Journal of Hydrology 306 (2005) 127–145

www.elsevier.com/locate/jhydrol

Sensitivity and uncertainty analysis coupled with automatic
calibration for a distributed watershed model
Misgana K. Muletaa,b,*, John W. Nicklowb,1
a
MWH Soft, Inc., 300 North Lake Avenue, Suite 1200, Pasadena, CA 91101, USA
b
Department of Civil and Environmental Engineering, Southern Illinois University at Carbondale, Carbondale, IL 62901-6603, USA
Received 20 February 2004; revised 31 August 2004; accepted 10 September 2004

Abstract
Distributed watershed models should pass through a careful calibration procedure before they are utilized as a decision
making aid in the planning and management of water resources. Although manual approaches are still frequently used for
calibration, they are tedious, time consuming, and require experienced personnel. This paper describes an automatic approach
for calibrating daily streamflow and daily sediment concentration values estimated by the US Department of Agriculture’s
distributed watershed simulation model, Soil and Water Assessment Tool (SWAT). The automatic calibration methodology
applies a hierarchy of three techniques, namely screening, parameterization, and parameter sensitivity analysis, at the parameter
identification stage of model calibration. The global parameter sensitivity analysis is conducted using a stepwise regression
analysis on rank-transformed input–output data pairs. Latin hypercube sampling is used to generate input data from the assigned
distributions and ranges, and parameter estimation is performed using genetic algorithm. The Generalized Likelihood
Uncertainty Estimation methodology is subsequently implemented to investigate uncertainty of model estimates, accounting
for errors due to model structure, input data and model parameters. To demonstrate their effectiveness, the parameter
identification, parameter estimation, model verification, and uncertainty analysis techniques are applied to a watershed located
in southern Illinois.
q 2004 Elsevier B.V. All rights reserved.

Keywords: Sensitivity analysis; Automatic calibration; Uncertainty analysis; Model verification; Genetic algorithms; Distributed watershed
model

1. Introduction

Hydrologic models are particularly useful tools in
that they enable us to investigate many practical and
* Corresponding author. Address: MWH Soft, Inc., 300 North pressing issues that arise during planning, design,
Lake Avenue, Suite 1200, Pasadena, CA 91101, USA.
E-mail addresses: misgana.k.muleta@mwhglobal.com
operation, and management of water resources systems.
(M.K. Muleta), nicklow@engr.siu.edu (J.W. Nicklow). Models are, however, simplifications of reality, and no
1
Tel.: C1 618 453 3325; fax: C1 618 453 3044. matter how sophisticated they may be, models undergo
0022-1694/$ - see front matter q 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2004.09.005

the type and complexity of the Refsgaard. black box) models and lumped.e. it is highly likely essential in order to properly simulate the effect of that results would be better than that which could be spatially varying properties. limited Unfortunately. The number of parameters and variables in a genetic algorithms (GAs) (Holland. land uses. values obtained during calibration and the sub- Refsgaard (1997). and weather parameters is the existing parameter possibilities.W. and Eckhardt sequent predictions made using the calibrated and Arnold (2001) are some of the most recent model are only as realistic as the validity of contributions that focus on this area of study. is time consuming and very previous calibration studies have dealt with lumped. distributed type (Refsgaard and Knudsen. 1992) search algorithm. Monte Carlo technique with Latin hypercube Parameter specification and parameter estimation sampling. among other factors. 1995. Refsgaard and Knudsen (1996).K. along with model assumptions and Gupta. 1996. The parameter accomplishing this useful task is to identify values of sensitivity analysis was performed using stepwise model parameters so that model simulations closely regression analysis. model calibration does not guar- work has been completed with regard to calibration of antee reliability of model predictions.128 M. Eckhardt and Arnold (2001) have implemented ment of Agriculture’s (USDA) Soil and Water automatic calibration for distributed models. and since it makes an extensive search of topographic features. However. and cing the number of parameters that should be carried parameter estimates. detail and quality of inputs. the modeler. based on observed information available to namely parameter screening. including those of the of conducting these two calibration stages depend on. Yet. which are capable of studies used a Shuffled Complex Evolution (Duan describing this spatial and temporal variability. 1997. Gupta et al. estimates. far more complex. (2000). it is imperative over to the parameter estimation stage. as well as temporally. manually obtained. Automatic calibration involves the use variables and inputs that may affect hydrologic of a search algorithm to determine best-fit parameters. The parameter distributed models. approach. much higher than that of resulting module is applied to calibrate streamflow a lumped model for the same watershed. long-term. Manual watershed simulation model being calibrated. Muleta.. 1998). making and sediment concentration estimates of SWAT using calibration of such models. Senarath et al. the stage at which parameters that need to be adjusted Manual calibration and automatic calibration are are selected. 1986. an generally be used for complex watershed simulation automatic calibration module is developed using problems.. Senarath et al. Distributed. such as the US Depart. of identifying the ‘optimal’ or ‘near-optimal’ values of Manual calibration is by far the most widely used the specified parameters. and its success highly depends on the empirical (i. Nicklow / Journal of Hydrology 306 (2005) 127–145 some aspect of conceptualization or empiricism. watershed its algorithms. 1995). Automatic calibration is fast. Sorooshian and study watershed. Approaches and difficulties approach for complex models. Most calibration. The former refers to parameters of the simulation model to calibrate. specification stage. subjective. and it offers a number of advantages over the manual Therefore.. which was carried out on ranks of match observed behavior of the study area. The distributed model is. subjective. however. however. Senarath et al. spatial parameterization. experience of the modeler and their knowledge of the conceptual models (Klemes. a process input–output data pairs that were generated based on a commonly referred to as calibration. In this study. (2000) and continuous simulation models. are the two most important stages of calibration Parameter estimation follows the decision of which (Sorooshian and Gupta. should et al. be implemented before using models for and global parameter sensitivity analysis to reduce their intended purposes. particularly the parameter data from a southern Illinois watershed. 2000). The common approach to calibrable parameters of SWAT. it is less mental variables such as soil types. responses may vary spatially. For most models. Both Assessment Tool (SWAT). the model assumptions for the study watershed . and Perhaps the most important task that needs to be their results are only as realistic as model assumptions accomplished during parameter specification is redu- and algorithms. This paper that a mechanism that improves accuracy of model reports on the use of three hierarchical methods. while the latter is the subsequent process two types of parameter estimation approaches. J. 1975). accounting for heterogeneity of environ.

