You are on page 1of 20

Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: https://www.tandfonline.com/loi/thsj20

Prediction of daily flow duration curves and


streamflow for ungauged catchments using
regional flow duration curves

YUSUF M. MOHAMOUD

To cite this article: YUSUF M. MOHAMOUD (2008) Prediction of daily flow duration curves and
streamflow for ungauged catchments using regional flow duration curves, Hydrological Sciences
Journal, 53:4, 706-724, DOI: 10.1623/hysj.53.4.706

To link to this article: https://doi.org/10.1623/hysj.53.4.706

Published online: 18 Jan 2010.

Submit your article to this journal

Article views: 4814

View related articles

Citing articles: 20 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
706 Hydrological Sciences–Journal–des Sciences Hydrologiques, 53(4) August 2008

Prediction of daily flow duration curves and streamflow for


ungauged catchments using regional flow duration curves

YUSUF M. MOHAMOUD
USEPA National Exposure Research Laboratory, 960 College Station Road, Athens, Georgia, USA
mohamoud.yusuf@epa.gov

Abstract A method is presented to predict flow duration curves (FDCs) and streamflow for ungauged
catchments in the Mid-Atlantic Region, USA. Twenty-nine catchments were selected from the Appalachian
Plateau, Ridge and Valley, and Piedmont physiographic provinces to develop and test the proposed method.
Using a step-wise multiple regression analysis, the dominant landscape and climate descriptors were identi-
fied and regional FDC models were developed for each province. Predictive performance of the proposed
method was estimated using data from three evaluation sites that were not included in parameter estimation
for the regional FDC models. The results of this study show that climate and geomorphological descriptors
strongly influence the hydrology of the Appalachian Plateau and some Ridge and Valley catchments,
whereas soil and geomorphological descriptors strongly influence the hydrology of the Piedmont and some
Ridge and Valley catchments. Streamflow values calculated by the drainage area ratio method and those
reconstructed by the FDC method were compared with the observed streamflow values. The results indicate
that the FDC-based method shows great promise for predicting streamflow in ungauged basins.
Key words flow duration curve; ungauged basins; streamflow; data transfer

Prévision de courbes de débits classés et de débit fluviatile pour des bassins versants non
jaugés à l’aide de courbes de débits classés régionales
Résumé Une méthode est présentée pour la prévision de courbes de débits classés (CDC) et de débit
fluviatile pour des bassins versants non-jaugés de la région Centre-Atlantique des Etats-Unis. Vingt neuf
bassins versants ont été sélectionnés au sein des provinces physiographiques du Plateau, de la Chaîne, de la
Vallée et du Piedmont Appalachiens afin de développer et de tester la méthode proposée. A l’aide d’une
analyse par régression multiple pas à pas, les descripteurs paysagers et climatiques dominants ont été
identifiés et des modèles régionaux de CDC ont été développés pour chaque province. La performance de
prévision de la méthode proposée a été estimée grâce aux données de trois sites d’évaluation non prises en
compte pour l’estimation des paramètres des modèles régionaux de CDC. Les résultats de cette étude
montrent que les descripteurs climatiques et géomorphologiques influencent fortement l’hydrologie des
bassins du Plateau et de quelques bassins de la Chaîne et de la Vallée, alors que les descripteurs
pédologiques et géomorphologiques influencent fortement l’hydrologie des bassins du Piedmont et de
certains bassins de la Chaîne et de la Vallée. Les valeurs de débits calculées par la méthode du ratio de l’aire
de drainage et les valeurs reconstruites par la méthode basée sur les CDC ont été comparées avec les valeurs
de débit observées. Les résultats indiquent que la méthode basée sur les CDC est prometteuse pour la
prévision de débit en bassin non jaugé.
Mots clefs courbe des débits classés; bassins versants non jaugés; débit fluviatile; transfert de données

INTRODUCTION
Complex interactions between precipitation inputs and landscape characteristics, such as geology,
soil, topography and vegetation, have an influence on hydrological responses (McNamara et al.,
1998; Post & Jones, 2001). Indeed, one can view hydrological response as an indicator of how
catchments transfer precipitation inputs into streamflow. Precipitation inputs, particularly rainfall
and snow, are widely recognized as forcing factors that control hydrological responses. These
important hydrometeorological descriptors are used as input variables in many hydrological
models. However, techniques to identify and quantify those landscape descriptors that strongly
influence hydrological response at the catchment scale are still lacking. To quantify the relation-
ship between basin characteristics and hydrological descriptors, many investigators have
developed single indices such as mean annual flow (Reimers, 1990), baseflow index (Nathan &
McMahon, 1990; Lacey & Grayson, 1998), Q95 (Laaha & Blöschl, 2007), runoff ratio (Berger &
Entekhabi, 2001), and baseflow recession curve characteristics (Zecharias & Brutsaert, 1988).
Prediction of continuous streamflow time series remains uncertain (Wagener & Wheater,
2006) although progress has been made for predicting indices like mean annual flow or extreme

Open for discussion until 1 February 2009 Copyright ” 2008 IAHS Press
Prediction of daily flow duration curves and streamflow for ungauged catchments 707

flow values. However, complete flow duration curves can provide more useful and detailed infor-
mation than such indices. Flow duration curves (FDCs) incorporate the relationship between the
frequency and magnitude of streamflow (Vogel & Fennessey, 1995). They also integrate the com-
bined impacts of climate, geology, geomorphology, soils and vegetation, as well as flow regulation
by dams and diversions; they are, therefore, useful in comparing runoff characteristics of different
land-use areas at the catchment scale (Linsley et al., 1949; Searcy 1959; Pearce, 1990; Sugiyama
et al., 2003). In general, FDCs sort out streamflow data by shifting high flows with high precipi-
tation signals to one end of the curve, medium flows to the middle, and low flows (presumably
with low precipitation signals) to the other end of the curve. According to Fennessey & Vogel
(1990), FDC applications should be limited to problems in which the sequential nature of stream-
flow is not important, thereby acknowledging FDC limitations for many reservoir operations
issues. Castellarin et al. (2004) reviewed regionalization approaches to predict FDCs and classified
these estimation procedures into statistical (Leboutillier & Waylen, 1993; Yu & Tang, 2000; Singh
et al., 2001; Yu et al., 2002; Croker, et al., 2003; Claps et al., 2005), parametric (Quimpo et al.,
1983; Mimikou & Kaemaki, 1985) and graphical approaches (Smakhtin et al., 1997).
The approach presented here differs from other FDC approaches in the scientific literature in
several ways. First, it employs extensive landscape–climate descriptors, including soil descriptors
from the top two layers, as explanatory variables, whereas earlier studies used far fewer descrip-
tors. Second, many of the earlier studies predicted only a few percentile flows representing low
flows, or used the same set of parameters to estimate different percentile flows. Third, this study
not only focuses on estimating FDCs and streamflows for ungauged basins, but also identifies
landscape–climate descriptors for 15 percentile flows. Fourth, this study presents a method to re-
construct streamflow from FDCs using a sequential generation scheme instead of the spatial inter-
polation method proposed by Hughes & Smakhtin (1996). And finally, this study views observed
streamflow time series data as a combination of streamflow magnitude and sequence. The FDC is
predominantly a representation of the magnitude component; it has no time sequence component.
Conversely, the streamflow sequence component has no magnitude; it only gives the FDC’s mag-
nitude component its time sequence. Specifying the sequence component is essential for conver-
sion of a FDC into a complete streamflow time series. The objective of this study is to develop and
test a method to predict flow duration curves and streamflow for ungauged catchments.

