Spatial Distributions and Stochastic Parameter Influences on SWAT Flow and Sediment Predictions

Transcription

1 Spatial Distributions and Stochastic Parameter Influences on SWAT Flow and Sediment Predictions Kati W. Migliaccio 1 and Indrajeet Chaubey 2 Abstract: The Soil and Water Assessment Tool SWAT was implemented in Northwest Arkansas to investigate flow and sediment predictive ability at multiple subbasin and hydrologic response units HRU distributions. The objectives of this study were to use SWAT and identify differences in annual predicted flow and sediment response of a watershed considering two subbasin delineations with six different HRU distributions each; quantify the uncertainty in SWAT output when sensitive model parameters are considered to have stochastic distribution; and evaluate the ability of the model to describe flow and sediment predictions of an ungauged watershed by stochastically validating the model. Flow results from the Single-Factor Between-Subjects Analysis of Variance test =0.05 indicated that predicted flow means were not significantly different from each other for all SWAT subbasin/hru combinations; however, predicted flow means and measured flow mean were significantly different. Sediment simulation results suggested significant differences were present amongst the different model subbasin/hru delineations and measured values. The Monte Carlo simulation of the model, using curve number CN, soil evaporation compensation factor ESCO, groundwater revap coefficient GW_REVAP, and peak rate adjustment factor for sediment routing in the subbasin AMP as uncertain parameters, indicated that generally ESCO induced most uncertainty in predicted flow. However, sediment prediction uncertainty was affected most by uncertainty in AMP. Results indicated that SWAT applications on ungauged watersheds should include small subbasin sizing 2% of watershed area, HRUs that reflect actual land cover composition, a check to evaluate surface runoff and ground water contributions and modification of parameters as needed to reflect site conditions, sensitivity analysis, and uncertainty analysis that includes several sensitive parameters. DOI: / ASCE :4 258 CE Database subject headings: Spatial distribution; Stochastic process; Parameters; Sediment; Arkansas. Introduction Watershed models are used to represent landscape processes that exhibit spatial and temporal heterogeneity. They describe a relatively large area by spatially designating land areas into hydrologically connected units with each having its own user defined characteristics. Watershed models generally provide some provision for the user to specify the spatial representation within the watershed in the form of cells, subbasins, or some other spatial unit Grayson and Blöschl Therefore, it is often to the user s discretion how the watershed is divided spatially. The manner in which a watershed is designated into unique units is important in hydrologic and water quality modeling because this generally is the smallest spatial unit for which characteristics can be entered and for which predicted outputs can be simulated. Hence, how the user chooses to delineate a watershed into smaller units will influence the ability of the model to mimic the natural 1 Assistant Professor, Univ. of Florida Tropical Research and Education Center, SW 280 St., Homestead, FL klwhite@ufl.edu 2 Associate Professor, Agricultural and Biological Engineering Dept., and Dept. of Earth and Atmospheric Sciences, Purdue Univ., 225 South University St., West Lafayette, IN Note. Discussion open until September 1, Separate discussions must be submitted for individual papers. To extend the closing date by one month, a written request must be filed with the ASCE Managing Editor. The manuscript for this paper was submitted for review and possible publication on March 28, 2006; approved on June 19, This paper is part of the Journal of Hydrologic Engineering, Vol. 13, No. 4, April 1, ASCE, ISSN /2008/ /$ system and predict representative output Grayson and Blöschl 2000; Jha et al. 2004; Lopes and Canfield In an ideal world, watershed modelers would like to have the smallest spatial units feasible when implementing a model, but at some point the increase in spatial resolution outweighs the benefits by increasing cost and time beyond reasonable limits. As model spatial resolution increases, model input requirements increase. The availability, cost, and obtainability of high resolution data must balance with the resources available for a particular modeling application Haverkamp et al In addition, there is a spatial resolution threshold where new information is not gained in regards to meeting modeling objectives by increasing the resolution of input data Cotter et al The discretization of watershed areas into smaller increments, such as subbasins, units, or cells, has been shown to influence model predictions e.g., Vieux and Needham 1993; Bingner et al. 1997; FitzHugh and Mackay 2000; Kalin et al. 2003; Arabi et al Whereas different models were considered in these listed studies, results indicated that flow, sediment, and/or nutrient predictions varied depending on the subdivisions used in defining the model. The identification of optimal spatial divisions of a model to predict a desired variable is not unique to any one model, but is an element influencing all models that allow for discretization. The Soil and Water Assessment Tool 2000 SWAT is a watershed model that computes hydrologic pathway characteristics on three spatial levels Arnold et al The first level is the watershed level where parameters are included that possess no spatial variability within the watershed. Some of the first level parameters found in SWAT include snow melt parameters, reach 258 / JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008