from the National Agricultural Statistics Service mized agricultural management practices. Data obtained included a land management practices to simulate watershed 10 m resolution Digital Elevation Model (DEM) from processes such as surface and subsurface flow. and such as weather. daily streamflow and sediment concen- hydrologic routines within SWAT simulate surface tration data were obtained from the Illinois State and subsurface runoff processes. lateral flows watershed. evaporation. topography. including both overland and 30 m-pixel land use maps for the years 1999 and 2000 channel sediment processes. as well as streamflow and 1999). spatially distributed the Lower Cache River. Arnold et al.. January 1990 to August 2001 were obtained from the logic response units (HRUs). shown in Fig.e. maximum either a daily or sub-daily time scale. is used throughout this study for the demonstration of the methodologies and the models developed in this 2. crop growth for custo. Beven and Freer. Application of SWAT to a on water. soil. even Sediment yield from each subbasin is generated using after calibration. 1. and peak runoff rate is calculated using tration data spanned from June 25. 1998. . land use.e. accounting for snow Water Survey (ISWS) for Perks Road Station (PRS). based on physical Midwest Climate Center (MCC) for nearby climate characteristics of the watershed (i. and water (NASS). Runoff station that drains about 18% of the watershed (see volume is estimated using the Curve Number Fig. Finally. equation on a daily basis using information from the precipitation. including various species of nitrogen and Southern Illinois District of the NRCS. and analysis of uncertainty. M. IL and was used to derive a monthly possible (Arnold and Allen. SWAT makes use of watershed information sediment data for calibration. vegetation. the Generalized Likelihood Uncertainty Estimation (GLUE) (Beven and Binley. and ground water flows. 1999 to August 26. humidity. the period January 1971–April 2002 was also obtained SWAT simulates major hydrologic components from the MCC for a station located at nearby and their interactions as simply and yet realistically as Murphysboro. a fall and snow melt. Nicklow / Journal of Hydrology 306 (2005) 127–145 129 and the quality and quantity of actual watershed data the Modified Rational Formula (Williams. 1975). and land use). the year. 1996). The demonstration watershed and data with predicting streamflow and sediment concen- tration for the study watershed. wind speed. Daily phosphorus. J. The model operates on historical data related to precipitation. topography. verification. and sedimentation. This study used erosion. The watershed simulation model study. vadose zone processes (i. and for Church Road Station (CRS). gauging station that drains approximately 65% of the infiltration. 2001. sediment and agricultural chemical yields basin such as Big Creek requires topographic. 2001).K. soil. and climate data. and potential evapotranspiration from potentially. streamflow. uncertainty in results that arises simply because it is The model updates the C factor of the MUSLE too unlikely to find error-free observational data (e. ASCE. In addition to its maximum half an hour rainfall for all the months of a crop growth and water quality components..g. but also carries a higher simulator developed to assist water resource managers sediment load than other tributaries located in the area in predicting impacts of land management practices (Demissie et al. the and minimum temperatures. there is potentially a great deal of the Modified Universal Soil Loss Equation (MUSLE). further delineates subbasins into hydro. model subdivides a watershed in to subbasins and. 2001) in order to investigate uncertainties involved 3.W. Big Creek watershed. and a 30 m-resolution soil map from the quality. erosion the Natural Resources Conservation Service (NRCS). among others. stations. Fifteen minute interval precipitation data for soil. this 133 km2 basin not only contributes significant amounts of water to SWAT is a continuous-time. solar radiation. 1992. Therefore. (Neitsch et al. Muleta. hence accounting for variation in no simulation model is an entirely true reflection of plant cover during its growth cycle and its effect on the physical process being modeled. topography) and because crop growth module. a and percolation).. plant uptake. Daily streamflow and daily sediment concen- technique. Spatially. used for calibration and simulation. 2). Located in southern Illinois.

K. Location of Big Creek watershed. Location of streamflow and sediment gauging stations. parameters in Table 1 that are integrally related to the model’s streamflow and sediment yield prediction. three hierarchical methods are applied to achieve this reduction. Fig. the watershed was divided into 78 subbasins.1. A detailed investigation of the literature classifications with SWAT’s soil5id system. a total of only 682 Screening. identification of model parameters that could be were available. J. Nicklow / Journal of Hydrology 306 (2005) 127–145 Fig. 2001 for PRS and April 20. SWAT allows for subdivi- sion of a watershed into smaller subwatersheds. Muleta. In this study. The sediment concentration record was intermittent. 2000 to August 26.130 M. 2001 4. whose estimation from readily available data alone may pose significant uncertainty. which related to SWAT has assisted in identifying the 35 was used for this study. . refers to and 413 daily records for PRS and CRS. the smallest spatial scale considered in this study. as applied in this study. 2. and over the spans.W. Determining parameter values that are both realistic and ‘optimal’ for such a large number of parameters is not feasible. 1. calling for a necessary reduction of the number of calibrable parameters. 2003). The original soils map obtained for estimated with reasonable accuracy based on field this study was preprocessed in order to match data alone. respectively. and for which 4. Screening for CRS. Each of these subbasins is represented by a number of parameters that could be derived by calibration. Parameter identification there exists insufficient information from which As a distributed model. Based on a detailed spatial sensitivity and model feasibility analysis (Muleta.

75 GWQMIN Threshold depth of water in the shallow aquifer required for return flow to occur (mm) 0 100 the parameters could be directly estimated.005 10 EVRCH Reach evaporation adjustment factor 0.5 2 PRF Adjustment factor for sediment routing in the main channel 0.2 0.001 15 APM Adjustment factor for sediment routing in tributary channels 0.02 0. and/or topographic features. ism used. M.05 1 CH_N2UP Manning’s n value for the main channels in uplands 0.01 0. Nicklow / Journal of Hydrology 306 (2005) 127–145 131 Table 1 Model parameters involved in the calibration process Name Description Range of values Min. holding capacity of a canopy. Thus.5 SMFMN Minimum melt rate for snow during the year (mm/8C-day) 1. the parameter estimation algorithm is left to identify best-fit values for over one thousand 4.2 REVAPMN Threshold depth of water in the shallow aquifer for ‘revap’ to occur (mm) 0 100 OV_NP Manning’s n value for pasture land for overland flow 0. a concept referred to in if screening is the only parameter reduction mechan. which is still quite a daunting task. land use.001 1 GW_REVAP Groundwater revap coefficient 0.15 CH_N1 Manning’s n value for the tributary channels 0.5 TIMP Snow pack temperature lag factor 0.001 0.8 SFTMP Snowfall temperature (8C) K10 5 SPCON A linear parameter used in channel sediment routing 0. J.4 7.g. ground water flow parameters) are forced to spatial units in the watershed.2 0.2. Max.5 1 ALPHA_BF Base flow alpha factor (days) 0.4 7. SMFMX Maximum melt rate for snow during the year (mm/8C-day) 1.6 ALPHA_BNK Base flow alpha factor for bank storage 0.15 CH_N2LW Manning’s n value for the main channels in lowlands 0.W.K.g. the watershed.005 10 MSK_CO2 Muskingum routing coefficient 2 0. depending on coefficient for overland flow.3 OV_NF Manning’s n value for forest land for overland flow 0. For several other parameters.01 SPEXP An exponent parameter used in channel sediment routing 1 2 SURLAG Surface runoff lag time (days) 0.001 1 MSK_CO1 Muskingum routing coefficient 1 0.8 LAT_TTIME Lateral flow travel time 0 1 CANMXP Maximum canopy storage for pasture land 1 3 CN2PA SCS runoff curve number for moisture condition II 39 68 CH_COV Channel cover factor K0.01 0.05 0.05 0.001 1 EPCO Plant uptake compensation factor 0.001 1 CH_EROD Channel erodibility factor K0. Muleta. upon which . a assume uniform values over the watershed and some ‘representative subbasin’ is selected.001 100 RCHRG_DP Deep aquifer percolation fraction. The first others (e. this study as parameterization was applied.01 0. while values for the remaining 19 including the Curve Number.5 GW_DELAYE Groundwater delay 0.15 CANMXF Maximum canopy storage for forest land 2 6. Manning’s coefficient for channels) are 16 of these 35 parameters assume uniform values over broadly grouped.5 1 SMTMP Snow melt base temperature (8C) K2 20 SNO50COV Snow water equivalent corresponding to 50 percent snow cover (mm) 0.5 USLE_CP Minimum value of MUSLE C factor applicable to the land cover 0.001 0. In an attempt to further reduce the number of Parameterization is a technique for transferring calibrable parameters. and maximum water soil type. 0. For this study. Parameterization parameters. Manning’s roughness could differ between the 78 subbasins.001 2 ESCO Soil evaporation compensation factor 0. some spatially varying inputs model parameters of a given spatial unit to other (e.01 0.

