial conditions precludes long-term ones. An early method for predic-
is considered to exhibit chaotic behavior, with the magnitude of the exponent tion of chaotic time series was proposed by Farmer and Sidorowich (1987).
reflecting the timescale on which system dynamics become unpredictable. A Since then, there have been many advances (e.g., Casdagli, 1989, 1992;
negative Lyapunov exponent indicates that the orbit is stable and periodic. A Sugihara and May, 1990), not only for prediction but also for chaos identifica-
zero Lyapunov exponent is an indication of a marginally or neutrally stable tion using the prediction results. All these methods are essentially based on
orbit, which often occurs near a point of bifurcation. An infinite Lyapunov “local approximation.”
exponent value is an indication of a stochastic system. The first step in chaos prediction method is phase space reconstruction of
Many algorithms have been formulated for calculation of the Lyapunov a time series according to Eq. (29.1) to reliably represent the underlying
exponents from a time series (e.g., Wolf et al., 1985; Eckmann et al., 1986; dynamics in the form of an m-dimensional map fT (Eq. 29.2). The problem
Rosenstein et al., 1993; Kantz, 1994). Among these, the algorithm by Wolf then is to find an appropriate expression for fT (e.g., FT ) to predict the future.
et al. (1985) has been most widely used. The algorithm tracks a pair of arbi- There are several possible approaches for determining FT . One promising
trarily close points over a trajectory to estimate the accumulated error per approach is the “local approximation method” (Farmer and Sidorowich,
timestep. The points are separated in time by at least one orbit on the attrac- 1987), which uses only nearby states to make prediction. The basic idea in the
tor. The trajectory is defined by the fiducial and test trajectories. They are local approximation method is to break up the domain FT into local neighbor-
tracked for a fixed time period or until the distance between the two compo- hoods and fit parameters in each neighborhood separately. In this way, the
nents of the trajectory exceeds some specific value. In sequence, another test underlying system dynamics are represented step by step locally in the phase
point near the fiducial trajectory is selected and the estimation proceeds. The space. To predict Xj+T based on Yj (an m-dimensional vector) and past history,
end product is that the stretching and squeezing are averaged. k nearest neighbors of Yj are found on the basis of the minimum values of
The computation of a Lyapunov exponent proposed by Wolf et al. (1985) ||Yj – Yj′||, with j¢ < j. If only one such neighbor is considered, the prediction
can be explained as follows. Let us assume an initial data point, Y1, and its of Xj+T would be X ′j +T . For k number of neighbors, the prediction of Xj+T could
neighbor, Z1, that are L1 units apart. Over ∆t, a series of timesteps from 1 to k, be taken as an average of the k values of X ′j +T . The value of k is determined by
29_Singh_ch29_p29.1-29.12.indd 4 8/22/16 1:51 PM

ISSUES IN CHAOS IDENTIFICATION AND PREDICTION 29-5
trial and error. It is also relevant to note that k = 1 for m = 1 is a very limited phase space is crucial for reliable invariant estimates and predictions. The
case of prediction, and is similar to the “method of analogues,” originally delay embedding techniques, such as those of Takens’ embedding theorem
proposed by Lorenz (1969). In general, however, both k and m are varied to (Takens, 1981), widely used for phase space reconstruction involve the use of
find out the optimum predictions. a delay time, t.
The prediction accuracy can be evaluated using a variety of measures, such An appropriate delay time, t, for phase space reconstruction is necessary,
as correlation coefficient, root mean square error, and coefficient of effi- because only an optimum t gives the best separation of neighboring trajectories
ciency. In addition, direct time series plots and scatter diagrams can also be within the minimum embedding phase space, whereas an inappropriate t may
used to choose the best prediction results among a large combination of lead to unreliable outcomes (e.g., underestimation or overestimation of invari-
results achieved with different embedding dimensions (m) and number of ants). For example, if t is too small, then there is little new information
neighbors (k), and also different delay times. contained in each subsequent datum and the reconstructed attractor is com-
While the local approximation method primarily serves as a prediction pressed along the identity line. This situation is termed as redundance (Casdagli
tool, the prediction results can also be used to distinguish between determin- et al., 1991), and the result of which is an inaccurate estimation of the invariants,
istic and stochastic systems, i.e., an inverse way to identify chaos. In general, such as an underestimation of the correlation dimension (e.g., Havstad and
a high prediction accuracy from the local approximation method may be Ehlers, 1989). On the other hand, if t is too large, and the dynamics happen to
considered as an indication of a deterministic system, whereas a low predic- be chaotic, then all relevant information for the phase space reconstruction is
tion accuracy is expected if the dynamics are stochastic. However, this is not lost, since neighboring trajectories diverge, and averaging in time and/or space
a very effective way, since the predictions are strongly influenced by the two is no longer useful (Sangoyomi et al., 1996). This situation is termed as irrele-
main parameters involved (m and k) as well as the lead time (T) considered. vance (Casdagli et al., 1991), and this may result in an inaccurate estimation of
It should be noted that while the parameter m may be fixed a priori based on the invariants, such as an overestimation of the correlation dimension.
reliable available information (e.g., results from correlation dimension and In view of the importance of t, several methods and guidelines have been
false nearest neighbor methods), there is no reliable way to fix the parameter proposed in the literature for the selection of an appropriate t. These
k in advance. Furthermore, even if m is reliably known a priori, it would still approaches are based on series correlation (e.g., autocorrelation, mutual
be helpful to make the predictions for different m values, so that the optimum information, high-order correlations), phase space extension (e.g., fill factor,
m can be compared and cross-verified against the dimension obtained from wavering product, average displacement), and multiple autocorrelation and
other methods. non-bias multiple autocorrelation (e.g., Holzfuss and Mayer-Kress, 1986;
With results from the nonlinear local approximation prediction method at Fraser and Swinney, 1986; Albano et al., 1991; Rosenstein et al., 1994; Judd
hand, the following guidelines may be adopted to distinguish between sto- and Mees, 1998). Two of the more widely used methods are briefly discussed
chastic and chaotic systems: here: the autocorrelation function method and the mutual information
• Embedding dimension (m): If the time series exhibits deterministic chaos, method.
then the prediction accuracy would increase to its best with the increase The autocorrelation function method is the most commonly used method
in the embedding dimension up to a certain point (low value of m), called for delay time selection, for at least two reasons: (1) its computation is simple;
the optimal embedding dimension (mopt), and would remain close to its and (2) it is one of the most fundamental and standard statistical tools in any
best for embedding dimensions higher than mopt. For stochastic time time series analysis. Within the autocorrelation function method, there are
series, there would be no increase in the prediction accuracy with an several guidelines for the selection of t. For instance, Holzfuss and Mayer-Kress
increase in the embedding dimension and the accuracy would remain the (1986) recommended using a value of lag time (or index lag) at which the
same for any value of m (e.g., Casdagli, 1989). autocorrelation function first crosses the zero line. Schuster (1988) suggested
• Neighbors (k): Smaller number of neighbors would give the best predic- the use of the lag time at which the autocorrelation function attains 0.5, while
tions if the system dynamics are deterministic, whereas for stochastic Tsonis and Elsner (1988) suggested the selection of the lag time at which the
systems the best predictions are achieved when the number of neighbors autocorrelation function crosses 0.1.
is large. This approach is also called the deterministic versus stochastic Despite its widespread use, the appropriateness of the autocorrelation func-
algorithm (Casdagli, 1992). The idea behind this approach is that since tion method for the selection of t has been seriously questioned. For example,
small k represents local models, it would be more appropriate for Fraser and Swinney (1986) pointed out that the autocorrelation function
deterministic systems, while global models (large k) would be more appro- method measures only the linear dependence between successive points and,
priate for stochastic systems. However, if the best prediction is obtained thus, may not be appropriate for nonlinear dynamics. They suggested the use
using neither deterministic nor stochastic models but intermediate mod- of the local minimum of the mutual information, which measures the general
els (i.e., intermediate number of neighbors), then such a condition can be dependence, not just the linear dependence, between successive points. They
taken as an indication of chaotic behavior with some amount of noise in reasoned that if t is chosen to coincide with the first minimum of the mutual
the data or chaos of moderate dimension (e.g., Casdagli, 1992). information, then the recovered state vector would consist of components
• Lead time (T): For a given embedding dimension and for a given number that possess minimal mutual information between them, i.e., the successive
of neighbors, predictions in deterministic systems deteriorate considerably values in the time series are statistically independent but (also) without any
faster than in stochastic systems when the lead time is increased. This is redundancy. For a discrete time series, with Xi and Xi-t as successive values, for
due to the sensitivity of deterministic chaotic systems to initial conditions instance, the mutual information function, I(t), is computed according to
(Sugihara and May, 1990).
 P( X , X ) 
I (τ ) = ∑ P( Xi , Xi−τ )log 2  i i −τ
 (29.10)
29.4 ISSUES IN CHAOS IDENTIFICATION i ,i −τ  P ( Xi ) P( Xi−τ ) 
AND PREDICTION
where P(Xi) and P(Xi–t) are the individual probabilities of Xi and Xi–t, respec-
The reliability of the aforementioned for chaos identification and prediction tively, and P(Xi, Xi–t) is the joint probability density. The mutual information
in real-time series, such as hydrologic time series, has been under consider- method is a more comprehensive method of determining proper delay time
able debate. This is mainly due to some of the basic assumptions in the values (e.g., Tsonis, 1992). However, the method has the disadvantage of
development of the above methods and, therefore, their potential limitations requiring a large number of data, unless the dimension is small, and is com-
when applied to real data. For instance, almost all of the above methods have putationally cumbersome.
been developed with an inherent assumption that the time series is infinite For some attractors, it may not really matter which method is used for the
and noise-free, but real time series are always finite and often contaminated selection of t. For example, when applied to the Rössler system (Rössler,
by noise. Much of the criticism on chaos analysis, however, has been directed 1976), the autocorrelation function, the mutual information, and the
at the correlation dimension method. Part of the criticisms has been due to a correlation integral methods all provide a value of t approximately equal to
number of studies carried out on data requirements for dimension estimation, one-fourth of the mean orbital period (Tsonis, 1992). However, for some
but the fact that the correlation dimension method has been the most widely other attractors, the estimation of t might depend strongly on the approach
used method for chaos detection has also contributed to such criticisms. In employed. An obvious way to have more confidence in the selection of t
what follows, some of the important issues associated with chaos methods are may be to use different methods and check the consistency of the resulting
discussed. t values.
A reliable alternative to address the issue of delay time selection is to try to
29.4.1 Delay Time
fix the delay time window tw = t(m – 1), rather than just the delay time itself,
Almost all of the chaos identification and prediction methods involve the since the delay time window is the one that is of actual interest at the end to
reconstruction of the time series in phase space. An appropriately constructed represent the dynamics. An early attempt in this regard was made by
29_Singh_ch29_p29.1-29.12.indd 5 8/22/16 1:51 PM

Martinerie et al. (1992). Comparing the delay time window and delay times when it comes to real time series, as the properties (e.g., correlation dimen-
estimated using the autocorrelation function and mutual information meth- sion) are not known a priori, it is nevertheless still useful, if sufficient caution
ods, Martinerie et al. (1992) did not observe a consistent agreement between is exercised in its implementation. This approach has been adopted by many
them. This is because, tw is basically the optimal time for representing the studies, including in hydrology (e.g., Jayawardena and Lai, 1994; Wang and
independence of the data, whereas autocorrelation function and mutual Gan, 1998; Sivakumar et al., 1999a; Sivakumar, 2005).
information methods determine only the first local optimal times in their
estimation of t. Kugiumtzis (1996) put emphasis on the relation between tw 29.4.3 Data Noise
and dynamics of the underlying chaotic system and suggested to set tw > tp, Another important assumption in the development of chaos identification
the mean orbital period, with tp approximated from the oscillations of the and prediction methods is that the data are noise-free. However, all real data
time series. Kim et al. (1998), through analysis of time series generated from are contaminated by noise. This means, we usually observe a noisy time series
the Lorenz system, the Rabinovich-Fabrikant system, and the three-torus, Xi, i = 1, 2, …, N, which is composed of a clean signal Yi from a deterministic
showed that with an increase in the embedding dimension, the correlation system and some amount of noise ni, according to
dimension converges more rapidly for the case of tw held fixed than for the
case of t held fixed. Kim et al. (1999) subsequently developed a new technique Xi = Yi + ni (29.11)
to estimate both t and tw. This technique, called the C–C method, uses the There are two types of noise: measurement noise and dynamical noise.
Brock–Dechert–Scheinkman (BDS) statistic (Brock et al., 1996), which has its Measurement noise refers to the corruption of observations by errors, which
base on the correlation integral for testing nonlinearity in a time series. are external and independent of the dynamics, and may be caused by, for
In the absence of clear-cut guidelines on t selection, a practical approach is example, the measuring device. Dynamical noise, in contrast to measurement
to experiment with different t values to ascertain its effect, for example, on noise, is a feedback process wherein the system is perturbed by a small ran-
the estimation of invariants. Such an exercise is particularly fruitful for syn- dom amount at each timestep.
thetic time series, since the dynamic properties (e.g., invariants) of such series Noise affects the performance of many techniques of identification, model-
are known a priori. One has to be careful, however, in adopting this approach ing, prediction, and control of deterministic systems (e.g., Schreiber and
for real time series, since their dynamic properties are not known, and deter- Kantz, 1996; Kantz and Schreiber, 1997; Sivakumar et al., 1999b). The severity
mination of which is indeed the task at hand. Nevertheless, the exercise can of the influence of noise depends largely on the level and the nature of noise.
offer some important clues. In general, most dynamical measures of determinism are reasonably robust to
small levels of noise, but as the noise level approaches a few percent, estimates
29.4.2 Data Size can become quite unreliable (Schreiber and Kantz, 1996; Kantz and Schreiber,
The issue of data size (or length) has perhaps attracted the fiercest criticism 1997). Therefore, the estimation of the level of noise in data is an important
on studies employing chaos theory-based methods to real time series. A com- first step to understand its possible effects. Several attempts have been made
mon criticism is that the correlation dimension, arguably the most widely to determine the level of noise in a time series (e.g., Schreiber, 1993a; Schouten
used indicator of chaos, is underestimated when the data size is small. The et al., 1994; Heald and Stark, 2000; Jayawardena et al., 2008; Xu et al., 2012),
basis of this criticism lies with a basic assumption of the method that the time and have made different assumptions about the nature of noise, even when
series is infinite. The recognition that the effects of data size may be different using somewhat similar approaches. If it is found that the level of noise is only
for different chaos identification and prediction methods has led to studies to moderate, and there are hints that there is a strong deterministic component
estimate the minimum data requirement for different methods. However, in the signal, then one can attempt the second step of separating the determin-
most of such studies have been directed at the correlation dimension method istic signal from the noise. While this two-step procedure is the most appro-
(e.g., Smith, 1988; Havstad and Ehlers, 1989; Nerenberg and Essex, 1990; priate approach to assess the effects of noise and their mitigation (see
Ramsey and Yuan, 1990; Lorenz, 1991; Tsonis et al., 1993; Sivakumar et al., Sivakumar et al., 1999b), it is often difficult to implement this for real data,
2002c; Sivakumar, 2005), many of them attempting to link the minimum data since determination of the level of noise for such data is a very complicated
size (Nmin) to the embedding dimension (m) or correlation dimension (d). process. Therefore, many studies have employed only the second step, i.e.,
Some examples are presented here. noise reduction. A large number of nonlinear noise reduction methods have
The study by Smith (1988) was the first to address the minimum data size been proposed in the literature (e.g., Kostelich and Yorke, 1988; Schreiber and
for correlation dimension estimation in terms of embedding dimension. Smith Grassberger, 1991; Schreiber, 1993b; Luo et al., 2005; Chelidze, 2014). The
(1988) concluded that the minimum data size was equal to 42m, where m is the different noise reduction methods differ in the way the dynamics are approx-
smallest integer above the dimension of the attractor. Nerenberg and Essex imated, how the trajectory is adjusted, and how the approximation and the
(1990) demonstrated that the procedure by Smith (1988) was flawed and that adjustment steps are linked to each other; see Kostelich and Schreiber (1993)
the data requirements might not be so extreme. They suggested that the mini- for an early survey. It has been reported that most of these methods reduce
mum number of points required for the dimension estimate is Nmin ~ 102+0.4m. noise by a similar amount and their performances do not differ much (Kantz
Havstad and Ehlers (1989) used a variant of the nearest neighbor dimen- and Schreiber, 1997). However, noise reduction algorithms are generally cho-
sion algorithm to compute the dimension of the time series generated from sen on the basis of their robustness, ease of use and implementation, and the
the Mackey–Glass equation (Mackey and Glass, 1977), whose actual dimen- computing resources needed.
sion is 7.5. Using a dataset of as small as 200 points, Havstad and Ehlers
29.4.4 Presence of Zeros
(1989) reported an underestimation of the dimension by about 11%. Ramsey
and Yuan (1990) concluded that for small sample sizes, dimension could be Most of the issues associated with chaos identification and prediction meth-
estimated with upward bias for chaotic systems and with downward bias for ods are common to almost all fields. For instance, delay time and embedding
random noise as the embedding dimension is increased. They proved that, dimension for phase space reconstruction are inherent issues in the applica-
due to these bias effects, a correlation dimension estimate of 0.214 could tion of chaos methods to any real time series. Similarly, smaller data size and
imply an actual correlation dimension value of as high as 1.68. Lorenz (1991) presence of noise are common problems in dealing with observed time series,
argued that, while underestimation of correlation dimension for small sample regardless of the field of study. These are the reasons why these issues have
sizes may occur, different [climatic] variables yield different estimates and received considerable attention in the analysis of chaos in real time series.
that a suitably selected variable could yield a fairly accurate estimate of However, there are also other problems that have not received the neces-
dimension even if the number of points were not large. sary attention in chaos analysis, although their effects can be as serious as
A number of studies have followed and/or supported one or more of the those of the above issues. An important reason for this is that such issues are
above-mentioned guidelines as to the potential underestimation of the correla- essentially encountered in just one or a few fields. For instance, the issue of
tion dimension when the data size is small (e.g., Tsonis et al., 1993; Wang and the presence of a large number of zeros in time series is perhaps unique to the
Gan, 1998; Schertzer et al., 2002). However, several counter-arguments to these field of hydrology and water resources, as is the case with rainfall (and
guidelines have also been made (e.g., Sivakumar et al., 2002a, c; Sivakumar, streamflow) observations, especially at high temporal resolutions. One pos-
2005); see below for some details. As of now, a clear-cut guideline on the sible influence of the presence of a large number of zeros is that the recon-
minimum data size for the correlation dimension estimation continues to be structed hyper-surface in the phase space will tend to a point and may result
elusive. in an underestimation of the correlation dimension (e.g., Tsonis et al., 1994;
In the absence of clear-cut guidelines, a practical way to address the mini- Sivakumar, 2001a).
mum data size requirement is by decreasing (or increasing) the length of the It is not yet clear how to deal with the issue of a large number of zeros in a
time series step-by-step and estimating the correlation dimension. The length time series. It is even questionable whether zero values should be eliminated
of data below which significant changes are observed can be taken as the from the time series in chaos analysis for more reliable results, since such
minimum data size required. While this procedure may have some drawbacks values are also often indicative of, and important to understand, how the
29_Singh_ch29_p29.1-29.12.indd 6 8/22/16 1:51 PM

FINAL REMARKS 29-7
dynamics of the system evolve; there can, however, be exceptions depending other techniques, such as stochastic methods and artificial neural networks
upon the task at hand (e.g., Sivakumar et al., 2001b). In fact, removal of the (e.g., Jayawardena and Gurung, 2000; Lambrakis et al., 2000; Sivakumar et al.,
zeros can have serious effects and result in unrealistic outcomes. What is 2002b, c; Laio et al., 2003). Attempts were also made to analyze hydrologic
needed, therefore, is a careful understanding of the potential influences of the time series based on multi-variable phase space reconstruction (e.g., Porporato
inclusion or exclusion of the zero values with respect to the particular task at and Ridolfi, 2001; Jin et al., 2005; Sivakumar et al., 2005b), rather than recon-
hand. On the other hand, there may be alternative means to address the prob- struction based on just a single variable. There were also attempts to use other
lem of zeros in chaos analysis, such as the verification of the results from one data analysis methods for prediction of hydrologic time series that exhibit
method with that of another. chaotic behavior (e.g., Karunasinghe and Liong, 2006).
The significant inroads made in the application of chaos theory in hydrol-
ogy over two decades led to an increasing realization in more recent years on
29.5 HYDROLOGIC APPLICATIONS
the opportunities and possibilities for studying large-scale problems. It is fair
Applications of the ideas of nonlinear dynamic and chaos theories in hydrol- to say that this realization has come at a critical juncture, as we are (and will
ogy have been one of the most exciting areas of hydrologic research during be in the future) facing some tremendous global-scale challenges in hydrol-
the past two decades or so. Extensive details of such applications and the ogy and water resources, such as assessment of impacts of global climate
associated issues are reported in Sivakumar (2000, 2004a, 2009); see also change, development of a generic modeling framework, study of interactions
Phillips (2003, 2006) for some details of nonlinearity, complexity, and chaos between hydrologic systems and other Earth and socio-economic systems,
in geomorphology. Here, only a brief overview of chaos studies in hydrology and study of issues related to transboundary waters (river basins as well as
is presented. aquifers), among others. Details of these challenges are available in, for
The first ever study on chaos theory application in a hydrologic context was example, Paola et al. (2006), Wagener et al. (2010), Sivakumar (2011a, b), and
probably the one conducted by Hense (1987) on rainfall time series, although Sivakumar and Singh (2012); see also Sivakumar and Singh (2015) for compi-
chaos in rainfall time series had already been investigated previously in the lation of some of the grand challenges in hydrology. During the last few years,
context of climate and weather (e.g., Fraedrich, 1986). It is fair to say, however, some initial attempts have been made to apply the ideas of chaos theory to
that it was the study by Rodriguez-Iturbe et al. (1989), which also investigated study these global-scale hydrologic challenges, especially for catchment clas-
the presence of chaos in rainfall time series, that attracted many researchers sification and global climate change. For instance, Sivakumar et al. (2007)
about chaos theory in hydrology and was a significant motivation to many explored the utility of phase space reconstruction for assessing the complexity
other studies that followed during the subsequent period. Studies during the of hydrologic systems and thus their classification, through analysis of several
early 1990s focused mainly on the investigation and prediction of chaos in real river-related hydrologic series observed around the world. Sivakumar
rainfall, river flow, and lake volume time series in a purely single-variable data and Singh (2012) offered a more comprehensive scientific background and
reconstruction sense (e.g., Sharifi et al., 1990; Wilcox et al., 1991; Berndtsson discussion for proposing system complexity as a basis for catchment classifi-
et al., 1994; Jayawardena and Lai, 1994; Abarbanel and Lall, 1996; Puente and cation framework and nonlinear dynamic concepts as a suitable methodology
Obregon, 1996; Sangoyomi et al., 1996; Porporato and Ridolfi, 1996). These for assessing system complexity. They also presented a classification of 117
early studies employed various chaos identification and prediction methods, catchments in the western United States, based on the application of correla-
including correlation dimension method, Lyapunov exponent method, tion dimension method to monthly streamflow data. Kyoung et al. (2011)
Kolmogorov entropy method, nonlinear prediction method, and Poincáre examined the dynamic characteristics of rainfall under conditions of climate
maps. Some of these studies and several others that followed also addressed change, through analysis of observed and global climate model (GCM)-
important methodological and data issues, including minimum data size simulated (present and future) monthly rainfall in the Korean Peninsula.
requirement for correlation dimension estimation (e.g., Tsonis et al., 1993; Based on the application of the correlation dimension and close returns plot,
Kim et al., 1998; Sivakumar et al., 1998, 1999a), effects of data noise on chaos among others, they reported that the nature of rainfall dynamics falls more on
identification and prediction (e.g., Berndtsson et al., 1994; Tsonis et al., 1994; the chaotic dynamic spectrum than on the linear stochastic spectrum and
Porporato and Ridolfi, 1997; Sivakumar et al., 1999b, c), effects of the pres- also that future GCM-simulated rainfall exhibits stronger nonlinearity and
ence of zeros on chaos identification and prediction (e.g., Tsonis et al., 1994; chaos compared to the present rainfall, thus emphasizing the need for a cha-
Koutsoyiannis and Pachakis, 1996; Wang and Gan, 1998; Sivakumar et al., otic dynamic-based framework for downscaling GCM outputs.
1999a), delay time/delay window selection for phase space reconstruction While studies on these global-scale hydrologic challenges using chaos
(e.g., Sangoyomi et al., 1996; Sivakumar et al., 1998, 1999a; Kim et al., 1998, theory and related concepts have been gaining momentum, other chaos stud-
1999; Pasternack, 1999), stochastic processes possibly leading to positive ies in hydrology have and continue to grow as well. These include studies on
chaos identification (e.g., Wang and Gan, 1998; Pasternack, 1999; Sivakumar rainfall (e.g., Jothiprakash and Fathima, 2013), regional streamflow variability
et al., 1999a), and others (e.g., Liu et al., 1998; Krasovskaia et al., 1999). and classification (e.g., Vignesh et al., 2015), sediment transport (e.g.,
At the very beginning of the twenty-first century, Sivakumar (2000) pub- Sivakumar and Chen, 2007), arsenic contamination in groundwater (e.g., Hill
lished the first ever review of chaos theory applications in hydrology, with et al., 2008), ensemble prediction of chaotic hydrologic time series (including
particular emphasis on addressing the important issues in chaos theory appli- multivariate prediction with climate inputs) (e.g., Dhanya and Nagesh Kumar,
cations in hydrology and also interpreting the reported results. The review 2010, 2011), river stage (Khatibi et al., 2012), and others (e.g., Sivakumar,
significantly helped allay many of the earlier fears and misgivings about chaos 2007; Kim et al., 2009), including advances in methodologies (e.g.,
studies in hydrology and their outcomes and completely changed the course of Jayawardena et al., 2008). A more recent account of chaos studies in hydrol-
chaos theory in hydrology. It also led to further advances in theory and appli- ogy is presented in Sivakumar (2009), which also reiterates the need for a
cation, with study of other hydrologic processes and associated problems. The middle-ground approach in hydrology and the role chaos theory can play in
processes studied include: rainfall-runoff (e.g., Sivakumar et al., 2000, 2001a; its formulation. Further philosophical and pragmatic arguments to this end
Dodov and Foufoula-Georgiou, 2005), sediment transport (e.g., Sivakumar, are made by Sivakumar (2011c).
2002; Sivakumar and Jayawardena, 2002, 2003; Sivakumar and Wallender,
2004, 2005), soil nutrient cycles (e.g., Manzoni et al., 2004), and subsurface 29.6 FINAL REMARKS
flow and solute transport (e.g., Faybishenko, 2002; Sivakumar et al., 2005a;
Hossain and Sivakumar, 2006). The hydrologic problems studied include: scal- Since its introduction in hydrology in the late 1980s, chaos theory has and
ing and data aggregation/disaggregation (e.g., Sivakumar, 2001a, b; Sivakumar continues to be a fascinating topic in hydrologic research. As a result, there
et al., 2001b, 2004; Regonda et al., 2004; Sivakumar and Wallender, 2004; Salas has been a tremendous growth both in our understanding of the concepts of
et al., 2005; Gaume et al., 2006) including the development of a new chaotic chaos theory and in their application areas. We have come a long way, starting
approach for disaggregation of hydrologic data (Sivakumar et al., 2001b), from simple (and perhaps sometimes blind) applications of chaos methods
missing data estimation (e.g., Elshorbagy et al., 2002a), reconstruction of sys- looking for something (or anything) to more-refined applications for better-
tem equations (e.g., Zhou et al., 2002), regional hydrology and river flow defined purposes at local scales and now with an aim to tackle global-scale
regimes (Sivakumar, 2003), parameter estimation (Hossain et al., 2004), and hydrologic challenges. Along the way, we have encountered several difficul-
model integration (e.g., Sivakumar, 2004b). A number of studies also ties, identified some potential problems, and resolved at least a few of them.
addressed the issues of data size, noise, zeros, selection of optimal parameters, There is also an increasing recognition that chaos theory, with its underpin-
and others (e.g., Jayawardena and Gurung, 2000; Sivakumar, 2001a, 2005; ning concepts of nonlinear interdependence, hidden determinism and order,
Islam and Sivakumar, 2002; Elshorbagy et al., 2002b; Jayawardena et al., 2002; and sensitivity to initial conditions, can encompass all the key ideas of our
Phoon et al., 2002; Schertzer et al., 2002; Sivakumar et al., 2002a, c, 2006; Khan currently dominant, yet extreme, deterministic and stochastic views.
et al., 2005; Salas et al., 2005; Koutsoyiannis, 2006). Some studies also com- Consequently, it can offer a more balanced and realistic approach for hydro-
pared hydrologic predictions based on chaos methods with those based on logic modeling, than either the deterministic or the stochastic approach can
29_Singh_ch29_p29.1-29.12.indd 7 8/22/16 1:51 PM

on their own. With issues around water resources starting to take a prominent Freeze, R. A. and R. L. Harlan, “Blueprint for a physically-based, digitally-
stage and as the need for a better, and more balanced, approach becoming simulated hydrologic response model,” Journal of Hydrology, 9: 237–258,
clear, now is indeed an exciting time for research on chaos in hydrology. 1969.
Gaume, E., B. Sivakumar, M. Kolasinski, and L. Hazoumé, “Identification
of chaos in rainfall disaggregation: application to a 5-minute point series,”
ACKNOWLEDGMENT
Journal of Hydrology, 328 (1–2): 56–64, 2006.
This work is supported by the Australian Research Council (ARC) Future Grassberger, P. and I. Procaccia, “Measuring the strangeness of strange
Fellowship grant (FT110100328). Bellie Sivakumar acknowledges the finan- attractors,” Physica D, 9: 189–208, 1983a.
cial support from ARC through this Future Fellowship grant. Grassberger, P. and I. Procaccia, “Estimation of the Kolmogorov entropy
from a chaotic signal,” Physical Review A, 28: 2591–2593, 1983b.
Gupta, V. K., B. M. Troutman, and D. R. Dawdy, “Towards a nonlinear geo-
REFERENCES
physical theory of floods in river networks: an overview of 20 years of progress,”
Abarbanel, H. D. I. and U. Lall, “Nonlinear dynamics of the Great Salt Lake: Nonlinear Dynamics in Geosciences, edited by A. A. Tsonis and J. B. Elsner
system identification and prediction,” Climate Dynamics, 12: 287–297, 1996. JB, Springer, New York, NY, 2007, pp. 121–152.
Albano, A. M., A. Passamante, and M. E. Farrell, “Using high-order correla- Havstad, J. W. and C. L. Ehlers, “Attractor dimension of nonstationary dynam-
tions to define an embedding window,” Physica D, 54 (1–2): 85–97, 1991. ical systems from small data sets,” Physical Review A, 39 (2): 845–853, 1989.
ASCE Task Committee on Applications of Artificial Neural Networks in Heald, J. P. M. and J. Stark, “Estimation of noise levels for models of chaotic
Hydrology, “Artificial neural networks in hydrology. II: Hydrologic applica- dynamical systems,” Physical Review Letters, 84 (11): 2366–2369, 2000.
tions,” Journal of Hydrologic Engineering, 5 (2): 124–137, 2000. Henon, M., “A two-dimensional mapping with a strange attractor,”
Amorocho, J., “The nonlinear prediction problems in the study of the run- Communications in Mathematical Physics, 50: 69–77, 1976.
off cycle,” Water Resources Research, 3 (3): 861–880, 1967. Hense, A., “On the possible existence of a strange attractor for the southern
Berndtsson, R., K. Jinno, A. Kawamura, J. Olsson, and S. Xu, “Dynamical oscillation,” Beitraege zur Physik der Atmosphaere, 60 (1): 34–47, 1987.
systems theory applied to long-term temperature and precipitation time Hill, J., F. Hossain, and B. Sivakumar, “Is correlation dimension a reliable
series,” Trends in Hydrology, 1: 291–297, 1994. proxy for the number of dominant influencing variables for modeling risk of
Bras, R. L. and I. Rodriguez-Iturbe, Random Functions and Hydrology, arsenic contamination in groundwater?” Stochastic Environmental Research
Addison-Wesley, Reading, MA, 1985. and Risk Assessment, 22 (1): 47–55, 2008.
Brock, W. A., W. A. Dechert, J. A. Scheinkman, and B. LeBaron, “A test for Holzfuss, J. and G. Mayer-Kress, “An approach to error-estimation in the
independence based on the correlation dimension,” Economic Review, 15 (3): application of dimension algorithms,” Dimensions and Entropies in Chaotic
197–235, 1996. Systems, edited by G. Mayer-Kress, Springer, New York, pp. 114–122.
Casdagli, M., “Nonlinear prediction of chaotic time series,” Physica D, 35: Hossain, F. and B. Sivakumar, “Spatial pattern of arsenic contamination in
335–356, 1989. shallow wells of Bangladesh: regional geology and nonlinear dynamics,”
Casdagli, M., “Chaos and deterministic versus stochastic nonlinear model- Stochastic Environmental Research and Risk Assessment, 20 (1–2): 66–76,
ing,” Journal of the Royal Statistical Society, Series B, 54 (2): 303–328, 1992. 2006.
Casdagli, M., S. Eubank, J. D. Farmer, and J. Gibson, “State space recon- Hossain, F., E. N. Anagnostou, and K. H. Lee, “A non-linear and stochastic
struction in the presence of noise,” Physica D, 51: 52–98, 1991. response surface method for Bayesian estimation of uncertainty in soil mois-
Chelidze, D., “Smooth local subspace projection for nonlinear noise reduc- ture simulation from a land surface model,” Nonlinear Processes in Geophysics,
tion,” Chaos, 24: 013121, 2014. 11: 427–440, 2004.
Crawford, N. H. and R. K. Linsley, “Digital simulation in hydrology,” Islam, M. N. and B. Sivakumar, “Characterization and prediction of runoff
Stanford Watershed Model IV, Department of Civil Engineering, Stanford dynamics: a nonlinear dynamical view,” Advances in Water Resources, 25 (2):
University, Technical Report 39, 1966. 179–190, 2002.
Dhanya, C. T. and D. Nagesh Kumar, “Nonlinear ensemble prediction of Izzard, C. F., “A mathematical model for nonlinear hydrologic systems,”
chaotic daily rainfall,” Advances in Water Resources, 33: 327–347, 2010. Journal of Geophysical Research, 71 (20): 4811–4824, 1966.
Dhanya, C. T. and D. Nagesh Kumar, “Multivariate nonlinear ensemble Jayawardena, A. W., Environmental and Hydrological Systems Modelling,
prediction of daily chaotic rainfall with climate inputs,” Journal of Hydrology, CRC Press, Taylor and Francis Group, Boca Raton, USA, 516 pp.
403: 292–306, 2011. Jayawardena, A. W. and A. B. Gurung, “Noise reduction and prediction of
Dibike, Y. B., S. Velickov, D. Solomatine, and M. B. Abbott, “Model induc- hydrometeorological time series: dynamical systems approach vs. stochastic
tion with support vector machines: introduction and applications,” Journal of approach,” Journal of Hydrology, 228: 242–264, 2000.
Computing in Civil Engineering, 15 (3): 208–216, 2001. Jayawardena, A. W. and F. Lai, “Analysis and prediction of chaos in rainfall
Dodov, B. and E. Foufoula-Georgiou, “Incorporating the spatio-temporal and stream flow time series,” Journal of Hydrology, 153: 23–52, 1994.
distribution of rainfall and basin geomorphology into nonlinear analysis of Jayawardena, A. W., W. K. Li, and P. Xu, “Neighborhood selection for local
streamflow dynamics,” Advances in Water Resources, 28 (7): 711–728, 2005. modeling and prediction of hydrological time series,” Journal of Hydrology,
Dooge, J. C. I., “A new approach to nonlinear problems in surface water 258: 40–57, 2002.
hydrology: hydrologic systems with uniform nonlinearity,” International Jayawardena, A. W., P. Xu, and W. K. Li, “A method of estimating the noise
Association of Scientific Hydrology Bulletin, 76: 409–413, 1967. level in a chaotic time series,” Chaos, 18 (2): 023115, 2008, doi:10.1063/
Eckmann, J. P., S. O. Kamphorst, D. Ruelle, and S. Ciliberto, “Lyapunov 1.2903757.
exponents from a time series,” Physical Review A, 34: 4971–4979, 1986. Jayawardena, A. W., P. Xu, and W. K. Li, “Modified correlation entropy
Elshorbagy, A., S. P. Simonovic, and U. S. Panu, “Estimation of missing estimation for a noisy chaotic time series,” Chaos, 20: 023104, 2010.
streamflow data using principles of chaos theory,” Journal of Hydrology, 255: Jin, Y. H., A. Kawamura, K. Jinno, and R. Berndtsson, “Nonlinear multi-
123–133, 2002a. variate analysis of SOI and local precipitation and temperature,” Nonlinear
Elshorbagy, A., S. P. Simonovic, and U. S. Panu, “Noise reduction in chaotic Processes in Geophysics, 12: 67–74, 2005.
hydrologic time series: facts and doubts,” Journal of Hydrology, 256: 147–165, Jothiprakash, V. and T. A. Fathima, “Chaotic analysis of daily rainfall series
2002b. in Koyna reservoir catchment area,” Stochastic Environmental Research and
Farmer, D. J. and J. J. Sidorowich, “Predicting chaotic time series,” Physical Risk Assessment, 27 (6): 1371–1381, 2013.
Review Letters, 59: 845–848, 1987. Judd, K. and A. I. Mees, “Embedding as a modelling problem,” Physica D,
Faybishenko, B., “Chaotic dynamics in flow through unsaturated fractured 120: 273–286, 1998.
media,” Advances in Water Resources, 25 (7): 793–816, 2002. Kantz, H., “A robust method to estimate the maximal Lyapunov exponent
Feigenbaum, M. J., “Quantitative universality for a class of nonlinear trans- from a time series,” Physics Letters A, 185: 77–87, 1994.
formations,” Journal of Statistical Physics, 19: 25–52, 1978. Kantz, H. and T. Schreiber, Nonlinear Time Series Analysis, Cambridge
Fleming, G., Computer Simulation Techniques in Hydrology, Elsevier, New University Press, Cambridge, 1997.
York, USA, 1975. Kavvas, M. L., “Nonlinear hydrologic processes: conservation equations for
Fraedrich, K., “Estimating the dimensions of weather and climate attrac- determining their means and probability distributions,” ASCE Journal of
tors,” Journal of the Atmospheric Sciences, 43: 419–432, 1986. Hydrologic Engineering, 8 (2): 44–53, 2003.
Fraser, A. M. and H. L. Swinney, “Independent coordinates for strange Karunasinghe, D. S. K. and S. Y. Liong, “Chaotic time series prediction with
attractors from mutual information,” Physical Review A, 33 (2): 1134–1140, a global model: artificial neural network,” Journal of Hydrology, 323: 92–105,
1986. 2006.
29_Singh_ch29_p29.1-29.12.indd 8 8/22/16 1:51 PM

REFERENCES 29-9
Kennel, M. B., R. Brown, and H. D. I. Abarbanel, “Determining embedding Paola, C., E. Foufoula-Georgiou, W. E. Dietrich, M. Hondzo, D. Mohrig,
dimension for phase space reconstruction using a geometric method,” G. Parker, M. E. Power, et al., “Toward a unified science of the Earth’s surface:
Physical Review A, 45: 3403–3411, 1992. opportunities for synthesis among hydrology, geomorphology, geochemistry,
Khan, S., A. R. Ganguly, and S. Saigal, “Detection and predictive modeling and ecology,” Water Resources Research, 42: W03S10, 2006, doi:10.1029/
of chaos in finite hydrological time series,” Nonlinear Processes in Geophysics, 2005WR004336.
12: 41–53, 2005. Pasternack, G. B., “Does the river run wild? Assessing chaos in hydrological
Khatibi, R., B. Sivakumar, M. A. Ghorbani, Ö Kisi, K. Kocak, and D. F. Zadeh, systems,” Advances in Water Resources, 23 (3): 253–260, 1999.
“Investigating chaos in river stage and discharge time series,” Journal of Phillips, J. D., “Sources of nonlinear complexity in geomorphic systems,”
Hydrology, 414–415: 108–117, 2012. Progress in Physical Geography, 26: 339–361, 2003.
Kim, H. S., R. Eykholt, and J. D. Salas, “Delay time window and plateau Phillips, J. D., “Deterministic chaos and historical geomorphology: a review
onset of the correlation dimension for small data sets,” Physical Review E, 58 and look forward,” Geomorphology, 76: 109–121, 2006.
(5): 5676–5682, 1998. Phoon, K. K., M. N. Islam, C. Y. Liaw, and S. Y. Liong, “A practical inverse
Kim, H. S., R. Eykholt, and J. D. Salas, “Nonlinear dynamics, delay times, approach for forecasting of nonlinear time series,” Journal of Hydrologic
and embedding windows,” Physica D, 127 (1–2): 48–60, 1999. Engineering, ASCE, 7 (2): 116–128, 2002.
Kim, H. S., K. H. Lee, M. S. Kyoung, B. Sivakumar, and E. T. Lee, Porporato, A. and L. Ridolfi, “Clues to the existence of deterministic chaos
“Measuring nonlinear dependence in hydrologic time series,” Stochastic in river flow,” International Journal of Modern Physics B, 10: 1821–1862, 1996.
Environmental Research and Risk Assessment, 23: 907–916, 2009. Porporato, A. and R. Ridolfi, “Nonlinear analysis of river flow time
Kostelich, E. J. and J. A. Yorke, “Noise reduction in dynamical systems,” sequences,” Water Resources Research, 33 (6): 1353–1367, 1997.
Physical Review A, 38 (3): 1649–1652, 1988. Porporato, A. and R. Ridolfi, “Multivariate nonlinear prediction of river
Kostelich, E. J. and T. Schreiber, “Noise reduction in chaotic time series data: flows,” Journal of Hydrology, 248 (1–4): 109–122, 2001.
a survey of common methods,” Physical Review E, 48 (3): 1752–1763, Puente, C. E. and N. Obregon, “A deterministic geometric representation of
1993. temporal rainfall. Results for a storm in Boston,” Water Resources Research, 32
Koutsoyiannis, D., “On the quest for chaotic attractors in hydrological (9): 2825–2839, 1996.
processes,” Hydrological Science Journal, 51 (6): 1065–1091, 2006. Ramsey, J. B. and H. J. Yuan, “The statistical properties of dimension calcu-
Koutsoyiannis, D. and D. Pachakis, “Deterministic chaos versus stochastic- lations using small data sets,” Nonlinearity, 3: 155–176, 1990.
ity in analysis and modeling of point rainfall series,” Journal of Geophysical Regonda, S., B. Sivakumar, and A. Jain, “Temporal scaling in river flow: can
Research, 101 (D21): 26441–26451, 1996. it be chaotic?” Hydrological Sciences Journal, 49 (3): 373–385, 2004.
Krasovskaia, I., L. Gottschalk, and Z. W. Kundzewicz, “Dimensionality of Rodriguez-Iturbe, I, F. B. De Power, M. B. Sharifi, and K. P. Georgakakos,
Scandinavian river flow regimes,” Hydrological Science Journal, 44 (5): “Chaos in rainfall,” Water Resources Research, 25 (7): 1667–1675, 1989.
705–723, 1999. Rosenstein, M. T., J. J. Collins, and C. J. De Luca, “A practical method for
Kugiumtzis, D., “State space reconstruction parameters in the analysis of calculating largest Lyapunov exponents from small data sets,” Physica D, 65:
chaotic time series—the role of the time window length,” Physica D, 95 (1): 117–134, 1993.
13–28, 1996. Rosenstein, M. T., J. J. Colins, and C. J. De Luca, “Reconstruction expansion
Kumar, P. and E. Foufoula-Georgiou, “Wavelet analysis for geophysical as a geometry-based framework for choosing proper delay times,” Physica D,
applications,” Reviews of Geophysics, 35 (4): 385–412, 1997. 73 (1): 82–98, 1994.
Kyoung, M. S., H. S. Kim, B. Sivakumar, V. P. Singh, and K. S. Ahn, Rössler, O. E., “An equation for continuous chaos,” Physics Letters A, 57:
“Dynamic characteristics of monthly rainfall in the Korean peninsula under 397–398, 1976.
climate change,” Stochastic Environmental Research and Risk Assessment, 25 Ruelle, D. and F. Takens, “On the nature of turbulence,” Communications in
(4): 613–625, 2011. Mathematical Physics, 20: 167–192, 1971.
Laio, F., A. Porporato, R. Revelli, and L. Ridolfi, “A comparison of nonlinear Salas, J. D. and R. A. Smith, “Physical basis of stochastic models of annual
flood forecasting methods,” Water Resources Research, 39 (5): 1129, 2003, flows,” Water Resources Research, 17 (2): 428–430, 1981.
doi:10.1029/2002WR001551. Salas, J. D., H. S. Kim, R. Eykholt, P. Burlando, and T. R. Green,
Lambrakis, N., A. S. Andreou, P. Polydoropoulos, E. Georgopoulos, and “Aggregation and sampling in deterministic chaos: implications for chaos
T. Bountis, “Nonlinear analysis and forecasting of a brackish karstic spring,” identification in hydrological processes,” Nonlinear Processes in Geophysics,
Water Resources Research, 36 (4): 875–884, 2000. 12: 557–567, 2005.
Linsay, P. S., “Period doubling and chaotic behaviour in a driven anhar- Sangoyomi, T. B., U. Lall, and H. D. I. Abarbanel, “Nonlinear dynamics of
monic oscillator,” Physical Review Letters, 47: 1349–1352, 1981. the Great Salt Lake: dimension estimation,” Water Resources Research, 32 (1):
Liu, Q., S. Islam, I. Rodriguez-Iturbe, and Y. Le, “Phase-space analysis of 149–159, 1996.
daily streamflow: characterization and prediction,” Advances in Water Sauer, T., J. A. Yorke, and M. Casdagli, “Embedology,” Journal of Statistical
Resources, 21: 463–475, 1998. Physics, 65 (3/4): 579–616, 1991.
Lorenz, E. N., “Deterministic nonperiodic flow,” Journal of the Atmospheric Schertzer, D., I. Tchiguirinskaia, S. Lovejoy, P. Hubert, and H. Bendjoudi,
Sciences, 20: 130–141, 1963. “Which chaos in the rainfall-runoff process? A discussion on ‘Evidence of
Lorenz, E. N., “Atmospheric predictability as revealed by naturally occur- chaos in the rainfall-runoff process’ by Sivakumar et al.” Hydrological Sciences
ring analogues,” Journal of the Atmospheric Sciences, 26: 636–646, 1969. Journal, 47 (1): 139–147, 2002.
Lorenz, E. N., “Dimension of weather and climate attractors,” Nature, 353: Schouten, J. C., F. Takens, and C. M. van den Bleek, “Estimation of the
241–244, 1991. dimension of a noisy attractor,” Physical Review E, 50 (3): 1851–1861, 1994.
Luo, X. D., J. Zhang, and M. Small, “Optimal phase space projection for Schreiber, T., “Determination of the noise level of chaotic time series,”
noise reduction,” Physical Review E, 72: 046710, 2005. Physical Review E, 48 (1): R13–R16, 1993a.
Mackey, M. C. and L. Glass, “Oscillation and chaos in physiological control Schreiber, T., “Extremely simple nonlinear noise reduction method,”
systems,” Science, 197 (4300): 287–289, 1977. Physical Review E, 47 (4): 2401–2404, 1993b.
Manzoni, S., A. Porporato, P. D’Odorico, F. Laio, and I. Rodriguez-Iturbe, Schreiber, T. and P. Grassberger, “A simple noise reduction method for real
“Soil nutrient cycles as a nonlinear dynamical system,” Nonlinear Processes in data,” Physics Letters A, 160: 411–418, 1991.
Geophysics, 11: 589–598, 2004. Schreiber, T. and H. Kantz, Observing and predicting chaotic signals: is 2%
Martinerie, J. M., A. M. Albano, A. I. Mees, and P. E. Rapp, “Mutual infor- noise too much?, Predictability of Complex Dynamical Systems, edited by
mation, strange attractors, and the optimal estimation of dimension,” Physical A. Kravtsov Yu and J. B. Kadtke, Springer Series in Synergetics, Springer,
Review A, 45: 7058–7064, 1992. Berlin, 1996, pp. 43–65.
May, R. M., “Simple mathematical models with very complicated dynam- Schuster, H. G., Deterministic Chaos: An Introduction, Physik Verlag,
ics,” Nature, 261: 459–467, 1976. Weinheim, 1988.
Minshall, N. E., “Predicting storm runoff on small experimental water- Sharifi, M. B., K. P. Georgakakos, and I. Rodriguez-Iturbe, “Evidence of
sheds,” Journal of the Hydraulics Division of American Society of Engineers, 86 deterministic chaos in the pulse of storm rainfall,” Journal of the Atmospheric
(HYB): 17–38, 1960. Sciences, 47: 888–893, 1990.
Nerenberg, M. A. H. and C. Essex, “Correlation dimension and systematic Singh, V. P., “A uniformly nonlinear hydrologic cascade model,” Irrigation
geometric effects,” Physical Review A, 42 (12): 7065–7074, 1990. Power, 36 (3): 301–317, 1979.
Packard, N. H., J. P. Crutchfield, J. D. Farmer, and R. S. Shaw, “Geometry Singh, V. P., Hydrologic Systems: Rainfall-Runoff Modelling, Prentice Hall,
from a time series,” Physical Review Letters, 45 (9): 712–716, 1980. Englewood Cliffs, New Jersey, Vol. 1, 1988.
29_Singh_ch29_p29.1-29.12.indd 9 8/22/16 1:51 PM

Singh, V. P., “The use of entropy in hydrology and water resources,” Sivakumar, B., S. Sorooshian, H. V. Gupta, and X. Gao, “A chaotic approach
Hydrological Processes, 11: 587–626, 1997. to rainfall disaggregation,” Water Resources Research, 37 (1): 61–72, 2001b.
Sivakumar, B., “Chaos theory in hydrology: important issues and interpre- Sivakumar, B., R. Berndtsson, J. Olsson, and Jinno, “Reply to ‘which chaos
tations,” Journal of Hydrology, 227 (1–4): 1–20, 2000. in the rainfall-runoff process?’ by Schertzer et al.” Hydrological Sciences
Sivakumar, B., “Rainfall dynamics at different temporal scales: a chaotic Journal, 47 (1): 149–158, 2002a.
perspective,” Hydrology and Earth System Sciences, 5 (4): 645–651, 2001a. Sivakumar, B., A. W. Jayawardena, and T. M. G. H. Fernando, “River flow
Sivakumar, B., “Is a chaotic multi-fractal approach for rainfall possible?” forecasting: use of phase-space reconstruction and artificial neural networks
Hydrological Processes, 15 (6): 943–955, 2001b. approaches,” Journal of Hydrology, 265 (1–4): 225–245, 2002b.
Sivakumar, B., “A phase-space reconstruction approach to prediction of Sivakumar, B., M. Persson, R. Berndtsson, and C. B. Uvo, “Is correlation
suspended sediment concentration in rivers,” Journal of Hydrology, 258: dimension a reliable indicator of low-dimensional chaos in short hydrological
149–162, 2002. time series?” Water Resources Research, 38 (2): 2002c, doi: 10.1029/
Sivakumar, B., “Forecasting monthly streamflow dynamics in the western 2001WR000333
United States: a nonlinear dynamical approach,” Environmental Modelling & Sivakumar, B., W. W. Wallender, C. E. Puente, and M. N. Islam, “Streamflow
Software, 18 (8–9): 721–728, 2003. disaggregation: a nonlinear deterministic approach,” Nonlinear Processes in
Sivakumar, B., “Chaos theory in geophysics: past, present and future,” Geophysics, 11: 383–392, 2004.
Chaos, Solitons & Fractals, 19 (2): 441–462, 2004a. Sivakumar, B., T. Harter, and H. Zhang, “Solute transport in a heteroge-
Sivakumar, B., “Dominant processes concept in hydrology: moving for- neous aquifer: a search for nonlinear deterministic dynamics,” Nonlinear
ward,” Hydrological Processes, 18 (12): 2349–2353, 2004b. Processes in Geophysics, 12: 211–218, 2005a.
Sivakumar, B., “Correlation dimension estimation of hydrologic series and Sivakumar, B., R. Berndtsson, M. Persson, and C. B. Uvo, “A multi-variable
data size requirement: myth and reality,” Hydrological Sciences Journal, 50 (4): time series phase-space reconstruction approach to investigation of chaos in
591–604, 2005. hydrological processes,” International Journal of Civil & Environmental
Sivakumar, B., “Nonlinear determinism in river flow: prediction as a pos- Engineering, 1 (1): 35–51, 2005b.
sible indicator,” Earth Surface Processes and Landforms, 32 (7): 969–979, 2007. Sivakumar, B., W. W. Wallender, Horwath, J. P. Mitchell, S. E. Prentice, and
Sivakumar, B., “Nonlinear dynamics and chaos in hydrologic systems: latest B. A. Joyce, “Nonlinear analysis of rainfall dynamics in California’s Sacramento
developments and a look forward,” Stochastic Environmental Research and Valley,” Hydrological Processes, 20 (8): 1723–1736, 2006.
Risk Assessment, 23: 1027–1036, 2009. Sivakumar, B., A. W. Jayawardena, and W. K. Li, “Hydrologic complexity
Sivakumar, B., “Global climate change and its impacts on water resources and classification: a simple data reconstruction approach,” Hydrological
planning and management: assessment and challenges,” Stochastic Processes, 21 (20): 2713–2728, 2007.
Environmental Research Risk Assessment, 25 (4): 583–600, 2011a. Sivakumar, B., F. M. Woldemeskel, and C. E. Puente, “Nonlinear analysis of
Sivakumar, B., “Water crisis: from conflict to cooperation—an overview,” rainfall variability in Australia,” Stochastic Environmental Research Risk
Hydrological Sciences Journal, 56 (4): 531–552, 2011b. Assessment, 28 (1): 17–27, 2014.
Sivakumar, B., Chaos theory for modeling environmental systems: philoso- Smith, L. A., “Intrinsic limits on dimension calculations,” Physics Letters A,
phy and pragmatism, System Identification, Environmental Modelling, and 133 (6): 283–288, 1988.
Control System Design, edited by L. Wang and H. Garnier, Springer-Verlag, Sugihara, G. and R. M. May, “Nonlinear forecasting as a way of distinguish-
London, 2011c, pp. 533–555. ing chaos from measurement error in time series,” Nature, 344: 734–741,
Sivakumar, B. and R. Berndtsson, Advances in Data-Based Approaches for 1990.
Hydrologic Modeling and Forecasting, World Scientific, Singapore, 2010. Sun, J., Y. Zhao, J. Zhang, X, Luo, and M. Small, “Reducing coloured noise
Sivakumar, B. and J. Chen, “Suspended sediment load transport in the for chaotic time series in the local phase space,” Physical Review E, 76: 026211,
Mississippi River basin at St. Louis: temporal scaling and nonlinear determin- 2007.
ism,” Earth Surface Processes and Landforms, 32 (2): 269–280, 2007. Swinney, H. L. and J. P. Gollub, “Hydrodynamic instabilities and the transi-
Sivakumar, B. and A. W. Jayawardena, “An investigation of the presence of tion to turbulence,” Physics Today, 31 (8): 41–49, 1978.
low-dimensional chaotic behavior in the sediment transport phenomenon,” Şen, Z., Fuzzy Logic and Hydrologic Modeling, CRC Press, Boca Raton, FL,
Hydrological Sciences Journal, 47 (3): 405–416, 2002. 2009.
Sivakumar, B. and A. W. Jayawardena, “Sediment transport phenomenon in Takens, F., Detecting strange attractors in turbulence, Dynamical Systems
rivers: an alternative perspective,” Environmental Modelling & Software, 18 and Turbulence, Lecture Notes in Mathematics 898, edited by D. A. Rand and
(8–9): 831–838, 2003. L. S. Young, Springer, Berlin, 1981, pp. 366–381.
Sivakumar, B. and V. P. Singh, “Hydrologic system complexity and nonlin- Thomas, Jr., H. A. and M. B. Fiering, Mathematical synthesis of streamflow
ear dynamic concepts for a catchment classification framework,” Hydrology sequences for the analysis of river basins by simulation, Design of Water
and Earth System Sciences, 16: 4119–4131, 2012. Resource-Systems: New Techniques for Relating Economic Objectives,
Sivakumar, B. and V. P. Singh, “Special issue on grand challenges in hydrol- Engineering Analysis, and Governmental Planning, edited by A. Mass, M. M.
ogy,” Journal of Hydrologic Engineering, ASCE, 20 (1): A2014001, 2015. Hufschmidt, R. Dorfman, H. A. Thomas, Jr., S. A. Marglin, and G. M. Fair,
Sivakumar, B. and W. W. Wallender, “Deriving high-resolution sediment Harvard University Press, Cambridge, MA, 1962, pp. 459–493.
load data using a nonlinear deterministic approach,” Water Resources Tsonis, A. A., Chaos: From Theory to Applications, Plenum Press, New York,
Research, 40: W05403, 2004, doi: 10.1029/2004WR003152. 1992.
Sivakumar, B. and W. W. Wallender, “Predictability of river flow and sedi- Tsonis, A. A. and J. B. Elsner, “The weather attractor over short timescales,”
ment transport in the Mississippi River basin: a nonlinear deterministic Nature, 333: 545–547, 1988.
approach,” Earth Surface Processes and Landforms, 30: 665–677, 2005. Tsonis, A. A., J. B. Elsner, and K. P. Georgakakos, “Estimating the dimen-
Sivakumar, B., S. Y. Liong, and C. Y. Liaw, “Evidence of chaotic behavior in sion of weather and climate attractors: important issues about the procedure
Singapore rainfall,” Journal of the American Water Resources Association, 34 and interpretation,” Journal of the Atmospheric Sciences, 50: 2549–2555, 1993.
(2): 301–310, 1998. Tsonis, A. A., G. N. Triantafyllou, J. B. Elsner, J. J. Holdzkom II, and
Sivakumar, B., S. Y. Liong, C. Y. Liaw, and K. K. Phoon, “Singapore rainfall A. D. Kirwan, Jr., “An investigation on the ability of nonlinear methods to infer
behavior: chaotic?” Journal of Hydrological Engineering, ASCE, 4 (1): 38–48, dynamics from observables,” Bulletin of the American Meteorological Society,
1999a. 75: 1623–1633, 1994.
Sivakumar, B., K. K. Phoon, S. Y. Liong, and C. Y. Liaw, “A systematic Valencia, D. R. and J. L. Schaake “Disaggregation processes in stochastic
approach to noise reduction in chaotic hydrological time series,” Journal of hydrology,” Water Resources Research, 9 (3): 211–219, 1973.
Hydrology, 219 (3–4): 103–135, 1999b. Vignesh, R., V. Jothiprakash, and B. Sivakumar, “Streamflow variability and
Sivakumar, B., K. K. Phoon, S. Y. Liong, and C. Y. Liaw, “Comment on classification using false neighbor neighbor method,” Journal of Hydrology,
‘Nonlinear analysis of river flow time sequences’ by Amilcare Porporato and 531: 706–715, 2015.
Luca Ridolfi,” Water Resources Research, 35 (3): 895–897, 1999c. Wagener, T., M. Sivapalan, P. A. Troch, B. L. McGlynn, C. J. Harman,
Sivakumar, B., R. Berndtsson, J. Olsson, K. Jinno, and A. Kawamura, H. V. Gupta, P. Kumar, et al., “The future of hydrology: an evolving science for
“Dynamics of monthly rainfall-runoff process at the Göta basin: a search for a changing world,” Water Resources Research, 46: W05301, 2010, doi:10.1029/
chaos,” Hydrology and Earth System Sciences, 4 (3): 407–417, 2000. 2009WR008906.
Sivakumar, B., R. Berndtsson, J. Olsson, and K. Jinno, “Evidence of chaos in Wang, Q. and T. Y. Gan, “Biases of correlation dimension estimates of
the rainfall-runoff process,” Hydrological Sciences Journal, 46 (1): 131–145, streamflow data in the Canadian prairies,” Water Resources Research, 34 (9):
2001a. 2329–2339, 1998.
29_Singh_ch29_p29.1-29.12.indd 10 8/22/16 1:51 PM

REFERENCES 29-11
Wilcox, B. P., M. S. Seyfried, and T. M. Matison, “Searching for chaotic dynam- Yevjevich, V. M., Stochastic Processes in Hydrology, Water Resources
ics in snowmelt runoff,” Water Resources Research, 27 (6): 1005–1010, 1991. Publications, Fort Collins, CO, 1972.
Wolf, A., J. B. Swift, H. L. Swinney, and A. Vastano, “Determining Young, P. C. and K. J. Beven, “Data-based mechanistic modeling and
Lyapunov exponents from a time series,” Physica D, 16: 285–317, 1985. rainfall-flow non-linearity,” Environmetrics, 5 (3): 335–363, 1994.
Xu, P., W. K. Li, and A. W. Jayawardena, “Noise level estimation for a cha- Zhou, Y., Z. Ma, and L. Wang, “Chaotic dynamics of the flood series in the
otic time series,” International Journal of Bifurcation and Chaos, 22 (3): Huaihe River Basin for the last 500 years,” Journal of Hydrology, 258: 100–110,
1250052, 2012. 2002.
29_Singh_ch29_p29.1-29.12.indd 11 8/22/16 1:51 PM

Chapter 30
Copula Modeling in Hydrologic
Frequency Analysis
BY
CHRISTIAN GENEST AND FATEH CHEBANA
ABSTRACT severity-duration-frequency curves and copula-based dissimilarity mea-

sures for streamflow prediction in ungaged catchments to models for
This chapter provides an introduction to the copula approach to multivariate
temporal sampling errors in satellite-derived rainfall estimates and for
data modeling and surveys some of its uses in hydrologic frequency analysis
checking dam spillway adequacy.
(HFA). The various steps involved in the construction and validation of a
copula model are outlined. Flood peak and volume data from the Fraser River
at Hope (British Columbia, Canada) are used to illustrate the methodology, 30.2 DESCRIPTION OF COPULA MODELS
and resources for carrying out the analysis are provided. Possible extensions
A copula is a multivariate distribution function whose margins are uniform
of these techniques to regional and nonstationary frequency analysis, as well
on the interval (0, 1). Given a random vector X = (X1, …, Xd) with joint dis-
as to spatial contexts, are briefly discussed.
tribution function H and margins F1,…, Fd, Sklar (1959) showed that there
always exists a copula C such that, for all x1, …, xd ∈ (−∞, +∞),
30.1 INTRODUCTION
Pr( X1 ≤ x1 , …, X d ≤ x d ) = H ( x1 , …, x d ) = C{F1 ( x1 ), …, Fd ( x d )}. (30.1)
Floods, droughts, and extreme precipitation can have serious social and
economic consequences, including widespread material destruction and loss If the marginal distribution functions F1, …, Fd are continuous, then C is
of human lives. It is crucial, therefore, to rely on appropriate models in order unique and can be retrieved from H by setting, for all u1, …, ud ∈(0, 1),
to predict the frequency and magnitude of such events. Hydrologic frequency
analysis is a set of statistical techniques designed for this purpose. C(u1 , …, ud ) = H {F1−1 (u1 ), …, Fd −1 (ud )} (30.2)
Hydrologic phenomena involve many interrelated variables, for example,
flood peak, volume, and duration. It is natural, therefore, to adopt a multivari- where, for each j ∈{1, …, d }, Fj–1 denotes the generalized inverse of Fj. In
ate framework for their analysis, and it is in fact often essential to do so for other words, C is the distribution function of the vector (U1, …, Ud) =
accurate risk assessment. In contrast, univariate HFA can provide only a (F1(X1), …, Fd (Xd )).
limited evaluation of extreme events and their probability of occurrence, When the distribution H of X is unknown, Eq. (30.1) remains valid but C and
because treating each event characteristic in isolation ignores possibly crucial F1, …, Fd are undetermined. A copula model for X then consists of assuming
relations between the variables at play. that
This chapter describes the use of copula models in the context of multi-
variate HFA. The merits of this approach in statistics are now well known C ∈C, F1 ∈ F1, …, Fd ∈Fd , (30.3)
(Genest and Favre, 2007; Joe, 2014) and references therein. In hydrology, where C is a class of copulas indexed by a (possibly vector-valued) parameter
copula modeling techniques have been used to study events such as floods, q while F1, …, Fd are classes of univariate distribution functions indexed by
storms, droughts, sediments, rainfalls, and sea storms (Grimaldi and parameters h1, …, hd, respectively.
Serinaldi, 2006; Kao and Govindaraju, 2007; Chebana and Ouarda, 2011; The main advantage of this approach is that each variable can be mod-
Vandenberghe et al., 2011; Requena et al., 2013; Volpi and Fiori, 2014). eled separately. Once suitable parametric classes of univariate distributions
Copulas are also useful to study a single feature, for example, flood peak, at (e.g., Gamma, Normal, or Student t) have been selected for each variable,
multiple sites; for an illustration, see El Adlouni and Ouarda (2008). they can be “glued” together with an appropriate choice of copula. Classical
Copula models are introduced in Sec. 30.2, where their advantages are multivariate models typically offer much less flexibility because the margins
highlighted. An overview of inference techniques for these models is then given are often all of the same type; for example, all margins of a multivariate
in Sec. 30.3, and their use for multivariate quantile and return period estima- Normal distribution are Normal. When they use copula models, analysts
tion is described in Sec. 30.4. The methodology is illustrated in Sec. 30.5 using face no restriction (other than practicality) in the choice of C and F1, …, Fd
flood peak and volume data from the Fraser River at Hope (British Columbia, in Eq. (30.1). This advantage is of particular interest in hydrology, where
Canada). Extensions are outlined in Sec. 30.6, and specialized resources are variables associated with a given event (e.g., flood peak and volume) may
listed in Sec. 30.7. vary widely in nature, scale, and distribution. Copula models eliminate the
Although copula modeling in hydrology only developed in the last need to transform the variables in the (often illusory) hope of making them
decade or so, the literature on the subject is already substantial. Only fit into a classical framework. This is also convenient because for inter-
selected references are provided here. The readers can consult the web- pretation purposes, practitioners tend to prefer models that involve the
site of the International Commission on Statistical Hydrology for an original variables. Another distinct advantage of copula modeling is its reliance
up-to-date list of publications related to copulas and their applications in on the common tools of univariate HFA. Before proceeding with the construc-
HFA and other hydrologic contexts, from the construction of drought tion of a univariate model, an analyst would first carry out an exploratory
30-1
30_Singh_ch30_p30.1-30.10.indd 1 8/22/16 1:53 PM

30-2 Copula Modeling in Hydrologic Frequency Analysis
analysis aimed at identifying outliers and checking the basic assumptions of Although EV copulas are not necessarily symmetric in their arguments, the
stationarity, homogeneity, and serial independence of the time series. The same EV dependence structure is inherited by all subsets of a vector whose copula
preliminary steps are prerequisite to copula modeling. As the procedures satisfies property (30.5). Members of the Gumbel family are the only copulas
required to accomplish this task and select suitable univariate parametric models that are both EV and Archimedean (Genest and Rivest, 1989).
are already well documented in the HFA literature (Rao and Hamed, 2000), it is
taken for granted henceforth that 30.2.3 Meta-Elliptical Copulas
Suppose that a random vector is of the form X = RAT, where R is a strictly
Assumption: The data at hand constitute a random sample from a joint positive continuous random variable, A is a d × d matrix of constants, and T
distribution H with continuous margins F1, …, Fd. is a random vector that is both independent of R and uniformly distributed
This implies, in particular, that the underlying dependence structure is on the unit sphere Td = {(t1 , , td ): t12 + + td2 = 1}. The vector X then has
represented by a unique copula C that does not change over time. The success an elliptical distribution with mean 0 and dispersion matrix S = AAT, where
of the multivariate modeling exercise then depends on the choice of a suitable T
denotes matrix transposition; the copula associated with the distribution of
copula C within some parametric class C. The basic procedure to be followed X is said to be meta-elliptical (Genest et al., 2007). For instance when R2 is
is described in Sec. 30.3. To fix ideas, some of the simple classes of copulas chi-square with d degrees of freedom, X has a multivariate Normal distribu-
commonly used to model low-dimensional hydrologic data are listed below. tion; its underlying copula, which is parameterized by S, is called the
For additional examples and book-length treatments of the mathematical Gaussian copula. The multivariate Student t copula with v degrees of freedom
properties of copulas, see Joe (2015) or Nelsen (2006). also belongs to this class. Similar to EV copulas, meta-elliptical copulas can be
asymmetric in their arguments but all subsets of a vector whose copula is
30.2.1 Archimedean Copulas meta-elliptical will inherit the same type of dependence structure (e.g., if X
A copula C is said to be Archimedean if it can be written in the form has a multivariate Student t copula with v degrees of freedom, then so do all
subsets of d ≥ 2 components of X).
C(u1 , …, ud ) = ψ {ψ −1 (u1 ) +  + ψ −1 (ud )} (30.4)
30.2.4 More Elaborate Copula Structures
in terms of a generator y: [0, ∞) → [0, 1] such that y(0) = 1, y(t) → 0 as Many parametric copula families available in the literature are such that all
t → ∞ and y is d-monotone in the sense of McNeil and NeŠlehová (2010). For subsets of variables share a common type of dependence structure. When this
instance, the generator of the Clayton copula with parameter q ∈ (0, ∞) is property is too limitative, one can resort to asymmetric extensions, for exam-
defined, for all t ∈ [0, ∞) by y(t) = (1+qt)−1/q. Many other examples, including ple, hierarchical Archimedean copulas (Joe, 2014) or Liouville copulas (McNeil
the Frank and Gumbel parametric families, can be found, for example, in and NeŠlehová, 2010). The state-of-the-art in this regard is vine copula models,
Genest and MacKay (1986), Nelsen (2006), Salvadori et al. (2007). built from a tree-like structure of bivariate copulas consisting of conditional
Archimedean copulas are symmetric in their arguments. Thus if a vector distributions beyond the first level of the hierarchy (Kurowicka and Joe, 2011).
(U1, …, Ud) is distributed as per Eq. (30.4), every subset of its components In such models, various subsets of variables can have widely different depen-
inherits the same dependence structure. dence structures. However, this added flexibility comes at the cost of more
30.2.2 Extreme-Value Copulas complex inference procedures than those that will be reviewed below.
A copula is said to be extreme-value (EV) if, for all t > 0 and u1, …, ud ∈ (0, 1),
30.3 OVERVIEW OF MODEL SELECTION
Ct(u11/t, …, ud1/t) = C(u1, …, ud) (30.5)
Given a random sample from some unknown multivariate distribution H,
This property is characteristic of the dependence structure in multivariate how can one select a suitable parametric copula family to model its depen-
extreme-value distributions; it is particularly attractive for data that are block dence, estimate its parameters, and validate this choice? A brief practical
maxima or that can be construed as such, for example, yearly maxima of daily guide is provided below, based largely on (Genest and Favre, 2007). It is
flow measurements (Salvadori and De Michele, 2011). In dimension d = 2, EV assumed henceforth that the margins F1, …, Fd are continuous to ensure that
copulas can be expressed simply in terms of a convex map A: [0, 1] → [0, 1] the unknown copula C in Eq. (30.1) is unique. A discussion of the discon-
such that, for all t ∈ [0, 1], A(t) ≥ max(t, 1 – t). To be specific, one has, for all tinuous case can be found in Genest and NeŠlehová (2007).
u1, u2 ∈ (0, 1), Figure 30.1 summarizes the key issues to be considered. The items in the
chart are not ordered because the various steps are interrelated and one may
C(u1 ,u2 ) = exp[ln(u1u2 )A{ln(u2 ) / ln(u1u2 )}] (30.6) have to go back and forth between them. A rough order is provided below and
illustrated in the data analysis of Sec. 30.4.
Examples of parametric families of EV copulas include the Galambos, The preliminary steps are those needed to ensure that the assumption
Gumbel, Hüsler–Reiß, Marshall–Olkin, and Student t v – EV (Joe, 2014; Kotz stated in Sec. 30.2 holds. As the required procedures are not specific to copula
and Nadarajah, 2000; Salvadori and De Michele, 2010; Tawn, 1988). modeling, the reader can refer, for example, to Chebana and Ouarda (2011)
Preliminary Preprocessing of the series at hand to ensure that the data form a random sample
steps from a distribution with continuous margins (in particular, stationarity, homogeneity,
and serial independence must hold, at least approximately)
Modeling steps Multivariate Fx,y
Margins Copula
Fx and Fy C (u, v)
Refer to U-HFA
for each variable
Graphics Dependence Tail

measures dependence
Goodness Parameter Selection

of testing estimation criteria
Inference steps Evaluation of return periods and multivariate quantile
Figure 30.1 Main steps of HFA and procedure for copula fitting and selection.
30_Singh_ch30_p30.1-30.10.indd 2 8/22/16 1:53 PM

Overview of model selection 30-3
or Chebana et al. (2013) for descriptions in the context of hydrologic flood is an important clue. Among other characteristics that should be taken into
data. In what follows, the (possibly de-trended or otherwise preprocessed) account are: the presence of symmetry or asymmetry, the strength of depen-
data are denoted X1 = (X11, …, X1d), …, Xn = (Xn1, …, Xnd) and are considered dence, signs of extreme-value behavior, etc. Various indices can be computed
to form (at least approximately) a random sample from distribution H with to help with this task.
margins F1, …, Fd and copula C.
30.3.2 Numerical Summaries
30.3.1 Visualizing Data from Copulas
There exist several measures of dependence, asymmetry, and tail behavior
If the margins F1, …, Fd were known, in view of Eq. (30.2), the observations whose population values depend on the underlying copula alone, and whose
could be transformed into a random sample from C by setting U1 = (F1(X11), …, sample values can be computed from the vectors of ranks derived from the
Fd (X1d)), …, Un = (F1(Xn1), …, Fd(Xnd)). The vectors U1, …, Un could then be observations. Prominent among them are Spearman’s rho (r) and Kendall’s
plotted if d = 2 (or two components at the time when d > 2) to get a sense of tau (t). In dimension d = 2, Spearman’s rho is simply Pearson’s correlation
the dependence embodied in C. In practice, however, the margins are coefficient, but computed from the ranks of the observations. Kendall’s tau is
unknown and must be estimated. As this process involves uncertainty and is defined theoretically as the difference between the probability of concor-
prone to error, inference about C is best based on conservative estimates of dance (positive slope) and the probability of discordance (negative slope)
these margins defined, for all j ∈{1, …, d} and x ∈ (−∞, ∞), by between two observations chosen at random. Its sample version is defined in
a similar way.
1 n In families of bivariate copulas indexed by a real-valued parameter q, there
Fnj ( x ) = ∑1( Xij ≤ x ) (30.7)
n + 1 i=1 is typically a one-to-one correspondence between q, r, and t; in the Clayton
model, for instance, t = q/(q + 2) for all q ∈ (0, ∞) and similarly for r, though
where 1(A) is the indicator of the event A, that is, 1(A) = 1 when the event A the latter cannot be expressed in closed form as a function of q. Some copula
occurs and 1(A) = 0 otherwise. One may then construct a pseudo-sample families have restricted ranges for the values of r or t, for example,
from C by letting, for each i ∈ {1, …, n}, t ∈ [−2/9, 2/9] in the Farlie–Gumbel–Morgenstern (FGM) bivariate copula
family; such models may thus be judged inadequate from the start in certain
Ûi = (Ûi1, …, Ûid) = (Fn1(Xi1), … Fnd(Xid)) (30.8) contexts. Multivariate definitions are available for many coefficients, includ-
ing Spearman’s rho and Kendall’s tau; see, for example, Genest et al. (2011b),
Note that, for each i ∈ {1, …, n} and j ∈ {1, …, d}, (n + 1) × Fnj(Xij) is the rank Joe (1993), and Schmid and Schmidt (2007).
of Xij among X1j, …, Xnj. Thus if the points Û1, …, Ûn could be visualized in Several other rank-based/copula-based coefficients are possible; see, for
d-dimensional space, each of their marginal distributions would be found to be example, Genest and Nešlehová (2012b) and Nelsen (2006). Among them,
exactly uniform on the set {1/(n + 1), …, n/(n + 1)} because ties do not occur Joe’s lower and upper tail dependence coefficients (Joe, 1993) are particularly
(almost surely) when F1, …, Fd are continuous. Statistical theory further important for hydrologic applications. In the bivariate case, they are given
guarantees that the joint pattern described by Û1, …, Ûn contains all the respectively by
information needed to identify C. In formal terms, consider the empirical distri-
bution of the pseudo-sample Û1, …, Ûn defined, for all u1, …, ud ∈ (0, 1), by λ l = lim C(q , q ) q , λu = 2 − lim{1 − C(q , q )} / (1 − q ) (30.11)
q↓0 q↑1
1 n whenever the limits exist. Rank-based estimates are defined analogously; for
Cˆn (u1 ,...,ud ) = ∑1(Uˆ i1 ≤ u1 ,...,Uˆ id ≤ ud ) (30.9) instance,
n i=1
This so-called empirical copula is a consistent estimator of C, that is, it λun = 2 − (n / k )Cˆn (1 − k / n,…,1 − k / n) (30.12)
converges [in distribution for all choices of u1, …, ud ∈ (0, 1)] to C as the
sample size n grows indefinitely; see, for example, Segers (2012). is associated to lu for an appropriate choice of k; more sophisticated
A plot of the vectors Û1, …, Ûn (in dimension d = 2, or else looking at pairs approaches are preferable, however; see, for example, Schmidt and Stadtmüller
of components) provides many clues as to the type of dependence embodied in (2006).
C; it can also assist in selecting a parametric form for C. This procedure called a Tail dependence coefficients aim at detecting and quantifying the presence
rank-plot can be used, for example, to check whether all pairs share the same of extreme dependence between X1 and X2. For instance when C is a bivariate
type of dependence. Alternatively, one could transform the pseudo-observations extreme-value copula with dependence function A, as per Eq. (30.6), one has
to the Normal scale by setting, for each i ∈ {1, …, n}, Ôi = (F-1(Ûi1), … F-1(Ûid)), ll = 0 and lu = 2 – 2A(0.5); other examples of calculation are given, for
where F denotes the N(0, 1) distribution. One could then plot Ô1, …, Ôn, called example, in Ghoudi et al. (1998). Computing ll and lu can sometimes reveal
a rankit plot. important differences between otherwise similarly looking dependence struc-
In large samples, rank-plots and rankit plots become too dense and hence tures. For instance, both coefficients vanish for the Gaussian copula but they
ineffective. They can then be replaced, for example, by rank-based estimates are strictly positive for the Student t copula. As a result, models based on the
of the copula density (Genest et al., 2009b). In high-dimensional problems, Gaussian copula can sometimes lead to an underestimation of joint extreme
however, rank-based plots of the various pairs of variables shed only partial risks (McNeil et al., 2005). Alternative multivariate extreme coefficients are
light into the overall dependence structure. Additional insight can be gained considered in Frahm (2006).
by looking at one-dimensional projections of the multivariate data, for
30.3.3 Model Fitting
example, by setting, for all i ∈ {1, …, n},
Suppose that a copula model (30.3) is being considered for a d-variate distri-
ˆ = Cˆ (Uˆ ,…,Uˆ )
W (30.10) bution H. Assume that following an exploratory data analysis, parametric
i n i1 id
families F1, …, Fd and C have been selected for the margins and the copula,
In words, Ŵi is the proportion of points among Û1, …, Ûn whose compo- respectively. Further assume that densities exist for all members of these
nents are all smaller than those of Ûi. The empirical distribution function K̂ n families and denote them by lower-case letters.
associated with the set W ˆ ,, Wˆ is an estimate of the distribution K of the A natural way to estimate the parameters is to maximize the joint log-
1 n
variable W = C(U1, …, Ud) = H(X1, …, Xd); the latter is unknown but can be likelihood, viz.
computed for any hypothesized choice of C or H. For example, in dimension
n n d
d = 2, the assumption that the variables X1, X2 (or equivalently U1, U2) are
independent implies that, for all w ∈ (0, 1), K(w) = w – w ln(w). One could (θ, η1 ,…, ηd ) = ∑ ln[cθ {F1,η1 ( Xi1 ),…, Fd ,ηd ( Xid )}] + ∑∑ ln{ f j ,η j ( Xij )} (30.13)
i =1 i =1 j =1
then check whether independence is reasonable by comparing visually K̂ n
and K or by drawing a Q-Q plot. This can be a useful complement to a rank-
In principle, this expression can be maximized, but this can be tedious and
plot, even when dealing with only two variables, but the point is that this
time-consuming, especially in high dimensions. To circumvent this problem,
procedure can be used whatever the number of variables because the pseudo-
ˆ ,, Wˆ are always real-valued and hence have a univariate Joe (2005) examined the merits of proceeding in two steps. In the first step,
observations W 1 n
estimates of the marginal distribution parameters h1, …, hd are obtained. For
distribution; see Genest and Boies (2003) for details.
each j ∈ {1, …, d}, this is done by maximizing
These and other graphical tools provide insight into the dependence structure
of the data and help to select suitable parametric copula families. As men- n
tioned earlier, the presence or not of similar patterns between different pairs ( η j ) = ∑ ln{f j ,η j ( Xij )} (30.14)
i =1
30_Singh_ch30_p30.1-30.10.indd 3 8/22/16 1:53 PM

In the second step, the resulting estimates of the marginal parameters are Both Sn and Tn have unwieldy limiting distributions that depend on the
plugged into the first summand of Eq. (30.13) to obtain a log-likelihood func- unknown value of the dependence parameter q under H0. For this reason, one
tion for the dependence parameter q, viz. must resort to resampling techniques to compute the p-values associated to these
test statistics. Two commonly used procedures are the parametric bootstrap
n (Genest and Rémillard, 2008) and the multiplier method (Kojadinovic et al., 2011).
(θ) = ∑ln[cθ {F1,ηˆ1 ( Xi1 ),…, Fd ,ηˆ d ( Xid )}] (30.15) In addition to being a function of sample size, the performance of all GoF
i =1 tests for copulas typically depends on the discrepancy between the true
underlying copula and the hypothesized one. For extensive power studies and
As shown by Joe (2005), this technique typically yields estimates of q that general recommendations, see, for example, Berg (2009) and Genest et al.
are consistent and asymptotically Normal, but somewhat less efficient than (2009c).
those derived from the full maximization of Eq. (30.13), except in the neigh- More elaborate GoF tests have been proposed based, for example, on the
borhood of independence. However, the estimator of q can be severely biased Rosenblatt transformation or on smooth nonparametric estimates of the
if the marginal models are specified incorrectly (Kim et al., 2007). copula density. Tests tailored for specific dependence structures such as
A safer, and yet conceptually simple, solution is to base the estimate of q on Clayton or Gaussian copulas are also available; see, for example, Berg and
the likelihood function derived from the pseudo-observation Û1, …, Ûn, viz. Quessy (2009) for general moment-based GoF test statistics. Tests of the more
general hypothesis that C belongs to some specific nonparametric class of
n n copulas are also beginning to emerge; the case of extreme-value copulas
(θ) = ∑ln[cθ {Fn1 ( Xi1 ),…, Fnd ( Xid )}] = ∑ln{cθ (Uˆ i1 ,…,Uˆ id )} (30.16) considered below provides an example.
i =1 i =1
30.3.5 Procedures for Extreme-Values Copulas
which is rank-based. The consistency and asymptotic Normality of this esti- Given their importance in modeling catastrophic events, extreme-value copu-
mator are established under mild regularity conditions in Genest et al. (1995), las have been the object of a great deal of attention in recent years. Let Ed
where a consistent estimate of the asymptotic variance is also proposed. The denotes the class of d-variate extreme-value copulas. To see whether an
maximum pseudo-likelihood estimator is generally less efficient than the extreme-value copula applies to a given situation, one should first test the
full maximum likelihood estimator when the margins are properly specified, hypothesis HE: C ∈ Ed. This is called a test of extremeness. The oldest and
but much preferable than either of the two previous methods in case of simplest procedure of this type, in the bivariate case, consists of setting Iij =
marginal misspecification (Kim et al., 2007). 1(Xj1 ≤ Xi1, Xj2 ≤ Xi2) for all i, j ∈ {1, …, n} and computing
Several commonly used bivariate copula families are indexed by a real-
valued parameter q that can be expressed as a simple algebraic function of 8 n
9 n
Spearman’s rho or Kendall’s tau, say q = g(r) or q = h(t). In a moment-based SnGKR = −1 + ∑Iij − n(n − 1)(n − 2) ∑ Iij I kj (30.19)
n(n − 1) i≠ j
method, a quick-and-dirty estimate qn of q is then given either by qn = g(rn) i≠ j ≠k
or qn = h(tn), where rn and tn are the classical estimators of r and t, respec-
tively. Both solutions are acceptable; they also provide reliable starting values This statistic is asymptotically Normal with zero mean under HE. To carry
for numerical maximization of the pseudo-likelihood, given that rn and tn are out the test, the variance of SnGKR must be estimated, either by the jackknife or
rank-based, consistent, and asymptotically Normal. Inversion of Spearman’s using approximations given in Ben Ghorbal et al. (2009) for its finite-sample
or Kendall’s tau can also be used to estimate dependence parameters in mul- and asymptotic variance under HE. This test performs very well against a wide
tivariate copula families. In d-variate meta-elliptical copula models, for range of alternatives; see Genest and Nešlehová (2012a) and Kojadinovic and
example, the (i, j)th element sij of the correlation matrix S can be estimated Yan (2010) for additional tests of extremeness.
via Kendall’s tau through the relation qij = (2/p) arcsin(sij); see, for example, When there is no evidence against HE, one may seek an adequate parametric
Genest et al. (2007). extreme-value copula family for the data. To validate a particular class A of
The above moment-based estimators are generally less efficient than maxi- Pickands dependence functions, one may then wish to test HA: A ∈ A using a
mum pseudo-likelihood estimators. In addition, they sometimes need to be GoF test. A natural way to accomplish this is to compare the best representative
adjusted to ensure that they remain within the admissible range of values for Aθn from the class A to a model-free, nonparametric estimator An. In the bivari-
q. In the bivariate FGM copula family, for instance, one has r = q/3 and ate case, the simplest such rank-based estimator is defined, for all t ∈ [0, 1], by
t = 2q/9, so that qn = 3rn and qn = 9tn /2, both of which can fall outside the
interval [−1, 1]. Similarly, the correlation matrix S of a d-variate meta-ellipti- 1 n 
cal copula obtained by inversion of Kendall’s tau may fail to be positive semi- AnP (t ) = 1  ∑ ξˆ i (1 − t ) (30.20)
definite. Alternative approaches are being developed; for example, a method  n i=1 
based on multivariate L-moments was recently proposed in Brahimi et al.
(2015) to account for various hydrologic constraints. where, for each i ∈ {1, …, n},
30.3.4 Model Validation ξˆ i (t ) = min{− ln(Uˆ i1 ) t , − ln(Uˆ i 2 ) (1 − t )} (30.21)

Before a copula model is used, for example, for prediction, it is important to This estimator is a rank-based version of the original estimator of A due to
check whether the selected copula family C is appropriate. This can be Pickands (1981). In practice, however, it is generally preferable to use another
accomplished by testing the hypothesis H0: C ∈ C. Goodness-of-fit (GoF) rank-based estimator Aˆ nCFG , called the CFG estimator after Capéraà,
tests for copula models are typically based on a comparison between the Fougères, and Genest (1997); the latter is based on a geometric mean (with
empirical copula Ĉn, which stands for the pseudo-sample, and the “best rep- offset) rather than a harmonic mean as in Eq. (30.20). As shown in Genest
resentative” Cθn from the parametric class C. Any convenient rank-based and Segers (2009), both estimators are consistent and asymptotically Normal
estimator qn of q can be used, provided that it is consistent. at every fixed t ∈ (0, 1). For multivariate extensions of the CFG estimator in
For moderate sample sizes, an informal graphical procedure consists of the known-margin case, see Gudendorf and Segers (2011) and Zhang et al.
comparing the rank-plot to a random sample of equal size from Cθn . The (2008).
simplest formal test (Genest et al., 2009c) is based on a measure of discrep- In the absence of evidence against HE, one can proceed with testing the
ancy between the best representative Cθn under H0 and the empirical copula hypothesis HA that the unknown Pickands dependence function belongs to
Ĉn , viz. some parametric family A. This idea is explored in Genest et al. (2011a) in
dimension d = 2, where tests are based on the statistics
n
Sn = ∑{Cˆn (Uˆ i1 ,…,Uˆ id ) − Cθn (Uˆ i1 ,…,Uˆ id )} (30.17)
2
1 1
i =1 Tn
GKNY −1
= ∫ { Aθn (t ) − Aˆ nCFG (t )}2 dt and TnGKNY−2 = ∫ { Aθn (t ) − Aˆ nP (t )}2 dt
0 0
This test is consistent for any copula family. An alternative test, which (30.22)
exhibits good power when dealing with Archimedean copulas, is based on a As the limiting null distributions of these test statistics depend on the
comparison between the empirical Kendall distribution and its parametric unknown parameter value q, one must resort to a parametric bootstrap or the
estimate under H0, viz. multiplier method to compute the associated p-values. From the simulation
n study in Genest et al. (2011a), it appears that the statistic TnGKNY−1 typically
n ∑ n î ) − K θn (Wî )}2 (30.18)
T = {Kˆ (W leads to a more powerful test than the statistic TnGKNY−2 . Moreover, the
i =1
30_Singh_ch30_p30.1-30.10.indd 4 8/22/16 1:53 PM

An illustration: The Fraser River at Hope 30-5
Furthermore, relationships between univariate return periods and the joint

procedure based on TnGKNY−1 outperforms the general-purpose blanket test
return period can be derived. The reader is referred to Salvadori et al. (2007)
based on the statistic
for more details about these concepts.
Comparing different multivariate return periods means comparing the
{Cˆn (u1 ,u2 ) − Cθn (u1 ,u2 )}2 dCˆn (u1 ,u2 ) (30.23) probabilities describing different sets of events corresponding to different
Tn = ∫ mechanisms of failure, only one of which describes the response of the system
[0,1]2
at hand (Serinaldi, 2015). Therefore, conclusions about supposed overestima-
tion or underestimation are illogical and misleading because every univariate,
as well as other similar procedures reviewed and studied in Genest et al.
multivariate, and conditional T and p correctly describes its own event set.
(2009c).
However, differences exist among the above multivariate return periods, not
30.3.6 Selection Criteria (Adapted AIC and BIC for Copula) only in terms of the selected design events, but also regarding their specific
statistical meaning, as pointed out in Requena et al. (2013) and Volpi and
Once GoF tests have been applied, one can rely on selection criteria to help
Fiori (2014), respectively. The reader is referred to Requena et al. (2015) for a
choose the best copula family by ranking them. Criteria such as the AICcop
recent summary. Mathematical relations can be found, for example,
and the BICcop can be used, viz.
T2 ≤ min(TX1, TX2) ≤ max(TX1, TX2) ≤ T1 ≤ T3 (30.27)
n
R S
AICcop = − 2∑ ln c  i , i ; θˆ   + 2kcop , BICcop = AICcop − {2 − ln(n)}kcop In the statistical literature, a number of multivariate extensions of the
  n + 1 n +1  
i =1
notion of quantile are available. The version often used in hydrology involves
(30.24)
the values of x1 and x2 such that H(x1, x2) = Pr(X1 ≤ x1, X2 ≤ x2) = p for a given
where c ( .,.; θ ) is the copula density function and kcop is the number of param- p in (0,1). This version is presented in Belzunce et al. (2007) and adapted to
eters of C. The selected copula is the one with the smallest criterion value. See, the hydrologic context in Chebana and Ouarda (2011). Given that there are
for example, De Michele et al. (2013) and Dias and Embrechts (2010) for typically infinitely many combinations (x1, x2) for which H(x1, x2) = p, the
illustrations in hydrology and finance, respectively. corresponding quantile is a curve. More precisely, for example, when consid-
ering the event { X1 ≤ x1 , X 2 ≤ x 2 }, the quantile curve can be expressed as
follows:
30.4 MULTIVARIATE QUANTILE AND RETURN PERIOD
Suppose that a multivariate distribution H with representation (30.1) has

{ }
QX ,Y ( p) = ( x , y ) ∈ 2 such that x = FX−1 (u), y = FY−1 (v ); u, v ∈[0,1]: C(u, v ) = p
been selected. This section describes how this model can then be used to (30.28)
estimate the quantiles and return periods that are commonly used, for This definition is simple, intuitive, interpretable, and probability-based. More
instance, as criteria in the design of hydraulic structures and water supply properties and applications of multivariate quantiles in hydrology can be
systems. In univariate HFA, the return period or recurrence interval of an found in Chebana and Ouarda (2011) and Salvadori et al. (2011b).
event of the form {X > x} or {X < x} is an estimate of its probability p of occur-
rence; the value x of the variable corresponding to a given return period p is
called a quantile. The connection between the return period (based on the 30.5 AN ILLUSTRATION: THE FRASER RIVER AT HOPE
probability of the event at hand) and the quantile (as the values of the vari-
ables defining the event) is valid both in the univariate and multivariate Yearly measurements of peak flow (Q) and volume (V) for the Fraser River at
frameworks. Indeed, multivariate extensions of these notions were recently Hope (British Columbia, Canada) are available from 1913 to 2011 without
introduced; see, for example, Chebana and Ouarda (2011) and Salvadori et al. any missing values, so the sample size is n = 99. Both series are stationary,
(2011b) for a discussion of the multivariate quantile and return period, homogeneous, and exhibit no serial dependence. The following analysis
respectively. focuses on the study of the pair (Q, V) for this station.
The notions of return period and quantile are both based on the selected A rank-plot and a K-plot of the pairs (Q, V) are given in Fig. 30.2. They
distribution and the probability of occurrence of the critical event P of inter- clearly reveal the presence of positive dependence between the two variables.
est. In contrast with the univariate case, however, structure failure may result This is confirmed by the values of Spearman’s rho (rn = 0.60) and Kendall’s
from various combinations of values for the hydrologic variables; see, for tau (tn = 0.42), which are both significantly different from 0, as easily verified
example, Volpi and Fiori (2014). with tests of independence based on these two statistics.
In the univariate case, the return period of {X > x} is defined by The issue then arises as to how best to model this dependence. For illustra-
TX = µ / {1 − Pr( X ≥ x )}, where m is the mean inter-arrival time between two tion purposes, six copula families were considered: the Clayton, Frank,
consecutive occurrences of X. By analogy, the multivariate version is thus Galambos, Gumbel, Hüsler–Reiß, and Plackett families of copulas. All of
given by them can achieve the observed values of r and t; in contrast, the FGM copula
family cannot model such a strong level of dependence and could be dis-
1 1 carded offhand.
T= = (30.25) Figure 30.3 shows scatter plots of simulated samples of size N = 500 from
p Pr(critical event) each of these copulas once their parameters have been estimated by inversion
of Kendall’s tau. It is clear from Fig. 30.3 that even for a fixed level of depen-
Various definitions of the multivariate return period have been derived dence (as measured here by Kendall’s tau), different copulas express different
from this notion. In the bivariate case, for instance, the event P could take the types of dependence between the variables: the Clayton copula exhibits lower-
forms tail dependence (accumulation of points in the lower left corner of the plot),
the three extreme-value copulas exhibit upper-tail dependence, and the
Π1 = { X1 ≥ x1 ∩ X 2 ≥ x 2 }, Π2 = { X1 ≥ x1 ∪ X 2 ≥ x 2 } remaining two copulas have neither.
Π3 = { X1 ≥ x1 | X 2 ≥ x 2 }, Π4 = { X1 ≥ x1 | X 2 ≤ x 2 } (30.26) A visual comparison of the plots gives only partial clues as to which of these
models provides the best fit of the data (represented in the rank-plot). Yet the
Π5 = { X1 ≥ x1 | X 2 = x 2 }, Π6 = {H ( X1 , X 2 ) ≥ t } choice of copula is crucial for predictive purposes, as illustrated in Fig. 30.4,
where the scatter plots of the synthetic pairs were converted to the original
for some choices of x1, x2, and t. Events P1 to P5 are all related to the prob- units, namely, a lognormal distribution with mean 9.04 and standard devia-
abilities of exceedance. Events of the type P6 are called Kendall return tion 0.19 for peak flow (Q), and a GEV distribution with parameters
periods. The choice between the multiple definitions depends on how the x = −0.22, m = 41,916.23, and b = 8,046.58 for volume (V).
system (for example, a hydraulic device) responds to a specific forcing. This As a first step in the selection of a copula structure, GoF tests can be
mechanism has a unique probabilistic description that results in a particular applied. This is done in Table 30.1 using the statistic Sn, whose p-value was
type of probability p (univariate, multivariate, conditional, etc.), which in estimated using R = 1000 parametric bootstrap samples. The only model that
turn corresponds to a unique choice of T according to the reciprocal trans- can be rejected this way is the Clayton family (p-value < 5%); all the others
formation 1/T = p. appear to be consistent with the dataset at the 5% level. Models that are not
The probability of each of the above events can be expressed in terms rejected should not be ranked in order of merit according to their p-value,
of the corresponding copula, for example, Pr(Π2 ) = 1– H ( x1 , x 2 ) = however. Adapted AIC and BIC criteria should be used instead to discrimin-
1– C{F1 ( x1 ), F2 ( x 2 )}. Definitions are also available for partial duration series. ate between them. Among the models listed in Table 30.1, the Frank copula
30_Singh_ch30_p30.1-30.10.indd 5 8/22/16 1:53 PM

Rank plot between Q and V

1.0 1.0
0.8 0.8
0.6 0.6
V
H
0.4 0.4
0.2 0.2
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Q W
Figure 30.2 Rank-plot and K-plot for the pair (Q, V).
looks like the best choice, given that it has the lowest criteria values. This is little to choose between them. In what follows, the Gumbel copula is used for
not the end of the story, however, because selecting this specific copula family convenience; the same choice was made, for example, in a recent study of
would be an implicit admission that there is no upper- or lower-tail depend- Spanish floods (Requena et al., 2013).
ence in the data. Contours associated to the fitted Gumbel copula corresponding to the
Given that joint extremes are of particular interest in HFA, one should events {U ≥ u, V ≥ v} and {U ≤ u, V ≤ v} with the same probability of
always pay attention to tail dependence. Given that the null hypothesis of occurrence are shown in Fig. 30.5.
extremeness cannot be rejected [for example, p-value = 0.17 for the origi- Figure 30.6 shows what happens to the left plot once the margins are fac-
nal test of Ghoudi et al. (1998)], a nonparametric estimate of the upper- tored in. As mentioned earlier, infinitely many pairs (Q, V) are associated to
tail dependence coefficient based on the CFG estimator is given by a given return period in the bivariate case. For comparison purposes, various
λˆ UCFG = 0.4725 . This value should be compared to the parametric estimates return periods are given in Table 30.3. One can see that the inequalities dis-
given in Table 30.2 for each of the copulas exhibiting upper-tail depen- played in Eq. (30.27) are satisfied here for each copula. This result suggests
dence. that ignoring the upper-tail dependence (which would have occurred if the
From Table 30.2, it looks as though either one of the Galambos, Gumbel, Frank copula had been preferred to the Gumbel copula) in joint extreme
and Hüsler–Reiß extreme-value copulas provides an adequate representation event modeling can lead to underestimation of the risk and becomes prob-
of the tail-dependence. In view of the AIC and BIC, there also appears to be lematic for high return periods.
Rank plot Clayton copula Frank copula Plackett copula

1.0 1.0 1.0 1.0
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6

V
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0.0 0.0 0.0 0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
U U U U
Galambos copula Gumbel copula Husler Reiss copula

1.0 1.0 1.0
0.8 0.8 0.8
0.6 0.6 0.6

V
0.4 0.4 0.4
0.2 0.2 0.2
0.0 0.0 0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
U U U
Figure 30.3 Scatter plots of simulated series from different copulas with N = 500.
30_Singh_ch30_p30.1-30.10.indd 6 8/22/16 1:53 PM

Extensions 30-7
Clayton copula Frank copula Plackett copula
60000
60000 60000
60000
Volume
Volume
Volume
Volume
40000
40000 40000
40000
20000 20000 20000

20000
6000 10000 14000 6000 10000 15000 4000 8000 12000 4000 8000 12000
Peak flow Peak flow Peak flow Peak flow
Galambos copula Gumbel copula Husler Reiss copula
60000 60000 60000

Volume
Volume
Volume
40000 40000 40000
20000
20000 20000
5000 10000 15000 4000 8000 12000 16000 4000 8000 14000
Peak flow Peak flow Peak flow
Figure 30.4 Scatter plots of simulated series from different copulas and margins with N = 500.
Table 30.1 Parameter Estimates Based on the Maximum Pseudo-Likelihood in Ben Aissia et al. (2015) and Chebana et al. (2009); for an application to
Approach, GoF Test Based on the Statistic Sn, and Criteria Results for the droughts, see Sadri and Burn (2011).
Fraser River Data at Hope
Copula Parameter Sn p-value AIC BIC 30.6.2 Nonstationary Multivariate HFA
Clayton 0.907 0.0914 0.0005 −26.22965 −23.63453 Proper assessment of risk factors requires that the statistical inference be valid
during the projected life span of the structure. However, such assumptions are
Frank 4.494 0.0228 0.2602 −40.34380 −37.74868 sometimes unrealistic. For instance, statistically significant trends have been
Plackett 6.739 0.0255 0.2133 −39.42036 −36.82524 identified in extreme values of different hydro-climatological series in differ-
Galambos 0.915 0.0351 0.0804
ent parts of the world. The reality of non-stationary hydro-meteorological
−37.29895 −34.70383
extremes needs to be properly addressed as probability distributions with
Gumbel 1.632 0.0353 0.0794 −36.46249 −33.86737 constant parameters may no longer be valid under non-stationary conditions.
Hüsler–Reiß 1.363 0.0364 0.0894 −37.50378 −34.90866 In the multivariate case, Ben Aissia et al. (2014) studied the evolution of the
dependence between flood variables through time for a long simulated cli-
mate change series; non-stationarity was showed to be present, and that it
could affect the marginal parameters as well as those of the copula. This topic
30.6 EXTENSIONS was recently considered in a hydrologic context by Bender et al. (2014) and
30.6.1 Multivariate Regional HFA Ouarda and Chebana (2014).
Regional HFA aims to assess hydrologic risks in sites where hydrologic infor- 30.6.3 Spatial Copulas
mation is unavailable or insufficient to perform an analysis. Similar informa-
tion from neighboring gaged sites can then be used to derive estimates at The motivation for using a copula to describe the dependence structure of
targeted, ungaged sites. Regional HFA consists of two main steps: (i) delinea- spatial data was first discussed by Bárdossy (2006). The objective was to
tion of homogeneous regions and (ii) regional estimation. In the multivariate develop a more general approach for characterizing spatial patterns that tra-
context, step (i) was treated in Chebana and Ouarda (2007) based on multi- ditional methods fail to capture. Building a copula that allows the formulation
variate L-moments; step (ii) was the object of Chebana et al. (2009) by extend- of the spatial structure according to the separating distance can be a difficult
ing the index-flood model. Applications to floods may be found, for example, task and not all copulas are suited for spatial analysis. The Gaussian copula is
one special case that insures continuity with existing methods. Alternatively,
Bárdossy (2006) and Bárdossy and Li (2008) proposed the ν -transformed
Table 30.2 Parametric Estimates of the Tail-Dependence Coefficient copulas that, in contrast to the Gaussian copula, account for radial asymme-
try, a property observed with variables having different interrelation between
Copula λUC ( θ ) λ̂UC the lower and upper quantiles of the marginal distribution. Recently, vine
Galambos 2 −1 θ
0.4686745 copulas have also been proposed as a tool to model various form of spatial
dependence Gräler and Pebesma (2011).
Gumbel 2 − 21 θ 0.4706775 For predictive purposes, Bárdossy and Li (2008) showed that copulas pro-
Hüsler–Reiß 2 – 2F(1/q) 0.4631934 vide a practical way of writing a predictive distribution at a new location.
Adopting a copula framework for spatial modeling is a way of generalizing
F is the univariate standard Normal distribution traditional kriging techniques such as indicator kriging, trans-Gaussian
30_Singh_ch30_p30.1-30.10.indd 7 8/22/16 1:53 PM

0.9 0.1 0.1

0.9 0.1
0.3
0.5
0 .7
0.9 0.8 0.2 0.2
0.1
0.8
0.4
0.2
0.6
0.8
0.2
0.7 0.3
0.3
0.7 0.7
0.6 0.4
0.5 0.6
0.1
0.6
0.5 0.5 0.
0.3
0.4 4
U2
0.5
U2
0.3
0.1
0.2
0.4
0.4 0.6 0.
0.4 5
0.
2
0.3 0.3
0.3 0.7
0.2 0.2
0.2
0.1
0.2 0.1
0.4
0.
0 .6
0.1 0.1 8
0.3
0.1 0.1
0.5
0.
0.2
0.7
9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
U1 U1
Simulated pairs Simulated pairs

Observed data Observed data
Contour Contours
Figure 30.5 Contours for the events {U ≤ u, V ≤ v} (left) and {U ≥ u, V ≥ v} (right) having the same probability based on a Gumbel copula fit of the Fraser River data at Hope.
80,000
70,579 0.999 (T=1000)

65,488 0.995
0.99 (T=100)
60,000 0.95
56,627 0.9 (T=10)
0.8
Volume
0.7
0.6
0.5
0.4
40,000 0.3
0.2
0.1
10,794
13,143
15,177
20,000
5000 10,000 15,000 20,000
Peak flow
Figure 30.6 Representation, on the original scales, of the contours for the events {U ≤ u, V ≤ v}
having the same probability based on a Gumbel copula fit of the Fraser River data at Hope.
Table 30.3 Joint Return Periods (TX∨,Y , TX∧,Y , and ρ∨t ) with Frank and Gumbel Copulas Corresponding to the Combinations of the Univariate Quantiles
Univariate quantiles TX∨,Y TX∧,Y ρt∨

Return periods
T = TX1 = TX2 qT (m3/s) vT (m3) Gumbel Frank Gumbel Frank Gumbel Frank
10 10,794 66,383 7 5 18 56 15 20
100 13,143 109,216 67 51 196 4,838 158 1,182
1000 15,177 181,030 670 501 1,970 476,381 1586 110,703
kriging, and normal-rank kriging Kazianka and Pilz (2010), as well as provid- qrmlib. For a historical overview of copula modeling in statistics and related
ing additional flexibility. In the context of physiographical kriging, Durocher fields, see Genest et al. (2009a). For further reading on copulas and
et al. (2016) showed that adopting the spatial copula framework to predict multivariate extremes, including threshold models, see, for example, McNeil
flood quantiles at ungaged location allows to reduce considerably the bias et al. (2005). For a current list of publications related to copulas and their
created by the failure of the normality assumption and leads to improved applications in hydrology, see www.stahy.org.
predictive power. In terms of specific references, for a short overview about multivariate HFA,
see Chebana (2012). More details about copula modeling and additional devel-
opments can be found in Genest and Nešlehová (2012b), and more specifically
30.7 RESOURCES AND FURTHER SPECIFIC REFERENCES
about extremes in Genest and Nešlehová (2012a). A number of recent papers
Most of the procedures described here are available in the R Project for dealt with extending the return period notion to the multivariate setting, such
Statistical Computing, for example, through the packages copula, envir, and as Salvadori et al. (2011a,b), Gräler et al. (2013), Requena et al. (2013).
30_Singh_ch30_p30.1-30.10.indd 8 8/22/16 1:53 PM

REFERENCES 30-9
Multivariate quantiles were adapted to the hydrologic context by Chebana Genest, C. and J-C. Boies, “Detecting dependence with Kendall plots,” The
and Ouarda (2011). All the above is related to at-site HFA. Regional estima- American Statistician, 57 (4): 275–284, 2003.
tion at ungaged sites was introduced in Chebana and Ouarda (2009); for Genest, C. and A-C. Favre, “Everything you always wanted to know about
recent applications, see, for example, Grimaldi and Serinaldi (2006), Sadri copula modeling but were afraid to ask,” Journal of Hydrologic Engineering, 12
and Burn (2011), Requena et al. (2016). In the presence of non-stationarity in (4): 347–368, 2007.
one or more of the multivariate distribution components (margins or copula), Genest, C., A-C. Favre, J. Béliveau, and C. Jacques, “Metaelliptical copulas
Bender et al. (2014) and Ouarda and Chebana (2014) proposed models to and their use in frequency analysis of multivariate hydrological data,” Water
take into account this trend. Resources Research, 43 (9): 2007.
Genest, C., M. Gendron, and M. Bourdeau-Brien, “The advent of copulas
in finance,” The European Journal of Finance, 15 (7–8): 609–618, 2009a.
ACKNOWLEDGMENTS
Genest, C., K. Ghoudi, and L-P. Rivest, “A semiparametric estimation pro-
Funding in partial support of this work was provided by the Canada Research cedure of dependence parameters in multivariate families of distributions,”
Chairs Program, the Natural Sciences and Engineering Research Council of Biometrika, 82 (3): 543–552, 1995.
Canada (RGPIN/39476-2011), the Canadian Statistical Sciences Institute, and Genest, C., I. Kojadinovic, J. Nešlehová, and J. Yan, “A goodness-of-fit test
the Fonds de recherche du Québec - Nature et technologies (2015-PR- for bivariate extreme-value copulas,” Bernoulli, 17 (1): 253–275, 2011a.
183236). Genest, C. and R. J. MacKay, “Copules archimédiennes et familles de lois
bidimensionnelles dont les marges sont données,” The Canadian Journal of
Statistics, 14 (2): 145–159, 1986.
REFERENCES
Genest, C., E. Masiello, and K. Tribouley, “Estimating copula densities
Bárdossy, A., “Copula-based geostatistical models for groundwater quality through wavelets,” Insurance: Mathematics & Economics, 44 (2): 170–181,
parameters,” Water Resources Research, 42 (11): 2006. 2009b.
Bárdossy, A. and J. Li, “Geostatistical interpolation using copulas,” Water Genest, C. and J. Nešlehová, “A primer on copulas for count data,” ASTIN
Resources Research, 44 (7): 2008. Bulletin, 37 (2): 475–515, 2007.
Belzunce, F., A. Castaño, A. Olvera-Cervantes, and A. Suárez-Llorens, Genest, C. and J. Nešlehová, “Copula modeling for extremes,” Encyclopedia
“Quantile curves and dependence structure for bivariate distributions,” Com- of Environmetrics, 2nd ed., Wiley, Chichester, 2012a, pp. 530–541.
putational Statistics & Data Analysis, 51 (10): 5112–5129, 2007. Genest, C. and J. Nešlehová, “Copulas and copula models,” Encyclopedia of
Ben Aissia, M-A., F. Chebana, T. B. M. J. Ouarda, P. Bruneau, and M. Environmetrics, 2nd ed., Wiley, Chichester, 2012b, pp. 541–553.
Barbet, “Bivariate index-flood model: case study in Québec, Canada,” Genest, C., J. Nešlehová, and N. Ben Ghorbal, “Estimators based on Ken-
Hydrological Sciences Journal, 60 (2): 247–268, 2015. dall’s tau in multivariate copula models,” Australian and New Zealand Journal
Ben Aissia, M-A., F. Chebana, T. B. M. J. Ouarda, L. Roy, P. Bruneau, and of Statistics, 53 (2): 157–177, 2011b.
M. Barbet, “Dependence evolution of hydrological characteristics, applied to Genest, C. and B. Rémillard, “Validity of the parametric bootstrap for
floods in a climate change context in Québec,” Journal of Hydrology, 519 (A): goodness-of-fit testing in semiparametric models,” Annales de l’Institut
148–163, 2014. Henri-Poincaré - Probabilités et Statistiques, 44 (6): 1096–1127, 2008.
Bender, J., T. Wahl, and J. Jensen, “Multivariate design in the presence of Genest, C., B. Rémillard, and D. Beaudoin, “Goodness-of-fit tests for
non-stationarity,” Journal of Hydrology, 514: 123–130, 2014. copulas: a review and a power study,” Insurance: Mathematics & Economics,
Ben Ghorbal, N., C. Genest, and J. Nešlehová, “On the Ghoudi, Khoudraji, 44 (2): 199–213, 2009c.
and Rivest test for extreme‐value dependence,” The Canadian Journal of Sta- Genest, C. and L-P. Rivest, “A characterization of Gumbel’s family of
tistics, 37 (4): 534–552, 2009. extreme value distributions,” Statistics & Probability Letters, 8 (3): 207–211,
Berg, D., “Copula goodness-of-fit testing: an overview and power com- 1989.
parison,” The European Journal of Finance, 15 (7–8): 675–701, 2009. Genest, C. and J. Segers, “Rank-based inference for bivariate extreme-
Berg, D. and J-F. Quessy, “Local power analyses of goodness-of-fit tests for value copulas,” The Annals of Statistics, 37 (5B): 2990–3022, 2009.
copulas,” Scandinavian Journal of Statistics, 36 (3): 389–412, 2009. Ghoudi, K., A. Khoudraji, and L-P. Rivest, “Propriétés statistiques des
Brahimi, B., F. Chebana, and A. Necir, “Copula representation of bivariate copules de valeurs extrêmes,” The Canadian Journal of Statistics, 26 (1):
L-moments: a new estimation method for multiparameter two-dimensional 187–197, 1998.
copula models,” Statistics, 49 (3): 497–521, 2015. Gräler, B. and E. Pebesma, “The pair-copula construction for spatial
Capéraà, P., A-L. Fougères, and C. Genest, “A nonparametric estimation pro- data: a new approach to model spatial dependency,” Procedia Environmental
cedure for bivariate extreme value copulas,” Biometrika, 84 (3): 567–577, 1997. Sciences, 7: 206–211, 2011.
Chebana, F., “Multivariate analysis of hydrological variables,” Encyclopedia Gräler, B., M. J. van den Berg, S. Vandenberghe, A. Petroselli, S. Grimaldi,
of Environmetrics, 2nd ed., Wiley, Chichester, 2012. B. De Baets, and N. E. C. Verhoest, “Multivariate return periods in hydrol-
Chebana, F. and T. B .M. J. Ouarda, “Multivariate L-moment homogeneity ogy: a critical and practical review focusing on synthetic design hydrograph
test,” Water Resources Research, 43 (8): 2007. estimation,” Hydrology and Earth System Sciences, 17 (4): 1281–1296, 2013.
Chebana, F. and T. B. M. J. Ouarda, “Index flood-based multivariate Grimaldi, S. and F. Serinaldi, “Design hyetograph analysis with 3-copula
regional frequency analysis,” Water Resources Research, 45 (10): 2009. function,” Hydrological Sciences Journal, 51 (2): 223–238, 2006.
Chebana, F. and T. B. M. J. Ouarda, “Multivariate quantiles in hydrological Gudendorf, G. and J. Segers, “Nonparametric estimation of an extreme-
frequency analysis,” Environmetrics, 22 (1): 63–78, 2011. value copula in arbitrary dimensions,” Journal of Multivariate Analysis, 102
Chebana, F., T. B. M. J. Ouarda, P. Bruneau, M. Barbet, S. El Adlouni, and (1): 37–47, 2011.
M. Latraverse, “Multivariate homogeneity testing in a northern case study in the Joe, H., “Multivariate dependence measures and data analysis,” Computa-
province of Québec, Canada,” Hydrological Processes, 23 (12): 1690–1700, 2009. tional Statistics & Data Analysis, 16 (3): 279–297, 1993.
Chebana, F., T. B. M. J. Ouarda, and T. C. Duong, “Testing for multivariate Joe, H., “Asymptotic efficiency of the two-stage estimation method for
trends in hydrologic frequency analysis,” Journal of Hydrology, 486: 519–530, copula-based models,” Journal of Multivariate Analysis, 94 (2): 401–419,
2013. 2005.
De Michele, C., G. Salvadori, R. Vezzoli, and S. Pecora, "Multivariate Joe, H., Dependence Modeling With Copulas, CRC Press, Boca Raton, FL,
assessment of droughts: frequency analysis and dynamic return period," 2015.
Water Resources Research, 49 (10): 6985–6994, 2013. Kao, S-C. and R. S. Govindaraju, “A bivariate frequency analysis of extreme
Dias, A. and P. Embrechts, “Modeling exchange rate dependence dynamics rainfall with implications for design,” Journal of Geophysical Research: Atmo-
at different time horizons,” Journal of International Money and Finance, 29 spheres, 112 (D13), 2007.
(8): 1687–1705, 2010. Kazianka, H. and J. Pilz, “Copula-based geostatistical modeling of contin-
Durocher, M., T. B. M. J. Ouarda, and F. Chebana, “On the prediction of uous and discrete data including covariates,” Stochastic Environmental
extreme flood quantiles at ungauged locations with spatial copula,” Journal of Research and Risk Assessment, 24 (5): 661–673, 2010.
Hydrology, 533:523–532, 2016. Kim, G., M. J. Silvapulle, and P. Silvapulle, “Comparison of semiparametric
El Adlouni, S. and T. B. M. J. Ouarda, “Study of the joint law flow-level by and parametric methods for estimating copulas,” Computational Statistics &
copulas: case of the Châteauguay river,” Canadian Journal of Civil Engineer- Data Analysis, 51 (6): 2836–2850, 2007.
ing, 35 (10): 1128–1137, 2008. Kojadinovic, I. and J. Yan, “Nonparametric rank-based tests of bivariate
Frahm, G., “On the extremal dependence coefficient of multivariate extreme-value dependence,” Journal of Multivariate Analysis, 101 (9): 2234–
distributions,” Statistics & Probability Letters, 76 (14): 1470–1481, 2006. 2249, 2010.
30_Singh_ch30_p30.1-30.10.indd 9 8/22/16 1:53 PM

Kojadinovic, I., J. Yan, and M. Holmes, “Fast large-sample goodness-of-fit Salvadori, G. and C. De Michele, “Estimating strategies for multiparameter
tests for copulas,” Statistica Sinica, 21 (2): 841–871, 2011. multivariate extreme value copulas,” Hydrology and Earth System Sciences, 15
Kotz, S. and S. Nadarajah, Extreme Value Distributions: Theory and Appli- (1): 141–150, 2011.
cations, Imperial College Press, London, 2000. Salvadori, G., C. De Michele, and F. Durante, “Multivariate design via copu-
Kurowicka, D. and H. Joe, Dependence Modeling: Handbook on Vine Copu- las,” Hydrology and Earth System Sciences Discussions, 8 (3): 5523–5558, 2011a.
lae, World Scientific, Singapore, 2011. Salvadori, G., C. De Michele, and F. Durante, “On the return period and
McNeil, A. J., R. Frey, and P. Embrechts, Quantitative Risk Management: design in a multivariate framework,” Hydrology and Earth System Sciences, 15
Concepts, Techniques and Tools, Princeton University Press, Princeton, New (11): 3293–3305, 2011b.
Jersey, 2005. Salvadori, G., C. De Michele, N. T. Kottegoda, and R. Rosso, Extremes in
McNeil, A. J. and J. Nešlehová, “From Archimedean to Liouville copulas,” Nature: An Approach Using Copulas, Springer, Dordrecht, 2007.
Journal of Multivariate Analysis, 101 (8): 1772–1790, 2010. Schmid, F. and R. Schmidt, “Multivariate extensions of Spearman’s rho and
Nelsen, R. B., An Introduction to Copulas, Springer, New York, 2006. related statistics,” Statistics & Probability Letters, 77 (4): 407–416, 2007.
Ouarda, T. B. M. J. and F. Chebana, “Multivariate non-stationarity hydro- Schmidt, R. and U. Stadtmüller, “Non-parametric estimation of tail depen-
logical frequency analysis,” Submitted, 2014. dence,” Scandinavian Journal of Statistics, 33 (2): 307–335, 2006.
Pickands III, J., “Multivariate extreme value distributions,” Proceedings of the Segers, J., “Asymptotics of empirical copula processes under nonrestrictive
43rd Session of the International Statistical Institute, Vol. 49, 1981, pp. 859–878, smoothness assumptions,” Bernoulli, 18 (3): 764–782, 2012.
894–902. Serinaldi, F., “Dismissing return periods!” Stochastic Environmental
Rao, A. R. and K. H. Hamed, Flood Frequency Analysis, CRC Press, Boca Research and Risk Assessment, 29 (4): 1179–1189, 2015.
Raton, FL, 2000. Sklar, A., “Fonctions de répartition à n dimensions et leurs marges,” Publi-
Requena, A. I., F. Chebana, and L. Mediero, “A complete procedure for cations de l’Institut de statistique de l’Université de Paris, 8: 229–231, 1959.
multivariate index-flood model application,” Journal of Hydrology, 535: 559– Tawn, J. A., “Bivariate extreme value theory: models and estimation,”
580, 2016. Biometrika, 75 (3): 397–415, 1988.
Requena, A. I., L. Mediero, and L. Garrote, “A bivariate return period Vandenberghe, S., N. E. C. Verhoest, C. Onof, and B. De Baets, “A com-
based on copulas for hydrologic dam design: accounting for reservoir routing parative copula-based bivariate frequency analysis of observed and simulated
in risk estimation,” Hydrology and Earth System Sciences, 17 (8): 3023–3038, storm events: a case study on Bartlett-Lewis modeled rainfall,” Water
2013. Resources Research, 47 (7): 2011.
Sadri, S. and D. H. Burn, “A fuzzy C-means approach for regionalization Volpi, E. and A. Fiori, “Hydraulic structures subject to bivariate hydro-
using a bivariate homogeneity and discordancy approach,” Journal of Hydrol- logical loads: return period, design, and risk assessment,” Water Resources
ogy, 401 (3–4): 231–239, 2011. Research, 50 (2): 885–897, 2014.
Salvadori, G. and C. De Michele, “Multivariate multiparameter extreme Zhang, D., M. T. Wells, and L. Peng, “Nonparametric estimation of the
value models and return periods: a copula approach,” Water Resources dependence function for a multivariate extreme value distribution,” Journal
Research, 46 (10): 2010. of Multivariate Analysis, 99 (4): 577–588, 2008.
30_Singh_ch30_p30.1-30.10.indd 10 8/22/16 1:53 PM

Chapter 31
Entropy Theory
BY
VIJAY P. SINGH
ABSTRACT 31.2.1 Real Domain
Since the development of mathematical foundations of informational entropy There are different types of entropy (Kapur, 1989; Singh, 2013), such as
by Shannon in 1948, there has been a proliferation of concepts related to Shannon, Tsallis, exponential, epsilon entropy, algorithmic, Hartley, Renyi,
entropy and their application in a multitude of areas, including hydrology. and Kapur. Here, we define only two types that are more commonly used:
Now there is a well-developed entropy theory. The theory is so versatile that Shannon and Tsallis, and the first will be detailed in the remainder of the
there is hardly any area where its application is not found. Over the past half chapter. Consider a discrete random variable X that takes on values
a century, applications of entropy have been multitudinous. The objective of x1, x2,…, xN occurring with probabilities p1, p2,…, pN, respectively. The
this chapter is to describe the entropy theory and show how it is applied in Shannon (1948) entropy is defined as the weighted average information, and
hydrology. is given as :
N
31.1 ORIGIN H = − k ∑ pi log pi (31.1)

i =1
The concept of entropy originated in statistical physics in the nineteenth
where H is Shannon entropy, and k is a parameter whose value depends on
century and is related to the second law of thermodynamics. Classical ther-
the base of the logarithm used. Depending on the base of the logarithm, the
modynamics is concerned with only macroscopic states of matter character-
unit of measurement of entropy changes. For example, the unit of entropy is
ized by observable properties, such as temperature, pressure, volume, etc. In
bit for base 2, Napier or nat for base e, and decibel or dogit for base 10. In
1850, Rudolf Clausius (Papalexiou and Koutsoyiannis, 2012) defined entropy
general, k can be taken as unity. The information gained from the occurrence
in terms of these macroscopic quantities and this is what is known as thermo-
of any event x with probability p(x) is –log p(x). Equation (31.1) satisfies a
dynamic entropy (Templeman, 1989). A macroscopic state may be made up
number of desiderata, such as continuity, symmetry, additivity, expansibility,
of various combinations of microscopic states. In order to infer on the nature
recursivity, and others (Shannon and Weaver, 1949). If X were a deterministic
of macroscopic state of a thermodynamic system, analyzing microscopic
variable, the probability that it will take on a certain value is one, and the
states of such a system, in 1872 Ludwig Boltzmann (Tamburo, 1992) formu-
probabilities of all other alternative values will be zero.
lated a new definition of entropy as a measure of disorder in the system, thus
Then, Eq. (31.1) shows that H(x) = 0 which can be viewed as the lower limit
constituting a basis for statistical thermodynamics. J. Willard Gibbs (cited in
of the values the entropy function may assume. This corresponds to the abso-
Papalexiou and Koutsoyiannis, 2012) further advanced the statistical concept
lute certainty; that is, there is no uncertainty and the system is completely
of entropy. Adopting this definition, Shannon (1948) developed the mathe-
ordered. On the other hand, when all xi’s are equally likely; that is, the variable
matical foundation of what is now known as informational entropy theory.
is uniformly distributed (pi = 1/N, i = 1, 2,…, N), then Eq. (31.1) yields
Koutsoyiannis (2013; 2014) has given a historical account of the entropy
concept.
H ( X ) = H max ( X ) = log N (31.2)
31.2 DEFINITION
This shows that the entropy function attains a maximum, and Eq. (31.2) thus
Entropy has been interpreted in many different ways. Often, it is considered defines the upper limit.
as a measure of uncertainty, information, or surprise. When an event of very If X is a continuous variable with probability density function (PDF), f(x),
low probability occurs, we are surprised because we did not anticipate its then the Shannon entropy can be written as:
occurrence. The occurrence of this event is more uncertain than an event of
higher probability. Clearly, more information will be needed to predict low- ∞ ∞
probability events than high-probability events. Information reduces uncer- H ( X ) = − ∫ f ( x )log f ( x )d x = − ∫ log[ f ( x )]d F ( x ) = E[− log f ( x )] (31.3)
tainty. If p(x) is the probability of occurrence of an event x, there are different 0 0
ways by which the uncertainty or information associated with the event can
be expressed. Likewise, if there are, say, n events, then entropy provides a Here F(x) is the cumulative (probability) distribution function (CDF) of
measure of uncertainty or information associated with these events. X, E[·] is the expectation of [·], and H(X) is the entropy of variable X.
Often, a random variable is defined in real domain, that is, in space or time, The Tsallis (1988) entropy can be expressed as:
but for certain problems, such as spectral analysis, a random variable may be N
characterized by frequency. Then, entropy is defined a little differently. In the 1 − ∑ pi m
real domain, it has been defined in different ways and hence the information k N
measure has been expressed differently.
Hm = k i =1
m −1
= ∑[ pi − pim ] (31.4)
m − 1 i=1
31-1
31_Singh_ch31_p31.1-31.8.indd 1 8/22/16 1:54 PM

31-2 Entropy Theory
where Hm is the Tsallis entropy and k is often taken as unity. For m → 1, the Clearly H(X) depends on the probability distribution of X and the same for Y.
Tsallis entropy reduces to the Shannon entropy. Exponent m can be positive H(X, Y) in Eq. (31.11) indicates the total amount of uncertainty that X and Y
or negative, characterizes the degree of nonlinearity, and is often referred to entail or the total amount of information they convey. Consider, for example,
as nonextensivity index or Tsallis entropy index. Tsallis (2002) noted that the flood process characterized by flood peak (X) and flood volume (Y) if
superextensivity, extensivity, and subextensivity occur when m < 1, m = 1, or they are independent. A practical implication is that more observations on X
m > 1, respectively. For m ≥ 0, m < 1 corresponds to the rare events and m>1 and Y would reduce the uncertainty more than more observations of only X
corresponds to frequent events (Tsallis et al., 1998). The information gain or Y.
from the occurrence of any event i can be expressed as: Let X take on values x1, x2, x3,…, xN with corresponding probabilities
N
P = {p1, p2, p3,…, pN}, and let Y take on values y1, y2, y3,…, yM with
1 corresponding probabilities Q={q1, q2, q3,…, qM}, such that
∆Ii =
m −1
(1 − pi m−1 ), ∑ pi = 1 (31.5)
i =1
N M
where ΔIi is the gain in information from an event i which occurs with prob- ∑ pi = 1, pi ≥ 0; ∑ q j = 1, q j ≥ 0 (31.12)
i =1 j =1
ability pi. For continuous variable X, the Tsallis entropy can be expressed as:
1
∞
1
∞ Note that each value of a random variable represents an event. The joint prob-
m − 1 ∫0
Hm ( X ) = Hm ( f ) = { f ( x ) − [ f ( x )]m }d x = {1 − ∫ [ f ( x )]m }d x (31.6) ability of xi and yj can be denoted as p(xi, yj) = pij; that is, p( xi , y j ) is the joint
m −1 0 probability of X = xi and Y = yj; N is the number of values X takes on; and M
is the number of values Y takes on. Then, the joint entropy of X and Y can be
If one guesses a PDF of X, referred to as a prior distribution {qi} and finds
defined as:
its a posterior distribution denoted as {pi}, then one defines relative entropy
or cross-entropy (Kullback and Leibler, 1951) as: N M
H ( X ,Y ) = − ∑ ∑ p( xi , y j )log p( xi , y j ) (31.13)
N
p
D = ∑ pi log  i  (31.7) i =1 j =1
i =1 q  i
From now onward only the Shannon entropy is discussed in this chapter. Then for the case of the flood process, the uncertainty would be represented
by Eq. (31.13). Equation (31.13) can be generalized to any number of
variables.
31.2.2 Frequency or Spectral Power Domain
Entropy is also defined in the frequency or spectral power domain for doing If p(xi, yj), i = 1, 2, 3,…, N; j = 1, 2,…, M, are the joint probabilities; and
spectral analysis wherein the random variable is frequency or spectral power. p(xi| yj) and p(yj | xi) are conditional probabilities, then the conditional entro-
Let frequency f be considered a random variable, and the normalized spectral pies H ( X Y ) and H (Y X ) can be expressed mathematically as:
density P(f) be considered its PDF. Burg (1975) defined entropy as: N M
W H ( X Y ) = − ∑ ∑ p( xi , y j )log p( xi y j ) (31.14a)
H( f ) = ∫ log[P( f )]d f (31.8) i =1 j =1
−W
where P( f ) is the power spectrum as a function of frequency related to the H (Y X ) = − ∑ p( xi , y j )log p( y j xi ) (31.14b)
autocovariance function, and W is the Nyquist frequency or band width. i, j
Equation (31.8) is often known as the Burg entropy.
The entropy of the power spectrum density can also be defined in the Equation (31.14a) or (31.14b) can be easily generalized to any number of
Shannon sense as: variables.
W Now consider the conditional entropy for two variables
denoted as H ( X Y ). The conditional entropy H(X|Y) is a measure of the
H( f ) = − ∫ P( f )ln[P( f )]d f (31.9) amount of uncertainty remaining in X even with the knowledge of Y; the same
−W
amount of information can be gained by observing X. The amount of reduc-
This definition is called the configurational entropy, which is an extension of tion in uncertainty in X equals the amount of information gained by observ-
the Burg entropy (Frieden, 1972; Gull and Daniell, 1978). ing Y. When X and Y are dependent, as may frequently be the case, say, for
The relative entropy of the spectral density can also be defined as: example, flood peak and flood volume, then their joint (or total) entropy
W
 P( f )  equals the sum of marginal entropy of the first variable (say X) and the
H ( P , q) = ∫ P( f )ln  q( f )  d f (31.10) entropy of the second variable (say Y) conditioned on the first variable (X)
−W (i.e., the uncertainty remaining in Y when a certain amount of information it
where q( f ) is the prior spectral density function and P( f ) is the posterior conveys is already present in X).
spectral density function. These three definitions are primarily used for fore-
casting. Configurational entropy can be considered as a special case of mini- H(X, Y) = H(X) + H(Y|X) (31.15)
mum relative entropy (MRE), when prior information is given as a uniform or
distribution.
H(X, Y) = H(Y) + H(X|Y) (31.16)
31.3 FORMS OF ENTROPY The conditional entropy can be expressed as

There are different types of entropy or measures of information: marginal N M
entropy, approximation entropy, sample entropy, conditional entropy, joint H ( X Y ) = − ∑ ∑ p( xi , y j )log p( xi y j ) = H ( X ,Y ) − H (Y ) (31.17)
entropy, transinformation, and interaction information. The marginal entropy i =1 j =1
is the entropy of a single variable and is defined by Eq. (31.2) if the variable is
continuous, or Eq. (31.1) if the variable is discrete. Other types of entropies N M
are defined when more than one variable is considered. H (Y X ) = − ∑ ∑ p( xi , y j )log p( y j xi ) = H ( X ,Y ) − H ( X ) (31.18)
For two random variables X and Y that are not necessarily independent, the i =1 j =1
joint entropy H ( X ,Y ) is the total information content contained in both X

and Y; that is, it is the sum of marginal entropy of one of the variables and the Transinformation represents the amount of information common to both
uncertainty that remains in the other variable when a certain amount of infor- X and Y or repeated in both or shared between X and Y, and is denoted as
mation that it can convey is already present in the first variable. This leads to T(X, Y). Transinformation is also referred to as mutual information and is a
the definition of the joint entropy of two independent variables as: measure of the dependence between X and Y and is always non-negative. It
is a measure of the amount of information random variable X contains
H(X, Y) = H(X) + H(Y) (31.11) about random variable Y; that is, if something is known about X, transinfor-
mation indicates the amount of information known about Y and vice versa.
31_Singh_ch31_p31.1-31.8.indd 2 8/22/16 1:54 PM

Total Correlation 31-3
It equals the difference between the sum of two marginal entropies and the Now the Shannon entropy can be written as:
total entropy: ∞
H (Y ) = − ∫ fY ( y )log[ fY ( y )]d y
T(X, Y) = H(X) + H(Y) — H(X, Y) (31.19) −∞
When X and Y are independent, T(X, Y) = 0. The information transmitted ∞

a a
from variable X to variable Y is represented by the mutual information = − ∫ (ay + b)(1−c )/c f X [(ay + b )1/c ]log[ (ay + b )1/c f X ((ay + b)1/c )]d y (31.26)
T(X, Y) and is given (Lathi, 1969) as: − b /a
c c
p( xi y j ) Let w = (ay + b)1/c . Then, Eq. (31.26) can be simplified as:

T ( X ,Y ) = ∑ ∑ p( xi , y j )log (31.20)
i j p( xi )
∞
a
or H (Y ) = − ∫ f X (w )[log x 1−c f X (w )]d w = log  a  + (c − 1)E[log X ] + H ( X )
c  c
p( xi , y j ) 0
T ( X ,Y ) = ∑ ∑ p( xi , y j )log (31.21) (31.27)
i j p( xi ) p( y j )
Similarly, if X is negative, that is, X ≤ 0 , and Y = [(− X )c − b]/ a, then follow-
ing the same method as previously, the entropy of Y can be expressed as:
31.4 DIRECTIONAL INFORMATION TRANSFER INDEX
Although transinformation indicates the dependence of two variables, its  a

H (Y ) = − log   + (c − 1)E[log(− X )] + H ( X ) (31.28)
upper bound varies from site to site (it varies from 0 to marginal entropy H).  c
Therefore, Yang and Burn (1984) normalized T by dividing by the joint
entropy as:
31.6 INFORMATIONAL CORRELATION COEFFICIENT
T ( H − H Lost ) H
= DIT = = 1 − Lost (31.22) The coefficient of determination r2 is a measure of the amount of variance
H H H
that can be explained by the regression relationship between random vari-
The ratio of T by H is called directional information transfer (DIT) index. ables X and Y, and 1–r2 represents the amount of unexplained variance. Here
Mogheir and Singh (2002) called it as Information Transfer Index (ITI). The r is the classical Pearson correlation coefficient under the assumption that X
physical meaning of DIT is the fraction of information transferred from one and Y are normally distributed and linearly correlated, and it is often used as
site to another. DIT varies from zero to unity when T varies from zero to H. a measure of transferred information by regression. In a similar manner, the
The zero value of DIT corresponds to the case where sites are independent informational correlation coefficient R measures the mutual dependence
and therefore no information is transmitted. A value of unity for DIT corre- between these two random variables and is a measure of transferable informa-
sponds to the case where sites are fully dependent and no information is lost. tion but does not assume any type of distributional relationship between
Any other value of DIT between zero and one corresponds to a case between them. It is a dimensionless quantity and is expressed in terms of transinfor-
fully dependent and fully independent. DIT is not symmetrical, since mation as:
DITXY = T/H(X) will not, in general, be equal to DITYX = T/H(Y). DITXY
describes the fractional information inferred by station X about station Y, R = 1 − exp(−2T0 ) (31.29)
whereas DITYX describes the fractional information inferred by station Y
about station X. where T0 is the transinformation or mutual information representing the
upper limit of transferable information between two variables X and Y. When
31.5 ENTROPY UNDER TRANSFORMATION OF X and Y are normally distributed and linearly correlated, R reduces to the
VARIABLES classical Pearson correlation coefficient. The coefficient of nontransferred
information t1 measures the percentage of information left in Y after transfer
Let there be two random variables X and Y with PDF f X ( x ) and fY ( y ) and
to X, and can be expressed as:
CDF FX ( x ) = P( X ≤ x ) and FY ( y ) = P(Y ≤ y ). Let Y = ( X − b)/ a or X = aY + b.
Then, one can write T0 − T1
t1 = (31.30)
T0
FY ( y ) = P( X ≤ ay + b ) = FX (ay + b ) and f Y ( y ) = af X (ay + b) (31.23)
where T1 is the transinformation computed for the relationship between X
and Y. For example, if the relationship is described by the bivariate normal
Following Saeb (2012), the Shannon entropy can be written as: distribution, T1 is given as:
1
∞ T1 = − ln(1 − r 2 ) (31.31)
2
H (Y ) = − ∫ fY ( y )log[ fY ( y )]d y
−∞ Otherwise, T1 is the transinformation between X and Y and T0 will be the
∞
marginal entropy of X. Quantity t1 basically describes the relative portion,
= − ∫ af X (ay + b)log[af X (ay + b)]d y T0 – T1, of the untransferred information. Likewise, 1 – t1 expresses the per-
−∞ centage of transferred information. Both R and t1 may be used to test the
∞
validity of the assumed relationship between two random variables X and Y.
= − ∫ af X (ay + b)[log(a) + log f X (ay + b)]dy
−∞
∞ 31.7 TOTAL CORRELATION
= − log a − ∫ f X (w )log f X (w )dw The total correlation can be considered as a measure of general dependence,
−∞
including both linear and nonlinear relationship, among multiple variables.
= − log a + H ( X ) or H ( X ) = log a + H (Y ) (31.24)
The total correlation for N variables ( X1 , X 2 ,, X N ) can be defined as
(McGill, 1954; Watanabe, 1960) as:
Likewise, let Y = ( X c − b)/ a , where X is a positive random variable, that is, N
x ≥ 0. Then, X = (aY + b)1/c , y > −b / a. In a similar fashion as mentioned earlier C( X1 , X 2 ,..., X N ) = ∑ H ( Xi ) − H ( X1 , X 2 ,..., X N ) (31.32)
i =1
a Equation (31.32) shows that the total correlation is always positive, because
FY ( y ) = P[ X ≤ (ay + b)1/c ] = FX [(ay + b)1/c ] , f Y ( y ) = (ay + b)(1−c )/c f X [(ay + b )1/c ] the sum of marginal entropies of the N variables will be greater than their
c
(31.25) joint entropy. It is symmetric with respect to its arguments. If these N
31_Singh_ch31_p31.1-31.8.indd 3 8/22/16 1:54 PM

31-4 Entropy Theory
variables as well as their combinations are independent, C will be zero. If N = distribution shape is gathered from empirical data, (4) constraints should be
2, Eq. (32.32) will reduce to the usual transinformation T. The total correla- such that they are more or less preserved in the future, and (5) specify
tion can be computed without resorting to the computation of multivariate constraints in terms of the laws of mathematical physics-mass conservation,
entropy by recalling the grouping property of total correlation and accord- momentum conservation, and energy conservation-or constitutive laws, if
ingly a systematic grouping of bivariate entropies (Kraskov et al., 2003). possible. It is convenient to express the information in terms of moments,
such as mean, variance, covariance, cross-covariance, or linear combinations
of these statistics. The constraints encode relevant information.
31.8 THEORY OF ENTROPY
If observations are available, then, for simplicity of analysis, let the infor-
The entropy theory can be composed of three main components: (1) defining mation on the random variable, Xε (a, b), be expressed in terms of constraints
entropy, (2) entropy optimizing, and (3) concentration theorem for testing the Cr, r = 0, 1, 2,…, n, defined as:
entropy-based probability distribution. The first component is already b
discussed and remaining two are now briefly discussed. C0 = ∫ f ( x )d x = 1 (31.33)
31.8.1 Entropy Maximizing a
Jaynes (1957) formulated the principle of maximum entropy (POME) that b

states that the minimally prejudiced assignment of probabilities is that which Cr = ∫ g r ( x ) f ( x )d x = g r ( x ), r = 1,2,..., n (31.34)
maximizes entropy subject to the given information; that is, POME takes into a
account all of the given information and at the same time avoids consider-
where gr(x), r =1, 2,…, n, represent some functions of x, n denotes the number
ation of any information that is not given. POME reasons that for given
of constraints, and g r ( x ) is the expectation of gr(x), with g0(x) = 1. Equation
information the best possible distribution that fits the data would be the one
(31.34) states that the PDF must satisfy the total probability. Other con-
with the maximum entropy, since this contains the most reliable assignment
straints, defined by Eq. (31.34), have physical meaning. For example, if r = 1
of probabilities. The information usually included in POME is specified as
and g1(x) = x, Eq. (31.34) would correspond to the mean x ; likewise, for r = 2
some statistics, including, for example, mean, variance, covariance, cross- 2
and g 2 ( x ) = ( x − x ) , it would denote the variance of X. For many hydrologic
variance, etc., or linear combinations of these statistics. Since the POME-
engineering problems, more than two or three constraints are not needed.
based distribution is favored over those with less entropy among those which
satisfy the given constraints, according to the Shannon entropy as an informa- 31.9.2 Maximization of Entropy
tion measure, entropy defines a kind of measure on the space of probability
In order to obtain the least-biased f(x), the entropy given by Eq. (31.1) is
distributions. Intuitively, distributions of higher entropy represent more dis-
maximized, subject to Eq. (31.34) and (31.34), and one simple way to achieve
order, are smoother, more probable, less predictable, or assume less. The
the maximization is the use of the method of Lagrange multipliers. To that
POME-based distribution is maximally noncommittal with regard to missing
end, the Lagrangean function L can be constructed as:
information and does not require invocation of ergodic hypotheses.
b b b
Constraints encode relevant information. POME leads to the distribution n
that is most conservative and hence most uninformative. If a distribution with L = − ∫ f ( x )ln f ( x )d x − (λ0 − 1)[ ∫ f ( x )d x − C0 ] − ∑ λr [ ∫ f ( x ) g r ( x )d x − Cr ]
r =1
lower entropy were chosen, it would mean that we would be assuming infor- a a a
(31.35)
mation that was not available, while a distribution with higher entropy would
violate the known constraints. The maximum entropy leads to a probability
distribution of particular macrostate occurring among all possible where λ1 , λ2 ,..., λn are the Lagrange multipliers. In order to obtain f(x) which
arrangements (or microstates) of the events under consideration. maximizes L, one may recall the Euler-Lagrange calculus of variation, and
The Shannon entropy is maximum when the probability distribution of the therefore one differentiates L with respect to f(x) (noting X as parameter and
random variable is that one which is as close to the a priori distribution as f as variable), equates the derivative to zero, and obtains
possible. This is referred to as the principle of minimum cross entropy
n
(POMCE) which minimizes the Bayesian entropy (Kullback and Leibler, ∂L
1951). This is equivalent to maximizing the Shannon entropy.
= 0 ⇒ −[1 + ln f ( x )] − (λ0 − 1) − ∑ λr g r ( x ) = 0 (31.36)
∂f r =1
31.8.2 Concentration Theorem

31.9.3 Probability Distribution
The concentration theorem, formulated by Jaynes (1958; 1968), has two
aspects. First, it shows the POME-based probability distribution best repre- Equation (31.36) leads to the PDF of X in terms of the given constraints:
sents our knowledge about the state of the system by showing the spread of n
lower entropies around the maximum entropy value. Second, POME is the f ( x ) = exp[− λ0 − ∑ λr g r ( x )] (31.37)
preferred method to obtain this distribution. The basis for these two aspects r =1
is contained in the Shannon inequality and the relation between entropy and
the Chi-square test. where the Lagrange multipliers λr, r = 0,1, 2,…, n, can be determined with the
use of Eqs. (31.34) and (31.34). Equation (31.37) is also written as:
n
1
31.9 METHODOLOGY FOR APPLICATION f (x ) = exp[− ∑ λr g r ( x )] (31.38)
Z (λ1 , λ2 ,..., λn ) r =1
The methodology for application of the entropy theory involves the following
steps: (1) specification of constraints, (2) maximization of entropy which can where Z = exp(λ0 ) is called the partition function. Integration of Eq. (31.37)
be accomplished using the method of Lagrange multipliers, (3) determination leads to the cumulative distribution function or simply probability distribu-
of the entropy-based PDF, (4) determination of the Lagrange multipliers in tion of X, F(x), as:
terms of constraints, (5) determination of entropy in terms of constraints, and x n
(6) derivation of the desired relation. These parts are now briefly discussed for
F ( x ) = ∫ exp[− λ0 − ∑ λr g r ( x )]d x (31.39)
a continuous variable case having a PDF as f(x) and CDF as F(x). r =1
a
31.9.1 Choice and Specification of Constraints Substitution of Eq. (31.37) in Eq. (31.1) results in the maximum entropy of
Entropy maximizing shows that there is a unique correspondence between a f(x) or X as:
probability distribution and the constraints that lead to it. Therefore, choos-
n n n
ing appropriate constraints is fundamental to the entropy formalism. The
H = λ0 + ∑ λr g r ( x ) = λ0 + ∑ λ1E[ g r ( x )] =λ 0 + ∑ λr Cr (31.40)
rationale for choosing different types of constraints in hydrology is discussed
r =1 r =1 r =1
by Papalexiou and Koutsoyiannis (2012). Constraints encode the information
or summarize the knowledge that can be garnered from empirical observa- where E[g(x)] is the expectation of g(x). Equation (31.40) shows that the
tions or theoretical considerations. entropy of the probability distribution of X depends only on the specified
There can be a large number of constraints that can perhaps summarize the constraints, since the Lagrange multipliers themselves can be expressed in
information on the random variable. Therefore, the following considerations, terms of the specified constraints. Equations (31.1), (31.34), (31.34), (31.37),
when choosing appropriate constraints, are (1) constraints should be simple, and (31.40) constitute the building blocks for application of the entropy the-
(2) constraints should be as few as absolutely needed, (3) an idea about the ory. Entropy maximization results in the probability distribution that is most
31_Singh_ch31_p31.1-31.8.indd 4 8/22/16 1:55 PM

Methodology for Application 31-5
conservative and hence most informative. If a distribution with lower entropy 31.9.5 Maximum Entropy
were chosen, it would mean the assumption of information which is not avail- Substitution of Eq. (31.37) into Eq. (31.1) yields the maximum entropy:
able. On the other hand, a distribution with higher entropy would violate the
n
known constraints. Thus, it can be stated that the maximum entropy leads to
H ( X ) = λ0 + ∑ λr g r ( x )] (31.51)
a probability distribution of a particular macro-state among all possible con- r =1
figurations of the states (or events) under consideration.
The maximum entropy is a function of the Lagrange multipliers and con-
31.9.4 Determination of Lagrange Multipliers straints. Because the Lagrange multipliers can be expressed in terms of con-
Equation (31.37) is the POME-based probability distribution containing straints, the entropy can be expressed in terms of constraints alone. It can be
Lagrange multipliers λ0, λ1, λ2,…, λn as parameters which can be determined shown that H(X) is a concave function of constraints.
by inserting Eq. (31.37) in Eqs. (31.34) and (31.34): To illustrate, consider a random variable X varying over a semi-infinite
Interval (0, ∞) and having a PDF f(x). From observations, the expected value
b n
of X, E(x), is known. The objective is to derive the PDF f(x) of X. In this case
exp(λ0 ) = ∫ exp[− ∑ λr g r ( x )]dx , r = 1,2,..., n (31.41)
the constraint equation is given as:
a r =1
∞
and
b n ∫ x f (x )dx = E[x] = x = k (31.52)
Cr = ∫ g r ( x )exp[− ∑ λr g r ( x )]dx , r = 1,2,..., n
0
(31.42)
a r =1 The least-biased f(x) is determined by maximizing Eq. (31.1), subject to Eqs.
(31.33) and (31.52). To that end, the Lagrangian L is constructed as:
Equation (31.41) can be written for the zeroth Lagrange multiplier as:
∞ ∞ ∞
b n L = − ∫ ( x ) In f ( x )dx − (λ0 − 1)[ ∫ f ( x )d x − 1] − λ1[ ∫ x f ( x )d x − k] (31.53)
λ0 = ln ∫ exp[− ∑ λr g r ( x )]dx , r = 1,2,..., n (31.43) 0 0 0
r =1
a Taking the derivative of L with respect to f(x) and equating it to 0, one obtains
Equation (31.43) shows that λ 0 is a function of λ1 , λ2 ,..., λn and expresses the ∂L
partition function Z as: = 0 ⇒ −[1 + In f ( x )] − (λ0 − 1) − λ1x = 0 (31.54)
∂ f (x )
b m
Z (λ1 , λ2 ,..., λm ) = ∫ exp[− ∑ λr g r ( x )]dx (31.44) Therefore,
a r =1 f(x) = exp(–λ0 – λ1x) (31.55)
Differentiating Eq. (31.43) with respect to λr, one gets Equation (31.55) is the POME-based distribution with λ0 and λ1 as parameters.
Substituting Eq. (31.55) in Eq. (31.34), one obtains
b n
∫ g r ( x )exp[−∑ λr g r ( x )]dx

∞

∂λ0
=− a b
r =1
, r = 1,2,..., n (31.45) ∫ exp (− λ0 − λ1x ) dx = λ1 exp (λ0 ) = 1 (31.56)
∂λr n 0
∫ exp[−∑ λr g r ( x )]dx
a r =1 Substituting Eq. (31.55) in Eq. (31.51), one gets
Multiplying the numerator as well as the denominator of Eq. (31.45) by
exp(–λ0), one obtains f ( x ) = λ1 exp (− λ1x ) (31.57)
b n
Inserting Eq. (31.57) in Eq. (31.34), one gets
∂λ0
∫ g r ( x )exp[− λ0 − ∑ λr g r ( x )]dx
=− a
r =1
, r = 1,2,..., n (31.46) ∞
1
∂λr b n ∫ λ1x exp (− λ1x ) dx = k or λ1 = k (31.58)
∫ exp[− λ0 − ∑ λr g r (x )]dxr =1
0
a
Thus Eq. (31.57) becomes
The denominator in Eq. (31.46) equals unity by virtue of Eq. (31.33).
Therefore, Eq. (31.46) becomes 1 x
f ( x ) = exp(− ), k = x (31.59)
b
k k
n
∂λ0
= − ∫ g r ( x )exp[− λ0 − ∑ λr g r ( x )]dx , r = 1,2,..., n (31.47) which is the exponential distribution.
∂λr a r =1
Note that substitution of Eq. (31.47) in Eq. (31.34) yields 31.9.6 Entropy Minimization
b n For a continuous random variable X in the range 0 and infinity, let q(x) be the
Cr = ∫ g r ( x )exp[− λ0 − ∑ λr g r ( x )]dx , r = 1,2,..., n (31.48) prior PDF and p(x) be the posterior PDF. Note q(x) is a prior estimate of p(x).
a r =1 The objective is to determine p(x) subject to specified constraints and given
the prior PDF. Then the relative entropy or cross-entropy of p(x) relative to
Therefore, Eqs. (31.47) and (31.48) yield q(x) is expressed by Eq. (31.58) where both the prior and the posterior PDFs
satisfy:
∂λ0
= −C r , r = 1, 2,..., n (31.49) ∞
∂λr ∫ q(x )dx = 1 (31.60)
0
Likewise, Eq. (31.43) can be written analytically as:
∞
λ0 = λ0 (λ1 , λ2 ,.., λn ) (31.50)

∫ p(x )dx = 1 (31.61)
0
In order to derive p(x) by applying POMCE, the following constraints can be

Equation (31.50) can be differentiated with respect to λr, r = 1, 2,…, n, and specified:
each derivative can be equated to the corresponding derivative in Eq. (31.49).
This would lead to a system of n–1 equations with n–1 unknowns, whose ∞
solution would lead to the expression of Lagrange multipliers in terms of

∫ g r ( x ) p( x )dx = g r = Cr , r = 1,2..., m (31.62)
constraints. 0
31_Singh_ch31_p31.1-31.8.indd 5 8/22/16 1:55 PM

31-6 Entropy Theory
The problem is to determine p(x), subject to the information given by Eqs. Equation (31.67) in light of Eqs. (31.60), (31.52) and (31.73) becomes
(31.60) through (31.62). Equation (31.58) does not lead to a unique p(x) but
does restrict the permissible densities that could be plausible. To obtain p(x) 1
p( x ) = exp (− x / k )exp (− λ0 − λ1x − λ2 In x ) (31.74)
uniquely, D(p, q) is minimized, subject to Eqs. (31.60)–(31.62), and k
minimization can be achieved using the method of Lagrange multipliers
where the Lagrangian function can be expressed as:
which can be expressed as:
∞ ∞
p( x )
L = ∫ p( x )In dx + (λ0 − 1)[ ∫ p( x )d x − 1] x − λ2 (31.75)
0
q( x ) 0 p( x ) = exp (− x / k )exp (− λ0 − λ1x )
∞
k
m
+ ∑ λr [ ∫ g r ( x ) p( x )d x − Cr ]
(31.63)
Substitution of Eq. (31.76) in Eqs. (31.60), (31.51), and (31.74) yield,
r =1 0 respectively,
∞
where λr , r = 1,2,..., m, are the Lagrange multipliers. Minimizing the variation x − λ2
∫ exp (− x / k )exp (− λ0 − λ1x )dx = 1 (31.76)
in L with respect to p(x) or differentiating Eq. (31.63) with respect to p(x) k
0
while recalling the calculus of variation and equating the derivative to zero,
the following is obtained: ∞
x 1−λ2
∞ ∫ exp (− x / k )exp (− λ0 − λ1x )dx = x (31.77)
p( x ) m k
∫ δ p{ln[ q(x ) ] + 1 + (λ0 − 1) + ∑
0
λr g r ( x )}dx = 0 (31.64)
0 r =1
∞
x 1−λ2
Since δp is arbitrary, terms inside the parentheses must vanish. Therefore,
∫ k
In x exp (− x / k )exp (− λ0 − λ1x )dx = In x
(31.78)
0
m
p( x ) Equations (31.76)–(31.78) can be solved numerically for the Lagrange
In + 1 + (λ0 − 1) + ∑ λr g r ( x ) = 0 (31.65) multipliers.
q( x ) r =1
31.9.7 Entropy Maximizing in Frequency Domain

Equation (31.65) leads to the posterior PDF p(x):
Let a time series (e.g., streamflow) y over time t be denoted as y(t) and let yt
m
(y1, y2,…, N) define a set of observations generated sequentially in time,
p( x ) = q( x )exp[− λ0 − ∑ λr g r ( x )] (31.66) t = 1,…, N, where N is the total number of observations. Then, the time series
r =1 can be expressed using the Fourier series as:
q
The Lagrange multipliers are determined using Eqs. (31.60) through (31.62). yt = α 0 + ∑ α k cos(2π f kt ) + βk sin(2π f kt ) (31.79)

Let exp(− λ0 ) = C . Then Eq. (31.63) can be written as: k =1
m where f k is the kth frequency defined as f k = k / N , and αk and βk are the

p( x ) = q( x )C exp[− ∑ λ r g r ( x )] (31.67) Fourier coefficients defined as:
r =1
N N
which can be written in a product or multiplicative form: α0 = y ; α k = 2 ∑ yt cos(2π f kt ); β = 2 ∑ y sin(2π f t ) (31.80)
k t k
N t =1 N t =1
m
p( x ) = Cq( x ) Π exp[− λr g r ( x )] (31.68)
r =1 where y is the mean of y. The coefficients can be computed by the least square
method. If yt is normalized, α0 = 0. Frequencies f k = k / N are called harmon-
Equation (31.68) is also conveniently written as ics of the fundamental frequency 1/N.
The spectral power, xk, at a given frequency f k , is defined as:
m
1
p( x ) = q( x )exp[− ∑ λr g r ( x )] (31.69) N
Z (λ0 ) r =1 xk = (α k 2 + βk 2 ) (31.81)
2
where Z is called the partition function obtained by substituting Eq. (31.69)
in Eq. (31.61): For each frequency f k, there is a one corresponding spectral power. The
∞
power spectrum can be defined as:
m
Z = exp(λ0 ) = ∫ q( x )exp[− ∑ λr g r ( x )]dx (31.70)
N
0 r =1
G( f ) = (a f 2 + b f 2 ) (31.82)
The Lagrange multipliers are determined using Eq. (31.70) with known values 2
of constraints as: where f is the frequency, and af and bf are Fourier coefficients defined in Eq.
(31.80) with f instead of f k. The spectral density of the time series is defined
1 ∂Z ∂In Z by dividing the power spectrum by the variance of observed series
− =− = g r ( x ) = Cr (31.71)
Z ∂λr ∂λr values, σ 2 , as:
Equation (31.71) does not lend itself to an analytical solution except for P( f ) = G( f ) / σ 2
simple cases, but numerical solution is not difficult. Equation (31.71) shows (31.83)
that specific forms of p(x) depend on the specification of q(x) and gr(x). Here
two simple cases of the prior q(x) are dealt with. The spectral density defined as above is equivalent to
The prior PDF is exponential, given as: ∞ 2
∑ y(t )e−2πtif
2
1 P( f ) = = Y ( f ) (31.84)
q( x ) = exp(− x / k ) (31.72) t =−∞
k
where i = −1 , and Y(f) is the Fourier transform of y(t). The power spectrum
and the constraints are mean and mean log specified as: can be expressed in terms of autocovariance function, Rk, as:
∞ N −1

∫ In xp(x )dx = In x (31.73) G( f ) = ∑ Rk exp(−i 2π fk ) (31.85)
0 k=− N +1
31_Singh_ch31_p31.1-31.8.indd 6 8/22/16 1:55 PM