2 evaporation adjustment factor, and surface runoff lag coefficient Neitsch et al. 2001b. The second spatial level of input and parametrization available in SWAT is the subbasin level. Subbasins basically represent spatial areas within the larger watershed that are hydrologically connected. Hence, subbasins are defined by an entry point and an outlet point for a stream reach so that the subbasin is considered to be the area contributing flow to the stream reach between these two points. Subbasins are initially defined in the SWAT watershed delineation process; however, there is some user ability to modify these designated areas. After the initial delineation process, users can add or remove subbasin outlets to change the stream reach area and therefore the subbasin area for a particular subbasin. Subbasins are characterized by input and parameter values in SWAT, such as weather data, bank storage coefficient, and stream channel hydraulic conductivity Neitsch et al. 2001b. The third spatial level available in SWAT is the hydrologic response unit HRU. HRUs represent the unique combinations of soil and land cover within each subbasin at a specified distribution and are considered to be hydrologically homogeneous. HRU distributions are set by the user as a percentage value for soil and land cover. The distribution percent is used by the model so that only combinations of soil and land cover that are greater than the set distribution percentage are considered. Soil and land cover combinations below the set HRU distribution percentage are lumped into the combinations above the set distribution percentage and therefore not represented in the model. HRUs provide the greatest resolution for parametrization in SWAT; however, HRUs do not possess spatial orientation with respect to each other within a subbasin. Parameters at the HRU level include such variables as the curve number, groundwater parameters, and Manning s roughness coefficient n Neitsch et al. 2001b. A complete list of parameters corresponding to watershed, subbasin, and HRU levels is provided in the SWAT User s Manual Neitsch et al. 2001b. Even though discretization of a watershed into small spatial units is one of the central characteristics of distributed parameter models, there is no universally accepted guideline available detailing how this should be achieved when using such a model for watershed response predictions. The characterization of a watershed into subbasins and HRUs for SWAT depends upon user specifications. As many SWAT parameters are derived from land cover, management, and soil distributions, it is possible that the criteria used for discretizing a watershed into subbasins and HRUs may influence parameter values and resulting uncertainty in model outputs. In the absence of data for model calibration, quantification of output uncertainty due to spatial representation of the watershed input data should be assessed and minimized to appropriately interpret modeling results. It should be noted that SWAT was developed to be used in ungauged watersheds Arnold and Fohrer Previous researchers have investigated the relationship between subbasin delineation and flow predictions using SWAT e.g., Bingner et al. 1997; FitzHugh and Mackay 2000; Jha et al Their results indicated that the subbasin delineation used did not influence SWAT annual and/or monthly flow predictions substantially. Alternatively, annual sediment predictions have been found by some to be dependent on the area of subbasins chosen during delineation Bingner et al. 1997; Jha et al Arabi et al suggested that subbasins should be 4% of total watershed area to improve model predictive abilities for evaluating Best Management Practices. These previous investigations into watershed delineation and respective predictive abilities did Fig. 1. Illinois River watershed boundary in Northwest Arkansas not consider HRU distributions within subbasin delineations. They also used deterministic parameter inputs and outputs; hence, they did not consider the stochastic nature of input parameters and resulting output variables. There are a vast number of parameters in SWAT and many of these parameters may be more accurately represented by a distribution than a deterministic value Chaubey et al. 2003; Chaubey and White This is primarily because of the stochastic characteristic of natural processes that prevents them from being properly described by a deterministic value. Therefore, parameters that describe natural processes and that greatly influence predicted model values should be modeled as a distribution to account for the uncertainty in their estimation Haan In any modeling application, uncertainty in model parameters propagates to output uncertainty. Parameter uncertainty can be minimized through model calibration; a process in which model parameters are adjusted until the model predictions match measured data within a predefined accuracy level or meet a predefined objective function. However, model calibration is not possible for ungauged watersheds. This implies that reductions in parameter uncertainty that result from calibration will not occur when modeling ungauged watersheds. Therefore, if models such as SWAT are to be used on ungauged watersheds, quantification of this