Unlike the local techniques. . local approaches do not representative subbasin. parameters and consider only the parameters to In Monte Carlo analysis. A relationship between required parameters yield. all can be broadly subbasins) is developed using available information grouped as local and global approaches (Saltelli et al. In local techniques. canmax..W.K. about the parameters. 1993). ship between CN of a subbasin having a given soil and all input factors can be simultaneously varied. While of this representative modeling unit and correspond. this research further integrates Latin reasonable to assign nominal estimates for those hypercube sampling into the SA. For parameterization of the Curve Number. and the capacity of a canopy. output responses are variables in the representative subbasin enables determined by sequentially varying each of the input determination of a corresponding parameter in other factors and by fixing all other factors to constant subbasins. Simple scatterplot Parameter sensitivity analysis (SA) is applied to analysis. it is essential to efficient. the watershed. the less reliable the with pasture lands. Parameter sensitivity analysis provide a qualitative or quantitative measure of output uncertainty caused by each input. there are a number of techniques available for ing parameters of other homogeneous units (e. methods explore the entire range of input factors. grouping and land use and CN value of a subbasin that allowing investigation of output variation as a result is covered by pasture land and the same soil grouping of all inputs and their possible interaction (i. regression analysis. for Carlo analysis. The further the perturbation moves CN. were derived based on CN values exists. possible. Monte evaluation of Manning’s roughness coefficient. such as reduce the number of calibrable parameters as much as importance sampling and Latin hypercube sampling. the major land use in the study analysis results become (Helton.3.. n. once a ‘sufficient’ number which the model is sensitive during the calibration of input–output pairs are sampled using any preferred effort. nominal values. the more difficult and unreliable it is to the relationship between the CN of the representative employ local techniques.132 M. J. Particularly for watersheds like Big Creek distributions. Iman and Conover. correlation and partial identify parameters of SWAT model that contribute correlation analysis. for example. 19 that were originally allowed to the Monte Carlo family of methods. if any soil group.g. However. Also. 2000). Nicklow / Journal of Hydrology 306 (2005) 127–145 the model assumes homogeneity of parameters and most to the variability of streamflow and sediment variables. In this way. global SA recommended in the literature. the response surface methodology. the relation. techniques.e. those that should be calibrated. which is typical in hydrologic group ‘A’ was considered to be representative. Muleta. Fortunately. 1979. and effective sampling approaches. Then. that is grown under treatment conditions more nonlinear the relationship between inputs and common for pasture lands in the watershed and soil output variables. If an output is not (McKay et al. 1980). output were developed.e. model outputs are not equally can significantly reduce computational demand sensitive to all parameters of a model. since sampling virtual subbasin and other subbasins that have the is performed for one input at a time by fixing all other same land use and treatment conditions as the inputs at constant values. A large vary spatially and 16 that assume uniform values over computational demand is typically a concern of the watershed). calling for a parameter sensitivity analysis. conducting SA (Saltelli. it may still not be necessary or these SAs and is a result of the random and wise to apply a search algorithm to all remaining 35 unsystematic generation of inputs from specified parameters. Fourier amplitude method are common global SA A combination of the screening and parameteriza. but belonging to a different account for any interaction between inputs. a hypothetical subbasin covered away from the nominal value. also known as a sampling-based overland flow and the maximum water holding method. and thus. sampling technique. appreciably sensitive to certain parameters. Similar approaches were applied for uncertainty is averaged over all input factors). and stepwise regression analysis. Yet. the definition of 1999). Likewise. that lack long years of recorded data. it would be Therefore. further analysis needs to be performed to explore the input–output mapping and to 4. models. the use of more strategic. Furthermore. tion reduced the number of parameters that need to be The global SA method used in this study belongs to calibrated to 35 (i.

These input–output data pairs were further analyzed. uniform or triangular) would suffice tasks: for exploratory SA studies. stepwise regression involves a step-by-step ranges. M.1. all of the remaining 35 calibrable SWAT i. and identify the highly correlated sample data from the specified distributions and input factor. suspect that output variable and an input factor highly triangular distribution may better represent some of correlated to the output variable. (1998) and Helton (1993) indicate considered. The Stepwise regression analysis (Helton and Davis. and data apportion output uncertainty to each of the input sampling factors and determine the most influential parameters Assignment of input ranges and specification of for each of the two outputs. climate data previously described for the study v. previous studies. Construct a simple regression model between the its simplicity. At each documentation. starting with a and its parameters. Any available knowledge about the watershed construction of multiple regressions. were iii. Furthermore. as described in the following sections. In brief. satisfactory when few parameters are involved. As the name samples of inputs to determine their ‘true’ PDFs and implies. In fact. which is only were generated for streamflow and sediment yield. Perform a correlation analysis between each of the parameters were assumed to follow a uniform input factors and an output variable.3. and even if measurable. SWAT lation analysis between output uncertainty not yet user documentation. Test the performance of the regression model and watershed.3. and select the distribution. Ranges. ranges. Three hundred of sampling based SA efforts in water resources related such Monte Carlo realizations (i. and to this point. however. and knowledge of Big Creek watershed. Also. another input that is highly correlated to output should be explored to make an ‘educated guess’ at this uncertainty unexplained by the regression model is stage. including information from model simple regression model and building upon it.g. in order to 4.K. Haan et al. associated probability distribution functions (PDFs) is the most difficult and subjective stage in application 4. Nicklow / Journal of Hydrology 306 (2005) 127–145 133 the latter of which is used in this study. was evaluated. no previous study in corresponding outputs were compared. both studies suggest that simple stepwise regression analysis involves the following distributions (e. a measure used as an output for sensitivity analysis. available. land use. For iv. random individual input factors on outputs. perform a corre- combination of a detailed literature review. for this study. the hydrologic simulation model is the output variable and the input factor(s) selected executed using the topographic.e. Determination of measure of sensitivity of Monte Carlo analysis to hydrologic studies. Time series of daily streamflow and daily the stopping criteria. most previous variable in the SA. PDFs. soil. until a stage appears at which inclusion of that proper assignment of input ranges is more another input factor does not appreciably improve influential on SA results than knowledge of actual performance of the regression model. Accordingly. The authors. reason is simply that many model parameters are not 2000) is used herein to determine a measure of directly measurable. the parameters. Here it should be noted that choice of input factor most related to the output. are the were extracted for the days on which observed data was common approaches for accomplishing this task. Muleta. previous studies that applied explained by the regression model and all input SWAT. J. uniform distribution over triangular is entirely due to ii. excluding those factors selected in In application of Latin hypercube sampling to previous steps. Construct a multiple regression model between each sample. The simulated and observed values of the Based on the literature. factors. determine a measure of sediment concentration estimates at the calibration site sensitivity for each of the input factors included .2. input–output pairs) studies used scatter plot analysis. it would sensitivity that is capable of describing the effect of be cost prohibitive to collect numerous. Determine output uncertainty not yet explained by assigned for each input and were based on a the current regression model.W. a number of samples are first generated. as shown in Table 1. and the sum of watershed modeling has applied stepwise regression the square of the residuals. Input ranges and distributions. and other literature step.