Hydrologic Modeling Using Entropy theory 31-7
and the spectral density in terms of autocorrelation function as: The second integral on the left side of Eq. (31.93) becomes the cepstrum of
N −1 autocorrelation defined in Eq. (31.92). Now the right side of Eq. (31.94) can
P( f ) = ∑ ρk exp(−i 2πfk ) (31.86) be cast as:
k =− N +1
W N N W N
where ρk is the autocorrelation of lag k, and i = −1 .
∫ (−n∑ )ei 2π fn∆t d f = − ∑ λs ∫e ∑ λsδ n−s (31.97)
i 2π fn∆t i 2π f (n− s ) ∆t
λne df =
One can also express the autocovariance and autocorrelation functions in −W =− N s =− N −W s =− N
terms of power spectrum and spectral density, respectively, by taking the
inverse Fourier transform of Eqs. (31.85) and (31.86) as:
Thus, summing Eqs. (31.95) and (31.97) and Equating to Eq. (31.96), the
W
result of integration of (31.92) is:
Rn = ∫ G( f )ei 2π fn∆t df (31.87)
−W
N
and δ n + e(n) = − ∑ λsδ n−s (31.98)
s =− N
W
ρn = ∫ P( f )ei 2π fn∆t df (31.88)
where e(n) is the cepstrum of the autocorrelation and σn is the delta function.
−W
Equation (31.98) can be expanded as a set of N linear equations:
W W
Thus, ∫ G( f )d f and ∫ P( f )d f are equivalent in the case n = 0, and the λ0 = −1 − e(0)
−W −W
λ1 = −e(1)
integral of G(f) equals R(0) and the integral of P(f) equals 1. (31.99)
The entropy theory can be applied to derive the least-biased spectral den- 
sity which is needed for forecasting. The spectral density can be derived using λ N = − e( N )
the Burg entropy, the configurational entropy, or relative entropy. To that end,
one can use either frequency f or spectral power as a random variable. The
procedure remains the same in either case and is illustrated by maximizing Equation (31.94) shows that the Lagrange multipliers can be determined from
the configurational entropy defined by Eq. (31.9), subject to constraints the values of cepstrum which entail the spectral density that is obtained from
defined by Eq. (31.88). Eq. (31.91).
Entropy maximizing can be done using the method of Lagrange multipliers For finite length of data, Nadeu (1992) developed a simple method for
in which the Lagrangian function can be formulated as: computing cepstrum based on the use of the causal part of autocorrelation,
W W
where ρ(n) is used only for 0 < n ≤ N instead of − N ≤ n ≤ N . Thus, cepstrum
N
can be estimated by the following recursive relation:
L( f ) = − ∫ P( f )ln[P( f )]d f − n∑
=− N
λn[ ∫ P( f )exp(i 2π fn∆t )d f − ρn ]
−W −W
(31.89) n−1
k
e(n) = 2[ρ(n) − ∑ e(k ) ρ(n − k )], n > 0 (31.100)
k=1 n
where λn, n = 0, 1, 2,…, N, are the Lagrange multipliers. Taking the partial
derivative of Eq. (31.80) with respect to P(f) and equating the derivative to In order to compute e(k), one needs autocorrelation from lag 0 to k. Thus, for
zero, one obtains given N lag autocorrelations, the cepstrum of autocorrelation can be com-
W
∂ L( f ) N puted up to lag N, and beyond this lag, the cepstrum is defined as 0. Then, the
= 0 = − ∫ {ln[P( f )] +1 + ∑ λn exp(i 2π fn∆t )}df (31.90) calculated cepstrum from lag 0 to T can be used to compute the Lagrange
∂ P( f ) n=− N
−W
multipliers using Eq. (31.99).
Equation (31.90) yields the least-biased P(f):
N 31.10 HYDROLOGIC MODELING USING ENTROPY THEORY
P( f ) = exp(−1 − ∑ λnei 2π fn∆t ) (31.91)
n=− N From a hydrologic point of view, applications of the entropy can be distin-
The next step is to compute the Lagrange multipliers which can be efficiently guished into three classes: (1) statistical or empirical, (2) physical, and (3)
done by cepstrum which, by definition, is the inverse Fourier transform of the mixed. Applications in the first category essentially entail analysis of data
logarithm of the spectrum: with little physics. Here the focus is on determining the probability distribu-
π
tion of the random variable under consideration, subject to specified con-
1 straints. Examples of such applications include derivation of probability
∫ log P( f ) e d f (31.92)
2 πnif
e(n) =
2π −π distributions, frequency analysis, parameter estimation, network evaluation
and design, flow forecasting, reliability assessment of water distribution sys-
It is a measure of the rate of change in the spectrum bands. Thus, taking the tems, spatial analysis, grain-size distribution, system complexity analysis,
logarithmic transform of Eq. (31.92) for computing the Lagrange multipliers, regionalization of watersheds, and clustering of data. Information is specified
one obtains in the form of constraints and the determination is based on POME or
N POMCE and entropy is maximized. Applications in this category are com-
1 + log[P( f )] = − ∑ λnei 2π fn∆t (31.93) mon in water engineering.
n=− N
In the second category, the focus is on deriving a functional relation in time
Taking the inverse Fourier transform of Eq. (31.93), one gets or space between a dependent variable and an independent variable. Here
three aspects are involved. First, the entropy theory is applied to determine
W W N
the probability distribution of the random variable under consideration.
∫ {1 + log[P( f )]}e ∫ (−n∑
i 2π fn∆t i 2π fn∆t
df = λne )ei 2π fn∆t d f (31.94) Second, the cumulative probability distribution relating the flux and concen-
−W −W =− N
tration is hypothesized. Third, the first and second aspects are connected to
where the first part of the left side of Eq. (31.94) can be denoted as: obtain the relation in time or space as the case may be. Examples in this cat-
egory include rainfall-runoff modeling, infiltration, soil moisture movement,
W W
sin(π n) velocity distribution, hydraulic geometry, channel cross-section, sediment
∫e ∫ cos(2π fn∆t )d f =
i 2π fn∆t
df = (31.95)
πn concentration and discharge, sediment yield, river bed profile, flow duration
−W −W
curve, and rating curve.
Thus, when n = 0, Eq. (31.95) is equivalent to 1; otherwise it equals 0. The third category is a mixture of the first two categories wherein analysis
Therefore, it can be written using the delta function as: is partly empirical and partly physical. Examples include geomorphic rela-
tions for elevation, slope, and fall; and reliability of water distribution systems.
 1, n = 0
W
Singh (1997; 2011a; 2011b) has reviewed entropy applications and numerous
∫e
i 2π fnt
d f = δn =  (31.96)
−W  0, n ≠ 0 examples of applications are described in Singh (2013; 2014; 2015).
31_Singh_ch31_p31.1-31.8.indd 7 8/22/16 1:55 PM

31-8 Entropy Theory
31.11 CONCLUSION Mogheir, Y. and V. P. Singh, “Application of information theory to

groundwater quality monitoring networks,” Water Resources Management, 16
The entropy theory is a versatile tool for modeling in hydrology. It can serve
(1): 37–49, 2002.
as a connecting link when modeling seemingly disparate hydrologic processes.
Nadeu, C., “Finite length cepstrum modeling—a simple spectrum estima-
In other words, it can be a unifying theory for hydrologic science and
tion technique,” Signal Process, 26 (1): 49–59, 1992.
engineering. A great advantage of the theory is that it provides a means for
Papalexiou, S. M. and D. Koutsoyiannis, “Entropy based derivation of
incorporating the known information into the derivation of probability distri-
probability distributions: a case study to daily rainfall,” Advances in Water
butions and to estimate the distribution parameters in terms of the known
Resources, 25: 51–57, 2012
information. In this manner, it obviates the need for using the usual parameter
Saeb, A., On Extreme Value Theory and Information Theory, Unpublished
estimation techniques. Also, it brings the model and the modeler closer.
Ph.D. dissertation, University of Mysore, Mysore, India, 2013.
Shannon, C. E., “A mathematical theory of communications, I and II,” Bell
REFERENCES
System Technical Journal, 27: 379–443, 1948.
Burg, J. P., Maximum entropy spectral analysis, Unpublished Ph.D. thesis, Shannon, C. E. and W. Weaver, The Mathematical Theory of Communication,
Stanford University, Palo Alto, CA, 1975, p. 123. University of Illinois Press, Urbana, IL, 1949.
Frieden, B. R., “Restoring with maximum likelihood and maximum Singh, V. P., “The use of entropy in hydrology and water resources,”
entropy,” Journal of the Optical Society of America, 62 (4): 511–517, 1972. Hydrological Processes, 11: 587–626, 1997.
Gull, S. F. and G. J. Daniell, “Image-reconstruction from incomplete and Singh, V. P., “Hydrologic synthesis using entropy theory: review,” Journal of
noisy data,” Nature, 272 (5655): 686–690, 1978. Hydrologic Engineering, 16 (5): 421–433, 2011a.
Jaynes, E. T., “Information theory and statistical mechanics,” I. Physical Singh, V. P., “Entropy theory for earth science modeling,” Indian Geological
Review, 106: 620–630, 1957. Congress Journal, 2 (2): 5–40, 2010, 2011b.
Jaynes, E. T., “Probability Theory in Science and Engineering,” Colloquium Singh, V. P., Entropy Theory in Environmental and Water Engineering,
Lectures in Pure and Applied Science, No. 4, Socony Mobil Oil Company, Wiley-Blackwell, Sussex, U.K., 2013
Dallas, TX, 1958. Singh, V. P., Entropy Theory in Hydraulic Engineering, ASCE Press, Reston,
Jaynes, E. T., “Prior probabilities,” IEEE Transactions on System Science and VA, 2014.
Cybernetics, 70: 939–952, 1968. Singh, V. P., Entropy Theory in Hydrologic Science and Engineering,
Kapur, J. N., Maximum Entropy Models in Science and Engineering, Wiley McGraw-Hill Education, New York, 2015.
Eastern, New Delhi, India, 1989. Tamburo, A. M., “Random walk between order and disorder,” Entropy and
Kapur, J. N. and H. K. Kesavan, Entropy Optimisation Principles with Energy Dissipation in Water Resources, edited by V. P. Singh and M. Fiorentino,
Applications, Academic Press, San Diego, p. 408. Kluwer Academic, Dordrecht, The Netherlands, 1992, pp. 131–136.
Koutsoyiannis, D., “Physics of uncertainty, the Gibbs paradox and Templeman, A. B., Entropy and Civil Engineering Optimization, NATO/ASI
indistinguishable particles,” Studies in History and Philosophy of Modern on Optimization and Decision Support Systems in Civil Engineering,
Physics, 44: 480–489, 2013. Edinburg, U.K., 1989, p. 17.
Koutsoyiannis, D., “Entropy: from thermodynamics to hydrology,” Entropy, Tsallis, C., “Possible generalization of Boltzmann-Gibbs statistics,” Journal
16: 1287–1314, 2014, doi: 10.3390/e16031287. of Statistical Physics, 32 (½): 479–487, 1988.
Kraskov, A., H. Stogbauer, and P. Grassberger, “Hierarchical clustering Tsallis, C., “Entropic nonextensivity: a possible measure of complexity,”
based on mutual information,” 2003, arXiv preprint q-bio. QM/0311039. Chaos, Solitons and Fractals, 12: 371–391, 2002.
Kullback, S. and R. A. Leibler, “On information and sufficiency,” Annals of Tsallis, C., R. S. Mendrs, and A. R. Plastino, “The role of constraints within
Mathematical Statistics, 22: 79–86, 1951. generalized nonextensive statistics,” Physica A, 261: 534–554, 1998.
Lathi, B. P., An Introduction to Random Signals and Communication Theory, Watanabe, S., “Information theoretical analysis of multivariate correlation,”
International Textbook Company, Scanton, PA, 1969. IBM Journal of Research and Development, 6 (1): 66–82, 1960.
McGill, W. J., “Multivariate information transmission,” Psychometrica, 19 Yang, Y. and D. H. Burn, “An entropy approach to data collection network
(2): 97–116, 1954. design,” Journal of Hydrology, 157: 307–324, 1984.
31_Singh_ch31_p31.1-31.8.indd 8 8/22/16 1:55 PM

Chapter 32
Entropy Production
Extremum Principles
BY
ROBERT K. NIVEN AND HISASHI OZAWA
ABSTRACT Paltridge MaxEP heuristic (and its conjugate)—as applied by many research-
We review a number of extremum principles of nonequilibrium thermody- ers—is that the user can omit most details of the dynamics. If correct, this
namics based on dissipation, power consumption or entropy production, and would suggest the operation of a new variational principle of science,
their significance to hydrology. The connections between these and various consistent with conservation laws but which acts outside the framework of
allied methods are reviewed and mapped using the rigorous framework sup- the four laws of equilibrium thermodynamics.
plied by maximum entropy analysis. Two conjugate maximum and minimum The aim of this chapter is to impart a more rigorous theoretical foundation
entropy production principles, originating from Paltridge, Gaggioli, and oth- and understanding of minimum and maximum entropy production (MinEP
ers, are found to be of considerable importance and warrant further investiga- and MaxEP) principles, and to review their current and possible future appli-
tions in hydrology. cations in hydrology. In Sec. 32.2, we examine the entropy production con-
cept for both macroscopic and infinitesimal dissipative systems, and then
review and attempt to classify the many variational methods based on dissipa-
32.1 INTRODUCTION tion, power consumption, or entropy production. From this review, we draw
For several decades, the maximum entropy production (MaxEP) principle of out the significance of the MaxEP principle of Paltridge for flow systems
Paltridge (1975, 1978) has been applied as a heuristic tool to infer the station- subject to particular constraints, as well as its conjugate MinEP principle for
ary state of a variety of far-from-equilibrium dissipative systems, including flow systems subject to different constraints. In Sec. 32.3, we apply the maxi-
many fluid and heat flow systems of interest to hydrologists. Examples include mum entropy method formulated in Chap. 31 to the analysis of flow systems,
the atmospheric, oceanic, and mantle circulation systems on the Earth and both locally and globally, in each case leading to an extremum principle based
other planets (e.g., Paltridge, 1975, 1978; Vanyo and Paltridge, 1981; Ozawa on a thermodynamic-like “flux potential.” Under different conditions, this
and Ohmura, 1997; Lorenz et al., 2001; Shimokawa and Ozawa, 2001, 2002; furnishes subsidiary MinEP and MaxEP principles, in the same fashion that
Ozawa et al., 2003; Kleidon and Lorenz, 2005); turbulent shear and heat con- the minimum free energy principle of thermodynamics gives rise to subsidiary
vection systems (Ozawa et al., 2001, 2003); global planetary cycles and the minimum and maximum internal energy (or enthalpy) principles. This
biosphere (Kleidon, 2004, 2009a-b, 2010a-b; Kleidon and Lorenz, 2005); analysis provides a theoretical justification of the two conjugate Paltridge
vegetation spatial distribution (Kleidon et al., 2007; del Jesus et al., 2012); principles. In Sec. 32.4, we review the existing (relatively sparse) literature on
ecosystem selection (Meysman and Bruers, 2007); photosynthesis and plant applications in hydrology, leading in Sec. 32.5 to our conclusions and recom-
optimization (Dewar et al., 2006; Dewar, 2010); fluvial geomorphology (Paik mendations for further research.
and Kumar, 2010; Beven, 2015); particle sedimentation (Chung and Vaidya,
2008, 2011); jumpwise colloidal processes (Kozvon et al., 2002); crystal 32.2 BACKGROUND AND REVIEW
growth (Martyushev and Axelrod, 2003); tropical cyclones (Ozawa and 32.2.1 Entropy Production
Shimokawa, 2015); plasma dynamics (Yoshida and Mahajan, 2008; Kawazura
From the second law of thermodynamics, the thermodynamic entropy is not
and Yoshida, 2010, 2012); and earthquake dynamics (Main and Naylor, 2008).
conserved but rather is preserved; that is, once created, it cannot be destroyed.
Several reviews and monographs have been published (Ozawa et al., 2003;
For a macroscopic open system subject to external flows and internal entro-
Kleidon and Lorenz, 2005; Martyushev and Seleznev, 2006; Kleidon, 2010a;
py-generating processes, the rate of increase in entropy is measured by the
Dewar et al., 2013). Numerous studies also reveal the existence of a “conjugate
total (thermodynamic) entropy production, as defined by (Jaumann, 1911; de
Paltridge” minimum entropy production (MinEP) principle (e.g., Paulus and
Groot and Mazur, 1984; Niven and Noack, 2014):
Gaggioli, 2004; Martyushev, 2007; Niven, 2010b; Kawazura and Yoshida,
2010, 2012), quite different to the MinEP principle of Prigogine (1967). Of ∂S
course, the application of variational methods to nonequilibrium systems has σ = + FSout in
,tot − FS ,tot ≥ 0 (32.1)
∂t
a long pedigree, with a variety of minimum or maximum dissipation, power
or entropy production principles proposed by Helmholtz, Rayleigh; Onsager, where S is the thermodynamic entropy within the system, t is time, and FSin,tot
and Machlup (1953); Prigogine (1967), and in turbulent fluid mechanics and FSin,tot are, respectively, the flow rates of thermodynamic entropy out of and
(upper-bound theory) (e.g., Malkus, 1956, 2003; Busse, 1970; Howard, 1972), into the system. For a fluid flow system represented by a control volume (CV)
finite-time thermodynamics (e.g., Salamon and Berry, 1983; Nulton et al., enclosed by its control surface (CS), this becomes (de Groot and Mazur, 1984;
1985), and engineering design (Bejan, 1996). One apparent advantage of the Niven and Noack, 2013):
32-1
32_Singh_ch32_p32.1-32.8.indd 1 8/22/16 1:55 PM

32-2 Entropy Production Extremum Principles
32.2.2 Entropy Production Extremum Principles

∂ ρs
σ = ∫∫∫ ∫∫ ( jS + ρsv ) ⋅ n dA ≥ 0
dV + (32.2) A schematic diagram is given in Fig. 32.1 of several variational methods pro-
CV
∂t CS
posed in equilibrium and nonequilibrium thermodynamics. In equilibrium
where ρ is the fluid density, s is the specific thermodynamic entropy (per unit thermodynamics, the purpose of an extremum principle is to predict the
mass of fluid), jS is the nonfluid entropy flux (per unit area of the control equilibrium position of the system. Analogously, for a nonequilibrium sys-
surface), v is the fluid velocity (whence ρsv is the fluid-borne entropy flux), tem, a useful extremum principle should facilitate the prediction of the sta-
n is the unit normal pointing out of the control surface, “⋅” is the vector scalar tionary state, usually termed the steady-state flow2. Some principles have also
product, and dV and dA are infinitesimal volume and area elements, respec- been proposed for transient flows. As evident from Fig. 32.1, a tremendous
assortment of variational methods—each denoted by a small roman letter—
tively. Using σ = ∫∫∫ σ̂ dV to define the local entropy production σ̂, expressed have been proposed. The validity of and borderlands between many of these
CV
per unit volume of fluid, we obtain from (32.2) using Gauss’ theorem: methods have not been adequately charted; indeed, the authors do not even
proclaim that the present representation is correct or complete. Furthermore,
∂ ρs Ds some methods are of relatively narrow applicability (indeed, some may not be
σ̂ = + ∇ ⋅ ( jS + ρ sv ) = ρ + ∇ ⋅ jS ≥ 0 (32.3)
∂t Dt valid at all). Briefly summarizing the essential points:
1. All of equilibrium thermodynamics and many branches of nonequilib-
where D/Dt is the substantial derivative. Whether local or global, from the
rium thermodynamics are underpinned by the maximum entropy (Max-
second law of thermodynamics the total entropy production must be non-
Ent) method (a) developed by Boltzmann (1877) and Planck (1901),
negative.1 By a standard analysis of a nonradiative system at local thermody-
rewritten in generic form by Jaynes (1957, 2003). This method is dis-
namic equilibrium, subject to flows of heat, fluid, chemical species, and with
cussed in detail in Chap. 31. For example, applying MaxEnt to the con-
chemical reactions, it can be shown that the non-fluid entropy flux and
tents of a thermodynamic system (b) gives the four laws of
entropy production are given by (de Groot and Mazur, 1984):
thermodynamics (c) and, for an isolated system, can be applied directly
1 µ (max S ) to infer its equilibrium state (d). For an open thermodynamic
jS =   jQ − ∑  c  jc (32.4) system, MaxEnt leads to a thermodynamic potential Φeq (analogous to a
T c
T
Planck potential or free energy/temperature)—representing the (nega-
1 µ τ : ∇v ˆ  G  tive) entropy of the universe—which is minimized (min Φeq ) at equilib-
σˆ = jQ ⋅∇   − ∑ jc ⋅∇  c  − − ∑ ξd ∆  d  (32.5) rium (e). Both isolated and open equilibrium formulations have strong
T c T T T 
d connections to empirical thermodynamics (e.g., Clausius 1876;
where T is the absolute temperature, jQ is the heat flux, jc and µc are respectively Gyftopoulos and Beretta, 2005) (f), finite-time thermodynamics (e.g.,
Salamon and Berry, 1983; Nulton et al., 1985) (g), and probabilistic
the flux and chemical potential of the cth chemical species, τ is the stress tensor dynamics based on the Liouville, Hamiltonian, Fokker-Planck, or Master
(here positive in compression), ξ̂d and G d are, respectively, the rate and molar equations (e.g., Lanczos, 1970; Risken, 1996) (h).
Gibbs free energy of the dth chemical reaction, and “:” is the tensor scalar product. However, while the framework of equilibrium thermodynamics indicates
the direction of spontaneous change of a system—from which the total
Equations (32.4) and (32.5) can be summarized respectively by jS = ∑ jr λr and global or local entropy production (32.1)–(32.3) must be positive—it
makes no statement concerning the rate of change. Many prominent
{ }
r
σ̂ = ∑ j ⋅ f , where j ∈ j , j , τ , ξ are generalized fluxes, λ ∈{1/ T , − µ / T }
r r r Q c d r c researchers have been confused on this point. For this reason, any justifica-
are intensive variables, and fr ∈{∇T , −∇( µc / T ), −∇v / T , −∆(G d / T )} are
r
−1 tion of an entropy production extremum principle must lie outside the
established framework of equilibrium thermodynamics.
thermodynamic forces conjugate to jr. Note that the fluid velocity (volumetric 2. The conservation of mass, momentum, energy, and charge gives rise to
flux) v does not appear as a flux in (32.5), but is connected to these relations via the conservation laws of fluid mechanics, thermodynamics, and chemi-
(32.3) or by the specific Gibbs-Duhem equality (de Groot and Mazur, 1984). For cal reactions (i), including the continuity, Navier-Stokes, energy and
systems with a potential field or interactions with electromagnetic radiation, more mass action equations (e.g., de Groot and Mazur, 1984; White, 2006).
complicated (not necessarily bilinear) relations are required (de Groot and Mazur, These can be used to directly model steady-state and time-variant flows
1984; Niven and Noack, 2014). (j). However, these analyses can involve considerable complexity, due to
Relations (32.4) and (32.5) are often extended to larger (macroscopic) the nonlinear Reynolds stress terms and chemical reaction kinetics,
systems with uniform flow rates and gradients: motivating the present search for direct variational methods.
3. A spectral representation of turbulent flow, in conjunction with dimen-
σ = ∫∫∫ σˆ dV = ∫∫∫ ∑ jr ⋅ fr dV = − ∑ Fr ∆λr (32.6) sional arguments, yields the Kolmogorov (1941) equation and related
uniform
r r
CV CV flows analyses of the turbulent energy cascade (k). Furthermore, the MaxEnt
analysis of spectral decompositions of flow systems (l), e.g., by Galerkin
where Fr ∈{ FQ , Fc } are generalized flow rates and ∆λr are their conjugate inten-
proper orthogonal decomposition, can also be used to infer spectral
sive variable differences. For example, the total steady-state entropy produc-
coefficients (modal amplitudes) in certain classes of flows (e.g., Noack
tion in an incompressible fluid flow with only turbulent dissipation—typical of
and Niven, 2012, 2013).
hydrological flows—can be calculated by (Adeyinka and Naterer, 2004):
4. As will be discussed, the application of MaxEnt to an open, flow-controlled
τ : ∇v P ρ gH L system (m) gives a thermodynamic-like potential Φst which is minimized
σ bulk = − ∫∫∫ dV = Q L = Q (32.7) at steady state (min Φst ; Niven, 2009, 2010a) (n). This behavior is analo-
CV T T T
gous to that of the thermodynamic potential of an open equilibrium sys-
where Q is the volumetric flow rate, PL is the loss in piezometric pressure, H L tem (e). Related analyses give steady-state analogs of the four laws of
is the head loss, and g is the gravitational acceleration. The sign of (32.7) thermodynamics (Niven, 2009) (o), as well as the Fluctuation Theorem (p)
indicates spontaneous flow in the direction of decreasing pressure or head. for the ratio of probabilities of forward and backward fluxes of the same
For open channel flow, H L is the negative change in total head magnitude (e.g., Evans et al., 1993). The MaxEnt analysis of hybrid systems
H L = −∆H = −∆( z + y + U 2 /2 g ) , where z is the channel base elevation, y is the constrained by both thermodynamic contents and flow rates gives the field
water surface elevation, and U is the mean cross-sectional velocity; further- of Extended Irreversible Thermodynamics (e.g., Jou et al., 1993) (x).
more, for uniform flow H L = SL ≈ Sx, where S is the bed slope, L is the down- 5. An important pair of variational principles consist of the MaxEP prin-
stream length, and x is the downstream horizontal distance. ciple of Paltridge (1975, 1978) (q) and its conjugate MinEP principle
Whether microscopic or macroscopic, for nonradiative systems (32.4) and (e.g., Paulus and Gaggioli, 2004; Martyushev, 2007; Niven, 2010b; Kawa-
(32.5) or (32.6) we can refer to pairs of conjugate fluxes and thermodynamic zura and Yoshida, 2010, 2012) (r). The former is commonly allied with a
forces, which combine to give individual components of the entropy produc- MaxEP orthogonality principle developed by Ziegler (1977) (u), often
tion. If a pair of extremum principles are identical in all respects but lead to applied to mechanical systems. These methods select the observed
opposing extrema (minimum or maximum), depending on whether they are steady state from a set of possible steady-state flows; they are therefore of
constrained by flux(es) or thermodynamic force(s), they can be termed con- considerable utility and, as discussed, have been widely applied. Exam-
jugate extremum principles. ples include the state of a heat convection system based on Rayleigh
2
The term “steady-state” is somewhat misleading, since it refers only to the mean flow
Individual components of the entropy production can be negative, for example heat
1
and not its fluctuations. A steady-state flow need not be steady in time, only in the
transfer and chemical reaction, provided they are coupled to be non-negative in total. mean.
32_Singh_ch32_p32.1-32.8.indd 2 8/22/16 1:56 PM

BACKGROUND AND REVIEW 32-3
Prob. dynamics (h)

(ab) (f) (Liouville, Hamitonian,
Extended (x) Fokker–Planck and
irreversible Steepest entropy Empirical Master eqs;
thermodynamics ascent (Beretta) thermodynamics Fisher information) ?
? ? Quantum
(b) (c) (d) (e)
? thermodynamics
MaxEnt on contents Zeroth,first, Isolated Open systems (Beretta)
second and third systems Suniv
(a) (aa)
laws of max Ssys = eq
Finite-time (g)
MaxEnt thermodynamics
thermodynamics
(Boltzmann, (m) (n)
Jaynes) Local
MaxEnt on fluxes Open (flow) systems Finite-time (z)
(variable , ) steady-state
Global MinEP limit
(o) univ st d
Steady-state
analogues of ?
(y) (q) (r)
four laws of
Upper bound MaxEP ? MinEP steady state MinEP
thermodynamics ? (ac)
theory steady state (e.g., Paulus/Gaggioli, design
Fluctuation (Malkus, etc.) (Paltridge) Yoshida, Niven) (Bejan)
theorem (p)
? ?
(steady state) (l) (u) (t) ? (w)
MaxEnt on modes Ziegler Linear regime Min/max power
(e.g., Fourier, Galerkin) orthog. (Onsager) (networks)
?
(k) (j) ? ? ?
Turbulence cascade Time-variant (s) (v)
(Kolmogorov) Onsager Prigogine
flows
(i) “min dissip” MinEP
Conservation laws
(mass, momentum,
energy, charge)
Figure 32.1 Schematic diagram of relationships between extremum and other allied principles in equilibrium and nonequilibrium systems.
number (Ozawa et al., 2001, 2003), of a turbulent flow system based on stationary state can be calculated by other methods, it is not very useful
Reynolds number (Niven, 2010b), or of a chemical degrading system (Jaynes, 1980); indeed, there does not appear to be any application of
based on chemical or biological processes (Meysman and Bruers, 2007). Prigogine’s method in any branch of engineering. Unfortunately, there is
Furthermore, the type of extremum (MinEP or MaxEP) appears to widespread confusion between Prigogine’s and other MinEP principles
depend on the choice of constraint from a given conjugate pair. For in the literature, and also an incorrect view that Prigogine’s principle
example: “contradicts” other principles, such as of Paltridge.
• For flow in series or parallel pipes, the choice of laminar or turbulent 8. For decades, a MinEP or minimum power principle has been applied to
flow regime is consistent with MaxEP for a fixed flow rate, but with the analysis of flow networks subject to fixed flow rates, primarily elec-
MinEP for a fixed head loss (Paulus and Gaggioli, 2004; Martyushev, trical circuits (e.g., Jeans, 1925) and more recently fluid flow networks
2007; Niven, 2010b). (Paulus and Gaggioli 2004) (w). A conjugate MaxEP or maximum power
• For a zonal flow heat transfer system, the choice of convection regime principle also exists under conjugate constraints of fixed potential
is consistent with MaxEP for a flux-driven system, but with MinEP for differences (e.g., Županović et al., 2004), often confused with the Pal-
a temperature-driven system (Kawazura and Yoshida, 2010, 2012). tridge principle (see Niven, 2010b). However, these two principles are of
• For a heat convection system with parallel connections such as a the same character, and can be proven to be valid only in networks with
Bénard cell, the choice of convection regime is consistent with MinEP a linear or power-law relationship between flow rates and potential
for a flux-driven system, but with MaxEP for a temperature-driven differences, in the power-low case with a common power exponent
system (Kawazura and Yoshida, 2012). (Niven, 2010b).
The full spectrum of this MaxEP/MinEP inversion for different flow 9. In the field of “upper bound theory” within turbulent fluid mechanics,
systems has not been properly delineated. Apart from the MaxEnt analy- the steady state is identified by maximizing a functional of the total,
sis based on fluxes given herein (see Sec. 32.3), few adherents of the mean, or turbulent dissipation under various constraints (e.g., Malkus,
Paltridge MaxEP principle have paid much attention to its observed 1956, 2003; Busse, 1970; Howard, 1972) (y). While enjoying some
conjugate MinEP form in the literature. success, such methods have been criticized for their seemingly ad hoc
6. A so-called “minimum dissipation” principle was developed by Onsager choice of extremum functional and lack of theoretical justification.
and Machlup (1953), building on work by Helmholtz and Rayleigh, 10. A minimum entropy or MinEP principle of rather different character
involving maximizing the entropy production less a dissipation function has been derived in finite-time thermodynamics, as a fundamental
(s). This method applies only in the linear transport regime (t), i.e., in bound to the “cost” of moving a thermodynamic system between two
which the fluxes are linear functions of the forces jr = ∑ Lrk fk , where Lrk equilibrium states at a specified (finite) rate (e.g., Salamon and Berry,
k 1983; Nulton et al., 1985) (g). The same theoretical framework has also
is the phenomenological coefficient between the rth and kth processes been applied to derive the MinEP bound for moving a flow system
(Onsager, 1931). between two steady states at a finite rate (Niven and Andresen, 2009) (z).
7. A well-known theorem was given by Prigogine and coworkers (e.g., 11. A quantum formulation of thermodynamics is given by some authors
Prigogine, 1967), involving MinEP with respect to certain fluxes or (e.g., Beretta et al., 1984) (aa). Using this formulation, or more direct
forces, constrained by their conjugate parameters (v). The theorem can approaches, a “steepest entropy ascent” principle has been formulated to
be derived in the linear transport regime (t), and selects the stationary predict the transient state of an isolated thermodynamic system
state from the set of transient states of the system. However, since the (Beretta, 2006) (ab).
32_Singh_ch32_p32.1-32.8.indd 3 8/22/16 1:56 PM

12. A quite different MinEP principle involves minimizing the entropy We see that for an open system, the differential dΦst accounts (in a negative
production of an engineered system to obtain the most efficient design sense) for the interplay between the change of (flux) entropy within the sys-
(e.g., Bejan, 1996) (ac). This principle serves a different purpose to the tem, dH* , and the transfer of (thermodynamic) entropy (in the mean) from
predictive principles examined previously. the system to the environment, d σ̂  . From this we conclude that (1) the flux
and thermodynamic entropy concepts, while distinct, are explicitly connected
through (32.15); and (2) we can interpret Φst = − ln Z as a thermodynamic-
32.3 MAXIMUM ENTROPY ANALYSIS
like potential which should be minimized at the stationary state of the system.
We now embark on MaxEnt analyses of a nonequilibrium flow system, The flux potential Φst therefore governs the state of the system. Furthermore,
respectively for an infinitesimal (local) or macroscopic control volume. depending on the states accessible to the system, the interplay between
changes in H* and σ̂  to minimize Φst can be manifested in three outcomes:
32.3.1 Analysis of an Infinitesimal Fluid Element
1. The system could increase both H*and σ̂, to give ∆H > 0 and ∆ σ̂  / κ > 0
Consider an open infinitesimal fluid element subject to instantaneous fluxes
2. The system could increase H* and decreaseσ̂, provided ∆H ≥| ∆ σˆ  / κ |≥ 0
of heat, chemical species, and momentum, as well as instantaneous rates of
various chemical reactions. The uncertainty in the system is expressed by the 3. The system could decrease H* and increase σ̂ , provided ∆ σˆ  / κ ≥| ∆H |≥ 0
joint probability of these quantities. Departing from previous treatments These outcomes are analogous to what is observed in chemical thermody-
(Niven 2009, 2010a), we here adopt the continuum assumption for each flux namics, in the competition between (so-called) “entropic” and “enthalpic”
or rate, giving the joint probability density function (pdf) defined by processes (Atkins, 1982).
Finally, we see that if one is unaware of the flux entropy H* concept, the first
 jQ ≤ ϒ Q ≤ jQ + djQ  and third outcomes above might be identified heuristically as “MaxEP pro-
  cesses,” while the middle outcome would be identified as a “MinEP process.”
 jc ≤ ϒ c ≤ jc + djc , ∀c  We also note these processes are formulated in the same manner as the
p( X )dX = Prob.   Paltridge MaxEP or conjugate MinEP form, based on products of mean
 τ ≤ ϒ τ ≤ τ + dτ  (32.8) fluxes and mean thermodynamic forces, rather than the total entropy produc-
 ˆ ˆ ˆ 
 ξd ≤ ϒ d ≤ ξd + dξd , ∀d  tion (Niven and Noack, 2014). We therefore establish a connection between
the MaxEnt analysis of an infinitesimal fluid element and the two Paltridge
extremum principles, united by a thermodynamic-like potential which is
ˆ minimized at the steady-state flow.
where X = { jQ ,{ jc }, τ ,{ξd }}, dX = djQ ∏ djc dτ ∏ dξ̂d , and ϒ x is the random
c d
variable of quantity x. The relative entropy, here termed the local flux entropy, 32.3.2 Analysis of a (Macroscopic) Control Volume
then is We now examine an open macroscopic control volume, in which instanta-
p( X ) neous fluxes of fluid, heat, and chemical species cross its control surface.
H = − ∫ d X p( X )ln
Ω
q( X ) (32.9) From (32.2) at the steady state, it is not necessary to account separately for
X
dissipation or internal chemical reactions, since their entropy production
where Ω X represents the domain of all variables. Adopting the global expec- must be manifested as heat or material fluxes through the boundary (Jaynes,
tation notation 1980). It is, however, necessary to account for the entropy carried by the
fluid velocity v through the boundary, since this is transported into or out
〈gr 〉 = ∫ d X p( X ) g r (X ) (32.10) of the system. Writing Xˆ = { j ,{ j }, v }, the total uncertainty is now expressed
Q c
ΩX by the joint probability of the fluxes, each as a function of position x on the
for any function g r , we see that maximizing (32.9) subject to normalization boundary:
(〈1〉 = 1) and specified values of each mean flux and reaction rate {〈 jQ 〉, 〈 jc 〉, 〈 τ 〉,
〈ξ̂d 〉 } gives  jQ ( x ) ≤ ϒ Q ( x ) ≤ jQ ( x ) + djQ ( x ) 
 
q ˆ p( X ˆ dx = Prob.  jc ( x ) ≤ ϒ c ( x ) ≤ jc ( x ) + djc ( x ), ∀c 
ˆ ( x ), x ) dX
  v ( x ) ≤ ϒ ( x ) ≤ v ( x ) + dv ( x )  (32.16)
p* ( X ) = exp  −ζ Q ⋅ jQ − ∑ζ c ⋅ jc − ζ τ : τ − ∑ζ d ⋅ ξd  (32.11)
 v 
Z  c d 
 x ≤ ϒ x ≤ x + dx 
 
where Z is the partition function and ζ x is the Lagrangian multiplier for
quantity x. From (32.9) and (32.11) the maximum entropy is
The control surface flux entropy can then be defined by integration around the
ˆ boundary:
H* = ln Z + ζ Q ⋅ 〈 jQ 〉 + ∑ζ c ⋅ 〈 jc 〉 + ζ τ : 〈 τ 〉 + ∑ζ d ⋅ 〈ξd 〉 (32.12)
ˆ ˆ
ˆ , x )ln p( X , x ) = − d A ˆ , x )ln p( X , x )
c d
HCV = − ∫ d x ˆ p( X ˆ p( X
Comparing the latter to the local entropy production (32.5), we see that
∫ dX
ˆ
q( X , x ) ∫∫
 ∫ dX
ˆ
q( X , x )
CS Ω Xˆ ( x ) CS Ωˆ
X(x )
each Lagrangian multiplier must be proportional to the mean gradient or
(32.17)
potential difference conjugate to its flux or rate (Niven, 2009):
where Ω Xˆ ( x ) is the domain of all fluxes at each position on the control surface.
 1 µ ∇v G ˆ 
H* = ln Z − κ −1 〈∇ 〉⋅ 〈 jQ 〉 − ∑ 〈∇ c 〉⋅ 〈 jc 〉 − 〈 〉 : 〈 τ 〉 − ∑ 〈∆ d 〉⋅ 〈ξd 〉 By an analogous argument to the previous section, maximizing (32.17) sub-
 T c
T T d
T  ject to normalization and constraints on the mean value of each flux around
ˆ the boundary {〈 jQ ( x )〉,〈 jc ( x )〉,〈 v ( x )〉}, the inferred pdf and maximum

−1 
= ln Z − κ σ 
 entropy are
(32.13)
ˆ , x ) = q exp  −η ( x ) ⋅ j ( x ) − ∑η ( x ) ⋅ j ( x ) − η ( x ) ⋅ v ( x ) (32.18)
p* ( X
where κ is a physical constant (of units J K−1 s−1 m−3), and σ̂  = ∑ 〈 jr 〉⋅〈fr 〉 is ZCS  Q Q c c v 
 c 
r
the local entropy production in the mean. Strictly, the latter is not the same as
( )
the total entropy production, since by the Reynolds decomposition
∫∫ dA ηQ (x ) ⋅〈 jQ (x )〉 + ∑ηc (x ) ⋅〈 jc (x )〉 + ηv (x ) ⋅〈 v (x )〉
*
HCS = ln ZCS + 
CS c
〈σˆ 〉 = ∑ 〈 jr ⋅ fr 〉 = ∑ 〈 jr 〉⋅〈fr 〉 + ∑ 〈 jr ' ⋅ fr '〉 = σˆ  + ∑ 〈 jr ' ⋅ fr '〉 (32.14) (32.19)

r r r r
where the last term is the mean-fluctuating component (Niven and Noack, 2014). where ηr are the Lagrangian multipliers and ZCS is the global partition func-
Equation (32.13) can be further rearranged in terms of a local flux potential: tion. Comparing (32.19) to (32.2) and (32.4), we recognize each multiplier as
the mean conjugate intensive variable or entropy density in the direction of the
Φst = − ln Z = − H* − κ −1 σˆ  unit normal, i.e., ηr ( x ) = n 〈λr ( x )〉 for jr ∈{ jQ , jc } and η v ( x ) = n 〈 ρ( x )s( x )〉.
(32.15)
From (32.19) and the steady-state form of (32.2)
32_Singh_ch32_p32.1-32.8.indd 4 8/22/16 1:56 PM

REFERENCES 32-5
This furnishes subsidiary (pseudo-) MinEP and MaxEP principles, consistent

 1 µ (x )
*
HCS ∫∫ dA n T (x ) ⋅ 〈 jQ (x )〉 − ∑ n Tc(x ) ⋅
= ln ZCS − κ −1 with the two conjugate Paltridge principles. The relatively sparse literature on
CS  c applications in hydrology and hydraulics is then reviewed. This chapter indi-
 (32.20) cates the need for further substantial, rigorous research on the theoretical
〈 jc ( x )〉+ n 〈 ρ ( x )s( x )〉⋅ 〈 v ( x )〉 development of nonequilibrium extremum principles, and their application
 throughout hydrology and hydraulics.

= ln ZCS − κ ∫∫ dAω ( x )
−1

CS
REFERENCES

where ω ( x ) is the mean half-boundary entropy production per unit area at Adeyinka, O. B. and G. F. Naterer, “Modeling of entropy production in
position x, due to the absolute fluxes out of the enclosed volume through the turbulent flows,” Journal of Fluids Engineering, 126: 893–899, 2004.
boundary [for more details see Niven and Noack (2014)]. Rearrangement Atkins, P. W., Physical Chemistry, 2nd ed., Oxford, U.K., 1982.
gives the control surface flux potential: Bejan, A., Entropy Generation Minimization, CRC Press, Boca Raton,
FL, 1996.
∫∫ dA ω (x )
− κ −1
*
ΦCS = − ln ZCS = − HCS (32.21) Beretta, G. P., “Nonlinear model dynamics for closed-system, constrained,
CS maximal-entropy-generation relaxation by energy redistribution,” Physical
Review E, 73: 026113, 2006.
As with (32.15), this again expresses the (negative) interplay between changes Beretta, G. P., E. P. Gyftopoulos, J. L. Park, and G. N. Hatsopoulos,
of (control surface flux) entropy within the system and the transfer of (ther- “Quantum thermodynamics. A new equation of motion for a single constitu-
modynamic) entropy from the system to the environment. We again recover ent of matter,” Il Nuovo Cimento, 82B (2): 169–191, 1984.
three possibilities, of which two involve apparent maximization of an entropy Beven, K., “What we see now: event-persistence and the predictability of
production term, while one involves its apparent minimization. Once again, hydro-eco-geomorphological systems,” Ecological Modelling, 298 (SI):
we obtain a connection between the MaxEnt analysis of a (macroscopic) 4–15, 2015.
control volume and the two conjugate Paltridge extremum principles, united Boltzmann, L., “Über die Beziehung zwischen dem zweiten Hauptsatze
by minimization of a thermodynamic-like potential at the steady-state flow. dewr mechanischen Wärmetheorie und der Wahrscheinlichkeitsrechnung,
respective den Sätzen über das Wärmegleichgewicht,” Wien. Ber, 76:
32.4 REVIEW OF APPLICATIONS IN HYDROLOGY AND 373–435, 1877.
HYDRAULICS Busse, F. H., “Bounds for turbulent shear flow,” Journal of Fluid Mechanics,
41 (1): 219–240, 1970.
Applications of the Paltridge MaxEP or MinEP principles to hydrology and Chung, B. J. and A. Vaidya, “An optimal principle in fluid-structure
hydraulics only commenced recently. Kleidon and Schymanski (2008), interaction,” Physica D, 237 (22): 2945–2951, 2008.
Kleidon et al., (2009), and Westhoff and Zehe (2013) developed detailed Chung, B. J. and A. Vaidya, “Non-equilibrium pattern selection in particle
MaxEP models for hydrological water balance, while Porada et al., (2011) and sedimentation,” Applied Mathematics and Computation, 218 (7):
Zehe et al., (2013) have applied MaxEP to hillslope drainage. Zehe et al., 3451–3465, 2011.
(2010) proposed a maximum dissipation model for soil infiltration. Westoff Clausius, R., Die Mechanische Wärmetheorie, F. Vieweg, Braunschwieg, 1876.
et al., (2014) and Wang et al., (2015) found the empirical relations for the de Groot, S. R., and P. Mazur, Non-Equilibrium Thermodynamics, Dover
partitioning of precipitation between runoff and evaporation—and subse- Publications, New York, 1984.
quent runoff and soil wetting—to be consistent with MaxEP. Paik and Kumar del Jesus, M., R. Foti, A. Rinaldo, and I. Rodriguez-Iturbe, “Maximum
(2010) and Beven (2015) examined the possibility of an optimality principle entropy production, carbon assimilation, and the spatial organization of veg-
for landscape evolution, while Hergarten et al., (2014) proposed a minimum etation in river basins,” Proceedings of the National Academy of Sciences U S A,
dissipation principle for surface drainage networks. Quijano and Lin (2014) 109: 20837–20841, 2012.
review a variety of optimality principles for the critical zone (rock-soil-water- Dewar, R. C., “Maximum entropy production and plant optimization theo-
air-organism interface). Other analyses based directly on MaxEnt are exam- ries,” Philosophical Transactions of the Royal Society B, 365: 1429–1435, 2010.
ined in Chap. 31. As mentioned, MaxEP may be important in controlling Dewar, R. C., D. Juretić, and P. Županović, “The functional design of the
spatial vegetation patterns (Kleidon et al., 2007; del Jesus et al., 2012), while rotary enzyme ATP synthase is consistent with maximum entropy produc-
the optimization of the photosynthesis cycle and various whole-plant pro- tion,” Chemical Physics Letters, 430: 177–182, 2006.
cesses is consistent with MaxEP (e.g., Dewar et al., 2006; Dewar, 2010). Dewar, R. C., C. Lineweaver, R. K. Niven, and K. Regenauer-Lieb (Eds.),
Further research is needed on the connection between these phenomena and Beyond the Second Law: Entropy Production and Non-Equilibrium Systems,
the MaxEnt analyses of Sec. 32.3. Springer-Verlag, Berlin, 2014.
As discussed, analyses of simple series or parallel pipe networks show that Evans, D. J., E. G. D. Cohen, and G. P. Morriss, “Probability of second law
the choice of laminar or turbulent flow is consistent with MaxEP if flow- violations in shearing steady states,” Physical Review Letters, 71 (15):
constrained or MinEP if head-constrained (Niven, 2010b). Chung and 2401–2404, 1993.
Vaidya (2011) and Vaidya (2013) found the settling of various particles at low Gyftopoulos, E. P. and G. P. Beretta, Thermodynamics, Foundations and
Reynolds number, including elongated particles, dual spheres or a sphere Applications, Dover Publications, New York, 2005.
near a wall, to be consistent with MaxEP. Shimokawa and Ozawa (2010) and Hergarten, S., G. Winkler, and S. Birk, “Transferring the concept of mini-
Ozawa and Shimokawa (2015) found that tropical cyclones evolve to a high- mum energy dissipation from river networks to subsurface flow patterns,”
EP steady state, and may tend to move along trajectories with higher rates of Hydrology and Earth System Sciences, 18 (10): 4277–4288, 2014.
entropy production. Finally, as discussed, there is now a sizeable literature on Howard, L. N., “Bounds on flow quantities,” Annual Review of Fluid
MaxEP as a driving force for the Earth’s climate system (e.g., Paltridge, 1975, Mechanics, 4: 473–494, 1972.
1978; Ozawa and Ohmura 1997; Lorenz et al., 2001; Shimokawa and Ozawa, Jaumann, G., “Geschlossenes System physikalischer und chemischer
2001, 2002; Ozawa et al., 2003; Kleidon and Lorenz, 2005), which could have Differentialgesetze,” Sitz. Akad. Wis. Wien, Math.-Nat. Klasse, Abt. 2a, 120:
many implications for global atmospheric and oceanic circulation patterns, 385–530, 1911.
water and energy cycles, and species distributions under climate change Jaynes, E. T., “Information theory and statistical mechanics,” Physical
(Kleidon, 2004, 2009a-b, 2010a-b; Kleidon et al., 2007). Review, 106: 620–630, 1957.
Jaynes, E. T., “The minimum entropy production principle,” Annual Review
of Physical Chemistry, 31: 579–601, 1980.
32.5 CONCLUSION
Jaynes, E. T., Probability Theory: The Logic of Science, Cambridge,
This chapter presents a review of extremum principles for nonequilibrium U. K., 2003.
systems, with emphasis on the MaxEP principle of Paltridge (1975, 1978) and Jeans, J., The Mathematical Theory of Electricity and Magnetism, 5th ed.,
its conjugate MinEP form. A crude “treasure map” of such principles, contain- Cambridge, U. K., 1925.
ing some hints on the known and unknown borders and connections between Jou, D., J. Casas-Vázquez, and G. Lebon, Extended Irreversible
different formulations, is provided in Fig. 32.1. The MaxEnt method of Jaynes Thermodynamics, Springer-Verlag, Berlin, 1993.
(1957) is then applied to local and global nonequilibrium systems, in each Kawazura, Y. and Z. Yoshida, “Entropy production rate in a flux-driven
case giving a flux potential which is minimized to determine the steady state. self-organizing system,” Physical Review E, 82 (6): 066403, 2010.
32_Singh_ch32_p32.1-32.8.indd 5 8/22/16 1:56 PM

Kawazura, Y. and Z. Yoshida, “Comparison of entropy production rates in Noack, B. R. and R. K. Niven, “A hierarchy of maximum entropy closures
two different types of self-organized flows: Bénard convection and zonal for Galerkin systems of incompressible flows,” Computers & Mathematics with
flow,” Physics of Plasmas, 19 (1): 012305, 2012. Applications, 65: 1558–1574, 2013.
Kleidon, A., “Beyond Gaia: thermodynamics of life and Earth system Nulton, J., P. Salamon, B. Andresen, and Q. Anmin, “Quasistatic processes
functioning,” Climatic Change, 66: 271–319, 2004. as step equilibrations,” Journal of Chemical Physics, 83: 334–338, 1985.
Kleidon, A. and S. Schymanski, “Thermodynamics and optimality of the water Onsager, L., “Reciprocal relations in irreversible processes,” Physical
budget on land: a review,” Geophysical Research Letters, 35 (20): L20404, 2008. Review, 37: 405–426 and 38: 2265–2279, 1931.
Kleidon, A., S. Schymanski, and M. Stieglitz, “Thermodynamics, Onsager, L. and S. Machlup, “Fluctuations and irreversible processes,”
irreversibility, and optimality in land surface hydrology,” edited by Physical Review, 91: 1505–1515, 1953.
K. Střelcová et al., Bioclimatology and Natural Hazards, 2009, pp. 107–118. Ozawa, H. and A. Ohmura, “Thermodynamics of a global-mean state of
Kleidon, A., “Nonequilibrium thermodynamics and maximum entropy the atmosphere—a state of maximum entropy increase,” Journal of Climate,
production in the Earth system,” Naturwissenschaften, 96 (6): 653–677, 2009a. 10: 441–445, 1997.
Kleidon, A., “Maximum entropy production and general trends in Ozawa, H., A. Ohmura, R. D. Lorenz, and T. Pujol, “The second law of
biospheric evolution,” Journal of Paleontology, 43 (8): 980–985, 2009b. thermodynamics and the global climate system: a review of the maximum
Kleidon, A., “Life, hierarchy, and the thermodynamic machinery of planet entropy production principle,” Reviews of Geophysics, 41 (4): 1–24, 2003.
Earth,” Physics of Life Reviews, 7: 424–460, 2010a. Ozawa, H. and S. Shimokawa, “Thermodynamics of a tropical cyclone:
Kleidon, A., “A basic introduction to the thermodynamics of the Earth generation and dissipation of mechanical energy in a self-driven convection
system far from equilibrium and maximum entropy production,” Philosophical system,” Tellus A, 67: 24216, 2015.
Transactions of the Royal Society B, 365: 1303–1315, 2010b. Ozawa, H, S. Shimokawa, and H. Sakuma, “Thermodynamics of fluid
Kleidon, A., K. Fraedrich, and C. Low, “Multiple steady-states in the turbulence: a unified approach to the maximum transport properties,”
terrestrial atmosphere-biosphere system: a result of a discrete vegetation Physical Review E, 64: 026303, 2001.
classification?,” Biogeosciences, 4: 707–714, 2007. Paik, K. and P. Kumar, “Optimality approaches to describe characteristic
Kleidon, A. and R. D. Lorenz (eds.), Non-Equilibrium Thermodynamics and fluvial patterns on landscapes,” Philosophical Transactions of the Royal Society
the Production of Entropy: Life, Earth and Beyond, Springer-Verlag, London B, 365 (1545): 1387–1395, 2010.
Berlin, 2005. Paltridge, G. W., “Global dynamics and climate—a system of minimum
Kovzun, I. G., N. V. Pertsov, and I. T. Protsenko, “Jumpwise development entropy exchange,” Quarterly Journal of the Royal Meteorological Society, 101:
of the processes with the participation of colloidal systems,” Colloid Journal, 475–484, 1975.
64 (3): 312–318, 2002. Paltridge, G. W., “The steady-state format of global climate,” Quarterly
Kolmogorov, A. N., “The local structure of turbulence in incompressible Journal of the Royal Meteorological Society, 104: 927–945, 1978.
viscous fluid for very large Reynolds numbers,” Doklady Akademii nauk SSSR, Paulus, D. M. and R. A. Gaggioli, “Some observations of entropy extrema
30: 299–303, 1941. in fluid flow,” Energy, 29: 2847–2500, 2004.
Lanczos, C., Variational Principles of Mechanics, 4th ed., Dover Publications, Planck, M, “Über das gesetz der Energieverteilung im Normalspektrum,”
New York, 1970. Annals of Physics, 4: 553–563, 1901.
Lorenz, R. D., J. I. Lunine, P. G. Withers, and C. P. McKay, “Titan, Mars and Porada, P., A. Kleidon, and S. J. Schymanski, “Entropy production of soil
Earth: Entropy production by latitudinal heat transport,” Geophysical Research hydrological processes and its maximization,” Earth System Dynamics, 2 (2):
Letters, 28: 415–418, 2001. 179–190, 2011.
Main, I. G. and M. Naylor, “Maximum entropy production and earthquake Prigogine, I., Introduction to Thermodynamics of Irreversible Processes, 3rd
dynamics,” Geophysical Research Letters, 35 (19): L19311, 2008. ed, Interscience, N.Y., 1967.
Malkus, W. V. R., “Outline of a theory of turbulent shear flow,” Journal of Quijano, J. and H. Lin, “Entropy in the critical zone: a comprehensive
Fluid Mechanics, 1 (5): 521–539, 1956. review,” Entropy, 16 (6): 3482–3536, 2014.
Malkus, W. V. R., “Borders of disorder: in turbulent channel flow,” Journal Risken, H., The Fokker-Planck Equation: Methods of Solution and
of Fluid Mechanics, 489: 185–198, 2003. Applications, 2nd ed., Springer-Verlag, Berlin, 1996.
Martyushev, L. M., “Some interesting consequences of the maximum Salamon, P. and R. S. Berry, “Thermodynamic length and dissipated
entropy production principle,” Journal of Experimental and Theoretical availability,” Physical Review Letters, 51: 1127–1130, 1983.
Physics, 10: 651–654, 2007. Shimokawa, S. and H. Ozawa, “On the thermodynamics of the oceanic
Martyushev, L. M. and E. G. Axelrod, “From dendrites and S-shaped general circulation: entropy increase rate of an open dissipative system and its
growth curves to the maximum entropy production principle,” JETP Letters, surroundings,” Tellus, 53A: 266–277, 2001.
78 (8): 476–479, 2003. Shimokawa, S. and H. Ozawa, “On the thermodynamics of the oceanic
Martyushev, L. M. and V. D. Seleznev, “Maximum entropy production general circulation: irreversible transition to a state with higher rate of
principle in physics, chemistry and biology,” Physics Reports, 426: 1–45, 2006. entropy production,” Quarterly Journal of the Royal Meteorological Society,
Meysman, F. J. R. and S. Bruers, “A thermodynamic perspective on food 128: 2115–2128, 2002.
webs: quantifying entropy production within detrital-based ecosystems,” Shimokawa S. and H. Ozawa, “On analysis of typhoon activities from a
Journal of Theoretical Biology, 249: 124–139, 2007. thermodynamic viewpoint,” CAS/JSC WGNE Report No. 40, World
Niven, R. K., “Steady state of a dissipative flow-controlled system and the Meteorological Organization, Geneva (Switzerland), 2.07–2.08, 2010.
maximum entropy production principle,” Physical Review E, 80 (2): 021113, 2009. Vaidya, A., MaxEP and stable configurations in fluid-solid interactions,
Niven, R. K., “Minimisation of a free-energy-like potential for edited by Dewar et al., 2013, pp. 257–276.
non-equilibrium systems at steady state,” Philosophical Transactions of the Vanjo, J. P. and G. W. Paltridge, “A model for energy dissipation at the
Royal Society B, 365: 1323–1331, 2010a. mantle-core boundary,” Geophysical Journal of the Royal Astronomical Society,
Niven, R. K., “Simultaneous extrema in the entropy production for steady- 66: 677–690, 1981.
state fluid flow in parallel pipes,” Journal of Non-Equilibrium Thermodynamics, Wang, D. B., J. S. Zhao, Y. Tang, and M. Sivapalan, “A thermodynamic
35: 347–378, 2010b. interpretation of Budyko and L’vovich formulations of annual water balance:
Niven, R. K. and B. Andresen, “Jaynes’ maximum entropy principle, proportionality hypothesis and maximum entropy production,” Water
Riemannian metrics and generalised least action bound,” edited by R. L. Resources Research, 51 (4): 3007–3016, 2015.
Dewar and F. Detering, Complex Physical, Biophysical and Econophysical Westhoff, M. C. and E. Zehe, “Maximum entropy production: can it be
Systems, World Scientific Lecture Notes in Complex Systems, World Scientific, used to constrain conceptual hydrological models?,” Hydrology and Earth
Hackensack, USA, Vol. 9, 2009. System Sciences, 17 (8): 3141–3157, 2013.
Niven, R. K. and B. R. Noack, Control volume analysis, entropy balance Westhoff, M. C., E. Zehe, and S. J. Schymanski, “Importance of temporal
and the entropy production in flow systems, In Beyond the Second Law: variability for hydrological predictions based on the maximum entropy pro-
Entropy Production and Non-Equilibrium Systems, edited by Dewar et al., duction principle,” Geophysical Research Letters, 41 (1): 67–73, 2014.
Springer-Verlag, Berlin, 129–162, 2014. White, F. M., Viscous Fluid Flow, 3rd ed., McGraw-Hill, New York, 2006.
Noack, B. R. and R. K. Niven, “Maximum-entropy closure for a Galerkin Yoshida, Z. and S. M. Mahajan, “Maximum entropy production in
model of an incompressible periodic wake,” Journal of Fluid Mechanics, self-organized plasma boundary layer: a thermodynamic discussion about
700: 187–213, 2012. turbulent heat transport,” Journal of Plasma Physics, 15 (3): 032307, 2008.
32_Singh_ch32_p32.1-32.8.indd 6 8/22/16 1:56 PM

REFERENCES 32-7
Zehe, E., T. Blume, and G. Bloschl, “The principle of ‘maximum energy Ziegler, H., An Introduction to Thermomechanics, North-Holland, N.Y.,
dissipation’: a novel thermodynamic perspective on rapid water flow in 1977.
connected soil structures,” Philosophical Transactions of the Royal Society, Županović, P., D. Juretić, and S. Botrić, “Kirchhoff›s loop law and the
B365 (1545): 1377–1386, 2010. maximum entropy production principle,” Physical Review E, 70: 056108,
Zehe, E., U. Ehret, T. Blume, A. Kleidon, U. Scherer, and M. Westhoff, 2004.
“A thermodynamic approach to link self-organization, preferential flow and
rainfall-runoff behavior,” Hydrology and Earth System Sciences, 17 (11):
4297–4322, 2013.
32_Singh_ch32_p32.1-32.8.indd 7 8/22/16 1:56 PM

Chapter 33
Data-Based Mechanistic
Modeling
BY
PETER C. YOUNG
ABSTRACT hypothetico–deductive approach to scientific research used by ordinary sci-

entists often tends to be too constrained by current paradigms: hypotheses are
There are numerous approaches to the mathematical modelling of hydrologi-
made within the current paradigm and do not often seek to question it.
cal systems. Many of these are based on an approach that the philosopher Karl
So, in a modeling context, which is better: the inductive approach that
Popper (1959) termed the “hypothetico–deductive” scientific method. Here,
concentrates on inference from experimental or monitored data, followed by
the model, or theory in more general terms, forms a hypothesis that is tested
its interpretation in physical terms; or the hypothetico–deductive approach
against data, usually obtained from experiments, often within a laboratory
that relies on the creation of a prior hypotheses that normally have a direct
where experiments can be carefully planned. Unfortunately, it is not easy to
physical interpretation and are then tested against such data? The answer is,
plan experiments in hydrology and the modeller is forced to rely on data col-
of course, that they are not mutually exclusive methodologies and should be
lected from the hydrological system during its “normal operation,” where the
combined in a constructive way to yield an array of models that satisfy differ-
planning of experiments is difficult or impossible. This chapter describes the
ent objectives. My own predilection is to concentrate on an inductive
data-based mechanistic (DBM) approach to modelling that follows an alter-
approach whenever the availability of suitable data allows for this. Such an
native “hypothetic–inductive” approach where the model structure is first
approach can often yield a useful, physically meaningful model rather quickly,
inferred from the data with the minimum of a priori assumptions and only
without being overly constrained by existing hypotheses; and, as we see later
then interpreted in physically meaningful terms that may relate to pre-
in Sec. 33.5, it may also be an aid in the falsification of hypothetico–deductive
existing hypotheses or may provide new insight into the nature of the
models.
hydrological system.
But suitable data are not always available and hypothetico–deductive simu-
lation models provide an obvious alternative in this data-scarce situation that
33.1 INTRODUCTION occurs so often in the natural sciences. Moreover, in making inferences about
the model structure, inductive analysis is normally guided by existing or new
Data-based mechanistic (DBM) modeling is a method theory that has been hypotheses concerning the physical interpretation of parsimonious model
developed over many years and exploits a largely inductive approach to scien- structures. In other words, inductive and hypothetico–deductive modelings
tific investigation and data-based modeling. The inductive approach to sci- are synergistic activities, the relative contributions of which will depend on
ence and mathematical modeling has a long history; indeed, it was the the system being modeled and the information of all types, not only time-
approach preferred by Newton (1643–1727) who wrote in his magnum opus, series data, that are available to the scientist and modeler.
the Principia Mathematica (Newton, 1687), Hypotheses non fingo, or “I frame With above caveats in mind, this chapter first outlines the main stages of
no hypotheses.” But the inductive approach was not Newton’s invention; it the DBM approach to the modeling of dynamic systems from observational
had been discussed by Francis Bacon (1561–1626) and, almost contempora- time series data. DBM modeling has been applied to the investigation of
neously with Newton, by Robert Boyle (1627–1691). dynamic processes in a wide variety of systems from engineering, through
During the twentieth century, scientific philosophers, such as Karl Popper economics to ecology. The DBM modeling philosophy recognizes that, in
(1959) and Thomas Kuhn (1962), looked at the philosophy of science in the contrast to human-designed and constructed dynamic systems, the nature of
context of the scientific research that they observed at the time; research that more evolutionary systems, particularly at the holistic or macro-level (e.g.,
was dominated by enormous advances that occurred in physics during the global climate, river catchment, macro-economy), is still not well understood.
latter part of the nineteenth century and the first half of the twentieth century. Reductionist approaches to modeling such systems, based on the aggregation
Popper, in particular, was a proponent of what he termed the hypothetico– of hypothetico–deductive models at the micro level, or the application of
deductive method. Here, the model, or theory in more general terms, forms a micro-scale laws at the macro level, often result in large models that suffer
hypothesis that is tested against data, usually obtained from experiments, from identifiability problems [sometimes termed equifinality in the hydro-
often within a laboratory where experiments could be carefully planned. And, logical literature (e.g., Beven, 1993), following von Bertalanffy (1968)], and
in Popper’s view, the aim is not to prove the hypothesis, but rather to attempt are not fully identifiable from the available data. In other words, the informa-
its “falsification” and consider it to be “conditionally valid” until falsified. tion content in the data is not sufficient to allow for the unambiguous estima-
Whether Kuhn subscribed to the hypothetico–deductive concept is not tion of all the model parameters.
clear. Rather, he viewed science from a paradigmatic standpoint in which By contrast, in DBM modeling the model is identified and estimated from
most ordinary science worked within and embroidered research according to the data in a minimally parameterized or parsimonious form that is, there-
defined paradigms, while the more fundamental achievements of science fore, clearly identifiable. Of course, this identifiable model may not answer all
were those that questioned or even overturned these current paradigms (as questions about the behavior of the dynamic system; rather it represents only
Einstein’s theories of relativity and the development of quantum mechanics that part of the system behavior that is identifiable from the available data. In
radically changed the Newtonian view of the World). In this regard, the the present world of science, this kind of inductive modeling is now more the
33-1
33_Singh_ch33_p33.1-33.12.indd 1 8/22/16 1:57 PM

33-2 Data-Based Mechanistic Modeling
domain of the statistician and time series analyst, as well as those engineers modes (see stage 5) of dynamic behavior that are so important in fore-
like myself who prefer to use efficiently parameterized, data-based models casting and control system design. Moreover, the parsimonious nature of
whose physical interpretation is inferred from the data-based model, taking a the DBM model facilitates such design because the most effective
balanced but questioning account of any existing physical interpretations and methods normally require an efficiently parameterized model.
models. 2. Inductive DBM modeling: If experimental time series data are available,
However, the DBM approach is rather different to that of the conventional the inductive stage of model synthesis can begin. Here, an appropriate
statistician and time series analyst in that it attempts to build a bridge between model structure is identified by a process of statistical inference applied
the hypothetico–deductive and inductive modeler. For example, most time directly to these real time-series data. This is normally based on a
series models are formulated as black-box, discrete-time, difference equations generic class of dynamic models. Although discrete-time models could
that do not have a very transparent connection with physical, biological, and be utilized, linear, stochastic models described by continuous-time
ecological dynamic processes of the kind that occur in nature. Whereas DBM transfer functions (i.e., lumped parameter ordinary differential
modeling encourages the identification and estimation of continuous-time, equations) are often preferred since these have advantages when used in
differential equation descriptions that normally have an immediate physical modeling physical systems (see, e.g., Sec. 33.3.1 and the examples given
interpretation, with parameters that have defined physical units (see, e.g., in Garnier and Young, 2014). If adequate time series data are available at
Garnier and Wang, 2008; Garnier and Young, 2014). In this way, the best the start of the study, this analysis will constitute the first and most
aspects of both approaches are melded together into a unified procedure that important stage of inductive DBM modeling. The result of the analysis
tries to reach the right balance between parsimony, identifiability, and realis- will be a functioning stochastic-dynamic model of the system based, at
tic descriptive ability. Of course, as we shall see in later practical examples, it this stage, entirely on inductive reasoning and avoiding any prejudicial
sometimes makes sense to use discrete-time models since these fit better into incorporation of existing hypothetico–deductive modeling results.
the context of the modeling problem being considered. However, based on the philosophy of DBM modeling, the model should
It will be argued that the DBM approach is often a more appropriate be credible to those working in the discipline or disciplines for whom the
method of scientific inference in research on natural systems, where the model is intended (here, hydrology). This credibility will, of course,
natural laws at the macro-level, as used in reductionist modeling, are depend on the discipline, but, in general terms, it will mean that the
normally untestable by planned experimentation, which is often difficult or model can be interpreted in “physically meaningful” terms, where this
impossible in the broad gamut of the natural sciences. In such applications, phrase should be construed widely to include not only physical but also
DBM modeling tends to question existing paradigms if they are not other factors that affect the hydrological system.
compatible with the measured data. In this manner, it not only helps to avoid Hypothetico–Inductive (HI-DBM) modeling: Normally, the DBM
3.
the possibility of false hypotheses and overly parameterized, poorly modeler will be well acquainted with other models and modeling
identifiable models, but also provides a compelling reason for exploiting the methods that have used previously to meet the objectives defined in
powerful and relatively novel tools of statistical inference that have been stage 1. Following the initial, purely inductive DBM analysis in stage 2,
developed to service the requirements of DBM modeling. therefore, it is necessary to consider these in relation to the DBM results.
These DBM modeling tools are conveniently collected together as Such existing models can take a range of different forms, from the large
computational routines (m-files) in the freely available CAPTAIN Toolbox for computer simulation models, as discussed later in stages 4 through 6, to
MatlabTM (website details in Conclusion). These are used to generate the relatively small conceptual models. In hydrology, for example, these have
modeling results for the examples presented in this chapter, as well as other been termed “hybrid metric-conceptual” models (Wheater et al., 1993).
hydrological examples in my book Recursive Estimation and Time Series Whatever their form, however, these models need to be evaluated
Analysis: an Introduction for the Student and Practitioner (Young, 2011) and carefully in relation to the results obtained at the inductive DBM
the many references cited therein. This latter book also provides a modeling stage 2. For example, smaller conceptual models may have
comprehensive and detailed description of the identification and estimation elements that can immediately and usefully replace any equivalent ele-
methods used for DBM modeling in this chapter. As a result, only the main ments in the inductively identified DBM model that are less satisfactory
aspects of the DBM modeling methodology are presented here, particularly in physically meaningful terms and require further elucidation in this
as they relate to the data-based modeling of hydrological systems. regard. An example of such hypothetico–inductive DBM (HI-DBM)
analysis is outlined in Sec. 33.5. In the case of large, high-order simula-
tion models, this HI-DBM analysis can lead on from any emulation
33.2 THE MAIN STAGES OF DBM MODELING
modeling that has been carried out, as discussed later in stage 6. Here,
Although the term “data-based mechanistic modeling” was first used in the dynamic behavior of the large model is emulated by a lower-order
Young and Lees (1993), the basic concepts of this DBM approach to modeling model that represents its dominant modes of dynamic behavior, and the
dynamic systems have been developed over many years. For example, they HI-DBM model may involve the consideration of conceptual elements
were first applied seriously within a hydrological context in the early 1970s, identified by such emulation.
with application to the modeling of water quality and flow in rivers (Young 4. Stochastic simulation modeling: In the initial phases of modeling, it
and Beck, 1974; Young, 1974) and set within a more general framework may be that real observational data will be scarce, so that major
shortly thereafter (Young, 1978). Since then, they have been applied to many modeling effort is often centered on simulation modeling, normally
different systems in diverse areas of application (see, e.g., Young, 1998; Lees, based on largely deterministic concepts, such as the conservation laws
2000; Young, 2006; Young and Ratto, 2009; and the prior references therein). (mass, energy momentum, etc.). One approach in this situation is to
In its latest form, the DBM modeling strategy consists of the 10 major convert these deterministic simulation equations into a stochastic form
methodological stages that are outlined as follows and considered in greater by assuming that the associated parameters and inputs are inherently
detail in later sections. While all of these procedures can be used in the uncertain and can only be characterized in some suitable stochastic
process of DBM model synthesis, they may not all be required in any specific form, such as a probability distribution function (pdf ) for the parame-
application; rather, they are possible stages and associated methodological ters and a continuous or discrete time-series model for the inputs. The
tools to be used at the discretion of the modeler. subsequent stochastic analysis normally uses Monte Carlo Simulation
1. M
odeling objectives: The important first stage in any modeling exercise (MCS) analysis to perform sensitivity analysis (see, e.g., Ratto et al.,
is to define the modeling objectives and to consider the types of model 2007, and the prior references therein) and identify the most important
that are most appropriate to meeting these objectives. Since the concept parameters which lead to the simulated model behavior. Such analysis
of DBM modeling requires adequate data if it is to be completely can be useful for DBM and HI-DBM modeling because it identifies
successful, this stage also includes considerations of scale and the likely those aspects of the large model that appear most important in defining
data availability at this scale, particularly as they relate to the defined its dominant dynamic behavior. This approach will not be discussed in
modeling objectives. Of course, the objectives of modeling are often this chapter, but a typical example of the methodology is described in
specific to the system being studied and it is difficult to generalize. In all Parkinson and Young (1998).
cases, however, the objectives normally include the enhancement of our 5. Dominant mode analysis (DMA): While sensitivity analysis can be
understanding of the system, most often in terms of its inherent dynamic useful in defining important behavioral mechanisms in a large, high-
behavior. But they may well also include the use of the model for order model, DMA has proven to be more important as a precursor to
forecasting the future behavior of the system and/or the design of control DBM modeling because it utilizes the same statistical modeling
and management systems that will ensure desired system response. methodology and tools employed in stages 2 and 3. Here, these tools are
Provided the necessary data are available, the DBM model is in an ideal applied to time-series data obtained from planned experimentation, not
form for performing such tasks since it seeks to identify the dominant on the system itself but on the high-order computer simulation model of
33_Singh_ch33_p33.1-33.12.indd 2 8/22/16 1:57 PM

Linear DBM Models 33-3
the system that, in effect, becomes a surrogate for the real system. In some manner, as well as the use of the model in a forecasting (predictive)
particular, the analysis yields low-order approximations to the high- sense; see, e.g., Young (2010b). The exact form of these DA procedures
order simulation model that are often able to explain its dynamic will depend to some extent on the nature of the application and the
response characteristics to a remarkably accurate degree with more than associated DBM model, but they can range from relatively simple, com-
99.99% of the large model output variance explained by the reduced putationally efficient methods based around techniques such as the KF,
order model output. Typical examples are the dominant mode emulation to computationally intensive numerical Bayesian procedures that exploit
of large climate simulation models in Young et al. (1996) and Young and Monte Carlo methods, such as the ensemble Kalman filter (EKF), particle
Parkinson (2002). Such a reduced-order model is referred to as nominal filter (PF), and unscented Kalman filter (UKF). This process of data
emulation model (NEM) because it emulates the dynamic behavior of the assimilation is often made simpler in the DBM case because the estima-
large simulation model for a nominal set of its parameter values. tion methods used in DBM modeling are all inherently recursive in form
6. Dynamic emulation modeling: A more complete understanding of the (Young, 2011) and so can be used directly to process monitored data in
links between a high-order simulation model and its reduced-order real-time, as it is acquired (Young, 2011, 2002; Romanowicz et al., 2006).
representation obtained in stage 5 can often be obtained by performing 10. Application: The ultimate utility of a DBM model is how well it attains
multiple DMA analysis over a user-specified range of simulation model its objectives. Normally these objectives are dominated by its ability to
parameter values. The mapping between the large and reduced-order perform defined practical tasks. In addition to the advancement of our
model parameters or responses then yields a full dynamic emulation understanding of the system being modeled, the most obvious practical
(or meta) model (DEM) that can replace the simulation model over a applications in hydrology are adaptive flow/level forecasting and control,
wide range of parameter values. This approach to high-order model as well as model-based flood warning and management systems. Nor-
emulation is based primarily on the previous success of DMA in emulat- mally, data assimilation includes forecasting rather naturally because the
ing large climate simulation models in the nominal sense mentioned updating of the model parameters and state variables includes both the
above. It was introduced in Young and Ratto (2009), while Young and interpolation and extrapolation of state variables that are important to
Ratto (2011) describe in detail two methods of emulation: stand-alone the achievement of the modeling objectives. Moreover, it is interesting to
parameter mapping and response mapping, with application to the note that forecasting is either an explicit or implicit aspect of many
emulation of the Nash-Cascade hydrological model and a large economic advanced control systems, such as those based on model predictive
model. A comprehensive discussion on emulation modeling of all types control (MPC), so that DBM models are in an ideal form for utilization
is available in Castelletti et al. (2012). in the design of such control systems within a hydrological context.
7. Hypothetico–deductive and inductive model reconciliation: If emula- As regards the implementation of the above stages in DBM modeling, it is
tion modeling has been carried out prior to the acquisition of data, then, important to note that the DBM approach is generic, so that any suitable
if at all possible, the DBM model obtained at the inductive DBM model- stochastic-dynamic model type and associated optimization procedure can be
ing stage 2 should be reconciled with the dynamic emulation version of utilized when it is applied in practice to observational time-series data.
the simulation model. Otherwise, if time-series data are available at the However, the models and statistical modeling methods outlined in the subse-
start of the study and a purely inductive DBM model is available, any quent sections have been developed specifically with the DBM type of
available and relevant emulation model that merits such investigation modeling in mind and are implemented in the CAPTAIN Toolbox.
should be considered at this stage and reconciled with the DBM model.
Although such reconciliation will depend on the nature of the application 33.3 LINEAR DBM MODELS
being considered, the DBM model obtained from the real data should
have strong similarities with the reduced-order dynamic emulation Linear models are obviously useful for modeling linearly behaving hydrologi-
model. If this is not the case, the differences need to be investigated, with cal systems. However, they can also be used for piece-wise linear systems,
the aim of linking the reduced-order model with the high-order where the behavior is linear under similar environmental conditions. For
simulation model via the parametric mapping of the dynamic emulation example, in the case of rainfall-flow data, they can describe single storm
model. An example of such analysis is described in Chap. 12 of events that occur under the same soil-moisture and flow conditions. And
Young (2011). following from this, they can form a significant component in a nonlinear
8. Conditional validation: An important stage of DBM model synthesis rainfall-flow model where the nonlinearity acts on the input rainfall, produc-
should always be an attempt at model validation (see, e.g., Young, ing the effective rainfall that provides an input to the underlying, predomi-
2001a). The word “attempt” is important since validation is a complex nantly linear system. We shall see later, in Sec. 33.4, that such a nonlinear
process and even its definition is controversial. Some academics (e.g., system is a special example of a state-dependent parameter model.
Konikow and Bredehoeft, 1992), within a ground-water context, and A typical, linear, continuous-time, DBM model is normally presented in a
Oreskes et al. (1994) in relation to the whole of the earth sciences, ques- useful transfer function (TF) form, but, to start with, these will be considered
tion even the possibility of validating models. However, one specific, also in the equivalent differential equation form that is more familiar to a
quantitative aspect of validation is widely accepted, namely “predictive wider audience. In the case where a single output variable x(t) (e.g., flow, level
validation” or “cross-validation,” in which the predictive potential of the pollution concentration, etc.) is related to nu input variables ui(t), i = 1,
model is evaluated on data other than that used in the identification and 2, . . . , nu, the simplest model takes the following general form [see, e.g.,
estimation stages of the analysis. When validated in this narrow sense, it Chap. 8 in Young (2011)]:
can be assumed that this conditionally valid model represents the best
theory of behavior currently available that has not yet been falsified in a
Popperian sense. Statistical evaluation of the model, by exploring d n x (t ) d n−1x (t ) nu
d mi ui (t − τ i )
+ a1 + . . . + an x (t ) = ∑ bi 0 + . . . + bimi u(t − τ i )
whether the statistical diagnostics are satisfactory (e.g., no significant dt n
dt n−1
i =1 dt mi
autocorrelation in the residuals or cross correlation between the residu- (33.1)
als and input variables; no evidence of un-modeled nonlinearity, etc.), is
always possible and can engender greater confidence in the conditional
validity of the model. Of course, the DBM modeler has to be careful where τi , i = 1, 2, . . . , nu are pure time delays to allow for any such delays in
when considering such diagnostics. Residual autocorrelation will almost the system dynamics, while aj , j = 1, 2, . . . , n and bij , i = 1, 2, . . . , nu; j = 0,
always be present in models obtained from the analysis of real data. 1, 2, . . . , mi are normally constant coefficients. A model such as this can be
Consequently, scientific judgment is necessary to decide whether it is transformed straightforwardly into an equivalent state-space form. More
acceptably described by a purely stochastic process (aleatory error), or importantly in the present DBM modeling context, however, it can also be
whether it is caused in part by model deficiency that needs further atten- written as a TF model in the equivalent common denominator operator form:
tion (epistemic error), see, e.g., Helton and Burmaster (1996).
9. Continuing data assimilation: Model validation is not the end of the nu bi 0s mi + bi1s mi −1 +  + bimi nu
Bi (s )
modeling process. As additional data are received, they should be used x (t ) = ∑ n n−1 ui (t − τ i ) = ∑ ui (t − τ i ) (33.2)
to evaluate further the model’s ability to meet its objectives. Then, if pos- i =1 s + a1s +  + an i =1 A(s )
sible, both the model parameters and structure can be modified if they
are inadequate in any way. This process, often referred to as data assimi- where sr is the derivative operator, i.e., sr = dr/dtr and {Bi(s); A(s)}, i = 1, 2,. . ., nu
lation, can be achieved in a variety of ways. Since most data assimilation are the appropriately defined polynomials in this operator. The model can be
methods attempt to mimic the Kalman filter (KF), however, it is likely to extended easily to additive TFs with different denominators, i.e., where A(s)
involve recursive updating of the model parameter and state estimates in in (33.2) is replaced by Ai(s). And a multi-output system can be defined as an
33_Singh_ch33_p33.1-33.12.indd 3 8/22/16 1:57 PM

amalgamation of single output models such as this. Given that state space (SS) Ai (z−1) = 1 + ai1 z−1 + ai2 z−2 + . . . + aini z−ni ; i = 1, 2, . . . , nu
models are so popular in recent hydrological systems analysis, the reader
might ask why such a multivariable TF model is favored in DBM modeling. Bi (z ) = bi0 + bi1 z + bi2 z + . . . + bimi z
−1 −1 −2 −mi
; i = 1, 2, . . . , nu
This is discussed in Young (2013), but the main reasons are (1) the unique-
(33.7)
ness and parametric efficiency of the TF model which, for example, has only
n + m parameters in the SISO case compared with n2 + n for the general SS
For convenience, the parametric nomenclature of the main transfer
alternative; and (2) the fact that it can be converted easily to any SS model
function model in Eq. (33.6a) is made the same as in (33.3a). However, it is
whose state variables best suite the nature of the hydrological system being
important to note that these parameters are quite different since, in the com-
modeled.
pletely discrete-time model (33.6a), they are dependent on the sampling
The generic model used for the analysis reported in the present paper is the
interval ∆t. In other words, for a unique continuous-time model such as
multi-input, single-output (MISO) hybrid Box-Jenkins model, which is simply
Eq. (33.3), there are an infinite number of discrete-time models (33.6). In
a stochastic version of the TF model (33.2) with different denominators and an
each case, the discrete-time model can be obtained from the continuous-time
additional, serially correlated or “colored” noise component, denoted at the
model by the continuous to discrete model transform, which is available in
sampling instants by ξ(k), k = 1, 2, . . . , N, where N is the total sample size and
Matlab as the routine c2d. This transformation is dependent on an assumption
the uniform sampling interval is denoted by ∆t time units.1 This model is
about the behavior of the input variable u(t) between its sampled values u(k),
composed of the following three component equations:
where the most common assumption is the zero-order hold (zoh), in which
the input is assumed to remain constant over the sampling interval. This
nu
Bi (s ) unique nature of the continuous-time model reveals another advantage of this
Deterministic Model: x (t ) = ∑ ui (t − τ i ) (33.3a) model form: once its structure has been identified and the associated param-
Ai (s )
i =1
eters have been estimated, it can be converted to a discrete-time model at any
sampling interval that best suits the modeling objectives.
D( z −1 )
Noise Model: ξ (k ) = e(k ); e(k ) = N (0, σ 2 ) (33.3b) The full model structure for both of the above models is defined by
C( z −1 )
[n1 . . . nnu m1 . . . mnu τ1 . . . τnu p q] (33.8)
Observation Equation: y (k ) = x (k ) + ξ (k ) (33.3c)
although, in the later practical examples, it is sometimes convenient to refer
only to the deterministic part of the model, which defines the structure of the
where the polynomials Ai (s) and Bi (s) are defined as follows: main hydrological relationship between the input and the output; i.e., p and q
are omitted, while the subscript 1, 2, . . . , nu is also omitted in the SISO case,
where the structure is simply [n m τ]. Identification of these model structure
Ai (s) = sni + ai1sni −1 + ai2 sni −2 + . . . + aini ; i = 1, 2, . . . , nu parameters, as well as the estimation of the associated polynomial parameters
(33.4) that characterize the model, is a well-known statistical problem. In DBM
Bi (s) = bi0smi + bi1smi −1 + bi2 smi −2 + . . . + bimi ; i = 1, 2, . . . , nu modeling, it is normally solved by maximum likelihood optimization using
the refined instrumental variable (RIV) methods for estimating continuous-

time (RIVC) and discrete-time (RIVD) models of this type (see Young, 2011,
and the i subscript is omitted in the single-input, single-output (SISO) case. 2015, and the prior references therein). These are available in the CAPTAIN
Note that this is a hybrid (continuous-discrete) model because the main, Toolbox as the rivcbjid and rivbjid routines, for model structure identification
underlying deterministic model in Eq. (33.3a) is represented by a continuous- and parameter estimation; and the rivcbj and rivbj routines for subsequent
time transfer function with output and input variables x(t) and final estimation of the parameters in the identified structure, together with
ui(t), i = 1, 2, . . . , nu, having a time argument t; and x(k) is the sampled their associated statistical uncertainty bounds.
measure of x(t) at the sampling instants k = 1, 2, . . . , N, with an argument k. The most common measure of how well the model explains the data is the
The y(k) observations are then x(k) contaminated by the colored noise ξ(k), simulation coefficient of determination, R2T , which is directly equivalent to the
as shown in Eq. (33.3c). This colored noise, which represents the effect of all Nash-Sutcliffe efficiency (Nash and Sutcliffe, 1970) used in hydrology, and is
stochastic influences on the system, including measurement and modeling defined as follows:
errors, as well as unmeasured or diffuse inputs, is assumed in Eq. (33.3c) to be
generated by passing the zero mean, serially uncorrelated sequence of ran-
dom variables (white noise) e(k), with variance σ2, through an autoregressive
{
R2T = 1 − σ ξ2 /σ 2
y } (33.9)
moving average (ARMA) filter (see Young, 2011), which converts the white where σ y2 is the mean square value of the flow y(k) about its mean value;
noise into the serially correlated, colored sequence of random variables ξ(k). and σ ξ2 is similarly the mean square value of the model simulation error
The associated polynomials C (z−1) and D(z−1) are defined as follows: ξ̂(k) = y(k) − x̂(k), where x̂(k) is the noise-free, deterministic output of the
model, i.e., model-generated value of the sampled output x(k) in the models
C (z−1) = 1 + c1 z−1 + c2 z−2 + . . . + cp z−p (33.3) and (33.6). There are numerous model structure (order) identification
(33.5) criteria, but only the most common of these (AIC, BIC, and YIC: see pages
D(z−1) = 1 + d1 z−1 + d2 z−2 + . . . + dq z−q 176–179 in Young, 2011) are available in rivcbjid and rivbjid. Here, the AIC
and BIC are based on a cost function that balances the variance of the residual
where z−r is the backward shift operator, i.e., z−r ξ(k) = ξ(k − r).
errors, ê(k) [or ξ̂(k) if C (z−1 ) = D(z−1 ) = 1.0], with the number of parameters
The completely discrete-time equivalent of the model 33.3 can be obtained
in the model, while the YIC is based on the properties of the instrumental
by discretization at the specified sampling interval of ∆T time units
product matrix (IPM) used in instrumental variable algorithms such as RIV
(see, e.g., App. B in Young, 2011) and takes the form:
estimation (see Wellstead, 1978; Young et al., 1980). Model order identification
nu
Bi ( z −1 ) can also be aided by recursive parameter estimation, which is available as an
System TF Model: x (k ) = ∑ u(k − δ i ) (33.6a)
Ai ( z −1 ) option in the CAPTAIN rivcbj and rivbj routines. In particular, it can help in
i =1
diagnosing whether there are any signs of either time variation (see Sec. 33.3),
poor identifiability, or the presence of epistemic error, where the recursive
D( z −1 ) estimates of the nominally constant parameters tend to vary, with large uncer-
Noise TF Model: ξ (k ) = e(k ); e(k ) = N (0, σ 2 ) (33.6b)
C( z −1 ) tainty bounds and no clear convergence (see, e.g., Young, 2010c or 2011).
Output Observation : y (k ) = x (k ) + ξ (k ) (33.6c) 33.3.1 A Solute Transport and Dispersion Example

The DBM approach in this context has been described in a number of papers
where δi are the time delays measured in sampling intervals and the main TF (see, e.g., Beer and Young, 1983; Wallis et al., 1989; Blazkova et al., 2012),
polynomials are defined as follows: where the resulting model has been termed the aggregated dead zone (ADZ)
1
Nonuniform sampling is possible in the case of continuous-time models (see, e.g., model for solute concentration changes in a reach of a river or channel sys-
Garnier and Young, 2014), but this is not considered here. tem. In the case of solute transport through soils (Beven and Young, 1988),
33_Singh_ch33_p33.1-33.12.indd 4 8/22/16 1:57 PM

Linear DBM Models 33-5
this name has been modified to aggregated mixing zone (AMZ), and for more 5
general applications (Young and Lees, 1993), the active mixing volume 95% confidence bounds
(AMV). The ADZ model is compared with the distributed parameter advec- 4.5
Model output
tion dispersion equation (ADE) model in Young and Wallis (1993). The term
4 Tracer data
“dead zone” is somewhat misleading because these are the zones where most
mixing occurs, but it derives from earlier literature on the subject (see 3.5
Concentration (mg/L)
Valentine and Wood, 1977).
Research on solute transport and dispersion is aided by the ability to con- 3
duct simple planned experiments using conservative tracer materials, such as
2.5
the fluorescent red dye, Rhodamine WT (see, e.g., Wallis et al., 1989), or
potassium bromide (KBr), as used in the tracer experiments of Martinez and 2
Wise (2003), who kindly provided me with the resultant data set utilized
below. These consist of the tracer concentration changes monitored at various 1.5
locations in a wetland area of Florida, following an impulsive or gulp injection
1
u(k) of the tracer at the upstream, input location. This study, which was
designed to investigate the effects of domestic wastewater passing through the 0.5
system, used a sampling interval of ∆T = 2 hours. With such an impulse input
of tracer material, one might expect there to be some initial mixing behavior 0
before the concentration changes y(k) become relatively smooth at any of the
–0.5
downstream output sites. This needs to be noted when interpreting the nature 0 100 200 300 400 500 600 700 800 900
of the estimated DBM model (see the following text).
Time (hours)
The standard DBM analysis of such tracer data is to use them to identify
and estimate a TF model, in either continuous or discrete-time form, based on Figure 33.1 The output of the DBM aggregated dead zone model compared with the
the measurements u(k) and y(k). Discrete-time DBM modeling is described tracer data (note: impulsive input of 186.3 applied after 82 hours).
in Young (2001b), where a [4 2 22 0 0] model is identified for the TF between
u(k) and y(k), with the latter measured at one of the output sites and the differential equation model representing the changes in the mass of solute M
model estimated in a constrained form to ensure four physically interpretable (t), based on simple conservation considerations:
real eigenvalues in the estimated TF model. However, it is clear that a contin-
uous-time DBM model makes more sense since most solute transport and
dispersion models are formulated in continuous-time, differential, or partial dM (t ) d {Vc2 (t )}
= = Q(t − τ )c1 (t ) − Q(t )c2 + R(t ) (33.12)
differential equation (PDE) terms that relate directly to mass conservation dt dt
(see later). Such a continuous-time analysis is described in Young (2011), but
here we will re-consider the data is a slightly different manner using hybrid where pure time delay τ is introduced to account for advection; R(t) accounts
continuous-time modeling with the rivcbjid/rivcbj routines in CAPTAIN. for any solute gains or losses within the reach; c1(t) and c2(t) are the concen-
As in the previous analysis of these data, the roots (poles) of this estimated trations of the solute at the input and output of the reach, respectively, and
continuous-time TF denominator polynomial are complex so, using the same V is the AMV, i.e., the volume of water in the reach that is estimated to be
arguments as those used in the above references, it is constrained to have real effective in the transport and dispersion of the solute. In general, because of
poles (see Special TF Optimization in CAPTAIN in the “Technical Matters” the inefficient mixing in the reach as a whole, V < Va, where Va is the actual
section of http://captaintoolbox.co.uk/Captain_Toolbox.html/Captain_ volume of water in the reach. Note that, for simplicity, the concentration
Toolbox.html). The basic TF model structure is identified well, with a fourth within the AMV is assumed to be the same as that measured at the output of
order denominator and first-order numerator, as in the previous analysis. the reach c2(t). Thus, if we let x(t) = c2(t), u(t) = c1(t), and R(t) = 0 (no inflow
However, there is some uncertainty in the pure time delay, which was identi- within the reach), the equation can be rearranged to the form
fied in Young (2001b) and Young (2011) as 22 and 20 sampling intervals,
dx (t ) Q Q
respectively. In order to see how selecting a different time delay affects the = u(t − τ ) − x (t ) (33.13)
continuous-time modeling results, therefore, let us consider a [4 2 21 20 0] dt V V
structure, in which the additive colored noise process is identified as an
AR(20) process: i.e., or, in TF terms,
b0
x (t ) = u(t − τ ) (33.14)
b0 s + b1 b0 s + b1 s + a1
x (t ) = u(t − 42) = 4 u(t − 42);
(s + α 1 )3 (s + α 2 ) s + a1s 3 + a2 s 2 + a3s + a4
y(k ) = x (k ) + ξ (k ) where a1, = b0 = Q/V, so that the steady state gain G = b0/a1 = 1.0 (fully conser-
vative with no solute loss over the reach) and the time constant, here the resi-
1 dence time, is T = V/Q time units, while the travel time is defined as τ + T. If
ξ (k ) = e(k ) = N (0,σ 2 )
e( k ) (33.10)
1 + c1z −1 +  + c20 z −20 R(t) ≠ 0, then G < 1.0 or > 1.0, representing loss or addition of solute, respec-
tively, over the reach. This model is a single-input, first-order example of the
general continuous-time model (33.3a). From a DBM perspective, however, it
where the reader can evaluate the values of a1 to a4 in terms of α1 and α2. Note is important to note that, as in this example, a single ADZ model may not be
here that the AR(20) noise model indicates considerable serial correlation in able to explain all of the concentration changes in the reach so a higher-order
ξ(k), although with very low variance (σ2 = 0.007, see below), and that the model, consisting of several first-order ADZ elements, such as Eq. (33.14), in
time delay of 21 samples becomes a temporal time delay of 42 hours because series or parallel, may be identified statistically from the tracer data, depend-
of the ∆t = 2-hour sampling interval. The parameter estimates are as follows, ing on the length and nature of the reach, as well as the type of input u(t). In
with the adjusted (see as follows) standard errors shown in parentheses: this more general case, the total travel time is the sum of the pure time delay
τ and the time constants of the ADZ elements. This is, of course, the type of
α̂ 1 = 0.0801(0.0013); α̂ 2 = 0.0101(0.00037); σ̂2 = 0.007 higher-order model identified in Eq. (33.10) for the Martinez-Wise data.
Returning to this identified model, the autocorrelation function of the
b̂0 = 2.040 × 10−4 (1.0213 × 10−5 ); b̂1 = 5.0929 × 10−6 (1.998 × 10−7) estimated residuals e(k) shows that they are reasonably white, with low
(33.11) variance, although they are somewhat heteroscedastic (i.e., their variance
changes over time, a quite common characteristic of hydrological systems). In
The tracer data are explained well by this model, as shown in Fig. 33.1, with order to allow for this changing variance, it has been estimated using time
R2T = 0.997% (i.e., 99.7% of the output variance explained by the model). variable parameter estimation; see Sec. 33.4 and Young and Pedregal (1996).
From the DBM modeling standpoint, it is now necessary to interpret the This is used to adjust the parametric error covariance matrices that define the
model (33.10) in physically meaningful terms. Assuming that there is a standard errors on the parameters, as well as the MCS analysis used to infer
constant flow rate Q m3/s out of a reach in a river system, the simplest contin- the uncertainties in the estimates of the derived time constants (T1 = 1/α1 and
uous-time ADZ model is the following lumped parameter, first-order T2 = 1/α2 ), the steady-state gain, G = b1 /a4, and the 95% confidence bounds
33_Singh_ch33_p33.1-33.12.indd 5 8/22/16 1:57 PM

on the model response in Fig. 33.1. This MCS analysis, based on 10,000 ran- parameter estimation (Young, 2011) within a data assimilation context.
dom realizations, yields the sample distributions for the three parameters The RIV class of estimation algorithms have the advantage that they are based
shown in Fig. 33.2 and the associated estimates of the parameters are as on “pseudo-linear regression” (Solo, 1978) and can, therefore, exploit the
follows: optimal IV version of the best known and simplest of recursive algorithms,
the recursive linear least squares (RLS) algorithm, thus allowing for the esti-
T̂ 1 = 12.49(0.196); T̂ 2 = 99.34(3.65); Ĝ = 0.987(0.084) (33.15) mation of time-variable parameters. Remarkably, RLS was developed origi-
nally by Gauss sometime before 1826 (see App. A of Young, 2011) and can be
interpreted as a simple example of the KF (see, e.g., Young, 2010b). There is
which compare well with the estimates obtained in the previous analysis
insufficient space available in this chapter to discuss recursive estimation in
(Young, 2011). Note how, in statistical terms, Ĝ = 0.987 ± 0.084 is insignifi-
detail, but the above references provide a full description of its development
cantly different from unity, so that no tracer appears to have been lost
and application.
between input and output, and the varying width of the 95% confidence
If the time variations in the parameters of an estimated TVP model are
bounds in Fig. 33.1 is a reflection of the changing variance in the residuals
found to be related to the variations of measured variables from the system
e(k). The sharp widening around 115 to 125 hours, for example, indicates the
under study, they can be assumed to be state dependent. Such state-depen-
uncertainty in the estimation of the advective time delay arising from the
dent parameter (SDP) models (Priestley, 1988; Young, 1993, 2000, 2001c;
lumped approximation of this delay, when the physical process is naturally
Ratto et al., 2007) have figured prominently in the development of nonlinear
more distributed, or possibly some variation in the time delay over the test
DBM models for rainfall-flow models from early contributions in Young
period. However, this limitation of the ADZ model has only a minor effect on
(1993) and Young and Beven (1994) to recent publications, such as
the overall dynamics of the estimated model.
Romanowicz et al. (2006), McIntyre et al. (2011), Young (2013), and Leedal
Finally, how can we interpret the higher-order nature of the model (33.10)?
et al. (2013).
Such an interpretation is likely to be somewhat subjective, but I feel that part
The complete SDP transfer function model is reviewed in Young (2013),
of the model is accounting for the initial nonuniform mixing following the
where it is then applied to the well-known data from the humid Leaf River
initial impulsive input of the tracer. This illustrates the well-known
basin, as outlined in Sec. 33.5. Here, for simplicity and to illustrate how the
experimental conclusion that it is best to base solute transport and dispersion
methodology is widely applicable, let us consider the analysis of daily data
modeling on an input measured someway down the system from the injection
from another, quite different semi-arid catchment, analysis that any reader
point, where mixing has stabilized. A DBM analysis of this kind, again using
who has access to Matlab and the CAPTAIN Toolbox can run through
the Martinez-Wise data, is described in Young (2011), where a model based
using the command-line demonstration example provided in the CAPTAIN
on inputs downstream of the injection site is identified, in which all of the TF
package. This rfdemo example is illuminating since it shows how both TVP
elements are now first order.
and SDP modeling can be exploited in this rainfall-flow context. It concerns
the DBM modeling of daily rainfall-flow data from the Canning, an ephem-
33.4 TIME VARIABLE AND STATE-DEPENDENT eral river in Western Australia (for further details, see Young et al., 1997;
PARAMETER MODELS Young, 2008). The discrete-time model is identified and estimated on the
basis of the data from 1985.2 to 1987.1, and it is then validated on two sets
Although many processes in hydrology are nonlinear, the range of application of data from 1977 to 1978.5 and 1979.4 to 1980. For those readers who are
of linear models in hydrology can be extended by considering time variable not able to use the demonstration example, however, let us briefly review
parameter (TVP) models, whose parameters and structure may change over the results obtained.
time. This leads to the idea of adaptive linear models whose dynamics can First, the analysis shows that although a linear model is unable to explain
change in response to the updating of the model parameters using recursive the data in a satisfactory manner, a first-order TVP model estimated using
Quick path time constant, T1 Slow path time constant, T2 Steady-state gain, G
1 1 1
0.8 0.8 0.8

Normalized frequency
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
11.5 12 12.5 13 13.5 90 100 110 120 0.6 0.8 1 1.2 1.4
Residence time (hours) Residence time (hours) Conservativity measure
Figure 33.2 MCS estimated sample distributions of the three derived parameters T1 , T2, and G.
33_Singh_ch33_p33.1-33.12.indd 6 8/22/16 1:57 PM

HypotheticO–Inductive DBM Modeling 33-7
the dtfm routine in CAPTAIN reveals that the main identified parameter 33.5 HYPOTHETICO–INDUCTIVE DBM MODELING
variation is in the numerator b0 parameter, with little estimated change in the
denominator a1 parameter. The resulting TVP model provides a large Hypothetic–inductive DBM (HI-DBM) modeling cannot be generalized
improvement in modeling ability, with R2T = 0.963. Moreover, the TVP esti- because it depends on the particular application and the available hypotheti-
mates of the numerator parameter correlate very well with the measured cal elements that can be evaluated in the context of the identified and esti-
flow. It is not surprising, therefore, that if flow is considered as the variable mated DBM model. However, a recent, comprehensive example (Young,
on which the numerator parameter is dependent, non-parametric SDP esti- 2013) illustrates well the nature of the HI-DBM modeling approach. Here, the
mation, using the sdp routine in CAPTAIN, also produces a quite good HI-DBM analysis is applied to daily mean areal rainfall (mm/d), potential
first-order SDP model, with R2T = 0.910. Final identification of the model evapo transpiration (PET) in mm/d, and streamflow (m3/d) data from the
structure using the rivbj routine suggests a [2 3 0 0 0] model and the resulting well-known, humid, Leaf River basin (1944 km2 ) located north of Collins,
estimated model has an R2T = 0.958. This model validates well on the periods Mississippi, over the 40 water-years from October 1948 to September 1988.
noted above, with R2T = 0.954 and 0.928, respectively. The associated HyMOD conceptual model (e.g., Vrugt et al., 2009, and the
The use of flow as the SDP variable may seem strange at first sight, but, prior references therein) provides the hypothetical element in the form of the
not surprisingly, it appears to be acting as a good surrogate measure of the probability distributed model (PDM) hypothesis; namely, a procedure for soil
soil moisture. Moreover, the estimated TF model can be decomposed by the moisture accounting and determining the value of runoff production accord-
residuez routine in Matlab to a parallel pathway model, where the instanta- ing to a probability-distributed storage capacity model that is driven by the
neous effect accounts for a small 5.84% of the flow; the quick pathway, prob- rainfall, r(k), and PET, Ev(k), inputs (see Moore, 2007).
ably representing the surface flow contributions to the flow changes, has a As in the case of the most rainfall-flow models, such as the Canning River
time constant of 2.57 days (contributing 52.53% of the flow), while the slow model considered in the previous section, HyMOD is of a Hammerstein form,
pathway, probably accounting mainly for the subsurface and groundwater with the PDM element converting the measured rainfall into an effective
effects (and contributing 41.63% of the flow), has a time constant of 18.7 rainfall measure. However, the model is unusual because this effective rainfall
days. These provide a reasonable physical interpretation of the model in provides the input to a fourth-order linear dynamic system, rather than the
hydrological terms and, therefore, satisfy the DBM modeling requirements more normal second-order system (as in the Canning River example). This
in this regard. fourth-order system consists of a third-order Nash Cascade (Nash, 1958), i.e.,
Note that, in the demonstration example, the simplified refined instrumen- a serial connection of three identical, quick-flow storage tanks, each modeled
tal variable (SRIV) is used for final model estimation because the noise ξ(k) as a first-order discrete-time system, in parallel with a single slow-flow tank,
is heteroscedastic and its modeling by an ARMA process is questionable, so again modeled as a first-order discrete-time system. There are two versions of
p and q are set to zero. However, the SRIV algorithm produces consistent the HyMOD model (see Young, 2013) and Fig. 33.3a compares the optimized
estimates with low variance, so the results are reliable. For completeness, responses of these models over a typical section of the Leaf River data.
however, full RIV identification suggests a [2 3 0 6 0] model with an AR(6) Version 2 explains the data best, with R2T = 0.814.
noise model. This model has an R2T = 0.953 and the residual variance esti- This atypical fourth-order structure of the HyMod model arises from the
mate is σ2 = 0.0031, i.e., standard deviation of 0.056 m3/s. The main reason somewhat unusual response characteristics of the Leaf River and is reflected
for estimating such a discrete-time, fully stochastic model is that it makes also in the initial inductive DBM modeling analysis, which identifies a
flow forecasting based on the state space, KF form of the model, with the Hammerstein model form, with a power-law nonlinearity, similar to that
state variables defined by the outputs of the quick and slow flow pathways, identified for the Canning River in Sec. 33.4, in series with a number of pos-
quite straightforward (see, e.g., Young, 2002; Romanowicz et al., 2006; sible linear model structures which explain the flow data better than HyMOD.
Young, 2011; and the prior references cited therein). In order to remove the One of these, like HyMod, is a fourth-order [4 4 0 1 5] model, with its eigen-
need for forecasting rainfall, which is always a daunting task, this requires values constrained to be real. However, the best identified linear model ele-
that there is a pure time delay between rainfall and flow and, in this example, ment is a simple, unconstrained, first-order [1 7 0 1 5] TF model, with an
the model does not have such a delay. Fortunately, however, a false pure time R2T = 0.871. This is an interesting result because it shows that the fast response
delay of 1 day can be introduced into the model, yielding an identified is complete in 7 days and so can be modeled very well by the seven coeffi-
[2 3 1 8 0] structure that leads to a reasonable 1-day-ahead forecasting error cients in the TF numerator polynomial, with the very low-amplitude slow
standard deviation of 0.060 m3/s. response modeled by the first-order TF denominator polynomial, with an
30
HYMOD 95% uncertainty bounds 95% uncertainty bounds
HYMOD model version 1 30 Version 1 HI-DBM model
Measured flow Measured flow
25 HYMOD model version 2 Version 1 HI-DBM ex-post predictions
HYMOD model version 2 (Vrugt)
25
20
20
Flow (mm)
Flow (mm)
15
15
10
10
5 5
0 0
1949.3 1949.35 1949.4 1949.45 1949.3 1949.35 1949.4 1949.45
Date Date
Figure 33.3 The simulated outputs of selected optimized HyMOD and DBM models compared with the flow data y(k) over a typical short segment of the Leaf River data: (a) left,
HYMOD; (b) right, HI-DBM.
33_Singh_ch33_p33.1-33.12.indd 7 8/22/16 1:57 PM

associated time constant of 31.9 days, within a standard error range of 33.6 DBM EMULATION MODELING OF HIGH-ORDER
between 28.4 and 36.2 days. Moreover, the response does not contain the SIMULATION MODELS
slower, second-order exponential recession that normally characterizes well A dynamic emulation model (DEM) is a special model form designed to over-
the response of most rainfall-flow hydrographs and is usually modeled by a come some of the identifiability issues associated with high-order and
second-order process, with quick and slow dynamic modes, as in the Canning complex hydrological simulation models and produce a low-order model that
River example and most other previously identified DBM models of rainfall- emulates the behavior of the simulation model as closely as possible over a
flow behavior. specified range of its parameter values (see, e.g., Castelletti et al., 2012). One
It is interesting to note that the unit hydrograph (unit impulse response) of approach to DEM synthesis that exploits DBM modeling methodology and
the [1 7 0 1 5] model is redolent of higher-order, serial connections of first- DMA (see stage 5 in Sec. 33.2) for model order reduction is described in
order dynamic processes and is probably the reason why the Nash Cascade of Young and Ratto (2009) and shown diagrammatically in Fig. 33.5. This uses
three quick-flow stores had to be used in the HyMOD model, rather than the non-parametric or parametric nonlinear function estimation to identify and
more usual, single store. In fact, if the number of quick-flow tanks in the map the multivariate relationship between the high-order model parameters
model is increased to 10 (with no increase in the number of estimated param- and those of the low-order model. The latter could involve a suitable known
eters since the tanks are identical), then the estimated model explains the flow nonlinear function, as in the later practical example; a function identified by
data quite well, with R2T = 0.860 compared with the standard HyMOD model’s SDP-type estimation, as in Young and Ratto (2009); or, in more complex
R2T = 0.814. Moreover all of the parameter estimates seem well defined, with- cases, smoothing spline ANOVA models (see, e.g., Gu, 2002). Normally, the
out any signs of poor identifiability, suggesting that the previously utilized emulation model obtained in this manner is in the form of either a constant
HyMOD model structure for the Leaf River data might be best modified to parameter linear transfer function relationship; or a state-dependent param-
this alternative, higher order form. This illustrates how the initial, simple, and eter, nonlinear transfer function model (SISO, MISO, or MIMO); and it is
computationally efficient DBM modeling can be useful in a diagnostic role, estimated from discrete-time, sampled data in either continuous or discrete-
quickly revealing unusual dynamic model structures, such as the [1 7 0 1 5] time, using the CAPTAIN tools or other equivalent estimation procedures.
DBM model and the 10-tank HyMOD model, models that are capable of Examples of DBM emulation modeling applied to hydrological models are
efficiently characterizing the Leaf River dynamics and yet might be unlikely reported in Young and Ratto (2009), Young (2010a), and Young and Ratto
products of conceptual model formulation. (2011). Another typical example (Young et al., 2009) is based on an HEC-RAS
The HI-DBM model is obtained by replacing the effective rainfall nonlin- model representation of the River Severn formed from 89 cross-sectional
earity in the DBM model by the conceptual PDM effective rainfall mecha- nodes between the Montford and Buildwas locations, which includes the
nism in the HyMOD model. As such, this change not only introduces a town of Shrewsbury. HEC-RAS solves the dynamic Saint-Venant partial dif-
conceptual nonlinear mechanism that has a clearer physical interpretation ferential equation (PDE) model using an implicit, Preissman-type finite dif-
than the SDP powerlaw nonlinearity, but it also converts the DBM model ference method. The main computational engine is based on the UNET
from one that is limited to forecasting applications (see Young, 2013) into a computational scheme originally developed by Barkau et al. (1989) and this is
stochastic simulation model that has a much wider application potential. run in an unsteady flow mode and forced with an upstream boundary condi-
Based on these considerations and the initial DBM modeling results, the tion defined by 15-minute observations of flow at Montford between
general HI-DBM model has the structure shown in Fig. 33.4. Although, in the December and March 2002.
DBM modeling analysis, the constrained [4 4 0] TF model does not yield as The DEM analysis was carried out by David Leedal at Lancaster and the six
good estimation and validation results as the unconstrained [1 7 0] model, it nodes through Shrewsbury were selected as the sites for the DEM output,
is, perhaps, more appealing to hydrologists because it has a structure that is using data from the HEC-RAS simulated water surface level for all the cross-
similar to that used in the HyMOD model. section nodes. These nodes (references cs: 82, 79, 77, 73, 70, and 68) represent
The main difference between the structure of the TF model in this figure sites that may be useful for a subsequent real-time flood forecasting system.
and the linear part of the HyMOD model is that the Nash-Cascade of three, The purpose of the research is then to develop initial, low-order, DBM sub-
identical, first-order stores in HyMOD is modified by the addition of a sim- models to emulate the relationship between the upstream boundary condition
ple, three-parameter, moving average prefilter PF(z−1) = 0.073 + 0.499z−1 + as an input time series and the water surface level at each of the chosen HEC-
0.344z−2 in the quick-flow routing equation. Since this prefilter polynomial RAS nodes as output series. The HEC-RAS simulated data with a nominal set
operates only on the PDM-generated effective rainfall input ue (k), it is simply of parameter values is shown in Fig. 33.6.
modifying ue (k) to better explain the flow data. And the fact that the model In this case, given the finite difference nature of the simulation model, it
then explains these data significantly better than the standard HyMOD model makes sense to consider discrete-time models. Consequently, the initial DBM
suggests that this is one other area where the HyMOD model hypothesis analysis of the data in Fig. 33.6 was carried out using the CAPTAIN rivbjid
might be reinvestigated. routine, which identifies the following first order, nonlinear SDP model:
Stochastic input ARMA(1,5) noise model

e(k) x(k)
D(z –1)
C(z –1)
Nash-Cascade of 3
PET α = 0.876 Prefilter quick-flow tanks xq(k)
Ev(k) PF(z –1) each time constant
Tq = 0.98 days
Rain 87.6% Flow
r(k) PDM rainfall y(k)
Effective rain
nonlinearity
{ } +
{
F r(k),Ev(k), qs } ue(k) = F r (k), Ev(k), qs
12.4%
Slow-flow tank
xs(k)
Time constant Ts = 25.4 days
1– a = 0.124
Figure 33.4 Block diagram of the HI-DBM model for the Leaf River (from Young, 2013).
33_Singh_ch33_p33.1-33.12.indd 8 8/22/16 1:57 PM

DBM Emulation Modeling of High-Order Simulation Models 33-9
High order Reduced order

model: model:
(e.g., 17th order and (e.g., 3rd order and
circa 180 parameters) 27 parameters)
Parameterized .
High dynamic . state-dependent . Reduced
order model . parameter . dynamic
estimated . regression . order model
parameters . (SDR) . parameters
. relationships
.
Figure 33.5 The process of dynamic emulation model synthesis: note that the order and number of parameters
referred to for the high and low order models relate to a typical example (from Young and Ratto, 2009).
b j ,0
y j (k ) = f j ( y j (k − 1)) u (k − δ ) + ξ j (k ); j = 1  6 (33.16) models of Secs. 33.4 and 33.5. The nonlinear SDP functions in each case were
1 + a j ,1z −1 parameterized and estimated using the adaptive neuro-fuzzy inference system
where the slightly modified nomenclature aj, 1 and bj, 0 is used to differentiate (ANFIS) algorithm from the Matlab Fuzzy Systems Toolbox package (see,
this single-input, multiple output (SIMO) model from the MISO models used e.g., Vernieuwe et al., 2005) and are plotted in Fig. 33.7a as a function of
heretofore. The input transformation function fj (yj (k − 1)) is the SDP for y(k–1) (denoted in the figure by yk–1). Table 33.1 displays the SRIV estimated
cross-section j, with the parameter dependent on the measured yj (k − 1), TF parameter values âj, 1, b̂j, 0, and δ for each cross-section. It also shows the
which acts as a time-variable gain on the input data, as in the rainfall-flow time constants derived from the aj,l parameters, which demonstrate that the
5.5
4.5
3.5
3
Level (m)
2.5
1.5
0.5
0
0 200 400 600 800 1000 1200
Time (hours)
input cs 82 cs 79 cs 77 cs 73 cs 70 cs 68
Figure 33.6 The data used for nominal emulation model identification and estimation. The bold line is
the upstream boundary level for the HEC-RAS simulation, the thin lines are levels at six cross-sections rep-
resenting the River Severn through Shrewsbury (this and subsequent Fig. 33.7 are from Young et al., 2009).
33_Singh_ch33_p33.1-33.12.indd 9 8/22/16 1:57 PM

0.7
–0.8
cs 82 cs 79 cs 77
0.6 0.5
0.6
State−dependent parameter, Cj,k
0.5 –0.85
TFa parameter
0.4
0.5 0.4
0.3 –0.9
0.4 0.3
1
0 2 0 2 0 1 2
cs 73 1 cs 70 1 cs 68 –0.95
0.95
0.95
0.9 0.8
0.9 –1
0.85
0.6 0.5
0.8 0.85 1.5
1 1
0.75 0.8 0.4
0 yk–1 4 0 yk–1 5 0 yk–1 5 Inbank mannings multiplier 1.4 0.6 Outbank mannings multiplier
SDP ANFIS fit ANFIS fit conf. int.
(a) (b) Parametric fit Data
Figure 33.7 The estimated nonlinear functions of the HEC-RAS emulation model: (a) SDP input nonlinearities; (b) the emulation mapping surface for the â70,1
DBM emulation model parameter.
identified first-order TF model structure provides a reasonable mechanistic mapped estimate from Eq. (33.17). Figure 33.7b compares the resulting esti-
interpretation, with the time constant increasing with distance from the mated surface defined by Eq. (33.17) for â70,1 with the estimates of this param-
input, ranging from 0.65 hours at cross-section 82, to 6.57 hours at cross- eter obtained from the MCS realizations.
section 68. Despite its very simple lumped parameter approximation to the distributed
Also, as required, the emulation model explains the data well, with R2T > 0.99 parameter HEC-RAS model, the resulting DEM is quite successful, with a
for all cross-sections, so that the noise terms ξj (k), which represent the model- coefficient of determination in a validation run, based on a new Monte Carlo
ing error since there is no measurement noise on the data, have a very low set generated using randomly selected values of the HEC-RAS model param-
variance, accounting for less than 1% of the simulated output variance. More eters, of R2T = 0.989 even though, for this illustrative exercise, no attempt was
importantly, the explanation is quite similar in the case of the validation exer- made to include the parameters of the input nonlinearity into the DEM
cises, where a completely new input series is used to drive the HEC-RAS model. analysis. In more general terms, the DBM emulation of PDE models such as
The results of the preceding analysis show that the nominal model can this has also produced promising results. For example, another more recent
accurately reproduce the level-to-level relationship of the HEC-RAS model study, similar to the previous (Tych and Young, 2012), has considered the
for the chosen cross-sections when the HEC-RAS model parameters are set at emulation of the popular OTIS model, which simulates the distributed
their nominal values. This good performance can be attributed to the rich parameter transient storage (TS) model for solute transport and dispersion
range of dynamics that can be reproduced by the surprisingly simple used by Martinez and Wise in their wetland area study (see Sec. 33.3.1).
nonlinear DBM model form. However, it is now necessary to map the rela- Figure 33.8 from the preceding reference shows a typical validation run for
tionship between the DBM model parameters and whichever HEC-RAS this DEM. Also, I have developed an emulation model for a semidistributed
model parameters are considered most important in the proposed applica- ADZ (SDADZ) model (see Sec. 12.5 of Chap. 12 in Young, 2011), which is a
tion. In order to identify and estimate these mapping characteristics, a Monte simpler alternative to the OTIS model, and compared the results with the
Carlo ensemble of 1000 HEC-RAS simulation realizations was generated by discrete-time DBM model of the Martinez and Wise data.
introducing variation into the Manning’s coefficients. In particular, the full set Finally, it is important to recall, from stages 5 through 7 of the DBM mod-
of 89 nominal coefficients were multiplied by a factor sampled from a uni- eling strategy in Sec. 33.2, that the DEM model is not in any way an alterna-
form probability distribution between 0.5 and 1.5. These scaling factors were tive to DBM modeling; it simply exploits DBM modeling methodology to
sampled independently (i.e., with zero covariance) and, in each of the 1000 evaluate how an existing large computer simulation model can be represented
realizations, the simulated levels generated by the HEC-RAS model were col- by a much lower-dimensional DBM-type of model in order to consider how
lected at each of the selected cross-sections. this can be reconciled with a DBM model identified and estimated using real
For simplicity, the input SDP nonlinearities obtained in the nominal emula- data from the same system.
tion model estimation were maintained for the full DEM identification, so that
the only DEM parameters that had to be reestimated, at each realization, were
the two, linear TF model parameters. The ensemble of these parameter estimates
was then used in the mapping analysis. For example, in the case of the estimated
âj,1 parameter, the resulting relationship with the two HEC-RAS parameters was
1 0.9
identified in the following nonlinear form for all of the submodels:
Tracer concentration
âj,1 = exp(− (pj,1 α p1j , 2 + pj,3 α p j , 4 )) + pj,5 (33.17)

2
0.8
p ,2 p , 4
where α 1j and α 2j are the HEC-RAS parameters (in-bank and out-of-
bank scaling factors, respectively) and pj,i, where j indicates the cross-section
identities from Fig. 33.7a, are the five mapping parameters, i = 1, 2,..., 5 opti- 0.5
mized to minimize the sum of squares of the errors between âj,1 and the 10 11
Table 33.1 SRIV Estimation Results from rivbj

Location 82 79 77 73 70 68
âj,1 –0.213 –0.330 –0.436 –0.681 –0.764 –0.859 0 5 10 15 20 25
^ Time in days
bj,O 0.7708 0.673 0.557 0.311 0.224 0.141
Figure 33.8 A typical validation result showing the OTIS and emulation model
δ 6 6 6 6 6 6 outputs (concentration in mg/l) for a randomly selected point in the full model
parameter space. The shaded area shows the 95% uncertainty band and the full
T (h) 0.65 0.90 1.21 2.60 3.72 6.57
model simulation line is practically indistinguishable from the emulator output.
33_Singh_ch33_p33.1-33.12.indd 10 8/22/16 1:57 PM

REFERENCE 33-11
33.7 CONCLUSION Martinez, C. J. and W. R. Wise, “Analysis of constructed treatment wetland

hydraulics with the transient storage model OTIS,” Ecological Engineering, 20
This chapter has outlined the data-based mechanistic (DBM) approach to
(3): 211–222, 2003.
modeling stochastic dynamic systems and, in particular, shown how it pro-
McIntyre, N. and M. Marshall, “Identification of rural land management
vides a useful methodological framework for the inductive modeling of
signals in runoff response,” Hydrological Processes, 24 (24): 3521–3534, 2010.
hydrological systems from observational time-series data. DBM modeling is
McIntyre, N., P. C. Young, B. Orellana, M. Marshall, B. Reynolds, and H.
a general method theory consisting of several analytical stages of model struc-
Wheater, “Identification of nonlinearity in rainfall-flow response using data-
ture identification and associated parameter estimation. These can also
based mechanistic modelling,” Water Resources Research, 47: W03515, 2011,
include the assimilation of hypothetical conceptual elements using a
doi:10.1029/2010WR009851.
hypothetico–inductive DBM (HI-DBM) modeling strategy, or the develop-
Moore, R. J., “The PDM rainfall-runoff model,” Hydrology and Earth
ment of a low-order dynamic emulation model (DEM) for a large, high-order
System Sciences Journal, 11: 483–499, 2007.
simulation model. In this manner, HI-DBM and DEM synthesis can provide
Nash, J. E., “The form of the instantaneous hydrograph,” Proceedings of
a bridge between hypothetico–deductive (and normally large), conceptual
IASH General Assembly of Toronto, IAHS Publ. No. 42, Vol. 3, 1958,
computer simulation models and the parsimonious (parametrically efficient,
pp. 114–118.
low-order) DBM models. Although it is possible to employ any statistically
Nash, J. E. and J. V. Sutcliffe, “River flow forecasting through conceptual
acceptable methods of time-series model identification and estimation for
models: discussion of principles,” Journal of Hydrology, 10: 282–290, 1970.
DBM modeling, the various statistical tools available in the CAPTAIN
Newton, I., Mathematical principles of natural philosophy, Philosophiae
Toolbox2 for Matlab have been designed specifically to service the DBM mod-
Naturalis Principia Mathematica, London, 1687.
eling requirements, thus providing a convenient set of routines for any reader
Ockenden, M. C. and N. A. Chappell, “Identification of the dominant
who has access to the Matlab software environment. Finally, other recent
runoff pathways from data-based mechanistic modelling of nested catch-
applications that have used DBM modeling methods successfully are McIntyre
ments in temperate UK,” Journal of Hydrology, 402 (1–2): 71–79, 2011.
and Marshall (2010), Ockenden and Chappell (2011), Wagener and McIntyre
Oreskes, N., K. Shrader-Frechette, and K. Belitz, “Verification, validation,
(2012), Vaughan and McIntyre (2012), Jones and Chappell (2014), and Smith
and confirmation of numerical models in the earth sciences,” Science, 263:
et al. (2014).
641–646, 1994.
Parkinson, S. and P. C. Young, “Uncertainty and sensitivity in global carbon
cycle modelling,” Climate Research, 9: 157–174, 1998.
REFERENCES Popper, K., The Logic of Scientific Discovery, Hutchinson, London, 1959.
Priestley, M. B., Nonlinear and Nonstationary Time Series Analysis,
Barkau, R., M. Johnson, and M. Jackson, “UNET: a model of unsteady flow Academic Press, London, 1988.
through a full network of open channels,” Proceedings of the 1989 National Ratto, M., A. Pagano, and P. Young, “State dependent parameter metamod-
Conference on Hydraulic Engineering, American Society of Civil Engineers, elling and sensitivity analysis,” Computer Physics Communications,
New York, 1989. 177: 863–876, 2007.
Beer, T. and P. C. Young, “Longitudinal dispersion in natural streams,” Romanowicz, R. J., P. C. Young, and K. J. Beven, “Data assimilation and
American Society of Civil Engineers, Journal of Environmental Engineering, adaptive forecasting of water levels in the River Severn catchment,” Water
109: 1049–1067, 1983. Resources Research, 42: W06407, 2006, doi:10.1029/2005WR004373.
Beven, K. J., “Prophecy, reality and uncertainty in distributed hydrological Smith, P. J., K. J. Beven, D. Leedal, A. H. Weerts, and P. C. Young, “Testing
modelling,” Advances in Water Resources, 16: 41–51, 1993. probabilistic adaptive real-time flood forecasting models,” Journal of Flood
Beven, K. J. and P. C. Young, “An aggregated mixing zone model of solute Risk Management, 7 (3): 265–279, 2014.
transport through porous media,” Journal of Contaminant Hydrology, 3: Solo, V., Time series recursions and stochastic approximation, Ph.D. thesis,
129–143, 1988. Australian National University, Canberra, Australia, 1978.
Blazkova, S. D., K. J. Beven, and P. J. Smith, “Transport and dispersion in Tych, W. and P. Young, “A Matlab software framework for dynamic model
large rivers: application of the aggregated dead zone model,” edited by L. emulation,” Environmental Modelling and Software, 34: 19–29, 2012.
Wang and H. Garnier, System Identification, Environmetric Modelling and Valentine, E. M. and I. R. Wood, “Longitudinal dispersion with dead
Control, Springer-Verlag, London, 2012, pp. 367–382. zones,” Journal of the Hydraulics Division, ASCE, 103 (9): 975–990, 1977.
Castelletti, A., S. Galelli, M. Ratto, R. Soncini-Sessa, and P. Young, “A gen- Vaughan, M. and N. McIntyre, “An assessment of DBM flood forecasting
eral framework for dynamic emulation modelling in environmental prob- models,” Proceedings of the Institution of Civil Engineers-Water Management,
lems,” Environmental Modelling & Software, 34: 5–18, 2012. 165 (2): 105–120, 2012.
Garnier, H. and P. C. Young, “The advantages of directly identifying con- Vernieuwe, H., O. Georgieva, B. de Baets, V. Pauwels, N. Verhoest, and F. P.
tinuous-time transfer function models in practical applications,” International D. Troch, “Comparison of data-driven Takagi-Sugeno models of rainfall-dis-
Journal of Control, 87 (7): 1319–1338, 2014. charge dynamics,” Journal of Hydrology, 302: 173–186, 2005.
Garnier, H. and L. Wang, editors. Identification of Continuous-Time mod- von Bertalanffy, L., General System Theory: Foundations, Development,
els from Sampled data. Springer, London, 2008. Applications, George Braziller (revised edition 1976), New York, 1968.
Gu, C., Smoothing Spline ANOVA Models, Springer-Verlag, Berlin, 2002. Vrugt, J. A., C. J. F. ter Braak, H. V. Gupta, and B. A. Robinson, “Equifinality
Helton, J. C. and D. E. Burmaster, “Guest editorial: treatment of aleatory of formal (DREAM) and informal (GLUE) Bayesian approaches in hydrologic
and epistemic uncertainty in performance assessments for complex systems,” modelling,” Stochastic Environmental Research and Risk Assessment, 23:
Reliability Engineering and Systems Safety, 54: 91–94, 1996. 1011–1026, 2009, doi: 10.1007/s00477–008–0274–y.
Jones, T. D. and N. A. Chappell, “Streamflow and hydrogen ion interrela- Wagener, T. and N. McIntyre, “Hydrological catchment classification using
tionships identified using data-based mechanistic modelling of high frequency a data-based mechanistic strategy system identification,” edited by L. Wang
observations through contiguous storms,” Hydrology Research, 45 (6): and H. Garnier, System Identification, Environmental Modelling, and Control
868–892, 2014. System Design, Springer-Verlag, London, UK, 2012, pp. 483–500.
Konikow, L. and J. Bredehoeft, “Ground water models cannot be validated,” Wallis, S. G., P. C. Young, and K. J. Beven, “Experimental investigation of
Advances in Water Resources, 15: 75–83, 1992. the aggregated dead zone model for longitudinal solute transport in stream
Kuhn, T., The Structure of Scientific Revolutions, University of Chicago, channels,” Proceedings of the Institution of Civil Engineers, Part 2, 87: 1–22,
Chicago, 1962. 1989.
Leedal, D., A. H. Weerts, P. J. Smith, and K. J. Beven, “Application of data- Wellstead, P. E., “An instrumental product moment test for model order
based mechanistic modelling for flood forecasting at multiple locations in the estimation,” Automatica, 14: 89–91, 1978.
Eden catchment in the National Flood Forecasting System (England and Wheater, H. S., A. J. Jakeman, and K. J. Beven, “Progress and directions in
Wales),” Hydrology and Earth System Sciences Journal, 17: 177–185, 2013. rainfall-runoff modelling,” edited by A. J. Jakeman, M. B. Beck, and M. J.
Lees, M. J., “Data-based mechanistic modelling and forecasting of hydro- McAleer, Modelling Change in Environmental Systems, Wiley, Chichester,
logical systems,” Journal of Hydroinformatics, 2: 15–34, 2000. Chap. 5, 1993, pp. 101–132.
Young, P. C., “Recursive approaches to time-series analysis,” Bulletin of
Institute of Mathematics and Its Applications, 10: 209–224, 1974.
2
This can be downloaded free via a link from my website http://captaintoolbox.co.uk/ Young, P. C., “A general theory of modeling for badly defined dynamic
Captain_Toolbox.html, which also provides information on CAPTAIN routines and systems,” edited by G. C. Vansteenkiste, Modeling, identification and Control
DBM modeling; or directly from http://www.lancaster.ac.uk/staff/taylorcj/tdc/down-
load.php
in Environmental Systems, North Holland, Amsterdam, 1978, pp. 103–135.
33_Singh_ch33_p33.1-33.12.indd 11 8/22/16 1:57 PM

Young, P. C., “Time variable and state dependent modelling of nonstation- Young, P. C., Recursive Estimation and Time-Series Analysis: An Introduction
ary and nonlinear time series,” edited by T. Subba Rao, Developments in Time for the Student and Practitioner, Springer-Verlag, Berlin, 2011.
Series Analysis, Chapman and Hall, London, 1993, pp. 374–413. Young, P. C., “Hypothetico-inductive data-based mechanistic modeling of
Young, P. C., “Data-based mechanistic modeling of environmental, ecohydrological systems,” Water Resources Research, 49 (2): 915–935, 2013.
logical, economic and engineering systems,” Environmental Modelling & Young, P. C., “Refined instrumental variable estimation: maximum likelihood
Software, 13: 105–122, 1998. optimization of a unified Box-Jenkins model,” Automatica, 52: 35–46, 2015.
Young, P. C., “Stochastic, dynamic modelling and signal processing: time Young, P. C. and M. B. Beck, “The modelling and control of water quality
variable and state dependent parameter estimation,” edited by W. J. Fitzgerald, in a river system,” Automatica, 10: 455–468, 1974.
A. Walden, R. Smith, and P. C. Young, Nonlinear and Nonstationary Signal Young, P. C. and K. J. Beven, “Data-based mechanistic (DBM) modelling
Processing, Cambridge University Press, Cambridge, 2000, pp. 74–114. and the rainfall-flow nonlinearity,” Environmetrics, 5: 335–363, 1994.
Young, P. C., “Data-based mechanistic modelling and validation of rainfall- Young, P. C., A. J. Jakeman, and R. McMurtrie, “An instrumental variable
flow processes,” edited by M. G. Anderson and P. D. Bates, Model Validation: method for model order identification,” Automatica, 16: 281–296, 1980.
Perspectives in Hydrological Science, John Wiley, Chichester, 2001a, pp. 117–161. Young, P. C., A. J. Jakeman, and D. A. Post, “Recent advances in the data-
Young, P. C., “Data-based mechanistic modelling of environmental systems,” based modelling and analysis of hydrological systems,” Water Science and
Proceedings, IFAC Workshop on Environmental Systems, First Plenary Session Technology, 36: 99–116, 1997.
Keynote Paper, Yokohama, Japan, 2001b. Young, P. C., D. Leedal, K. J. Beven, and C. Szczypta, “Reduced order emu-
Young, P. C., “The identification and estimation of nonlinear stochastic lation of distributed hydraulic models,” Proceedings 15th IFAC Symposium on
systems,” edited by A. I. Mees, Nonlinear Dynamics and Statistics, Birkhauser, System Identification SYSID09, St. Malo, France, 2009.
Boston, 2001c, pp. 127–166. Young, P. C. and M. J. Lees, “The active mixing volume: a new concept in
Young, P. C., “Advances in real-time flood forecasting.” Philosophical modelling environmental systems,” edited by V. Barnett and K. Turkman,
Transactions of the Royal Society: Physical and Engineering Sciences, 360 (9): Statistics for the Environment, John Wiley, Chichester, 1993, pp. 3–43.
1433–1450, 2002. Young, P. C. and S. Parkinson, “Simplicity out of complexity,” edited by M.
Young, P. C., “The data-based mechanistic approach to the modelling, B. Beck, Environmental Foresight and Models: A Manifesto, Elsevier, Oxford,
forecasting and control of environmental systems,” Annual Reviews in 2002, pp. 251–294.
Control, 30: 169–182, 2006. Young, P. C., S. Parkinson, and M. J. Lees, “Simplicity out of complexity:
Young, P. C., “Real-time flow forecasting,” edited by H. Wheater, S. Occam’s razor revisited,” Journal of Applied Statistics, 23: 165–210, 1996.
Sarooshian, and K. D. Sharma, Hydrological modelling in arid and semi-arid Young, P. C. and D. J. Pedregal, “Recursive fixed interval smoothing and the
areas, Cambridge University Press, Cambridge, U.K., 2008, pp. 113–138. evaluation of LIDAR measurements,” Environmetrics, 7: 417–427, 1996.
Young, P. C., “The data-based mechanistic approach to the emulation of Young, P. C. and M. Ratto, “A unified approach to environmental systems
large dynamic hydrological process models,” edited by C. Walshe, Proceedings modelling,” Stochastic Environmental Research and Risk Assessment, 23:
British Hydrological Society 3rd International Symposium, British Hydrological 1037–1057, 2009.
Society, London, 2010a, ISBN 1 903741 17 3. Young, P. C. and M. Ratto, “Statistical emulation of large linear dynamic
Young, P. C., “Gauss, Kalman and advances in recursive parameter estima- models,” Technometrics, 53: 29–43, 2011.
tion,” Journal of Forecasting (special issue celebrating 50 years of the Kalman Young, P. C. and S. G. Wallis, “Solute transport and dispersion in channels,”
Filter), 30: 104–146, 2010b. edited by K. J. Beven and M. J. Kirkby, Channel Network Hydrology, John
Young, P. C., “Real-time updating in flood forecasting and warning,” edited Wiley, Chichester, 1993, pp. 129–173.
by G. J. Pender and H. Faulkner, Flood Risk Science and Management, Wiley-
Blackwell, Oxford, U.K., 2010c, pp. 163–195.
33_Singh_ch33_p33.1-33.12.indd 12 8/22/16 1:57 PM

Chapter 34
Decomposition Methods
BY
SERGIO E. SERRANO
ABSTRACT Rach, 2008; Wazwaz, 2000; Adomian, 1994, 1991). For nonlinear equations
in particular, decomposition is one of the few systematic solution proce-
Adomian’s decomposition method (ADM) constitutes a useful alternative to
dures available. Other analytical methods to solve nonlinear equations have
model linear, and especially nonlinear, equations in surface, subsurface, and
been proposed, such as the homotopy perturbation method (e.g., He, 2006),
contaminant hydrology. ADM offers the advantages of both analytical and
and the variational iterations method (e.g., He, 1999), with few published
numerical procedures, while minimizing many of their disadvantages. It
applications in hydrology as yet.
exhibits the simplicity, stability, and spatial and temporal continuity of ana-
ADM consists in deriving an infinite series that in many cases converge to
lytical solutions, in addition to the ability to handle irregularly shaped
an exact solution. The convergence of decomposition series has been rigor-
domains, and nonlinearity typical of numerical solutions. The most impor-
ously established in the mathematical community (Gabet, 1994, 1993, 1992;
tant feature in ADM is its simplicity; many problems that were considered
Abbaoui and Cherruault, 1994; Cherruault, 1989; Cherruault et al., 1992), and
tractable by complex numerical methods only, are now easily approached
in the hydrologic community (Serrano, 2003c, 1998). In many complex linear
with ADM. This chapter presents a simple introduction to the features of
and nonlinear problems, an exact closed form solution is difficult to obtain.
ADM with some application examples to the modeling of various phenomena
However, in these cases the usual fast convergence rate of decomposition
in hydrology. These include nonlinear kinematic and dynamic flood waves,
series provides the modeler with a sufficiently accurate approximate solution.
infiltration under variable rainfall, nonlinear soil moisture propagation,
A convergent decomposition series made of the first few terms usually pro-
groundwater flow in irregularly shaped aquifers, groundwater flow in hetero-
vides an effective model in practical applications. With the concepts of partial
geneous aquifers subject to pumping and recharge, stream-aquifer interac-
decomposition and of double decomposition (Adomian, 1994, 1991), the
tion, propagation of nonlinear hydraulic transients in aquifers, contaminant
process of obtaining an approximate solution was further simplified. Also,
transport in soils and aquifers subject to nonlinear reactions, and stochastic
recent contributions suggest that the choice of the initial term greatly influ-
analysis in hydrology without small perturbation restrictions.
ences the rate of convergence and the complexity in the calculation of indi-
vidual terms, especially for nonlinear equations (Wazwaz and Gorguiz, 2004;
34.1 INTRODUCTION: ADOMIAN’S DECOMPOSITIONS METHOD
Wazwaz, 2000). Thus, as long as the initial term in a decomposition series,
usually the forcing function or the initial condition, is described in analytic
The fundamental building blocks of hydrologic models are usually made of form, a partial decomposition procedure may offer a simplified approximate
partial differential equations governing fundamental laws in a watershed, a solution to many modeling problems.
soil, or an aquifer. Traditionally, these equations must be solved analytically In hydrology, several fundamental works using decomposition have been
or numerically. Classical analytical solutions (e.g., Fourier series and published on groundwater flow (Moutsopoulos, 2007; Serrano, 1995; Serrano
Laplace transform) are simple to program, are usually stable, and offer a and Unny, 1987); analytical steady and transient groundwater modeling in
spatially and temporally continuous description of hydrologic variables irregularly shaped aquifers (Serrano, 2013, 2012; Patel and Serrano, 2011;
(e.g., Steward et al., 2009; Read, 2007; Read and Volker, 1993; Chapman and Tiaiff and Serrano, 2014); fracture flow (Moutsopoulos, 2009); contaminant
Dressler, 1984; Kirkham, 1957; and Philip, 1957). However, classical ana- transport (Serrano, 1988; Adomian and Serrano, 1998; Serrano and Adomian,
lytical solutions require regular domain shapes (e.g., simple one-dimension- 1996); special problems involving non-Fickian and scale-dependent contami-
al or rectangular geometries), and cannot handle nonlinearities. On the nant transport (Serrano, 1997b, 1996, 1995b); hydraulics of wells in heteroge-
other hand, numerical solutions (e.g., finite differences, finite elements, and neous aquifers (Serrano, 1997a); stream-aquifer interaction (Serrano and
finite volume) reduce the differential equations to a simpler system of alge- Workman, 2008, 1998; Serrano et al., 2007; Srivastava et al., 2006); modeling
braic equations, can manage irregular domain shapes, and may deal with in heterogeneous aquifers (Srivastava and Serrano, 2007; Serrano, 1995a);
numerical approximations of nonlinearities. However, numerical solutions nonlinear moving boundaries in unconfined aquifers (Serrano, 2003a); infil-
yield the state variables at discrete nodes only (i.e., they require a grid), are tration in unsaturated and hysteretic soils (Serrano, 2004, 2003b, 2001, 1998,
difficult to program (i.e., they require specialized computer software), and 1990a); catchment hydrology and nonlinear flood propagation (Serrano,
often have difficulties with instabilities and round-off errors. An alternative 2006; Sarino and Serrano, 1990); nonlinear reactive contaminant transport
method, Adomian’s decomposition method (ADM; Adomian, 1994, 1991, (Serrano, 2003c); and sediment transport in alluvial streams (Tayfur and
1986, 1983), offers the simplicity, stability, and spatial continuity of analyti- Singh, 2011). Several fundamental problems that were considered tractable
cal solutions, in addition to the ability to handle system nonlinearities and with numerical methods only have been easily approached with decomposi-
irregular geometries typical of numerical solutions. Many studies have tion. Besides simplicity, an analytical solution offers a continuous spatio
reported new solutions to a wide class of equations (ordinary, partial, dif- temporal distribution in heads, gradients, velocities and fluxes, thus reducing
ferential, integral, integrodifferential, linear, nonlinear, deterministic, or instability. Combination of analytical decomposition with numerical methods
stochastic) in a variety of fields of mathematical physics, science, and engi- offers an ideal modeling scenario that exhibits the advantages of both ana-
neering (see, e.g., Rach, Wazwaz, and Duan, 2013; Duan and Rach, 2011; lytical and numerical procedures (Serrano, 1992).
34-1
34_Singh_ch34_p34.1-34.6.indd 1 8/22/16 1:58 PM

34-2 Decomposition Methods
34.2 REGIONAL FLOW IN AN UNCONFINED AQUIFER where the integration “constants,” k3 and k4, are found from the y boundary
conditions (34.2)
This section illustrates the basis of ADM with a simple problem. For detailed
introduction, practical examples in hydrology, and computer programs, see  f ( x ) − f 2 ( x ) Rg l y  Rg y 2
Serrano (2011, 2010). Consider the two-dimensional regional groundwater hy 0 = f 2 ( x ) +  3 + y− (34.11)
flow in an unconfined aquifer with Dupuit assumptions. The governing equa-  ly 2T  2T

tion is given by
The y-partial solution satisfies the differential Eq. (34.1) and the y bound-
∂2 h ∂2 h Rg ary conditions in Eq. (34.2), but not necessarily those in the x direction. We
2 + 2 =− , 0 ≤ x ≤ lx , 0 ≤ y ≤ l y (34.1)
∂x ∂y T now have two partial solutions to Eq. (34.1): the x-partial solution (34.7), and

the y-partial solution (34.11). Since both are solutions to h, a combination of
where h(x, y) is the hydraulic head; x and y are the planar coordinates with the two partial solutions yields
respect to an origin; the aquifer dimensions are lx = 860 m and ly = 2000 m,
respectively; the aquifer transmissivity is T = 700 m2/month, and the mean  hx 0 ( x , y ) + hy 0 ( x , y ) 
h0 ( x , y ) =   (34.12)
recharge from rainfall is Rg = 10 mm/month. Equation (34.1) is subject to a  2 

mixed set of boundary conditions given by
where h0 is the first combined term. Higher-order terms are derived similarly.
∂h The ith order term in the x-partial solution, hxi, may be derived from
h(0, y ) = f1 ( y ), (lx , y ) = 0, h( x ,0) = f 2 ( x ), h( x , l y ) = f3 ( x ) (34.2)
∂x Eq. (34.5) as:
where f1( y) = 241 – 0.001y, f2(x) = Cx2 + Ax + f1(0), f3(x) = Ex2 + Bx + f1(ly), hxi = k4i+1 ( y ) + k4i+2 ( y )x − L−x1L y hi−1 (34.13)
C = –Rg /(2T), A = –2Clx, B = –2Elx, and E = –Rg /T. The functions f1, f2, and f3
have been derived from field measurements of the stage of the boundary where hi – 1 is the previous combined term in the decomposition series, and
rivers. Now define the operators Lx = ∂2/∂x2 and Ly = ∂2/∂y2. The inverse k4i + 1 k4i + 2 are such that homogeneous x boundary conditions in Eq. (34.2) are
−1
operators L−1
x and L y are the corresponding twofold indefinite integrals with satisfied. Similarly, the ith-order term in the y-partial solution, hyi, may be
respect to x and y, respectively. Thus, Eq. (34.1) reduces to derived from Eq. (34.9)
Rg hyi = k4i+3 ( x ) + k4i+4 ( x ) y − L−y1Lx hi−1 (34.14)
Lx h + L y h = − (34.3)
T

where hi – 1 is the previous combined term in the decomposition series, and
Two partial decomposition expansions are possible: The x-partial solution k4i + 3 and k4i + 4 are such that homogeneous y boundary conditions in
and the y-partial solution. The x-partial solution, hx, results from operating Eq. (34.2) are satisfied. Similarly to Eq. (34.12), the ith combined term is given
with L−1
x on Eq. (34.3) and rearranging. by
Rg
hx = − L−x1 − L−x1L y hx (34.4)  hxi ( x , y ) + hyi ( x , y ) 
T hi ( x , y ) =   (34.15)
 2 

Expanding hx in the right side as an infinite series hx = hx0 + hx1 + hx2 +… ,
Eq. (34.4) becomes Lastly, approximate the final solution with N terms, h ≈ h0 + h1 + …+ hN ,
where each term in the series is a combination of two partial solutions, one in
Rg x and one in y. The aforementioned procedure may be extended to three-
hx = − L−x1 − L−x1L y (hx 0 + hx1 + hx 2 + ) (34.5) dimensional transient problems. In such case, each combined term in the
T
decomposition series has four components, one in each spatial coordinate,
The choice of hx0 often determines the level of difficulty in calculating and one in the temporal coordinate. Due the high rate of convergence of
subsequent terms and the rate of convergence (Adomian, 1994; Wazwaz, decomposition solutions, the hydrologist often finds that few terms in the
2000). A simple choice is to set hx0 as equal to the first term in the right side above iteration might be reasonably accurate in many practical applications.
of Eq. (34.5). Thus, the first approximation to the solutions is In this example, the first three terms at the center of the aquifer x = lx/2,
y = ly/2 are h0 = 248.52 m, h1 = 5.46 m, and h2 = 2.13 m. It is easy to program
Rg Rg x 2 the aforementioned solution in any standard mathematics software, such as
hx 0 = − L−x1 = k1 ( y ) + k2 ( y )x − (34.6) Maple. Figure 34.1 shows the regional groundwater flow distribution. Patel
T 2T
and Serrano (2011) showed that the maximum relative error of a four-term
where the integration “constants,” k1 and k2, must be found from the x bound- decomposition solution with respect to the exact analytical solution of a
ary conditions in Eq. (34.2). Hence, Eq. (34.6) becomes similar problem is less than 2%. Serrano (2012) also showed similar results
when comparing a transient regional groundwater flow with respect to its
R g lx x Rg x 2 exact analytical solution. Consideration of pumping wells at known coordi-
hx 0 = f1 ( y ) + − (34.7) nates in the aquifer is an easy extension to the aforementioned description.
T 2T
Similarly, aquifer heterogeneity, nonlinearity in the differential equation, and
Equation (34.7) satisfies the governing Eq. (34.1) and the x boundary con- irregular aquifer geometries can be incorporated into the analysis (Tiaif and
ditions in Eq. (34.2), but not necessarily the ones in the y direction. Now, to Serrano, 2014, 2013; Patel and Serrano, 2011). For practical examples with
obtain the y-partial solution to Eq. (34.3), hy, operate with L−1 detailed programs, see Serrano (2010).
y on Eq. (34.3)
and rearrange
Rg 34.3 PROPAGATION OF NONLINEAR KINEMATIC FLOOD WAVES IN RIVERS
hy = − L−y1 − L−y1Lx hy (34.8)
T This section illustrates the application of ADM series, double decomposition,
and successive approximation to the solution of nonlinear hydrologic prob-
Expanding hy in the right side as an infinite series hy = hy0 + hy1 + hy2 +…, lems. Consider the following nonlinear kinematic flood routing equation that
Eq. (34.8) becomes combines the continuity and momentum equations, when the acceleration
and the pressure terms in the latter are neglected:
Rg
hy = − L−y1 − L−y1Lx (hy 0 + hy1 + hy 2 + ) (34.9)
T ∂Q  ∂Q 
+ αβQ β −1  = QL , Q(0, t ) = QI (t ), Q( x ,0) = QI (0) (34.16)
∂x  ∂T 
Taking hy0 as the first term in the right side of Eq. (34.9) one obtains the
first approximation where Q(x, t) is the flow rate (m3/s); QL is the lateral flow into the channel per
2 unit length (m3/s); x is the distance from a streamflow station with a known
Rg Rg y hydrograph; t is time (hour); a is a constant with dimensions (m2 – 3bsb); QI(t)
hy 0 = − L−y1 = k3 ( x ) + k4 ( x ) y − (34.10)
T 2T is the inflow hydrograph at t = 0; and b is a dimensionless constant. Using the
34_Singh_ch34_p34.1-34.6.indd 2 8/22/16 1:58 PM

Propagation of Nonlinear Kinematic Flood Waves in Rivers 34-3
100
ADM
90 Finite elements
80
70
247
60
248
Q (m3/s)
245 50
244
40
243
h (m)
30
242
241 800 20
700
240 600
500 10
400
2000 300 x (m)
1500 200 0 10 20 30 40 50 60
1000 100 Time (h)
500 0
y (m) 0 Figure 34.2 Analytical versus numerical hydrograph at x = 75 km. Finite elements
data kindly provided by Szymkiewicz (2010).
Figure 34.1 Regional groundwater head distribution with contours every 0.5 m,
beginning at 239.5 m. As stated, the convergence rate of ADM series is so high that only a few
terms are needed to assure an accurate solution. Thus, if, for instance, Q02 is a
reasonable approximation, then
Q ≈ Q0 ≈ QI (t − x[F0 ( f0 ) + F1 ( f0 ) + F2 ( f0 )]) + QL x , f0 = Q02 ( x , t ) − QL x

method of characteristics, an implicit solution to Eq. (34.16) was obtained by
Serrano (2006) (34.21)
As an illustration, assume that a = 5/3, b = 3/5, QL = 0, and an inflow

Q( x , t ) = QI (t − F (Q − QL x )x ) + QL x , F (Q ) = αβQ β −1 (34.17)
hydrograph given by
To approximate an explicit solution define the decomposition series  t  (1− tmax

t
) (34.22)
Q = ∑ i∞=0 Qi , and use the concept of double decomposition to expand the initial QI (t ) = q0 + (qmax − q0 ) e
 t max 
term Q0 = ∑ ∞j=0 Q0 j in the right side of Eq. (34.17) (Serrano, 2006; Adomian,
1994). Thus, the first term in Eq. (34.17) is
where q0 = 5 m3/s, qmax = 100 m3/s, and tmax = 4 h. Figure 34.2 shows a com-
parison of a three-term analytical nonlinear hydrograph, Eqs. (34.19)–
  ∞   (34.21), at x = 75 km and the numerical hydrograph calculated by the
Q0 = Ql  t − F  ∑ Q0 j − QL x  x  + QL x (34.18)
  j =0   modified finite element method according to Szymkiewicz (2010). The time
to peak calculated by the two methods is in excellent agreement. The analyti-
cal solution seems to better preserve the peak magnitude, in agreement with
Now expand F = ∑ k∞=1 Fk , where Fk is calculated using one of the many algo- kinematic wave theory. There appears to be some minor differences in the
rithms for the Adomian polynomials (Duan and Rach, 2011; Rach, 2008; recession limb, possibly because of numerical dissipation. However, the cal-
Wazwaz, 2000). Using the traditional Adomian polynomials about an initial culation of the analytical hydrograph is much simpler than the numerical one.
term f0 = Q00 − QL x = QI (t ) − QL x (Adomian, 1994), the first few terms are The analytical solution lends itself easily to extended simulations in rivers.
Figure 34.3 shows observed and predicted nonlinear hydrographs at the
F0 ( f0 ) = F0 ( f0 )
d F ( f0 )
F1 ( f0 ) = F0 ( f0 ) 1200 Observed at Philadelphia
d f0
Nonlinear
d F ( f0 ) F0 ( f0 )2 d 2 F ( f0 ) 1000
F2 ( f0 ) = F1 ( f0 ) + (34.19)
d f0 2! d f02
d F ( f0 ) d 2 F ( f ) F 3 ( f ) d 3 F0 ( f0 ) 800
Q (m3/s)
F3 ( f0 ) = F2 ( f0 ) + F1 ( f0 )F2 ( f0 ) 0 2 0 + 1 0
d f0 d f0 3! d f03
600

400
Combine Eqs. (34.18) and (34.19) to successively approximate Q0:
200
Q00 = QI (t − xF0 ( f0 )) + QL x , f0 = QI (t ) − QL x
Q01 = QI (t − x[F0 ( f0 ) + F1 ( f0 )]) + QL x , f0 = Q00 ( x , t ) − QL x 100 150 200 250 300
(34.20) Time (days)
Q02 = QI (t − x[F0 ( f0 ) + F1 ( f0 ) + F2 ( f0 )]) + QL x , f0 = Q01 ( x , t ) − QL x
 Figure 34.3 Observed and predicted nonlinear flow rates at the Schuylkill River,
September 2004–May 2005.
34_Singh_ch34_p34.1-34.6.indd 3 8/22/16 1:58 PM

Philadelphia station in the Schuylkill River (Southeast Pennsylvania), given 0.50

an inflow hydrograph at the Norristown station (Serrano, 2006). The simula- Philip (1955) Observed
tions included variable lateral flow that included groundwater flow and vari- Parlange (1971) ADM
able effective precipitation. The standard deviation of the absolute error 0.45
between observed and predicted was only 13.21 m3/s.
0.40
34.4 NONLINEAR INFILTRATION IN UNSATURATED SOILS
0.35
This section presents a combination of ADM with a successive approximation
(Serrano, 2004) that yields a simple approximate solution to a highly nonlin-
0.30
ear equation, when traditional numerical solutions present numerous
q
accuracy, complexity, and instability problems. Consider the horizontal infil-
tration equation in a semi-infinite homogeneous soil with a constant bound- 0.25
ary condition maintained on one end:
0.20
∂θ ∂  ∂θ 
− D(θ )  = 0, 0 < x < ∞ , 0 < t , θ (0, t ) = θb , θ (∞ , t ) = θi , θ ( x ,0) = θi
∂t ∂ x  ∂x  0.15
(34.23)
0.10
where q = soil volumetric water content; x = horizontal distance (m); t = time
(h); qb = 0.458 is the water content at the left boundary; qi = 0.086 is the initial 0
α
water content; D(θ ) = c1e λθ is the soil-water diffusivity (m2/h); c1 = 1 m2/h; 0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40
l = 500; and a = 11 (Serrano, 1998). Expanding the nonlinear diffusivity as Distance, x (m)
the series D(θ ) = ∑ i∞=0 Di , Eq. (34.23) becomes
Figure 34.4 Water content versus distance at t = 1 h according to ADM, two numeri-
∂θ ∂  ∞  ∂θ  cal solutions, and laboratory observation.
−  ∑ Di  = 0 (34.24)
∂t ∂ x  i=0  ∂ x 

Similar to Eq. (34.19), the Adomian polynomials about the first term, q0, This process may be continued: from Eq. (34.25), calculate D2, obtain an
are given by (Adomian, 1994) improved diffusivity D(q) ≈ D0 + D1 + D2, then similar to Eq. (34.28)
derive θ 3′ , then similar to Eq. (34.29) derive q3. Many studies report that
D0 = D(θb ) ADM series converges fast and only a few terms are needed. Thus, q ≈ q3 may
be a good approximation to the water content. Figure 34.4 shows profiles of
dD(θ 0 )
D1 = θ1 the water content versus distance profiles at t = 1 h according to four sources:
dθ 0 ADM, the classical numerical solution of Philip (1955), the numerical solu-
dD(θ 0 ) θ 02 d 2D(θ 0 ) tion of Parlange (1971), and laboratory observations (Serrano, 2004). The
D2 = θ 2 + (34.25) ADM solution appears to better predict the position of the wetting front, and
dθ 0 2! dθ 02
the subsequent shape of the tail, than Philip’s (1955) or Parlange’s (1971)
dD(θ 0 ) d 2D(θ 0 ) θ13 d 3D(θ 0 ) solutions. The decomposition solution is simpler; it provides a continuous
D3 = θ 3 + θ1θ 2 +
dθ 0 dθ 02 3! dθ 03 spatio-temporal description, and it does not exhibit the stability and dis-
cretization restrictions of numerical solutions. Extensions of ADM to physi-
 cally based vertical infiltration and distribution were derived by Serrano
(2004). Simple physically based ADM models of infiltration in watersheds
Now recursively calculate each Di, which is used to approximate the next subject to variable rainfall have been proposed (Serrano, 2010, 2004).
qi. Calculations end when the water content distribution reaches a desired Explicit ADM solutions to the Green and Ampt equation were derived by
accuracy. From Eq. (34.25), use the left boundary condition, q ≈ qb, as an Serrano (2003b, 2001).
initial estimate of the water content. With the approximation D(q) ≈ D0 =
D(q b), Eq. (34.24) reduces to the classical heat flow equation with constant
coefficient whose solution is (Zauderer, 1983) 34.5 SUMMARY AND CONCLUSION
The basic features of hydrologic modeling with ADM have been presented
 x  via three simple examples: regional groundwater flow in aquifers, the
θ 0 = θi + (θb − θi )erfc   (34.26)
 4 D0t  propagation of nonlinear kinematic flood waves in rivers, and nonlinear
infiltration in unsaturated soils. ADM offers the advantages of both ana-
lytical and numerical procedures, while minimizing many of their disad-
where erfc( ) denotes the “complementary error function.” With q ≈ q0, an vantages. It exhibits the simplicity, stability, and spatial and temporal
improved estimate of the water content, calculate continuity of analytical solutions, in addition to the ability to handle
irregularly shaped domains, and nonlinearity typical of numerical solu-
 x  tions. ADM may be applied to a wide class of problems in hydrology,
θ1 = θi + (θb − θi )erfc   (34.27)
 4 D(θ 0 )t  including, infiltration under variable rainfall, groundwater flow in irregu-
larly shaped aquifers, groundwater flow in heterogeneous aquifers subject
to pumping and recharge, stream-aquifer interaction, propagation of non-
Now from Eq. (34.25), calculate D1 and obtain an improved diffusivity
linear hydraulic transients in aquifers, contaminant transport in soils and
D(q) ≈ D0 + D1. The improved
aquifers subject to nonlinear reactions, and stochastic analysis in hydrology
without small perturbation restrictions. Expanded references to such work,
 x 
θ 2′ = θi + (θb − θi )erfc   (34.28) including practical computer programs, have been included. The most
 4(D0 + D1 )t  important feature in ADM is its simplicity; many problems that were con-

sidered tractable by complex numerical methods only, are now easily
solution to Eq. (34.24) becomes approached with ADM. The method has proved invaluable to hydrologic
With θ ≈ θ 2′ , an improved estimate of the water content, calculate analysis and preliminary design under scarce data. Future work should be
devoted to the application of ADM to unresolved problems in nonlinear
 x  hydrology. As new algorithms for the Adomian polynomials continue to
θ 2 ≈ θi + (θb − θi )erfc   (34.29) appear in the research literature, the possibilities to solved increasingly
 4( D (θ ′ )t 
2 complex problems are promising.
34_Singh_ch34_p34.1-34.6.indd 4 8/22/16 1:58 PM