uncertainty is necessary to appropriately interpret the results. Model outputs presented as probability density functions PDF under uncertain parameter conditions would help evaluate watershed models where there are no observed data available for model calibration Haan et al The objectives of this study were to use SWAT and: 1 identify differences in annual predicted flow and sediment response of a watershed considering two subbasin delineations with six different HRU distributions each; 2 quantify the uncertainty in SWAT output when sensitive model parameters are considered to have stochastic distribution; and 3 evaluate the ability of the model to describe flow and sediment predictions of an ungauged watershed by stochastically validating the model. Our primary goal is to provide guidance for using SWAT on ungauged watersheds. Study Site The study site is the Illinois River watershed in Northwest Arkansas Fig. 1, which encompasses approximately 1,470 km 2. The headwaters of the Illinois River are located in Northwest Arkansas and flow west into Oklahoma, eventually entering Lake JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008 / 259

3 Tenkiller. The watershed is a portion of the eight-digit U.S. Geological Survey USGS designated hydrologic unit code HUC and contains predominantly pasture 55%, forest 37%, and urban 8% land cover CAST Four waste water treatment plants WWTP are located within the watershed and discharge into tributaries of the Illinois River. Soils in the watershed were determined using U.S. Department of Agricultural-Natural Resources and Conservation Service USDA-NRCS U.S. General Soil Map STATSGO. Three soil series were identified: Clarksville, Linker, and Enders. They are described as gravelly silt loam, fine sandy loam, and gravelly fine sandy loam, respectively. Clarksville series was the most prevalent in the watershed. Surface runoff, interflow, and ground water flow are dominant flow paths of water into the stream system. The proportion of each differs depending on precipitation events. During storm events, the proportion of flow from surface runoff and interflow increases; whereas during dry periods, the flow is mostly comprised of ground water. Table 1. SWAT Configurations with Nomenclature, Subbasin Delineation, and HRU Distributions Defined Model nomenclature Subbasin delineation HRU distribution % land cover/ % soil Average subbasin area ha Average HRU area ha D-05 Default 5/ 30 4, D-07 Default 7/ 30 4, D-10 Default 10/ 30 4,740 1,090 D-20 Default 20/ 30 4,740 1,650 D-25 Default 25/ 30 4,740 2,260 D-00 Default Dominant 4,740 4,740 D2-05 Default/2 5/ 30 2, D2-07 Default/2 7/ 30 2, D2-10 Default/2 10/ 30 2, D2-20 Default/2 20/ 30 2,580 1,030 D2-25 Default/2 25/ 30 2,580 1,280 D2-00 Default/2 Dominant 2,580 2,580 Methods SWAT SWAT is a physically based, watershed model developed by U.S. Department of Agriculture Agriculture Research Service USDA-ARS Arnold et al. 1998; Srinivasan et al It simulates watershed response on a continuous time step with input options for hydrology, nutrients, erosion, land management, main channel processes, water bodies, and climate data. SWAT predicts the influence of land management practices on constituent yields from a watershed and includes agricultural components such as fertilizer, crops, tillage options, and grazing. SWAT also has the capability to include point source discharges Neitsch et al. 2001a. Hydrology in SWAT is based on a water balance that includes soil water content, precipitation, surface runoff, evapotranspiration, percolation and bypass flow, and return flow. Precipitation data may be input by the user or may be generated using SWAT s weather generator, WXGEN Sharpley and Williams Surface runoff is estimated using the Soil Conservation Services SCS curve number procedure or the Green & Ampt infiltration method. SWAT has three different options for determining evapotranspiration: Penman Monteith, Priestly Taylor, and Hargreaves. Percolation and bypass flow are a function of soil properties and soil water content. SWAT estimates erosion using the Modified Universal Soil Loss Equation MUSLE. The watershed model also includes considerations for snow cover effects on erosion, sediment lag in surface runoff, and sediment in lateral and groundwater flows. More information on model components and functionality can be obtained from the SWAT Theoretical Documentation Neitsch et al. 2001a. Data Used We used AvSWAT 2000 with the following GIS data to parametrize SWAT to simulate response for Illinois watershed from 1998 to 2002: 30-m DEM U.S. Geological Survey, USGS, 28.5-m 1999 land use and land cover image file CAST 2004, and STATSGO soils shape file USEPA AvSWAT 2000 is the version of the SWAT 2000 model that is interfaced through ESRI ArcView software. Weather data from two stations within the region were incorporated to provide the most representative precipitation and temperature data available Fig. 1. This simulation time period was selected due to available data at the time of study initiation. Also, to include a much longer time period would not be advised for this watershed due to land cover changes in subsequent years. As land cover is a primary component in the model and change of this layer would result in modification of HRUs and related parameters, significant change in land cover would require a different SWAT configuration. For all model applications in our study, the model was actually simulated from 1995 to Years prior to 1998 were included as warm up for the model so that