If any of these input parameters are not acceptance of the hypothesis that the regression any more significant. and M is the total number of logic processes being investigated. to determine the p-value. which could occur if the coefficient is nonzero. bi for The SA model used in this study extensively iZ1.   1 1 The remedy to this problem is often to use the rank of QF ðFO F=v1 . output uncertainty. Convinced by the nonlinearity of the hydro. R2 values lie between zero and implying that addition of the latest input factor did not one. 1981). therefore. If the stopping criteria are not satisfied.. and the difference in R2 between using only the remaining input factors. 1990) was applied to test the hypothesis that all coefficients of the regression model (i. If the p-value under this assumption p-value. reject the The R2 value is evaluated for every regression factor and construct another regression model model developed. rank transform. output variables. is satisfied. The probability QF ðFO ~ F=L. for description of these two statistical methods. bi. The F-statistic (Myers. The p-value is the probability that a either criterion. in reality. Sub.. stopping stopping criteria.e. been implemented. and defined threshold. 2000). can be estimated by (Helton and QF ðFO relationships are nonlinear (Iman and Conover. a value close to one indicates that the regression significantly improve performance of the model. The use of v2 rank-transformed data results in an analysis based on KZ (2) v2 C v1 F the strength of monotonic relationships rather than on the strength of linear relationships (Helton and Davis.. v2 Þ Z IK ~ v . A small p-value biZ0 is rejected. model is accounting for most of the uncertainty in the Therefore. regression and correlation analyses are based on The probability of exceeding an F-statistic value of developing linear relationships between input and F calculated with (v 1. go to improve model performance. (Press et al. there was no threshold. where L is the number of input factors involves construction of regression models and included in the regression model being tested) are not correlation analysis. step iii.134 M. a p-value corresponding to bZ0 regression coefficient. Davis. which was one used as stopping criteria. 1992). the apparent relationship that led to regression coefficients (i.K. the stepwise regression model is halted if observed output.. samples. two successive regression models is used as one of the sequently. J. they often perform poorly when the ~ F=v1 . Myers (1990) may be referred significantly different from zero (i. are the statistics exceeds the user-defined threshold. as a result. the stepwise regression procedure significance of each of the factors in the new would be terminated.W. with an absolute value as exceeds a threshold value. Muleta. ation of the latest input factor did not significantly vi. Since hence. input factors are naturally correlated. or the difference in R2 of large as or larger than the one constructed in the successive regression models is lower than a analysis would be obtained if.v 2) degrees of freedom. evaluate the performance. 1979. a 2 2 2 1 concept known as rank transformation (Iman and and Conover. the same as a p-value corresponding to the assumption The coefficient of multiple determinations (R2) and that bZ0. relationship between the input and the output variable. MK LK 1Þ is ation has. measure of sensitivity for the factors. 2000). 1979. Nicklow / Journal of Hydrology 306 (2005) 127–145 and test the significance of the individual input have occurred simply by chance and suggests factors. Helton and Davis.L.e. bZ0) and. v (1) the individual data rather than the actual data. the hypothesis that bZ0 would extent to which the regression model explains the be accepted and the stepwise regression is terminated. otherwise the input factor would be indicates that the regression coefficient is less likely to removed and another regression model with LK1 input . Conover and Iman. R2 is an indicator of the percent in this study. A test was also performed for the individual and. bi for iZ1. 1981. If the difference is less than a user- criteria. also known as the a-value. where IK(a1.e. v2 Þ. 2000) Conover and Iman..a2) denotes the incomplete beta function 2000). indicating that added consider- model.L) included the constructed regression coefficient was due entirely in the regression model to ensure that the hypothesis to chance (Helton and Davis.

farm land covered by input factor is removed from the regression model. or Qt ðjt~jO jtj=M K LK 1Þ. and soybean sensitivity. corn. the computational subroutines of SWAT † The absolute values of the standardized regression were executed and the output variable used in the SA coefficients (SRC) in the regression models are for streamflow (i. the sampled values were assumed input factor into the regression model. Input data collected for the demonstration watershed were where used for application of the new SA model. negative sign).W. Application results and discussion QF ðjt~jOjtj=M KLK1Þ Z 1 KIO ðM KLK1Þ. J. and Davis. sum of the square of residuals quantitative indicators of an input factor’s import.2.a2] denotes the incomplete beta function Urbanized subbasins and other subbasins that were (Press et al. or measure of office of the NRCS. For spatially amount of output accounted for by including the varying parameters.K. 1992).e. pasture. 1993). Provided to be valid for the representative subbasin previously that the input factors are uncorrelated. Based on interviews conducted used as qualitative and/or quantitative indicators of with personnel from the Southern Illinois District each input variable’s importance. Generation of Latin hypercube samples. . watershed. the rotate on an annual basis (i. SRC provides a measure of importance based extracted.e. rank-transformation was per- expected value by a fixed fraction of its standard formed. defined here as 2. between observed and estimated values) was ance. were assumed to have been grown without tillage. If the p-value of an individual input factor exceeds a simulation period. † The order in which the variables are selected and such as planting and harvesting/cutting dates of enter into the regression model is a qualitative agricultural crops.e. or obtaining a value t~ from the t-distribution for which decrease. After obtaining input–output pairs for all on the effect of moving each variable away from its of the 300 samples.. or tend to jt~jO jtj. Muleta.2. while retaining all other variables at fed to subroutines of the SA model for stepwise their expected values. 2 2 The SA model was developed and incorporated as (3) a submodule within the source code of SWAT. were assumed to be consistent indicator of their importance. Once model parameters for all subbasins variables being added at each step (Helton. The sign of a standardized regression. M. and the resulting rank-transformed data were deviation. The probability of input and output variable tends to increase. were known. forest. The land use data obtained was for years 1999 and 2000.e. and M KL K1 OZ (4) land uses for the remaining simulation years (i. the differ.5%. together (i. 1990). described with regard to parameterization. This was accomplished regression coefficient (SRC) indicates whether the using the t-statistic (Myers. † The difference in R 2 values of subsequent Three hundred Latin hypercube samples were regression models provides a measure of import. The probability Qt ðjt~jOjtj=MK covered primarily by pasture. other farm management operations. positive sign). Corn and soybean were assumed to user-defined probability. and hay were LK1Þ is the same as the p-value corresponding to biZ assumed to have the same land use for the whole 0. The corn in 1999 was assumed to have been covered by following aspects of stepwise regression analysis are soybean in 1998). The ences in the R2 values for the regression models relationships developed to link parameter values of constructed at successive steps equals the fraction the representative subbasin to other subbasins were of the total variability in the output variable that used to derive corresponding values for other can be accounted for by the individual input subbasins.e. Nicklow / Journal of Hydrology 306 (2005) 127–145 135 factors would be constructed. hay. 2000)   1 1 4. is given by (Helton inversely related (i. and IO[a1. with the most with dates commonly used in the demonstration important variable being selected first. Furthermore. generated from the input ranges shown in Table 1 ance for the latest variable by indicating the and the assumed uniform distribution. 1998 M KL K1 Ct2 and 2001) were derived from the available land uses.