METHOD
Study area description and data sets
The study region covers the Appalachian Plateau, Ridge and Valley, and Piedmont physiographic
provinces of Maryland, Pennsylvania, Virginia and West Virginia (Fig. 1). This study region offers
an opportunity for comparative hydrological studies because of its marked differences in geo-
morphology, vegetation, soil and bedrock geology. Fenneman (1938) classified the Mid-Atlantic
region into several physiographic provinces that have nearly similar landscape characteristics. In
this study, the physiographic province was used as a basis for grouping the 29 headwater
catchments into three groups. About 31% of these are located in the Appalachian Plateau, 38% are
in the Ridge and Valley, and 31% are in the Piedmont. Three evaluation sites were randomly
selected from each physiographic province and the remaining 26 catchments were used for
percentile flow model development and model parameter estimation.
Table 1 lists physiographic provinces, drainage areas, major river systems and information on
whether data from a particular catchment were used for parameter estimation, percentile flow
model development, or model evaluation. All 29 catchments were selected from the US Geological
Survey (USGS) Hydro-Climatic Data Network (HCDN) (Slack & Landwehr, 1992). Most of the
catchments have a high percentage of forest cover and limited human-induced disturbances, such
as urbanization and flow regulation. Soil, land use, land cover and geomorphological parameters
were extracted from geographic information system (GIS) databases for each catchment.
Geomorphological data were extracted from digital elevation model (DEM) databases and soil

Copyright ” 2008 IAHS Press


708 Yusuf M. Mohamoud

Fig. 1 Location of study catchments.

descriptors were extracted from digital State Soil Geographic (STATSGO) databases. Climate and
streamflow data were also compiled for each catchment. Table 2 lists all the landscape, climate
and hydrology descriptors developed.
Land use and land cover descriptors Dominant forest types in the study region are deciduous,
mixed and evergreen. In mountainous areas of the Appalachian Plateau and Ridge and Valley catch-
ments, deciduous forest is the dominant type. In low-lying Ridge and Valley catchments, mixed
forest is dominant. In the Piedmont province, the dominant forest type is evergreen. Land use and land
cover data were obtained from the US EPA Better Assessment Science Integrating Point and
Non-point Sources (BASINS) land use and land cover database (USEPA, 2004). The data reflect land
cover conditions existing from the mid-1970s to the early 1980s. Some of the quantitative land-use
descriptors included percentage of forest cover. Forest cover influences important hydrological pro-
cesses such as rainfall interception and evapotranspiration, but climate descriptors can mask its
influence.

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 709

Table 1 Study catchments arranged by physiographic province and major river system with additional
information about drainage area, and whether used for parameter estimation or model evaluation.
ID Catchment Major river basin Drainage area Record period Data used for estima-
(km2) tion (E) or testing (T)
Appalachian Plateau:
1 Greenbrier River Ohio 1598.00 1980–2000 E
2 Little Mahoning Creek, PA Ohio 226.26 1980–2000 T
3 Little Shenango Creek, PA Ohio 269.36 1980–2000 E
4 West Branch Clarion Creek, PA Ohio 163.00 1980–2000 E
5 Young Womans Creek, PA Susquehanna 119.65 1980–2000 E
6 Pine Creek, PA Susquehanna 563.68 1980–2000 E
7 Blockhouse Creek, PA Susquehanna 97.60 1980–2000 E
8 Towanda Creek, PA Susquehanna 556.61 1980–2000 E
9 Lehigh River, PA Delaware 237.40 1980–2000 E
Ridge and Valley:
10 Wallopen Creek, PA Susquehanna 115.39 1980–2000 E
11 Marsh Creek, PA Susquehanna 114.17 1980–2000 E
12 Little Juniata River, PA Susquehanna 569.55 1980–2000 E
13 Sherman Creek, PA Susquehanna 517.78 1980–2000 E
14 Back Creek, WV Potomac 629.10 1965–1975 E
15 Cacapon River, WV Potomac 1752.67 1980–1995 E
16 South Branch Potomac, WV Potomac 3808.24 1980–2000 E
17 Patterson Creek, WV Potomac 566.96 1980–2000 E
18 S F Shenandoah, VA Potomac 4250.94 1980–2000 E
19 Bullpasture River, VA James 284.78 1980–2000 E
20 Johns Creek, VA James 269.24 1980–2000 T
Piedmont :
21 Holiday Creek, VA James 23.30 1980–2000 E
22 Slate River, VA James 585.09 1980–1995 E
23 Hardware River, VA James 300.31 1980–1995 E
24 Robinson Creek, VA Rappahannock 463.40 1980–2000 T
25 Hazel Creek, VA Rappahannock 743.40 1980–2000 E
26 Seneca Creek , MD Potomac 261.50 1980–2000 E
27 Bennett Creek, MD Potomac 162.60 1980–2000 E
28 Deer Creek, MD Susquehanna 244.39 1980–2000 E
29 Tulpehocken Creek, PA Delaware 546.25 1980–2000 E
PA: Pennsylvania; WV: West Virginia; VA: Virginia; MD: Maryland.

Geomorphology descriptors GIS spatial analysis tools were used to extract quantitative
measures of geomorphological parameters from DEM data obtained from the USGS website
http://edc.usgs.gov/geodata. For large study catchments, several digital coverages were merged to
cover the entire area. All the geomorphological descriptors extracted from the USGS 30-m DEM
data were used to develop relationships between landscape–climate and hydrological descriptors
(Table 2). Most geomophological descriptors describe catchment, channel, or drainage network
characteristics.
Soil descriptors GIS analysis tools available within the BASINS modelling system (USEPA,
2004) were used to extract generalized soil properties from STATSGO data (USDA, 1993). For
each study catchment, descriptors from the top two layers of the soil profile were extracted. In
general, soils of the Mid-Atlantic Region vary due to their topographic position in the landscape
and underlying bedrock geology. For example, ridge-dominated Ridge and Valley catchments
have soils derived from sandstone and shale bedrock, but valley-dominated catchments have soils
derived from limestone and shale. In contrast, most soils found in the Appalachian Plateau
province formed from glacial till deposits, and soils found in the Piedmont province are derived
from metamorphic rocks. To obtain representative soil properties for each study catchment,
spatially-weighted soil descriptors, d , over each catchment were calculated as the sum of the

Copyright ” 2008 IAHS Press


710 Yusuf M. Mohamoud

Table 2 List of landscape, climate and hydrological descriptors.


Symbol Variable description Units
Land use and land cover:
AGRC Agriculture %
URBN Urban %
FRSD Deciduous forest %
FRST Mixed forest %
FRSE Evergreen forest %
Geomorphology:
HMIN Minimum elevation m
HMAX Maximum elevation m
HAVG Average elevation m
HMED Median elevation m
BREL Basin relief m
RRAT Relief ratio m/km
SAVG Average slope m/km
SMED Median slope m/km
MCSL Main stream channel length km
TCHL Total length of streams km
CSLP Main channel slope m/km
DDEN Drainage density km/km2
DARE Drainage area km2
HPC10 Proportion of hypsometric curve height corresponding to area proportion = 0.1
HPC50 Proportion of hypsometric curve height corresponding to area proportion = 0.5
Soil: (top two layers; suffixes 1 and 2 refer to layers 1 and 2)
STDP Total soil depth mm
SOLD12 Soil depth of layers one and two mm
AWC12 Available water capacity mm
KSAT12 Saturated hydraulic conductivity mm/h
KEFC Effective hydraulic conductivity (depth weighted for layers 1 and 2) mm/h
SBLD12 Bulk density g/cm3
SORC12 Organic carbon content %
CLAYl2 Clay %
SILT12 Silt %
SAND12 Sand %
ROCK12 Rock fragment %
AWCR Available water capacity ratio mm/mm
STOR12 Available moisture storage mm
TSTOR Total storage of both layers mm
SDRA Depth ratio (layer 2 / layer 1) mm/mm
Geology:
DLIT Dominant lithology (qualitative)
BFI Baseflow index obtained by hydrograph separation of daily streamflow
Climate:
MAP Mean annual precipitation for period of record mm
MMP Mean monthly precipitation for period of record (12 months) mm
(for August –
AUGPREC)
PET Mean annual potential evapotranspiration for period of record mm
MPET Mean monthly potential evapotranspiration for period of record (12 months) mm
(for August –AUGPET)
PET/PREC Mean annual dryness index for period of record mm/mm
Hydrology:
Qp (e.g. for p = 0.1%, Normalized percentile flow equalled or exceeded p% of the time, where L s-1 km-2
Q1) p = 0.1, 0.5, 1, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 95, 99%.

product of area weighted averages determined by dividing the area covered by each soil series by
the total area of the entire catchment and the specific parameter values of each soil series.