more accurate initial values could be generated. Spatial Discretizations To evaluate selected subbasin/hru distribution influences on predicted flow and sediment values, twelve different SWAT configurations were created using two subbasin delineations with six HRU distributions in each Table 1. The first subbasin delineation was the default delineation provided in the SWAT interface, which was designated with a threshold of 3,300 ha Fig. 2. The second subbasin delineation was obtained by setting the threshold to 1,650 ha or one half of the default threshold Fig. 3. These subbasin delineations result in subbasin sizes that are 3 and 2% of the total watershed area, corresponding to watershed area sizes suggested for subbasin delineation when simulating sediment by Jha et al and Arabi et al For both subbasin delineations, four subbasin outlets were added for the four WWTP dischargers. The WWTP contributions were input using the.fig file. The stream network generated using a subbasin delineation set at 3,300 ha threshold did not extend to the tributary branches where the WWTPs were located; therefore, outlets were placed at the closest stream network location to the WWTP discharge. This was considered an acceptable compromise as we were only considering output at the annual scale and WWTPs were located in headwater streams which were a reasonable distance from the watershed outlet Fig. 1. Six different HRU distributions were designated within the two subbasin delineations. HRUs were designated as dominant 260 / JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008

4 Fig. 2. Soil and Water Assessment Tool SWAT subbasin delineation with subbasin delineation threshold set to 3,300 ha Fig. 3. SWAT subbasin delineation with subbasin delineation threshold set to 1,650 ha land cover and soil at 5, 7, 10, 20, and 25% land cover with consistently 30% soils Table 1. Soil thresholds were not changed because of the low resolution of STATSGO soils data. Sensitivity Analysis Sensitivity analysis was conducted to determine the influence a set of parameters had on predicting annual flow volume and sediment yield using model configuration D2-07. Sensitivity was approximated using the relative sensitivity S r S r = x y 2 y 1 y x 2 x 1 where x=parameter and y=predicted output. x 1, x 2 and y 1, y 2 correspond to ±10% of the initial parameter and output values, respectively James and Burges The greater the absolute value of S r the more sensitive a model output variable was to that particular parameter. Parameters were selected for sensitivity analysis by reviewing previously used calibration parameters and by reviewing documentation from SWAT manuals. This procedure was previously employed to identify parameters for sensitivity analysis for another watershed within the same region White and Chaubey 2005 ; as our watershed and the watershed used by White and Chaubey 2005 are in similar ecoregions, we used the parameters identified by White and Chaubey 2005 to conduct sensitivity analysis. The four parameters with the greatest absolute S r value were selected for stochastic analyses. 1 Monte Carlo Simulations The four parameters found to contribute the most uncertainty having the greatest absolute S r values for flow and sediment were simulated as distributions instead of deterministic values using Monte Carlo analysis. Monte Carlo analysis was conducted by developing a Microsoft Visual Basic VB program that would vary input randomly based on a user defined distribution. The VB program SWAT simulator was created with a GUI interface for ease of operation. The VB program allowed a user specified number of simulations and for congregation of output to minimize computational time. Output was generated using the Monte Carlo approach for 500 model simulations and evaluated as a distribution for each subbasin/hru model combination Haan Output was generated on a daily time step and aggregated for annual output values. Evaluation of Results Results from each subbasin/hru combination were compared to each other and measured data to determine significant differences amongst the groups Objective 1. Significant differences between the twelve different subbasin/hru configurations for annual flow volume and annual sediment predictions were evaluated using a single-factor between-subjects analysis of variance SFBSAOV with =0.05. The SFBSAOV tests the null hypothesis that the mean of each population is equal to each other with an alternative hypothesis that the means of each population are not equal to each other Sheskin We used measured data from the USGS gauge Illinois River South of Siloam Springs, Ark. Fig. 1 USGS Mean flow at the gauge from 1998 to 2002 was 18.1 m 3 s 1 with a standard deviation of 1.8 m 3 s 1. About twice-a-month water quality sampling, including suspended sediment concentration, occurred at the USGS gauge; daily measured constituent concentrations were not available. The sediment load was estimated from collected samples using LOADEST2 software Crawford 1991; Results indicated a mean sediment load of 4.15 kg s 1 with a median, 25th percentile, and 75th percentile values of 0.58, 0.26, and 1.63 kg s 1 for 1998 to Relative error RE, % in mean predicted output was calculated using RE % = z y z where z=measured value and y=predicted value. Stochastic distributions were developed using results from Monte Carlo simulations for each of the twelve SWAT configurations. Stochastic validation was pursued by further analyzing modeling results; the D2-07 SWAT configuration results are presented in this paper for brevity Objective 2. Model predictions were evaluated using the coefficient of variability CoV in simulated flow and sediment for each year based on 500 model runs JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008 / 261