The relative importance of each R2 of the regression model constructed using the input of these input factors could be judged using the order factor(s) selected at each stage. These numbers can be better CH_N2LW K0.1.35 0 5.6966 through combinations of parameters to achieve the set CH_N2LW K0.61 0 0.60 0 0.e. J.e. data quantity and CH_N2LW K0.0067 parameters of the simulation model to calibrate.62 0 0.28 0 SPCON 0.34 0 CH_EROD 0.e.27 0 SPCON 0. and the SRC and in which the parameters were selected.08 0.0104 a GA is used to systematically and extensively search 8 PRF 0. 12 for streamflow and 8 for 5 PRF 0. parameterization of physically CH_N2LW K0. APM 0.6514 based on calibration.16 0 6 PRF 0. are among the input Parameter sensitivity results for sediment yield factors obtained for both streamflow and sediment Step no.e.09 0. Muleta.W. and 2.0172 From the perspective of data quantity and quality. sediment yield.684 CH_N2LW K0.691 addition to parameter identification.61 0 0.5131 In conclusion of the parameter identification stage CH_N2LW K0.60 0 0. respectively. and the p-value used the original developers of SWAT for inclusion into calibration efforts. and application of the stepwise biZ0) were 0.5%.09 0. objective function). a recommendation that may not be Table 2 valid for every study watershed. there is a general agreement that the amount . The criterion used here is the sum of the CH_COV 0. In 7 PRF 0.27 0 4 PRF 0. In this study. p-value based on the test that considers all input factors Interestingly.3842 available in Muleta (2003).26 0 (i.0207 though. the SA identified 12 parameters that play a yield. The Table provides the input factors streamflow and eight inputs from the perspective of selected at each step of the stepwise regression model.35 0 quality. rank-transformation. The threshold values used R2 that was achieved due to inclusion of the input factor as a stopping/parameter removal criteria for the into the regression model and the SRC coefficient of the difference in R2 of successive regression models. CH_COV 0.09 0.676 sediment yield). APM 0. most of the parameters recommended by included to that point (i.16 0 square of differences between corresponding simu- SPEXP K0. to test significance of the individual input factors (i.27 0 CH_COV 0.136 M. Nicklow / Journal of Hydrology 306 (2005) 127–145 model execution and extraction of output variable. Parameter estimation CH_EROD 0.0058 lated and observed values at a given time scale.K.08 0. Parameter SRC P-value R2 yield. screening of which model parameters to estimate CH_N2LW K0.27 0 parameter estimation.16 0 search mechanism are the most decisive factors in SPEXP K0.28 0 managed in the subsequent stage of calibration.5842 (i. Results of the SA model for sediment yield are significant role in explaining the uncertainty of given in Table 2.34 0 immeasurable parameters.26 0 ameters to 20 (i. improvement in p-value of each input factor. For these regression tasks were also accomplished for sediment criteria.36 0 of this study. bZ0). and application of a proper CH_COV 0. both of which are listed in the Table. proper selection of a performance measure CH_EROD 0. the input factor.16 0 Parameter estimation follows the decision of which SPEXP K0. SPCON 0. and sensitivity analysis) CH_EROD 0.60 0 0. CN2PA 0. 2 PRF 0. 1. Note that a similar table for streamflow is 1 PRF 0.61 0 0.35 0 CH_EROD 0. a combination of the three techniques 3 PRF 0.34 0 based on field data alone and which to determine CH_EROD 0.28 0 has reduced the number of calibrable SWAT par- SPCON 0.e.26 0 accuracy.28 0 that is ‘best’ in terms of satisfying the criterion of SPCON 0.08 0.61 0 0.0077 successful automatic model calibration.

a continuous GA.W. have been mate. sedi. and Sorooshian and guaranteed to locate global optima (Cieniawski et al. Each model. Nevertheless. The objective function used should ameters. known as genes. GAs search a wide portion of the solution number of observations) should be at least 20 times space and. a solution alternative identified and the output variables being calibrated is that has the lowest RMSE was the fittest set of too complicated to derive simple (i. 1998. that require no derivative information the fittest of these five chromosomes was selected as a about the objective function or constraints. In fact. 12 for sediment yield are being calibrated. or residuals (RMSE) tion (RMSE) was repeated until a user-defined between observed and simulated values of streamflow number of initial chromosomes had been reached. Sample generation are simulated correctly. The procedure was repeated.g. For calibration of each (e. there have been nonconvex. respectively.g. Though not ultimately necessity of 3–5 years of data. Muleta. 1995) of these two watershed responses. Best-fit values obtained for the parameters and Gupta (1995). is used in this study. Klemes (1986) argued on the theory optimization problems. 200 to optimal solutions. have a better capability of locating the number of parameters to be estimated (e.. and such as GAs. In this study. GAs are a technique that applies Darwinian length. hence. for example. 1995). streamflow and 8 for sediment yield) being calibrated. and two . A sediment yield cannot be simulated without correct gene is essentially a value for one of the influential simulation of the streamflow. unimodal) mathematical functions and. This implies that gradient. In Big GA.. and sediment concentration. Accordingly. was used. and multimodal problems (Schwefel. the root of the and evaluation of the corresponding objective func- mean of square of errors. known as chromosome. thus. For example. their The initial solution candidates were ranked in derivative information. five complex problems like calibration of distributed solution candidates among those ranked within the models. some studies calibration of sediment yield. values yielding a simulation result that best resembles The relationship between the model parameters the observed data. also known as a real coded generated and transported during flood events. Therefore. the overall objective is to achieve sample. Gupta (1995) argued that the length of data (i. (1990) have computational subroutines of SWAT were executed to reported that 96. during calibration of streamflow were used during For selection of an objective function. were the be consistent with the anticipated application of the initial solution candidates supplied to the GA. literature consistently demonstrates an ability to the length of available data at PRS is about 793 identify global or very near global optima for a samples for streamflow and 682 samples for sediment range of complicated problems (Nicklow. collectively known as a population. 1500 random have indicated that calibration success is very samples.e. and parent alternatives were selected based optimization methods have little to offer for using multiple tournament selection. Once a chromosome is selected. Gupta et al. J. M. Here. an objective function The objective was to determine a set of parameter biased towards peak flows. it is essential that flood peaks longer periods of historical data. Since sequence of parameter values. heuristic search algorithms. some recommendations regarding minimum data 1995). However. the available data length is sufficient ducted for streamflow and subsequently for sediment with respect to the recommendation by Sorooshian yield.e. Therefore.3% of the sediment moves during five provide the corresponding RMSE using observed and percent of the time. Nicklow / Journal of Hydrology 306 (2005) 127–145 137 of information contained in the data is more important practically proven to work well on nonlinear. As an alternative. Since the number of calibrable The parameter estimation effort was first con- parameters is 20. based on data collected during the simulated data at PRS. generated from the ranges and the uniform dependent on the objective function used as a distribution assigned for the most influential par- selection criterion. 2000). and the majority of sediment is usually number. Demissie et al. both streamflow and parameters selected for the output variable (i. the Creek. Sorooshian and Gupta. smooth and parameters. top half of the population were randomly selected. the majority of GA calibrate 10 parameters). is defined by a a reasonable simulation of sediment yield. ascending order. ment yield is a seasonally fluctuating watershed Since the values of genes could assume any real response. than the amount of data. the station with relatively late 1980s. yield.e.K. For the current application.