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 711

Bedrock geology descriptors Two major rock types cover the study area: consolidated
crystalline and consolidated to unconsolidated sedimentary. Crystalline rocks found in the
Piedmont physiographic province consist of metamorphic rocks. Appalachian Plateau and Ridge
and Valley provinces have consolidated to unconsolidated sedimentary rocks that consist of
sandstone, siltstone and shale (USGS, 1992). Lithological and structural geological properties
influence the hydrological response of a catchment (Maxey, 1964), but no quantitative bedrock
geology descriptor was extracted for each catchment. In the absence of a bedrock geology
descriptor, baseflow index was used as a surrogate. Baseflow index was determined using an
automated baseflow separation and recession analysis technique (Arnold et al., 1995). Some
inferences about geological descriptors were also made from qualitative bedrock geology data
extracted from digital USGS maps (Shruben et al., 1998).
Climate descriptors A problem that hinders the identification of landscape descriptors at the
catchment scale is the dominance of precipitation on hydrological response at short time scales and
the dominance of evapotranspiration at longer time scales, i.e. periods between rainy days. Such
dominance masks any influences of the landscape descriptors, making identification of important
landscape descriptors more difficult. Removing short-term precipitation signals from streamflow
time series offers a way to adjust for climate influence and, thus, to identify dominant landscape
descriptors. These climate descriptors included long-term mean annual precipitation (MAP),
annual potential evapotranspiration (PET), mean monthly potential evapotranspiration (MPET),
mean monthly precipitation (MMP) and mean annual and monthly dryness index described as the
ratio of potential evapotranspiration to precipitation (PET/MAP or MPET/MMP) (Budyko, 1974;
Milly, 1994). For each catchment, potential evapotranspiration was estimated with the equation of
Hamon (1961). Climate data were obtained from the National Climatic Data Center in Asheville,
North Carolina and the BASINS database (USEPA, 2004).
Hydrological descriptors A major challenge in developing landscape–climate and
hydrological descriptor relationships stems from the selection of representative hydrological
descriptors. In this study, the FDC, which integrates landscape–climate and hydrological
influences, was selected as the hydrological descriptor. For most catchments, the period of record
was from 1980 to 2000 (Table 1). The FDCs are period-of-record or composite, and reflect the
integrated hydrologic responses of the study catchments during that period. A normalized FDC
was constructed from normalized streamflow data (L s-1 km-2), and 15 percentile flow exceedence
percentages (Table 2) were selected from the normalized FDC for each catchment. Throughout
this paper, these 15 fixed points on the FDC are referred to as percentile flows. Each one
represents a segment of the flow duration curve and these 15 points approximately define the
shape of the FDC (Fig. 2). Figure 3 illustrates the water balance components of catchments from
the three physiographic provinces. As shown in Fig. 3, catchment elevation seems to have strong
influence on important hydrological processes. High-altitude Appalachian Plateau catchments
have annual hydrographs with higher peaks than those of the Ridge and Valley whereas the Ridge
and Valley catchments have annual hydrographs with peaks higher than those of the Piedmont
catchments.

Data analysis
A sequential data analysis approach was used that consists of the following steps: (1) construct
normalized FDCs from daily streamflow time series data for each study catchment; (2) select the
15 percentile flow points from the normalized FDC; (3) identify landscape–climate descriptors that
are good percentile flow predictors using a step-wise regression method; (4) build regional
percentile flow models for Appalachian Plateau, Ridge and Valley, and Piedmont province; (5) test
the method by predicting the 15 percentile flows for the ungauged evaluation sites, reconstruct
complete FDCs for the ungauged evaluation sites, and then reconstruct streamflow time series
from the reconstructed FDCs for the ungauged evaluation sites; and, finally, (6) evaluate the

Copyright ” 2008 IAHS Press


712 Yusuf M. Mohamoud

Flow duration curve Q0.1 Q0.5 Q1


Q5 Q10 Q20 Q30
Q40 Q50 Q60 Q70
Q80 Q90 Q95 Q99
Streamflow

Percent of time flow equaled or execeeded

12.40%
16.50%
21.20%
24.80%
29.70%
33.00%
37.40%
42.10%
45.40%
49.50%
53.90%
57.70%
62.40%
66.30%
70.90%
74.80%
78.30%
83.30%
89.60%
90.70%
95.10%
99.20%
0.00%
4.20%
8.30%
Normalized streamflow, duration curve, and percentile

1000

100
flows

10

1
1/1/1975

2/1/1975

3/1/1975

4/1/1975

5/1/1975

6/1/1975

7/1/1975

8/1/1975

9/1/1975
10/1/1974

11/1/1974

12/1/1974

Day
Fig. 2 Relationships between normalized flow duration curve, daily streamflow and the 15 percentile
flow points (Q0.1 to Q99) generated from a single year streamflow time series data.

140.00
Precipitation for a typical watershed

Streamflow for a typical Appalachian Plateau

120.00
Streamflow for a typical ridge-dominated Ridge and Valley

Streamflow for a typical Piedmont

100.00 Potential Evapotranspiration for a typical watershed


Water balance components (mm)

80.00

60.00

40.00

20.00

0.00
O N D J F M A M J J A S

Fig. 3 Regional water balance components.

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 713

prediction performance of the method by comparing reconstructed FDCs and streamflow with
observed FDCs and streamflow time series data.
Construction of normalized FDCs A normalized FDC was constructed from normalized
daily streamflow for each study catchment, following the Weibull plotting formula (Sugiyama et
al., 2003) (p = r/(N + 1) u 100), where p is the percentage of time that a given flow is equalled or
exceeded, N is the total number of data points in the period of record, and r is the rank assigned to
each streamflow value in the period of record.
Figure 2 illustrates how normalized daily streamflow, FDC and percentile flows are related,
and where each percentile flow point crosses the daily streamflow and the FDC. Specifically, it
illustrates how this relationship facilitates the reconstruction of FDCs and streamflow time series
from the 15 predicted percentile flow points.

Identification of landscape–climate descriptors Each regional FDC model consists of 15


percentile flows. Conceptually, each percentile flow represents a different segment of the FDC.
The proposed approach subdivides streamflow time series data into segments with nearly similar
magnitudes, thus developing separate models with separate parameters for each segment for most
of the time.
A step-wise regression method identified the landscape–climate descriptors that strongly
influence the dominant hydrological processes for the different segments. The landscape–climate
descriptor data described in Table 2 were used as independent variables and the 15 percentile
flows representing specific segments of the FDC were used as dependent variables. Equation (1)
relates each percentile flow to landscape and climate descriptors as:
Qn ax1b x2c (1)
where Qn is the drainage area-normalized percentile flow expressed as L s-1 km-2; a, b and c are
coefficient and exponents, respectively; and x1 and x2 are the landscape and climate descriptors,
respectively. A multiplicative model structure similar to those presented by Nathan & McMahon
(1990) and Vogel & Kroll (1992) was employed to represent each percentile flow. The multi-
plicative model structure was selected because it has been widely used to develop relationships
between specific flow indices, such peak and low flows and catchment descriptors. A hierarchical
approach that starts with an analysis of the correlation matrix of independent variables, followed
by an analysis of the correlation matrix of the dependent and the independent variables, followed
by step-wise regression analysis and ending with the development of equation (1) using a
nonlinear regression method was employed.
The step-wise method searched landscape–climate descriptors to select suitable percentile
flow predictors on the basis of pre-selected variable entry and removal criteria known as the
stepping method (SPSS Inc., 2003). The objective of using the automated method was to ensure
that the two independent variables selected for each model are not highly correlated with each
other. A tolerance statistic close to 1 ensured that the two descriptors in each model are not highly
correlated. Tolerance is a measure of collinearity among descriptors. In most cases, the final
parameters identified by the step-wise regression were those identified by the analysis of the
correlation matrix. For each percentile flow model, only the first two descriptors selected by the
step-wise regression method were retained (e.g. SOLD, KSAT). In equation (1), the descriptor that
explained most of the variation in percentile flow was expressed as x1 and the second best as x2.
The results of the analysis and the regional FDC models developed herein are presented in Table 3.
Reconstruction of a complete FDC and streamflow time series from percentile flow
models A sequential generation scheme was used to estimate all the unknown points on the FDC,
segment by segment, from the 15 known percentile flows. Note that all FDC points have ranks
already established and their corresponding percentile flows are determined when constructing the
FDCs. For example, for the period between 1 January 1980 and 31 December 2000, constructed

Copyright ” 2008 IAHS Press


714 Yusuf M. Mohamoud

Table 3 Percentile flow models developed for the Mid-Atlantic Region.