References 34-5
REFERENCES Serrano, S. E., “New approaches to the propagation of nonlinear transients

in porous media,” Transport in Porous Media, 93 (2): 331–346, 2012.
Abbaoui, K. and Y. Cherruault, “Convergence of Adomian’s method
Serrano, S. E., Engineering Uncertainty and Risk Analysis, 2nd ed., A
applied to differential equations,” Computers & Mathematics with Applications,
Balanced Approach to Probability, Statistics, Stochastic Models and Stochastic
28 (5): 103–109, 1994.
Differential Equations, HydroScience Inc., Ambler, PA, 2011.
Adomian, G., Solving Frontier Problems in Physics: The Decomposition
Serrano, S. E., Hydrology for Engineers, Geologists, and Environmental
Method, Kluwer Academic Publishers, Dordrecht, The Netherlands, 1994.
Professionals, 2nd ed., An Integrated Treatment of Surface, Subsurface, and
Adomian, G., “A review of the decomposition method and some recent
Contaminant Hydrology, HydroScience Inc., Ambler, PA, 2010.
results for nonlinear equations,” Computers & Mathematics with Applications,
Serrano, S. E., “Development and verification of an analytical solution for
21 (5): 101–127, 1991.
forecasting nonlinear kinematic flood waves,” Journal of Hydrologic
Adomian, G., Non-Linear Stochastic Operator Equations, Academic Press,
Engineering, ASCE, 11 (4): 347–353, 2006.
New York, 1986.
Serrano, S. E., “Modeling infiltration with approximate solutions to
Adomian, G., Stochastic Systems, Academic Press, New York, 1983.
Richard’s equation.” Journal of Hydrologic Engineering, ASCE, 9 (5): 421–432,
Adomian, G. and S. E. Serrano, “Stochastic contaminant transport equation
2004.
in porous media,” Applied Mathematics Letters, 1191: 53–55, 1998.
Serrano, S. E., “Modeling groundwater flow under a transient non-linear
Chapman, T. G. and R. F. Dressler, “Unsteady shallow groundwater flow
free surface,” Journal of Hydrologic Engineering, ASCE, 8 (3): 123–132,
over a curved impermeable boundary,” Water Resources Research, 20: 1427–
2003a.
1434, 1984.
Serrano, S. E., “Improved decomposition solution to Green and Ampt
Cherruault, Y., “Convergence of Adomian’s method,” Kybernetes, 18 (2):
equation,” Journal of Hydrologic Engineering, ASCE, 8 (3): 158–160, 2003b.
31–38, 1989.
Serrano, S. E., “Propagation of nonlinear reactive contaminants in porous
Cherruault, Y., G. Saccomardi, and B. Some, “New results for convergence
media,” Water Resources Research, 39 (8): 1228–1242, 2003c.
of Adomian’s method applied to integral equations,” Mathematical and
Serrano, S. E., “An explicit solution to the Green and Ampt infiltration
Computer Modelling, 16 (2): 85–93, 1992.
equation,” Journal of Hydrologic Engineering, ASCE, 6 (4): 336–340, 2001.
Duan, J.-S. and R. Rach, “A new modification of the Adomian decomposi-
Serrano, S. E., “Analytical decomposition of the non-linear infiltration
tion method for solving boundary value problems for higher-order nonlinear
equation,” Water Resources Research, 34 (3): 397–407, 1998.
differential equations,”Applied Mathematics and Computation, 218: 4090–
Serrano, S. E., “The Theis solution in heterogeneous aquifers,” Ground
4118, 2011.
Water, 35 (3): 463–467, 1997a.
Gabet, L., “The decomposition method and distributions,” Computers and
Serrano, S. E., “Non-Fickian transport in heterogeneous saturated porous
Mathematics with Applications, 27 (3): 41–49, 1994.
media,” Journal of Engineering Mechanics, ASCE, 123 (1): 70–76, 1997b.
Gabet, L., “The decomposition method and linear partial differential equa-
Serrano, S. E., “Hydrologic theory of dispersion in heterogeneous aquifers,”
tions,” Mathematical and Computer Modelling, 17 (6): 11–22, 1993.
Journal of Hydrologic Engineering, ASCE, 1 (4): 144–151, 1996.
Gabet, L., Equissed’uneThéorieDécompositionnelleet Application aux
Serrano, S. E., “Analytical solutions of the nonlinear groundwater flow
Equations aux DérivéesPartialles, Dissertation, Ecole Centrale de Paris,
equation in unconfined aquifers and the effect of heterogeneity,” Water
France, 1992.
Resources Research, 31 (11): 2733–2742, 1995a.
He, J-H.,“Homotopy perturbation method for solving boundary value
Serrano, S. E., “Forecasting scale-dependent dispersion from spills in het-
problems,” Physics Letters A, 350: 87–88, 2006.
erogeneous aquifers,” Journal of Hydrology, 169: 151–169, 1995b.
He, J-H.,“Variational iteration method—a kind of non-linear analytical
Serrano. S. E., “Semi-analytical methods in stochastic groundwater trans-
technique: some examples,” International Journal of Non-Linear Mechanics,
port,” Applied Mathematical Modelling, 16: 181–191, 1992.
34: 699–708, 1999.
Serrano, S. E., “Modeling infiltration in hysteretic,” Soil and Advanced
Kirkham, D., Theory of land drainage, Drainage of Agricultural Lands,
Water Resources, 13 (1): 12–23, 1990a.
AgronomyMonograph7, ASA, CSSA, and SSSA, Madison, WI, 1957.
Serrano, S. E., “Stochastic differential equation models of erratic infiltra-
Moutsopoulos, K. N., “Exact and approximate analytical solutions for
tion,” Water Resources Research, 26 (4): 703–711, 1990b.
unsteady fully developed turbulent flow in porous media and fractures for
Serrano, S. E., “General solution to random advective-dispersive equation
time dependent boundary conditions,” Journal of Hydrology, 369 (1–2):
in porous media. Part II: Stochasticity in the parameters,” Stochastic Hydrology
78–89, 2009.
and Hydraulics, 2 (2): 99–112, 1988.
Moutsopoulos, K. N., “One-dimensional unsteady inertial flow in phreatic
Serrano, S. E. and S. R. Workman, “Experimental verification of models of
aquifers, induced by a sudden change of the boundary head,” Transport in
nonlinear stream aquifer transients,” Journal of Hydrologic Engineering, ASCE,
Porous Media, 70: 97–125, 2007.
13 (12): 1119–1124, 2008.
Parlange, J. Y., “Theory of water movement in soils: I. One-dimensional
Serrano, S. E., S. R. Workman, K. Srivastava, and B. Miller-Van Cleave,
absorption,” Soil Science, 111 (2): 134–137, 1971.
“Models of nonlinear stream aquifer transients,” Journal of Hydrology, 336
Patel, A. and S. E. Serrano, “Decomposition solution of multidimensional
(1–2): 199–205, 2007.
groundwater equations,” Journal of Hydrology, 397: 202–2109, 2011.
Serrano, S. E. and S. R. Workman, “Modeling transient stream/aquifer
Philip, J. R., “The theory of infiltration: 1. The infiltration equation and its
interaction with the nonlinear Boussinesq equation and its analytical solu-
solution,” Soil Science, 83: 345–358, 1957.
tion,” Journal of Hydrology, 206: 245–255, 1998.
Philip, J. R., “Numerical solution of equations of the diffusion type with
Serrano, S. E. and G. Adomian, “New contributions to the solution of trans-
diffusivity concentration-dependent,” Transactions of the Faraday Society, 51
port equations in porous media,” Mathematical and Computer Modelling, 24
(7): 391, 1995.
(4): 15–25, 1996.
Rach, R., “A new definition of the Adomian polynomials,” Kybernetes, 37
Serrano, S. E. and T. E. Unny, “Semi group solutions of the unsteady
(7): 910–955, 2008.
groundwater flow equation with stochastic parameters,” Stochastic Hydrology
Read, W. W., “An analytical series method for Laplacian problems with
and Hydraulics, 1 (4): 281–296, 1987.
mixed boundary conditions,” Journal of Computational and Applied
Srivastava, K. and S. E. Serrano, “Uncertainty analysis of linear and nonlin-
Mathematics, 209: 22–32, 2007.
ear groundwater flow in a heterogeneous aquifer,” Journal of Hydrologic
Read, W. W. and R. E. Volker, “Series solutions for steady seepage through
Engineering, ASCE, 12 (3): 306–318, 2007, doi:10.1061/(ASCE)1084-
hillsides with arbitrary flow boundaries,” Water Resources Research, 29:
0699(2007)12:3(306).
2871–2880, 1993.
Srivastava, K., S. E. Serrano, and S. R. Workman, “Stochastic modeling of
Sarino, and Serrano, S. E., “Development of the instantaneous unit hydro-
stream-aquifer interaction with the nonlinear Boussinesq equation,” Journal
graph using stochastic differential equations,” Stochastic Hydrology and
of Hydrology, 328: 538–547, 2006, doi:10.1016/j.jhydrol.2005.12.035.
Hydraulics, 4 (2): 151–160, 1990.
Steward, D. R., X. Yang, and S. Chacon, “Groundwater response to chang-
Tiaiff, S. and Serrano, S. E., “Regional groundwater flow in the Louisville
ing water-use practices in sloping aquifers using convolution of transient
aquifer: new analytical solution in irregular domains,” Ground Water, 53:
response functions,” Water Resources Research, 45: 2009, doi:10.1029/
550–557, 2014, doi:10.1111/gwat.12242.
2007WR006775.
Serrano, S. E., “A simple approach to groundwater modeling with decom-
Szymkiewicz, R., Numerical modeling in open channel hydraulics, Water
position,” Hydrological Sciences Journal, 58 (1): 1–9, 2003, doi:10.1080/02626
Science and Technology Library, Springer, New York, 2010.
667.2012.745938.
34_Singh_ch34_p34.1-34.6.indd 5 8/22/16 1:58 PM

Tayfur, G. and V. P. Singh, “Simulating transient sediment waves in aggrad- Wazwaz, A.M. and A. Gorguis A., “Exact solutions for heat like and wave
ed alluvial channels by double-decomposition method,” Journal of Hydrologic like equations with variable coefficients,” Applied Mathematics and
Engineering, ASCE, 16 (4): 362–370, 2011. Computation, 149, 15–29, 2004.
Workman, S. R., S. E. Serrano, and K. Liberty, “Development and applica- Zauderer, E., Partial Differential Equations of Applied Mathematics. John
tion of analytical model of stream/aquifer interaction,” Journal of Hydrology, Wiley, New York, 1983.
200: 149–63, 1997.
Wazwaz, A-M.,“A new algorithm for calculating Adomian polynomials for
nonlinear operators,” Applied Mathematics and Computation, 111: 53–69, 2000.
34_Singh_ch34_p34.1-34.6.indd 6 8/22/16 1:58 PM

Chapter 35
Network Theory
BY
BELLIE SIVAKUMAR, FITSUM M. WOLDEMESKEL, KOREN FANG, AND VIJAY P. SINGH
ABSTRACT However, a key reason is the absence of a strong scientific theory that is suit-
able for studying all types of connections encountered in hydrology. This
Connections are everywhere in hydrology. Unraveling the nature and extent
deficiency has increasingly been realized in recent times, with many calls for
of connections in hydrologic systems, as well as their interactions with other
a generic theory or a broader framework (e.g., Dooge, 1986; Paola et al., 2006;
Earth systems, has always been a great challenge. Numerous scientific con-
Sivakumar, 2008; Young and Ratto, 2009). A host of yet new challenges, such
cepts and mathematical methods have been proposed and applied for study-
as impacts of global climate change (e.g., Sivakumar, 2011a), issues associated
ing the various connections in hydrology. Despite such efforts and progress,
with water planning, management, and conflicts (e.g., Sivakumar, 2011b),
an adequate knowledge of such connections continues to be elusive. A key
predictions in ungaged basins (e.g., Hrachowitz et al., 2013), and interactions
reason for this situation is the absence of a strong scientific theory that is suit-
between hydrology and society (e.g., Montanari et al., 2013), puts further
able for studying all types of connections encountered in hydrology. To this
emphasis on the importance of a generic theory for hydrology.
end, modern developments in network theory, and complex networks in
In the context of connections, network theory or graph theory can offer
particular, seem to provide new avenues. Applications of the concepts of
useful ideas. The concept of networks originated in the mid-eighteenth cen-
complex networks have recently started to emerge and are gaining significant
tury (Euler, 1741), and witnessed significant developments over the next three
momentum at the current time. This chapter provides an overview of network
centuries, including topology (Listing, 1848), trees (Cayley, 1857), and ran-
theory, especially the theory of complex networks, and its applications in
dom graph theory (Erdös and Rényi, 1959, 1960); see Harary (1969), Bondy
hydrology. The chapter: (1) reviews the basic concepts of a network and the
and Murty (1976), and Bollobás (1998) for extensive details of these develop-
history of the development of network theory (including graph theory, topol-
ments and their applications. However, developments since the late 1990s,
ogy, random graph theory, and complex networks); (2) describes different
including small-world networks (Watts and Strogatz, 1998), scale-free net-
types of networks (regular, random, small-world, and scale-free) and different
works (Barabási and Albert, 1999), network motifs (Milo et al., 2002), and
measures of their properties (e.g., centrality, clustering, adjacency, distance,
community structure (Girvan and Newman, 2002), put under the broad
and community structure) that have been widely studied; and (3) a brief
umbrella of the science of complex networks, have offered a whole new
review of their applications in hydrology, including studies on rainfall moni-
dimension and various avenues for studying connections encountered in
toring networks, streamflow monitoring networks, river networks, and vir-
natural and social systems; see, for example, Watts (1999), Barabási (2002),
tual water trade networks.
and Estrada (2012) for extensive details of these concepts and their applica-
tions in various fields.
35.1 INTRODUCTION
Applications of such complex networks-based concepts in hydrology and
water resources are just starting to emerge, including for rainfall monitoring
Connections are everywhere in hydrologic systems. The hydrologic cycle or networks (e.g., Malik et al., 2012; Boers et al., 2013; Scarsoglio et al., 2013;
water cycle is perhaps the best example of connections, with every component Sivakumar and Woldemeskel, 2015; Jha et al., 2015), streamflow monitoring
of the hydrologic cycle connected to every other component. Depending networks (e.g., Sivakumar and Woldemeskel, 2014; Halverson and Fleming,
upon the place they occupy in the hydrologic cycle and the spatial and 2014; Fang et al., 2015), river networks (Rinaldo et al., 2006; Zaliapin et al.,
temporal scales of interest, the components can be connected directly or 2010; Czuba and Foufoula-Georgiou, 2014, 2015), and virtual water net-
indirectly, and strongly or weakly. Similar to the hydrologic cycle, connections works (e.g., Suweis et al., 2011; Konar et al., 2011; Carr et al., 2012; Dalin et
exist within each component as well (e.g., spatial and temporal connections in al., 2012; D’Odorico et al., 2012; Tamea et al., 2013). The outcomes of these
rainfall). Unraveling the nature and extent of connections in hydrologic studies are certainly encouraging, as they offer important information about
systems, and also their interactions with other Earth systems (e.g., climatic, the connections in these networks and also suggest the potential of the
geomorphic, ecologic, and environmental), has always been a major challenge network-based concepts for studying other hydrologic networks. Indeed, in
in hydrology. light of the recent calls for a broader framework in hydrology, Sivakumar
Until now, a plethora of scientific concepts and mathematical methods (2015) has also made an argument in favor of network theory as a suitable
have been proposed and applied for studying the connections associated with tool to study all types of connections in hydrology and, hence, to serve as a
hydrologic systems. Such concepts and methods are based on, among others, generic theory.
time, distance, correlation, variability, scale, patterns, information, and many The purpose of this chapter is to provide an overview of network theory,
other properties/measures as well as their combinations and variants; see, for especially in the context of complex networks, and its applications in hydrol-
example, Gupta et al. (1986), Salas et al. (1995), Grayson and Blöschl (2000), ogy. The rest of the chapter is organized as follows. Section 35.2 reviews the
Yang et al. (2004), Mishra and Coulibaly (2009), Li et al. (2012), Sivakumar basic concept of a network and the history of the development of network
and Singh (2012), and Niu (2013) for some details. Despite such efforts and theory. Section 35.3 presents a brief description of the different types of
progress, our understanding of connections in hydrologic systems remains networks, and Sec. 35.4 describes the different measures for identification of
largely inadequate. Several major factors contribute to this situation, such as their properties. Section 35.5 presents a review of the applications of the
the inherent complexity of the hydrologic cycle and catchment systems, natu- ideas of network theory in hydrology. Section 35.6 offers some closing
ral and anthropogenic influences, and data and computational constraints. remarks.
35-1
35_Singh_ch35_p35.1-35.10.indd 1 8/22/16 1:59 PM

35-2 Network Theory
35.2 NETWORK THEORY: CONCEPT AND HISTORY 35.2.3 Topology

35.2.1 Concept of a Network Topology is the study of geometric properties and topological spaces, through
analysis of such concepts as space, dimension, and transformation. The basic
A network or a graph is a set of points connected together by a set of lines, as
ideas of topology go back to the works of Gottfried Wilhelm von Leibniz in
shown in Fig. 35.1. The points are referred to as vertices or nodes and the lines
the seventeenth century, with suggestions establishing the geometry of posi-
are referred to as edges or links; here, we use the terms nodes for points and
tions as a field without any measurements or calculations. However, the term
links for lines. Therefore, mathematically, a network can be represented as
“topology” was first introduced only in the nineteenth century, as topologie
G = {P, E}, where P is a set of N nodes (P1, P2,…,PN) and E is a set of n links.
(in German) by Johann Benedict Listing (Listing, 1848). Significant develop-
Figure 35.1 shows a network with N = 7 nodes and n = 8 links. In this
ments in topology were made during late nineteenth century and early twen-
network, P = {1, 2, 3, 4, 5, 6, 7} and the set of links is E = {{1, 7},{2, 3},{2, 5},{2, 7},
tieth century, with important contributions in the areas of set theory by Georg
{3,7 },{4, 7},{5, 6},{6, 7}}.
Cantor (Cantor, 1874), homotopy and homology (now part of algebraic topol-
Figure 35.1a, consisting of a set of identical nodes connected by identical
ogy) by Henri Poincaré (Poincaré, 1895), and on topological space by Felix
links, is perhaps the simplest form of network. This kind of network, however,
Hausdorff (Hausdorff, 1919). Topology had become a major branch of math-
is rarely seen in nature and society, since natural and social networks are often
ematics by the middle of the last century.
far more complex. There are many ways in which networks may be more
Topology has many sub-fields, including:
complex. For instance, a network may: (1) have more than one different type
• general topology, which deals with the basic set-theoretic definitions and
of node and/or link (Fig. 35.1b); (2) contain nodes and links with a variety of
constructions used in topology;
properties associated with them, such as different weights for different nodes
• algebraic topology, which uses tools from abstract algebra to study topo-
and links depending on the strength of nodes and connections (Fig. 35.1c);
logical spaces;
(3) have links that can be directed (pointing in only one direction Fig. 35.1d),
• differential topology, which deals with differentiable functions on dif-
with either cyclic (i.e., containing closed loops of links) or acyclic form; (4)
ferentiable manifolds; and
have multilinks (i.e., repeated links between the same pair of nodes), self-links
• geometric topology, which not only focuses on low-dimensional mani-
(i.e., links connecting a node to itself), and hyperlinks (i.e., links connecting
folds (i.e., dimensions 2, 3, and 4) and their interaction with geometry,
more than two nodes together); and (5) bipartite (containing nodes of two
but also includes higher-dimensional ones.
distinct types, with links running only between unlike types).
Extensive details on the concepts of topology and their applications are
35.2.2 Origin of Network Theory (Graph Theory) available in Bourbaki (1966), Mendelson (1990), and Adams and Franzosa
(2007), among others. In hydrology, the concept of topology has been exten-
The origin of the concept of networks can be traced back to the works of the
sively discussed with reference to river/channel networks; see, for example,
Swiss mathematician Leonhard Euler, during the first half of the eighteenth
Scheidegger (1967), Shreve (1967), Smart (1970), Coffman and Turner
century, on one of the most famous problems in mathematics, the Seven
(1971), Kirkby (1976), and Mark and Goodchild (1982) for some earlier
Bridges of Königsberg (Euler, 1741). The problem studied by Euler is as fol-
studies.
lows. The city of Königsberg in Prussia (now Kaliningrad, Russia) was set on
both sides of the Pregel River, and included two large islands which were 35.2.4 Trees
connected to each other and the mainland by seven bridges, thus making it a
network (for transportation). The problem Euler attempted to solve was to In graph theory, a tree (Cayley, 1857) is an undirected connected graph with-
find a route through the city that would cross each bridge once and only once out cycles. Therefore, a tree has to be a simple graph, since self-loops and
(including complete crossing of each bridge at once). While Euler proved that parallel links both form cycles. A graph G with n nodes is a tree, when it satis-
the problem had no solution, he also found that the choice of route inside fies the following equivalent conditions: (1) the graph is connected without
each land mass is irrelevant and that the only important feature of a route is cycles; (2) the graph has n – 1 links and no cycle; (3) the graph is connected
the sequence of bridges crossed. These findings allowed Euler to reformulate and has n – 1 links; (4) there exists a unique path between any two nodes of
the problem in abstract terms, eliminating all features except the list of land the graph; and (5) the graph becomes a nontree if a link is removed (i.e., the
masses and the bridges connecting them. This reformulation laid the founda- graph becomes disconnected) or added (i.e., the graph creates cycle).
tions of what would become popularly known as graph theory. The graph There are different types of trees, including:
theory has historically been a branch of the broader field of discrete mathe- • rooted tree—one node serves as a starting point and all the branches
matics, but is now a branch of combinatorics. Since this initial work by Euler, (i.e., links) fan out from this;
graph theory has witnessed a number of important theoretical developments, • binary tree—a rooted tree in which there are at most two descending
including topology (Listing, 1848), trees (Cayley, 1857), random graph theory branches at any node;
(Erdös and Rényi, 1959), and complex networks (Watts and Strogatz, 1998). • polytree—a directed acyclic graph whose underlying undirected graph is
A brief description of these developments is presented next. a tree; and
• spanning tree—a partial graph G′ = {P, E′} of a connected graph
G = {P, E} is a spanning tree of G if G′ is a tree. There is at least one span-
Node ning tree for any connected graph. A spanning tree of the graph that
optimizes the weight or cost (compared to all spanning trees of the
1 Link graph) is called an optimal spanning tree.
2
The concept of tree is highly relevant in hydrology, especially in river net-
6 works; see, for example, Horton (1945), Strahler (1957), Tokunaga (1978),
7 3 Rodriguez-Iturbe and Rinaldo (1997), Rinaldo et al. (2006), and Zaliapin et
al. (2010) for details.
4 35.2.5 Random Graph Theory
5
(a) (b) A random graph is, as the name suggests, a network of nodes connected by
links in a purely random fashion. The roots of random graph theory can be
traced back to the work of Solomonoff and Rapoport (1951). However, the
model introduced by Erdös and Rényi (1959, 1960) is the best-known version
today. This model—which is now known as the Erdös–Rényi model—is also
sometimes known as the “classical” random graph. Therefore, assuming a
network consisting of N nodes, the n links that connect these N nodes are
chosen randomly from all the N(N – 1)/2 possible links. Therefore, there are
a total of C[nN ( N −1)/2] graphs with N nodes and n links, forming a probability
space in which every realization is equiprobable. Considering that every pair
of nodes is connected with probability p, the total number of links is a random
(c) (d) variable with the expectation value E(n) = p[N(N – 1)/2]. If G0 is a graph with
nodes P1,P2, …,PN and n links, then the probability of obtaining such a graph
Figure 35.1 Networks: (a) an undirected network with only a single type of node
and a single type of link; (b) a network with a number of discrete node and link types; by the graph construction process is P(G0) = pn(1 – p)(N – 1)/2 – n.
(c) a network with varying node and link weights; and (d) a directed network in which As random graph theory basically examines the properties of the probabil-
each node has a direction. ity space associated with graphs with N nodes as N → ∞ , many properties of
35_Singh_ch35_p35.1-35.10.indd 2 8/22/16 1:59 PM

NETWORK TYPES 35-3
random graphs can be determined in a probabilistic sense. The classical ran-

dom graph model of Erdös and Rényi (1959, 1960) assumed that almost every
graph has a property Q if the probability of having Q approaches 1 as N → ∞ .
The primary purpose of random graph theory is to determine the connection
probability p at which a particular property of a graph will most likely arise.
Erdös and Rényi (1959, 1960) discovered that many important properties of
random graphs appear quite suddenly. In other words, at a given probability p,
either almost every graph has some property Q (e.g., every pair of nodes is
connected by a path of consecutive links) or almost no graph does. The transi-
tion from a property’s being very unlikely to its being very likely is usually
swift. For many such properties, there is a critical probability pc(N). If p(N)
grows more slowly than pc(N) as N → ∞, then almost every graph with con- (a) (b)
nection probability p(N) fails to have Q. If p(N) grows somewhat faster than
pc(N), then almost every graph has the property Q. Therefore, the probability
that a graph with N nodes and connection probability p = p(N) has property
Q satisfies
 p( N )
 0 if → 0
 pc ( N )
lim PN , p (Q ) =  (35.1)
N →∞
 1 if p( N ) → ∞
 pc ( N )

The random graph theory addressed a number of important questions that

are relevant for studying real complex systems, such as: (1) Is a typical graph (c) (d)
connected? (2) Does it contain a triangle of connected nodes? and (3) How
Figure 35.2 Network types: (a) regular network; (b) random network; (c) small-
does its diameter depend on its size? Consequently, the theory gained consid- world network; and (d) scale-free network.
erable attention and applications in many different fields until the end of the
last century, including in natural and physical sciences, engineering and
technology, economics, and social sciences; see, for example, Harary (1969),
Bondy and Murty (1976), and Bollobás (1998) for extensive details. The ran- shows, for example, four different types of networks: a completely regular
dom graph theory has found numerous applications in hydrology, especially network (Fig. 35.2a), a completely random network (Fig. 35.2b), a small-
in river/channel networks; see, for instance, Shreve (1966, 1967, 1969), Kirkby world network (Fig. 35.2c), and a scale-free network (Fig. 35.2d). The last
(1976), Smart and Werner (1976), Smart (1978), Moon (1980), and Werner two networks belong to random networks as well, but have different proper-
(1982) for some earlier studies. ties when compared to classical random networks. They are also generally
called complex networks. A brief description of these types of networks is
35.2.6 Complex Networks presented here.
While random graph theory has and continues to be applied for studying a
35.3.1 Regular Networks
broad range of networks in nature and society, including in hydrology, it also
has certain important deficiencies. For instance, the assumption that all net- Regular or ordered networks (Fig. 35.2a) are simple types of networks.
works are wired randomly together makes the application of this theory to Regular networks have a fixed number of nodes, with each node having the
real networks questionable, since order and determinism are inherent in real same number of links connecting it in a specific way to a number of neighbor-
systems and networks. Indeed, advances in some other areas of complex ing nodes. A simple example of a regular network is a 3D Cartesian grid. If
systems science, which revealed nonlinear deterministic dynamics, self- each node is linked to all other nodes in the network, then the network is a
organization, and scale invariance as inherent properties of complex systems fully connected one.
(e.g., Lorenz, 1963; Mandelbrot, 1983; Bak, 1996), also led to reconsideration Regular networks display a wide range of properties, because there are
of the assumption of random connections in complex networks. In addition, many ways to construct them while keeping the degree uniform across all
significant changes in technology, economy, and society during the last few nodes. In general, however, regular networks are highly clustered, with clus-
decades resulted in real networks that are highly irregular, complex, large, and tering coefficient C (a measure of local density—see Sec. 35.4.2 for details)
dynamically evolving, which necessitated the development of new methods close to 1.0 and, therefore, such networks are stable. However, they also have
and measures for studying network properties. long average path lengths L (a global measure of separation—see Sec. 35.4.4
All these led to a renewed and fresh perspective of the study of complex for details), and so often have a high degree of inefficiency. In the context of
networks in the late 1990s (Watts and Strogatz, 1998; Barabási and Albert, complex networks, stability means that the removal of any randomly chosen
1999), under the new science of networks. Such studies also led to new node will have little effect on the network as a whole, while efficiency means
discoveries about complex networks, such as small-world networks (Watts that information may easily be propagated across the network because the
and Strogatz, 1998), scale-free networks (Barabási and Albert, 1999), network average path length is small.
motifs (Milo et al., 2002), as well as other notable advances, such as a new
35.3.2 Random Networks
method for identifying community structure (Girvan and Newman, 2002).
Since then, the science of complex networks has found applications in many In random networks or classical random networks (e.g., Erdös and Rényi,
different fields, including natural and physical sciences, social sciences, 1959, 1960), the nodes are connected at random (Fig. 35.2b). In this case, the
medical sciences, economics, and engineering and technology (e.g., Albert et degree distribution pk (a measure of network spread—see Sec. 35.4.3 for
al., 1999; Bouchaud and Mézard, 2000; Newman, 2001a; Liljeros et al., 2001; details) is a Poisson distribution. The problem with these networks is that
Montoya and Solé, 2002; Tsonis and Roebber, 2004; Miguens and Mendes, they are not very stable, and removal of a number of nodes at random may
2008; Suweis et al., 2011; Sivakumar and Woldemeskel, 2014). Further details fracture the network to noncommunicating parts.
of complex network theory and its applications can be found in, for example, As the name suggests, classical random networks have no patterns. They
Watts (1999), Barabási (2002), and Estrada (2012). The concepts of complex have a small clustering coefficient (often close to zero) and a small average
networks, the measures to identify their properties, and their applications in path length. These mean that random networks tend to be unstable but effi-
hydrology are the focus of the rest of this chapter. cient. Some other important details about the random networks have already
been presented in Sec. 35.2.5 (random graph theory).
35.3 NETWORK TYPES 35.3.3 Small-World Networks
The recent concepts of complex networks and the new measures to identify While the concepts of regular networks and random networks have strong
their properties can be better explained by comparing such networks with theoretical underpinnings, they nevertheless have only limited application in
the previously known types (i.e., regular and random). To this end, Fig. 35.2 real networks encountered in nature and society. This is because, most
35_Singh_ch35_p35.1-35.10.indd 3 8/22/16 1:59 PM

35-4 Network Theory
real-world networks are neither completely ordered nor completely random, 35.4 NETWORK MEASURES
but rather exhibit important properties of both (Watts and Strogatz, 1998).
Over the past two decades or so, a large number of measures have been devel-
Indeed, developments in some other areas of complex systems science also
oped to study the properties of complex networks. The measures include
emphasize the influence of both order and randomness in real systems, as
centrality, clustering, adjacency, distance, community structure, bipartivity,
revealed by, for example, nonlinear dynamic and chaos theories (e.g., Lorenz,
fragments (or subgraphs), communicability, and global invariants, among
1963; May, 1976). According to Watts and Strogatz (1998), some properties of
others; see, for example, Estrada (2012) for details. Such measures not only
real networks can be embodied by simple mathematical models that
identify/quantify different properties of networks, but also often offer cross-
interpolate between order and randomness. They represented order by a
verification, and possible confirmation, of results. For some measures, there
uniform one-dimensional lattice (where each node is connected to its k
are also different definitions, submeasures, and corresponding methods, as
nearest neighbors on the lattice) and characterized randomness by a tunable
appropriate. In what follows, a brief description of some of these network
parameter p (which specifies the fraction of randomly rewired links). They
measures is presented: centrality (degree centrality), clustering (clustering
showed that these properties can be quantified with simple statistics, such as
coefficient), adjacency (degree distribution), distance (average shortest path
the clustering coefficient C (see Sec. 35.4.2) and the average shortest path
length), and community structure.
length L (see Sec. 35.4.4). In their model, both C and L can be measured as a
function of p. When p = 0 (completely ordered), the network is large [L(0) ∼ 35.4.1 Centrality (Degree Centrality)
N/2k] and highly clustered [C(0) ∼ ¾], as mentioned earlier (Sec. 35.3.1). On
Centrality is one of the most basic and intuitive measures of a network, as it
the other hand, when p = 1 (completely random), the network is small [L(1)
identifies the “most important” or “dominant” nodes. The concept of central-
∼ ln(N)/ln(k)] and poorly clustered [C(1) ∼ k/N], suggesting that path lengths
ity goes back to the studies of Bavelas (1948) and Leavitt (1951) for commu-
are short only when clustering is low, as highlighted earlier (Sec. 35.3.2). The
nication networks; see Freeman (1979) for an early comprehensive review.
model by Watts and Strogatz (1998) indeed exhibits a broad region of p values
However, Jeong et al. (2001) and Newman (2001b) were among the first to use
in which C(p) is high relative to its random limit C(l), yet L(p) is as small as
the concept in the context of complex networks. A number of centrality-based
possible. Watts and Strogatz (1998) coined the term small-world networks to
measures have been proposed in the network literature, such as degree cen-
refer to networks in this class, after similar concepts as manifestation of small
trality, centrality beyond nearest neighbors (e.g., Katz centrality, eigenvector
worlds; see Karinthy (1929) and Milgram (1967). The small-world network
centrality, subgraph centrality, pagerank centrality, and vibrational centrality),
concept of Watts and Strogatz (1998) highlights the fact that most networks,
closeness centrality, betweenness centrality, and information centrality; see,
despite their often large size, have a relatively short path between any two
for example, Estrada (2012) for details. Among these, the degree centrality has
nodes (see Fig. 35.2c). Therefore, small-world networks are one form of ran-
been one of the most widely used measures.
dom networks, but different from classical random networks: they are stable
The idea behind the use of degree centrality as a network measure is that it
(unlike classical random networks) and efficient (like classical random net-
identifies whether a given node, say i in a network, is more central or more
works).
influential than another node in the network. The degree centrality of node i
35.3.4 Scale-Free Networks
in a network of N nodes is defined as the number of first neighbors (or simply
neighbors) of node i divided by the total number of possible neighbors (N – 1)
The random graph of Erdös and Rényi (1959, 1960) and the small-world in the network. Let us consider a selected node i in a network of N nodes,
network of Watts and Strogatz (1998) display Poissonian degree distributions. having ki links which connect it to ki other nodes. For illustration, Fig. 35.3
However, there are real-world networks that display a very different degree presents a network consisting of nine nodes (i.e., N = 9), with the node i hav-
distribution, which is characterized by a few nodes of high degree and a large ing four links (i.e., with four other nodes) (see Fig. 35.3, left panel). In this
proportion of nodes with relatively low degree (Barabási and Albert, 1999). case, the four nodes corresponding to the four links are the neighbors of
Although such networks can display a wide range of “fat-tail” degree distribu- node i, which are identified based on some conditions (e.g., correlation
tions, the easiest way of conceptualizing such topological characteristics is to between node i and other nodes in the network), while the total number of
consider a model in which P(k) ∼ k–γ, where γ is the degree exponent. In other possible neighbors for node i is eight (i.e., N – 1). The procedure is repeated
words, a model in which the probability of finding a node with degree k for each and every node of the network.
decreases as a power-law of its degree.
Barabási and Albert (1999) were the first ones to address the origin of the 35.4.2 Clustering (Clustering Coefficient)
power-law degree distribution observed in networks. They argued that the One of the most fundamental properties of networks is their tendency to
scale-free nature of real networks is rooted in two generic mechanisms shared cluster. The concept of clustering has its origin in sociology, under the name
by many real networks as follows: fraction of transitive triples (Wasserman and Faust, 1994), but Watts and
1. The random graph and small-world network assume that the network Strogatz (1998) were the first to use this concept in the context of complex
starts with a fixed number N of nodes that are then randomly connected networks. The tendency of a network to cluster is quantified by the clustering
or rewired, without modifying N. However, real-world networks gener- coefficient. There exist several definitions of clustering coefficient; see Watts
ally describe open systems that grow by the continuous addition of new and Strogatz (1998), Barrat and Weigt (2000), and Newman (2001a) for
nodes. The networks start with only a few nodes and the number of details. However, the clustering coefficient method proposed by Watts and
nodes increases throughout the lifetime of the networks by the subse- Strogatz (1998), which measures the local density, is very widely used. A brief
quent addition of new nodes (e.g., World Wide Web). description of its calculation is presented here.
2. The random graph and small-world network assume that the probability Let us consider first a selected node i in the network, having ki links which
that two nodes are connected is independent of the nodes’ degree, that connect it to ki other nodes (Fig. 35.3, left panel), as mentioned earlier. If the
is, new links are always placed randomly. However, most real networks neighbors of the original node i were part of a cluster, there would be
exhibit preferential attachment, in such a way that the likelihood of con- ki(ki – 1)/2 links between them. As shown in Fig. 35.3 (right), with four neigh-
necting to a node basically depends on the node’s degree. bors of the node i part of the cluster, there are 4(4 – 1)/2 = 6 links in the
These two mechanisms inspired the introduction of the Barabási–Albert

model (Barabási and Albert, 1999) and led, for the first time, to a network
Neighbors = 4 Actual connections = 3
with a power-law degree distribution.
The dynamic properties of the scale-free model can be addressed using
various analytic approaches. Barabási and Albert (1999) proposed the con-
tinuum theory, which focuses on the dynamics of node degrees. Dorogovtsev
et al. (2000) introduced the master-equation approach, while Krapivsky et al. i i
(2000) proposed the rate-equation approach.
Finally, from an overall perspective, the goal of random graph and small-
world network is to construct a graph with correct topological features.
However, the modeling of scale-free networks puts the emphasis on capturing
the network dynamics. The underlying assumption behind evolving or
dynamic networks is that if the processes that assembled the networks are
captured correctly, then their topology can also be obtained correctly.
Therefore, dynamics takes the driving role, and topology is only a by-product Figure 35.3 Connections in networks and calculation of clustering coefficient: nearest
of this modeling philosophy. neighbors and actual connections.
35_Singh_ch35_p35.1-35.10.indd 4 8/22/16 1:59 PM

NETWORK MEASURES 35-5
cluster of node i. The clustering coefficient of node i is then given by the ratio 35.4.4 Distance (Average Shortest Path Length)
between the number Ei of links that actually exist between these ki nodes Several distance-based metrics are used for studying the topology of net-
(shown as solid lines on Fig. 35.3, right) and the total number ki(ki – 1)/2 (i.e., works, including the average shortest path length, resistance length, and
all lines on Fig. 35.3, right), generalized network length. The average path length (or distance) is consid-
ered as one of the most robust measures of network topology, along with
2 Ei clustering coefficient and degree distribution. In a network, the shortest path
Ci = (35.2)
ki (ki − 1) length of a node pair i and j is the number of links on the shortest path con-
necting the node pair. If the node pair is unconnected, then the value of the
The procedure is repeated for each and every node of the network. The aver- shortest path length is set to infinite. The average path length (L) of a network
age of the clustering coefficients Ci’s of all the individual nodes is the cluster- with N nodes is the average over all nodes of the shortest path between every
ing coefficient of the whole network C. combination of node pairs, and is given by
The clustering coefficient of the individual nodes and of the entire network
can be used to obtain important information about the type of network, group- 1
ing (or classification) of nodes, and identification of dominant nodes (e.g.,
L=
N ( N − 1)
∑ dij (35.7)
super nodes), among others. For instance, as mentioned earlier, a high cluster-
ing coefficient (close to 1.0) indicates a regular network (Fig. 35.2a), while a where dij is the distance between pair i and j.
very low clustering coefficient (close to 0), with C = p, indicates a random This definition for average shortest path length, however, diverges if there
network (Fig. 35.2b). The clustering coefficient of a small-world network (Fig. are unconnected nodes in the network, since the distance between such nodes
35.2c) and a scale-free network (Fig. 35.2d) is generally not only smaller than is set to infinite; see also da F. Costa et al. (2007). Consideration of only the
that of the regular network, but also considerably larger than that of a compa- connected node pairs avoids this divergence problem, but such also intro-
rable random network (i.e., having the same number of nodes and links). duces a distortion for networks with many unconnected pairs of nodes. The
consequence of this is a small value of the average path length, which is
35.4.3 Adjacency (Degree Distribution) expected only for networks with a high number of connections.
In a network, several structural properties are related to adjacency relation- A closely related measurement is the global efficiency (E), proposed by
ships between nodes. There are different ways to measure the adjacency Latora and Marchiori (2001):
relations in a network, including node adjacency, degree distribution, degree–
degree correlation, and link adjacency. Among these, the degree distribution 1 1
is a particularly useful measure (e.g., Barabási and Albert, 1999), especially for E= ∑ (35.8)
N ( N − 1) dij
the identification of the type of network, and is widely used in network
studies. A brief description is as follows.
where the sum takes all pairs of nodes into account. The global efficiency
In a network, different nodes may have different number of links. The
quantifies the efficiency of the network in sending information between
number of links (k) of a node is called as node degree. The degree is an impor-
nodes, with the assumption that the efficiency for sending information
tant characteristic of a node, as it allows one to derive many measurements for
between two nodes i and j is proportional to the reciprocal of their distance.
the network. The spread in the node degrees is characterized by a distribution
The reciprocal of the global efficiency is the harmonic mean of the geodesic
function p(k), which expresses the fraction of nodes in a network with
distances [see Eq. (35.7)], given by
degree k. This distribution, called degree distribution, is often a reliable indica-
tor of the type of network. In a random graph, since the links are placed
randomly, the majority of nodes have approximately the same degree, and 1
h = (35.9)
close to the average degree 〈k〉 of the network. Therefore, the degree distribu- E
tion of a completely random graph is a Poisson distribution with a peak at
P(〈k〉), and is given by: The fact that the harmonic mean eliminates the divergence problem other-
wise encountered in the average shortest path length makes is a more appro-
priate measurement for networks with more than one connected component.
e−k k k The average shortest path length offers important information about the
p( k ) = (35.3)
k! type of network. For instance, as mentioned earlier, regular networks, with
their high clustering (i.e., stable), have long average path lengths (i.e., ineffi-
Similarly, depending upon the properties of networks, degree distribution can cient). Random networks have short average path lengths (i.e., efficient) but
be Gaussian, given by: have low clustering (i.e., unstable). Small-world networks have short path
lengths and have high clustering, and so are both stable and efficient.
 ( k − k )2 
−
1 2 σ k2 

p( k ) = e  (35.4) 35.4.5 Community Structure
√ 2πσ k In many complex networks, nodes cluster together into distinct groups, each
of which is more densely linked together when compared to the rest of the
exponential, given by: network. The properties of these groups are more or less independent of the
properties of individual nodes and of the network as a while. These groups are
p( k ) ~ e − k/k (35.5) known as communities and this kind of network structure is known as com-
munity structure.
Identification of communities in networks, especially large networks, is
power-law or scale-free, given by: particularly useful, since nodes belonging to the same community are more
likely to share properties and dynamics. Further, the number and character-
istics of the existing communities provide subsidies for identifying the type of
p( k ) ~ k − γ (35.6) a network and in understanding its dynamic evolution and organization.
Community identification is highly relevant and useful in hydrology, includ-
or other. ing in the contexts of catchment classification and data interpolation/
Among these distributions, the power-law or scale-free distribution has extrapolation.
attracted the most attention in the literature on complex networks, as such has Despite the importance of the concept of community in networks, there is
been found in a large number of natural and social networks (e.g., Barabási no consensus about its definition, and there exist several different formal
and Albert, 1999; Kim et al., 2004; Keller, 2005; Clauset et al., 2010). The definitions. The strictest definition is a clique, a subset where all nodes are
fractal or scale-free nature of numerous natural and social systems and their neighbors. In other words, the shortest path between any node pair within a
ability also to self-organize themselves, already well-documented in the clique is 1. By relaxing this requirement, the definition of n-cliques is found,
literature (e.g., Mandelbrot, 1983; Bak, 1996; Barnsley, 2012), give both where the maximum shortest path between any node pair is n. An alternative
credence and motivation to further advance research on scale-free networks. set of community definitions is based upon the relative frequency of links.
Although it is true that some scale-free networks display an exponential tail, One definition in this category is a LS set, in which nodes have more links to
the functional form of P(k) still deviates significantly from the Poisson nodes within the set than to other nodes. Radicchi et al. (2004) provided a
distribution expected for a random graph. simple and intuitive definition of a community based on the comparison of
35_Singh_ch35_p35.1-35.10.indd 5 8/22/16 1:59 PM

35-6 Network Theory
the link density among nodes. As per this definition, a community is defined the intracommunity connections are sparse. Other methods include the
as a set of nodes which has more internal links than external links. Silhouette index, performance, and Davies-Bouldin Validation Index.
A number of methods have been developed for community detection in Still other methods to compare whether the communities identified by one
networks, based on the different definitions of a community and on the need algorithm is similar to those found by another algorithm exist. The normal-
for faster approximations in large networks. Some of these methods rely on ized mutual information (NMI) measure (Danon et al., 2005) is one such
the modularity, Q, which quantifies the quality or strength of a community method. The NMI utilizes a confusion matrix where the rows represent the
(see later for details). The community detection methods include edge communities identified in one method and the columns represent the com-
betweenness centrality, greedy optimization, multilevel modularity optimiza- munities from another method. The measure of similarity between methods
tion, leading eigenvector, label propagation, walktrap, info map, optimal, and is given by
several others. The first six methods are briefly described here, as they are
more widely used.  
N ij N
1. Edge betweenness centrality: It is a measure of how central a particular −2∑ ∀i ∑ ∀jN ij log  
 ∑ iN ij ∑ jN ij 
link is in a network. It measures the number of shortest paths which pass NMI( A, B ) = (35.12)
through the link. The links with the highest centrality are thought to act  ∑ N ij   ∑ N ij 
as bridges joining the communities together. Removing these links will ∑ ∀i∑ jNij log  N  + ∑ ∀j∑ iNij log  N 
j i
split the network into more densely connected communities. As the    

algorithm (Newman and Girvan, 2004) removes links, a dendrogram is
constructed for later analysis to determine the communities. where i corresponds to the communities found with identification method
2. Greedy algorithm: The greedy algorithm, formulated by Clauset et al. A, j corresponds to the communities found with identification method B, N
(2004), attempts to optimize the modularity (i.e., quality or strength of a is the total number of nodes in the network, and Nij is the number of nodes
community—see later for further details) directly by first placing each that appear in both community i and community j.
node into its own community. The change in property Q from joining If the identified communities are identical, then the NMI will have the
any two communities together is then calculated. The community pairs maximum value of 1, while a value of 0 means that the identified communi-
that will produce the largest change in Q are merged together and this is ties are completely independent of each other.
repeated until a maximum value of Q is obtained.
3. Multilevel modularity optimization: This algorithm was proposed by
Blondel et al. (2008) and utilizes a similar greedy approach with each 35.5 APPLICATIONS IN HYDROLOGY
node first assigned to its own community. The nodes are then reassigned
Successful applications of the concepts of complex networks in various fields
one by one into a community that will provide the highest increase in
have inspired applications of such concepts in hydrology and water resources
modularity. When no more nodes can be reassigned, then each com-
as well. Studies in this direction are still at a very early stage, and the applica-
munity is considered a node on its own and the process is repeated. This
tions have thus far been limited to rainfall monitoring networks, streamflow
process continues until there is only one node left or if the modularity
monitoring networks, river networks, and virtual water networks (e.g.,
cannot be increased further.
Rinaldo et al., 2006; Zaliapin et al., 2010; Suweis et al., 2011; D’Odorico et al.,
4. Leading eigenvector method: This method (Newman, 2006) takes
2012; Scarsoglio et al., 2013; Sivakumar and Woldemeskel, 2014, 2015).
advantage of the algebraic properties from the matrix representations of
Sivakumar (2015) discussed the general relevance of the concepts of complex
networks in order to optimize the modularity. It defines the modularity
networks in hydrology and presented three specific examples (the hydrologic
matrix as:
cycle, streamflow monitoring network, and parameter estimation in hydro-
B = A − P (35.10) logic models) as to the treatment of hydrologic systems as complex networks.
In light of the generality of the concept and early promising outcomes,
where A is the adjacency matrix of the network and P is the probability matrix Sivakumar (2015) also argued that network theory can serve as a generic
of links existing according to the configuration model. The eigenvector of the theory for hydrology. In what follows, a brief account of the applications of
modularity matrix for the largest positive eigenvalue is calculated and nodes network theory in hydrology is presented. For details, the interested reader is
are then separated based on the sign of the corresponding element in the directed to the respective studies.
eigenvector.
35.5.1 Rainfall Monitoring Networks
5. Label propagation method: This method (Raghavan et al., 2007) first To our knowledge, the study by Malik et al. (2012) was the first one to apply
assigns each node with its own unique label. Each node will then adopt the concepts of network theory to rainfall. Malik et al. (2012) examined the
the label which the majority of its neighbors have in an iterative process. spatial and temporal characteristics of extreme (summer) monsoonal rainfall
Densely connected nodes will reach a consensus on a unique label for over South Asia, by analyzing daily gridded rainfall data over the period
the group. When the labels cannot propagate any further, the process 1951–2007 from the APHRODITE (Asian Rainfall Highly Resolved
ends and nodes with the same label are grouped as one community. Observational Data Integration Towards the Evaluation of Water Resources)
6. Walktrap method: This algorithm, proposed by Pons and Latapy (2006), project. They applied a host of network-based methods, including degree
uses short random walks to define a distance between nodes and com- centrality, degree distribution, clustering coefficient, and closeness centrality.
munities. Nodes that are within the same community should have short Following up on the study by Malik et al. (2012), Boers et al. (2013) applied
distances between them, while nodes from different communities would similar network concepts to investigate the spatial characteristics of extreme
have long distances. Similar to the algorithms mentioned earlier, the rainfall synchronicity of the South American Monsoon System (SAMS),
initial step separates each node into its own community. Communities through analysis of a 15-year long (January 1998–December 2012) gridded
are then joined on the basis of minimizing the mean of the squared daily rainfall data with a spatial resolution of 0.25° × 0.25°, obtained from the
distances between each node and its community. This iterates until a Tropical Rainfall Measuring Mission (TRMM) 3B42 V7 satellite product.
single community consisting of all the nodes is formed. The dendrogram Scarsoglio et al. (2013) applied the complex networks concepts to examine the
formed during this process is then used to decide on the final commu- spatial dynamics of annual precipitation around the globe. They analyzed a
nity partitions. 70-year long (January 1941–December 2010) gridded precipitation data from
Given that there are many methods for the identification of communities, the Global Precipitation Climatology Center (GPCC) Database using a num-
with each providing differing results, there is a need to evaluate the quality (or ber of methods, including degree centrality, clustering coefficient, degree
strength) of the identified communities as well as the similarities between distribution, and shortest path length.
identification methods. One of the most popular methods for quantifying the Sivakumar and Woldemeskel (2015) examined the spatial connections in
quality of a community division is modularity, which is defined as: rainfall in Australia using concepts of complex networks. They employed the
clustering coefficient method and the degree distribution method to monthly
Q = ∑(eii − ai2 ) (35.11) rainfall observed over a period of 68 years (1940–2007) at 230 raingage sta-

i
tions across Australia. Their study was the first study that employed complex
networks concepts to ground-measured rainfall data, as opposed to the above
where ai is the fraction of links in the network which connect to community studies that used gridded rainfall data. They also considered the influence of
i and eii is the fraction of links that exist within the community. A high rainfall correlation threshold on network properties, by carrying out the
modularity community will have dense intercommunity connections while analysis for several different thresholds. They identified regions/stations of
35_Singh_ch35_p35.1-35.10.indd 6 8/22/16 1:59 PM

APPLICATIONS IN HYDROLOGY 35-7
high and low connectivities, which have important implications for optimal many related properties (e.g., Mandelbrot, 1983; Rodriguez-Iturbe and
raingage density and locations. Their results also suggested that the rainfall Rinaldo, 1997; Peckham and Gupta, 1999).
monitoring network is not a classical random network, but more likely an Rinaldo et al. (2006) introduced the concepts of complex networks in
exponentially truncated power-law network. Following up on the study by hydrology, through a review of theoretical and observational developments
Sivakumar and Woldemeskel (2015), Jha et al. (2015) attempted to offer on the form and function of natural networks in different contexts in differ-
hydrologic explanation for the outcomes of network-based methods. For this ent fields and their relevance in hydrology. They discussed the properties
purpose, they applied the clustering coefficient method to two different rain- and dynamic origin of the scale-invariant structure of river patterns and its
gage networks in Australia (57 stations from Western Australia and 45 sta- relation to optimal network selection, and argued that purely random or
tions from the Sydney region) and interpreted the results in terms of deterministic constructs are unsuitable for a proper description of river net-
topographic properties of raingage stations (latitude, longitude, and eleva- works and other natural network forms. They reported, through application
tion) and characteristics of rainfall data (mean, standard deviation, and coef- of degree distribution, clustering coefficient, and average path length meth-
ficient of variation). ods, the emergence of nontrivial phase transitions with increasing links-to-
nodes ratios and that different features like scale-free or small-world
35.5.2 Streamflow Monitoring Networks networks are obtained for particular cases; see also Colizza et al. (2004).
They concluded that the emergence of the structural properties in river (and
The study by Sivakumar and Woldemeskel (2014) was the first ever that
other natural) networks may not necessarily be due to embedded rules for
applied the concepts of complex networks to streamflow monitoring net-
growth, but may rather reflect the interplay of dynamic mechanisms with an
works. The study employed the degree centrality method and the clustering
evolutionary selective process. Rinaldo et al. (2014) discuss further on the
coefficient method to examine spatial connections in monthly streamflow
statics, dynamics, and complexity of the evolution and optimal selection of
observed over a period of 52 years (1951–2002) at 639 streamflow gaging sta-
river networks.
tions in the contiguous United States. The study also investigated the influ-
Zaliapin et al. (2010) applied the concepts of network theory to study envi-
ence of streamflow correlation threshold on the aforementioned properties of
ronmental transport problems in river networks. Specifically, they studied the
the network. One of the important outcomes of the study is that even nearest
dynamic topology of directed trees. They described the static geometric
stations can have significantly different properties as part of a network and
structure of a drainage network by a tree, referred to as the static tree, and
even distant stations can have similar properties. This has important implica-
introduced an associated dynamic tree that describes the transport along the
tions for interpolation and extrapolation of data, including for predictions in
static tree. Applying connectivity concepts, such as hierarchical aggregation
ungaged basins. The study also revealed that the streamflow monitoring net-
and clustering, they showed that dynamic trees are also self-similar just as
work is not a classical random graph, but of some other nature, which has
their corresponding static trees, but that their properties differ systematically
implications for model development, including identification of the nature
from those of the corresponding static trees. They also reported an unex-
and complexity of the model. The outcomes of the study also led to a series of
pected phase transition in the dynamics of river networks (one from
other studies, including the role of hydrologic regions on network properties
California and two from Italy) and demonstrated universal features of this
(i.e., “regional” vs. “global” network), optimum number of gaging stations,
transition, with interpretation in a hydrologic context.
and catchment classification using community structure, details of which will
Czuba and Foufoula-Georgiou (2014) proposed a simplified network-
be reported elsewhere.
based predictive framework of sedimentological response in a basin, incor-
Halverson and Fleming (2015) applied the concepts of complex networks
porating network topology, channel characteristics, and transport-process
to a network of 127 streamflow monitoring stations in Canada. In addition to
dynamics to perform a nonlinear process-based scaling of the river-network
the investigation of whether regional streamflow hydrology might be quanti-
width function to a time-response function. They developed the process-
tatively represented as a formal network, their study aimed at assessing
scaling formulation for transport of mud, sand, and gravel, using simplify-
whether the results from the network-based methods might offer important
ing assumptions including neglecting long-term storage. They applied the
information as to the optimal design of streamflow monitoring systems. They
methodology to the Minnesota River basin in the United States. Their
employed a host of network-based methods, including clustering coefficient,
results indicated, among others, that the network topology and sediment-
degree distribution, average shortest path length, betweenness, and commu-
transport dynamics combined to produce a double peaked response func-
nity structure. The community structure analysis identified 10 separate com-
tion for sand, suggesting that there exists a resonant frequency of sediment
munities, each of which defined by the combination of its median seasonal
supply that could lead to an unexpected downstream amplification of sedi-
flow regime and geographic proximity to other communities, with important
mentological response. Subsequently, Czuba and Foufoula-Georgiou (2015)
implications for catchment classification. Based on such results, they pro-
extended this framework to understand the internal dynamics of the basin
posed that an idealized sampling network should sample high-betweenness
for sediment transport, that is, how sediment is organized and where sedi-
stations as well as small-membership communities which are, by definition,
ment accumulates due to the combined effects of river network structure
rare or undersampled relative to other communities, while at the same time
(topology and associated geometry) and transport dynamics (accounting
retaining some degree of redundancy to maintain network robustness.
for slopes, channel morphology, bed shear stress, grain size, etc.).
Fang et al. (2015) introduced the concept of complex networks and com-
Specifically, they developed a dynamic connectivity framework and applied
munity structure to classify catchments in large-scale river basins, consider-
it to understand sand transport in the Greater Blue Earth River Network in
ing the Mississippi River basin as a representative basin. They applied six
the Minnesota River basin.
community detection methods: edge betweenness, greedy algorithm, multi-
level modularity optimization, leading eigenvector, label propagation, and 35.5.4 Virtual Water Trade Networks
walktrap. They examined the influence of correlation threshold on classifica-
tion. The consistency among the methods in classifying catchments was Virtual water trade networks have been the most studied networks, thus far,
assessed using the NMI index. Their results suggested the role of geography in the context of complex networks in hydrology and water resources. Suweis
and river branching on classification and also a notable influence of correla- et al. (2011) applied the complex network theory to examine the connections
tion threshold on the number and size of communities identified. They also in virtual water trade networks. They used a simple model to describe the
reported high degree of consistency in the performance among the six meth- topological and weighted properties of the virtual water trade network
ods, except for the leading eigenvector method at lower thresholds. Detailed between world nations for staple food products, using degree distribution,
investigations as to the specific role of river branching on classification are average nearest neighbor degree, and clustering coefficient methods.
currently underway. Assuming the gross domestic product and yearly rainfall on agricultural areas
in each country as sole drivers, they identified and quantified the high degree
of globalization of water trade and showed that only a small group of nations
35.5.3 River Networks and Processes play a key role in the connectivity of the network and in the global redistribu-
The relevance of the concepts of networks to the structure and dynamics of tion of virtual water. Konar et al. (2011) applied the concepts of complex
rivers has been known for many decades now, especially from the point of networks to quantify the topology of the virtual water network associated
view of topology, trees, and random networks (e.g., Horton, 1945; Strahler, with international food trade (58 commodities from five major crops and
1957; Shreve 1966, 1967; Scheidegger, 1967; Kirkby, 1976; Tokunaga, 1978; three livestock types). They analyzed both directed and weighted properties
Werner, 1982; Rodriguez-Iturbe and Rinaldo, 1997). It is no surprise, thereof the networks. They reported that the number of trade connections follows
fore, that the first applications of complex networks concepts in hydrology an exponential distribution, except for the case of import trade relationships,
were to study the river networks. Among the new network concepts, scale- while the volume of water that each nation trades compares well with a
free networks and network motifs (subgraphs) are readily applicable to river stretched exponential distribution, indicating high heterogeneity of flows
networks, since river networks exhibit scale-invariance, self-organization, and between nations. They also reported that there is a power-law relationship
35_Singh_ch35_p35.1-35.10.indd 7 8/22/16 1:59 PM

35-8 Network Theory
between the volume of water traded and the number of trade connections of a generic theory for hydrology. The future of network theory in hydrology
each nation. looks bright!
Carr et al. (2012) employed the degree distribution method to the virtual
water network associated with international food trade over the period
ACKNOWLEDGMENTS
1986–2008. They found that the total flow over the period had more than
doubled, and the number of links had increased by 92%. The network grew This work is supported by the Australian Research Council (ARC) Future
fast by establishing new links during 1990–2001 and then by increasing the Fellowship grant (FT110100328). Bellie Sivakumar acknowledges the finan-
fluxes through existing links during 2002–2008. The virtual water network cial support from ARC through this Future Fellowship grant. We thank
was found to have become more homogeneous but most of the flow was Dawen Yang and Jun Niu for their review comments and suggestions on an
concentrated in a few links and hubs, while several countries exhibited only earlier version of the manuscript.
few, and mostly weak, connections. Dalin et al. (2012) used the concepts of
complex networks, including the degree distribution method, to study the REFERENCES
virtual water network associated with international food trade during 1986–
2007 and linked it to trade policies, socioeconomic circumstances, and agri- Adams, C. C. and R. D. Franzosa, Introduction to Topology: Pure and
cultural efficiency. They found that both regional and national virtual water Applied, Pearson Prentice Hall, Harlow, UK, 2007.
trade patterns significantly changed, with Asia increasing its virtual water Albert, R., H. Jeong, and A-L. Barabási, “Internet: diameter of the World
imports by more than 170%, switching from North America to South Wide Web,” Nature, 401: 130–131, 1999.
America as its main partner, whereas North America oriented to a growing Bak, P., How Nature Works: The Science of Self-Organized Criticality,
intraregional trade. Springer-Verlag, New York, 1996, p. 212.
D’Odorico et al. (2012) studied the community structure of the virtual Barabási, A-L., Linked: The New Science of Networks, Perseus, Cambridge,
water network. They employed a method based on the maximization of the MA, 2002.
modularity (Newman and Girvan, 2004) to the virtual water network over Barabási, A-L. and R. Albert, “Emergence of scaling in random networks,”
the period 1986–2008. Identifying an increase in the ratio between virtual Science, 286: 509–512, 1999.
water flows within communities and the total global trade of virtual water, Barnsley, F. M., Fractals Everywhere, Dover, New York, NY, 2012.
they reported the existence of well-defined clusters of virtual water transfers. Barrat, A. and M. Weigt, “On the properties of small-world networks,”
However, geographic proximity was found to only partly explain the com- European Physical Journal B, 13: 547–560, 2000.
munity structure of virtual water trade. Tamea et al. (2013) introduced the Bavelas, A., “A mathematical model for group structure,” Human Organiza-
average distance traveled by virtual water as a measure for the analysis of tion, 7: 16–30, 1948.
globalization in the virtual water trade. They visualized the fluxes between a Blondel, V. D., J-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfold-
country and its importing/exporting partners with a geographical represen- ing of communities in large networks,” Journal of Statistical Mechanics, 2008
tation shaping the trade network as a virtual river/delta. Using Italy as a (10): P10008, 2008.
country of interest, they found that food trade has a steadily growing impor- Boers, N., B. Bookhagen, N. Marwan, J. Kurths, and J. Marengo, “Complex
tance compared to domestic production, with a major component repre- networks identify spatial patterns of extreme rainfall events of the South
sented by plant-based products, and luxury products taking an increasingly American Monsoon System,” Geophysical Research Letters, 40: 1–7, 2013,
larger share. Replicating the analysis to 10 other countries, which triggered doi:10.1002/grl.50681.
similar investigations on different socioeconomic actualities, they showed Bollobás, B., “Modern Graph Theory,” Springer, New York, 1998.
that integrating the local (country) perspective and the global perspective Bondy, J. A. and U. S. R. Murty, Graph Theory with Applications, Elsevier
can shed new light on the structure and dynamics of virtual water trade Science, New York, NY, 1976.
network. Bouchaud, J-P. and M. Mézard, “Wealth condensation in a simple model of
Several other studies have also examined the structure and dynamics of economy,” Physica A, 282: 536–540, 2000.
virtual water trade networks, either in the context of complex networks or in Bourbaki, N., Elements of Mathematics: General Topology, Addison-Wesley,
other contexts that support the results reported by the above studies. Reading, MA, 1966.
Examples of such studies are those by D’Odorico et al. (2010), Konar and Cantor, G., “Üeber eine Eigenschaft des Inbegriffes aller reellen alge-
Caylor (2013), Konar et al. (2013), O’Bannon et al. (2014), Dalin et al. (2014), braischen Zahlen,” Journal fur die Reine und Angewandte Mathematik, 77:
and Tamea et al. (2014). 258–262, 1874.
Carr, J., P. D’Odorico, F. Laio, and L. Ridolfi, “On the temporal variability
of the virtual water network,” Geophysical Research Letters, 39: L06404, 2012,
35.6 CONCLUSION
doi:10.1029/2012GL051247.
A fundamental requirement in studying the dynamics of hydrologic systems Cayley, A., “On the theory of the analytical forms called trees,” Philosophi-
and phenomena is the identification of the nature and extent of connections cal Magazine, Series IV, 13 (85): 172–176, 1857.
in space and time. Research over the past century on hydrologic connections, Clauset, A., M. E. J. Newman, and C. Moore, “Finding community struc-
through numerous scientific concepts and mathematical techniques, has ture in very large networks,” Physical Review E, 70 (6): P066111, 2004.
indeed resulted in significant progress. Nevertheless, our understanding of Clauset, A., C. Rohilla Shalizi, and M. E. J. Newman, “Power-law distribu-
such connections still remains far from adequate. To this end, the concepts of tion in empirical data,” SIAM Review, 51: 661–703, 2010.
network theory, and especially the recent concepts of complex networks, offer Coffman, D. M. and A. K. Turner, “Computer determination of the geom-
a whole different dimension and new avenues, for their holistic perspective of etry and topology of stream networks,” Water Resources Research, 7 (2): 419–
systems and problems encountered in hydrology and water resources. For 423, 1971.
instance, small-world, scale-free, community structure, and several other new Colizza, V., V. R. Banavar, A. Maritan, and A. Rinaldo, “Network structures
concepts are highly relevant for hydrologic systems, whose structure and from selection principles,” Physical Review Letters, 92 (19): 198701, 2004.
dynamics are a combination of order and randomness. Czuba, J. A. and E. Foufoula-Georgiou, “A network-based framework for
As the aforementioned review, albeit brief, suggests, the relevance of the identifying potential synchronizations and amplifications of sediment deliv-
concepts of complex networks in hydrology as well as the potential areas for ery in river basins,” Water Resources Research, 50: 3826–3851, 2014.
their applications are obvious; see also Sivakumar (2015). These new con- Czuba, J. A. and E. Foufoula-Georgiou, “Dynamic connectivity in a fluvial
cepts help obtain important information toward addressing a wide range of network for identifying hotspots of geomorphic change,” Water Resources
problems in hydrology. A few important ones among such problems are: Research, 51: 1401–1421, 2015.
interpolation and extrapolation of data, development of a catchment clas- da F. Costa, L., F. A. Rodriguez, G. Traviesco, and P. R. Villas Boas, “Char-
sification framework, and formulation of an integrated framework for acterization of complex networks: a survey of measurements,” Advances in
water planning and management (taking into account also other Earth Physics, 56 (1): 167–242, 2007.
systems and socioeconomic systems). As such problems are at the core of Dalin, C., S. Suweis, M. Konar, N. Hanasaki, and I. Rodriguez-Iturbe,
many of the most important recent and current initiatives and activities in “Modeling past and future structure of the global virtual water trade network,”
hydrology and water resources, such as predictions in ungaged basins, Geophysical Research Letters, 39: L24402, 2012, doi:10.1029/2012GL053871.
downscaling of outputs from global climate models for hydrologic analysis, Dalin, C., N. Hanasaki, H. Qui, D. L. Mauzerall, and I. Rodriguez-Iturbe,
and study of human-water interactions, network concepts can go a long “Water resources transfers through Chinese interprovincial and foreign food
way in hydrology. Indeed, as Sivakumar (2015) argued, network theory is a trade,” Proceedings of the National Academy of Science of the United States of
suitable tool for studying all types of connections and, hence, can serve as America, 111 (27): 9774–9779, 2014.
35_Singh_ch35_p35.1-35.10.indd 8 8/22/16 1:59 PM

REFERENCES 35-9
Danon, L., A. Díaz-Guilera, J. Duch, and A. Arenas, “Comparing commu- Liljeros, F., C. Edling, L. N. Amaral, H. E. Stanley, and Y. Åberg, “The web
nity structure identification,” Journal of Statistical Mechanics, 2005 (9): of human sexual contacts,” Nature, 411: 907–908, 2001.
P09008, 2005. Listing, J. B., Vorstudien zur Topologie, Vandenhoeck und Ruprecht, Göt-
Dooge, J. C. I., “Looking for hydrologic laws,” Water Resources Research, 22 tingen, 1848, pp. 811–875.
(9): 46S–58S, 1986. Lorenz, E. N., “Deterministic nonperiodic flow,” Journal of the Atmospheric
Dorogovtsev, S. N., J. F. F. Mendes, and A. N. Samukhin, “Structure of Sciences, 20 (2): 130–141, 1963.
growing networks with preferential linking,” Physical Review Letters, 85: Malik, N., B. Bookhagen, N. Marwan, and J. Kurths, “Analysis of spatial and
4633–4640, 2000. temporal extreme monsoonal rainfall over South Asia using complex net-
D’Odorico, P., F. Laio, and L. Ridolfi, “Does globalization of water reduce works,” Climate Dynamics, 39: 971–987, 2012.
societal resilience to drought?” Geophysical Research Letters, 37: L13403, Mandelbrot, B. B., The Fractal Geometry of Nature, W. H. Freeman, New
2010, doi:10.1029/2010GL041637. York, 1983.
D’Odorico, P., J. Carr, F. Laio, and L. Ridolfi, “Spatial organization and driv- Mark, D. M. and M. F. Goodchild, “Topologic model for drainage networks
ers of the virtual water trade: a community-structure analysis,” Environmental with lakes,” Water Resources Research, 18 (2): 275–280, 1982.
Research Letters, 7: 034007, 2012, doi:10.1088/1748–9326/7/3/034007. May, R. M., “Simple mathematical models with very complicated dynam-
Erdös, P. and A. Rényi, “On random graphs,” Publicationes Mathematicae, ics,” Nature, 261 (5560): 459–467, 1976.
6: 290–297, 1959. Mendelson, B., Introduction to Topology, Dover, New York, NY, 1990.
Erdös, P. and A. Rényi, “On the evolution of random graphs,” Publications Miguens, J. and J. Mendes, “Weighted and directed network on traveling
of the Mathematical Institute of the Hungarian Academy of Sciences 5: 17–61, patterns,” Biowire, 5151: 145–154, 2008.
1960. Milgram, S., “The small world problem,” Psychology Today, 2: 60–67, 1967.
Euler, L., “Solutio problematis ad geometriam situs pertinentis,” Commen- Milo, R., S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon,
tarii Academiae Scientiarum Imperialis Petropolitanae, 8: 128–140, 1741. “Network motifs: simple building blocks of complex networks,” Science, 298:
Fang, F., B. Sivakumar, and F. M. Woldemeskel, “Complex networks, com- 824–827, 2002.
munity structure, and catchment classification,” Environmental Modelling & Mishra, A. K. and P. Coulibaly, “Developments in hydrometric network
Software (submitted), 2015. design: a review,” Reviews of Geophysics, 47: RG2001, 2009,
Freeman, L. C., “Centrality in networks: I. Conceptual clarification,” Social doi:10.1029/2007RG000243.
Networks, 1: 215–239, 1979. Montanari, A., G. Young, H. H. G. Savenije, D. Hughes, T. Wagener, L. L.
Grayson, R. B. and G. Blöschl, Spatial Patterns in Catchment Hydrology: Ren, D. Koutsoyiannis, et al., “‘Panta Rhei—Everything Flows’: change in
Observations and Modeling, Cambridge University Press, Cambridge, 2000. hydrology and society—The IAHS Scientific Decade 2013–2022,” Hydrologi-
Gupta, V. K., I. Rodriguez-Iturbe, and E. F. Wood, “Scale problems in cal Sciences Journal, 58 (6): 1256–1275, 2013.
hydrology: runoff generation and basin response,” Water Science and Technol- Montoya, J. M. and R. V. Solé, “Small world patterns in food webs,” Journal
ogy Library Series, Springer, Dordrecht, The Netherlands, 1986. of Theoretical Biology, 214: 405–412, 2002.
Halverson, M. and S. Fleming, “Complex networks, streamflow, and hydro- Moon, J. W., “On the expected diameter of random channel networks,”
metric monitoring system design,” Hydrology and Earth System Sciences, 19: Water Resources Research, 16 (6): 1119–1120, 1980.
3301–3318, 2015. Newman, M. E. J., “The structure of scientific collaboration networks,”
Harary, F., Graph Theory, Addison-Wesley, Reading, Massachusetts, USA, Proceedings of the National Academy of Sciences of the United States of Amer-
1969. ica, 98: 404–409, 2001a.
Hausdorff, F., “Dimension und äußeres Maß,” Mathematische Annalen, 79: Newman, M. E. J., “Scientific collaboration networks: II. Shortest paths,
157–179, 1919. weighted networks, and centrality,” Physical Review E, 64: 016132, 2001b.
Horton, R. E., “Erosional development of streams and their drainage Newman, M. E. J., “Finding community structure in networks using the
basins: hydrophysical approach to quantitative morphology,” Geological Soci- eigenvectors of matrices,” Physical Review E, 74: 036104, 2006.
ety of America Bulletin, 56: 275–370, 1945. Newman, M. E. J. and M. Girvan, “Finding and evaluating community
Hrachowitz, M., H. H. G. Savenije, G. Blöschl, J. J. McDonnell, M. Sivapalan, structure in networks,” Physical Review E, 69: 026113, 2004.
J. W. Pomeroy, B. Arheimer, et al., “A decade of predictions in ungaged basins Niu, J., “Precipitation in the Pearl River basin, South China: scaling,
(PUB)—a review,” Hydrological Sciences Journal, 58 (6): 1198–1255, 2013. regional patterns, and influence of large-scale climate anomalies,” Stochastic
Jeong, H., S. Mason, A-L. Barabási, and Z. N. Oltvai, “Lethality and central- Environmental Research and Risk Assessment, 27 (5): 1253–1268, 2013.
ity in protein networks,” Nature, 411: 41–42, 2001. O’Bannon, C., J. Carr, D. A. Seekell, and P. D’Odorico, “Globalization of
Jha, S. K., H. Zhao, F. M. Woldemeskel, and B. Sivakumar, “Network theory agricultural pollution due to international trade,” Hydrology and Earth System
and spatial rainfall connections: an interpretation,” Journal of Hydrology 527: Sciences, 18: 503–510, 2014.
13–19, 2015. Paola, C., E. Foufoula-Georgiou, W. E. Dietrich, M. Hondzo, D. Mohrig,
Karinthy, F., Minden másképpen van (Everything is the Other Way), Athe- G. Parker, M. E. Power, et al., “Toward a unified science of the Earth’s sur-
neum Press, Budapest, Hungary, 1929. face: opportunities for synthesis among hydrology, geomorphology, geo-
Keller, E. F., “Revisiting ‘scale-free’ networks,” BioEssay, 27: 1060–1068, 2005. chemistry, and ecology” Water Resources Research, 42: W03S10, 2006,
Kim, D-H., J. D. Noh, and H. Jeong, “Scale-free trees: the skeletons of com- doi:10.1029/2005WR004336.
plex networks,” Physical Review E, 70: 046126, 2004. Peckham, S., V. Gupta, “A reformulation of Horton’s laws for large river
Kirkby, M. J., “Tests of the random network model, and its application to networks in terms of statistical self‐similarity,” Water Resources Research, 35:
basin hydrology,” Earth Surfaces Processes and Landforms, 1 (3): 197–212, 2763–2777, 1999.
1976. Poincaré, H., “Analysis situs,” Journal de l’École Polytechnique, 2 (1): 1–123,
Konar, M. and K. K. Caylor, “Virtual water trade and development in 1895.
Africa,” Hydrology and Earth System Sciences, 17: 3969–3982, 2013. Pons, P. and M. Latapy, “Computing communities in large networks using
Konar, M., C. Dalin, S. Suweis, N. Hanasaki, A. Rinaldo, and I. Rodriguez- random walks,” Journal of Graph Algorithms and Applications, 10 (2): 191–
Iturbe, “Water for food: the global virtual water trade network,” Water 218, 2006.
Resources Research, 47: W05520, 2011, doi:10.1029/2010WR010307. Radicchi, F., C. Castellano, F. Cecconi, V. Loreto, and D. Parisi, “Defining
Konar, M., Z. Hussein, N. Hanasaki, D. L. Mauzerall, and I. Rodriguez- and identifying communities in networks,” Proceedings of the National Acad-
Iturbe, “Virtual water trade flows and savings under climate change,” Hydrol- emy of Sciences of the United States of America, 101: 2658–2663, 2004.
ogy and Earth System Sciences, 17: 3219–3234, 2013. Raghavan, U. N., R. Albert, and S. Kumara, “Near linear time algorithm to
Krapivsky, P. L., S. Redner, and F. Leyvraz, “Connectivity of growing ran- detect community structures in large-scale networks,” Physical Review E, 76:
dom networks,” Physical Review Letters, 85: 4629–4632, 2000. 036106, 2007.
Latora, V. and M. Marchiori, “Efficient behavior of small-world networks,” Rinaldo, A., J. R. Banavar, and A. Maritan, “Trees, networks, and hydrol-
Physical Review Letters, 87 (19): 198701, 2001. ogy,” Water Resources Research, 42: W06D07, 2006, doi:10.1029/2005WR004108.
Leavitt, H. J., “Some effects of certain communication patterns on group Rinaldo, A., R. Rigon, J. R. Banavar, A. Maritan, and I. Rodriguez-Iturbe,
performance,” Journal of Abnormal and Social Psychology, 46: 38–50, 1951. “Evolution and selection of river networks: statics, dynamics, and complexity,”
Li, C., V. P. Singh, and A. K. Mishra, “Entropy theory-based criterion for Proceedings of the National Academy of Sciences of the United States of Amer-
hydrometric network evaluation and design: maximum information mini- ica, 111 (7): 2417–2424, 2014.
mum redundancy,” Water Resources Research, 48: W05521, 2012, Rodriguez‐Iturbe, I. and A. Rinaldo, Fractal River Networks: Chance and
doi:10.1029/2011WR011251. Self‐Organization, Cambridge University Press, New York, 1997.
35_Singh_ch35_p35.1-35.10.indd 9 8/22/16 1:59 PM

35-10 Network Theory
Salas, J. D., J. W. Delleur, V. Yevjevich, and W. L. Lane, Applied Modeling composition,” Earth Surface Processes and Landforms, 1: 219–233, 1976.
of Hydrologic Time Series, Water Resources Publications, Littleton, CO, 1995. Solomonoff, R. and A. Rapoport. “Connectivity of random nets,” Bulletin of
Scarsoglio, S., F. Laio, and L. Ridolfi, “Climate dynamics: a network-based Mathematical Biophysics, 13: 107–117, 1951.
approach for the analysis of global precipitation,” PloS One, 8 (8): e71129, Strahler, A. N., “Quantitative analysis of watershed geomorphology,” Eos
2013, doi:10.1371/journal.pone.0071129. Transactions American Geophysical Union, 38: 913–920, 1957.
Scheidegger, A. E., “On the topology of river nets,” Water Resources Suweis, S., M. Konar, C. Dalin, N. Hanasaki, A. Rinaldo, and I. Rodriguez-
Research, 3 (1): 103–106, 1967. Iturbe, “Structure and controls of the global virtual water trade network,”
Shreve, R. L., “Statistical law of stream numbers,” Journal of Geology, 74: Geophysical Research Letters, 38: L10403, 2011, doi:10.1029/2011GL046837.
17–37, 1966. Tamea, S., P. Allamano, J. Carr, P. Claps, F. Laio, and L. Ridolfi.” Local and
Shreve, R. L., “Infinite topologically random channel networks,” Journal of global perspectives on the virtual water trade,” Hydrology and Earth System
Geology, 75: 178–186, 1967. Sciences, 17: 1205–1215, 2013.
Shreve, R. L., “Stream lengths and basin areas in topologically random Tamea, S., J. A. Carr, F. Laio, and L. Ridolfi, “Drivers of the virtual water
channel networks,” Journal of Geology, 77: 397–414, 1969. trade,” Water Resources Research, 50: 17–28, 2014.
Sivakumar, B., “Dominant processes concept, model simplification and Tokunaga, E., “Consideration on the composition of drainage networks and
classification framework in catchment hydrology,” Stochastic Environmental their evolution,” Geographical Reports of Tokyo Metropolitan University, 13:
Research and Risk Assessment, 22 (6): 737–748, 2008. 1–27, 1978.
Sivakumar, B., “Global climate change and its impacts on water resources Tsonis, A. A. and P. J. Roebber, “The architecture of the climate network,”
planning and management: assessment and challenges,” Stochastic Environ- Physics A, 333: 497–504, 2004.
mental Research and Risk Assessment, 25 (4): 583–600, 2011a. Wasserman, S. and K. Faust, Social Network Analysis, Cambridge Univer-
Sivakumar, B., “Water crisis: from conflict to cooperation—an overview,” sity Press, Cambridge, 1994.
Hydrological Sciences Journal, 56 (4): 531–552, 2011b. Watts, D. J., Small Worlds: The Dynamics of Networks Between Order and
Sivakumar, B. “Networks: a generic theory for hydrology?” Stochastic Envi- Randomness, Princeton University Press, Princeton, 1999.
ronmental Research and Risk Assessment, 29: 761–771, 2015. Watts, D. J. and S. H. Strogatz, “Collective dynamics of ‘small-world’ net-
Sivakumar, B. and V. P. Singh, “Hydrologic system complexity and nonlin- works,” Nature, 393: 440–442, 1998.
ear dynamic concepts for a catchment classification framework,” Hydrology Werner, C., “Analysis of length distribution of drainage basin parameter,”
and Earth System Sciences, 16: 4119–4131, 2012. Water Resources Research, 18 (4): 997–1005, 1982.
Sivakumar, B. and Woldemeskel, F. M., “Complex networks for streamflow Yang, D., C. Li, H. Hu, Z. Lei, S. Yang, T. Kusuda, T. Koike, and K. Musiake,
dynamics,” Hydrology and Earth System Sciences, 18: 4565–4578, 2014. “Analysis of water resources variability in the Yellow River of China during
Sivakumar, B. and F. M. Woldemeskel, “A network-based analysis of spatial the last half century using historical data,” Water Resources Research, 40:
rainfall connections,” Environmental Modelling and Software, 69: 55–62, W06502, 2004, doi:10.1029/2003WR002763.
2015. Young, P. C. and M. Ratto, “A unified approach to environmental systems
Smart, J. S., “Use of topologic information in processing data for channel modeling,” Stochastic Environmental Research and Risk Assessment, 23:
networks,” Water Resources Research, 6 (3): 932–936, 1970. 1037–1057, 2009.
Smart, J. S., “The analysis of drainage network composition,” Earth Surface Zaliapin, I., F. Foufoula‐Georgiou, and M. Ghil, “Transport on river net-
Processes and Landforms, 3 (2): 129–170, 1978. works: a dynamic tree approach,” Journal of Geophysical Research, 115:
Smart, J. S. and C. Werner, “Applications of the random model of drainage F00A15, 2010, doi:10.1029/2009JF001281.
35_Singh_ch35_p35.1-35.10.indd 10 8/22/16 1:59 PM

Chapter 36
Hydroeconomic Analysis
BY
PETER BAUER-GOTTWEIN, NIELS RIEGELS, MANUEL PULIDO-VELAZQUEZ, JULIEN J. HAROU, XIMING CAI,
CLAUS DAVIDSEN, AND SILVIO J. PEREIRA-CARDENAL
ABSTRACT prior appropriation, can be evaluated using a simulation approach. In the

optimization approach, water is allocated to maximize net economic value, or,
Hydroeconomic analysis and modeling provides a consistent and quantitative
alternatively, to minimize the total costs of water allocation and water-supply
framework to assess the links between water resources systems and economic
curtailment, occurring at all use locations in the system. The management
activities related to water use, simultaneously modeling water supply and
problem is cast into an optimization framework where the objective is to
water demand. It supports water managers and decision makers in assessing
maximize net value and the constraints reflect various physical and institu-
trade-offs between different water uses, different geographic regions, and
tional characteristics of the system as well as environmental flow requirements.
various economic sectors and between the present and the future.
The applications of hydroeconomic analysis are numerous and diverse.
Hydroeconomic analysis provides consistent economic performance criteria
One class of applications involves planning problems. Often, the economic
for infrastructure development and institutional reform in water policies and
effects of planned changes (e.g., new hydraulic infrastructure, institutional
management organizations. This chapter presents an introduction to hydro-
reform) have to be assessed a priori. Hydroeconomic analysis and modeling
economic analysis and modeling, and reviews the state of the art in the field.
can quantify differences in costs and benefits between the with and without-
We review available economic water-valuation techniques and summarize the
project scenarios to analyze if a certain water project is deemed economically
main types of decision problems encountered in hydroeconomic analysis.
acceptable (cost-benefit analysis; e.g., Griffin, 1998). One example is the
Popular solution strategies for hydroeconomic problems are introduced.
design of a new surface water reservoir in a multireservoir system for water
Typical problems and issues in real-world applications of hydroeconomic
supply. In order to assess the benefits that the reservoir would generate (to be
analysis and modeling are discussed. The chapter ends with a discussion of
compared with the costs of the project), we need to estimate how this addi-
challenges, limitations, and future research directions in this field.
tional storage capacity would change water deliveries to the competing uses
during all periods. The policy may involve new operating rules for all reser-
36.1 INTRODUCTION voirs in the systems, which need to be defined during the design process. A
Around the globe, natural resources and ecosystems are coming under increas- hydroeconomic model with value-sensitive demand functions can be used to
ing pressure due to human development (Steffen et al., 2007). In many parts of assess the welfare changes in the system for developing the cost-benefit analy-
the world, demands for freshwater and pollution loads to receiving water bod- sis and for finding an economically efficient design.
ies are increasing (Postel et al., 1996; Rockstrom et al., 2009). Moreover, future Another application of hydroeconomic modeling and analysis is decision sup-
renewable freshwater supplies are uncertain due to the stochastic nature of port in water resources management. Operators are confronted with trade-offs
climate/weather and the insufficiently understood effects of anthropogenic in water allocation in the spatial dimension, in terms of intersectoral competi-
climate change (Oki and Kanae, 2006). In this context, water resources man- tion at a certain time, and in terms of intertemporal decisions affecting current
agement over all spatial and temporal scales has had to reconsider the demand and future benefits/costs. Economic efficiency in water management requires
fulfillment paradigm and to consider whether society might be better served considering opportunity costs in water-allocation decisions. Hydroeconomic
by redistributing water among sectors, for example, from irrigated agriculture analysis tools can reveal those opportunity costs, which can be used to inform
to aquatic ecosystems (Booker et al., 2012). Hydroeconomic analysis provides water-allocation decisions and in the design of efficient economic instruments,
a systematic and quantitative framework to assess the links between water such as water pricing (Pulido-Velazquez et al., 2013a). Dynamic optimization
resources systems and economic activities and reveal economic trade-offs algorithms can support the development of operating rules and scheduling
between different water users (Harou et al., 2009; Maass et al., 1962). Complex decisions and reveal intertemporal trade-offs to identify the long-term optimal
real-world water resources systems are mapped into simplified conceptual management. Hydroeconomic analysis has also been used to guide and inform
models that represent the topology and connectivity of the flow network as reallocation policies. For instance, in Europe, the EU Water Framework
well as the main water-storage options in the system. Water uses are classified Directive required reallocation of water to ecosystems. Reallocation can be
into a number of sectors and use locations, and are represented by economi- achieved using economic tools such as flexible water pricing (e.g., Riegels et al.,
cally respectful demands—true value-sensitive functions rather than historical 2013) and water markets (Erfani et al., 2014). The performance of these tools
use amounts. Hydroeconomic analysis and modeling has been carried out over can be partially assessed a priori using hydroeconomic modeling and analysis.
a range of spatial scales from single reservoir, irrigation district, and sub-basin, This chapter starts with an overview of available economic water-valuation
to basin-scale, regional, and even global models. methods. We proceed with a review of typical decision problems in hydroeco-
Hydrologic and economic submodels have been coupled in two principal nomic analysis and modeling, and of common solution strategies for these
ways: simulation and optimization approaches. While early studies tradition- problems. Subsequently, we discuss applications and implementation of
ally focused on costs, new era hydroeconomic studies generally use the net hydroeconomic models for informing real-world problems. We discuss limi-
present value (NPV), that is, the sum of discounted benefits net of costs, to tations of current methodologies and future research directions and provide
express economic performance of projects and operational strategies. In the conclusions.
simulation approach, NPV of observed or modeled water allocations is quanti-
fied. Simulation approaches are useful when the focus is on the evaluation of 36.2 ESTIMATING THE ECONOMIC VALUE OF WATER
water resource interventions such as new policies or infrastructure develop-
ments or combinations of these. For instance, the economic implications of A key task in hydroeconomic analysis is the determination of the water users’
widely used decision rules in water resources management, such as the rule of water demand function, especially the relationship between the economic
36-1
36_Singh_ch36_p36.1-36.10.indd 1 8/22/16 2:01 PM

36-2 Hydroeconomic Analysis
value of water and the amount of water delivered. A broad range of valuation in real-world watershed management situations. Decentralized modeling
approaches have been presented (e.g., Booker et al., 2012, provide a recent approaches may be partially free of those assumptions (Wallace et al., 2003;
and comprehensive overview). Applicability and suitability depend on the Yang et al., 2009).
specific water services being valued as well as on the purpose of the valuation For example, Yang et al. (2009) present a decentralized optimization
exercise and the data/information availability (Young, 2005). Microeconomic method that depicts a watershed as a multiagent system, which is expected to
theory differentiates two types of resource users: producers and consumers support stakeholder-driven, bottom-up watershed management. This method
(e.g., Varian, 2010). Producers use water to produce a product and their is applied to the Yellow River Basin for water allocation by incorporating a
objective is to maximize the profits obtained from selling the product on the more realistic institutional setting of the basin (Yang et al., 2012). Other
market. Consumers use water to derive utility and their objective is to maxi- examples in this modeling direction include Barreteau et al. (2004), Berger
mize the utility they obtain from consuming a mix of water and other goods, et al. (2006), and Ng et al. (2011). Britz et al. (2013) presented the MOPEC
subject to a budget constraint (Griffin, 2006). In hydroeconomic analysis, the method (multiple optimization problems with equilibrium constraints) to
challenges are to determine the relationship between the amount of water solve multiuser allocation problems under the assumption that each user
delivered and the amount of product produced or utility obtained (the pro- maximizes her individual welfare, while users are connected to each other
duction and utility functions) in order to estimate the marginal value of water. through flow network topology and market equilibria.
Numerical efforts have been put on the development of functional relation-
ships between water use output (e.g., crop yield, industrial product, profit, 36.3.1 Estimating Water Demand Functions
etc.) and water and other inputs (McKinney et al., 1999), and demand func- Most water uses can be defined as agriculture, industry, power generation,
tions, which connect water demand by sector or by individual water user and recreation, navigation, domestic, or ecosystem uses. In real-world applications,
their willingness to pay. The estimation of economic value is particularly there is typically insufficient information to parameterize water production
challenging and controversial for ecosystem water uses. Challenges exist with and utility functions for different uses. Different methods and approaches
the valuation of more comprehensive ecosystem services especially biodiver- have been proposed for the estimation of water demand functions as well as
sity and water quality, which is the concern of environmental economics (de point estimates of the marginal economic value of water, using varying levels
Groot et al., 2002). Furthermore, the lack of quantified relationships between of complexity and data requirements. Obtaining reliable data at the appropri-
ecological indicators and hydrologic variables often limits the incorporation ate spatial aggregation level is an important challenge in hydroeconomic
of ecological benefits into hydroeconomic analysis (e.g., Yang et al., 2008). analysis.
Water can be either an intermediate (production) or a final good (con-
36.3 WATER DEMAND FUNCTIONS sumption). When water is an input to a production process (e.g., irrigation,
hydropower generation, commercial, or industrial uses), water demand is
If production or utility functions are known for a given water user, the water derived from the demand for the final output and depends on the production
demand function can be derived by solving the profit maximization problem process. When water is a final good (e.g., residential or recreational water
for the producer and the constrained utility maximization problem for the uses), water provides direct utility to consumers; this utility translates into a
consumer (Griffin, 2006). Most common production functions show decreas- certain WTP for it, influenced by the preferences of consumers and their
ing returns to scale, that is, the marginal benefit decreases with increasing budget constraints. Different economic theories (of consumer’s and produc-
water allocation. At the point of operation that is optimal from the perspec- er’s demands) (Hanemann, 1999; Young, 2005) are applicable to each case,
tive of the individual user, the marginal net benefit (= the first derivative of affecting the selection of the valuation method.
the benefit with respect to water allocation) is zero. Producer and consumer For intermediate goods (derived demand, e.g., agriculture and industry), a
surplus are defined as the area between the user’s willingness to pay (WTP) widely used method for making point estimates of marginal economic values
function and the cost function. At water prices higher than the choke price, is residual imputation (RI; Young, 2005). The basic idea in RI is to subtract all
water users are assumed to switch to alternative supplies (e.g., bottled water). non–water-related costs from the observed gross benefit and to attribute the
Surplus is considered as the standard measure of welfare in microeconomic residual value to the water. The challenge with RI is how to deal with
theory. non–water-production factors that contribute to the production but are not
Figure 36.1 illustrates these basic concepts of water resources economics. easily quantifiable, such as, for instance, entrepreneurial skill and experience.
Conventional water resources management models use a centralized approach If such factors are ignored in RI, the resulting value of water will be inflated.
to explore the system-wide optimal solution, assuming that (1) there is a top- Estimates of marginal values of water using the RI approach also depend on
down management process in which all watershed stakeholders completely the time horizon of the planning problem. For short-term planning problems,
obey a “super brain,” (2) complete information exchange exists among the marginal value may be higher because producers do not have the ability to
stakeholders in different locations and production sectors, and (3) perfect eco- shift sunk investments in machinery or other physical capital (so-called “sunk
nomic efficiency is realized, that is, the marginal welfare of natural/raw/source costs”) to other production processes. In seasonal irrigation planning, even
water use is identical for all agents. These assumptions are usually not fulfilled costs of seeds and fertilizers may be considered sunk costs if marginal value
is being estimated, for example, in the middle of a crop growth cycle. In the
long term, however, all inputs can be reallocated to other uses and should be
subtracted from the residual value used to estimate marginal value of water.
Water price Econometric and especially mathematical programming approaches have
been widely applied to define demand functions for water in irrigation
(Booker et al., 2012; Young, 2005). Irrigation water demands are usually
Choke price simulated in hydroeconomic models using exogenous linear or quadratic
equations, relating water application to economic benefits (e.g., Pulido-
Velazquez et al., 2007a; Pulido-Velázquez et al., 2006). In other cases, crop
Consumers’ surplus + Market value = Gross benefits yield functions are explicitly included in a river basin hydroeconomic model
(e.g. Cai et al., 2003b) so that the marginal values are obtained from multiply-
ing marginal physical productivities by crop prices.
Howitt (1995) introduced a calibration method for estimating agricultural
water demand functions, positive mathematical programming (PMP), which
has since become popular in hydroeconomic modeling studies. The central
A idea is that observed allocations of land and water represent profit-maximiz-
ing behavior by farmers. Shadow prices on observed land and water use are
used to parameterize production functions, with the assumption that these
Demand function shadow prices incorporate unobserved information, such as soil characteris-
tics, consideration of risk and uncertainty, and farmer expertise. The
Water demand approach can be implemented using cross-sectional surveys of land use, water
use, crop yields, prices, and input costs, making it suitable for use with
B datasets that are commonly available in hydroeconomic studies. The PMP
Figure 36.1 Water demand function for a water user, based on Harou et al. (2009). approach estimates a production function, not a demand function, although
As the water price increases, the users demand for water decreases. For a situation with it can be easily used to estimate a water demand function by solving the
a water price equal to A (e.g., $/m3), the water demand (e.g., m3) will be B. profit maximization problem at different water price increments.
36_Singh_ch36_p36.1-36.10.indd 2 8/22/16 2:01 PM

Considerations in the design of hydroeconomic analysis studies 36-3
For determining water demand for final uses (e.g., urban demands), econo- change the power price on the market (Pereira-Cardenal et al., 2014). For
metric approaches are dominant (Dalhuisen, 2003; Renzetti, 2002). For urban system-scale applications, it is important to internalize output prices and
demands, a simplified but popular method used to estimate water demand other exogenous variables to achieve a more realistic simulation of the effects
functions is the point expansion method. In this approach a pair of observed occurring throughout the entire economy. Several attempts to couple water
water price and water use is taken as one point on the demand function. resources models with general equilibrium models of the economy have been
Extrapolation to different allocations or different user prices is possible if reported (e.g., Luckmann et al., 2014).
assumptions are adopted about the slope of the demand function or the elas-
ticity of the demand (Griffin, 2006). This method is applicable for any sector
where a price-quantity pair can be determined for anchoring the demand 36.4 CONSIDERATIONS IN THE DESIGN OF
function. HYDROECONOMIC ANALYSIS STUDIES
When estimating water demand functions or WTP, it is important to dif- A hydroeconomic analysis study generally (although not always) combines a
ferentiate between “at site” and the “at source” value of water. For example, if hydrologic or river basin model with some form of economic analysis. In
a demand function is estimated from observed consumer water use using the some cases, economic water demand functions are integrated directly into the
point expansion method, this reflects demand for water that has already been hydrologic model and optimization is used to allocate water across space and
abstracted from natural waters, treated, and delivered to the point of use (i.e., time in order to maximize water values or other criteria. Typical water
the at-site value). However, in an investigation of trade-offs between water sources include surface water, groundwater, reclaimed water, desalinated
uses in a river basin, the value of water at the point of abstraction is the rele- water, or transfer water from other basins. Each water source is characterized
vant value for comparison to competing uses (the at-source value). The at- by its availability, storage capability, quality, and pumping and conveyance
source value should include any producer-side benefits to the supplying water cost. The allocation problem is often complicated further by the presence of
utility (Young, 2005). However, the nonprofit status of most utilities implies multiple and contradicting management objectives. While economic effi-
that producer-side benefits are often transferred to consumers in some fash- ciency can be an important objective in water resources management, other
ion (especially by underpricing). objectives, such as reliability, equity, and sustainability may also be important.
The estimation of economic water demand functions for ecosystem uses is
a new area for research and professional practice. In this context, the emerg-
ing “ecosystems services” paradigm can be useful for defining what are gener- 36.4.1 Multiple User Allocation
ally understood as “ecosystem water uses” (for an overview of the ecosystem Figure 36.2 illustrates a representative multiuser allocation problem. The fig-
services concept, see Braat and de Groot, 2012). The ecosystem services para- ure shows three water use locations, each with water demand and each pro-
digm argues that ecosystems provide a number of benefits to humans and that ducing net benefit (NB), aligned along a river reach with supply Q and an
the best way to ensure efficient conservation of ecosystems is to measure the in-stream flow requirement R at the downstream end of the reach. If the
contributions of ecosystems to human welfare. supply Q is insufficient to meet all three demands and the in-stream flow
A number of methods have been applied to estimating ecosystem water use requirement, then the diversions x must be curtailed to be less than the
values. These methods can be broadly grouped into two categories: revealed demands (or the amount of water allocated to the in-stream flow requirement
preference methods and stated preference methods. Revealed preference must be reduced). Hydroeconomic analysis is often concerned with the ques-
methods estimate ecosystem values from observations of market transactions tion of what these diversions or flow reductions should be.
involving related private goods, such as prices for homes located near ecosys- If the net benefits at the use locations are linear functions of the water
tem resources, or costs associated with travel to visit natural areas. Stated allocations and if only one single time period is considered, this optimization
preference methods estimate ecosystem values by asking people what they problem becomes a simple linear program:
would be willing to pay for proposed or hypothetical changes in ecosystem
service provision. A useful overview of the uses of stated and revealed prefer-
ence methods in water resource economics is provided in Young (2005). max(WV1 ⋅ x1 + WV2 ⋅ x 2 + WV3 ⋅ x 3 )
Estimating the value of ecosystem water uses is only one of two steps subject to
needed to estimate a water demand function. It is also necessary to develop a x1 + x1 + x 3 ≤ Q − R
relationship between water supply and the associated level of ecosystem ser-
vice provision. For example, if river restoration activities are under consider- xi ≤ Di
ation in order to provide recreational and aesthetic benefits, it may be
necessary to estimate a relationship between the amount of flow in the river
where xi are the water allocations to the three users, WVi are the water values
and the amount of benefits provided. Estimation of this relationship may be
at the three use locations, Di are the demands, Q is the available river flow, and
complicated by the fact that, in the case of river flows, the ecosystem service
R is an in-stream water requirement.
level depends on many features of the flow regime, including the timing of
In real-world applications, of course, river network topology and connec-
flows and the amount of variability in the flow regime (for a discussion of the
tivity are typically much more complex. Moreover, as discussed earlier, uses
hydrological regime, see Arthington et al., 2006). A number of operational
are not always rival and return flows generated by one user can be reused
approaches to link measures of the river flow regime to levels of ecosystem
further downstream in the system (Griffin, 2006). The economic theory and
service provision have been proposed (e.g., Poff et al., 2010; Smakhtin et al.,
modeling of multiple users connected by their withdrawals, return flows, and
2004). Because estimation of ecosystem values and modeling of ecosystem
instream flow demands is presented and discussed in Griffin and Hsu (1993).
service provision are complicated and uncertain, almost all hydroeconomic
Net benefits involve a marginal value of water that increases with lower deliv-
studies represent demand for ecosystem water uses as hard constraints (for
eries, which is better represented by nonlinear economic demand functions.
an example of the use of an economic demand function for recreational
Water-quality requirements by various users and reactive transport processes
ecosystem service in a hydroeconomic optimization study, see Ward and
in the river network may introduce additional constraints into the allocation
Pulido-Velazquez, 2008).
problem.
Because of generalized lack of data, scaling problems and methodological
Hydrologic models are mostly built for simulation experiments, while eco-
issues, economic values of water are always uncertain. At the same time, water
nomic models are often of the optimization type. Compartment models
values are among the most sensitive parameters in hydroeconomic analysis
couple the simulation and optimization models usually through genetic algo-
and modeling (Riegels et al., 2011). Rigorous uncertainty analysis and analysis
rithms (GA; Nicklow et al., 2010), while holistic models combine physical
of robustness of results is presently not standard practice in hydroeconomic
analysis and modeling. Some hydroeconomic optimization studies include
the uncertainty of economic performance into the objective function, penal-
izing highly variable performance with a risk aversion term (e.g., Blanco-
Gutiérrez et al., 2013). R
In most hydroeconomic modeling studies, economic effects are quantified Q X1
as surplus changes based on prespecified water demand functions. In reality, X2 X3
the perturbations generated by water management decisions will propagate NB1
further into the economy and may change output prices, labor costs, and NB2 NB3
other parameters used in the derivation of the demand functions. For
instance, if hydropower has a significant share of total power supply in a given Figure 36.2 Basic river basin allocation problem. Q is the inflow to the system, xi are
the water allocations to the individual users, NBi are the net benefits obtained by the
system, then the scheduling decisions of the hydropower operators can users, and R is an instream flow requirement.
36_Singh_ch36_p36.1-36.10.indd 3 8/22/16 2:01 PM

relationships, economic constraints, and objectives in a so-called simulation- (Rani and Moreira, 2009). One of the most promising methods is stochastic
embedded optimization model. Such a model is often complicated due to dual dynamic programming (SDDP; Pereira and Pinto, 1991), which has been
nonlinearity and discontinuity involved in the simulation module. Both types successfully applied to optimize multi-purpose reservoir systems (e.g., Goor
of models meet computational difficulties when they are applied to large et al., 2011; Tilmant and Kelman, 2007). An alternative to dynamic program-
water resources systems. A hybrid GA and linear programming (LP) ming methods are heuristic optimization techniques, which do not guarantee
approach was developed to find approximate global solutions or feasible solu- the optimal solution, but provide a feasible solution often close to the optimal
tions for large hydroeconomic models formulated as nonlinear programming one. A popular class of such methods are evolutionary algorithms, which have
models with high nonlinearity and nonconvexity (Cai et al., 2001a); a “piece- been used extensively in water resource management (Cai et al., 2001a;
by-piece” approach (Cai et al., 2001b) was used to solve models that can be Nicklow et al., 2010).
naturally decomposed into components (e.g., hydrologic and economic com-
ponents). Network-flow optimization was used to solve other large hydroeco- 36.4.3 Conjunctive Use of Multiple Water Resources
nomic models (e.g., Draper et al., 2003). In many allocation problems, managers have access to more than one water
source to satisfy demands. Sources may include surface water, groundwater,
36.4.2 Intertemporal Allocation reclaimed water, desalinated water, and transfer water from other basins (e.g.,
When allocation decisions today have an impact on future water availability, Luckmann et al., 2014). Often, water can also be supplied from multiple res-
the allocation problem must be handled using a dynamic framework. Many ervoirs which are connected and managed jointly as a multireservoir system
river system allocation problems are characterized by intertemporal trade- (Labadie, 2004). In systems with important groundwater components, the
offs because of the possibility of storing water for future use, including reser- model should include surface and ground water systems, and stream-aquifer
voir and groundwater storage (Fig. 36.3, Gisser and Sanchez, 1980; Loucks interaction. Otherwise, significant economic opportunities, opportunity
and van Beek, 2005). In the case of surface reservoirs, hedging operating rules costs, and externalities may be ignored (e.g., Pulido-Velazquez et al., 2008).
are usually designed to minimize the risk of high losses due to large water Flexible management of conjunctive use facilities under flexible water alloca-
deficits in severe future drought periods (Bayazit and Unal, 1990; You and tion can generate substantial economic benefits in water resource systems
Cai, 2008a; You and Cai, 2008b). In finance and economics, the term hedging operation, which can be assessed through hydroeconomic models
is used to describe an action or instrument that reduces the exposure of an (e.g., Harou and Lund, 2008; Pulido-Velazquez et al., 2004, 2007b).
individual or organization to any significant losses/gains due to uncertain If different water sources have to be managed jointly, managers have to
future developments. In irrigation agriculture, water is allocated over the consider differences in fixed and variable costs between the various water
course of the entire growing season, while harvest and associated economic sources. Such costs may depend on the state of the system (e.g., depth to the
benefits materialize at the end of the growing season. Such “delayed yields” in groundwater table; Riegels et al., 2013), on the rate of abstraction (e.g., regional
irrigated agriculture (Tilmant et al., 2008), storage in water supply systems, and local drawdown in groundwater pumping) and on the water quality
time delays in network routing and water quality effects can also contribute (e.g., desalination; Karagiannis and Soldatos, 2008). Often, cost functions are
to intertemporal trade-offs. complex and introduce nonlinearity and nonconvexity into the optimization
One important issue is how to discount costs and benefits occurring in the problems, which requires more advanced and CPU-intensive optimization
future in order to trade them off against costs and benefits in the present. A algorithms (e.g., Pulido-Velázquez et al., 2006, Davidsen et al., 2016).
large body of literature discusses discounting and the choice of appropriate
discount rates in environmental studies (e.g., Heal, 1998). 36.4.4 Conflicting Objectives and Pareto Fronts
Hydroeconomic analysis studies that aim to provide advice about In hydroeconomic optimization, managers and analysts often encounter mul-
intertemporal allocation in real-world operations, such as reservoir operating tiple and contradicting management objectives (Xevi and Khan, 2005).
rules, should address the role of uncertainty. Future water availability is Contradicting objectives may be related to competing users which can be
uncertain, particularly in the case of surface water resources, and hydroeco- geographically defined (transboundary river systems, Fig. 36.4) or defined in
nomic analysis of intertemporal trade-offs should consider the value of hedg- terms of water use (e.g., irrigation agriculture vs. aquatic ecosystems). A tra-
ing against uncertainty. Hydroeconomic studies to support real-world ditional hydroeconomic approach is to value all benefits and costs so they can
operations are often formulated as stochastic dynamic optimization prob- be aggregated into one commensurate monetized objective function for an
lems. A starting point for this type of problems is stochastic dynamic pro- optimization model.
gramming (SDP; Stedinger et al., 1984, Loucks and van Beek, 2005), although However, sometimes, objectives cannot be easily expressed in monetary
applications are limited by the well-known curse of dimensionality in com- units and the requirement to monetize can weaken the analysis or underap-
plex water resources systems. A widely used variant of SDP known as the preciate stakeholder interests. Such criteria may include reliability metrics,
water value method, determines and archives the shadow prices for any measures of the spatial and sectorial distribution of impacts, and sustainabil-
combination of time and system state (Stage and Larsson, 1961, Wolfgang et ity metrics. The inherently multiobjective nature of most water resources
al., 2009). The water values can subsequently be used as decision rules in a management problems has led to further development of multiobjective
more highly resolved simulation model (e.g., Wolfgang et al., 2009; Davidsen optimization approaches (Maier et al., 2014; Nicklow et al., 2010; Reed et al.,
et al., 2015a). The SDP method has many advantages: it can handle discon- 2013). Such approaches do not provide one single best solution, but a whole
tinuous and nonlinear functions; it provides the optimal solution from any set of Pareto-optimal or noninferior solutions, the so-called Pareto frontier. A
given stage, and can handle a large amount of stages. However, computing Pareto-optimal policy is characterized by the fact that there is no alternative
and memory requirements increase exponentially with the number of states, policy that is better in terms of at least one objective and not worse in terms
which limits the amount of states that can be included in the problem. Hence, of all other objectives. A selection between two Pareto-optimal policies will
only three or four reservoirs can be optimized at a time (Labadie, 2004). always have to involve higher-level information, subjective preferences, and/
Several extensions to DP have been proposed to overcome this challenge or political choice.
Country A Country B
R
R
Q X1
X2 X3 Q X1
NB1 X2 X3
NB2 NB3 NB1
g1 NB2 NB3
g2 g3 g1
g2 g3
q Groundwater aquifer
q Groundwater aquifer
Figure 36.3 River basin allocation problem with surface and groundwater storage. Q
is the surface water inflow to the system, q is the groundwater recharge, xi are the sur- Figure 36.4 Transboundary river basin allocation problem. Q is the surface water
face water allocations to the individual users, gi the groundwater allocations to the indi- inflow to the system, q is the groundwater recharge, xi are the surface water allocations
vidual users, NBi are the net benefits obtained by the users, and R is an instream flow to the individual users, gi are the groundwater allocations to the individual users, NBi
requirement. are the net benefits obtained by the users, and R is an instream flow requirement.
36_Singh_ch36_p36.1-36.10.indd 4 8/22/16 2:01 PM

Applications and implementation of hydroeconomic analysis for management and decision support 36-5
Country A Country B “holistic modeling” (Braat and van Lierop, 1987). The compartment approach
treats hydrologic and economic components as separate submodels, which
R interact by data exchange. In holistic modeling, the submodels are combined
Q into a single consistent model, which is typically solved in its entirety, and
X1 r1 r3 information between components is transferred endogenously (Cai et al.,
X2 r2 X3
NB1 2003a; Draper et al., 2003). Holistic models depict the coupled human-natural
Water NB2 NB3 interrelationships and mimic the impact of driving forces and feedbacks from
treatment g1 the environment (Cai et al., 2006).
g2 g3
Data requirements for hydroeconomic analysis refer to the type of data
q Groundwater aquifer and the level of aggregation over spatial and temporal scales in order to
develop either a holistic model or a compartment model. The divergent
Figure 36.5 Transboundary river basin allocation problem with water quality deci- numerical scales of these items and uncertainties will raise difficulties in
sion variables. Q is the surface water inflow to the system, q is the groundwater judging the reliability of model outputs. Especially in calculus-based nonlin-
recharge, xi are the surface water allocations to the individual users, gi the groundwater ear programming (NLP) models, water resources and economic data need to
allocations to the individual users, ri the polluted return flows to the river, NBi are the use appropriate units so that their numerical magnitudes appear similar (Cai
net benefits obtained by the users, and R is an instream flow requirement. Pollution can
be removed before or after use at each of the use locations.
et al., 2006). The correlations between some economic variables and hydro-
logic ones can make uncertainty analysis with a hydroeconomic model com-
plex.
Both lumped and distributed models are applied to water resources and
36.4.5 Water Quality economic modeling. Hydroeconomic modeling requires an appropriate
Another challenge for the hydroeconomic modeling and analysis community matching of the spatial aggregation with both components. The spatial aggre-
is joint management of water quantity and quality. Different water users and gation of water resources modeling needs to facilitate economic analysis,
ecosystems have different water quality requirements. Water-quality manage- while, at the same time, it has to be effective enough in generating impacts to
ment can operate with both pretreatment (water is purified prior to delivery hydrologic system operation and water allocation. The ideal match-up would
to the users) and posttreatment (water is purified prior to discharge into the provide effective information transfer between the two components, while
environment) as illustrated in Fig. 36.5. The optimal policy will depend on economizing on the complexity of the modeling framework and reducing
complex interactions of different water quality processes (advective and disaggregation error. One typical challenge is that the boundaries of the eco-
persive transport, reactions, sedimentation, etc.), water flow in the river (dilu- nomic systems of a specific resources problem may not a priori be the same
tion), available treatment technologies and water-quality requirements for as those of the hydrologic systems. Hydrologic simulation usually adopts a
human use and ecosystems. Hydroeconomic models have commonly targeted top-down structure with the spatial scale starting from the river basin and
pure water quality problems (e.g., Cools et al., 2011; Hasler et al., 2014), while going down to the crop land. Water balance is examined crossing multiple
salinization problems (e.g., Cai et al., 2003a) have dominated the joint water- scales, for example, at the basin level deciding water allocation among various
quantity and water-quality-management studies. There are only a few hydro- demand sites, at the demand site level determining water allocation to crop
economic modeling and optimization studies on other basin-scale lands; and finally at the crop land level determining irrigation schedule and
water-quality problems [e.g., biological oxygen demand, nutrients and heavy crop acreage. Meanwhile economic modeling usually adopts a “bottom-up”
metals (e.g., de Moraes et al., 2010; Davidson et al., 2015b)]. One class of structure, for example, with the spatial scale starting from crop lands and
approaches focuses on stream water quality (e.g., Cardwell and Ellis, 1993; going up to the whole basin; and profit from irrigation water use is calculated
Karamouz et al., 2010). Peña-Haro et al. (2009) presented a hydroeconomic at the crop land scale, and is aggregated to the demand site and the basin to
modeling framework for optimal management of groundwater nitrate pollu- optimize basin-level water use profit. The top-down and bottom-up struc-
tion from agriculture. Ejaz and Peralta (1995) used a response matrix optimi- tures should be coupled to allow interactions between the two processes
zation-simulation approach to maximize waste loads from a sewage treatment crossing all spatial levels (Cai, 2008).
plant while complying with water-quality constraints (dissolved oxygen, TDS, Economic analysis generally uses larger time intervals (seasonal or annual)
chlorophyll-a, and nutrients). Other studies include also upstream reservoir and a longer time horizon (e.g., for long-term forecasts) than hydrologic models.
releases in the decision problem (e.g., de Azevedo et al., 2000; Hayes et al., This is because some production activities (for instance, crop production) are
1998). associated with certain time periods. Moreover, economic damage caused by
Inclusion of a reservoir in the water-quality problem (illustrated in Fig. some environmental consequences of production activities, for example, crop
36.5) couples the water-quality decisions in time, that is, pollutants remain or damage caused by soil salinity accumulation (Cai et al., 2003a), need a long-
slowly degrade in the reservoir. Castelletti et al. (2014) used a reinforcement term projection. Other economic activities, such as water leasing usually
learning algorithm coupled with a hydrodynamic-ecological model to find involve relatively short time intervals, such as months. For hydrologic simula-
Pareto-optimal operation policies for a reservoir equipped with a multipur- tion, the time interval should be small enough to reflect real-world processes
pose selective withdrawal system (SWS) under both water quantity and qual- and capture the transition change of physical systems, which will affect eco-
ity constraints. The water-quality constraints were formulated as multiple nomic costs and benefits. The time horizon should be long to reflect the
operating objectives to avoid algal bloom, sedimentation, and excessive water regional hydroclimatic cycle and economic and environmental consequences
temperature. Ahmadi et al. (2012) used a fuzzy multiobjective GA-based resulting from flow regulation and water and land uses. Special model formu-
approach to maximize upstream agricultural production and mitigate unem- lation is often needed to match the temporal scales particularly in holistic
ployment impacts of land use changes while complying with selected water hydroeconomic models (see examples in Cai, 2008; Foster et al., 2014).
quality and quantity constraints. Davidsen et al. (2015b) used the water value Most applications of hydroeconomic modeling and analysis fall into three
method to find optional water allocation and water-treatment at basin scale broad categories: planning problems, operational problems, and assessments
water downstream water quality restraints. A hybrid GA-LP implementation of water policies.
(Cai et al., 2001) was used to handle nonlinearity and nonconvexity in the
objective function. A GA-based approach was also used in the context of a 36.5.1 Planning Problems
simplified SDP framework to resolve water quantity and quality conflicts Hydroeconomic models can be used to evaluate the impact of future changes
summarized in a Nash bargaining setup (Karamouz et al., 2008; Kerachian in supply and demand as can occur due to infrastructure investment or imple-
and Karamouz, 2007). mentation of demand management (water conservation) schemes.
Hydroeconomic models can inform cost-benefit analyses (Griffin, 1998)
valuing the sectoral economic impacts of new assets such as reservoirs, trans-
fers, or other hydraulic infrastructure. Whereas traditionally cost-benefit
36.5 APPLICATIONS AND IMPLEMENTATION OF analysis is used to evaluate single assets, hydroeconomic modeling is a sys-
HYDROECONOMIC ANALYSIS FOR MANAGEMENT AND tems approach (also referred to as “integrated assessment”) seeking to evalu-
DECISION SUPPORT
ate how changes propagate through spatially distributed engineered-natural
Implementation of hydroeconomic models faces many difficulties in terms of water systems. With/without studies can quantify the benefits and/or oppor-
model formulation, temporal and spatial scale issues, model calibration, solu- tunity costs to different sectors of one or more interventions. For example,
tion, and result interpretation as well as extensive data requirements in mul- Pulido-Velázquez et al. (2006) and Harou and Lund (2008) quantify the
tiple areas (Cai, 2008). In general, there are two approaches to combining opportunity costs of infrastructure bottlenecks using shadow values (appro-
hydrologic and economic components—“compartment modeling” and priate for water systems with water markets).
36_Singh_ch36_p36.1-36.10.indd 5 8/22/16 2:01 PM

Capacity-expansion problems are a special class of planning problems water resources management. Several studies have had a considerable impact
consisting of phasing investments over time. In capacity-expansion problems, on the water-management debate through the simulation of different assump-
we need to select, size, and schedule options to meet the predicted demands tions about the economic use of water (e.g., Cai et al., 2003b; Draper et al.,
at maximum benefit/minimum cost over a multiyear period (O’Laoghaire 2003). However, few, if any, hydroeconomic models formulate, test, and make
and Himmelblau, 1974). For example, a least-cost deterministic capacity decisions about economic policy. Reasons include: the perception that water
expansion optimization model was formulated by Padula et al. (2013) for siz- is essential for many of its uses and therefore not appropriate for economic
ing and scheduling of water supply and demand options in a large regional analysis; lack of conceptual and/or technical clarity about how to simulate
water supply system in South East England. economic policies in a hydroeconomic modeling framework; and, as with
other types of hydrologic and water resources planning models, the percep-
36.5.2 Operational Problems tion that model results may be too uncertain to formulate serious policy
Operation of an existing water resources system requires managers to take conclusions. One source of uncertainty is coefficients and parameters in the
day-to-day decisions. Decisions today typically have an impact on what is hydrologic as well as the economic compartment of hydroeconomic models
possible in the future and operational management must therefore balance that are unobserved and cannot be verified empirically. Prompted by chal-
present and future performance. The most important operational problem in lenges associated with model uncertainty, a literature has emerged describing
water resources management is reservoir scheduling. Many studies address a variety of systematic approaches to support decision making in planning
optimal management of single or multipurpose reservoirs and reservoir sys- conditions characterized by scientific uncertainty. Two approaches that have
tems (see review by Labadie, 2004). Typical objectives in reservoir manage- found applications in water resources planning are the Info-Gap approach
ment include water demand fulfillment, hydropower production, flood (introduced by Ben-Haim, 2004; recent water-management applications
protection, and recreational value (Loucks and van Beek, 2005). A hydroeco- include Hine and Hall, 2010, and Riegels et al., 2011) and Robust Decision
nomic modeling and optimization approach reveals the shadow price of Making (introduced by Lempert et al., 2006; applications include Groves and
stored water in the reservoir as a function of time, present inflow, and reser- Lempert, 2007, and Lempert and Groves, 2010). The general goal of the
voir level (e.g., Davidsen et al., 2015a; Stage and Larsson, 1961). Shadow approaches is to use ensemble computer simulation to characterize a range of
prices can subsequently be used as decision rules in operational management. plausible uncertain futures and use this information to design water resource
In the power systems literature, many studies have addressed hydropower plans that perform reasonably well across this range of future conditions.
scheduling applications (e.g., Wolfgang et al., 2009). From the power systems Matrosov et al. (2013) provide a recent comparison of the Info-Gap and
perspective, hydropower is a unique power source because of its low variable Robust Decision Making approaches.
cost, its dispatch flexibility, and low ramp on/ramp off costs. These character- Although hydroeconomic optimization models have the capacity to simu-
istics of hydropower become increasingly important with the increased pen- late economic policies directly, the level of aggregation and simplification
etration of highly intermittent renewable power sources in most power used in some optimization schemes also limits the uptake of these models for
systems. Groundwater management also involves operational real-time deci- real-world policy making. In many hydroeconomic models, network topolo-
sions, for instance, to minimize groundwater pumping cost, to minimize gy is simplified, storage facilities are aggregated, process descriptions are
energy footprint of delivered groundwater, and to minimize the risk of pollu- lumped, and the system representation in the model leaves out aspects of the
tion (e.g., Ahlfeld and Laverty, 2011; Bauer-Gottwein et al., 2016; Fowler et al., physical system. If model results contain aggregated variables, such as aggre-
2008; Hansen et al., 2012; Mayer et al., 2002). gated fluxes and storages, decision makers and managers may have difficulty
understanding simulated quantities. In a deterministic optimization approach,
36.5.3 Assessment of Water Policies where the water balance is solved through a solution of a single optimization
New water policies typically change the way in which water is allocated or problem that assumes perfect foresight of future water availability, the net-
controlled. Examples include water rights or water licensing changes, water work representation can be more complex; however, model credibility suffers
markets, water pricing, and water allocation. Hydroeconomic models help from unrealistic operations resulting from the perfect foresight assumption.
assess new water policies by quantifying how policies lead to economic out- Stochastic optimization offers the possibility to simulate more realistic opera-
comes at the system-wide level. tions, but typically requires a highly simplified system representation because
An important problem in hydroeconomic analysis is to quantitatively of computational challenges. Other optimization techniques (e.g., genetic
specify policy instruments that provide incentives to steer water- allocation algorithms, SDDP) together with the increasing availability of computational
patterns into a more desirable direction. Changing water pricing schemes is resources may help combine a detailed system representation with realistic
one approach to reallocating water resources (Pulido-Velazquez et al., operations in optimization models. Combined simulation-optimization
2013a; Riegels et al., 2013). Hydroeconomic modeling and analysis can approaches, where a simplified optimization model is used to formulate poli-
inform innovative pricing policies by providing information on the oppor- cies and operating rules that are then tested in a detailed simulation model
tunity cost of water at any point in space and time (Pulido-Velazquez et al., (e.g., Hurford et al., 2014; Pulido-Velazquez et al., 2013b) are another
2013a). approach to reconciling optimization with the need for realistic detail, and
Some countries have introduced markets for water or mechanisms for offer the advantage of using simulation models that are trusted by decision
water right exchange through markets. If one assumes the economic value of makers (for a related argument for coupled optimization-simulation, see
water is the exclusive driver to water exchanges, hydroeconomic models can Labadie, 2004).
be used to estimate the benefits and water allocations that would occur if a Unlike in other areas of modeling, it is not straightforward to develop base-
water market functioned without transaction costs. Examples of water mar- line simulations representing observed conditions in hydroeconomic model-
kets simulated with hydroeconomic models include Chile (Bauer, 1997; Cai et ing. This is not always necessary when models are used, for example, to
al., 2006), California (Draper et al., 2003; Jenkins et al., 2004), Australia, and identify system constraints limiting the efficient allocation of water (e.g.,
the United Kingdom (Erfani et al., 2014). Although the market mechanism Draper et al., 2003; Pulido-Velázquez et al., 2006). Calibrating hydroeconomic
ensures efficient allocation of water under ideal conditions, real-world water models is challenging because of the large number of parameters involved from
markets are often dysfunctional and inactive because of high transaction costs different components and also because of interdependence among parameters
and market failures due to the public good character of water (Rosegrant and to be calibrated. In addition, time series of economic data may not be available
Binswanger, 1994). Real-world water markets also tend to be limited because to the same extent as hydrological variables, such as streamflow and rainfall.
quantified and enforced water rights are often absent or transferability of such Changes in hydrologic and economic parameters affect not only hydrologic
water rights is illegal. Recent advances in hydroeconomic optimization of model outputs, but also the economic outputs, and vice versa. The PMP meth-
flow path models (Cheng et al., 2009; Erfani et al., 2013) enable tracking of odology described earlier appears to have some promise as a method for cali-
individual transactions of water in the optimal allocation. This approach brating hydroeconomic models. The method was used, for example, by
allows for a detailed and realistic representation of transaction costs in Medellín-Azuara et al. (2012) to simulate farmer’s adoption of irrigation tech-
models of water markets (Erfani et al., 2014). nology. Cai and Wang (2006), adapted the PMP method by using a numerical
approach based on a hybrid genetic algorithm (GA; Goldberg, 1989) and
nonlinear programming, and developed a calibrated baseline scenario of a
36.6 DISCUSSION OF CHALLENGES, LIMITATIONS, hydroeconomic model for the Maipo River Basin in Chile.
AND FUTURE DIRECTIONS
Another problem for the application of hydroeconomic models in practical
Most hydroeconomic analysis and modeling work present to date has been decision-making is that, given their complexity and specificity, the models
undertaken with the ambition to inform decision-making. In contrast to other often have to be adapted for a certain system. Usually there is a lack
types of hydrologic and water resources analysis, hydroeconomic modeling of generic decision-support system (DSS) tools that could facilitate the pro-
offers the opportunity to simulate and test the impact of economic policies on cess of implementation and application of the model. DSSs often involve
36_Singh_ch36_p36.1-36.10.indd 6 8/22/16 2:01 PM

References 36-7
capabilities of computer assisted graphical design, geographically referenced marginal value of water throughout the system, and therefore, would allow us
databases, and interactive and user-friendly graphical interfaces and tools for to assess the economic gains or losses coming from different water policies,
input management, results display, and analysis. Some examples of DSS which new infrastructures, new operating rules, or water allocation decisions to
incorporate hydroeconomic analysis capabilities are CALVIN (Draper et al., cope with water scarcity.
2003; Jenkins et al., 2004) or AQUATOOL (Andreu et al., 1996; Pulido- Despite the significant potential of hydroeconomic analysis, its practical
Velazquez et al., 2007b, 2013c). On the other hand, model platforms allow application for real decision-making processes has so far been quite limited.
model developers to link a generic user interface and data-management sys- To this end, uncertainties inherent in hydroeconomic models must be ana-
tem to external model codes. One example is HydroPlatform, an open-source lyzed, quantified, and reported, and the models should be fused with observa-
model platform that facilitates data introduction and management, allowing tions to obtain reliable and sharp predictions. Decentralized hydroeconomic
data transfer to external water-management network flow models (Harou et modeling approaches may enable analysts to relax some of the restrictive
al., 2010). SIMGAMS and OPTIGAMS are generic tools for hydroeconomic assumptions about human behavior inherent in microeconomic theory, allow
modelling that link HydroPlatform to a generic hydroeconomic code for more modeling flexibility, and yield more realistic results. Multicriteria
simulation and optimization in GAMS, respectively (Lopez-Nicolas and approaches will increasingly enable hydroeconomic analysis to include
Pulido-Velazquez, 2014; Pulido-Velazquez et al., 2013c), while SDP_GAMS is important benefits that are usually hard to monetize, such as supply reliabil-
a generic tool oriented for stochastic programming models (Macian-Sorribes ity, and resilience and ecological value. Finally, water resources systems will
and Pulido-Velazquez, 2014). increasingly have to be managed jointly with other resources systems (food,
When used for operational support or to support ongoing planning pro- power, etc.) to overcome the challenges of population growth, economic
cesses (e.g., adaptive management), a static representation of economic development, and global environmental change.
behavior may be inappropriate. Factors influencing economic behavior are
usually changing, for example, because of changes in the physical environ- REFERENCES
ment, new legislation, or changes in risk perception. Like other fields of sci-
ence that deal with complex and dynamic systems (e.g., climate science), Ahlfeld, D. P. and M. M. Laverty, “Analytical solutions for minimization of
operational hydroeconomic modeling can benefit from methods to update energy use for groundwater pumping,” Water Resources Research, 47: W06508:
system state and model parameters, including economic parameters. 2011, doi:10.1029/2010WR009752.
Data-assimilation methods that combine model predictions with observa- Ahmadi, A., M. F. Karamouz, A. Moridi, and D. Han, “Integrated planning
tions have had a significant impact in other fields of modeling, particularly in of land use and water allocation on a watershed scale considering social and
data-rich environments. Maneta and Howitt (2014) presented an approach to water quality issues,” Journal of Water Resources Planning and Management,
update agricultural production functions in real-time using data-assimilation 138 (6): 671–681, 2012, doi:10.1061/(ASCE)WR.1943-5452.0000212.
techniques, an approach that could be used, for example, with remotely Andreu, J., J. Capilla, and E. Sanchís, “AQUATOOL, a generalized decision-
sensed agricultural land use. support system for water-resources planning and operational management,”
Hydroeconomic modeling methodologies have been used in the simulation Journal of Hydrology, 177: 269–291, 1996, doi:10.1016/0022-1694(95)02963-X.
of coupled hydrohuman systems to understand the complexity of the systems, Armstrong, M. and S. Huck, “Behavioral economics as applied to firms: a
using agent-based modeling approaches. In some applications, agent-based primer,” Competition Policy International, 6: 3–45, 2010.
modeling is used to simulate rational economic behavior, but with a decen- Arthington, A. H., S. E. Bunn, N. L. Poff, and R. J. Naiman, “The challenge
tralized approach that allows system behavior to emerge from the interactions of providing environmental flow rules to sustain river ecosystems,” Ecological
of individual, rational economic agents (e.g., Zhao et al., 2013). Agent-based Applications, 16: 1311–1318, 2006, doi:10.1890/1051-0761(2006)016 {[}1311:
modeling has also been proposed as a way to relax assumptions about the TCOPEF]2.0.CO;2.
behavior of economic agents and replace them to some extent by other behav- Barreteau, O., F. Bousquet, C. Millier, and J. Weber, “Suitability of multi-
ioral theories such as bounded rationality (Armstrong and Huck, 2010). agent simulations to study irrigated system viability: application to case stud-
Instead of postulating profit or utility maximizing behavior, economic agents ies in the Senegal River Valley,” Agricultural Systems, 80: 255–275, 2004,
are assumed to behave according to a set of rules which can, in some cases, doi:10.1016/j.agsy.2003.07.005.
evolve as agents learn about the environment and the behavior of other agents Bauer, C., “Bringing water markets down to earth: the political economy of
(e.g., Ng et al., 2011; Pahl-Wostl, 2002). In some cases, for example, when water rights in Chile, 1976–95,” World Development, 25: 639–656, 1997.
modeling markets, rules can be added to economic drivers to improve upon Bauer-Gottwein, P., R. Schneider, and C. Davidsen, “Optimizing wellfield
the realism of agent-behaviors in models (Erfani et al., 2014). operation in a variable power price regime,” Groundwater, 54 (1): 92–103,
Developing integrated system-scale approaches to management of food, 2016, doi:10.1111/gwat.12341.
energy, and water resources under climate change, the so-called “water- Bayazit, M. and N. E. Unal, “Effects of hedging on reservoir performance,”
energy-food nexus” is one of today’s key challenges in environmental engi- Water Resources Research, 26: 713–719, 1990.
neering. Joint management of water, energy, and agricultural resources has Ben-Haim, Y., “Uncertainty, probability and information-gaps,” Reliability
been receiving increasing attention from researchers, operators, and decision Engineering & System Safety, 85: 249–266, 2004, doi:10.1016/j.ress.2004.03.015.
makers because recent technological and institutional developments have Berger, T., R. Birner, N. Mccarthy, J. DíAz, and H. Wittmer, “Capturing the
tightened the coupling links between these resource systems (e.g., Olsson, complexity of water uses and water users within a multi-agent framework,” Water
2012): (1) with increasing penetration of highly intermittent renewable Resource Management, 21: 129–148, 2006, doi:10.1007/s11269-006-9045-z.
energy sources (solar and wind), the importance of hydropower as balancing Blanco-Gutiérrez, I., C. Varela-Ortega, and D. R. Purkey, “Integrated
and storage facility in the power market has been increasing. (2) It is becom- assessment of policy interventions for promoting sustainable irrigation in
ing more and more difficult to meet cooling water requirements from thermal semi-arid environments: a hydro-economic modeling approach,” Journal of
power plants, given tighter environmental regulation and climatic extremes. Environmental Management, 128: 144–160, 2013, doi:10.1016/j.jenvman.
(3) The advent of biofuels has resulted in direct competition for scarce water 2013.04.037.
between the food and energy sectors. (4) Irrigation agriculture is responsible Booker, J. F., R. E. Howitt, A. M. Michelsen, and R. A. Young, “Economics
for an increasing share of food production to meet growing global food and the modeling of water resources and policies,” Natural Resources
demand. (5) Freshwater ecosystems are a significant water user and are Modelling, 25: 168–218, 2012, doi:10.1111/j.1939-7445.2011.00105.x.
among the most vulnerable and endangered ecosystems on the planet. Britz, W., M. Ferris, and A. Kuhn, “Modeling water allocating institutions
Hydroeconomic modeling concepts provide a consistent and quantitative based on Multiple Optimization Problems with Equilibrium Constraints,”
framework to inform the water-energy-food nexus debate and to provide Environmental Modelling & Software, 46: 196–207, 2013, doi:10.1016/j.env-
decision support for joint management of multiple resource systems (Pereira- soft.2013.03.010.
Cardenal et al., 2015, 2014) and the trade-offs involved (Hurford and Harou, Braat, L. C. and R. de Groot, “The ecosystem services agenda: bridging the
2014). worlds of natural science and economics, conservation and development, and
public and private policy,” Ecosystem Services, 1: 4–15, 2012, doi:10.1016/j.
ecoser.2012.07.011.
36.7 CONCLUSION
Braat, L. C. and W. J. F. van Lierop, Integrated Economic-Ecological
The main advantage of hydroeconomic modeling is its ability to track the full Modeling, Elsevier Science, Amsterdam, 1987.
temporal and spatial distribution of economic benefits that result from water Cai, X., “Implementation of holistic water resources-economic optimiza-
use and costs (losses) that result from water scarcity. No other approach has tion models for river basin management—reflective experiences,”
similar capabilities. By representing spatially distributed economic water Environmental Modelling & Software, 23: 2–18, 2008, doi:10.1016/j.envsoft.
demands throughout the water resources network, such models quantify the 2007.03.005.
36_Singh_ch36_p36.1-36.10.indd 7 8/22/16 2:01 PM

Cai, X., D. C. McKinney, and L. S. Lasdon, “Solving nonlinear water man- Erfani, T., I. Huskova, and J. J. Harou, “Tracking trade transactions in water
agement models using a combined genetic algorithm and linear program- resource systems: a node-arc optimization formulation,” Water Resources
ming approach,” Advacnes in Water Resources, 24: 667–676, 2001a, Research, 49: 3038–3043, 2013, doi:10.1002/wrcr.20211.
doi:10.1016/S0309-1708(00)00069-5. Foster, T., N. Brozović, and A. P. Butler, “Modeling irrigation behavior in
Cai, X., D. C. McKinney, and L. S. Lasdon, “Piece-by-piece approach to groundwater systems,” Water Resources Research, 50: 6370–6389, 2014,
solving large nonlinear water resources management models,” Journal of. doi:10.1002/2014WR015620.
Water Resources Planning and Management, 127: 363–368, 2001b, doi:10.1061/ Fowler, K. R., J. P. Reese, C. E. Kees, J. E. Dennis, C. T. Kelley, C. T. Miller,
(ASCE)0733-9496(2001)127:6(363). C. Audet, et al., “Comparison of derivative-free optimization methods for
Cai, X., D. C. McKinney, and L. S. Lasdon, “Integrated hydrologic-agro- groundwater supply and hydraulic capture community problems,” Advances
nomic-economic model for river basin management,” Journal of Water in Water Resources, 31: 743–757, 2008, doi:10.1016/j.advwatres.2008.01.010.
Resources Planning and Management, 129: 4–17, 2003a, doi:10.1061/ Gisser, M. and D. A. Sanchez, “Competition versus optimal control in
(ASCE)0733-9496(2003)129:1(4). groundwater pumping,” Water Resources Research, 16: 638–642, 1980,
Cai, X., D. C. McKinney, and M. W. Rosegrant, “Sustainability analysis for doi:10.1029/WR016i004p00638.
irrigation water management in the Aral Sea region,” Agricultural System, 76: Goldberg, D. E., Genetic Algorithms in Search, Optimization and Machine
1043–1066, 2003b, doi:10.1016/S0308-521X(02)00028-8. Learning, Addison-Wesley Longman, Boston, MA, 1989.
Cai, X., C. Ringler, and M. Rosegrant, “Modeling water resources manage- Goor, Q., R. Kelman, and A. Tilmant, “Optimal multipurpose-multireser-
ment at the basin level: methodology and application to the Maipo River voir operation model with variable productivity of hydropower plants,”
Basin,” Research Report 149, International Food Policy Research Institute, Journal of Water Resources Planning and Management, 137: 258–267, 2011,
Washington, D.C., 2006. doi:10.1061/(ASCE)WR.1943-5452.0000117.
Cai, X. and D. Wang, “Calibrating holistic water resources–economic mod- Griffin, R., “The fundamental principles of cost-benefit analysis,” Water
els,” Journal of Water Resources Planning and Management, 132: 414–423, Resources Research, 34: 2063–2071, 1998.
2006, doi:10.1061/(ASCE)0733-9496(2006)132:6(414). Griffin, R., Water Resource Economics: The Analysis of Scarcity, Policies and
Cardwell, H. and H. Ellis, “Stochastic dynamic programming models for Projects, MIT Press, Cambridge, MA, 2006.
water quality management,” Water Resources Research, 29: 803, 1993, Griffin, R. and S-H. Hsu, “The potential for water market-efficiency when
doi:10.1029/93WR00182. instream flows have value,” American Journal of Agricultural Economics, 75:
Castelletti, A., H. Yajima, M. Guiliani, R. Soncini-Sessa, and E. Weber, 292–303, 1993.
“Planning the optimal operation of a multioutlet water reservoir with water Groves, D. G. and R. J. Lempert, “A new analytic method for finding policy-
quality and quantity targets,” Journal of Water Resources Planning and relevant scenarios,” Global Environmental Change, 17: 73–85, 2007,
Management, 140: 496–510, 2014, doi:10.1061/(ASCE)WR.1943-5452. doi:10.1016/j.gloenvcha.2006.11.006.
Cheng, W-C., N-S. Hsu, W-M. Cheng, and W. W-G. Yeh, “A flow path Hanemann, W. M., The economic theory of WTP and WTA, Valuing
model for regional water distribution optimization,” Water Resources Research, Environmental Preferences: Theory and Practice of the Contingent Valuation
45: 2009, doi:10.1029/2009WR007826. Method in the U.S., EU and Developing Countries, edited by L. J. Bateman and
Cools, J., S. Broekx, V. Vandenberghe, H. Sels, E. Meynaerts, P. Vercaemst, K. G. Willis, Oxford University Press, Oxford, 1999, pp. 42–96.
P. Seuntjens, et al., “Coupling a hydrological water quality model and an eco- Hansen, A. K., H. Madsen, P. Bauer-Gottwein, A. K. V. Falk, and D.
nomic optimization model to set up a cost-effective emission reduction sce- Rosbjerg, “Multi-objective optimization of the management of a waterworks
nario for nitrogen,” Environmental Modelling & Software, 26: 44–51, 2011, using an integrated well field model,” Hydrology Research, 43: 430, 2012,
doi:10.1016/j.envsoft.2010.04.017. doi:10.2166/nh.2012.142.
Dalhuisen, J., “Price and income elasticities of residential water demand: a Harou, J., D. Pinte, A. Tilmant, D. E. R. D. E. Rosenberg, K. Hansen, P. M.
meta-analysis,” Land Economics, 79: 292–308, 2003. Reed, A. Reynaud et al., “An open-source model platform for water manage-
Davidsen, C., S. J. Pereira-Cardenal, S. Liu, X. Mo, D. Rosbjerg, and P. ment that links models to a generic user-interface and data-manager,”
Bauer-Gottwein, “Using stochastic dynamic programming to support water International Environmental Modelling and Software Society (iEMSs) 2010
resources management in the Ziya River Basin, China,” Journal of Water International Congress on Environmental Modelling and Software Modelling for
Resources Planning and Management, 141 (7): 04014086, 2015a, doi:10.1061/ Environment’s Sake, Ottawa, Canada, 2010.
(ASCE)WR.1943-5452.0000482. Harou, J. J. and R. R. Lund, “Ending groundwater overdraft in hydrologic-
Davidsen, C., S. Liu, X. Mo, P.E. Holm, S. Trapp, D. Rosbjerg, and P. Bauer- economic systems,” Hydrogeology Journal, 16: 1039–1055, 2008, doi:10.1007/
Gottwein “Hydroeconomic optimization of reservoir management under s10040-008-0300-7.
downstream water quality constraints,” Journal of Hydrology, 529: 1679–1689, Harou, J. J., M. Pulido-Velazquez, D. E. Rosenberg, J. Medellín-Azuara, J. R.
2015b, doi:10.1016/j.jhydrol.2015.08.018. Lund, and R. E. Howitt, “Hydro-economic models: concepts, design, applica-
Davidsen, C., S. Liu, X. Mo, D. Rosbjerg, and P. Bauer-Gottwein, “The cost tions, and future prospects,” Journal of Hydrology, 375: 627–643, 2009,
of ending groundwater overdraft on the North China Plain”, Hydrology and doi:10.1016/j.jhydrol.2009.06.037.
Earth System Sciences, 20:771-785, 2016, doi:10.5194/hess-20-771-2016, Hasler, B., J. C. R. Smart, A. Fonnesbech-Wulff, H. E. Andersen, H.
2016. Thodsen, G. Blicher Mathiesen, E. Smedberg et al., “Hydro-economic mod-
De Azevedo, L. G. T., T. K. Gates, D. G. Fontane, J. W. Labadie, and R. L. elling of cost-effective transboundary water quality management in the
Porto, “Integration of water quantity and quality in strategic river basin plan- Baltic Sea,” Water Resources Economics, 5: 1–23, 2014, doi:10.1016/j.
ning,” Journal of Water Resources Planning and Management, 126: 85–97, wre.2014.05.001.
2000. Hayes, D. F., J. W. Labadie, T. G. Sanders, and J. K. Brown, “Enhancing
De Groot, R. S., M. A. Wilson, and R. M. Boumans, “A typology for the water quality in hydropower system operations,” Water Resources Research,
classification, description and valuation of ecosystem functions, goods and 34: 471, 1998, doi:10.1029/97WR03038.
services,” Ecological Economics, 41: 393–408, 2002, doi:10.1016/S0921- Heal, G., Valuing the Future: Economic Theory and Sustainability, Columbia
8009(02)00089-7. University Press, New York, 1998.
De Moraes, M. M., X. Cai, C. Ringler, B. E. Albuquerque, S. P. da Rocha, Hine, D. and J. W. Hall, “Information gap analysis of flood model uncer-
and C. A. Amorim, “Joint water quantity-quality management in a biofuel tainties and regional frequency analysis,” Water Resources Research, 46: 2010,
production area-integrated economic-hydrologic modeling analysis,” Journal doi:10.1029/2008WR007620.
of Water Resources Planning and Management, 136: 502–511, 2010, Howitt, R. E., “Positive mathematical programming,” American Journal of
doi:10.1061/(ASCE)WR.1943-5452.0000049. Agricultural Economics, 77: 329–342, 1995, doi:10.2307/1243543.
Draper, A. J., M. W. Jenkins, K. W. Kirby, J. R. Lund, and R. E. Howitt, Hurford, A. P. and J. J. Harou, “Balancing ecosystem services with energy
“Economic-engineering optimization for California water management,” and food security—assessing trade-offs from reservoir operation and irriga-
Journal of Water Resources Planning and Management, 129: 155–164, 2003, tion investments in Kenya’s Tana Basin,” Hydrological Earth System Sciences,
doi:10.1061/(ASCE)0733-9496(2003)129:3(155). 18: 3259–3277, 2014, doi:10.5194/hess-18-3259-2014.
Ejaz, M. S. and R. C. Peralta, “Maximizing conjunctive use of surface and Hurford, A. P., I. Huskova, and J. J. Harou, “Using many-objective trade-off
ground water under surface water quality constraints,” Advances in Water analysis to help dams promote economic development, protect the poor and
Resources, 18: 67–75, 1995, doi:10.1016/0309-1708(95)00004-3. enhance ecological health,” Environmental Science Policy, 38: 72–86, 2014,
Erfani, T., O. Binions, and J. J. Harou, “Simulating water markets with doi:10.1016/j.envsci.2013.10.003.
transaction costs,” Water Resources Research, 50: 4726–4745, 2014, Jenkins, M. W., J. R. Lund, R. E. Howitt, A. J. Draper, S. M. Msangi, S. K. Tanaka,
doi:10.1002/2013WR014493. R. S. Ritzema, and G. F. Marques, “Optimization of California’s water supply
36_Singh_ch36_p36.1-36.10.indd 8 8/22/16 2:01 PM

References 36-9
system: results and insights,” Journal of Water Resources Planning and Management, O’Laoghaire, D. T. and D. M. Himmelblau, Optimal Expansion of a Water
130: 271–280, 2004, doi:10.1061/(ASCE)0733-9496(2004)130:4(271). Resources System, Academic Press, New York and London, 1974.
Karagiannis, I. C. and P. G. Soldatos, “Water desalination cost literature: Oki, T. and S. Kanae, “Global hydrological cycles and world water resources,”
review and assessment,” Desalination, 223: 448–456, 2008, doi:10.1016/j. Science, 313: 1068–1072, 2006, doi:10.1126/science.1128845.
desal.2007.02.071. Olsson, G., Water and Energy. Threats and Opportunities, IWA Publishing,
Karamouz, M., S. A. Mojahedi, and A. Ahmadi, “Interbasin water transfer: London, UK, 2012.
economic water quality-based model,” Journal of Irrigation and Drainage Padula, S., J. J. Harou, L. G. Papageorgiou, Y. Ji, M. Ahmad, and N.
Engineering, 136: 90–98, 2010, doi:10.1061/(ASCE)IR.1943-4774.0000140. Hepworth, “Least economic cost regional water supply planning—optimising
Karamouz, M., A. Moridi, and H. M. Fayyazi, “Dealing with conflict over infrastructure investments and demand management for South East England’s
water quality and quantity allocation: a case study,” Scientia Iranica, Sharif 17.6 million people,” Water Resources Management, 27: 5017–5044, 2013,
University of Technology, Iran, 15: 34–49, 2008. doi:10.1007/s11269-013-0437-6.
Kerachian, R. and M. Karamouz, “A stochastic conflict resolution model for Pahl-Wostl, C., “Towards sustainability in the water sector—the impor-
water quality management in reservoir-river systems,” Advances in Water tance of human actors and processes of social learning,” Aquatic Sciences, 64:
Resources, 30: 866–882, 2007, doi:10.1016/j.advwatres.2006.07.005. 394–411, 2002, doi:10.1007/PL00012594.
Labadie, J. W., “Optimal operation of multireservoir systems: state-of-the- Peña-Haro, S., M. Pulido-Velazquez, and A. Sahuquillo, “A hydro-economic
art review,” Journal of Water Resources Planning and Management, 130: modelling framework for optimal management of groundwater nitrate pollu-
93–111, 2004, doi:10.1061/(ASCE)0733-9496(2004)130:2(93). tion from agriculture,” Journal of Hydrology, 373: 193–203, 2009, doi:10.1016/j.
Lempert, R. J. and D. G. Groves, “Identifying and evaluating robust adap- jhydrol.2009.04.024.
tive policy responses to climate change for water management agencies in the Pereira, M. V. F. and L. Pinto, “Multistage stochastic optimization applied
American west,” Technological Forecasting and Social Change, 77: 960–974, to energy planning,” Mathematical Programming, 52: 359–375, 1991,
2010, doi:10.1016/j.techfore.2010.04.007. doi:10.1007/BF01582895.
Lempert, R. J., D. G. Groves, S. W. Popper, and S. C. Bankes, “A general, Pereira-Cardenal, S. J., H. Madsen, K. Arnbjerg-Nielsen, N. Riegels, R.
analytic method for generating robust strategies and narrative scenarios,” Jensen, B. Mo, I. Wangensteen, and P. Bauer-Gottwein, “Assessing climate
Management Science, 52: 514–528, 2006, doi:10.1287/mnsc.1050.0472. change impacts on the Iberian power system using a coupled water-power
Lopez-Nicolas, A. and M. Pulido-Velazquez, “Methods and tools to simu- model,” Climatic Change, 126: 351–364, 2014, doi:10.1007/s10584-014-
late the effect of economic instruments in complex water resources systems. 1221-1.
Application to the Jucar river basin, EGU General Assembly. Geophysical Pereira-Cardenal, S. J., B. Mo, N. D. Riegels, K. Arnbjerg-Nielsen, and P.
Research Abstracts, Copernicus Publications, Göttingen, Germany, Vol. 16, Bauer-Gottwein, “Optimization of multipurpose reservoir systems using
2014, pp. EGU2014–743–2. power market models,” Journal of Water Resources Planning and Management,
Loucks, D. P. and E. van Beek, “Water resources systems planning and 141 (8): 04014100 2015.
management, An Introduction to Methods, Models and Applications, UNESCO, Poff, N. L., B. D. Richter, A. H. Arthington, S. E. Bunn, R. J. Naiman, E.
Paris, 2005. Kendy, M. Acreman et al., “The ecological limits of hydrologic alteration
Luckmann, J., H. Grethe, S. McDonald, A. Orlov, and K. Siddig, “An inte- (ELOHA): a new framework for developing regional environmental flow
grated economic model of multiple types and uses of water,” Water Resources standards,” Freshwater Biology, 55: 147–170, 2010, doi:10.1111/j.1365-2427.
Research, 50: 3875–3892, 2014, doi:10.1002/2013WR014750. 2009.02204.x.
Macian-Sorribes, H. and M. Pulido-Velazquez, “Hydro-economic optimi- Postel, S. L., G. C. Daily, and P. R. Ehrlich, “Human appropriation of renew-
zation under inflow uncertainty using the SDP_GAMS generalized optimiza- able fresh water,” Science, 271: 785–788, 1996, doi:10.1126/science.
tion tool,” Proceedings of ICWRS2014, Evolving Water Resources Systems: 271.5250.785.
Understanding, Predicting and Managing Water–Society Interactions, Bologna, Pulido-Velazquez, D., A. Sahuquillo, J. Andreu, and M. Pulido-Velazquez,
Italy, 2014. “An efficient conceptual model to simulate surface water body-aquifer inter-
Maier, H. R., Z. Kapelan, J. Kasprzyk, J. Kollat, L. S. Matott, M. C. Cunha, action in conjunctive use management models,” Water Resources Research, 43
G. C. Dandy et al., “Evolutionary algorithms and other metaheuristics in (7): W07407: 2007a, doi:10.1029/2006WR005064.
water resources: current status, research challenges and future directions,” Pulido-Velazquez, D., A. Sahuquillo, J. Andreu, and M. Pulido-Velazquez,
Environmental Modelling & Software, 62: 271–299, 2014, doi:10.1016/j.env- “A general methodology to simulate groundwater flow of unconfined aquifers
soft.2014.09.013. with a reduced computational cost,” Journal of Hydrology, 338: 42–56, 2007b,
Maneta, M. P. and R. E. Howitt, “Stochastic calibration and learning in doi:10.1016/j.jhydrol.2007.02.009.
nonstationary hydroeconomic models,” Water Resources Research, 50: 3976– Pulido-Velazquez, M., E. Alvarez-Mendiola, and J. Andreu, “Design of
3993, 2014, doi:10.1002/2013WR015196.Received. efficient water pricing policies integrating basinwide resource opportunity
Matrosov, E. S., A. M. Woods, and J. J. Harou, “Robust decision making and costs,” Journal of Water Resources Planning and Management, 139: 583–592,
info-gap decision theory for water resource system planning,” Journal of 2013a, doi:10.1061/(ASCE)WR.1943-5452.0000262.
Hydrology, 494: 43–58, 2013, doi:10.1016/j.jhydrol.2013.03.006. Pulido-Velazquez, M., E. Alvarez-Mendiola, and J. Andreu, “Design of
Mayer, A. S., C. T. Kelley, and C. T. Miller, “Optimal design for problems efficient water pricing policies integrating basinwide resource opportunity
involving flow and transport phenomena in saturated subsurface systems,” costs,” Journal of Water Resources Planning and Management, 139: 583–592,
Advances in Water Resources, 25: 1233–1256, 2002, doi:10.1016/S0309- 2013b, doi:10.1061/(ASCE)WR.1943-5452.0000262.
1708(02)00054-4. Pulido-Velázquez, M., J. Andreu, and A. Sahuquillo, “Economic optimiza-
McKinney, C. D., X. Cai, M. R. Rosegrant, C. Ringler, and C. A., Scott, tion of conjunctive use of surface water and groundwater at the basin scale,”
Integrated basin-scale water resources management modeling: review and Journal of Water Resources Planning and Management, 132: 454–467, 2006,
future directions, Research Paper No. 6, International Water Management doi:10.1061/(ASCE)0733-9496(2006)132:6(454).
Institute, Colombo, Sri Lanka, 1999. Pulido-Velazquez, M., J. Andreu, A. Sahuquillo, and D. Pulido-Velazquez,
Medellín-Azuara, J., R. E. Howitt, and J. J. Harou, “Predicting farmer “Hydro-economic river basin modelling: the application of a holistic surface–
responses to water pricing, rationing and subsidies assuming profit maximiz- groundwater model to assess opportunity costs of water use in Spain,”
ing investment in irrigation technology,” Agricultural Water Management, Ecological Economy, 66: 51–65, 2008, doi:10.1016/j.ecolecon.2007.12.016.
108: 73–82, 2012, doi:10.1016/j.agwat.2011.12.017. Pulido-Velazquez, M., M. W. Jenkins, and J. R. Lund, “Economic values for
Maass, A., M. M. Hufschmidt, R. Dorfman, H. A. Thomas, S. A. Marglin, conjunctive use and water banking in southern California,” Water Resources
and G. M. Fair, “Design of Water-Resource Systems—New Techniques for Research, 40: 2004, doi:10.1029/2003WR002626.
Relating Economic Objectives, Engineering Analysis, and Governmental Pulido-Velazquez, M., A. Lopez-Nicolas, J. Andreu, and J. Harou,
Planning, McMillan, London, 1962. “Generalized DSS shell for developing simulation and optimization hydro-
Ng, T. L., J. W. Eheart, X. Cai, and J. B. Braden, “An agent-based model of economic models of complex water resources systems, EGU General Assembly.
farmer decision-making and water quality impacts at the watershed scale Geophysical Research Abstracts, Copernicus Publications, Göttingen,
under markets for carbon allowances and a second-generation biofuel crop,” Germany, Vol. 15, 2013c, pp. EGU2013–8920.
Water Resources Research, 47: 2011, doi:10.1029/2011WR010399. Rani, D. and M. M. Moreira, “Simulation–optimization modeling: a survey
Nicklow, J., P. Reed, D. Savic, T. Dessalegne, L. Harrell, A. Chan-Hilton, M. and potential application in reservoir systems operation,” Water Resources
Karamouz et al., “State of the art for genetic algorithms and beyond in water Management, 24: 1107–1138, 2009, doi:10.1007/s11269-009-9488-0.
resources planning and management,” Journal of Water Resources Planning and Reed, P. M., D. Hadka, J. D. Herman, J. R. Kasprzyk, and J. B. Kollat,
Management, 136: 412–432, 2010, doi:10.1061/(ASCE)WR.1943-5452.0000053. “Evolutionary multiobjective optimization in water resources: the past,
36_Singh_ch36_p36.1-36.10.indd 9 8/22/16 2:01 PM

present, and future,” Advances in Water Resources, 51: 438–456, 2013, Wallace, J. S., M. C. Acreman, and C. A. Sullivan, “The sharing of water
doi:10.1016/j.advwatres.2012.01.005. between society and ecosystems: from conflict to catchment-based co-man-
Renzetti, S., The Economics of Water Demands, Kluwer Academic, agement,” Philosophical Transactions of the Royal Society of London B:
Dordrecht, The Netherlands, 2002. Biological Sciences, 358: 2011–2026, 2003, doi:10.1098/rstb.2003.1383.
Riegels, N., R. Jensen, L. Bensasson, S. Banou, F. Møller, and P. Bauer- Ward, F. A. and M. Pulido-Velazquez, “Water conservation in irrigation
Gottwein, “Estimating resource costs of compliance with EU WFD ecological can increase water use,” Proceedings of the National Academy of Sciences of the
status requirements at the river basin scale,” Journal of Hydrology, 396: 197– United States of America, 105: 18215–18220, 2008, doi:10.1073/
214, 2011, doi:10.1016/j.jhydrol.2010.11.005. pnas.0805554105.
Riegels, N., M. Pulido-Velazquez, C. Doulgeris, V. Sturm, R. Jensen, F. Wolfgang, O., A. Haugstad, B. Mo, A. Gjelsvik, I. Wangensteen, and G.
Møller, and P. Bauer-Gottwein, “Systems analysis approach to the design of Doorman, “Hydro reservoir handling in Norway before and after deregula-
efficient water pricing policies under the EU water framework directive,” tion,” Energy, 34: 1642–1651, 2009, doi:10.1016/j.energy.2009.07.025.
Journal of Water Resources Planning and Management, 139: 574–582, 2013, Xevi, E. and S. Khan, “A multi-objective optimisation approach to water
doi:10.1061/(ASCE)WR.1943-5452.0000284. management,” Journal of Environmental Management, 77: 269–277, 2005,
Rockstrom, J., W. Steffen, K. Noone, A. Persson, F. S. Chapin III, E. Lambin, doi:10.1016/j.jenvman.2005.06.013.
T. M. Lenton et al., “Planetary boundaries: Exploring the safe operating space Yang, Y-C.E., X. Cai, and E. E. Herricks, “Identification of hydrologic indi-
for humanity,” Ecology and Society, 14 (2), Article Number 32: 2009. cators related to fish diversity and abundance: a data mining approach for fish
Rosegrant, M. and H. Binswanger, “Markets in tradable water rights— community analysis,” Water Resources Research, 44: 2008, doi:10.1029/
potential for efficiency gains in developing-country water-resource alloca- 2006WR005764.
tion,” World Development, 22: 1613–1625, 1994. Yang, Y-C.E., X. Cai, and D. M. Stipanović, “A decentralized optimization
Smakhtin, V., C. Revenga, and P. Doll, “A pilot global assessment of envi- algorithm for multiagent system-based watershed management,” Water
ronmental water requirements and scarcity,” Water International, 29: 307– Resources Research, 45: 2009, doi:10.1029/2008WR007634.
317, 2004. Yang, Y-C.E., J. Zhao, and X. Cai, “Decentralized optimization method for
Stage, S. and Y. Larsson, “Incremental cost of water power,” Transactions of water allocation management in the Yellow River Basin,” Journal of Water
the American Institute of Electrical Engineering, 80: 361–364, 1961. Resources Planning and Management, 138: 313–325, 2012, doi:10.1061/
Stedinger, J., B. Sule, and D. Loucks, “Stochastic dynamic programming (ASCE)WR.1943-5452.0000199.
models for reservoir operation optimization,” Water Resources Research, 20: You, J-Y. and X. Cai, “Determining forecast and decision horizons for res-
359–375, 1984. ervoir operations under hedging policies,” Water Resources Research, 44:
Steffen, W., P. J. Crutzen, and J. R. Mcneill, “The anthropocene : Are 2008a, doi:10.1029/2008WR006978.
humans now overwhelming the great forces of nature?” Ambio, 36: 614–621, You, J-Y. and X. Cai, “Hedging rule for reservoir operations: 1. A theoreti-
2007. cal analysis,” Water Resources Research, 44: 2008b, doi:10.1029/2006WR005481.
Tilmant, A. and R. Kelman, “A stochastic approach to analyze trade-offs Young, R. A., Determining the economic value of water: concepts and
and risks associated with large-scale water resources systems,” Water Resources methods, Resources for the Future, An RFF Press book published by Resources
Research, 43: 2007, doi:10.1029/2006WR005094. of the Future, Washington, D.C., 2005.
Tilmant, A., D. Pinte, and Q. Goor, “Assessing marginal water values in Zhao, J., X. Cai, and Z. Wang, “Comparing administered and market-based
multipurpose multireservoir systems via stochastic programming,” Water water allocation systems through a consistent agent-based modeling frame-
Resources Research, 44: 2008, doi:10.1029/2008WR007024. work,” Journal of Environmental Management, 123: 120–130, 2013,
Varian, H. R., Intermediate Microeconomics—A Modern Approach, 8th ed., doi:10.1016/j.jenvman.2013.03.005.
W. W. Norton, New York, 2010.
36_Singh_ch36_p36.1-36.10.indd 10 8/22/16 2:01 PM