3652 simulation REVAPMN 28. 31) Year Observed Calibrated Default 2000 65. for the watershed gence by introducing new traits or characteristics into responses simulated using the best-fit parameter the search process.138 M.865 2001 131. forming pairs. CPU time on a 1.40 simulation 2001 54. Ef (default) Parameter Calibrated Ef (calibrated) Ef (default) value brated) Value CN2PA 59. half of the current population) generations were considered to be ‘optimal’ par- had been obtained.e. J.743 191. ameter values.569 516.138 2001 131. Results for sediment yield are were known.06 196.0345 MSK_CO2 0.0247 CH_N2LW 0.69 GHz. selecting mating pairs.1549 ESCO 0.5877 APM 1. The number of solution parents that ultimately produce two offspring alterna. Muleta.77 mg/l 25–Dec. injecting new traits by summary of the annual outputs is presented in mutation.082 29. 4.0031 CANMXP 2. candidates was. In addition.919 88. For these GA candidates that were selected randomly. or values that yield the ‘best’ match Mutation was performed on 15% of the solution between observed and simulated data.582 Average daily sediment concentration (mg/l) GWDELAYE 47.52 Annual Streamflow (mm) Annual streamflow obtained using these parameter values (mm) 1999 (June 31.8488 CH_COV 0.2478 RCHRG_DP 0.3122 1999 (June 25.585 516. which corre- Ranking solutions according to the ascending order sponds to the ideal condition when measured and of their fitness values. as well as default values of the model mined for the mutated solution candidates and for the parameters without calibration.005 PRF 0.686 . and creating related to computational time. a value randomly selected from the given in Table 4 and Fig. reduced from 1500 to 150 tives through the process of uniform crossover.K.1963 EPCO 1 CH_EROD 0. the replaced the existing value of the gene. The from the third generation forwards due to issues process of choosing mates. a offspring by crossover.0882 SPCON 0. Pentium IV processor. The Tables provide values range and distribution assigned for the parameter of the selected parameters and values of Ef.01 Year Observed Calibrated Default GWQMN 37.897 2000 262.1 SURLAG 1. Fitness values were then deter. This mutation coefficient of efficiency (Nash and Sutcliffe.082 24.W. 4.38 CH_N2LW 0.744 K0.9 mm/yr 25–Dec.6129 CH_N2UP 0. or iterations. 1970) process minimizes the chance of premature conver. Once the Calibration results for streamflow are given in solution candidate and the specific gene to be mutated Table 3 and Fig. The genes correspond- offspring were continued until a user-defined number ing to the fittest chromosome at the end of 75 of new candidates (i. however.9097 0.77 50. values. and calibrated and noncalibrated mutated and new candidates was repeated for 75 Table 4 Table 3 Summary of calibration results for sediment concentration Summary of streamflow calibration results Parameter Calibrated Ef (Cali. creating estimated values perfectly match.9 mm/yr 1999 (June 31. vary from negative infinity to unity. 31) 2000 262.919 93. Nicklow / Journal of Hydrology 306 (2005) 127–145 successively selected mates were paired to form generations.461 K4. used to test goodness-of-fit.496 CN2PA 52. the search process required about 50 h of mutant candidates.9512 0.743 220.31) 25–Dec. The value of Ef can newly born alternatives.0824 SPEXP 1. the specific gene to be manipu.95 23. lated was also decided by random selection. and evaluating fitness values for the the Tables observed. For each of parameters.49 60.

W. some sort of model calibration. Figs. Results for streamflow respectively. and Arnold. Comparison of observed and calibrated sediment yield. 3. Eckhardt simulations by the calibration effort for both stream. J. 4. Comparison of observed and calibrated streamflow. simulations. 2001. are very good when From observation of these results. 3 and 4 provides a graphical underestimate both streamflow and sediment yield.g. Santhi et al. with an Ef of 0. . comparison of the observed data and calibrated values These figures alone demonstrate the importance of on a daily basis for streamflow and sediment yield. The results obtained herein for flow as well as sediment yield. M. Muleta.74. the reader should compared to that of previous calibration efforts acknowledge the improvement achieved in the model attempted for SWAT (e. the calibrated model seems to ‘slightly’ to the default simulation. which highly overestimates watershed considerable improvement is achieved when compared responses. simulations. Nicklow / Journal of Hydrology 306 (2005) 127–145 139 Fig. 2001). but simulation.K. Contrary to the default sediment yield may not be considered impressive. Note that no previous work Fig..

and since it drains most of the watershed. For distributed models. adequate information for answering the questions that Verification results at CRS are graphically presented face decision-makers. Fig. 5 and 6 for streamflow and sediment yield. This requires multiple gauging stations in ment for streamflow (i. Owing to availability of relatively 6. confidence. PRS and CRS. Verification result for sediment concentration. The sediment from subbasins can be determined with reasonable calibration effort has achieved a significant improve. Table 4) when compared to the watershed to measure model outputs of interest.140 M. it is in Figs. Verification results for streamflow. has been conducted in relation to calibrating SWAT for predictions are satisfactory so that model results sediment yield on a daily time scale.W. . Muleta. The calibrated model was then tested the best values for model parameters. as described previously. the outcome for its capability of simulating a reasonable response needs to be verified to determine if the results provide at the gauging station not used in calibration. For Big Creek watershed. Nicklow / Journal of Hydrology 306 (2005) 127–145 Fig. or CRS. J. 5. Model verification longer periods of data at the PRS station. the results given in Table 3. 6.K. there exists two flow and sediment gauging stations. data from PRS was used Once calibration has been conducted to estimate for calibration. important to determine if spatially distributed respectively.e.

An Ef value of 0. and parameters on reliability of technique of determining reliability of model predic. known as for sediment concentration. data in the model calibration effort. and Monte Carlo based default simulation.233 is obtained approaches have been used to investigate reliability for streamflow at CRS.g. Muleta..W. it is tion implemented in this research (i. second-moment method methods such as Latin hypercube sampling are used. it model to a watershed with longer data records may argues that for a different set of data (e. calibration. first-order. less than three years) used being modeled under the given conditions (i. simplifications and approximations intro. is computationally tolerable if systematic sampling The mean-value. there are many reasons for data could predict the watershed behavior reasonably the modeler to be suspicious of model results. Using a The GLUE procedure argues that given the inherent physically based. (Melching.there is no reason a model specifically designed to accurately model to expect that any one set of parameter values will runoff alone. Beven and Freer. calibrating grounded on a conception that conflicts with model using data at one gauging station and conducting calibration procedures. but no similar work parameter set to satisfy a user-defined fitness function. (2000) attempted the type of verifica. 2001).e. (2000) obtained an represent a true parameter set (within a given model absolute error (i. The study by (GLUE) (Beven and Binley. Harr’s point estimation ging. a data. better better reveal the capability of the simulation model. this seemingly optimal parameter set may perform poorly and many other parameter sets may perform 7. the meth- respectively. absolute deviation between structure) to be found by calibration procedures” observed and estimated values) for runoff volume (Beven and Binley. J. The poor performance of the by the calibration procedure simply as parameter verification effort could be attributed to a relatively values that are more likely to represent the watershed short duration of data (i. is. default simulation. essential that every modeling effort be it accounts for the effects of uncertainties due to input accompanied by an uncertainty analysis (UA).e. quality data) or even for a different simulation period. and calibration procedure applied). 1992). 1995). a concept known as the equifinality problem Uncertainty arises due to incomplete information (Beven and Freer. Application of the and model used. two-dimensional. The method is based on the ranging from 28 to 42% for four internal gauging premise that. as compared to K0. 1995).e. such as the one applied in this verification using data from other gauging station(s) study. has been completed for sediment yield. the tainties is acknowledged.e. therefore. The approach thus recognizes that many model parameters within a given model structure and input Even after calibration.K. the method tions. It are that the method considers global uncertainty. method (Rosenblueth. Uncertainty analysis better. as compared to K4. 1975). well. 1992. Alternatively stated. model structure. once the unavoidability of the uncer- stations used for verification. accounting for various sources of uncertainty.14 for the of hydrologic model outputs. Rosenblueth’s point estimation and the method is conceptually simple. watershed behavior. Senarath et al. but indicates a significant improvement over the method (Harr. For this study. The major . 1989). a result quite comparable with that of odology considers the optimal parameter set derived Senarath et al.005 is obtained The UA method used in this study. 2001). However. uncertainties associated with distributed watershed ameter model known as CASC2D (Julien et al. that strive to derive a single best optimal in the watershed) for streamflow. Among the major advantages of using the GLUE duced into the modeling exercise. quantity and quality of input data) used for to predict bounds of uncertainty for the model outputs. M. 2001) is a Monte Carlo based approach.26 Generalized Likelihood Uncertainty Estimation obtained using the default simulation.g. the model output (Beven and Freer. Nicklow / Journal of Hydrology 306 (2005) 127–145 141 The verification result is not particularly encoura.g. and parameter methodology for UA of a distributed watershed model estimates (e. models and modeling exercises. capability of the search algorithm). and an Ef of K0. Senarath et al. (2000). “. distributed par. and applies this conception (e. it is only possible to make absolute deviation between the observed data and an assessment of the likelihood of a particular simulated estimates for streamflow and sediment yield parameter set being an acceptable simulator of the at CRS station is found to be 33 and 26 percent.