Percentile flow Regional FDC models R2
Appalachian Plateau:
Q0.1 7.31SOLD10.63 KSAT20.13 0.74
Q0.5 1251.5HPC500.04STDP-0.28 0.74
Q1 6578.93STDP-0.52KSAT1-0.055 0.88
Q5 5.3HMIN0.32AWC2-0.26 0.95
Q10 18.22(PET/MAP)-0.69ROCK10.09 0.91
Q20 8.55(PET/ MAP)-0.85AWC1-0.21 0.94
Q30 10.83(PET/MAP)-0.82KSAT20.087 0.93
Q40 1.17(PET/MAP)-1.32SAND20.36 0.99
Q50 0.84(PET/MAP)-1.57SAND20.32 0.98
Q60 0.50(PET/MAP)-1.89SAND20.31 0.98
Q70 0.243(PET/MAP)-2.17SAND20.34 0.97
Q80 3.88(PET/MAP)-2.12BREL-0.33 0.94
Q90 4.86(PET/MAP)-2.25BREL-0.48 0.94
Q95 4.73(PET/MAP)-2.2BREL-0.54 0.91
Q99 2.26(PET/MAP)-2.7BREL-0.58 0.90
Ridge and Valley:
Q0.1 4.40FRSD0.37AUGPREC0.58 0.81
Q0.5 0.15KSAT20.13FEBPREC1.58 0.95
Q1 0.22FEBPREC1.4KSAT20.14 0.89
Q5 0.54STDP0.45RRAT0.46 0.84
Q10 19.39HPC100.9RRAT0.27 0.82
Q20 17.63HPC101.27RRAT0.17 0.84
Q30 21.89HPC101.38AWCR-0.10 0.93
Q40 16.87HPC101.49AWCR0.682 0.94
Q50 13.00HPC101.56AWCR0.82 0.93
Q60 10.03HPC101.59AWCR0.99 0.91
Q70 7.21HPC101.54AWCR1.2 0.87
Q80 0.007AWCR2.15HPC501.64 0.85
Q90 0.0013AWCR2.67HPC502.02 0.85
Q95 0.0058AWCR2.54ROCK21.67 0.81
Q99 0.00082AWCR3.42ROCK22.14 0.86
Piedmont :
Q0.1 305.68(PET/MAP)2.01RRAT0.14 0.90
Q0.5 49.69KSAT10.29KSAT20.17 0.98
Q1 2549.79HMED-0.80FRSD0.25 0.80
Q5 24893.89SAND1-1.00SILT2-0.76 0.95
Q10 738.69SAND1-1.11BREL0.12 0.94
Q20 0.056SAND1-0.93SEPPREC1.94 0.87
Q30 11.14SAND1-1.53AUGPREC1.24 0.94
Q40 3.64SAND1-1.58AUGPREC1.48 0.95
Q50 1.44SAND1-1.58AUGPREC1.64 0.95
Q60 2.77x10-4SOLD11.84RRAT0.24 0.96
Q70 1.1x10-4SOLD11.98RRAT0.22 0.98
Q80 4.96x10-5SOLD12.11RRAT0.163 0.99
Q90 1.5x10-4SOLD11.94KSAT1-0.08 0.97
Q95 9.6x10-5SOLD12.05KSAT1-0.23 0.97
Q99 3.6x10-4SOLD11.89KSAT1-0.61 0.84

FDCs consist of 7671 ranked and normalized percentile flow values, of which only 15 are
predicted (known). The sequential generation for any segment can be written as:
Qr  Qrn
qrk 1 qrk  k (2)
rn  rk
where q rk 1 is the first sequentially estimated unknown percentile flow within the segment, Qrk is

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 715

the known percentile flow at the upper end of the segment starting point, Qrn is a known
percentage flow at the lower end of the segment end point, and rk and rn are their respective ranks.
Note that qrk = Qrk when calculating the first unknown percentile flow within each segment and
the ratio (Qrk  Qrn ) /( rn  rk ) is constant within each segment. The scheme iteratively estimates the
next percentile flow by using the percentile flow estimated in the previous step. For example, qrk 1
substitutes qrk when calculating qr k  2 in the second iteration. All unknown (in-between) predicted
percentile flows of the FDC were reconstructed by applying this scheme to each segment. After all
the FDCs were reconstructed from the normalized FDCs, the normalized FDCs were multiplied by
the drainage area of the catchment to convert the FDC into m3/s. The FDC conversion to a
complete streamflow time series was achieved by reshuffling the predicted FDC values using
streamflow time sequence from a nearby gauged catchment. The reshuffling procedure sorts the
predicted FDC and the streamflow time series from the source site by magnitude. Note that both
have date and magnitude columns and both are not ranked in chronological order, but rather in
magnitude order. The predicted FDC magnitude column was copied to the magnitude column of
the streamflow time series column of the source site leaving the date column unchanged. Once
copied, the two columns were re-sorted, this time using the date column. The assumption here is
that neighbouring sites have similar streamflow sequences since they receive similar precipitation
inputs and had similar catchment characteristics; and that days with high flows for the source site
correspond to days with high flows for the destination site and vice versa.
The FDC reconstruction must ensure that percentile flows estimated for higher magnitudes,
i.e. Q5, are always larger than those estimated for lower magnitude percentile flows, i.e. Q10.
However, in some cases, percentile flows estimated for Q5 could be lower than those estimated for
Q10. This can happen if the coefficients of determination for the percentile flow model is low, or if
the regional model was developed from catchments with diverse landscape characteristics. When
this happens, one should ignore that specific percentile flow and instead use the next lower
segment percentile flow using a two-segment step instead of a one-segment step sequential
generation scheme.
Evaluation method Three evaluation sites that were not used for the regional FDC parameter
estimation were randomly selected, one each from the Appalachian, Ridge and Valley, and
Piedmont province. All three sites have long records of streamflow data, but for FDC method
evaluation purposes, they were considered to be ungauged. The ungauged site is denoted as the
destination site throughout the paper. Downstream from, or adjacent to, each destination site, a
source site with long-term daily streamflow record was also selected, and streamflow from the
source site was transferred to the destination site using the FDC and drainage area ratio methods.
Evaluation consisted of comparing the reconstructed (i.e. predicted) FDCs and streamflow
time series data, plus the streamflows transferred by drainage area ratio methods, with the
observed FDCs and streamflow time series data from the three selected evaluation catchment sites.
These comparisons were intended to demonstrate the predictive performance of the FDC method
and its suitability for predicting streamflow in ungauged basins. Performance evaluation criteria,
or goodness-of-fit, were the Nash-Sutcliffe efficiency (E) (Nash & Sutcliffe, 1970), the coefficient
of determination (R2) and the root mean square error (RMSE). The predictive performance of the
FDC and the drainage area ratio methods, for both daily and monthly simulations, was evaluated
for the 21-year period of record. In addition to goodness-of-fit comparisons, serial correlation
(lag 1), mean, standard deviation, skewness coefficient of streamflow predicted by the FDC,
drainage area ratio methods and observed streamflow were calculated and compared. A brief
discussion of the modified or nonlinear drainage area ratio method developed by Mohamoud &
Parmar (2006) for streamflow extrapolation from gauged to ungauged sites is presented in the
following section.