5 Table 2. HRU Distribution Results with Land Cover Areas for Each Subbasin/HRU SWAT Configuration Model nomenclature No. of subbasins No. of HRUs Resulting land cover area ha % of total Forest Pasture Urban Other Measured 54, , , D , , , D , , , D , , , D , , , D , , , D , , , D , , , D , , , D , , , D , , , D , , , D , , , x i x 2 i CoV = x 500 x i i=1 x = where x i =ith predicted value for each year. Confidence intervals were tabulated based on Monte Carlo simulation results. The fraction of total variance in the model output F i attributable to the ith parameter was estimated Haan 2002 F i = 2 i n 2 i i=1 0.5 where i =correlation between ith parameter and the output. The ability of the D2-07 SWAT configuration to describe flow and sediment predictions of an ungauged watershed was determined using stochastic validation Objective 3. In summary, the measured means were compared to predicted Monte Carlo distributions using a PDF plotting technique. Results were separated annually from 1998 to 2002 and compared to respective measured values. The model was considered to predict watershed response satisfactorily if the measured values were within the 95% confidence interval of the predicted distribution. Results and Discussion 3 5 Comparison of Multiple Subbasin Delineations and HRU Distributions Each unique SWAT delineation is presented with resulting land cover representation in Table 2. The composition of each SWAT subbasin/hru combination is further described by the average size of subbasins and HRUs in Table 1. Sensitivity analysis of the model indicated that many parameters affected flow and sediment predictions Table 3. Flow and sediment predictions were most sensitive to curve number CN and the soil evaporation compensation factor ESCO. In addition, ground water revap coefficient GW_REVAP is another parameter affecting flow predictions only and is often used in SWAT flow calibrations. Similarly, peak rate adjustment factor for sediment routing in the subbasin AMP is another parameter that affects only sediment predictions. These four parameters CN, ESCO, GW_REVAP, and AMP were described by a distribution rather than a deterministic value to account for their stochastic nature during Monte Carlo model simulations. Determining an input distribution for the CN parameter was somewhat complex due to the derived nature of the parameter. The CN value is based on a combination of the hydrologic soil group, land use description, and soil antecedent moisture and is used to calculate runoff volume where Haan et al P 0.2S 2 Q =, P 0.2S 6 P + 0.8S S = 25, CN and Q, P, and S are in millimeters where Q=accumulated runoff volume or rainfall excess; P = accumulated precipitation; and S = maximum soil water retention parameter. As values of CN are based on physical properties, a distribution cannot be derived directly and instead are generated using S. The S parameter has been shown to be log normally distributed with a standard deviation of 0.5 times the mean of S Haan and Schulze The expected values of S were calculated using the expected values of CN derived from soils and land use for each HRU Chaubey et al Note that this is a standard procedure for deriving the expected values of CN in all modeling applications. The ESCO, GW_REVAP, and AMP parameters were considered to be described by a uniform distribution. A uniform distribution was chosen as the only information available was the range of values for each parameter Santhi et al ; for ESCO, for GW_REVAP, and for AMP. All four parameters were varied using a Monte Carlo technique resulting in 500 predictions. Results from the model simulations indicated that flow was overpredicted for all subbasin/ HRU combinations and was characterized by minimal variance among SWAT configurations except for Model D-00 Table 4. Although the D-00 Model predicted flow is closer in value to the measured flow, this does not necessarily imply that the D-00 spa- 262 / JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008

6 Table 3. Parameters Evaluated in SWAT Sensitivity Analysis with Relative Sensitivity S r Absolute Values Parameter abbreviation Parameter definition parameter range S r flow S r sediment ALPHA_BF Baseflow alpha factor a AMP Peak rate adjustment factor for sediment routing in the subbasin NA CH_N1 Manning s n value for the tributaries a CN2 Initial SCS runoff curve number for moisture condition II ±10% b ESCO Soil evaporation compensation factor b GW_DELAY Groundwater delay time NA GW_QMN Threshold depth of water in the shallow aquifer required for return flow to occur NA GW_REVAP Groundwater revap coefficient a OVN Manning s n value for overland flow a PRF Peak rate adjustment factor for sediment routing in the main channel NA SOL_AWC Available water capacity of the soil layer NA SPCON Linear parameter for calculating the maximum amount of sediment that can be reentrained during channel sediment routing a SURLAG Surface runoff lag coefficient NA Note: NA=not available. From SWAT documentation. b Santhi et al tial representation is the best. Further statistical investigation was needed to assess the reason such a different value was obtained for the D-00 model. The SFBSAOV test was conducted for flow using two different groups of samples: 1 all predicted data and measured data and 2 