and SWAT was executed to predict and used to determine various statistics such as sediment yield and streamflow values at PRS for each sample. Following the approach used by Brazier et al. were evaluated for both sediment yield and stream- flow. or likelihood measure. and a iZ1 discussion of these tasks is not repeated here. (2000) to derive a combined measure of likelihood 7. Cumulative distribution of the SCML is derived were generated. All of the input factors were assumed to follow a simulations were ranked in their ascending order for uniform distribution. it is important to emphasize that there is a considerable Fig.Fi and Ef. it involves determining a user. and ranges of the factors were each time step (i. streamflow and sediment yield. parameterization of spatial varying CMLi SCMLi Z (6) input factors. Latin (SCMLi) sum to unity and. it requires derivation of model simulations for the where Ef.Si are coefficients of efficiencies of generated samples. and annually). or acceptable. for defined measure of fitness. 7. respect- ively.K. yield a probability hypercube sampling is used to generate input samples distribution function for the parameter sets. Again. 5000 Latin hypercube samples run. All of the parameters used during the parameter In order to derive measures of uncertainty. the threshold values and were involved in the subsequent reader is refereed to Beven and Binley (1992) and steps of the UA. daily. only 887 parameter sets into the acceptable and nonacceptable behavioral simulations were accepted according to the categories. such that the rescaled likelihood measures measures of uncertainty. hence.1. estimation stage (see Table 3) were considered in the simulations results from all of the accepted behavioral UA. simulator of the system for streamflow and sediment yield. monthly. Muleta. The Ef values obtained for streamflow and sediment yield during the calibration procedure aided in the choice of these threshold values. Therefore. Beven and Freer (2001). the pro. The for the UA model.W. Efvalues The GLUE methodology has similarities to the SA obtained for the two variables were combined as model developed in this study in the sense that GLUE follows for each of the accepted behavioral simu- also requires generation of numerous input samples lations from ranges and distributions assigned for the input CMLi Z Ef. As in the SA model.Yi (5) factors involved in the uncertainty estimation. and preserved to the assignment made during the SA. function for model outputs in order to derive measures grammatic codes developed for the SA model to of uncertainty accomplish tasks related to Latin hyper sampling of input factors. used as a measure of fitness. respectively.45 and 0. Application results and discussion (CMLi) that simultaneously serves as a goodness of fit for both streamflow and sediment yield. For details on the GLUE methodology. J. the distribution function of the predictions is calcu- Using the input ranges and distributions assigned to lated using the SCMLi of the corresponding model the eight parameters.e. 95% confidence limit with observed and calibrated degree of subjectivity involved in the choice of these streamflow. Values of Ef.25 were used as threshold likelihood values to consider a parameter set that is a behavioral. and simulation i.Fi !Ef. Nicklow / Journal of Hydrology 306 (2005) 127–145 drawback of the GLUE methodology is subjectivity of threshold values. hence improving the computational function ultimately serves as a probabilistic weighting efficiency of the methodology. Of the 5000 parameter sets. CMLi values are then rescaled from it applies input–output information to determine zero to one. a major shortcoming of the GLUE the likelihood level assignment that groups the methodology. and extraction of model predictions at X N CMLi PRS are directly used for the UA as well.142 M. Ef values of 0. .

The automatic technique presented allows one to avoid the limitations of existing ‘trial-and-error’ calibration techniques. 8 clearly shows the over. Application of the UA model to mates. 8. Muleta. however. complex having a ratio between the upper and lower bounds of control problems.K.W. spatial parameterization. Fig. as well as . 95% confidence limit with observed and calibrated sediment based approach that uses Latin hypercube sampling concentration. and parameters esti- of the SWAT model. The model is demonstrated using a watershed located in southern Illinois. In addition. 8) are very wide. strategy along with stepwise regression analysis of the uncertainty bounds (e. respectively.g. The par- of the 95% confidence limits). The upper bound exceeds the flow and sediment concentration. and a detailed parameter sensitivity analysis. Summary and conclusions relatively short duration of recorded data due to the lack of sufficient data for the study watershed. Nicklow / Journal of Hydrology 306 (2005) 127–145 143 estimated by the USDA’s distributed watershed model known as SWAT. popular in recent years for solving realistic. model given here should not be a generalized characteristic algorithms and assumptions. J. this of streamflow and sediment yield estimates. The parameter sensi- tivity analysis was conducted using Monte Carlo Fig. identification) stage of model calibration is accom- plished using hierarchy of three techniques: parameter screening. are ‘narrow’. In contrast. as well as its inability to envelop the watershed simulation model to analyze uncertainty observed data for 2000 and 2001.e. 7 indicate that Parameter estimation was performed using a the model’s streamflow prediction is fairly consistent genetic algorithm. sediment yield prediction of the SWAT model was found to involve a great deal of uncertainty. a search method that has become in the sense that the uncertainty bounds are ‘narrow’. The verification exercise indicated poor confidence limits for sediment concentration (i. the Generalized Likelihood Uncertainty Estimate predicting behavior of the model for sediment (GLUE) is developed and incorporated into the concentration. The study did. and the result including those arising from input data. Application results indicate signifi- less than a factor of two. Unlike streamflow simulation. The An automatic calibration model is developed in capability of the automatic calibration algorithm and order to improve streamflow and sediment yield the uncertainty analysis methodology. a model that uses sediment prediction. number of calibrable parameters of SWAT’s stream- sponding to the 95% confidence limit. Application of the uncertainty estimation other watersheds that have a relatively longer data model indicates that the model’s streamflow predic- record could yield additional insight regarding the tion is fairly consistent in that the uncertainty bounds behavior of SWAT. estimation phase of model calibration. indicating a great other distributed hydrologic models. The bounds given in Fig. rely upon a 8. The parameter specification (i. the 95% estimates. a behavior shared by lower bound by up to a factor of six. The work is conducted using a watershed with limited model accounts for the major sources of uncertainty recorded sediment and streamflow data.e. the bounds cant improvement over default simulations and results obtained have embraced the observed data and of previous works that have reported on calibration of the calibrated model simulations for the year 1999 SWAT’s streamflow and/or sediment concentration and 2000. Once again. 7 and 8 provide uncertainty bounds corre. M. In addition to the deal of uncertainty associated with the model’s calibration and verification study. ameter identification exercise effectively reduced the Figs. observed data flow and sediment concentration variables to a and calibrated model simulation data for annual number that can be easily managed at the parameter streamflow and average daily sediment concentration. performance of the calibrated model for both stream- Fig. lower and upper bounds rank-transformed input–output data pairs.