Copyright ” 2008 IAHS Press


716 Yusuf M. Mohamoud

Drainage area ratio methods The standard drainage area ratio method is the most straight-
forward technique used for transferring streamflow from gauged sites to ungauged sites (Smakhtin
& Masse, 2000). In this study, the standard drainage area ratio method (Stedinger et al., 1993) and
the nonlinear drainage ratio methods proposed by Mohamoud & Parmar (2006) were used. The
standard method is based on the assumption that the ratio of streamflows of the gauged and the
ungauged sites are equal to the ratios of their drainage areas (equation (3)). The nonlinear drainage
area ratio methods are not based on the linearity assumption and are intended to address some
limitations of the standard method. The drainage area ratio equations are written as follows:
Qu = Qg (Au/Ag) (3)
Qu = Qg tan(Au/Ag) (4)
Qu = Qg arctan(Au/Ag) (5)
where Qu is the estimated daily streamflow at the ungauged or destination site, Qg is daily
streamflow at the gauged site or source site, and Ag and Au are the drainage areas of the gauged and
the ungauged sites, respectively. The standard method works well when the ratio of the source and
the destination site drainage areas is between 0.5 and 1.5 (Hortness, 2006). If the area ratio is less
than 0.5 or the linearity assumption is invalid, equations (4) and (5) can be used as alternative
methods. The nonlinear methods estimate streamflow as either single-step or multi-step calcula-
tions. For example, when using the Rappahannock River as the source site and the Robinson River
as the destination site, one can calculate flow using equations (4) and (5) in a single step, as
tan(213/737), or in multi-steps first as tan(400/737) to calculate predicted flow at an intermediate
point and then as tan(213/400) to calculate flow from the intermediate flow to the final flow. A
detailed discussion of the nonlinear drainage area ratio methods is given in Mohamoud & Parmar
(2006). Note that the single step uses the entire drainage area of a catchment in one calculation,
whereas the multi-step method uses intermediate drainage ratio to calculate an intermediate flow
and then uses the intermediate flow to calculate a final flow.

RESULTS AND DISCUSSION


How meaningful are the landscape–climate descriptors identified by the step-wise regression
method?
Flow duration curves represent the magnitude component of streamflow over the period of record.
The hypothesis here is that each percentile flow point is influenced by a set of unknown
landscape–climate parameters. That is to say, each percentile flow point characterizes a hydro-
logical condition governed by identifiable hydrological processes, which operate in a catchment on
days when it experiences very wet, wet, moist, dry, or very dry conditions. A major challenge is
how to quantify the physical relationships between landscape–climate descriptors and percentile
flows.
Evaluating the relevance of the landscape–climate descriptors identified for the different
percentile flows requires some a priori knowledge of these relationships. Unfortunately, these
relationships are not well established at the catchment scale. In the absence of this knowledge of
relationships, generalizations were made based on limited information gleaned from the scientific
literature and a preliminary analysis of the results of this study. In the following sections,
relationships between landscape–climate descriptors and percentile flows are examined. Examina-
tion of the descriptors for the different physiographic provinces revealed that climate descriptors
are the dominant predictors for the Appalachian Plateau whereas soil descriptors are the dominant
predictors for the Piedmont; geomorphological descriptors are the best percentile flow predictors
for the Ridge and Valley catchments.
Step-wise regression analysis identified landscape–climate descriptors that are good predictors
of percentile flows (Table 3). However, as Sefton & Howarth (1998) pointed out, relationships
identified by step-wise regression are purely empirical and, even though statistically significant,

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 717

some of the relationships may be an accident of the data. While some of the relationships have
strong physical significance, others may be surrogate predictors or may represent process
interactions that cannot be explained by current knowledge of the relationships between hydro-
logical processes and parameters controlling them at the catchment scale. When examining the
possible physical significance of the landscape–climate descriptors identified by the step-wise
regression analysis for the 15 percentile flows, the approach followed was to group the descriptors
into high flows (Q0.1, Q0.5, Q1, Q5, Q10), median flows (Q20, Q30, Q40, Q50, Q60) and low flows (Q70,
Q80, Q90, Q95, Q99).
Q0.1, Q0.5, Q1, Q5 and Q10 percentile flows High flows are most often caused by intense
rainstorms or snow-melt, but climate descriptors that represent short-term precipitation signals
were not included in the list of descriptors (Table 2). Consequently, the descriptors identified by
the step-wise regression for high flows are most likely surrogate descriptors because high flows
are closely linked to short-term rainfall or snow-melt. Descriptors identified for the high flows
included soil depth, hypsometric curve characteristics, saturated hydraulic conductivity, precipita-
tion in the months of August and February, percentage of catchment area as deciduous forest,
relief ratio, dryness index, percentage of soil sand and silt and minimum and median elevation.
Q20, Q30, Q40, Q50 and Q60 percentile flows The dominant predictors identified for the median
flows varied from province to province. The best predictors for the Appalachian Plateau, the Ridge
and Valley and the Piedmont catchments were dryness index, that proportion of hypsometric
elevation that corresponds to 0.1 proportion of the catchment area, and the percent sand in soil
layer 1, respectively. These three descriptors were repeatedly selected throughout the median flow
range (Table 3). In general, those descriptors identified first by the regression analysis explain
much of the variability of the percentile flows. One can, therefore, assume that those repeatedly
selected as first predictors have more physical relevance than those selected as second predictors,
or selected just once or twice as first predictors. Other descriptors that are good predictors of
median flows included relief ratio, ratio of available water capacity of soil layers 1 and 2, August
precipitation and percentage of sand in soil layer 2.
Q70, Q80, Q90, Q95 and Q99 percentile flows The best low flow predictors identified for the
Appalachian Plateau, the Ridge and Valley and the Piedmont catchments were dryness index, ratio
of available water capacity of the two soil layers and the soil depth of layer 1, respectively. Other
important low flow predictors included baseflow index, basin relief, hypsometric characteristics,
percentage of rock in soil layer 2, hydraulic conductivity of soil layer 1 and relief ratio. Generally,
precipitation influence may be low since low flows most likely come from soil, groundwater or
channel storage releases. As a result, baseflow index was found to be a good low flow predictor.
However, baseflow index was not included in any of the regional FDC models because its value is
determined from observed streamflow that is not available in ungauged catchments.

Evaluation results
Predictive performance Predictive performance evaluations were conducted for both FDC
and streamflow predictions. The FDC predictions concerned only magnitude, whereas the stream-
flow predictions concerned both magnitude and time sequence. Using the regional FDC models
developed in Table 3, an FDC was reconstructed from the predicted 15 percentile flows for each of
the three evaluation catchment sites. Figure 4 shows comparisons of observed and predicted flow
percentiles for three evaluation sites, while Fig. 5 shows a comparison of the reconstructed and
observed FDC for one of the evaluation catchment sites. The overall accuracy in predicting
streamflow for ungauged basins greatly depends on the accuracies of the predicted FDC magnitude
and time sequence.
The periodic properties of the reconstructed streamflow statistics, such as mean, standard
deviation and skewness coefficient representing streamflow magnitude, generally closely matched
the corresponding observed streamflow statistical properties (Table 4). Comparisons of lag 1 serial

Copyright ” 2008 IAHS Press


718 Yusuf M. Mohamoud

440

Observed Flow Percentiles (l/sec/km )


390
2
Normalized streamflow (L s-1 km-2)

340

290

240

190 2
R = 0.99

140

90

40

-10
-10 90 190 290 390
-1 -2
Predicted flow percentiles (L s km )
Fig. 4 Predicted and observed 15 flow percentiles for three evaluation sites