PART 4
HYDROLOGIC PROCESSES
AND MODELING
37_Singh_ch37_p37.1-37.10.indd 1 8/22/16 2:07 PM

Chapter 37
Weather and Climate
BY
MICHAEL ANDERSON
ABSTRACT concepts will be presented. An essential part of weather includes the monitor-
ing methods that provide the data inputs for hydrologic engineering applica-
Weather and climate play a foundational role in hydrologic engineering.
tions, which are covered highlighting important considerations when using
Weather reflects the short-term conditions of the atmosphere while climate
such data including gridded products. Climate concepts are covered next,
reflects the long-term statistical distribution of atmospheric conditions driv-
including variability across space and time scales. Some of the fundamental
en by variability in the atmospheric composition, ocean, land surface and
hydrologic engineering applications are based on assumptions of a stationary
solar input conditions. Knowledge of weather and climate and the physical
climate. However, these assumptions are being challenged under an increas-
processes that drive them can improve the development and use of hydro-
ing awareness of a changing climate. With that in mind, climate change
logic engineering tools and practice. Such knowledge will be critical in the
impacts on hydrology and hydrologic engineering will be introduced to close
coming decades as impacts from climate change materialize.
out the chapter.
This chapter starts with a description of the intersection of weather and
climate with hydrologic engineering. Elements of weather and monitoring of
weather that impact hydrologic engineering is presented next. Climate con- 37.2 HYDROLOGIC ENGINEERING AND INTERSECTION
cepts follow including a discussion of the role of variability. The chapter WITH WEATHER AND CLIMATE
closes with some thoughts on the impacts of climate change on hydrologic Hydrologic engineering methods for planning and design of storm-water
engineering. runoff management, reservoir sizing, forecasting of stream flow for
water supply, or flood management all rely on observations and statistics of
37.1 INTRODUCTION weather variables like precipitation and temperature. Engineering tools, such
as the rational method, univariate and multivariate regression, and rainfall-
Weather and climate play a foundational role in hydrologic engineering. runoff or watershed models all make use of weather data and climate con-
Weather is the short-term conditions of the atmosphere (i.e., daily) whereas cepts. Underlying the methods used to create these tools are assumptions
climate is the long-term average conditions of the atmosphere (i.e., decadal). about the processes that drive weather and climate for a given location over
Weather systems leading to flood and drought, the seasonal cycles of weather time. A better understanding of weather and climate principles will improve
over the course of a year, and annual to decadal variability all impact an area’s an engineer’s judgment on when to use which tool for the job at hand. In this
hydrology. Knowledge of weather and climate and the physical processes that section, selected engineering applications are used to demonstrate how
drive them can improve the development and use of hydrologic engineering weather and climate underlie those applications. Foundational documents for
tools and practice. This knowledge will be critical in the coming decades as these methods are referenced. The section starts with planning applications,
impacts from climate change appear. moves to operational applications such as snowmelt forecasting, and ends
Figure 37.1 illustrates the hydrologic cycle that is foundational to hydro- with a look at watershed modeling.
logic engineering. Precipitation falls on the landscape either as liquid (rain) or
solid (snow, sleet, hail, etc.). Water, as rain or melting snow and ice, can infil- 37.2.1 Planning
trate into the ground or runoff via overland and channel flow. Channel flow Engineering planning utilizes design concepts to estimate the size of a struc-
moves water from the land surface to terminal reservoirs, lakes, seas, and ture like a spillway or storm-water detention basin. These concepts are used
oceans. Evaporation and transpiration from plants return water to the atmo- to establish thresholds for flood protection that go into project development
sphere where it is moved by the winds and condenses into clouds leading back and planning documents. Concepts of how much rain or runoff, how long
to precipitation. Texts like Bedient and Huber (1988) and Brutsaert (2005) the event lasts, and how frequently one can expect events of that magnitude
cover introductory concepts to the hydrologic cycle as it relates to hydrologic have been developed based on statistics of observed quantities. Modifications
engineering. are made when new observations exceed past events.
Weather and climate are critical elements to understanding the role of the One of the most widely used engineering applications using these concepts
hydrologic cycle in hydrologic engineering. In general, the relation between is the Rational Method. The method developed in 1889 by Kuichling relates
the earth’s atmosphere and solar radiation drives the current climate pro- the magnitude of runoff (peak discharge, Q) to the intensity of rainfall i over
cesses that set the stage for day-to-day weather variability that drives the the area A that the rain falls upon with runoff coefficient c, which is a function
hydrologic cycle. Climate processes establish the spatial and temporal vari- of soil type, drainage basin slope, and ground cover, as shown in Eq. (37.1).
ability of weather and lead to changes in the statistical properties of weather
variables (i.e., temperature, precipitation, wind, atmospheric pressure, etc.) Q = ciA (37.1)

that impact hydrologic engineering. Climate change impacts can modulate
the relationships between different flow paths in the hydrologic cycle chang- In this method, weather and climate are tied to the estimate of the inten-
ing the timing and magnitude of runoff, recharge, evapotranspiration, and sity, i. Federal documents such as the NOAA Atlas 2 (1973) and updates like
snow accumulation and melt. In this chapter the introductory concepts of NOAA Atlas 14 (2006) provide estimates of intensities over different dura-
weather, climate, and the hydrologic cycle that are central to civil engineering tions for different return periods for different locations. For a given location,
methods are presented. tables of intensities for different durations are compiled. Different volume
The chapter starts with a description of the intersection of hydrologic engi- numbers are used to organize the data for different regions or states. As an
neering with weather and climate. Next, a general overview of weather example, California data can be found in NOAA Atlas 14 Volume 6 (2011).
37-3
37_Singh_ch37_p37.1-37.10.indd 3 8/22/16 2:07 PM

37-4 Weather and Climate
Evapotranspiration
Reservoir
Water
Runoff treatment
plant
Municipal/
Stream industrial use
Diversion
canal River Treatment
plant
Wildlife
refuge
Irrigated Municipal/
Recharge basin Injection industrial
agriculture well supply well
Shallow
monitoring
Unconfined well
aquifer Groundwater Deep
Agricultural table
Confined supply well monitoring
aquifer well
Aquitard
Figure 37.1 Schematic of the hydrologic cycle [Source: California Department of Water Resources].
To develop such tables of intensity-duration-frequency data, historical flood events do not change in time. Efforts are made to sort through historical
rainfall totals over different durations are analyzed and fit to probability dis- records to sort out different flood peaks associated with rainfall or snowmelt.
tributions which relate the magnitude to the return period. Statistical meth- If spatial independence can be established a space for time exchange is used
ods such as L-Moments (Hosking and Wallis, 2005) or empirical fitting to extend short records.
functions (Shorack and Wellner, 1986) are used for the distribution fitting. The assumption of stationarity of hydrologic variables has recently been
The observed rainfall peaks are evaluated to make sure that they meet the questioned as evidence of anthropogenic climate change mounts.
assumptions of independence and that they come from a homogeneous Nonstationarity has also arisen from changes to the landscape such as forest
population to satisfy the requirements of the statistical theory used to develop clearing or from channel changes upstream for different water management
the method. purposes. Workshops on addressing nonstationarity issues have been held
Another important engineering application that uses statistical methods to (Colorado Water Institute, 2010) and it remains an active area of research.
fit observational data is the Bulletin 17B (USGS, 1982) method for sizing
flood peaks and volumes. The Bulletin 17B process was started to provide a
nationwide standard (log-Pearson Type III) to sizing flood peaks and volumes 37.2.2 Forecasting
for flood management projects. In this application, the flood peak or volume Hydrologic forecasting can be either event driven estimates of how much
for a given return period Q is described in Eq. (37.2): runoff is generated from observed or predicted precipitation or how much
runoff is expected over the spring from a melting snow pack. Here we will
logQ = logX + KSlogX (37.2) focus on the seasonal snow pack melt forecasts used to estimate water vol-

umes used for water supply in the Western United States. For watersheds in
where X is the mean of the annual peak flow time series, S is the standard 11 western states, snowmelt forecasts are provided by the Natural Resources
deviation of the time series of the log of the annual peak flows, and K is a Conservation Service at the National Water and Climate Center (http://www.
parameter based on the skew and return period. Tables of the coefficient K wcc.nrcs.usda.gov/). In California, the forecasts are provided by the
can be found in the Bulletin 17B documentation. Department of Water Resources as part of the California Cooperative Snow
In this application, the mean, standard deviation, and skew–the first three Surveys (http://www.water.ca.gov/floodmgmt/hafoo/hb/sss/).
moments of the statistical distribution of observed values–are used to esti- In the mountainous areas of the western United States, snow typically accu-
mate the size flood for a given return period. Federal programs like the mulates from December through March. From April through July, the accu-
National Flood Insurance Program (https://www.fema.gov/media-library/ mulated snow melts as the sun angle moves higher in the sky and temperatures
assets/documents/101759) and the Army Corps of Engineers Levee warm (Lundquist and Flint, 2006). April 1 is the historical average date of
Certification Program (http://www.usace.army.mil/Missions/CivilWorks/ maximum accumulation. Some years have heavy snow packs that build into
LeveeSafetyProgram.aspx) use the threshold of a 1% exceedance or 100-year May and have melt that extends past July. On the other end of the spectrum dry
return period event at the time of publication. years tend to have smaller snow packs that peak earlier and melt out before the
The foundational assumptions for this method are that the largest annual end of July. An average snowpack trace for three regions in California is shown
flood peaks for a given location are independent events that come from the in Fig. 37.2. Also shown in Fig. 37.2 are the historical maximum and minimum
same statistical distribution. This assumption is necessary to satisfy the sam- traces of each region illustrating the wide range of possible outcomes.
ple statistics theory used to develop the method. It is further assumed that the Statistical methods have been used that relate runoff volumes with select
longer the data record, the closer the observed statistical properties approach observations of the amount of water in the snow in the watershed and
the real distribution. Such concepts presume that the same weather types expected precipitation over the forecast period. As long as the runoff in the
generate the flood peaks each year and that the statistical properties of the forecast period stays within the historical bounds of observations, these
37_Singh_ch37_p37.1-37.10.indd 4 8/22/16 2:07 PM

Weather 37-5
California snow water content–percent of April 1 average

250
North
% April 1 average
200
150
100
50
0
250
% April 1 average Central
200
150
100
50
0
250
Soutn
% April 1 average
200
150
100
50
0
Dec Jan Feb Mar Apr May Jun Jul
Average 1976–1977 (min) 1982–1983 (max)
Figure 37.2 Regional plots of snow accumulation and melt for California [Source: California Data
Exchange Center].
methods tend to work fairly well. Another alternative is to use watershed oceans and the subsequent thermodynamic balancing of the atmosphere.
models to simulate the physical processes of the watershed (see, e.g., Kavvas There are several components of the atmosphere that help set the stage for
et al., 2013). Watershed modeling in hydrologic engineering is described weather events and are described in the following paragraphs. For more infor-
briefly in the next section. mation on weather and the atmosphere, see Carlson (1991), Gill (1982),
Holton (1992), and Pedlosky (1987).
37.2.3 Watershed Modeling The atmosphere is made up of a series of layers defined by physical charac-
teristics. The weather that impacts the hydrologic cycle occurs in the lowest
Watershed models are collections of mathematical representations of hydro-
layer of the atmosphere known as the troposphere. It extends from the surface
logic processes to simulate the movement of water through a watershed. The
to about 12 km. In this layer, temperature decreases with altitude. The rate of
level of complexity of the models varies from simple, lumped parameter
temperature decrease with altitude is known as the lapse rate. At the top of the
models to detailed, spatially distributed models using some variation of math-
troposphere is the tropopause which is a layer of atmosphere with little to no
ematical representations of the different physical processes. A review of the
change in temperature with altitude. It separates the troposphere from the
history of watershed models can be found in Singh and Frevert (2002a 2002b).
stratosphere and tends to cap the vertical extent of moisture-bearing weather
The level of complexity of the model often corresponds with the level of
systems. See Holton (1992) for more information on atmospheric structure.
detail required of the input data. Parameterized relationships to represent dif-
In general, more heat through solar radiation is input to the atmosphere in
ferent watershed processes may not require much input data, but need to be
the tropics, from the equator to 23° N/S latitude. The warmer temperatures
calibrated and verified against observed runoff data. These calibrations “fit”
and vast areas of ocean in the tropics also create high concentrations of water
the model to the rainfall and runoff data. The underlying relationships being
vapor. In the atmosphere, heat (in latent and sensible forms) and water vapor
simulated are presumed to be constant through time and under different cli-
are moved from the tropics toward the polar regions, latitude greater than 66°,
mate conditions, such as droughts or floods. Lack of performance by these
through the middle latitudes, the region from 23° latitude to 66° latitude by
models is usually tied to the violation of the aforementioned assumption.
wind patterns called the atmospheric global circulation. These regions have
More detailed distributed models require detailed data at many points in
distinct characteristics to their weather patterns and the resulting temperature
the watershed. Variables, such as maximum and minimum temperature, pre-
and precipitation that drive hydrology. Differences in solar input are also
cipitation, wind, and solar or net radiation, are used in the computation of the
affected by the rotation of the earth, which results in day and night and tilt of
equations to follow water through the watershed. For many watersheds, such
the earth which results in the seasonal cycle.
detailed data are not available. This is where derived data products, such as
One important component of the atmospheric global circulation for hydrol-
gridded data fields, have been very helpful.
ogy is the jet stream. The jet stream is a narrow band of fast moving air near
Gridded datasets are regular-spaced values of a variable like temperature or
the top of the troposphere that moves eastward across the middle latitudes.
precipitation that are derived from point observations. Interpolation schemes
There are areas of high pressure and low pressure that form during the evolu-
are used to translate point observations to all grid spaces in the domain. An
tion of the atmosphere as the earth spins and tilts and heat is transferred from
example of this process is the Precipitation-Elevation Regressions on
equator to poles. These areas of high and low pressure form and dissipate over
Indipendent Stopes Model (PRISM) method developed at Oregon State
the course of days to weeks. The space and time scale of these pressure regions
University (Daley et al., 1994). Gridded datasets are models themselves and
is called the synoptic scale. Synoptic scale evolution of the atmosphere helps
are discussed further in the next section which looks at weather.
mix the air between equator and poles and in the vertical. An observer on the
ground would note different weather conditions as the synoptic systems
37.3 WEATHER
evolve and pass overhead. More information on atmospheric dynamics can
Weather drives the atmospheric component of the hydrologic cycle. It is be found in Holton (1992) and information on global circulation of the
driven by solar energy from the sun to the atmosphere, land surface, and atmosphere can be found in Grotjahn (1993), and Cushman-Roisin (1994).
37_Singh_ch37_p37.1-37.10.indd 5 8/22/16 2:07 PM

The areas of high and low pressure have different characteristics that set the (b) 7 Nov 2006 a.m. composite 625 mm
stage for the weather that is observed on the ground. Areas of high pressure 50N
in 72 h
are associated with sinking air. Sinking air is warming and is associated with Atmospheric
40N river
warmer temperatures, less relatively humid air, and no precipitation. The
areas of sinking motion tend to be large with little pressure differences result- 30N
ing in little horizontal wind. Air rotates clockwise around high-pressure cen-
ters. A weather observer underneath a high-pressure system would note 20N
sunny skies and warm temperatures.
Areas of low pressure are where air is rising and cooling. Air rotates counter- 10N
clockwise around low-pressure systems. Rising, air cools and thus can hold less
water. As the air cools and has less capacity to hold the water vapor, it was car- Eq
rying condensation (the transition in water phase from gas to liquid) occurs Figure 37.3 Microwave imagery of water vapor over the Pacific Ocean with atmo-
and clouds form. If the temperatures drop below freezing (0°C) then the con- spheric river annotated.
densed droplets can freeze and become ice particles if the right nuclei are pres-
ent. The relationship between temperature and the amount of water vapor the
atmosphere can hold is defined by the Clausius Clapyron equation (see Haltiner on atmospheric rivers and their relation to extreme precipitation events can be
and Williams, 1980). The saturation vapor pressure (mbar) over a water surface found in Ralph et al. (2010) and Ralph and Dettinger (2012).
as a function of temperature (°C) (Gill, 1982) can be approximated as: As noted earlier, the land surface plays an important role in precipitation
processes. Winds encountering mountains are forced up and over them. The
Log10 ew(t) = (0.7859+0.03477t)/(1+0.00412t) (37.3) forcing of the air by the topography causes increased convergence of water
In this equation, t is temperature (°C), and ew is water vapor (mbar). As can vapor along with cooling. The combined effects of rising and cooling during
be seen from Eq. (37.3), the relationship between temperature and the a frontal precipitation event result in enhancement of precipitation. This is
amount of water vapor is highly nonlinear. This equation describes how called orographic precipitation and is why mountainous areas tend to receive
warmer tropical air will hold much more water vapor than colder, polar air. higher amounts of rain and snow and is described in detail by Rhea (1978).
The transition from warmer temperatures to colder temperatures from equa- Another important process that can happen along weather fronts is called
tor to pole and their mixing in the middle latitudes through the synoptic scale convection. Warm, moist air is lighter than colder, dryer air. Because of this
pressure systems are where most precipitation form and fall to the ground. difference in weight, warmer moister air will want to rise and the colder, dryer
These transition zones between the high-and low-pressure systems that air will want to sink. This buoyant motion is called convection and can result
evolve in the middle latitudes are anything but gradual and smooth. Areas of in strong vertical motions in the atmosphere. As the warm, moist air rises
strong gradients in pressure and temperature can form which are called fron- buoyantly it cools rapidly and water vapor condenses releasing heat that fur-
tal zones. As the pressure gradient increases, the winds increase in speed ther accelerates the rising and further condensation. The rapid vertical
along the pressure gradient due to the spinning of the earth. This effect, the motions can result in the development of thunderstorms which can drop
Coriolis Effect, enables the gradients to persist and possibly intensify. intense precipitation and hail. The winds of synoptic systems help to move
The frontal zone also has a strong temperature gradient. Frontal zones are thunderstorms across the landscape. The scale of individual cells of convec-
described as a warm front, cold front, or stationary front depending on tion is on the order of tens of kilometers. Sometimes multiple thunderstorms
the relative temperature of the advancing front. Each has a different vertical form along a frontal boundary and multiple thunderstorms are moved over
structure and relationship to the location of the low pressure. See Carlson an observer on the ground. This is called training and can result in extended
(1991) for more details. periods of heavy precipitation and ultimately flooding. Carlson (1991) pro-
The winds along the foot of frontal zones, especially cold fronts, can be vides more information on convective processes.
quite intense and if supplied by a source of water vapor typically from the Convective precipitation is also associated with tropical systems known as
tropics can form narrow features called atmospheric rivers (Zhu and Newell, hurricanes, typhoons, or cyclones. These tropical weather systems form when
1998). These narrow bands of intense atmospheric water vapor are responsible tropical convection becomes self-sustaining at large scales and organized to
for 90% of the equator to pole transport of moisture. Figure 37.3 shows an start rotating around a low pressure center. The low pressure center of a hur-
image of water vapor concentration and atmospheric rivers over oceans from ricane is called the eye of the hurricane. As a hurricane develops, the pressure
satellite data. The bands are only hundreds of kilometers wide, but can extend at the center can drop, which increases the strength of the winds rotating
thousands of kilometers from the tropics and across the middle latitudes. around the eye. Warm ocean waters help support convection providing the
When atmospheric rivers are entrained into middle latitude low-pressure hurricane with vast amounts of water vapor that condenses in the thunder-
systems, heavy precipitation occurs ahead of the passing of the cold front as storms generating more energy for the storm. Landfalling hurricanes result in
the intense moisture, which is located in the lowest 5 km of the atmosphere, very strong winds and heavy precipitation that can cause significant damage.
interacts with the dynamics of the front and the land surface. Further information Figure 37.4 shows that hurricanes produce rainfall totals over 3 days over the
Figure 37.4 Three-day accumulation ranges of precipitation associated with atmospheric rivers and hurricanes over the
Continental United States [Source: From Ralph and Dettinger, 2012].
37_Singh_ch37_p37.1-37.10.indd 6 8/22/16 2:07 PM

Climate 37-7
Gulf and Atlantic states equal to the precipitation of landfalling atmospheric estimation of values. While work continues to generate improved values, the
rivers on the Pacific Coast (Ralph and Dettinger, 2012). These are the most products are still widely accepted and used. These gridded products have also
intense precipitation processes in the United States. helped in the advancement of the description of a location’s climate.
37.4 OBSERVING WEATHER 37.5 CLIMATE
As noted earlier, an observer on the ground watches weather change as syn- Climate can be thought of as the expected state of the atmosphere at a given
optic systems, fronts, thunderstorms, and hurricanes pass overhead. The time. In summer, a location may be expected to be warm and dry or warm
observer records the impact of these events by measuring atmospheric vari- with the chance of thunderstorms. In winter, expectations may be for colder
ables, like pressure, temperature, amount and form of precipitation, and wind weather and rain or snow. The description of these expectations can be done
speed and direction. Other observations, like solar and terrestrial radiation, using sampling statistics methods. Each weather observation is a sample that
atmospheric water vapor content, snow water equivalent, and soil moisture helps to inform the statistical moments (mean, variance, and skew) of the
help inform the hydrologic response. distributions of weather variables that describe a location’s climate.
There are many different ways in which weather observations are gathered. Historically temperature and precipitation have been the primary variables
The National Weather Service uses trained volunteers to record daily tem- used to describe the climate of a location. The time scale such as daily,
perature, maximum and minimum temperature, snowfall, and snow depth monthly, seasonal, annual, decadal, or longer can be used to characterize the
through their Cooperative Observer network. Grass-roots weather observers, observed statistical distributions that represent a location’s climate. The
such as the Community Collaborative Rain, Hail and Snow (CoCoRaHS) National Climatic Data Center of the National Oceanographic and
group (http://www.cocorahs.org) or Weather Underground (http://www. Atmospheric Administration uses a 30-year window to construct the statistical
wunderground.com/), use volunteer observers to record weather data that measures of climate for a location that are referred to as the 30-year normal
can be useful for hydrologic engineering projects. Care must be taken to values. They are updated once per decade and account for new methodologies
review the data for quality control and completeness. in computing such statistics accommodating issues, like missing or bad data,
In remote areas, weather instruments on towers are used. Data are relayed data from multiple sources with differing types of observation errors (see
by satellite communications to provide near–real-time information for hydro- Arguez et al., 2012). These values are often the values used in Hydrologic
logic forecasting. The data can also be used to drive hydrologic models for Engineering planning and design.
other engineering activities. Using the same statistical methods with different 30-year time periods, one
While the data gathered in this manner is very important, it is a point can evaluate how the climate distributions are evolving in time. The change in
measurement in a large and variable landscape. To address this issue, new these distributions through time due to the evolution of the state of the atmo-
instrumentation using radar and satellite technology have been invented. sphere, land surface, and ocean systems is called climate variability. Human-
Radar technology has been successfully applied to determine spatial rain induced changes to the atmosphere and land surface that modulate the
fields and the intensity and form of precipitation. The National Weather statistical distributions of temperature and precipitation are referred to as
Service has deployed a network of radar systems across the United States for anthropogenic climate change (IPCC, 1995).
this purpose (see Anagnostou et al., 2004; Baeck and Smith, 1998; Austin, Figure 37.5 shows a depiction of the year-to-year variability in precipitation
1987). In the near future, aircraft measurements including drones will be used for the continental United States. The values depicted are the coefficient of
for weather observations (Painter et al., 2015). variation, the standard deviation divided by the mean. This value provides a
Vertically pointing radar has been used to identify the altitude at which sense of how far each annual precipitation totals typically deviates from their
falling precipitation changes from snow to rain to help determine freezing long-term mean values. Low values imply that any one year’s total is not that
elevation. This metric has been identified as an important factor in determin- different from the average value. In this case, the mean value may be a reason-
ing the amount of the watershed producing direct runoff from a weather able representation of the annual precipitation. Larger values imply that each
event which impacts the size and volume of the hydrograph. California has year’s accumulated precipitation can be quite different from the mean and
deployed a picket fence of these radar units down the length of the state to that the mean is not necessarily a reasonable depiction of what transpires in
monitoring the freezing elevation (White et al., 2013). any given year. In such cases, it is advantageous to know about the variance in
Data is also being gathered from satellites. The Geostationary Operational the data and its possible sources, such as the Pacific Decadal Oscillation,
Environmental Satellite is perhaps the most common source of satellite data Atlantic Multi-decadal Oscillation, or El Nino Southern Oscillation.
tracking clouds, cloud heights, and temperatures in the visible and infrared These “oscillations” are identified recurring changes in the state of the
spectrums respectively. Other satellites, which carry the Special Sensor ocean-atmosphere-land surface system. For example, the El Nino Southern
Microwave/Imager provide a more direct estimate total water vapor, but only Oscillation refers to a coupled ocean-atmosphere phenomena in the tropical
work with water vapor at the lower boundary of the atmosphere. Other data Pacific (see Philander, 1989). During its warm phase or El Nino, the eastern
are needed to track this information over land. It was the application of some tropical Pacific sea-surface temperatures are warmer than normal and sea
of these satellite data led to the discovery of atmospheric river phenomena. level pressure tends to be high over Tahiti and low over Darwin, Australia.
The visualization of water vapor movement over the globe from satellites
showed how the abundant moisture of the tropics is drawn into the middle
latitudes in narrow filaments that while hundreds of kilometers wide can be
thousands of kilometers in length. Such work has helped advance and inform
flood forecasting efforts and improved our understanding of how the atmo-
sphere works.
It is important to note that the point scale measurements gathered from the
ground are vital to the operation of the radar and satellite observing systems.
Data from the point-scale measurements are used to calibrate and validate the
data and image-processing routines used to transform satellite and radar
energy readings into the desired hydrometeorological observations. Radar
and satellite products can also be challenged by rapidly varying terrain, such
as mountains. Mountains can block or distort the signals which impact the
data-transformation routines and resultant data. Used together, suites of
observations from stations, radar, satellites, and aircraft, can provide a wealth
of valuable information for hydrologic engineering applications.
While this information is valuable, its uneven coverage over the landscape
can lead to challenges in hydrologic simulation. To that end, effort has been
undertaken to create gridded interpolations of the observed data fields. One
Fraction
of the most prominent gridded datasets is the PRISM product. PRISM was
developed at Oregon State University in 1994 (Daley et al, 1994). Currently,
PRISM maps are available at 4 km and 800 m for annual, monthly, and daily 0.00 0.15 0.30 0.45 0.60 0.75 0.90
values. Such data can be very helpful for distributed hydrologic models as Figure 37.5 Interannual variability of precipitation depicted through the coefficient
noted earlier. It should be noted that gridded data products are not perfect. of variation (mean divided by standard deviation) of National Weather Service
Lack of data in key locations, like mountain tops can lead to errors in the Cooperative Observer sites [Source: M. Dettinger 2011].
37_Singh_ch37_p37.1-37.10.indd 7 8/22/16 2:07 PM

The northern hemisphere jet stream and storm tracks tend to be displaced to Such fluctuations can be difficult to analyze with periods of records that
the south during winter and in some instances split into two paths impacting extend only for 100 years or so. To address this challenge, reconstruction of
winter precipitation patterns. Temperatures over the United States tend to be hydrologic time series has been made using paleo-proxies, such as tree-rings,
warmer than average. In its cold phase, or La Nina, the eastern tropical Pacific lake sediments, and chemical isotope signatures (see Bradley, 1985 or Cronin,
is cooler than normal and the pressure relationship between Darwin and 2010). Examples of such reconstructions include the works of Woodhouse et
Tahiti flips. The wintertime jet stream over the northern hemisphere develops al. (2006) for the Colorado River, and Meko et al. (2014) for the Sacramento,
a looping pattern with a high pressure system located somewhere over the San Joaquin, and Klamath Rivers.
Pacific Basin. The looping jet stream brings colder air into North America. If Along with variability over different time scales in time series of precipita-
the high-pressure system is located over the eastern Pacific or the west coast tion and temperature data sets is the appearance of trends. These trends are
of North America, drought conditions can set in as storms are diverted up long term, often monotonic changes in the distribution indicating a response
and over the high-pressure system away from much of the west coast. Such to changes in the atmosphere-ocean-land surface system. In urbanized areas,
expectations have been used to try to inform hydrologic engineering applica- the elevated night-time temperatures associated with the heat retention of the
tions (e.g., Dracup and Kahya, 1994; Piechota and Dracup, 1996). built environment is known as the urban heat island (Kim, 1992). Such
Variability can also occur over longer periods. Figure 37.6 depicts an changes have significant influence on the populations of those urbanized
annual time series of precipitation with an 11-year running average of pre- areas for health, energy use, etc. This has been documented in EPA (http://
cipitation for the state of California. These smoothed time series damp out the epa.gov/heatislands/resources/pdf/BasicsCompendium.pdf).
large year-to-year variability and indicate a fluctuation in precipitation out-
comes on the order of a decade. This same fluctuation also shows up in runoff 37.6 CLIMATE CHANGE
records both in California and the Colorado River Basins. While recognizable
in the smoothed observations, associations to specific processes in the climate Another type of change that has garnered increasing concern and attention
system have yet to be made. relates to changes in Earth’s radiation balance associated with the burning of
California statewide precipitation Oct–Sep
40
30
Inches
20
10
0
1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Solid line denotes 11-year running mean Year Western regional
climate center
Linear trend 1895–present + 2.45 ± 3.09 in. (+ 10 ± 13%) per 100 year
Linear trend 1949–present – 2.06 ± 8.75 in. (– 8 ± 38%) per 100 year
Linear trend 1975–present – 8.99 ± 21.46 in. (– 39 ± 93%) per 100 year
Wettest year 40.44 in. ( 176%) in 1983 Mean 22.90 in.
Driest year 9.23 in. ( 40%) in 1924 Stdev 6.53 in.
Oct–Sep 2014 12.05 in. ( 52%) Rank 3 of 119

Figure 37.6 Time series of precipitation with 11 year moving average from the California Climate Tracker (http://www.wrcc.dri.edu/monitor/cal-mon/index.html).
37_Singh_ch37_p37.1-37.10.indd 8 8/22/16 2:07 PM

Climate Change 37-9
Climate change effects on water resources
Total precipitation may increase or decrease
Increased
air temperature
Less snowpack
More precipitation as rain than snow

due to higher temperatures
Earlier runoff Increased evaporative demand on

from snow melt lakes and landscapes
Changes in timing and

amount of river flows
Changes in water resource

system operations
Sea level rise
Figure 37.7 Schematic of expected hydrologic impacts from climate change.
fossil fuels that began with the industrial revolution. The burning of fossil these efforts national and state-level efforts have been undertaken to attempt
fuels has increased the concentration of carbon dioxide and other green- to document and plan for a changing climate and its associated impacts.
house gases in the atmosphere which is effective in trapping earth’s heat in Examples include the United States third national assessment competed in
the atmosphere. As a result, temperatures have risen over the twentieth cen- 2014 (USGCRP, 2014) and the California Department of Water Resources
tury and look to continue to do so through the twenty-first century. These report in (2006). Efforts will continue as more impacts manifest themselves.
changes in temperature will impact the way weather events develop and Hydrologic Engineering will need to respond to the challenge by developing
evolve in space and time. As a consequence of those changes, significant new approaches to designing infrastructure that will be resilient to climate
hydrologic changes are expected which will profoundly impact hydrologic change impacts and work across a range of possible outcomes.
engineering.
A summary of expected hydrologic impacts due to climate change are illus- 37.6.1 Concluding Remarks
trated in Fig. 37.7. Higher temperatures will lead to more precipitation falling Weather and climate play important roles as drivers of hydrologic engineering
as rain than snow which will change the timing and magnitude of watershed design and applications. Climate sets the stage for weather events to unfold
runoff as well as increase evaporative demand on the landscape. The relation- across the landscape and is driven by energy balances in the land-ocean-atmo-
ships between infiltration and groundwater recharge, and surface and subsur- sphere system. Weather is the time evolution of the atmosphere across the earth.
face flow may change too as snow pack accumulation and melt patterns Precipitation, temperature, wind, solar radiation, and relative humidity are
change, vegetation changes, and soil moisture amounts and timing change. weather variables observed to provide input into hydrologic tools and applica-
Inflows into reservoirs will change which will impact water resources man- tions. Knowledge of how the processes work that control the magnitude and
agement in basins served by those reservoirs. Reservoir operations rules variation of these variables enables the engineer to accurately judge the right
derived from past hydrologic relationships may not be as effective. At the method for the job at hand.
downstream end, sea level rise will impact coastal infrastructure, increase Understanding how climate modulates the weather and evolves in response
seawater intrusion into coastal aquifers, and may cause backwater effects dur- to changes in the ocean-atmosphere-land system can provide insight into the
ing floods. conditions hydrologic systems will face and how the hydrologic system
Much of the work documenting observed change and anticipating future responds. This can be beneficial for planning for and managing extreme
change has been done in the past 25 years. The Intergovernmental Panel on events like floods and droughts as well as for identifying important thresholds
Climate Change (IPCC) has issued a series of reports beginning in 1990 with for different resource management objectives. Anticipating changes in hydro-
updates approximately every 5 years (http://www.ipcc.ch/index.htm). These logic conditions in response to climate change enables adaptive strategies to
reports are the collaborative efforts of scientists from many nations. Using be developed for a more robust and resilient civilization.
37_Singh_ch37_p37.1-37.10.indd 9 8/22/16 2:07 PM

REFERENCES Kuichling, E., “The relation between the rainfall and the discharge of sew-
ers in populous districts,” Transactions, American Society of Civil Engineers,
Anagnostou, E. N., M. N. Anagnostou, W. F. Krajewski, A. Kruger, and
20: 1–56, 1889.
B. J. Miriovsky, “High-resolution rainfall estimation from X-band polarimet-
Lundquist, J. and A. Flint, “Onset of snowmelt and streamflow in 2004 in
ric radar measurements,” Journal of Hydrometeorology, 5: 110–128, 2004.
the western United States: how shading may affect spring streamflow timing
Arguez, A., I. Durre, S. Applequist, R. S. Vose, M. F. Squires, X. Yin, R. R.
in a warmer world,” Journal of Hydrometeorology, Vol. 7: 1199–1216, 2006.
Heim, Jr., et al., “NOAA’s 1981–2010 U.S. Climate Normals: An Overview,”
Meko, David M., Connie A. Woodhouse, and Ramzi Touchan, Klamath/
Bulletin of the American Meteorological Society, 93: 1687–1697, 2012.
San Joaquin/Sacramento Hydroclimatic Reconstructions from Tree Rings.
Austin, P. M., “Relation between measured radar reflectivity and surface
Report to California Department of Water Resources, 2014, 117 pages.
rainfall,” Monthly Weather Review, 115: 1053–1071, 1987.
NOAA Atlas 2, Precipitation Frequency Atlas of the United States, 1973.
Baeck, M. L. and J. A. Smith, “Rainfall estimation for the WSR-88D for
NOAA Atlas 14, Precipitation Frequency Atlas of the United States, 2006.
heavy rainfall events,” Weather and Forecasting, 13: 416–436, 1998.
NOAA Atlas 14, Precipitation Frequency Atlas of the United States Volume 6
Bedient, P. B. and W. C. Huber, Hydrology and Floodplain Analysis, Addison
California, 2011, Vol. 6.
Wesley, Reading, MA, 1988.
Painter, T. H., D. F. Berisford, J. W. Boardman, K. J. Bormann, J. S. Deems,
Bradley, Raymond S., Quaternary Paleoclimatology: Methods of Paleoclimatic
F. Gehrke, P. Kirchner, et al., “The airborne snow observatory: fusion of imag-
Reconstruction, Allen & Unwin, Boston, 1985, ISBN 0-04-551067-9.
ing spectrometer and scanning LiDAR for studies of mountain snow cover,”
Brutsaert, W., Hydrology—An Introduction, Cambridge University Press,
Remote Sensing of Environment, 2015, in review.
Cambridge, UK, 2005, ISBN-13: 978-0521824798
Pedlosky, J., Geophysical Fluid Dynamics, Springer-Verlag, New York, NY,
California Department of Water Resources, “Progress on incorporating
1987.
climate change into management of California’s water resources,” Technical
Philander, S., El Niño, La Niña and the Southern Oscillation, Academic
Memorandum Report, 2006.
Press, London, UK, 1989.
Carlson, T., Mid-latitude Weather Systems, Routledge, London, UK, 1991.
Piechota, T. and J. Dracup, “Drought and regional hydrologic variation in
Colorado Water Institute, Workshop on Nonstationarity, Hydrologic
the United States: Associations with El Niño-Southern Oscillation,” Water
Frequency Analysis, and Water Management, January 13–15, 2010 Boulder,
Resources Research, 32 (5): 1359–1373, 1996.
CO. Colorado Water Institute Information Series No. 109, 2010.
Ralph, F. M., E. Sukovich, D. Reynolds, M. Dettinger, S. Weagle, W. Clark,
Cronin, Thomas N., Paleoclimates: Understanding Climate Change Past and
and P. J. Neiman, “Assessment of extreme quantitative precipitation forecasts
Present, Columbia University Press, New York, 2010, ISBN 978-0-231-
and development of regional extreme event thresholds using data from HMT-
14494-0.
2006 and COOP observers,” Journal of Hydrometeorology, 11: 1286–1304, 2010.
Cushman-Roisin, B., Introduction to Geophysical Fluid Dynamics, Prentice
Ralph, F. M. and M. Dettinger, “Historical and national perspectives on
Hall, New York, NY, 1994.
extreme west coast precipitation associated with atmospheric rivers during
Daly, C., R. P. Neilson, and D. L. Phillips, “A statistical-topographic model
December 2010,” Bulletin of the American Meteorological Society, 93: 783–790,
for mapping climatological precipitation over mountainous terrain,” Journal
2012.
of Applied Meteorology, 33: 140–158, 1994.
Rhea, O., Orographic precipitation model for hydrometeorological use.
Dracup, J. A. and E. Kahya, “The relationships between U.S. streamflow
Atmospheric Paper 278, Colorado State University, Fort Collins, CO, 198
and La Niña events,” Water Resources Research, 30 (7): 2133–2141, 1994.
pages, 1978.
Gill, A., “Atmosphere-ocean dynamics,” International Geophysics Series,
Singh, V. P. and D. K. Frevert (Eds.), Mathematical Models of Small
Vol. 30, Academic Press, London, UK, 1982.
Watershed Hydrology and Applications. 950 pp., Water Resources
Grotjahn, R., Global Atmospheric Circulations Observations and Theories,
Publications, Highlands Ranch, Colorado, 2002a.
Oxford Press, Oxford, UK, 1993.
Singh, V. P. and D. K. Frevert (Eds.) Mathematical Models of Large
Haltiner, G. and R. Williams, Numerical Prediction and Dynamic
Watershed Hydrology. 891 pp., Water Resources Publications, Highlands
Meteorology, 2nd ed., John Wiley & Sons, New York, NY, 1980.
Ranch, Colorado, 2002b.
Holton, J. R., An Introduction to Dynamic Meteorology, 3rd ed., Academic
Shorack, G. R. and J. A. Wellner, Empirical Processes with Applications to
Press, London, UK, 1992.
Statistics. New York: John Wiley & Sons, 1986, ISBN 0-471-86725-X.
Hosking, J. R. M. and J. R. Wallis, Regional Frequency Analysis: An
USGS, Guidelines for determining flood flow frequency, Bulletin 17B of
Approach Based on L-moments, Cambridge University Press, Cambridge, UK,
the Hydrology Subcommittee Interagency Advisory Committee on Water
2005, p. 3, ISBN 0521019400.
Data, US Department of the Interior, 1982.
IPCC, Climate Change 1995 The Science of Climate Change Contribution
USGCRP, Climate Change Impacts in the United States. US National
of Working Group I to the Second Assessment Report of the Intergovernmental
Climate Change Assessment, 2014.
Panel on Climate Change, Cambridge University Press, 1995.
White, A. B., M. L. Anderson, M. D. Dettinger, F. M. Ralph, A. Hinojosa,
Interagency Advisory Committee on Water Data, March 1982. Bulletin
D. R. Cayan, R. K. Hartman, et al., “A 21st century California observing net-
17B, “Guidelines for Determining Flood Flow Frequency,” Published by the
work for monitoring extreme weather events,” Journal of Atmospheric and
U.S. Department of the Interior, Geologic Survey.
Oceanic Technology, Vol. 30: 1585–1603, 2013.
Kavvas, M. L., S. Kure, Z. Chen, N. Ohara, and S. Jang, “WEHY-HCM for
Woodhouse C. A., S. T. Gray, and D. M. Meko, “Updated streamflow recon-
modeling interactive atmospheric-hydrologic processes at watershed scale. I:
structions for the Upper Colorado River basin,” Water Resources Research, 42:
Model Description,” Journal of Hydrologic Engineering, 18 (10): 1262–1271,
W05415, 2006, doi:10.1029/2005WR004455.
2013.
Zhu, Y. and R. E. Newell, “A proposed algorithm for moisture fluxes from
Kim, H. H., “Urban heat island,” International Journal of Remote Sensing, 13
atmospheric rivers,” Monthly Weather Review, 126: 725–735, 1998.
(12): 2319–2336, 1992.
37_Singh_ch37_p37.1-37.10.indd 10 8/22/16 2:07 PM

Chapter 38
Hydroclimatology: Global
Warming and Climate Change
BY
JOHN W. NIELSEN-GAMMON
ABSTRACT year. The Climate Prediction Center has developed the concept of optimal
climate normals (Huang et al., 1996), whereby the averaging interval is
Many factors cause the global and local climate to change, leading to ongoing
selected to be the interval (usually 10–15 years) whose average conditions
and future impacts to many aspects of the hydrologic cycle such as peak
provides the most accurate prediction of future average conditions. If the rate
intensity and spatial distribution of precipitation. Climate models are able to
of change is large and the natural variability is small, the optimal averaging
simulate many of the observed changes, particularly at larger scales, and are
interval will tend to be small too. Temperature typically has a systematic trend
an important tool for estimating future changes according to projections of
that is fairly large compared to interannual variability, so short averaging
future greenhouse gas concentrations and other change agents. Using model
periods work best for temperature. Precipitation has relatively large natural
projections for hydrologic applications requires applying appropriate down-
variability and thus requires longer averaging times.
scaling techniques to convert the model-simulated variables to hydrologic
Shortening the averaging time is a luxury not available to one who wishes
information specific to a particular information need at a particular location.
to estimate the probability of rare events such as major floods. Such events are
so rare that data from as many years as possible are needed to accurately
estimate the magnitude of extreme events. This is one important reason that
38.1 INTRODUCTION: THE AMBIGUITY OF CLIMATE rare event probabilities are particularly subject to inaccuracies in the case of
Climate can be succinctly defined as the statistical properties of the weather. climate change. Another important reason is that rare event probabilities can
Climate change, in turn, is defined by the American Meteorological Society change dramatically if there is merely a slight shift in the central value of the
(AMS, 2015) as “any systematic change in the long-term statistics of climate probability distribution (Fig. 38.1).
elements (such as temperature, pressure, or winds) sustained over several We can now define climate change from an engineering perspective: a
decades or longer.” change in the statistical properties of the weather such that perfect past obser-
These definitions leave plenty of room for ambiguity. Climate statistics vations would be insufficiently representative of statistical properties over a
incorporate the variability associated with such phenomena as El Niño, which design lifetime. The general statistical term for such a situation is nonstation-
is a coupled ocean-atmosphere phenomenon with a period ranging from less arity. Climate change is presently proceeding at a sufficiently rapid pace that,
than 2 years up to 8 years (Wittenberg, 2009). Some slower modes of vari- for many engineering applications, stationarity is no longer an appropriate
ability, such as the Atlantic Multidecadal Oscillation, have characteristic assumption (Milly et al., 2008).
periods greater than 50 years (Gray et al., 2004), and the weather variations The remainder of this chapter assumes that nonstationarity is a concern
associated with them may thus persist for several decades. This would techni- and discusses the physical processes that are causing present-day climate
cally meet the preceding definition of climate change. change, observations of climate change impacts already present in the hydro-
Major volcanic eruptions spew tiny solid and aqueous particles, collec- logic cycle, and the use of global climate model (GCM) output for estimating
tively known as aerosols, into the stratosphere, where they reduce incoming possible future climate changes.
solar radiation for several years. A single volcanic eruption is not generally
38.2 NATURAL AND HUMAN INFLUENCES ON
said to cause climate change, but a period of enhanced volcanic activity
PRESENT-DAY CLIMATE
appears to have helped bring about the Little Ice Age. If the Little Ice Age can
be regarded as an example of climate change, it is difficult to see where the The knowledge that the climate is presently changing at an historically rapid
transition from ordinary climate variability to climate change occurs. rate comes not just from comparisons between recent climate observations
An engineer designing a weather-sensitive project such as a flood control and reconstructions of past climate, but also from observations and calcula-
structure typically has a closely related definition of climate: the expected tions of the magnitude of processes driving imbalances in the climate system.
statistics of weather throughout the design lifetime of the project. As long as The combined rate and magnitude of warming is apparently larger than has
climate change is not taking place, such statistics may be obtained through happened for several millennia, and the magnitude of those processes driving
historical observations and related information. Even with climate change, climate change are collectively much larger than any such drivers over the
the change may be slow enough that the historical observations may still past several millennia as well (Masson-Delmotte et al., 2013).
provide a suitable estimate of future statistics. Global-scale changes in climate are most directly produced by changes in
The conventional period established by the World Meteorological the radiative balance of the climate system. The Earth receives electromagnetic
Organization for calculating climatic averages is 30 years (Arguez and Vose, radiation from the Sun and emits electromagnetic radiation into space. The
2011). If the climate is changing, the weather observations taken 30 years ago majority of the received radiation is absorbed by the Earth’s surface and the
are less relevant to present-day and future climate than observations from last vast majority of emitted radiation escaping to space is emitted by greenhouse
38-1
38_Singh_ch38_p38.1-38.8.indd 1 8/22/16 2:07 PM

38-2 Hydroclimatology: Global Warming and Climate Change
0.7 Land use changes appear to have had a small net cooling effect. Some warm-
0.6 ing was contributed by an increase in solar intensity, though intensity seems
to have peaked around the middle of the twentieth century and has declined
0.5 since the 1970s.
0.4 Of these energy imbalances, the easiest to quantify are imbalances due to
atmospheric gaseous composition changes. Most difficult are imbalances due
0.3
to suspended particles, because these particles may strongly affect the radia-
0.2 tive properties of clouds and top-down global observations of cloud proper-
0.1 ties are only available for the past few decades. Because gas and suspended
particle concentrations have increased in tandem, it is nearly impossible to
0
–4 –3 –2 –1 0 1 2 3 4
separate out their influences through statistical analysis of changes in past
–0.1 temperatures. Likewise, because the direct and indirect radiative impacts of
–0.2 suspended particles is only known to within a factor of 10, the total energy
imbalance driving global warming is not known very accurately (Trenberth et
–0.3 al., 2014b). This makes estimating the magnitude of future climate changes
–0.4 from observations of recent changes difficult.
Original PDF PDF with new mean
Other methods of estimating climate sensitivity have their own limitations.
Any GCM can be driven by specified processes that create an energy imbal-
Change in low extremes (*10) Change in high extremes (*10)
ance and the size of the simulated climate response can be determined from
Figure 38.1 The change in the frequency of extreme events (plus or minus two stan- the model simulation. At present, different climate models exhibit different
dard deviations) associated with a change in the mean. Note that the probability change is climate sensitivities, and there is no guarantee that the actual climate sensitiv-
asymmetric, with roughly twice as large an increase in high extreme events as the decrease ity lies within the range of sensitivities of the different climate models. The
in low extreme events.
differences in sensitivity are primarily due to the effects of simulated clouds,
which for a number of reasons are particularly challenging for GCMs
(Boucher et al., 2013). Climate sensitivity may also be estimated from past
gases and clouds; the energy loop is closed by net transfer energy from the climate changes, but both the size of the energy imbalance and the size of the
surface to the atmosphere by primarily nonradiative processes. resulting climate change must be estimated from the limited geological evi-
Any overall difference between what is absorbed and what is emitted will dence available.
cause the climate system to gain or lose energy. Over multiple years, this bal- The overall climate change response to an energy imbalance involves
ance is so exact that the present global warming is estimated to be caused by changes to the temperature, humidity, cloud cover, snow cover, atmospheric
an energy imbalance smaller than one part in 300 (Trenberth et al., 2014b). and ocean circulations, and other aspects of the climate system. Atmospheric
The fundamental reason for this stability is a local equilibrium in the cli- changes happen rapidly, within hours to weeks of a sudden energy balance
mate system: Any gain in energy will cause the temperature of the atmosphere change. The response of the ocean is more gradual, and the atmosphere con-
to rise, thereby increasing the rate at which energy is emitted to space. tinues to change in response to ocean surface changes. Other changes are even
Likewise, a drop in temperature produces a decrease in emitted energy. Any slower, such as the increase in energy absorption caused by the retreat of the
energy imbalance causes a response in the climate system that reduces that major ice sheets of Greenland and Antarctica. These changes are called feed-
imbalance. backs, and their individual impacts can also be separately quantified, though
Past long-term climate changes all seem to have involved changes in one or for our purposes it is more useful to think of the climate system as an inte-
more of the following factors affecting the planetary energy balance: atmo- grated whole with a complex set of time-dependent responses on the way
spheric gaseous composition, snow and ice cover, solar intensity, volcanic toward a new energy equilibrium.
activity, vegetation distribution, and atmospheric dust. For example, during Many of the important time-dependent responses involve changes to the
the last glacial maximum about 21,000 years ago, when much of North hydrological cycle, and such changes are discussed next.
America and Europe was covered by thick ice sheets, snow and ice cover and
atmospheric gaseous composition were both having a strong cooling effect 38.3 IMPACTS OF CLIMATE CHANGE ON THE
compared to present-day climate, while vegetation distribution and atmo- HYDROLOGICAL CYCLE IN THE TWENTIETH AND
spheric dust were having a weak cooling effect (though the actual size of the TWENTY-FIRST CENTURIES
dust effect is difficult to pin down) (Schmittner et al., 2011).
Although multiple factors are usually at work, the root cause of a climate 38.3.1 Evapotranspiration
change is typically a single process or event. For example, the recovery from The rate of evaporation from the oceans into the atmosphere depends primar-
the glaciation 21,000 years ago is thought to have been triggered by orbit- ily on the wind speed and on the saturation vapor pressure deficit, which is
caused changes in the relative amount of sunlight received in summer versus the difference between the actual partial pressure of water vapor and the
winter, which caused an increase in summertime melting of the Northern partial pressure of water vapor at 100% relative humidity at the actual tem-
Hemisphere ice sheets. Because of their sheer size after nearly 100,000 years perature. According to basic thermodynamics, an increase in surface tem-
of accumulation, it appears that they pushed the continents down far enough perature would lead directly to an increase in saturation vapor pressure and
that melting could take place rapidly at relatively low elevations (Abe-Ouchi hence an increase in saturation vapor pressure deficit, which in turn would
et al., 2013). Atmospheric carbon dioxide then began rising through some imply an increase in the evaporation rate, all else being equal.
combination of outgassing due to warmer temperatures and changes in the GCMs predict an increase in evaporation rate over the oceans averaging a
ocean circulation that brought carbon dioxide–rich waters to the surface. As bit less than 2% per degree of warming (Held and Soden, 2006), which is
rainfall patterns changed, vegetation became more widespread and atmo- much smaller than the 7% increase in saturation vapor pressure per degree.
spheric dust decreased. By the time the last of the continental North American As is discussed in the following text, models tend to increase tropospheric
ice sheet melted, this particular climate change had been complete. humidity at 7% per degree. This difference implies an overall reduction in
Of those four factors, the three that have changed substantially in the past wind speed. Although wind speed variations can be highly localized and
few hundred years are atmospheric composition, vegetation, and suspended noisy, a comprehensive survey of papers documenting wind speed changes
particles. All three changes are in this instance due to humanity’s activities finds that globally averaged surface wind speeds seem to have decreased
rather than orbital triggering of ice sheet changes. The specific causal mecha- (McVicar et al., 2012).
nisms are fossil fuel burning (affecting composition and suspended particles) Over land, much of the flux of water into the atmosphere is via transpira-
and agriculture (affecting all three). tion. In addition to the same constraints on water vapor fluxes as over the
The largest present-day imbalance by far is caused by changes in atmo- ocean, transpiration is also affected by changes in plant cover and the partly
spheric composition, specifically increases in the concentrations of carbon temperature-dependent regulation of transpiration by plants. Atmospheric
dioxide, methane, chlorofluorocarbons, nitrous oxide, ozone, and other com- increases in carbon dioxide additionally have two primary direct but compet-
pounds (Myhre et al., 2013). These increases have a warming effect because ing effects on transpiration. First, stomata need not open as much to intake
they decrease the amount of energy emitted to space. Next in present-day the same amount of carbon dioxide, thereby allowing plants to lose less water
importance is the increase in suspended particles, which by itself would have vapor through transpiration. Second, the resulting increased plant growth
a cooling effect due to a reduction in incoming solar radiation. Some sus- would create larger plants, which in turn would transpire more. Direct mea-
pended particles absorb solar radiation and have their own warming effect. surements of evapotranspiration over land are sparse and quite spatially vari-
38_Singh_ch38_p38.1-38.8.indd 2 8/22/16 2:07 PM

Global Climate Models 38-3
able, but overall they indicate an increase in evapotranspiration (Hartmann et declines are particularly important for humans (implying a longer dry season)
al., 2013). The average rate of change in evapotranspiration is not known with and for the climate system (the change in reflectivity of the Earth’s surface
sufficient precision to compare against the evapotranspiration increases sim- near the summer solstice affects Earth’s radiation balance).
ulated by climate models, but observations and climate models are consistent Changes in extreme precipitation can be distinct from those in mean pre-
with respect to the spatial pattern of evapotranspiration changes. cipitation. Because extreme precipitation, by its nature, is more erratic,
Over water, the various specific humidity data sets are too inconsistent with observed trends are less spatially coherent than trends in mean precipitation.
each other to permit a statement on humidity trends there (Kent et al., 2014). However, increasing temperatures directly imply an increase in the maximum
In the free troposphere (between about 1 km and 12–15 km above ground water vapor in the atmosphere and hence the maximum amount of rainfall,
level), and at ground level over land, average humidity seems to be increasing so an increase in extreme precipitation is a robust expectation from rising
at something close to the 7% per degree warming change in saturation water temperatures. Globally, there are more places where heavy precipitation is
vapor content (Hartmann et al., 2013; Willett et al., 2010). Climate models increasing than decreasing. Across much of the United States, for example,
consistently predict that relative humidity at the surface and troposphere 1- and 2-day extreme rainfall events are becoming much more frequent
ought to increase at about 7% per degree as well (Collins et al., 2013; Willett (DeGaetano, 2009; Janssen et al., 2014; Walsh et al., 2014). GCMs tend to
et al., 2010). underestimate this particular increase (Janssen et al., 2014). Overall, models
Wind patterns in the atmosphere can alter the distribution of water vapor have wide spread but are reasonably centered around the average observed
within the atmosphere, and are closely related to the production of precipita- rate of increase in extreme 1-day rainfall in the Northern Hemisphere mid-
tion. In the tropics, the primary large-scale circulation features are the Hadley latitudes, but they tend to underestimate the rate of increase in extreme
cells, which feature upward motion in convective regions near the equator and rainfall in the tropics (Asadieh and Krakauer, 2015).
downward motion over the subtropics. During the past several decades, the
Hadley cells have become taller and the poleward edges of the Hadley cells 38.3.3 Land Surface Hydrology
have migrated toward the poles (Hartmann et al., 2013), implying an expan- In the absence of a global soil moisture network or long-term satellite-based
sion of the Hadley cells and associated expansion of areas of low relative estimates of soil moisture, model-simulated changes in soil moisture are
humidity from the subtropics toward the midlatitudes. Climate models mostly unverified. The most prominent decreases in soil moisture are pro-
likewise predict an expansion of the Hadley cells, but the observed expansion jected to occur in the Mediterranean, southwestern North America, and
has been greater than that predicted by model simulations. The discrepancy southern Africa (Collins et al., 2013). All three of these changes are consistent
has been attributed to a specific form of multi-decadal natural variability in with the expansion of the Hadley cell and resulting changes in precipitation
the Pacific Ocean, which alters the pole-to-equator temperature gradient and and humidity.
thus affects the extent of the Hadley cells (Allen et al., 2014; Adam et al., 2014). Drought seems to be increasing globally, but with large regional variations.
Over land, the surface relative humidity has stayed roughly constant. The lack of confidence regarding globally averaged drought trends is caused
Climate models predict a decrease, which can be understood as caused by the in part by the divergent trends in global precipitation analyses and in part by
enhanced warming over land due to radiative changes (Sherwood et al., 2010) challenges with estimating potential evapotranspiration in a world with a
and, particularly in the subtropics, by the increased desiccation of air as it is changing climate and changing atmospheric composition (Trenberth et al.,
carried to higher altitudes within the Hadley cell before subsiding back down 2014a).
to the surface (Lau and Kim, 2015). Global observed streamflow trends are mixed, and a comprehensive esti-
mate of streamflow trends is difficult because many major rivers are ungauged
38.3.2 Precipitation (Hartmann et al., 2013). A modeled increase in streamflow can be inferred,
Precipitation on a global and annual basis is constrained to equal evaporation, even for those models without a hydrologic routing component, by the simu-
since the atmosphere only stores about 1 or 2 weeks’ worth of precipitation. lated increase in precipitation minus evaporation over land (Collins et al.,
Thus, if one is known precisely, so is the other. Unfortunately, neither is 2013). There is no detected trend in flood frequency globally, which is consis-
known precisely, due to the difficulty of measuring evaporation over land and tent with the lack of a streamflow trend but inconsistent with the observed
of measuring precipitation over water. New satellite-based precipitation mea- increase in extreme rainfall. Presumably, getting from extreme rainfall to
surements will help close the budget, but it will take decades before the data flooding involves mediation from flood control measures and better local
are extensive enough to pin down the globally averaged trend. water management. Substantially decreased streamflow is simulated for the
Model simulations of precipitation resemble the observed global pattern in Mediterranean, southern Africa, and the Middle East. Despite the decline in
most areas (Flato et al., 2013). Notable exceptions are a tendency to simulate simulated soil moisture and earlier findings, southwestern North America is
too little precipitation in the western tropical Pacific and to simulate too not projected to have a substantial decline in streamflow in the latest model
much precipitation over the tropical oceans south of the equator. simulations (Collins et al., 2013).
The general consensus from the various global rainfall data sets is that There is a substantial observed and simulated increase in streamflow at high
rainfall seems to be increasing globally on average (Hartmann et al., 2013). latitudes, with contributions from increased precipitation and melting of semi-
There are particular places in the Northern Hemisphere, such as the central permanent ice in glaciers and ice sheets. This is one of the main components of
United States, where there is a clear positive trend. Together with the 2% global sea level rise, along with thermal expansion. Melting of ice from
increase in evaporation per degree of warming in climate models is a 2% Greenland and Antarctica is an increasing component of sea level rise, and
increase in precipitation per degree of warming (Collins et al., 2013). These there is an additional contribution from aquifer depletion (Church et al., 2013).
trends are consistent with observations, but the range of estimated observed
precipitation increases is so large that the observations do not constrain the 38.3.4 A Comment on Extremes
models. The dependence of observed rainfall trends on latitude seems to be The preceding discussion of droughts, floods, and extreme rainfall should
simulated fairly well by models. Both observations and models indicate a make it clear that it is difficult to generalize about extreme weather event
decrease of precipitation in the subtropics (except for summertime) and an trends. For example, the general statement “global warming causes more
increase at higher latitudes that is largest in wintertime. Although this pattern droughts and floods” is quite misleading. While droughts are increasing on
is frequently generalized as “wetter places get wetter and drier places get average worldwide, the increase is by no means uniform, nor is the rate of
drier,” there are many places where that generality does not hold. increase known. Meanwhile, floods are not increasing on a global average
In the models, the subtropical change is associated with the changes in basis, and extreme rainfall increases are only guaranteed to lead to flooding
Hadley cell discussed earlier, while the change at higher latitudes is due to the increases if societal or infrastructure changes are not at the same time reduc-
increased moisture in a warmer world. Regional variations in trends are in ing the flooding potential. Many actions taken by society, either intentionally
many areas considerably different in observations compared to models. or inadvertently, affect flood risk, and one shouldn’t generalize to an overall
Multidecadal natural variability explains some, but not all, of the differences consequence when only one contributing factor is being considered. Even if
(Gu and Adler, 2015). floods increase in the future, the area of overlap, where both droughts and
Because of warmer temperatures, some of what used to fall as snow is now floods are increasing, will certainly be less than half of the globe.
falling as rain. This has implications, for example, for the snow pack in the
western mountains of North America in places where the rain-snow line has 38.4 GLOBAL CLIMATE MODELS
a strong influence on snow accumulation, and observed changes in spring-
time mountain snowpack are consistent with warming temperatures. The The preceding section compared observed trends in the water cycle to global
overall Northern Hemisphere snow cover extent is decreasing at a statistically climate model (GCM) output. This section provides some perspective on
significant rate in springtime; trends in fall and winter are for the most part climate model performance by discussing the components of a climate model
indistinguishable from zero (Hernández-Henríquez et al., 2015). Springtime and the sources of discrepancies between model output and reality.
38_Singh_ch38_p38.1-38.8.indd 3 8/22/16 2:07 PM

A state-of-the-art GCM is actually a set of models that exchange informa- asymptotic relaxation to a prescribed vertical distribution of heat and mois-
tion as they run. A minimal GCM will include an atmospheric dynamical ture or as complicated as a one-dimensional column model of convection
model, an ocean model, and a land surface model. Common additional com- complete with calculated entrainment/detrainment and updraft and down-
ponents include a snow model, a sea ice model, a hydrologic model, a vegeta- draft strengths at every level.
tion model, and an atmospheric chemistry model. GCMs are typically It has been a difficult challenge to create convective parameterizations that
designed so that some component models are optional and can be replaced by perform well in both the tropics and higher latitudes. Tropical convection is
steady-state assumptions or simplistic approximations. typically very deep and develops from weak instability. Midlatitude convec-
The atmospheric dynamical and ocean model both solve a set of equations tion is often much stronger because instability can increase substantially
known as the Navier-Stokes equations or an approximate version of them. above a capping inversion before the onset of convection. It is also common
These partial differential equations govern fluid flow and represent Newton’s for models to have a specialized convective parameterization for shallow
second law of motion, the first law of thermodynamics, conservation of mass, convection, the type of convection that typically forms at the top of the
and an equation of state, which for the atmosphere is the Ideal Gas Law. The boundary layer and produces little if any precipitation. Unfortunately,
equations describe how mass, momentum, and energy vary smoothly the convective parameterization is a critically important component of a
throughout the fluid, with the fluid continuously creating and responding to climate model, because the general circulation of the atmosphere is driven by
changes in forces or internal energy. The equations are exact if the atmo- the atmospheric convection and heat redistribution that takes place in the
sphere and ocean behave as an ideal fluid. tropics in response to solar heating. Models have difficulty reproducing even
To make numerical integration possible, the fluid state is approximated by such seemingly simple characteristics of convection as the local time of day at
a set of numbers that collectively describe variations in the state of the fluid which convection is most likely to occur.
(wind, temperature, pressure, moisture, or salinity) that are larger than a The different models have different strengths and weaknesses, but they also
certain spatial scale. This is done by dividing the atmosphere into a handful share some common strengths and weaknesses. The various GCMs cannot be
of levels (typically 25–40) and either creating a horizontal grid or using a assumed to span the range of reality, because reality does not include small-
truncated expansion of continuous functions such as spherical harmonics. scale processes that are entirely predictable, as is assumed by most parameter-
The neglect of finer-scale processes creates errors, in part because energy is izations. Even model ensembles in which the various parameters controlling
constantly cascading to finer scales—there is no clean scale separation the parameterizations are varied throughout their plausible numerical ranges
between phenomena that can be resolved and phenomena that cannot. Other cannot by their nature encompass reality. On the whole, then, climate models
errors are introduced when the equations are integrated forward over a finite produce simulations of weather and climate on imaginary planets that are
time interval. Various techniques are available for numerical integration in very much like our own.
time and derivative calculation in space, but tradeoffs must be made among
accuracy, desirable properties (such as conservation of energy or conserva- 38.5 WORKING WITH CLIMATE MODEL PROJECTIONS
tion of mass), and speed. Ultimately, the accuracy of the simulation is limited
by the available computer power. There are many sources of difficulty and inaccuracy in estimating future
Configured in this way, the GCM is not rigged or preprogrammed to climate states. For example, for estimating future flood frequency in
include specific large-scale phenomena such as cold fronts, El Niño, the England, Kay et al. (2009) list human emissions, GCM structure, internal
Hadley cells, the jet streams, the Gulf Stream, and so forth, except that some variability, regional climate model structure, downscaling, hydrologic
of them will exist in the starting conditions. Instead, the model continuously model structure, and hydrologic model parameters. The previous section
creates and regenerates these and other weather and climate features as emer- discussed uncertainties caused by GCM structure. This section considers
gent phenomena resulting from the laws of physics and thermodynamics uncertainties associated with human emissions, regional climate model
acting on a quasi-continuous fluid. Climate models are validated by how well structure, and/or downscaling.
they reproduce the climate of the past and present.
Some processes occurring at smaller scales have important, fundamental 38.5.1 Projections
influences on the larger-scale motions and thus cannot be completely neglected. For forecasts multiple decades into the future, neither initial atmosphere nor
To include the influences of these processes, modelers rely on relationships initial ocean conditions are relevant. The primary source of predictability of
between the local larger-scale conditions and the smaller-scale processes, in future climate states arises from knowledge of future changes in atmospheric
effect predicting or estimating the processes that the model cannot simulate radiative forcing. A true climate model forecast would consist of a GCM
directly based on what the model can simulate directly. These relationships, driven by a forecast of radiative forcing changes.
programmed into a climate model, are called parameterizations. Comprehensive forecasts of radiative forcing changes do not exist.
The most important parameterizations from a hydrological perspective are Volcanic eruptions are unpredictable, and solar variability is presently
those involving clouds and precipitation. Clouds are too small in scale to be unpredictable more than a few years in advance. Changes in the atmo-
simulated directly by climate models. Climate models must infer the presence spheric composition caused by human activities are believed to be much
and properties of clouds within a grid cell from the grid cell’s average values of more predictable over the near term, in part because predicting human
water vapor, temperature, and vertical motion. In many current versions of emissions is much like predicting the course of a supertanker: because it
climate models, the properties of aerosol particles that can serve as cloud con- takes so long to institute a significant change, continuation of the present
densation nuclei are also simulated and are taken into account by the cloud evolution for quite some time is likely. By the end of the twenty-first cen-
parameterization. The cloud parameterization also generates precipitation if the tury, though, a very wide range of emissions and subsequent atmospheric
clouds and the condensation rate within them are determined to be large conditions is possible.
enough. Condensation occurs if there is on average upward motion at a par- Estimates of possible ways that climate-relevant human activities might
ticular level in the grid box and at least a portion of that level within the grid change over time are called scenarios. In general, scenarios are based on
box is determined to be at or above saturation with respect to water vapor. assumptions of population, global wealth, commerce, and technology.
The cloud parameterization produces only part of the precipitation simulated So-called “business as usual” scenarios, such as the SRES (Special Report on
by a climate model. The remainder, including almost all of the precipitation in Emission Scenarios; Nakicenovic and Swart, 2000) scenarios, assume that
the tropics, is simulated by the convective parameterization. The convective there will be no intentional efforts to reduce emission for climate change
parameterization represents the effect of moist convection, which visually takes mitigation purposes, though emissions reductions due to air quality regula-
the form of cumulus or cumulonimbus clouds, including thunderstorms. The tions or technological improvements are included in the scenarios. Other
effects of moist convection include a vertical redistribution of heat, moisture, scenarios assume emissions are reduced as a result of policy decisions to limit
and, in some parameterizations, momentum; the possible generation of pre- human influence on climate.
cipitation; and the possible generation of hydrometeors that are tracked at the As of 2015, the two most recent coordinated sets of GCM runs for the next
grid scale and will be dealt with by the cloud parameterization. century are CMIP3 and CMIP5. (CMIP stands for Coupled Model
Convective parameterizations can be quite complicated. In general, most of Intercomparison Project.) The six SRES scenarios in CMIP3 were all busi-
them include two interacting components: the trigger function, which deter- ness-as-usual scenarios: A1fi, A2, A1B, A1T, B2, and B1 (Nakicenovic and
mines whether subgrid-scale convection is present, and the cloud model, Swart, 2000). They differed in their assumptions regarding demographics,
which determines the consequences of subgrid-scale convection. The trigger international cooperation, and so forth. Three of these scenarios have emis-
function is based on a calculation of the amount of instability or on the rate sions peaking and then declining by the end of the twenty-first century
at which instability is being generated, and it may also include an assessment through a combination of reduced population growth and reduced carbon
of the extent of inhibition to convection due to such features as lower-tropo- intensity. The IPCC report putting forth these scenarios advised that they
sphere temperature inversions. The cloud model may be as simple as an were all to be considered as equally likely.
38_Singh_ch38_p38.1-38.8.indd 4 8/22/16 2:07 PM

Working with Climate Model Projections 38-5
The CMIP5 group took a different approach (van Vuuren et al., 2011). or monthly precipitation over an externally defined 5000-km2 grid area.
Instead of specifying scenarios and calculating emissions, they specified sev- Instead, what will be of interest will be precipitation at a single gauge location,
eral sets of postulated emissions (called representative concentration path- or at a particular spatial resolution, or averaged over a particular drainage
ways, or RCPs) and then had scenarios created that were consistent with those basin, over a range of accumulation intervals that may include intervals
emissions. The RCPs were designed to span a wide range of possibilities. The shorter than the gaps between model output. The conversion of climate
RCP with the greatest greenhouse gas concentrations by the end of the model output from a coarse spatial and temporal scale to a fine spatial and
twenty-first century, RCP 8.5, is virtually identical to the higher SRES A1fi temporal scale is called downscaling.
business-as-usual scenario; most SRES scenarios and other business-as-usual There are two types of downscaling: statistical and dynamical. Statistical
scenarios have lower emissions than RCP 6.0 and RCP 8.5 through the middle downscaling is based on empirical relationships between modeled quantities
of the twenty-first century and finish the century with net emissions some- and the quantities of interest. Statistical downscaling is only possible if a
where between RCP 6.0 and RCP 8.5. At the low end, RCP 2.6 envisions such direct relationship exists between the desired model output variable and the
aggressive action to combat greenhouse gas emissions that by the last few actual output variables. Such a relationship can be established using observa-
decades of the twenty-first century there is net removal of carbon dioxide tions from the climate record; the task becomes much harder if such an
from the atmosphere. The final RCP, 4.5, lies in between RCP 2.6 and RCP 6.0 observational relationship is not readily available due to lack of appropriate
and includes fairly strong greenhouse gas emission reductions. observations. If observations exist, the statistical downscaling problem
There is no single correct scenario to use. Or rather, it is not known which becomes similar to the bias correction problem. We know the probability
scenario, if any, will turn out to be the correct one. One way to deal with this distribution of the desired variable, the present-day simulated distribution of
uncertainty is to use model output generated from several different scenarios. a corresponding model output variable, and the future simulated distribution
Another approach is to use model output generated from one of the high-end of the model output variable. As with bias correction, the task becomes one of
emissions scenarios (such as A1FI or RCP 8.5) and present the results in computing a transfer function between the present-day model output variable
terms of the amount of change per degree of global warming by a certain year distribution and the observed distribution of the desired variable, and apply-
(e.g., Collins et al., 2013; Swain and Hayhoe, 2015). To be confident with this ing that transfer function to the simulated future output variable distribution.
approach, it is useful to determine the extent to which the local hydrological Since the process is similar, the concerns are similar. An important addi-
parameters of interest change linearly with global mean surface temperature tional concern also arises: the stability of the transfer function between the
and are independent of the rate of change of surface temperature. observed variable and the model variable. One needs the two variables to be
so closely related that a certain value of the model variable would be expected
38.5.2 Bias Adjustment to correspond to the same value of the observed value in the present-day cli-
Ideally, a GCM will simulate climatic conditions exactly like those that are mate and in a climate that has undergone all the changes associated with
observed. In practice, this is impossible, partly because natural variability global warming. An example of two variables that would seem to satisfy this
prevents the true long-term average climate conditions from ever being per- condition is observed gauge precipitation and model-simulated grid-box
fectly known, but primarily because the models themselves are imperfect. average precipitation. But if the increased moisture content of a warmer
Some aspects of the climate system can be reproduced very accurately by atmosphere leads to greater concentration of rainfall in smaller areas, the
global climate models. Other aspects give climate models greater difficulty. mapping of peak grid-box average precipitation onto peak rain gauge pre-
Global precipitation distributions and amounts fall into the latter category. cipitation will change over time.
Suppose a GCM’s present-day simulation is 200 mm of precipitation, its Rather than relying on mapping of a single simulated variable to the observed
future simulation is 150 mm and present-day observations show 100 mm. variable of interest, one can use information from multiple simulated variables.
Even though the model’s future climate is wetter than the present-day climate Taken to its logical extreme, this leads to an approach known as weather typing,
we observe, should we interpret the model projection as indicating that pre- in which the observed weather pattern or patterns most closely corresponding
cipitation will decrease? If so, by how much? Bias adjustment says that the to the model-simulated situation is identified and the corresponding desired
model’s precipitation decreased by 25%, from 200 to 150 mm, so that we input variables are drawn from the observations for that day or days.
should apply a 25% adjustment to the observed climatic conditions, yielding The literature on statistical downscaling and bias adjustment is rapidly
a projected precipitation amount of 75 mm? The latter approach, applying a expanding as new techniques are developed and tested; a sampling of papers
multiplicative bias adjustment, is the most common one. relevant to hydrologic downscaling is listed here. Wilby et al. (2004) provides
In general, one is not interested merely in the future mean conditions, but an early but reasonably comprehensive guide to addressing many issues
also (or exclusively) in the future variability or extremes. Thus the problem related to statistical downscaling, including those mentioned previously.
can be stated: given an observed present-day probability distribution, a simu- Maraun et al. (2010) provides a more up-to-date and detailed overview
lated present-day probability distribution, and a projected probability distri- of downscaling issues specifically as applied to precipitation variables.
bution, how should one use the knowledge of the difference between the Teutschbein et al. (2011) provide a more detailed comparison of three
observed and simulated present-day probability distribution to improve the downscaling techniques as applied to hydrological impacts in a specific
accuracy of the projected probability distribution? There are a number of Scandinavian catchment. Ho et al. (2012) compare the bias adjustment
possible approaches. One common one is called empirical quantile mapping: paradigm described previously with a different but equally valid approach
Assume that the first percentile of the observed distribution corresponds to known as change factors and shows that large differences can result in the
the first percentile of the simulated present-day distribution and in turn to the downscaled output. Estrada et al. (2013) caution against the danger that the
first percentile of the simulated future distribution, and apply the same statistical assumptions inherent in the techniques applied to statistical
assumption to all percentiles. Then the bias between the observed and simu- downscaling may not commonly hold.
lated present-day percentiles may be computed and applied as an adjustment
to the future percentile values. 38.5.4 Dynamical Downscaling
The quantile approach seems fine in principle, but it easy to create patho- An alternative approach to obtaining the needed information from large-scale
logical cases. For example, consider observed precipitation ranging from 40 climate simulations is dynamical downscaling. Dynamical downscaling con-
to 60 mm, simulated present-day precipitation ranging from 20 to 80 mm, sists of using the output from the large-scale climate simulation to drive a
and simulated future precipitation ranging from 40 to 90 mm. At first glance, meteorological (or coupled) model over a limited area at much higher spatial
these seem like plausible numbers rather than pathological ones. At the high resolution, with output available at much higher temporal resolution.
end, the present-day observations are 75% of the simulated value, so the high The challenge of running a limited-area model can be daunting and usually
end in the future climate ought to be reduced to 90 × 0.75 = 67.5 mm. requires considerable computing power and expertise to avoid detrimental
Meanwhile, at the low end, the present-day observations are 200% of the impacts from phenomena such as model climate mismatch. Fortunately for
simulated value, so the low end in the future climate ought to be increased to the nonatmospheric scientist, several sets of model output from dynamical
40 × 2 = 80 mm. The low end of the adjusted precipitation distribution is downscaling are available online from various sources, so a new model run
greater than the high end of the adjusted precipitation distribution. An alter- may not be necessary.
native common approach is to separately consider changes in the mean and Dynamical downscaling provides a way forward when observations are not
standard deviation, but that too can lead to inconsistencies. available with which to apply statistical downscaling techniques. Even so, just
as the GCM will generally not accurately simulate the real climate, neither will
38.5.3 Statistical Downscaling the limited-area model. Also, if one wishes to produce point data, a statistical
The preceding discussion has assumed that one is interested in the future state downscaling step must be applied to the limited-area model output. So it is
of some variable that is both observable and appears in the model output. common for dynamical downscaling to include bias adjustment and statistical
Generally this will not be the case; it is rare to want to work with average daily downscaling components too.
38_Singh_ch38_p38.1-38.8.indd 5 8/22/16 2:07 PM

38.6 CONCLUSION A.D.,” Geophysical Research Letters, 31: L12205, 2004, doi:10.1029/2004

GL019932.
Hydrologists attempting to deal with future conditions are presently stuck in
Gu, G. and R. F. Adler, “Spatial patterns of global precipitation change and
a gray area. On the one hand, it is no longer possible in certain circumstances
variability during 1901-2010,” Journal of Climate, 28: 4431–4453, 2015,
to assume stationarity and rely exclusively on historical observations for esti-
doi:10.1175/JCLI-D-14-00201.1.
mating future conditions. On the other hand, techniques for estimating future
Hartmann, D. L., A. M. G. Klein Tank, M. Rusticucci, L. V. Alexander,
conditions under a changing climate are still in their infancy, and it is not
S. Brönnimann, Y. Charabi, F. J. Dentener, et al., “Observations:
clear whether all the important issues involved with such estimations have yet
Atmosphere and Surface,” Climate Change 2013: The Physical Science
been identified, let alone resolved into best practices. There do at least exist
Basis. Contribution of Working Group I to the Fifth Assessment Report of
standard approaches, such as those utilized by the National Climate
the Intergovernmental Panel on Climate Change, edited by T. F. Stocker, D.
Assessments to identify policy-relevant current and future climate changes on
Qin, G-K. Plattner, M. Tignor, S. K. Allen, J. Boschung, A. Nauels, et al.,
a subnational scale (Melillo et al., 2014). While it is tempting to treat down-
Cambridge University Press, Cambridge, United Kingdom and New York,
scaled climate information available online as coming from a black box, it is
New York, USA, 2013.
essential to be aware of the assumptions, tradeoffs, and sources of error inher-
Held, I. M. and B. J. Soden, “Robust responses of the hydrological cycle to
ent to each downscaled data set and to understand how those issues might
global warming,” Journal of Climate, 19: 5686–5699, 2006.
apply to the particular question and location at hand.
Hernández-Henríquez, M. A., S. J. Dery, and C. Derksen, “Polar amplifica-
REFERENCES
tion and elevation-dependence in trends of Northern Hemisphere snow cover
extent, 1971–2014,” Environmental Research Letters, 10: 044010, 2015,
Abe-Ouchi, A., F. Saito, K. Kawamura, M. E. Raymo, J. Okuno, K. Takahasi, doi:10.1088/1748-9326/10/4/044010.
and H. Blatter, “Insolation-driven 100,000-year glacial cycles and hysteresis of Ho, C. K., D. B. Stephenson, M. Collins, C. A. T. Ferro, and S. J. Brown,
ice-sheet volume,” Nature, 500: 190–193, 2013, doi:10.1038/nature12374. “Calibration strategies: A source of additional uncertainty in climate change
Adam, O., T. Schneider, and N. Harnik, “Role of changes in mean tempera- projections,” Bulletin of the American Meteorological Society, 93: 21–26, 2012,
tures versus temperature gradients in the recent widening of the Hadley cir- doi:10.1175/2011bams3110.1.
culation,” Journal of Climate, 27: 7450–7461, 2014, doi:10.1175/ Huang, J., H. M. van den Dool, and A. G. Barnston, “Long-lead seasonal
JCLI-D-14-00140.1. temperature prediction using optimal climate normals.” Journal of Climate,
Allen, R. J., J. R. Norris, and M. Kovilakam, “Influence of anthropogenic 9: 809–817, 1996, doi:10.1175/1520-0442(1996)009<0809:LLSTPU>2.0.CO;2.
aerosols and the Pacific Decadal Oscillation on tropical belt width,” Nature Janssen, E., D. J. Wuebbles, K. E. Kunkel, S. C. Olsen, and A. Goodman,
Geoscience, 7: 270–274, 2014, doi:10.1038/ngeo2091. “Observational- and model-based trends and projections of extreme precipi-
American Meteorological Society, Climate Change, Glossary of Meteorology, tation over the contiguous United States,” Earth’s Future, 2: 2014,
2015, Available online at http://glossary.ametsoc.org/wiki/Climate_change. doi:10.1002/2013EF000185.
Arguez, A. and R. S. Vose, “The definition of the standard WMO climate Kay, A. L., H. N. Davies, V. A. Bell, and R. G. Jones, “Comparison of uncer-
normal: The key to deriving alternative climate normals,” Bulletin of the tainty sources for climate change impacts: Flood frequency in England,”
American Meteorological Society, 92: 699–704, 2011, doi:10.1175/ 2010BAMS Climatic Change, 92, 41–63, 2009, doi:10.1007/s10584-008-9471-4.
2955.1. Kent, E. C., D. I. Berry, J. Prytherch, and J. B. Roberts, “A comparison of
Asadieh, B. and N. Y. Krakauer, “Global trends in extreme precipitation: global marine surface-specific humidity datasets from in situ observations
climate models versus observations,” Hydrology and Earth System Sciences, 19: and atmospheric reanalysis,” International Journal of Climatology, 34: 355–
877–891, 2015, doi:10.5194/hess-19-877-2015. 376, 2014, doi:10.1002/JOC.3691.
Boucher, O., D. Randall, P. Artaxo, C. Bretherton, G. Feingold, P. Forster, Lau, W. K. M. and K-M. Kim, “Robust Hadley Circulation changes and
V-M. Kerminen, et al., “Clouds and Aerosols,” Climate Change 2013: The increasing global dryness due to CO2 warming from CMIP5 model projec-
Physical Science Basis. Contribution of Working Group I to the Fifth Assessment tions,” Proceedings of the National Academy of Sciences, 112: 30630–3635,
Report of the Intergovernmental Panel on Climate Change, edited by T. F. 2015, doi:10.1073/pnas.1418682112.
Stocker, D. Qin, G-K. Plattner, M. Tignor, S. K. Allen, J. Boschung, A. Nauels, Maraun, D., F. Wetterhall, A. M. Ireson, R. E. Chandler, E. J. Kendon,
et al., Cambridge University Press, Cambridge, United Kingdom and New M. Widmann, S. Brienen, et al., “Precipitation downscaling under climate
York, New York, USA, 2013. change: Recent developments to bridge the gap between dynamical models
Church, J. A., P. U. Clark, A. Cazenave, J. M. Gregory, S. Jevrejeva, and the end user,” Reviews of Geophysics, 48: Rg3003, 2010,
A. Levermann, M. A. Merrifield, et al., “Sea Level Change,” Climate Change doi:10.1029/2009rg000314.
2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Masson-Delmotte, V., M. Schulz, A. Abe-Ouchi, J. Beer, A. Ganopolski, J. F.
Assessment Report of the Intergovernmental Panel on Climate Change, edited Gonz.lez Rouco, E. Jansen, et al., “Information from Paleoclimate Archives,”
by Stocker, T. F., D. Qin, G-K. Plattner, M. Tignor, S. K. Allen, J. Boschung, Climate Change 2013: The Physical Science Basis. Contribution of Working
A. Nauels, Cambridge University Press, Cambridge, United Kingdom and Group I to the Fifth Assessment Report of the Intergovernmental Panel on
New York, New York, USA, 2013. Climate Change, edited by T. F. Stocker, D. Qin, G-K. Plattner, M. Tignor,
Collins, M., R. Knutti, J. Arblaster, J-L. Dufresne, T. Fichefet, S. K. Allen, J. Boschung, A. Nauels, et al., Cambridge University Press,
P. Friedlingstein, X. Gao, et al., “Long-term Climate Change: Projections, Cambridge, United Kingdom and New York, New Year, USA, 2013.
Commitments and Irreversibility,” Climate Change 2013: The Physical Science McVicar, T. R., M. L. Roderick, R. J. Donohue, L. T. Li, T. G. Van Niel,
Basis. Contribution of Working Group I to the Fifth Assessment Report of the A. Thomas, J. Grieser, et al., “Global review and synthesis of trends in
Intergovernmental Panel on Climate Change, edited by T. F. Stocker, D. Qin, observed terrestrial near-surface wind speeds: Implications for evaporation,”
G-K. Plattner, M. Tignor, S. K. Allen, J. Boschung, A. Nauels, et al., Cambridge Journal of Hydrology, 416: 182–205, 2012, doi:10.1016/j.jhydrol.2011.10.024.
University Press, Cambridge, United Kingdom and New York, New York, Melillo, J. M., T. C. Richmond, and G. W. Yohe (Eds.), “Climate Change
USA, 2013. Impacts in the United States: The Third National Climate Assessment,” U.S.
DeGaetano, A. T., “Time-dependent changes in extreme-precipitation Global Change Research Program, 841, 2014, doi:10.7930/J0Z31WJ2.
return-period amounts in the continental United States,” Journal of Applied Milly, P. C. D., J. Betancourt, M. Falkenmark, R. M. Hirsch, Z. W.
Meteorology and Climatology, 48: 2086–2099, 2009, doi:10. 1175/2009 Kundzewicz, D. P. Lettenmaier, and R. J. Stouffer, “Stationarity is dead:
JAMC2179.1. Whither water management?,” Science, 319: 573–574, 2008, doi:10.1126/sci-
Estrada, F., V. M. Guerrero, C. Gay-Garcia, and B. Martinez-Lopez, ence.1151915.
“A cautionary note on automated statistical downscaling methods for climate Myhre, G., D. Shindell, F-M. Bréon, W. Collins, J. Fuglestvedt, J. Huang,
change,” Climatic Change, 120: 263–274, 2013, doi:10.1007/s10584-013-0791-7. D. Koch, et al., “Anthropogenic and Natural Radiative Forcing,” Climate
Flato, G., J. Marotzke, B. Abiodun, P. Braconnot, S. C. Chou, W. Collins, Change 2013: The Physical Science Basis. Contribution of Working Group I to
P. Cox, et al., “Evaluation of Climate Models,” Climate Change 2013: The the Fifth Assessment Report of the Intergovernmental Panel on Climate Change,
Physical Science Basis. Contribution of Working Group I to the Fifth Assessment edited by T. F. Stocker, D. Qin, G-K. Plattner, M. Tignor, S. K. Allen, J.
Report of the Intergovernmental Panel on Climate Change, edited by T. F. Boschung, et al., Cambridge University Press, Cambridge, United Kingdom
Stocker, D. Qin, G-K. Plattner, M. Tignor, S. K. Allen, J. Boschung, A. Nauels, and New York, New York, USA, 2013.
et al., Cambridge University Press, Cambridge, United Kingdom and New Nakicenovic, N. and R. Swart (Eds.), Emissions Scenarios, IPCC Special
York, New York, USA, 2013. Report, Cambridge University Press, U. K., 2000, p. 570.
Gray, S. T., L. J. Graumlich, J. L. Betancourt, and G. T. Pederson, “A tree- Sherwood, S. C., W. Ingram, Y. Tsushima, M. Satoh, M. Roberts, P. L.
ring based reconstruction of the Atlantic Multidecadal Oscillation since 1567 Vidale, and P. A. O’Gorman, “Relative humidity changes in a warmer climate,”
38_Singh_ch38_p38.1-38.8.indd 6 8/22/16 2:07 PM

References 38-7
Journal of Geophysical Research, 115: D09104, 2010, doi:10. 1029/2009 Van Vuuren, D. P., J. Edmonds, M. Kainuma, K. Riahi, A. Thomson,
JD012585. K. Hibbard, G. C. Hurtt, et al., “The Representative Concentration Pathways:
Schmittner, A., N. M. Urban, J. D. Shakun, N. M. Mahowald, P. U. Clark, An overview,” Climatic Change, 109: 148, 2011, doi:10.1007/s10584-011-0148-z.
P. J. Bartlein, A. C. Mix, et al., “Climate sensitivity estimated from temperature Walsh, J., D. Wuebbles, K. Hayhoe, J. Kossin, K. Kunkel, G. Stephens,
reconstructions of the Last Glacial Maximum,” Science, 334: 1385–1388, 2011, P. Thorne, et al., “Chapter 2: Our Changing Climate,” Climate Change Impacts
doi:10.1126/science.1203513. in the United States: The Third National Climate Assessment, edited by J. M.
Swain, S. and K. Hayhoe, “CMIP5 projected changes in spring and summer Melillo, T. C. Richmond, and G. W. Yohe, U.S. Global Change Research
drought and wet conditions over North America,” Climate Dynamics, 44: Program, 2014, 19–67, doi:10.7930/J0KW5CXT.
2737–2750, 2015, doi:10.1007/s00382-014-2255-9. Wilby, R. L., S. P. Charles, E. Zorita, B. Timbal, P. H. Whetton, and L. O.
Teutschbein, C., F. Wetterhall, and J. Seibert, “Evaluation of different down- Mearns, Guidelines for Use of Climate Scenarios Developed from Statistical
scaling techniques for hydrological climate-change impact studies at the Downscaling Methods, Intergovernmental Panel on Climate Change, Zurich,
catchment scale,” Climate Dynamics, 37: 2087–2105, 2011, doi:10.1007/ Switzerland, 2014, p. 27.
s00382-010-0989-8. Willett, K. M., P. D. Jones, P. W. Thorne, and N. P. Gillett, “A comparison of
Trenberth, K. E., A. Dai, G. van der Schrier, P. D. Jones, J. Barichivich, K. R. large scale changes in surface humidity over land in observations and CMIP3
Briffa, and J. Sheffield, “Global warming and changes in drought,” Nature general circulation models,” Environmental Research Letters, 5: 025210, 2010,
Climate Change, 4: 17–22, 2014, doi:10.1038/nclimate2067. doi:10.1088/1748-3926/5/025210.
Trenberth, K. E., J. T. Fasullo, and M. A. Balmaseda, “Earth’s energy imbal- Wittenberg, A. T., “Are historical records sufficient to constrain ENSO
ance,” Journal of Climate, 27: 3129–3144, 2014b, doi:10.1175/ simulations?,” Geophysical Research Letters, 36: L12702, 2009, doi:10.1029/
GCLI-D-13-00294.1. 2009GL038710.
38_Singh_ch38_p38.1-38.8.indd 7 8/22/16 2:07 PM

Chapter 39
Spatial and Temporal Estimation
and Analysis of Precipitation
BY
RAMESH S. V. TEEGAVARAPU
ABSTRACT rain-runoff (RR) models. Several conceptually simple methods with low com-
putational complexity to methods that require enormous computational
Spatial and temporal estimation and analysis of precipitation are critical and
effort are available in literature (WMO, 1994; ASCE, 1996; Dingman, 2011).
essential tasks for hydrologic modeling and water resources management.
A few commonly used methods are discussed briefly in the next sections.
This chapter presents a comprehensive overview of methods for estimation of
mean areal precipitation (MAP) and missing data. MAP estimation methods 39.2.1 Gauge Mean Method
ranging from conceptually simple methods using rain gauge observations to
The gauge mean method (GMM) is simplest method for estimating mean
computationally intensive methods that use multi-sensor observations are
precipitation depth over an area of interest (i.e., watershed or basin) using an
presented and discussed. Methods to obtain serially complete precipitation
arithmetic average of all the observations at the gauges for a given time inter-
datasets when missing data exist are needed for many hydrologic, climate
val. All the gauges that are inside an area of interest are used for calculation of
variability and change studies. Deterministic and stochastic spatial interpola-
mean precipitation. The GMM can also be applied by considering rain gauges
tion methods and their variants for estimation of missing precipitation data
(or stations) outside the region of interest provided the observations recorded
are described and reviewed. Recently developed variants of spatial interpola-
by these gauges are representative of precipitation inside the region. The
tion methods that benefit from advanced data-driven, data mining, and
method as an arithmetic average of all the available observations is objective.
optimization techniques are also discussed. The chapter presents issues and
However, applicability of this method for estimation of MAP is recommended
challenges related to the estimation of MAP and filling of missing precipita-
when the region has: (1) uniform precipitation with little spatial variability of
tion data, and points to new methods that can overcome the limitations of
rainfall; (2) precipitation patterns that are not influenced by any topographic
some of the existing methods.
or physiographic features; and (3) the gauges are uniformly distributed across
the region.
39.1 INTRODUCTION
39.2.2 Thiessen Polygon-Based Weighting Method
The estimation of mean rainfall or mean areal precipitation (MAP) over a
region or watershed and the development of gap-free serially complete pre- The Thiessen (or Voronoi) polygon method (Thiessen, 1911; Brassel and Reif,
cipitation data are two critical tasks for hydrologic modeling and water 1979; Boots, 1986) provides an area-weighted average of all precipitation
resources management. The accuracy of MAP estimation depends on the rain observations based on development of polygons with an assumption that any
gauge network or precipitation estimates from any other sources (e.g., radar point in the polygon will have similar rainfall characteristics to those recorded
and satellite) used. While this chapter deals with quantification of variability at a rain gauge located in that polygon. Thiessen polygons are constructed by
of precipitation in space and time based on observations, instruments used to connecting gauges by straight lines and drawing perpendicular bisections that
measure precipitation, measurement errors and uncertainty in measurement meet to form irregular polygons. If polygon areas extend beyond the bound-
process that are relevant to estimation are not discussed. Readers are advised ary of the region, only partial areas within the region are considered. This
to refer to works of WMO (1983), Linacre (1992), Strangeways (2007), and method is an objective approach for estimation of MAP. The polygons need
Sene (2010) for an exhaustive discussion on precipitation measurement to be recreated every time the gauging network changes. Also, the method
devices, networks, radar-based precipitation estimation, and advances in does not include or benefit from physiographic information relevant to the
hydrometeorology. Deterministic weighting and stochastic interpolation region that influences the spatial variability of the precipitation. A graphical
methods have been used for the creation of rainfall surfaces (fields) or grid- illustration of Thiessen polygons created based on five rain gauges in a water-
ded precipitation data and also estimating missing rainfall data at points in shed is shown in Fig. 39.1. The MAP (qA) is estimated used using Eqs. (39.1)
space. Some of these methods are also used in bias correction procedures for and (39.2). The variables wj, qj, and Aj represent the weight, precipitation
radar data (Teegavarapu, 2013b). Methods for MAP estimation and missing amount, and polygon area associated with rain gauge j, respectively. The
data are discussed in several sections of this chapter. A summary of issues and number of rain gauges is referred by variable ng.
considerations given for selection of MAP and missing data estimation meth-
ng
ods are also provided.
θ A = ∑w jθ j
(39.1)
j =1

39.2 ESTIMATES OF MEAN AREAL PRECIPITATION
−1
MAP, also referred to as equivalent uniform depth of rainfall over a region or 
ng

areal average rainfall, is required in many hydrologic modeling studies as a w j = A j ∑A j  ∀j (39.2)
main precipitation input to lumped hydrologic simulation models or  j =1 

39-1
39_Singh_ch39_p39.1-39.10.indd 1 8/22/16 2:08 PM

39-2 Spatial and Temporal Estimation and Analysis of Precipitation
3
5
1
Centroid
Figure 39.2 Illustration of centroid distance calculations for estimation of MAP.

Figure 39.1 Location of rain gauges and Thiessen polygons.
39.2.3 Isohyetal Method 39.2.7 Line Method

The isohyetal method or eyeball method (Linsley et al., 1949; France, 1985; Line method for calculation of mean average precipitation was introduced by
Dingman, 2011) requires the creation of isohyets (i.e., lines or contours of Goel and Aldabagh (1979). In this method, lines are drawn between any two
equal rainfall amounts) and calculation of area in-between the isohyets. points with precipitation measurements as shown in Fig. 39.3. The lines are
Isohyets are created using available rain gauge measures at different points in not allowed to cross other lines. Once the lines are drawn, their lengths are
a region by linear interpolation between any two points and adopting some measured and are used in a weighted average estimation of MAP. Gauges
form of smoothing or surface fitting method to generate smooth curves. MAP lying outside the region can be used for the line method. The MAP is esti-
is based on area-weighted average precipitation value calculated based on the mated using Eq. (39.5) and the variable Li,j in Eq. (39.5) is the length of line
isohyet values. from point i to point j (i.e., distance between gauges).
∑ i=1∑ j=1Li , j (0.5 *[θi + θ j ]) (39.5)

ng ng
39.2.4 Inverse Centroidal Distance Method
θA =
The inverse centroidal distance method belongs to a class of weighting tech-
∑ i=1∑ j=1Li , j
ng ng
niques for estimation of mean precipitation over an area. Initially, the centroid
of the basin or watershed or any region of interest is found. Distances of the
rain gauges from the centroid are calculated and the inverses of these dis- 39.2.8 Triangle Method
tances (dc−, jf ) raised to an exponent (f ) are used as weights (wc,j) in a distance
weighted method along with precipitation measures at these gauges. The The triangle method, as the name suggests, involves creation of triangles
MAP is estimated using Eqs. (39.3) and (39.4) and is illustrated in Fig. 39.2. using the rain gauges (or control points) as vertices. The triangle method
is discussed in detail by Sumner (1988) and Shuttleworth (2012). The
method involves development of triangles connecting any three points and
ng making sure that they are equilateral as possible. Once triangles are devel-
θ A = ∑wc , jθ j
(39.3) oped the area covered by each triangle is calculated and the average of
j =1

−1
 
ng
wc , j = dc−, jf ∑dc−, jf  (39.4)
 j=1 

39.2.5 Multiquadratic Surface Method

2
The multiquadratic method (Shaw and Lynn, 1972; Shaw et al., 2011) uses a
three-dimensional description of the rainfall surface. The coefficients used to 3
define the surface are estimated using available equations using different
values of precipitation amounts at different locations in space. Once the coef-
5
ficients are estimated, the equation can be used to obtain volume of rainfall
1
by integration over the area of the catchment or basin and finally dividing the
volume of rainfall by catchment area. More details of this method can be
found in work by Shaw et al. (2011).
39.2.6 Percent Normal method
The percent normal method (WMO, 1994) is beneficial in regions where the
physiography influences the spatial variability of rainfall at different temporal
resolutions. If average precipitation for a storm event over a region is 4
required, then the storm precipitation is expressed as a percentage of mean
annual or mean seasonal precipitation, and iso-percental maps are used for
preparing isohyetal maps (WMO, 1994). Figure 39.3 Illustration of line method for estimation of MAP.
39_Singh_ch39_p39.1-39.10.indd 2 8/22/16 2:08 PM

Estimates of Mean Areal Precipitation 39-3
precipitation value for each triangle is obtained using the average of the estimation of rainfall in South Florida. They conclude that multiquadric, krig-
three gauges. The MAP is calculated by obtaining the weighted average ing, and optimal interpolation are best three methods among those evaluated.
using the product of mean precipitation value and the area of each triangle Inverse-distance weighting method with optimal parameters (exponent and
and divide the product by total area from all the triangles. Sumner (1988) search radius) was used to estimation of spatial rainfall distribution by Chen
suggests that gauges that are outside the region of interest can also be used and Liu (2012). A similar exercise was reported by Noori et al. (2014) for
in this method. spatial interpolation of rainfall. Mair and Fares (2011) used Thiessen polygon,
inverse distance weighting (IDW), linear regression, OK, and simple kriging
39.2.9 Two-Axis and Hyposometric Methods with varying local means to estimate MAP in mountainous regions. Use of
The two-axis method, often referred to as Bethlahmy’s two-axis method, is artificial neural networks (ANNs) for evaluation of spatial distribution of
another way of obtaining mean annual precipitation for a region. The method is precipitation was reported by Tsangaratos et al. (2014) and for MAP by Bryan
developed by and described in detail by Dingman (2011) and WMO (1994). and Adams (2002). Gridded precipitation data developed using rain gauge
The hyposometric method is a surface fitting method and is particularly suitable data (referred to MAPO) with the help of spatial interpolation methods and
for mountainous regions and requires four curves, and they are: (1) area eleva- radar-based data (referred to as MAPX) are used by the National Weather
tion; (2) station precipitation; (3) station elevation; and (4) area precipitation. Service (NWS) of the United States (NOAA, 2015) for estimation of MAP.
The procedures for construction of curves, use of these curves, and estimation The reader is advised to refer to comprehensive reviews of methods for esti-
of MAP are available from WMO (1994). mation of MAP in mountainous and non-mountainous provided by Anderson
(2002) and Jones (1983). Lebel et al. (1987) used scaled estimation error vari-
39.2.10 Kriging Method ance computed from a scaled climatological variogram model of the rainfall
Variance-dependent stochastic interpolation techniques based on geostatistical field as a criterion to evaluation three MAP estimation methods. Distance-
approaches (Isaaks and Srivastava, 1989), such as kriging are also often used based methods are not only used for estimation of MAP, but also for under-
for estimation of MAP. Cokriging of radar and rain gauge data has been standing spatial variability of precipitation extremes (Teegavarapu et al.,
employed by Krajewski (1987) to estimate MAP. Seo et al. (1990a; 1990b) and 2013). Radar- (Raghavan, 2003; Sene, 2010; Teegavarapu 2013b, 2015a) and
Seo (1996) described the use of cokriging and indicator kriging for interpolat- satellite-based [e.g., Tropical Rainfall Monitoring Mission (TRMM),
ing rainfall data. Chua and Bras (1982) report an application of kriging to Teegavarapu, 2013b] and Global Precipitation Measurement (GPM), Sene
estimate MAP for mountainous regions considering drift and spatial depen- (2010) precipitation data available at a prespecified spatial resolution are now
dency of precipitation influenced by orographic effects. Goovaerts (2000) available for calculation of MAP. Details of radar-based precipitation estima-
report the use of simple kriging with varying local means, kriging with an tion and bias correction procedures are discussed exhaustively by Teegavarapu
external drift; and colocated kriging for incorporating a digital elevation (2013b) and Sene (2010). Transformations of precipitation data from one grid
model into the spatial prediction of rainfall. Phillips et al. (1992) evaluated the size to another using spatial interpolation methods (Teegavarapu et al., 2012a)
performance of three geostatistical methods for obtaining mean annual pre- are essential in many situations to estimate MAP. Grimes et al. (1999) esti-
cipitation estimates on a regular grid of points in mountainous terrain. The mated MAP using optimal merging of the estimates provided by satellite
methods evaluated include: (1) kriging; (2) kriging elevation-detrended data; information and estimates obtained from rain gauges. Sokol (2003) reported
and (3) cokriging with elevation as an auxiliary variable. The last two meth- the use of rain gauge and bias-corrected radar-based precipitation estimates
ods provided better estimates compared to kriging. Rogelis and Werner for estimation of MAP. Several other spatial interpolation methods such as
(2013) investigated the use of ordinary kriging (OK), universal kriging, and natural neighbors, nearest neighbors, triangular irregular networks (TINs),
kriging with external drift with individual and pooled variograms for real- trend surface models, and thin plate splines (TPS) can be used for estimation
time rainfall field estimation in areas with complex topography. In their study, of MAP. An exhaustive review of several interpolation methods along with
they found that interpolators using pooled variograms provided results com- discussions about the assumptions need, complexities encountered, and com-
parable to those obtained when the interpolators were applied to the storms putational requirements for accurate surface generation is provided by
individually. Burrough and McDonnell (1998) and Li and Heap (2008). Table 39.1 provides
main features and limitations of a few select MAP estimation methods. The
39.2.11 Other MAP Estimation Methods methods and references listed and the features and limitations documented
This section provides a brief survey of MAP estimation methods and their should not be considered complete.
applications that consider and use basin-specific characteristics, physio-
graphic features, and orographic effects, and merged data from different 39.2.12 Summary of Issues for MAP Estimation
measurement and estimation sources of precipitation (i.e., multisensor esti- Advancements in radar and satellite-based meteorological measurements
mates). Objective analysis (OA) proposed by Gandin (1965) can be used for (Sene, 2010) provide two alternative sources to rain gauge-based precipitation
interpolation of precipitation data for MAP. OA uses an optimal interpolation measurements for accurate estimation of MAP. Radar-based precipitation
analysis scheme to obtain best possible estimate of a meteorological field (e.g., estimates, even while becoming more reliable and accurate, require appropri-
precipitation in this context) at a regular network of grid points using obser- ate bias corrections before use (Teegavarapu et al., 2015a). Moulin et al. (2009)
vations at sites available in space by minimizing the mean square interpola- indicate that major source of input uncertainty in hydrologic models comes
tion error for a large ensemble of analysis situations (Bergman, 1978). A few from the lack of representativeness of a discrete set of gauges of a network and
other methods listed by Dingman (2011), such as polynomial, Lagrange from the necessity to interpolate the rain rates between these points. A num-
polynomial and spline surface, optimal interpolation via kriging, and empiri- ber of factors need to be considered when selecting suitable methods for MAP
cal orthogonal functions (EOFs) can also be used for estimation of MAP. estimates and they include: (1) rain gauge density; (2) spatial location of the
Application of bicubic spline to estimate MAP was discussed by Shaw and gauges including inter rain gauge distances; (3) availability of optimal net-
Lynn (1972). Finite element method with altitude corrections and approach work of gauges for characterizing spatial and temporal variability of precipita-
using product of distances enclosing or radiating from a rain gauge as weights tion in a region; (4) topographic and physiographic variations of the region
was developed and evaluated for MAP estimation by Hutchinson and Walley including existence of separable homogeneous rain areas; (5) utility of rain
(1972) and Lal and Al-Mashidani (1978), respectively. Sugawara (1992) dis- gauge(s) that are outside of the domain of interest; (6) long-term temporal
cusses the limitations of Thiessen polygon-based weights and reports a rain gauge-based datasets; (7) accuracy of rainfall measurements from rain
method wherein weights are calculated based on observed discharge. Variants gauges considering random and systematic errors; (8) availability of radar or
of regression models (Daly et al., 1994, 2002) that use elevation, topography satellite-based precipitation datasets at a reasonable spatial resolution (i.e.,
and proximity to coastal area, and distances as independent variables are also tessellation) considering the watershed size; (9) availability of bias corrected
available for estimation of MAP. Descriptions and application of other meth- radar and satellite-based precipitation data; (10) validated and quality assured
ods suitable for desert environments and varying topography can be found in and quality controlled (QAQC) rainfall data that is error, anomaly, and gap
works of Sen and Eljadid (2000) and Dawdy and Langbein (1960), respec- free; and (11) the spatial interpolation method used to generate rainfall fields
tively. Sen (1988) and Akin (1971) provided different weighting methods for or surfaces. While radar and satellite estimated precipitation data are ideally
estimation of MAP using polygons defined by rain gauge networks. Trend suited for MAP estimation, ground-based reliable rain gauge network of suf-
surface analyses using linear, quadratic, and cubic functions for estimation of ficient density and with spatially uniform placement of gauges in a region is
MAP were reported by Mandeville and Rodda (1973). They conclude that critical as the data from rain gauges are used for bias corrections of radar
trend surface methods provide better estimates of MAP compared to those (Teegavarapu et al., 2015a) and satellite data. Exhaustive comparison of differ-
from Thiessen polygon or isohyetal methods. Abtew et al. (1993) evaluated ent methods to select the best method for estimation of MAP is recom-
Thiessen polygon, inverse-distance, multiquadratic, polynomial, optimal and mended in several studies (e.g., Tabios and Salas, 1985; Singh and Chowdhury,
point interpolation, ordinary and universal kriging for point, and areal 1986; Teegavarapu, 2012). Improved monitoring network designs based on
39_Singh_ch39_p39.1-39.10.indd 3 8/22/16 2:08 PM

Table 39.1 Main Features and Limitations of Different MAP Estimation Methods
MAP estimation method Features of the method Limitations/issues Reference
Arithmetic average Conceptually simple No consideration for topography or other features McCuen (2004)
Bicubic spline Objective and simple method Computationally intensive Shaw and Lynn (1972)
Centroidal distance and variants Simple, objective Rain gauge network dependent Shuttleworth (2012)
EOF Conceptually sound Computationally intensive Dingman (2011)
Finite element Conceptually sound Computationally intensive Hutchinson and Walley (1972)
Gridded precipitation (radar) Objective, dependent Needs bias corrected data, good spatial resolution required Sokol (2013), Teegavarapu (2013b), NOAA
on good quality data (2015)
Gridded precipitation (rain gauge) Spatially interpolated data Underestimation of higher-end extremes Teegavarapu et al. (2013), Goly and
Teegavarapu (2014), NOAA (2015)
Hyposometric Conceptually sound WMO (1994), Dingman (2011)
Inverse-distance weighting Simple and objective No consideration for topography or other features Chen and Liu (2012), Mair and Fares (2011)
Isohyetal Objective and simple method Requires meteorological expertise in creation of isohyets Linsley et al. (1949), France (1985)
Kriging Most widely used, objective, Selection of variogram, anisotropy issues Chua and Bras (1982), Krajewski (1987),
provides uncertainty in Goovaerts (2000), Rogelis and Werner (2013)
estimates
Line Simple and objective No consideration for topography or other features Goel and Aldabagh (1979)
Multiquadratic surface Surface generation method Requires good rain gage network density Shaw and Lynn (1972), Shaw et al. (2011)
Percent normal Simple and objective Needs long-term data WMO (1994)
Polygon weighting Simple and objective No consideration for topography or other features Akin (1971), Sen (1988)
Regression Fixed functional form Rigid functional form, validation of assumptions about Daly et al. (1994, 2002)
residuals
Satellite data Spatial variation is captured Needs bias corrected data, good spatial resolution required Grimes et al. (1999) Ding man (2011)
well
Thiessen polygon Objective and widely used Rain gauge network dependent, no physiographic Thiessen (1911), Brassel and Reif (1979)
information considered
Triangle Simple and objective No consideration for topography or other features Sumner (1988) and Shuttleworth (2012)
Two axis Conceptually simple No consideration for topography or other features WMO (1994)
geostatistical approaches (Teegavarapu et al., 2015b) information-theoretic the data is referred to also as last observation carry forward (LOCF) proce-
methods (Xing et al., 2013), entropy and multiobjective-based techniques dure (Buuren, 2012). A high serial autocorrelation value at lag one is essential
(Mogheir et al., 2013) and fuzzy theory and multiple criteria analysis methods for success of this method. If several missing data exist in a series of continu-
(Chang and Lin, 2014) can help in the MAP estimation. Ly et al. (2013) pro- ous time intervals, a constant or a mean value can be used as values in these
vided a survey commonly used spatial interpolation methods for areal rainfall time intervals. This process is referred as the baseline observation carry for-
estimation and discuss advantages and limitations of different methods. A ward (BOCF) approach (Buuren, 2012). BOCF method is also referred to as
summary of conceptually simple techniques used methods for MAP estima- single imputation. Serially complete precipitation data without any gaps is
tion was provided by Rainbird (1967). needed for climate variability and change studies (Teegavarapu, 2013a,
2013b) and therefore spatial or temporal interpolation becomes essential to
fill the gaps.
39.3 MISSING PRECIPITATION DATA ESTIMATION
METHODS 39.3.1 Conceptually Simple or Naïve Estimation
Existence of serially complete, continuous precipitation data from rain gauge Methods
measurements is practically impossible due to a number of reasons. These methods are conceptually simple and use rainfall information in space
Instrument malfunctions, systematic and random errors, transcription and time from nearby rain gauges to estimate missing data at base station (i.e.,
errors, and extreme weather events sometimes disturb the measurement site or gauge at which missing data exist). Arithmetic average referred to as
process and also introduce gaps in precipitation measurements. Understanding gauge mean estimator (GME), single best estimator (SBE), and climatological
the mechanisms by which missing data occur is crucial for development of mean estimator (CME) use an average value of observations from the nearby
missing data estimation methods. A few mechanisms are defined and they are rain gauges, observations from one rain gauge, and mean of historical obser-
missing at random (MAR), missing completely at random (MCAR), and vations available for that specific temporal slice at the rain gauge which is
missing not at random (MNAR) by Little and Rubin (2002). In many missing known to be having the missing data, respectively. The optimal number of
precipitation data-estimation studies, gaps can be attributed to be as MCAR rain gauges in GME and most representative rain gauge in SBE can be
suggesting that number and temporal occurrence of gaps (missing) in pre- obtained by optimization or by use of any proximity metric that is Euclidean
cipitation data a site (i.e., rain gauge) are not dependent on the data at the site or predefined distance metric or using correlation between base station and
or any other sites. Estimation of missing precipitation data can be carried any other station. A special case of using only three nearby gauges for estima-
using temporal and spatial interpolation methods. The former relies on the tion was referred to as three-station average method by Paulhus and Kohler
persistence with strong serial autocorrelation and the latter depends on reli- (1952). These methods are often used as benchmark methods against whom
able time consistent observations at different rain gauges. Chronological more advanced spatial interpolation methods are compared and evaluated.
pairing (CP) of data is needed to estimate missing values using spatial inter-
polation. At low temporal resolutions (e.g., 15 min or 1 h), rainfall observa- 39.3.2 Weighting Methods
tions often show strong persistence allowing for infilling a missing Traditional weighting and data-driven methods generally are used for esti-
observation in a time interval using observation from the immediate previ- mating missing precipitation. Inverse distance (Wei and McGuinness, 1973;
ous time interval (referred to as temporal interpolation). This way of infilling Simanton and Osborn, 1980) is one such method that is widely used in
39_Singh_ch39_p39.1-39.10.indd 4 8/22/16 2:08 PM

New Methods for Missing Data Estimation 39-5
hydrological applications. The inverse distance weighting method (IDWM) is for spatial interpolation (O’Unwin and Sullivan, 2003). Brimicombe (2003)
the most commonly used approach for estimating missing data in the hydro- indicated that the main point of contention in applying IDWM to spatial
logical and geographical sciences. In the United States especially in the opera- interpolation is selecting the number of relevant observation points used for
tional hydrology literature, IDWM is often referred to as the NWS method. the spatial interpolation. Many spatial interpolation methods for missing data
The IDWM is routinely used for estimating missing rainfall data (ASCE, estimation lack: (1) a procedure for objective selection of optimal number of
1996). In the field of quantitative geography, IDWM and inverse exponential neighbors (or gauges) and neighborhood size (in reference to local or global
weighting methods are commonly used for spatial interpolation (O’Sullivan interpolation); (2) optimal weights in distance or correlation based or other
and Unwin, 2003). Several variants of IDWM have been developed. Hodgson weighting schemes; (3) the ability to preserve site-specific statistics of pre-
(1989) modified IDWM to include a learned search approach that reduces the cipitation data; (4) ability to provide uncertainty in the estimates; (5) mecha-
number of distance calculations. To incorporate topographical aspects, nism to define rain or no-rain conditions at the site of interest or correct the
Shepard (1968) proposed a modified IDWM that is referred to as a barrier estimates based on these conditions; (6) ability to preserve all the precipita-
method. The quadrant method (McCuen, 2004) uses the nearest gauge to the tion characteristics [e.g., dry and wet spells and extremes, transitions (e.g.,
base gauge in each quadrant for estimation of missing precipitation data in wet to dry), autocorrelation, and several others]; and (7) the ability to pre-
weighting scheme similar to IDWM. Nearest gauge in each quadrant is serve spatial variance and regional statistics.
defined by Euclidean distance. Normal ratio method (Paulhus and Kohler,
1952; Tung, 1983; McCuen, 2004; Teegavarapu and Chandramouli, 2005) is
another conceptually simple method for estimation of missing data that relies 39.5 NEW METHODS FOR MISSING DATA ESTIMATION
on the assumption related to average annual precipitation values at sites. Distance weighting methods often used for estimating missing rainfall
records are revisited by Teegavarapu and Chandramouli (2005). Conceptual
39.3.3 Surface Interpolation Methods revisions are incorporated into their methods and the revised methods are
Variance-dependent surface interpolation methods, belonging to the general tested for estimation of missing rainfall values. The revisions addressed two
family of kriging, have been applied to hydrological interpolation problems main issues relating to the definition of distance used in the calculations and
(Grayson and Bloschl, 2001; Vieux, 2001). These interpolation schemes are selection process of the nearby gauges.
based on the principle of minimizing the estimate of variance at points where
measurements are unavailable. Kriging in various forms has been used to 39.5.1 Variants of Weighting, Data-Driven, and
estimate missing precipitation data at stations (Teegavarapu and Copula-Based Methods
Chandramouli, 2005) as well as to interpolate precipitation from point mea- Teegavarapu and Chandramouli (2005) developed and evaluated several vari-
surements (Vieux, 2001; Dingman, 2011). Ashraf et al. (1997) compared ants of inverse distance method and data-driven methods for estimation of
interpolation methods (kriging, inverse distance, and cokriging) to estimate missing precipitation records at a site. These methods include coefficient of
missing precipitation values. They indicated that kriging provided the small- correlation weighing method (CCWM), artificial neural network estimation
est root mean square error (RMSE). Application of kriging for areal mapping method, modified Thiessen polygon method, nearest neighbor weighting
of precipitation was reported by Karnieli (1990). Regression and time series method, and its revised version and kriging estimation method (KEM). All
analysis methods (Salas, 1993) belong to data-driven approaches. Global the methods are data intensive and sensitive, with KEM being computation-
interpolation methods that use trend surface analysis and regression (Chang, ally most intensive of all the methods. Use of conceptually acceptable surro-
2006) provide several advantages compared to deterministic weighting tech- gate measures for distances and improvised weighting factors improved the
niques. Trend surface methods use polynomial equations of spatial coordi- estimates from these variants. CCWM has proven to be a better weighting
nates to approximate points with known values. Regression models using method compared to other interpolation methods in several studies (e.g., Kim
elevation information (Lloyd, 2009), locally weighted polynomials (Regonda et al., 2008; Westerberg et al., 2009). The kriging method is considered a reli-
et al., 2006), thin splines (Xia et al., 1999), and local and global trend surface able interpolation technique (O’Sullivan and Unwin, 2003), but is plagued by
models (Teegavarapu, 2012) are available for estimation of missing precipita- several limitations. These include selecting the appropriate semivariogram,
tion data. Guillermo et al. (1985) evaluated polynomial interpolation by least- assignment of arbitrary values to sill and nugget parameters and distance
squares and Lagrange approach, inverse distance, multiquadric interpolation, intervals, observation value-insensitive variance estimates, and the computa-
optimal interpolation, and kriging techniques for estimation of missing data. tional burden to interpolate the surfaces. Universal functional approximators
They indicated kriging and optimal interpolation techniques are superior to such as ANNs are used for fitting a semivariogram model using the raw data
the other techniques. in OK to estimate missing precipitation data by Teegavarapu (2007). The use
of ANN eliminates the need for the predefined authorized semivariogram
models to capture the spatial variation of data, and the trial-and-error process
39.4 LIMITATIONS OF ESTIMATION METHODS
involved in estimation of semivariogram parameters. Use of modular neural
Teegavarapu and Chandramouli (2005) reported several limitations and networks for estimation of missing precipitation is reported by Kajornrit et al.
advantages of using deterministic and stochastic spatial interpolation tech- (2012). Kuligowski and Barros (1998) report improved estimates of missing
niques to estimate missing precipitation data at a base station using data at all data using ANN compared to those from distance-weighted and linear regres-
other stations. They indicated that all interpolation techniques will fail to sion models at a number of rain gauge sites. Similar conclusion about
provide accurate estimates of missing precipitation data in two situations: improvements in estimates was reported by Teegavarapu and Chandramouli
(1) when precipitation is measured at all or a few other stations, but no pre- (2005). Association rule mining (ARM)-based spatial interpolation
cipitation actually occurred at the base station, and (2) when precipitation is (Teegavarapu, 2009) is used to improve the precipitation estimates provided
measured at the base station but no precipitation is measured or occurred at by traditional and improved deterministic and stochastic spatial interpolation
all other stations. In case 1, all spatial interpolation techniques will produce a techniques. The ARM methodology besides offering insights into the spatio-
positive estimate whereas in reality zero precipitation is recorded at the base temporal precipitation data patterns and the associations among observa-
station. It is impossible to estimate missing precipitation data in the second tions, also helped in addressing one major ubiquitous limitation of all spatial
case since point observations are used to estimate the missing value at the interpolation techniques in accurately estimating missing precipitation
base station using spatial interpolation algorithms alone. All interpolation records. Considerable improvements in the estimates were achieved when
techniques produce a zero estimate for situations encountered in case (2). ARM was used in conjunction with the interpolation techniques. Copula-
Data from other sources (e.g., radar or satellite-based estimates) can be used based methods for infilling are found to be superior compared to a number
in these situations to estimate the missing values. However, the reliability of of interpolation methods [nearest neighbor, linear regression, fuzzy rules,
radar-based precipitation measurements is always a contentious issue OK, multiple linear regression (MLR), and MLR using the expectation maxi-
(Teegavarapu, 2008) considering several radar-based estimation errors (Sene, mization (EM) algorithm] in a study by Bardossy and Pegram (2014). Landot
2010). Limitations of spatial interpolation methods have been reported in et al. (2008) used a copula-based method for infilling long-term gaps in pre-
recent studies. Vieux (2001) pointed out several limitations of the IDWM, cipitation data in Florida with the help of support vector machine (SVM).
with a major one being the “tent pole effect” that leads to greater estimates However, the copula-based method reported by Landot et al. (2008) was
closer to the point of interest. Grayson and Bloschl (2000) list several limita- found to be inferior to CCWM and linear optimization method in a study by
tions of Thiessen polygons and IDWMs. They have suggested that these Aly et al. (2009). Presti et al. (2010) investigated simple substitution, parametric
methods should not be recommended for spatial interpolation considering regression, ranked regression, and rank-invariant method of linear and poly-
their limitations. However, they recommend thin splines and kriging for nomial regression analysis (Theil method). They concluded that Theil meth-
interpolating hydrologic variables. The Thiessen polygon approach has the od and simple substitution method provided better results among all those
major limitation of not providing a continuous field of estimates when used compared.
39_Singh_ch39_p39.1-39.10.indd 5 8/22/16 2:08 PM

39.5.2 Optimal Spatial Interpolation Methods User-specified weights can be attached to performance metrics or measures
Range and cluster-based optimization methods in space and time were devel- as numerical values (Teegavarapu, 2012) or fuzzy membership function-
oped by Teegavarapu (2012). Optimization models using mixed integer linear based values (Teegavarapu and Elshorbagy, 2005) to quantify the importance
and nonlinear programming formulations are developed for estimation of of one measure over the other. In addition to these performance measures, a
missing precipitation data at a gauge. The formulations use binary variables bias plot or analysis of bias is also recommended as a means of understanding
for selection of rain gauges that participate in the spatial interpolation process the structure of errors, model-specific bias, and distribution of errors over a
and also in the process of selection of the optimum cluster of rain gauges for specific range of values or time. The methods can be qualified using the cri-
estimation of missing data. Several variants of these mathematical program- teria of accuracy and precision, which are related to the bias and variance of
ming formulations are proposed by Teegavarapu (2012) to improve the esti- the residuals (Kanevski and Maignan, 2004). Small variations in rainfall
mates of missing precipitation data. The variants include: (1) IDWM with intensity can introduce significant changes in the runoff values generated
optimal number of gauges; (2) optimal quadrant method; (3) IDWM with from distributed rainfall-runoff (RR) models (Vieux, 2001). Any improve-
optimal parameters and number of gauges; and (4) linear weighting method ment in the rainfall magnitude estimation, however minute it may be, can be
with optimal clusters of rain gauges. These variants investigated in their study considered significant, as rainfall is a crucial input that governs the response
provided improved estimates of precipitation data compared to those of hydrologic systems and the results of continuous simulation models.
obtained from different traditional distance-based weighting and kriging Similar arguments were made by Xu and Vandewiele (1994) to suggest that
methods. A nonlinear mathematical programming model using binary vari- errors in precipitation values may lead to significant effects on the model
ables is proposed and investigated by Teegavarapu and Pathak (2008) and performance and also on parameters. Two error measures (RMSE and abso-
Teegavarapu et al. (2010) to infill missing precipitation records using radar- lute error) are generally used to assess and compare the performance of
based rainfall estimates. The model identifies the cluster of radar values that interpolation techniques. However, these global error assessment measures
can be used for infilling the rain gauge records. The application of the model may not provide a complete assessment of methods as they are average mea-
also revealed several interesting insights, which suggest that optimal selection sures calculated for a specific period of time. The temporal distribution and
of radar-based precipitation values using weights in the grid surrounding a magnitude of the precipitation affects the outputs of RR models, which use
rain gauge is essential for the infilling process. Spatial and temporal variabil- precipitation as an input. One of the best ways to evaluate spatial interpola-
ity of weights for different clusters are evident. Linear programming method tion methods is to use the estimated precipitation values in a hydrologic
using topographical information was discussed by Yoo (2013). simulation model and assess the impact of the major input (i.e., rainfall) on
the estimation of one or several hydrological variables. A recent study by
39.5.3 Nearest-Neighbor and Proximity Metric-Based Teegavarapu (2014a) provides a number of performance measures, error
Methods measures, and statistical hypothesis test to evaluate different spatial interpola-
Similarity of two observation datasets based on real or categorical attribute tion methods. The measures included summary statistics, quantile–quantile
values can be evaluated using proximity metrics based on numerical taxono- (Q–Q) plots, statistical hypothesis tests (e.g., two sample Kolmogorov–
my concepts (Ling, 2010). These metrics provide a measure of proximity Smirnov and Chi-square tests) to check the similarity of probability distribu-
between any two observed datasets and, in turn, can be used to assign weights tions, differences in two state first-order Markov chain transition probabilities
in the interpolation schemes to estimate missing data. Ten different binary for dry–wet, dry–dry, wet–dry, and wet–wet conditions, autocorrelations at
and real-valued distance metrics were used as proximity measures by several temporal lags and deviations in extreme precipitation indices (WMO,
Teegavarapu (2014b) to obtain optimal weights in spatial interpolation meth- 2009; Teegavarapu et al., 2012b) derived based on observed and filled pre-
ods for the estimation of missing data. In a recent study, Pappas et al. (2014) cipitation datasets. Contingency measures (i.e., concordance, error rate, sen-
provided a conceptually simple method using nearest neighbors (observa- sitivity, and specificity) and several forecast-verification measures and skill
tions in time at one site) for estimation of missing data. scores are used by Teegavarapu (2014a) for comparison of different spatial
interpolation methods. Finally, the homogeneity of the filled datasets should
39.5.4 Corrections of Spatially Interpolated Estimates be tested using one of the tests: Pettitt’s (Pettitt, 1979), Buishand’s (Buishand,
Corrections of spatial interpolation-based precipitation estimates considering 1982), or Alexandersson’s standard normal homogeneity test (SNHT)
rain or no-rain conditions at a nearest rain gauge selected by a proximity (Alexandersson, 1986), or Von Neumann ratio test (Von Neumann, 1941).
metric or correlation can be used. Corrections are suggested by Teegavarapu Table 39.2 provides a list of missing data estimation methods their main fea-
(2009) using data mining-based association rules. Improved estimates can be tures and limitations. The list and discussion of features and limitations pro-
obtained compared to use of methods with no corrections. However, vided is not considered to be exhaustive.
overestimation of precipitation magnitudes in many instances is common.
The SVM-based approach was used to select rain gauges for correction by Aly
39.6 SUMMARY OF ISSUES FOR MISSING PRECIPITATION
et al. (2009). The single best estimator (SBE) method and its optimal variant DATA ESTIMATION
are developed by Teegavarapu (2011) and Teegavarapu (2014b), respectively,
for postcorrections of missing data estimates at a gauge. An exhaustive survey of the literature indicates several methods ranging from
conceptually simple to computationally intensive are available that can be
39.5.5 Site and Regional Statistics Preserving used for estimation of missing precipitation data. Selection and use of one or
Methods more methods for estimation is an exhaustive task and will depend on sev-
Preserving site-specific statistics along with accurate estimation of missing eral constraints and is also dictated by numerous objectives related to any
data is essential when data gaps at a gauge are filled. Postcorrection methods specific study. The focus, in general will be on the selection of those methods
using quantile–quantile (Q–Q) and equiratio quantile matching and single that are conceptually sound, field tested, and robust. The geographical region
best estimation methods were suggested by Teegavarapu (2014a). The methods of study, ease of method development by modelers, transferability, and utility
preserved site statistics and also provide improved estimates of several WMO- of a part or complete computational code from a similar model available
based extreme precipitation indices that are closer to observed series than elsewhere, practical applicability, data availability, and time constraints are a
those obtained without these corrections. Teegavarapu (2013b) and El-Sharif few important factors that are considered in the selection process. Also, the
and Teegavarapu (2012) proposed and evaluated several multiobjective opti- objectives and goals set forth by study and agency needs will primarily direct
mization formulations for preserving site-specific statistics and correlations the identification and finally the selection process. Computationally efficient
among different sites and spatial variability of precipitation data evaluated algorithms are required for implementation of improved estimation methods
using a variogram. based on auxiliary data other than precipitation datasets. Surrogate measures
of correlation among observations are required for applying weighting tech-
39.5.6 Evaluation Measures and Selection of niques. Use of correlation coefficients (Teegavarapu and Chandramouli,
Interpolation Methods
2005), statistical distance (Ahrens, 2006), and proximity metrics (Teegavarapu,
The spatial interpolation methods can be evaluated using error measures and 2013) are few recommendations for development of conceptually simple
performance measures such as: RMSE, mean absolute error MAE, and weighting methods. Universal functional approximation methods, such as
goodness-of-fit measure criterion, coefficient of correlation (ρ), or determi- (ANNs), are viable techniques for estimation of missing precipitation.
nation (R2), based on observed and estimated rainfall values at the base sta- However, overfitting, lack of generalization, unrealistic numeric results, and
tion. Factors such as intuitive reasonableness of model or approach, replicability and repeatability are some of the main concerns of these tech-
conceptual accuracy, and simplicity of the model are considered to make an niques. Reliable data-driven, artificial intelligence, or statistical techniques are
objective assessment, and finally selection of the models or approaches used required for prediction of ‘‘rain’’ or ‘‘no-rain’’ conditions. These techniques
for estimation of missing data (Teegavarapu and Chandramouli, 2005). precede the application of spatial interpolation techniques for estimating
39_Singh_ch39_p39.1-39.10.indd 6 8/22/16 2:08 PM

Conclusion 39-7
Table 39.2 Main Features and Limitations of Different Missing Precipitation Data Estimation Methods
Missing data-estimation method Features of the method Limitations/issues Reference
ANN based Data driven, function approximation Generalization and issues with unrealistic esti- Kuligowski and Barros (1998),
mates Teegavarapu and Chandramouli
(2005), Kajornrit et al. (2012)
ARM based Data- and rule-driven method Long-term data is required Teegavarapu (2009)
Climatological mean estimator Conceptually simple Last resort method, not accurate Aly et al. (2011)
Correlation coefficient weighting Improvised inverse distance method Data driven, needs long-term data Teegavarapu and Chandramouli
(2005), Kim et al. (2008),
Westerberg et al. (2009)
CME Conceptually simple, benchmark method Underestimation of variance Aly et al. (2011), Teegavarapu
(2012)
Inverse-distance weighting Conceptually simple, widely used Network density dependent, no meteorological Tung (1983), ASCE (1996)
and topographical influences considered
K-means cluster-based optimization Provides ability to select clusters in space Subjective parameters (e.g., number of clusters) Teegavarapu (2014b)
K-nearest neighbor-based optimization Provides ability to select clusters in time Subjective parameters (e.g., number of clusters) Teegavarapu (2014b)
Kriging Widely used, provides estimation uncer- Selection of variogram models, anisotropy con- Ashraf et al. (1997), Vieux
tainty siderations (2001), Dingman (2011)
Local and global polynomial Surface fitting Variability of rainfall is not captured if rain Regonda et al. (2006),
gauge density is not sufficient Teegavarapu (2012)
Normal ratio Simple, objective Long-term data is required Paulhus and Kohler (1952),
Teegavarapu and
McCuen (2004), Chandramouli
(2005)
Optimal spatial interpolation and vari- Objective, efficient method Needs optimization solvers, nonglobal optimal Yoo (2013), Teegavarapu (2014b)
ants solutions
Optimal spatial interpolation—neighbor Objective, efficient method, provides opti- Computational issues with optimization solvers Teegavarapu (2012), Teegavarapu
selection mal selection of neighbors (2014b)
Proximity-metric–based optimization Objective method with weights estimated Relies heavily on data length Teegavarapu (2009)
based on proximity metrics
Quadrant Simple, objective Proximity of gauges with similar data character- McCuen (2004)
istics important, availability of long-term data
SBE Conceptually simple, benchmark method Needs a highly correlated station Teegavarapu (2009)
Spatial interpolation—with post correc- Objective method, improvement over Bias may be introduced due to corrections Teegavarapu (2014a)
tions other methods in preserving site specific
statistics
Thin plate splines (TPS) Assumption of smooth surface Variability of rainfall not captured based on the Xia et al. (1999), Chang (2006)
assumption
Time series models Autoregressive models, objective Strong serial autocorrelation at few lags is Dingman (2011)
needed
Universal function approximation–based Modified version of kriging Function approximation of semi variogram, Teegavarapu (2009)
kriging generalization issues
missing precipitation records. Rain-radar relationships need to be character- A comprehensive survey of interpolation methods for environmental data
ized via functional forms to effectively utilize them in estimating missing applications is provided by Li and Heap (2008), and Akkala et al. (2010) and
precipitation. Optimum clustering of gauges for efficient and robust estima- kriging-based methods is provided by USEPA (2004). Practical application of
tion of missing precipitation records can be identified. Mathematical pro- all spatial interpolation methods require time consistent precipitation mea-
gramming formulations with binary variables can be used for this purpose for surements at all the rain gauges that are used for development of parameter
determination of optimal number of sites (i.e., neighborhood size) and clus- values (e.g., weights, exponents, etc.) in these methods. If one or more gauges
ters. Spatial and temporal variability of rainfall processes, and local and are missing data, they need to be filled first before interpolation can be used
regional climatology are essential factors that need to be considered when for estimation. The filling process can be carried out using a conceptually
estimating missing data at a site. In a few data-sparse situations, naïve meth- simple estimation method, temporal interpolation, or data from sources other
ods described in this chapter are the only way out. In general, filling missing than rain gauge if available. If one or more rain gauges have substantial
precipitation data more than 5% of the entire dataset leads to underestimation amount of missing data, then parameters of the interpolation method need to
of higher-end extremes and an increase in bias in the estimation of precipita- be re-estimated using the available rain gauges with serially complete precipi-
tion extremes (Teegavarapu, 2014a). Multiple imputation (infilling) methods tation records. If the amount of data missing is negligible and missingness can
(Buuren, 2012) based on bootstrap sampling and optimal interpolation be explained as MCAR, listwise and pairwise deletion of data methods com-
(Teegavarapu, 2013b) are useful in quantifying the uncertainties associated mon in social sciences arena (Enders, 2010) can be used. However, if missing
with estimates provided by spatial interpolation methods. Software focused data are not estimated or gaps in data are not filled, meaningful conclusions
on estimation of missing data is rarely generic. Variance deflation, over- and about the data cannot be made for any hydrologic or water resources model-
underestimation of lower- and higher-end extremes, respectively, and altera- ing studies.
tion of statistical distributions are inevitable consequences of spatial interpo-
lation methods (Goly and Teegavarapu, 2014; Teegavarapu, 2014a) often used 39.7 CONCLUSION
for generation of gridded datasets. Point and spatially complete gridded pre-
cipitation datasets are equally valuable for analysis of precipitation extremes MAP and missing data estimation are the two topics of focus in this chapter.
and characteristics. Justifiable inferences about variability in precipitation This chapter provides a brief summary of available methods. Different con-
characteristics and extremes can be drawn when both the datasets are used ceptually simple methods available for MAP along with computationally
independently (Goly and Teegavarapu, 2014). demanding approaches are discussed. Increasing availability of radar- and
39_Singh_ch39_p39.1-39.10.indd 7 8/22/16 2:08 PM

satellite-based precipitation datasets is leading to improved estimates of MAP Chua, S-H. and R. L. Bras, “Optimal estimators of mean areal precipitation
when compared with those obtained from rain gauges. Combined use of rain in regions of orographic influence,” Journal of Hydrology, 57 (1–2): 23–48, 1982.
gauge and radar or satellite precipitation datasets can also lead to improve- Daly, C., W. P. Gibson, G. H. Taylor, G. H. Johnson, and P. Pasteris, “A
ments in estimates. Availability of good quality and bias-corrected radar or knowledge-based approach to the statistical mapping of climate,” Climate
satellite-based precipitation data is essential especially in those regions with Research, 22: 99–113, 2002.
very low rain gauge density. Missing precipitation data at a site or multiple Daly, C., R. P. Neilson, and D. L. Phillips, “A statistical topographic model
sites is a ubiquitous problem that can be solved by using data and region- for mapping climatological precipitation over mountainous terrain,” Journal
specific traditional spatial interpolation, geostatistical, or surface generation of Applied Meteorology, 33: 140–158, 1994.
methods. Recent developments in estimation of missing data methods have Dawdy, D. R. and W. B. Langbein, “Mapping mean areal precipitation,”
also been documented in this chapter. Methods that preserve site and region- International Association of Scientific Hydrology, Bulletin, 5 (3): 16–23, 1960.
al summary statistics, probability distributions, reduce under- and overesti- Dingman, S. L., Physical Hydrology, Prentice Hall, New York, NY, 2011.
mation of higher- and lower-end extremes, maintain various time El-Sharif, H. and R. S. V. Teegavarapu, “Evaluation of spatial interpolation
series-specific properties, and correctly estimate the dry and wet states are methods for missing precipitation data: preservation of spatial statistics,”
required. Variants of spatial interpolation methods to incorporate correc- Proceedings of World Environmental and Water Resources Congress 2012,
tions to spatially interpolated precipitation estimates, optimal neighborhood American Society of Civil Engineers (ASCE), 3822–3832, 2012,
and number of sites selections, and local as opposed global interpolations are doi:10.1061/9780784412312.384.
proven to have potential for improving missing precipitation data estimates. Enders, C. K., Applied Missing Data Analysis, The Guilford Press, New
York, 2010.
France, P. W., “A comparison of various techniques for computation of areal
REFERENCES
rainfall,” Journal of Institution of Engineers, 66: 74–78, 1985.
Abtew, W., J. Obeysekera, and G. Shih, “Spatial analysis for monthly rainfall Gandin, L. S., Objective analysis of meteorological fields. Translated from
analysis in South Florida,” Journal of American Water Resources Association, the Russian. Jerusalem (Israel Program for Scientific Translations), Report,
29 (2): 179–188, 1993. 1965, p. 242.
Ahrens, B., “Distance in spatial interpolation of daily rain gauge data,” Goel, S. M. and A. S. Aldabagh, “A distance weighted method for comput-
Hydrology and Earth System Sciences, 10: 197–208, 2006. ing average precipitation,” Journal of the Institute of Water Engineers and
Akkala, A., V. Devabhaktuni, and A. Kumar, “Interpolation techniques and Science, 33: 451–454, 1979.
associated software for environmental data,” Environmental Progress and Goly, A. and R. S. V. Teegavarapu, “Individual and coupled influences of
Sustainable Energy, 29 (2): 134–141, 2010. AMO and ENSO on regional precipitation characteristics and extremes,”
Akin, J. E., “Calculation of areal depth of precipitation,” Journal of Water Resources Research, 50 (6): 4686–4079, 2014.
Hydrology, 12: 363–376, 1971. Goovaerts, P., “Geostatistical approaches for incorporating elevation into the
Alexandersson, H., “A homogeneity test applied to precipitation data,” spatial interpolation of rainfall,” Journal of Hydrology, 228 (1–2): 113–129, 2000.
Journal of Climatology, 6: 661–675, 1986. Grayson, R. and G. Bloschl, Spatial Patterns in Catchment Hydrology:
Aly, A., C. Pathak, R. Teegavarapu, J. Ahlquist, and H. Fuelberg, “Evaluation Observations and Modeling, Cambridge University Press, Cambridge, UK,
of improvised spatial interpolation methods for infilling missing precipitation 2001.
records,” World Environmental and Water Resources Congress 2009, pp. 1–10, Grimes, D. I. F., E. Pardo-Igúzquiza, and R. Bonifacio, “Optimal areal
2009, doi:10.1061/41036(342)598. rainfall estimation using rain gauges and satellite data,” Journal of Hydrology,
Anderson, E. A, Calibration of conceptual hydrologic models for use in 222 (1999): 93–108, 1999.
river forecasting, Technical Report, NOAA Technical Report, NWS 45, Guillermo, Q., G. Q. Tabios III, and J. D. Salas, “A comparative analysis of
Hydrology Laboratory, Silver Spring, MA, 2002. techniques for spatial interpolation of precipitation,” Journal of American
ASCE, Hydrology Handbook, 2nd ed., American Society of Civil Engineers Water Resources Association, 21 (3): 365–380, 1985.
ASCE, New York, 1996. Hodgson, M. E., “Searching methods for rapid grid interpolation,”
Ashraf, M., J. C. Loftis, and K. G. Hubbard, “Application of geostatisticals Professional Geographer, 411: 51–61, 1989.
to evaluate partial weather station network,” Agricultural Forest Meteorology, Hutchinson, P. and W. J. Walley, “Calculation of areal rainfall using finite
84: 255–271, 1997. element techniques with altitude corrections,” Bulletin on International
Bardossy, A. and G. Pegram, “Infilling missing precipitation records—a Association of Hydrological Sciences, 17 (3): 259–272, 1972.
comparison of a new copula-based method with other techniques,” Journal of Isaaks, H. E. and R. M. Srivastava, An Introduction to Applied Geostatistics,
Hydrology, 519: 1162–1170, 2014. Oxford University Press, New York, 1989.
Bergman, K. H. The role of observational errors in optimum interpolation Jones, S. B., The estimation of catchment average point rainfall profiles,
analysis, Office Note 1975, U.S. Department of Commerce, National Oceanic and Report No. 87, Institute of Hydrology, Wallingford, UK, 1983.
Atmospheric Administration (NOAA), National Weather Service (NWS), National Kajornrit, J., W. K. Wong, and C. C. Fung, Estimation of missing precipita-
Meteorological Center, Report # 013BAF1, 1978. Available at: http://www.lib.ncep. tion records using modular artificial neural networks, edited by T. Huang
noaa.gov/ncepofficenotes/files/013BA1F1.pdf (accessed January 2015). et al., ICONIP 2012, Part IV, LNCS 7666, 2012, pp. 52–59.
Boots, B. N., “Voronoi Thiessen polygons,” Concepts and Techniques in Kanevski, M. and Maignan, M., Analysis and Modelling of Spatial
Modern Geography, Geo Book, Norwich, 1986, p. 45. Environmental Data, EPFL Press, Lausanne, Switzerland, 2004.
Brassel, K. E. and D. Reif, A procedure to generate Thiessen polygons. Karnieli, A., Application of kriging technique to areal precipitation map-
“Geographical Analysis,” 2 (3): 298–303, 1979. ping in Arizona, GeoJournal, 22 (4): 391–398, 1990.
Brimicombe, A., GIS, Environmental Modeling and Engineering, Taylor & Kim, U., J. J. Kaluarachchi, and V. U. Smakhtin, “Generation of monthly
Francis, London, 2003. precipitation under climate change for the Upper Blue Nile River basin,
Bryan, B. A. and J. A. Adams, Three-dimensional neurointerpolation of Ethiopia,” Journal of the American Water Resources Association, 44 (5): 1752–
annual mean precipitation and temperature surfaces for China, “Geographical 1688, 2008.
Analysis,” 34 (2): 93–111, 2002. Krajewski, W. F., “Co-kriging of radar and rain gauge data,” Journal of
Buishand, T. A., Some methods for testing the homogeneity of rainfall Geophysics Research, 92D8: 9571–9580, 1987.
records, “Journal of Hydrology,” 58: 11–27, 1982. Kuligowski, R. J. and A. P. Barros, “Using artificial neural networks to esti-
Burrough, P. A. and R. A. McDonnell, Principles of Geographical Information mate missing rainfall data,” Journal of American Water Resources Association,
Systems, Oxford University Press, New York, NY, 1998. 34 (6): 1437–1447, 1998.
Buuren, V. S., Flexible Imputation of Missing Data, CRC Press, Boca Raton, Lal, P. B. B., and G. Al-Mashidani, “A technique for determination of areal
FL, 2012. average rainfall,” Hydrological Sciences Bulletin, 23 (4): 445–453, 1978.
Chang, K-T., Introduction to Geographic Information Systems, McGraw Hill, Landot, T., S. Sgellari, C. Lima, and U. Lall, In-filling missing historical
New York, 2006. daily rainfall data study, Final Report submitted to South Florida Water
Chang, C. L. and Y. T. Lin, “A water quality monitoring network design Management District, Columbia University, New York, 2008.
using fuzzy theory and multiple criteria analysis,” Environmental Monitoring Lebel, T., G. Bastin, C. Obled, and J. D. Creutin, “On the accuracy of areal
Assessment, 186 (10): 6459–6469, 2014. rainfall estimation: a case study,” Water Resources Research, 23 (11): 2123–
Chen, F-W. and C-W. Liu, “Estimation of the spatial rainfall distribution 2134, 1987.
using inverse distance weighting (IDW) in the middle of Taiwan,” Paddy and Li, J. and A. D. Heap, A Review of Spatial Interpolation Methods for
Water Environment, 10 (3): 209–222, 2012. Environmental Scientists, Geoscience Australia, Canberra, 2008.
39_Singh_ch39_p39.1-39.10.indd 8 8/22/16 2:08 PM

References 39-9
Linacre, E., “Climate data and resources,” A Reference and Guide, Routledge, Shaw, E. M. and P. P. Lynn, “Areal rainfall evaluation using two surface fit-
London, 1992. ting techniques,” Bulletin of International Association of Hydrological Sciences,
Ling, H. T. M., “Distance coefficients between two lists or sets,” The Python 27: 419–433, 1972.
Papers, 2: 1–31, 2010. Shepard, D., “A two-dimensional interpolation function for irregularly
Linsley, R. K., Jr., M. A. Kohler, and J. L. Paulhus, Applied Hydrology, spaced data,” Proceedings of the Twenty-Third National Conference of the
McGraw-Hill, New York, 1949. Association for Computing Machinery, New York, NY, 517–524, 1968.
Little, R. J. A., and D. B. Rubin, Statistical Analysis with Missing Data, Wiley, Shuttleworth, J. W., Terrestrial Hydrometeorology, John Wiley & Sons,
New York, 2002. Hoboken, NJ, 2012.
Lloyd C. D., “Multivariate interpolation of monthly precipitation amount Simanton, J. R. and H. B. Osborn, “Reciprocal-distance estimate of point
in the United Kingdom,” Proceedings of the Seventh European Conference on rainfall,” Journal of Hydraulic Engineering Division, 106 (HY7): 1242–1246, 1980.
Geostatistics for Environmental Applications, edited by P. M. Atkinson and C. Singh, V. P. and K. Chowdhury, “Comparing some methods of estimating
D. Lloyd, Springer, Berlin, 2009. mean aerial rainfall,” Water Resources Bulletin, 222: 275–282, 1986.
Ly, S., C. Charles, and A. Degré, “Different methods for spatial interpola- Sokol, Z., “The use of radar and gauge measurements to estimate areal
tion of rainfall data for operational hydrology and hydrological modeling at precipitation for several Czech river basins,” Studia Geophysica et Geodaetica,
watershed scale: a review,” Biotechnologie, Agronomie, Société et Environnement, 47 (3): 587–604, 2003.
17 (2): 392–406, 2013. Strangeways, I., Precipitation: Theory, Measurement and Distribution,
Mair, A. and A. Fares, “Comparison of rainfall interpolation methods in a Cambridge University Press, London, 2007.
mountainous region of a tropical island,” Journal of Hydrologic Engineering, Sugawara, M., On the weights of precipitation stations, Advances in
ASCE, 16 (4): 371–383, 2011. Theoretical Hydrology: A Tribute to James Dooge, edited by J. P. O’Kane,
Mandeville, A. N. and J. C. Rodda, “A contribution to the objective assess- European Geophysical Sciences Society Series on Hydrological Sciences 1,
ment of areal rainfall amounts: results of research on representative and Elsevier, Amsterdam, The Netherlands, Vol. 59, Chap. 5, 1992.
experimental basins, Proceedings of the Wellington Symposium, IAHS, Paris, Sumner, G. N., Precipitation: Process and Analysis, John Wiley & Sons, New
Vol. 2, 1973, pp. 120–127. York, NY, 1988.
McCuen, R. H., Hydrologic Analysis and Design, Prentice-Hall, Upper Tabios, G. O., III and J. D. Salas, “A comparative analysis of techniques for
Saddle River, NJ, 2004. spatial interpolation of precipitation,” Water Resources Bulletin, 21: 365–380,
Moulin, L., E. Gaume, and C. Obled, “Uncertainties on mean areal precipi- 1985.
tation: assessment and impact on streamflow simulations,” Hydrology and Teegavarapu, R. S. V. and V. Chandramouli, “Improved weighting methods,
Earth System Sciences, 13: 99–114, 2009. deterministic and stochastic data-driven models for estimation of missing
NOAA, 2015, http://www.srh.noaa.gov/abrfc/?n=map (accessed January precipitation records,” Journal of Hydrology, 312: 191–206, 2005.
2015). Teegavarapu, R. S. V. and A. Elshorbagy, “Fuzzy set-based error measure for
Noori, J. M., H. H. Hussein, and Y. T. Mustafa, “Spatial estimation of rain- hydrologic model evaluation,” Journal of Hydroinformatics, 7: 199–208, 2005.
fall distribution and its classification in Duhok governorate using GIS,” Teegavarapu, R. S. V., “Use of universal function approximation in vari-
Journal of Water Resource and Protection, 6: 75–82, 2014. ance-dependent interpolation technique: an application in Hydrology,”
O’Sullivan, D. and D. J. Unwin, Geographical Information Analysis, John Journal of Hydrology, 332: 16–29, 2007.
Wiley & Sons, New York, NY, 2003. Teegavarapu, R. S. V., “Innovative spatial interpolation methods for estima-
Pappas, C., S. M. Papalexiou, and D. Koutsoyiannis, “A quick gap filling of tion of missing precipitation records: concepts and applications,” edited by
missing hydrometeorological data,” Journal of Geophysical Research: J. Bruthans, K. Kovar, and Z. Hrkal, Proceedings of HydroPredict-2008,
Atmospheres, 119 (15): 1920–1930, 2014. Association of Hydrogeologists, Prague, Czech Republic, 2008, pp. 79–82.
Paulhus, J. L. H. and M. A. Kohler, “Interpolation of missing precipitation Teegavarapu, R. S. V. and C. Pathak, “Infilling of rain gauge records using
records,” Monthly Weather Review, 80 (8): 129–133, 1952. radar (NEXRAD) data: influence of spatial and temporal variability of rainfall
Pettitt, A. N., “A non-parametric approach to the change-point problem,” processes, ASCE,” World Environmental and Water Resources Congress, 2008,
Applied Statistics, 28 (2): 126–135, 1979. pp. 1–9, doi:10.1061/40976(316) 406.
Phillips, D. L., J. Dolph, and D. Marks, “A comparison of geostatistical Teegavarapu, R. S. V., D. Peters, and C. Pathak, “Evaluation of functional
procedures for spatial analysis of precipitation in mountainous terrain,” forms of rain gauge–radar (NEXRAD) data relationships,” CD-ROM
Journal of Hydrology, 58 (1–2): 119–141, 1992. Proceedings of World Environmental and Water Resources Congress, edited by
Presti, R. L., E. Barca, and G. Passarella, “A methodology for treating miss- R. W. Babcock and R. Walton, 2008, pp. 1–9, doi:10.1061/40976(316)413.
ing data applied to daily rainfall data in the Candelaro River Basin (Italy),” Teegavarapu, R. S. V., “Estimation of missing precipitation records integrat-
Environmental Monitoring and Assessment, 160 (1–4): 1–22, 2010. ing surface interpolation techniques and spatio-temporal association rules,”
Raghavan, S., Radar Meteorology, Springer, The Netherlands, 2003. Journal of Hydroinformatics, 11 (2): 133–146, 2009.
Rainbird, A. F., “Methods of estimating average areal precipitation,” WMO- Teegavarapu, R. S. V., M. Tufail, and L. Ormsbee, “Optimal functional
IHD Report 3, WMO, Geneva, Switzerland, 1967. forms for estimation of missing precipitation records,” Journal of Hydrology,
Regonda S. K., B. Rajagopalan, M. Clark, and E. Zagona, “A multi-model 374: 106–115, 2009.
ensemble forecast framework: application to spring seasonal flows in the Teegavarapu, R., C. Pathak, and S. Chintalapudi, “Infilling missing precipi-
Gunnison River Basin,” Water Resources Research, 42 (9): 1–14, 2006. tation data using NEXRAD Data: use of optimal spatial interpolation and
Rogelis, M. C. and M. G. F. Werner, “Spatial interpolation for real-time data-driven methods,” Proceedings of World Environmental and Water
rainfall field estimation in areas with complex topography,” Journal of Resources Congress, 4659–4667, American Society of Civil Engineers (ASCE),
Hydrometeorology, 14: 85–104, 2013. Reston, Virginia, USA, 2010, doi:10.1061/41114(371)473.
Salas, J. D-J., Analysis and modeling of hydrological time Series, Handbook Teegavarapu, R. S. V., “Spatial interpolation using nonlinear mathematical
of Hydrology, edited by D. R. Maidment, McGraw-Hill, New York, 1993. programming models for estimation of missing precipitation records,”
Sen, Z., “Average areal precipitation by percentage weighted polygon Hydrological Sciences Journal, 57 (3): 383–406, 2012.
method,” Journal of Hydrologic Engineering, ASCE, 3 (1): 69–76, 1988. Teegavarapu, R. S. V., T. Meskele, and C. Pathak, “Geo-spatial grid-based
Sen, Z. and A. G. Eljadid, “Automated average rainfall calculation in Libya,” transformation of precipitation estimates using spatial interpolation meth-
Water Resources Management, 14 (5): 405–416, 2000. ods,” Computers and Geosciences, 40: 28–41, 2012a.
Sene, K., Hydrometeorology: Forecasting and Applications, Springer, The Teegavarapu, R. S. V., A. Goly, C. Viswanathan, and P. Behera, “Precipitation
Netherlands, 2010. extremes and climate change: evaluation using descriptive WMO indices,”
Seo, D-J., “Nonlinear estimation of spatial distribution of rainfall—an indica- Proceedings of World Environmental and Water Resources Congress 2012:
tor cokriging approach,” Stochastic Hydrology and Hydraulics, 10:127–150, 1996. Crossing Boundaries, 1927–1936, American Society of Civil Engineers
Seo, D-J., W. F. Krajewski, and D. S. Bowles, “Stochastic interpolation of (ASCE), Reston, Virginia, USA, 2012b, doi: http://dx.doi.
rainfall data from rain gauges and radar using cokriging-1. Design of org/10.1061/9780784412312.193.
Experiments,” Water Resources Research, 26 (3): 469–477, 1990a. Teegavarapu, R. S. V., A. Goly, and J. Obeysekera, “Influences of Atlantic
Seo, D-J., W. F. Krajewski, and D. S. Bowles, “Stochastic interpolation of multi-decadal oscillation on regional precipitation extremes,” Journal of
rainfall data from rain gauges and radar using cokriging-2. Results,” Water Hydrology, 495: 74–93, 2013.
Resources Research, 26 (5): 915–924, 1990b. Teegavarapu, R. S. V., “Climate change-sensitive hydrologic design under
Shaw, E., K. J. Beven, and N. A. Chappell, Hydrology in Practice, CRC Press, uncertain future precipitation extremes,” Water Resources Research, 49 (11):
Boca Raton, FL, 2011. 7804–7814, 2013.
39_Singh_ch39_p39.1-39.10.indd 9 8/22/16 2:08 PM