J.E. Uncertainty and sensitivity analysis techniques provided support for this research. J.M. W. Williams.. Arnold. Publications. H. Beven. J. Equifinality. M. Conover. S. 1992. W. 1993. Toward improved and the uncertainty analysis efforts could be applied to calibration of hydrologic models: multiple and noncommensur- a ‘data rich’ watershed. R.. K. Keefer..H. Temple. P. Small sample sensitivity analysis Association 34 (1). B. and Klemes..J. Saghafian. Cambridge University Water Resources and Research 28 (4). 1975. Journal of Hydrology 251. 2000. New York. between parametric and non-parametric statistics. E. Reston.. Contract Report 484. Vetterling.... J. 103–109. Champaign. 1981. Hydrological hydrologic modeling of spatially-varied surface runoff. Earth Computer Models of Watershed Hydrology. V.... Edwards. J. MA. The future of distributed models: Julien. W. 55–77..R. M. The Amer. New York.D. which has Helton. 1989. 1998. K. and Analysis. PWS-KENT Publishing Company.. Conover. F. Kiniry. W.E. PhD Thesis. B. R.H.. T. problem. CO.. Illinois State Water Survey. Sorooshian.. Probability estimates for multivariate analyses. 1979. Makowski.R. Ann Arbor.R. L. IL. in: References Saltelli. Conover. 1990. Version Demissie. Communications in Statistics: Theory and Methods Watersheds.... Numerical Recipes in Fortran. V. Technometrics 21. Champaign. Iman. Q. 2000. The use of rank transform in Large area hydrologic modeling and assessment part I: model regression. 73–89.W. Cache River Basin: Hydrology.. S. Grassland. K. Illinois State resources engineering and management. W.O. Parmer. Demissie. Muleta. T. (Eds. Using genetic Muleta. Effective and efficient Press. VA. Freer. C..J. Wiley. J. Nash. Sabbagh.T. Water Resources and Research 34 (4).. Cieniawski. A 9. the verification. 1015–1031.. V. 1995. Sutcliffe.. A.L. Adaptation in Natural and Artificial Systems.. for use in performance assessment for radioactive waste disposal. J. 277–291.. 1998. S. 124–129. soil and water research service.G. College of Engineering. erosion prediction project—for sites in the UK and USA. Neitsch. Transactions of ASAE 41 (1). 825–845. Water Assessment Tool Theoretical Documentation...T. Hydrological Sciences Journal 31 (1). R. Journal McKay. Rowan.. A comparison of Hydrology 249. 1996.. 313–318.. J. Flannery.. environmental systems using the GLUE methodology.J. Discrete-time optimal control for water Lower Cache River.. J. three Illinois watersheds. Hydrology 125. 1995.H. Harr..M.. 2003. 13–24. Iman. with an application to risk ASCE. Sensitivity Analysis.G. IL. Data Collection. Southern Illinois University at Carbon- Conover. K. Myers. P. 751–763. D. 399–409. Arnold. V.. River flow forecasting through 1990. A decision support system for the management algorithms to solve a multiobjective groundwater monitoring of non-point source pollution from watersheds. Resources Bulletin 31 (5). J. Press. Prabhu. Al-Issa. Estimating hydrologic budgets for Holland. Water Resources and Research 31 (2). 1999... MI..L. Eheart. The authors wish to thank the Illinois Council for Applied Mathematical Modelling 13. Williams.144 M. catchment model. S.. 1970.H. Classical and Modern Regression with ican Statistician 35. Freer.. K. Reliability Estimation.K. Arnold. J. Equifinality analysis of output from a computer code. application of the GLUE methodology to WEPP—the water Melching. Applications...J.. Technometrics 21.Anon.S. global optimization for conceptual rainfall–runoff models. in: Singh.P. C.. 1: Background.L. R.E. 1999. (1).. 2001. Part I—a discussion of principles.. Knapp. 11–29. Carbondale.S. Gupta. (Ed. M. 499–509.. TX. J.J.. Allen. and uncertainty in physically based soil erosion models: 239–245.C. Hydrology of the Big Creek Watershed and its Influence on the Nicklow. J. Duan. Food and Agricultural Research (CFAR).S. 1998.J. S.A. Binley. 2000.M. Rank transformations as a bridge dale. Boston. P. R. J.. 89–95. Soil and IL.E. Automatic calibration of distributed effectively tested if the calibration. 2001. Reliability Engineering and System Safety 42. 279–298..W. 2001..L. Storm. 1992. R. Ogden...).K.J. Journal of the American Water Resources Iman. A. Beven..G. Vol. J. Sorooshian.. 1980.V.W. M. Operational testing of hydrological simulation uncertainty estimation in mechanistic modeling of complex models. development.. Arnold. Scott. data assimilation.. 1979.. Sampling-based methods.. 1749–1874. 2000. M. University of Michigan Press. Haan. R. S.R. Nicklow / Journal of Hydrology 306 (2005) 127–145 the behavior of the simulation model could be more Eckhardt. Yapo..G. S. Soong.. Srinivasan. Water Processing 6..C. Water International 25 Water Survey.L. GIS Modules and Distributed Models of assessment. 1992. D.. Chan. P. Effect of parameter distribution on Acknowledgements uncertainty analysis of hydrologic models. Kriesant.).. Water Resources Surface Processes and Landforms 25.K..Y.. Journal of Transport. 1986. D. Teukolsky.V. 1995. Helton.. R... NY. Beven.J.E.. Davis... Journal of Hydrology 176. 523–536. Beckman. 327– 367. W. Raster based model calibration and uncertainty prediction. R. Ranjithan. P. J. J.P. Muttah. 65–70. J. Contract Report 2001-06. American Society of Civil Engineers. Hydraulics and Sediment conceptual models. G. F.. of three methods for selecting values of input variables in the Brazier. .. techniques for computer models. Highlands Ranch. Gupta. able measures of information.J. Allgire. 2001. J.

river basin with point and non point sources. (Ed.). S. 2001. Model calibartaion. J. intercomparison of different types of hydrologic models.-P.. Proceedings of the National Academy of Sciences USA 72 (10). Technometrics 41 (1).. 965–974. R. A. Journal of Hydrology 198. A.W. New York. Chan. A quantitative Water Resources Publication. S. continuous watershed models. Muleta. M. What is sensitivity analysis?. A. S. Validation of the SWAT model on a large 69–97..C... Senarath. Saltelli. J. K. in: Saltelli.R. Parameterization. (Eds..K. Nicklow / Journal of Hydrology 306 (2005) 127–145 145 Refsgaard. Highlands Ranch. Computer Models of Watershed Hydrology.. Gupta. Resources and Research 32 (7).P.. J. Downer. hortonian. On the calibration and verification of two-dimensional. H. J.S. Water Saltelli. 1996.. 3812–3814. V.. E. L.. Torantola. 2000. Resources and Research 36 (6). 1169–1188. calibration and validation Santhi. . 2000. Water Schwefel.C. Arnold. of distributed hydrologic models. Water Resources Bulletin 11 (5).R.. in: York...P. 1997. 1995.. Operational validation and American Water Resources Association 37 (5). K. V.). Knudsen. 1995. CO. model output. Scott. F..U..O. New Sorooshian. 2189–2202.K. Wiley..L. Journal of the Refsgaard. Point estimates for probability moments. 1975. 1975.. Singh. Evolution and Optimum Seeking.. E. J. distributed. Sediment routing for agricultural watersheds. Ogden. J. 39–56. Sensitivity Analysis. Sharif. 1999.G.M. Wiley.. model-independent method for global sensitivity analysis of Williams. Rosenblueth.. C. Williams.S. C.W. Srinivasan. H.. Hauck. 1495–1510. Chan. J.M.