1000
Constructed 21 Year Flow Duration Curve
) s-1 km-2)

Observed 21 year Flow Duration Curve

Estimated 15 Percentile Flow Points Predicted by Regional Regression


100
(l/sec/km(L
2
flow percentiles

Robinson River, Virginia


Piedmont Physiographic Province
Streamflow

10
Observed
Normalized

1
0 10 20 30 40 50 60 70 80 90 100

0.1
Percentage of time flow equalled or exceeded
Fig. 5 Comparison of observed and predicted flow duration curves of a Piedmont catchment.

correlation was used to compare streamflow sequence related statistics generated from the FDC
and drainage area ratio prediction methods with those of the observed streamflow. The FDC
method predictions generally agreed with observed streamflow and also with streamflow predicted
by the drainage area ratio methods. Note that the FDC method uses only streamflow sequence
from a source site to reconstruct streamflow time series, but the drainage area ratio method uses
both streamflow sequence and magnitude from a source site. For the drainage area ratio method,
therefore, both the predicted sequence and magnitude related statistical properties change
whenever a new source site is used (Table 4). Conversely, the FDC method retains its own
magnitude that is determined from the regional FDC models; however, its predicted sequence
changes whenever a different source site is used as shown by changes in lag 1 serial correlation.

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 719

Table 4 Statistical properties of a 21-year observed and predicted daily streamflow data from test and source
sites.
Catchment Run USGS gauge Area km2 Lag 1 serial Mean St. Dev. Skewness
correlation coefficient
Johns Creek Observed 02017500 124 0.74 3.80 6.33 5.9
(Destination)
Craig Creek Observed 02018000 391 0.69 3.53 6.80 9.9
(Source)
Craig Creek Standard 02018000 391 0.69 3.53 6.80 9.9
(Source)
Craig Creek FDC 02018000 391 0.76 3.69 6.73 4.8
(Source)
Johns Creek FDC 02017500 124 0.76 3.69 6.73 4.8
(Source)
Little Mohaning Observed 03034500 104 0.66 4.15 6.47 5.7
(Destination)
Mahoning Creek Observed 03034000 189 0.71 4.34 6.15 7.7
(Source)
Mahoning Creek Tangent 03034000 189 0.71 4.34 6.15 7.7
(Source)
Mahoning Creek FDC 03034000 189 0.80 3.87 6.43 3.5
(Source)
Little Mahoning FDC 03034500 104 0.74 3.87 6.43 3.5
(Source)
Robinson River Observed 01664000 213 0.57 6.71 13.53 12.9
(Destination)
Rapidan River Observed 01667500 561 0.60 6.22 12.06 15.7
(Source)
Rappahannock Observed 01666500 737 0.65 6.00 10.48 8.7
River (Source)
Rapidan River Tangent 01667500 561 0.60 6.22 12.06 15.7
(Source)
Rappahannock Tangent 01667500 561 0.65 6.00 10.48 8.9
River (Source)
Rappahannock FDC 01666500 737 0.62 6.97 11.69 8.2
River (Source)
Rapidan River FDC 01667500 561 0.65 6.97 11.69 8.2
(Source)
Robinson Creek FDC 01664000 213 0.58 6.97 11.69 8.2
(Source)

The FDC method’s streamflow predictive performance is greatly influenced by the separation
distance between source (gauged) and destination (ungauged) sites. The Rappahannock and the
Rapidan sites were used to demonstrate the effect of separation distance on the predictive
performance of streamflow as predicted by the drainage area ratio and FDC methods. As shown in
Table 5, the predictive performance of the FDC and drainage area ratio methods for the Robinson
River was low when the Rappahannock was the source site, but increased when the source site was
switched to the Rapidan. The Rappahannock site was not used for any predictive performance
comparisons included herein since it is far from the Robinson site; it was used only to illustrate the
role that separation distance can play in the predictive performance of the two methods. Separation
distance may act as a surrogate for precipitation over the source and destination catchments. When
gauged sites are available near the ungauged site of interest, practitioners must select the source
site that is nearest to the destination site. However, when gauged sites are not available near the
ungauged site, practitioners are forced to select sites that have long separation distances. In such
cases, the method proposed herein may not accurately predict streamflow.
The predictive performance relative to the reconstructed streamflows of the three evaluation
catchments using the Rapidan, Craig and Mahoning as source sites was high (Table 5). The
coefficients of determination R2 ranged from 0.80 to 0.86 for the 21-year daily simulations and
from R2 = 0.88 to 0.98 for the 21-year monthly simulations (Table 5). The predictive performance

Copyright ” 2008 IAHS Press


720 Yusuf M. Mohamoud

Table 5 Performance evaluation of the FDC based method predictions and comparisons with other stream-
flow estimation methods for ungauged catchment for a 21-year period.
Source site Transfer method Daily predictions: Monthly predictions:
E R2 RMSE E R2 RMSE
Robinson River, Virginia (destination site)
Rappahannock River Tangent 0.34 0.61 8.49 0.83 0.86 2.42
Rappahannock FDC 0.46 0.60 8.50 0.85 0.85 2.47
Rapidan River Tangent 0.80 0.82 5.71 0.97 0.98 0.96
Rapidan FDC 0.77 0.83 5.58 0.95 0.95 1.35
Robinson (destination) FDC 0.88 0.92 4.08 0.97 0.97 1.11
Little Mahoning Creek, Pennsylvania (destination site)
Mahoning Tangent 0.81 0.82 2.97 0.86 0.90 1.37
Mahoning FDC 0.79 0.80 3.03 0.87 0.88 1.27
Little Mahoning (destination) FDC 0.90 0.91 2.01 0.97 0.99 0.52
Johns Creek, Virginia (destination site)
Craig Creek Standard 0.82 0.82 2.87 0.95 0.97 0.79
Craig Creek FDC 0.85 0.86 2.56 0.98 0.98 0.58
Johns Creek (destination) FDC 0.96 0.97 1.28 0.99 1.0 0.34

400
Observed 21-year streamflow of Robinson River, Virginia
350 Converted streamflow using time sequence from the Rapidan River
300 Nas h - Sutcliffe
Streamflow (m 3/sec)

Efficie ncy E = 0.77


250
R2 = 0.83
200

150

100

50

0
80

82

84

86

88

90

92

94

96

98

00
19

19

19

19

19

19

19

19

19

19

20
1/

1/

1/

1/

1/

1/

1/

1/

1/

1/

1/
1/

1/

1/

1/

1/

1/

1/

1/

1/

1/

1/

Day

Fig. 6 Comparison of observed and predicted streamflow for a 21-year period.

of the drainage ratio method was comparable to that of the FDC method when the same source site
was used. If, instead of using the source site sequence, the sequence of the destination site
(formerly assumed to be ungauged) was used, all sequence related errors were eliminated and the
predictive performance of the FDC method was improved.
Figure 6 plots a comparison of FDC method reconstructed vs observed daily streamflow over
a 21-year period using the Robinson River as the destination site and the Rapidan River as the
source site. This comparison demonstrates the general suitability of the FDC method for long-term
streamflow prediction for ungauged catchments. Figure 7 illustrates how the FDC and drainage
area ratio methods work for short-duration streamflow hydrograph simulations. When both
methods used Mahoning Creek as the source site and Little Mahoning as the destination site, both
methods showed magnitude and sequence errors. The FDC method exhibited less predictive

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 721

70

Observed (Little Mahoning) Little Mahoning Creek, Pennsylvania


60 Appalachian Plateau Province
Tangent Method (Mahoning)

FDC (Mahoning Sequence)


50
Streamflow (m /sec)

FDC - Little Mahoning Sequence


3

40

30

20

10

0
7/ 0
4/ 80

4/ 80

10 0

11 0

17 0
12 0

13 0

14 0

16 0

18 0

19 0

20 0

21 0

22 0

23 0

0
15 0
8

4/ 19 8

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98

4/ 98
98
4/ 98
19

19

19

/1

/1

/1

/1

/1

/1

/1

/1

/1

/1

/1

/1

/1

/1
8/

9/
6/
4/

4/

Day
Fig. 7 Comparison of observed and predicted streamflow hydrographs showing the effect of time
sequence on prediction performance of the FDC method.

uncertainties than the drainage area ratio-methods because magnitude and sequence are separately
determined. To illustrate, when the streamflow sequence of the Little Mahoning (formerly the
destination site) was used as the source site, the predicted hydrograph shifted to the left to match
the observed streamflow, thus correcting the sequence error associated with using the Mahoning as
the source site sequence (Fig. 7).