Teegavarapu, R. S. V., Floods in a Changing Climate: Extreme Precipitation, Wei, T. C. and J. L. McGuinness, “Reciprocal distance squared method, a
Cambridge University Press, London, 2013b. Computer Technique for Estimating Area Precipitation, Technical Report
Teegavarapu, R. S. V., “Statistical corrections of spatially interpolated mis- ARS-Nc-8.” U.S. Agricultural Research Service, North Central Region, OH,
sing precipitation data estimates,” Hydrological Processes, 28 (11): 3789–3808, 1973.
2014a, doi:10.1002/hyp.9906. Westerberg, I., A. Walther, J-L. Guerrero, Z. Coello, S. Halldin, C-Y. Xu,
Teegavarapu, R. S. V., “Missing precipitation data estimation using optimal D. Chen, et al., “Precipitation data in a mountainous catchment in Honduras:
proximity metric-based imputation, nearest neighbor classification and quality assessment and spatiotemporal characteristics,” Theoretical and
cluster-based interpolation methods,” Hydrological Sciences Journal, 59 (11): Applied Climatology, 101 (3–4): 381–296, 2009.
2009–2026, 2014b, dx.doi.org/10.1080/02626667.2013.862334. WMO, World Meteorological Organization: Guide to Meteorological
Teegavarapu, R. S. V., A. Goly, and Q. Wu, “Comprehensive framework for Instruments and Methods of Observation, 6th edition, Part I, WMO-No. 8,
assessment of radar-based precipitation estimates,” Journal of Hydrologic Geneva, Switzerland, Chap. 6, 1983.
Engineering, ASCE, 2015a, dx.doi.org/ 10.1061/(ASCE)HE.1943- WMO, Guide to Hydrological Practices, WMO-No. 168, World
5584.0001277. Meteorological Organization, Geneva, Switzerland, 1994.
Teegavarapu, R. S. V., C. S. Pathak, R. Mecikalski, and J. Srikishen,“Optimal WMO, Guidelines on Analysis of Extremes in a Changing Climate in Support
solar radiation sensor network design using spatial and geostatistical analy- of Informed Decisions for Adaptation, World Meteorological Organization,
ses,” 2015b, dx.doi.org/10.1080/14498596.2015.1051147. Geneva, Switzerland, 2009.
Tsangaratos, P., R. Dimitrios, and B. Andreas, “Use of artificial neural net- Xia, Y., P. Fabian, A. Stohl, and M. Winterhalter, “Forest climatology: esti-
work for spatial rainfall analysis,” Journal of Earth System Science, 123 (3): mation of missing values for Bavaria, Germany,” Agricultural and Forest
457–465, 2014. Meteorology, 96: 131–144, 1999.
Thiessen, A. J. and J. C. Alter, “Precipitation for large areas,” Monthly Xing, T., Z. Xuesong, and J. Taylor, “Designing heterogeneous sensor net-
Weather Review, 39: 1082–1084, 1911. works for estimating and predicting path travel time dynamics: an informa-
Tung, Y., “Point rainfall estimation for a mountainous region,” Journal of tion-theoretic modeling approach,” Transportation Research Part B:
Hydraulic Engineering, 109 (10): 1386–1393, 1983. Methodological, 57: 66–90, 2013.
USEPA, United States Environmental Protection Agency, Developing spa- Xu, C-Y. and G. L. Vandewiele, “Sensitivity of monthly rainfall runoff mod-
tially interpolated surfaces and estimating uncertainty, Report, els to input errors and data length,” Hydrological Sciences Journal, 39 (2):
EPA-454/R-04-004, Environmental protection Agency, Cincinnati, 2004. 157–176, 1994.
Vieux, B. E., Distributed Hydrologic Modeling using GIS, Water Science and Yoo, J., “Linear programming method considering topographical factors
Technology Library, Kluwer Academic Publishers, The Netherlands, 2001. used for estimating missing precipitation,” Journal of Hydrologic Engineering,
Von Neumann, J., “Distribution of the ratio of the mean square successive 18 (5): 542–551, 2013.
difference to the variance,” Annals of Mathematical Statistics, 12: 367–395,
1941.
39_Singh_ch39_p39.1-39.10.indd 10 8/22/16 2:08 PM

Chapter 40
Snow Distribution and
Snowpack Characteristics
BY
MING-KO WOO AND PHILIP MARSH
ABSTRACT capacity is reduced to precipitate snow. This gives rise to lake-effect snowfall.
The same effect can develop when warm air flows over an open sea, and
In cold regions, snow is not only hydrologically important, but it also exerts
deposits snow on the cold land downwind.
influences on ecosystems and human activities. The distribution of snow is
In cold regions including the cold temperate zone, the Arctic and the
controlled by the location and amount of snowfall, redistribution after depo-
Antarctic, snowfall tends to be larger in areas of high precipitation. Partition
sition, and losses through sublimation and melt. Snowcover exhibits varia-
of precipitation into rainfall and snowfall is predicated largely upon air
tions at several spatial scales, from patterns that vary globally and regionally
temperature, but both forms of precipitation can coexist around 0°C. In gen-
to unevenness at local and micro-scales. Once snowfall is deposited on the
eral, a larger portion of annual precipitation falls as snow at high latitudes and
ground, snowpack properties undergo modifications through snow meta-
at high altitudes compared with the lower latitudes and lower elevations.
morphism that alters the ice crystals, water content, and density of the pack.
Examples of mean monthly precipitation at low-elevation stations located in
Snow properties affect the movement of liquid water through the snowpack.
different climatic regions of Canada demonstrate the effects of latitude and
Of equal importance are the optical and thermal properties of the snow; the
continentality on the proportion of snowfall relative to rainfall (Fig. 42.1).
former influences the albedo therefore the energy balance and the latter
Arctic regions where the winters are long (lasting for over 7 months or longer)
affects insulation of the underlying ground. Both the distribution and charac-
but where precipitation is often of low intensity, the annual snowfall is small.
teristics of snow undergo changes during the cold season. Snow condition in
Coastal belts of the temperate region, such as Chile, Norway, and coastal
the melt period plays a major role in runoff generation.
mountains of western North America, snowfall is much higher than the con-
tinental interior, such as the prairies. Large snowfall nourishes mountain
40.1 INTRODUCTION glaciers and ice caps, like those of British Columbia and Alaska (Meier, 1990).
Snow is hydrologically and climatically important in cold regions, and it also 40.2.2 Snow Redistribution
influences terrestrial ecosystems and human activities. Physical properties of
snow such as its reflectivity or albedo, and its low temperatures and insulating Once fallen, some of the snow undergoes redistribution. In vegetated areas,
capabilities, have feedback on the climate. The amount of snow and the precipitation is intercepted by plants. The stand closure of the vegetation plays
pattern of the snowcover, as well as the internal condition of the snowpack, a role as it limits the amount of snow that can reach the ground as throughfall
are important considerations for melt and for runoff generation that delivers (Pomeroy and Gray, 1995). Coniferous and deciduous trees before their leaves
water for use but creates flood hazards. have fallen are notably effective in intercepting snowfall. For the latter, the
weight of snow can break the leaf-clad branches, but the branches of the
conifers are more pliant and can bend without breaking under heavy snow.
40.2 PROCESSES CONTROLLING SNOW DISTRIBUTION In open areas including the tundra, grasslands, and open fields, wind is the
principal agent that redistributes snow. Topography and vegetation play
40.2.1 Snowfall
important roles (Essery and Pomeroy, 2004; Sturm et al., 2001). Snow is lifted
Latitude, altitude, and proximity to large water bodies affect the amount of from its place of rest and entrained with the wind flow. The snow moves by
precipitation. For winter precipitation, atmospheric moisture content and the creeping close to the surface, by saltation in which snow particles skip along
tracks of storms are also major considerations. The Arctic, with a relatively the surface, or by suspension in the air. These processes transfer the snow
dry atmosphere, has lower precipitation than the subarctic and temperate from exposed sites to locations sheltered from the wind.
regions where active winter cyclones bring about frequent and intense pre- Large amounts of snow accumulated on upper hillslopes are subject to
cipitation events. In addition to latitudinal differences in precipitation, topog- avalanches in which the snow mass moves rapidly down the slopes (Schaerer,
raphy gives rise to notable contrasts in snowfall pattern. As onshore wind 1981). When the snow is saturated with rain-water or meltwater, the mixture
encounters a topographic barrier such as a mountain range, the moisture- of snow and water, known as slush, can flow quickly down the slopes or along
laden air is forced to rise and cooling of the rising air yields high orographic valleys (Gude and Scherer, 1995).
precipitation on the windward slope, usually increasing with elevation. The Sublimation from the snowpack occurs when ice crystals are converted to
western side of Norwegian mountains, for example, receives 2000 to 3000 mm water vapor, which is then lost to the atmosphere. This process is favored by
of precipitation from the onshore westerlies, whereas on the leeward side such an atmosphere of low humidity and by high wind speed, and is further
as around Oslo, annual precipitation drops to around 1000 mm (Killingtveit, enhanced during blowing snow events. In vegetated areas, the snow
1994). Another situation is where winter air flows over a large body of open intercepted by leaves and branches is also subject to sublimation. An
water, such as the Great Lakes of North America. Moisture is picked up by the experiment that continuously weighed a cut tree in the boreal forest, with
air during its passage over the water body, but as the air encounters the cold intercepted snow on this tree, showed that about one-third of the intercepted
land downwind of the water body, the air is chilled and its moisture-holding snow could be sublimated in the winter (Pomeroy et al., 1998).
40-1
40_Singh_ch40_p40.1-40.6.indd 1 8/22/16 2:08 PM

40-2 Snow Distribution and Snowpack Characteristics
Figure 40.1 Partitioning of mean monthly precipitation into rainfall and snowfall for stations representative of different geographical settings in Canada.
40.3 SPATIAL PATTERNS OF SNOW AT VARIOUS SCALES ment, interception and sublimation loss leave the forest floors with less snow
than open areas. The amount of snow on the forest floor is uneven due to
The distribution of snow on the ground is highly uneven and the unevenness
throughfall, sublimation, and shedding of the intercepted canopy snow. When
increases as the melt season progresses. Depending on the duration that the
melt begins, shadows cut down short-wave radiation, but tree trunks, being
snow occupies the ground, the snowcover may be considered as ephemeral,
warmer than the snow, emits long-wave radiation to melt the snow
persistent, or prolonged. Areas with ephemeral snowcover, including the warm
preferentially around the trees to form tree-wells, which are bowl-shaped
and cool temperate zones, are subject to frequent above-freezing winter tem-
depressions in the snow surrounding the trunks (Sturm 1992; Woo and Steer
peratures, or have multiple episodes of rain events, both of which lead to the
1986). Snowmelt rate is lower in the forest than the open areas that are more
removal of snow only days or weeks after its deposition. Most cold temperate
exposed to high radiation and wind. The snow then lingers in the forest even
regions have a persistent snowcover that lasts for several months. At higher lati-
after the open sites become snow-free. Logging and forest fire not only
tudes, the snowcover is prolonged for over 6 months every year and some snow
modify the timing and the rate of snowmelt, but the presence of wood debris
has acquired semipermanent status or is incorporated into glaciers and ice caps.
and subsequent growth of grasses and brushes also alter the snow accumula-
Within a climatic region, spatial variation in snow pattern reflects the
tion pattern, as does the plantation of trees in the disturbed woodlands.
influence of storm tracks, topography, and proximity to water bodies.
Urban areas exhibit snow patterns distinct from their surrounding countryside.
Windward and leeward slopes have different amount of snow. Differential
Buildings and other structures alter the wind direction and speed to affect snow-
snowmelt rates further the variation in snowcover on slopes of different eleva-
fall and snow deposition. Snow removal from roofs and streets create bare sur-
tions and orientations due to temperature, radiation, and wind. In extensive
faces during the winter. Urban heat island effect and heat advection from bare
open areas with sparse vegetation, such as the Arctic and open fields of the
surfaces render the cities and towns clear of snow much before their rural
prairies and steppes, terrain enables local acceleration or reduction in wind
neighbors. From the air, cities present a mixed pattern of snow on rooftops, open
flow, leading to the development of typical snow patterns indicative of snow
grounds, and linear features where the snow plows along highways.
scour and deposition (Rees et al., 2013). On a local scale (102–103m in length
scale), minor snow features in the forms of ripples, dunes, and sastrugi are 40.4 SNOWPACK CHARACTERISTICS
commonly superimposed on the broad snowfields. One special snow phe-
nomenon is created by lake-effect snowfall in the Great Lakes area where the 40.4.1 Snowpack Metamorphism
wind picks up moisture from the open water surface of the lakes and deposits Falling snowflakes occur in many shapes and sizes, including stellars,
snow on the downwind sides of the lakes (Phillips and McCulloch, 1972). As columns, plates, and grauples, for example (Fierz et al., 2009), but after a few
the winds are predominantly from the west during the winter, repeated snow- days or even hours, they undergo rapid metamorphic changes that can occur
fall events accumulate much snow to form snowbelts with snowcovers that extremely quickly since the snow is typically close to the freezing point. The
last longer than their nearby areas. resulting snowpack (referring to the accumulation of snow at a specific site)
A difference in vegetation affects the snowcover distribution. Marsh et al. consists of snow grains (or particles) with impurities and pore spaces filled
(2003), for example, found that end-of-winter snow water equivalent (SWE) with air, water vapor, and liquid water when the snow is wet. Snowpacks are
near the treeline in the Mackenzie Valley, NWT, was 98 mm on the tundra, typically layered as the result of individual storms (either snowfall or blowing
141 mm for the shrubs, and 155 mm for the woodland. In a forest environ- snow deposition) (Colbeck, 1991). Understanding snow stratigraphy, and
40_Singh_ch40_p40.1-40.6.indd 2 8/22/16 2:08 PM

Discussion and Conclusion 40-3
how it evolves in time, provides key information required for understanding increased in the visible when trace amounts of impurities are present in the
snow hydrology, avalanche formation, and heat and moisture exchange snow. As clouds absorb radiation in similar ways to snow, they increase both
between the underlying soil, snowpack, and the atmosphere. An international the fraction of shortwave radiation reaching the snow surface as well as the
snow on the ground classification has been provided by Fierz et al. (2009), a diffuse radiation, leading to an increase the snow albedo.
regional classification by Sturm et al. (1995), and observational methods by
Pielmeier and Schneebeli (2003). 40.4.3 Thermal Properties of Snow
Rapid metamorphic changes described by Colbeck (1997) include changes in
grain size, grain shape, snow density, and specific surface area (SSA), defined as Snow is an excellent insulator, and hence strongly controls the snowpack tem-
the surface area per unit mass (Proksch et al., 2015) and degree of sintering perature gradient. Sturm et al. (1997) reviewed the many attempts to estimate
between particles. Each change significantly impacts the optical, thermal, and snow thermal conductivity and presented data that showed a strong relation-
hydrologic properties of the snowpack. For example, optical properties are ship of conductivity for rounded grains or wind-blown fragments with snow
strongly related to changes in snow grain size, mechanical properties to the density. Sturm and Johnson (1992) noted that snow texture (size, shape, bond-
bonds between grains, and thermal properties to microstructure (Kaempfer ing, and spatial arrangement of grains) also impacts thermal conductivity, but
and Schneebeli, 2007). In snowpacks with a temperature below 0°C and with such observations are difficult to make, and it is believed that for faceted grains
little liquid water, metamorphism depends on whether the snowpack is isother- (i.e., depth hoar) there is little connection between density and thermal conduc-
mal or has a temperature gradient, while in snowpacks with a temperature of tivity. To demonstrate such changes, Sturm et al. (2002) suggested a timeline of
0°C, metamorphism depends on the liquid water content. changes in thermal conductivity, where after a snowfall, the thermal conductiv-
With equitemperature metamorphism (ETM) the snow is isothermal at a ity would increase as the snow hardens, then decrease with increasing depth
scale that includes a significant portion of the snowpack, but at the grain scale hoar formation, and finally if the snow is wetted by melting and then refreezes,
there are small temperature and vapor pressure gradients related to particle the conductivity would again increase. Calonne et al. (2011) used both experi-
size and curvature around and between snow particles. These gradients lead to mental observations and modeling to study various snow properties, one of the
sublimation of ice from the convex portions of the particles and condensation results being the finding of anisotropy, with the vertical thermal conductivity of
on the concave areas, with large grains growing at the expense of small grains facetted grains being 1.5 times higher than the horizontal values, and the oppo-
and producing well-rounded snow particles (Colbeck, 1982). Deposition also site occurring for rounded grains.
occurs at the concave connections where two particles touch (referred to as
sintering) to result in grain bonding (Kaempfer and Schneebeli, 2007). 40.4.4 Liquid Water Movement Through Snow
Temperature gradient metamorphism (TGM) occurs when large gradients of Melting snowpack consists of a complex mixture of rounded and sintered
greater than 10 to 20°C m–1 exists, often where the base of the snowpack is grains, or faceted particles, of low density and low thermal conductivity.
warmer than the snow surface, but also when temperature of the snow surface When the snow is melting and freely draining, water saturation is typically
layer is reduced by radiative heating and cooling. Moisture is transferred along less than 7% for coarse grained snow and 7 to 15% for finer grained snow
this temperature-induced vapor pressure gradient, with sublimation and con- (Marsh, 2005). When melt stops and a snowpack drains, the water saturation
densation from one snow particle to the next, described by Yosida (1955) as declines to the irreducible value, between 2 and 7% depending on snow prop-
“hand-to-hand transport,” and the resulting recrystalization produces kinetic erties. As snow grains grow in size when wetted, the irreducible water satura-
growth crystal forms (Colbeck, 1997). Pinzer et al. (2012) demonstrated that the tion decreases with time since wetting. The permeability of melting snowpacks
rapid flux of vapor can lead to a turnover of up to 60% of the total snow mass is related to both grain size and density, with a commonly used parameteriza-
per day. A typical kinetic growth form, referred to as depth hoar, is in form of tion provided by Shimizu (1970).
faceted snow crystals (Fierz et al., 2009) that typically have large grain size, low When meltwater enters into a dry snowpack, a wetting front forms that
density, and low strength. Collapse of depth hoar layers under loading (storm divides the snowpack into a zone with an upper layer of wet snow at 0oC and
or human) is a common avalanche trigger, while the low thermal conductivity a lower zone with no liquid water (Marsh and Woo, 1984). However, the wet-
of these layers retards heat transfer through the snow and decouples atmo- ting front quickly becomes heterogeneous and preferential flow paths form
spheric conditions from soil freezing (Sturm and Benson, 1997). (Marsh and Woo, 1984; Schneebeli, 1995; Waldner et al., 2004). Such prefer-
In snow at 0°C and with liquid water fully or partially filling the pore spaces, ential flow paths play an important role in moving meltwater quickly through
melt-freeze metamorphism (MFM) and grain growth rate depend on water the snowpack, but modeling of these features remains a challenge. The inter-
content. In typical freely draining snowpacks with a low liquid water content, action of flow within these flow paths and stratigraphic boundaries in cold
the air in the pores occupies continuous paths throughout the snowpack and snowpacks results in the formation of large ice lenses, and in sufficiently cold
liquid saturation is below 14% of the pore volume. Grain growth is much faster snowpacks water also freezes in the vertical flow paths to develop ice columns
than during ETM and the resulting snow is characterized by clusters of rounded (Marsh and Woo, 1984). Lateral flow of water, especially through deep snow-
grains (Colbeck, 1982). With higher saturation, the liquid is continuous covers (Eriksson et al., 2013) also plays an important role in the movement of
throughout the pores and air only occurs in isolated bubbles. Grain growth is meltwater to the stream channel.
much higher than in nonsaturated snow (Colbeck, 1982; Marsh, 1987). Snow
grains develop into a true equilibrium form, with rounded and individual
grains. In such “slush,” the snow has very low strength and is cohesionless. 40.5 DISCUSSION AND CONCLUSION
Both the spatial distribution and characteristics of snowcovers are altered as

snowfall is deposited and as the snowpack undergoes accumulation, blowing
40.4.2 Optical Properties of Snow and metamorphic processes. Sturm et al. (1995) presented a seasonal snow
Snowmelt, a key driver of streamflow in cold regions, is controlled by turbu- classification for snowcovers that evolve over the winter and melt periods in
lent fluxes of heat and absorption of shortwave radiation, and in many cases major environments across the globe. With the warmest snow, ephemeral
the shortwave energy balance dominates melt (Male and Gray, 1982). The snowcovers are typically less than 0.5 m deep, often formed of a single snow-
actual absorption of shortwave radiation is controlled by various factors fall, with melt metamorphism beginning soon after deposition, and melt from
(Warren, 1982). Snow albedo is low (less than 0.4) in the near infrared (IR) both the surface and base leading to complete melting of the snow over a
(i.e., wavelengths > 0.7 μm) due to the strong absorption by ice. As a result, short period of time. A snowcover may form again after the next snowfall. In
near-IR radiation plays an important role in snowmelt. Albedo in the visible cooler maritime climates, a deeper (up to 3 m) snowcover with multiple layers
wavelengths (0.3 to 0.7 μm) ranges from near 0.4 to 1.0. deposited by individual storms is typical. At warm temperatures, TGM is rare
Snow albedo decreases with increasing grain size, ranging from 0.9 for a while ETM occurs over much of the winter. However, winter melt events are
radius of 0.1 mm to 0.68 at 2.5 mm. As snow grains enlarge during melting, common and MFM dominates, producing ice layers and large melt-rounded
albedo decreases and melt rate increases. Snow albedo is also very sensitive to snow grains. A typically protracted spring melt season results in gradual
impurities in the snow, especially for strongly absorbent impurities, including removal of the snow. Under increasingly cold conditions, such as those of the
soot, dust and volcanic ash, with the greatest effect for snow with large grain prairies and tundra, winter melt events become infrequent and total snowfall
size (Warren and Wiscombe, 1980). Especially important is black carbon is usually low. The snow is normally shallow (less than 1 m) and consists of a
(BC) (Doherty et al., 2010) produced by incomplete combustion by diesel small numbers of layers. Substantial blowing snow produces wind slabs, sas-
engines, coal burning, forest fires, agricultural fires, and residential wood trugi, and deep snowdrifts. Although ETM occurs, large temperature gradi-
burning. Hadley and Kirchstetter (2012) noted that BC of only 10 to 100 parts ents usually lead to dominance of TGM. Both the magnitude and duration of
per billion by mass reduces albedo by 1 to 5%, sufficient to increase melt and the temperature gradients and TGM increase in the tundra regions, causing a
have a positive radiative climate forcing. Albedo, especially in the near-IR large portion the snowcover to be composed of depth hoar by the end of
(Warren, 1982), is higher as the sun approaches the horizon, and this effect is winter. Taiga snowcovers are also dominated by low snowfalls, but low winds
40_Singh_ch40_p40.1-40.6.indd 3 8/22/16 2:08 PM

40-4 Snow Distribution and Snowpack Characteristics
Bare ground
Snow cover
Water/Ice
Cloud
Sep. 30, 2010 Oct. 16, 2010 Nov. 17, 2010
Elevation (m)
2700
30
0 50 100 200 Km
May 1, 2011 May 17, 2011
Figure 40.2 Seasonal change in snow distribution in Liard basin in northern Canada during the winter of 2010–11. [Source: MODIS Snow Cover 8-Day L3 Global 500m
Grid (MOD10A2), obtained from the NSIDC.]
limit the frequency of blowing events. Under cold temperatures, large tem- REFERENCES
perature gradients and TGM prevail, ending with snow that is of low density
Calonne, N., F. Flin, S. Morin, B. Lesaffre, S. R. Du Roscoat, and C.
and dominated by depth hoar. Due to their shallow depth, prairie, tundra, and
Geindreau, “Numerical and experimental investigations of the effective ther-
taiga snowcovers often disappear quickly during spring melt. In contrast,
mal conductivity of snow,” Geophysical Research Letters, 38: 1–6, 2011,
alpine and mountain areas have highly variable snowcovers depending on
doi:10.1029/2011GL049234.
orientation, steepness, and elevation. Consequently, the snowpack varies from
Colbeck, S. C., A review of sintering in seasonal snow, CRREL Rep., 97–10,
being cold and deep, commonly with many layers and a base of depth hoar, to
Cold Regions Research and Engineering Laboratory, U.S. Army Corps of
being shallow and ephemeral. The melt period is similarly variable, ranging
Engineers, 1997, p. 12.
from protracted melt at high elevations with deep snow, to short and early
Colbeck, S. C., “An overview of seasonal snow metamorphism,” Reviews of
melt in low elevations where the snow is relatively shallow and warm.
Geophysics, 20 (1): 45, 1982, doi:10.1029/RG020i001p00045.
Snow distribution pattern evolves continually from the time of snowcover
Colbeck, S., “The layered character of snow covers,” Reviews of Geophysics,
formation to the time when snow disappears, as is illustrated by the melt-
29 (1): 81–96, 1991.
season snowcover of Liard River basin (area 275,000 km2) in northwestern
Doherty, S. J., S. G. Warren, T. C. Grenfell, A. D. Clarke, and R. E. Brandt,
Canada, where large-scale topography exerts prominent control. Western
“Light-absorbing impurities in Arctic snow,” Atmospheric Chemistry and
Liard basin lies in the Western Cordillera, reaching more than 3000 m and
Physics, 10 (23): 11647–11680, 2010, doi:10.5194/acp-10-11647-2010.
dissected by many headwater tributaries that unite and drain eastward. The
Eiriksson, D., M. Whitson, C. H. Luce, H. P. Marshall, J. Bradford, S. G.
eastern section of the basins is part of the Interior Plains, rising from 170 to
Benner, T. Black, et al., “An evaluation of the hydrologic relevance of lateral
800 m at the foothills. Figure 40.2 shows the changing snowcover pattern in
flow in snow at hillslope and catchment scales,” Hydrological Processes, 27 (5):
the winter of 2010–11. Topographic effect is demonstrated by late buildup but
640–654, 2013, doi:10.1002/hyp.9666.
early disappearance of snow on lowlands in the east, branching out into the
Essery, R. and J. Pomeroy, “Vegetation and topographic control of wind-
valleys that extend into the western mountain chains. High mountains
blown snow distributions in distributed and aggregated simulations for an
remained snow-clad in mid-May. Within the highland zone, the southern
Arctic tundra basin,” J. Hydrometeorology, 5 (5): 735–744, 2004.
mountains lost their snow earlier than those in the northern end of the basin,
Fierz, C., R. L. Armstrong, Y. Durand, P. Etchevers, E. Greene, D. M.
suggesting the influence of latitude.
McClung, K. Nishimura, et al., The international classification for seasonal
The changing mosaic of snow distribution and continuous alteration of
snow on the ground, IHP-VII Technical Documents in Hydrology No. 83, IACS
snow properties significantly affect snowmelt but are themselves modified by
Contribution No. 1, UNESCO-IHP, Paris, 2009.
melt events. These considerations prominently control melt runoff genera-
Gude, M. and D. Scherer, “Snowmelt and slush torrents–preliminary report
tion, which is the topic for another chapter of this handbook.
from a field campaign in Kärkevagge, Swedish Lappland,” Geografiska
Annaler, 77A: 199–206, 1995.
Hadley, O. L. and T. W. Kirchstetter, “Black-carbon reduction of snow
ACKNOWLEDGMENT
albedo,” Nature Climate Change, 5 (March), 2012, doi:10.1038/
We thank Dr. Laura Brown for preparing the figures. NCLIMATE1433.
40_Singh_ch40_p40.1-40.6.indd 4 8/22/16 2:08 PM

REFERENCES 40-5
Killingtveit, Å., Hydrology in northern Norway, Northern Hydrology: Journal of Geophysical Research: Earth Surface, 120: 346, 2015, doi:
International Perspectives, edited by T. D. Prowse, C. S. L. Ommanney, and L. 10.1002/2014JF003266.
E. Watson, National Hydrology Research Institute Science Report No. 3, Rees, A., M. English, C. Derksen, P. Toose, and A. Silis, “Observations of
Saskatoon, Canada, 1994, pp. 81–107. late winter Canadian tundra snow cover properties,” Hydrological Processes,
Kaempfer, T. U. and M. Schneebeli, “Observation of isothermal metamor- 28 (12): 3962–3977, 2013, doi: 10.1002/hyp.9931.
phism of new snow and interpretation as a sintering process,” Journal of Schaerer, P. A., Avalanches, Handbook of Snow: Principles, Processes,
Geophysical Research: Atmospheres, 112 (D24): 24101 2007, doi:10.1029/2007J Management and Use, edited by D. M. Gray and D. H. Male, Pergamum Press,
D009047. Toronto, 1981, pp. 475–518.
Male, D. H. and R. J. Granger, “Snow surface energy exchange,” Water Schneebeli, M., “Development and stability of preferential flow paths in a
Resources Research, 17 (3): 609–627, 1981. layered snowpack,” IAHS Publications, 228 (228): 89–95, 1995.
Marsh, P., “Grain growth in a wet arctic snow cover,” Cold Regions Science Shimizu, H., “Air permeability of deposited snow,” Contributions No. 1053,
and Technology Journal, 14 (1): 23–31, 1987, doi:10.1016/0165-232X Institute of Low Temperature Science, Hokkaido, Japan, pp. 1–32, 1970.
(87)90041-3. Sturm, M. and C. Benson, “Vapor transport, grain growth, and depth-hoar
Marsh, P., “Water flow through snow and firn,” Encyclopedia of Hydrological development in the subarctic snow,” Journal of Glaciology, 43 (143): 42–59,
Sciences, Vol. 4, Part 14, John Wiley & Sons, Chichester, UK, 2005, pp. 97–123. 1997.
Marsh, P., C. Onclin, M. Russell, and S. Pohl, “Effects of shrubs on snow Sturm, M., “Snow distribution and heat flow in the taiga,” Arctic Alpine
processes in the vicinity of the Arctic treeline in NW Canada,” Proceedings of Research, 24: 145–152, 1992.
the 14th Northern Research Basins Symposium and Workshop, Kangerlussuaq, Sturm, M., “Thermal conductivity and heat transfer through the snow on
Greenland, 2003, pp. 113–118. the ice of the Beaufort Sea,” Journal of Geophysical Research, 107 (C10), 2002,
Marsh, P. and M. K. Woo, “Wetting front advance and freezing of meltwater doi:10.1029/2000JC000409.
within a snow cover 1,” Observations in the Canadian Arctic, Water Resources Sturm, M., J. Holmgren, and G. E. Liston, “A seasonal snow cover classifica-
Research, 20: 1853–1864, 1984. tion system for local to global applications,” Journal of Climate, 8: 1261–1283,
Meier, M. F., Snow and ice, The Geology of North America, edited by M. G. 1995.
Wolman and H. C. Riggs, Geological Society of America, Boulder, CO, Vol. Sturm, M. and J. B. Johnson, “Thermal conductivity measurements of
O-1: Surface Water Hydrology, 1990, pp. 131–158. depth hoar,” Journal of Geophysical Research, 97 (Figure 2): 2129, 1992,
Phillips, D. W. and J. A. W. McCulloch, The Climate of the Great Lakes doi:10.1029/91JB02685.
Basin, Atmospheric Environment Service, Toronto, 1972, p. 40. Sturm, M., J. McFadden, G. Liston, S. Chapin, C. Racine, and J. Holmgren,
Pielmeier, C. and M. Schneebeli, “Developments in the stratigra- “Snow-shrub interactions in Arctic tundra: a hypothesis with climatic impli-
phy of snow,” Surveys in Geophysics, 24: 389–416, 2003, cations,” Journal of Climate, 14: 336–344, 2001.
doi:10.1023/B:GEOP.0000006073.25155.b0. Waldner, P. A., M. Schneebeli, U. Schultze-Zimmermann, and H. Flühler,
Pinzer, B. R., M. Schneebeli, and T. U. Kaempfer, “Vapor flux and recrystal- “Effect of snow structure on water flow and solute transport,” Hydrological
lization during dry snow metamorphism under a steady temperature gradient Processes, 18 (April 2002): 1271–1290, 2004, doi:10.1002/hyp.1401.
as observed by time-lapse micro-tomography,” Cryosphere, 6: 1141–1155, Warren, S. G., “Optical properties of snow,” Reviews of Geophysics and
2012, doi:10.5194/tc-6-1141-2012. Space Physics, 20 (1): 67–89, 1982.
Pomeroy, J. W. and D. M. Gray, Snow Cover: Accumulation, Relocation and Warren, S. G. and W. J. Wiscombe, “A model for the spectral albedo of
Management, National Hydrology Research Institute Science Report No. 7, snow. II: Snow containing atmospheric aerosols,” Journal of Atmospheric
Saskatoon, Canada, 1995, p. 144. Sciences, 37: 2734–2745, 1980.
Pomeroy, J. W., J. Parvianinen, N. Hedstrom, and D. M. Gray, “Coupled Woo, M. K. and P. Steer, “Monte Carlo simulation of snow depth in a for-
modeling of forest snow interception and sublimation,” Hydrological Processes, est,” Water Resources Research, 22: 864–868, 1986.
12: 2317–2337, 1998. Yosida, Z. and Colleagues, “Physical studies on deposited snow 1: Thermal
Proksch, M., H. Lowe, and M. Schneebeli, “Density, specific surface area, properties,” Institute of Low Temperature Science, 7: 19–74, 1955.
and correlation length of snow measured by high-resolution penetrometry,”
40_Singh_ch40_p40.1-40.6.indd 5 8/22/16 2:08 PM

Chapter 41
Time-Space Modeling of
Precipitation
BY
M.L. KAVVAS, N. OHARA, K. ISHIDA, AND A. ERCAN
ABSTRACT least 10 × 100 year projections of the precipitation sequences, ideally at

hourly increments, at either some specified locations or over the whole of the
For the water resources planning and management studies during the last
target region. Until recently, it was not practical to perform simulations of the
century, the necessary long-term simulations of the precipitation fields over a
precipitation fields at a target region at sufficiently fine spatial (less than
target region were performed by stochastic precipitation models. After the
10-km grid) and temporal (less than or equal to 1 h) resolutions by the deter-
necessary advances in computational resources and modeling technologies
ministic numerical atmospheric models. Therefore, for the water resource
were achieved toward the end of the 20th century, in the 21st century it has
planning/management studies during the last century, the necessary long-
been possible to simulate the precipitation fields at a target region also by
term simulations of the precipitation fields over a target region were per-
deterministic numerical atmospheric models at sufficiently fine spatial and
formed by stochastic precipitation models.
temporal resolutions. This chapter reports on the existing approaches to the
In the case of the design of ordinary hydraulic structures, it is necessary to
stochastic and deterministic numerical modeling of precipitation, as well as
estimate the design flood for a specified return period which is typically taken
the remote sensing approaches. Modeling of time-space precipitation is
to be 100 years. In order to develop such an estimate, it is then necessary to
reported with respect to three purposes: the water resources planning, the
estimate the flood peak discharge that would correspond to the 100-year
design of hydraulic structures, and the real-time management of water
return period. However, only in few locations around the world flow records
resources systems. Numerical modeling approaches to regional-scale or
that exceed 100 years exist. In most locations, it is necessary to extend the flow
watershed-scale time-space precipitation to simulate historical, real-time, and
record-based empirical flood frequency curve to 100-year or longer return
future conditions are presented. A recently developed methodology for the
periods either by means of a fitted theoretical frequency distribution or by
physically-based Probable Maximum Precipitation (PhysPMP) estimation, or
performing the simulation of an ensemble of flow projections (described
more technically Maximum Precipitation estimation (MPE), is described.
earlier) from which one can then estimate the 100-year or longer return
MPE approach, which does not require the assumption of stationarity, can
period flood hydrographs that can then be used for the design of the hydrau-
account for the physical features of precipitation modeling, including the
lic structures. Within this framework, it is again necessary to perform long-
orographic effect and the blocking effect of mountains, that cannot be prop-
term simulations of the precipitation fields as the necessary input to the target
erly handled by the traditional PMP approach.
region’s hydrologic model’s simulations. Due to the aforementioned problems
with the numerical atmospheric models, during the last century simulations
of the precipitation fields for hydrologic design were again performed by
41.1 INTRODUCTION
stochastic methods.
In the planning and long-term management of the water resources of a geo- Within the aforementioned framework, first the existing approaches to the
graphical region, the engineer needs to consider a wide range of possibilities stochastic modeling of precipitation will be discussed. Since starting around
for the water conditions during the future planning/operational horizon at the end of the twentieth century and the beginning of the twenty-first cen-
the target water resources system. This requires the projection of the uncer- tury, the necessary advances in computational resources and modeling tech-
tain but possible hydrologic conditions (streamflows, soil water, evapotrans- nologies for the simulation of the precipitation processes by deterministic
piration, groundwater recharge rates, and levels, etc.) during the planning/ numerical atmospheric models for long time periods at fine time-space reso-
operational horizon at the target region. However, since the hydroclimate lutions has been achieved, the deterministic numerical atmospheric modeling
conditions during the future are uncertain, one needs to consider a wide approaches to the simulation of the precipitation fields will then be reported.
range of future hydrologic water supply and water demand conditions.
Typically, the future hydrologic conditions are simulated by numerical hydro- 41.2 STOCHASTIC MODELING OF PRECIPITATION
logic models of the target region with the projected atmospheric inputs.
Precipitation is the fundamental one among these inputs, since it provides the Precipitation is a complicated process which is the outcome of the interaction
essential water input to the hydrologic model of the target region. Since a of dynamical, thermodynamical, and microphysical atmospheric processes
typical planning/operational horizon is in the order of a century, it is neces- occurring in continuous time and space. However, the precipitation process,
sary to simulate the hydrologic conditions, and, thereby, the necessary pre- as observed at a gauging station or at a group of stations at hourly or daily
cipitation inputs, at fine time increments over the horizon of a century for increments at a study region, emerges as a temporal stochastic process with
each individual projection that represents one possible future realization. In uncertain times for the occurrence of wet periods, uncertain dry periods, and
order to develop sound policies for the target water resources system, it is uncertain precipitation amounts. Within this context, first the approaches to
necessary to simulate a sufficiently large ensemble of future hydrologic the modeling of precipitation as a temporal stochastic process will be dis-
projections to be able to develop sound statistical inferences for the out- cussed. Then the more comprehensive, yet more complicated models of time-
comes of possible policies for this system. As such, it is desirable to have at space precipitation will be reported.
41-1
41_Singh_ch41_p41-1-41-14.indd 1 8/22/16 2:09 PM

41-2 Time-Space Modeling of Precipitation
41.2.1 Stochastic Modeling of Precipitation in Time at stationary. Hence, they need to be applied to predefined time periods within
Fixed Spatial Locations which the stationarity of the precipitation process can be assumed.
As observed at a gauging station, the precipitation occurrence process can be Precipitation being basically a continuous time-space process, a different
modeled as a two-state process, consisting of a wet and a dry state. Once a wet approach can also be taken to modeling precipitation in time by means of
state occurs, this state can be marked by a random precipitation quantity for continuous-time point process stochastic models. Point process stochastic
the considered time increment (daily or hourly). This two-state precipitation models consider the occurrence locations of precipitation events in time or in
occurrence process can be modeled conveniently as a two-state Markov chain time space as random points. After modeling, the occurrence of precipitation
with a two-by-two transition probability matrix that describes the dry-to-dry, event origin locations, then a mark, either as a random value, or a random
dry-to-wet, wet-to-dry, and wet-to-wet state transition probabilities when rectangular pulse or a random signal can be attached to each origin location
moving from one time increment to the next. Once the state probabilities of in order to describe the precipitation amounts as accumulated values at
being in a wet state and in a dry state at the time origin of the process are desired time increments. The simplest of such point process stochastic mod-
defined, it is straightforward to estimate the probability of being in a specified els of precipitation is the Poisson process model. The Poisson process has the
state (wet or dry) at any specified time of the process. Assuming that the unique property that the number of event occurrences in nonoverlapping
outcome of the process (wet or dry) at any given day is only dependent on the time-space intervals are mutually independent with Poisson probability dis-
state of the process in the previous day, the Markov chain model is a practical tribution. The time intervals between consecutive events are also independent
model that was applied by many researchers during the twentieth century to with exponential distribution. The Poisson process can be described by a
the modeling of precipitation occurrences (Gabriel and Neumann, 1957; single parameter function, the occurrence rate function, that may be either a
Wiser, 1965; Feyerherm and Bark, 1967; Smith and Schreiber, 1973; Roldan constant value, or a function of time, depending on whether the studied pro-
and Woolhiser, 1982; Woolhiser and Osborne, 1985). It is also possible to cess is considered as stationary or nonstationary in time. Accordingly, the
generalize the two-state Markov chain for wet and dry state occurrences to a Poisson process was employed by various researchers for modeling precipita-
multistate Markov chain where the precipitation amounts can be classified tion fields (Todorovic and Yevjevich, 1969; Gupta and Duckstein, 1975;
into a number of interval classes during a wet time increment. This way one Eagleson, 1978; Rodriguez-Iturbe et al., 1984, 1987; Valdes et al., 1985;
ends up with an n by n transition probability matrix, where n becomes the Foufoula-Georgiou and Guttorp, 1987, among others). In order to model
number of precipitation depth intervals plus one for the dry state. Various precipitation depths at various time intervals, Rodriguez-Iturbe et al. (1984)
researchers applied the multistate Markov chain model to describe the occur- marked the Poisson occurrence time locations of the precipitation events
rences and amounts during a precipitation process at discrete time increments either by an independent random depth value [Poisson white noise model
(Hopkins and Robillard, 1964; Haan et al., 1976; Srikanthan and McMahon, (PWNM)] or by a rectangular pulse with random independent precipitation
1983). Also, higher-order Markov chain models for winter time daily rainfall rate and independent duration [Poisson rectangular pulse model (PRPM)]. In
over continental United States were developed by Chin (1977). While being their application of these models to hourly and daily precipitation sequences
easy to apply, the Markov chain models suffer from having a short memory, in the United States and Venezuela, they observed that the parameter esti-
being unable to account for the observed persistence during long wet periods mates in these models vary with the time increment sizes. They also observed
due, for example, to the passage of a synoptic extratropical cyclonic system, that these models cannot preserve the correlation structure of the hourly
and also during long dry periods. A promising generalization to the Markov precipitation sequences. Foufoula-Georgiou and Guttorp (1987) also report-
chain models is the Markov renewal process (MRP) model of Foufoula- ed the inadequacy of the PWNM in modeling short time increment rainfall.
Georgiou and Lettenmaier (1986) for daily rainfall occurrences. The funda- Based on their studies of the extratropical cyclonic precipitation systems in
mental advantage of the MRP model over a Markov chain is that on a given the United States, Austin and Houze (1972) and Hobbs and Locatelli (1978)
day the probability of being on a wet state (rainy day) or a dry state does not observed that precipitation is organized in distinct regions within the synop-
depend only on the previous day being wet or dry, but on the number of days tic scale (>104 km2) cyclones with respect to temporal and spatial scales as
since a wet day. It has a Markovian renewal structure in that the length of the follows: (a) synoptic areas which have lifetimes of one to several days and
interval in between the occurrences of a particular state has a probability horizontal spatial extents > 104 km2, (b) large mesoscale areas (LMSAs), also
distribution (taken as a geometric distribution) determined by Markovian called “rainbands,” which last several hours and have spatial horizontal
transition probabilities among wet and dry states. Within this framework, the extents in the range 103–104 km2; (c) small mesoscale areas (SMSAs) or pre-
MRP can account for the effect of different rainfall mechanisms. cipitation cores which last from several minutes to an hour and have horizon-
Since, as observed at a gauging station, the precipitation process can be tal spatial extents in the 100–800 km2 range; and (d) rain cells which last from
interpreted as a wet-dry two-state process, it is possible to also model the several minutes to an hour and have horizontal spatial extents in the order of
process as an alternating renewal process of dry and wet periods that alternate 10 km2. They observed that rainbands are located within synoptic areas of the
randomly in time (Grace and Eagleson, 1966; Todorovic and Yevjevich, 1969; cyclonic system while the rain cores and cells are mainly located within rain-
Galloy et al., 1981; Roldan and Woolhiser, 1982). The basic assumption of the bands. Within this framework, the extratropical cyclonic systems demon-
alternating renewal process model is the independence of the wet and dry strate a clustered precipitation field structure both in time and space. This
period durations. As such, different probability distributions for the wet and cluster structure of the precipitation systems are also typical for the severe
dry period durations can be assigned. If the process is considered in discrete precipitation-producing mesoscale convective systems (MCSs; Parker and
time increments, a negative binomial distribution for both the wet and the dry Johnson, 2004; Trapp, 2013) and for tropical cyclones (TCs; Lonfat et al.,
period durations (Galley et al., 1981), or a truncated geometric distribution 2004). Based on a statistical analysis of rainfall sequences in Indiana, the
for wet period durations and a truncated negative binomial distribution for United States, Kavvas and Delleur (1976, 1981) detected the clustering struc-
dry period durations (Roldan and Woolhiser, 1982) can be taken. Furthermore, ture of the precipitation sequences in time by means of the precipitation
given that the process is in a wet period, a precipitation depth needs to be occurrence counts spectrum and counts variance-time functions. Hence, they
assigned to that wet period independently. Aside from the restrictive inde- developed a Poisson cluster point process model of Neyman–Scott-type
pendence assumption among the sequential dry and wet period durations, (NSPCM) (Neyman and Scott, 1958) for the precipitation occurrence process
another serious issue with the alternating renewal process model is how to in time by means of its various moments. The NSPCM is a two-level Poisson
disaggregate the total precipitation depth during a wet period among the cluster model, where in the primary level, there is a Poisson process for the
individual wet days. A simple solution to this problem is to consider either a arrivals of rainfall generating mechanisms (RGMs), representing cyclonic
sequence of independent rain depths for each of the individual time incre- fronts or convective systems. At the secondary level of NSPCM, to each RGM
ments of a wet period, or to consider a multistate Markov chain in order to of the primary level process there corresponds a random number (geometri-
account for the autocorrelation among the precipitation depths of the sequen- cally distributed) of individual rain cells that are independently located (with
tial wet days during a wet period. exponential distribution) with respect to the RGM arrival time, forming a
As a way to relax the independence assumption in the wet and dry period cluster around that RGM. Rodriguez-Iturbe et al. (1984) expanded the
durations of the alternating renewal process model, and to expand the short NSPCM by marking each of the individual rain cell occurrences with an
memory structure of the Markov chain model for the wet–dry day occur- independent random rain depth and deriving the moment of the rainfall
rences, discrete autoregressive moving average models (DARMA, Jacobs and depths for various time intervals [Neyman–Scott white noise marked Poisson
Lewis, 1978) were also applied to daily rainfall sequences (Buishand, 1978; cluster process (NSWNPCP)]. Their application of the NSWNPCP model to
Chang et al., 1984). The appeal of DARMA models is their flexible structure hourly and daily rainfall depths in Colorado and Venezuela showed that
in modeling discrete-state time series with zero values, accommodating vari- NSWNPCP is superior to Poisson occurrence process-based marked rainfall
ous memory lengths in a parsimonious manner. Hence, they can be useful in depth models. However, a separate application by Valdes et al. (1985) showed
modeling droughts with long-term memories in a parsimonious way. that NSWNPCP could not model the statistics of extreme precipitation satis-
However, while the precipitation process is seasonal, DARMA models are factorily. In order to overcome the shortcomings of NSWNPCP, Rodriguez-Iturbe
41_Singh_ch41_p41-1-41-14.indd 2 8/22/16 2:09 PM

Deterministic numerical modeling of time-space precipitation 41-3
et al. (1987a, 1987b) introduced the Bartlett–Lewis process for rain cell arriv- fields by Waymire et al.’s (1984) model, Valdes et al. (1985) contended that the
als with each of the rain cells then marked by a rectangular pulse with an model can simulate the clustering, kinematic, and life evolution of rain cells.
independent rain rate and independent rain duration [Bartlett–Lewis rectan- Kavvas et al. (1987, 1988) also developed a kinematic, time space Poisson
gular pulse model (BLRPM)]. Two types of probability distributions were cluster random field model for ground rainfall in time space. This time-space
selected for rectangular rain pulses in order to differentiate among the strati- nonstationary, kinematic two-level filtered Poisson cluster model was devel-
form rain and the convective rain. Application of BLRPM to Colorado rainfall oped in terms of its characteristic function and mean function of time-space
at various intervals yielded satisfactory results for rain depths at various time rain intensity. At the primary level of this model, cyclone centers are born by
intervals. Also, the statistics of extremes were deemed to be modeled satisfac- a time-space nonstationary Poisson process in order to mimic the preferred
torily. It may be noted that Bartlett–Lewis process is again a Poisson cluster birth locations of extratropical cyclone centers over the United States. Once
process with both the primary level and the secondary level arrival processes born, each of the cyclone centers move as a nonhomogeneous Markov pro-
taken as Poisson processes. The BLRPM was later generalized by Hanaish cess. At the secondary level of the model, rain cores and rain cells are born
et al. (2011) to rectangular pulses where the rain intensity distribution is according to another time-space nonstationary Poisson process with respect
conditional on the random rain cell duration. The Bartlett–Lewis two level to each cyclone center time-space location at preferred time-space locations.
Poisson cluster model was also extended by Cowpertwait et al. (2007) to a Each born rain core/cell assumes a random circular area, random rain inten-
three-level Poisson cluster process of instantaneous rain pulse arrivals, sity, random life length, and random velocity. The ground rain fields are
accounting for both the stratiform and convective rainfall. Yet another type of formed from the time-space superimposition of the circular rain core/cell
Poisson model, the doubly stochastic Poisson process (DSPP) was introduced areas. The simulated rain fields by Saquib et al. (1988) resembled the radar-
by Smith and Karr (1983) to the modeling of daily rainfall with application to observed characteristics of the rainbands. While, the Poisson cluster-type
Potomac river basin. In this model, the occurrence rate becomes random with stochastic precipitation field models advanced the realism in the description
a zero state and a positive state. Given that the process is in the positive state, of precipitation fields in time and space, they suffered from parameter esti-
it takes on values according to a Markov chain. The aforementioned Poisson mation and validation problems (Georgakakos and Kavvas, 1987; Smith and
cluster models have rich structures to account for the correlation structure of Krajewski, 1987). Since at the time of their development routine concurrent
the precipitation occurrences and statistics of the extremes. However, they all rain gauge-radar rainfall data (the aforementioned Stage-IV data) were not
suffer from parameter estimation problems since the rain gauge observations available, the calibration of these continuous time-space models could only
they use at hourly or daily intervals provide only the aggregated products be performed indirectly in terms of fitting certain rainfall moments of these
(precipitation depths) of a fundamentally continuous-time process, as first models to ground rain gauge observations. No rigorous validation of these
observed by Foufoula-Georgiou and Guttorp (1987). Also, they have many models could be performed. However, with the current availability of Stage-
parameters whose estimation are sensitive to the size of the time increments IV multisensor precipitation data (Lin and Mitchell, 2005), it may be possible
being considered. One way to avoid these problems is to utilize concurrent to calibrate and validate Poisson cluster-type precipitation random field
weather radar and rain gauge data for the modeling of sequential precipitation models in time and space.
depths at a selected number of ground locations. In this direction, Kavvas and A paradigm shift in time-space precipitation modeling has happened in
Herd (1985) developed a filtered Poisson cluster process model (FPCPM), mid-1980s by the recognition that the precipitation fields are scaling (Lovejoy,
where the time-varying precipitation depths at any selected time interval can 1982; Lovejoy and Mandelbrot, 1985; Lovejoy and Schertzer, 1985; Waymire,
be obtained from the trajectories that are drawn by the radar-detected rain 1985; Schertzer and Lovejoy, 1987, etc.). Scaling transformations stretch or
fields that pass over the selected ground location. The basic advantage of this contract processes or geometric objects/fields in time and/or space. A process
modeling approach is that its parameters can be estimated directly from the or an object is self-similar if it is invariant under a subset of scaling transfor-
radar-observed rain field arrivals, and the time-varying rainfall intensities can mations. Within the framework of a physical process, self-similarity is the
be constructed by the simultaneous use of the radar reflectivity fields and rain case where the solution of the process at a specified time-space scale can be
gauge observations. The model can also accommodate the nonstationary related to the solution of the process at another time-space scale. Meanwhile,
characteristics of the temporal precipitation process by means of its nonsta- a fractal may be thought as a geometric object/field that is self-similar at
tionary Poisson cluster process structure. Since December 2001 Stage-IV various scales. About three decades ago, Mandelbrot (1983) introduced the
multisensor precipitation analyses at as fine as one hour intervals by U.S. fractal geometry of nature, the degree of irregularity, or fragmentation of
National Weather Service (NWS) (Lin and Mitchell, 2005) became available. which is identical at all geometric scales. Meanwhile, if the studied geometric
Stage-IV multisensor analyses combine rain gauge data with radar-estimated field is self-similar with different scale factors at different scales, then that
rainfall, and are quality controlled by the U.S. NWS’ river forecast centers. field becomes a multifractal, and needs to be modeled accordingly (Schertzer
Stage-IV data are ideally suited to the application of FPCPM for the identifi- and Lovejoy, 1987). In mid-1980s Lovejoy and Mandelbrot (1985) proposed a
cation of the arrival times of the rain fields, and for the observation of the fractal model for stochastic precipitation fields that scale with the same scale
precipitation depths as they evolve in time. factor at all scales. Later on, starting with Schertzer and Lovejoy (1987), the
stochastic precipitation fields were modeled as multifractal cascades by vari-
41.2.2 Stochastic Modeling of Precipitation in Time ous researchers (Gupta and Waymire, 1990, 1993; Hubert et al., 1993; Over
and Space and Gupta, 1994; Burlando and Rosso, 1996; Veneziano et al., 1996; Menabde
Building upon the reported conceptualizations of the cyclonic precipitation et al., 1997; Deidda, 2000; Lovejoy and Schertzer, 2005; Lovejoy and de Lima,
fields by Austin and Houze (1972) and Hobbs and Locatelli (1978), and 2015, etc.). Based on various observations of the precipitation fields in time
exploiting the property of a time-space Poisson process that the number of and space (e.g., Lovejoy and Schertzer, 2005; Lovejoy and de Lima, 2015), the
events in nonoverlapping time-space boxes are mutually independent, sev- fractional-integrated flux model (FIF) which was originally developed in
eral researchers developed various forms of the Poisson cluster random field Schertzer and Lovejoy (1987), emerged as a viable model for the time-space
model for precipitation in time and space (Waymire et al., 1984; Smith and stochastic precipitation fields that does preserve the multifractal characteris-
Karr, 1985; Valdes et al., 1985; Rodriquez-Iturbe et al., 1986; Kavvas et al., tics of precipitation at various scales (Lovejoy and Schertzer, 2005; Lovejoy
1987). Among these studies, the ones that also account for the observed and de Lima, 2015). The fundamental advantage of the multifractal models
kinematic behavior of the rain fields shall be discussed here. The first Poisson over the point process models in modeling stochastic time-space precipita-
cluster random field model for precipitation in time space was by Waymire tion fields is that they can describe these fields with very few parameters. For
et al. (1984). In this time-space filtered three-level Poisson cluster process example, the FIF model can describe the evolution of precipitation fields at
model for ground rainfall in time space, which was developed in terms of the multiple scales with three parameters. Its details are given in Schertzer and
mean and covariance functions of the time-space ground rainfall intensity, Lovejoy (1987).
the rainband arrivals in time, described by a stationary Poisson process, make
up the primary level. At the secondary level, small mesoscale rain areas 41.3 DETERMINISTIC NUMERICAL MODELING OF
(SMSAs or rain cores) are born according to a spatially stationary planar TIME-SPACE PRECIPITATION
Poisson process at the arrival time of each rainband. Once the time-space
birth location of an SMSA is realized, a time × two dimensional-space sta- While significant advances in the stochastic modeling of precipitation fields
tionary stochastic process is taken for the births of rain cells within the disk have been achieved in the last 50 years, there are still some fundamental issues
of the SMSA. Each cell is also assigned a random lifetime. As such, the births to be resolved in modeling precipitation with respect to its short-term (with
of rain cells is modeled as a time-space stationary Neyman–Scott process. A several hours to several days lead time) forecasting in time and space over a
common constant velocity is taken for all cells. Then the rain intensity con- target region, and the estimation of maximum precipitation at the drainage
tributions of all born cells are superimposed in time and space to compose basin of a designed large hydraulics structure toward the estimation of maxi-
the ground rainfall field. Based on their simulation of the time-space rain mum flood at the site of the structure.
41_Singh_ch41_p41-1-41-14.indd 3 8/22/16 2:09 PM

In the Americas and elsewhere around the globe, one of the mesoscale NWP are different, but a regional atmospheric model is frequently used both
atmospheric systems that produce the heaviest precipitation are the MCSs as a RCM as well as in NWP in practice. Unlike currently operational GCMs
(Schumacher and Johnson, 2005). The precipitation in MCSs is organized in and some GWFSs, the mesoscale models solve the nonhydrostatic equations,
distinct convective and stratiform regions with distinct intensity textures, which tend to be more computationally intensive, to capture the details of
locations and dynamics (Parker and Johnson, 2004; Schumacher and Johnson, vertical atmospheric dynamics. While the application of a global-scale model,
2005; Trapp, 2013). Another very important mesoscale atmospheric system such as a GCM or a GWFS, essentially solves an initial value problem, a
that renders extreme precipitation is a TC. The TCs that form hurricanes in regional atmospheric model application deals with an initial and boundary
the United States and typhoons in Asia again have their precipitation fields value problem, as it needs external boundary conditions typically supplied
organized in distinct patterns in intensity and location with distinctly defined from a global model or a larger-scale regional model.
dynamics that play important roles in the time-space distribution of precipita- A regional atmospheric model can be used for atmospheric simulation
tion over a region (Marks Jr., 1985; Powell, 1990a, 1990b; Lonfat et al., 2004; with fine spatial resolution in a limited area. Sometimes simulation at a fine
Lin et al., 2010). In order to be able to estimate the maximum precipitation resolution is required to obtain detailed information on precipitation fields.
over a target watershed due to MCSs or to TCs, or to be able to forecast the For example, if a study region or watershed is in a mountainous area, a fine
precipitation in short term due to these systems over a target region, it is resolution is required to account for the effect of topography (e.g., Kanamitsu
necessary to use deterministic numerical atmospheric models. and Kanamaru, 2007; Maraun et al., 2010; Soares et al., 2012). The application
Time-space precipitation can be described by a numerical atmospheric of a hydrologic model frequently focuses on a geographical region or a water-
model, which simulates atmospheric processes to predict weather or climate shed. Such an application would require less than 10-km spatial grid resolu-
based on initial conditions and/or boundary conditions, depending upon tion (e.g., Boé et al., 2007; Lin et al., 2015). A global model can literally cover
time horizons. A numerical atmospheric model numerically solves the gov- the whole globe and does not require external boundary conditions from
erning equations of the atmosphere, including Navier–Stokes equations and another model simulation, but requires very substantial computational
heat and mass conservation equations for air, water, and various atmospheric resources in order to be able to simulate the atmospheric phenomena at fine
substances. These equations are nonlinear partial differential equations for spatial resolution (such as less than or equal to 10 km). Thus, a regional atmo-
which exact solutions are unobtainable, and even numerical methods often spheric model is useful for simulating precipitation at a fine resolution that is
suffer from instability due to the nonlinearity. Furthermore, to handle subgrid necessary for a regional-scale or a watershed-scale hydrologic analysis and
variability within a computational grid cell, the numerical atmospheric mod- modeling.
els at present require parameterizations for solar radiation, clouds and pre- A regional atmospheric model requires external boundary conditions from
cipitation, heat exchange, soil, vegetation, surface water, and the effects of a global model or a larger-scale regional model within which it is nested. Most
terrain. The complexity level of the model, which determines the computa- regional atmospheric models are built with domain nesting functionality.
tional requirements, depends on the scale of the process typically character- Such regional atmospheric models can gradually downscale the coarse spatial
ized by a grid resolution. resolution atmospheric data, supplied by a global model, into a fine spatial
A general circulation model (GCM) is a class of numerical atmospheric resolution over the target model region through a sequence of nestings of
models for the general circulation of the planetary atmosphere. Nowadays, regional models with gradually decreasing domains and spatial grid resolu-
some GCMs have a component to model the ocean, and solve the general tions. Such a process is called dynamical downscaling. Many meteorological
circulation of the planetary atmosphere and ocean together. When specifying centers and agencies provide atmospheric data simulated by a global model
a model, a GCM that models the atmosphere, the land, and the ocean inter- that can be used by a regional atmospheric model. Such simulated global
actively is called an atmospheric-oceanic GCM (AOGCM), whereas a GCM atmospheric data can be divided into three categories; reanalysis data, fore-
that models only the atmosphere and land interactively is called an atmo- casting data, and future climate change projections. Reanalysis data are atmo-
spheric GCM (AGCM), while a GCM that models only the ocean is called an spheric simulations by a global atmospheric model that are fitted to
oceanic GCM (OGCM). A GCM typically solves a set of nonlinear equations observation data by a DAS, such as the three-dimensional variational data
called primitive equations that consist of momentum equations, thermal assimilation system (3D-VAR), the four-dimensional variational data assimi-
energy equations, and various mass conservation equations under the hydro- lation system (4D-VAR), the ensemble Kalman filter (EnKF; Evensen, 1994)
static approximation. GCMs are used to integrate the dynamics of the etc. Therefore, reanalysis data, after being dynamically downscaled by a
physical components of the earth’s climate system including atmosphere, regional atmospheric model over the target study region, are frequently used
ocean, land, and sea ice. Since they cover the whole earth as their modeling as inputs to hydrologic models in order to reconstruct historical hydrological
domain, the GCMs are typically run under initial conditions for long-term conditions over the target region. Table 42.1 summarizes the currently avail-
climate projections in the order of many years, or for medium term forecasts able reanalysis data.
typically with seasonal or 6-month horizons, or for short-term weather fore- With respect to short-term forecasting, a GWFS provides global weather
casts with a typical horizon of several days to 2 weeks. As such, in mathemat- forecasts at coarse grid resolution (~0.5o) over a target study region with
ical terms they solve initial value problems. horizon from several days to several weeks ahead in real time. Such real-time
For short-term forecasting of weather at global scale, a class of numerical weather forecast data are provided every 6 h or every 12 h. These global fore-
atmospheric models, called the global weather forecasting system (GWFS), casts at coarse resolution are then dynamically downscaled by a regional
are utilized. The main purpose of a GWFS is the real-time weather forecast- atmospheric model at fine spatial resolution (< 6 km) and at hourly intervals
ing. Since the results of real-time weather forecasting strongly depend on the as input to a hydrologic model for its hydrologic forecasts in real-time over
initial state, GWFS typically uses a data assimilation system (DAS) to fit the the target region.
initial state to observation data, such as surface observations, balloon data, Nowadays, there is another type of forecasting system to predict climate
wind profiler data, aircraft reports, buoy observations, radar observations, instead of weather. It is called seasonal forecasting system, and its forecasting
and satellite observations. Over the last several decades, the effectiveness of period is more than several months. Such global coarse resolution forecasts
the GWFSs has been significantly improved even in terms of rainfall (e.g., with seasonal to 6-month time horizons, when downscaled by a regional
Hamill et al., 2013). There are several major meteorological centers opera- atmospheric model at fine grid resolution (<6 km) and at hourly intervals
tionally running GWFSs, including the environmental modeling center of the over a target geographical region, provide the necessary atmospheric inputs
U.S. National Weather Service, a member of the U.S. National Centers for to a hydrologic model for its seasonal hydrologic forecasts over the target
Environmental Prediction (NCEP), the European Center for Medium Range region.
Weather Forecasting (ECMWF), the United Kingdom Met Office, the Finally, future long-term climate projections with time horizons of a cen-
Canadian Meteorological Center, the Japanese Meteorological Agency (JMA), tury or longer are simulated by a GCM based on a future climate scenario
the U.S. Navy, the Commonwealth Scientific and Industrial Research based on the future composition of greenhouse gases in the Earth’s atmo-
Organization, and few others. The GWFS modeling has become increasingly sphere. Unlike forecasts, the aim of a future climate projection is not to obtain
essential for daily weather forecasting, aviation traffic controls as well as the best estimate of the future climate, but to investigate the response of the
hydrological analyses (e.g., Lynch, 2008). earth’s climate system to a possible future climate scenario [Intergovernmental
A mesoscale or a regional atmospheric model is a limited-area model with Panel On Climate Change (IPCC), 1996].
fundamentally the same representations of atmospheric processes as those in Three important practical applications of numerical modeling of time-
the GCMs or the GWFSs. The regional climate model (RCM) and the numer- space precipitation with the aforementioned dynamical downscaling tech-
ical weather prediction (NWP) model belong to this class of atmospheric nique are introduced in the subsections as follows; for water resources
models. A RCM is used for long-term climate assessment, and a NWP model planning, for the design of hydraulic structures, and for the real-time man-
is used for short-term weather simulation. Thus, the purposes of an RCM and agement of water resources systems.
41_Singh_ch41_p41-1-41-14.indd 4 8/22/16 2:09 PM

41_Singh_ch41_p41-1-41-14.indd 5
Table 41.1 List of Meteorological Reanalysis Data
Assimilation Spatial Spatial Temporal
Data name Full database name Source Model resolution Precipitation Period References
method coverage resolution resolution
NNRP NCEP/NCAR reanalysis NCEP/NCAR T62L28 3D-VAR No Global 2.5° 1948–present 6 hourly Kalnay et al., 1996
Reanalysis-2 NCEP/DOE reanalysis AMIP-II (R2) NCEP/DOE T62L28 3D-VAR No Global 2.5° 1979–present 6 hourly Kanamitsu et al., 2002
CFSR Climate forecast system reanalysis NCAR T382 L64 3D-VAR Yes Global 0.5° 1979–present 6 hourly Saha et al., 2010
MERRA Modern-era retrospective analysis for NASA 0.67 × 0.5 deg 3D-VAR Yes Global 0.67° × 0.5° 1979–present 3 hourly Bosilovich et al., 2008
research and applications
20CR NOAA-CIRES twentieth century reanalysis NOAA/ESRL PSD T62 L28 EnKF Yes Global 2.0° 1871–2012 6 hourly Compo et al., 2011
ERA-40 ERA 40-year Reanalysis ECMWF T159L60 3D-VAR Yes Global 1.125° 1958–2001 3 hourly Uppala et al., 2005
ERA-Interim ERA interim reanalysis ECMWF T255L60 4D-VAR Yes Global 0.7° 1979–present 3 hourly Berrisford et al., 2009
JRA-25 Japanese 25-year reanalysis JMA/CRIEPI T106L40 3D-VAR Yes Global 1.25° 1979–2004 6 hourly ONOGI et al., 2007
JRA-55 Japanese 55-year reanalysis JMA TL319L60 4D-VAR Yes Global 1.25° 1958–2012 6 hourly Ebita et al., 2011
ASR Arctic system reanalysis BPRC 10 and 30 km 3D-VAR Yes Arctic region 10 km 2000–2010 3 hourly Bromwich et al., 2010
NARR North America regional reanalysis NCAR 32 km RDAS Yes North America 0.3° 1979–present 3 hourly Mesinger et al., 2006
41-5
8/22/16 2:09 PM
41.3.1 Numerical Modeling of Time-Space and the estimates strongly depend on the observation data. Flood–fre-
Precipitation for Water Resources Planning quency analysis may allow a severe flood realization to exceed the esti-
A water-resource-planning process requires information on the flow regime mates. In fact, several storms caused more flood flow discharge in several
in the target geographical region. However, the historical flow data may be areas in the United States than 100- or 200-year return period floods
insufficient in the target region because the observation period may be too obtained from the historical observation data before the storm event (e.g.,
short, or some of the observation data may be missing. Furthermore, since Engman et al., 1974; Baker et al., 1979). On the other hand, the most
there is evidence of global climate change (IPCC, 2013), the statistical equi- widely accepted definition of the PMP is the meteorologically possible
librium of the historical hydrologic conditions including precipitation over a upper limit of precipitation over a given size of storm area at a particular
target region may be quite different than the one in the future. Since the sta- location at a certain time of the year without long-term climatic trends
tistical equilibrium of the hydroclimate system will be changing within the (World Meteorological Organization, 1986). This approach tries to con-
horizon of the next century, a planning process for the water resources of a sider the physics of precipitation processes, but the physics are very simpli-
target region for the next century can no longer be based on the hydrologic fied, even linearized. The standard PMP (as defined by WMO, 1986)
conditions of the past, even in the statistical sense. Consequently, the plan- estimates strongly depend on available observation data. The PMP esti-
ning of a water resources system for the next century over a target region will mates that were obtained within this framework have been exceeded by
require information on the future flow conditions that are not available from recent storms (e.g., Dooge, 1986; Koutsoyiannis, 1999). In addition, both
historical observations. Meanwhile, various meteorological centers and agen- flood frequency analysis and the PMP approaches are based on the assump-
cies around the world provide global atmospheric data for the next century tion of stationarity of observation data, or on the statistical equilibrium of
from the outputs of GCM simulations, and historical atmospheric data from the hydroclimate system of the target region. In other words, these
reanalysis data. By downscaling these coarse-resolution data by a numerical approaches cannot consider nonstationarity under changing climate. In this
atmospheric model over the target region, which is called “dynamical down- context, new methodologies are urgently needed to approach these issues
scaling,” one can simulate the historical and future time-space precipitation at for the realistic design of hydraulic structures. Accordingly, numerical
fine spatial and temporal resolutions over the target region. Since the physical modeling of time-space precipitation by regional atmospheric models has
conditions of the global hydroclimate system are changing in the next century recently been performed as input for the design of hydraulic structures
(IPCC, 2013), the resulting nonstationarity in the hydroclimate conditions (Ishida et al., 2015a, 2015b; Ohara et al., 2011).
over the target region can best be quantified by the dynamical downscaling A numerical modeling approach to obtain time-space precipitation for the
technique with a RCM, since the numerical modeling of the coupled atmo- design of hydraulic structures can be conducted by means of a physically
spheric-land hydrologic conditions by an RCM can account for the changing based regional atmospheric model, such as RCM and NWP. A new methodol-
physical conditions at the target region during the next century. ogy based on this numerical modeling approach was recently introduced by
Numerical modeling of regional-scale or watershed-scale time-space pre- Ohara et al. (2011) and demonstrated by Ishida et al. (2015a, 2015b) in order
cipitation for a historical period requires a RCM and a reanalysis dataset. to estimate the upper bound of time-space precipitation over a target water-
Most of the reanalysis datasets in Table 42.1 cover the globe, but they have a shed or region by a regional atmospheric model. This methodology will be
6-hourly temporal resolution and more than 0.5° horizontal resolution, referred to as physically based PMP (PhysPMP) approach or more technically
which is too coarse for inputs to a regional-scale or watershed-scale hydro- maximum precipitation estimation (MPE) approach, and the estimate of the
logical model. Hence, it is still necessary to downscale the reanalysis data. precipitation upper bound will be referred to as the MP estimate, in order to
The selection of a reanalysis dataset for downscaling should depend on distinguish it from the traditional PMP approach and its estimate. MPE
purpose. The temporal coverage of the reanalysis datasets ranges approxi- approach first obtains initial and boundary conditions of the regional atmo-
mately from 35 to 140 years. A reanalysis dataset covering a longer period spheric model from global atmospheric data such as historical reanalysis data
uses less observation data for its DAS. For example, most of the reanalysis or GCM future climate projection data. After this step, initial and boundary
datasets covering a period from around 1980, which spans the last 35 years conditions are artificially adjusted in order to maximize precipitation over the
and which may be appropriate for short-term modeling purposes, use remote target area. Note that the interior of the domains of the regional atmospheric
sensing data for the data assimilation. More detailed information of reanalysis model except the initial conditions should not be artificially changed in order
data can be obtained in the references in Table 42.1. Hence, they are more to conserve mass, energy, and momentum within each of the modeling
accurate than a reanalysis dataset that may span a longer period back into the domains so that atmospheric processes in the domains will be simulated
past, but may be constructed mainly from the synthesis of various global based on physics by the regional atmospheric model. Since a physically based
observations by a GCM during the period. Once a reanalysis dataset is regional atmospheric model can simulate 3D wind fields and moisture flow
selected, it can be dynamically downscaled by a RCM. Note that most RCMs along slopes of mountains, the simulated atmospheric processes with this
have many schemes for their physics options such as cloud microphysics, framework can account for all the physical features including the orographic
cumulus parameterization, radiation physics, planetary boundary layer phys- effect and the blocking effect by mountains, the effects that cannot be prop-
ics, and so on. As such, before its use the RCM needs to be configured for the erly handled by the traditional PMP approach. Since MPE approach is based
best combination of its physical components that yields the best fit to on physics, it does not require the assumption of stationarity, and can be
the historical observations. Then it is necessary to validate the RCM over the applied under changing climate conditions. In addition, MPE approach does
target area before its use for climate simulations. Once validated by historical not require observation data except for calibration and validation processes of
data, the RCM can then reconstruct precipitation without any observation the model. Therefore, the estimates of the upper bounds of the precipitation
data. do not depend on any observation data. MPE approach can keep the reliabil-
Future time-space precipitation at fine temporal and spatial resolution ity of the estimates even when applied to a sparsely gauged or ungauged area.
under a changing climate can be simulated by downscaling a future climate Another noteworthy feature of MPE approach is that the regional atmo-
projection from a GCM by means of a RCM. There are many sets of future spheric model simulation can provide all the atmospheric variable values
global climate projection data simulated by GCMs based on IPCC future cli- required by a watershed-scale or regional-scale hydrological model, not only
mate change scenarios from special report on emissions scenarios (IPCC., precipitation and temperature, but also wind fields, radiation, humidity, and
2000) or from representative concentration pathways (RCPs; Moss et al., so on.
2010). Future global climate projection data based on the RCP scenarios can MPE approach can be applied under changing climate conditions as men-
be found on the Earth System Grid Federation website (http://pcmdi9.llnl tioned earlier. To obtain MPE estimates with the consideration of climate
.gov/esgf-web-fe/). It may be noted that dynamical downscaling by a RCM change, a procedure for the MPE approach was recommended by Ohara et al.
requires at least 6-hourly temporal resolution to reflect diurnal fluctuations. (2011), and is modified here as follows:
Therefore, the future coarse-resolution climate projection data from a GCM 1. Validate a regional atmospheric model over a specified area (watershed
must be with less than or equal to 6-hourly time intervals. or region)
2. Extract extreme storm events over the study area from historical pre-
41.3.2 Numerical Modeling of Time-Space cipitation observations or reconstructed data
Precipitation for the Design of Hydraulic Structures
3. Reconstruct the extracted historical storm events over the study area by
Designing hydraulic structures, such as dams and levees, are important top- means of a regional atmospheric model that will use a historical reanalysis
ics of hydrologic engineering. Traditionally, flood–frequency analysis or dataset for its synoptic initial and boundary atmospheric conditions if obser-
the probable maximum precipitation (PMP) approaches are used for the vation data are used in the previous step
design of hydraulic structures. Flood–frequency analysis is a technique to 4. Maximize the extracted historical storms over the study watershed by
fit a frequency curve to a series of annual maximum peaks of flow dis- adjusting the corresponding synoptic initial and boundary conditions
charge. However, flood frequency analysis does not refer to any physics, obtained from the historical reanalysis dataset
41_Singh_ch41_p41-1-41-14.indd 6 8/22/16 2:09 PM

Remote Sensing for the Modeling of Time-Space Precipitation 41-7
5. Construct precipitation over the study area under future climate condi- The RFS consists of preprocessing programs, a DAS, and a NWP model.
tions by means of the selected regional atmospheric model that will use all the Figure 42.1 shows a sample of the forecasting cycle. Initial conditions of the
available GCM future climate projections data for its synoptic initial and NWP are assimilated to the observation data at the beginning of each fore-
boundary conditions casting cycle. Then the NWP is run for forecasting. The simulated data of the
6. Extract future extreme storm events over the study area from the con- previous forecasting cycle are used as the initial conditions of the current
structed precipitation data under future climate conditions forecasting cycle. This forecasting cycle is continuously repeated in the RFS.
7. Maximize the extracted future storms over the study area by adjusting Figure 42.2 shows the details of a RFS. The RFS requires three types of inputs:
the corresponding synoptic atmospheric initial and boundary conditions GWFS global forecasting data, geographical information system (GIS) data,
obtained from the GCM climate projection datasets and sounding and surface observation data. GWFS global forecasting data are
8. Develop a MPE estimate from the analysis of all the maximized histori- required to provide the boundary conditions for the NWP. GIS data include
cal and future storm events over the study area surface terrain, land use/ land cover, and so on. Weather sounding and surface
In Steps 2–4, the MPE value is estimated based on a historical reanalysis observation data are utilized for the DAS after a quality control. Currently,
dataset. Then, in Steps 5–7, future projected climate data are used to take remote sensing data may also be available for use by the DAS. By using those
changing climate conditions into consideration for MPE estimation. In the inputs, the RFS first creates boundary conditions from the global forecasting
final step, the maximized results for all the extracted storms in the historical data. The lower boundary conditions are updated based on the initial condi-
and future periods are analyzed altogether to reach an estimate of the upper tions that are obtained from the previous forecasting cycle. After this step, the
bound of precipitation over the study area. initial conditions are assimilated to the observation data. The lateral bound-
MPE approach requires methods to actually maximize the contribution of ary conditions are updated based on the assimilated initial conditions. With
each storm to precipitation over the study area. These methods should be the assimilated initial conditions and the updated boundary conditions, the
specified or customized to the study area after analyzing the atmospheric forecasting is conducted by the NWP model. Then the forecasting results are
phenomena of each of the extracted storms for the study area. However, a used to create the initial conditions for the next forecasting cycle.
combination of two methods proposed by Ishida et al. (2015a, 2015b) could Some meteorological centers around the world provide global forecasting
form a basis for methods that need to be specified for a selected study area. outputs that can be used as inputs to a RFS. Among weather forecasting sys-
First, Ishida et al. (2015a) shifted the atmospheric boundary conditions of a tems whose forecasting period is from several days to several weeks, the
regional atmospheric model in space with latitude and longitude so that the global forecasting system (GFS) may be the most popular global forecasting
high moisture flow of a storm strikes the study area in an optimal direction system because it provides outputs online in real time for the real-time opera-
and location to maximize the precipitation over the study area. Second, Ishida tion anywhere around the world without charge. The real-time GFS outputs
et al. (2015b) maximized the moisture of the storm by setting the relative can be obtained at a website of U.S. NCEP (http://nomads.ncep.noaa.gov/). The
humidity at the boundaries of the modeling domain at 100% after shifting the ensemble version of NCEP forecasting system, called global ensemble fore-
atmospheric boundary conditions. The first step above is called the atmo- casting system, is also available online in real time. Moreover, some other
spheric boundary condition shifting (ABCS) method, and the second the meteorological centers such as ECMWF (http://www.ecmwf.int/) and JMA
relative humidity maximization (RHM) method. Generally, the RHM method (http://www.jmbsc.or.jp/) provide global weather forecasting outputs that
should be applied after the ABCS method because the moisture of the storm may be able to be used for inputs to a RFS. Meanwhile, U.S. NCEP also oper-
should be maximized. Otherwise, the study area can end up receiving unreal- ates a seasonal forecasting system, called climate forecast system (CFS; Saha
istically heavy precipitation even during a dry period by the RHM method. et al., 2010), to predict climate with more than a half year horizon. The out-
puts of the CFS can also be inputs to an RFS although the RFS may need to
use a RCM instead of a NWP. The real-time products by the CFS can be
41.3.3 Numerical Modeling of Time-Space
obtained at its website (http://cfs.ncep.noaa.gov/). The CFS climate forecast
Precipitation for the Real-Time Management of Water
Resources Systems data can then be downscaled by a RCM over the target region in order to be
utilized as input to the hydrologic model of the region that can then forecast
In this section, the last topic on the use of numerical time-space precipitation the seasonal flows to the water resources system of the region in order to
modeling for water resources engineering practice is the real-time manage- manage the system at seasonal scale.
ment of water resources systems. Although the real-time management of
water resources systems requires inflow data to such systems, observation
data are not available for the future. Future flow conditions need to be pre- 41.4 REMOTE SENSING FOR THE MODELING OF
dicted in order to develop the real-time management of the operation of water TIME-SPACE PRECIPITATION
resources systems during the twenty first century. Meanwhile, for short-term Measurement of precipitation is necessary for water resources management,
operation management horizons (daily to weekly in terms of hourly intervals) flood/drought forecasting, climate monitoring, and hydroclimate modeling.
global numerical weather forecast data are now provided by some meteoro- Although rain gauges provide a direct measurement of precipitation, precipi-
logical centers by means of a GWFS. In this context, a numerical forecasting tation estimates are adversely affected by the gauge undercatch bias of point
approach can be used by downscaling the coarse-grid weather forecasts of a measurements (UNESCO, 1978; Groisman and Legates, 1994). When areal
GWFS by a numerical atmospheric model over the target region as input to averages are estimated from these point measurements, additional biases may
the hydrologic model for the region in order to predict the flow conditions for be introduced at mountainous regions since rain gauges are usually located at
the real-time management of the water resources system of the region. lower elevations. Remotely sensed precipitation estimates provide useful
Accordingly, in the following we introduce a numerical forecasting approach information at basins with limited rain gauges or at un-gauged basins. After
by a regional-scale forecasting system (RFS) to predict precipitation fields as the installation of U.S. National Weather Service Next Generation Weather
input to the hydrologic model of the target region. Radar (NEXRAD) network (Klazura and Imy, 1993), weather radars started
0 UTC 6 UTC 12 UTC 18 UTC
Forecasting
Assimilation ICs Forecasting
Assimilation ICs
Forecasting
Assimilation
Figure 41.1 Forecasting cycle.
41_Singh_ch41_p41-1-41-14.indd 7 8/22/16 2:09 PM

GWFS global GIS data Sounding and

forecasting data surface
observations
Spatial interpolation GIS data processing
Surface and 3D pressure-level

data on model grids
Quality-controlled
sounding and Quality control
Data assimilation
surface data
Assimilated initial
conditions
Lower BCs
update
Initial Lateral BCs
conditions update
Lower and lateral BCs NWP
Create BCs Forecasting

outputs
Figure 41.2 Regional-scale forecasting system.
to play an important role in meteorological and hydrological forecasting. infrared and visible instruments allow empirical precipitation estimations
Compared to rain gauges, radars provide higher spatial, and temporal resolution (e.g., Hydro-Estimator of US National Weather Service and GOES
with larger areal coverage. However, radars suffer from the transformation of Precipitation Index) using the cloud top observations, and are utilized as a
reflectivity to rainfall rate, spatial, and temporal representation of the true part of both Geostationary Orbit satellites (e.g., Spinning Enhanced Visible
rainfall field (Pessoa et al., 1993; Winchell et al., 1998), range-dependent bias, and Infrared Imager of EUMETSAT’s METEOSAT Second Generation) and
bright band, anomalous propagation, incomplete beam filling, and overshoot- Low Earth Orbit satellites (e.g., Moderate-Resolution Imaging
ing of precipitation (Smith et al., 1996). In spite of the uncertainties involved, Spectroradiometer on Aqua and Terra satellites, and Visible and InfraRed
the radar-based precipitation estimates have been utilized in flood and storm Scanner on TRMM satellites). PMW instruments are utilized mostly as a part
forecasting applications (Kouwen and Garland, 1989; Schell et al., 1992; of low earth orbit satellites. Electrically scanning microwave radiometer,
Dixon and Wiener, 1993; James et al., 1993; Becchi et al., 1994; Mimikou and advanced microwave scanning radiometer–Earth observing system (Kawanishi
Baltas, 1996; Bedient et al., 2003; Knebl et al., 2005). et al., 2003), special sensor microwave imager (Barrett et al., 1988), special
Due to limited spatial coverage of the ground gauges and weather radars, sensor microwave imager/sounder (Kunkee et al., 2008), advanced microwave
satellite-based precipitation estimates are also necessary for regional or global sounding unit-B, TRMM microwave imager, and global precipitation mea-
scale precipitation estimates. As a major step forward in remotely sensed surement microwave imager are examples of PMW instruments of low Earth
observations, the first Earth-orbiting satellites, SPUTNIK-1 and SPUTNIK-2, orbit satellites. TRMM precipitation radar (Iguchi et al., 2000) and GMP
were launched by the USSR in 1957, followed by the launch of EXPLORER-1 (global precipitation measurement) and dual-frequency precipitation radar
by the United States. By the initiation of TIROS-1 Program (Television (Nakamura et al., 2005) are the only satellite precipitation radars. Summary
Infrared Observation Satellite Program; Fritz and Wexler, 1960), NASA start- information (spatial resolution, spatial coverage, temporal resolution, and
ed collecting meteorological observations of the Earth from space in 1960. temporal coverage) on the satellite-based precipitation estimates by several
Meteorological satellites can be divided into geostationary satellites and providers are listed in Table 41.2.
low Earth orbiting satellites, which include polar-orbiting satellites. The challenges in the satellite-based precipitation observations to be dealt
Geostationary satellites, which orbit the Earth at about 35,800 km above the with in the future research efforts can be listed as (Sorooshian et al., 2011):
ground, remain stationary with respect to the rotating Earth. Examples of (1) quantification of uncertainties of individual satellite sensors and their
geostationary satellites include the EUMETSAT’s Meteosat Second Generation propagation into multisensor products; (2) optimization of precipitation
satellites (Schmetz et al., 2002), the United States’ Geostationary Operational algorithms for retrieval, downscaling, and data fusion, considering the error
Environmental Satellite system (GOES satellites, Menzel and Purdom, 1994), properties in different climate regions, storm regimes, surface conditions,
China’s Feng-Yun-2 satellites (http://www.fas.org/spp/guide/china/earth/fy-2. seasons, and altitudes; (3) development and validation of precipitation esti-
htm#ref625), Japan’s Multifunctional Transport Satellite system (Yoshiro, mates in fine spatial and temporal resolutions; (4) development of reliable
2002), and Russia’s Elektro-L 1 (Asmus et al., 2012). Polar-orbiting satellites bias removal techniques, particularly the probability distribution function-
orbit the Earth at about 850 km above the ground. Examples of polar-orbiting based adjustment techniques; (5) near-real-time bias adjustment and the
satellites include the National Oceanic and Atmospheric Administration integration of heterogeneous airborne and ground-based precipitation esti-
(NOAA) satellites (Goodrum et al., 2000), EUMETSAT’s MetOp series (Klaes mates; (6) development of a reliable and near-real-time probabilistic uncer-
et al., 2007), China’s FY-1D (Liu et al., 2004), and FY-3A (Dong et al., 2009), tainty model for satellite-based precipitation estimates; (7) improving the
and Russia’s METEOR-3M (http://www.russianspaceweb.com/). timeliness of receiving satellite precipitation products; and (8) development of
There are three main groups of satellite instruments utilized for precipita- near-real-time precipitation ensemble generation to be used in ensemble
tion estimates: (1) infrared and visible instruments, (2) passive microwave streamflow modeling based on remotely sensed data, particularly when
(PMW) instruments, and (3) active microwave instruments or radars. The uncertainty in streamflow needs to be quantified.
41_Singh_ch41_p41-1-41-14.indd 8 8/22/16 2:09 PM

41_Singh_ch41_p41-1-41-14.indd 9
Table 41.2 Summary Information on the Satellite-Based Precipitation Estimates by Several Providers
Temporal
Short name Full name Provider Spatial resolution Spatial coverage resolution Temporal coverage References
TMPA-RT Real Time TRMM Multi-Satellite National Aeronautics and Space 0.25° 50° S – 50° N 3h January 2000–present Huffman et al. (2007)
(3B42RT) Precipitation Analysis Administration (NASA) Goddard Space
Flight Center (GSFC)
TMPA (3B42) TRMM Multi-Satellite National Aeronautics and Space 0.25° 50° S – 50° N 3h March 2000–present Huffman et al. (2007)
Precipitation Analysis Administration (NASA) Goddard Space
Flight Center (GSFC)
NRLgeo Naval Research Laboratory (NRL) Naval Research Laboratory (NRL) 0.10° Global 3h September 2004–present Turk and Miller (2005)
Global Blended-Statistical
Precipitation Analysis
PERSIANN Precipitation Estimation from University of California, Irvine 0.25° 60° S – 60° N 3h March 2000–present Sorooshian et al. (2000)
Remotely Sensed Information
using Artificial Neural Networks
PERSIANN- PERSIANN-Cloud Classification University of California, Irvine 0.04° Africa and India 3h January 2006–present Hong et al., 2004
CCS System
CMORPH2 Climate Prediction Center NOAA/CPC 0.07277° 60° S – 60° N 30 min December 2002–present Joyce et al. (2004)
Morphing Method It was recently repro-
cessed back to 1998 by
Kidd and Huffman (2011)
PMIR The Combined Passive University of Birmingham 0.10° 60° S – 60° N 30 min January 2002–present Kidd et al. (2003); Kidd
Microwave-Infrared algorithm and Muller (2009)
HYDRO Hydro-Estimator NOAA/National Environmental Satellite, 4 km 60° S – 60° N 15 min Online archive beginning 22 Scofield and Kuligowski
Data, and Information Service (NESDIS) August 2006; CONUS archive (2003)
beginning 1 April, 2003
GPI GOES Precipitation Index NOAA/CPC GSFC 1° 40° S – 40° N 1 day Since 1 January 1997 by 1 day; Arkin and Meisner (1987);
January 1986 to February 2004 Joyce and Arkin (1997)
by 5 days and 1 month
CST Convective Stratiform Technique NASA GSFC ECMWF 4 km 60° N – 60° S 30 min November 1, 2004–present Negri et al. (2002)
MPE Multi sensor Precipitation Estimate European Organisation for the 4 km 60° S – 60° N 15–30 min August 2006–present Heinemann et al. (2002)
Exploitation of Meteorological
Satellites (EUMETSAT)
2
CMORPH was recently reprocessed back to 1998 by Kidd and Huffman (2011).
41-9
8/22/16 2:09 PM
Gauge and satellite observations are the existing sources of large-scale for Community Involvement,” Eos, Transactions American Geophysical Union,
precipitation estimates, and have advantages and weaknesses (Barret and 91 (2): 13, 2010.
Martin, 1981; and Arkin and Ardanuy, 1989). Xie and Arkin (1996) devel- Buishand, T. A., “Some remarks on the use of daily rainfall models,” Journal
oped an algorithm to construct global gridded fields of monthly precipitation of Hydrology, 36: 295–308, 1978.
by merging estimates of five different sources which have different character- Burlando, P., and R. Rosso, “Scaling and multiscaling models of depth-
istics: gauge-based monthly analysis from Global Precipitation Climatology duration frequency curves for storm precipitation,” Journal of Hydrology, 187:
Centre, three types of satellite estimates (the microwave scattering based 45–64, 1996.
Grody, microwave emission-based Chang estimates, and infrared-based Chang, T. J., M. L. Kavvas, and J. W. Delleur, “Daily precipitation modeling
GOES Precipitation Index), and predictions of the operational forecast model by discrete autoregressive moving average processes,” Water Resources
of European Centre for Medium-Range Weather Forecasts. Ebert et al. (2007) Research, 20 (5): 565–580, 1984.
evaluated the performance of near real-time precipitation estimates from Chin, E. H., “Modeling daily precipitation occurrence process with Markov
satellite observations and numerical models and found that the performance chain,” Water Resources Research, 13 (6): 949–956, 1977.
of both approaches are highly dependent on the rainfall regime. Moreover, Compo, G. P., J. S. Whitaker, P. D. Sardeshmukh, N. Matsui, R. J. Allan, X.
Ebert et al. (2007) concluded that the satellite estimates of rainfall occurrence Yin, B. E. Gleason, R. S. Vose, G. Rutledge, P. Bessemoulin, S. BroNnimann,
and amount are most accurate during summer and at lower latitudes, while M. Brunet, R. I. Crouthamel, A. N. Grant, P. Y. Groisman, P. D. Jones, M. C.
the numerical models show better skill during winter and at higher latitudes. Kruk, A. C. Kruger, G. J. Marshall, M. Maugeri, H. Y. Mok, O. Nordli, T. F.
Despite global coverage and continuous availability, satellite-based precipi- Ross, R. M. Trigo, X. L. Wang, S. D. Woodruff, S. J. Worley, “The twentieth
tation estimates are not usually integrated into operational hydrologic model- century reanalysis project,” Quarterly Journal of the Royal Meteorological
ing mainly due to lack of information on their reliability at basin scale Society, 137 (654):1–28, 2011.
(Behrangi et al., 2011). Behrangi et al. (2011) evaluated 6 years (2003–2008) of Cowpertwait, P., V. Isham, and C. Onof, “Point process models of rainfall:
five satellite-based precipitation products (TMPA-RT, TMPA-V6, CMORPH, developments for fine-scale structure,” Research Report No. 277, Department
PERSIANN, and PERSIANN-adj) with respect to multisensor (NEXRAD and of Statistical Science, University College London, London, 2007.
gauge) datasets, and utilized Sacramento soil moisture accounting model for Deidda, R., “Rainfall downscaling in a space-time multifractal framework,”
streamflow simulation over the Illinois River Basin. Streamflow simulations Water Resources Research, 36:1779–1794, 2000.
were then compared with the streamflow gauge data. They found that the Dixon, M. and G. Wiener, “TITAN: Thunderstorm Identification, Tracking,
bias-adjusted satellite precipitation products agree well with the gauge- Analysis, and Nowcasting—a radar-based methodology,” Journal of
adjusted radar, compared to their counterparts with no bias-adjustment Atmospheric and Oceanic Technology, 10: 785–797, 1993.
(TMPA-RT, CMORPH, and PERSIANN). Behrangi et al. (2011) concluded Dong C., et al., “An overview of a new Chinese weather satellite FY-3A,”
that the bias-adjustment of satellite precipitation data is very important to Bulletin of the American Meteorological Society, 90 (10): 1531–1544, 2009.
improve the precipitation estimates in streamflow simulations. However, the Dooge, J. C. I., “Looking for hydrologic laws,” Water Resources Research,
bias-adjusted precipitation products could still result in poor estimation of 22 (9S): 46S–58S, 1986 (retrieved March 7, 2013).
the precipitation extent, precipitation intensity, and the resulting streamflow Eagleson, P. S., “Climate, soil and vegetation. 2. The distribution of annual
for extreme events. precipitation derived from observed storm sequences,” Water Resources
Research, 14 (5): 713–721, 1978.
Ebert, E. E., J. E. Janowiak, and C. Kidd. “Comparison of near-real-time
REFERENCES
precipitation estimates from satellite observations and numerical models,”
Arkin, P. A. and E. Ardanuy, “Estimating climatic-scale precipitation from Bull. Amer. Meteor. Soc., 88: 47–64, 2007.
space: a review,” Journal of Climate, 2: 1229–1238, 1989. Ebita, A., Shinya Kobayashi, Yukinari Ota, Masami Moriya, Ryoji Kumabe,
Arkin P. A. and B. N. Meisner, “The relationship between large-scale con- Kazutoshi Onogi, Yayoi Harada, Soichiro Yasui, Kengo Miyaoka, Kiyotoshi
vective rainfall and cold cloud over the western hemisphere during 1982– Takahashi, Hirotaka Kamahori, Chiaki Kobayashi, Hirokazu Endo, Motomu
1984,” Monthly Weather Review, 115: 51–74, 1987. Soma, Yoshinori Oikawa, Takahisa Ishimizu, “The Japanese 55-year reanaly-
Asmus V. V., V. N. Dyadyuchenko, and V. A. Zagrebaev, et al., “Development sis ‘JRA-55’: an interim report,” Scientific Online Letters on the Atmosphere, 7:
of the space complex of hydrometeorological support on the basis of 149–52, 2011.
Elektro_L series geostationary satellites,” Vestn. FGUP NPO im. S. A. Engman, E. T., L. H. Parmele, and W. J. Gburek, “Hydrologic impact of
Lavochkina, 2012 (no. 1): 3–14. tropical storm agnes,” Journal of Hydrology, 22 (1–2): 179–93, 1974.
Austin, P. M. and R. A. Houze, Jr., “Analysis of the structure of precipita- Evensen, G., “Sequential data assimilation with a nonlinear quasi-geostrophic
tion patterns in New England,” Journal of Applied Meteorology, 11: 926–935, model using Monte Carlo methods to forecast error statistics,” Journal of
1972. Geophysical Research, 99 (C5): 10143, 1994.
Baker, V. R., R. C. Kochel, and P. C. Patton, “Long-term flood frequency Feyerherm, A. M. and L. D. Bark, “Goodness of fit of a Markov chain model for
analysis using geological data,” International Association of Hydrological sequences of wet and dry days,” Journal of Applied Meteorology and Climatology, 6:
Sciences, 128: 3–9, 1979. 770–773, 1967.
Barrett, E. C. and D. W. Martin, The Use of Satellite Data in Rainfall Foufoula-Georgiou, E. and P. Guttorp, “Assessment of a class of Neyman-
Monitoring, Academic Press , London, 1981, p. 340. Scott models for temporal rainfall,” Journal of Geophysical Research, 92 (D8):
Barrett E. C., C. Kidd, and J. O. Bailey, “The special sensor microwave/ 9679–9682, 1987.
imager: a new instrument with rainfall monitoring potential,” International Foufoula-Georgiou, E. and D. P. Lettenmaler, “Continuous-time versus
Journal of Remote Sensing, 9: 1943–1950, 1988. discrete-time point process models for rainfall occurrence series,” Water
Becchi, I., E. Caporali, and E. Palmisano, “Hydrologic response to radar Resources Research, 22 (4): 531–542, 1986.
rainfall maps through a distributed model,” Natural Hazards, 9: 95–108, 1994. Fritz S. and H. Wexler, “Cloud pictures from satellite TIROS I,” Monthly
Bedient, P., A. Holder , J. Benavides, and B. Vieux, “Radar-based flood Weather Review, 88: 79–87, 1960.
warning system applied to tropical storm allison,” Journal of Hydrologic Gabriel, K. R. and J. Neumann, “On a distribution of weather cycles by
Engineering, 8 (6): 308–318, 2003. lengths,” Quarterly Journal of the Royal Meteorological Society, 83: 375–380,
Behrangi, A., B. Khakbaz, T. C. Jaw, A. AghaKouchak, K. Hsu, and S. 1957.
Sorooshian, “Hydrologic evaluation of satellite precipitation products over a Galloy, E., A. Le Breton, and S. Martin, “A model for weather cycles based
mid-size basin,” Journal of Hydrology, 397: 225–237, 2011. on daily rainfall occurrences,” Rhythms in Biology and Other Fields of
Berrisford, P., Dick Dee, Keith Fielding, Manuel Fuentes, Per Kållberg, Application, Springer Lecture Notes in Biomathematics, edited by Cosnard, M.,
Shinya Kobayashi, Sakari Uppala, ERA Report Series-The ERA-Interim Demongeot, Jacques, LeBreton, A, Vol. 49, 1981, pp. 303–318.
Archive Version 1.0, 2009. Georgakakos, K. and M. L. Kavvas, “Precipitation analysis, modeling, and
Boé, J., L. Terray, F. Habets, and E. Martin, “Statistical and dynamical prediction in hydrology,” Reviews of Geophysics, 25 (2): 163–178, 1987.
downscaling of the seine basin climate for hydro-meteorological studies,” Goodrum, G, K. B. Kidwell, and W. Winston, NOAA KLM Users Guide:
International Journal of Climatology, 27 (12):1643–55, 2007. September 2000 Revision, U.S. Department of Commerce, NOAA, NESDIS,
Bosilovich, M. G., C. Junye, R. R. Franklin, and F. A. Robert, “Evaluation of NCDC, Climate Services Division, Satellite Services Branch, Suitland, MD,
global precipitation in reanalyses,” Journal of Applied Meteorology and 2000.
Climatology, 47 (9): 2279–2299, 2008. Grace, R. A. and P. S. Eagleson, “The synthesis of short-time-increment
Bromwich, D., Ying Hwa Kuo, Mark Serreze, John Walsh, Le Sheng Bai, rainfall sequences,” Hydrodynam. Laboratory Report 91, Department of Civil
Michael Barlage, Keith Hines, Andrew Slater, “Arctic System Reanalysis: Call Engineering, Massachusetts Institute of Technology, Cambridge, MA, 1966.
41_Singh_ch41_p41-1-41-14.indd 10 8/22/16 2:09 PM

REFERENCES 41-11
Groisman, P. Y. and D. R. Legates, “The accuracy of United States precipita- W. Higgins, J. Janowiak, K. C. Mo, C. Ropelewski, J. Wang, A. Leetmaa,
tion data,” Bulletin of the American Meteorological Society, 75: 215–227, 1994. R. Reynolds, Roy Jenne, Dennis Joseph, “The NCEP/NCAR 40-year reanalysis
Gupta, V. K. and L. Duckstein, “A stochastic analysis of extreme droughts,” project,” Bulletin of the American Meteorological Society, 77 (3): 437–71,
Water Resources Research, 12 (2): 221–228, 1975. 1996.
Gupta, V. and E. Waymire, “Multiscaling properties of spatial rainfall and Kanamitsu, M., Wesley Ebisuzaki, Jack Woollen, Shi Keng Yang, J. J. Hnilo,
river flow distributions,” Journal of Geophysical Research, 95: 1999–2009, M. Fiorino, G. L. Potter, “NCEP-DOE AMIP-II reanalysis (R-2),” Bulletin of
1990. the American Meteorological Society, 83 (11): 1631–43, 1559, 2002.
Gupta, V. and E. Waymire, “A statistical analysis of mesoscale rainfall as a Kanamitsu, M. and H. Kanamaru, “Fifty-seven-year California reanalysis
random cascade,” Journal of Applied Meteorology, 32 (2): 251–267, 1993. downscaling at 10 km (CaRD10), Part I: System detail and validation with
Haan, T. N., D. M. Allen, and J. 0. Street, “A Markov chain model of daily observations,” Journal of Climate, 20 (22): 5553–71, 2007.
rainfall,” Water Resources Research, 12 (3): 443–449, 1976. Kavvas, M. L. and J. W. Delleur, “The stochastic and chronologic structure
Hamill, T. M., Gary T. Bates, Jeffrey S. Whitaker, Donald R. Murray, Michael of rainfall sequences-application to Indiana,” Technical Report 57, Water
Fiorino, Thomas J. Galarneau, Yuejian Zhu, William Lapenta, “NOAA’s Resources Research Center Purdue University, West Lafayette, Indiana, 1975.
Second-Generation Global Medium-Range Ensemble Reforecast Dataset,” Kavvas, M. L. and J. W. Delleur, “A stochastic cluster model for daily rainfall
Bulletin of the American Meteorological Society, 94 (10): 1553–65, 2013. sequences,” Water Resources Research, 17 (4): 1151–1160, 1981.
Heinemann, T., A. Latanzio, and F. Roveda, “The EUMETSAT multi-sensor Kavvas, M. L., and K. R. Herd, “A radar-based stochastic model for short-
precipitation estimate (MPE),” Proceedings of the second International time-increment rainfall,” Water Resources Research, 21 (9): 1437–1455, 1985.
Precipitation Working Group (IPWG) Meeting, Madrid, Spain, 2002. Kavvas, M. L., M. N. Saquib, and P. S. Puri, “On a stochastic description of the
Hanaish, I. S., K. Ibrahim, and A. A. Jemain, “Stochastic modeling of rainfall time-space behavior of extratropical cyclonic precipitation fields,” Journal of
in Peninsular Malaysia using Bartlett-Lewis rectangular pulse models,” Stochastic Hydrology and Hydraulics, 1: 37–52, 1987.
Modeling and Simulation in Engineering, 2011. Kavvas, M. L., P. S. Puri, and M. N. Saquib, “Analytical extensions on a
Hobbs, P. V. and J. D. Locatelli, “Rainbands, precipitation cores and description of cyclonic precipitation fields in time and space,” Journal of
generating cells in a cyclonic storm,” Journal of the Atmospheric Sciences, 35: Stochastic Hydrology and Hydraulics, 2: 263–279, 1988.
230–240, 1978. Kawanishi T., T. Sezai, Y. Ito, K. Imaoka, T. Takeshima, Y. Ishido, A. Shibata,
Hong, Y., K-L. Hsu, S. Sorooshian, and X. Gao, “Precipitation estimation et al., “The advanced microwave scanning radiometer for the earth observing
from remotely sensed imagery using an artificial neural network cloud clas- system (AMSR-E), NASDA’s contribution to the EOS for global energy and
sification system,” Journal of Applied Meteorology and Climatology, 43: 1834– water cycle studies,” IEEE Transactions on Geoscience and Remote Sensing,
1853, 2004. 41: 184–194, 2003.
Hopkins, J. W. and P. Robillard, “Some statistics from the Canadian Prairie Kidd, C. and G. Huffman, “Global precipitation measurement,”
provinces,” Journal of Applied Meteorology and Climatology, 3: 600–602, 1964. Meteorological Applications, 18 (3): 334–353, 2011.
Hubert, P., Y. Tessier, S. Lovejoy, D. Schertzer, F. Schmitt, P. Ladoy, J. P. Kidd, C. K., D. R. Kniveton, M. C. Todd, and T. J. Bellerby, “Satellite rainfall
Carbonnel, et al., “Multifractals and extreme rainfall events,” Geophysical estimation using combined passive microwave and infrared algorithms,”
Research Letters, 20 (10): 931–934, 1993. Journal of Hydrometeorology, 4: 1088–1104, 2003.
Huffman, G. J., R. F. Adler, D. T. Bolvin, G. Gu, E. J. Nelkin, K. P. Bowman, Kidd, C. and C. Muller, “Passive microwave—infrared combined algo-
Y. Hong, et al., “The TRMM multi-satellite precipitation analysis (TMPA): rithm,” Satellite Applications for Surface Hydrology, edited by F. Houssain and
quasi-global, multiyear, combined-sensor precipitation estimates at fine M. Gebremichael, Springer-Verlag, Dortrecht, The Netherlands, 2009, p. 327.
scales,” Journal of Hydrometeorology, 8: 38–55, 2007, doi:10.1175/JHM560.1. Klaes, K. D., M. Cohen, Y. Buhler, P. Schlussel, R. Munro, J-P. Luntama, A.
Iguchi T, T. Kozu, R. Meneghini, J. Awaka, and K. Okamoto, “Rainprofiling von Engeln, et al., “An introduction to the EUMETSAT polar system,” Bulletin
algorithm for the TRMM precipitation radar,” Journal of Applied Meteorology, of the American Meteorological Society, 88: 1085–96, 2007.
39: 2038–2052, 2000. Klazura, G. E. and D. A. Imy, “A description of the initial set of analysis
Intergovernmental Panel On Climate Change (IPCC), Climate Change products available from the NEXRAD WSR-88D system,” Bulletin of the
1995: The Science of Climate Change, edited by J. T. Houghton, L. G. Meiro American Meteorological Society, 74 (7): 1293–1311, 1993.
Filho, B. A. Callander, N. Harris, A. Kattenburg, K. Maskell, Cambridge Knebl, M. R., Z-L. Yang, K. Hutchison, and D. R. Maidment, “Regional
University Press, Cambridge, New York, 1996. scale flood modeling using NEXRAD rainfall, GIS, and HEC-HMS/RAS: a
Intergovernmental Panel on Climate Change (IPCC), IPCC Special Report case study for the San Antonio River Basin summer 2002 storm event,”
on Emissions Scenarios, edited by N. Nebojsa and R. Swart, Cambridge Journal of Environmental Management, 75 (4): 325–336, 2005.
University Press, Cambridge, 2000. Koutsoyiannis, D., “A probabilistic view of Hershfield’s method for estimat-
Intergovernmental Panel on Climate Change (IPCC), Climate Change ing probable maximum precipitation,” Water Resources Research, 35 (4): 1313,
2013: The Physical Science Basis, edited by Stocker, T.F., D. Qin, G.-K. Plattner, 1999.
M. Tignor, S.K. Allen, J. Boschung, A. Nauels, Y. Xia, V. Bex, and P.M. Midgley, Kouwen, N. and G. Garland, “Resolution considerations in using radar
Cambridge University Press, Cambridge, United Kingdom and New York, rainfall data for flood forecasting,” Canadian Journal of Civil Engineering, 16:
NY, USA, 1535 pp. 279–289, 1989.
Ishida, K., Kavvas, M., Jang, S., Chen, Z., Ohara, N., and Anderson, M, Kunkee, D. B., G. A. Poe, D. J. Boucher, S. D. Swadley, Y. Hong, J. E. Wessel,
“Physically based estimation of maximum precipitation over three watersheds and E. A. Uliana, “Design and evaluation of the first special sensor microwave
in Northern California: atmospheric boundary condition shifting,” Journal of imager/sounder,” IEEE Transactions on Geoscience and Remote Sensing,
Hydrologic Engineering, 20 (4): 04014052, 2015a. 46: 863–883, 2008.
Ishida, K., Kavvas, M., Jang, S., Chen, Z., Ohara, N., and Anderson, M., Lin, L-F., M. E. Ardeshir, L. B. Rafael, N. F. Alejandro, and W. Jingfeng,
“Physically based estimation of maximum precipitation over three watersheds “Dynamical precipitation downscaling for hydrologic applications using
in Northern California: relative humidity maximization method,” Journal of WRF 4D-Var data assimilation: implications for GPM era,” Journal of
Hydrologic Engineering, 04015014, 2015b. Hydrometeorology, 16 (2): 811–29, 2015.
Jacobs, P. A. and P. A. W. Lewis, “A discrete time series generated by mix- Lin, N., J. A. Smith, G. Villarini, T. P. Marchok, and M. L. Baeck, “Modeling
tures I: Correlational and run properties,” Journal of the Royal Statistical extreme rainfall, winds, and surge from Hurricane Isabel (2003),” Weather
Society. Series B (Methodological), 40 (1): 94–105, 1978. and Forecasting, 25 (5): 1342–1361, 2010.
James, W. P., C. G. Robinson, and J. F. Bell, “Radar-assisted real-time flood Lin, Y., and K. Mitchell, “The NCEP stage II/IV hourly precipitation
forecasting,” Journal of Water Resources Planning and Management, ASCE, analyses: Development and applications,” Proc., 19th Hydrology Conf.,
119 (1): 32–44, 1993. American Meteorological Society, Boston, 1.2, 2005.
Joyce, R. and P. A. Arkin, “Improved estimates of tropical and subtropical Liu, J. J., Z. Li, Y. L. Qiao, Y. J. Liu, and Y. X. Zhang, “A new method for
precipitation using the GOES Precipitation Index,” Journal of Atmospheric and cross-calibration of two satellite sensors,” International Journal of Remote
Oceanic Technology, 10: 997–1011, 1997. Sensing, 25: 5267–5281, 2004.
Joyce, R. J., J. E. Janowiak, P. A. Arkin, and P. Xie, “CMORPH: a method Lonfat, M., F. D. Marks, Jr., and S. S. Chen, “Precipitation distribution in
that produces global precipitation estimates from passive microwave and tropical cyclones using the Tropical Rainfall Measuring Mission (TRMM)
infrared data at high spatial and temporal resolution,” Journal of microwave imager: a global perspective,” Monthly Weather Review, 132 (7):
Hydrometeorology, 5: 487–503, 2004. 1645–1660, 2004.
Kalnay, E., M. Kanamitsu, R. Kistler, W. Collins, D. Deaven, L. Gandin, M. Lovejoy, S., “Area perimeter relations for rain and cloud areas,” Science, 187:
Iredell, S. Saha, G. White, J. Woollen, Y. Zhu, M. Chelliah, W. Ebisuzaki, 1035–1037, 1982.
41_Singh_ch41_p41-1-41-14.indd 11 8/22/16 2:09 PM

Lovejoy, S. and D. Schertzer, “Generalized scale invariance in the atmo- Rodriguez-Iturbe, I., V. K. Gupta, and E. Waymire, “Scale considerations in
sphere and fractal models of rain,” Water Resources Research, 21: 1233–1250, the modeling of temporal rainfall,” Water Resources Research, 20 (11): 1611–
1985. 1619, 1984.
Lovejoy, S. and B. B. Mandelbrot, “Fractal properties of rain, and a fractal Rodriguez-Iturbe, I., D. R. Cox, and P. S. Eagleson, “Spatial modeling of total
model,” Tellus, 37A: 209–232, 1985. storm rainfall,” Proceedings of the Royal Society of London A, 403: 27–50, 1986.
Lovejoy, S. and D. Schertzer, “Multifractals, cloud radiances and rain,” Rodriguez-Iturbe, I., D. R. Cox, and V. Isham, “Some models for rainfall
Journal. of Hydrology, 1–30, 2005. based on stochastic point processes,” Proceedings of the Royal Society of
Lovejoy, S. and I. P. de Lima, “The joint space-time statistics of macroweather London A, 410: 269–288, 1987a.
precipitation, space-time factorization and macroweather models,” Chaos: An Rodriguez-Iturbe, I., B. Febres de Power, and J. B. Valdes, “Rectangular
Interdisciplinary Journal of Nonlinear Science, 25 (7), 075410, 2015. pulses point process models for rainfall: analysis of empirical data,” Journal of
Lynch, P., “The origins of computer weather prediction and climate model- Geophysical Research, 92 (D8): 9645–9656, 1987b.
ing,” Journal of Computational Physics, 227 (7): 3431–44, 2008. Roldan, J. and D. A. Woolhiser, “Stochastic daily precipitation models, I. A
Mandelbrot, B. “The Fractal Geometry of Nature,” Freeman, New York, comparison of occurrence processes,” Water Resources Research, 18 (5): 1451–1459,
1983. 1982.
Maraun, D., F. Wetterhall, A. M. Ireson, R. E. Chandler, E. J. Kendon, M. Pessoa, M. L., R. L. Bras, and E. R. Williams, “Use of weather radar for flood
Widmann, S. Brienen, H. W. Rust, T. Sauter, M. Themel, V. K. C. Venema, K. P. forecasting in the Sieve River basin: a sensitivity analysis,” Journal of Applied
Chun, C. M. Goodess, R. G. Jones, C. Onof, M. Vrac, I. Thiele-Eich, Meteorology, 32 (3): 462–475, 1993.
“Precipitation downscaling under climate change: recent developments to Saha, S., Shrinivas Moorthi, Hua Lu Pan, Xingren Wu, Jiande Wang, Sudhir
bridge the gap between dynamical models and the end user,” Reviews of Nadiga, Patrick Tripp, Robert Kistler, John Woollen, David Behringer, Haixia
Geophysics, 48 (3): RG3003, 2010. Liu, Diane Stokes, Robert Grumbine, George Gayno, Jun Wang, Yu Tai Hou,
Marks, Jr., F. D., “Evolution of the structure of precipitation in Hurricane Hui Ya Chuang, Hann Ming H. Juang, Joe Sela, Mark Iredell, Russ Treadon,
Allen (1980),” Monthly Weather Review, 113 (6): 909–930, 1985. Daryl Kleist, Paul Van Delst, Dennis Keyser, John Derber, Michael Ek, Jesse
Menabde, M., D. Harris, A. Seed, G. Austin, and D. Stow, “Multiscaling Meng, Helin Wei, Rongqian Yang, Stephen Lord, Huug Van Den Dool, Arun
properties of rainfall and bounded random cascades,” Water Resources Kumar, Wanqiu Wang, Craig Long, Muthuvel Chelliah, Yan Xue, Boyin
Research, 33 (12): 2823–2830, 1997. Huang, Jae Kyung Schemm, Wesley Ebisuzaki, Roger Lin, Pingping Xie,
Menzel, W. P. and J. F. W. Purdom, “Introducing GOES-I: the first of new Mingyue Chen, Shuntai Zhou, Wayne Higgins, Cheng Zhi Zou, Quanhua Liu,
generation of Geostationary Operational Environmental Satellites,” Bulletin of Yong Chen, Yong Han, Lidia Cucurull, Richard W. Reynolds, Glenn Rutledge,
the American Meteorological Society, 75: 757–781, 1994. Mitch Goldberg, “The NCEP climate forecast system reanalysis,” Bulletin of
Mesinger, F., Fedor Mesinger, Geoff DiMego, Eugenia Kalnay, Kenneth the American Meteorological Society, 91 (8): 1015–57, 2010.
Mitchell, Perry C. Shafran, Wesley Ebisuzaki, Dušan Jović, Jack Woollen, Eric Saquib,M. N., M. L. Kavvas, and P. S. Puri, “Simulation of the time-space
Rogers, Ernesto H. Berbery, Michael B. Ek, Yun Fan, Robert Grumbine, evolution of an extratropical cyclonic precipitation field over USA,” Journal of
Wayne Higgins, Hong Li, Ying Lin, Geoff Manikin, David Parrish, Wei Shi, Stochastic Hydrology and Hydraulics, 2: 281–294, 1988.
“North American regional reanalysis,” Bulletin of the American Meteorological Schell, G. S., C. A. Madramootoo, G. L. Austin, and R. S. Broughton, “Use
Society, 87 (3): 343–60, 2006. of radar measured rainfall for hydrologic modeling,” Canadian Agricultural
Mimikou, M. A. and E. A. Baltas, “Flood forecasting based on radar rainfall Engineering, 34: 041–048, 1992.
measurements,” Journal of Water Resources Planning and Management, ASCE, Schertzer, D. and S. Lovejoy, “Physical modeling and analysis of rain and
122 (3): 151–156, 1996. clouds by anisotropic scaling multiplicative processes,” Journal of Geophysical
Moss, R. H., Richard H. Moss, Jae a Edmonds, Kathy a Hibbard, Martin R. Research, 92 (D8): 9693–9714, 1987.
Manning, Steven K. Rose, Detlef P. van Vuuren, Timothy R. Carter, Seita Scofield, R. A. and R. J. Kuligowski, “Status and outlook of operational
Emori, Mikiko Kainuma, Tom Kram, Gerald a Meehl, John F. B. Mitchell, satellite precipitation algorithms for extreme-precipitation events,” Weather
Nebojsa Nakicenovic, Keywan Riahi, Steven J. Smith, Ronald J. Stouffer, and Forecasting, 18: 1037–1051, 2003.
Allison M. Thomson, John P. Weyant, Thomas J. Wilbanks, “The next genera- Schmertz, J., P. Pili, S. Tjemkes, D. Just, J. Kerkmann, S. Rota, and A. Ratier,
tion of scenarios for climate change research and assessment,” Nature, 463 “An introduction to meteosat second generation (MSG),” Bulletin of the
(7282): 747–56, 2010. American Meteorological Society, 83: 977–992, 2002.
Nakamura K., T. Iguchi, M. Kojima, and E. A. Smith, “Global precipitation Schumacher, R. S. and R. H. Johnson, “Organization and environmental
mission (GPM) and dual-wavelength radar (DPR),” Proceedings of the URSI properties of extreme-rain-producing mesoscale convective systems,” Monthly
General Assembly, New Delhi, 2005, pp. 23–29, October, http://www.ursi. weather review, 133 (4): 961–976, 2005.
org/Proceedings/ProcGA05/pdf/F10.1(0803).pdf. Smith, J. A., D. J. Seo, M. L. Baeck, and M. D. Hudlow, “An intercomparison
Negri. A. J., L. Xu, and R. F. Adler, “A TRMM-calibrated infrared rainfall study of NEXRAD precipitation estimates,” Water Resources Research, 32:
algorithm applied over Brazil,” Journal of Geophysical Research, 107 [D20]: 2035–2045, 1996.
8048–8062, 2002. Smith, E. R. and H. A. Schreiber, “Point processes of seasonal thunder-
Neyman, J. E. and E. L. Scott, “A statistical approach to problems of cosmol- storm rainfall. Part 1, distribution of rainfall events,” Water Resources
ogy,” Journal of the Royal Statistical Society: Series B, 20: 1–43, 1958. Research, 9 (4): 871–884, 1973.
Ohara, N., Kavvas, M., Kure, S., Chen, Z., Jang, S., and Tan, E., “Physically Smith, J. A. and A. F. Karr, “A point process model of summer season rainfall
based estimation of maximum precipitation over American river watershed, occurrence,” Water Resources Research, 19 (1): 95–103, 1983.
California,” Journal of Hydrologic Engineering, 16 (4): 351–61, 2011. Smith, J. A. and A. F. Karr, “Parameter estimation for a model of space-time
Onogi, K., Junichi Tsutsui, Hiroshi Koide, Masami Sakamoto, Shinya rainfall,” Water Resources Research, 21 (8): 1251–1257, 1985.
Kobayashi, Hiroaki Hatsushika, Takanori Matsumoto, Nobuo Yamazaki, Smith, J. A. and W. F. Krajewski, “Statistical modeling of space-time rainfall
Hirotaka Kamahori, Kiyotoshi Takahashi, Shinji Kadokura, Koji Wada, Koji using radar and rain gage observations,” Water Resources Research, 10: 1893–
Kato, Ryo Oyama, Tomoaki Ose, Nobutaka Mannoji, Ryusuke Taira, “The 1900, 1987.
JRA-25 reanalysis,” Journal of the Meteorological Society of Japan, 85 (3): Soares, P. M. M., Rita M. Cardoso, Pedro M. A. Miranda, Joana de Medeiros,
369–432, 2007A. Margarida Belo-Pereira, Fátima Espirito-Santo, “WRF high resolution
Over, T. M. and V. K. Gupta, “Statistical analysis of mesoscale rainfall: dynamical downscaling of ERA-interim for portugal,” Climate Dynamics, 39
dependence of a random cascade generator on the large-scale forcing,” (9–10): 2497–2522, 2012.
Journal of Applied Meteorology and Climatology, 33: 1526–1542, 1994. Sorooshian S., K-L. Hsu, X. Gao, H. V. Gupta, B. Imam, and D. Braithwaite,
Parker, M. D. and R. H. Johnson, “Structures and dynamics of quasi-2D “Evaluation of PERSIANN system satellite-based estimates of tropical rain-
mesoscale convective systems,” Journal of the atmospheric sciences, 61 (5): fall,” Bulletin of the American Meteorological Society, 81: 2035–2046, 2000.
545–567, 2004. Sorooshian S., A. AghaKouchak, P. Arkin, J. Eylander, E. Foufoula-
Powell, M. D., “Boundary layer structure and dynamics in outer hurricane Georgiou, R. Harmon, J. Hendrickx J., et al., “Advancing the remote sensing
rainbands. Part I: Mesoscale rainfall and kinematic structure,” Monthly of precipitation,” Bulletin of the American Meteorological Society Nowcast, 92
Weather Review, 118 (4): 891–917, 1990a. (10): 1271–1272, 2011, doi: 10.1175/BAMS-D-11-00116.1.
Powell, M. D., “Boundary layer structure and dynamics in outer hurricane Srikanthan, R. and T. A. McMahon, “Stochastic Simulation of Daily
rainbands. Part II: Downdraft modification and mixed layer recovery,” Rainfall for Australian Stations”, Transactions of the American Society of
Monthly Weather Review, 118 (4): 918–938, 1990b. Agricultural and Biological Engineers, 26 (3): 0754-0759, 1983.
41_Singh_ch41_p41-1-41-14.indd 12 8/22/16 2:09 PM

REFERENCES 41-13
Todorovic, P. and V. Yevjevich, “Stochastic processes of precipitation,” Veneziano, D., R. L. Bras, and J. D. Niemann, “Nonlinearity and self-simi-
Hydrology Paper, 35: 1–61, 1969. larity of rainfall in time and a stochastic model,” Journal of Geophysical
Trapp, R. J., Mesoscale-Convective Processes in the Atmosphere, Cambridge Research, 101: 26371–26392, 1996.
University Press, Cambridge, 2013. Waymire, E., “Scaling limits and self-similarity in precipitation fields,”
Turk, F. J. and S. Miller, “Toward improving estimates of remotely sensed Water Resources Research, 21 (8): 1271–1281, 1985.
precipitation with MODIS/AMSR-E blended data techniques,” IEEE Waymire, E., V. K. Gupta, and I. Rodriguez-Iturbe, “A spectral theory of
Transactions Geoscience Remote Sensing, 43: 1059–1069, 2005, doi:10.1109/ rainfall intensity at the Meso-β scale,” Water Resources Research, 20 (10):
TGRS.2004.841627. 1453–1465, 1984.
UNESCO, World Water Balance and Water Resources of the Earth, Winchell, M., H. V. Gupta, and S. Sorooshian, “On the simulation of
UNESCO, Leningrad, 1978. infiltration- and saturation-excess runoff using radar-based rainfall estimates:
Uppala, S. M., P. W. KÅllberg, A. J. Simmons, U. Andrae, V. Da Costa effects of algorithm uncertainty and pixel aggregation,” Water Resources
Bechtold, M. Fiorino, J. K. Gibson, J. Haseler, A. Hernandez, G. A. Kelly, X. Li, Research, 34, 2655–2670, 1998.
K. Onogi, S. Saarinen, N. Sokka, R. P. Allan, E. Andersson, K. Arpe, M. A. Wiser, E. H., “Modified Markov probability models of sequences of pre-
Balmaseda, A. C. M. Beljaars, L. Van De Berg, J. Bidlot, N. Bormann, S. Caires, cipitation events,” Monthly Weather Review, 93: 511–516, 1965.
F. Chevallier, A. Dethof, M. Dragosavac, M. Fisher, M. Fuentes, S. Hagemann, Woolhiser, D. A. and H. B. Osborn, “A stochastic model of dimensionless
E. Hólm, B. J. Hoskins, L. Isaksen, P a. E. M. Janssen, R. Jenne, A. P. Mcnally, thunderstorm rainfall,” Water Resources Research, 21 (4): 511–522, 1985.
J.-F. Mahfouf, J.-J. Morcrette, N. A. Rayner, R. W. Saunders, P. Simon, A. Sterl, World Meteorological Organization (WMO), Manual for Estimation of
K. E. Trenberth, A. Untch, D. Vasiljevic, P. Viterbo, J. Woollen, “The ERA-40 Probable Maximum Precipitation, WMO, Geneva, 1986.
Re-Analysis,” Quarterly Journal of the Royal Meteorological Society, 131 (612): Xie, P. and P. A. Arkin, “Analyses of global monthly precipitation using
2961–3012, 2005. gauge observations, satellite estimates, and numerical model predictions,”
Valdes, J. B., I. Rodriguez-Iturbe, and V. K. Gupta, “Approximations of Journal of Climate, 9: 840–858, 1996.
temporal rainfall from a multidimensional model,” Water Resources Research, Yoshiro K., “Meteorological mission of MTSAT,” Institute of Electronics,
21 (8): 1259–1270, 1985. Information and Communication Engineers Technical Report, 102: 55–58, 2002.
41_Singh_ch41_p41-1-41-14.indd 13 8/22/16 2:09 PM

Chapter 42
Evapotranspiration
and Evaporative Demand
BY
MIKE HOBBINS AND JUSTIN HUNTINGTON
ABSTRACT must be addressed to make accurate, physically based estimates of ET and E0

that are supported by the best available science and that avoid the pitfalls of
Evapotranspiration is the transfer of moisture from sources at the surface into
accepted, and often dangerous, practices.
the atmosphere; evaporative demand is its maximum value assuming unlim-
Other essential references that extend the study and application of ET and
ited moisture availability and ambient atmospheric conditions. Despite
E0 beyond the scope of this chapter include McMahon et al. (2013) for a prac-
evapotranspiration being an essential flux in the water balance at all scales
tical look at current practice estimating ET and E0 on daily and monthly time
and its importance to many sectors of human endeavor, it remains difficult to
scales; Brutsaert (1982) for a complete scientific (and historical) background
observe directly and is most often modeled from the starting point of evapo-
on evaporation; Brutsaert (2005) and Dingman (2015) are excellent hydrology
rative demand. Evaporative demand is estimated using a wide array of obser-
textbooks, with Dingman (2015) particularly useful as a primer on the evapo-
vations or parameterizations, many of which are outdated or of dubious value.
ration dynamics of snowpacks; the Shuttleworth (1992) chapter on evapora-
Approaches to estimation of evapotranspiration vary across sectors and
tion in the Handbook of Hydrology is an excellent desk reference; the UN Food
scales of time and space: for example, inferred from the water balance across
and Agriculture Irrigation and Drainage Paper no. 56, commonly (and here-
basins scales, estimated at near-point scales from advanced instrumentation,
in) referred to as “FAO-56” (Allen et al., 1998), is a complete and accessible
or, increasingly, remotely sensed at scales that bridge from field to regional.
report on the most accurate and scientifically representative estimation pro-
Evaporative demand is observed or estimated from meteorological and radia-
cedure on reference ET and crop water requirements; Huntington (2006)
tive variables, yet many obstacles remain to its use in practice and in future
provides a review of trends in ET and E0; Wang and Dickinson (2012) provide
climate scenarios.
a review of global terrestrial evapotranspiration; and Kalma et al. (2008)
This chapter summarizes these issues with an eye to applicability, closing
provide a review of the remote sensing of ET.
the gap between operations and science in the observation, estimation, and
prediction of evapotranspiration and evaporative demand, and posing future
research questions for researchers and practitioners. 42.1.1 Definitions
In order to enhance the readability of the chapter, we here provide a list of
some essential definitions:
42.1 INTRODUCTION AND HISTORY OF THEORY
Evapotranspiration, ET—the actual transfer of moisture from the sur-
Evapotranspiration (ET)—the fluxes of evaporation from soil moisture, water, face—evaporation from bare soil, intercepted water on vegetation, and
and ice, and transpiration from plants—is essential to life on earth, playing a open water (and snow and ice) and transpiration from within vegetation—
critical role in regulating the hydrologic, climatic, and ecological dynamics at to the atmosphere under ambient conditions.
most time and space scales. Yet its estimation presents both research and Evaporative demand, E0—the maximal rate of evapotranspiration given
operational hydrologists significant challenges. To simplify the estimation of ambient atmospheric conditions (i.e., radiation and advective drivers) but
ET, hydrologists have long relied on the simpler concept of evaporative with an unlimited moisture supply; the “thirst of the atmosphere” which
demand (E0) to measure instead the thirst of the atmosphere as ET’s upper can be measured at a point by pan evaporation (Epan) or lysimetry; or
limit. Here, we examine both ET and E0. modeled areally by potential evaporation (Ep), reference ET (ET0), or wet
The many uses of, and estimation techniques for, ET and E0 vary along environment evaporation (Ew).
many axes. Users need ET and E0 estimates at timescales varying from daily Pan evaporation, Epan—the amount of water that evaporates from an
operations (e.g., reservoir operations) to weekly (e.g., drought monitoring evaporation pan over a given length of time; a physical, point-source mea-
and irrigation scheduling) to seasonal (e.g., demand analysis by water utili- sure of E0.
ties). Scientific scales run the gamut from instantaneous to secular and point Potential evapotranspiration, Ep—ET that would take place from a well-
to global. There is a clear need for operationally and physically sound mea- watered surface under ambient atmospheric conditions; a synthetic mea-
sures of ET and E0—measures that rely on appropriate forcings, are well cali- sure of E0.
brated to reasonable parameter values, and yield accurate results across all Reference evapotranspiration, ET0—ET that would take place from a
relevant time and space scales. The purpose of this chapter is to close the gap well-watered reference crop under specific surface moisture conditions but
between the science of ET and E0 estimation and its applications in opera- ambient atmospheric conditions; a synthetic measure of E0.
tional settings and to support climate change impact studies and resource Wet environment evaporation, Ew—ET that would take place from an
management in increasingly uncertain hydroclimates. While this is not an areally extensive wet surface under ambient atmospheric conditions; a
operations manual, we will attempt to inform operations as to what questions synthetic measure of E0.
42-1
42_Singh_ch42_p42.1-42.18.indd 1 8/26/16 4:36 PM

42-2 Evapotranspiration and Evaporative Demand
Latent heat, lET—the heat energy transferred by evapotranspiration (or Equations). Further, the Bowen ratio may be used to obtain a near-direct
sublimation); the energetic equivalent of ET; ET multiplied by the latent physical estimate of ET from both water and land surfaces in the Energy
heat of vaporization l (or the latent heat of fusion for evaporation from ice). Budget Bowen Ratio (EBBR) method. Measurements over land surfaces are
Crop evapotranspiration, ETc—ET estimated from a crop surface assum- generally performed in research-oriented or nonoperational settings: Ts and T
ing well-watered and stress-free conditions. and esat and ea are replaced by T and ea measured at two reference heights to
Fraction of reference evapotranspiration, ET0F, Kc—ratio of ET to ET0. establish the temperature and vapor gradients. Using EBBR for water bodies
is more commonly applied for nonoperational E estimates and is discussed
42.1.2 Ancient Paradigms later (see Sec. Energy Budget).
As far back as the ancient Greeks, ET was understood to represent an integral
part of a natural system that links the earth, the atmosphere, and the seas— Penman and other Combination-Based Equations
what we now know as the hydrologic cycle. Yet significant gaps still remain The year of Thornthwaite’s equation also saw one of the greatest break-
between this long-held scientific understanding and current operational throughs in estimation of E0, and consequently of ET: the development of the
practice. For instance, the ancient Greeks also understood that ET was driven potential ET (Ep) concept by Penman (1948). Prior to this point, ET had been
by the combined effects of winds, sunshine, humidity, and temperature, and, calculated using aerodynamic methods when the supply of energy was not
indeed, there are many current formulations that describe the evaporative limiting, and by energy budget methods when the vapor transport mecha-
process in terms of all of these drivers; however, many modern estimation nism was not limiting. Instead, Penman (1948) examined pan evaporation
procedures rely instead solely on temperature, and this disjoint from the state and described the meteorologic limits on ET as a sink strength that metered
of the science lingers yet in current operational estimation of both ET and E0. the ability of the atmosphere to absorb and bear away moisture (the advective
42.1.3 Climatologic Estimators of ET limit) and the source strength that metered the energy available for ET (the
radiative limit). He posited that in normal situations both the advective and
Absent direct estimators, a useful start for the modern era in ET estimation radiative limits apply, but that, further, the availability of moisture to evapo-
may be the climatological estimates of Ol’Dekop (1911), Turc (1954), and rate is essentially unknowable at useful scales. To circumvent the difficulties
Budyko (1974), who viewed ET from the perspective of long-term analyses of replicating the latter constraint, he assumed “a reproducible surface of
and as a function of its hydrologic and energetic limitations: distinguishing known properties”—that is, a surface sufficiently moist that ET is constrained
landscapes between those whose hydrologic cycle is predominantly energy only by the remaining radiative and advective limits (this has since become
limited or water limited, and formulating a framework that is still useful in the general definition of Ep). His approach takes into consideration the simi-
ecological and hydrological analyses (see Sec. 42.2.2), for example, for estima- larity of flux profiles of wind speed, specific humidity, and potential tempera-
tion of ET in ungaged basins. ture in the dynamic sublayer above a saturated surface. (Confusion regarding
Thornthwaite (1948) took another tack: a geographer, he demonstrated a the precise definition of Ep and its place among the variety of E0 estimators is
simple approach to climatic classification that used an E0 parameterized addressed in Sec. 42.3.6.)
solely on air temperature (T). While a useful engineering approach for an era The Penman equation takes the form of a linear combination (often called
when the availability of data for physical drivers was limited, this has since led a “convex linear combination” as its coefficients are non-negative and sum to
to much of today’s disjoint between science and operations (see Sec. 42.3.4). one) of ET rates driven by the surface energy availability (i.e., the radiative
driver in the first term on the LHS) and by the vapor transfer (i.e., the advec-
42.1.4 Physical Estimators of E0 tive driver in the second term), as follows:
Evaporation Pans
The first widespread physical estimator of evaporation that was of direct ∆ γ
operational use was the evaporation pan. The basic premise of pans—or their λEp = Qn + EA (42.2)
∆ +γ ∆ +γ
relatives, the lysimeter and sunken (Colorado) pan—is that evaporation from
a freely available water source exposed to the atmosphere represents the
maximum possible evaporation under ambient meteorological and radiative where lEp is the potential ET (W m−2); Qn is the net available energy (W m−2);
conditions. This “potential” evaporation rate is measured simply over a given EA is the drying power of the air (W m−2); Δ (Pa K−1) is the slope of the satu-
period as the water depth change in a pan, or the weight in a weighing lysim- rated vapor pressure curve (desat/dT) at 2-m air temperature T (K); and g is
eter. Its conceptual and practical simplicity leant itself to early adoption as a the psychrometric constant (Pa K−1) (see Shuttleworth (1992) for formula-
direct measure of evaporative demand with uses in irrigation scheduling in
tions for Δ, g, and l). EA is parameterized as f(U z )(esat − ea ), and the wind
agriculture (see Sec. 42.2.4) and estimation of evaporation from lakes in water
function f(Uz) (W m–2 Pa–1) is open to regional and seasonal recalibration.
resources (see Sec. 42.2.9). While in many regions of the world evaporation
Penman’s “combination equation” has given rise to many physically based
pans are no longer considered essential metrics in hydrometeorologic sys-
equations still used under various conditions today, to either supply Ep for use
tems, the availability of widespread, long-term records has recently fostered
in land surface models (LSMs, e.g., in basin budget analyses), in irrigation
new interest in changes in E0 (see Sec. 42.3.2), and it has even been proposed
scheduling, water management decisions (e.g., in reservoir scheduling), or for
that they act as surrogate radiometers (Roderick et al., 2007).
its own use (e.g., in climatologic analysis). Examples are the Penman-
Bowen Ratio Monteith equation (Monteith, 1965), which replaces the empirical advective
driver (wind function), particularly to improve the characterization of the
Bowen (1926) recognized the physical similarity between the conduction and
diffusion process, with uses in agriculture (see Sec. 42.2.4); the Priestley–
diffusion of sensible heat (H) and evaporation (E) and diffusion of water
Taylor equation (Priestley and Taylor, 1972), which replaces the advective
vapor from a water body, and combined the two in a physical expression that
term by a constant multiplier (> 1) of the radiative term yielding an estimate
defines the dimensionless Bowen Ratio β:
of E0 that is primarily a function of radiation; the PenPan equation (Rotstayn
H T −T et al., 2006), which uses meteorological and radiative data to synthesize pan
β= ≈γ s (42.1) evaporation observations, with uses in streamflow simulation and forecasting
λE esat − ea
and climatological studies and predictions.
where H and λE are in the same units (commonly W m–2); l is the latent heat
Reference ET
of vaporization of water (J kg−1); T is the air temperature (°C) at a reference
height usually of 1–3 m above the water surface; g is the pyschrometric con- In no small part due to the uncertainty about the definition of Ep (see Sec.
stant (Pa °C−1), which is a function of atmospheric pressure and T; Ts is the 42.3.6), and to meet the ET-estimation needs of agricultural users, the con-
water surface “skin” temperature (°C); and esat and ea are the saturated and cept of reference ET was formulated (Doorenbos and Pruitt, 1977). In this
actual vapor pressures (Pa), respectively. b can also be defined over nonwater concept, the prevailing meteorologic and radiative conditions are used to
surfaces, and reflects their wetness: values of b less than one indicate ample estimate an ET rate for a heavily specified, well-watered surface: this estimate
moisture availability (lE exceeds H); values greater than one indicate drier is called reference ET (ET0). To estimate actual ET from a cropped surface
surfaces (H exceeding lE). However, as lE approaches zero, b becomes (ETc), ET0 is then multiplied by coefficients that account for variations from
unbounded, leading to a preference for use of the evaporative fraction [EF =1/ the reference conditions due to such factors as crop type, phenology, salinity,
(1+ b)]. and stress. Most current estimation of ET0 relies on a specific parameteriza-
The similarity assumptions underlying the Bowen Ratio are an integral part tion of the Penman–Monteith (PM) model (Monteith, 1965), itself an off-
of the derivation of the Penman (1948) combination equation central to so much shoot of Penman’s (1948) seminal approach. The general PM formulation
of current ET estimation (see Sec. Penman and Other Combination-Based parameterizes the moisture availability by two resistances to the diffusion of
42_Singh_ch42_p42.1-42.18.indd 2 8/26/16 4:36 PM

Relevant Concepts and Methods 42-3
water vapor from within the leaves of a vegetative canopy out to the dynamic where ∂Θ / ∂t is the time rate of change of moisture avaialability in units of
boundary layer. mass flux, the negative sign is due to ∂Θ/ ∂t itself being negative; and the
42.1.5 Satellite Remote Sensing of ET
subscript “max” indicates the fastest rate at which the soil can exfiltrate water
to the surface for evaporation.
Optical and thermal satellite imagery of earth’s surface is commonly used to Crucially, moisture availability is seldom knowable at scales useful to prac-
estimate ET from field to landscape scales, and is especially useful for water titioners, and this has led to the development of the more tractable concept of
rights, water resource decisions, and global water budget accounting. Remote evaporative demand (E0), which is the maximal ET possible subject to only
sensing of ET has allowed scientists and water managers to investigate and the remaining radiative and advective limits. The idealized flux of E0 has
make management decisions based on figures that incorporate the spatial and become central to hydrologic science and practice. It is primarily used in the
temporal variability of ET. Advances in ET estimation that incorporate estimation of ET, which may be derived from E0 by adding a further con-
remotely sensed imagery with local weather station or gridded weather data straint derived from some parameterization of Θ. While ET has long been
allow for these important spatial and temporal ET characteristics to be esti- scaled from E0 using a linear function of Θ, feedbacks between ET and E0 have
mated. For a review of recent remote sensing techniques for estimating ET, see also been recognized for some time (Bouchet, 1963), and they have been
Kalma et al. (2008) and Gonzalez-Dugo et al. (2009). Recent case studies and coupled in a nonlinear fashion in what is now known as the Complementary
highlights on the uses and benefits of field scale remote sensing applications Relationship (CR; see Sec. 42.2.3).
for water resource decision making can be found in Serbina and Miller (2014).
Radiative Limit
42.1.6 Emerging and Scientific Methodologies
The net available energy (Qn) is the energy available for turbulent heat trans-
Advances in instrumentation, data collection, and processing techniques have fers to the dynamic boundary layer as either latent heat lET or sensible heat
permitted previously research-oriented ET-estimation technologies, such as H and it is derived from the instantaneous balance of the energy storages and
eddy correlation (EC) and scintillometry techniques, to gain traction as fluxes shown in Fig. 42.1 and described as follows:
operational methodologies.
∂W
Eddy Correlation = (1 − α )Rd + Ld − Lu − λ ET − H − G − C − Ad (42.4)
∂t
The eddy correlation (EC) technique for estimating ET has become common
among researchers and practitioners alike over the last decade. The EC tech- where all terms are in flux units (W m−2, multiply by 86,400/l for mm day−1).
nique is considered one of the most accurate and direct approaches to study On the LHS, ∂W/ ∂t is the time rate of change of heat storage in the evaporat-
micrometeorology, land surface energy balance, land and atmospheric inter- ing layer, positive for increasing energy storage. On the RHS, positive fluxes
actions and feedbacks, remote sensing models of ET, and vegetation physio- are into the evaporating surface, negative out: α is the surface albedo (-); Rd is
logical and ecological processes. Practitioners commonly rely on the EC the downward shortwave radiation incident at the surface; Ld and Lu are the
technique for vegetation monitoring, water budget estimation, and precision longwave radiation fluxes inward to and outward from the surface, respec-
agriculture. The technique estimates vertical fluxes of heat, water vapor, and tively; λ ET is the latent heat flux (the energy equivalent of ET); H is the
other gases by measuring instantaneous deviations in vertical wind speed and sensible heat flux; G is the net heat flux conducted from the evaporating
the flux of interest. The development of eddy flux networks over the last surface into the soil (or ground heat flux) below; C is the energy absorbed by
decade, such as Ameriflux, CarboEurope, FLUXNET, ICOS, iLEAPS, NEON, biochemical processes in vegetation in the control volume; and Ad is heat
and OzFlux, has led to more unification of data-collection and processing gained by advection to the control volume.
steps. However, differences in experimental purpose, design, and station Although Eq. (42.4) is defined explicitly for water evaporating from the liq-
instrumentation make EC-processing steps difficult to standardize for opera- uid state, its broad concept still applies for evaporation from snow and ice—that
tional applications. is, sublimation—although some of the terms change their relative importance
and, indeed, sign. The phase change of solid to vapor requires a further ener-
Scintillometry getic input, so the latent heat of fusion lf must be added to the latent heat of
Scintillometry is an emerging technique for indirectly estimating integrated vaporization l, noting that l exceeds lf by a factor of about 7.5. The conduction
ET over large areas through estimation of H and available energy (Qn). The of H with the ground below the snowpack G is usually negative (i.e., heat trav-
technique relies on measuring the transmission and scattering of electromag- els to the snowpack from the ground), Ad accounts for heat advected from
netic radiation through the turbulent atmosphere. H is estimated by measur- precipitation (Prcp) onto the snowpack, and is always positive. Lu usually
ing atmospheric scintillations caused by turbulence-induced refractive index exceeds Ld, while ET and H are mostly negative. The thermodynamics of the
fluctuations of the air, in combination with measurements of site-specific snowpack itself vary both seasonally from the accumulation phase through the
meteorological conditions and application of Monin-Obukhov similarity snowmelt phase (itself consisting of warming, ripening, and finally liquid run-
theory (MOST). Scintillometer systems consist of an optical or radio trans- off phases) and diurnally with an internal freeze/thaw cycle. The change in
mitter, receiver, data logger, and signal-processing unit. Optical transmitters energy storage term ∂W / ∂t dominates across seasonal time-scales, and ET
and receivers are most commonly used in ET studies. Path-averaged H esti- may become a minor flux at short time-scales during snowmelt periods when,
mates can be made over path lengths of hundreds to a few thousand meters. in fact, the opposite flux—condensation of moisture from the atmosphere onto
Common research applications include estimating ET and H over large areas, the snowpack—often occurs (Hood et al., 1999); in fact, over a year, condensa-
validating airborne and satellite remote sensing estimates of ET and H, esti- tion and evaporation are of the same order of magnitude, leaving snowmelt the
mating water budgets, and validating mesoscale numerical models. Costs of largest water balance term. It is for this reason and the complications pertaining
scintillometry systems are currently equivalent to those of well-equipped EC to its measurement and modeling that we do not further examine sublimation
and available energy measurement systems. Scintillometry is well suited for in this chapter: interested readers are encourage to seek out Dingman (2015).
long-term measurement due to the system being fairly insensitive to environ- As the net available energy (Qn) for lET and H is dominated by radiative
mental conditions such as rain, snow, and dust, and requires little power and terms it is also commonly referred to as the “radiative driver”; it may be better
maintenance, making the system ideal for measurements in remote areas. understood by rearranging Eq. 42.4 as follows:
∂W
λ ET + H = Qn = (1 − α) Rd + Ld − Lu − G − C − Ad − (42.4a)
42.2 RELEVANT CONCEPTS AND METHODS ∂t
42.2.1 Constraints and Drivers of ET and E0 which elements of Eq. 42.4a appear in the estimation of Qn depends on the
ET is subject to three physical constraints: the hydrologic limit is the avail- surface assumptions, time step, and data availability: C, Ad, and ∂W/∂t are
ability of water to evaporate and/or transpire from the soil, plants, open water, generally ignored, as is G on daily or annual time steps.
and ice; the radiative limit is the availability of energy to drive the evaporative For a water body, evaporation lE is expressed as a residual of the energy
process; and the advective limit is the ability of the atmosphere to absorb and balance of the entire water body as a control volume from:
bear away moisture.
λ E = Rn + Ln + Qv + Qb − Qw − H − Qx (42.5)
Hydrologic Limit
The hydrologic limit can be described as: where lE is the latent heat flux (for open-water evaporation, we no longer dis-
cuss lET, as there is no transpiration); Rn is the net shortwave radiation [or
 ∂Θ  (1 – α)Rd from Eq. (42.4)]; Ln is the net longwave radiation [or Ld – Lu from Eq.
ET ≤ −  (42.3)
 ∂t  max (42.4)]; Qv is the net advected energy to the water body from surface and
42_Singh_ch42_p42.1-42.18.indd 3 8/26/16 4:36 PM

Figure 42.1 Instantaneous energy balance at an evaporating surfaces. All fluxes and heat-storage changes are
in flux units (W m−2). The grey box represents the control volume to and from which all fluxes pass and within
which all heat-storage changes are considered.
groundwater inflows and outflows and direct precipitation (Prcp); Qb is the where the function g(…) represents a Penman-like combination of the advec-
energy exchange from bottom sediments to the water body; Qw is the energy tive and radiative drivers expressed in terms of the input variables. ET is then
advected away from the body of water to the atmosphere by the evaporated always constrained to the lowest of the hydrologic limit [Eq. (42.3)] and the
water; H is the sensible heat flux convected and conducted from the water body; function g(…) in Eq. (42.8).
and Qx is the change in stored energy in the water body [∂W /∂t in Eq. (42.4)]
(all terms are W m−2). The flux Qw occurs over land but is rarely estimated Uses of E0
because it is small, and the temperature of the evaporating water is usually not There are three traditional uses of E0, and three general types: Epan; Ep; and
known. For open-water applications Qw can be estimated based on the difference ET0. The primary use of E0 is as a starting point in the estimation of ET, gen-
between the water surface temperature and a base temperature [see Eq. (42.29)]. erally through some parameterization of Θ that depends on the specific use
Complete knowledge of Qn requires extensive instrumentation not gener- for the required ET estimate: traditionally, LSMs scale ET from Ep as a func-
ally available on an operational basis: much parameterization is therefore tion of a model state that tracks soil moisture, or, where vegetation is present,
required, and this often relies on, or is mixed with, observations of surrogate the transpiration (the dominant term in ET) is parameterized with a bulk
variables (Morton 1983a; Allen et al., 1998; Brutsaert, 2005). stomatal resistance that reflects water availability; remote sensing models
commonly scale ET from Ep or ET0 based on modeled surface energy balances
Advective Limit or vegetation indices. A second use is in irrigation scheduling and consump-
The advective limit describes the ability of the dynamic boundary layer to tive use estimation in agriculture, wherein ET0 estimates are scaled by apply-
absorb and bear away moisture, and is described following Dalton’s (1802) ing coefficients specific to crops, their phenology, soil stress, etc. (see Sec.
general form of the mass transfer equation for evaporation from water: 42.2.4). The third, more recent use of E0 has developed within the climato-
E = M(esat − ea ) (42.6) logical community, which has recently focused on its long-term trends (par-
ticularly in Epan, due to its long-standing and widespread global data record),
where E is the evaporation rate; esat and ea are the saturation vapor pressure at and as a diagnostic indicator of the dryness of the lower atmosphere and its
the temperature of the water surface and the actual vapor pressure of the air, causes (see Sec. 42.3.2).
respectively; and M is a mass transfer coefficient that is a function of wind
speed, atmospheric stability, surface roughness, thermally induced turbu- 42.2.2 Climatologic Estimators of ET
lence, barometric pressure, and the density and viscosity of the air. Expressing Budyko (1974) realized that ET couples the water and energy cycles, thus
M as a “vapor transfer function” or “wind function” f (Uz) of wind speed Uz examining ET in the context of its water and energy limits (see Sec. 42.2.1)
(m s−1) at height z (m) above the surface that describes the variation of vapor would permit an ecologically useful analysis of the variation of large-scale,
transfer of the air as a function of Uz gives us a generic expression of the long-term hydroclimatology. This is best expressed graphically (see Fig. 42.2),
advective limit to ET from all surfaces: by displaying the relation between two indices representing hydroclimatic
λ ET ≤ f(U z )(esat − ea ) (42.7) behavior across temporal scales long enough to allow steady-state conditions

to prevail and spatial scales large enough for influences of climate on the land-
where λET is in W m ; esat and ea (Pa) are estimated from the Clausius–
–2
surface/climate interactions to fully develop. The dryness index Φ (on the
Clapeyron equation at T and dewpoint temperature (Tdew), respectively (both x-axis) is the ratio of E0 [where E0 is generally estimated from a radiation-
in °C); and f(Uz) is in W m–2 Pa–1. [Note the different temperatures at which based parameterization, commonly the Priestley–Taylor equation; Priestley
esat is estimated in Eqs. (42.6) and (42.7): in Eq. (42.6), esat is estimated at the and Taylor (1972)] and Prcp; the evaporative index ε (on the y-axis) is the
water surface temperature, whereas in Eq. (42.7), it is at the air temperature, ratio of ET to Prcp. He proposed a relation between the two indices through
commonly measured at a 2-m height.] The RHS of Eq. (42.7) is also known as an empirical relationship now known as the “Budyko curve,” as follows:
the drying power of the air (EA). The accurate parameterization of the advec-
tive component is essential to any physically based ET or E0 model (Hobbins
et al., 2001a, 2001b; Sugita et al., 2001). f(Uz) has many functional forms  1
ε B = Φtanh   (1 − cosh Φ + sinh Φ) (42.9)
depending on site-specific factors and experimental design; a complete  Φ
expression for f(Uz) under neutral conditions—that is, a stable atmospheric
boundary layer—is given by Brutsaert and Stricker (1979), Eq. (42.17). Figure 42.2 shows the relation between Φ and ε in theory and as applied to
However, due to the complexities resulting from the effects of atmospheric 229 Australian basins from Donohue et al. (2010). The horizontal extent rep-
instability acting over short time periods and onerous data requirements, such resents the hydroclimatic spectrum: water-limited to the right; energy-limited
theoretical formulations are generally dropped in favor of empirical relation- to the left. In water-limited (arid-tending) areas, ET is limited by water avail-
ships, such as Penman’s (1948) simple, empirical, linear U2 function (m s−1). ability: it is asymptotic to, but generally cannot exceed, Prcp. In energy-limit-
The relationship between ET, E0, and their meteorological and radiative ed (humid-tending) areas, ET is asymptotic to, but cannot exceed, E0. The
drivers may be summarized thus: Budyko curve represents the variation in the behavior of ET between these
two limits, and so can be used to define long-term, large-scale ET in regions
λ ET ≤ λ E0 = g (Qn ,U z ,T , ea ) (42.8)
where E0 and Prcp—both generally more available than ET—are known. Not
42_Singh_ch42_p42.1-42.18.indd 4 8/26/16 4:36 PM

Figure 42.2 The Budyko (1974) concept in theory and practice. Solid straight lines denote the
water and energy limits on ET (as labeled), to which the behavior of ET is asymptotic. The curve
shows Budyko’s idealized relation between Φ and ε. The vertical arrows represent the ET and Runoff
portions of Prcp, respectively. The circles represent the behavior as observed in 229 Australian basins
by Donohue et al. (2010).
only can ET be estimated (lower vertical arrow in Fig. 42.2) in an ungaged the energy available goes to H, leading to smaller increases in T, VPD, or EA,
region, but so too can Runoff (i.e., Runoff = E0 - ε. Prcp), as the angled dashed and thus smaller increases in E0 above Ew. At the limit when moisture is com-
line represents the totality of the hydrologic cycle (recall the large-scale, long- pletely available (i.e., Θ = 1 in Fig. 42.3), ET is at its maximum, and converges
term constraints in which the paradigm operates). with E0 on Ew. Required for the relationship are a regional-scale homogeneity
While Budyko (1974) was able to close water budgets to within 10%, Fig. of land surface and water availability and a time-scale sufficient for complete
42.2 also indicates scatter of basins’ behavior around the curve (there are development of the aforementioned feedbacks between surface moisture and
other curves developed). This so-called “Budyko scatter” (δ) is the difference the dynamic boundary layer: Brutsaert and Stricker (1979) state that scales on
between the evaporative index predicted by the Budyko curve (εB) and that the order of thee days and hundreds of meters are sufficient.
observed, or δ = ε - εB = (ET - ETB)/Prcp, where ETB is the Budyko-predicted The original hypothesis (Bouchet, 1963) for the CR held that:
ET. The absolute value of δ increases at lower time and space scales that
increasingly contravene the central assumption of long-term regionality. ET + E0 = kEw (42.10)
Attempts to increase the ET-predictive power of these relations (i.e., to
decrease δ) at smaller time and space scales involve characterizing the basin where the value of the dimensionless constant k is 2. This value implied that
surface, including land cover, soil, and topography; but vegetation dynamics all the energy released due to moisture limitations is taken up by E0, leading
and particularly storage effects must also be accounted for (Milly, 1994; to a direct, symmetrical feedback between the energy released at the surface
Donohue et al., 2010). The Budyko (1974) framework remains a useful para- (q2, from ET decreasing) and the increase in the demand for latent heat flux
digm for hydrologists: for example, Szilagyi and Jozsa (2009a) used meteorological (q1, leading to E0 increasing)—that is, q1/l = q2/l in Fig. 42.3. However, recent
and radiative data in the combined contexts of the CR (Sec. 42.2.3) and Budyko’s work has uncovered different feedback strengths depending on the particular
coupled water/energy balance to derive various ecosystem characteristics— E0 (i.e., ET0, Ep, and Epan) and Ew used, leading to discussion of asymmetrical
mean effective relative soil moisture and maximum soil water storage available (q1/l > q2/l) complementarity (Szilagyi, 2007; Huntington et al., 2011).
to plants—useful for predicting ecosystem responses to climate change. A generic expression for the CR that permits both symmetry and asym-
metry can be given as:
42.2.3 Complementarity of ET and E0
bET + E0 = (1 + b)Ew (42.11)
The first considerations of feedbacks across the land surface-atmosphere
interface linking both the supply of and demand for evaporated moisture
from the land surface led Bouchet (1963) to hypothesize a complementarity where b is the scaling factor (dimensionless), taking a value of 1 for symmetry,
between variations in regional-scale ET and those in point-scale E0. While his and exceeding 1 for asymmetry (see Fig. 42.3).
CR hypothesis still lacks a rigorous proof, it has nevertheless been observed The CR hypothesis has led to a family of ET models that break with the
across many hydroclimates and regions (e.g., Hobbins et al., 2004) and mod- traditional conception of E0: that is, one that ignored feedbacks and interac-
els predicated upon it have proved useful in estimating regional-scale ET in tions across the land surface-atmosphere interface and held E0 to be indepen-
areas where moisture data are scarce (e.g., Morton, 1983a; Hobbins et al., dent of ET. In this traditional paradigm, ET is generally modeled either as a
2001a; 2001b). simple, often empirical function of E0 derived from pan evaporation (Epan)
Briefly, Bouchet (1963) proposed that, considering a specified energy bud- observed at nearby weather stations, or by ground-based models that tend to
get, the passage of air masses over land surfaces modifies them according to rely on gross assumptions regarding processes in the soil, vegetative, and
the proportion of net radiative energy transferred back to the atmosphere as atmospheric systems. Instead, CR models rely on these surface-atmosphere
latent heat (lET). Decreases in ET due to reductions in water availability lead feedbacks and thus bypass the poorly understood dynamics within each com-
to an energy release as an increase in sensible heat (H in Fig. 42.1) and hence ponent, and that incur minimal data requirements as to the nature of the land
to temperature increases and humidity decreases of the dynamic boundary surface that bedevils traditional ET estimation. They can be applied across
layer, which lead to an increase in the drying power of the air (EA) through an unmetered basins at a regional spatial scale and so can be of particular use to
increase of vapor pressure deficit (VPD, or esat – ea) and ultimately to an water managers. Widely used examples of such models are the WREVAP suite
increase in E0. Less-important effects relate to changes in albedo (which of models for areal and lake evaporation (Morton et al., 1985) and the
determines the shortwave radiation balance) and the surface temperature Advection-Aridity approach (Brutsaert and Stricker, 1979). In the field of
(which affects the longwave radiation balance). The CR is indicated in drought-monitoring, attention is turning to the use of the CR to uncover
Fig. 42.3, where the Ew term is the theoretical ET rate for a large-scale wet ET-related drought dynamics that may be more easily detected as a signal in
environment. As moisture availability increases, the opposite occurs: less of E0 (see Sec. 42.3.3). Further, as we show in Sec. 42.3.2, consideration of the CR
42_Singh_ch42_p42.1-42.18.indd 5 8/26/16 4:36 PM