all predicted data. Flow results from the SFBSAOV test =0.05 level indicated a rejection of the null hypothesis that the means were equal to each other for the group that included all predicted and measured flow data F=2.44, p=0.01. Results for predicted flows indicated that the means were not significantly different amongst all predicted data F=1.00, p=0.46. Therefore, even though on visual inspection of the predicted flow values the D-00 Model appeared to predict a different value from other Table 4. Results from 500 Simulations per Model Using Annual Predictions from 1998 to 2002 Model Average flow m 3 s 1 Standard deviation m 3 s 1 Average sediment kg s 1 Standard deviation kg s 1 Measured NA D D D D D D D D D D D D Note: NA=not available. subbasin/hru model flow predictions, it statistically was not different. In addition, the flow results indicated that the measured data mean was significantly different from the twelve configurations predicted means. Whereas flow predictions were generally similar between models, sediment predictions were found to have more variation between subbasin/hru delineations. Average sediment predictions ranged from 3.39 to 6.53 kg s 1. SFBSAOV was also used to evaluate sediment predictions and measured values. Two groupings were used: 1 all predicted and measured data and 2 all predicted data. Results indicated that the means were significantly different considering all predicted and measure data F=1.99, p=0.04 and all predicted data F=1.98, p=0.05. The variations observed in predicted sediment yields with different HRU configurations are likely a result of the land covers included in each configuration. Each land cover is assigned a Universal Soil Loss Equation cover and management factor USLE C value from the model database. This value is a component of the MUSLE equation used to calculate sediment yield. The default USLE C values for pasture and forest are and 0.001, respectively. Generally, model predictions indicate that sediment yield increases Table 4 as the portion of pasture increases and the portion of forest decreases Table 2. Minimal to no differences were observed amongst the twelve SWAT configurations and predicted flows, whereas differences were observed amongst SWAT predicted sediment values. These differences amongst various HRU/subbasin configurations, while focusing on HRU distributions, reflect similar findings that indicated minimal influence of subbasin delineation on flow and significant influence of subbasin delineation on sediment yield Bingner et al. 1997; FitzHugh and Mackay 2000; Jha et al. 2004; Arabi et al However, our study was limited to an uncalibrated model. It would be interesting to see if the limits imposed by different subbasin delineations and HRU distributions would JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008 / 263

7 Table 5. Effect of CN, ESCO, GW_REVAP, and AMP Parameter Uncertainty on Predicted Flow for D2-07 SWAT Year Mean m 3 s 1 CoV a % 95% CI b Measured m 3 s 1 CN ESCO GW_REVAP AMP All parameters a Coefficient of variability. b Confidence interval. influence the ability of the model user to calibrate. It is likely that this would influence calibration as parametrization is available only at these subbasin and HRU spatial levels. Quantifying Uncertainty in SWAT Predicted Flow and Sediment Response Effects of uncertainty in CN, ESCO, GW_REVAP, and AMP on SWAT predicted flow are shown in Table 5 and Fig. 4 for the D2-07 SWAT configuration. Results for other models are not presented here for brevity. The PDF for each year was derived from 500 model runs using Monte Carlo analyses. The CoV and 95% confidence interval CI provide a range for the output uncertainty. When CN, ESCO, GW_REVAP, and AMP were individually treated as uncertain parameter alone, the greatest uncertainty in predicted flow resulted due to ESCO, and least uncertainty due to AMP Table 5 and Fig. 4. These results were in agreement with the relative sensitivity values obtained for each parameter. The mean predicted flow for uncertain CN was relatively higher than that with other parameters Table 5. Results from varying all four parameters in the Monte Carlo simulation indicated that both uncertainty and mean flow values were similar to results obtained from the flow uncertainty due to ESCO. Table 6 and Fig. 5 illustrate uncertainty in SWAT predicted sediment transport due to variability in CN, ESCO, and AMP. Uncertainty in predicted sediment is not shown in Table 6 because GW_REVAP had negligible uncertainty compared to other parameters. This was also evident from the model sensitivity analysis Table 3. A greater uncertainty measured as % CoV was present in predicted sediment transport compared to flow uncertainty, with AMP having the greatest impact on sediment prediction uncertainty Table 6. Also, mean predicted sediment transport under uncertain CN was generally greater compared to other parameters. As sediment transport predictions are greatly influenced by the accuracy of flow prediction, results from this study indicate that a smaller uncertainty in flow prediction may lead to considerably higher uncertainty in predicted sediment values. We conducted uncertainty analysis by identifying specific parameters for Monte Carlo simulation. These parameters were selected using sensitivity analysis. Although this is a common approach for selecting parameters for uncertainty analysis Haan et al and for SWAT model calibration Santhi et al. 2001, others have suggested that the most sensitive parameters may not correspond