Sources of predictive uncertainty The proposed FDC method is a new approach that views
the predicted streamflow time series as a combination of magnitude and sequence components. In
this approach, the FDC predominantly represents the magnitude component, estimated from time-
invariant, landscape–climate descriptors. Unlike the magnitude component, the sequence com-
ponent represents short-term stochastic factors, such as rainfall, and gives the magnitude com-
ponent its time sequence. Accordingly, the FDC method divides streamflow predictive uncer-
tainties into magnitude and sequence-related uncertainties. For instance, magnitude-related
uncertainties stem from errors related to regional FDC model parameter estimation or errors
associated with the use of the sequential generation scheme. Streamflow sequence-related uncer-
tainties are mainly caused by a possible mismatch between the sequential orders of the source
(gauged) and destination (ungauged) catchments. The sequence-related uncertainties are indepen-
dent of the magnitude component and therefore of the regression method. Streamflow sequence
errors associated with different source sites can be quantified by using the destination site as the
source site (Table 5). As shown in the destination rows of the three evaluation sites in Table 5,
using the destination site as the source site removes all the streamflow sequence errors and
quantifies the errors associated with the use of any source site. All remaining predictive
uncertainties, after the destination sites were used as source sites, are magnitude-related. Hence,
when sequence errors are eliminated, the coefficients of determination are 0.92, 0.91 and 0.97 for
Robinson, Mahoning and Johns Creek, respectively. The remaining unexplained variances
between predicted and observed streamflow is magnitude error.

Copyright ” 2008 IAHS Press


722 Yusuf M. Mohamoud

Future research Two research areas would greatly improve the proposed method. First,
research on statistical methods that can identify landscape–climate descriptors with physical
relevance at the catchment scale is needed. As shown by Wallis (1965), regression methods may
not identify descriptors that have physical relevance without prior knowledge of the relationships
between the landscape–climate descriptors and hydrological processes. Lack of knowledge of
those descriptors that strongly influence hydrological processes represented by specific percentile
flows casts doubt on the physical significance of the predictors identified for a given percentile
flow. Research that links the regression equations to conceptual models, similar to the approach
used by Vogel & Kroll (1992), to suggest the key variables (descriptors) and the functional form
for the regional regression equations, or the water balance-based approaches proposed by Woods
(2003) and Jothityangkoon et al. (2001), may, indeed, reveal the physical significance of some
descriptors identified herein for the 15 percentile flows.
Second, field research is needed to identify new descriptors that closely represent the
hydrological conditions characterized by the 15 percentile flows. For instance, the approach
employed here to select the best predictors for the 15 percentile flows can only select descriptors
that are included in the list of explanatory variables. It is important to search for new descriptors or
parameters that might better reflect how landscape–climate properties influence hydrological
processes at the catchment scale. Unless the descriptors reflect actual hydrological processes as
they operate at the catchment scale, the use of available landscape–climate descriptors will not
resolve the parameter identifiability problem.
For example, saturated hydraulic conductivity values obtained from county survey reports
represent the effects of interstices on water transmission, but may not account for the effect of
macropore flow (Zacharias & Brutseart, 1988) and lateral hydraulic conductivity. Lin (2006)
reported that approximately two-thirds of the soil horizon measured in a Ridge and Valley
catchment in Pennsylvania had lateral hydraulic conductivity (KSAT) values 1.5 to 142.5 times
greater than vertical values. Elsewhere, Brooks et al. (2004) estimated lateral saturated hydraulic
conductivity at the hillslope to be about one or two magnitudes larger than the small-scale, vertical
saturated hydraulic conductivity measured from soil samples. Clearly, the list of descriptors in
Table 2 is incomplete and many relevant ones, such as those that influence subsurface flow
processes including lateral hydraulic conductivity, are absent.

CONCLUSION
This study presents a method to predict FDC and streamflow time series for ungauged streams in
the Mid-Atlantic Region, USA. The results indicated that climate, geomorphologic and soil
descriptors strongly influence the hydrology of the Appalachian Plateau, Ridge and Valley and
Piedmont provinces. The dominant effective climate descriptors were annual dryness index, and
mean monthly precipitation for April, August and February; hypsometric elevation ratios, channel
slope and median catchment elevation were the dominant geomorphological descriptors that were
good predictors of percentile flows. The dominant soil descriptors identified were available water
capacity, soil depth and soil texture classes. Another percentile flow predictor revealed was
baseflow index that was considered as a surrogate descriptor. The non-uniqueness problem that
affects calibrated rainfall–runoff models is also a problem in the proposed method since many
landscape–climate descriptors can be good predictors of a given percentile flow.
Reconstructed (predicted) FDCs and streamflow time series for test catchments closely
matched observed values and predictions using drainage area ratio methods. This suggests that the
proposed FDC method is suitable for FDC and streamflow time series prediction in ungauged
basins. Although the approach was developed using data from the Mid-Atlantic Region, USA, the
method can be used anywhere in the world if regional regression models are established following
the approach presented herein.
The separation of predicted streamflow into magnitude and sequence components and
evaluation of the associated uncertainties in each have important implications for advancing

Copyright ” 2008 IAHS Press


Prediction of daily flow duration curves and streamflow for ungauged catchments 723

process-based hydrological models at the catchment scale. For example, streamflow time series
sequence errors can be interpreted as errors related to model input variables, such as rainfall input
data. Similarly, magnitude prediction errors can be interpreted as model structure and model
parameter uncertainties. The FDC method only takes streamflow sequence from a source site if
available; if streamflow is unavailable from a nearby source site, the FDC method may obtain its
sequence from daily rainfall or soil moisture content time series. Obtaining time sequences from
rainfall or soil moisture is beyond the scope of this paper.
This study has potential implications for the International Association of Hydrological
Sciences (IAHS) decadal initiative on Prediction in Ungauged Basins (PUB). The FDC method’s
ability to separate streamflow into magnitude and sequence components makes it highly flexible
and suitable for many water resources planning and management applications, including
hydrological forecasting and streamflow prediction in ungauged basins.

Acknowledgements The author acknowledges the helpful comments provided by Dr Robert


Swank. This paper has been reviewed in accordance with the US Environmental Protection
Agency’s peer and administrative review policies and approved for publication. Mention of trade
names or commercial products does not constitute endorsement or recommendation.