ET and E0 rates (normalized by Ew)

4 Increasing b
3
E0
q1/λ Ew
1
q2/λ
ET
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Notional moisture availability, Θ (saturation fraction)
Figure 42.3 The CR hypothesis showing ET and E0 converging on Ew at high moisture availability,
diverging at low moisture availability. The depth equivalent of the energy released by decreasing moisture
availability limiting ET below Ew is indicated as q2/l, and that released to the overpassing air as q1l. b var-
ies from 1 (for symmetrical relationship, indicated by solid grey line for the resulting E0) to 5 (b from 2 to
5 indicated by dashed grey lines for resulting E0).
resolves the long-standing “evaporation paradox,” wherein Epan is observed to internationally (FAO-56) and for the United States (ASCE05); here we cite
decline around the globe in the face of a global ET rate that must be rising to their specific sections where applicable.
supply the increasing global Prcp and cloudiness (Brutsaert and Parlange, In these approaches, ET0 is derived using the PM formulation for ambient
1998). In an analysis of long-term trends in basin-budget derived ET (and meteorological and radiative conditions either measured over the cropped
therefore independent of the CR) in 655 basins representing 19.4% of the surface in question, at a nearby agrometeorological station, or, increasingly,
continental US (CONUS) by area, Hobbins (2004) found that the CR hypoth- derived from meteorological reanalyses. ET0 assumes a reference surface: a
esis explained trend directions in 92% of the basins, compared to only 57% healthy, well-fertilized, extensive crop of either short grass (0.12 m high) or
basins explained by the traditional paradigm. Finally, McMahon et al. (2013) alfalfa (0.5 m high), completely shading the ground, actively growing under
provides an excellent literature survey of the development and current state of well-watered conditions, of specific albedo (0.23), and with a stomatal resis-
the science and application of the CR in the estimation of multiday ET and tance rs of 70 s m−1 and aerodynamic resistance ra of 208/U2 s m−1 (U2 is 2-m
lake evaporation and in studies of hydroclimatic change. wind speed). It further assumes no advected energy from water or heat stor-
age effects. This physically based formulation retains the Penman (1948)-
42.2.4 Reference ET Concept based combination form but includes in its advective driver a parameterization
The most popular method to estimate ET from cropped surfaces (crop ET, of the fine-scale diffusive characteristics of the plants, replacing Penman’s
ETc) is via the concept of reference evapotranspiration (ET0). This concept (1948) linear approximation of the complex vapor transfer process by a “big
obviates direct knowledge of moisture availability at operational scales, asking leaf ” approximation. In this approximation, bulk resistances parameterize
instead “assuming well-defined, ideal surface conditions, what would be the moisture availability and vegetation roughness: water vapor overcomes first
moisture supply from the land surface to the atmosphere?” Answering this the rs to diffusion from the leaf stomates into the canopy, thereby capturing
question provides ET0, the theoretical ET for an idealized, well-watered refer- moisture availability, and then overcomes the ra from the canopy into the
ence surface—an upper limit constrained only by the advective and radiative overpassing air, thereby capturing vegetation roughness. Many smaller terms
drivers. Additional assumptions as to the prevailing soil- and vegetation- drop away from the radiative driver Qn (see Fig. 42.1), leaving Rn, Ln, and G,
moisture conditions and the vegetation mix and phenology are then applied which are variously observed, ignored at various time steps (e.g., G at a daily
to ET0 to account for the divergence from the well-defined reference condi- time step), or parameterized as functions of other variables [e.g., Ln as a func-
tions and thereby scale ET0 to ETc. This divergence may arise due to a variety tion of T, ea, and extraterrestrial solar radiation, itself a function of solar and
of factors: crop factors (e.g., type, variety, and development stage), manage- orbital geometry as described in Eq. (4.4.4) in Shuttleworth (1992)].
ment factors (e.g., management of disease, pests, and soil), and environmental The ASCE05 approach standardizes both ET0 estimation and the transfer-
factors (e.g., salinity and water stress). The ETc so derived is primarily used in ability of crop coefficients for agriculture and landscape use. Its general
irrigation scheduling and in water demand models. ETc may be met by water expression is
stored in the soil profile, Prcp and, when ETc exceeds available soil water, by
irrigation scheduled to meet the deficit, so that ET reaches ETc. Beyond these C
0.408∆ 86400 γ n e −e
traditional uses, ET0 (the starting estimate for ETc) is now used in combina- T
ET0 = ( Rn + Ln − G) + U 2 sat 3 a
tion with remotely sensed vegetation indices and land-surface temperature ∆ + γ (1 + CdU 2 ) 106 ∆ + γ (1 + CdU 2 ) 10
(LST) to estimate actual ET from crops and natural landscapes.
(42.12)
In this section, we summarize the main points of the concept of ET0 and its
use to derive ETc so as to raise the issues that need to be addressed by users.
We do not attempt to detail the procedures, as this is well covered in the where ET0 is the standardized reference ET (mm day–1) for 0.12-m grass (in
seminal material—particularly the U.N. Food and Agriculture Organization which case it is commonly abbreviated as ETos) or 0.5-m alfalfa (commonly
FAO-56 standard (FAO-56; Allen et al., 1998) and the American Society of ETrs) crop surfaces. The 0.408 coefficient (m2 mm MJ−1) represents the
Civil Engineers (ASCE) Environmental Water Resources Institute (EWRI) inverse of the latent heat of vaporization; Δ and γ are in the same units (com-
Standardized Reference ET Equation report (ASCE05; Allen et al., 2005). The monly Pa K–1); the 86,400/106 term converts Rn, Ln, and G from W m–2 to
intent of these reports is to resolve uncertainties arising from the variety of MJ m−2 day−1; the 1/103 term converts the last term (in ea and esat ) from Pa to
ET0 model inputs and applications and to codify approaches to ET0 estimation kPa. Values of the “numerator constant” Cn (here in K mm s3 Mg−1 day−1 for
42_Singh_ch42_p42.1-42.18.indd 6 8/26/16 4:36 PM

daily time steps) and the “denominator constant” Cd (s m−1) specific to time thermal radiance and optical reflectance, incoming solar radiation, vegetation
step and the choice of grass or alfalfa, as well as the derivation of all other indices, and other surface parameters (Allen et al., 2007; Anderson et al., 2012).
variables, are detailed in Allen et al. (2005). The FAO-56 equation, which Net radiation is estimated from incoming longwave and shortwave radiation,
yields identical results on a daily time step, is shown as Eq. (6) in Allen et al. while satellite-derived estimates of surface temperature, emissivity, and surface
(1998). As we discuss in Secs. 42.3.4 and 42.3.6, simpler formulations for ET0 albedo are used to calculate outgoing longwave and shortwave radiation.
(e.g., Blaney–Criddle and Hargreaves) are no longer recommended for use. Ground heat flux is commonly estimated using satellite-derived net radiation,
To estimate ETc from ET0 various user-selected coefficients that vary from surface temperature, and vegetation indices. In the widely used Mapping
~0 to 1.4 are applied to account for the difference between the idealized refer- EvapoTranspiration at high Resolution using Internalized Calibration
ence condition (i.e., ETos or ETrs) and field crops under “standard conditions” (METRIC) model, sensible heat flux is estimated with LST and the Calibration
(i.e., free from environmental or water stresses) must be accounted for. These using Inverse Modeling at Extreme Conditions (CIMEC) procedure (Allen
differences may arise due to variations in field crop height (which affects ra et al., 2007). Once Rn, Ln, G, and H are estimated, lET is calculated as a resid-
and vapor transfer), canopy-soil albedo (which affects Qn), the stomatal resis- ual of the energy balance (Eq. 42.4). The instantaneous ET rate at satellite
tance rs, and evaporation from soil; they are accounted for using the “single overpass time (ETinst) is calculated by dividing lET by l. ETinst is then used to
crop coefficient” approach, as follows: compute ET0F per satellite image pixel, where estimates of ET0 at the time of
ETc = K c ET0 (42.13) the satellite overpass are derived from a local weather station or gridded

weather dataset. The ET0F is analogous to an instantaneous crop coefficient.
where Kc is the crop coefficient (dimensionless), which varies with crop type, ET for the 24-h period (ET24) for day i is typically estimated as:
its location, and development stage (see Fig. 42.4). The calibration of Kc values ET24i = ET0 Fi × ET0_ 24i (42.16)

for particular crop types, regions, and development stages is the subject of a
vast literature, but representative values are summarized in Chapter 6 of the where ET0_24i is the 24-h ET0 for day i; and ET0Fi is the fraction of ET0 for the
the FAO-56 report (Allen et al., 1998) and in an earlier FAO report known as ith day. Time integration of ET over an extended time period is estimated
FAO-24 (Doorenbos and Pruitt, 1977). from:
For standard conditions but when the proportion of evaporation from bare m
ET = ∑ ET0 Fi × ET0_ 24i (42.17)

soil may be significant, such as for capturing specific wetting events (irriga- i =n
tion or Prcp), the “dual crop coefficient” approach is used: where n and m are the first and last days of the desired time period, respec-
ETc = ( K cb + K e )ET0 (42.14) tively.
In irrigated and natural, well-watered environments (i.e., riparian and
where Kcb is the basal crop coefficient, which accounts for transpiration from wetland areas) the use of a PM -based ET0 such as ETrs is recommended for
the crop; and Ke is the soil evaporation coefficient (both Kcb and Kc are dimen- time integration due to the ability of ET0 to capture the potential effects of
sionless). The estimation of ETc using a dual crop coefficient approach is advection on ET.
covered in detail in Chap. 7 of the FAO-56 report (Allen et al., 1998). For vegetated surfaces where no net transfer of advected energy exists, such
For crops grown under nonstandard conditions—whether due to soil- as deserts and rangelands, an evaporative fraction (EF) approach is recom-
water stress or salinity, nontypical or nonpristine vegetation types, manage- mended for estimating ET24 and time integration similar to that employed in
the Surface Energy Balance for Land (SEBAL, Bastiaanssen et al., 1998). In
Kc 1.4 this case, the instantaneous EF is defined as:
Kcmid
λ ETinst
1.2 EFinst = (42.18)
( Rn + Ln − G )inst
1 where lETinst and (Rn + Ln - G)inst are estimated through the satellite energy
balance model; and 24-h ET (ET24i) is estimated as:
0.8
ET24i = EFinst ( Rn + Ln − G)24i (42.19)
0.6 Daily time series of per-pixel ET0F and EF are commonly estimated between
Kcini image-acquisition dates through linear or spline interpolation of ET0Finst and
0.4 Kcend EFinst values. Time-interpolated ET0Finst and EFinst values are multiplied by
daily ET0 or empirical 24-h average Rn + Ln and G estimates (Brunt, 1932; de
0.2 Bruin, 1987) to estimate per-pixel daily ET time series.
Empirical models for estimating ET and ET0F are commonly based on
0 satellite-derived vegetation indices. Heilman et al. (1982) proposed a linear
Time (days) relationship between a vegetation index and ET0F, which has been supported
by various other studies (Tasumi et al., 2005; Calera-Belmonte et al., 2005). If
Initial Crop development Mid season Late season no local calibration data exists, Allen et al. (2011a) suggest that ET0F can be
generally estimated as:
Figure 42.4 The development of Kc through the growing season for a notional crop, ETrs F = 1.25 NDVI (42.20)
showing crop growth stages. [Source: http://www.fao.org/nr/water/cropinfo_sugarcane. or
html]
ETos F = 1.25 NDVI + 0.2 (42.21)
ment practices, or simply during the non-growing period—further coefficients where ETrsF and ETosF are the relative fractions of alfalfa and grass reference ET,
apply to account for deviations from the standard, well-watered conditions. respectively. NDVI is the normalized difference vegetation index, defined as:
For example, under soil water-stressed conditions, Kc in Eq. (42.13) and Kcb in
Eq. (42.14) are multiplied by the water stress coefficient Ks, which ranges from NDVI = ( ρNIR − ρRed ) ( ρNIR + ρRed ) (42.22)
0 for zero water availability to 1 for fully irrigated conditions. ETc-estimation
procedures for nonstandard conditions are detailed in Part C of the FAO-56 where ρ is the at-surface or at-satellite (i.e., top of atmosphere) reflectance; NIR
report (Allen et al., 1998). is near-infrared waveband from ~0.7 to 1.3 μm; and Red is the visible waveband
42.2.5 Satellite Remote Sensing of ET from ~0.60 to 0.70 μm. Due to differences in satellite wave bandwidths and the
use of at-surface or at-satellite reflectances for computing NDVI, Eqs. (42.20)
Satellite remote sensing is commonly used to estimate ET from agricultural and (42.21) are general relationships and should be used with caution, espe-
areas by scaling grass or alfalfa standardized reference ET0 as: cially in areas of low vegetation cover, or where bare soil evaporation is a sig-
ET = ET0 F × ET0 (42.15) nificant component of ET.
where ET0 is the reference ET; and ET0F is the fraction of the respective ET0. 42.2.6 Eddy Correlation
ET0F is typically estimated from remotely sensed vegetation indices or a LST- Up to tens of meters above the land or water surface, vertical transport is
based surface energy balance (SEB) model. Most SEB models estimate mostly driven by turbulence. Turbulent eddies near the surface cause devia-
required variables from a series of calculations involving weather data, tions about the mean values of vertical wind w’, specific humidity q’, and
42_Singh_ch42_p42.1-42.18.indd 7 8/26/16 4:36 PM

temperature θ’. On average, there are positive correlations between high verti- sent the sum of surface and subsurface runoff fluxes from all areas within the
cal wind and humidity deviations, negative correlations between low vertical basin upstream of the gage (surface-water outflow less surface-water inflow in Fig.
wind and humidity deviations. By exploiting these correlations, estimates of 42.5). Streamflow is generally assumed to be one of the most accurate hydrologic
ET can be made through direct measurements of the deviations in vertical measurements, but any differences between the stream gage contributing area and
wind speed and specific humidity at frequencies of 5–20 Hz, and computing the basin boundaries must be accounted for. GW is the surface water transferred
the cross-correlation between q’ and w’ for averaging periods that usually to regional groundwater and is unrecorded at the streamgage; it is generally
range from 15 min to 1 h depending on surface conditions and the surround- ignored (and difficult to quantify), but can be a significant unmeasured quantity
ing environment. The integrated product of the means of instantaneous depending on the system (groundwater outflow less groundwater inflow in Fig.
deviations q’ and w’ results in the expression of ET from advective vapor 42.5). D is the net diversion of water into the basin from pipelines and canals;
transport by turbulence (Swinbank, 1951) as: depending on the basin, this may be zero or it may contribute a large portion of
the water balance (imported water less exported water in Fig. 42.5). ΔS is the net
ET = ρa q′w ′ (42.23) basin-wide increase in water storage within the basin across the time step. This

where ra is the density of air; q’ and w’ are, respectively, deviations of specific storage may be in lakes (or reservoirs), rivers, snowpack, and both the saturated
humidity and vertical wind speed about their mean values; and the overbar and unsaturated soil zones (depth below and height above the water table in Fig.
denotes averages taken over the time period considered. Sensible heat flux 42.5, respectively), and it may be a function of the hydrologic, climatic, geologic,
may also be computed following the same theory as: and pedologic characteristics of the basin. This term is generally ignored at the
annual time step, but will generally be significant at subannual time scales. The
H = ρac p θ ′w ′ (42.24) concept of water years (in the United States, they are defined as starting October
1 of the previous year) is used to minimize the effects of interannual ΔS.
where cp is the specific heat of air. Measurements of w’, q’, and θ’ require spe- Evaporation from a water body E can be estimated through the water bal-
cialized instruments but are widely available and used in practice, such as a ance, as follows:
three-dimensional sonic anemometer, high-frequency hygrometers and gas
analyzers, and fine wire thermocouples to measure w’, q’, and θ’, respectively. E = Prcp + SWin + GWin - SWout - GWout - B - ΔS (42.27)
Numerous corrections are required to account for temperature and humidity where Prcp is precipitation onto the water surface; SWin and GWin are surface
effects on air density and eddy structure, instrument separation, frequency and groundwater inflows, respectively; SWout and GWout are surface and
response, coordinate rotation, and hygrometer type (Fuehrer and Friehe, groundwater outflows, respectively; B is bank storage; and ∆S here is the
2002). Net radiometers, ground heat flux plates, and soil moisture and tem- change in water body storage. 
perature sensors are typically collocated with EC instrumentation to evaluate The uncertainty in water balance E and ET estimates is a summative func-
the land-surface energy imbalance (see Sec. 42.3.5). For details on the EC tion of the uncertainties in the input variables, so the fewer input variables the
theory, see Foken (2008a) and Brutsaert (2005), and for EC networks, such as better. The process is simpler for basins that are minimally disturbed, that is,
FLUXNET and EUROFLUX, see Baldocchi et al. (2001) and Aubinet et al. with minimal interyear, in-basin storage and no trans-basin water transfers
(2000). For detailed discussions on measurement accuracy, and recommend- [e.g., the basins of the Hydro-Climatic Data Network of Slack and Landwehr
ed reporting requirements and documentation, see Allen et al. (2011a; 2011b). (1992)]. Then further assuming no net groundwater losses (i.e., GW = 0) and
that the gage is sited at the natural outlet, it is possible to reduce the water bal-
42.2.7 Scintillometry ance down to ET = Prcp – Runoff. The primary drawback of the water budget
The scintillometry technique estimates H by relating scattered electromag- approach is the time step required to reduce uncertainties to an acceptable
netic radiation due to atmospheric turbulence to the structure parameter of level and that, in the case of basin budgets, the ET so derived is a lumped value
the refractive index of air, Cn2 (Wang et al. 1978). Scintillometry-estimated H across the spatial extent of the basin. Often the primary use of the mass bal-
(Hsc) is obtained from: ance approach is to provide a physical ET (or E) estimate against which other
H sc = ρac pT *u* (42.25) estimation procedures may be calibrated (such as for groundwater recharge).

where T * is the temperature scale; and u* is the friction velocity. T * can be 42.2.9 Open-Water Evaporation
derived relying on MOST relationships and the structure parameter of tem- Open-water evaporation E is one of the most difficult surface energy fluxes to
perature following Wyngaard et al. (1971). Several variants for estimating T * quantify and is rarely directly measured in the natural environment.
are commonly used depending on the study purpose and design. For more Common indirect techniques include pan evaporation measurements com-
detailed background and recent developments in scintillometry see Moene bined with application of pan coefficients, water budget, energy budget, mass
et al. (2009). Recent novel applications of scintillometry include the following: transfer, and a combination of energy and mass-transfer techniques. The EC
estimation of open-water evaporation (McGloin et al., 2014; McJannet et al., technique is considered the most accurate direct approach, given ideal envi-
2011); comparison to EC measurements of H and ET (Hoedjes et al., 2007; ronmental conditions, water body physical setting, and experimental design
Liu et al., 2013); validation of remotely sensed ET (Marx et al., 2008); valida- (Brutsaert, 2005). The water budget technique is considered the most accu-
tion of catchment-scale hydrologic modeling (Samain et al., 2011); and devel- rate indirect approach, especially in arid environments, where evaporation is
opment of a scintillometer network designed to support remote sensing, a relatively large water budget component (Morton, 1994). Primary factors
hydrologic, and meteorological models (Kleissl et al., 2009). The main advan- governing E include net radiation, heat storage, air temperature, water surface
tage of scintillometry is that it can be used to estimate aggregated H fluxes “skin” temperature, humidity, wind speed, stability of the atmosphere, advec-
over different scales, and can therefore be used to estimate aggregated ET if tion of water and heat into and out of the water body, and salinity.
reliable available energy fluxes are acquired over the same scales. Additionally, the aerodynamics of the water surface, turbidity of the water,
and inflow and outflow rates control the rate of transfer between energy bal-
ance variables. All of these factors are important to consider when deciding
42.2.8 Water Balance Estimates of ET
which technique is most appropriate for application. In most cases, the lack of
Mass balance ET may be estimated for a basin or a water body by taking adequate hydrologic and meteorological data limits the application of physi-
advantage of the law of conservation of mass across a naturally defined con- cally robust techniques for estimating open-water evaporation.
trol volume. “Basin-budget” or “water balance” ET is estimated as a residual
of the water budget—that is, the expression of all fluxes and state-changes (ET, Pan Evaporation
Prcp, groundwater recharge, surface and subsurface runoff, diversions, and The most common operational technique for estimating E is the pan evapora-
storage changes) estimated or observed—across the spatial extent of a basin tion approach:
or control volume of a water body (see Sec. 42.2.9) at a tractable time step E = Kp Epan (42.28)
(time and space scales for which fluxes and storage changes are quantifiable).
For a basin, the expression of the natural water balance that yields the where Kp is a pan coefficient (dimensionless); and Epan is measured pan evapo-
estimate of ET (shown in Fig. 42.5 as the sum of E from water surfaces and ET ration, typically from a standardized pan such as the U.S. Weather Bureau Class
from vegetation and bare soil, as well as the sublimation from snowpack that A pan (this is the standard in the United States; China has a micropan standard).
is not shown) is as follows: Epan measurements reflect the combined effects of net radiation, air and water
temperature, humidity, and wind speed on a point-scale reference open water
ET = Prcp - Runoff - GW + D - ΔS (42.26) surface. Kp adjusts Epan to more accurately represent the physical conditions of
where all fluxes and storage changes are in units of time-rate changes of depth the water body of interest and to account for the fact that the pan has limited
(e.g., mm day-1) and in the directions specified. Prcp is spatially averaged over the heat storage potential and is exposed to the air on its sides and bottom, allowing
basin. Runoff is the streamflow recorded by streamgage, and is assumed to repre- for heat exchange and thus affecting the energy balance. Annual Kp values
42_Singh_ch42_p42.1-42.18.indd 8 8/26/16 4:36 PM

Figure 42.5 Components of a water balance for part of a watershed, indicating inputs, outputs, and storages. [Source: Healy et al. (2007). Used with permission.]
typically range from 0.65 to 0.85, while monthly Kp values can range from 0.3 to many issues including the need to estimate net advected energy to the water
1.7 depending on water body characteristics, such as depth, turbidity, and body, energy advected from evaporating water, energy exchange from bottom
potential for heat storage. In addition to considering potential effects of heat sediments, and–most difficult to estimate–the change in heat storage.
storage for the water body, appropriate Kp values depend on pan type, ground Solar radiation penetrating and absorbed by the water body affects the mag-
cover, evaporative conditions within the fetch of the pan, presence of aquatic nitude and timing of E. Energy storage depends on climate and physical char-
plants, and salinity of the water body. While Epan measurements are still widely acteristics of the water body, such as volume, depth, geometry, clarity, and
used, they are very susceptible to microclimatic conditions surrounding the pan surrounding environment. For shallow and turbid water bodies, energy storage
(including freezing, which limits their operations in time and space) and subject effects on E are small compared to deep, clear water bodies. Stored energy can
to the care of station maintenance (Thom et al., 1981). Further, some limitations be partitioned into E, heating of the air, emission of longwave radiation, or
apply as to the appropriate minimum time scale of observations (Roderick et al., advection of heat in the discharging water. Most energy storage occurs during
2009a). Ideally, the pan technique relies on locally calibrated monthly and spring and summer months and is not readily available for E until conditions
annual Kp values. For a review of considerations and techniques for selecting Kp are such that it is transferred to the surface by conduction or convection, at
values, see Grayson et al. (1996) and Alvarez et al. (2007). which point the water surface temperature may exceed the air temperature,
thereby causing a large portion of energy to be partitioned from stored heat into
Water Budget sensible heat or longwave emission rather than E, reducing the total evapora-
Evaporation can be expressed as a residual of the water budget volume or tion from the water body. The change in energy storage is typically estimated
depth per unit time following the continuity equation for a generalized water through repeated thermal profile measurements of the water column; smaller
body as shown in Sec. 42.2.8. Estimating open-water evaporation as a residu- energy budget periods require more frequent thermal profile measurements.
al of the water budget is sometimes considered the most accurate approach For water bodies with large depth fluctuations and thermal stratification,
for large water bodies in arid environments due to the fact evaporation is a heat-storage estimates can be obtained using discrete layer volumes and
large component of the water budget in such cases, and errors in estimated respective average temperatures to compute the integrated energy storage
terms are small and insignificant relative to E. changes for each layer (Saur and Anderson, 1956; Moreo and Swancar, 2013).
Since both H and lET are unknown in the energy budget equation they are
Energy Budget expressed in the Bowen ratio [Bowen, 1926; Eq. (42.1) in Sec. 42.1.4], where Ts
This approach has been most widely used in research, and it is likely the most (the water surface “skin” temperature, °C) is preferably measured with a ther-
data-intensive approach for estimating evaporation due to the need to con- mal infrared sensor, and T is measured at a reference height usually 1–3 m
sider the entire water body as a control volume rather than as just the surface above the water surface [Eq. (42.4)]. b is also commonly estimated over water
in the case of a land surface energy balance. This requirement brings about at two reference heights above the water surface with automatic sensor
42_Singh_ch42_p42.1-42.18.indd 9 8/26/16 4:36 PM

exchange mechanisms, which reduces sensor bias and drift effects on estimated only limited application unless calibrated. Morton (1983b) developed a
T and e gradients. Substituting H = blET and Qw into the energy balance numerical energy-aerodynamic approach by iteratively solving energy bal-
equation (Eq. 42.5) yields the commonly used EBBR expression of open-water ance and vapor transfer equations to estimate ambient Ep, which is used to
evaporation: obtain an estimate of the equilibrium wet environment surface temperature
Rn + Ln + Qv − Qx at which to estimate the slope of the saturation vapor pressure curve (∆) and
E= (42.29) Priestley–Taylor potential evaporation for a spatially extensive wet surface
ρw [λ (1 + β ) + c w (Ts − Tb )] (i.e., Ew). The relationship between Ep and E is generally complementary
(discussed in Sec. 42.2.3); hence, the model is called the Complementary
where E is the evaporation rate (m s−1); Rn, Ln, Qv, and Qx are in W m–2; cw is Relationship Lake Evaporation (CRLE) model. The modified Priestley–
the volumetric heat capacity of water (4186 J kg−1 °C−1 at 15 °C); ρ is the den- Taylor equation (Morton, 1983b) used in the CRLE is
sity of water (1000 kg m−3); Ts is the water surface temperature (°C); and Tb is
a base temperature usually assumed to be 0°C for convenience. Extended ∆w ( Rw − Qx )
discussions on the EBBR for open water can be found in Winter et al. (2003) E = b1 + b2 (42.32)
and Brutsaert (2005). There is some debate as to whether the EBBR technique ∆w + γ
for estimating E is useful operationally or for secular analysis: some view it as
one of the more accurate techniques for long-term monitoring; others point where Rw is the net radiation for a water surface; b1 and b2 are calibration
to many limitations regarding the practicality of accurately estimating the constants [Morton (1983b) suggests 13 W m−2 and 1.12, respectively]; and Δw
most sensitive variables (Morton, 1994). More accurate instrumentation and is the wet-environment slope of the saturation vapor pressure curve (kPa °C−1)
advanced automated measurement techniques have become widely used in computed with the wet-environment surface temperature. The energy storage
the last few decades, making research applications of the EBBR especially term (Qx) is solved for using an approach outlined by Morton (1983b) and
useful for understanding evaporation processes, developing semiempirical Morton (1986) that simulates the hypothetical, linear, heat storage by lagging
approaches, and comparing and combining energy fluxes with those obtained absorbed shortwave radiation. Further details on the combination equation
by other techniques such as EC and aerodynamic mass transfer techniques are in Brutsaert (2005). For background and development of the CRLE, see
(Allen and Tasumi, 2005; Moreo and Swancar, 2013). Morton (1983b; 1986), with recent applications by Jones et al. (2001) and
Huntington et al. (2015).
Aerodynamic Mass Transfer
The aerodynamic mass-transfer technique is one of the oldest and most com- 42.3 OUTSTANDING PROBLEMS AND DIRECTIONS FOR
mon techniques for operational estimation of E due to its simplicity and relative FUTURE WORK
accuracy. A detailed evaluation and sensitivity analysis of commonly used
In this section, we examine ongoing research themes and outline directions
aerodynamic mass transfer equations is given by Singh and Xu (1997a; 1997b).
for future work, including advances in the data availability for ET and E0 and
More recent applications of the aerodynamic mass-transfer approach for esti-
the science of their estimation, with the aim of closing the gap between the
mating evaporation rely on turbulent transport theory for estimating vapor flux
well-established science and its practice in operational settings that so
where
often plagues the study and application of ET and E0. In the following sections,
λ ET= λρaCEu1 (qs* − q2 ) (42.30) we examine an often-obscured practice still rooted in the past (T-based

parameterizations of E0), and one that is being dragged forwards into the
science (representation of ET and E0 in drought monitoring).
and CE is the Dalton number, a bulk transfer coefficient of water vapor; qs* is
the specific humidity at the water surface temperature (kg kg−1); q2 is the 42.3.1 Reanalyses
specific humidity of air at reference level z2 (kg kg−1); and u1 is wind speed at
reference level z1 (m s−1). The coefficient CE is only constant for neutral atmo- In the hydrological context, reanalyses are spatially and temporally compre-
spheric conditions and constant surface roughness. CE for neutral conditions hensive datasets of meteorological, radiative, land-surface, and atmosphere
can be expressed for a water surface following Brutsaert (2005) from: fluxes and states describing the earth-atmosphere system. They are used in
long-term (generally several decades), large-scale (continental to global)
k2 analyses of weather and climate. Observations of fluxes and states at the earth’s
CE = , (42.31) surface and various levels in the atmosphere are objectively combined in a
ln( z 2 / z 0m )ln( z1 / z 0h ) numerical model to synthesize the overall state of the earth-atmosphere sys-
tem (at several levels from the surface to above the stratosphere). Reanalyses
where k is the von Kármán constant (0.41); and z2 and z1 are measurement are used to characterize and compare current and past climate, to assist in
heights of wind speed and humidity (m), respectively; and z0m and z0h are predictions and near-real-time monitoring (e.g., as in drought) across sectors
roughness lengths of momentum and humidity (m), respectively. For lakes of (drought monitoring, agriculture, water resource management, energy, and
medium size with fetches of 1–10 km CE values of ranging from 0.0011 to insurance) in climatology, and in providing climate services. A significant
0.0013 can be used as a first approximation (Brutsaert, 2005; Tasumi, 2005). example of a reanalysis dataset is the North American Land Data Assimilation
For stable and unstable conditions, stability correction factors must be con- System (NLDAS; Mitchell et al., 2004), which is commonly used—for example,
sidered. Allen and Tasumi (2005) used Bowen ratio exchange arm measure- to verify ET-based drought monitoring tools by Anderson et al. (2013).
ments of q at two different heights in Eq. (42.30), and results compared Figure 42.6 demonstrates a reanalysis dataset (the CRU CL 2.0 Global
favorably to EC estimates of E. Climate Dataset) using as examples maps of global mean annual ET and evapo-
A common challenge in applying the aerodynamic mass transfer approach rative demand (using ET0). The two fluxes are plotted to the same range so as
is obtaining accurate weather data for q and u representative of water surface to demonstrate how ET and ET0 vary and interrelate across the water- and
conditions. An advantage of the approach is that the required variables are energy-limited spectrum. The spatial pattern of ET reflects the regional avail-
fairly easily obtainable using floating rafts or buoys, and can be used in an abilities of both water and energy: ET is lowest in dry regions, both hot and
automated, near-real-time fashion to compute operational estimates of E. cold, and is highest in warm, wet regions. It follows energy availability (mostly
latitudinal) when water is available, and water availability otherwise. The spatial
Combination Methods pattern of ET0, which assumes ample water availability, reflects both energy
Combination energy-aerodynamic mass transfer methods are commonly availability and its complementarity with ET in water-limited regions. As an
used for land applications (e.g., see Sec. 42.2.4) and are typically based on example of this interrelation, note that in the water-limited Saharan, Arabian,
Penman (1948; 1956) and Penman-Monteith (Monteith, 1965) formulations. and Australian deserts, ET is very low whereas ET0 (the water that would
When applied to water bodies, the same limitations of the energy balance evaporate given such dry atmospheric conditions) is very high; contrast that
approach apply to the combination approach in that estimates of Rn, Ln, Qx, with the energy-limited areas (e.g., Indonesia and the basins of Amazon and
and Qv are required, as are representative open-water estimates of es, ea, and Congo rivers), where the availability of water is high and ET and ET0 rates
Uz. Available energy is often used for open-water evaporation estimates converge, with ET higher than its global mean and ET0 lower than at similar
through the use of the Priestley–Taylor equation. Because of the absence of latitudes.
an aerodynamic term in the Priestley–Taylor equation (Priestley and Taylor,
1972) it is considered to represent equilibrium evaporation (i.e., Ew), some- 42.3.2 Trends in ET and E0
thing that rarely occurs from water bodies surrounded by water-limited Understanding the impacts of climate change on ET and E0 is crucial for the
environments. For this reason, the use of the Priestley–Taylor equation has operational implications for future regional water budgets and, from a
42_Singh_ch42_p42.1-42.18.indd 10 8/26/16 4:36 PM

Outstanding Problems and Directions for Future Work 42-11
Figure 42.6 Mean annual (top) ET and (bottom) ET0 across the globe (1961–1990), as determined from the CRU CL
2.0 Global Climate Dataset (University of East Anglia, United Kingdom) reanalysis dataset. ET is estimated using a
simple 1-dimensional water balance model, ET0 is estimated using the Penman Monteith formulation of the FAO-56.
[Source: FAO. Used with permission.]
research perspective, to understand changing E0 as indicative of lower atmo- than Prcp [from increased ocean evaporation and in places, on land, increased
sphere changes and to provide physical evidence-based global climate model Prcp and/or decreased ET (Labat et al. 2004)], what is suppressing terrestrial
(GCM) constraints. Here we summarize ET and E0 trends, identify their driv- ET? An exception to these declining terrestrial Prcp and ET is North America.
ers, and outline primary research questions. Observations over CONUS show that Prcp is increasing faster than Runoff; the
implied acceleration in ET is supported by observations of basin-budget-
Trends in ET derived ET over the last half of the twentieth century (Milly and Dunne 2001;
Observations of hydrologic storages and fluxes (including Prcp and ET) and Brutsaert, 2006; Huntington, 2006) and by ET modeling using CR models
simulations by GCMs (Huntington, 2006) all point to an ongoing and future (Szilagyi, 2001; Hobbins, 2004) and LSMs. Trends vary with model- and
intensification of the global hydrologic cycle under warming [e.g., Jung et al. observation-type, period, and region, but generally do not exceed 2 mm year−2.
(2010) note an increase in globally averaged ET of +0.71±0.1 mm year−2 from
1982 to 1997], while atmospheric water vapor’s residence time (the atmo- Trends in E0
spheric reservoir) is increasing as atmospheric circulation weakens
Long-standing, worldwide observations of Epan yield the most widely distrib-
(Bosilovich et al., 2005; Held and Soden, 2006). Prcp affects ET rates through
uted set of E0 trend analyses and have revealed phenomena central both to E0
its direct effects on water availability: in water-limited hydroclimates, ET will
and to its drivers. Epan trend studies are notable for their consensus: Roderick
vary directly with Prcp, whereas in energy-limited hydroclimates the effects
et al. (2009a; 2009b) and Fu et al. (2009) summarize studies showing declin-
of Prcp on ET will be through indirect effects such as changes in cloudiness,
ing Epan almost globally over the last 30–50 years, with a conservatively esti-
albedo, and vegetation effects. Long-term global mean Prcp and ET must
mated overall trend of about -2 mm year−2, equivalent to a total decline in
balance, but there are spatial distinctions: oceanic Prcp and ET are both
radiative forcing of about 4.8 W m−2 over a 30-year period, or an order of
increasing; but terrestrial Prcp (mainly in the tropics) and ET are both
magnitude higher than the 0.02 W m−2 year−1 trend in the top-of-atmosphere
(mostly) declining. Terrestrially, wet regions are getting wetter and dry
(TOA) imbalance (Hansen et al., 2005).
regions are getting drier, while an increasingly variable ET–Prcp residual
increases the intensity of both floods and droughts. There are also temporal
nuances in the hydrologic intensification: Jung et al. (2010) note no hydro- Drivers of trends in ET and E0
logic intensification after 1997, and suggest that soil moisture limitation in ET and E0 trends are driven by trends in their hydroclimatic drivers and relate
the southern hemisphere drives global ET to decline (though nonsignifi- to each other within the context of the CR (see Sec. 42.2.3). Both ET and E0
cantly) since then, with other possible mechanisms, such as stomatal closure are subject therefore to trends in T, humidity, Uz, and radiation, with ET
under increased CO2, land-use change, and decreasing Uz (stilling, see later). further subject to water availability trends. Of all driver trends, those in T are
Other change drivers include Θ, seasonality, land cover changes, and con- the most easily estimated: all else equal, rising T raises E0 through an increase
sumptive use. in VPD, which increases ET if water is available. But all else is seldom equal
Few influential studies estimate global terrestrial ET [interested readers and the effects of T-trends are often overshadowed by other drivers’ trends
should turn to Wang and Dickinson’s (2012) comprehensive review of observa- (e.g., Roderick et al., 2007). Indeed, relying on secular analyses of T alone has
tion, modeling, and variabilities], thus many vital research questions remain. If led to oversimplistic, incorrect conclusions about E0 and ET trends and to
a limit to the energy- and T-driven acceleration of the global hydrologic cycle discussion of an “Evaporation Paradox,” which has also been cited in observa-
has indeed been reached, will this lead to productivity and terrestrial carbon tions that were once difficult to explain: ET and E0 trending in opposite direc-
sink declines and so to a growing surface energy budget increasingly partitions in some regions, but parallel in others; decreasing Epan in the face of
tioned in favor of H, thereby increasing land-surface warming and intensifying increasing Prcp (e.g., Cong et al., 2009), cloudiness, GCM-derived ET (e.g.,
regional land-atmosphere feedbacks? As global Runoff has increased more Brutsaert and Parlange, 1998), and T (e.g., Roderick and Farquhar, 2002).
42_Singh_ch42_p42.1-42.18.indd 11 8/26/16 4:36 PM

Table 42.1 Directions of Long-Term Trends in Annual ET, Qn, and EA for Stilling
42 Years across 655 Hydrologically Undisturbed Basins across CONUS.
Possible Combinations Represent Whether the Combined Trends in Qn Globally, observational studies—for example, in Australia, China, Italy, the
and EA Explain the ET Trend Direction Within the Context of (CR) the CR United States, Canada, and the Tibetan Plateau [see Roderick et al. (2009b)
Hypothesis, or (Trad) a Traditional Relationship Between ET and E0, for a bibliography]—show declining land-surface Uz, or “stilling.” Decreasing
Respectively: � = Explained,  = Not Explained. [Source: Hobbins (2004)] Uz leads to decreasing eddy diffusivity and decreasing turbulence, which, all
else equal, decrease E0 and reflect a complementary ET increase. Possible
mechanisms for stilling are: the oasis effect, whereby irrigation affects local
Physically possible
Trend directions
combinations?
wind circulation in both magnitude and direction; irrigation development
% of basins in resulting in increased surface roughness over large distances; Uz instrumenta-
combination CR Trad tion affected by local shielding, either by crop growth or by trees planted
EA Qn ET complementary traditional specifically to reduce Uz over crops; climatological factors acting on a regional
relationship paradigm or supra-regional scale unrelated to local ET changes. However, the ubiquity
0.3 + + +   of observations of declining Uz argues against stilling being either solely an
observational quality problem or an isolated, regional one. Neither reanalyses
2.8 - + +   nor gradients derived from daily Pa surfaces have proved useful in attributing
Uz trends to large atmospheric circulation changes, suggesting that daily Uz is
7.6 + - + ×  not dominated by large-scale atmospheric patterns. Clearly, more research
into identifying and quantifying the drivers of declining Uz and its effects on
43.2 - - +  × E0 and ET is necessary.
0 + + -  × 42.3.3 ET, E0, and Drought Monitoring and Forecasting
0.5 - + - ×  Evaporative dynamics are generally not well represented in current opera-
tional drought monitoring, due to ET being poorly parameterized, whether
2.4 + - -   independently of E0 or derived from E0 by LSMs. Of the direct use of E0 itself
as an indicator of the state of the lower atmosphere, there is no representation.
43.2 - - -   When E0 is used, it is as a driver of ET and is mostly derived from T (see Sec.
42.3.4) for reasons of convenience and operational inertia. The state of the
lower atmosphere in drought monitors is instead represented simply by Prcp
and T; typical of this operational underparameterization of drought character-
This paradox has been resolved as the interrelations of E0 and ET and of istics is the Palmer Drought Severity Index (PDSI; Palmer, 1965) that is used
their drivers have become better understood. Brutsaert and Parlange (1998) widely across the world and underpins much of the U.S. Drought Monitor
propose the CR to explain opposite ET and Epan trends in water-limited envi- (USDM), the foremost drought-monitoring tool across CONUS. Many authors
ronments. Roderick and Farquhar (2002) further explain that, in non–water- have described the shortcomings of the E0 parameterization in the PDSI (e.g.,
limiting environments, Epan and ET may decline together due to decreasing Hobbins et al., 2008; Sheffield et al., 2012). Basically, it relies on an oversimpli-
Qn (dimming, see later). Hobbins et al. (2004) show that both relations apply: fied relationship between E0 and T that cannot capture physical dynamics in
(i) ET and Epan are complementary under constant radiative component (Qn) energy-limited areas, or areas where advection is crucial to E0 variability (e.g.,
due to variations in the advective component (EA); and (ii) under declining the desert Southwest CONUS). In long-term drought-trend analyses it forces
(or increasing) Qn, both Epan and ET decline (or increase) together without more droughtiness than is observed in reanalyses (simply, due to rising T forcing
contravening their complementarity. In examining whether the directions of rising E0, whereas E0 has been shown to be falling in general).
trends in independently derived ET (i.e., through a water budget) are physi- There is a move in drought monitoring to include ET-based products,
cally resolvable with those of its components Qn and EA, Hobbins (2004) which would represent a significant narrowing of the science/operations gap.
found that the CR explained the directions of over 90% of ET trends from 655 These are based on a physical understanding of ET processes and recent
basins ( in the CR column in Table 42.1), including 43% whose ET trend remote-sensing products (Sec. 42.2.5). While there is as yet no inclusion of
directions would have been described as paradoxical under the traditional E0-based products, the current near-real-time availability of land-based,
paradigm ( in the Trad column in Table 42.1), wherein decreasing EA and remotely sensed, and reanalyses-based driving data (see Sec. 42.3.1) is pro-
decreasing Qn together result in increasing ET. voking their development. One example is the Evaporative Demand Drought
Index (EDDI; Hobbins et al., 2016; McEvoy et al., 2016a), which is showing
Global Dimming promise as a leading indicator of agricultural drought across many parts of
Changes in the radiation budget result from changes in both the longwave CONUS at time frames pertaining to both flash (i.e., fast-developing) and
flux [increasing concentrations of greenhouse gasses such as methane, CO2, extended droughts. Such E0-based indices must be physically based (EDDI is
and water vapor contribute to a TOA radiative imbalance of 0.9 ± 0.5 W m−2 based on ASCE05 ET0). ET-based metrics directly measure a hydrologic flux
(Trenberth and Fasullo, 2010)] and the shortwave flux (due to solar variability (ET should decrease in sustained droughts, and may in fact increase at the
and changing transmission and reflection due to atmospheric dust, aerosols, onset of flash droughts in the period between increasing Qn and/or EA and
clouds, and humidity). Stanhill and Cohen (2001) found evidence of “global water availability becoming limiting), but are often only available with con-
dimming” across the second half of the twentieth century: globally averaged siderable latency due to the data-QA/QC procedures necessary to incorpo-
Rd fell by 2.7% decade−1 for a total (in 2000) of 20 W m−2. They conclude that rate remotely sensed drivers (e.g., to eliminate cloud effects) and interpolation
aerosols and cloud cover and their indirect interactions are the most likely between satellite overpasses (in the case of nongeostationary satellites).
causes of this dimming; however, they are the most difficult to quantify. E0-based measures, while they do not directly measure a hydrologic flux
A reverse of the dimming—or rebrightening—has been observed in many (recall that E0 is an idealized flux) nevertheless offer some advantages over
regions (Pinker et al., 2005; Wild et al., 2005). This turn-around may be due ET-based metrics: they are ground-based, and so do not require R/S data
to economic factors and/or clean-air legislation. Wild et al. (2005) anticipate streams; they have limited latency (NLDAS data are currently available after
a stronger greenhouse signal being observed from a brighter atmosphere, as ~100 h); and their dynamics respond positively to both flash droughts and
the aerosol direct and indirect dimming effects are alleviated. Brightening sustained droughts. E0 rises both in response to drivers of flash drought
would have the same effect on surface warming as greenhouse gas-forcing, (increased T, Rd, or Uz, or decreased ea) and via the CR in response to drivers
but would double the impact on the hydrologic cycle (Romanou et al., 2007). of sustained drought (decreased Prcp), while the response of ET to drought
The main research questions provoked by dimming point to the need for depends on drought type: ET (like E0) rises in response to drivers of flash
resolution of the issue of aerosol effects on the hydrologic cycle: what is the drought, but falls (opposite to E0) in response to sustained drought. There is
direction of the aerosol indirect effect on Rd? Are the dimming and rebright- also a potential role for near-real-time E0 estimates in drought monitors (e.g.,
ening phenomena global or regional? What are the effects of rebrightening the U.S. Drought Monitor) as physically based inputs to LSMs and drought
due to anthropogenic aerosol mitigation: will the hydrologic cycle further models (e.g., PDSI), to replace the T-based E0 estimators currently in vogue
intensify as clean-air legislation and energy production reduce regional aero- (see Sec. 42.3.4). E0 and ET-based measures may act as a complement to each
sol loads such as in India and China? What will be their effects on the crucial other in drought indices: this is clearly an avenue ripe for further research.
Asian monsoon? Beyond minimizing observational uncertainties, we also As forecast drivers become available at the short term (e.g., weather ele-
need improved parameterizations of aerosol dynamics and effects in GCMs, ments from numerical weather predictors used by weather forecasters, such
particularly of aerosol-cloud interactions. as the National Weather Service), and at the mid-term (e.g., the Climate
42_Singh_ch42_p42.1-42.18.indd 12 8/26/16 4:36 PM

Outstanding Problems and Directions for Future Work 42-13
Forecast System version 2, CFSv2), forecasting of E0 becomes possible are often deeply obscured within LSMs (such as the PDSI) to such a degree
(McEvoy et al., 2016b). This is a highly desired outcome, with calls from many that users of such LSMs may be completely oblivious to their presence and,
sectors and at many time scales; for example, irrigation scheduling and thus, that they are relying on poor models.
drought monitoring at the shortest time scale, or drought forecasting and However, not only has this underlying assumption been shown not to hold,
water resource planning at seasonal time scales. Such a product is the Forecast but its effects in estimating both trends and short-term variability of E0 are
Reference ET product of the U.S. National Weather Service, which forecasts dangerous. Hobbins et al. (2012) showed that across much of CONUS, T and
ET0 at 1- to 7-day lead times, driven by meteorological and radiative forcing Rd are, in fact, negatively correlated across much of the year (and positively
fields that are already operationally forecast by numerical weather predictors. correlated at other times and in other regions), giving the lie to the central
assumption itself. Regarding E0-trends, Roderick et al. (2007) showed that
trends in observed Epan in Australia are almost completely a function of
42.3.4 Warning on Temperature-Based E0 Parameterizations (declining) U2, not in T or Rd, while Hobbins et al. (2008) showed (also in
A persistent, significant problem in the operational estimation of E0 (and Australia) that using T-based E0 in the PDSI in secular drought analyses
through LSMs, ET) is the continued use of outdated E0 parameterizations that results in almost uniformly increasing PDSI trends (increasing droughtiness),
are typically based on T alone (e.g., Thornthwaite, Hargreaves and Samani, contrary to both observations and the results when a more-realistic, physi-
Hamon, and Blaney–Criddle). The motivation for using such T-based param- cally based E0 is used. Also note that each driver’s relative contributions to the
eterizations is that they are simple to use, they appear intuitive, and their data variability of both Epan and ET0 varieties of E0 across CONUS is dynamic, with
requirements are minimal (often T alone). Their underlying assumption is different drivers dominating regionally at different ties of the year and at dif-
that there is a strong relationship between the radiative budget Qn (or the ferent time scales (Hobbins et al., 2012; Hobbins, 2016).
component of Qn that is most commonly measured, Rd) and either 2-m T or Figure 42.7 illustrates the dangers of estimating long-term E0 from T-based
the diurnal temperature range (DTR = Tmax - Tmin). This encourages the use parameterizations. While the T-based Ep from the PDSI inhabits much the
of T (or DTR) as a sole input for parameterization of Rd (or Qn). In estimating same T-response space as Ep from observations of Epan (see Fig. 42.7a), it is
E0, representing the effects of its other physical drivers—humidity and wind constrained to a much-reduced variability and exhibits physically unrealistic
speed—is assumed not to be as important in estimating E0 as is capturing the T-responses in that (i) it is zero below 0°C, and (ii) it declines above 38°C
effects of the T and Rd drivers, so one or both of these minor drivers is (100°F; not shown). When used in trends analysis, the almost uniformly
ignored. Let us call this the “T-Rd assumption.” upwards trends in T (see Fig. 42.7b) result in similarly upward trends in Ep
The problem is that these T-based parameterizations are being used out- from PDSI (see Fig. 42.7c), in marked contrast to observed trends, which are,
side of the geography and, particularly, time-frame appropriate to their at the majority of stations, downwards (see Fig. 42.7d).
underlying assumptions. For instance, Thornthwaite (1948) derived his Regarding the operational estimation of E0, our advice here is that the
approach to estimating E0 as a taxonomic aid in classifying climates around drivers of evaporation from hydrologic models must be carefully examined:
the globe: he needed a simple technique for which the data were available at particularly the parameterization of E0. A priori, we should assume that the
his required global extent and climatologic time frame. However, due to its best E0 parameterizations are physically based and are driven by observations
simplicity, his T-based E0-estimation approach and its descendants are now (or predictions) of all drivers; if T-based E0 estimators are used, their use
used in dynamic modeling of E0 (a particularly egregious example of which is should be supported within the context of the purpose to which the E0 is
to estimate ET from the simple hydrology bucket model that underpins the being put, and for the particular region, season, and time scale. A variability
PDSI, see Sec. 42.3.3)—a purpose for which it was never intended and for analysis (e.g., Hobbins, 2016) or a trends analysis (e.g., Roderick et al., 2007)
which it is ill-suited (Hobbins et al., 2008). Such T-based parameterizations will assist in answering these questions. Clearly, however, in many places,
(b)
(a)
(c) (d)
Figure 42.7 Comparing the effects of a T-based versus a physically based parameterization of E0 at 27 Australian and eight New Zealand stations, 1975–2004: (a) shows the
T-response of monthly E0 (mm mon-1 on the y-axis) estimated by the T-based Ep from the PDSI [grey triangles; see Hobbins et al. (2008); Supporting Information] and contempora-
neously observed Epan (dark circles) to mean monthly T (°C, on the x-axis); (b), (c), and (d) show station trends in T (°C a−1), Ep from the PDSI (mm a-1); and Ep trends from Epan
observations using U.S. class-A pans (mm a-2), respectively. [Source: Hobbins et al. (2008). Used with permission.]
42_Singh_ch42_p42.1-42.18.indd 13 8/26/16 4:36 PM

T-based E0 estimators should be avoided at all time scales, and they are not Because constants and empirical relationships inherent in Ep and ET0 equa-
recommended everywhere for any season, or for all seasons in most places. tions assume a well-watered environment, and the fact that Kc ratios were
developed under well-watered conditions, it is therefore required that weather
42.3.5 Energy Imbalance Problem data conditioned by the well-watered environment be used to compute ET0
Many land surface energy balance experiments using EC techniques show for estimating ETc. This requirement has been the motivation for developing
that available energy [see Eq. (42.4a)] is larger than the sum of turbulent irrigated environment weather station networks, such as AgriMet, CIMIS,
fluxes lET and H by 10–20% (Foken, 2008b), leading to an energy balance NICE Net, AZMET, to more accurately estimate ET0 and provide benchmark
residual Res, defined as: datasets for correcting water-limited weather data.
In efforts to reflect the weather that would exist over a well-watered sur-
Res = Rn + Ln − G − λ ET − H (42.33) face, various adjustment techniques have been developed and applied to
weather collected in water-limited regions. Temesgen et al. (1999) developed
Energy imbalance has been a significant topic of research and discussion and proposed empirical T and humidity adjustments using agricultural
among the micrometeorological community over the last decade (Foken weather data. Abatzoglou (2011) showed that bias correction of gridded
2008b, Leuning et al., 2012). There are many reasons for energy imbalance, weather data can improve well-watered ET0 estimates, but that biases still
but the most obvious is the disparity in measurement height and horizontal exist. More recently, a physically based strategy using theoretical blending
scales (i.e., footprint) between available energy and turbulent flux sensors. height-profiling procedures was developed and applied to water-limited
Footprints for measured Rn, Ln, and G are ~10 m and 0.1–1 m, respectively, weather station and ambient gridded weather data. Results were shown to
while the footprint for lET and H is ~100 m at measurement heights of better reflect weather of a well-watered reference environment (Allen et al.,
2–10 m (Foken, 2008b). Errors in measured Rn, Ln, and incorrect estimates of 2013a).
G and heat storage due to phase-lag errors associated with heat storage in For practical applications, it is suggested that weather data collected from
soils, air, and biomass are emerging as significant contributors to energy ET0 networks be used for estimating ETsz and ETc. If no representative agro-
imbalance (Liebethal, 2006; Higgins, 2012; Leuning et al., 2012). While mea- meteorological data exist in the area of interest, gridded weather data can be
surement errors clearly contribute to energy imbalance they alone cannot used provided some bias corrections are applied (e.g., Temesgen et al., 1999;
solve the closure problem, nor can the many processing steps and corrections Abatzoglou, 2011). Monthly spatially distributed climatologies of dewpoint
required to solve the closure problem. Other sources of imbalance are flux depression (i.e., daily average dew point temperature minus daily Tmin) and Uz
divergence due to terrain and vegetation heterogeneities, interaction of scales, derived from agricultural weather data can also be applied to arid weather
and low-frequency mesoscale eddies not being captured by the instrumenta- data to better reflect agricultural conditions (Huntington and Allen, 2010;
tion (Foken, 2008b). Huntington et al., 2015). Developing agricultural representative weather data
Energy imbalance corrections are commonly applied by researchers and for operational ETc estimation continues to be an outstanding problem within
practitioners even though sources of errors or the energy balance terms in the ET community. However, recent and future advancements will hopefully
question are not fully understood. Historically, turbulent fluxes have been soon be well established and used for operational and practical applications.
targeted through increasing lET by Res (Stannard et al., 1994), or distributing
Res across lET and H according to b (Twine et al., 2000). More recently, Allen 42.3.7 Remote Sensing of ET
(2008) and Higgins (2012) used least squares linear regression to identify and
correct terms that describe large variances in Res. These studies found that G Developing accurate ET maps through the use of satellite or airborne imagery
and soil heat storage accounted for the largest variance. This technique is that are useful for water management remains a challenge. While estimating
attractive due to the statistical identification of systematic bias and coeffi- ‘snapshot’ ETinst or anomalies of ETinst operationally is fairly common, given
cients that are relatively unique due to the lack of correlation among the peaks freely available satellite imagery, such as Landsat and MODIS, developing
and timing of energy balance terms. Methodological uncertainties and mea- field-scale seasonal and annual ET maps that are physically based, fairly accu-
surement errors are inevitable, but uncertainties and errors can be avoided rate, and usable for water rights and management still requires expert knowl-
with intelligent experimental design and general guidelines. Tools and guide- edge, human oversight, and research codes. The main challenges are
lines for measurement, assessment, and reporting of flux measurements are accounting for Prcp and ET from wetting events prior to or after image dates,
summarized in detail by Foken and Wichura (1996), Allen (2008), and Allen detection and treatment of clouded areas of the image, estimating representa-
et al. (2011a; 2011b). tive daily ET0 and EF to account for daily variability of weather between image
dates, and the fusion of data from multiple satellite sensors and gridded
weather datasets to increase the frequency of useable images, spatial resolu-
42.3.6 Potential and Reference ET tion, and representation of local weather conditions. Consideration of these
Confusion regarding the concepts of potential ET (Ep) and reference ET (ET0) issues significantly improves the quality and accuracy of resulting seasonal
stems from misunderstandings of their basic assumptions; when applied in and annual ET estimates. In arid regions, where annual ET from natural veg-
dry climates, the impacts of these assumptions can be important to under- etation is nearly equivalent to annual Prcp, an empirical single-image
stand and consider (Sellers, 1964). As discussed under Penman and other approach has proven useful for estimating annual ET and ET from ground-
Combination-Based Equations, Penman (1948) described the concept of Ep as water (ETg) as it avoids the difficulties and associated errors with time inte-
the ET from a surface sufficiently moist where ET is constrained only by the gration between image dates (Groeneveld et al., 2007; Beamer et al., 2013).
remaining radiative and advective limits. The concept of ET0 is also traced to Application of remotely sensed surface energy balance approaches for esti-
Penman (1948), who moved beyond the definition of Ep by further specifying mating ETinst and seasonal ET in arid regions remains a challenge because
a cropped surface—usually grass or alfalfa. Standardizing the term “reference values of Qn and H are typically large, commonly leading to ET errors that
ET” to reduce confusion on the assumed “surface of known properties” was a exceed ET itself—this is a particular problem in estimating annual ETg
primary motivation of the ASCE-EWRI Evapotranspiration Task Committee (annual ET - Prcp) for phreatophyte shrub areas. For a review of outstanding
that developed Eq. (42.12) for estimating ETsz (Allen et al., 2005). Penman’s issues related to ET mapping using optical and thermal satellite imagery, see
assumption of a sufficiently moist surface stems from the fact that Ts and ea Gowda et al. (2008) and Irmak et al. (2012). Some recent advances in satellite-
are commonly unknown in defining the T and ea gradients over the surface. based remote sensing of ET include consideration of hydrologic states and
For a wet surface, these two quantities are related by the Clausius–Clapeyron fluxes between image dates to improve monthly and seasonal ET (Kjaersgaard
equation, and Penman used the finite difference form of this equation to et al., 2011; Pereira et al., 2015), automated calibration of the CIMIC process
eliminate Ts-T from the energy balance (as Ts is generally unknown). Penman (Allen et al., 2013b; Morton et al., 2013), use of the CR with remotely sensed
also assumed zero G, and that Ln was a function of T and ea over the surface. surface temperature (Szilagyi and Jozsa, 2009b), development of representa-
Based on these assumptions, it is obvious that the use of data collected over tive ET0 (Allen et al., 2013a) and wind fields in mountainous terrain (Allen
insufficiently moist surfaces directly violates both concepts, unless explicitly et al., 2013c), integration of soil-vegetation-atmosphere transfer models (Kustas
stated that the estimates represent the ambient Ep or ET0 from a hypothetical and Anderson, 2009), and fusion of data from multiple satellite sensors to
wet surface small in size. Larger wet environments invoke surface-atmosphere map field-scale ET (Cammalleri et al., 2013). Future directions for practical
interactions that condition the near surface boundary layer, resulting in local- applications of remotely sensed ET for water management include developing
scale complementarity, as discussed in Sec. 42.2.3. Even field-scale feedbacks frameworks to integrate recent advancements into operational workflows,
have been well documented in irrigated areas surrounded by water-limited and testing the sensitivity of different automated calibration and data-fusion
regions (Allen et al., 1983; Temesgen et al., 1999; Szilagyi and Schepers, 2014). techniques. Using cloud computing, parallel processing, and data-fusion
Despite this common knowledge, practitioners and researchers alike rou- techniques on systems such as the Google Earth Engine cloud computing and
tinely and erroneously apply ET0 equations to estimate well-watered ETc environmental monitoring platform shows exceptional promise for integrat-
using arid and nonconditioned weather data. ing recent advances for efficient, practical, and operational remotely sensed ET
42_Singh_ch42_p42.1-42.18.indd 14 8/26/16 4:36 PM

REFERENCES 42-15
applications (Allen et al., 2014). Future field-scale research and applications remote sensing and NLDAS-2 simulations with U.S. Drought Monitor clas-
that rely on reflectance and surface energy balance-based modeling rests on sifications,” Journal of Hydrometeorology, 14: 1035–1056, 2013, doi:10.1175/
the availability of Landsat-like paired optical and thermal infrared data that JHM-D-12-0140.1.
are at least at the 100-m resolution, with at least 4- to 16-day return intervals Aubinet, M., A. Grelle, A. Ibrom, Ü. Rannik, J. Moncrieff, T. Foken, A. S.
(Miller et al., 2013; Serbina and Miller, 2014). Kowalski, et al., “Estimates of the annual net carbon and water exchange of
forests: the EUROFLUX methodology,” Advances in Ecological Research, 30:
ACKNOWLEDGMENTS 113–175, 1999, doi:10.1016/S0065-2504(08)60018-5.
Baldocchi, D., E. Falge, L. Gu, R. Olson, D. Hollinger, S, Running, P. Anthoni,
M. H. was supported by the NOAA-National Integrated Drought Information et al., “FLUXNET: a new tool to study the temporal and spatial variability of
System (NIDIS) and the Famine Early Warning System Network (FEWS ecosystem-scale carbon dioxide, water vapor, and energy flux densities,”
NET). J. H. was partially funded by the U.S. Geological Survey Landsat Bulletin of the American Meteorological Society, 82 (11): 2415–2434, 2001.
Science Team, a Nevada NASA EPSCoR Space Grant, and U.S. Bureau of Bastiaanssen, W. G. M., M. Menenti, R. A. Feddes, and A. A. M. Holtslag,
Reclamation WaterSMART and Science and Technology grants for improved “A remote sensing surface energy balance algorithm for land (SEBAL): 1.
irrigation demands and open-water evaporation monitoring. Formulation,” Journal of Hydrology, 212–213: 198–212, 1998, doi:10.1016/
S0022-1694(98)00253-4.
REFERENCES Beamer, J. P., J. L. Huntington, C. G. Morton, and G. M. Pohll, “Estimating
Abatzoglou, J. T., “Development of gridded surface meteorological data for annual groundwater evapotranspiration from phreatophytes in the Great
ecological applications and modeling,” International Journal of Climatology, Basin using Landsat and flux tower measurements,” Journal of the American
33 (1): 121–131, 2011, doi:10.1002/joc.3413. Water Resources Association, 49 (3): 518–533, 2013, doi:10.1111/jawr.12058.
Allen, R., “Quality assessment of weather data and micrometeorological Bosilovich, M. G., S. D. Schubert, and G. K. Walker, “Global changes of the
flux: impacts on evapotranspiration calculation,” Journal of Agricultural water cycle intensity,” Journal of Climate, 18: 1591–1608, 2005, doi:10.1175/
Meteorology (Japan), 64 (4): 191–204, 2008. JCLI3357.1.
Allen, R. G., C. E. Brockway, and J. L. Wright, “Weather station siting and Bouchet, R. J., “Évapotranspiration réelle et potentielle, signification clima-
consumptive use estimates,” Journal of Water Resources Planning and tique,” Proceedings of International Association Scientific Hydrology,
Management, 109 (2): 134–136, 1983. Publication No. 62, IAHS, Berkeley, CA, 1963, pp. 134–142.
Allen, R. G., L. S. Pereira, D. Raes, and M. Smith, Crop evapotranspiration: Bowen, I. S., “The ratio of heat losses by conduction and by evaporation
guidelines for computing crop water requirements, FAO Irrigation and from any water surface,” Physical Review, 27: 779–787, 1926.
Drainage Paper 56, Rome: FAO, 1998. Brunt, D., “Notes on radiation in the atmosphere,” Quarterly Journal of the
Allen, R. G., I. A. Walter, R. Elliott, T. Howell, D. Itenfisu, and M. Jensen, Royal Meteorological Society, 58 (247): 389–420, 1932.
“The ASCE standardized reference evapotranspiration equation,” American Brutsaert, W. and H. Stricker, “An Advection-Aridity approach to estimate
Society of Civil Engineers-Environmental Water Resources Institute Task actual regional evapotranspiration,” Water Resources Research, 15 (2): 443–
Committee on Standardization of Reference Evapotranspiration, Baltimore, 450, 1979.
MD, 2005, p. 59. Brutsaert, W., “Evaporation into the atmosphere: theory, history, and appli-
Allen, R. G. and M. Tasumi, “Evaporation from American Falls reservoir in cations,” D. Reidel, Dordrecht, The Netherlands, 1982, p. 299.
Idaho via a combination of Bowen Ratio and Eddy Covariance,” Proceedings, Brutsaert, W. and M. Parlange, “Hydrologic cycle explains the evaporation
2005 Environmental and Water Resources Institute Conference, Anchorage, paradox,” Nature, 396: 30, 1998, doi:10.1038/23845.
AK, May 15–19, 2005. Brutsaert, W., Hydrology: An Introduction, Cambridge University Press,
Allen, R. G., M. Tasumi, and R. Trezza, “Satellite-based energy balance for New York, 2005, p. 605.
mapping evapotranspiration with internalized calibration (METRIC)- Brutsaert, W., “Indications of increasing land surface evaporation during
model,” Journal of Irrigation and Drainage Engineering, 133 (4): 380–394, the second half of the 20th century,” Geophysical Research Letters, 33: L20403,
2007, doi:10.1061/(ASCE)0733-9437. 2006, doi:10.1029/2006GL027532.
Allen, R. G., L. S. Pereira, T. A. Howell, and M. E. Jensen, “Evapotranspiration Budyko, M. I., “Climate and life,” International Geophysics Series, 18: 508, 1974.
information reporting: I. Factors governing measurement accuracy,” Calera Belmonte, A., A. M. Jochum, A. Cuesta-Garcia, A. Montoro-
Agricultural Water Management, 98 (6): 899–920, 2011a, doi:10.1016/j. Rodriguez, and P. Lopez-Fuster, “Irrigation management from space: towards
agwat.2010.12.015. user-friendly products,” Irrigation Drainage Systems, 19: 337–353, 2005.
Allen, R. G., L. S. Pereira, T. A. Howell, and M. E. Jensen, “Evapotranspiration Cammalleri, C., M. C., Anderson, F. Gao, C. R. Hain, and W. P. Kustas, “A
information reporting: II. Recommended documentation,” Agricultural Water data fusion approach for mapping daily evapotranspiration at field scale,”
Management, 98 (6): 921–929, 2011b, doi:10.1016/j.agwat.2010.12.016. Water Resources Research, 49 (8): 4672–4686, 2013, doi:10.1002/wrcr.20349.
Allen, R. G., J. L. Huntington, A. Kilic, H. deBruin, and A. Joros, Cong, Z. T., D. W. Yang, and G. H. Ni, “Does evaporation paradox exist in
“Conditioning of NLDAS, NARR and arid weather station data to improve China?,” Hydrology and Earth Systems Sciences, 13: 357–366, 2009, doi:10.5194/
their representation of well-watered (reference) environments associated with hess-13-357-2009.
irrigated agriculture,” Proceedings of the 27th Conference on Hydrology, Dalton, J., “Experimental essays on the constitution of mixed gases; on the force
American Meteorological Society, Austin, TX, January 5–10, 2013a. of steam or vapour from water and other liquids in different temperatures, both
Allen, R. G., B Burnett, W. Kramber, J. L. Huntington, J. Kjaersgaard, A. in a Torricellian vacuum and in air; on evaporation; and on the expansion of
Kilic, C. Kelly, et al., “Automated calibration of the METRIC‐Landsat gases,” The Manchester Literary and Philosophical Society, Part 2: 574–594, 1802.
evapotranspiration process,” Journal of the American Water Resources de Bruin, H. A. R., “From Penman to Makink,” Proceedings of the
Association, 49 (3): 563–576, 2013b, doi:10.1111/jawr.12056. Information: TNO Committee on Hydrological Research, Vol. 39, edited by J. C.
Allen, R. G., R. Trezza, A. Kilic, M. Tasumi, and H. Li, “Sensitivity of Hoghart, Gravenhage, The Netherlands, 1987, pp. 5–31.
Landsat-scale energy balance to aerodynamic variability in mountains and Dingman, S. L., Physical Hydrology, 3rd ed., Waveland Press, Long Grove,
complex terrain,” Journal of the American Water Resources Association, 49 (3): IL, 2015, p. 631.
592–604, 2013c, doi:10.1111/jawr.12055. Donohue, R. J., M. L. Roderick, and T. R. McVicar, “Can dynamic vegeta-
Allen, R., B. Kamble, C. Morton, A. Kilic, J. L. Huntington, C. Robison, I. tion information improve the accuracy of Budyko’s hydrological model?,”
Rattcliffe, et al., “An evapotranspiration mapping tool at Landsat resolution on Journal of Hydrology, 390: 23–34, 2010, doi:10.1016/j.jhydrol.2010.06.025.
the Google Earth Engine: EEFlux,” Proceedings of Pecora 19, Sustaining Land Doorenbos, J. and W. O. Pruitt, “Crop water requirements,” FAO Irrigation and
Imaging, Denver, CO, November 17–20, 2014. Drainage Paper 24, Land and Water Development Division, FAO, Rome, 1977.
Alvarez, M. V., M. M. Gonzalez-Real, A. Baille, and J. M. Molina Martinez, Foken, T. and B. Wichura, “Tools for quality assessment of surface-based
“A novel approach for estimating the pan coefficient of irrigation water reser- flux measurements,” Agricultural Forest Meteorology, 78 (1–2): 83–105, 1996,
voirs: application to South Eastern Spain,” Agricultural Water Management, doi:10.1016/0168-1923(95)02248-1.
92: 29–40, 2007, doi:10.1016/j.agwat.2007.04.011. Foken, T., Micrometeorology, Springer–Verlag, Berlin Heidelberg, 2008a,
Anderson, M. C., R. G. Allen, A. Morse, and W. P. Kustas, “Use of Landsat p. 307, ISBN 978-3-540-74665-2.
thermal imagery in monitoring evapotranspiration and managing water Foken, T., “The energy balance closure problem: an overview,” Ecological
resources,” Remote Sensing of Environment, 122: 50–65, 2012, doi:10.1016/j. Applications, 18: 1351–1367, 2008b, doi:10.1890/06-0922.1.
rse.2011.08.025. Fu, G., S. P. Charles, and J. Yu, “A critical overview of pan evaporation
Anderson, M. C., C. Hain, J. Otkin, X. Zhan, K. Mo, M. Svoboda, B. trends over the last 50 years,” Climatic Change, 97: 193–214, 2009, doi:10.1007/
Wardlow, et al., “An intercomparison of drought indicators based on thermal s10584-009-9579-1.
42_Singh_ch42_p42.1-42.18.indd 15 8/26/16 4:36 PM

Fuehrer, P. L. and C. A. Friehe, “Flux corrections revisited,” Boundary-Layer Huntington, J. L., S. Gangopadhyay, J. M. Spears, R. Allen, D. King, C.
Meteorology, 102: 415–457, 2002. Morton, A. Harrison, et al., “West-wide climate risk assessments: Bias-
Gonzalez-Dugo, M. P., C. M. U. Neale, L. Mateos, W. P. Kustas, J. H. corrected and spatially downscaled irrigation demand and reservoir evapora-
Prueger, M. C. Anderson, et al., “A comparison of operational remote sensing- tion projections,” Technical memorandum No. 68–68210-2014-01, U.S. Bureau
based models for estimating crop evapotranspiration,” Agricultural and Forest of Reclamation, 2015, p. 196, 841 app.
Meteorology, 149 (11): 1843–1853, 2009, doi:10.1016/j.agrformet.2009.06.012. Huntington, T. G., “Evidence for intensification of the global water cycle:
Gowda, P. H., J. L. Chávez, P. D. Colaizzi, S. R. Evett, T. A. Howell, and J. A. review and synthesis,” Journal of Hydrology, 319: 83–95, 2006, doi:10.1016/j.
Tolk, “ET mapping for agricultural water management: present status and chal- jhydrol.2005.07.003.
lenges,” Irrigation Science, 26: 223–237, 2008, doi:10.1007/s00271-007-0088-6. Irmak, A., R. G., Allen, J. L. Kjaersgaard, J. L. Huntington, J., B. Kamble, R.
Grayson, R. B., R. M. Argent, R. J. Nathan, T. A. McMahon, and R. G. Mein, Trezza, and I. Ratcliffe, “Operational remote sensing of ET and challenges,”
“Hydrological recipes–estimation techniques in Australian hydrology,” Evapotranspiration—Remote Sensing and Modeling, edited by A. Irmak,
Cooperative Research Centre for Catchment Hydrology, Clayton, Australia, INTECH: Manhattan, New York, Chap. 21, 2012.
1996. Jones, R. N., T. A. McMahon, and J. M. Bowler, “Modelling historical lake
Groeneveld, D. P., W. M. Baugh, J. S. Sanderson, and D. J. Cooper, “Annual levels and recent climate change at three closed lakes, Western Victoria,”
groundwater evapotranspiration mapped from single satellite scenes,” Journal Journal of Hydrology, 246: 159–180, 2001, doi:10.1016/S0022-1694(01)00369-9.
of Hydrology, 344: 146–156, 2007, doi:10.1016/j.jhydrol.2007.07.002. Jung, M., M. Reichstein, P. Ciais, S. I. Seneviratne, J. Sheffield, M. L.
Hansen, J., L. Nazarenko, R. Ruedy, M. Sato, J. Willis, A. Del Genio, D. Goulden, G. Bonan, et al., “Recent decline in the global land evapotranspira-
Koch, et al., “Earth’s energy imbalance: confirmation and implications,” tion trend due to limited moisture supply,” Nature, 467: 951–954, 2010,
Science, 308: 1431–1435, 2005, doi:10.1126/science.1110252. doi:10.1038/nature09396.
Healy, R. W., T. C. Winter, J. W. LaBaugh, and O. L. Franke, “Water bud- Kalma, J. D., T. R. McVicar, and M. F. McCabe, “Estimating land surface
gets—foundations for effective water-resources and environmental manage- evaporation: a review of methods using remotely sensed surface temperature
ment,” U.S. Geological Survey Circular 1308, Reston, VA, 2007, p. 90. data,” Surveys in Geophysics, 29: 421–469, 2008, doi:10.1007/s10712-008-
Heilman, J. L., W. E. Heilman, and D. G. Moore, “Evaluating the crop coef- 9037-z.
ficient using spectral radiance,” Agronomy Journal, 74: 967–971, 1982. Kjaersgaard, J., R. G. Allen, and A. Irmak, “Improved methods for estimat-
Held, I. M. and B. J. Soden, “Robust responses of the hydrological cycle to ing monthly and growing season ET using METRIC applied to moderate reso-
global warming,” Journal of Climate, 19: 5686–5699, 2006, doi:10.1175/ lution satellite imagery,” Hydrological Processes, 25 (26): 4028–4036, 2011,
JCLI3990.1. doi:10.1002/hyp.8394.
Higgins, C. W., “A-posteriori analysis of surface energy budget closure to Kleissl, J., S. Hong, and J. Hendrickx, “New Mexico scintillometer network:
determine missed energy pathways,” Geophysical Research Letters, 39: L19403, supporting remote sensing and hydrologic and meteorological models,”
2012, doi:10.1029/2012GL052918. Bulletin of the American Meteorological Society, 90: 207–218, 2009,
Hobbins, M. T., J. A. Ramírez, T. C. Brown, and L. H. J. M. Claessens, “The doi:10.1175/2008BAMS2480.1.
complementary relationship in estimation of regional evapotranspiration: the Kustas, W. and M. Anderson, “Advances in thermal infrared remote sens-
CRAE and Advection-Aridity models,” Water Resources Research, 37 (5): ing for land surface modeling,” Agricultural and Forest Meteorology, 149 (12):
1367–1388, 2001a. 2071–2081, 2009, doi:10.1016/j.agrformet.2009.05.016.
Hobbins, M. T., J. A. Ramírez, and T. C. Brown, “The complementary rela- Labat, D., Y. Goddéris, J. L. Probst, and J. L. Guyot, “Evidence for global
tionship in estimation of regional evapotranspiration: an enhanced Advection- runoff increase related to climate warming,” Advances in Water Resources, 27:
Aridity model,” Water Resources Research, 37 (5): 1389–1404, 2001b. 631–642, 2004, doi:10.1016/j.advwatres.2004.02.020.
Hobbins, M. T., “Regional evapotranspiration and pan evaporation: com- Liu, S. M., Z. W. Xu, Z. L. Zhu, Z. Z. Jia, and M. J. Zhu, “Measurements of
plementary interactions and long-term trends across the conterminous evapotranspiration from eddy-covariance systems and large aperture scintil-
United States,” Ph.D. dissertation, Hydrologic Science and Engineering lometers in the Hai River Basin, China,” Journal of Hydrology, 487: 24–38, 2013,
Program, Civil Engineering Department, Colorado State University, Fort doi:10.1016/j.hydrol.2013.02.025.
Collins, CO, 2004, p. 232. Leuning, R., E. Van Gorsel, W. J. Massman, and P. R. Isaac, “Reflections on
Hobbins, M. T., J. A. Ramírez, and T. C. Brown, “Trends in pan evaporation the surface energy imbalance problem, Agricultural and Forest Meteorology,
and actual evapotranspiration across the conterminous U.S.: paradoxical or 156: 65–74, 2012, doi:10.1016/j.agrformet.2011.12.002.
complementary?,” Geophysical Research Letters, 31: L13503, 2004, Liebethal, C., “On the determination of the ground heat flux in microme-
doi:10.1029/2004GL019846. teorology and its influence on the energy balance closure,” Dissertation,
Hobbins, M. T., A. Dai, M. L. Roderick, and G. D. Farquhar, “Revisiting the University of Bayreuth, Bayreuth, Germany, 2006.
parameterization of potential evaporation as a driver of long-term water bal- Marx, A., H. Kunstmann, D. Schuttemeyer, and A. F. Moene, “Uncertainty
ance trends,” Geophysical Research Letters, 35: L12403, 2008, analysis for satellite derived sensible heat fluxes and scintillometer measure-
doi:10.1029/2008GL033840. ments over Savannah environment and comparison to mesoscale
Hobbins, M. T., A. Wood, D. Streubel, and K. Werner, “What drives the vari- meteorological simulation results,” Agriculture and Forest Meteorology, 148
ability of evaporative demand across the conterminous United States?,” Journal (4): 656–667, 2008, doi:10.1016/j.agrformet.2007.11.009.
of Hydrometeorology, 13: 1185–1214, 2012, doi:10.1175/JHM-D-11-0101.1. McEvoy, D., J. Huntington, M. T. Hobbins, A. Wood, C. Morton, M.
Hobbins, M. T., A. Wood, D. McEvoy, J. Huntington, C. Morton, M. Anderson, Anderson, and C. Hain, “The Evaporative Demand Drought Index: Part II -
and C. Hain, “The Evaporative Demand Drought Index: Part I - Linking drought CONUS-wide assessment against common drought indicators,” Journal of
evolution to variations in evaporative demand,” Journal of Hydrometeorology, Hydrometeorology, 2016a, doi:10.1175/JHM-D-15-0122.1.
2016, doi:10.1175/JHM-D-15-0121.1. McEvoy, D. J., J. L. Huntington, J. F. Mejia, and M. T. Hobbins, “Improved
Hobbins, M. T., “The variability of ASCE Standardized Reference seasonal drought forecasts using reference evapotranspiration anomalies,”
Evapotranspiration: a rigorous, CONUS-wide decomposition and attribu- Geophysical Research Letters, 43, 377–385, 2016b, doi:10.1002/2015GL067009.
tion,” Transactions of the ASABE, 59 (2): 1–16, 2016, doi:10.13031/ McGloin, R., H. McGowan, D. McJannet, F. Cook, A. Sogachev, and S.
trans.59.10975. Burn, “Quantification of surface energy fluxes from a small water body using
Hoedjes, J. C. B., A. Chehbouni, J. Ezzahar, R. Escadafal, and H. A. R. De scintillometry and eddy covariance,” Water Resources Research, 50: 494–513,
Bruin, “Comparison of large aperture scintillometer and eddy covariance 2014, doi:10.1002/2013WR013899.
measurements: can thermal infrared data be used to capture footprint- McJannet, D. L., F. J. Cook, R. P. McGloin, H. A. McGowan, and L. S. Burn,
induced differences?,” Journal of Hydrometeorology, 8: 144–159, 2007, “Estimation of evaporation and sensible heat flux from open water using a
doi:10.1175/JHM561.1. large-aperture scintillometer,” Water Resources Research, 47: W05545, 2011,
Hood, E., M. Williams, and D. Cline, “Sublimation from a seasonal snow- doi:10.1029/2010WR010155.
pack at a continental, mid-latitude alpine site,” Hydrological Processes, 13: McMahon, T., M. C. Peel, L. Lowe, R. Srikanthan, and T. R. McVicar,
1781–1797, 1999. “Estimating actual, potential, reference crop and pan evaporation using stan-
Huntington, J. L. and R. Allen, Evapotranspiration and Net Irrigation Water dard meteorological data: a pragmatic synthesis,” Hydrology and Earth System
Requirements for Nevada, Nevada State Engineer’s Office Publication, 2010, p. 266. Sciences, 17: 1331–1363, 2013, doi:10.5194/hess-17-1331-2013.
Huntington, J. L., J. Szilagyi, S. W. Tyler, and G. M. Pohll, “Evaluating the Miller, H., L. Richardson, S. R. Koontz, J. Loomis, and L. Koontz, “Users,
complementary relationship for estimating evapotranspiration from arid uses, and value of Landsat satellite imagery—results from the 2012 survey of
shrublands,” Water Resources Research, 47 (5): W05533, 2011, doi:10.1029/ users,” U.S. Geological Survey Open-File Report 2013–1269, 2013, p. 51,
2010WR009874. doi:10.3133/ofr/20131269.
42_Singh_ch42_p42.1-42.18.indd 16 8/26/16 4:36 PM

REFERENCES 42-17
Milly, P. C. D., “Climate, soil water storage, and the average annual water Rotstayn, L. D., M. L. Roderick, and G. D. Farquhar, “A simple pan-evap-
balance,” Water Resources Research, 30 (7): 2143–2156, 1994, doi:10.1029/ oration model for analysis of climate simulations: evaluation over Australia,”
94WR00586. Geophysical Research Letters, 33: L17715, 2006, doi:10.1029/2006GL027114.
Milly, P. C. D. and K. A. Dunne, “Trends in evaporation and surface cooling Samain, B., B. V. A. Ferket, W. Defloor, and V. R. N. Pauwels, “Estimation
in the Mississippi River basin,” Geophysical Research Letters, 28 (7): 1219– of catchment averaged sensible heat fluxes using a large aperture scintillom-
1222, 2001, doi:10.1029/2000GL012321. eter,” Water Resources Research, 47: W05536, 2011, doi:10.1029/2009WR009032.
Mitchell, K. E., D. Lohmann, P. R. Houser, E. F. Wood, J. C. Schaake, A. Saur, J. F. T. and E. R. Anderson, “The heat budget of a body of water of
Robock, B. A. Cosgrove, et al., “The multi-institution North American varying volume,” Limnology and Oceanography, 1(4): 247–251, 1956.
Land Data Assimilation System (NLDAS): utilizing multiple GCIP prod- Sellers, W. D., “Potential evapotranspiration in arid regions,” Journal of
ucts and partners in a continental distributed hydrological modeling sys- Applied Meteorology, 3(1): 98–104, 1964.
tem,” Journal of Geophysical Research, 109: D07S90, 2004, Serbina, L. and H. M., Miller, “Landsat uses and benefits—case studies by
doi:10.1029/2003JD003823. application area,” U.S. Geological Survey Open-File Report 2014–1108, 2014,
Moene, A. F., O. K. Hartogensis, and F. Beyrich, “Developments in scintil- p. 61, doi:10.3133/ofr20141108/.
lometry,” Bulletin of the American Meteorological Society, 90: 694–698, 2009, Sheffield, J., E. F. Wood, and M. L. Roderick, “Little change in global
doi:10.1175/2008BAMS2672.1. drought over the past 60 years,” Nature, 491: 435–438, 2012, doi:10.1038/
Monteith, J. L., “Evaporation and environment,” Symposia of the Society for nature11575.
Experimental Biology, XIX: 205–234, 1965. Shuttleworth, W. J., “Evaporation,” Handbook of Hydrology, edited by D. R.
Moreo, M. T. and A. Swancar, “Evaporation from Lake Mead, Nevada and Maidment, McGraw-Hill, New York, 1992, p. 551.
Arizona, March 2010 through February 2012,” U.S. Geological Survey Singh, V. P. and C. Y., Xu, “Evaluation and generalization of 13 mass-
Scientific Investigations Report 2013–5229, 2013, p. 40, doi:10.3133/ transfer equations for determining free water evaporation,” Hydrological
sir20135229. Processes, 11: 311–324, 1997a, doi:10.1002/(SICI)1099-1085(19970315)11:3<
Morton, C. G., J. L. Huntington, G. M. Pohll, R. G. Allen, K. C. McGwire, 311::AID-HYP446>3.0.CO;2-Y.
and S. D. Bassett, “Assessing calibration uncertainty and automation for esti- Singh, V. P. and C. Y., Xu, “Sensitivity of mass transfer-based evaporation
mating evapotranspiration from agricultural areas using METRIC,” Journal of equations to errors in daily and monthly input data,” Hydrological Processes,
the American Water Resources Association, 49 (3): 549–562, 2013, doi:10.1111/ 11: 1465–1473, 1997b, doi:10.1002/(SICI)1099-1085(199709)11:11<
jawr.12054. 1465::AID-HYP452>3.0.CO;2-X.
Morton, F. I., “Operational estimates of areal evapotranspiration and their Slack, J. R. and J. M. Landwehr, “Hydro-Climatic Data Network (HCDN):
significance to the science and practice of hydrology,” Journal of Hydrology, a US Geological Survey streamflow data set for the United States for the study
66: 1–76, 1983a. of climate variations (1874–1988),” U.S. Geological Survey Open-File Report
Morton, F. I., “Operational estimates of lake evaporation,” Journal of 92–129, 1992.
Hydrology, 66: 77–100, 1983b. Stanhill, G. and S. Cohen, “Global dimming: a review of the evidence for a
Morton, F. I., S. Fogarasi, and F. Ricard, “Operational estimates of areal widespread and significant reduction in global radiation with discussion of its
evapotranspiration and lake evaporation-Program WREVAP,” NHRI Paper probable causes and possible agricultural consequences,” Agricultural and
No. 24, Inland Waters Directorate, Environment Canada, Ottawa, Ontario, Forest Meteorology, 107: 255–278, 2001, doi:10.1016/S0168-1923(00)00241-0.
1985. Stannard, D. I., J. H. Blanford, W. P. Kustas, W. D. Nichols, S. A. Amer, T. J.
Morton, F. I., “Practical estimates of lake evaporation,” Journal of Climate Schmugge, and M. A. Weltz, “Interpretation of surface flux measurements in
and Applied Meteorology, 25 (3): 371–387, 1986. heterogeneous terrain during the Monsoon 90 experiment,” Water Resources
Morton, F. I., “Evaporation research—a critical review and its lessons for Research, 30 (5): 1227–1239, 1994, doi:10.1029/93WR03037.
the environmental sciences,” Critical Reviews in Environmental Science and Sugita, M., J. Usui, I. Tamagawa, and I. Kaihotsu, “Complementary rela-
Technology, 24 (3): 237–280, 1994. tionship with a convective boundary layer to estimate regional evapotranspi-
Oldekop, E. M., “On the evaporation from the surface of river basins,” ration,” Water Resources Research, 37 (2): 353–365, 2001, doi:10.1029/
Trans. Meteorol. Obs. Univ. Tartu, 4: 200, 1911. 2000WR900299.
Palmer, W. C., “Meteorological drought,” Research Paper 45, U.S. Swinbank, W. C., “The measurement of vertical transfer of heat and water vapor
Department of Commerce, Washington, DC, 1965, p. 58. by eddies in the lower atmosphere,” Journal of Meteorology, 8(3): 135–145, 1951.
Penman, H. L., “Natural evaporation from open water, bare soil, and grass,” Szilagyi, J., “Modeled areal evaporation trends over the conterminous United
Proceedings of the Royal Society, London, A (193): 120–145, 1948. States,” Journal of Irrigation Drainage and Engineering, 127 (4): 196–200, 2001,
Penman, H. L., “Estimating evaporation,” Transactions, American doi:10.1061/(ASCE)0733-9437.
Geophysical Union, 37: 43–50, 1956. Szilagyi, J., “On the inherent asymmetric nature of the complementary
Pereira, L. S., Allen, R. G., Smith, M., and Raes, D. Crop evapotranspiration relationship of evaporation,” Geophysical Research Letters, 34 (2): L02405,
estimation with FAO56: past and future,” Agricultural Water Management, 2007, doi:10.1029/2006GL028708.
147: 4–20, 2015, doi:10.1016/j.agwat.2014.07.031. Szilagyi, J. and J. Jozsa, “Complementary relationship of evaporation and
Pinker, R.T., B. Zhang, and E.G. Dutton, “Do satellites detect trends in the mean annual water-energy balance,” Water Resources Research, 45:
surface solar radiation?,” Science, 308: 850–854, 2005, doi:10.1126/sci- W09201, 2009a, doi:10.1029/2009WR008129.
ence.1103159. Szilagyi, J. and J. Jozsa, “Estimating spatially distributed monthly evapo-
Priestley, C. H. B. and R. J. Taylor, “On the assessment of surface heat flux transpiration rates by linear transformations of MODIS daytime land surface
and evaporation using large-scale parameters,” Monthly Weather Review, 100: temperature data,” Hydrology and Earth System Sciences Discussions, 6 (2):
81–92, 1972. 1433–1462, 2009b, doi:10.5194/hess-13-629-2009.
Roderick, M. L., M. T. Hobbins, and G. D. Farquhar, “Pan evaporation Szilagyi, J. and A. Schepers, “Coupled heat and vapor transport: the ther-
trends and the terrestrial water balance, I. Principles and observations,” mostat effect of a freely evaporating land surface,” Geophysical Research
Geography Compass, 3/2: 746–760, 2009a, doi:10.1111/j.1749-8198 Letters, 41: 2014, doi:10.1002/2013GL058979.
.2008.00213.x. Tasumi, M., “A review of evaporation research on Japanese lakes,”
Roderick, M. L., M. T. Hobbins, and G. D. Farquhar, “Pan evaporation Proceedings, 2005 Environmental and Water Resources Institute Conference,
trends and the terrestrial water balance, II. Energy balance and interpreta- Anchorage, Alaska, 2005, May 15–19.
tion,” Geography Compass, 3/2: 761–780, 2009b, doi:10.1111/ Tasumi, M., R. G. Allen, R. Trezza, and J. L. Wright, “Satellite-based energy
j.1749-8198.2008.00214.x. balance to assess within-population variance of crop coefficient curves,”
Roderick, M. L. and G. D. Farquhar, “The cause of decreased pan evapora- Journal of Irrigation Drainage Engineering, ASCE, 131 (1): 94–109, 2005,
tion over the past 50 years,” Science, 298: 1410–1411, 2002, doi:10.1126/ doi:10.1061/(ASCE)0733-9437(2005)131:1(94).
science.1075390-a. Temesgen, B., R. G. Allen, and D. T. Jensen, “Adjusting temperature param-
Roderick, M. L., L. D. Rotstayn, G. D. Farquhar, and M. T. Hobbins, “On eters to reflect well-watered conditions,” Journal of Irrigation and Drainage
the attribution of changing pan evaporation,” Geophysical Research Letters, 34: Engineering, ASCE, 125 (1): 26–33, 1999, doi:10.1061/(ASCE)0733-
L17403, 2007, doi:10.1029/2007GL031166. 9437(1999)125:1(26).
Romanou, A., B. Liepert, G. A. Schmidt, W. B. Rossow, R. A. Ruedy, and Y. Thom, A. S., J-L. Thony, and M. Vauclin, “On the proper employment
Zhang, “20th century changes in surface solar irradiance in simulations and of evaporation pans and atmometers in estimating potential transpira-
observations,” Geophysical Research Letters, 34: L05713, 2007, tion,” Quarterly Journal of the Royal Meteorological Society, 107:
doi:10.1029/2006GL028356. 711–736, 1981.
42_Singh_ch42_p42.1-42.18.indd 17 8/26/16 4:36 PM

Thornthwaite, C. W., “An approach toward a rational classification of cli- Wang, T., G. R. Ochs, and S. F. Clifford, “A saturation resistant optical scintillom-
mate,” Geographical Review, 38: 55–94, 1948. eter to measure C2n,” Journal of the Optical Society of America, 68: 334–338, 1978.
Trenberth, K. E. and J. T. Fasullo, “Tracking Earth’s energy,” Science, 328: Wild, M., H. Gilgen, A. Roesch, A. Ohmura, C. N. Long, E. G. Dutton, B.
316, 2010, doi:10.1126/science.1187272. Forgan, et al., “From dimming to brightening: decadal changes in solar radia-
Turc, L., “Le bilan d’eau des sols: relation entre les précipitations, tion at Earth’s surface,” Science, 308: 847–850, 2005, 6 May, doi:10.1126/sci-
l’évaporation et l’écoulement,” Annales Agronomiques Série A, 5: 491–596, ence.1103215.
1954 (in French). Winter, T. C., D. C. Buso, D. O. Rosenberry, G. E. Likens, A. M. Sturrock,
Twine, T. E., W. P. Kustas, J. M. Norman, D. R. Cook, P. Houser, T. P. and D. P. Mau, “Evaporation determined by the energy-budget method for
Meyers, J. H. Prueger, “Correcting eddy-covariance flux underestimates over Mirror Lake, New Hampshire,” Limnology and Oceanography, 48(3): 995–
a grassland,” Agricultural and Forest Meteorology, 103(3): 279–300, 2000, 1009, 2003, doi:10.4319/lo.2003.48.3.09950.
doi:10.1016/s0168-1923(00)00123–4. Wyngaard, J. C., Y. Izumi, and J. S. A. Collins, “Behavior of the refractive-
Wang, K. and R. E. Dickinson, “A review of global terrestrial evapotranspi- index-structure parameter near the ground,” Journal of the Optical Society of
ration: observation, modeling, climatology, and climatic variability,” Reviews America, 61 (12): 1646–1650, 1971, doi:10.1364/JOSA.61.001646.
of Geophysics, 50: RG2005, 2012, doi:10.1029/2011RG000373.
42_Singh_ch42_p42.1-42.18.indd 18 8/26/16 4:36 PM

Chapter 43
Rainfall Interception, Detention,
and Depression Storage
BY
DARRYL E. CARLYLE-MOSES
ABSTRACT Forest and tree characteristics influence Ic by controlling above-ground

water storage capacities and regulating during- and post-event evaporation
Rainfall depth is an important consideration for estimating the amount of water
from that storage. Canopy cover fraction is an important control on the quan-
available for hydrologic processes such as infiltration, transpiration, and runoff
titative importance of Ic as it represents the effective area from which rain can
generation; nevertheless, not all rain is available to take part in the terrestrial
be stored upon and evaporated (Gash et al., 1995), while leaf area index (LAI)
portion of the hydrologic cycle. Interception of rain by vegetation canopies and
is considered to be an important factor influencing the canopy storage capac-
the forest floor layer, as well as storage of water in surface depressions, represent
ity (Deguchi et al., 2006). However, differing foliage characteristics, including
avenues by which rain may evaporate directly back to the atmosphere. Canopy
leaf hydrophobicity (Holder, 2013) and morphology (Licata et al., 2011), may
interception loss Ic denotes the portion of rain that strikes vegetation canopies
result in dissimilar storage and, by extension, Ic under similar canopy cover
and is stored upon that vegetation and subsequently evaporated, while forest
and LAI conditions. In addition, the proportion of cover that is comprised of
floor interception loss If is the fraction of rain that is retained by and evapo-
branches and boles or epiphytes as opposed to leaf foliage should also be
rated from the litter-layer, bryophytes, lichens, and/or grasses that compose the
considered. Wood has a much greater storage capacity per unit surface area
understory. Rain may also be temporarily stored directly on the soil surface or
than foliage (Herwitz, 1985), while epiphytes may have large storage capaci-
on artificial surfaces such as concrete and asphalt, constituting detention or
ties and retain water over long periods that may bridge events (Pypker et al.,
depression storage SD. The goal of this chapter is to provide an overview of
2011; Allen et al., 2014). Interception loss typically increases with forest age,
interception loss, both from vegetation canopies and from the forest floor, as
with positive relationships between Ic and stand basal area, diameter at breast
well as detention and depression storage, and to highlight the quantitative
height, and tree height being reported; however, stand density may not serve
importance of these hydrologic components, the physical processes involved,
as a good indicator of the Ic efficiency of a stand since forests self-thin with
and the current measurement techniques and modeling approaches applied.
age (see Llorens and Domingo, 2007; Carlyle-Moses and Gash, 2011).
Although much variability exists among and within forest types, coniferous
43.1 CANOPY INTERCEPTION LOSS
forests typically partition a greater amount of annual rainfall as Ic than their
Tree canopies and their associated water storage capacities coupled with broad-leaved deciduous and broad-leaved evergreen counterparts. Miralles et
evaporation during and after a storm event can substantially diminish the al. (2010), using multisatellite observations coupled with the climate predic-
quantity of rain that reaches the ground. For example, Wallace and McJannet tion center morphing technique (CMORPH) rainfall product and the Gash
(2006) found Ic accounted for 25% of the rain that fell at a lowland tropical analytical interception model (see as follows), derived global estimates of
rainforest in Australia over a 3-year period, while Licata et al. (2011) esti- annual Ic. Canopy interception loss of rainfall for coniferous forests was found
mated annual Ic over 2 consecutive years to range from 35 to 44% of rainfall to average 22%, while for deciduous and broad-leaved evergreen forests Ic
in stands of cordilleran cypress (Austrocedrus chilensis) in Argentina. Annual averaged 19 and 13% of annual rainfall, respectively. However, as Miralles
and season-long Ic can exceed 50% of rainfall (e.g., McMinn, 1960) and often et al. (2010) note, differences in Ic among the forest types may have more to
represents an important and even dominant component of evapotranspira- do with the rainfall regime of a region than the characteristics of the forest.
tion from natural forests and plantations (Benyon and Doody, 2014). For example, coniferous forests are commonly found in areas subjected to
At the storm-event scale, the percentage of rain partitioned into Ic typically long-duration synoptic rainfall, while tropical broad-leaved forests are
decreases asymptotically with rainfall depth (Carlyle-Moses and Gash, 2011). exposed to short-duration, high-intensity rainfalls and large annual rain
For trees and forests with complete canopy cover, Ic may be equated with the depths (Miralles et al., 2010). In addition, forests with widely differing fea-
total depth of rain during small events that do not saturate the canopy (i.e., no tures, but located in the same region, may have very similar Ic (see Krämer and
rain reaches the ground). However, with progressively larger rain events dif- Hölscher, 2009), highlighting the importance of an area’s rainfall regime in
fering areas within the canopy saturate with additional rain input becoming evaluating the relative importance of this component of evapotranspiration.
canopy-drip throughfall or flow along branches and down the tree boles as During the dormant-season temperate, deciduous forests may be efficient
stemflow. For relatively large events that entirely saturate the canopy Ic should at partitioning rainfall into Ic despite having no leaf cover. Intuitively, the rela-
not be considered negligible. For example, the fraction of rainfall partitioned tively large storage capacities associated with the wood component of the
into Ic for a cedar (Juniperus flaccida)—white oak (Quercus glaucoides)—ash canopy and the increased exposure of this retained water to wind in the
(Fraxinus cuspida) forest of northeastern Mexico remained quasi-constant at absence of leaves may result in relatively large Ic, especially when considered
approximately 15% of rainfall for events ranging in depth from ≈10 to on a per unit of cover basis. Deguchi et al. (2006), in a Japanese deciduous
120 mm (Carlyle-Moses and Gash, 2011). forest, found that Ic over a 3-year period decreased from 17.6% of rainfall
43-1
43_Singh_ch43_p43.1-43.4.indd 1 8/22/16 2:11 PM

43-2 Rainfall Interception, Detention, and Depression Storage
under growing-season conditions to 14.3% under dormant conditions; how- complicated by the spatial variability exhibited by understory precipitation.
ever, the decrease in the Ic efficiency of the stand (24.2%) was much lower Evaluations of sampling and scaling methodologies of understory precipita-
than the 41.1% decrease in LAI. tion are provided by Hanchi and Rapp (1997), Ritter and Regalado (2014),
As aforementioned, Ic is composed of during-event evaporation E [L T-1] and Zimmermann and Zimmermann (2014).
and the evaporation of water held in storage at the end of an event Ce [L]: Several physically based Ic simulation models exist (Muzylo et al., 2009);
I c = Et + Ce (43.1) however, the Gash model (Gash, 1979) and reformulations of that model

(Gash et al., 1995; Valente et al., 1997) are the most commonly used (e.g., Jian
where t [T] is the storm duration. et al., 2014). The reformulated Gash model is an analytical model that
The relationship between event Ic [L] and rainfall depth P [L] is often combines the conceptual framework of the sparse numerical Rutter model
expressed as a linear equation (e.g., Helvey and Patric, 1965) of the form: with the simplicity of empirical equations (Valente et al., 1997). The model
I c = aP + b (43.2) considers Ic in three distinct phases: (1) the wetting-up of the canopy phase,

(2) the saturated canopy phase, and (3) the canopy drying phase. The param-
Estimates of E and Ce may be obtained from Eq. (43.2). Using rain events eters required for the model include Sc, E c, and R . In addition, estimates of the
that are assumed to be large enough to satisfy the canopy storage capacity, the ratio of the mean evaporation rate from the trunks and the canopy e, the
slope a represents the ratio of mean E from the saturated canopy to the mean trunk storage capacity St [L], and the trunk partitioning coefficient pd—which
intensity of rain falling on that saturated canopy, R [L T-1], that is, a = E / R . represents the fraction of total drainage under saturated conditions that flows
If R is known (see methods described by Valente et al., 1997), E may be down the trunks as stemflow—are required to run the model. Methods of
derived as (Gash, 1979): deriving the various parameter values are provided by Valente et al. (1997).
E = aR (43.3) Reformulated Gash model components of rainfall partitioning by a vegetation

canopy, including Ic, throughfall and stemflow, and their formulation are
The intercept b of Eq. (43.2) represents a first approximation of Ce as well provided in Table 43.1.
as its storage capacity S [L], since Ce = S under saturation conditions (Gash,
1979). However, as a consequence of evaporation and drainage prior to the
Table 43.1 Components of Rainfall Partitioning by Forests Canopies
canopy reaching saturation, b may not be a good indicator of Ce or S (see in the Sparse Gash Model
Gash, 1979). Instead, other regression-based methods, such as the use of
envelope curves (see Valente et al., 1997), are recommended. Miralles et al. Canopy partitioning component Formulation
(2010) suggest, based on reviews of field studies from around the globe, that Rainfall thresholds
the E and S, scaled to the fraction of canopy cover, Ec and Sc, have a fairly nar-
Depth of rainfall required to  (1 − ε ) Ec 
row range, averaging 0.3 ± 0.1 mm h-1 and 1.2 ± 0.4 mm, respectively, across R
saturate canopy, P ′ − Sc ln1 − 
a wide array of forest types. (1 − ε ) Ec  R 
The evaporation of intercepted rainfall differs from transpiration in that E
is only depended on aerodynamic conductance, whereas transpiration is a
function of both aerodynamic conductance and canopy conductance. This Depth of rainfall required to
R St
has important implications in that E, and thus Ic may occur during periods saturate tree boles P ′′ + P′
R − (1 − ε ) Ec pd c
when the stomata of leaves are closed (e.g., at night for many species) or dur-
ing the dormant period for deciduous trees and forests (Helvey and Patric, Interception loss components
1965). Although E is often simulated using the Penman-Monteith equation,
Carlyle-Moses and Gash (2011) suggest that this approach should be aban- Interception loss from the canopy m
doned. The evaporation of rainfall stored on vegetation canopies is often

for m events of insufficient depth c∑P, j
to saturate the canopy, P < P ′ j =1
greater than that of transpiration under the same environmental conditions
and as such Ic needs to be modeled separately. In addition, it has long been
recognized that net radiation, the energy source in the Penman-Monteith Interception loss from the canopy
 
(1 − ε ) Ec n
approach, plays a minor role in the evaporation of intercepted rainfall from
forest and tree canopies (Stewart, 1977). Assuming a mean Ec rate of 0.3 ± 0.1
for n events of sufficient depth to
saturate the canopy, P ≥ P ′ c nP′ +
 R
(
∑ P , j − P′ )
j =1 
mm h-1 from forest canopies (Miralles et al., 2010), the average supply of
energy required to sustain Ec is 206 ± 69 W m-2. The energy involved in the Ic
process is considered to be advected sensible heat with the evaporation from Interception loss from the trunks qSt
for q events that saturate the trunks,
and the associated cooling of wetted canopies creating a sensible heat sink
P ≥ P ′′
(Stewart 1977; Carlyle-Moses and Gash, 2011). Sources of sensible heat
involved in the Ic process include the relatively warmer understory, surround-
ing vegetation not wetted by the rain event, and oceans (Stewart, 1977; Interception loss from trunks for  (1 − ε ) Ec  n
Carlyle-Moses and Gash, 2011). Asadian and Weiler (2009) suggest that one
reason Ic percentages from urban trees may be so great is due to the enhanced
n – q events that do not saturate the
canopy, P′′ > P ≥ P ′
pd c 1 −
 R  j=1
∑ P , j − P′ ( )
sensible heat generated by the urban heat island effect. For some vegetation
the source of advected energy may vary seasonally. Ringgaard et al. (2014), for
example, conclude that the sensible heat source for E in a Norway spruce Understory precipitation components
(Picea abies) plantation in Denmark is at the regional scale (North Sea) over Throughfall m+n
the winter and at the local scale during the summer.
With net radiation playing a minor role and canopy conductance not appli-
(1 − c ) ∑ P , j +
j =1
cable under saturated canopy conditions, Pereira et al. (2009) suggest that E
be derived using a diffusion equation with canopy surface temperature
equated with the wet-bulb temperature of the air:
(1 − p )c 1 − (1 −Rε ) Ec  ∑  1 − (1 −Rε ) Ec 
n
ρaC p d
λE = g [VPD ] (43.4) j =1
γ b
where l [L2 T-2] is the latent heat of vaporization, ra [M L-3] is the density of
Stemflow
air, cp is the heat capacity of air [L2 T-2 Θ-1], γ represents the psychrometric  (1 − ε ) Ec  q
constant [M L-1 T-2 Θ-1], g b [L T-1] is the aerodynamic conductance for water pd c 1 −
 R  j=1
(
∑ P , j − P′ − qSt )
vapor integrated over the distance between the foliage surface and the adja-
cent air, and VPD [M L-1 T-2 ] is the vapor pressure deficit between the satu-
rated surface and the overlying air. Definition of terms: R [L T-1] mean rate of rainfall on a saturated canopy; Ec [L T-1] mean evapo-
Canopy interception loss is almost always measured indirectly by taking ration rate from a saturated canopy on a per unit canopy basis; Sc [L] canopy storage capacity per unit
the difference between precipitation incident on the canopy, either measured area of canopy; St [L] the storage capacity of the tree boles; ε (dimensionless) the ratio between the
evaporation rate from the tree boles and the canopy; c (dimensionless) the canopy cover fraction; and
above the canopy or in a near-by clearing, and the understory precipitation pd (dimensionless) the trunk partitioning coefficient.
inputs of throughfall and stemflow. Obtaining an accurate estimate of Ic is (Source: Valente et al., 1997.)
43_Singh_ch43_p43.1-43.4.indd 2 8/22/16 2:11 PM

References 43-3
43.2 FOREST FLOOR INTERCEPTION LOSS importance of SD diminishes (Mishra and Singh, 2010). Following Linsley
et al. (1949), the depth of water held in surface depressions at any given time
Not all throughfall and stemflow inputs reach the soil surface. Below the for-
can be approximated as
est or tree canopy the litter layer, mosses, and grasses intercept and store a
proportion of understory precipitation. This forest floor interception loss If ( )
VS = SD−Max 1 − e − kPe (43.6)
may be appreciable. Price et al. (1997), for example, found that the feather
moss (Pleurozium schreberi) carpet below the canopy of a black spruce (Picea where VS [L] is the equivalent depth of storage at some time of interest, SD-Max
mariana) stand in northern Manitoba stored and subsequently evaporated [L] is the maximum depression storage of the surface, Pe [L] is the effective
23% of the summer throughfall input, while Gerrits et al. (2010) found that If rainfall depth, and k [L-1] is a constant. The constant k may be found by
for a beech (Fagus sylvatica) forest in Luxemborg accounted for 10 to 35% of assuming that for small values of Pe all effective rainfall is used to satisfy
rainfall. SD–Max (dVS/dPe = 1), giving k = 1/SD–Max (Viessman and Lewis, 2002).
Forest floor interception loss may be defined as follows (Gerrits and The maximum depression storage is an important property for runoff assess-
Savenje, 2011): ment in response to rain events. Singh (1989) provides reference values of
SD–Max for large impervious and pervious areas. Several empirical models have
dS f been developed relating SD–Max to an index of surface roughness (e.g., Onstad,
If = + E f = Pu − F (43.5)
dt 1984). Borselli and Torri (2010) suggest an empirical model of the form:
where Sf [L] and Ef [L T-1] are water storage by and evaporation from the for- SD−Max = c + de kX e − gSL (43.7)
est floor, respectively, Pu [L] represents understory precipitation (throughfall
+ stemflow) input, F [L] is the total infiltration into the ground, and t is time. where c is the minimum depression storage [L]; X is a generic roughness
Evaporation rates from forest floors may be considerably lower than those index; SL [L L-1, %] is the slope gradient expressed as a percentage; and d, k,
of wetted overstorey canopies due to a decrease in wind speed and the associ- and g are empirical coefficients. Borselli and Torri (2010) provide the coeffi-
ated aerodynamic conductance below the canopy. Additionally, net radiation, cient values associated with Eq. (43.7) for a variety of roughness indices.
which may be an important source of energy for Ef, is often much less for the
forest floor layer compared to the overstorey canopy (Gerrits and Savenje, 43.4 SUMMARY
2011). Although Ef may be relatively small, If may be appreciable, especially
where the thickness (or mass) of that forest floor is ample and there is enough Important pathways of rainfall occur before infiltration, or surface runoff can
time between rain events to allow for sufficient drying. Gerrits and Savenje take place and should not be considered negligible even for large rainfall
(2011) state that the frequency of rain events is the dominant factor influenc- events. This is especially true for canopy interception loss, whose quantitative
ing the magnitude of If. The importance of If may also vary by season, espe- importance is owed, in part, to relatively high during-event evaporation rates
cially for deciduous forests where the thickness of the litter layer increases that are sustained by advected sensible heat. Forest floor interception loss and
during leaf fall and, as a consequence of the reduction in LAI, short-wave depression storage, where the bulk of evaporation occurs after the storm has
radiation more readily penetrates to the forest floor (Gerrits and Savenje, ceased, may not be inconsequential, depending on not only surface character-
2011). Additionally, If for a given rainfall may differ before and after the istics but also the rainfall regime of the area.
dormant season since snow cover compresses and flattens leaf litter reducing
its storage capacity (Gerrits et al., 2010). REFERENCES
Measurement techniques concerning If include lab methods, which focus Allen, S. T., J. R. Brooks, R. F. Keim, B. J. Bond, and J. J. McDonnell, “The
on determining Sf using sample submersion or artificial rainfall, and field role of pre-event canopy storage in throughfall and stemflow by using isotopic
methods, which typically use lysimeters. Gerrits et al. (2010) developed a tracers,” Ecohydrology, 7: 858–868, 2014.
method that allows for both Ef and Sf to be evaluated separately in the field. Asadian, Y. and M. Weiler, “A new approach in measuring rainfall intercep-
Two aluminum basins, with one mounted on top of the other, are measured tion by urban trees in coastal British Columbia,” Water Quality Resources
continuously using strain gauges with the upper permeable basin containing Journal of Canada, 44: 16–25, 2009.
forest floor material and the lower basin acting as a reservoir for the drainage Benyon, R. G. and T. M. Doody, “Comparison of interception, forest floor
from the upper basin. A simple water balance approach is applied [see Eq. evaporation and transpiration in Pinus radiata and Eucalyptus globulus plan-
(43.5)] in deriving the components of If. Gerrits et al. (2010) provide a tations,” Hydrological Processes (in press), 2014, doi: 10.1002/hyp.10237.
modified version of the original Rutter model (Rutter et al., 1971) that Borselli, L. and D. Torri, “Soil roughness, slope and surface storage relation-
includes a forest floor interception reservoir. ship for impervious areas,” Journal of Hydrology, 393: 389–400, 2010.
Carlyle-Moses, D. E. and J. H. C. Gash, “Rainfall interception loss by forest
43.3 DETENTION AND DEPRESSION STORAGE canopies,” Forest Hydrology and Biogeochemistry: Synthesis of Past Research
and Future Directions, Ecological Series 216, edited by D. F. Levia, D. E.
In addition to canopy and forest floor interception, a portion of rainfall may Carlyle-Moses, and T. Tanaka, Springer-Verlag, Heidelberg, Germany, 2011,
be held in storage on the surface of soil (or other media) where it is also sub- pp. 407–423.
jected to evaporation. Although detention storage and depression storage Deguchi, A., S. Hattori, and H-T. Park “The influence of seasonal changes
have been differentiated by some (e.g., Viessman and Lewis, 2002), here the in canopy structure on interception loss: application of the revised Gash
two terms are considered synonymous with each other (Ponce, 1989; Mishra model,” Journal of Hydrology, 318: 80–102, 2006.
and Singh, 2010) and the term “depth of ponding” has also been used to Dingman, S. L., Physical Hydrology, 2nd ed., Waveland Press, Long Grove,
describe this surface storage component (e.g., Dingman, 2008). Depression IL, 2008, p. 656.
storage SD is the amount of rain that is retained on the surface in micro- Gash, J. H. C., “An analytical model of rainfall interception by forests,”
depressions, ditches, and other terrain irregularities where water is allowed to Quarterly Journal of the Royal Meteorological Society, 105: 43–55, 1979.
collect and pond. In response to rainfall, SD is present to some degree for all Gash, J. H. C., C. R. Lloyd, and G. Lachaud, “Estimating sparse forest rain-
impervious surfaces, including concrete, asphalt, and highly compacted soils, fall interception with an analytical model,” Journal of Hydrology, 170: 79–86,
with the depth of SD being dependent on surface roughness and the local 1995.
slope gradient (Borselli and Torri, 2010). Depression storage may develop on Gerrits, A. M. J., L. Pfister, and H. H. G. Savenije, “Spatial and temporal
pervious surfaces, such as soils, if either the rainfall rate exceeds the infiltra- variability of canopy and forest floor interception in a beech forest,”
tion capacity of the surface (saturation from above) or if infiltration ceases as Hydrological Processes, 24: 3011–3025, 2010.
a result of the surface pores becoming saturated (saturation from below). Gerrits A. M. J. and H. H. G. Savenije, Forest Floor Interception. Forest
Thus, for soils and other pervious surfaces, the development and depth of SD Hydrology and Biogeochemistry: Synthesis of Past Research and Future
are dependent not only on surface roughness and slope but also on the char- Directions, Ecological Series 216, edited by D. F. Levia, D. E. Carlyle-Moses,
acteristics of the surface, including its infiltration capacity and hydraulic and T. Tanaka, Springer-Verlag, Heidelberg, Germany, 445–454, 2011.
conductivity (Dingman, 2008). Water held in SD either evaporates or infil- Hanchi A. and M. Rapp, “Stemflow determination in forest stands,” Forest
trates into the soil while rainfall that exceeds the SD capacity of the surface Ecology and Management, 97: 231–235, 1997.
may contribute to surface runoff. Helvey J. D. and J. H. Patric, “Canopy and litter interception of rainfall by
The importance of SD varies as a function of time and thus the duration of hardwoods of eastern United States,” Water Resources Research, 1: 193–206,
rainfall (Mishra and Singh, 2010). At the beginning of a rain event SD may 1965.
play an important role. However, as the storm event progresses maximum SD Herwitz S. R., “Interception storage capacities of tropical rainforest canopy
is satisfied with any additional rainfall contributing to runoff and the relative trees,” Journal of Hydrology, 77: 237–252, 1985.
43_Singh_ch43_p43.1-43.4.indd 3 8/22/16 2:11 PM

43-4 Rainfall Interception, Detention, and Depression Storage
Holder C. D., “Effects of leaf hydrophobicity and water droplet retention on Ponce V. M., Engineering Hydrology: Principles and Practices, Prentice Hall,
canopy storage capacity,” Ecohydrology, 6: 483–490, 2013. Englewood Cliffs, NJ, 1989, p. 640.
Jian S., C. Zhao, S. Fang, and K. Yu, “Soil water content and water balance Price A. G., K. Dunham, T. Carleton, and L. Band, “Variability of water
simulation of Caragana korshinskii Kom. In the semiarid Chinease Loess fluxes through the black spruce (Picea mariana) canopy and feather moss
Plateau,” Journal of Hydrology and Hydromechanics, 62: 89–96, 2014. (Pleurozium schreberi) carpet in the boreal forest of northern Manitoba,”
Krämer I. and D. Hölscher, “Rainfall partitioning along a tree diversity Journal of Hydrology, 196: 310–323, 1997.
gradient in a deciduous old-growth forest in Central Germany,” Ecohydrology, Pypker T. G., D. F. Levia, J. Stailens, and J. T. Van Stan II, Canopy structure
2: 102–114, 2009. in relation to hydrological and biogeochemical fluxes, Forest Hydrology and
Licata J. A., T. G. Pypker, M. Weigandt, M. H. Unsworth, J. E. Gyenge, M. Biogeochemistry: Synthesis of Past Research and Future Directions, Ecological
E. Fernandez, T. M. Schlichter, et al., “Decreased rainfall interception bal- Series 216, edited by D. F. Levia, D. E. Carlyle-Moses, and T. Tanaka, Springer-
ances increased transpiration in exotic ponderosa pine plantations compared Verlag, Heidelberg, Germany, 2011, pp. 371–388.
with native cypress stands in Patagonia, Argentina,” Ecohydrology, 4: 83–93, Ringgaard R., M. Herbst, and T. Friborg, “Partitioning forest evapotranspi-
2011. ration: interception evaporation and impact of canopy structure, local and
Linsley Jr., R. K., M. A. Kohler, and J. L. H. Paulhus, Hydrology for Engineers regional advection,” Journal of Hydrology, 517: 677–690, 2014.
(3rd ed., McGraw-Hill College Division, New York, 1949, p. 508. Ritter A. and C. M. Regalado, “Roving revisited, towards an optimum
Llorens P. and F. Domingo, “Rainfall partitioning by vegetation under throughfall sampling design,” Hydrological Processes, 28: 123–133, 2014.
Mediterranean conditions. A review of studies in Europe,” Journal of Rutter A. J., K. A. Kershaw, P. C. Robins, and A. J. Morton, “A predictive
Hydrology, 335: 37–54, 2007. model of rainfall interception in forests I. Derivation of the model from
McMinn R. G., Water Relations and Forest Distribution in the Douglas-fir observations in a plantation of Corsican pine,” Agricultural and Forest
Region on Vancouver Island, Publication 1091 of the Canadian Department of Meteorology, 9: 367–384, 1971.
Agriculture, Science Service, Forest Biology Division, Ottawa, ON, 1960, p. 71. Singh V. P., Hydrologic Systems. Watershed Modelling, Prentice Hall,
Miralles D. G., J. H. Gash, T. R. H. Holmes, R. A. M. de Jeu, and A. J. Inglewood Cliffs, NJ, Vol. II, 1989, p. 448.
Dolman “Global canopy interception from satellite observations,” Journal of Stewart, J. B., “Evaporation from a wet canopy of a pine forest,” Water
Geophysical Research: Atmosphere, 115: D16122, 2010. Resources Research, 13: 915–921, 1977.
Mishra S. K. and V. Singh, “Soil Conservation Service Curve Number (SCS- Valente, F., J. S. David, and J. H. C. Gash, “Modelling interception loss for
CN) Methodology,” Water Science and Technology Library, Springer, The two sparse eucalypt and pine forests in central Portugal using reformulated
Netherlands, Vol. 42, 2010, p. 516. Rutter and Gash analytical models,” Journal of Hydrology, 190: 141–162, 1997.
Muzylo A., P. Llorens, F. Valente, J. J. Keizer, F. Domingo, and J. H. C. Gash, Viessman Jr., W. and G. L. Lewis, Introduction to Hydrology, 5th ed.,
“A review of rainfall interception modelling,” Journal of Hydrology, 370: 191– Prentice Hall–Pearson Education, Upper Saddle River, NJ, 2002, p. 612.
206, 2009. Wallace J. and D. McJannet, “On interception modelling of a lowland
Onstad C. A., “Depressional storage on tilled soil surfaces,” Transaction of coastal rainforest in northern Queensland, Australia,” Journal of Hydrology,
the American Society of Agricultural Engineers, 27: 729–732, 1984. 329: 477–488, 2006.
Pereira F. L., J. H. C. Gash, J. S. David, and F. Valente, “Evaporation of Zimmermann A. and B. Zimmermann, “Requirements for throughfall
intercepted rainfall from isolated evergreen oak trees: do the crowns behave monitoring: the roles of temporal scale and canopy complexity,” Agricultural
as wet bulbs?” Agricultural and Forest Meteorology, 149: 667–679, 2009. and Forest Meteorology, 189–190: 125–139, 2014.
43_Singh_ch43_p43.1-43.4.indd 4 8/22/16 2:11 PM

Chapter 44
Watershed Geomorphological
Characteristics
BY
FAITH A. FITZPATRICK
ABSTRACT The descriptive and conceptual models (Davis, 1899) became process based
This chapter describes commonly used geomorphological characteristics that (Gilbert, 1877; Horton, 1945), which then ultimately developed into dynamic
are useful for analyzing watershed-scale hydrology and sediment dynamics. It quantitative geomorphology (Strahler, 1952a; 1957).
includes calculations and measurements for stream network features and Schumm and Lichty’s (1965) framework for time, space, and causality
areal basin characteristics that cover a range of spatial and temporal scales encompasses how watershed geomorphological characteristics fit into a range
and dimensions of watersheds. Construction and application of longitudinal of temporal scales (Table 44.1). For example, if the time span of interest is a
profiles are described in terms of understanding the three-dimensional devel- few hundred years (between geologic scales and typical engineering scales),
opment of stream networks. A brief discussion of outstanding problems and watershed runoff and sediment yield (as affected by geologic setting, climate,
directions for future work, particularly as they relate to water-resources vegetation, and relief) influence drainage network morphology, hillslope
management, is provided. Notations with preferred units are given. morphology, channel hydrology, and hydrology at a point along a stream
(water and sediment discharge). At engineering scales, the drainage network
44.1 INTRODUCTION AND LITERATURE REVIEW morphology, hillslope morphology, and channel morphology influence water
This chapter represents an update to Strahler’s description of watershed and sediment discharge at a particular point along a stream.
geomorphological characteristics in “Quantitative Geomorphology of At the time of the second edition of the Handbook (1964), another treatise
Drainage Basins and Channel Networks” (Strahler, 1964). Watershed geomor- by Leopold, Wolman, and Miller (1964) was published. Leopold et al. (1964)
phological characteristics are formed by the downhill movement of water and is still considered a key reference book for fluvial geomorphologists, espe-
how the land is eroded into different forms or morphology because of this cially for its treatment of “The Drainage Basin as a Geomorphologic Unit”
movement. Strahler’s career encompassed a time period during which the (Chap. 5), which, combined with Strahler’s and Horton’s work forms the basis
methods of quantitative geomorphology emerged, and advanced fluvial of much of the watershed measurements that are still done today. A quote
geomorphology from “purely qualitative and deductive” to “rigorous quanti- from Leopold, Wolman, and Miller (1964) gives that sense of interconnection
tative science capable of providing hydrologists with numerical data of of watersheds with the streams that drain them as well as interconnection of
practical value” (Strahler, 1964, pp. 4–40). Strahler’s quantitative geomorphol- geomorphological principles with river and hydraulic engineering:
ogy had its roots in Horton’s (1932; 1945) quantification of geomorphological “Each denudational agent, depending upon its density, gradient, and mass
characteristics of drainage basins, stream networks, and erosion processes at a particular place, is capable of applying a given stress on the materials
over time. Previous to the quantitative approach, landscape evolution models available. A certain amount of work may be performed by the application of
were more qualitative or descriptive in nature, such as William Morris Davis’s this stress, and the results of this work are the landforms that we see
analysis framework and classification of cycles of erosion (Davis, 1899). developed in various parts of the world.” (p. 4)
Table 44.1 Factors that Influence Watersheds and Stream Network Characteristics at Three Temporal Scales of Decreasing Duration
Drainage basin variables Cyclic (geologic, millions of years) Graded (centuries to millennia) Steady (days to decades)
Time I NA NA
Initial relief I NA NA
Geology (lithology, structure) I I I
Climate I I I
Vegetation (type and density) D I I
Relief or volume of system above base level D I I
Hydrology (runoff and sediment yield per unit drainage area) D I I
Drainage network morphology D D I
Hillslope morphology D D I
Channel morphology D D I
Hydrology (discharge of water and sediment from system) D D D
I, independent; D, dependent; NA, not applicable.
[Source: Schumm and Lichty, 1965] 44-1
44_Singh_ch44_p44.1-44.12.indd 1 8/22/16 2:13 PM

44-2 Watershed Geomorphological Characteristics
A holistic fluvial systems approach is now popular that interconnects drainage patterns, and three-dimensional development and evolution of
multiple spatial and temporal scales of quantitative and qualitative descrip- stream networks. Watershed geomorphological characteristics are completely
tors and physical mechanistic processes (e.g., see Sear et al., 2003; Brierley dependent on data from maps or remote sensing techniques, thus it is of
and Fryirs, 2005; Rosgen, 2006). This systems approach has a backbone of utmost importance to document the spatial resolution, scale, and source of all
classical geomorphological concepts including geomorphological work, map data. Comparative studies cannot be done successfully unless the same
equilibrium, thresholds, feedbacks, and relaxation and response time associ- method of measurement, map source, and scale are used.
ated with watershed changes, sediment dynamics, and disturbance mecha-
nisms (Mackin, 1948; Chorley, 1957, 1962; Schumm, 1956, 1973; Schumm 44.2.1 Watershed Boundary Delineation
and Lichty, 1965; and Bull, 1991). Currently, we are at a state of integration of A watershed is defined as the area of land that drains water, sediment, and
scales of geomorphologic processes, knowing that we cannot treat watersheds dissolved materials toward a common outlet. The term watershed carries the
and stream networks as simply as the laws of gravitational and molecular same meaning as “drainage basin” and “catchment,” and usage of one term
stresses and chemical processes (Strahler, 1952a). In addition, geomorpho- over another is determined by regional and disciplinary preferences. The
logical principles help to link hydraulic and chemical-based processes with bounds of a watershed are determined by a drainage divide, also known as an
the larger spatial and temporal variations that shape the development of interfluve. Topographic-derived drainage boundaries are delineated by draw-
watersheds and fluvial landforms from perturbations in climate, tectonics, ing lines connecting the highest points of elevation along a land surface that
base level, and humans (Church, 2002; Brierley and Fryirs, 2005). separate diverging flow paths. Usually, this looks like a ridge of high ground
Watershed geomorphological characteristics are measured for multiple that falls off either side into a series of ravines and valleys.
reasons that support theoretical, empirical, and statistical analyses. An The delineation can be done by hand with maps of elevation contour lines
example of theoretical analyses—formulation and testing of specific hypoth- or automatically with a geographic information system (GIS) and digital ele-
eses based on established principles—is Schumm’s (1973) paper on geomor- vation model (DEM) data. Historically, watershed areas were measured with
phological thresholds for gully formation and river meandering. Empirical the use of planimeters (http://persweb.wabash.edu/facstaff/footer/Planimeter/
applications involve collection of field data to form mathematical relations planimeter.htm). In the United States, methods were standardized by the U.S.
between watershed variables and stream characteristics, such as Emmett’s Geological Survey (USGS) to measure the area of watersheds upstream of
(1975) classic work on the relations of stream network characteristics to streamgaging stations. An example of drainage divide lines drawn on a
channel characteristics for the Upper Salmon River Area, Idaho,” Watershed 1:24,000 USGS topographic quadrangle map is provided in Fig. 44.1. Before
geomorphological characteristics have been used successfully to help GIS, each USGS Water Science Center maintained an official set of paper
describe watershed hydrology and sediment delivery across varying physi- copies of these quadrangle maps with drainage boundaries.
ography (Langbein and Schumm, 1958) and also have been used in models Topographic divides are easiest to delineate in well-developed, geologically
of surface water sediment and nutrient modeling (Robertson et al., 2009). old watersheds. In geologically young landscapes, wetlands and interconnected
Other applications using watershed characteristics include studies of chan- lakes may form watershed divides making it difficult to determine the flow
nel and hillslope instability, effects of dams and dam removals on fluvial direction. In urban areas, storm sewers often transcend topographic boundaries
systems, flood and debris flow hazards, wetland classification (Brinson et al., and additional data on storm-sewer networks are needed. In arid areas, water is
1995), groundwater susceptibility (http://www.dnr.state.mn.us/whaf/about/ often artificially routed through pipes or canals across topographic divides into
scores/geomorphology/gw_contamination.html), ecology (Gordon et al., adjacent watersheds for storage or supply for agricultural and industrial uses
1992), habitat studies (Fitzpatrick et al., 1998), and climate- or vegetation- and drinking water supplies. Some watersheds contain depressions or closed
change effects on river meandering, incision, and sedimentation (Knox et basins with no surface water outlet, such as kettle ponds in glaciated terrain or
al., 1974; Knox, 1977; Fitzpatrick et al., 1999; Fitzpatrick and Peppler, 2010). playas in the arid western United States. These areas are referred to as noncon-
tributing areas by the USGS, meaning they do not directly contribute to surface
44.2 WATERSHEDS AND DRAINAGE NETWORKS water drainage. In sandy areas, infiltration rates may exceed any known
precipitation rates, leaving no trace of surficial drainage because all precipitation
Quantitative measurements and calculations of watershed geomorphological either evaporates or infiltrates to groundwater aquifers.
characteristics described in this section include watershed boundary Watershed boundaries are not just only delineated for surface-water drain-
delineations, stream network characteristics, watershed areal characteristics, ages, but can also be defined along groundwater divides, which may be different
Figure 44.1 Example of U.S. Geological Survey traditional method of delineating drainage divides on 1:24,000-scale topographic
maps. Delineations with crosshatching are noncontributing areas.
44_Singh_ch44_p44.1-44.12.indd 2 8/22/16 2:13 PM

Watersheds and Drainage Networks 44-3
Table 44.2 Hydrologic Unit Codes and Levels of Classification for to trace the stream in upstream or downstream directions. The blue lines
U.S. Watersheds originally were drawn with 5- to 10-ft contour intervals of elevations as a base
map and are a simplification of all the possible drainage lines that could be
HUC code number of digits Description drawn from contour crenulations, especially for ephemeral streams. A stream
2 Region (first level) network generated in a GIS from airborne light detection and ranging
4 Subregion (second level)
(LiDAR)-based 1-ft contours may be denser, as shown in Fig. 46.2, and help-
ful for studies of small watersheds; however, additional time and resources are
6 Basin (third level) required for field verification and establishing continuous flow networks. The
8 Subbasin (fourth level) NHD remains the preferred dataset in the United States for many agencies on
10 Watershed (fifth level)
which regional stream network geomorphological measurements are made
because of its wide availability and consistent national coverage, consistent
12 Subwatershed (sixth level) flow networks, and comprehensive history of field verification and updates.
[Source: http://nhd.usgs.gov/wbd_facts.html]
Stream Order
than surface-water divides because they depend on the subsurface horizontal
and vertical extent of the aquifer and groundwater gradients (Fitzpatrick et al., A hierarchical system for describing the spatial position of a stream in a stream
2015). It is not uncommon for the surface water and groundwater divides to be network was introduced in the United States by Horton (1945) and later
different in areas with nonhomogeneous rock types and glacial deposits. modified by Strahler (1952b; 1957; Fig. 44.3). The Laws of Horton’s stream
Topographic drainage divides form the framework of the hydrologic unit orders summarize mathematical empirical relations that describe how stream
code (HUC) system in the United States and the Caribbean, which uses a networks are arranged and related to hydrologic characteristics (Horton, 1945;
national hierarchical system of stream categorization into six levels as shown Melton, 1959). The mathematical relations reflect the watershed size; geo-
in Table 44.2 (Seaber et al., 1987). The newest and most up-to-date delinea- graphic relations among the number, length, and size of channels; stream
tions for the United States are included in the watershed boundary dataset branching; and physiography of the landscape drained by the stream network.
(WBD). The WBD contains boundaries for the 8-, 10-, and 12-digit HUCS These empirical relations were especially useful before the use of GIS for mea-
and is available at http://www.nrcs.usda.gov/wps/portal/nrcs/main/national/ suring the number of streams and stream lengths. Perennial, intermittent, and
water/watersheds/dataset/. ephemeral streams can be assigned a stream order number.
In Strahler’s stream ordering system (Fig. 44.3a), the smallest tributaries are
44.2.2 Stream Network Characteristics given an order number of 1. Where two streams of order 1 join, a stream of order
This section describes four commonly used stream network measurements of 2 is formed. Where two streams of order 2 join a stream of order 3 is formed,
stream order, bifurcation ratio, drainage density, and overland flow. A brief and so on. If a stream of order 1 joins a stream of order 2, the downstream seg-
overview of drainage patterns is included. ment stays at order 2. The order of the stream at the outlet is the “Principle
The first step before measuring geomorphological characteristics of stream Order.” Thus, in the example shown in Fig. 44.3, the principle order is 3.
networks is mapping the flow patterns of water, or the science of hydrography. Once the ordering of a stream network is defined using the Strahler
If not already available, the stream network is usually mapped during the method, each of the orders can be summarized for the watershed in terms of
delineation of drainage divides. It is important to know the source and the number of stream segments in each order No (Table 44.3). As the order
resolution of the line work representing the stream network, along with the increases, the number of stream segments, No, decreases as shown in Fig. 44.4.
assumptions that went into their mapping. The number of segments in the principle stream order is always equal to 1.
In the United States, stream networks are commonly derived from blue Conversely, mean stream length and drainage area increase in an orderly
lines on USGS 7.5-min 1:24,000-scale topographic quadrangle maps. The fashion that can be explained by a power function.
quadrangle maps show perennial (solid blue lines) as well as ephemeral Another ordering system uses the Shreve method (1967) and is typically
(dashed blue lines) streams. These streams form the basis for the national used to describe the relation of a stream to the next waterbody downstream.
hydrography dataset (NHD) (http://nhd.usgs.gov/; Simley and Carswell, 2009). This measurement is important in the aquatic ecology fields for explaining
Each line segment in the NHD contains networking codes that allow a user longitudinal connection among organisms whose habitat requirements differ
(a) (c)
(b) (d)
Figure 44.2 Examples of detailed stream networks delineated from 1-ft LiDAR overlain with elevation
contours from different map sources for a 1.5-km reach of the Bark River, Wis. (A) 1:24,000-scale topo-
graphic map, where the stream was represented by a single blue line with no tributaries. (B) Contours from a
30-m DEM that are too coarse to show a valley. (C) Contours from a 10-m DEM show a valley. (D) LiDAR
based 1-ft contours show multiple tributaries and valleys.
44_Singh_ch44_p44.1-44.12.indd 3 8/22/16 2:13 PM

Strahler 1
1 (b)
1
1 (a)
1 1
1 2 2
1
2
1 1
1
2 3
1
(a)
Rb = 5
Shreve 1 1
1
2 B
1
1 1 Shreve number = 12
2 A
1 3 Q
3 1
2 4
1
5 9 1 Rb = 17 t
1
2 11 Figure 44.5 Bifurcation ratio examples. Stream networks with (A) high bifurcation
1 12 ratios have low flood peaks compared to (B) stream networks with low bifurcation
(b) ratios.
Figure 44.3 Stream ordering examples: (A) Strahler (modified from Strahler, 1954,
1957) and (B) Shreve (modified from Shreve, 1967). It is usually calculated for the second highest stream order or the mean
value of all orders measured. The number is usually between 2 and 4, and
indicates the peakedness of runoff. Low ratios are associated with flashy
Table 44.3 Strahler Stream Order and Number of Stream Segments hydrographs with high flood peaks, as shown in Fig. 44.5 for a small
Shown in Example Stream Network in Fig 44.3a
watershed. When the log of the number of streams is plotted against order, the
Stream order Number of stream segments, No Bifurcation ratio, Rb relation is usually described by a linear regression equation.
1 12 3 Drainage Density
2 4 4 Drainage density reflects runoff characteristics; for example, dense stream
3 1 networks have flashy high peak flows (Horton, 1945; Leopold et al., 1964)
because the reduced friction in closely spaced channels produce faster flows
compared to rough terrain and vegetation encountered by overland flow in
over life stages. Each stream segment is counted and all the segments are less dense networks. The density of a stream network (D) is the quotient of
summed at a confluence (Fig. 44.3b). A downstream link number for a stream the cumulative length of streams (L) and the total drainage area (A):
segment is the Shreve order calculated for the next downstream stream
segment. In this example, the tributary with a Strahler stream order of 1 has a D = ∑L/ A
downstream link number of 12, indicating that it drains to a much larger river. Drainage density has units of length per unit area such as km/km2.
Bifurcation Ratio Drainage density is a result of a watershed’s geology, climate, time, and human
impacts (Table 44.4). Two watersheds may have similar drainage densities but
The bifurcation ratio, Rb, is the ratio of the number of stream segments of any
still have different bifurcation ratios, which will further define the hydrologic
order to the number of stream segments of the highest order (Horton, 1945;
characteristics (Leopold et al., 1964). Because the calculation involves stream
Strahler, 1957).
Rb = N u /N u+1 length, it can be quite tedious to calculate by hand, especially for large
watersheds, but relatively easy to do in a GIS.
– –
Nu = 615e–1.33u Lu = 0.21e0.97u Au = 0.18e1.48u
100 RB = 3.78 RL = 2.64 100 RA = 4.39
Mean drainage area (km2)
Mean stream length (km)
50 10 50
Number of streams
10 10
5 5
1
1 0.5 1
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
(a) Stream order (b) Stream order (c) Stream order
– –
Nu = Number of streams Lu = Average length of Au = Average catchment area
RB = Bifurcation ratio streams of order u of streams of order u
RL = Stream length ratio RA = Catchment area ratio
Figure 44.4 Horton’s Laws of stream networks for (A) number of streams, (B) mean stream length, and (C) mean drainage area (from Horton, 1945).
44_Singh_ch44_p44.1-44.12.indd 4 8/22/16 2:13 PM

Table 44.4 Driving Variable Effects on Drainage Density Leopold et al. (1964) and Hack (1957) found in calculations of many
stream lengths and drainage areas that total stream length can be estimated
Watershed characteristic Low density High density from drainage area by the relation (in English units of square miles for area):
Infiltration High Low
Geology Resistive Erodible ∑ L = 1.4 A∧ 0.6
Age Young Old This equation was more commonly used before GIS was widely available
Climate Humid Arid because it saved time by not having to measure stream lengths by hand. In
Flood size Small Large
general, there is about 1.3 km of stream length for every square km of
drainage area. Emmett’s (1975) analyses for tributaries of the East Fork
Sediment production Low High Salmon River yielded a similar relation as shown in Fig. 44.6a.
Relief Gentle Steep
100
10
Stream length, in miles
L = 1.5 DA0.55
Explanation
Compilation by stream order
Compilation by drainage area
0.1
0.1 1 10 100 1000
(a) Drainage area, in square miles
10,000
Bankful discharge
QB = 28.3 DA0.69
Discharge: cubic feet/second
1000
Average annual
discharge
100
10
1 10 100 1000 10,000
(b) Drainage area: square miles
Figure 44.6 Relations among drainage area and (A) stream length (from Emmett, 1975) and (B) discharge for tributaries to
East Fork Salmon River (from Emmett, 1975; Leopold, 1994).
44_Singh_ch44_p44.1-44.12.indd 5 8/22/16 2:13 PM

Length of Overland Flow climate (Leopold et al., 1964). Like other measurements, map resolution
Once the total drainage area and cumulative length of streams have been affects the accuracy of a drainage area calculation, but the most important
computed, the average length of overland flow in a watershed can be calcu- factor is the correct identification of the drainage divide, and the knowledge
lated (Horton, 1945). The length of overland flow is used in surface-water of other factors that may affect the contributing area of a watershed, such as
modeling to help predict the quantity of water needed to exceed a threshold diversions and withdrawals for irrigation and drinking water supplies; and
of erosion that begins rill and channel development. It is also related to the storm sewer networks in urban areas. The presence of these features is
constant of channel maintenance (C), the reciprocal of drainage density which important because many watershed geomorphologic measurements are
has units of length (km), which is the average distance between channels. normalized by drainage area before comparisons are done among streams.
One-half of channel maintenance length is the length of overland flow. Basin Length and Shape
C = A/∑L There are many ways to describe the areal shapes and dimensions of water-
Drainage Patterns sheds. For basin length, upon which many descriptors rely, one of the most
common methods is to measure the distance along a line centered through
Drainage patterns of a watershed reflect rock type, geologic structure and the watershed from its mouth to where the main channel meets its drainage
tectonics, and fracturing patterns (Howard, 1967). Some common drainage divide (Schumm, 1956; Gardiner, 1975; Harvey and Eash, 1996; Fitzpatrick
patterns are shown in Fig. 44.9 and originate from Howard’s (1967) detailed et al., 1998; Fig. 44.7). Sometimes a straight line from the mouth to the high-
overview. Common terms used to describe drainage patterns include est point in the basin is used, but this method may not be as accurate because,
dendritic (tree-like), rectangular or trellis, radial, and parallel (Horton, 1945). depending on the orientation of the basin, this line may fall outside of the
If the geologic setting is uniformly resistant to erosion, the land surface is actual drainage basin, and is usually shorter than the actual basin length.
gently sloped, or surficial deposits are horizontally layered, a dendritic pattern Basin shape has many descriptors including form, circularity, elongation,
develops, similar to the branching of a tree. Parallel patterns develop on and compactness. These features, among others equally important, help to
steeply sloping surfaces or in areas with parallel elongate landforms, where describe the geologic setting and resultant watershed hydrology and sediment
long tributaries are oriented parallel to, and eventually join, the parent stream regime (Horton, 1932; Morisawa, 1958; Strahler, 1964). Basin shape was
from opposite sides. Trellis drainages develop on land surfaces with nonuni- described by Horton (1932; 1945) as usually a pear-shaped ovoid. Horton
form geology, alternation of resistant and less-resistant rocks, dipping, folded thought that the predominance of this pear-shape form was evidence that
or faulted rocks, or exposed lake floors or sea beds with beach ridges. Rectan drainage basins are formed by sheet erosion processes acting on an inclined
gular patterns indicate repeated jointing and fracturing at right angles. surface (Strahler, 1964). Horton defined basin form (Rf) as the quotient of the
A radial pattern would be found on a volcano or dome with uniform geology, drainage area (A) divided by the basin length squared (L2b ):
whereas an annular pattern would develop on a dome with concentric rings
of usually layered sedimentary bedrock units with differing resistance to ero- R f = A/L2b
sion. Multibasinal patterns develop on glacial deposits with hummocky
surfaces, limestone solution or karst landscapes, and permafrost. Contorted The units of basin shape are dimensionless, but usually the area is in km2 and
drainages represent multiple tectonic events that altered the course of already the basin length is in km. This measurement is the same as the elongation
developed drainage patterns, metamorphic rocks, dikes, and veins. ratio (Re; Fryirs and Brierley, 2013):
44.2.3 Watershed Areal Characteristics Re = A0.5 / Lb

Watershed areal characteristics are major drivers in forming stream network pat- A ratio of 1.0 is a round basin. A ratio closer to 0.6 is an elongated basin, and
terns, watershed hydrology, longitudinal sediment dynamics, and ultimately slope the more elongate a basin the slower the runoff and less flashy the floods
and channel morphology over a range of time scales (Leopold et al., 1964; Schumm compared to a more round basin. Elongated basins tend to have relatively
and Lichty, 1965; Fryirs and Brierley, 2013). These characteristics are easily calcu- high bifurcation ratios as well (Fig. 44.5).
lated in a GIS with high-resolution DEMs. Traditionally, characteristics were Another basin shape measurement, the circularity ratio (Rc), compares a
calculated by hand from 1:24,000-scale USGS quadrangle maps. basin’s drainage area (A) to the area of a circle (Ac) with the same
Drainage Area circumference (Miller, 1953; Fig. 44.8):
The drainage area forms one of the most important features of a watershed, as Rc = A/Ac
it provides a scaling factor for many other characteristics related to hydrology
and sediment (Leopold et al., 1964; http://streamstatsags.cr.usgs.gov/ss_defs/ Using a circle as a standard reference for drainage basin, however, is a less-
basin_char_defs.aspx). Streamflow characteristics, including base flow, favorable method because of the sharply defined point at the mouth (Strahler,
bankfull flows, and flood frequencies are empirically related to drainage area 1964). In general, watersheds with low ratios of about 0.4 are more elongate
(Fig. 44.8b), usually by physiographic and climatic regions (Reis, 2006) and shapes are likely controlled by geologic setting and structure. More round
(http://water.usgs.gov/osw/streamstats/ssinfo1.html). For example, stream- basins have ratios of 0.6–0.7 (Fryirs and Brierley, 2013). The compactness
flow can be estimated using drainage area as follows: ratio (Cc) is related to Rc in that it compares a basin’s perimeter to the
Q = aAb circumference of a circle of equal area (Reddy, 2005):
where Q is discharge or streamflow in units of volume per unit time, A is
Cc = P / (4π A)
drainage area, and coefficient a and exponent b relate to the physiography and
Highest altitude
Longest from mouth,

following main
streamline
Most direct length from

mouth to highest point
Figure 44.7 Basin length measurements and definitions.
44_Singh_ch44_p44.1-44.12.indd 6 8/22/16 2:13 PM

Basin Relief and Slope

C2
C1 Basin relief (H) is the measure of the difference in altitude between the
highest point along a drainage divide and the altitude at the river mouth.
Figure 44.7 shows how the highest point in a watershed along a drainage
divide might be slightly offset from the extension of the main streamline used
to measure basin length. The basin relief ratio (Rh) is a dimensionless value of
the quotient of basin relief divided by basin length (Lb), which provides a
measure of how fast the elevation drops across a watershed:
Rh = H/Lb
The relief ratio is a measure of the steepness of a watershed and is an indicator

of the potential erosion of the landscape (Strahler, 1964). Schumm (1956)
C1 = C2 used the same basin length measurement used for basin shape ratio—the
overall longest distance along the basin length from the mouth to the drainage
Figure 44.8 Comparison of a common pear-shaped watershed with a circle with the divide following the principal drainage line. Because the measurements
same circumference (C) (adapted from Fryirs and Brierley, 2013).
(a) Dendritic (b) Parallel
(c) Trellis (d) Rectangular
(e) Radial (f) Annular
(g) Multi-basinal (h) Contorted
Figure 44.9 Examples of drainage patterns (from Howard, 1967). AAPG © 1967. Reprinted by permis-
sion of the AAPG whose permission is required for further use.
44_Singh_ch44_p44.1-44.12.indd 7 8/22/16 2:13 PM

included in the basin relief ratio make up the sides of a right triangle, it can rills, water starts to flow slightly in a lateral direction that eventually erodes the
also be referred to as basin slope. As part of estimating streamflow statistics at divides between rills and forms a master rill that is lower in elevation than
ungaged streams in the United States using a National Streamflow Statistics surrounding rills. With additional erosion from overland flow, the divides
Program (Reis, 2006; Turnipseed and Ries, 2007), channel slope (S) is calcu- between the rills break down, and relatively smooth surfaces are formed
lated by dividing the elevation difference by stream length between points between rills or channels in a process called “cross grading” (Horton, 1945).
that are 10 and 85% of the total stream length. Permanent and more continuous channels form and deepen as flows continue
44.2.4 Drainage Development and Three-Dimensional to erode in upstream and lateral directions and capture smaller channels.
Evolution of Watersheds Glock (1931) described the development of stream networks as a concep-
tual model of initiation, growth, extension, and elaboration, followed by a stage
The descriptions of watershed geomorphological characteristics in the afore- of integration marked as a partial simplification of the network as time
mentioned sections have focused on two-dimensional, spatially related char- progresses (Clifford, 2011). This process describes the branches and tree-like
acteristics, and how they might differ in terms of the major driving framework of most drainage systems, with many small branches connected to
factors—climate, geology and tectonics, human disturbance, and time. This fewer large branches. Horton’s (1945) laws on stream orders reflect this process.
section expands on the temporal and spatial scales for describing drainage Leopold et al. (1964) noted the random nature of cross grading and micropiracy,
development and evolution. and that most landscapes have inconsistencies in erodibility and topography,
Drainage Development and Evolution which adds variability to the regular geometric patterns proposed by Horton.
Stream networks develop where erosive forces of water overcome resistant These same principles that reflect the conservation of energy and the effects of
forces (friction and gravitational) of particles (Horton, 1945; Leopold et al., gravity are responsible for the relation between slope and stream order, where
1964). The fundamental requirement for channel initiation is that precipitation larger stream orders have decreasing slopes (Leopold et al., 1964).
exceeds infiltration and produces overland flow. Rills form almost immediately The Longitudinal Profile
when overland flow erodes soil on bare hillslopes leaving small, mostly
temporary, intermittent channels that are no more than a few centimeters wide Longitudinal profiles offer vertical views of stream network development and
and deep. When overland flow further reaches a depth that fills and overtops evolution (Fig. 44.10) and help to describe the continuum of streamflow and
R Reference level
Hg O Elevation of stream head
Convex up
Yo
Concave up A1
Elevation
A2
Local base level M
Yc Sea level N
(a)
500
Sandy glacial till,
Marengo Bad river poorly developed
450
river drainage network, no valley
400
Altitute, in meters
350
Bedrock outcrop
300
Sandy post-glacial
shorelines,
250
entrenched valley
200
Clay plain,
entrenched/alluvial valley
150
120 100 80 60 40 20 0
(b) River kilometer, from mouth
Figure 44.10 Features and examples of longitudinal profiles. (A) Characteristics of longitudinal profiles. (B) Longitudinal
profile of the Bad River and its tributary the Marengo River. The streams are steep in the upper part of their profiles but
have poorly developed networks with common wetlands that slow sediment production and reduce streamflow flashiness.
Most of the sediment in the watershed is generated from 60-m high eroding valley sides from river kilometer 55 to 65.
From river kilometer 55 to the mouth, the river intersects highly erodible clayey glacial lacustrine deposits.
44_Singh_ch44_p44.1-44.12.indd 8 8/22/16 2:13 PM

Original channel
New channel
n
Aggradation/new alluvium sio Drainage
Ero
divide
Aggradation Tip of new
channel
Tip of original channel
Stage I premodified Stage II channelized Stage III degradation
Stage IV degradation Stage V aggradation and Stage VI quasi-equilibrium

and widening widening
Aggradation/new alluvium
Degradation/erosion Stages I, II
Stage III
Stage V Stage IV
Stage VI
Aggraded Aggradation
Figure 44.11 Spatial and temporal changes in stream networks and effects on longitudinal profiles and
erosion and deposition: (A) headward expansion of a stream network from land clearing and erosion
(modified from Strahler, 1958), (B) temporal changes due to headward knickpoint migration along chan-
nelized streams (from Simon and Hupp, 1986).
sediment dynamics (erosion, transfer, and storage) from headwaters to the flow and sediment load provided by the watershed. Equilibrium profiles are
mouth of a watershed (Schumm, 1977). The profile consists of a graph expected to have consistent increases in discharge and channel size with
showing elevation changes along a length of a stream. The elevation of its decreases in slope and bed material size (Fryirs and Brierley, 2013). Gilbert
mouth (M)—at a lake, ocean, or the valley bottom of a larger river—is called (1877) theorized that upward concavity could be explained by increasing
its local base level, which may be at various altitudes above sea level (Z) and discharge, and as discharge increases and channel cross-sectional area
may vary over time. Similar to Horton’s mathematical relations and Strahler’s increases, along with relative decreases in frictional losses, the channel is able
stream orders, characteristics of discharge, slope, and sediment characteristics to transport more bedload on a lesser slope. Steep long profiles result in
progress along a predictable longitudinal continuum that is useful for under- flashier hydrographs than profiles with low relief.
standing aquatic ecology/physical habitat connections (Allan and Castillo, Variations in lithology, geologic structure, and tectonic uplift can disrupt
2007) as well as river engineering and management (Thorne et al., 1998). the profile (Hack, 1957; Radoane et al., 2003). Convex sections of the longitu-
Streams generally develop smooth concave up profiles that reflect the over- dinal profile with steep slopes are usually where the stream intersects erosion
all longitudinal energy expenditure of a river system (Hack, 1957). Streams resistant bedrock. These may be local sources of sediment or geologic controls
with smooth concave up profiles are considered graded or in equilibrium— on sediment supply. Knickpoints are discontinuities in the longitudinal pro-
the channel slope has adjusted to be able to efficiently transport the stream- file, the most obvious being falls over bedrock outcroppings. Tectonic uplift,
44_Singh_ch44_p44.1-44.12.indd 9 8/22/16 2:13 PM

tributary junctions, and changes in base level can also cause profile irregu- geospatial data more accessible and usable, especially for hazards and human
larities. Dams, culverts, and other manmade structures function as artificial safety issues (https://daac.ornl.gov/spatial_data_access.shtml).
base level and grade controls. Young (generally less than 15,000 years old) In the United States, the widely available and georeferenced NHD digital
tributaries draining into the Great Lakes in the United States have breaks in hydrography data are based on blue lines (dashed and solid) drawn by cartog-
their somewhat convex longitudinal slopes coincident with ancient glacial raphers on the nationally available USGS 7.5-min 1:24,000-scale quadrangle
lake shorelines (Fitzpatrick et al., 1999; Fitzpatrick et al., 2006; Fig. 44.10). maps. Leopold (1994) summarized the problems inherent in a hydrography
Underlying stratigraphy not present on the land surface may be affecting dataset based on “….blue lines that do not reflect any statistical characteristics
sediment characteristics, channel slopes, and profile shapes, even in alluvial of streamflow occurrence… Rather, the choice of what is to be shown as an
sections. For the Bad River example in Fig. 44.10, the lower knickpoint at interrupted blue line is based on consistent portrayal...” Hydrology and geo-
river kilometer 35 is a set of falls over a bedrock ridge of Precambrian sand- morphology students and young practitioners that use the U.S. digital
stone within thick glaciolacustrine clay deposits. hydrography may not have seen or held a paper copy of a 7.5-min quadrangle
The shape of longitudinal profiles can be described quantitatively in mul- map, let alone thought about why a streamline exists in one swale and not
tiple fashions. Regression equations can be constructed that are based on another. Even so, as mentioned earlier, the NHD remains the preferred data-
simple linear, logarithmic, exponential, and power functions (Strahler, 1964; set on which regional stream network geomorphological measurements are
Radoane et al., 2003). Another method that is useful for comparing longitu- made by many agencies in the United States because of its consistent flow
dinal profiles on a per unit basis (important for comparing streams of differ- network and thorough metadata, even though more detailed networks can be
ent sizes), is to determine the concavity ratio (Ca) of two measured areas on a quickly generated from remotely sensed DEM with airborne LiDAR sensors.
longitudinal profile graphic: Like all geodetic data, however, it is important to document the accuracy not
only of the LiDAR data, but also the generated DEM data (Hodgson and
Ca = A1 /A 2 Bresnahan, 2005).
Directions for future work for application of watershed geomorphological
where A1 is the numerically integrated area between the curve of the profile
characteristics lie in watershed and longitudinal connections of hydrology
and a straight line connecting the endpoints of the curve, and A2 is the area of
and sediment dynamics for river and floodplain restoration (Brookes, 1995).
a right triangle with the hypotenuse as the same straight line (Radoane et al.,
Emerging environmental issues need interdisciplinary research that includes
2003). The Ca ranges from 0 to 1.0. If the profile is close to a straight line the
a combination of hydrology, geomorphology, and ecology, including priority
value is close to 0 and if the profiles are more L-shaped the Ca is closer to 1.0.
areas of landscape ecology and landscape evolution modeling, ecological
In general, the older the river the more concave its longitudinal profile and
restoration, and climate change adaption (Renschler et al., 2007). Research
closer the Ca to 1.0.
needs include understanding problems of erosion and sedimentation, and
Accelerated Erosion and Drainage Evolution identifying and fingerprinting sources of fine-sediment-related contaminants
and turbidity. Additional research is needed in applying watershed geomor-
Over the time span of human alterations, the drainage network density can phological principles to ecological integrity studies, land conservation man-
change from having a dependent to independent relation with watershed agement, flood planning and management, fisheries and habitat enhancement,
hydrology and downstream streamflow and sediment characteristics and and integration of aesthetic and recreational opportunities. Watershed geo-
move from steady state to unstable conditions (Table 44.1). One prominent morphological characteristics are important in assessments of drinking water
example of this phenomenon is the denudation of vegetative cover for sources, flood reservoir lifespans, and navigation planning (Binnie Black and
watersheds that have been subject to intensive agricultural uses (Strahler, Veatch, 2001). Ultimately, human alterations to water generation and supply
1958). Agriculture-associated excess runoff from erodible soils and uncon- can interfere with the application of these measurement. This is especially
solidated deposits leads to gullying, headcutting, and permanent headward important in semiarid or mountainous regions, where interbasin transfers,
expansion of drainage networks (Lowdermilk, 1932; Happ et al., 1940; diversions, and withdrawals are common.
Strahler, 1958; Costa, 1975; Knox, 1977; Knox, 2006). Even if the vegetative Watershed responses to changes in sediment dynamics occur along a
cover was returned to its predisturbance condition, the expanded drainage much longer time continuum than hydrologic changes because of legacy
network remains. Knickpoints can continue to migrate gradually and epi- effects on the storage and transport of sediment from upstream to down-
sodically in an upstream direction for decades, centuries, and millennia. stream (Church, 2002). Common factors that affect sediment dynamics may
Accelerated erosion associated with headward expansion leads to accelerated positively interact or counteract each other, making it important to under-
aggradation in main stem valley bottoms (Fig. 44.11a) (Strahler, 1958). This stand historical as well as current perturbations, such as dam building,
phenomenon of a continuum of aggradation along a longitudinal profile is agricultural drainage, ditching, or mining (Table 44.5). Understanding time
also observed from headward incision caused by channelization (Fig. scales of change and channel evolution in terms of a continuum of sediment
44.11b) (Simon and Hupp, 1986). In the first example, aggradation is caused transport from headwaters to mouth areas is important for management
by simultaneous watershed-wide soil and gully erosion. In the second considerations that seek to reduce sediment loads by controlling upland
example in Fig. 44.11 for channelization, aggradation proceeds over time erosion (Sear et al., 2003). This is especially pertinent when it comes to fine
from downstream to upstream following upstream progression of erosion. sediment quantity and quality, and understanding stream network sources of
Methods for measuring erosion and aggradation along longitudinal profiles sediment, such as how increased runoff not only causes soil erosion but
are described in detail in Fitzpatrick (2014), and require field measurements irreversible stream network extension, gully erosion, and landslides
and coring. (Owens et al., 2005).
In conclusion, defining spatial, longitudinal, and vertical continuums in a
44.3 OUTSTANDING PROBLEMS AND DIRECTIONS FOR watershed is an important first step to defining any hydrologically related
FUTURE WORK problem or issue, no matter how local or widespread. Watershed
The use of watershed geomorphological characteristics for hydrologic analy- geomorphological characteristics will help give perspective, depth, and
ses continues to grow and expand as high-resolution DEM and stream understanding to complex hydrologic, water quality, and ecological studies.
network data become more readily available at a global scale. Watershed geo-
morphological characteristics are almost ubiquitously measured and
calculated from paper or digital maps, making it a necessity to keep track of
spatial scale, resolution, and source of the original data. Linking land conser-
vation, management, and planning to stream hydrology, sediment dynamics, Table 44.5 Common Watershed-Scale Perturbations
and contaminant transport can easily cover large regions and river basins Increased sediment supply Decreased sediment supply
through virtually unlimited computing power for geospatial analyses in a GIS.
Climate change (increase in rainfall) Climate change (decrease in rainfall)
However, with large computing power and massive digital datasets comes a
potential disconnect between the user and the limitations of the source data Upland drainage Dams/river regulation
and standards used for mapping. This is probably the largest outstanding Deforestation Reduced cropping and grazing
problem for watershed geomorphology—standard documentation and assem- Mining spoil additions Cessation of mining
bly of metadata associated with large geospatial data sets. Technical standard
methods for calculating metrics consistently are still needed (Brookes, 1995). Urban development (construction phase) Revegetation, paving over bare ground
The Open Geospatial Consortium, an international multidisciplinary consor- Agricultural drainage Sediment management
tium of private companies, government agencies, and academia, is working
[Source: Sear et al., 2003]
toward a consensus-based process and development of standards to make
44_Singh_ch44_p44.1-44.12.indd 10 8/22/16 2:13 PM