to parameters that introduce the most uncertainty in model predictions. Hantush and Kalin 2005 suggest that uncertainty analysis should not necessarily include only the most sensitive parameters, as their study indicated that the most sensitive parameters were not always the ones contributing the most uncertainty. Result from this study confirm the conclusions reported by Hantush and Kalin 2005 by showing that AMP had most influence on the predicted sediment uncertainly Table 6 even though it was not the most sensitive parameter, as indicated by relative sensitivity value Table 3. However, conducting uncertainty analysis on all possible parameter combinations for a complex watershed model such as SWAT can be quite challenging and time consuming due to a larger number of parameters involved. Uncertainty analysis using a Monte Carlo type technique requires a distribution and range for a parameter. Many parameters that are available in distributed watershed models are accompanied by a default value and a range. In addition, the model may assign parameter values based on a set of input data such as soils maps, land cover, elevation, etc.. The use of model assigned values for parameters based on some type of input data will likely reduce the uncertainty in model outputs associated with the range of a particular parameter. However, this assigned value will likely not reduce uncertainty that is associated with sensitivity. Parameters that are identified during sensitivity analysis as most influencing model outputs of interest should always be considered in uncertainty evaluations. We feel this is important, especially when modeling natural systems, due to the random nature of the system and the inability to capture this with deterministic parameters. Evaluation of Stochastic Simulation Results from the Monte Carlo simulation can be used to evaluate SWAT and to assess its suitability in making watershed response predictions for ungauged watersheds. For the model to be stochastically valid, the measured watershed response data must fall within the 95% CI of the model output Haan et al The measured flow for the watershed is outside 95% CI of the predicted flow indicating that technically SWAT could not be validated when CN, ESCO, GW_REVAP, and AMP are considered uncertain. It should be emphasized that the SWAT configuration used here was not calibrated; hence, all of the model parameter values were based on available input data. The RE in the mean 264 / JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008

8 Fig. 4. Effect of uncertainty in CN, ESCO, and GW_REVAP, on predicted flow for D2-07 SWAT predicted flow ranged from +12 to 36% and can be considered reasonable given the uncalibrated nature of the model. When measured sediment transport data were compared to the 95% CI of the model output, the measured sediment load fell within 95% CI of simulated loads for 1998, 1999, and 2001 Table 6 and Fig. 5. For other years, the RE in model predictions ranged from 59to+52%. The range of the 95% CI provides an estimate for the practical applicability of the model on ungauged watersheds. A watershed model should have some range in predicted output due to uncertain model parameters showing that model is flexible to be used in watersheds with different response characteristics. However, a very large range in the 95% CI suggests that model outputs are uncertain, making any useful watershed management decisions JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008 / 265

9 Table 6. Effect of CN, ESCO, and AMP Parameter Uncertainty on Predicted Sediment Transport for D2-07 SWAT Year Mean t year 1 a CoV b Measured % 95% CI c t year 1 CN , , , , , , , , , , , , , , , , , , ,900 95,250 ESCO , , , , , ,680 96, , , ,790 80, , , , , , , , ,120 95,250 AMP , , , , , , , , , , , , , , , , , , ,180 95,250 All parameters , , , , , , , , , , , , , , , , , , ,880 95,250 a Tons refers to metric tons equal to 1,000 kg. b Coefficient of variability. c Confidence interval. difficult. It should be noted that only four parameters were selected for the uncertainty analysis in this study and inclusion of more parameters in the analysis would likely increase the 95% CI. The fraction of the total variance in the predicted flow and sediment due to uncertainty in CN, ESCO, GW_REVAP, and AMP are presented in Table 7. In addition, interactions of CN, ESCO, and GW_REVAP uncertainty on predicted 1998 flow are shown in Fig. 6. Similarly, interactions of CN, ESCO, and AMP uncertainty on predicted 1998 sediment are shown in Fig. 7. Effects of uncertain AMP on predicted flow in Fig. 6 and uncertain GW_REVAP in Fig. 7 are not shown because these parameters induced negligible uncertainty in the respective model outputs. When all four parameters were considered uncertain simultaneously, most of the variance in predicted flow resulted from uncertainty in ESCO Table 7. However, the fraction of variability for predicted sediment due to ESCO was not as dominant; AMP induced the most uncertainty in all years, except in A similar conclusion can be made from Fig. 7 when uncertainty by all four parameters is very similar to uncertainty induced by AMP alone. It is important to note that this observation is biased by the expanse of the tails of the distribution that are difficult to visualize in the Fig. 7. Implications for Application of SWAT on Ungauged Watersheds Delineation of subbasins within a watershed was already shown to influence model sediment predictions Bingner et al. 1997; Jha et al and a suggested subbasin