REFERENCES
Arnold, J .G., Allen, P. M., Muttiah, R. & Bernhardt, G. (1995) Automated base-flow separation and recession analysis
techniques. Ground Water 33, 1010–1018.
Berger, K. P. & Entekhabi, D. (2001) Basin hydrological response relations to distributed physiographic descriptors and
climate. J. Hydrol. 247, 169–182.
Brooks, E. S., Boll, J. & McDaniel, P. A. (2004) A hillslope-scale experiment to measure lateral hydraulic conductivity. Water
Resour. Res. 40, W04208.
Budyko, M. I. (1974) Climate and Life. Academic Press, San Diego, California, USA.
Castellarin, A., Galeati, G., Brandimarte, L., Montanari, A. & Brath, A. (2004) Regional flow-duration curves: reliability for
ungauged basins. Adv. Water Resour. 27, 953–965.
Claps, P., Giordano, A. & Laio, F. (2005) Advances in shot noise modeling of daily streamflows. Adv. Water Resour. 28,
992–1000.
Croker, K. M., Young, A. R., Zaidman, M. D. & Rees, H. G. (2003) Flow duration curve extreme events. Hydrol. Sci. J. 48(3),
427–439.
Fenneman, N. M. (1938) Physiography of Eastern United States, McGraw-Hill, New York, USA.
Fennessey, N. & Vogel, R. M. (1990) Regional flow-duration curves for ungauged sites in Massachusetts. J. Water Resour.
Plann. Manag. 116, 530–549.
Hamon, W. R. (1961) Estimating potential evapotranspiration. J. Hydraul. Div. Proc. ASCE 87, 107–120.
Hortness, J. E. (2006) Estimating low flow frequency statistics for unregulated streams in Idaho. US Geol. Survey. Sci. Invest.
Report 2006-5035.
Hughes, D. A. & Smakhtin, V. Y. (1996) Daily flow time series patching or extension: spatial interpolation approach based on
flow duration curves. Hydrol. Sci. J. 41(6), 851–871.
Jothityangkoon, C., Sivapalan, M. & Farmer, D. (2001) Process controls of water balance variability in a large semi-arid
catchment: downward approach to hydrological model development. J. Hydrol. 254, 174–198.
Laaha, G. & Blöschl, G. (2007) A national low flow estimation procedure for Austria. Hydrol. Sci. J. 52(4), 625-644.
Leboutillier, D. W. & Waylen, P. R. (1993) A stochastic-model of flow duration curves. Water Resour. Res. 29, 3535–3541.
Lin, H. (2006) Temporal stability of soil moisture spatial pattern and subsurface preferential flow pathways in shale hillslope.
Vadose Zone J. 5, 317–340.
Linsley, R. K., Jr, Kohler, M. A. & Paulhus, L. J. H. (1949) Applied Hydrology. McGraw-Hill, New York, USA.
Maxey, G. B. (1964) Handbook of Hydrology. McGraw-Hill, New York, USA.
McNamara, J. P., Kane, D. L. & Hinzman, L. D. (1998) An analysis of streamflow hydrology in the Kuparuk River basin,
Arctic Alaska: a nested catchment approach. J. Hydrol. 206, 39–57.
Milly, P. C. D. (1994) Climate, soil-water storage, and the average annual water-balance. Water Resour. Res. 30, 2143–2156.
Mimikou, M. & Kaemaki, S. (1985) Regionalization of flow duration characteristics. J. Hydrol. 82, 77–91.
Mohamoud, Y. M. & Parmar, R. S. (2006) Estimating streamflow and associated hydraulic geometry, the Mid-Atlantic Region,
USA. J. Water Resour. Assoc. 42, 755–768.
Nash, J. E. & Sutcliffe, J. V. (1970) River forecasting through conceptual model, Part 1: A discussion of principles. J. Hydrol.
10, 282–290.
Nathan, R. J. & McMahon, T .A. (1990) Identification of homogeneous regions for the purposes of regionalisation. J. Hydrol.
121, 217–238.
Pearce, A. J. (1990) Streamflow generation processes: an austral view. Water Resour. Res. 26, 3037–3047.
Post, D. A. & Jones, J. A. (2001) Hydrologic regimes at four long-term ecological research sites in New Hampshire, North
Carolina, Oregon, and Puerto Rico. Adv. Water Resour. 24, 1195–1210.

Copyright ” 2008 IAHS Press


724 Yusuf M. Mohamoud

Quimpo, R. G., Alejandrino, A. A. & McNally, T. A. (1983) Regionalized flow duration for Philippines. J. Water Resour.
Plann. Manag. 109, 320–330.
Reimers, W. (1990) Estimation hydrological parameters from basin characteristics for large semi-arid catchments; In:
Regionalization in Hydrology (ed. by M. A. Beran, M. Brilly, A. Becker & O. Bonacci) (Proc. Ljubljana Symp.,
April 1990), 187–194. IAHS Publ. 191. IAHS Press, Wallingford, UK. http://iahs.info/redbooks/a191/iahs_191_0187.pdf.
Searcy, J. K. (1959) Flow duration curves, manual of hydrology. Part 2, Low flow techniques. US Geol. Survey Water Supply
Paper 1542– A.
Sefton, C. E. M. & Howarth, S. M. (1998) Relationships between characteristics and physical descriptors of catchments Wales.
J. Hydrol. 211, 1–16.
Shruben, P. G., Arndt, R. E. & Bawiec, W. J. (1998) Geology of the conterminous United States at 1, 250,000 scale, a digital
representation of the 1974 P. B. King & H. M. Beikman Map. US Geol. Survey Data Series 001, Release 2.
Singh, R. D., Mishra, S. K. & Chowdhary, H. (2001) Regional flow-duration models for large number of ungauged Himalayan
catchments for planning microhydro projects. J. Hydrol. Engng 6, 310–316.
Slack, J. R. & Landwehr, J. M. (1992) Hydro-Climatic Data Network: A US Geological Survey streamflow data set for the
United States for the study of climate variations 1874–1988. US Geol. Survey Open-File Report 91-129.
Smakhtin, V. Y. & Masse, B. (2000) Continuous daily hydrograph simulation using duration curves of a precipitation index.
Hydrol. Processes 14, 1083–1100.
Smakhtin, V. Y., Hughes, D. A. & Creuse-Naudin, E. (1997) Regionalization of daily flow characteristics in part of the Eastern
Cape, South Africa. Hydrol. Sci. J. 42, 919–936.
SPSS Inc. (2003) SPSS for Windows Version 11.5 User’s Guide. SPSS Inc. Chicago, USA.
Stedinger, J. R., Vogel, R. M. & Foufoula-Georgiou, E. (1993) Frequency analysis of extreme events. In: Handbook of
Hydrology (ed. by D. Maidment), 18.1–18.66. McGraw-Hill, New York, USA.
Sugiyama, H., Vudhivanich, V., Whitaker, A. C. & Lorsirirat, K. (2003) Stochastic flow duration curves for evaluation of flow
regimes in rivers. J. Am. Water Resour. Assoc. 39, 47–58.
USGS (United States Geological Survey) (1992–1997) Groundwater Atlas of the United States. USGS Reston, Virginia, USA.
USDA (United States Department of Agriculture) (1993) State soils geographic database users guide. Soil Conservation Service
Misc. Publ. 1492.
USEPA (United States Environmental Protection Agency) (2004) BASINS: Better Assessment Science Integrating Point and
Nonpoint Sources: available at: http://www.epa.gov/OST/BASINS/
Vogel, R. M. & Fennessey, N. M. (1995) Flow duration curves. 2: A review of applications in water resources planning. Water
Resour. Bull. 31, 1029–1039.
Vogel, R. M. & Kroll, C. N. (1992) Regional geohydrologic-geomorphic relationships for the estimation of low-flow statistics.
Water Resour. Res. 28, 2451–2458.
Wagener, T. & Wheater, H. S. (2006) Parameter estimation and regionalization for continuous rainfall–runoff models including
uncertainty. J. Hydrol. 320, 152–154.
Wallis, J. R. (1965) Multivariate statistical methods in hydrology: a comparison using data of known functional relationship.
Water Resour. Res. 1(4), 447–461.
Woods, R. (2003) The relative roles of climate, soil, vegetation and topography in determining seasonal and long-term
catchment dynamics. Adv. Water Resour. 26, 295–309.
Yu, P. S. & Tang, T. C. (2000) Using synthetic flow duration curves for rainfall–runoff model calibration at ungauged sites.
Hydrol. Processes 14, 117–153.
Yu, P. S., Yang, T. C. & Wang, Y. C. (2002) Uncertainty analysis of regional flow duration curves. J. Water Resour. Plann.
Manag. ASCE 128, 424–430.
Zacharias, Y. B. & Brutsaert, W. (1988) The influence of basin morphology on groundwater outflow. Water Resour. Res. 24,
1645–1650.

Received 23 May 2007; accepted 22 February 2008

Copyright ” 2008 IAHS Press

You might also like