References 44-11
REFERENCES Horton, R. E., “Drainage basin characteristics,” Transactions of the American

Geophysical Union, 13: 350–361, 1932.
Allan, J. D. and M. M. Castillo, Stream Ecology, Structure and Function of
Horton, R. E., “Erosional development of streams and their drainage
Running waters, 2nd ed., Springer, Dordrecht, The Netherlands, 2007.
basins: hydrophysical approach to quantitative morphology,” Geological
Binnie Black and Veatch, Fluvial Design Guide: R&D Technical Report W109
Society of America Bulletin, 56: 275–370, 1945.
(W5-027 Version 11/01), Surrey, Redhill, 2001, https://www.gov.uk/govern-
Howard, A. D., “Drainage analyses in geologic interpretation: a summa-
ment/uploads/system/uploads/attachment_data/file/290794/str-w109a-e-e.pdf.
tion,” The American Association of Petroleum Geologists Bulletin, 51 (11):
Brierley, G. J. and K. A. Fryirs, Geomorphology and River Management: Appli-
2246–2259, 1967.
cations of the River Styles Framework, Blackwell Publications, Oxford, 2005.
Knox, J. C., “Human impacts on Wisconsin stream channels,” Annals of the
Brinson, M. M., F. R. Hauer, L. C. Lee, W. L. Nutter, R. D. Rheinhardt, R. D.
Association of American Geographers, 67 (3): 323–342, 1977.
Smith, and D. Whigham, “A guidebook for application of hydrogeomorpho-
Knox, J. C., “Floodplain sedimentation in the Upper Mississippi Valley—
logic assessments to riverine wetlands,” Technical Report WRP-DE-11, U.S.
Natural versus human accelerated,” Geomorphology, 79: 286–310, 2006.
Army Engineer Waterways Experiment Station, Vicksburg, MS. NTIS No. AD
Knox, J. C., P. J. Bartlein, and W. C. Johnson, “Environmental assessment of
A308 365, 1995.
sediment sources and sedimentation distributions for the Lake LaFarge
Brookes, A., “Challenges and objectives for geomorphology in U.K. river
watershed and impoundment,” Environmental Analysis of the Kickapoo River
management,” Earth Surface Processes and Landforms, 20: 593–610, 1995.
Impoundment: Institute for Environmental Studies, Report 28, University of
W. B. Bull, Geomorphologic Responses to Climate Change, Oxford University
Wisconsin, Madison, WI, 1974, pp. 77–116.
Press, New York, 1991, p. 326.
Langbein, W. B. and S. A. Schumm, “Yield of sediment in relation to mean
Chorley, R. J., Illustrating the Laws of Morphometry, Geological Magazine,
annual precipitation,” American Geophysical Union Transactions, 39:
Vol. 94, 1957, pp. 140–150.
1076–1084, 1958.
Chorley, R. J., Geomorphology and General Systems Theory, U.S. Geological
Leopold, L. B., A View of the River, Harvard University Press, Cambridge,
Survey Professional Paper 500-B, Washington, DC, 1962.
MA, 1994.
Church, M., “Geomorphologic thresholds in riverine landscapes,”
Leopold, L. B., M. G. Wolman, and J. P. Miller, Fluvial Processes in
Freshwater Biology, 47: 541–557, 2002.
Geomorphology, Dover Publications, New York, 1964.
Clifford, N., “Rivers and drainage basins,” edited by J. A. Agnew, and D. N.
Lowdermilk, W. C., “Acceleration of erosion above geologic norms,”
Livingstone, The SAGE Handbook of Geographical Knowledge, SAGE, London,
American Geophysical Union, 15th Transactions, 2: 505–509, 1932.
2011, pp. 502–527.
Mackin, J. H., “Concept of a graded river,” Geological Society of America
Costa, J. E., “Effects of agriculture on erosion and sedimentation in the
Bulletin, 59: 463–512, 1948.
piedmont province, Maryland,” Geological Society of America Bulletin, 86:
Melton, M. A., “A derivation of Strahler’s channel-ordering system,” Journal
1281–1286, 1975.
of Geology, 67: 345–346, 1959.
Davis, W. M., “The geographic cycle,” Geographical Journal, 14: 481–504, 1899.
Morisawa, M. E., “Measurement of drainage-basin outline form,” Journal of
Emmett, W., “The channels and waters of the upper Salmon River area,
Geology, 66: 587–591, 1958.
Idaho,” U.S. Geological Survey Professional Paper 870-A, 1975.
Miller, V. C., “A quantitative geomorphologic study of drainage basin
Fitzpatrick, F. A., “A geologic approach to field methods in fluvial
characteristics in the Clinch Mountain area, Virginia and Tennessee,” Project
geomorphology,” edited by M. J. Thornbush, C. D. Allen, and F. A. Fitzpatrick,
NR 389-042. Technical Report 3, Columbia University, Department of
Geomorphological Fieldwork: Developments in Earth Surface Processes, Vol. 18,
Geology, New York, 1953.
Chap. 5.2, 2014, pp. 201–230.
Owens, P. N., R. J. Batalla, A. J. Collins, B. Gomez, D. M. Hicks, A. J.
Fitzpatrick, F. A. and M. P. Peppler, “Relation of urbanization to stream
Horowitz, G. M. Kondolf, et al., “Fine-grained sediment in river systems:
geomorphologic and habitat characteristics in nine metropolitan areas of the
environmental significance and manage issues,” River Research Application,
United States,” U.S. Geological Survey Scientific Investigations Report
21: 693–717, 2005.
2010–5056, 2010, p. 29.
Radoane, M., N. Radoane, and D. Dumitriu, “Geomorphological evolution
Fitzpatrick, F. A., J. C. Knox, and H. E. Whitman, “Effects of historical land-
of longitudinal profiles in the Carpathians,” Geomorphology, 50: 293–306,
cover changes on flooding and sedimentation, North Fish Creek, Wisconsin,”
2003.
U.S. Geological Survey Water-Resources Investigations Report 99–4083, 1999, p. 12.
Reddy, J. R., A Textbook of Hydrology, Laxmi Publications, New Delhi,
Fitzpatrick, F. A., M. C. Peppler, M. M. DePhilip, and K. E. Lee, “Geomor-
India, 2005, p. 93.
phologic characteristics and classification of Duluth-area streams, Minnesota,”
Renschler, C. S., M. W. Doyle, and M. Thoms, “Geomorphology and
U.S. Geological Survey Scientific Investigations Report 2006–5029, 2006,
ecosystems: challenges and keys for success in bridging disciplines,”
accessed May 15, 2014, http://pubs.usgs.gov/sir/2006/5029/.
Geomorphology, 89: 1–8, 2007.
Fitzpatrick, F. A., M. C. Peppler, D. A. Saad, D. M. Pratt, and B. N. Lenz,
Ries, K. G., III, “The national streamflow statistics program: a computer
“Geomorphologic, flood, and groundwater-flow characteristics of Bayfield
program for estimating streamflow statistics for ungaged sites,” U.S. Geologi-
Peninsula streams, Wisconsin, and implications for brook-trout habitat,” U.S.
cal Techniques and Methods Report TM Book 5, Chap. A6, 2006, p. 45.
Geological Survey Scientific Investigations Report 2014–5007, 2015.
Ritter, D. F., R. C. Kochel, and J. R. Miller, Process Geomorphology, 4th ed.,
Fitzpatrick, F. A., I. R. Waite, P. J. D’Arconte, M. R. Meador, M. A. Maupin,
McGraw Hill, Boston, MA, 2002.
and M. E. Gurtz, “Revised methods for characterizing stream habitat in the
Robertson, D. M., G. E. Schwarz, D. A. Saad, and R. B. Alexander,
National Water-Quality Assessment Program,” U.S. Geological Survey
“Incorporating uncertainty into the ranking of Sparrow model nutrient yields
Water-Resources Investigations Report 98–4052, 1998, p. 67.
from Mississippi/Atchafalaya River Basin watersheds,” Journal of the American
Fryirs, K. A. and G. J. Brierley, “Geomorphologic Analysis of River Systems,”
Water Resources Association, 45 (2): 534–549, 2009.
An Approach to Reading the Landscape, John Wiley & Sons, Chichester, 2013.
Rosgen, D. L., “Watershed assessment of river stability and sediment supply
Gardiner, V., “Drainage basin morphometry,” Technical Bulletin 14, British
(WARSSS),” Wildland Hydrology, Fort Collins, CO, 2006.
Geomorphological Research Group, Norwich, 1975, p. 48.
Schumm, S. A., “Evolution of drainage systems and slopes in badlands at
Gilbert, G. K., “Report on the geology of the Henry Mountains,” U.S.
Perth Amboy, New Jersey,” Geological Society of America Bulletin,
Geographical and Geological Survey of the Rocky Mountain Region, Washington,
67: 597–646, 1956.
DC, 1877.
Schumm, S. A., “Geomorphologic thresholds and complex response of
Glock, W. S., “The development of drainage systems: a synoptic view,”
drainage systems,” edited by M. Morisawa, Fluvial Geomorphology, S.U.N.Y.
Geographical Review, 21: 475–482, 1931.
Binghamton Pubs., Binghamton, NY, 1973, pp. 299–310.
Gordon, N. D., T. A. McMahon, and B. L. Finlayson, Stream Hydrology, an
Schumm, S. A., The Fluvial System, John Wiley and Sons, Chichester, 1977.
Introduction For Ecologists, John Wiley & Sons, Chichester, 1992.
Schumm, S. A. and R. W. Lichty, “Time, space, and causality in geomor-
Hack, J., “Studies of longitudinal stream profiles in Virginia and Maryland,”
phology,” American Journal of Science, 263: 110–119, 1965.
U.S. Geological Survey Professional Paper, 294–B: 59 P, 1957.
Seaber, P. R., F. P. Kapinos, and G. L. Knapp, “Hydrologic unit maps,” U.S.
Happ, S. C., G. Rittenhouse, and G. C. Dobson, “Some principles of
Geological Survey Water Supply Paper 2294, 1987, p. 63.
accelerated stream and valley sedimentation,” Technical Bulletin 695, U.S.
Sear, D. A., M. D. Newson, and C. R. Thorne, “Guidebook of applied fluvial
Department of Agriculture, Washington, DC, 1940.
geomorphology: Defra/Environment Agency,” Flood and Coastal Defense
Harvey, C. A. and D. A. Eash, “Description, instructions, and verification
R&D Programme, R&D Technical Report FD1914, London, UK, 2003.
for Basinsoft, a computer program to quantify drainage-basin characteristics,”
Shreve, R. L., “Infinite topologically random channel networks,” Journal of
U.S. Geological Survey Water-Resources Investigations Report 95–4287, 1996,
Geology, 75: 178–186, 1967.
http://pubs.usgs.gov/wri/1995/4287/report.pdf.
44_Singh_ch44_p44.1-44.12.indd 11 8/22/16 2:13 PM

Simon, A. and C. R. Hupp, “Channel evolution in modified Tennessee Strahler, A. N., “Dimensional analysis applied to fluvially eroded
channels,” Proceedings of the 4th Federation of Interagency Sediment Conference, landforms,” Geological Society of America Bulletin, 69: 279–300, 1958.
2 (5): 71–82, 1986. Retrieved from https://www.sciencebase.gov/catalog/ Strahler, A. N., “Quantitative geomorphology of drainage basins and
item/50579e95e4b01ad7e02863d8. channel networks, Section 4-II,” edited by V. T. Chow, Handbook of Applied
Smith, R. A., G. E. Schwarz, and R. B. Alexander, “Regional interpretation of Hydrology, McGraw-Hill Book Company, New York, NY, 1964, pp. 4-39–
water-quality monitoring data,” Water Resources Research, 33: 2781–2798, 1997. 4-76.
Strahler, A. N., “Dynamic basis of geomorphology,” Bulletin of the Thorne, C. R., R. D. Hey, and M. D. Newson, Applied Fluvial Geomorphology
Geological Society of America, 63 (9): 923–938, 1952a. for River Engineering and Management, John Wiley & Sons, 1998.
Strahler, A. N., “Hypsometric (area-altitude) analysis of erosional topography,” Turnipseed, D. P. and K. G. Ries, III, “The national streamflow statistics
Geological Society of America Bulletin, 63 (11): 1117–1142, 1952b. program: estimating high and low streamflow statistics for ungaged sites,” U.S.
Strahler, A. N., “Quantitative analyses of watershed geomorphology,” Geological Survey Fact Sheet 2007–3010, 2007, p. 4.
Transactions American Geophysical Union Journal, 38 (6): 913–920, 1957.
44_Singh_ch44_p44.1-44.12.indd 12 8/22/16 2:13 PM

Chapter 45
Infiltration Modeling
BY
C. CORRADINI, R. MORBIDELLI, AND R. S. GOVINDARAJU
ABSTRACT saturated hydraulic conductivity (Russo and Bresler, 1981; 1982) that may be
assumed as a random field with a lognormal univariate probability distribu-
Many mathematical models of vertical infiltration of rainfall into a soil matrix
tion (Smith and Goodrich, 2000). In addition to the heterogeneity of
with horizontal surface are described. Firstly, classical approaches developed
saturated hydraulic conductivity, rainfall is also characterized by spatial vari-
for regular rainfalls which keep continuous saturation at the soil surface and
ability (Goodrich et al., 1995; Krajewski et al., 2003) and their interaction
a more general formulation suitable for any type of rainfall pattern are con-
makes in principle the estimate of infiltration at the field scale very complex
sidered for applications at the local (point) scale in homogeneous soils. Then,
(Morbidelli et al., 2006).
a simple model for point infiltration into a two-layered soil with a more per-
A model of the expected areal-average infiltration into a much more perme-
meable upper layer and a more complex model for any two-layered soil type
able upper layer of a two-layered soil was also proposed by Corradini et al.
are presented. Two field-scale infiltration models are also analyzed. One is of
(2011b) considering only a spatial horizontal random field of the saturated
semi-empirical type and represents the spatial variability of saturated hydrau-
hydraulic conductivity. It involves the solution of a set of algebraic equations
lic conductivity, Ks, the other relies upon a semi-analytical/conceptual frame-
obtained by upscaling simple local infiltration equations to the field scale. The
work and combines the spatial heterogeneity of Ks with that of rainfall rate.
same quantity, but for a vertically nonuniform soil characterized by a saturated
Open problems linked with rainfall infiltration into surfaces with significant
hydraulic conductivity decreasing with depth according to a power law, was
slopes are also discussed.
derived by Govindaraju et al. (2012) and upscaled to the field-scale by the same
semianalytical technique used in previous papers (Govindaraju et al., 2001; 2006).
45.1 INTRODUCTION Most of the aforementioned models consider a surface placed horizontally
or with a low slope that does not affect the infiltration process. However, in
Rainfall infiltration plays an important role in the reduction of surface runoff
most real situations, infiltration occurs in surfaces characterized by different
and provides subsurface water that governs water supply for agriculture,
gradients (Beven, 2002; Fiori et al., 2007) and the role of surface slope on
transport of pollutants through the vadose zone, and recharge of aquifers. The
infiltration is not clear. In fact, the results obtained by some theoretical and
spatiotemporal evolution of infiltration rate under natural conditions cannot
experimental investigations (Sharma et al., 1983; Poesen, 1984; Philip, 1991;
be currently deduced by direct measurements at any scale of interest in
Chen and Young, 2006; Essig et al., 2009; Morbidelli et al., 2015) lead to
applied hydrology, therefore, the use of infiltration modeling allowing to
rather contrasting conclusions, suggesting an improved understanding and
express it through measurable quantities is of fundamental importance.
modeling of infiltration on sloping surfaces are required.
Even though the representation of the main natural processes in applied
When macropore flow plays a significant role in determining infiltration,
hydrology requires areal infiltration modeling for both flat and sloping sur-
then other simplified approaches may be used (e.g., Arabi et al., 2006;
faces, research activity has been limited for many years to the development of
Stillman et al., 2006).
local, or point, infiltration models for vertically homogeneous soils (Green
Finally, a method widely used for the representation of infiltration in
and Ampt, 1911; Philip, 1957a; 1957b; 1957c; Smith and Parlange, 1978;
ungaged basins, together with interception and depression storage, was devel-
Corradini et al., 1997).
oped by the United States Department of Agriculture (USDA)-soil-conversion
Natural soils are rarely vertically homogeneous. In hydrological simula-
service (SCS) to estimate rainfall excess (Soil Conservation Service, 1972).
tions, local infiltration for the estimate of effective rainfall can be reasonably
schematized by a two-layered vertical profile (Mualem et al., 1993; Taha et al.,
1997). Soils representable by a sealing layer over the parent soil or a vertical 45.2 BASIC EQUATIONS FOR VERTICAL INFILTRATION
profile with a more permeable upper layer are frequently found. Some models Under conditions of horizontally homogeneous soils, the water movement in
for infiltration into crusted soils were developed under significant approxima- the vertical direction is governed by one-dimensional soil water flow and
tions by adapting the Green-Ampt model. A general semianalytical/concep- continuity equations. The flow rate, q, per unit cross-sectional area is
tual model for crusted soils was later formulated by Smith et al. (1999), and described by Darcy’s law as:
was extended by Corradini et al. (2000) to represent infiltration and reinfiltra-
tion after a redistribution period under any rainfall pattern and for any two-  ∂ψ 
layered soil, where either layer may be more or less permeable than the other. q = −K  − 1 (45.1)
 ∂z 
For a much more permeable upper layer, and under more restrictive rainfall
patterns, a simpler semiempirical/conceptual model was presented by where K is the hydraulic conductivity, ψ the soil water matric capillary head,
Corradini et al. (2011a). Under conditions of surface saturation, a simple and z the vertical soil depth assumed positive downward. The infiltration rate,
Green-Ampt–based model was also proposed (Chow et al., 1988). q0, is given by Eq. (45.1) applied at the soil surface.
In applied hydrology, upscaling of point infiltration modeling to the field The mass conservation equation, in the absence of changes in the water
scale is required to estimate the areal-average infiltration. This is a complex density and porosity, ϕ, as well as of sinks and sources, is expressed by:
task because of the natural spatial heterogeneity of hydraulic soil properties
(Nielsen et al., 1973; Warrick and Nielsen, 1980; Greminger et al., 1985; ∂θ ∂q
=− (45.2)
Sharma et al., 1987; Loague and Gander, 1990) and particularly of the soil ∂t ∂z
45-1
45_Singh_ch45_p45.1-45.10.indd 1 8/22/16 2:14 PM

45-2 Infiltration Modeling
where θ is the volumetric water content and t the time. Substitution of Equation (45.8) applies starting from the assumption of immediate pond-
Eq. (45.1) in Eq. (45.2) leads to the well-known Richards equation: ing. Under more general conditions, with a constant rainfall rate r > Ks, that
begins at the time t = 0, surface saturation is reached at a time tp > 0. For t ≤ tp
∂ψ ∂  ∂ψ  ∂ψ
C1 = K − C2 (45.3) the infiltration rate q0 is equal to r and later to the infiltration capacity. Mein
∂t ∂z  ∂z  ∂z and Larson (1973) formulated this process through Eq. (45.7) as:
where C1 = dθ/dψ and C2 = dK/dψ under the assumption that θ and K are K sψav (θ s − θi )
unique functions of ψ with hysteresis in these functions being neglected (see r − Ks = − tp t = tp(45.9)
also Dagan and Bresler, 1983; Smith, 2002). The initial condition at time t = 0
for z ≥ 0 is ψ = ψi, the upper boundary conditions at the soil surface are: ∫ r dt
0
z = 0 0 < t ≤ tp q0 = r(45.4a) which leads to determine tp as:

z = 0 tp < t ≤ tr θ0 = θs saturation stage (45.4b)
K sψav (θ s − θi )
z = 0 tr < t q0 = 0 redistribution stage (45.4c) tp = − (45.10)
r (r − K s )
where r is the rainfall rate, tp the time to ponding, tr the duration of rainfall;
hereafter, the subscripts i and s denote initial and saturation quantities, Furthermore Eq. (45.8) becomes
respectively, and subscript 0 stands for quantities at the soil surface. The lower
F − ψav (θ s − θi )
boundary conditions at a depth zb, which is not reached by the wetting front F = Fp − ψav (θ s − θi ) ln + K s (t − t p ) t > tp(45.11)
is ψ(zb) = ψi for t > 0. The soil water hydraulic properties can be represented Fp − ψav (θ s − θi )
by the following parameterized forms (Smith et al., 1993):
Equation (45.11) can be solved at each time, for example, by successive
 − c /λ 1/c
θ −θ  substitutions of F which is then substituted in Eq. (45.7) to obtain the corre-
ψ = ψb  * r  − 1  + d (45.5a) sponding value of fc. Alternatively, through a series expansion of the log-term
 θ s − θ r  
and simplifying, the following explicit approximation to F is obtained
  c −(bλ+a)/c (Govindaraju et al., 2001):
ψ −d  
K = K s∗ 1 +   (45.5b) 12
  ψb   12 2 1  2 K s3  3 2 3 2
F = Fp +2 K sψav (θ s − θi ) (t 1 2 − t 1p 2 )+ K s (t − t p )+   (t − t p )
3 18 ψ (θ s − θi ) 
where θ s∗ and K s∗ are used as scaling quantities, ψb is the air entry head, θr is
the residual volumetric water content, c, λ, and d are empirical coefficients, t > tp(45.12)
b = 3 and a = 2 according to Burdine’s method (Brooks and Corey, 1964). For
particular values of parameters, Eqs. (45.5a and 45.5b) reduce to the well- Indicative values of |ψav | and Ks, together with porosity, φ, are shown in
known equations proposed by Brooks and Corey (1964) and van Genucthen Table 45.1 (see also Rawls and Brakensiek, 1983).
(1980). For two-layered soils two additional conditions are required at the
interface between the two layers: 45.3.2 Extended Philip Model
Philip (1957a; 1957b; 1957c; 1969) proposed an infiltration model obtained
ψ1 ( Zc ) = ψ2 ( Zc ) = ψc (45.6a) through an analytical series solution of the Richards equation under the con-
 ∂ψ    ∂ψ   ditions of vertically homogeneous soil, constant initial moisture content, and
K1  1  − 1 = K 2  2  − 1 (45.6b) saturated soil surface with immediate ponding. For early to intermediate
 ∂z Zc   ∂z Zc  times, the series solution can be truncated after the first two terms and the
infiltration capacity can be expressed as:
where hereafter the subscripts 1, 2, and c denote variables in the upper layer,
1
lower layer, and at the interface, respectively, and Zc is the interface depth. fc = St −1 2 + A (45.13)
2
45.3 CLASSICAL MODELS FOR POINT INFILTRATION INTO where S is the sorptivity, depending on soil properties and initial moisture
VERTICALLY HOMOGENEOUS SOILS content, and A is a quantity ranging from 0.38 Ks to 0.66 Ks. For t→∞
Eq. (45.13) is replaced by fc = Ks. Integration of Eq. (45.13) yields the cumula-
A variety of local infiltration models for vertically homogeneous soils with tive infiltration:
constant initial soil water content and over horizontal surfaces have been
proposed (Green and Ampt, 1911; Kostiakov, 1932; Horton, 1940; Holtan,
F = St 1 2 + At (45.14)
1961; Swartzendruber, 1987; Philip, 1957a; 1957b; 1957c; Smith and Parlange,
1978; Broadbridge and White, 1988; Dagan and Bresler, 1983; Smith et al.,
Philip’s model was extended for applications to less–restrictive c onditions.
1993; Corradini et al., 1994; Corradini et al., 1997, Kacimov and Obnosov,
For constant rainfall rate r > Ks, surface saturation occurs at a time tp > 0 and
2013). Furthermore, for isolated regular storms and when ponding is not
achieved instantly, extended forms of the Philip model (Chow et al., 1988), Table 45.1 Green–Ampt Infiltration Parameters for Various USDA
Green–Ampt model (Mein and Larson, 1973; Chu, 1978) and the Smith and Soil Classes
Parlange model (Parlange et al., 1982) have been widely used, while for
erratic rainfalls the Corradini et al. (1997) model can be properly selected. Wetting front soil Saturated hydraulic
suction head, ψav conductivity, Ks
These four models have been extensively used in applied hydrology and as Soil texture class Porosity, ϕ (mm) (mmh-1)
building blocks in the development of infiltration approaches at the field scale.
Sand 0.437 49.5 235.6
45.3.1 Extended Green–Ampt Model
Loamy sand 0.437 61.3 59.8
The original Green–Ampt model represents infiltration into homogeneous Sandy loam 0.453 110.1 21.8
soils under the conditions of continuously saturated soil surface and uniform
initial soil moisture as: Loam 0.463 88.9 13.2
 ψ (θ − θ )  Silt loam 0.501 166.8 6.8
fc = K s 1 − av s i  (45.7)
 F  Sandy clay loam 0.398 218.5 3.0
where ψav is the soil water matric capillary head at the wetting front, F the Clay loam 0.464 208.8 2.0
cumulative depth of infiltrated water, and fc the soil infiltration capacity. To Silty clay loam 0.471 273.0 2.0
express the infiltration as a function of time, this equation can be solved after Sandy clay 0.430 239.0 1.2
the substitution fc = dF/dt. The resulting equation (Chow et al., 1988) is:
Silty clay 0.479 292.2 1.0
 F  Clay 0.475 316.3 0.6
F = K st − ψav (θ s − θi ) ln1 −  (45.8)
 ψav (θ s − θi )  [Source: Rawls and Brakensiek, 1983]
45_Singh_ch45_p45.1-45.10.indd 2 8/22/16 2:14 PM

Classical Models For Point Infiltration Into Vertically Homogeneous Soils 45-3
following Chow et al. (1988) infiltration can be described through an equiva- A more general model was formulated by Corradini et al. (1997) starting from
lent time origin, t0, for potential infiltration after p
onding as: the same integrated equations then combined with a conceptual representa-
tion of the wetting soil moisture profile.
 A
S 2 r −  To highlight the structure of the last model, a specific rainfall pattern which
 2 allows to describe all the involved components is used here. The application
tp = 2 (45.15)
2r (r − A) to erratic rainfall is then a straightforward problem.
1  2 2 Let us consider a stepwise rainfall pattern involving successive periods of
( ) − S
12
t0 = t p − S + 4 AFp (45.16) rainfall with constant r > Ks, separated by periods with r = 0 (see Fig. 45.1).
4 A2  
We denote by t1 the duration of the first pulse, t2 and t3, the beginning and end
1 −1 2 of the second pulse, respectively, and t4 the beginning of the third pulse. The
fc = S (t − t0 ) + A t > tp(45.17) model was derived considering a soil with a constant value of θi and combin-
2
ing the depth-integrated forms of the Darcy law and continuity equation. In
For unsteady rainfall under the condition of a continuously saturated sur- addition, as the event progresses in time, a dynamic wetting profile, of lowest
face for t > tp, the infiltration process can be represented adopting a similar depth Z and represented by a distorted rectangle through a shape factor
procedure. Generally, S and A are derived from calibration of hydrological β(θ0) ≤ 1, was assumed. The resulting ordinary differential equation is
models; however, S can also be approximated through the parameters of the
Green–Ampt equation (see Youngs, 1964) as: dθ 0
=
(θ0 − θi ) β (θ0 )  (θ0 − θi ) G (θi ,θ0 ) β (θ0 ) p K 0 
 q0−K 0 − 
12
dt
F ′ (θ 0 − θi )
dβ (θ 0 ) 
+ β (θ 0 )
 F′ 
S = 2 (φ − θi ) K s ψav  (45.18)  dθ 0 
45.3.3 A Three-Parameter Semianalytical Model (45.22)

Parlange et al. (1982) presented a three-parameter model obtained through an where p is a parameter linked with the profile shape of θ, and G(θi,θ0) is
analytical integration of the Richards equation and expressed as: expressed by Eq. (45.20) modified by the substitution of θs with θ0 and Ks with
  K0. Equation (45.22) can be applied for 0 < t < t2, the profile shape of θ(z) is
  approximated (Corradini et al., 1994) by
 α 
fc = K s 1 +  θ ( z ) − θi  β z (θ − θ ) − F ′ 
 αF′  (45.19) = 1 − exp 0 i
 (45.23)
 exp − 1  θ 0 − θi 2
 (β − β ) F 
′
 
  G (θ s − θi )  
Functional forms for β and p were obtained by calibration using results
where F' = F – Kit is the cumulative dynamic infiltration rate, α a parameter provided by the Richards equation applied to a silty loam soil, specifically:
linked with the behavior of hydraulic conductivity and soil water diffusivity,
D, as functions of θ, and G is the integral capillary drive defined by θ s − θi
β (θ 0 ) = 0.6 + 0.4 (45.24a)
1
θs θ s −θr
G=
Ks
∫ D (θ ) dθ (45.20) β p = 0.98 − 0.87exp (−r K s ) dθ0/dt ≥ 0
θi (45.24b)
Equation (45.19) includes as limiting forms (Smith and Goodrich, 2000)
the Smith and Parlange (1978) equation (α = 1) and the Green–Ampt equa- β p =1.7 dθ0/dt < 0 (45.24c)
tion (α→0). It can be applied to determine tp and fc for any rainfall pattern
and for t > tp can be rewritten under the condition of surface saturation in a Equation (45.22) can be solved numerically. For q0 = r, with F' = (r - Ki)t, it
time dependent form as: gives θ0(t) until time to ponding, t ′p, corresponding to θ0 = θs and dθ0/dt = 0,
  ′   then for t ′p < t ≤ t1, with θ0 = θs and dθ0/dt = 0, it provides the infiltration
 exp α F − 1 +
αKs 
′ (θ s − θi ) K sG
  (θ s − θi ) G  Kd 
[(1 − α ) K s − Ki ](t − t p )= F − Fp −
′
ln 
 α Fp′  
Kd  αKs 
 exp (θ − θ ) G  − 1 + K 
  s i  d 
(45.21)
Rainfall rate
where Kd = Ks – Ki and Fp′ = F'(tp). The quantities Fp′ and tp are the values of F'
and t, respectively, at which Eq. (45.19) with fc = r(tp) is first satisfied. The
value of α usually ranges from 0.8 to 0.85 (Smith, 2002).
45.3.4 A Semianalytical/Conceptual Model
For complex rainfall patterns involving rainfall hiatus periods or a rainfall rate
after time to ponding less than soil-infiltration capacity, the models presented
in the previous subsections cannot be applied in principle because the (a) tp’ t1 t2 tp” t3 t4 Time
assumption of uniform initial soil moisture cannot be met for successive
storms. Alternatively, an approach for the application of the aforementioned Soil water content Soil water content
classical models was developed (Mls, 1980; Péschke and Kutílek, 1982; and θi θs θi θs
Verma, 1982) starting from the time compression approximation proposed by
Reeves and Miller (1975) for post-hiatus rainfall producing immediate pond- tp”
ing (see also Brutsaert, 2005). However, Smith et al. (1993) by comparison
with results of the Richards equation showed that the last approach was not tp’
sufficiently accurate because it neglects the soil water redistribution process,
Depth
Depth
which is particularly important when long periods with a light rainfall or a t1 t1

rainfall hiatus occur. Models combining infiltration and redistribution are,
therefore, the best solution when complex rainfall patterns, which are rather
common under natural conditions, have to be considered (see also Basha, t2
2011a; 2011b). Dagan and Bresler (1983) developed an analytical model along t3
this line starting from depth integrated forms of the Darcy and continuity (b) (c)
equations and using simplifications in the initial and surface boundary condi- Figure 45.1 (a) Rainfall pattern selected to describe the semianalytical/conceptual
tions that make easier areal investigations, but reduce practical applications at model for point infiltration. [(b) and (c)] Profiles of soil water content at various times
the local scale. In any case, the model is not addressed toward local studies indicated in (a) and associated with different infiltration-redistribution stages. For
with erratic rainfalls producing successive infiltration-redistribution cycles. symbol, see text.
45_Singh_ch45_p45.1-45.10.indd 3 8/22/16 2:15 PM

capacity (q0 = fc) and for t1 < t < t2, with q0 = 0, it gives dθ0/dt < 0 thus conditions, the model by Corradini et al. (2000) appears to be accurate and in
describing the redistribution process. any case requires a fairly limited computational effort.
The second rainfall pulse leads to a new time to ponding, t ′′p, but reinfiltra-
tion occurs according to two alternative approaches determined by a com- 45.4.1 Green–Ampt-Based Model for a More
Permeable Upper Layer
parison of r and the downward redistribution rate, DF(t = t2), expressed by:
Chow et al. (1988) presented a model for infiltration into a two-layered soil
 1 dβ 1  ′ dθ 0 with a more permeable upper layer under the condition of continuously satu-
DF =  − − F − Ki t = t2(45.25)
 β dθ 0 θ 0 − θi  dt rated soil surface. The classical Green–Ampt equations are applied until the
wetting front is in the upper layer, then the following equations are used:
Specifically, for r ≤ DF the reinfiltrated water is distributed to the whole
(θ 2s − θ 2i ) + 1 
dynamic profile and θ0(t) can be still computed by Eq. (45.22), while for L2 (θ 2 s − θ 2i ) Zc K 2 s − (θ 2 s − θ 2i ) K1s (ψav 2 + Zc )
r > DF the profile of θ(z, t = t2) is assumed temporarily invariant and starts a K 2s K1 s K 2 s

superimposed secondary wetting profile which advances alongside the preex-  L2 
isting profile according to Eq. (45.22) modified substituting θi with θ0(t2) and ln1 + =t (45.26)
Fʹ with Fʹ2t accumulated for t ≥ t2. If the secondary profile reaches the depth  ψav 2 + Zc 
of the steady one, the compound profile reduces to a single profile and then F = Zc (θ1s − θ1i ) + L2 (θ 2 s − θ 2i ) (45.27)
Eq. (45.22) can be again applied [see Fig. 45.1 for a schematic representation
of θ(z,t)]. On the other hand, if at t = t3 the secondary profiles has not caught K1 s K 2 s
up with the first one, in the successive rainfall hiatus redistribution is first
fc = (ψav 2 + Zc + L2 ) (45.28)
Zc K 2 s + L2 K1s
applied to the secondary profile and then reestablished to the single profile.
Finally, in the case at t = t4 the θ(z) profile is still compound and r is larger where L2 is the depth of the wetting front below the interface. Equation (45.26)
than DF(t4), a procedure of consolidation that merges the composite profile is can be solved at each time for L2 by successive substitutions; L2 is then used
applied early to avoid the formation of a further additional profile. in Eqs. (45.27) and (45.28) to determine F and fc, respectively.
The model incorporates all the components required for the application to 45.4.2 A Semianalytical/Conceptual Model for Any
any natural rainfall pattern. It was calibrated by Corradini et al. (1997) for a Two-Layered Soil
silty loam soil (denoted as Soil C), and then tested using different soils
(denoted as A, B, J, D, E, and SL) from clay loam to sandy loam soil types. Corradini et al. (2000) formulated a semianalytical/conceptual model applicable
Table 45.2 shows the parameters required to define the hydraulic soil proper- to any horizontal two-layered soil, where either layer may be more permeable. It
ties of these soils. Weighted implicit finite difference solutions of the Richards relies on the same elements previously used by Corradini et al. (1997), but
equation were used as a benchmark. For each soil, the model accuracy was adopted here in each layer and integrated at the interface between the two layers
found to be acceptable in terms of both infiltration rate and soil water content, by the boundary conditions expressing continuity of flow rate, and capillary
even though better results were obtained for fine-textured soils. In addition head [Eqs. (45.6a and 45.6b)]. In addition at the lower boundary for t > 0, we
the model was found to simulate fairly well the θ(z,t) profiles observed in have ψ2 = ψi. The initial condition is ψ1 = ψ2 = ψi constant at t = 0, and at the
laboratory and field experiments (Melone et al., 2006; 2008). interface q(Zc) is approximated through the downward flux in the upper layer as:
G1 (ψc ,ψ10 )
q ( Z c ) = K1 s + K1c (45.29)
45.4 MODELING OF POINT INFILTRATION INTO Zc
VERTICALLY NONUNIFORM SOILS
where G(ψc,ψ10) is expressed by Eq. (45.20) modified by the substitutions of
A two-layer approximation, with each layer being schematized as homoge- D(θ)dθ with K(ψ)dψ, θs with ψ10 and θi with ψc. Until water does not infiltrate
neous, is frequently used to set up models of infiltration for natural soils. The in the lower layer, the Corradini et al. (1997) model is used, then starting from
process of formation of a sealing layer was accurately examined by Mualem the time tc when the wetting front enters the lower layer, the following system
and Assouline (1989) and Mualem et al. (1993), and that of disruption was of two ordinary differential equations may be applied:
considered by Bullock et al. (1988), Emmerich (2003) and Morbidelli et al.
(2011). Evidence of the role of crusted soils in semiarid regions has been dψ10 1  K1s G1 (ψc ,ψ10 ) 
recently provided by Chen et al. (2013). Green–Ampt-based models for infil- = q10 − K1c − 
tration into stable crusted soils were proposed by Hillel and Gardner (1970), dt γ Zc C1 (ψ10 )  Zc 
Ahuja (1983), Vandervaere et al. (1998), and Moore (1981). An efficient (1 − γ )C1 (ψc ) dψc for t ≥ tc(45.30)
approach which represents transient infiltration into crusted soils was pro- −
γ C1 (ψ10 ) dt
posed by Smith (1990), while Philip (1998) formulated a model, which
describes upper layer dynamics, but in the limits of a ponded upper bound- dψ c 1
=
ary. On the other hand, vertical profiles with a more permeable upper layer dt PL (ψ c ,t )
are observed in the hydrological practice and can be also used, for example,
as a first approximation in the representation of infiltration into homoge-  K G (ψ ,ψ ) β (θ ) p (θ − θ 2i ) K 2 s G2 (ψ i ,ψ c ) 
×  K1c + 1s 1 c 10 − K 2c − 2 2c 2 2c 
neous soils with grassy vegetation (Morbidelli et al., 2014). In the last layering Z F2
 c 
type, the simple model presented by Chow et al. (1988) can be usefully
applied for infiltration into a saturated soil surface. Under more general for t ≥ tc(45.31)
Table 45.2 Hydraulic Properties and Parameters for a Variety of Representative Soils
Characteristic property Soil A (silty) Soil B (clay loam) Soil C (silty loam) Soil J (sandy) Soil D (silty loam) Soil E (clay loam) Soil SL (sandy loam)
Ks (mmh )
* -1 0.4000 0.4000 1.6000 0.4000 0.4000 1.6000 25.0000
Ks (mmh-1) 0.3979 0.4000 1.6000 0.2785 0.3998 1.6000 25.0000
Φ 0.3500 0.3500 0.3500 0.3500 0.3500 0.3500 0.4530
θs* 0.3325 0.3325 0.3325 0.3325 0.3325 0.3325 0.4120
θs 0.3324 0.3325 0.3325 0.3219 0.3325 0.3325 0.4120
θr 0.1225 0.1225 0.1225 0.1225 0.1225 0.1225 0.0410
Ψb (mm) –25 –800 –400 –200 –400 –800 –300
λ 0.2 0.2 0.4 0.5 0.4 0.2 0.5
c 5 5 5 5 5 5 5
d (mm) 10 100 0 185 100 0 0
G(θr,θs) (mm) 27.8 1103.9 526.6 110.7 426.9 1203.8 376.1
[Source: Corradini et al., 1997]
45_Singh_ch45_p45.1-45.10.indd 4 8/22/16 2:15 PM

Models for Rainfall Infiltration Over Heterogeneous Areas 45-5
with PL (ψc , t ) and F2 defined as: Two models characterized by significant differences in complexity and appli-
cation area are presented here.
 dβ  F2
PL (ψc , t ) = β 2 (θ 2c ) + 2 (θ 2c − θ 2i ) C2 (ψc ) (45.32) 45.5.1 A Semiempirical Approach
 dθ 2c  (θ 2c − θ 2i ) β 2 (θ 2c )
Smith and Goodrich (2000) proposed a semiempirical model to determine the
F2 = F − Zc γ (θ1s − θ1i ) + (1 − γ ) (θ1c − θ1i ) − K 2it (45.33) areal-average infiltration rate into areas with random spatial variability of Ks.
They assumed a lognormal probability density function, PDF, of Ks with a mean
and C1(ψ10) = dθ10/dψ10, C1(ψc) = dθ1c/dψ1c. The quantity γ represents a con- value <Ks> and a coefficient of variation CV(Ks), and considered one realization
ceptual proportion of the upper layer, where θ is increasing due to rainfall and of the random variable. Then adopting the Parlange et al. (1982) model (see also
is assumed equal to 0.85 (Smith et al., 1999), β2 and p2 are determined by Smith et al., 1993) and the Latin Hypercube sampling method, through a large
Eqs. (45.24a–45.24c), but applied using θ20, θ2s, θ2i, θ2r , and substituting r with number of simulations performed for many values of CV(Ks) and rainfall rates,
they developed the following effective areal relation for the scaled areal-average

q(Zc). On the basis of the same stepwise rainfall pattern earlier adopted to
explain the Corradini et al. (1997) model, Eqs. (45.30) and (45.31) may be infiltration rate, I e∗, linked with the corresponding scaled cumulative depth, Fe∗:
used for tc < t < t2. Then, the two-layer model has to be applied in each layer
1 − cc
by analogy with the procedure described for homogeneous soils, in particular,   ∗ cc 
compound and consolidated profiles develop in each layer. In the underlying
soil, the generation of additional profiles occurs through q(Zc) in substitution
I e∗ = 1+ ( re∗ − 1 ) 1 + 
  α
(
re − 1 α Fe∗
e −1  
 
) re∗ > 1 (45.34)
of r. On the basis of the described steps, model application to erratic rainfall

patterns is straightforward. The solution of the above system, Eqs. (45.30) and with
(45.31), may be obtained by a library routine for the Runge–Kutta–Verner
ca ≅ 1 +
0.8  (−0.85(rb∗ −1)) 
fifth-order method with a variable time step. Calibration and testing of the 1 − e  (45.35a)
model were performed through a comparison with numerical solutions of the CV ( K s )1.3  
Richards equation. Three soils (clay loam, silty loam, and sandy loam) with a
variety of thicknesses were combined to realize two-layered soils, where either
where I e∗ = I e K e , Fe∗ = Fe G (θ s − θi ) , re∗ = r K e and rb∗ = r < K s > . The quan-
layer was more permeable, that were selected as test cases. In all instances, the
simulations involved the cycle of infiltration-redistribution-reinfiltration. The tity Ke denotes the areal effective value of Ks given by
infiltration rate as well as the water content at the surface and interface were r  r 
found to be very accurately estimated by the semianalytical/conceptual model. Ke = ∫ K f K ( K )d K + 1 − ∫ f K ( K ) d K r (45.35b)
s s
0  0 
45.5 MODELS FOR RAINFALL INFILTRATION OVER
where f K s ( K ) is the PDF of Ks. Finally, Eq. (45.34) may be also applied for r
HETEROGENEOUS AREAS
variable with time.
The mathematical problem of infiltration at the field scale is not analytically
tractable, while using accurate Monte-Carlo (M-C) simulation techniques
45.5.2 A Semianalytical/Conceptual Model
imposes an enormous computational burden for routine applications.
Considering the spatial heterogeneity of saturated hydraulic conductivity, M-C Govindaraju et al. (2006) formulated a semi-analytical model to estimate the
simulations were used, for instance, by Sharma and Seely (1979), Maller and expected field-scale infiltration rate <In(F)> under the condition of negligible
Sharma (1981), and Saghafian et al. (1995) in specific studies addressed to effects of the run-on process. The model incorporates heterogeneity of both Ks
describe field-scale infiltration. M-C simulations were also used by Sivapalan and r assumed as random variables with a log-normal and a uniform PDF,
and Wood (1986) to validate a relation they developed for areal average and respectively. The quantity <In(F)> is estimated through the averaging proce-
variance of infiltration rate under a time-invariant rainfall rate; however, the dure over the ensemble of two-dimensional realizations of Ks and r. For the
averaging procedure was applied in space over a single realization and the sake of simplicity, we first examine the model under a steady-rainfall condition,
behavior of the resulting errors was not specified. An alternative, but not much and then provide the guidelines for applications to a time-varying rainfall rate.
used, method to M-C sampling for the representation of the random variability Starting from the extended Green–Ampt model and choosing F as the
of a soil property is the Latin Hypercube sampling (McKay et al., 1979) adopted independent variable, <In(F)> at a given F can be written as:
by Smith and Goodrich (2000) to develop a simple parameterized approach for
areal-average infiltration. Even though M-C simulations, performed by many
∞ ∞ ∞ Kc 
ψ (θ s − θi ) 
realizations of the random variable, are too expensive in terms of computa-
= ∫ ∫ r fr (r ) f K ( K ) dr d K + ∫ ∫ 1 +
s
F
 K fr (r ) f K s ( K ) dr dK

0 Kc 0 0
tional effort for practical applications, they are useful as a tool to parameterize
simple semiempirical approaches or to serve as a benchmark for validating (45.36)
semianalytical models. Along these lines, Govindaraju et al. (2001) developed where fr(r) and f Ks ( K ) are the PDFs of r and Ks, respectively, with Kc which
three versions of a semianalytical/conceptual models for estimating the expect- denotes the maximum value of Ks leading to surface saturation in the ith cell,
ed areal-average infiltration into vertically homogeneous soils under a uniform determined by
rainfall spatial distribution, but with random horizontal values of the saturated
F ri
hydraulic conductivity (see also Corradini et al., 2002). A shortcoming of this Kc = = Fc ri (45.37)
model is that the process of infiltration of overland flow running over pervious ψ (θ s − θi ) + F
downstream areas (run-on process) is neglected. On the other hand, the impor- Equation (45.36) may be expressed as:
tance of run-on was shown in a few investigations concerning the effects of
1
horizontal variability of saturated hydraulic conductivity on Hortonian over-
land flow (Smith and Hebbert, 1979; Saghafian et al., 1995; Woolhiser et al.,
< In ( F ) > =
2 R Fc2
{
M K s ( rmin + R ) Fc ,2  − M K s [ rmin Fc ,2 ] }
1996; Corradini et al., 1998; Nahar et al., 2004), but the process has generally 2
rmin
been disregarded in hydrological models.
Some studies combining the random variability of saturated hydraulic
−
2R
{
M K s ( rmin + R ) Fc ,0  − M K s [ rmin Fc ,0 ] }
conductivity and rainfall rate were performed by Wood et al. (1986), Castelli R
(1996), and Govindaraju et al. (2006). The last paper presented a semiana-

{
+  rmin +  1 − M K s ( rmin + R ) Fc ,0 
 2
}
lytical model developed under less-restrictive conditions, even though run-
1  rmin + R 
on was not incorporated, and based upon the use of cumulative infiltration as
the independent variable that was linked with an expected time. Subsequently,
+ 
Fc  R 
{
 M K s ( rmin + R ) Fc ,1 − M K s ( rmin Fc ) ,1 }
Morbidelli et al. (2006) formulated a more complete mathematical model for
1 1 
the expected areal-average infiltration, which considers both the saturated
hydraulic conductivity and rainfall rate as random variables and combines the
−
Fc  R Fc 
{
M K s ( rmin + R ) Fc ,2  − M K s ( rmin Fc ) ,2  }
aforementioned semianalytical approach with a semiempirical/conceptual 1
component to represent the run-on process. M K s [ rmin Fc ,1] +
Some models for infiltration at the field-scale have recently been developed Fc
and represent a useful support for practical hydrological purposes. (45.38)
45_Singh_ch45_p45.1-45.10.indd 5 8/22/16 2:15 PM

with rmin and rmin + R extreme values of the PDF of r and M K s given by the effective rainfall. The depth of effective rainfall from a storm as a whole is
expressed by (see also Chow et al., 1988):
Ka
M K s ( K a ,ω ) = ∫ K ω fK (K )d K (45.39) ( R − 0.2 S )2
0 RE = R > 0.2 S(45.47)
R + 0.8 S
where Ka and ω stand for the first and the second argument, respectively, of with initial abstraction assumed to be 0.2 S, R storm rainfall depth, and S
the M K s function. To relate time to F an implicit relation between the e xpected potential maximum retention which is computable, in mm, through the
value of t, <t>, and F is provided as: dimensionless CN as:
  ψ (θ − θ 
{1 − M Ks [<r > Fc ,0 ]} + F + ψ (θ s − θi ) ln  ψ (θ −s θ i+) F 
F  1000 
< t (F ) > = S = 25.4  − 10 (45.48)
<r >   s i)   CN 
∞
1 where CN (in the range 0–100) can be associated with three antecedent mois-
× { M K s [<r > Fc , −1 ]} + ψ ∆θ ∑ { M K [< r > Fc , j ]} (45.40) ture classes denoted by I, II, III, and determined using the 5-day antecedent
j=1 (
j + 1) <r > j+1 s
rainfall depth from Table 45.3. Then, for class II, the value of CN (II) can be
To extend the model by incorporating the run-on effect, an additional deduced from Table 45.4 on the basis of soil type and land use. For dry and
empirical term is used in the form (Morbidelli et al., 2006): wet conditions CN(I) and CN(III), respectively, can be estimated as:
 t b  t  4.2 CN( II)

CN( I) = (45.49a)
 ≅ < In (t ) > + < r > a   exp−c  (45.41) 10 − 0.058CN( II)
t
 pa   t pa 
23 CN (II)
where tpa is the time to ponding [see Eq. (45.10)] associated with <r> and CN (III) = (45.49b)
10 + 0.13CN (II)
<Ks>. The parameters a, b, and c are expressed by
The value of CN corresponding to the determined class is then used in
a = 2.8 CV ( r ) + CV ( K s ) 
0.36
(45.42) Eq. (45.48).
In applied hydrology, the temporal distribution of effective rainfall depth
b = 5.35 − 6.32 CV (r ) CV ( K s ) (45.43) within a storm is generally required. The described approach can be extended
for this purpose. Once the value of CN has been determined, Eq. (45.47) can
 <r > < K > 
0.3 be rewritten by substituting RE and R with the corresponding cumulative
c = 2.7 + 0.3 s
 (45.44) values RE(t) and R(t) at a time between the beginning, and the end of the
 CV (r ) CV ( K s )  storm. Then the effective rainfall depth within a time interval, Δt, is obtained
by the difference RE(t + Δt) – RE(t).
where Eq. (45.42) holds for θi<<θs and for θi→θs, we have a→0. In Eqs. Finally, a physical support for the basic equations of the model was pro-
(45.41)–(45.44), length units are in millimeters and time is expressed in vided by Moore (1987) and Steenhuis et al. (1995), and a model adaption for
hours. the initial condition was proposed by Michel et al. (2005).
When the spatial heterogeneity of r can be neglected, with spatially uni-
form and steady rainfall and negligible run-on Eqs. (45.38) and (45.40) can be
replaced by (Govindaraju et al., 2001): 45.7 OPEN PROBLEMS
ψ (θ s − θi ) + F In spite of the continuous developments of infiltration modeling, the estimate

= r 1 − M K s ( K c ,0) + M K s ( K c ,1) (45.45) of infiltration at different spatial scales, that is, from the local to watershed
F scales, is a complex problem because of the natural spatial variability of soil
  ψ (θ − θ )  hydraulic characteristics and that of rainfall. An important issue to be
F
< t > = 1 − M K s ( K c ,0) + F + ψ (θ s − θi ) ln  s i
 addressed when areal estimates are involved is that concerning the determi-
r   ψ (θ s − θi ) + F  nation of <Ks>, CV(Ks), <r>, and CV(r) together with the corresponding
∞
M Ks ( K c ,i ) quantities for soil moisture content (Morbidelli et al., 2012).
+ M K s ( K c , −1) + ψ (θ s − θi ) ∑ (45.46) The models presented here apply to infiltration into a soil matrix, when
i=1 (i + 1) r i+1 macropore flow is significant the problem becomes much more complicated
even though simplified approaches have been proposed. Two practical
The last formulation was also extended for applications involving r variable approximations to describe infiltration into soils with macropores rely upon
in time. The same method can be used to adapt Eqs. (45.38) and (45.40) for the use of modified values of the saturated hydraulic conductivity of the soil
unsteady rainfall patterns. Furthermore, the additional empirical term for matrix (Maidment, 1993) or on the representation of the two processes of
run-on could be adapted for unsteady rainfalls following the guidelines indi- infiltration controlled by the matrix potential and the macropore volume
cated by Morbidelli et al. (2006). (Bronstert and Bardossy, 1999).
The solution of the model even in the conditions of coupled spatial vari- Finally, all these models are formulated for horizontal land surfaces.
ability of r and Ks is fairly simple and requires limited computational effort. Extensions of the classical infiltration theory to inclined surfaces were pro-
The model for coupled heterogeneity of r and Ks [Eqs. (45.38), (45.40), and posed by Philip (1991) and Chen and Young (2006); however, these theories
(45.41)] was validated by comparison with the results derived starting from do not explain the results of laboratory experiments, for example, those per-
M-C sampling and using a combination of the extended Green–Ampt formula- formed on bare soils in the absence of erosion and a sealing layer (Essig et al.,
tion at the local scale with the kinematic wave approximation (Singh, 1996) that 2009; Morbidelli et al., 2015). The modeling of the slope effects has to be,
is required to represent run-on. Through a wide variety of simulations it was therefore, considered as an open problem, in particular when surfaces with
shown that: (1) the model produced very accurate estimates of over a clay vegetation are involved.
loam soil and a sandy loam soil that are very similar to Soil B and Soil SL of
Table 45.2, respectively; (2) the spatial heterogeneity of both r and Ks can be
neglected only when <r> >> <Ks> or for storm durations much greater than tpa; Table 45.3 Classification of Antecedent Moisture Classes (AMC) for
(3) the effects on produced by significant values of CV(Ks) and CV(r) are the SCS-CN Method
similar; (4) run-on has considerable role for moderate storms and high values Total 5-day antecedent rainfall (cm)
of CV(Ks) and CV(r); and (5) the model can be simplified using Eqs. (45.45)
and (45.46) for CV(r) considerably less than CV(Ks) and steady rainfalls. AMC group Dormant season Growing season
I <1.27 <3.56
45.6 SOIL CONSERVATION SERVICE RUNOFF CURVE II 1.27–2.80 3.56–5.33
NUMBER MODEL
III >2.80 >5.33
The SCS-curve number (CN) model involves an empirical formulation which [Source: Soil Conservation Service, 1972]
provides an overall representation of the losses to be considered in estimating
45_Singh_ch45_p45.1-45.10.indd 6 8/22/16 2:15 PM

references 45-7
Table 45.4 Runoff Curve Numbers for Selected Land Uses and Antecedent Moisture Class II
Hydrologic soil group
Land use description A B C D
Cultivated land:
without conservation treatment 72 81 88 91
with conservation treatment 62 71 78 81
Pasture or range land:
poor condition 68 79 86 89
good condition 39 61 74 80
Meadow: good condition 30 58 71 78
Wood or forest land:
thin stand, poor cover, no mulch 45 66 77 83
good cover 25 55 70 77
Open spaces, lawns, parks, golf courses, cemeteries, etc.
good condition: grass cover on 75% or more of the area 39 61 74 80
fair condition: grass cover on 50–75% of the area 49 69 79 84
Commercial and business areas (85% impervious) 89 92 94 95
Industrial districts (72% impervious) 81 88 91 93
Residential (average impervious)
65% 77 85 90 92
38% 61 75 83 87
30% 57 72 81 86
25% 54 70 80 85
20% 51 68 79 84
Paved parking lots, roofs, driveways, etc. 98 98 98 98
Streets and roads
paved with curbs and storm sewers 98 98 98 98
gravel 76 85 89 91
dirt 72 82 87 89
Legend: Group A—Deep sand, deep loess, aggregated silts
Group B—Shallow loess, sandy loam
Group C—Clay loams, shallow sandy loam, soils low in organic content, soils usually high in clay
Group D—Soils that swell significantly when wet, heavy plastic clays, certain saline soils
[Sources: Soil Conservation Service, 1972; Chow et al., 1988]
REFERENCES Chen, L. and M. H. Young, “Green-Ampt infiltration model for sloping

surfaces,” Water Resources Research, 42: W07420, 2006.
Arabi, M., J. S. Stillman, and R. S. Govindaraju, “A process-based transfer Chen, L., S. Sela, T. Svoray, and S. Assouline, “The role of soil-surface seal-
function approach to model tile drain hydrographs,” Hydrological Processes, ing, microtopography, and vegetation patches in rainfall-runoff processes in
20 (14): 3105–3117, 2006. semiarid areas,” Water Resources Research, 49 (9): 5585–5599, 2013.
Ahuja, L. R., “Modeling infiltration into surface-sealed soils by the Green- Chow, V. T., D. R. Maidment, and L. W. Mays, Applied Hydrology, McGraw-
Ampt approach,” Soil Science Society of America Journal, 47: 412–418, 1983. Hill, New York, 1988.
Basha, H. A., “Infiltration models for semi-infinite soil profiles,” Water Chu T. C., “Infiltration during an unsteady rain,” Water Resources Research,
Resources Research, 47: W08516, 2011a. 14 (3): 461–466, 1978.
Basha, H. A., “Infiltration models for soil profiles bounded by a water Corradini, C., A. Flammini, R. Morbidelli, and R. S. Govindaraju, “A con-
table,” Water Resources Research, 47: W10527, 2011b. ceptual model for infiltration in two-layered soils with a more permeable
Beven, K. J., Rainfall-Runoff Modelling, John Wiley & Sons, Chichester, UK, upper layer: from local to field scale,” Journal of Hydrology, 410: 62–72, 2011b.
2002. Corradini, C., R. S. Govindaraju, and R. Morbidelli, “Simplified modelling
Broadbridge, P. and I. White, “Constant rate rainfall infiltration: a versatile of areal average infiltration at the hillslope scale,” Hydrological Processes, 16:
nonlinear model. 1. Analytic solution,” Water Resources Research, 24: 145–154, 1757–1770, 2002.
1988. Corradini, C., F. Melone, and R. E. Smith, “Modeling infiltration during
Bronstert, A. and A. Bardossy, “The role of spatial variability of soil mois- complex rainfall sequences,” Water Resources Research, 30 (10): 2777–2784,
ture for modelling surface runoff generation at the small catchment scale,” 1994.
Hydrology and Earth System Sciences, 3 (4): 505–516, 1999. Corradini, C., F. Melone, and R. E. Smith, “A unified model for infiltration
Brooks, R. H. and A. T. Corey, Hydraulic properties of porous media, and redistribution during complex rainfall patterns,” Journal of Hydrology,
Hydrology Paper, 3, Colorado State University, Fort Collins, CO, 1964, p. 27. 192: 104–124, 1997.
Brutsaert, W., Hydrology—An Introduction, Cambridge University Press, Corradini, C., F. Melone, and R. E. Smith, “Modeling local infiltration for a
Cambridge, UK, 2005. two layered soil under complex rainfall patterns,” Journal of Hydrology, 237:
Bullock, M. S., W. D. Temper, and S. D. Nelson, “Soil cohesion as affected 58–73, 2000.
by freezing, water content, time and tillage,” Soil Science Society of America Corradini, C., R. Morbidelli, and F. Melone, “On the interaction between
Journal, 52: 770–776, 1988. infiltration and Hortonian runoff,” Journal of Hydrology, 204: 52–67, 1998.
Castelli, F., “A simplified stochastic model for infiltration into a heteroge- Corradini, C., R. Morbidelli, A. Flammini, and R. S. Govindaraju, “A
neous soil forced by random precipitation,” Advances in Water Resources, 19 parameterized model for local infiltration in two-layered soils with a more
(3): 133–144, 1996. permeable upper layer,” Journal of Hydrology, 396 (3–4): 221–232, 2011a.
45_Singh_ch45_p45.1-45.10.indd 7 8/22/16 2:15 PM

Dagan, G. and E. Bresler, “Unsaturated flow in spatially variable fields. Moore, R. J., “Towards a more effective use of radar data for flood forecast-
1. Derivation of models of infiltration and redistribution,” Water Resources ing,” Weather Radar and Flood Forecasting, edited by V. K. Colling and
Research, 19 (2): 413–420, 1983. C. Kirby, John Wiley & Sons, Chichester, Chap. 15, 1987, pp. 222–238.
Emmerich, W. E., “Season and erosion pavement influence on saturated Morbidelli, R., C. Corradini, and R. S. Govindaraju, “A field-scale infiltra-
soil hydraulic conductivity,” Soil Science, 168: 637–645, 2003. tion model accounting for spatial heterogeneity of rainfall and soil saturated
Essig, E. T., C. Corradini, R. Morbidelli, and R. S. Govindaraju, “Infiltration hydraulic conductivity,” Hydrological Processes, 20: 1465–1481, 2006.
and deep flow over sloping surfaces: comparison of numerical and experi- Morbidelli, R., C. Corradini, C. Saltalippi, and L. Brocca, “Initial soil water
mental results,” Journal of Hydrology, 374: 30–42, 2009. content as input to field-scale infiltration and surface runoff models,” Water
Fiori A., M. Romanelli, D. J. Cavalli, and D. Russo, “Numerical experiments Resources Management, 26: 1793–1807, 2012.
of streamflow generation in steep catchments,” Journal of Hydrology, 339: Morbidelli, R., C. Corradini, C. Saltalippi, A. Flammini, and E. Rossi,
183–192, 2007. “Infiltration-soil moisture redistribution under natural conditions: experi-
Goodrich, D. C., J-M. Faurès, D. A. Woolhiser, L. J. Lane, and S. Sorooshian, mental evidence as a guideline for realizing simulation models,” Hydrology
“Measurement and analysis of small-scale convective storm rainfall variabil- and Earth System Sciences, 15: 1–9, 2011.
ity,” Journal of Hydrology, 173: 283–308, 1995. Morbidelli, R., C. Saltalippi, A. Flammini, M. Cifrodelli, C. Corradini, and
Govindaraju, R. S., C. Corradini, and R. Morbidelli, “A semi-analytical R. S. Govindaraju, “Infiltration on sloping surfaces: Laboratory experimental
model of expected areal-average infiltration under spatial heterogeneity of evidence and implications for infiltration modelling,” Journal of Hydrology,
rainfall and soil saturated hydraulic conductivity,” Journal of Hydrology, 316 523: 79–85, 2015.
(1–4): 184–194, 2006. Morbidelli, R., C. Saltalippi, A. Flammini, E. Rossi, and C. Corradini, “Soil
Govindaraju, R. S., R. Morbidelli, and C. Corradini, “Areal infiltration water content vertical profiles under natural conditions: matching of experi-
modeling over soils with spatially-correlated hydraulic conductivities,” Jour- ments and simulations by a conceptual model,” Hydrological Processes, 28:
nal of Hydrologic Engineering: ASCE, 6 (2): 150–158, 2001. 4732–4742: 2014.
Govindaraju, R. S., C. Corradini, and R. Morbidelli, “Local and field-scale Mualem, Y. and S. Assouline, “Modeling soil seal as a nonuniform layer,”
infiltration into vertically non-uniform soils with spatially-variable surface Water Resources Research, 25 (10): 2101–2108, 1989.
hydraulic conductivities,” Hydrological Processes, 26 (21): 3293–3301, 2012. Mualem, Y., S. Assouline, and D. Eltahan, “Effect of rainfall-induced soil
Green, W. A. and G. A. Ampt, “Studies on soil physics: 1. The flow of air seals on soil water regime: wetting processes,” Water Resources Research, 29
and water through soils,” Journal of Agricultural Science, 4: 1–24, 1911. (6): 1651–1659, 1993.
Greminger, P. J., Y. K. Sud, and D. R. Nielsen, “Spatial variability of field Nahar, N., R. S. Govindaraju, C. Corradini, and R. Morbidelli, “Role of run-
measured soil-water characteristics,” Soil Science Society of America Journal, on for describing field-scale infiltration and overland flow over spatially vari-
49 (5): 1075–1082, 1985. able soils,” Journal of Hydrology, 286: 36–51, 2004.
Hillel, D. and W. R. Gardner, “Transient infiltration into crust-topped pro- Nielsen, D. R., J. W. Biggar, and K. T. Erh, “Spatial variability of field mea-
files,” Soil Science, 109, 64–76, 1970. sured soil-water properties,” Hilgardia, 42 (7): 215–259, 1973.
Holtan, H. N., “A concept for infiltration estimates in watershed engineer- Parlange, J-Y., I. Lisle, R. D. Braddock, and R. E. Smith, “The three-param-
ing,” USDA Bulletin, 41–51, 1961. eter infiltration equation,” Soil Science, 133 (6): 337–341, 1982.
Horton, R. E., “An approach toward a physical interpretation of infiltration- Péschke, G. and M. Kutilek, “Infiltration model in simulated hydrographs,”
capacity,” Soil Science Society of America Journal, 5: 399–417, 1940. Journal of Hydrology, 56: 369–379, 1982.
Kacimov, A. and Y. Obnosov, “Pseudo-hysteretic double-front hiatus-stage Philip, J. R., “The theory of infiltration: 1. The infiltration equation and its
soil water parcels supplying a plant-root continuum: the Green-Ampt-Youngs solution,” Soil Science, 83: 345–357, 1957a.
model revisited,” Hydrological Sciences Journal, 58 (1): 237–248, 2013. Philip, J. R., “The theory of infiltration: 2. The profile at infinity,” Soil Sci-
Kostiakov, A. N., “On the dynamics of the coefficient of water-percolation ence, 83: 435–448, 1957b.
in soils and on the necessity for studying it from a dynamic point of view for Philip, J. R., “The theory of infiltration: 4. Sorptivity algebraic infiltration
purposes of amelioration,” Transactions of 6th Committee International Soci- equation,” Soil Science, 84: 257–264, 1957c.
ety of Soil Science, Russia, part A, 1932, pp. 17–21. Philip, J. R., “Theory of infiltration,” Advances in Hydroscience, edited by W.
Krajewski, W. F., G. J. Ciach, and E. Habib, “An analysis of small-scale rain- T. Chow, Academic Press, New York, Vol. 5, 1969, pp. 215–296.
fall variability in different climatic regimes,” Hydrological Sciences Journal, 48 Philip, J. R., “Hillslope infiltration: planar slopes,” Water Resources Research,
(2): 151–162, 2003. 27 (1): 109–117, 1991.
Loague, K. and G. A. Gander, “R-5 revisited, 1. Spatial variability of infiltra- Philip, J. R., “Infiltration into crusted soils,” Water Resources Research, 34
tion on a small rangeland catchment,” Water Resources Research, 26 (5): (8): 1914–1927, 1998.
957–971, 1990. Poesen, J., “The influence of slope angle on infiltration rate and hortonian
Maidment, D. R., Handbook of Hydrology, McGraw-Hill, New York, NY, overland flow volume,” Zeitschrift fur Geomorphologie, 49: 117–131, 1984.
1993. Rawls, W. J. and D. L. Brakensiek, A procedure to predict Green and Ampt
Maller, R. A. and M. L. Sharma, “An analysis of areal infiltration consider- infiltration parameters, Proceedings of the American Society of the Agricultural
ing spatial variability,” Journal of Hydrology, 52: 25–37, 1981. Engineers Conference on Advances in Infiltration, American Society of Agri-
McKay, M. D., R. J. Beckman, and W. J. Connover, “A comparison of three cultural Engineers, St. Joseph, MI, 102–112, 1983.
methods for selecting values of input variables in the analysis of output from Reeves, M. and E. E. Miller, “Estimating infiltration for erratic rainfall,”
a computer code,” Technometrics, 21 (2): 239–245, 1979. Water Resources Research, 11 (1): 102–110, 1975.
Mein, R. G. and C. L. Larson, “Modeling infiltration during a steady rain,” Russo, D. and E. Bresler, “Soil hydraulic properties as stochastic processes:
Water Resources Research, 9 (2): 384–394, 1973. 1. Analysis of field spatial variability,” Soil Science Society of America Journal,
Melone, F., C. Corradini, R. Morbidelli, and C. Saltalippi, “Laboratory 45: 682–687, 1981.
experimental check of a conceptual model for infiltration under complex Russo, D. and E. Bresler, “A univariate versus a multivariate parameter
rainfall patterns,” Hydrological Processes, 20: 439–452, 2006. distribution in a stochastic-conceptual analysis of unsaturated flow,” Water
Melone, F., C. Corradini, R. Morbidelli, C. Saltalippi, and A. Flammini, Resources Research, 18 (3): 483–488, 1982.
“Comparison of theoretical and experimental soil moisture profiles under Saghafian, B., P. Y. Julien, and F. L. Ogden, “Similarity in catchment response:
complex rainfall patterns,” Journal of Hydrologic Engineering: ASCE, 13 (12): 1. Stationary rainstorms,” Water Resources Research, 31 (6): 1533–1541,
1170–1176, 2008. 1995.
Michel, C., V. Andréassian, and C. Perrin, “Soil conservation service curve Sharma, K., H. Singh, and O. Pareek, “Rain water infiltration into a bar
number method: how to mend a wrong soil moisture accounting procedure?” loamy sand,” Hydrological Sciences Journal, 28: 417–424, 1983.
Water Resources Research, 41: W02011, 2005. Sharma, M. L., R. J. W. Barron, and M. S. Fernie, “Areal distribution of
Mls, J., “Effective rainfall estimation,” Journal of Hydrology, 45: 305–311, infiltration parameters and some soil physical properties in lateritic catch-
1980. ments,” Journal of Hydrology, 94: 109–127, 1987.
Mockus, V., “Estimation of direct runoff from storm rainfall,” National Sharma, M. L. and E. Seely, “Spatial variability and its effect on infiltration,”
Engineering Handbook: Section 4: Hydrology, United States Soil Conservation Proceedings of the Hydrology Water Resources Symposium, Institute of Engi-
Service, Washington, D.C., USA, 1972. neering, Perth, Australia, 1979, pp. 69–73.
Moore, I. D., “Infiltration equations modified for surface effects,” Journal of Singh, V. P., Kinematic Wave Modeling in Water Resources: Surface Water
Irrigation and Drainage Engineering, 107: 71–86, 1981. Hydrology, John Wiley & Sons, New York, 1996.
45_Singh_ch45_p45.1-45.10.indd 8 8/22/16 2:15 PM

references 45-9
Sivapalan, M. and E. F. Wood, “Spatial heterogeneity and scale in the infiltra- Swartzendruber, D., “A quasi solution of Richard equation for down-
tion response of catchments,” Scale Problems in Hydrology, Volume 6 of the series ward infiltration of water into soil,” Water Resources Research, 5: 809–817,
Water Science and Technology Library, edited by V. K. Gupta, I. Rodríguez-Iturbe 1987.
and E. F. Wood, D. Reidel Publishing, Dordrecht, Holland, 1986, pp. 81–106. Taha, A., J. M. Gresillon, and B. E. Clothier, “Modelling the link between
Smith, R. E., “Analysis of infiltration through a two-layer soil profile,” Soil hillslope water movement and stream flow: application to a small Mediterra-
Science Society of America Journal, 54 (5): 1219–1227, 1990. nean forest watershed,” Journal of Hydrology, 203: 11–20, 1997.
Smith, R. E., Infiltration theory for hydrologic applications, Water Resources Vandervaere, J-P., M. Vauclin, R. Haverkamp, C. Peugeot, J-L. Thony, and
Monograph 15, American Geophysical Union, Washington, D.C., 2002. M. Gilfedder, “A simple model of infiltration into crusted soils,” Soil Science,
Smith, R. E., C. Corradini, and F. Melone, “Modeling infiltration for multi- 163: 9–21, 1998.
storm runoff events,” Water Resources Research, 29 (1): 133–144, 1993. van Genuchten, M. Th., “A closed-form equation for predicting the hydrau-
Smith, R. E., C. Corradini, and F. Melone, “A conceptual model for infiltra- lic conductivity of unsaturated soils,” Soil Science Society of America Journal,
tion and redistribution in crusted soils,” Water Resources Research, 35 (5): 44: 892–898, 1980.
1385–1393, 1999. Verma, S. C., “Modified Horton’s infiltration equation,” Journal of Hydrol-
Smith, R. E. and D. C. Goodrich, “Model for rainfall excess patterns on ogy, 58: 383–388, 1982.
randomly heterogeneous area,” Journal of Hydrologic Engineering: ASCE, 5: Warrick, A. W. and D. R. Nielsen, “Spatial variability of soil physical prop-
355–362, 2000. erties in the field,” Applications of Soil Physics, edited by D. Hillel, Academic
Smith, R. E. and R. H. B. Hebbert, “A Monte Carlo analysis of the hydro- Press, New York, 1980, pp. 319–344.
logic effects of spatial variability of infiltration,” Water Resources Research, 15 Wood, E. F., M. Sivapalan, and K. Beven, Scale effects in infiltration and
(2): 419–429, 1979. runoff production, Proceedings of the Symposium on Conjunctive Water Use,
Smith, R. E. and J-Y. Parlange, “A parameter-efficient hydrologic infiltra- IAHS Publication No. 156, Budapest, 1986.
tion model,” Water Resources Research, 14 (3): 533–538, 1978. Woolhiser, D. A., R. E. Smith, and J-V. Giraldez, “Effects of spatial vari-
Steenhuis, T., M. Winchell, J. Rossing, J. Zollweg, and M. Walter, “SCS ability of saturated hydraulic conductivity on Hortonian overland flow,” Water
runoff equation revisited for variable-source runoff area,” Journal of Irrigation Resources Research, 32 (3): 671–678, 1996.
and Drainage Engineering, 121 (3): 234–238, 1995. Youngs, E. G., “An infiltration method measuring the hydraulic conductiv-
Stillman, J. S., N. W. Haws, R. S. Govindaraju, and P. S. C. Rao, “A model ity of unsaturated porous materials,” Soil Science, 97: 307–311, 1964.
for transient flow to a subsurface tile drain under macropore-dominated flow
conditions,” Journal of Hydrology, 317: 49–62, 2006.
45_Singh_ch45_p45.1-45.10.indd 9 8/22/16 2:15 PM

Chapter 46
Soil Moisture and Vadose
Zone Modeling
BY
BINAYAK P. MOHANTY
ABSTRACT
Vadose Zone, connecting surface, ground, and atmospheric water, serves as
the gate keeper for terrestrial water cycle. Nonlinear partially-saturated soil Transpiration
water flow and retention based on heterogeneous soil texture, structure, Irrigation
Rain
organic matter, and other large scale land attributes contribute to the spatio-
temporal dynamics of vadose zone hydrology and resultant soil moisture. In
Soil evaporation
this chapter, we summarized the developments in vadose zone hydrology
over the past three decades. New modeling concepts including single poro Run-on Run-off
sity, dual porosity, dual permeability, and equivalent continuum approaches Near
have emerged to address the challenges related to soil heterogeneity at con- surface soil
tinuum scale. Based on these concepts, several multi-dimensional vadose (0–5 cm) moisture
zone hydrology and land surface (connecting soil-vegetation-atmosphere- data
transfer) numerical models are now available for different applications,
including agriculture, hydrology, environment, weather, and climate assess-
ment and forecast. As one of the primary variables of interest, near-surface
soil moisture measurement and modeling at various spatio-temporal scales Downward
has drawn much attention in the Earth science community. Using inistu and Soil flux q(h) q(zi, t > 0)
remotely sensed soil moisture data, evolution of dominant physical controls profile Z K(h)
on soil moisture at different spatial scales has been unraveled. In addition, depth
various bottom-up aggregation (upscaling) or top-down disaggregation Shallow water
(downscaling) schemes in simple and complex vadose zone scenarios were Upward table
developed that reflect the need and the challenges. In addition, a number of flux
forward and inverse data assimilation schemes have been formulated to
translate the space-time soil moisture (modeled and observed) data series for Deep
generating effective unsaturated soil water retention and hydraulic conduc- percolation
tivity functions for vadose zone at the scale of interest, key inputs to any land
surface models. Figure 46.1 Exchange of fluxes through the boundaries in a one-dimensional (1D) soil
water balance model across the land-atmosphere interface and vadose zone-groundwater
46.1 BACKGROUND interface (Ines and Mohanty, 2008a). θ(Zi, t) reflects the near-surface soil moisture time
series, θ(h) defines the soil water retention function, and K(h) defines the soil hydraulic
The “Critical Zone” extending from the top of the vegetative canopy conductivity function of the vadose zone (between land surface and groundwater).
through the depth of weathered bedrock plays a central role in Earth’s cli-
mate system and provides essential ecosystem services, thereby affecting
virtually every aspect of society and the environment. This zone is where air,
water, carbon, minerals, and organisms interact, forming and evolving the
soil and plant ecosystem. The soil or “Vadose Zone” (Fig. 46.1) is the heart Its temporal and spatial variability over catchment areas affects surface and
of the Critical Zone; it is the medium through which all relevant fluxes pass, subsurface runoff, modulates evaporation and transpiration, determines the
and its pulse regulates the ecosystem as a whole. Humans are dependent on extent of groundwater recharge, and initiates or sustains feedback between
soil ecosystem services, particularly soil fertility and water holding capacity, the land surface and the atmosphere (NRC, 1991). At a particular point in
for their roles in food and fiber production (Palm et al., 2007; Janzen et al., time, root zone soil moisture content is influenced by: (1) the precipitation
2010), the hydrologic cycle (Loescher et al., 2007; Ines and Mohanty 2008a; history, (2) the texture of the soil, which determines the water-holding capa
2009), the carbon cycle (Amundson, 2001), and the climate system (Lal, city, (3) the slope of the land surface, which affects runoff and infiltration, and
2004; Brevik, 2012). (4) the vegetation and land cover, which influences evapotranspiration (ET)
Soil moisture in the partially saturated root zone (approximately top ~1–3 m and deep percolation. In other terms, the partitioning of vadose zone soil
and loosely defined as shallow vadose zone in this chapter) is the natural state moisture to recharge to the groundwater, ET to the atmosphere, and surface/
variable of the land surface and subsurface critical to climate feedback, subsurface runoff to the streams at different spatiotemporal scales and under
hydrology, and agriculture, and a key component of global water cycle. different hydroclimatic conditions pose one of the greatest challenges to
46-1
46_Singh_ch46_p46.1-46.14.indd 1 8/22/16 2:15 PM

46-2 Soil Moisture and Vadose Zone Modeling
weather and climate prediction, water resources availability, sustainability, in the vadose zone is discussed in more detail in a separate chapter. In the
quality, and variability in agricultural, range and forested watersheds, and following sections, as different models may use same mathematical nota-
hydroclimatic conditions. tions for different variables, in occasions we redefine the variables for
Soil system or vadose zone reflects a three-phase system including: better clarity.
(1) solid phase constituting soil matrix or minerals and organic matters,
(2) liquid phase consists of soil water or soil solution with dissolved sub- 46.2.1 Single Porosity Model
stances, and (3) gaseous phase or soil air. Soil water and soil air vary in
composition in space and time. Soil water or moisture can be represented A unimodal pore size distribution is assumed sufficient in describing the
gravimetrically (as a ratio of mass of soil water to mass of solid) or volu- closed form expressions for the hydraulic conductivity functions for the
metrically (as a ratio of volume of soil water to total or bulk volume of soil). equilibrium SPM. In the 1D SPM, Richards’ Eq. (46.1) is used for describing
Degree of saturation in the soil system reflects the percentage of porosity variably saturated flow and Convection Dispersion Equation CDE (46.2) for
occupied by soil water. One way to represent the soil system is as a pore modeling solute transport:
bundle, which forms the basis for using Poiseuille’s Law of capillarity to
describe unsaturated flow. In most cases, soil pores are highly irregular, ∂θ ∂  ∂h 
= K (h) + 1 − S (46.1)
having an intricate geometry which prohibits microscopic description of ∂t ∂z   ∂z 
flow pathways. For this reason, flow in soils is generally described using
macroscopic or averaging terms. Although in the past three decades, ∂θ c ∂ρ s ∂  ∂c  ∂qc
advancements have been made to measure and model pore scale soil water + = θ D  − − µ (θ c + ρ s ) + γθ + γρ (46.2)
phenomena, soil water retention and hydraulic fluxes through bulk soil ∂t ∂t ∂z  ∂z  ∂z
system are mostly modeled at the continuum (Representative Elementary
Volume) scale. where t is time (T), z is the vertical coordinate positive upward (L), θ is the
water content (L3L–3), h is the pressure head (L), K is the unsaturated hydrau-
lic conductivity (LT–1), S is a sink term (e.g., root water uptake rate by plants),
46.2 CONTINUUM-SCALE MODELS FOR PARTIALLY c and s are solute concentrations in the liquid (ML–3) and solid phases (MM–1),
SATURATED FLOW IN THE VADOSE ZONE respectively, ρ is the soil bulk density (ML–3), q is the volumetric flux density
(LT–1), μ is a first-order rate constant (T–1), γ is a zero-order rate constant
In past decades, significant progress has been made in vadose zone hydrologic
(ML–3T–1), and D is the dispersion coefficient (L2T–1). This formulation allows
process understanding, conceptual modeling (Fig. 46.2), unsaturated hydrau-
SPM to describe flow and transport that is uniform and at local equilibrium
lic property characterization, and predictive vadose zone/soil vegetation
(Šimůnek et al., 2003; 2006; Köhne et al., 2009).
atmosphere transfer (SVAT) modeling. The most popular model to describe
flow in vadose zone has been pore-bundle capillary theory. While vadose
zone hydraulics can be described with a number of closed-form expressions 46.2.2 Mobile Immobile Model
for soil water retention and hydraulic conductivity (Leij et al., 1997), few Pore space heterogeneity in the subsurface also leads to nonequilibrium pref-
equations (such as Richards’ equation and Green and Ampt equation) for erential flow movement at the continuum scale. Biological macropores and
variably saturated flow remain as some of the most popular governing equa- structural fractures provide preferential flow paths for water and solutes to
tions at the continuum scale. reach groundwater (Mohanty et al., 1997; 1998; Jarvis et al., 2007). Soil- and
Among others, Arora et al. (2011) evaluated the characteristic features of crop-management practices (e.g., tillage operations, multiple cropping, etc.)
different conceptual continuum scale models [single porosity model (SPM), have been found to modify agricultural soil structure and alter macropore
mobile immobile model (MIM), and dual porosity model] for a hypothetical densities. Variations in density, connectivity, and geometry of soil pores
infiltration scenario of a soil system with matrix and macropore/fracture change preferential flow and transport characteristics of structured soils
(Fig. 46.2). While the details can be found in Arora et al. (2011) and (Arora et al., 2011). The MIM approach represents the flow field through the
Simunek et al. (2012), a summarized description of the popular concepts macropore (mobile) domain and allows for water and solute transfer between
used for continuum scale flow and related transport (for n onreactive mobile and immobile regions. As opposed to SPM, the mobile-immobile
solutes) models is given as follows. Note, however, solute transport modeling model describes soil hydraulic functions using the macropore (mobile)
domain parameters, and utilizes information on matrix (immobile) domain
for quantifying the interdomain mass transfer. Richards’ equation is used to
Model concepts simulate mobile water, and a source/sink term is used to account for water
exchange with the soil matrix (immobile region) (Šimůnek et al., 2001; Köhne
Are: DPM: ECM: et al., 2006):
(axisymmetrical-)3D (2 domains-) 1D (1 domain) 1D
∂θm ∂  ∂h 
= K (hm ) m + 1 − G wMIM (46.3)
∂t ∂z   ∂z 

∂θ im
= G wMIM = w Sem − Seim  (46.4)
∂t
Z Z Z
where ΓwMIM is the water-transfer rate from mobile to immobile region (T–1),
w is a first-order rate coefficient (T–1), and Sem and Seim are effective fluid
f m f m (f, m) saturations in the mobile and immobile regions, respectively. Convective-
dispersive solute transport is assumed for the mobile region and analogous
to water flow, first-order solute exchange process is employed between the
r
two regions (Šimůnek et al., 2003):
(f, m) ∂θmcm ∂ρm sm ∂ ∂c  ∂q c
f f + = θm Dm m  − m m − µθmcm − G sMIM (46.5)
K ∂t ∂t ∂z  ∂z  ∂z
K K
m m
∂θ imcim ∂ρim sim
+ = − µθ imcim + G sMIM (46.6)
∂t ∂t
h h h
G sMIM = w s (cm − cim ) + G wMIMc* (46.7)
Figure 46.2 Some conceptual models [discrete fracture and axisymmetrical flow,
dual permeability model (DPM), equivalent continuum model (ECM)] for describing
soil hydraulics in biporous (contrasting textures, fracture/matrix) media (Sources: Kohne where Γ sMIM is the solute transfer rate between the two regions (ML–3T–1),
and Mohanty, 2005; Arora et al., 2011). Mohanty et al. (1997, 1998) showed the ws is the constant first-order diffusive solute mass transfer coefficient (T–1),
effectiveness of ECM concept for field-scale preferential flow. and c* is equal to cm for ΓwMIM > 0 and cim for ΓwMIM < 0.
46_Singh_ch46_p46.1-46.14.indd 2 8/22/16 2:15 PM

Numerical Vadose Zoneand Land Surface Models 46-3
46.2.3 Dual Permeability Model preferential flow and solute transport in complex multidimensional field
Another concept to describe water flow in macroporus media is DPM. This conditions.
approach uses two different hydraulic functions, one for each domain, for
describing flow through the macroporous media. Exchange between the  θ s ,c − θ r ,c 
θ (h) = Σc wcθ c (h) = Σc wc θ r ,c +  h ≤ hθ* (46.18)
matrix and macropore domains is established through a first- or second-  n mc
 1 + (α c h) c  
order coupling term. In the dual permeability model, water flow in both
macropore (subscript f ) and matrix (subscript m) domains is described by
two coupled Richards’ equations (Gerke and van Genuchten, 1993): θ = θs(c = 1) h > hθ* (46.19)

∂θ f ∂  ∂h f  G
= K f + K f  − w (46.8) K (h) = ∑wc K c (h)
∂t ∂z  ∂z  wf c

(1 − (αch)n −1[ 1 + (αch)n ]−m )
2
c c c
∂θm ∂  ∂h  Gw = ∑c w c , where (mc = 1 − 1/ nc ) h ≤ hK*
=  Km m + Km  + (46.9) mc /2
1 + (α c h)nc 
∂t ∂z  ∂z  1 − wf

(46.20)
where wf is the dimensionless volume factor defined as the ratio of the
macropore domain volume (Vf) relative to the total soil volume (Vt):  (|h−h*|)δ 
K nc (h) = K *+ K * exp − 1 hK* < h ≤ 0 (46.21)
 
Vf
wf = (46.10)
K nc (h) = K *+ K *  exp(−h )δ − 1 h > 0 (46.22)
*
Vt

Γw is the rate of water exchange between the two domains (T–1) described where
with first-order mass transfer for DPM1 as: Kc the hydraulic conductivity for capillary dominated flow
domain c (LT–1);
G wDPM1 = α w (h f − hm ) (46.11) Knc the hydraulic conductivity for noncapillary dominated flow

domain nc (LT–1);
in which aw is a first-order mass transfer coefficient for water (L–1T–1) θs,c the saturated water content for capillary dominated flow
given by domain c (L3L–3);
θr,c the residual water content for capillary dominated flow
β domain c (L3L–3);
αW = K aγw (46.12)
a2 h the equilibrium soil water pressure head for bulk soil (across
all flow domains) (L);
where b is a dimensionless geometry-dependent shape factor, a is the charac- hK* ≈ hθ* = h* the critical or break-point soil water pressure head where flow
teristic length of the aggregate (L), Ka is the hydraulic conductivity of the changes from capillary dominated to noncapillary dominated
fracture/matrix interface region (LT–1), and gw is a dimensionless scaling flow or vice-versa (L);
factor. Without fracture coatings for artificial macropores, Ka can be evalu- K the hydraulic conductivity corresponding to h* (LT–1);
*
ated as follows: δ a fitting parameter representing effective macroporosity or

other structural features contributing to noncapillary domi-
K a = 0.5[K f (h f ) + K m(hm )] (46.13) nated flow (L–1);

αcnc the van Genuchten fitting parameters (van Genuchten, 1980)
DPM with a second-order term (DPM2) for inter-domain mass transfer for the capillary dominated flow domain c (L–1, -);
of water can also be considered (Köhne et al., 2004): c the number of capillary dominated flow domains, where for
c = 1 the sum type multimodal van Genuchten–Mualem
β K a (h f − hm )[|hm − hi | − |h f − hi |] hydraulic functions Eqs. [(18) and (20)] reduce to the original
G wDPM2 = (46.14)
2a 2 |hm − hi | unimodal van Genuchten–Mualem-type functions;
nc the noncapillary dominated flow domain; and
wc the weighting factor for capillary dominated flow domain c(-),
where hi is the initial pressure head assumed to be equal for matrix and subjected to ∑wc = 1 and 0 < wc < 1.
macropore (L). For DPM2, Ka is evaluated as: Although aforementioned conceptual modeling techniques at the contin-
pKm(hm ) + Km(h f ) uum scale for soil systems are more suitable for various hydrologic applications,
Ka = (46.15) pore network models, percolation theory, and other thermodynamics-based
p+1
multiphase flow and transport approaches have been used in specific environ-
mental applications. Besides these process-based models, other statistical and
where p is a weighting factor for which an average value of 17 was found to pattern recognition models for soil hydrology (e.g., pedo transfer functions,
be suitable for a range of hydraulic properties and initial conditions (Köhne Sharma et al., 2006; Jana et al., 2008) are being developed and used for simpli
et al., 2004). For both DPM1 and DPM2, geometrical parameters can be city in various occasions.
derived according to Gerke and van Genuchten (1996) as:
1
β= ; 1 < ς < 100 (46.16)
[0.19ln(16ς )]2 46.3 NUMERICAL VADOSE ZONE
AND LAND SURFACE MODELS
a+b Several vadose zone and land surface models [e.g., HYDRUS, Soil-Water-
with ς = (46.17)
b Atmosphere-Plant (SWAP), Noah, and Community Land Model (CLM)]
have been developed to simulate flow in shallow subsurface. Note, however,
where, b is the radius of the cylindrical macropore (L). some of these models were developed keeping various applications in
In addition to the MIM, DPM, and other multidomain model concepts, mind. For example, HYDRUS has been extensively applied for various
Durner et al. (1994) came up with multiple superimposed (combination) complex flow and transport processes at local scale, while SWAP has been
hydraulic functions to describe nonequilibrium flow in the vadose zone. used for many field scale water-management applications, and Noah and
Furthermore, by superimposing and joining different hydraulic functions CLM land surface models have been used in the context of large scale
across different ranges of pressure head (h) for capillary-dominated versus hydrology and weather and climate models. As readers are referred to the
noncapillary dominated flow, Mohanty et al. (1997, 1998) formulated a original sources for details, we provide brief summaries and salient features
novel piecewise-continuous hydraulic function efficient for simulating of some of these models.
46_Singh_ch46_p46.1-46.14.indd 3 8/22/16 2:15 PM

46.3.1 HYDRUS Suite of Models unsaturated hydraulic conductivity (LT–1) at matric potential h. SWAP simu-
A suite of software packages [HYDRUS _1D, 2D, and 3D (Simunek et al., lates water flow, solute transport, heat flow, and crop growth simultaneously
2006, 2012)] has been developed over the past two decades by accounting for at field scales.
the movement of water, heat, and multiple solutes in the vadose zone. These
finite element numerical models solve multidimensional mixed-form 46.3.3 Noah Land Surface Model
Richards’ equation with sink or source term for water flow and Convection Noah land surface model [Noah LSM v2.7.1 (Ek et al., 2003)] has been widely
Dispersion equation with various decay chain reaction terms for solute trans- used in both coupled (integrated with other models) and uncoupled (stand-
port processes, and can adopt various uniform or time-dependent boundary alone) modes for simulating water and energy fluxes at various spatial scales.
conditions, and hydraulic conductivity/soil water retention functions (Brooks In uncoupled mode, it can be used as a 1D physically based model for esti-
and Corey, 1964; van Genuchten, 1980; Vogel and Cislerova, 1988; Durner, mating soil moisture dynamics at field-scale. Noah LSM calculates the total
1994; Kosugi, 1996). Conceptual models, such as SPM, MIM, and DPM ET by summing the direct evaporation from top soil layer, canopy evapora-
(described in the previous section), are included in the HYDRUS package. tion, and potential Penman–Monteith transpiration (Rosero et al., 2010).
Although uncoupled, water flow in HYDRUS is given as a sum of isothermal The model has typically four soil layers with the thicknesses of 10, 30, 60,
liquid flow, isothermal vapor flow, gravitational liquid flow, thermal liquid and 100 cm (total soil depth of 200 cm). It adapts a diffusion form of the
flow, and thermal vapor flow. System-independent and system-dependent Richards’ Eq. (46.25) for estimating soil moisture. Hydraulic conductivity and
boundary conditions for water, heat, and solute transport are featured in this soil matric potential are calculated based on the Clapp and Hornberger (1978)
model. Root water and nutrient uptake with various compensation terms, [Eqs. (46.26) and (46.27)],
root growth, hysteresis in soil hydraulic properties, ET using radiation
budget, subsurface drainage, seepage, and surface runoff are the salient fea- ∂θ ∂ ∂θ  ∂K (θ )
=  D(θ )  + + Q (46.25)
tures of this popular vadose zone model. Other important features of the
∂t ∂z  ∂z  ∂z
HYDRUS model include carbon dioxide transport and production, major
ion, and carbonate chemistry involving mass and charge balance equations,  θ 
−b
complexation reactions, cation exchange and selectivity, precipitation- ψ = ψsat   (46.26)

dissolution, and various kinetic models. While water, solute, and heat trans-  θ sat 

port modeling features are quite comprehensive in HYDRUS suite of models,
some tools are available for colloid, virus, and bacteria transport in the vadose  θ 
2b + 3
zone. In addition to conducting simulation in the forward mode, HYDRUS K (θ ) = K sat   (46.27)
model is capable of inverse modeling for estimating optimum soil hydraulic  θ sat 

and transport parameters, described in Sec. 46.2. Advanced numerical
schemes to solve governing equations used in HYDRUS minimize mass bal- where θ is the volumetric soil water content (L3L–3), z is the soil depth (L)
ance errors and numerical instability. Space-time discretization based on a  ∂ψ 
fully implicit discretization of the time derivative is solved with a Picard taken positive upward, D(θ) is the soil water diffusivity (L2T–1) K (θ ) ∂θ , K(θ)
iterative solution facilitated by Gaussian elimination. Peclet number and is the unsaturated hydraulic conductivity (LT–1), Q is a soil moisture sink
Courant number are used as guiding principles for optimizing the space-time term, ψ and ψsat are the soil matric potential and saturated soil matric potential
discretization minimizing the tradeoff between accuracy of the numerical (L), b is the curve fitting parameter related to the pore size distribution (-),
solutions and computational burden. Atmospheric boundaries are simulated and θsat and Ksat are the saturated soil moisture content (L3L–3) and saturated
by applying prescribed head, flux, or gradient boundary conditions as well as hydraulic conductivity (LT–1), respectively. Noah LSM has been enhanced to
other system-dependent conditions as surface ponding including run on/ achieve the better performance incorporating complex canopy resistance, bare
runoff. Various hydrologic conditions including seepage face, deep drainage soil evaporation, surface runoff, and higher-order time integration schemes.
based on ground water table position, free drainage, and tile drainage can be
simulated using different boundary conditions. Typical applications of 46.3.4 Community Land Model
HYDRUS model includes agricultural (irrigation design, drainage design, Community land model [CLM, Oleson et al. (2010)] is the LSM that provides
water management for crop production, root and crop growth simulation, the land surface forcing as the physical boundary for atmospheric model in
salinization and reclamation processes, nitrogen dynamics and leaching to the community climate system model. This model estimates bare soil evapo-
groundwater, pesticide transport and chain products, nonpoint source pollu- ration based on the Philip and de Vries (1957) diffusion model and calculates
tion, etc.), and nonagricultural (land-atmosphere interaction, surface water- transpiration using an aerodynamic approach using the biosphere atmo-
ground water interaction, environmental solute transport and contamination sphere transfer scheme model (Dickinson et al., 1993). CLM has a 10-layered
assessment, capillary barrier design and waste disposal, landfill cover design, soil column with the fixed layer thickness of 1.75, 2.76, 4.55, 7.5, 12.36, 20.38,
tunnel and highway design, recharge basin analysis, etc.). 33.60, 55.39, 91.33, and 113.7 cm (total depth of 343 cm). The vertical soil
water flow can be solved by the modified Richards’ Eq. (46.28) (Zeng and
46.3.2 Soil-Water-Atmosphere-Plant Model Decker, 2009). This equation was derived by subtracting the hydrostatic equi-
SWAP (Van Dam et al., 1997) has been used for simulating soil water flow librium soil moisture distribution from the original Richards’ equation for
between the soil, water, plant, and atmosphere system (Agnese et al., 2007; improving the mass-conservative numerical scheme when the water table is
Ying et al., 2011). This model contains physical processes for soil water flow, within the soil column,
potential and actual ET, crop growth, and irrigation. Daily potential ET is ∂θ ∂   ∂(ψ − ψe ) 
estimated using the Penman–Monteith method with daily meteorological = K   − Q (46.28)
data or crop factors (i.e., minimum resistance, leaf area index, and crop ∂t ∂z   ∂z 

height), and the actual ET rate is calculated using the root water uptake
reduction and maximum soil evaporation flux (Van Dam et al., 1997). As where ψe is the equilibrium soil matric potential (L). The hydraulic conduc-
HYDRUS, this model simulates soil moisture dynamics in the soil profile tivity is derived from Eq. (46.27), and equilibrium soil matric potential and
using the Richards’ Eq. (46.1), and the soil hydraulic properties of Mualem equilibrium volumetric water content were shown in Eqs. (46.29) and (46.30)
and van Genuchten (Mualem, 1976; van Genuchten, 1980), based on Clapp and Hornberger (1978),
−b
m θ (z ) 
θ (h) − θ res  1  (46.23) ψe = ψsat  e  (46.29)
Se = =   θ sat 
θ sat − θ res 1 + α h n 

1
−
m2 ψ + z − z  b
K (h) = K sat Seλ 1 − 1 − Se1/m
( )  (46.24) θ e ( z ) = θ sat  sat ∇  (46.30)
  ψsat 

where h is soil matric potential (L), n(-), m(-), l(-), and α(cm–1) are the
empirical shape factors of the soil water retention and hydraulic conductivity where z is the soil depth of interface of two adjacent layers from the soil sur-
functions, m = 1-1/n, Se is the relative saturation (-), θres is the residual water face (L), and θe(z) is the equilibrium (e) volumetric water content (L3L–3) at
content (L3L–3), Ksat is the saturated hydraulic conductivity (LT–1), K(h) is the depth z (z∇ is the water table depth). In CLM, 10 soil layers discretized
46_Singh_ch46_p46.1-46.14.indd 4 8/22/16 2:15 PM

Soil Moisture AcrossSpatial-Temporal Scales 46-5
Oklahoma
Central Little Washita watershed
LB-81
facility
1-44
SGP97
flight
line LW21
area
Scale
km
EI
Reno
Little
Washita
Figure 46.3 Hierarchy of spatial scales (pore, field, landscape, watershed, region, and continent) important for understanding fundamental processes in soil hydrology and root
zone soil moisture dynamics. (Source: Zhu et al., 2006.)
unevenly include a thin top soil layer (1.75 cm) needed to better simulate ultiscale dynamics of root zone soil moisture in land-surface hydrologic
m
infiltration and evaporation fluxes (Sahoo et al., 2008). Furthermore, CLM systems (Fig. 46.3). Traditionally, soil moisture spatial variability studies
considers the dynamics of water table as the lower boundary using the SIMple using ground-based point-scale (i.e., sampling area of cm2) measurements
Groundwater Model [SIMGM, Niu et al. (2007)]. A groundwater component are limited to small fields with uniform soil characteristics, topographic
is defined as an unconfined aquifer below the soil column (343 cm). features, and vegetative conditions. Lately, the use of various active and
Other land surface models, such as variable infiltration capacity (Liang passive microwave remote sensors has enhanced the capability to monitor
et al., 1994), Mosaic land surface model (Mosaic LSM) (Koster and Suarez, near-surface soil moisture in large land areas (i.e., hundreds of square
1996), transport of unsaturated groundwater and heat (Pruess et al., 2012), meters to thousands of square kilometers) encompassing various soil types
parallel watershed flow model (Maxwell, 2013), and subsurface transport over (e.g., texture), topographic features (e.g., slope), vegetation/land cover, and
multiple phases (White et al., 2012) are also capable of simulating multiphase climatic conditions. These remote-sensing signals give average values over
vadose zone flow processes. The global land data assimilation systems uses an area usually known as a footprint. For larger footprints, only predomi-
these hydrological models for validating pixel-scale soil moisture from satel- nant soil and vegetation types are used for calibration purposes. An inher-
lite platforms and evaluating water/energy cycle and fluxes near the land ent difficulty of remotely sensed soil moisture measurement is relating soil
surface (Liu et al., 2009). The North American land data assimilation system moisture variability at the scale of the footprint to larger or smaller scale
has monitored and predicted hydrological drought conditions using state soil moisture variability (Stewart et al., 1996; Kumar, 1999). In the recent
variables (e.g., soil moisture dynamics, runoff, evaporation, etc.) estimated past, concern has been raised that footprint-scale measurements may be
from various hydrological models (Ek et al., 2011). However, these models too crude for providing a good understanding of hydrologic systems at the
incorporated with their own parameterization schemes and simplified pro- subgrid (field, catchment, basin, and watershed) scales. Hence errors at
cesses that might not consider adequately the real-world conditions indicat- the level of catchments, basins, and watersheds may add up to providing
ing that each model has its own strengths and drawbacks for certain processes inaccurate regional-scale hydrologic fluxes important for water resources
(Hsu et al., 2009). Thus, as discussed in Kirchener (2006), inherent model assessment, agriculture, and hydroclimatic predictions. The temporal and
structures might produce different model outputs and cause uncertainties due spatial distribution of soil moisture and the resultant fluxes to different
to different model structures and input parameters (i.e., atmospheric forcings, hydrologic reservoirs are affected by interactions between soil, topography,
soil textures, vegetation covers, initial and bottom boundary conditions, etc.). vegetation, and climate (Reynolds, 1970; Sharma and Luxmoore, 1979;
To circumvent some of these model-specific uncertainties, multimodel aver- Moore et al., 1988; Loague, 1992; Charpentier and Groffman, 1992;
aging techniques with different weighting schemes (Kim et al., 2015) have Rodriguez-Iturbe et al., 1999; Mohanty and Skaggs, 2001; Joshi and
been developed in the recent years. Mohanty, 2010; Gaur and Mohanty, 2013). In-depth reviews of soil mois-
ture spatiotemporal variability-related studies are given by Mohanty and
Skaggs (2001), Western et al. (2002), Das and Mohanty (2006), Joshi and
46.4 SOIL MOISTURE ACROSS Mohanty (2010), Crow et al. (2012), Caldwell et al. (2012), and Gaur
SPATIAL-TEMPORAL SCALES
and Mohanty (2013; 2015, submitted to Water Resources Research), while
While challenges persist to provide in-depth process understanding or to a brief summary is given as follows.
upscale the characteristic properties for describing soil moisture and vadose
zone flow processes at various societal applications (e.g., agricultural manage- 46.4.1 Root Zone Soil
ment, hydrologic prediction, environmental protection, weather, and climate Heterogeneity in root zone soil affects the distribution of soil moisture
feedback) scales (e.g., field, catchment, watershed, river basin, or region), through variations in texture, organic matter content, porosity, structure, and
modern ground based and remotely sensed platforms encompassing large macroporosity (Fig. 46.4). Each of these factors affects the fluid transmission
land areas on a time-continuous basis provide tools to fill in some of these and retention properties in different ways (i.e., Entekhabi, 1994; Ek and
gaps lately. Cuenca, 1994; Kim et al. 1997; Kim and Stricker 1996; Mohanty et al., 1997;
To date, very few studies (e.g., Das and Mohanty, 2006, 2008; Das et al., Liu and Dickinson 2003; Joshi et al., 2011; Shin et al., 2012; Pollacco and
2008a, 2008b, 2010) have been made to quantitatively understand the Mohanty, 2012).
46_Singh_ch46_p46.1-46.14.indd 5 8/22/16 2:15 PM

radiative (aspect) effects, and heterogeneity in vegetation and soil character-

istics. Charpentier and Groffman (1992) studied the effects of topography and
moisture content on the variability of soil moisture within remote sensing
pixels during the First International Satellite Land Surface Climatology
Project field experiment. They showed that within-pixel (66 m × 66 m) soil
moisture variability increased with increased topographic heterogeneity.
A flat pixel had significantly lower standard deviations and fewer outlier
points than a sloping or a valley pixel. Furthermore, they observed that
remote sensing reflected soil moisture conditions less accurately on pixels
Biological pores Different soil textures with increased topographic variability and less precisely when soil is dry.
Mohanty et al. (2000a) showed the dominance of a slope-effect on the diurnal
soil moisture distribution in a gentle slope within the Little Washita agricul-
tural watershed, Chickasha, Oklahoma. Several other studies also showed that
the slope position is very important in determining soil moisture variation
(Joshi and Mohanty, 2010; Joshi et al., 2011), suggesting that a simple averag-
ing of soil moisture values over the slope may lead to errors at different time
scales. While a number of studies concluded that in some landscapes topo
graphy is a relatively poor indicator of soil moisture patterns and variability
[Western et al. (1999), Kim and Barros (2002), Bindlish and Barros (2002),
Structural cracks Karst geology and Chang and Islam (2003), Jawson and Niemann (2007)], other studies,
including Jacobs et al. (2004), Joshi et al. (2011), and Jana and Mohanty
Figure 46.4 Various biotic or abiotic preferential flow paths in vadose zone including:
(1) biological pores, (2) structural cracks, (3) varying soil textures and their adjacency,
(2012a; 2012b) show that soil properties (i.e., percent silt, percent sand, and
and (4) karst geology. soil texture), and topography (elevation and slope) are significant physical
controls jointly affecting the spatiotemporal evolution and time stability of
soil moisture at both point- and remote sensing footprint-scale.
Jana and Mohanty (2012a; 2012b; 2012c) conducted a comprehensive
Variability in soil hydraulic conductivity and soil water retention in the numerical study using a three-dimensional (3D) soil hydrologic model
presence of contrasting textures/macropores/fractures greatly influences the (HYDRUS_3D with surface flow routing) under same atmospheric forcing
vertical and lateral water transmission properties (Mohanty et al., 1997). followed by field testing to investigate various topographic variations (e.g.,
Heterogeneities in soil particles and pore sizes can lead to significant soil microheterogeneity with mild slope, steep slope, concave slope, and convex
moisture variations over very small distances. At the field plot scale, Das slope) and the consequential soil moisture distributions across hill slope, and
Gupta et al. (2006a) showed the significant effect of small scale soil texture and watershed. They suggested that topographic variations result in different
macroporosity variations on soil moisture distribution across depth as well as behavior of soil moisture distribution at subgrid-scale, and boundary fluxes,
partitioning of precipitation between surface runoff, subsurface baseflow, ET, such as seepage and recharge (see Fig. 46.5). This finding has significant
and groundwater recharge in a limestone Karst environment (see Fig. 46.4). implications for regional hydrology as boundary fluxes at the subgrid-scale
Just as soil moisture variation depends on soil properties, spatial variation in will propagate to adjacent units and accentuate larger differences and scale
the properties and thickness of the organic and mineral soil layers depends on behavior of root zone soil moisture and hydrologic fluxes across field, hill
long-term local soil moisture conditions. This may further be complicated by slope, catchment, to watershed.
the dynamic nature of preferential flow paths caused by swelling and shrink-
ing clay (e.g., Das Gupta et al., 2006b). The dual permeability and equivalent 46.4.3 Vegetation
continuum modeling (discussed earlier) that have been applied to simulate Land cover is also critical for understanding the soil moisture regimes, as it
preferential flow systems in contrasting soil textural or fractured media have affects infiltration, runoff, and ET through root water uptake. Vegetation type,
demonstrated the importance of these processes in soils. density, and uniformity are some of the associated features that contribute to
soil moisture variation at different space and time scales. Furthermore, the
46.4.2 Topography influence of vegetation on soil moisture and temperature is more dynamic as
Topography (macro or micro) also plays an important role in the spatial compared to soil and topographic factors. Previous studies show that the
organization of soil moisture at different scales. Variations in slope, aspect, variability of soil moisture is lowest with full canopy cover and highest with
curvature, upslope contributing area, and relative elevation all affect the dis- partial coverage. Hawley et al. (1983) demonstrated that various vegetation-
tribution of soil moisture near the land surface. At the small catchment and topography-soil combinations lead to temporal persistence (clustering) of soil
hill-slope scales, soil moisture varies as a result of water-routing processes, moisture patterns in complex terrains with mixed vegetation. They also
Figure 46.5 Soil moisture distributions (modeled using 3D vadose zone model, HYDRUS_3D) under different topographical conditions, clockwise from top left
(a) microtopography with mild slope, (b) steep slope, (c) concave slope, and (d) convex slope under same boundary conditions. (Source: Jana and Mohanty, 2012a.) Results signify
soil moisture and local/boundary fluxes will vary for different landscape configurations for the same atmospheric forcing.
46_Singh_ch46_p46.1-46.14.indd 6 8/22/16 2:15 PM

Soil Moisture AcrossSpatial-Temporal Scales 46-7
Iowa (Point scale-soil moisture) Oklahoma (Point scale-soil moisture)

0.6
SMEX02 SMEX03
0.6 SMEX05 Vegetation dominated 1 SMEX07 Topography dominated
0.4
0.5
∆ Entropy
∆ Entropy
0.2
0
0
–0.5
–0.2
–0.4 –1
Soil dominated Soil dominated
–1.5
160 165 170 175 180 185 190 195 155 160 165 170 175 180 185 190
Doy Doy
Iowa (Airborne scale-soil moisture) Oklahoma (Airborne scale-soil moisture)

1 1
SMEX02 (PSR) SGP 97 (ESTAR) SGP 99 (ESTAR)
0.5 Vegetation dominated 0.5 Topography dominated

∆ Entropy
∆ Entropy
0 0
–0.5 –0.5
Soil dominated Soil dominated

–1 –1
175 180 185 190 195 165 170 175 180 185 190 195 200 205
Doy Doy
Figure 46.6 Time series of entropy difference of soil moisture with error bars for (a) point support scale, and (b) airborne footprint scale. For Iowa: ∆ Entropy = Soil-based
entropy − Vegetation-based entropy, for Oklahoma: ∆ Entropy = Soil-based entropy − Topography-based entropy. Entropy analysis shows the dominance of different physical
factors (soil, vegetation, and topography) under different land covers (agricultural vs. natural), wetness conditions, and measurement scales (point scale vs. airborne footprint).
(Source: Gaur and Mohanty, 2013.)
s uggested that the presence of vegetation tends to diminish the soil moisture vegetation atmosphere transfer (SVAT) modeling, Salvucci (2001) showed
variations caused by topography. Mohanty et al. (2000b) showed that the veg- conditional dependence of soil moisture storage, drainage+runoff, and ET
etation dynamics (growth/decay), land management (tillage), and precipita- with amount of precipitation in Illinois. A multiscale analysis by Crow and
tion events controlled the intraseasonal soil moisture spatial structure for the Wood (1999) revealed a qualitatively different relationship between soil mois-
pixel with flat topography and uniform soil texture. ture means and soil moisture spatial variances when variability is sampled at
Past studies in temperate climate have suggested that soil, topography, and fine (<1 km) versus coarse (>10 km) spatial scales. The threshold between
vegetation conjointly control the root zone soil moisture spatiotemporal vari- these two scale regimes may represent a transition between organized coarse
ability and temporal stability. Furthermore, the influence of vegetation on soil scale spatial heterogeneity imposed by land surface response to rainfall and
moisture was found to be more dynamic across season and plant growth disorganized fine scale variability produced by local variations in soil, topog-
stages (Mohanty et al., 2000b; Pollacco and Mohanty, 2012; Gaur and raphy, and vegetation. Peters-Lidard et al. (2008), Pollacco and Mohanty
Mohanty, 2013) as compared to soil and topographic factors. In an agricul- (2012), and Gaur and Mohanty (2015) further reflected the significance of
tural watershed in Iowa, using entropy based analysis (see Fig. 46.6), Gaur and quality and accuracy of precipitation data in soil hydrology.
Mohanty (2013) showed that at the point support scale during the relatively
wet Soil Moisture Experiment (SMEX 2005), vegetation appeared to domi- 46.4.5 Dominant Physical Controls on
nate the root zone soil moisture spatial distribution. However, during the Soil Moisture and Scale
drier year (SMEX 2002), the controls shifted between soil and vegetation Based on the aforementioned discussion of various physical controls of
during precipitation events. On the other hand, in a natural (grassland) water- soil moisture, a general conclusion could be drawn that soil moisture spatio-
shed in Oklahoma, topography and soil played dominant roles for root zone temporal patterns reflect a conjoint variability of soil, topography, vegetation,
soil moisture spatiotemporal distribution. and precipitation attributes (Fig. 46.7). Jana (2010) suggested that soil moisture
variability is “dominated” by soil properties at the field scale, topographic fea-
46.4.4 Climate tures at the catchment/watershed scale, vegetation characteristics, and precipi-
Precipitation, solar radiation, wind, and humidity are some of the important tation patterns at the regional scale and beyond. Ensemble hydrologic fluxes
climatic factors that contribute to the space-time dynamics of root zone soil (including ET, infiltration, surface runoff, subsurface baseflow, streamflow, and
moisture (Das and Mohanty, 2008; Joshi and Mohanty, submitted; Shin et al., groundwater recharge) within and across the root zone reflect the evolution of
2013b) and resultant fluxes. Precipitation is the single most important cli- soil moisture at a particular spatial scale (point, field, watershed, or region) and
matic forcing for soil moisture content and its distribution. As shown by can be “effectively” represented by one or more linear/nonlinear hydrologic
Sivapalan et al. (1987), the dominant runoff producing mechanism may vary scale parameters reflecting dominant heterogeneity of the landscape.
with storm characteristics and antecedent soil moisture conditions resulting Mohanty and Skaggs (2001), Jacobs et al. (2004), Joshi and Mohanty (2010),
in the spatiotemporal variability in soil moisture. During the Southern Great Joshi et al. (2011), and Gaur and Mohanty (2013) conducted analyses of
Plains 1997 (SGP97) hydrology campaign, Famiglietti et al. (1999) found a ground and remote sensing soil moisture data collected during the Southern
distinct trend in mean soil moisture for Little Washita (in southern Great Plains Hydrology Campaigns in 1997 and 1999 (SGP97 and GP99), Soil
Oklahoma), El Reno (in central Oklahoma), and DOE-ARM Central Facility Moisture Experiments 2002, 2003, and 2005 (SMEX02, SMEX03, and
(in north Oklahoma) locations with a south-to-north precipitation gradient. SMEX05), and Cloud and Land-Surface Interaction Campaign (CLASIC
Within an analytical framework, Kim and Stricker (1996) showed the signifi- 2007) and concluded that characteristic differences were observed in the
cance of rainfall pattern on partitioning of water over the budget terms for space-time dynamics of soil moisture within selected remote sensing footprints
different climatic conditions. In a study, using statistical distribution and soil with various combinations of soil texture, slope, vegetation, and precipitation.
46_Singh_ch46_p46.1-46.14.indd 7 8/22/16 2:15 PM

Watershed scale
Regional scale
Field scale
Climate forcings
Land cover patterns
Topographic features
Soil texture
and structure
Point Coarsening Continental

scale scale scale
Figure 46.7 Multiple/dominant physical controls for soil moisture at different spatial scales ranging from local, field, watershed, regional, to continental scale by Jana (2010).
From a physical control perspective, soil moisture dynamics is related to tran- Zhang et al., 2009; Shin et al., 2012; Shin et al., 2013a). Zhang et al. (2008)
sient upward (ET), downward (infiltration, groundwater recharge), and lateral integrated several global optimization algorithms (i.e., GA, SCE-UA, PSO,
(surface runoff, subsurface baseflow) fluxes. etc.) with soil and water assessment tool and compared their performances in
Brutsaert (1982) recognized two-stages of soil profile drying: (1) an atmo- calibrating model input parameters. They showed that the GA found better
sphere-controlled stage followed by (2) a soil-controlled stage. In the first- optimized model parameters than others, although a large number of compu-
stage, the moist soil profile can fully supply all the water demanded by the tational resources were required. Further the near-surface (Ines and Mohanty,
atmosphere. As the soil near the surface dries out, the moisture delivery rate 2008) and layer-specific data assimilation (Shin et al., 2012) approaches using
is limited by the properties of the soil profile (Pollacco and Mohanty, 2012; the GA coupled with SWAP based on inversion model were developed for
Shin et al., 2012). Similarly, following any precipitation event, vertical and quantifying effective soil hydraulic properties in the homogeneous and het-
lateral subsurface flow and its connectivity is driven by soil hydraulic proper- erogeneous soil profiles. Their findings indicated that the estimated effective
ties and their anisotropy resulting in root zone moisture distribution. With soil parameters at the near-surface and subsurface layers can be adequately
the recent launch of Soil Moisture Active Passive (SMAP) satellite, near-sur- conditioned by the GA. However, although model parameter uncertainties for
face and root zone soil moisture products will be available globally. While a single model can be minimized by simulation-optimization schemes
near-surface (top 5 cm) soil moisture (L2 and L3 products) will be available (e.g., GA-SWAP, etc.), bias due to different model structures still remain
at 3, 9, and 36 km spatial resolutions on 1–2 days intervals, root zone (top (considerably) in model outputs.
1 m) soil moisture (L4) products will be available weekly at 9 km resolution.
These SMAP soil moisture products will provide the basis for better under- 46.5.1 Top-Down Aggregated Approach
standing of land surface hydrologic stores and fluxes (ET, infiltration, surface Remote sensing data have been used successfully in extracting biophysical
runoff, subsurface baseflow, stream flow, and groundwater recharge) leading data for regional hydrological modeling (Kite, 2001). Recently Ines and
to improved weather, climate, hydrologic, and agricultural models. Mohanty (2008a) and Shin et al. (2012), respectively, presented near-surface
soil moisture assimilation scheme for homogenous (no soil layers) and het-
erogeneous (with soil layers) soil texture, flat topography, and uniform land
46.5 INVERSE MODELING—SOIL HYDRAULIC
cover that can be used to quantify effective soil hydraulic parameters in the
PROPERTIES AT THE MODEL GRID SCALE
root zone soil profile. The (top-down) approach (review by Mohanty, 2013)
Vadose zone and land surface models described above have played a vital role included an indirect way of assimilating the near-surface soil moisture data
in understanding the processes of the SVAT system. However, the success of into soil hydrologic models to derive saturated and unsaturated hydraulic
our SVAT models is highly dependent on the quality (appropriateness for conductivity K(h) and soil water retention θ(h) properties of the soil at the
working scale) and availability of data required for the model applications in footprint/grid/pixel-scale.
a specific hydrologic domain. Both the physical (e.g., soil hydraulic data) and
nonphysical (management practices) properties of a hydrologic system are 46.5.2 Top-Down Stochastic Approach
necessary to be defined when applying models for simulation studies. The Bresler and Dagan (1983) concluded that in spatially variable fields, stochastic
root zone soil hydraulic properties, defined mainly by the water retention θ(h) modeling represents the actual flow phenomena in the porous media more
and hydraulic conductivity K(h) functions, are necessary inputs to hydrologic realistically and provides the main statistical moments (mean and variances)
models; where θ is the volumetric soil moisture, K the hydraulic conductivity by employing simplified flow models, which can be used with confidence in
and h the soil water pressure. They describe the potential of the soil to store the applications. Along the line, using a Noisy Monte Carlo Genetic
and release, and to transmit water under different environmental and bound- Algorithm (NMCGA) in conjunction with a SWAP model, Ines and Mohanty
ary conditions. Among various soil hydraulic and water retention functions, (2008b) suggested that the main assumption used in the (top-down) inverse
Brooks and Corey (1964), Clapp and Hornberger (1978), van Genuchten modeling of RS near-surface soil moisture data is that the temporal RS data
(1980), and Russo (1988) are some of the most popular in SVAT models. contain (enough) information that can possibly describe the effective (root
Parameter optimization, minimizing the difference between model predic- zone) hydrologic conditions of a pixel, such that when inverted, it (i.e., the RS
tions and observations, lies in the heart of inverse modeling. Genetic algo- data) will provide a set of soil hydraulic parameters representative of that
rithms [GAs, Holland (1975)], Shuffled Complex Evolution-University of pixel. This resulted in a more general approach of a scale-dependent para
Arizona [SCE-UA, Duan et al. (1992)], and Particle Swarm Optimization meter estimation concept (Fig. 46.8) such that both the effective soil hydraulic
[PSO, Kennedy and Eberhart (2001)] have been applied in estimating effec- parameters and their uncertainties could be determined simultaneously at a
tive model parameters in hydrology. Bayesian Model Averaging [BMA, remote sensing footprint/model grid.
Hoetting et al. (1999)], Hydrological Uncertainty Processor (Krzysztofowicz, Ines and Mohanty (2008b) determined the set of effective van Genuchten
1999; Krzysztofowicz and Kelly, 2000), Ensemble Model Output Statistics soil hydraulic parameters p*= {α*, n*, qres*, qsat*, Ksat*} to accommodate the
E-MOS (Gneiting et al., 2005), DiffeRential Evolution Adaptive Metropolis— uncertainty terms (e.g., analogous to horizontal heterogeneity in a pixel) of
Markov Chain Monte Carlo (DREAM-MCMC) algorithm (Vrugt et al., each parameter. However, soil hydraulic properties at the near-surface may
2009), and Model Conditional Processor (Todini, 2008; Coccia and Todini, not be always representative of the sub-surface (root zone) unless the soil
2011) have been used to account for the model structural uncertainties. GAs profile is statistically homogenous. Ines and Mohanty (2008a, 2008b) showed
have been used to minimize errors in searching optimized model parameters that effective soil hydraulic parameters derived from the near-surface (0–6 cm)
based on inversion model (Reed et al., 2000; Ines and Mohanty, 2008a, 2009; are applicable to some degree in representing the soil hydrologic processes in
46_Singh_ch46_p46.1-46.14.indd 8 8/22/16 2:15 PM

Inverse Modeling—Soil Hydraulic Properties at the Model Grid Scale 46-9
Remote sensing (km scale)
h
RS Scale 3
footprint scale
θ
Bottom-up approach
Top-down approach
h
Scale 2
h
In situ (cm scale)
Point scale Scale 1
θ
Figure 46.8 Top-down approach: panels show the concept of effective soil water retention properties (θ - h) and their uncertainties using soil moisture time series [θ(z, t)]
at different scales using stochastic NMCGA. (Source: Ines and Mohanty, 2008b.)
the subsurface (root zone). But under highly heterogeneous soils, they are less (Fig. 46.10). In more recent efforts, based on Efendiev and Hou (2009),
successful. Recently, Shin et al. (2013b) recommended that additional infor- multiscale finite element schemes with nonlinear basis function are being
mation (e.g., deeper layer soil moisture data and ET from RS) should be used developed for linking soil water flow in finer grids within coarse grid in a
in addition to the near-surface soil moisture to characterize better the effective computationally cost effective manner, while preserving the small as well as
soil hydraulic properties of the soil profile. Pollacco et al. (2013) formulated large scale landscape (heterogeneity) features in the flow domain.
a selector algorithm by devising multiobjective functions for improved para
meter optimization using remotely sensed soil moisture and ET. 46.5.4 Multiscale Modeling
Flexible multiscale modeling and uncertainty quantification tools play a criti-
46.5.3 Bottom-Up Approach cal role in performing predictive multiscale simulations. Traditionally, con-
In a series of study, using a combination of stochastic and aggregated ceptual flow models at various scales, such as pore, continuum, field,
approach, Zhu and Mohanty (2002a; 2002b; 2003; 2004; 2006), Zhu et al. catchment, watershed, and region, are considered in hydrology. Bridging
(2004; 2006), and Mohanty and Zhu (2007) investigated the suitability of scales is often accomplished via up/downscaling. Upscaling techniques form
various soil hydraulic parameter upscaling schemes (bottom-up approach) by coarse-scale equations with a prescribed analytical form that may differ from
matching their prediction performances with ensemble vadose zone fluxes underlying fine-scale equations. Coarse-scale parameters or functional rela-
(evaporation and infiltration) under different hydroclimatic scenarios for tions are evaluated by averaging appropriate variables obtained from local
horizontally and vertically heterogeneous soil systems (see Fig. 46.9). solutions.
Their synthetic experiment results showed that soil texture, geological layer- Multiscale methods have emerged as powerful simulation techniques that
ing, groundwater depth, surface/profile soil moisture status, vertical flux direc- can be used to link models at different scales. Multiscale methods share simi-
tion (evaporation vs. infiltration), hydraulic parameter statistics, correlations, larities with upscaling methods, though there are important differences. In
and spatial structures, root distribution in soil profile, and topographic features/ multiscale methods, the fine-scale information is carried throughout the simu
arrangements conjointly determines the “upscaled” pixel-scale soil hydraulic lation, while coarse-scale equations are generally not expressed analytically,
parameters for the equivalent homogeneous medium that delivers the same but rather are formed and solved numerically. Multiscale methods allow
amount of flux (evaporation or infiltration) as the natural h eterogeneous downscaling of the processes at the regions of interest and extracting the fine-
medium. Thus, different homogenization algorithms (rules) for different scale features of the processes. Moreover, multiscale methods are scalable and
hydrologic scenarios and land attribute complexities were suggested for para the simulations can be performed perfectly parallel. Predictive multiscale
meter upscaling. simulation techniques are needed, which can pass the information about sub-
The link between fine scale and coarse scale soil hydraulic conductivity for scale effects and associated uncertainties from one scale to another. Some of
root zone soil water flow processes is nontrivial and one needs to take into these approaches are developed for simpler applications and can be extended
account the effects of all the heterogeneities present at the fine scale to more complex critical zone.
q–
1.20 4.00
h=0.1
h=0.1 q h=0.2
0.90 h=0.3 3.00
h=0.3
h=0.5 Ksn, an
h=0.4
ECA
h=0.7
0.60 2.00 h=0.5
ECA
h=1.1
h=0.7
h=2
q1 q2 q3 qn 0.30 h=0.99
1.00
h=2
Ks1, a1
Ks2, a2
Ks3, a3
Ksn, an
Ks3, a3
0.00 0.00
0.00 0.30 0.60 0.90 Ks2, a2 0.00 0.30 0.60 0.90
CVA Ks1, a1 CVA
Figure 46.9 Bottom-up approach: soil hydraulic properties heterogeneity orientation (horizontal vs. vertical) and its impact on upscaled effective parameter coefficient (ECA) for
Gardner’s α for different land surface pressure head (h) and coefficient of variation of α (CVA). (Source: Mohanty and Zhu, 2007.)
46_Singh_ch46_p46.1-46.14.indd 9 8/22/16 2:15 PM

0.52 0.5 0.5 0.4

1 0.5 1
5 0.5
1.5 1.5 0.35
0.48
0.48 2 2
10 0.3
2.5 0.46 2.5
0.46
3 3 0.25
0.44 0.44
15 3.5 3.5
4 0.42 0.2
0.42 4
20 4.5 4.5
0.4 0.4 0.15
5 5
25 0.38 5.5 0.38 5.5 0.1
5 10 15 20 25 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5
Figure 46.10 Example of multiscale soil water flow modeling using upscaled soil hydraulic parameters, (left panel) soil moisture on the fine grid, (middle panel) fine-scale soil
moisture averaged over a coarse grid, and (right panel) coarse-scale soil moisture using coarse-scale model with upscaled soil hydraulic parameters. Results signify the differences in
soil moisture and hydrologic flux distribution based on grid size and model parameter treatment. (Source: Mohanty, unpublished work.)
46.5.5 Soil Moisture Upscaling Based on Multiscale 46.5.6 Soil Moisture Downscaling
Observation, Modeling, and Data Assimilation
Satellite-based soil moisture footprints provide available large-scale applications,
Data-assimilation schemes have been gaining ground over the past decade as because of their spatiotemporal extents. However, modeling for catchment/
a means for improving soil moisture predictability using process-based SVAT watershed scale hydrology, agricultural water management, environmental
models in conjunction with spatiotemporal observations from different in fate and transport, and ecosystems services particularly occur at finer-scale
situ/remote sensing platforms. Data from different sources are injected at ranging from several hundred meters to several kilometers. Ines et al. (2013)
different times into model outputs and, moreover, the boundary and initial developed a stochastic unmixing method for soil moisture using a simulation-
conditions as well as forcing terms. Consequently, our computations become assimilation scheme. This approach extracts soil-type identification (repre-
dynamic data-driven simulations. Data assimilation can be broadly divided senting soil hydraulic properties) and subarea fractions corresponding to
into two categories: variational techniques with adjoint model and (ensemble soil-vegetation combinations within a remote sensing footprint scale soil
or extended) Kalman filter techniques. Dunne and Entekhabi (2005) pro- moisture product. However, the stochastic disaggregation method estimates
vides a review of applications of these techniques in land data assimilation only the soil characteristics (soil texture–based ID values) and subarea frac-
including comparative advantages and limitations. Among others, Das and tions (% by the soil-vegetation combinations) within a pixel in a probabilistic
Mohanty (2006), Das et al. (2008a; 2008b; 2010), and Pollacco and Mohanty sense without their specific locations practically recognized. In order to realize
(2012) report on some of the significant data-assimilation studies related to the full potential, a new deterministic downscaling algorithm was developed
land surface hydrologic modeling including soil moisture dynamics in gen- for estimating fine-scale soil moisture within remotely sensed soil moisture
eral and several in the Great Plains region (with legacy data of many soil and ET products using genetic algorithm (Shin and Mohanty, 2013a, 2013b).
moisture field campaigns) in particular. The main idea of the procedure This approach was evaluated under various synthetic and field (Little Washita,
consists of updating model parameters, such as fine-scale hydraulic conduc- Oklahoma) conditions including homogeneous and heterogeneous land sur-
tivity field at each time step when fresh data is injected (Das et al., 2008a, see face composed of different soil textures and vegetation. Analysis for various
Fig. 46.11). For the data representing large (coarse) scale information, ade- landscapes provides deeper insight to dominant hydrologic controls for root
quate upscaled model is used and upscaled parameters are estimated. zone soil moisture at different scales and landscape complexities.
Further, the fine-scale model is updated using the coarse-scale information
and hierarchical Bayesian methods. Data-assimilation techniques with 46.6 SUMMARY
Markov Chain Monte Carlo (MCMC) schemes provide an efficient platform
for deriving upscaled hydrologic parameters using process-based models and In summary, modeling unsaturated flow and transport in vadose zone has
remote sensing data. Bayesian scheme further provide uncertainty bound for made significant progress in the past three decades. However, as in any other
these parameters. subdisciplines of hydrology, challenges remain in terms of dealing with scale,
Multivariate AMSR-E
distribution measurements
Time series θ
Relaxed PDF chosen from dominant soils
θr Posterior distribution
Likelihood of upscaled hydraulic
θs
type within AMSR-E footprint
MCMC
α Prior Z Upscaling parameter
1
0.8
Beta
n
0.6
Markov random
0 1 2 3 4 5
Ksat
Iterations x 104
Mean posterior plot
β 0.04
p (mean)
0.02
0
0.4 0.6 0.8 1
Mean
Figure 46.11 MCMC-based scheme for deriving upscaled soil hydraulic parameters that provides a basis for multiscale soil moisture data-assimilation. (Source: Das et al., 2008a.)
46_Singh_ch46_p46.1-46.14.indd 10 8/22/16 2:15 PM

References 46-11
heterogeneity, and uncertainty in the vadose zone. Although the formalized Das Gupta, S., B. P. Mohanty, and J. M. Kohne, “Impacts of juniper vegeta-
capillary-based governing flow equation and related numerical forward/ tion and karst geology on subsurface flow processes in the Edwards Plateau,
inverse models may be best suited for local flow processes, the physical basis Texas,” Vadose Zone Journal, 5: 1076–1085, 2006a.
of their adoption at the larger scale is debated. Alternative formulations Das Gupta, S., B. P. Mohanty, and J. M. Kohne, “Soil hydraulic conductivi-
based on thermodynamic principles, hydrologic functionality rules, or scale- ties and their spatial and temporal variations in a vertisol,” Soil Science Society
appropriate water balance models across different irregular hydrologic units of America Journal, 70: 1872–1881, 2006b.
(e.g., catchments, watersheds, and river basins) may prove to be more effec- Dickinson, R. E., A. Henderson-Sellers, and P. J. Kennedy, “Biosphere
tive for describing large scale soil hydrology. With the advent of new ground- atmosphere transfer scheme (BATS) version 1e as coupled to the NCAR
based sensor networks and remote sensing (satellite) measurement techniques Community Climate Model,” NCAR Technical Note, NCAR/TN-378+STR,
at multiple resolutions and frequency, development of novel scale-appropriate National Center for Atmospheric Research, Boulder, CO, 1993.
process modeling and data assimilation/fusion tools will be the key for Duan, Q., S. Sorooshian, and V. Gupta, “Effective and efficient global opti-
improved understanding of water, carbon, nutrient, and energy cycle in the mization for conceptual Rainfall-Runoff models,” Water Resources Research,
terrestrial system in the coming decades. 28 (4): 265–284, 1992, doi: 10.1029/91WR02985.
Dunne, S. and D. Entekhabi, “An ensemble-based reanalysis approach to
land data assimilation,” Water Resources Research, 41: W02013, 2005,
REFERENCES
doi:10.1029/2004WR003449.
Agnese, C., F. Blanda, A. Drago, M. Iovino, M. Minacapilli. G. Provenzano, Durner, W., “Hydraulic conductivity estimation for soils with heterogenous
Rallo, et al., “Assessing the agro-hydrological SWAP model to simulate soil pore structure,” Water Resources Research, 30: 211–223, 1994.
water balance in typical Mediterranean crops,” Geophysical Research Efendiev, Y. and T. Hou, Multiscale Finite Element Methods: Theory and
Abstract, 9: 08146, 2007. Applications, Springer LLC, New York, NY, 2009.
Amundson, R., “The carbon budget in soils,” Annual Review of Earth and Ek, M. and R. H. Cuenca, “Variation in soil parameters: implications for
Planetary Sciences, 29: 535–562, 2001. modeling surface fluxes and atmospheric boundary-layer development,”
Arora, B., B. P. Mohanty, and J. T. McGuire, “Inverse estimation of parameters Boundary-Layer Meteorology, 70: 369–383, 1994.
for multi-domain flow models in soil columns with different macropore densi- Ek, M., K. E. Mitchell, Y. Lin, E. Rogers, P. Grunmann, V. Koren, G.
ties,” Water Resources Research, 47: W04512, 2011, doi:10.1029/2009WR009451. Gayno, et al., “Implementation of Noah land surface model advances in the
Bresler, E. and G. Dagan, “Unsaturated flow in spatially variable fields: 2. National Centers for Environmental Prediction operational mesoscale Eta
Application of water flow models to various fields,” Water Resources Research, Model,” Journal of Geophysical Research, 108 (D22): 8851, 2003,
19: 421–428, 1983. doi:10.1029/2002JD003296.
Brevik, E. C., “Soils and climate change: gas fluxes and soil processes,” Soil Ek, M., Y. Xia, E. F. Wood, J. Sheffield, B. Cosgrove, and K. Mo, “The North
Horizons, 53 (4): 12–23, 2012, doi:10.2136/sh12-04-0012. American land data assimilation system: application to drought over
Brooks, R. H. and A. T. Corey, Hydraulic Properties of Porous Media, CONUS,” AGU, Fall Meeting, Abstracts, 1: 1015, GC31A-1015, 2011.
Hydrology Paper No. 3, Colorado State University, Fort Collins, CO, 1964. Entekhabi, D., “A simple model of the hydrologic cycle and climate:
Brutsaert, W., Evaporation into the Atmosphere, D. Reidel, Norwell, MA, 1. Model construct and sensitivity to the land surface boundary,” Advances in
1982, p. 299. Water Resources, 17: 79–91, 1994.
Bindlish, R. and A. P. Barros, “Subpixel variability of remotely sensed soil Famiglietti, J. S., J. A. Devereaux, C. A. Laymon, T. Tsegaye, P. R. Houser,
moisture: an intercomparison study of SAR and ESTAR,” IEEE Transaction on T. J. Jackson, S. T. Graham, et al., “Ground-based investigation of soil
Geoscience and Remote Sensing, 40 (2): 326–337, 2002. moisture variability within remote sensing footprints during the Southern
Caldwell, T. G., M. H. Young, E. V. McDonald, and J. Zhu, “Soil heterogene- Great Plains (1997) Hydrology Experiment,” Water Resources Research,
ity in Mojave desert shrublands: biotic and abiotic processes,” Water Resources 35: 1839–1851, 1999.
Research, 48: W09551, 2012, doi:10.1029/2012WR011963. Gaur, N. and B. P. Mohanty, “Evolution of physical controls for soil
Chang, D. H. and S. Islam, “Effects of topography, soil properties and mean moisture in humid and sub-humid watersheds,” Water Resources Research,
soil moisture on the spatial distribution of soil moisture: a stochastic analysis,” 49: 1–15, 2013, doi:10.1002/wrcr.20069.
Scale Issues in Soil Physics, edited by Pachepsky, Y., Radcliffe, D. E., and Selim, Gneiting, T., A. E. Raftery, A. H. Westveld, and T. Goldman, “Calibrated
H. M., CRC press, Boca Raton, FL, 2003, pp. 193–225. probabilistic forecasting using ensemble model output statistics and mini-
Charpentier, M. A. and P. M. Groffman, “Soil moisture variability within mum CRPS estimation,” Monthly Weather Review, 133: 1098–1118, 2005.
remote sensing pixels,” Journal of Geophysical Research, 97: 18987–18995, 1992. Gerke, H. H. and M. T. van Genuchten, “A dual-porosity model for simulat-
Clapp, R. B. and G. M. Hornberger, “Empirical equations for some ing the preferential movement of water and solutes in structured porous
hydraulic properties,” Water Resources Research, 14: 601–604, 1978. media,” Water Resources Research, 29: 305–319, 1993.
Coccia, G. and E. Todini, “Recent developments in predictive uncertainty Gerke, H. H. and M. T. van Genuchten, “Macroscopic representation of
assessment based on the model conditional processor approach,” Hydrology structural geometry for simulating water and solute mass transfer in dual-
and Earth System Sciences, 15 (10): 3253–3274, 2011. porosity media,” Advances in Water Resources, 19: 343–357, 1996.
Crow, W. T. and E. F. Wood, “Multiscale dynamics of soil moisture vari- Hawley, M. E., T. J. Jackson, and R. H. McCuen, “Surface soil moisture varia-
ability observed during SGP’97,” Geophysical Research Letters, 26 (23): tion on small agricultural watersheds,” Journal of Hydrology, 62: 179–200, 1983.
3485–3488, 1999. Hoetting, J. A., D. Madigan, A. E. Raftery, and C. T. Volinsky, “Bayesian
Crow, W. T., A. Berg, M. H. Cosh, A. Low, B. P. Mohanty, R. Panciera, P. De modeling averaging: A tutorial,” Statistical Science, 14 (4): 382–417, 1999.
Rosnay, et al., “Upscaling sparse ground-based soil moisture observations for Holland, J. H., Adaptation in Natural and Artificial Systems, University of
the validation of satellite surface soil moisture products,” Reviews of Michigan Press, Ann Arbor, MI, 1975, p. 183.
Geophysics, 50: RG2002, 2012, doi:10.1029/2011RG000372. Hsu, K. L., H. Moradkhani, and S. Sorooshian, “A sequential bayesian
Das, N. N. and B. P. Mohanty, “Root zone soil moisture assessment using remote approach for hydrologic model selection and prediction,” Water Resources
sensing and vadose zone modeling,” Vadose Zone Journal, 5: 296–307, 2006. Research, 45: W00B12, 2009, doi:10.1029/2008WR006824.
Das, N. N., B. P. Mohanty, and E. G. Njoku, “A Markov Chain Monte Carlo Ines, A. V. M. and B. P. Mohanty, “Near-surface soil moisture assimilation
algorithm for upscaled soil-vegetation-atmosphere-transfer modeling to to quantify effective soil hydraulic properties using genetic algorithm.
evaluate satellite-based soil moisture measurements,” Water Resources 1. Conceptual modeling,” Water Resources Research, 44: W06422, 2008a,
Research, 44: W05416, 2008, doi:10.1029/2008WR006472. doi:10.1029/2007WR005990.
Das, N. N., B. P. Mohanty, M. H. Cosh, and T. J. Jackson, “Modeling and Ines, A. V. M. and B. P. Mohanty, “Parameter conditioning with a noisy
assimilation of root zone soil moisture using remote sensing observations in montecarlo genetic algorithm for estimating effective soil hydraulic proper-
Walnut Gulch watershed during SMEX04,” Remote Sensing of Environment, ties from space,” Water Resources Research, 44: W08441, 2008b,
112 (2): 415–429, 2008. doi:10.1029/2008WR006125.
Das, N. N. and B. P. Mohanty, “Temporal dynamics of PSR-based soil Ines, A. V. M. and B. P. Mohanty, “Near-surface soil moisture assimilation
moisture across spatial scales in an agricultural landscape during SMEX02: to quantify effective soil hydraulic properties using genetic algorithm. 2. With
a wavelet approach,” Remote Sensing of Environment, 112 (2): 522–534, air-borne remote sensing during SGP97 and SMEX02,” Water Resources
2008. Research, 44: W06422, 2009, doi:10.1029/2007WR007022.
Das, N. N., B. P. Mohanty, and E. G. Njoku, “Profile soil moisture across Ines, A. V. M., B. P. Mohanty, and Y. Shin, “An unmixing algorithm for
spatial scales under different hydroclimatic conditions,” Soil Science, 175 (7): remotely sensed soil moisture,” Water Resources Research, 49: 408–425, 2013,
315–319, 2010. doi: 10.1029/2012WR012379.
46_Singh_ch46_p46.1-46.14.indd 11 8/22/16 2:15 PM

Jacobs, J., B. P. Mohanty, E-C. Hsu, and D. Miller, “Field scale variability Krzysztofowicz, R. and K. S. Kelly, “Hydrologic uncertainty processor for
and similarity of soil moisture during SMEX02,” Remote Sensing and probabilistic river stage forecasting,” Water Resources Research, 36 (11):
Environment, 92: 436–446, 2004. 3265–3277, 2000.
Jana, R. B., Scaling Characteristics of Soil Hydraulic Parameters at Varying Kumar, P., “A multiple scale state-space model for characterizing subgrid
Spatial Resolutions, PhD dissertation, Department of Biology and Agricultural scale variability of near surface soil moisture,” IEEE Transactions on Geoscience
Engineering, Texas A&M University, College Station, TX, 2010. and Remote Sensing, 37 (1): 182–197, 1999.
Jana, R., B. P. Mohanty, and E. Springer, “Multiscale Bayesian neural net- Lal, R., “Soil carbon sequestration impacts on global climate change and
works for soil water content estimation,” Water Resources Research, W08408, food security,” Science, 304: 1623–1627, 2004.
2008, doi:10.1029/2008WR006879. Leij, F. J., W. B. Russell, and S. M. Lesch, “Closed-form expressions for water
Jana, R. and B. P. Mohanty, “On topographic controls of soil hydraulic retention and conductivity data,” Groundwater, 35 (5): 848–858, 1997.
parameter scaling at hillslope scales,” Water Resources Research, 48: W02518, Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges, “A simple hydro-
2012a, doi:10.1029/2011WR011204. logically based model of land surface water, energy fluxes for general circula-
Jana, R. and B. P. Mohanty, “A topography-based scaling algorithm for soil tion models,” Journal of Geophysical Research, 99 (D7): 14415–14428, 1994,
hydraulic parameters at hillslope scales: field testing,” Water Resources doi:10.1029/94JD00483.
Research, 48: W02519, 2012b, doi:10.1029/2011WR011205. Liu, Q. and Dickinson, R. E., “Use of a two-mode soil pore size distribution
Jana, R. and B. P. Mohanty, “A comparative study of multiple approaches to to estimate soil water transport in a land surface model,” Geophysical Research
soil hydraulic parameter scaling applied at the hillslope scale,” Water Letter, 30 (6): 1331, 2003, doi:10.1029/2002GL016562.
Resources Research, 48: W02520, 2012c, doi:10.1029/2010WR010185. Liu, Y. Y., M. F. McCabe, J. P. Evans, A. I. J. M. van Dijk, R. A. M. de Jeu,
Janzen, H. H., P. E. Fixen, A. J. Franzluebbers, J. Hattey, R. C. Izaurralde, and H. Su, “Comparison of soil moisture in GLDAS model simulations and
Q. M. Ketterings, D. A. Lobb, et al., “Global prospects rooted in soil science,” satellite observations over the Murray Darling Basin,” Proceedings of the 18th
Soil Science Society of America Journal, 75: 1–8, 2010. World IMACS Congress and MODSIM09 International Congress on Modelling
Jarvis, N. J., “A review of non-equilibrium water flow and solute transport and Simulation, edited by Anderssen, R. S., Braddock, R. D., & Newham,
in soil macropores: principles, controlling factors and consequences for water L. T. H. Modelling and Simulation Society of Australia and New Zealand and
quality,” European Journal of Soil Science, 58: 523–546, 2007, International Association for Mathematics and Computers in Simulation,
doi:10.1111/j.1365-2389.2007.00915.x. Cairns, 2009, pp. 2798–2804.
Jawson, S. D. and J. D. Niemann, “Spatial patterns from EOF analysis of soil Loague, K., “Soil water content at R-5: Part-1. Spatial and temporal vari-
moisture at a large scale and their dependence on soil, land-use, and topo- ability,” Journal of Hydrology, 139: 233–251, 1992.
graphic properties,” Advances in Water Resources, 30 (3): 366–381, 2007. Loescher, H., J. Jacobs, W. Krajewski, R. Mason, K. McGuire, B. P. Mohanty,
Joshi, C. and B. P. Mohanty, “Physical controls of near‐surface soil G. Poulos, et al., “Enhancing water cycle measurements for future hydrologic
moisture across varying spatial scales in an agricultural landscape
research,” Bulletin of the American Meteorological Society, 2007, doi:10.1175/
during SMEX02,” Water Resources Research, 46: W12503, 2010, BAMS-88-5-669.
doi:10.1029/2010WR009152. Maxwell, R. M., “A terrain-following grid transform and preconditioner
Joshi, C., B. P. Mohanty, J. Jacobs, and A. V. M. Ines, “Spatio-temporal for parallel, large-scale, integrated hydrologic model,” Advances in Water
analyses of soil moisture from point to footprint scale in two different hydro- Resources, 53: 109–117, 2013.
climatic regions,” Water Resources Research, 47, 2011, W01508, Mohanty, B. P., “Soil hydraulic property estimation using remote sensing:
doi:10.1029/2010WR009002. a review,” Vadose Zone Journal, 12 (4), 2013, doi: 10.2136/vzj2013.06.0100.
Kennedy J. and R. C. Eberhart, Swarm Intelligence, Morgan Kaufmann, San Mohanty, B. P., M. D. Ankeny, R. Horton, and R. S. Kanwar, “Spatial analy-
Mateo, CA, 2001. sis of hydraulic conductivity measured using disc infiltrometer,” Water
Kim, G. and A. P. Barros, “Space-time characterization of soil moisture Resources Research, 30 (9): 2489–2498, 1994.
from passive microwave remotely sensed imagery and ancillary data,” Remote Mohanty, B. P., R. S. Bowman, J. M. H. Hendrickx, J. Simunek, and M. Th.
Sensing of Environment, 81: 393–403, 2002. van Genuchten, “New piecewise-continuous hydraulic functions for model-
Kim, C. P. and J. N. M. Stricker, “Influence of variable soil hydraulic proper- ing preferential flow in an intermittent-flood-irrigated field,” Water Resources
ties and rainfall intensity on the water budget,” Water Resources Research, Research, 33 (9): 2049–2063, 1997.
32: 1699–1712, 1996. Mohanty, B. P., R. S. Bowman, J. M. H. Hendrickx, J. Simunek, and M. Th.
Kim, C.P., J. N. M. Stricker, and R. A. Feddes, “Impact of soil heterogeneity van Genuchten, “Preferential transport of nitrate to a tile drain in an intermit-
on the water budget of the unsaturated zone,” Water Resources Research, 33: tent-flood-irrigated field: model development and experimental evaluation,”
991–999, 1997. Water Resources Research, 34 (5): 1061–1076, 1998.
Kim, J., B. P. Mohanty, and Y. Shin, “Effective soil moisture estimates and Mohanty, B. P., T. H. Skaggs, and J. S. Famiglietti, “Analysis and mapping of
its uncertainty using multi-model simulation based on Bayesian model field-scale soil moisture variability using high-resolution ground based data
averaging,” Journal of Geophysical Research-Atmosphere, 120, 2015, during the Southern Great Plains 1997 (SGP97) hydrology experiment,”
doi:10.1002/2014JD0022905. Water Resources Research, 36: 1023–1032, 2000a.
Kirchener, J. W., “Getting the right answers for the right reasons: linking Mohanty, B. P., J. S. Famiglietti, and T. H. Skaggs, “Evolution of soil mois-
measurements, analysis, and models to advance the science of hydrology,” ture spatial structure in a mixed-vegetation pixel during the SGP97 hydrology
Water Resources Research, 42: W03S04, 2006, doi:10.1029/2005WR004362. experiment,” Water Resources Research, 36 (12): 3675–3686, 2000b.
Kite, G, “Modeling the mekong: hydrological simulation for environmental Mohanty, B. P. and T. H. Skaggs, “Spatio-Temporal evolution and time
impact studies,” Journal of Hydrology, 253: 1–13, 2001. stable characteristics of soil moisture within remote sensing footprints with
Kohne, J. M. and B. P. Mohanty, “Water flow processes in a soil column with varying soils, slopes, and vegetation,” Advances in Water Resources, 24:
a cylindrical macropore: experiment and hierarchical modeling,” Water 1051–1067, 2001.
Resources Research, 41, W03010, 2005, doi:10.1029/2004WR003303. Mohanty, B. P., P. J. Shouse, D. A. Miller, and M. Th. van Genuchten, “Soil
Köhne, J. M., B. P. Mohanty, J. Šimůnek, and H. H. Gerke, “Numerical property database: Southern Great Plains 1997 hydrology experiment,” Water
evaluation of a second-order water transfer term for variably-saturated dual Resources Research, 38 (5), 2002, doi:10.1029/2000WR00076.
permeability model,” Water Resources Research, 40: W07409, 2004, Mohanty, B. P. and J. Zhu, “Effective hydraulic parameters in horizontally
doi:10.1029/2004WR003285. and vertically heterogeneous soils for steady-state land–atmosphere interac-
Köhne, J. M., S. Köhne, and J. Šimůnek, “Multi-process herbicide transport tion,” Journal of Hydrometeorology. 8 (4): 715–729, 2004.
in structured soil columns: experiments and model analysis,” Journal of Moore, I. D., G. J. Burch, D. H. Mackenzie, “Topographic effects on the
Contaminant Hydrology, 85: 1–32, 2006, doi:10.1016/j.jconhyd.2006.01.001. distribution of surface water and the location of ephemeral gullies,”
Köhne, J. M., S. Köhne, and J. Šimůnek, “A review of model applications for Transactions of the ASAE, 31: 1098–1107, 1988.
structured soils: water flow and solute transport,” Journal of Contaminant Mualem, Y., “A new model for predicting the hydraulic conductivity of
Hydrology, 104: 4–35, 2009, doi:10.1016/j.jconhyd.2008.10.002. unsaturated soils,” Water Resources Research, 12: 513–522, 1976.
Koster, R. and M. Suarez, “Energy and water balance calculations in the National Research Council, Opportunities in Hydrologic Sciences, National
Mosaic LSM,” NASA Technical Memorandum, 9: 104606, 1996. Academy Press, Washington, D.C., 1991; p. 348.
Kosugi, K., “Log-Normal distribution model for unsaturated soil hydraulic Niu, G., Y. Yang, Z. L. Dickinson, R. E. Gulden, and H. Su, “Development
properties,” Water Resources Research, 32: 2697–2703, 1996. of a simple groundwater model for use in climate models and evaluation with
Krzysztofowicz, R., “Bayesian theory of probabilistic forecasting via deter- gravity recovery and climate experiment data,” Journal of Geophysical
ministic hydrologic model,” Water Resources Research, 35: 2739–2750, 1999. Research, 112: D07103, 2007, doi:10.1029/2006JD007522.
46_Singh_ch46_p46.1-46.14.indd 12 8/22/16 2:15 PM

REFERENCES 13
Oleson, K. W., D. M. Lawrence, G. B. Bonan, M. G. Flanner, E. Kluzek, and multiple solutes in variably-saturated media,” User Manual, Version 1.0,
P. J. Lawrence, S. Levis, et al., “Technical description of version 4.0 of the PC Progress, Prague, Czech Republic, 2012, p. 303.
Community Land Model (CLM),” NCAR Technical Notes (NCAR/ Sivapalan, M., K. Beven, and E. F. Wood, “On hydrologic similarity 2.
TN-478+STR), 2010, p. 257. A scaled model of storm runoff production,” Water Resources Research, 23:
Palm, C., P. Sanchez, S. Ahamed, and A. Awiti, “Soils: a contemporary 2266–2278, 1987.
perspective,” Annual Review of Environment and Resources, 32: 99–129, 2007. Stewart, J. B., E. T. Engman, R. A. Feddes, and Y. Kerr, “Scaling up in
Peters-Lidard, C. D., D. M. Mocko, M. Garcia, J. A. Santanello, Jr., M. A. hydrology using remote sensing,” Wiley, Chischester, UK, 1996, p. 255.
Tischler, M.S. Moran, and Y. Wu, “Role of precipitation uncertainty in the Todini, E., “A model conditional processor to assess predictive uncertainty
estimation of hydrologic soil properties using remotely sensed soil moisture in flood forecasting,” International Journal of River Basin Management, 6 (2):
in a semi-arid environment,” Water Resources Research, 44, 2008, 123–137, 2008.
doi:10.1029/2007WR005884. Van Dam, J. C., J. Huygen, J. G. Wesseling, R. A. Feddes, P. Kabat, P. E. V.
Philip, J. R. and D. de Vries, “Moisture movement in porous materials van Waslum, P. Groenendjik, et al., “Theory of SWAP version 2.0: simulation
under temperature gradients,” Transactions of the American Geophysical of water flow and plant growth in the soil-water-atmosphere-plant environ-
Union, 38: 222–232, 1957. ment,” Technical Document, DLO Winand Staring Center, Wageningen
Pollacco, J. A. and B. P. Mohanty, “Uncertainties of water fluxes in SVAT Agricultural University, Wageningen, The Netherlands, 1997, p. 45.
models: inverting surface soil moisture and evapotranspiration retrieved from Van Genuchten, M. Th., “A closed-form equation for predicting the
remote sensing,” Vadose Zone Journal, 11 (3), 2012, doi:10.2136/vzj2011.0167. hydraulic conductivity of unsaturated soils,” Soil Science Society of America
Pollacco, J. A. P., B. P. Mohanty, and A. Efstratiadis, “The weighted objective Journal, 44: 892–898, 1980.
function selector algorithm and its application to invert surface soil moisture Vinnikov, K .Y., A. Robock, N. Speranskaya, and C. A. Schlosser, “Scales of
and evapotranspiration in SVAT models,” Water Resources Research, 49, 6959- temporal and spatial variability of midlatitude soil moisture,” Journal of
6978, 2013, doi: 10.1002/wrcr.20554. Geophysical Research, 101: 7163–7174, 1996.
Pruess, K., C. Oldenburg, and G. Moridis, TOUGH2 User’s Guide, Version 2.1, Vrugt, J. A., C. J. F. ter Braak, C. G. H. Diks, B. A. Robinson, J. M. Hyman,
Report LBNL-43134, Lawrence Berkeley National Laboratory, Berkeley, and D. Higdon, “Accelerating Markov chain Monte Carlo simulation by
CA, 2012. differential evolution with self-adaptive randomized subspace sampling,”

Reed, P., B. Minsker, and D. E. Goldberg, “Designing a competent simple International Journal of Nonlinear Sciences and Numerical Simulation, 10 (3):
genetic algorithm for search and optimization,” Water Resources Research, 271–288, 2009.
36 (12): 3757–3761, 2000, doi:10.1029/2000WR900231. Vogel, T. and M. Císlerová, “On the reliability of unsaturated hydraulic
Reynolds, S. G., “The gravimetric method of soil moisture determination, conductivity calculated from the moisture retention curve,” Transport in
Part III: an examination of factors influencing soil moisture variability,” Porous Media, 3: 1–15, 1988.
Journal of Hydrology, 11: 288–300, 1970. Western, A., R. B. Grayson, G. Bloschl, G. R. Willgoose, and T. A.
Rodriguez-Iturbe, I., P. D’Odorico, A. Porporato, and L. Ridolfi, “On the McMahon, “Observed spatial organization of soil moisture and its relation to
spatial and temporal links between vegetation, climate, and soil moisture,” terrain indices,” Water Resources Research, 35: 797–810, 1999.
Water Resources Research, 25 (12): 3709–3722, 1999. Western, A., R. B. Grayson, and G. Bloschl, “Scaling of soil moisture:
Rosero, E., Z-L. Yang, T. Wagener, L. E. Gulden, S. Yatheendradas, and a hydrologic perspective,” Annual Review of Earth and Planetary Sciences,
G.-Y. Niu, “Quantifying parameter sensitivity, interaction, and transferability 30: 149–180, 2002.
in hydrologically enhanced versions of the Noah land surface model over Western, A., R. B. Grayson, G. Bloschl, and D. J. Wilson, “Spatial variability
transition zones during the warm season,” Journal of Geophysical Research, of soil moisture and its implications for scaling,” Scale Issues in Soil Physics,
115: D03106, 2010, doi:10.1029/2009JD012035. edited by Pachepsky, Y., Radcliffe, D. E., and Selim, H. M. CRC Press, Boca
Russo, D., “Determining soil hydraulic properties by parameter estimation: Raton, FL, 2003, pp. 119–142.
on the selection of a model for the hydraulic properties,” Water Resources White, M. D., D. H. Bacon, B. P. McGrail, D. J. Watson, S. K. White, and
Research, 24: 453–459, 1988. Z. F. Zhang, “STOMP: Subsurface transport over multiple phases: STOMP-CO2
Sahoo, A. K., P. A. Dirmeyer, P. R. Houser, and M. Kafatos, “A study of land and STOMP-Coe Guide,” Version 1.0 PNNL-21268, Pacific Northwest
surface processes using land surface models over the Little River Experimental National Laboratory, Richland, WA, 2012.
Watershed, Georgia,” Journal of Geophysical Research, 113: D20121, 2008, Ying, M., S. Feng, Z. Huo, and X. Song, “Application of the SWAP model to
doi:10.1029/2007JD009671. simulate the field water cycle under deficit irrigation in Beijing, China,”
Salvucci, G. D., “Estimating the moisture dependence of root zone water Mathematical and Computer Modelling, 54: 1044–1052, 2011.
loss using conditionally averaged precipitation,” Water Resources Research, Zeng, X. B. and M. Decker, “Improving the numerical solution of soil
37 (5): 1357–1365, 2001. moisture-based Richards equation for land models with a deep or shallow
Sharma, M. L. and R. J. Luxmoore, “Soil spatial variability and its consequences water table,” Journal of Hydrometeorology, 10: 308–319, 2009.
on simulated water balance,” Water Resources Research, 15: 1567–1573, 1979. Zhang, X., R. Srinivasan, K. Zhao, and M. V. Liew, “Evaluation of global
Sharma, S. K., B. P. Mohanty, J. Zhu, “Including topography and vegetation optimization algorithms for parameter calibration of a computationally
attributes for developing pedo-transfer functions,” Soil Science Society of intensive hydrologic model,” Hydrological Processes, 2008, doi:10.1002/
America Journal, 70: 1430–1440, 2006. hyp.7152.
Shin, Y. and B. P. Mohanty, and A. V. M. Ines, “Soil hydraulic properties in one- Zhang, X., R. Srinivasan, and D. Bosch, “Calibration and uncertainty analy-
dimensional layered soil profile using layer-specific soil moisture assimilation sis of the SWAT model using Genetic Algorithms and Bayesian Model
scheme,” Water Resources Research, 48: W06529, 2012, doi:10.1029/2010WR009581. Averaging,” Journal of Hydrology, 374: 307–317, 2009.
Shin, Y., B. P. Mohanty, and A. V. M. Ines, “Estimating effective soil hydrau- Zhu, J., B. P. Mohanty, “Upscaling of soil hydraulic properties under steady
lic properties using spatially distributed soil moisture and evapotranspira- state evaporation and infiltration,” Water Resources Research, 38, 2002a,
tion,” Vadose Zone Journal, 12 (3), 2013a, doi:10.2136/vzj2012.0094. doi:10.1029/2001WR000704.
Shin, Y. and B. P. Mohanty, “Development of a deterministic downscaling Zhu, J., B. P. Mohanty, “Spatial averaging of van Genuchten hydraulic
algorithm for remote sensing soil moisture footprint using soil and vegetation parameters for steady state flow in heterogeneous soils,” Vadose Zone Journal,
classifications,” Water Resources Research, 49, 2013b, doi:10.1002/wrcr.20495. 1: 261–271, 2000b.
Šimůnek, J., O. Wendroth, N. Wypler, and M. Th. van Genuchten, “Non- Zhu, J., B. P. Mohanty, “Effective hydraulic parameters for steady state verti-
equilibrium water flow characterized by means of upward infiltration experi- cal flow in heterogeneous soils,” Water Resources Research, 39 (8), 2003,
ments,” European Journal of Soil Science, 52 (1): 13–24, 2001. doi:10.1029/2002WR001831.
Šimůnek, J., N. J. Jarvis, M. T. van Genuchten, and A. Gärdenäs, “Review Zhu, J., B. P. Mohanty, “Soil hydraulic parameter upscaling for steady
and comparison of models for describing non-equilibrium and preferential state flow with root water uptake,” Vadose Zone Journal, 3: 1464–1470, 2004.
flow and transport in the vadose zone,” Journal of Hydrology, 272: 14–35, Zhu, J., B. P. Mohanty, A. W. Warrick, and M. Th. van Genuchten,
2003, doi:10.1016/S0022-1694(02)00252-4. “Correspondence and upscaling of hydraulic functions for steady-state flow
Šimůnek, J., M. Th. van Genuchten, and M. Šejna, “The HYDRUS software in heterogeneous soils,” Vadose Zone Journal, 3: 527–533, 2004.
package for simulating two- and three-dimensional movement of water, heat, Zhu, J. and B. P. Mohanty, “Effective scaling factor for transient infiltration
and multiple solutes in variably-saturated media,” User Manual, Version 1.0, in heterogeneous soils,” Journal of Hydrology. 319: 96–108, 2006.
PC Progress, Prague, Czech Republic, 2006, p. 241. Zhu, J., B. P. Mohanty, and N. N. Das, “On the effective averaging schemes
Šimůnek, J., M. Th. van Genuchten, and M. Šejna, “The HYDRUS software of hydraulic properties at the landscape scale,” Vadose Zone Journal, 5:
package for simulating two- and three-dimensional movement of water, heat, 308–318, 2006.
46_Singh_ch46_p46.1-46.14.indd 13 8/22/16 2:15 PM

Chapter 47
Hydrogeologic Characterization
BY
KAVEH ZAMANI, TIMOTHY R. GINN, AND MOHAMED K. NASSAR
ABSTRACT substances being monitored. For more information on monitoring well

design, drilling, installation, and maintenance, see ASTM D 5876 and 5978
The subsurface is the unseen part of the hydrologic cycle that is critical for
standards (ASTM, 2011, 2012a; Aller et al., 1991).
municipal, industrial, agricultural, and/or ecosystem water use especially in
times of drought. The occurrence and mobility of subsurface water depends
on a host of hydrogeological properties that are typically very difficult to 47.3 INVESTIGATION OF BOREHOLE DRILLING
measure and sometimes not even well-defined. Furthermore, these properties
vary with location, often on multiple scales. In this chapter, we summarize Site investigation with boreholes provides one-dimensional (1D) information
techniques for characterization of the primary properties controlling on geologic structure and material properties via collection of core samples
groundwater flow in the saturated subsurface. Our focus is on only the most (e.g., ASTM, 2012a; Hunkeler, 2010). Rotary boring is a method of advancing
basic methods, with measurement scale defined by the sampling technique, a borehole in both rock and sedimentary formations up to depths of 350 m,
in order to provide a simple background for field work approaches to quanti- and generates core samples of diameters 10–25 cm. The drill shaft and bit are
fying the properties governing large-scale groundwater flow. advanced within a metal core barrel that is advanced in sections to provide a
temporary or permanent borehold casing. In direct rotary drilling, drilling
mud or fluid is pumped down the drilling core to the bits and brings up the
47.1 INTRODUCTION cuttings to the surface. In reverse-circulation rotary drilling, used for larger
Understanding groundwater for engineering, agricultural, or environmental boreholes, the drilling fluid is pumped down the annulus between the borehole
activities require sufficient knowledge of subsurface geological and hydro- and the casing and returns through the drill barrel. This provides the advan-
logical characteristics. Difficulties arise quantifying hydrogeological pro- tage of positive pressure to maintain the stability of the borehole during drill-
cesses mainly due to incomplete characterization of subsurface properties at ing. The method is called air-rotary drilling if instead of water, compressed air
the required spatial scale. This scale ranges from the order of tens of meters is being used. Subsurface materials returning to the surface with the drilling
for geothermal applications to kilometers for contamination problems to fluid provide qualitative if disturbed sample that can be used to characterize
hundreds of kilometers for regional groundwater sustainability studies. strata at depth. To acquire core samples, the drilling rod is raised and the drill-
Historically, many subsurface characterization techniques in hydrogeology ing shaft and bit is replaced with a core sampler. Some advanced bits are also
were developed originally for petroleum and mining industries and subse- able to obtaining a core sample when they are used for drilling. Disadvantages
quently adopted for groundwater engineering. Detailed references from these of this method compared to auger drilling and direct push techniques include
and related fields about hydrogeologic characterization exist, for example, potential contamination by drilling mud/fluid that may impact groundwater
Charbeneau (2006), EPA (2015), Fetter (1999), Mitchell and Soga (2005), chemistry, and the relatively higher cost and infrastructure required.
Rubin and Hubbard (2005), and Todd and Mays (2005). The purpose of this Auger boring involves a spiral drill (augur) and is commonly used in rela-
chapter is to give a summary overview of some of the practicable techniques tively shallow and unconsolidated formations to depths up to 100 m (e.g.,
to characterize geological facies, structure, and permeability as well as Hackett, 1987). A conventional auger would bring materials to the surface
groundwater occurrence hydraulic head, and flow velocity. Due to scope generally disturbed. The hollow-stem auger involves placement of the spiral
restrictions many related techniques, such as ground-penetrating radar, self- drill inside a metal barrel, and core samples can be obtained as with rotary
potential, remote sensing and GRACE data, inclinometry, environmental rigs by retrieving the drill and temporarily replacing it with a single or double
tracers, and seismic techniques are not covered here, and the treatment is sampling core. Advantages of this method include no involvement of drilling
restricted to saturated subsurface materials. mud/fluid, lighter infrastructure, and faster sampling at lower costs than
rotary methods. Direct-push sampling is an even lighter-footprint approach
that involves driving a metal barrel into the subsurface often using high-
47.2 BOREHOLE SAMPLES AND GROUNDWATER
MONITORING WELLS frequency hammering or vibration. Core sample retrieval may be done conti
nuously using this method, by driving core barrel segments that are retrieved,
Drilled wells are main source of subsurface hydrogeological information. In or at intervals wherein a driven penetration tip is replaced at intervals with a
general, location selection, design, well drilling, and sampling phases of wells segment of sample core barrel that is retrieved after it is driven till it is filled.
have to be conducted based on the purpose of the well, geologic structure of Advantages of this method are the light equipment and rapid advancement,
the site, expected groundwater occurrence and contamination, and location while disadvantages include the small diameter core samples obtained and the
of the site. Drilling techniques used for well construction include hollow-stem requirement for penetrable subsurface materials.
augers, solid-stem augers, water/mud rotary, pneumatic rotary, sonic (vibra- Percussion boring (or cable-tool drilling) is a method in which soil and
tor), direct push, and casing or cable. While hydraulic/pneumatic rotary and rock formations are broken by repeated blows of heavy chisel or bit hanging
cable tool are the more common approaches for well construction, the inves- from a cable or drill rod. A steel casing is driven into the borehole behind the
tigation borehole drilling methods useful for obtaining core samples are bit utile the rock formation is reached. When the bottom of the hole is filled
hollow-stem auguring, rotary drilling, and direct push sampling. At contami- with broken rocks and sediments, water is added to the hole, if not already
nated sites, well casing and screens must be chemically compatible with the below the water table and the slurry of pulverized material is removed with
47-1
47_Singh_ch47_p47.1-47.4.indd 1 8/22/16 2:16 PM

47-2 Hydrogeologic Characterization
bottom-loading bailer. An advantage of percussion drilling is unlike rotary electrodes emplaced at the surface and monitoring electrodes measuring the
drilling, no drilling fluid/mud circulated through the well, hence, induced difference in electrical potential (i.e., voltage), nearby and usually between the
contamination issue does not exist. The other advantage is that the percussion inducing electrodes. The current used is direct current, commutated direct
drilling is suitable for remote location due to low fuel consumption and small current (i.e., a square-wave alternating current), or alternating current with
needs for water. This method is generally slower than other alternatives drilling low frequency (about 20 Hz). The resulting potential (volt) distribution can
techniques; however, the drilling equipment are less expensive than rotary be related directly to electrical resistivity in several idealized cases, such as the
methods (ASTM, 2012a). case of perfectly layered strata or the case of homogeneous materials sepa-
rated by a vertical dike (e.g., EPA, 2015). In the 1D case, the electrical resisti
vity ρ of soil material is expressed as:
47.4 CONE PENETROMETRY, PERMEAMETRY, AND
ELECTRICAL CONDUCTIVITY LOGGING A ∆V
ρ= (47.1)
The static penetration or cone penetration test (CPT) is a method to deter- L I
mine geotechnical properties of relatively shallow subsurface materials and to
delineate stratigraphy of fine sand, clays and organic soils in particular (e.g., where, A is the cross-section area of current path, L is the length of flow path,
Rogers, 2006). The standard test apparatus consists of a still cone with an apex ∆V is voltage, that is, change in electrical potential in volts, and I is the electri-
angle of 55° to 60° and cross-section area of 10 cm2 (other variations exist) that cal current. Electrical resistivity is measured in the units of Ω. m or Ω. cm. In
is driven at a speed of 1–2 cm/s, while the tip resistance qc and friction fc are case measurement is conducted over a real heterogeneous earth, the subsur-
recorded as the instrument moves down at intervals, which can be as small as face profile may consist of various layers with different resistivity and the
5 mm. Generally, two types of friction-cone penetometers are used to measure symbol ρ is replaced with apparent resistivity ρa. The electrical resistivity of
qc and fc : mechanical and electrical, the latter provides more continues profile various soils and rocks depends largely on the pore water content; however,
compare to the former (cf. below, e.g., Harrington and Hendry, 2006). Several ion concentration is an important factor too. The general electrical resistivity
empirical curves that are useful in estimating the properties of subsurface have ranges from less than 100 for clays to 200–3000 for sands, to 1500–2500 for
been suggested for the point resistance qc and the friction ratio Fr, defined as fractured rock, to 3000–30,000 for coarse gravels (e.g., Mitchell and Soga,
the ratio of frictional resistance over cone resistance (e.g., Lapham et al., 1997; 2005). There are several common procedures for measuring electrical resis-
ASTM, 2012b). Some electrical CPT has passive pressure transducer to record tivity of a soil profile with electrodes that are driven into the ground. Some of
water pressure that is termed piezocone penetration test or “CPTU.” Advantages those include: Wenner array, Schlumberger array, dipole–dipole, Lee, half-
of this technology are low cost and rapid characterization of strata and resistive Schlumberger, polar dipole, bipole–dipole, and gradient arrays. Here, we
force of geomaterials; disadvantages include indirect relation of the measured briefly describe the first three methods which are widely utilized, relying
quantity to hydrogeological properties and penetrometer blocking by strongly heavily on EPA (2015). The Wenner array method consists of four in-line,
consolidated or hard-rock materials. For more information on conduction and equally spaced electrodes. The outer two electrodes are typically the current
interpretation of CPT for site investigation, see ASTM (1995). source and the inner two electrodes are the receiver electrodes (Loke et al.,
Butler and co-workers have extended the passive measurement of hydrau- 1996; ASTM, 2010). In the Schlumberger array, the potential electrodes are
lic pressure using direct-push equipment classically associated with cone installed at the center of the electrode array with a small separation relative to
penetrometry as done in the CPTU to active permeametry by inducing a the spacing between the current-inducing electrodes. The third common
hydraulic pressure at the porous cone tip and measuring the resulting pres- resistivity method involves the dipole–dipole array, that involves dipoles
sure at different distances along the pushed shaft (Butler et al., 2002, 2007; (two closely spaced electrodes), to measure the changes of electrical proper-
Dietrich et al., 2008; Zshornack et al., 2013). This technique, termed DPP for ties with depth. In depth discussion of implementations, limitations, and
direct push permeametry, measures the resulting pressure at can yield data on interpretation of electrical resistivity methods can be found in ASTM (2010)
vertical variations in hydraulic conductivity with a resolution scale of 0.4 m and in EPA (2015).
vertical separation for conductivity values associated with aquifer materials.
The advantages of this technique are the low cost and minimal disturbance of
the subsurface with high-resolution data. Disadvantages include penetrometer/ 47.6 HYDRAULIC METHODS FOR IN SITU CONDUCTIVITY
permeametry refusal in tight materials, errors induced in the presence of very MEASUREMENT
thin (less than 0.5 m) layering structure and errors associated with bypass In situ hydraulic conductivity is classically viewed as the primary control on
flow-up along the penetrometer shaft; however, these latter errors can be groundwater flow in response to gradients in hydraulic head, and essentially
controlled by careful push strategies that take into account the structural all groundwater textbooks include chapters on its characterization in place by
properties expected to be encountered in the materials (e.g., Zshornack et al., perturbation of the hydraulic head in one borehold and its observation in the
2013). Chapuis and Chenai (2010) provide a list of improvements derived same or adjacent boreholes screened in the same aquifer material where the
from a geotechnical context for the hydrogeologic extensions of direct-push perturbation takes place, when identifiable (e.g., Fetter, 2000; Todd and Mays,
technology including using a cone tip that is the same or smaller diameter 2005; Charbeneau, 2006). These techniques can be categorized into slug tests
than the shaft in order to minimize bypass flow, using shape factors in order to and pumping tests. A slug test involves the instantaneous change in the water
more accurately determine the local conductivity values, and allowing suffi- level in a borehole followed by monitoring its return to preperturbation con-
cient time for hydraulic equilibration to be attained prior to the collection of ditions. Most commonly a cylinder (slug) is emplaced (or removed) from the
pressure data for identification of conductivity. borehole, causing a sudden increase (or decrease) in the water level in the
Direct-push electrical conductivity logging is an alternative direct push borehole. The return of water level in the borehole to initial conditions is
technology that involves driving an electrical conductivity meter at sequential monitored (water level in the borehole as a function of time, best obtained
depths to obtain a vertical record of subsurface electrical conductivity. The through a pressure transducer deep in the borehole) and analyzed to deter-
device involves a metal rod with a pair of electrodes to incur an electrical mine the local hydraulic conductivity. The local hydraulic conductivity is
current and two inner electrodes used to monitor the resulting voltage inferred by inversion of an approximation of Darcy’s law that results in the
(e.g., Harrington and Hendry, 2006; Hunkeler, 2010). The advantages of this conductivity being proportional to the slope of the natural log of hydraulic
characterization approach are that electrical conductivity reflects water con- head versus time. The advantages of this approach are that only one borehole
tent and material type, with silts and clays and high water contents associated is needed, and no power source is required; the primary disadvantage is that
with a higher electrical conductivity than sands or gravels and lower water the hydraulic property measured is particular to the local neighborhood of
contents. When correlated with direct samples of the subsurface materials, the borehole being used for the test.
this technique can be used to distinguish identifiable stratigraphic units and Pumping tests involve extension of the slug test concept to one or more
thus to infer hydrogeologic properties. Combination of direct-push Electrical observation wells and continuous pumping at a fixed rate at a control well.
Conductivity (EC) profiling with DPP or with small-scale slug tests provide Observation wells for this experiment are best as narrow-diameter wells that
combined data for the same bore that is useful for cross-validation of inter- are screened only in the aquifer being pumped, termed “piezometers,” in
preted stratigraphy (e.g., Sellwood et al., 2005). order to eliminate artifacts due to storage in the monitoring well and adjacent
flows. Because the hydraulic head change in this approach is typically mea-
sured in separate monitoring wells, the hydraulic conductivity values
47.5 ELECTRICAL RESISTIVITY SURVEY
obtained are representative of the well-to-well separation scale, significantly
Surface electrical resistivity (reciprocal of conductivity) test is a geophysical larger than that of the slug test. The analyses of the hydraulic head drop
method for hydrogeological exploration that applies the same electrostatics of (drawdown) data observed as a function of time in the monitoring well
the direct push EC probe to larger scales, with current induced between two involves different equations depending on the nature of the pumping
47_Singh_ch47_p47.1-47.4.indd 2 8/22/16 2:16 PM

references 47-3
(constant or pulsed in steps), the confined versus unconfined nature of the are known for steady-state flow systems, then only relative conductivity
aquifer studies, and the impact of hydraulic boundaries (e.g., Charbeneau, information is needed to allocate the flow to different portions of the aquifer,
2000). For instance, in the confined case with constant pumping, assuming and if that aquifer is homogeneous with a constant conductivity then only the
homogeneous and perfectly layered aquifer materials, the Theis solution is aquifer geometry is needed to complete the understanding of the flow.
used, and this can be simplified to the Jacob solution for later time parts of the Recharge-estimation methods are many, beyond the current scope, and
observed drawdown. The more complex unconfined case is fortunately solved include the broad field of environmental tracer interpretation. USGS (2014)
in Neuman (1974), and these and a suite of related solution approaches for provides a summary tabulation of methods with references, advantages, and
pumping test data interpretation is available in numerous available codes limitations. Here, we briefly summarize two basic approaches to quantifying
(e.g., www.aqtesolve.com). areal recharge resulting from a surface influx of water. The unsaturated zone
water balance method uses unsaturated zone water saturation and tension
(negative of pressure head) measurements before and after a recharge event
47.7 CHARACTERIZATION OF THE HYDRAULIC (rainfall or irrigation) to quantify the increase in water content that occurs
GRADIENT AND FLOW RATES
below the depth above which water is extracted by evapotranspiration. This
As important as the stratigraphic understanding of the aquifer-aquitard struc- depth is identified by changes in the water tension with depth, and the net
ture of the subsurface and the corresponding hydraulic conductivities is the increase in water content fellow this depth is areally averaged to determine
hydraulic gradient occurring within these formations, because the gradient net recharge. In the water table fluctuation method, the data collected is the
and the hydraulic conductivity together dictate the natural groundwater flow increase in water table elevation and the recharge is estimated as the areally
direction and magnitude. The hydraulic gradient is the slope of the hydraulic averaged change in saturated thickness multiplied by the specific yield.
head surface. In confined aquifers, the hydraulic head is the level to which Advantages of these methods are the high-resolution identification of event-
water will rise in a piezometer screened in the aquifer, and in unconfined based recharge that is a big improvement over annual water-balance esti-
aquifers the hydraulic head is generally well approximated by the water table. mates, and disadvantages are the requirement for multiple soil water data or
In the two-dimensional (2D) case with perfectly layered aquifer materials, the water table data over the extent of the area in question. This latter require-
hydraulic head can be visualized as a surface just like topographic elevation, ment is the main limiter on the utility of the method for identification of
and the gradient is the slope of this surface. A convenient means of evaluating event-based recharge on regional scales.
the slope in a given region is to emplace three piezometers in a right triangle
layout, ABC, with B being the piezometer located at the right angle corner.
47.9 SUMMARY
Then the gradient is given by the two slopes, the head difference between
piezometers at A and B, divided by the distance between the two, and the head This chapter provides only the briefest survey of selected methods associated
difference between piezometers at B and C, divided by the distance between with the range of quantities requiring characterization for understanding
those two, respectively. It must be remembered that the 2D assumption allows subsurface hydrological cycles. Continued innovation in the numerous fields
an approximation of what may be in reality be a 3D gradient. A means of providing such data, including geophysical, hydrological, electrical, is neces-
evaluating the vertical component of the gradient (that is, the third dimension sary to overcome our inability to characterize subsurface hydrogeological
when the first two are horizontal), is via a “piezometer nest.” This is a series processes. The inaccessibility of the subsurface that makes hydrogeology the
of piezometers installed in the same location, but each screened at different most challenging of the areas of environmental fluid mechanics.
depths. Vertical components of hydraulic gradients are particularly important
in determining vertical fluxes across low-conductivity formations and in
recharge areas. REFERENCES
Tracer testing provides an alternative approach to characterizing flow rates. Aller, L., R. J. Petty, J. H. Lehr, H. Sedoris, D. M. Nielsen, and J. E. Denne,
The single-well point dilution tracer test involves emplacement of a tracer Handbook of Suggested Practices for the Design and Installation of Ground-
(typically dye or salt) in a known concentration in a well followed by monitor- Water Monitoring Wells, Environmental Monitoring Systems Laboratory,
ing of the change in concentration of the tracer in the well as a function of time. Office of Research and Development, US Environmental Protection Agency,
These data can be analyzed to estimate the ambient flow rate through its flow Las Vegas, NV, 1991.
component through the well itself. This test typically takes a day and the analy- ASTM, Designation D 5778-95(2000): Standard Test Method for Performing
ses should take account of the well geometry and the artificially high hydraulic Electronic Friction Cone and Piezocone Penetration Testing of Soils, ASTM
conductivity associated with the gravel pack associated with the well installa- International, West Conshohocken, PA, 2000.
tion (e.g., Hunkeler, 2010). When more observation wells are available, this test ASTM, Designation D 6431: Standard Guide for Using the Direct Current
can be generalized to a multiwell tracer test, in which the tracer is injected in Resistivity Method for Subsurface Investigation, ASTM International, West
one well and observed at other downstream wells. The disadvantages of this Conshohocken, PA, 2010.
approach are the increased costs associated with installation of the additional ASTM, Designation D 5978: Standard Guide for Maintenance and
monitoring wells and the challenge in capturing the tracer when the flow is Rehabilitation of Groundwater Monitoring Wells, ASTM International, West
relatively fast compared to the rate of dispersion of the salt around the injection Conshohocken, PA, 2011.
well. This latter aspect can lead to the tracer plume following a narrow path ASTM, Designation D 6286: Standard Guide for Selection of Drilling
that may not be intercepted by the downstream well. Also this method requires Methods for Environmental Site Characterization, ASTM International, West
knowledge of the downstream direction a priori. Conshohocken, PA, 2012a.
Finally, vertical flows monitored within boreholes can provide useful ASTM, Designation D 5876: Standard Guide for Use of Direct Rotary
information on hydraulic head and on conductivity profiles along the bore- Wireline Casing Advancement Drilling Methods for Geoenvironmental
hole (e.g., Molz et al., 1989; Paillet, 2000; USGS, 2015). In the absence of Exploration and Installation of Subsurface Water-Quality Monitoring Devices,
pumping, passive flow may occur in a borehole screened in multiple aquifer ASTM International, West Conshohocken, PA, 2012b. www.aqtesolve.com,
formations (or in one very thick one) due to vertical hydraulic gradients. In accessed August 2015.
this case, the measurement of vertical flow rates, as a function of elevation Butler J. J., J. M. Healey, G. W. McCall, E. J. Garnett, and S. P. Loheide,
within the b orehole, reveals the difference in hydraulic head between the “Hydraulic tests with direct-push equipment,” Ground Water, 40: 25–36, 2002.
contributing and receiving formations or the hydraulic head difference Butler, J. J., P. Dietrich, V. Wittig, and T. Christy, “Characterizing hydraulic
between them, given the other. Such monitoring can be combined with pump- conductivity with the direct-push permeameter,” Ground Water, 45 (4):
ing of a given well, in which case the flowmeter gives the relative flux rates 409–419, 2007.
entering (or leaving) the borehole as a function of distance. Given the knowl- Chapuis, R. P. and D. Chenai, “Comment on ‘Characterizing hydraulic
edge of the hydraulic head within the formation, such data give the hydraulic conductivity with the direct-push permeameter,’ by James J Butler Jr. Peter
conductivity variations along the borehole. Flowmetering in these cases can Dietrich, Volker Witting, and Tom Christy, July–August 2007, 45 (4): 409–
be done using a simple impeller, a heat tracer, or electromagnetic induction 419,” Ground Water, 48 (6): 792–795, 2010.
monitoring (e.g., USGS, 2015). Charbeneau, R. J., Groundwater Hydraulics and Pollutant Transport,
Waveland Press, Long Grove, IL, 2006, p. 593.
Dietrich P, J. J. Butler, and K. Faiss, “A rapid method for hydraulic profiling
47.8 RECHARGE ESTIMATION
in unconsolidated formations,” Ground Water, 46: 323–328, 2008.
Characterization of the large-scale behavior of subsurface flow systems is Environmental Protection Agency, Environmental Geophysics (8/8/2015),
incomplete without quantification of the boundary fluxes of water entering 2015, accessed online at: http://www.epa.gov/esd/cmb/GeophysicsWebsite/
or leaving the aquifer system. In fact, if flow rates at all inflow boundaries pages/basicInformation.htm.
47_Singh_ch47_p47.1-47.4.indd 3 8/22/16 2:16 PM

47-4 Hydrogeologic Characterization
Fetter, C. W., Contaminant Hydrogeology (2nd ed.), Prentice Hall, Upper and comparisons with other tests,” Water Resources Research, 25: 1677–
Saddle River, NJ, 1999, p. 500. 1683, 1989.
Fetter, C. W., Applied Hydrogeology (4th ed.), Prentice Hall, Englewood Neuman, S. P., “Effect of partial penetration on flow in unconfined aquifers
Cliffs, NJ, 2000, p. 691. considering delayed gravity response,” Water Resources Research, 10 (2): 303–
Hackett, G., “Drilling and constructing monitoring wells with hollow‐stem 312, 1974.
augers. Part 1: Drilling considerations,” Groundwater Monitoring & Remediation, Paillet, F. L., “A field technique for estimating aquifer parameters using flow
7 (4): 51–62, 1987. log data,” Ground Water, 38 (4): 510–521, 2000.
Harrington G. A. and M. J. Hendry, “Using direct-push EC logging to Rogers, J. D., “Subsurface exploration using the standard penetration test
delineate heterogeneity in a clay-rich aquitard,” Ground Water Monitoring and the cone penetrometer test,” Environmental & Engineering Geoscience,
Review, 26: 92–100, 2006. 12 (2): 161–179, 2006.
Hunkeler, D., “Geological and hydrogeological characterization of the Rubin, Y. and S. Hubbard, Hydrogeophysics, Springer, Delft, The Netherlands,
subsurface,” Handbook of Hydrocarbon & Lipid Microbiology, edited by 2005, p. 405.
K. N. Timmis, Springer, Berlin, 2010. Sellwood, S. M., J. M. Healey, S. Birk, and J. J. Butler, “Direct-push hydro-
Jones, S. B., M. W. Jon, and O. Dani, “Time domain reflectometry measure- stratigraphic profiling: coupling electrical logging and slug tests,” Ground
ment principles and applications,” Hydrological Processes, 16: 141–153, 2002. Water, 43 (1): 19–29, 2005.
Lapham, W. W., F. D. Wilde, and M. T. Koterba, “Guide-lines and standard Todd, D. K. and L. W. Mays, Groundwater Hydrology (3rd ed.), John-
procedures for studies of groundwater quality selection and installation of Wiley & Sons, New York, 2005, p. 636.
wells, and supporting documentation,” U.S.G.S. Water-Resources Investigations USGS, 2014, www.water.usgs.gov/ogw/gwrp/methods, accessed August 2015.
Report, 96–4233, 1997, p. 110. USGS, 2015, www.water.usgs.gov/ogw/bgas/flowmeter, accessed August 2015.
Mitchell, J. K., and K. Soga, Fundamentals of Soil Behavior, 3rd Edition Zshornack, L., G. C. Bohling, J. J. Butler Jr., and P. Dietrich, “Hydraulic
John Wiley & Sons, New York, 2005, p. 592. profiling with the direct-push permeameter: assessment of probe configura-
Molz, F. J., R. H. Morin, A. E. Hess, J. G. Melville, and O. Guben, “The tion and analysis methodology,” Journal of Hydrology, 496: 195–204, 2013.
impeller meter for measuring aquifer permeability variations: Evaluations
47_Singh_ch47_p47.1-47.4.indd 4 8/22/16 2:16 PM

Chapter 48
Groundwater Modeling
BY
HAI V. PHAM AND FRANK T.-C. TSAI
ABSTRACT where h is the groundwater head (L) in the porous medium; Q is the volu-
metric flux at sources or sinks per unit volume of the porous medium (T–1);
This chapter discusses a general procedure to conduct groundwater modeling.
Ss is the specific storage (L–1) for the porous medium; Kx, Ky , and Kz are
One of the most challenging steps in groundwater model development is gener-
anisotropic hydraulic conductivity (LT–1) for the porous medium in x, y, and
ating valid computational grids to be consistent with geological information. The
z directions, respectively; and t is time (T).
second challenge is model calibration, which can be very time consuming and
Equation (48.1) is derived under isothermal condition, compressible
become intractable for complex groundwater models. We introduce a grid gen-
groundwater and soil matrix, and no spatial variation in water density. The
eration technique to create MODFLOW grids from complex hydrostratigraphic
coordinates are aligned with principal directions of anisotropy of the porous
architectures. The study first uses well logs to construct the hydrostratigraphic
medium. Variations of Eq. (48.1) including formulation in terms of water
architecture with fine vertical discretization through indicator kriging. Then, an
pressure, full anisotropy of hydraulic conductivity, and density-
upscaling procedure is introduced to convert hydrostratigraphic architecture to
dependent flow can be found in Bear and Cheng (2010).
MODFLOW grids. To improve model calibration, we introduce the covariance
Groundwater modeling aims at solving the groundwater equation along
matrix adaptation-evolution strategy (CMA-ES) to estimate groundwater model
with an initial condition, boundary conditions, sinks, and sources. Many
parameters. The CMA-ES is a stochastic optimization method, which provides a
numerical methods have been successfully implemented, including the finite
covariance matrix of estimated parameters for uncertainty analysis. Moreover,
difference method (Harbaugh, 2005), the finite element method (Voss, 1984;
the CMA-ES uses a global-local derivative-free algorithm, which is easy to be
Lin et al., 1997; Diersch, 2002), the finite volume method (Panday et al.,
parallelized for calibrating computationally demanding groundwater models.
2013), the analytic element method (Haitjema, 1995), and the lattice
The techniques are applied to groundwater modeling in the Baton Rouge area,
Boltzmann method (Servan-Camas and Tsai, 2009; 2010). The USGS
southwestern Louisiana. Reconstruction of the Baton Rouge aquifer system
MODFLOW (Harbaugh, 2005) is one of the most popular groundwater solv-
using a vast amount of well logs is unprecedented. Specific groundwater models
ers to be used for modeling groundwater flows. Many graphical user inter-
for the “1,200-foot” sand to the “2,000-foot” sand are developed and calibrated
faces, commercials [e.g., Groundwater Modeling System (GMS) (Aquaveo,
by the parallel CMA-ES. Groundwater head prediction uncertainty is derived
2014), Visual MODFLOW (Waterloo, 2005), Groundwater Vistas (Rumbaugh
through the Monte Carlo realizations of model parameters.
and Rumbaugh, 2001), and Argus One (Liskov et al., 1987)] and noncom-
mercials [e.g., ModelMuse (Winston, 2009) and PMWIN (Chiang and
48.1 INTRODUCTION
Chiang, 2001)], have been developed to facilitate required MODLFOW pack-
Groundwater resources are vital for sustainable economic and demographic ages generation for real-world case studies.
growth (Gleeson et al., 2012). Reliable groundwater modeling is necessary for
successful management of groundwater resources. However, groundwater
48.2 GROUNDWATER MODEL DEVELOPMENT
simulation models are subjected to various uncertainties in their predictions.
The uncertainties may come from naturally occurring randomness (aleatory Developing a groundwater model requires geological and geophysical data as
uncertainty), insufficient data and knowledge (epistemic uncertainty), and well as hydrologic and hydrogeologic data. Figure 48.1 shows a flowchart of
lack of technology to translate known data/knowledge into valid groundwater groundwater model development. Some components may switch the order
models (technological uncertainty). Aleatory uncertainty is generally small without loss of generality. A numerical model may be selected before a com-
and not possible to eliminate. Epistemic uncertainty caused by a lack of data putational grid is generated.
and knowledge (Hora, 1996; Senge et al., 2014) is significant and can be effec- Errors and uncertainties exist in every component in the flowchart.
tively reducing by collecting informative data from the field. Technological Propagation of the errors and uncertainties through the components will
uncertainty can be significantly minimized by using better numerical meth- eventually affect groundwater calibration and prediction results.
ods and computing resources. This chapter focuses on groundwater modeling One of the most challenging components is grid generation, which governs
and introduces a general procedure for groundwater model development most of the technological uncertainty. Grid generation is the first step to
based on existing data. Aleatory and epistemic uncertainties will remain the construct a groundwater model structure after data are collected. A valid
same throughout the model-development process. However, technological groundwater model needs to have a valid computational grid that is consis-
uncertainty can be reduced by using a better grid-generation technique and tent with geological information whenever available. Otherwise, error from
high-performance computing for model calibration. geologically unsupported computational grid can produce significant system-
A groundwater equation describing flows in saturated sedimentary rocks atic error in groundwater prediction. Another challenging component is
can be derived by introducing Darcy’s law to the mass conservation law (Bear, model calibration, which is the last step in groundwater model development
1972). A widely accepted groundwater equation based on groundwater head before the developed model is applied to tasks. Model calibration can be very
(hydraulic head of groundwater) is (Freeze and Cherry, 1979) time consuming for complex groundwater models. Without advanced search
algorithms and computing resources, searching for better parameter values
∂h ∂  ∂h  ∂  ∂h  ∂  ∂h 
Ss =  Kx  + K y  +  K z  + Q (48.1) can end prematurely. Moreover, grid error amplifies parameter estimation
∂t ∂ x  ∂ x  ∂ y  ∂ y  ∂z  ∂z  error. Model calibration may result in overparameterization and unrealistic
48-1
48_Singh_ch48_p48.1-48.8.indd 1 8/22/16 2:17 PM

48-2 Groundwater Modeling
Figure 48.1 Flowchart of groundwater model development.
model parameter values in order to reduce fitting error produced by model 48.3.1 Well Log Interpretation
structure error due to an invalid grid. This chapter focuses on discussions of The primary sources of information used to establish hydrofacies geometries
grid generation and model calibration for groundwater modeling in order to are wire-line spontaneous potential (SP) and electrical resistivity logs for
reduce model structure uncertainty and better model parameter estimation. boreholes. Spontaneous potential and resistivity log responses are controlled
largely by the ratio of sand to clay minerals. They have long been used to
interpret sedimentary depositional environments. Galloway (1977) used SP
48.3 GRID GENERATION and resistivity curve morphologies to identify fluvial facies for channel fill,
The literature shows two approaches commonly used to construct a compu- levee, crevasse splay and floodplain, and established a meandering stream
tational grid from borehole data: one is the solid approach and the other is the facies. Kerr and Jirik (1990) adapted Galloway’s (1977) facies model and pro-
predefined grid approach. Before implementing either approach, one needs to vided examples of SP and resistivity responses that match known fluvial facies
obtain geological information (e.g., lithology, bed boundary elevation, forma- for the middle Frio formation, South Texas. Sands deposited by braided
tion dip, etc.) from well logs. Readers are referred to some classic books for streams produce jagged, wedge-shaped curve morphologies (Miall, 2010).
well-log-interpretation techniques (Schlumberger, 1972; Hilchie, 1982; Based on these established relationships between log responses and fluvial
Bassiouni, 1994). facies, Chamberlain et al. (2013) used SP and resistivity data to study deposi-
Using the solid approach, one needs to manually correlate well logs and tional environments of siliciclastic sediments in the Baton Rouge area,
label distinct hydrofacies for each well log. Once the correlation between well Louisiana.
logs is established, interpolation methods are used to generate the surfaces of Following Chamberlain et al. (2013), SP, resistivity, and gamma ray (when
hydrofacies. These surfaces represent the hydrofacies boundaries and con- available) are used to identify the location of sand facies at depth. Figure 48.2a
struct a solid model. Jones et al. (2002) and Lemon and Jones (2003) devel- shows a typical SP-resistivity log from fluvial depositional environments.
oped a grid generator to generate computational grids for MODFLOW Based on deviations from a visually estimated shale baseline, boundaries of
(Harbaugh et al., 2000; Harbaugh, 2005) from solid models. The beauty of this sands can be drawn on inflection points of SP curves. A cut-off value gener-
approach is the creation of nonuniform computational layers that match well ally fell between 10 and 35 ohm-m for resistivity curves is assigned to deter-
with the generated surfaces, including pinch-outs. Due to its simplicity, the mine boundaries of sands. Low long-normal resistivity generally indicates the
algorithm was adopted in several commercial software packages, for example, occurrence of salty water. Low gamma ray response generally indicates a sand
GMS (Aquaveo, 2014) and RockWorks (RockWare, 2010). facies. Sand boundaries can be well identified by correlating SP, resistivity, and
The biggest challenge in using the solid model approach is to perform gamma ray curves (Schlumberger, 1972; Hilchie, 1982; Bassiouni, 1994). For
manual correlation between well logs, which is subjective and can become example, four sand facies are picked and many thin sands are ignored as
laborious and impractical when dealing with a vast amount of well logs in shown in Fig. 48.2a. Nonsand intervals are assumed clay (shale or mudstone)
areas known to be highly complex (e.g., fluvial depositional environments). facies.
Manually correlating well logs often results in inconsistency with geological Well log interpretation is inevitably subjected to an individual’s experience
deposition, forces correlation of unrelated hydrofacies, and produces errone- and the purpose of the work. This chapter does not intend to discuss the
ous hydraulic connections of discontinuous hydrofacies. uncertainty of computational grids due to different log interpretations.
The predefined grid approach usually generates uniform, relatively coarse Moreover, it is possible to use the established relationships between log
layers directly to be used for groundwater modeling. Examples of this responses and fluvial facies from Galloway (1977), Kerr and Jirik (1990), and
approach include using T-PROGS (Carle, 1999) and geostatistical tools [e.g., Miall (2010) to infer different fluvial facies. For example, Fig. 48.2a shows
GSLIB (Deutsch and Journel, 1997)]. This approach does not force generating some identified fluvial facies based on the established relationships. However,
surfaces of hydrofacies, and therefore, avoids the issues caused by manual identifying specific fluvial facies is not the scope of this chapter. Instead, this
correlation. Since geostatistical approaches have been well received in the chapter focuses on sand and clay facies identification in well logs.
literature, they are commonly adopted by commercial software, for example,
T-PROGS with GMS (Jones et al., 2005; Faulkner et al., 2012; Aquaveo, 2014). 48.3.2 Indicator Kriging
However, the greatest concern of using predefined grids is of losing the verti- The study constructs hydrostratigraphic architecture of sand and clay facies
cal resolution of hydrofacies geometries if layers are not fine enough. Using using information from well logs. Well log data are first transformed into
very fine layers can improve vertical resolution to better capture vertical binary indicator values. The indicator value for sand facies is 1 and for clay
hydrofacies geometry, but will significantly increase computation time, since facies is 0. To honor a regional geological dip as shown in Fig. 48.2b, indicator
predefined grids are directly used for the groundwater modeling purpose. kriging needs to be performed on inclined surfaces where indicator data are
This chapter presents a grid-generation technique that maintains high obtained at the intersections with boreholes. To make it easier for operating
vertical resolution of hydrofacies geometries with a reasonable number of indicator kriging, all boreholes are translated vertically to a nondipped domain.
nonuniform boundary-fitted layers. This is done by vertically upscaling a To do so, the vertical translation distance depends on the dip angle and the
very-fine hydrostratigraphic architecture into a computational grid for distance from well log location to a strike that serves as a pivot as illustrated in
groundwater modeling. The method avoids correlation error by using Fig. 48.2c. Then, indicator kriging is performed on horizontal surfaces given a
manual correlation and avoids gridded error by using a predefined grid. The two-dimensional grid. The two-dimensional grid can have irregular boundar-
grid-generation technique is able to facilitate simulation model development ies and its cells can be generated by any methods available in the literature. This
and reduce technological uncertainty by constructing geologically valid com- chapter focuses on generating computational grids to be used by MODFLOW.
putational grids. While this chapter presents computational grid generation The same discretization is applied to all horizontal surfaces at different depths.
for MODFLOW, the technique is not restricted to any numerical models. Then, a detailed three-dimensional hydrostratigraphic architecture can be
48_Singh_ch48_p48.1-48.8.indd 2 8/22/16 2:17 PM

Case study: Groundwater modeling in Baton Rouge, southeastern Louisiana 48-3
Depth (ft) SP (mv) Resistivity (ohm-m)

–150–100–50 1 10 100
1000
Clay
1100 A Sand
Imaginary horizontal plane
Shale baseline
1200
Clay
Dip angle
1300
Clay
1400 B Sand Clay Dip d
irectio
n
1500 Clay Sand
S trike
C Sand Clay
1600 Sand Sand Clay
1700 D Horizontal
E Clay Clay Clay
Inclined surface
1800
Sand surface
1900
(a) (b) (c)
Figure 48.2 (a) Interpretation of fluvial facies for a well log: A—amalgamated braided channel-fill with brackish water, B—channel-fill point bar sand with brackish water, C—
stacked/amalgamated channel-fill with very salty water, D—floodplain, and E—natural levee (Kerr and Jirik, 1990; Miall, 2010); and translation of borehole positions from (b) dipped
domain to (c) nondipped domain.
achieved by assembling a large number of horizontal surfaces with a fine inter- 2009) run a global heuristic algorithm for exploring the search landscape fol-
val. lowed by a local derivative algorithm for exploiting favorable search regions.
Indicator kriging is adopted to derive sand-clay distribution in horizontal We adopt the CMA-ES (Hansen and Ostermeier, 2001; Hansen et al., 2003)
surfaces of one-foot intervals. The resulting indicator data from horizontal as a global-local stochastic derivative-free algorithm for groundwater model
surfaces are used to compute experimental variograms. Then, a variogram calibration. The important features in the CMA-ES are global-local search
model can be derived by fitting to the experimental variograms. It is noted capability, parallelization, and uncertainty quantification for estimated
that the grid generation technique is not limited to indicator kriging. Any parameters, which are especially great for calibrating computationally
geostatistical method can be used to estimate hydrofacies for a surface. demanding groundwater flow models (Elshall et al., 2015).
The expected value of an indicator function at an unobserved location is The enhanced search properties of the CMA-ES stems from its complex
obtained by learning techniques with high level of abstract description. The CMA-ES
S
adapts a covariance matrix representing the pair-wise dependency between
v( x 0 ) = ∑λi I ( x i ), (48.2) decision variables, which approximates the inverse of the Hessian matrix up
i =1 to a certain factor. The solution is updated with the covariance matrix and an
adaptable step size, which are adapted by two conjugates that implement
where v( x 0 ) is the expected value at unobserved location x 0 , S is the num- heuristic control terms. The covariance matrix adaptation uses information
ber of boreholes for a horizontal surface, and λi are the indicator kriging from the current population and from the previous search path.
weights. Indicator kriging has been well documented in the literature. CMA-ES utilizes multiple solutions in iteration that do not exchange infor-
Readers are referred to Olea (1999) for more information. mation and allows for embarrassingly parallel computation. This is the most
The expected value of indicator function represents the probability that efficient parallel technique since the solutions in iteration do not communi-
facies at a location x fall into sand facies or clay facies. By giving a cut-off a as cate. The parallel CMA-ES superiorly improves the calibration speed over the
follows, distributed sand and clay facies on a horizontal surface can be sequential CMA-ES. In addition, the speedup of parallel runs scales variably
achieved: with increasing the number of processors up to a certain limit.
In addition to the global-local search capabilities and parallelization,
 1 : sand if v ( x ) ≥ α
I (x) =  (48.3) another favorable feature of CMA-ES is to quantify model parameter uncer-
 0 : clay if v ( x ) < α tainty due to estimation error. The solutions of the CMA-ES, which consist of
a maximum likelihood estimate and a full covariance matrix, can be used for
Determination of a defensible cut-off value is challenging. A value of 0.5 is Monte Carlo sampling. Several algorithms have utilized the covariance matrix
commonly used for a neutral selection. However, a better cutoff can be deter- for Monte Carlo sampling (Haario et al., 1999; Haario et al., 2001; Kavetski et
mined in a calibration process where facies estimates are subject to addi- al., 2006; Smith and Marshall, 2008; Cui et al., 2011; Zhang and Sutton, 2011).
tional information, for example, driller’s logs, total volume of sand or clay As pointed out by Müller and Sbalzarini (2010) and Mueller (2010), the
facies from electrical logs, etc. (Elshall et al., 2013). CMA-ES shares many common concepts and features with the derivative-free
Once a hydrostratigraphic architecture with very fine vertical discretization Markov chain Monte Carlo (MCMC) sampling algorithms (Haario et al.,
is generated by indicator kriging, the hydrostratigraphic architecture is 1999; Haario et al., 2001; Haario et al., 2006; Andrieu and Thoms, 2008;
upscaled to a MODFLOW grid by merging the same hydrofacies in the verti- Müller and Sbalzarini, 2010). Elshall et al. (2015) showed that the adapted
cal direction to reduce the number of layers. covariance matrix of the maximum-likelihood estimates is precise and can be
used for Monte Carlo sampling. Their work is one of a few studies that used
CMA-ES to quantify model parameter uncertainty.
48.4 MODEL CALIBRATION
The use of optimization algorithms for model calibration in groundwater

48.5 CASE STUDY: GROUNDWATER MODELING IN BATON
model is a common practice. The optimization algorithms include local
ROUGE, SOUTHEASTERN LOUISIANA
derivative algorithms, global heuristic algorithms, and hybrid global-local
algorithms. Local derivative algorithms are computationally efficient, but at This section illustrates the implementation of the flowchart of the groundwa-
the cost of trapping at local solutions. Global heuristic algorithms are gener- ter model development by using a case study in the Baton Rouge area, south-
ally time consuming, but are able to escape local optima and improve solu- eastern Louisiana. Figure 48.3 shows the Baton Rouge area and the
tions through a simple to complex forms of learning, such as genetic groundwater modeling domain. The Baton Rouge aquifer system consists of
algorithm (Harrouni et al., 1996; Solomatine et al., 1999; Karpouzos et al., a succession of south-dipping siliciclastic sand units and mudstones of Upper
2001; Bastani et al., 2010) and particle swarm optimization (Scheerlinck et al., Miocene through Pleistocene age, is extended to a depth of 3000 ft (914 m),
2009; Jiang et al., 2010). Hybrid global-heuristic local-derivative algorithms and is highly complex as a result of fluvial deposition (Chamberlain et al.,
(Tsai et al., 2003; Blasone et al., 2007; Matott and Rabideau, 2008; Zhang et al., 2013). Sand deposition is nonuniform due to spatial and temporal variations
48_Singh_ch48_p48.1-48.8.indd 3 8/22/16 2:17 PM

Figure 48.3 Map of the Baton Rouge area and borehole locations. The box in the middle is the groundwater model-
ing area.
in fluvial processes as well as large amount of missing sand possibly due to kept the same as in the hydrostratigraphic architecture. Using the upscaling
erosional unconformity. From the region-scale study (Griffith, 2003), the approach, a MODFLOW grid of 968,316 cells given by 93 rows, 137 columns,
Baton Rouge aquifers dip south towards the Gulf of Mexico. The east–west and 76 layers accurately matches the complex hydrostratigraphic architecture
trending Baton Rouge fault and Denham Springs–Scotlandville fault crosscut and preserves layer continuity. Each cell is 200 m by 200 m with cell thickness
the aquifer/aquitard sequence in the study area (McCulloh and Heinrich, ranges from 3.05 to 13.4 m. The average thickness of the layers is 5.2 m. The
2013). Bense and Person (2006) suggested that the Baton Rouge fault is a low- “1200-ft,” the “1500-ft,” and the “1700-ft” sands are from layer 6 to layer 46.
permeability fault and acts as a horizontal flow barrier (HFB) that separates The “2000-ft” sand is from layer 47 to layer 76. The hydrostratigraphic archi-
the freshwater to the north from the saline water to the south. tecture shows the “1200-ft” sand, the “1500-ft” sand, and the “1700-ft” sand
between the two faults are interconnected. Therefore, these sands were mod-
48.5.1 Hydrostratigraphic Architecture eled together. The “2000-ft” sand is separated from the “1500-ft” sand and
“1700-ft” sand by a thick clay layer and was simulated separately.
We analyzed wireline well logs from 583 boreholes shown in Fig. 48.3. The Monthly pumping data are available from the Louisiana Capital Area
well log data were used to construct a hydrostratigraphic architecture of the Ground Water Conservation Commission (CAGWCC) from 1975 to 2010.
Baton Rouge aquifer system, which includes the Mississippi River Alluvial The “1200-1500-1700-ft” sands model has 87 pumping wells. The average
Aquifer (MRAA), the “400-ft” sand, the “600-ft” sand, the “800-ft” sand, the pumping rate of these wells from 1975 to 2010 was 112,556 m3/day (29.73
“1000-ft” sand, the “1200-ft” sand, the “1500-ft” sand, the “1700-ft” sand, the million gallons per day, MGD). The connector well (EB-1293) was injecting
“2000-ft” sand, the “2400-ft” sand, and the “2800-ft” sand. These sand units about 2600 m3/day (0.69 MGD) of groundwater from the “800-ft” sand to the
were classified and named by their approximate depth below ground level in “1500-ft” sand. The “2000-ft” sand model has 29 pumping wells extracting
the Baton Rouge Industrial District (Meyer and Turcan, 1955). about 78,457 m3/day (20.73 MGD) in December 2010.
The indicator kriging was adopted to construct a hydrostratigraphic archi- The initial head on January 1, 1975, and the time-varied head boundary
tecture with a dip angle of 0.29° and a cutoff value of 0.40 (Elshall et al., 2013). condition were estimated through extrapolation of the nearby head observa-
The resulted hydrostratigraphic architecture is illustrated in Fig. 48.4. The tion data.
model domain in the planar direction is discretized into 93 rows and 137 The “1500-ft” sand and the “1700-ft” sand are considered to have the same
columns with a cell size 200 m by 200 m. The hydrostratigraphy shows the hydrogeological parameter values, which are different from the “1200-ft”
complexity of the fluvial depositional environment in the Baton Rouge area. sand. The two faults are considered as HFBs and their permeability values are
There are 166 pumping wells in the model area. The majority of the pumping characterized by the hydraulic characteristics (HCs) (Hsieh and Freckleton,
wells in the Industrial District were screened at the “1200-ft” sand and the 1993). The “1200-ft” sand and the combination of the “1500-ft” and the
“2000-ft” sand. Public supply wells area scattered, except for the cluster of “1700-ft” sands have four homogeneous model parameters (hydraulic con-
Lula wells that were screened at the “1500-ft” sand. Saltwater intrusion has ductivity, specific storage, and two HCs for the two faults) to be estimated.
been a serious issue in the “1500-ft” sand and the “2000-ft” sand (Tomaszewski, Thus, the “1200-1500-1700-ft” sands model has eight unknown model
1996). The connector well (EB-1293) was built in 1999 to raise hydraulic head parameters. In addition to the four homogeneous model parameters, the
in the “1500-ft” sand by injecting groundwater of the “800-ft” sand through “2000-ft” sand model considers a vertical anisotropy ratio of hydraulic con-
pressure difference. A newly scavenger well (EB-1424) was built in 2014 to ductivity and a boundary head adjustment factor for the eastern boundary
extract salt water out of the “1500-ft” sand. Evaluation of the effectiveness of between the two faults to be estimated. The “2000-ft” sand model has six
such groundwater remediation approaches requires a valid geological model unknown model parameters.
and later a valid computational grid. The validity of the Baton Rouge geo- The MODFLOW Well (WEL) package is used to take into account the
logical model was verified by the up-to-date well log data and the depths of pumping rates of the pumping wells and the injection rate of the connector
the screens of the pumping wells as shown in Fig. 48.4. well. The Time-Variant Specified-Head (CHD) package is used to simulate
the time-dependent boundary condition, which is assigned to all active cells
48.5.2 Groundwater Modeling at the model boundaries. Inactive cells represent the clay unit. The Baton
The constructed hydrostratigraphic architecture from the previous section is Rouge fault and the Denham Springs–Scotlandville fault are represented as
used to generate a MODFLOW grid. The interested sands for groundwater HFBs using the Horizontal Flow Barrier package. The model calibration
modeling are the “1200-ft” sand, the “1500-ft” sand, the “1700-ft” sand, and period is from January 1, 1975, to January 1, 2010, 432 monthly stress periods.
the “2000-ft” sand. The model domain and the horizontal discretization are The time step is 1 month.
48_Singh_ch48_p48.1-48.8.indd 4 8/22/16 2:17 PM

Summary 48-5
Scavenger well (EB-1424)

Lula wells
Industrial district wells Connector well (EB-1293)
Mississippi river
0.0
–0.5
Elevation (x103) (ft)

–1.0
–1.5
–2.0
–2.5
690
3.380 685
680
3.378
3.376
3.374 3
.372 3.3
70 675 3) (m)
UTM_Y 3.368 3
670 10
(x10 6)
.366
3.364 3
UTM _Y (x
(m) .362 665
Figure 48.4 Constructed architecture of the Baton Rouge aquifer system. Cells underneath land surface are sand. Clay is
blanked. Vertical white columns are pumping wells with screens at the bottom.
48.5.3 Model Calibration Using Parallel CMA-ES station is the center of the cone of depression in the “1500-ft” sand. Heavy
To calibrate a groundwater model and estimate model parameters, the CMA- pumping has caused significant groundwater level decline north of the Baton
ES (Hansen and Ostermeier, 2001; Hansen et al., 2003) was adopted to mini- Rouge fault and has caused saline groundwater flowing northward across the
mize the root mean square error (RMSE) between the calculated and Baton Rouge fault toward the Lula pumping station and the Industrial District.
observed groundwater heads. The execution time for a single model simula- 48.5.5 Uncertainty Quantification
tion is around 1.89 ± 0.10 h for the “1200-1500-1700-ft” sands model and 1.28
± 0.1 h for the “2000-ft” sand model by a personal computer equipped with a Monte Carlo simulation was adopted to quantify groundwater head uncer-
Pentium Dual Core CPU 3.0 GHz and 2GB RAM. It is impractical to conduct tainty due to model parameter estimation uncertainty. Parameter realizations
model calibration in a single processor. For example, using a population size were generated using the mean and the covariance matrix of the estimated
of 80 and 40 iterations in CMA-ES, the computation time would be about 252 model parameters obtained by the CMA-ES after model calibration was com-
days for calibrating the “1200-1500-1700-ft” sands model. pleted. For the illustration purpose, the study analyzed the simulated ground-
To accelerate the model calibration process, a parallel CMA-ES (Elshall et water head at a USGS observation well EB-146 in December 14, 2010. EB-146
al., 2015) was implemented to SuperMike-II, a supercomputer at Louisiana was screened at the “1200-ft” sand, close to the Government St. pumping
State University. To maximize the efficiency of the parallel implementation, station. Figure 48.6a shows the convergence of the RMSE using CMA-ES with
the number of core-based processors used is the same as the population size the population size of 80. The RMSE was not improved after 60 iterations; yet
λ. For the best performance, Hansen and Ostermeier (2001) and Hansen et al. the calibration process was not terminated to ensure convergence of the cova-
(2003) recommended the optimal population size 4 + [ 3ln(n)] ≤ λ ≤ 10n, riance matrix. The standard deviation of simulated groundwater head was
where n is the number of unknown parameters. Thus, the number of proces- calculated based on 290 realizations at 10, 30, 50, and so forth up to 170
sors for the “1200-1500-1700-ft” sands model is 80 ( λ = 80 = 10n) and for the iterations as shown in Fig. 48.6a. The result indicates that the covariance
“2000-ft” sand model is 64 ( λ = 64 ≈ 10.67n ). For both groundwater models, matrix was converged within 170 iterations. Figure 48.6b shows the mean
the optimal population size is at the upper limit λ = 10n. Given that the com- simulated head and its one-standard deviation bounds at EB-146 for the last
munication overhead is less than 1 s per iteration, the speedup of the parallel iteration. After reaching the optimal solution, the head variance quickly con-
CMA-ES is roughly the population size λ. verged in less than 150 realizations.
48.5.4 Model Calibration Results 48.6 SUMMARY

The “1200-1500-1700-ft” sands model was calibrated using 2805 groundwater This chapter presents a general procedure for groundwater model develop-
head data from 20 USGS observation wells from January 1975 to December ment with a focus on grid generation and model calibration, which are two
2010. The “2000-ft” sand model was calibrated using 1285 head data from 18 crucial modeling steps in order to reduce model structure uncertainty and to
USGS observation wells for the same time period. The model calibration improve model parameter estimation. To do so, a better grid generation tech-
results show a good matching between the observed and the simulated ground- nique is needed in order to create a valid computational grid supported by
water heads. The RMSEs are 1.44 and 2.95 m for the “1200-1500-1700-ft” geological data. Second, a better optimization algorithm is needed for effi-
sands model and the “2000-ft” sand model, respectively. These RMSEs are cient model calibration and parameter uncertainty estimation.
acquired after 59 and 40 iterations for the “1200-1500-1700-ft” sands model In this chapter, we present a grid generation technique using indicator krig-
and the “2000-ft” sand model, respectively. The good fitting results in ground- ing to work on a vast amount of well log data. The technique can preserve
water head data although using homogeneous model parameters are because a facies geometries of a complex hydrostratigraphic architecture by using fine
valid geological model was built and correctly converted to a MODFLOW grid, vertical discretization and regional geological dip. In addition, the technique
without which complex heterogeneous model parameter fields may be needed. avoids a possible overwhelming number of computational cells by using an
Table 48.1 presents the estimated parameter values for all sands. The Baton upscaling approach.
Rouge fault and the Denham Springs–Scotlandville fault are found to be low- This chapter introduces a parallel CMA-ES as an efficient optimization
permeability faults that restrict horizontal flow. The “2000-ft” sand has hydrau- method for model calibration. The important features in the CMA-ES include
lic conductivity and specific storage higher than other three sands. Figure 48.5 global-local search capability, parallelization, and uncertainty quantification
presents the groundwater head distributions at the end of the model calibra- for estimated parameters, which are especially great for calibrating computa-
tion period, December 31, 2010, using the estimated parameters in Table 48.1. tionally demanding groundwater flow models.
The results show that the Industrial District is the center of the cone of The procedure for groundwater model development was illustrated by the
depression in the “1200-ft” sand and the “2000-ft” sand and the Lula pumping groundwater modeling study in Baton Rouge, southeastern Louisiana. A
48_Singh_ch48_p48.1-48.8.indd 5 8/22/16 2:17 PM

Figure 48.5 Simulated groundwater head distributions on December 31, 2010 for (a) “1,200-ft” sand, (b) “1,500-ft” sand
and “1,700-ft” sand, and (c) “2,000-ft” sand.
48_Singh_ch48_p48.1-48.8.indd 6 8/22/16 2:17 PM

References 48-7
Figure 48.6 Convergence profiles for (a) RMSE (meters) and standard deviation (meters) of simulated groundwater heads at EB-146, and (b) sim-
ulated groundwater head and its one-standard deviation bounds at EB-146.
Bassiouni, Z., “Theory, measurement, and interpretation of well logs,”

Table 48.1 Estimated Model Parameter Values for Different Sands
Memorial Fund of AIME, edited by Henry L. Doherty, Society of Petroleum
“1200-ft”
“1500-ft” and
“2000-ft”
Engineers, Richardson, TX, 1994.
Parameter “1700-ft” Bastani, M., M. Kholghi, and G. R. Rakhshandehroo, “Inverse modeling of
sand sand
sands variable-density groundwater flow in a semi-arid area in Iran using a genetic
Hydraulic conductivity (m/d) 23.13 25.64 144.86 algorithm,” Hydrogeology Journal, 18: 1191–1203, 2010, doi: 10.1007/s10040-
010-0599-8.
Specific storage (1/m) 5.27×10–6 2.82×10–6 1.86×10–5 Bear, J., Dynamics of Fluids in Porous Media, Courier Corporation, New
York, NY, 1972.
Vertical anisotropy ratio 1 1 1 Bear, J. and A.H-D. Cheng, Modeling Groundwater Flow and Contaminant
Transport, Springer Science & Business Media, The Netherlands, 2010.
HC of the Baton Rouge fault (1/d) 2.64×10–3 2.48×10–4 4.20×10–3
Bense, V. F. and M. A. Person, “Faults as conduit-barrier systems to fluid
HC of the Denham Springs– flow in siliciclastic sedimentary aquifers,” Water Resources Research, 42:
6.08×10–3 5.00×10–2 1.34×10–6
Scotlandville fault (1/d) W05421, 2006, doi: 10.1029/2005WR004480.
Boundary condition adjustment Blasone, R-S., H. Madsen, and D. Rosbjerg, “Parameter estimation in dis-
– – 1.36 tributed hydrological modelling: comparison of global and local optimisation
factor (m)
techniques,” Nordic Hydrology, 38: 451–476, 2007.
Carle, S. F., T-PROGS: Transition Probability Geostatistical Software,
significant amount of effort was spent on interpreting a great number of well University of California, Davis, CA, 1999.
logs in order to reconstruct a valid geological model before groundwater Chamberlain, E. L., J. S. Hanor, and F.T-C. Tsai, “Sequence stratigraphic
models were developed. The grid generation technique successfully reveals characterization of the Baton Rouge aquifer system, southeastern Louisiana,”
the complexity of the fluvial depositional environment in the Baton Rouge Transactions of the Gulf Coast Association of Geological Societies, 63: 125–136,
area including two geological faults. The validity of the Baton Rouge geologi- 2013.
cal model was verified by the up-to-date well log data and the depths of the Chiang, W-H. and W. H. Chiang, 3D-Groundwater Modeling with PMWIN,
screens of the pumping wells. Springer, Heidelberg, Germany, 2001.
With a focus on the “1200-ft” sand, “1500-ft” sand, “1700-ft” sand, and Cui, T., C. Fox, and M. J. O’Sullivan, “Bayesian calibration of a large-scale
“2000-ft” sand groundwater modeling, this chapter demonstrated the advan- geothermal reservoir model by a new adaptive delayed acceptance Metropolis
tages of using parallel CMA-ES for not only model calibration, but also Hastings algorithm,” Water Resources Research, 47: W10521, 2011, doi:
parameter uncertainty quantification. Using parallel computing, we were able 10.1029/2010WR010352.
to make model calibration time tractable for our computationally heavy Deutsch, C. V. and A. G. Journel, GSLIB: Geostatistical Software Library and
groundwater models. The covariance matrix provided by the CMA-ES User’s Guide, 2nd ed., Oxford University Press, New York, 1997.
enables us to quantify groundwater head uncertainty through Monte Carlo Diersch, H. J. G., FEFLOW Reference Manual, DHI-WASY, Berlin,
realizations of model parameters. Germany, 2002, p. 116.
Finally, we iterate the importance of developing a valid computational grid Elshall, A. S., H. V. Pham, F.T-C. Tsai, L. Yan, and M. Ye, “Parallel inverse
to reduce systematical uncertainty in groundwater modeling caused by a lack modeling and uncertainty quantification for computationally demanding
of advanced technology to translate geophysical data into valid groundwater groundwater-flow models using covariance matrix adaptation,” Journal of
models. As demonstrated in the case study, highly parameterized groundwa- Hydrological Engineering, 20: 04014087, 2015, doi: 10.1061/(ASCE)HE.1943-
ter models are not needed and model calibration can be simplified as long as 5584.0001126.
the groundwater model structure is supported by geological data. Elshall, A. S., F.T-C. Tsai, and J. S. Hanor, “Indicator geostatistics for
reconstructing Baton Rouge aquifer-fault hydrostratigraphy, Louisiana,
USA,” Hydrogeology Journal, 21: 1731–1747, 2013, doi: 10.1007/s10040-013-
ACKNOWLEDGMENTS 1037-5.
The study was supported in part by the LSU Economic Development Faulkner, B. R., J. Renée Brooks, K. J. Forshay, and S. P. Cline, “Hypothetic
Assistantship (EDA), the National Science Foundation under Grant No. EAR- flow patterns in relation to large river floodplain attributes,” The Journal of
1045064, and the United States Geological Survey under Grant/Cooperative Hydrology, 448–449: 161–173, 2012, doi: 10.1016/j.jhydrol.2012.04.039.
Agreement No. G10AP00136. LSU High Performance Computing is acknowl- Freeze, R. A. and J. A. Cherry, Groundwater, Prentice-Hall, Englewood
edged for providing access to supercomputers. Cliffs, NJ, 1979.
Galloway, W. E., Catahoula Formation of the Texas Coastal Plain:
Depositional Systems, Composition, Structural Development, Ground-Water
REFERENCES Flow History, and Uranium Distribution, Bureau of Economic Geology, Texas
Andrieu, C. and J. Thoms, “A tutorial on adaptive MCMC,” Statistics and University, Austin, TX, 1977.
Computing, 18: 343–373, 2008, doi: 10.1007/s11222-008-9110-y. Gleeson, T., Y. Wada, M. F. P Bierkens, and L. P. H. van Beek, “Water bal-
Aquaveo, Groundwater Modeling System (GMS), Aquaveo, LLC, Provo, UT, ance of global aquifers revealed by groundwater footprint,” Nature, 488:
2014. 197–200, 2012, doi: 10.1038/nature11295.
48_Singh_ch48_p48.1-48.8.indd 7 8/22/16 2:17 PM

Griffith, J. M., Hydrogeologic framework of Southeastern Louisiana, Water McCulloh R. P. and P. V. Heinrich, “Surface faults of the south Louisiana
Resources Technical Report 72, Louisiana Department of Transportation and growth-fault province,” Geological Society of America, Special Paper 493:
Development, Baton Rouge, LA, 2003, p. 21. 37–49, 2013, doi: 10.1130/2012.2493(03).
Haario, H., M. Laine, A. Mira, and E. Saksman, “DRAM: efficient adaptive Meyer, R. R. and A. N. Turcan, Geology and Ground-Water Resources of the
MCMC,” Statistics and Computing, 16: 339–354, 2006, doi: 10.1007/s11222- Baton Rouge Area, Louisiana, US Government Printing Office, Washington:
006-9438-0. U.S. Geological Survey Water-Supply Paper 1296, 1955, p. 144.
Haario, H., E. Saksman, and J. Tamminen, “Adaptive proposal distribution Miall, A. D., Alluvial deposits, Facies Models 4, edited by N. P. James and
for random walk Metropolis algorithm,” Computational Statistics, 14: 375– R. W. Dalrymple, 2010, Geological Association of Canada, Newfoundland,
395, 1999, doi: 10.1007/s001800050022. Canada, pp. 105–137.
Haario, H., E. Saksman, and J. Tamminen, “An adaptive Metropolis algo- Mueller, C. L., “Exploring the Common Concepts of Adaptive MCMC and
rithm,” Bernoulli, 7: 223–242, 2001, doi: 10.2307/3318737. Covariance Matrix Adaptation Schemes,” Dagstuhl Seminar Proceedings
Haitjema, H. M., Analytic Element Modeling of Groundwater Flow, 10361 on Theory of Evolutionary Algorithms, Schloss Dagstuhl—Leibniz-
Academic Press, London, UK, 1995. Zentrum fuer Informatik, Germany, 2010.
Hansen, N., S. D. Müller, and P. Koumoutsakos, “Reducing the time com- Müller, C. L. and I. F. Sbalzarini, Gaussian adaptation as a unifying frame-
plexity of the derandomized evolution strategy with covariance matrix adap- work for continuous black-box optimization and adaptive Monte Carlo sam-
tation (CMA-ES),” Evolutionary Computation, 11: 1–18, 2003. pling, 2010 IEEE Congress on Evolutionary Computation (CEC), 2010, p. 1–8.
Hansen, N., A. Ostermeier, “Completely derandomized self-adaptation in Olea, R. A., Geostatistics for Engineers and Earth Scientists, Springer,
evolution strategies,” Evolutionary Computation, 9: 159–195, 2001, doi: http:// Boston, 1999.
dx.doi.org/10.1162/106365601750190398. Panday, S., C. Langevin, R. Niswonger, M. Ibaraki, and J. D. Hughes,
Harbaugh, A. W., MODFLOW-2005, The U.S. Geological Survey Modular MODFLOW–USG Version 1: An Unstructured Grid Version of MODFLOW for
Ground-Water Model—the Ground-Water Flow Process: U.S. Geological Survey Simulating Groundwater Flow and Tightly Coupled Processes Using a Control
Techniques and Methods 6-A16, Reston, Virginia, 2005, p. 253. Volume Finite-Difference Formulation: U.S. Geological Survey Techniques and
Harbaugh, A. W., R. B. Edward, M. C. Hill, and M. G. McDonald, Methods, Book 6, Chap. A45, U.S. Geological Survey, Reston, VA, 2013, p. 66.
MODFLOW-2000, The US Geological Survey Modular Ground-Water Model: RockWare, RockWorks15. RockWare Incorporated, Golden, CO, 2010,
User Guide to Modularization Concepts and the Ground-Water Flow Process, p. 316.
US Department of the Interior, US Geological Survey Open-File Report Rumbaugh, O. J. and O. D. Rumbaugh, Groundwater Vistas User’s Manual,
00-92, Reston, Virginia, 2000, p. 130p. Environmental Simulations, Reinholds, PA, 2001, p. 258.
Harrouni, K. El., D. Ouazar, G. A. Walters, and A.H-D. Cheng, Scheerlinck, K., V. R. N. Pauwels, H. Vernieuwe, and B. De Baets,
“Groundwater optimization and parameter estimation by genetic algorithm “Calibration of a water and energy balance model: recursive parameter esti-
and dual reciprocity boundary element method,” Engineering Analysis with mation versus particle swarm optimization,” Water Resources Research, 45:
Boundary Elements, 18: 287–296, 1996, doi: 10.1016/S0955-7997(96)00037-9. W10422, 2009, doi: 10.1029/2009WR008051.
Hilchie, D. W., Advanced Well Log Interpretation, D. W. Hilchie, Golden, Schlumberger, Log Interpretation: Volume I—Principles, Schlumberger,
CO, 1982. New York, NY, 1972.
Hora, S. C., “Aleatory and epistemic uncertainty in probability elicitation Senge, R., S. Bösner, K. Dembczyński, J. Haasenritterb, O. Hirsch,
with an example from hazardous waste management,” Reliability Engineering N. Donner-Banzhoff, and E. Hüllermeier, “Reliable classification: learning
and System Safety, 54: 217–223, 1996, doi: 10.1016/S0951-8320(96)00077- classifiers that distinguish aleatoric and epistemic uncertainty,” Information
4. Science, 255: 16–29, 2014, doi: 10.1016/j.ins.2013.07.030.
Hsieh, P. A. and J. R. Freckleton, Documentation of a Computer Program to Servan-Camas, B. and FT-C. Tsai, “Saltwater intrusion modeling in hetero-
Simulate Horizontal-Flow Barriers Using the U.S. Geological Survey’s Modular geneous confined aquifers using two-relaxation-time lattice Boltzmann
Three-Dimensional Finite-Difference Ground-Water Flow Model, United States method,” Advances in Water Resources, 32: 620–631, 2009, doi: 10.1016/j.
Geological Survey Open-File Report 92-477, Sacramento, CA, 1993, p. 37. advwatres.2009.02.001.
Jiang, Y., C. Liu, C. Huang, and X. Wu, “Improved particle swarm algorithm Servan-Camas, B. and FT-C. Tsai, “Two-relaxation-time lattice Boltzmann
for hydrological parameter optimization,” Applied Mathematics and method for the anisotropic dispersive Henry problem,” Water Resources
Computation, 217: 3207–3215, 2010, doi: 10.1016/j.amc.2010.08.053. Research, 46: W02515, 2010, doi: 10.1029/2009WR007837.
Jones, N. L., T. J. Budge, A. M. Lemon, and A. K. Zundel, “Generating Smith, T. J. and L. A. Marshall, “Bayesian methods in hydrologic modeling:
MODFLOW grids from boundary representation solid models,” Ground a study of recent advancements in Markov chain Monte Carlo techniques,”
Water, 40: 194–200, 2002, doi: 10.1111/j.1745-6584.2002.tb02504.x. Water Resources Research, 44: W00B05, 2008, doi: 10.1029/2007WR006705.
Jones, N. L., J. R. Walker, and S. F. Carle, “Hydrogeologic unit flow charac- Solomatine, D. P., Y. B. Dibike, and N. Kukiric, “Automatic calibration of
terization using transition probability geostatistics,” Ground Water, 43: 285– groundwater models using global optimization techniques,” Hydrological
289, 2005, doi: 10.1111/j.1745-6584.2005.0007.x. Sciences Journal, 44: 879–894, 1999, doi: 10.1080/02626669909492287.
Karpouzos, D. K., F. Delay, K. L. Katsifarakis, and G. de Marsily, “A multi- Tomaszewski, D. J., Distribution and movement of saltwater in aquifers in
population genetic algorithm to solve the inverse problem in hydrogeology,” the Baton Rouge area, Louisiana, 1990–92, Louisiana Department of
Water Resources Research, 37: 2291–2302, doi: 10.1029/2000WR900411, 2001. Transportation and Development: Water Resources Technical Report No. 59,
Kavetski, D., G. Kuczera, and S. W. Franks, “Bayesian analysis of input Baton Rouge, LA, 1996, p. 50.
uncertainty in hydrological modeling: 1. Theory,” Water Resources Research, Tsai, FT-C., N-Z. Sun, and WW-G. Yeh, “Global-local optimization for
42: W03407, 2006, doi: 10.1029/2005WR004368. parameter structure identification in three-dimensional groundwater model-
Kerr, D. R. and L. A. Jirik, “Fluvial architecture and reservoir compartmen- ing,” Water Resources Research, 39: 1043, 2003, doi: 10.1029/2001WR001135.
talization in the Oligocene Middle Frio formation, south Texas,” Gulf Coast Voss, C. I. and A. M. Provost, SUTRA: a model for saturated-unsaturated,
Association of Geological Societies, Transactions, 40: 373–380, 1990. variable-density ground-water flow with solute or energy transport, US
Lemon, A. M. and N. L. Jones, “Building solid models from boreholes and Geological Survey Water Resources Investigation Reports 02-4231, Reston, VA,
user-defined cross-sections,” Computers and Geosciences, 29: 547–555, 2003, 2010, p. 300.
doi: 10.1016/S0098-3004(03)00051-7. Waterloo, Visual MODFLOW v. 4. User’s Manual, Waterloo Hydrogeologic,
Lin, H-C. J., D. R. Richards, G-T. Yeh, J. R. Cheng, H. P. Cheng, and N. L. Ontario, Canada, 2005, p. 654.
Jones, FEMWATER: A Three-Dimensional Finite Element Computer Model for Winston, R. B., ModelMuse—a graphical user interface for
Simulating Density-Dependent Flow and Transport in Variably Saturated MODFLOW–2005 and PHAST, Techniques and Methods 6–A29, U.S.
Media: Technical Report CHL-97-12, prepared for U.S. Army Environmental Geological Survey, Reston, VA, 2009, p. 52.
Center, Vicksburg, MS, 1997, p. 142. Zhang, Y., D. Gallipoli, and C. E. Augarde, “Simulation-based calibration of
Liskov, B., M. Day, M. Herlihy, P. Johnson, G. Leavens, R. Scheifler, and W. geotechnical parameters using parallel hybrid moving boundary particle
Weihl, Argus Reference Manual, MIT Laboratory for Computer Science, swarm optimization,” Computers and Geotechnics, 36: 604–615, 2009, doi:
Arlington, VA, 1987, p. 165. 10.1016/j.compgeo.2008.09.005.
Matott, L. S. and A. J. Rabideau, “Calibration of subsurface batch and Zhang, Y. and C. A. Sutton, Quasi-Newton Methods for Markov Chain
reactive-transport models involving complex biogeochemical processes,” Monte Carlo, Advances in Neural Information Processing Systems, edited by
Advances in Water Resources, 31: 269–286, 2008, doi: 10.1016/j.advwa- J. Shawe-Taylor, R. S. Zemel, P. L. Bartlett, F. Pereira and K.Q. Weinberger. 24,
tres.2007.08.005. Curran Associates, Granada, Spain, 2011, pp. 2393–2401.
48_Singh_ch48_p48.1-48.8.indd 8 8/22/16 2:17 PM

Chapter 49
Watershed Runoff, Streamflow
Generation, and Hydrologic
Flow Regimes
BY
FRED L. OGDEN AND GLENN S. WARNER
ABSTRACT which is known as ecohydrology. Although statistical estimation of peak runoff

can be straightforward given data from stream gauging stations with a substan-
This chapter is intended to educate the reader regarding processes that pro-
tial period of record, prediction of the variation of runoff, and water quality
duce flowing water on the surface of the earth. It emphasizes drivers that
over time as a function of land use changes, biological factors, and climate
influence these processes, namely: climate, weather, groundwater, and
change is much more complex. Difficulties arise because these changes intro-
includes some discussions on the mathematical methods used to predict
duce nonstationarities that invalidate common statistical approaches, and pre-
runoff generation. The relative importance of physical system properties such
dictions require an in-depth understanding of dynamic processes, ecological
as topography, land-cover, and soils is presented in the context of the different
connections, and feedback involved in runoff generation. Hence, the emphasis
models of runoff generation, and evidence from the literature is used to draw
in this chapter is on understanding fundamental processes that cause runoff to
conclusions on those important factors. Some classic examples from the lit-
occur, not just being able to quantify peak runoff rate.
erature are shown to illustrate longstanding unsolved problems in runoff
Given the strong links between climate and floods and droughts, we can
generation. The chapter concludes with a discussion of those poorly under-
expect water resources to change with climate changes, particularly with
stood aspects of runoff generation that are ripe for further study.
increasing human population and associated demands for fresh water
(Huntington, 2006; Vörösmarty et al., 2013). Furthermore, human-driven
49.1 INTRODUCTION land use changes can produce significant impacts on runoff generation and
the hydrological cycle. Often, land use, population, and climate change effects
This chapter deals with surface water, not groundwater. Surface water is spe- are confounded. It is difficult to separate the natural influence of climate
cial. Any water can be groundwater (GW) as the downward pull of gravity is variation from the effects of land use and water-use changes, particularly
omnipresent, and porous media are ubiquitous on the Earths’ surface. against the backdrop of anthropogenic climate change that is occurring
Something special must happen for water to become surface water. Either it (United States National Academy of Sciences, 2014).
must be returned to the surface from GW, or something must have prevented The accuracy of water-quality predictions with changing inputs of both water
this water from becoming GW once it fell as rain, or melted from soil ice, hail, and chemical or biological agents depends on the physical processes and flow path
graupel, snow, or ice. The flow path can affect the geochemical makeup and that water follows to the point of interest. Accurate predictions of the amount and
ratios of light stable isotopes in surface water, which can serve as a diagnostic. timing of flow, as well as flow path, are prerequisite to identification of sources,
Surface water is inherently more useful than GW because it is more easily concentrations, and the ultimate fate of chemicals, sediment, and biological mate-
accessed and quantified, and little to no pumping is required to use it. It is also rials in watersheds. Predictions of runoff in response to precipitation and other
inherently more dangerous because of the hazards posed by extreme hydro- forcings, such as evapotranspiration (ET) in the context of varying pre-event or
logic events, such as floods and droughts, which lead to dangerous excesses or antecedent conditions, requires understanding of mechanisms/processes within
shortages and associated negative socioeconomic consequences. the system that are generating runoff and streamflow over a range of spatial and
Runoff is loosely defined as water from rain or snow melt that flows over the temporal scales. Our understanding of the important factors that affect runoff
surface and into streams. In this chapter, we strictly define runoff as the vol- generation is limited (Beven and Germann, 2013). There are factors that we know
ume of water (L3) that flows from a given area (L2). Therefore, runoff has the are important, but we do not understand their effects on the physical mechanisms
units of depth of water (L). The rate of runoff has units of (L T–1). Runoff is a that control runoff generation to include them in a quantitative way in our models
convenient metric because it allows comparison of hydrological response per without introducing a large amount of uncertainty into the models as well. This is
unit area. Engineers and hydrologists predict the amount or rate of runoff in one reason why some claim that parsimony or stinginess in model parameters is
response to real or hypothetical water input events. We use the generic term desirable. However, the model complexity and hence the number of model
“water input” because water can be introduced from a variety of sources such parameters depends on the question being asked and parsimony should not inter-
as: rainfall, melting snow and ice, thaw of frozen soils, and irrigation. fere with predictive ability afforded by additional process descriptions or param-
Hydrologic design requires the peak runoff rate for a given flood probability eters when possible. Some seemingly parsimonious models have been
for sizing structures or establishing flood plain limits. Increasingly, predictions demonstrated as inadequate in certain circumstances and there are important
of runoff over a season or longer are required, including the changes in the tim- differences between conceptual and physics-based models (Schoups et al., 2008).
ing and shape of the hydrographs for assessments of water availability, in-stream
flows for aquatic habitat or for water rights considerations. In the larger context, 49.1.1 Runoff Ratio
the varied factors that affect runoff generation often depend on human activi- Surface water results from the process of runoff generation. The primary
ties as expressed by land use, water use, and by b iological factors, the study of objective of this chapter is to describe the factors that cause runoff, which
49-1
49_Singh_ch49_p49.1-49.12.indd 1 8/22/16 2:20 PM

49-2 Watershed Runoff, Streamflow Generation, and Hydrologic Flow Regimes
include a number of theories that are complicated by poorly understood Table 49.1 Hydroclimatic Definitions as a Function of Aridity Index
factors known to be important, but that do not have determinant m athematical
Aridity index range Hydroclimatic regime
descriptions. This is one of the reasons why the study of runoff generation
processes and mathematical process descriptions is an active and compelling
area of research. 12 > φ >= 5 Arid
Runoff generation is a nonlinear function of water input and soil proper- 5 > φ >= 2 Semiarid
ties, and is often discontinuous in space and time. This makes analytical 2 > φ >= 0.75 Subhumid
predictions of runoff difficult except in simple situations, such as from small
impervious areas over short time periods. Surface runoff sometimes follows a 0.75 > φ >= 0.375 Humid
tortuous path, involving multiple exchanges between surface and GW. Surface Source: Ponce et al., 2000.
flow paths include diffuse overland flow, and concentrated flow in rills,
gullies, and small streams before reaching a river.
One way to quantify runoff generation is the runoff ratio ρ, which repre-
intermediate ρ values because of high PET. The northwest aspect of the
sents the fraction of basin-average precipitation P that becomes surface
Appalachian mountains in the mid-Atlantic region has values of ρ > 0.6 that
runoff Rs:
extend through Pennsylvania and upstate New York and into New England
because of the increasing amount of annual precipitation that falls as snow,
Rs and reduced PET. Southeast of the Appalachian mountains, annual ρ values
ρ= (49.1)
p are generally between 0.25 and 0.35 because of high PET and a lower percent-
age of annual precipitation falling as snow.
Figure 49.1 shows average annual ρ computed using data up to 2013 from the Watershed area is the dominant physiographic factor affecting runoff
United States. Geological Survey (USGS) Historical Climatological Data behavior (Gupta and Dawdy, 1995). However, the aridity index f (Budyko,
Network (HCDN) stream gauges (Slack et al., 1993), for watersheds where the 1974), which is defined as the ratio of annual PET to annual precipitation, is
USGS computes annual mean areal precipitation (AMAP). HCDN stream another important climate related factor:
gauge data come from watersheds with minimal human impacts on flows. The
ρ values shown in Fig. 49.1 illustrate some very important spatial variations in PET
φ= (49.2)
runoff generation over the continental United States. at the annual time scale. P
The first observation is the region of low ρ (<0.10) that extends from south
Texas to North Dakota. Immediately to the east of this region, ρ increases Ponce et al. (2000) provided hydroclimatological definitions (Table 49.1) as a
toward 0.3. To the west of this region, in the high plains to the front range of function of the aridity index (f).
the Rockies, are a large number of ephemeral rivers in catchments that have ρ Figure 49.2 shows a plot of the aridity index (f) vs. drainage area (km2) for
values near 0 in most years, hence no dots. Three other pronounced regions the HCDN stations used in this analysis, assuming annual PET is 1400 mm
of low ρ values occur in peninsular Florida, south Texas, and in the south- for all watersheds. Also shown on Fig. 49.2 are the aridity index boundaries
western continental United States. The low ρ values in Florida occur because for different hydroclimatic regimes. Figure 49.2 shows that most of the sta-
despite large annual rainfall amounts, much of the rainfall contributes to GW tions used in the analysis are in the subhumid hydroclimatic regime, and have
flows to the ocean and does not become stream flow. In south Texas, the geol- drainage areas from 10 to 10,000 km2.
ogy is such that most streams lose flow to aquifers as they flow toward the gulf Drainage areas represented in the data shown in Fig. 49.3 range from 1.5 to
coast. In the southwestern United States, annual precipitation amounts are 25,600 km2 (Fig. 49.2). The mean and median length of record for these
small (<500 mm), and potential evapotranspiration (PET) is very high—two stations are 62.5 and 61 years, respectively, with a standard deviation of
factors that tend to combine to limit runoff generation. 18.5 years. The longest period of record is 156 years, and the shortest is
The highest ρ values exceed 0.8 in the Pacific northwest because of high 24 years. Figure 49.3a shows the annual average ρ versus AMAP. Figure 49.3a
annual rainfall and reduced PET due to cloud cover and reduced tempera- shows that ρ has a large variance for AMAP < 1600 mm. For AMAP >
tures during much of the year. High ρ values are also observed in portions of 1600 mm, ρ values generally exceed 0.5 with few exceptions.
western mountain ranges including the Cascade, Sierra Nevada, and northern Figure 49.3b shows annual precipitation minus annual runoff versus
and central Rocky Mountains, due to the dominance of high-altitude winter AMAP. The difference between annual precipitation and runoff in a water
snow accumulation in amounts that greatly exceed PET. Relatively rapid melt balance context represents losses to ET or changes in storage. In this
also reduces the opportunity for evaporation losses. analysis, each point in Fig. 49.3 represents the average of the period of
Despite receiving some of the highest amounts of rainfall in the continental record for each gauging station. Because of the length of record at most
United States, the gulf coast from east Texas to the Florida panhandle has stations, changes in storage should be nearly zero; and changes in storage
Runoff ratio
0.0−0.1
0.1−0.2
0.2−0.3
0.3−0.4
0.4−0.5
0.5−0.6
0.6−0.7
0.7−0.8
0.8−0.9
0.9−1.0
Figure 49.1 Mean annual runoff ratio for approximately 600 stream gauging stations from the USGS HCDN for which AMAP estimates are
available in the continental United States. [Source: USGS.]
49_Singh_ch49_p49.1-49.12.indd 2 8/22/16 2:20 PM

Dominant Runoff Generation Mechanisms 49-3
49.2 DOMINANT RUNOFF GENERATION MECHANISMS
This section discusses the mechanisms that determine runoff at the time scale
6 of individual precipitation or water input events. There is a continuum of
runoff generation mechanisms depending on limiting values of the parameter
Arid i Ks−1, where i is the maximum rain rate (L T−1) that persists for a time and Ks
is the saturated hydraulic conductivity (L T−1) of the soil. If this parameter is
Aridity index φ (−)
near 0, it is likely because of consistent low rain rates, such as those experi-
4 enced in cloud forests, or because of very large values of Ks such as those
found in the sand hills of Nebraska, or for many undisturbed, well-vegetated
areas. When i Ks−1 is near 0 then subsurface stormflow will dominate the
response of a catchment. Whether or not this subsurface stormflow causes
Semi saturation at the soil surface and direct runoff will depend upon the ratio of
arid the subsurface capacity to convey water relative to the input flux at a point.
2
One parameter that describes this likelihood is the Topographic Index (Beven
Sub and Kirkby, 1979). Subsurface stormflow is one end-member of the runoff-
humid generation continuum. Saturation excess runoff generation (Dunne and
Black, 1970) is a special case of subsurface stormflow, where the capacity of
Humid the subsurface to convey water becomes less than the flux at a locus of points
0 along a hillslope. When the cumulative amount of water input exceeds the
100 101 102 103 104 105
available soil storage such that all pore space is saturated, the runoff-
Watershed area (km2) generation mechanism is called saturation excess. In some cases when satura-
Figure 49.2 Aridity index versus catchment area for 633 HCDN stations. Data tion excess runoff-generation occurs, GW can flow back onto the land
courtesy of USGS. The ρ values and data used to calculate them are plotted in Fig. 49.3 surface in lower topographic regions, a process called exfiltration or return
against annual mean areal precipitation (AMAP). Data source: USGS HCDN stations flow. This water becomes surface runoff at seepage faces and fills surface
throughout the continental United States, Alaska, Hawaii, and Puerto Rico.
depressions.
At the other end of the runoff-generation continuum are cases where the
parameter i Ks–1 > 1 and the infiltration-excess mechanism (Horton, 1933) can
should be small compared to total amounts of rainfall, streamflow, and ET occur. Whether or not it does indeed occur depends upon a host of factors
over the long period of record. Therefore, Fig. 49.3b represents actual ET including the time series and duration of precipitation, antecedent soil moisture
plus losses to deep GW that underflows the site of the stream gauging conditions, and other factors such as microtopography, vegetation, etc.
station. Losses to deep GW have been documented, and Schaller and Fan In between these two limiting cases of: i Ks–1 ≈ 0 and i Ks–1 > 1, are a variety
(2009) called these GW exporting basins. Interbasin GW exchanges of other behaviors that complicate evaluation of the runoff generation. Most
identified by Schaller and Fan (2009) were primarily a function of geology, often, these are a function of soil heterogeneity or layering. We call those factors
and to a lesser extent, basin size, with smaller basins more likely to gain or “profile-related” runoff-generation mechanisms. The following sections discuss
lose GW than larger ones. the infiltration-excess, saturation-excess, and profile-related runoff generation
Using data from nested stream gauges on the Contoocook River in New mechanisms.
Hampshire, Dingman (2015, p. 440) showed that GW flow either as under-
49.2.1 Overland and Channel Phases of Runoff Generation
flow through the alluvial aquifer at the gauge site or as general recharge that
leaves the surface watershed as GW can be significant. Defining underflow as Runoff–generation processes can be divided into “land” and “channel” phases.
G/Q, where G and Q represent the underflow and measured surface discharge Runoff in the land phase is conceptualized as diffuse or non-concentrated
values, respectively, Dingman (2015) reported underflow values of 0.35, 0.18, flow over a plane or hillslope. The channel phase consists of topographically
0.19, and 0.07 for nested watersheds having drainage areas of 176, 953, 1106, confined or channelized flow. Channelized flow occurs over a large range of
and 1984 km2, respectively. This finding supports the notion that underflow scales: from rills, small ephemeral gullies, and small tributaries/first-order
decreases with increasing catchment size. streams up to large rivers. Channelized flow is most often measured because
The quantity of actual ET in Fig. 49.3b is tightly constrained to the amount the shallow and transient nature of overland flow makes it hard to measure.
of precipitation for AMAP less than about 1400 mm. Beyond this, actual ET Direct runoff is defined as runoff of water that never entered the soil or GW
is limited to about 1400 mm, except in the case of a small number of outliers. domain before reaching a channel. The amount of direct runoff is primarily
Note that 1400 mm year−1 corresponds to about 3.8 mm day-1. controlled by the land phase of runoff generation where “losses”, such as inter-
ception, infiltration, and evaporation reduce rates of water input and conse-
quently reduce the amount of water that becomes runoff.
1.0 At larger scales, the channel phase of runoff generation dominates the
(a) runoff rate and shape of the hydrograph due to attenuation of the discharge
through in-channel diffusive processes, bank storage, and attenuation of
0.5 flood peaks by overbank flow. Changes in runoff volume within channel flow
are usually small compared to the infiltration losses during the land phase,
Average annual runoff ratio (−) although seepage out of the channel (losing stream) or into the channel (gain-
ing stream) do occur depending on the geology and climate as well as the
0.0 4000
(b) Precipitation – runoff (mm) permeability of nonalluvial aquifers and depth to water table. For the most
3000 part, the effects of channel phase processes are better understood compared
to land phase processes.
2000
The attenuation of discharge with increasing area is shown in Fig. 49.4 (b)–(e)
1000 (modified from Dunne and Leopold, 1978). As the drainage area increases in
4000 0 the case of nested watersheds, the lag time to peak increases, and the peak
(c) Mean annual runoff (mm) runoff rate decreases. These changes can be attributed to diffusive attenuation
3000 as the flood wave moves downstream. The spatial extent of rainfall may have
2000 been a factor, but the absolute peak discharges shown on the right hand-side of
each hydrograph figure increases with the drainage area, but with a scaling
1000 exponent of 0.59. Comparing watersheds of different sizes, scaling exponents
0 less than 1.0 are the rule, and indicate that peak flows do not increase linearly
0 1000 2000 3000 4000 5000 with increasing area. Rather they increase at a rate less than linearly.
Annual mean areal precipitation (mm)
49.2.2 Subsurface Stormflow Runoff Generation
Figure 49.3 (a) Mean annual runoff ratio, (b) excess precipitation defined as mean
annual precipitation minus mean annual runoff, and (c) mean annual runoff, all plotted Knowledge of flow paths and residence times is essential for predicting the
against mean annual rainfall for 635 USGS HCDN stream gauges where mean annual biogeochemical and geomorphological processes and feedbacks in a water-
precipitation estimates are available. shed. Many different pathways exist for water to travel from upland areas to
49_Singh_ch49_p49.1-49.12.indd 3 8/22/16 2:20 PM

Runoff rate (cm h−1) 15 49.2.3 Saturation-Excess Runoff-Generation

Mechanism
(a) Total rain: 7.13 cm
10 Rainfall amounts in excess of the available pore space in the presence of a near-
surface water table can cause the water table to rise to the land surface. When
5 this happens, additional rainfall on those sites becomes direct runoff through
the saturation-excess runoff-generation mechanism. Saturation excess runoff-
0 generation is a special case of the subsurface stormflow mechanism where the
0.3 subsurface does not have the capacity to convey the flow at a point, which
(b) 0.52 km2
Qpeak causes the water table to rise up to the land surface. For this reason, the satura-
0.37 m3 s−1 tion-excess mechanism is sometimes called saturation from below.
Sklash and Farvolden (1979) and O’Brien (1982) wrote that precipitation
0.2 on soils near streams can cause a rapid water table rise known as groundwater
ridging, which can be considered a special case of saturation-excess runoff
generation. Groundwater ridging has been observed in riparian zones where
0.1 the capillary fringe is near the land surface at the beginning of a water input
event. However, this phenomena is less likely to occur in the cases of high
permeability soils (Zaltsberg, 1983), soils with preferential flow paths such as
Runoff: 1.38 cm
macropores (Beven and Germann, 1982; German, 1990; Youngs et al., 1996),
0 forested hillslopes (Buttle and Sami, 1992), and channels incised substantially
Qpeak
0.1 more than the height of the capillary fringe (Bonnell, 1993).
(c) 8.3 km2 1.92 m3 s−1
Runoff rate (cm h−1)
The setting for saturation-excess runoff generation is shown in Fig 49.6.

0.05
Runoff: 0.60 cm Preferential flow or percolation shown in Fig. 49.6 represents flow through
0 contiguous large diameter pore spaces in the medium called macropores.
Qpeak Macropores in some situations represent the dominant flow path through
0.05 5.35 m3 s−1 soils (Nimmo, 2012). Vertical macropore flow occurs in soils when the Bond
(d) 42.9 km2 Runoff: 0.35 cm number Bo, which is defined as the ratio of gravity to capillary forces at the
0
Qpeak pore scale exceeds approximately 0.05, or when the pore diameter exceeds
0.05
(e) 111 km2 8.50 m3 s−1 about 1.2 mm (Or, 2008). Biological actors, such as roots, microbes, micor-
0 Runoff: 0.23 cm rhyzea and fungi, and insects and abiotic processes such as freeze/thaw and
0 5 10 15 shrink/swell cycles routinely produce contiguous, interconnected pores or
Time (h) apertures that exceed this dimension. Water can enter macropores through
local ponding at the land surface in microtopography that funnels ponded
Figure 49.4 Changes in runoff amount and timing of nested watersheds in water (Trojan and Linden, 1992) or as pore water that is forced into macro-
saturation-excess Sleepers River watershed in Vermont. Panel (a) shows 15-minute
rainfall. Panels (b) through (e) show hydrographs at different scales. [Source: Modified
pores by positive pressure. Because macropores are so common in soils,
from Dunne and Leopold (1978).] (Weiler and Naef, 2003; Bastardie et al. 2005) preferential non-Darcian flow
is ubiquitous in undisturbed soils.
a stream. Figure 49.5 depicts a hillslope with a shallow soil profile in a

semihumid or humid hydroclimatic zone. During times of high water input 49.2.4 Infiltration Excess Mechanism
and low ET a near-surface water table can exist. Surface water, such as vernal In soils with reduced infiltration capacities, such as disturbed soils, surface
pools, seepage faces, and streamflow, are all in contact with the GW table. runoff is generated whenever the water input rate from rainfall, snowmelt,
Infiltration in upland areas may percolate through the soil to deeper geologic irrigation, or runon from upland areas exceeds the infiltration capacity of the
strata where it becomes GW. These deeper strata typically offer more resis- soil. This runoff generation mechanism is called infiltration excess.
tance to flow and provide a slower pathway to the stream, which is often the Infiltration excess runoff generation at the hillslope scale is shown in
source of baseflow in streams, that is, the persistent flow during periods Fig. 49.7. Runoff is produced when the water input rate exceeds the potential
without water input events to the upland areas. infiltration rate fp of the soil. The potential infiltration rate fp is transient, and
Figure 49.5 Hillslope-scale cross-section of potential pathways for water movement from upland areas
to streams for a humid or subhumid hydroclimatic regime, with soil profile. Note groundwater
contributions to stream flow.
49_Singh_ch49_p49.1-49.12.indd 4 8/22/16 2:20 PM

Dominant Runoff Generation Mechanisms 49-5
from above is sometimes used to describe this mechanism. A wetting front

(Zf in Fig. 49.8) develops and moves downward in the soil as the total amount
of infiltration increases as the water input continues. The wetting front in this
case can be quite sharp in that above the wetting front the soil is saturated,
with a very rapid decrease in water content across the front. The sharpest
wetting fronts are observed in coarse-textured soils (Hillel, 1971). The sharp
wetting front is the conceptual basis for the development of infiltration equa-
tions, such as the Green–Ampt (1911) method. However, in coarser soils, or
heterogeneous soils with regions of coarseness, the wetting front can become
unstable resulting in fingered flow—a jagged, nonuniform type of wetting
front (Parlange and Hill, 1976).
49.2.5 Profile-Related Infiltration Mechanisms

Most soils are heterogeneous and anisotropic. Therefore, critical soil parameters
are seldom uniform either vertically or laterally, even within well-defined lay-
ered soils. Two critical parameters are: porosity, that is, the amount of pore
space compared to solid particles, and hydraulic conductivity. Both of these
parameters are likely to decrease with depth from the surface of the soil under
natural conditions because of gravitational forces that cause soil consolidation.
This leads to decreasing permeability with depth in the soil.
There are several possible factors that can overcome this consolidation effect.
Figure 49.6 Conceptual drawing showing details of the saturation excess runoff Included in these are freeze/thaw, shrinkage cracking upon drying, biological
generation mechanism. Note that when water table rises to the land surface, infiltration
will stop.
activity or bioturbation, soil creep and downslope motion, etc. Freeze/thaw
processes as well as soil shrink/swell cause soil cracking to various depths.
While these cracks are open, sand, vegetation, and other detritus fall into them,
as discussed later depends upon the rainfall rate and the slope of the and create anisotropy that is often conducive to flow in the subsurface.
land surface. If the water input rate is less than the fp, then all water enters the Bioturbation is the process by which biological actors overturn the soil over a
soil; otherwise, the runoff rate at that point equals the difference between the long time span, and with decreasing intensity with increasing depth (Kirkby,
water input rate and fp. This mechanism is sometimes called Hortonian Runoff 1988). Tree fall is a dominant form of bioturbation. Burrowing animals are a
after Horton (1933) who was one of the first to document infiltration and dominant factor in temperate and arid regions (Yair, 1995), while in the tropics,
runoff from water inputs on agricultural lands. ants are a significant actor (Wilkinson et al., 2009). In soils that do not freeze
Two forces drive the vertical infiltration process: gravity and capillarity. earthworms and ants are responsible for up to 10–15 tonnes per hectare per
Capillarity is a very nonlinear function of the water content, going from year of soil movement (Madge, 1965; 1969). Niedzialek and Ogden (2012)
nearly zero near saturation, to very high values in the case of dry soils. Near identified rapid subsurface flow generating streamflow when the rainfall rate on
saturation, the hydraulic conductivity approaches its maximum value of satu- a 17 ha tropical catchment was at least 40 mm h–1 for 20 min, with no observed
rated hydraulic conductivity Ks, and decreases rapidly as the soil becomes less surface runoff except for some minor return flow. This despite the fact that the
than saturated. Hence, at the beginning of rainfall, capillarity provides most soils in the study catchment were primarily weathered clays.
of the drive, while as more and more rainfall infiltrates, hydraulic conductiv- Because of either natural layering, consolidation effects, or bioturbation,
ity limits the infiltration rate. In the battle between capillarity and hydraulic the infiltration rate may be limited at some depth in the soil. This can lead to
conductivity, the hydraulic conductivity is the most important parameter, saturated conditions and positive pore pressures developing at that depth,
because its effect is not only on the capillary driven flow, but also on the grav- creating a transient perched water table. There are two possible outcomes
ity driven flow component, whereas capillarity does not affect the gravity from this condition. The first is that the perched water table rises to the land
driven flow. Rainfall on the nearly impervious rock outcrops shown in surface and results in saturation-excess overland flow. The second condition
Fig. 49.7 is a limiting case of the infiltration excess mechanism where the ratio is that the positive pore pressures activate downslope macropores and soil
of the parameter i Ks–1 → • because the saturated hydraulic conductivity of pipes (Nieber and Warner, 1991; Jones 2010), leading to rapid subsurface
impervious surfaces is 0 or very nearly so. storm flow through preferential flow paths.
Conceptually, under the infiltration-excess mechanism, water completely We refer to this condition as a profile-related runoff mechanism. Unlike
saturates the soil as it moves vertically downward hence, the phrase saturation saturation excess runoff where the water table is created by bedrock or some
Figure 49.7 Infiltration excess runoff-generation mechanism at the hillslope scale in a semiarid region.
Note streamflow losses to groundwater. Infiltration excess runoff generation mechanism at the hillslope
scale in a semiarid region.
49_Singh_ch49_p49.1-49.12.indd 5 8/22/16 2:20 PM

Figure 49.8 Schematic drawing of infiltration excess runoff generation mechanism. Note: Zf is the depth
to the wetting front.
other impeding layer, this profile-related mechanism is temporary and can be simulated period was only 7 mm. The finite water-content method of Ogden
rain rate dependent. Because significant infiltration capacity remains in the et al. (2015) is equivalent to Richards (1931) equation minus a term which is
impeding layer, the transient perched water table will dissipate soon after the equal to the diffusive flux of capillary head divided by the slope of the wetting
end of precipitation or water input. This mechanism was suggested by Ogden front: D(θ)(∂2θ/∂θ 2) (∂θ/∂ θ)–1 (Seo et al., in review). Note that the diffusive
et al. (2013) as the primary cause of differences in runoff behaviors between flux has a zero mean and therefore does not affect the net flux, and that the
two adjacent tropical watersheds with identical geomorphology and soils, but slope of the wetting front is infinite for a sharp front, which is common in
with different land uses. hydrologically important coarser soil textures, which renders this term
unimportant.
Beven and Germann (2013) raised important questions on the adequacy of
49.3 INFILTRATION EQUATIONS
Darcy’s law and Richards (1931) equation in describing flow phenomena in
Given the importance of infiltration as it affects the runoff generation mecha- macroporous soils. They suggest that a viscosity based Stokes law approach
nism, it is perhaps a bit surprising that there is not a widely accepted solution has merit to describe film flow in pores. The finite water-content solution
method to calculate infiltration that is generally applicable. Dozen of infiltra- method (Ogden et al., 2015) is guaranteed to converge and to conserve mass,
tion calculation methods and several laws have been proposed since 1911. and it has the potential to simulate non-Darcian flow in a portion of the
Some of them are empirical equations that allow fitting of a mathematical porous medium. The need to develop methods to characterize the number,
equation to observed infiltration data. Those methods are not discussed here. size distribution, and location distribution of macropores in soils is apparent.
The most fundamental developments, in order of their development were: Developing methods to accurately simulate flows through macroporous soils
Green and Ampt (1911), Richards (1931), Philip (1957a; 1957b), Parlange et al. is similarly important.
(1982), Haverkamp et al. (1990), Celia et al. (1990), and Ogden et al. (2015).
The Green–Ampt (1911) method assumed a delta-function soil water 49.4 FACTORS AFFECTING RUNOFF
diffusivity, homogeneous soil, and uniform initial water content profile.
Green and Ampt (1911) derived their method from saturated flow We have posited that the dominant type of runoff generation mechanism at a
conservation of mass using a Lagrangian description of wetting front particular point, hillslope, or catchment is a function of a combination of
dynamics with gravity and capillary-driven flow. However, the Green and climatological, physiographical, and biological factors. Because of our poor
Ampt method was put into an applicable formulation by Mein and Larson understanding of biological factors and how they affect runoff, we focus here
(1973). The Richards (1931) equation is the standard by which all other on climatic and physiographic factors.
methods are compared. It is an infiltration law that is valid assuming that the
flow of water through the soil is dominated by capillarity and gravity, in that 49.4.1 Climatic, Pedologic, and Vegatative Factors
order. A general numerical solution of the Richards (1931) equation was not Climate, pedologic, and vegetative effects on runoff include: (1) the storm
published until Celia et al. (1990). However, this numerical solution, while total precipitation Ps, peak rainfall intensity i, and duration tr of rainfall of a
general, is not guaranteed to converge or to conserve mass. The Richards given frequency, (2) soil development over many millenia, (3) the variety and
(1931) equation has been criticized as overemphasizing capillarity (Germann, density of vegetation, and (4) the magnitude of soil water losses by ET.
2010), and for its general computational expense, particularly when simulating Furthermore, the largest effect of climate may be on potential infiltration
sharp fronts (Tocci et al., 1997). Another criticism of the numerical solution rates. Dunne (1978) showed that a saturation-excess runoff is considerably
of Richards (1931) equation is that its performance in general soil-water more likely to occur in a humid climate, while infiltration-excess runoff is
dynamics modeling is unpredictable and computationally expensive with more likely in arid regions. This factor can be at least partially explained by
widely varied soil constituative relations (Short et al., 1995). the aridity index f. Canopy cover on well-vegetated soils prevents raindrop
Philip (1957a; 1957b) made mathematical advances under specific initial impact on bare soils with associated surface sealing and loss of infiltration
and boundary conditions towards analytical solution of Richards (1931) capacity. Arid environments typically have lower plant densities with land
equation. The true utility of those solutions is limited by those required surfaces exposed to the rain drop impact and surface sealing by wind-
assumptions. Parlange et al. (1982) derived a three parameter infiltration deposited clay particles. Some of the highest runoff rates and therefore ero-
equation that encompasses both the Green and Ampt (1911) solution, and the sion rates result from Hortonian runoff due to bare or sparse vegetated cover.
Talsma and Parlange (1972) limit of a rapidly increasing soil water diffusivity
and unsaturated hydraulic conductivity function. Ogden et al. (2015) derived 49.4.2 Physiographic and Hydraulic Factors
a novel “finite water-content” infiltration solution method. Compared to the At the hillslope scale, we assume that the dominant factors are a function of
numerical solution of Richards (1931) equation in a numerical test involving geomorphology and soil hydraulic characteristics. In terms of geomorphol-
2600 mm of rainfall on a loam soil with a fixed water table at 1 m below the ogy, the hillslope length L and plan-view radius of curvature of the elevation
land surface, the difference in cumulative infiltration over the 8 month contours R, as well as the land-surface slope So are important. In terms of soil
49_Singh_ch49_p49.1-49.12.indd 6 8/22/16 2:20 PM

Poorly Understood Factors Affecting Runoff Generation 49-7
The factor f3 is the aridity index previously defined (Ponce et al., 2000):
104
Rainfall PDFs f3 = f.(49.6)
Northern Mississippi The factor f4 is the ratio of the storm total precipitation ps to the average pore
103 Central Panama space available to store water, which is calculated from the soil porosity n,
Gravel
average initial water content θi and thickness of the vadose zone D:
P
Clean sand
f4 = (49.7)
102 (n − θi )D
Ks , i (cm h−1)
The factor f5 is the slope plan view concavity, where R is the plan-view radius
of curvature of the elevation contours, and L is the slope length.:
Clay with macropores

101 R
f5 = .(49.8)
Peat L
Fine sand
The sixth and final factor f6 is the ratio of the average rainfall flux during a
100
precipitation event LP/tr, with tr equal to the rainfall duration, divided by the
Silt loam/loess
ability of the soil to convey water downslope KsSoD (Ogden and Watts, 2000):
LP
10−1 f6 = .
(49.9)
tr K s So D
Clay
Figure 49.10 shows a Bayesian synthesis of runoff generating factors
expressed as dimensionless quantities. The distribution of points for f4 and f6
10−2 were informed from results in the papers by Ogden and Watts (2000) and
0.0 0.5
Niedzialek and Ogden (2004), respectively. The other distribution functions
P(i) were estimated based on physical understanding of the system. The abscissa
Figure 49.9 Ranges of saturated hydraulic conductivity values for on each plot ranges from 0 to 1 with 0 indicating subsurface stormflow
common soil textures together with rainfall probability density functions dominated, and 1 indicating infiltration excess dominant. The spread of the
from northern Mississippi and central Panama. [Source: adapted from data points are indicative of the uncertainty in the effect of a particular
Freeze (1972).] dimensionless parameter value on a particular runoff generation mechanism.
Note that χ = 0 does not necessarily indicate the occurrence of saturation
properties, the thickness D, porosity n, average initial water content θi, and
excess runoff, as it is a special case of subsurface stormflow where the topog-
the mean saturated hydraulic conductivity of the soil, Ks, are important.
raphy and/or soil variability promotes exfiltration and/or the water table
Laboratory analyses performed using packed disturbed columns to mea-
intersecting the land surface.
sure Ks can demonstrate a wide range of values. When infiltration capacity is
Figure 49.11 shows results of a principal component analysis (PCA) of the
measured in situ or using intact columns, Ks and the infiltration rate are much
distributional means of the points shown in Fig. 49.10. This revealed that
less dependent on soil texture than on recent disturbance and vegetation
concavity R L–1 and slope So are colinear but acting in opposite directions,
conditions. An undisturbed, vegetated soil can exhibit much greater infiltra-
while both are orthogonal to the runoff vector, giving great significance to
bility than a tilled sandy textured soil. Although changes in soil structure play
these two parameters. The parameter i Ks–1 and the transport parameter:
a role, the major differences in Ks are due to the presence or lack of macro-
L P tr–1 Ks–1 So–1D–1 are equivalent, but are conceptually different. This differ-
pores. Figure 49.9 shows ranges of Ks for some typical soil types along with
ence arises from the fact that these two terms operate at time scales. The term
probablity distribution functions for rainfall calculated using data from
iKs–1 operates at the time scale of individual rainfall events while the term
northern Mississippi (temperate zone) and central Panama (seasonal tropics).
LPtr–1Ks–1So–1D–1 is more meaningful at longer time scales.
Figure 49.9 provides an indication of the likelihood of infiltration excess
The affects of the aridity index PET P–1 and the ratio of storm total rainfall
runoff generation for several soils and for clay with macropores.
Ps to soil available storage capacity (n-θi)–1 D–1 are similar, but they also work
When coupled with disturbances such as tillage, compaction from human
at very different time scales. These two dimensions of the PCA shown in
activities, etc., wide ranges of Ks can be expected, even over 2 or 3 orders of
Fig. 49.11, describe 85% of the variance in the runoff dynamics expectation.
magnitude. For example, data from Starr (1990) indicate ranges of Ks from
This exercise demonstrates the equivalence of similar terms that operate on
40 to about 110 mm hr–1, and 10 to about 220 mm hr–1 during a summer
different time scales, and illustrates the complexity of the multidimensional
season for a field under conservation tillage and plow tillage, respectively.
hydrologic response surface. Given this complexity across scales, we should
Gupta et al. (2006) found a range in Ks from about 0.1 to over 5 × 10−5 m s−1
not expect a lumped one- or two-parameter model to correctly predict runoff,
in a vertisol soil over a 2-year period. Burwell et al. (2011) showed that as
especially when the parameters are static in nature.
newly seeded Bermuda grass emerged and grew over a 3-month period, run-
off rates decreased by a factor of 5 or more during 2 different years.
49.5 POORLY UNDERSTOOD FACTORS AFFECTING
49.4.3 Dimensionless Parameters RUNOFF GENERATION
From the previously described factors that were identified as important, we In this section, we discuss factors that are known to be important in runoff
can hypothesize some functions that might help describe whether or not a generation, including tillage, land use/land cover, soil wetness, slope, and
particular hillslope will predominantly produce subsurface stormflow, or infil- rainfall intensity. Some of these factors have been recognized as important, at
tration excess (surface) runoff as two end-members of the previously discussed least in a qualitative way. However, the actual mechanisms/processes of how
continuum of runoff generation mechanisms, which we call χ. In this analysis, they affect runoff generation are poorly understood or have so far escaped
P varies from 0 (subsurface stormflow) to 1 (infiltration excess) mechanisms. quantitative description.
We assume the following:
49.5.1 Agricultural Effects of Tillage Practices
X = f1 f 2 f3 f 4 f5 f6 (49.3)
. Plants and plant residue affect fp in two major ways: (1) they provide protec-
tion to the soil surface from raindrop impact and (2) they provide mecha-
Where the factor f1 is the ratio of peak rainfall rate to the soil saturated nisms for macropore development. The macropore development may be
hydraulic conductivity (Saghafian et al., 1995): direct, for example, in the form of root channels, or indirect due to the
i formation of organic matter which in turn promotes soil fauna activity (e.g.
f1 = . (49.4) earthworms, ants, beetles, etc.) which promotes macropore development.
Ks The plant canopy of a well-vegetated soil surface intercepts rainfall and
The factor f2 is the slope So: provides protection of the soil from surface sealing. The potential for soil
sealing due to raindrop impact is highly associated with the amount and
f 2 = So (49.5)
. type of clay in the soil. When the surface is exposed, large raindrops having
very high kinetic energy will dislodge and disintegrate larger aggregates of
49_Singh_ch49_p49.1-49.12.indd 7 8/22/16 2:20 PM

Peak rain rate/soil conductivity Slope
Surface
Dominant flow path
Subsurface
0 2 4 6 8 10 0 10 20 30
i KS−1 So (percent)
Aridity index Storm total rain/soil storage capy.

Surface
Dominant flow path
Subsurface
0 5 10 10–2 100 102 104

PET P−1 P (n− q i)–1 D –1
Slope concavity Rain flux/subsurface transp. capy.

Surface
Dominant flow path
Subsurface
10–1 100 101 102 103 0 2 4 6

R L−1 L P tr −1 K−1 So −1 D−1
Figure 49.10 Bayesian posterior distributions of runoff-generation mechanism as affected by
dimensionless parameters.
1.0 clay particles. These small particles then form barriers to water entry to the
soil, greatly reducing the fp of the soil. The fp of a surface sealed soil has been
Ps shown (e.g., Baumhardt et al., 1990; Assouline and Mualem, 1997; Cerdan
(n − θi )D et al., 2001) to be an order of magnitude or more lower than for the same
−1
PET P soil under vegetative canopy.
0.5 So Agricultural operations may greatly impact fp through tillage, compaction,
harvesting of crops and removal/additions of mulches. The impacts of tillage
Runoff
practice may include important seasonal changes, e.g. amount of wintertime
Dim 2 (38.42%)
infiltration (Dao, 1993). Changes due to tillage may dramatically change the
soil surface from a vegetated to bare condition if the operation includes mold-
0.0 board plowing or similar tillage that create enormous retention storage. Some
i Ks−1 types of tillage, such as conservation or no-till operations, may leave much of
L Ps the residue of a previous crop on the surface as a mulch and thereby continue
t r K s So D to provide a relatively high fp. Repeated agricultural tillage often results in
formation of what is called a “plowpan” due to compaction by repeatedly driv-
−0.5
ing farm implements over the soil, and the fact that farming implements have
a limited tillage depth of 12–15 in (30.5–38 cm). The plowpan has a low Ks
R L−1 that creates an artificial barrier to an infiltration wetting front and tends to
elevate nongrowing-season soil moistures (Downer and Ogden, 2003).
−1.0 Figure 49.12 is a photograph taken down a fence line that clearly shows the
effect of tillage practice on hydrologic response of farm land. The field on the
−1.0 −0.5 0.0 0.5 1.0 left side of the fence is farmed using standard tillage practices. The field on
Dim 1 (46.60%) the right-hand side of the fence line was farmed using no-till practices for
three consecutive years (David Kusel, personal communication, May 20,
Figure 49.11 Variables factor map from principal component analysis on 2015). This is evident by the presence of vegetation in a portion of the no-till
posterior distributions shown in Fig. 49.10.
49_Singh_ch49_p49.1-49.12.indd 8 8/22/16 2:20 PM

Poorly Understood Factors Affecting Runoff Generation 49-9
49.5.4 Effect of Slope on Runoff

Most textbooks have some statement to the effect of “the steeper the slope, the
more runoff.” This axiom has been readily accepted by most lay people as well
as professional hydrologists. Yet is it true? The answer depends on the runoff
mechanism.
When considering the saturation-excess runoff mechanism, Dunne (1978)
and Dunne and Leopold (1978) showed that this mechanism favors areas of
low slope. Conversely, they showed that saturation-excess runoff seldom
occurs on steep slopes, but that subsurface stormflow occurs quite rapidly on
steep slopes. In the case of saturation-excess runoff production, the axiom is
NOT true.
In the case of infiltration-excess runoff, Nassif and Wilson (1975), Fox et al.
(1997), and Huang et al. (2013) performed rainfall simulator experiments on
different slopes. Figure 49.13 shows up to a 24% decrease in infiltration
capacity with increasing slope. This result supports the previously stated
maxim. The results are based on a laboratory study for bare and grass covered
“standard” soil that is classified as a loam, based on soil properties provided.
Nassif and Wilson (1975) reported little effect of slope on infiltration into a
peat or clayey sand soils that had much lower permeabilities than the loam
soil. A thorough explanation was not included for the increase in runoff with
increasing slope. Fox et al. (1997) found a declining rate of infiltration with
slope, up to a slope of about 12%, after which there was no effect. They also
Figure 49.12 Two fallow fields after an Iowa rain storm on May 11, 2002, with measured surface storage volume (depression storage) and found that it had a
50 mm (2 in) of rainfall in 2 h. Field on the left is cultivated using standard very similar shaped response curve with slope as infiltration. Huang et al.
tillage practices, while field on the right is no-till. [Source: David Kusel, Carroll County, (2013) found a slight initial decrease in runoff with slope, but then a signifi-
Iowa.]
cant increase in runoff at larger slopes. They attributed the initial decrease to
possible faster and larger surface crust formation on the bare soil for the
smallest slope. Both Nassif and Wilson (1975) and Huang et al. (2013) found
field. Notice that there is considerably more runoff from the field that is the biggest factor was vegetative cover in predicting runoff rates.
farmed using standard tillage practices, and runon infiltration as runoff from Laboratory studies do not necessarily represent the processes in the real
the field on the left flows onto the no-till field on the right. world. First, they use packed trays of sifted soil in which soil structure and
macropores have been destroyed; and second, they ignore the field/hillslope
49.5.2 Land Use and Land Cover scale processes that involve runon, subsurface flow, and natural surface
Land use and land cover impact the potential infiltration rate fp and runoff in microtopography. Some field studies have assessed runoff with slope, for
two fundamental ways: (1) how important is ET, which has a large affect on example, Ribolzi et al. (2011) who found in a study on very steep slopes (30
the evolution of soil water content between water input events, and (2) what vs. 75%) in Laos, that the steeper slope produced less surface runoff. They
are the timing and magnitude of human activities associated with a particular attributed this to the formation of microterraces on the steeper slope that
land use and land cover. Of course, intense development often results in were more pervious than the tilled soil on the 1 m2 plots. The possible role of
extensive impervious surfaces associated with buildings and roads. In agricul- subsurface flow and differences with slope on the small plots was not
tural operations, each crop will have a different time when planting, harvest- mentioned.
ing, etc. occurs and may result in temporary loss of vegetative cover or soil
49.5.5 Effect of Rainfall Intensity on Runoff
compaction from machinery. Due to growth stages (e.g., deciduous trees) and
seasonal variations in ET by plants owing to growing season, some land cov- The common concept of infiltration excess runoff generation is that once
ers may result in drier or wetter soil conditions than others at different times runoff begins fp will decrease and therefore runoff will increase with time as
of the year. For example, Zhang et al. (2001) have documented that mature long as the water input rate is higher than fp. In other words, the fp of the soil
forests have higher ET than grasses in semiarid Australia, while Kelliher et al. is not a function of the water input rate after the initiation of runoff. Yet sev-
(1993) found similar ET rates for coniferous forests and grasslands world- eral observers have shown increasing fp with increased rainfall intensities.
wide, despite the significant differences in canopy morphology and leaf area Nassif and Wilson (1975; Fig. 49.14) show that fp increases as rainfall rates
indices. The seasonality of the ET demand coupled with the seasonal patterns increase, which is contrary to the generally accepted theory that fp can only
of the intensity and amount of precipitation will determine the overall influ- decrease as time proceeds within a storm after runoff begins. Observations by
ence of a particular land use.
Urban soils are extensively modified by construction and compaction, then
by placement of imported topsoils that have nothing to do with the base soils Bare soil Soil with grass cover
(Smith et al., 2005). This has two confounding effects. First, the hydraulic 300 300
behavior of the topsoil is very unlike the underlying base soils. Second, the (a) Slope So (b)
predevelopment soil texture map has very little in common with the actual 0%
postdevelopment soils. For this reason, the use of standard soils datasets may 8%
Infiltration rate (mm h−1)
lead to large simulation errors, while the incorporation of dual-layer infiltra- 16%
tion and specific urban soil hydraulic parameters can lead to significant 200 24% 200
improvements in process-based modeling (Ogden et al., 2013). 36%
49.5.3 Antecedent Soil Wetness
The initial soil water content at the start of a precipitation event (θi) is one
of the most important factors affecting runoff generation (Mishra et al.,
100 100
2003). High θi can reduce infiltration and enhance infiltration excess run-
off generation in finer soil textures. In coarser soil textures, high initial soil
water is usually associated with a near surface water table, corresponding
with less available pore space, and increased runoff. Under both mecha-
nisms higher initial soil moisture leads to more runoff in less time.
Calculation of θi at the start of a water input event using a water budget 0 0
0 5 10 15 0 5 10 15
approach will improve runoff predictions, although verification, especially
by position in the landscape, is difficult (Downer and Ogden, 2004). In Time (min) Time (min)
areas subject to freezing soils, water budget predictions of soil moisture Figure 49.13 Effect of land surface slope on infiltration rate for rainfall simulator
over an entire year are complicated by soil water movement upward into a tests on soil pans (a) without grass cover; and (b) with grass cover. Legend is the same
freezing front. for both figures. [Source: Nassif and Wilson,1975.]
49_Singh_ch49_p49.1-49.12.indd 9 8/22/16 2:20 PM

300 300 (2003) performed a flood quantile scaling study using annual peak flow data
(a) (b) (1981–1996) from the 21 km2 USDA Goodwin Creek Experimental (GCEW)
Water input
rate (mm h−1) Watershed and found simple scaling up to 21 km2. The data from Dunne
(1978) are shown in Fig. 49.15 together with the results shown by Ogden and
Infiltration rate (mm h−1)
312
Dawdy (2003), and a subsequent reanalysis of peak runoff quantiles up to
234
200 200 50-year recurrence interval using GCEW data from 1981 through 2006
156
(Ogden, unpublished data). The envelope drawn on the data by Dunne (1978)
78
declines slightly until a drainage area of about 0.1 km2 (25 acres) is reached,
at which point it declines rather sharply with an exponent of –0.38. This
indicates a flood quantile scaling exponent of 0.62. Ogden and Dawdy (2003)
100 100 reported a flood peak scaling exponent for the 20-year recurrence interval of
0.77 using 16-year data from the GCEW between 1981 and 1996. The ICOLD
envelope curve of maximum observed flood peaks in watersheds less than
300 km2 has an exponent of 0.78 (ICOLD, 2002). A reanalysis value for the
50-year recurrence interval using 26 years data from 1981 to 2006 revealed a
simple scaling exponent of 0.71. Scaling breaks have been identified by other
0 0
0 5 10 15 0 5 10 15 researchers (Smith, 1992; Gupta and Dawdy, 1995). One potential reason for
the break in the relationship at this point is that for at smaller scales, high-
Time (min) Time (min)
intensity convective rainfall is probably the critical precipitation type; and this
Figure 49.14 Infiltration rates for different rainfall rates for (a) bare and, (b) grass type of storm may have a typical area of coverage corresponding to the break
covered loam soils in laboratory study. Slope = 16%. Legend is the same for both point on the x-axis. Another possible reason is the attenuating effects of stor-
figures. [Source: Nassif and wilson, 1975.] age (e.g., channel and overbank) may start to become dominant around the
observed break point, reflecting the different dominant hydraulic processes in
the land versus channel phases. As noted in the title of Fig. 49.15, the data are
Yu et al. (1997; 1998) based on field studies in Australia also documented from USDA watersheds where agriculture is the dominant land use, and
increasing fp with rainfall intensities; they postulated that it is due to the spa- where Hortonian-type runoff is the expected runoff mechanism. Goodrich
tial variation in Ks at the plot scale, and developed a one-parameter exponen- et al. (1997) showed that in the semiarid Walnut Gulch watershed in southern
tial distribution to characterize the spatial variation of maximum fp. Arizona, channel infiltration caused a scale break between 37 and 60 ha.
Several researchers have provided hypotheses on the mechanisms of how 49.6 Conclusions
infiltration capacity might increase with rainfall intensity. One hypothesis is
that longer detention times and higher depths of flow on flatter slopes could This chapter started by stating that hydrologists and engineers calculate peak
lead to the phenomenon, but the difference in depths of flow are very small flows for hypothetical storms, as well as flows over extended time periods for
and would add little to the overall pressure gradient driving fp. Nachabe et al. varied forcing, complex terrain and changing land use and climate conditions.
(1997) provided a hypothetical rationale of the phenomenon by showing how Doing this requires broad predictive understanding of complex interrelated
different parts (e.g., upper, middle, and lower) portions of a watershed will processes, particularly if the predictive model is to get the right answer for the
produce ponding and runoff at different parts of a storm depending on right reasons (Klemes, 1986; Grayson et al., 1992; Kirchner, 2006; Paniconi
whether Ks increases or decreases with elevation, that is, for the different and Putti, 2015; Fatichi et al., 2016). Model complexity aside, it is clear from
zones. At the other end of the scale, Fox et al. (1997) advance the theory that the observations presented in this chapter that runoff generation is complex
variations in fp at the microscale can explain the phenomenon. They state with numerous poorly understood and difficult to quantify processes. Many
that, even for their “smooth” packed soil, the surface was never entirely cov- of the process-level unknowns are biological actors and their effects, includ-
ered with water, and that small mounds and aggregates protruding above the ing human land use change effects. Demonstrated lack of widely applicable
flow may have higher Ks than where raindrop impact has created a more predictive ability, in general, and a lack of progress in increasing the predictive
dense, less permeable surface seal. As indicated above, Yu et al. (1997; 1998) ability of physics-based models even in “well-constrained” situations hints at
attribute it to spatial variations in Ks at a smaller, undefined scale. a need for a fundamentally different approach (Semenova and Beven, 2015).
Extent and density of vegetation may be a key in explaining the phenome- Sivakumar et al. (2014) suggest that our lack of understanding of system-
non. In Fig. 49.14, the vegetated loam soil has much higher and greater range level complexity is manifested by our lack of ability to classify differences
of fp than the bare soil for different rainfall intensities. As indicated earlier, among catchments. We agree with this assertion and believe that improved
many others have documented the connection between vegetation and fp, for process-level understanding is paramount. However, progress will only come
example, Huang et al. (2013) and Dunne et al. (1991) who showed that a when we acknowledge that the tools that are currently being used in standard
deeper wetting front formed directly beneath plants in an arid region of Kenya. practice, and that are loved by regulators because of their simplicity, are inad-
At an intermediate scale, there may be a strong connection between extent of equate. These tools include the Curve Number method (Mockus, 1949), Unit
ponding in depressions and the role of macropores. Rainfall simulator experi- Hydrograph (e.g., Sherman, 1932; McCarthy, 1940), and Lag time equations
ments conducted in the semiarid Wyoming steppe indicate increasing infiltra- (e.g., Kirpich, 1940; Kerby, 1959). The problem is not the existence of these
tion with increasing rainfall rates that visually corresponds to rapid ponding of
soil pans between vegetation clumps, followed by observed high infiltration
down the shafts of wheatgrass into the soil (Ogden, unpublished data).
In the real world, no surface is perfectly smooth. Rather, natural topo-
graphic depressions fill and grow in area as the rainfall intensity exceeds the
Peak runoff rate (cm h )
−1
1
fp at a point. As the area of ponding grows, the number of macropores 10
encountered associated with vegetation and animals increases. On steeper
slopes, the extent of ponding decreases for the same rainfall intensity and
runoff will begin earlier compared to flat slopes. Increases in ponded area
expose more surface to ponded water and allows more macropore activation,
0
increasing the fp. Because ponded areas expand more on flatter topography 10
than on steeper topography, fp will not increase as much on steep versus flat
USDA watershed data (from Dunne, 1978)
slopes and result in more runoff. So, we think there is a logical explanation for
Envelope (from Dunne, 1978)
why “the steeper the slope, the more runoff ” is true, but only if the Hortonian 20−year GCEW recurrence (Ogden and Dawdy, 2003)
runoff mechanism governs the runoff process. Further research is needed to −1
50−year GCEW recurrence (based on 1981−2006 data)
prove this hypothesis. If true, there may be a need to describe the extent of 10 −4 −3 −2 −1 0 1 2
ponding based on the microtopography of a slope in order to predict the 10 10 10 10 10 10 10
2
ponded area as a function of the resultant rainfall minus fp. Drainage area (km )
49.5.6 Relationship of Runoff to Watershed Area Figure 49.15 Observed peak runoff rates for the Hortonian runoff-generation
mechanism versus. catchment area. Data points from USDA study watersheds
Dunne (1978) graphed the relationship of peak runoff rate per area versus assembled by Dunne (1978). Curves from USDA-ARS GCEW at scales form 0.1 to
drainage area using data from USDA research watersheds. Ogden and Dawdy 21 km2 after Ogden and Dawdy (2003).
49_Singh_ch49_p49.1-49.12.indd 10 8/22/16 2:20 PM

References 49-11
methods. The problem is that their entrenched status as “standard practice” Dunne, T., W. Zhang, and B. F. Aubry, “Effects of rainfall, vegetation and
and regulatory requirements that they be used has stopped both researchers microtopograhy on infiltration and runoff,” Water Resources Research, 27:
and practitioners from asking the question “How can we do this better?” 2271–2285, 1991.
Perhaps the more relevant question is: “When we evaluate their performance, Fox, D. M., R. B. Bryan, and A. G. Price, “The influence of slope angle on
why do our current tools perform so poorly?” final infiltration rate for interrill conditions,” Geoderma, 80: 181–194, 1997.
Finally, we must acknowledge that advances in prediction will come from Freeze, R. A., “Role of subsurface flow in generating surface runoff: 2.
advances in predictive understanding made from well-designed process-level Upstream source areas,” Water Resources Research, 8: 1272–1283, 1972.
studies. Normally, exhaustive data-collection efforts tend to follow techno- Germann, P. F., Macropores and hydrologic hillslope processes, Process
logical developments. Large systematic data collection campaigns such as Studies in Hillslope Hydrology, edited by M. G. Anderson and T. P. Burt, 1990.
county-level soil surveys undertaken in the 1950s through the 1970s in sup- Germann, P. F., “Comment on ‘Theory for source-responsive and free-
port of Curve Number hydrology. We need contemporary studies using surface film modeling of unsaturated flow,’” Vadose Zone Journal, 9 (4),
technologies such as geophysical methods, remote sensing, light stable iso- 1000–1101, 2010.
topes, geochemistry, and in situ testing of soil hydraulic behavior at meaning- Goodrich, D. C., L. J. Lane, R. M. Shillito, S. N. Miller, K. H. Syed, and
ful scales. The need for improved prediction capabilities warrants a network D. A. Woolhiser, Linearity of basin response as a function of scale in a semi-
of existing and newly established research catchments aimed at collecting arid watershed,” Water Resources Research, 33 (12): 2951–2965, 1997.
high-quality full featured hydrologic data for formulating and testing process- Grayson, R. B., I. D. Moore, and T. A. McMahon, “Physically based hydro-
based hydrologic prediction tools. logic modeling: 2. Is the concept realistic?,” Water Resources Research, 28:
2659–2666, 1992.
Green, W. H. and G. A. Ampt, “Studies on soil physics: 1. The flow of air
ACKNOWLEDGMENTS
and water through soils,” Journal of Agricultural Science, 4 (1): 1–24, 1911.
The USGS HDCN data were downloaded by Nels Frazier using his superior Gupta, D. S., B. P. Mohanty, and J. M. Kohne,. “Soil hydraulic conductivities
web services skills. Hernan Moreno assisted with statistical analysis of the and their spatial and temporal variations in a vertisol,” Soil Science Society of
Bayesian runoff factors using the R software package. Dave Kusel, Rick Cruse America Journal, 70: 1872–1881, 2006.
and Witold Krajewski all contributed to our acquisition of the photograph Gupta, V. K. and D. R. Dawdy, “Physical interpretations of regional varia-
shown in Fig. 49.12. We also acknowledge the input of four anonymous tions in the scaling exponents of flood quantiles,” Hydrological Processes, 9:
reviewers. 347–361, 1995.
Haverkamp, R., J-Y. Parlange, J. L. Starr, G. Schmitz, and C. Fuentes,
“Infiltration under ponded conditions: 3. A predictive equation based on
REFERENCES
physical parameters,” Soil Science, 149 (5), 292–300, 1990.
Assouline, S. and Y. Mualem, “Modeling the dynamics of seal formation Hillel, D., Soil and Water: Physical Principles and Processes, Academic Press,
and its effect on infiltration as related to soil and rainfall characteristics,” 1971.
Water Resources Research, 33 (7): 1527–1536, 1997. Horton, R. E., “The role of infiltration in the hydrologic cycle,” Eos,
Baumhardt, R. L., M. J. M. Römkens, F. D. Whisler, and J-Y. Parlange, Transactions of American Geophysical Union, 14: 446–460, 1933, doi:10.1029/
“Modeling infiltration into a sealing soil,” Water Resources Research, 26 (10): TR014i001p00446
2497–2505, 1990. Huang, J., P. Wu and X. Zhao, “Effects of rainfall intensity, underlying sur-
Beven, K. and P. Germann, “Macropores and water flow in soils,” Water face and slope gradient on soil infiltration under simulated rainfall experi-
Resources Research, 18: 1311–1325, 1982. ments,” Catena, 104: 93–102, 2013.
Beven, K. and P. Germann, “Macropores and water flow in soils revisited,” Huntington, T. G., “Evidence for intensification of the global water cycle:
Water Resources Research, 49: 3071–3092, 2013, doi:10.1002/wrcr.20156. review and synthesis,” Journal of Hydrology, 380: 83–95, 2006.
Beven, K. and M. J. Kirkby, “A physically-based variable contributing-area Jones, J. A. A.,“Soil piping and catchment response,” Hydrological Processes,
model of catchment hydrology,” Hydrological Science Bulletin, 24 (1): 43–69, 24: 1548–1566, 2010.
1979. Kelliher, F. M., R. Leuning, and E-D. Schulze, “Evaporation and canopy
Bonell, M., “Progress in the understanding of runoff generation in forests,” characteristics from coniferous forests and grasslands,” Oecologia, 95:
Journal of Hydrology, 150: 217–275, 1993 153–163, 1993.
Budyko, M. I., Climate and Life, Academic Press, Orlando, FL, 1974, p. 508. Kerby, W. S., Time of concentration for overland flow, Civil Engineering,
Burwell, R. W. et al., “Losses of surface runoff, total solids, and nitrogen ASCE, Reston, VA, Vol. 26, Suppl. 3, 1959, p. 60.
during Bermudagrass establishment on levee embankments,” 2011, Kirchner J. W., “Getting the right answers for the right reasons: linking
doi:10.2134/jeq2010.0458. measurements, analyses, and models to advance the science of hydrology,”
Buttle, J. M. and K. Sami, “Testing the groundwater ridging hypothesis of Water Resources Research, 42: W03S04, 2006, doi:10.1029/2005WR004362.
streamflow generation during snowmelt in a forested catchment,” Journal of Kirkby, M. J., “Hillslope runoff processes and models,” Journal of Hydrology,
Hydrology, 135: 53–72, 1992. 100: 315–339, 1988.
Celia, M. A., E. T. Bouloutas, and R. L. Zarbo, “A general mass-conservative Kirpich, Z. P., Time of concentration of small agricultural watersheds, Civil
numerical solution for the unsaturated flow equation,” Water Resources Engineering, Vol. 10, Suppl. 6, 1940, p. 362.
Research, 26 (7): 1483–1496, 1990. Klemes, V., “Dilettantism in hydrology: transition or destiny?,” Water
Cerdan, O., V. Souchere, V. Lecomte, A. Couturier, and Y. Le Bissounnais, Resources Research, 22: S177–S188, 1986.
“Incorporating soil surface crusting processes in an expert-based runoff Madge, D. S., “Leaf fall and litter disappearance in a tropical forest,”
model: sealing and transfer by runoff and erosion related to agricultural man- Pedobiologia, 5: 273–288, 1965.
agement,” Catena, 46: 189–205, 2001. Madge, D. S., “Field and laboratory studies on the activities of two species
Dao, T. H., “Tillage and winter wheat residue management effects on water of tropical earthworms,” Pedobiologia, 9: 188–214, 1969.
infiltration and storage,” Soil Science Society of America Journal, 57: 1586–1595, McCarthy, G. T., Chap. V of “Flood Control,” The Engineer School, Fort
1993. Belvoir, VA, 1940, pp. 127–177.
Dingman, S. L., Physical Hydrology, 3rd. ed., Waveland Press, 2015. Mein, R. G. and C. L. Larson, “Modeling infiltration during a steady rain,”
Downer, C. W. and F. L. Ogden, “Prediction of runoff and soil moistures at Water Resources Research, 9: 384–392, 1973.
the watershed scale: Effects of model complexity and parameter assignment,” Mishra, S. K., J. V. Tyagi, and V. P. Singh, “Comparison of infiltration mod-
Water Resources Research, 39 (3), 2003, doi:10.1029/2002WR001439. els,” Hydrological Processes, 17 (13): 2629–2652, 2003.
Downer, C. W. and F. L. Ogden, “GSSHA: model to simulate diverse Mockus, V., “Estimation of total (and peak rates of) surface runoff for
streamflow producing processes,” Journal of Hydrologic Engineering, 9 (3): individual storms. Exhibit A in Appendix B,” Interim Survey Report (Neosho)
161–174, 2004, doi:10.1061/(ASCE)1084-0699(2004)9:3(161) River Watershed USDA, 1949.
Dunne T. and L. B. Leopold, Water in Environmental Planning, Freeman Nachabe, M. H., T. H. Illangasekare, H. J. Morel-Seytoux, and L. R. Ahuja,
Publishers, 1978. “Infiltration over heterogenous watershed: influence of rain excess,” Journal of
Dunne, T., Field studies of hillslope flow processes, Hillslope Hydrology, Hydrologic Engineering, 2: 140–143, 1997.
edited by M. J. Kirkby, John Wiley & Sons, 1978. Nassif, S. H. and E. M. Wilson, “The influence of slope and rain intensity
Dunne, T. and R. D. Black, “Partial area contributions to storm runoff in a on runoff and infiltration,” Hydrological Sciences Bulletin, 20: 539–553, 1975.
small New England watershed,” Water Resources Research, 6: 1296–1311, Nieber, J. L. and G. S. Warner, “Soil pipe contributions to steady subsurface
1970. stormflow,” Hydrological Processes, 5: 329–344, 1991.
49_Singh_ch49_p49.1-49.12.indd 11 8/22/16 2:20 PM

Niedzialek, J. M. and F. L. Ogden., 2004. “Numerical investigation of satu- Sherman, L. K., “Streamflow from rainfall by the unit-graph method,”
rated source area behavior at the small catchment scale,” Advances in Water Engineering News Record, 108: 501–505, 1932.
Resources, 27: 925–936, 2004. doi:10.1016/j.advwatres.2004.06.005. Short, D., W. R. Dawes, and I. White, “The practicability of using Richards’
Niedzialek, J. M. and F. L. Ogden, “First-order catchment mass balance equation for general purpose soil-water dynamics models,” Environment
during the wet season in the Panama Canal watershed,” Journal of Hydrology, International, 21 (5): 723–730, 1995.
462: 77–86, 2012. Sivakumar, B., V. P. Singh, R. Berndtsson, and S. K. Khan, “Catchment clas-
Nimmo, J. R., “Preferential flow occurs in unsaturated soils,” Hydrologicl sification framework in hydrology: challenges and directions,” Journal of
Processes, 26: 786–789, 2012. Hydrologic Engineering, 2014, doi:10.1061/(ASCE)HE.1943-5584.0000837.
O’Brien, A. L., “Rapid water table rise,” Water Resources Bulletin, 18 (4): Sklash, M. G. and R. N. Farvolden, “The role of groundwater in storm
713–715, 1982, doi:10.1111/j.1752-1688.1982.tb00056.x. runoff,” Journal of Hydrology,43: 45–65, 1979.
Ogden, F. L., W. Lai, R. C. Steinke, J. Zhu, C. A. Talbot, and J. L. Wilson, “A Slack, J. R., A. M. Lumb, and J. M. Landwehr, “HCDN: streamflow data set,
new general vadose zone solution method,” Water Resources Research, 51: 1874–1988,” U.S. Geological Survey Water-Resources Investigations Report,
4282–4300, 2015, doi:10.1002/2015WR017126. 93–4076, 1993.
Ogden, F. L., T. D. Crouch, R. F. Stallard, and J. S. Hall, “Effect of land cover and Smith, J. A., “Representation of basin scale in flood peak distribution,”
use on dry season river runoff, runoff efficiency, and peak runoff in the seasonal Water Resources Research, 28 (11): 2993–2999, 1992.
tropics of Central Panama,” Water Resources Research, 49 (12): 8443–8462, 2013. Smith, J. A., M. L. Baeck, K. L. Meierdiercks, P. A. Nelson, A. J. Miller, and
Ogden F. L. and D. R. Dawdy, “Peak discharge scaling in a small Hortonian E. J. Holland, “Field studies of the storm event hydrologic response in an
watershed,” Journal of Hydrologic Engineering, 8 (2): 64–73, 2003. urbanizing watershed,” Water Resources Research, 41 (10), 2005,
Ogden F. L. and B. A. Watts, “Saturated area formation on nonconvergent doi:10.1029/2004WR003712.
hillslope topography with shallow soils: a numerical investigation,” Water Starr, J. L., “Spatial and temporal variation of ponded infiltration,” Soil
Resources Research, 36 (7): 1795–1804, 2000. Science Society of America Journal, 54: 629–636, 1990.
Or, D., “Scaling of capillary, gravity and viscous forces affecting flow mor- Talsma T. and J-Y. Parlange, “One-dimensional vertical infiltration,”
phology in unsaturated porous media,” Advances in Water Resources, 31: Australian Journal of Soil Research, 10 (2): 143–150, 1972.
1129–1136, 2008. Tocci, M. D., C. T. Kelley, and C. T. Miller, “Accurate and economical solu-
Parlange J-Y., I. Lisle, R. D. Braddock, and R. E. Smith, “The three param- tion of the pressure-head form of Richards’ equation by the method of lines,”
eter infiltration equation,” Soil Science, 133 (6): 337–341, 1982. Advances.in Water Resources, 20 (1): 1–14, 1997.
Parlange, J-Y. and D. Hill, “Theoretical analysis of wetting front instability Trojan, M. D. and D. R. Linden, “Microrelief and rainfall effects on water
in soils,” Soil Science, 122 (4): 236–239, 1976. and solute movement in earthworm burrows,” Soil Science Society of America
Philip, J. R., “The theory of infiltration: 1. The infiltration equation and its Journal, 56: 727–733, 1992.
solution,” Soil Science, 83 (5): 345–357, 1957a. U.S. National Academy of Sciences, Climate Change: Evidence and Choices,
Philip, J. R., “The theory of infiltration: 2. The profile of infinity, Soil 2014, available online at: http://www.nap/edu/catalog.php?record_id=18730,
Science, 83 (5): 435–448, 1957b. accessed April 26, 2015.
Ponce, V. M., R. P. Pandey, and S. Ercan, “Characterization of drought Vörösmarty, C. J., P. Green, J. Salisbury, and R. B. Lammers, “Global water
across the climate spectrum,” Journal of Hydrologic Engineering, ASCE, 5 (2): resources: vulnerability from climate change and population growth,” Science,
222–224, 2000. 289: 284(2000), 2013, doi:10.1126/science.289.5477.284.
Ribolzi, O., J. Patin, L. M. Bresson, K. O. Latsachack, E. Mouche, O. Wilkinson, M. T., P. J. Richards, and G. S. Humphreys, “Breaking ground:
Sengtaheuanghoung, N. Silvera, et al., “Impact of slope gradient on soil sur- pedological, geological, and ecological implications of soil bioturbation,” Earth-
face features and infiltration on steep slopes in northern Laos,” Geomorphology, Science Reviews, 97 (1–4): 257–272, 2009, doi:10.1016/j.earscirev. 2009.09.005.
127: 53–63, 2011. Yair, A., “Short and long term effects of bioturbation on soil erosion, water
Richards, L. A., “Capillary conduction of liquids in porous mediums,” resources, and soil development in an arid environment,” Geomorphology, 13
Journal of Applied Physics, 1 (5): 318–333, 1931. (1): 1995.
Saghafian, B., P. Y. Julien, and F. L. Ogden, “Similarity in catchment Youngs, E. G., G. Spoor, and G. R. Goodall, “Infiltration from surface
response: 1. Stationary rainstorms,” Water Resources Research, 1995, ponds into soils overlying a very permeable substratum,” Journal of Hydrology,
doi:10.1029/95WR00518. 186: 327–334, 1996.
Schaller, M. F. and Y. Fan, “River basins as groundwater exporters and Yu, B., C. W. Rose, K. J. Coughlin, and B. Fentie, “Plot scale rainfall-runoff
importers: implications for water cycle and climate modeling,” Journal of characteristics and modeling at six sites in Australia and Southeast Asia,”
Geophysical Research, 114: D04103, 2009, doi:10.1029/2008JD010636. Transactions of the American Society of Agricultural Engineering, 40 (5):
Schoups, G., N. C. van de Giesen, and H. H. G. Savenije, “Model complex- 1295–1303, 1997.
ity control for hydrologic prediction,” Water Resources Research, 44 (12): 2008, Yu, B., U. Cakurs, and C. W. Rose, “An assessment of methods for estimat-
doi:10.1029/2008WR006836.

Handbook of Applied Hydrology, Second Edition by Vijay P. Singh

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Handbook of Applied Hydrology, Second Edition by Vijay P. Singh

Uploaded by

Copyright:

Available Formats

Handbook of Applied Hydrology

00_Singh_FM_pi-lvi.indd 1 8/26/16 4:35 PM

00_Singh_FM_pi-lvi.indd 2 8/26/16 4:35 PM

New York Chicago San Francisco Athens London

00_Singh_FM_pi-lvi.indd 3 8/26/16 4:35 PM

Handbook of Applied Hydrology, Second Edition

The pages within this book were printed on acid-free paper.

Sponsoring Editor Copy Editor

00_Singh_FM_pi-lvi.indd 4 8/26/16 4:35 PM

00_Singh_FM_pi-lvi.indd 5 8/26/16 4:35 PM

Part 1 Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1

Part 2 Data Collection and Processing . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1

Part 3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-1

Part 4 Hydrologic Processes and Modeling . . . . . . . . . . . . . . . . . . . . . 37-1

Part 5 Sediment and Pollutant Transport . . . . . . . . . . . . . . . . . . . . . . . 63-1

Part 6 Hydrometeorologic and Hydrologic Extremes . . . . . . . . . . . . 72-1

Part 7 Systems Hydrology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81-1

Part 8 Hydrology of Large Rivers and Lake Basins . . . . . . . . . . . . . . . 93-1

Part 9 Applications and Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125-1

Part 10 Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151-1

00_Singh_FM_pi-lvi.indd 7 8/26/16 4:35 PM

00_Singh_FM_pi-lvi.indd 8 8/26/16 4:35 PM

Part 1. Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1

Chapter 1. The Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3

Chapter 2. Watersheds, River Basins, and Land Use . . . . . . . . . . . . . . . . 2-1

Chapter 3. Water Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1

PART 2. Data Collection and Processing . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1

Chapter 4. Hydrometeors and Quantitative Precipitation Estimation 4-3

00_Singh_FM_pi-lvi.indd 9 8/26/16 4:35 PM

Chapter 5. Streamflow Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1

Chapter 6. Streamflow Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1

Chapter 7. Hydrologic Information Systems . . . . . . . . . . . . . . . . . . . . . . 7-1

Chapter 8. Remote Sensing Techniques and Data Assimilation for

Chapter 9. Geographic Information Systems . . . . . . . . . . . . . . . . . . . . . . 9-1

Chapter 10. Design of Hydrologic Networks . . . . . . . . . . . . . . . . . . . . . . 10-1

Part 3. Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-1

Chapter 11. Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-3

00_Singh_FM_pi-lvi.indd 10 8/26/16 4:35 PM

11.5 Drawbacks of ANN Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-5

Chapter 12. Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-1

Chapter 13. Evolutionary Computing: Genetic Algorithms . . . . . . . . . 13-1

Chapter 14. Relevance Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-1

Chapter 15. Harmonic Analysis and Wavelets . . . . . . . . . . . . . . . . . . . . . 15-1

Chapter 16. Outlier Analysis and Infilling of Missing Records

Chapter 17. Linear and Nonlinear Regression . . . . . . . . . . . . . . . . . . . . . 17-1

Chapter 18. Time Series Analysis and Models . . . . . . . . . . . . . . . . . . . . . 18-1

00_Singh_FM_pi-lvi.indd 11 8/26/16 4:35 PM

18.10 Stochastic Simulation, Forecasting, and Uncertainty . . . . . . . . . . . . . . . . . . . . . . 18-8

Chapter 19. Statistical Detection of Nonstationarity: Issues

Chapter 20. Spatial Analysis and Geostatistical Methods . . . . . . . . . . 20-1

Chapter 21. Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21-1

Chapter 22. Calibration, Parameter Estimation, Uncertainty, Data

Chapter 23. Bayesian Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23-1

Chapter 24. Optimization Approaches for Integrated Water Resources

Chapter 25. Nonparametric Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25-1

00_Singh_FM_pi-lvi.indd 12 8/26/16 4:35 PM

Chapter 26. Predictive Uncertainty Assessment

Chapter 27. Risk-Reliability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27-1

Chapter 28. Scaling and Fractals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28-1

Chapter 29. Nonlinear Dynamics and Chaos . . . . . . . . . . . . . . . . . . . . . . 29-1

Chapter 30. Copula Modeling in Hydrologic Frequency Analysis . . . 30-1

Chapter 31. Entropy Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31-1

00_Singh_FM_pi-lvi.indd 13 8/26/16 4:35 PM

Chapter 32. Entropy Production Extremum Principles . . . . . . . . . . . . . 32-1

Chapter 1. The Hydrologic Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3

Chapter 2. Watersheds, River Basins, and Land Use . . . . . . . . . . . . . . . . 2-1

Chapter 8. Remote Sensing Techniques and Data Assimilation for

Chapter 16. Outlier Analysis and Infilling of Missing Records

Chapter 19. Statistical Detection of Nonstationarity: Issues

Chapter 22. Calibration, Parameter Estimation, Uncertainty, Data

Chapter 24. Optimization Approaches for Integrated Water Resources

Chapter 26. Predictive Uncertainty Assessment

Chapter 39. Spatial and Temporal Estimation and

Chapter 43. Rainfall Interception, Detention, and

Chapter 49. Watershed Runoff, Streamflow Generation,

Chapter 54. Surface Water–Groundwater Interactions:

Chapter 55. Seawater Intrusion in Coastal Aquifers: Concepts,

Chapter 56. Regional Land Subsidence Caused by the Compaction

Chapter 62. Feasibility, Engineering, and Operations Models: Using

Chapter 73. Hydrometeorological Extremes

Chapter 78. Risk, Reliability, and Return Periods and