delineation of 4% of the total watershed area was proposed by Arabi et al in their study. Our study indicated an even finer subbasin size, 2% of total watershed area, might provide a better representation of the sediment load for ungauged watersheds. In addition, increase in the number of HRUs by decreasing the land cover percentage required to designate an HRU generally resulted in more accurate sediment predictions. Percentage settings for HRU formation in an ungauged watershed should be completed with comparisons between measured land cover and modeled land cover to minimize errors generated due to misrepresentation of the watershed s land cover composition. Measured values and predicted distributions displayed in PDF plots Fig. 4 indicate that SWAT predictions for all years except 2000 were overpredicted for flow. One scenario that would explain this pattern is that surface runoff contributions are overestimated. This would justify the overpredictions in the greater rainfall years and the underprediction in the drier year This would also suggest that percolation to the deep aquifer is not appropriately simulated. These conclusions would not be possible with an ungauged watershed, due to the lack of measured data for comparison purposes. Thus, simulation of hydrology in ungauged watershed does require some knowledge of the system. This knowledge should be used to adjust parameters to ensure base flow and surface runoff flow components are representative of watershed conditions. Interestingly, sediment did not show a similar pattern between predicted and measured values Fig. 5. Measured values varied in range and were likely a response to physical changes in the 266 / JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008

10 Fig. 5. Effect of uncertainty in CN, ESCO, and AMP on predicted sediment for D2-07 SWAT watershed related to spatial modifications, along with hydrological influences. Simulating sediment is always challenging, particularly in an ungauged watershed. We suggest that integration of as much detail as possible regarding land cover and stream degradation is needed to use SWAT for sediment predictions without proper calibration e.g., on ungauged watersheds. Sensitivity analysis is essential for any modeling investigation. This is also true when evaluating an ungauged watershed. Uncertainty analysis is also crucial. Sensitivity analysis and uncertainty analysis are described as a two step process with sensitivity occurring first, followed by uncertainty. Due to the vast number of parameters in SWAT, we advise that parameters be identified that JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008 / 267

11 Table 7. Fraction of the Total Variance in the Model Output Attributable to Each Parameter CN, ESCO, GW_REVAP, and AMP Parameter Flow Sediment Flow Sediment Flow Sediment Flow Sediment Flow Sediment CN ESCO GW_REVAP AMP influence outputs of interest, and these identified parameters be included in uncertainty analysis. We only investigated four parameters; other parameters affecting flow and sediment prediction would also need to be included to evaluate their relative contributions to the 95% CI of the model predictions. Conclusions Sensitivity analysis suggested that several parameters contributed to uncertainty in model predictions of flow and sediment. Four of these parameters CN, ESCO, GW_REVAP, and AMP were described using a distribution and input into twelve SWAT configurations that were differentiated by two subbasin delineations with six HRU distributions each. Results indicated that predicted flow values were not influenced by the different subbasin/hru spatial resolutions that we selected, while predicted sediment values were influenced by different subbasin/hru spatial resolutions selected. Stochastic evaluation of the D2-07 Model suggested that the model could not be validated for flow for all years. Although measured sediment loads fell within 95% CI of the simulated values for 3 out of 5 years of data evaluated in this study, these results should be interpreted in light of the watershed specific nature of the model and the limited nature of the measured data. It should be noted that no continuous sediment loads were measured in the watershed; the sediment loads were based on biweekly concentrations data and measured flow data along with a regression model. Therefore, the measured sediment data can be expected to have large uncertainty. Applicability of the SWAT model to reliably predict sediment loads needs to be evaluated in other watersheds to generalize results obtained in this study. Often complex watershed models are criticized based on their ability to represent true watershed characteristics as they have sufficient number of parameters that can be manipulated to match the model predictions with the measured response data. A stochastic validation does not suffer from this limitation as it is not biased by the manipulation of model parameters and, is therefore, a more powerful methodology to assess the applicability of a model on an ungauged watershed. Most of the variance in the flow prediction was due to ESCO. However, the greatest fraction in sediment uncertainty was explained by AMP. The range of uncertainty in the model predictions was relatively small and relative errors in model predications were reasonable indicating that the model could be applied in ungauged watersheds for flow and sediment response predictions when calibration data were not available. However, when applied in ungauged watersheds SWAT applications should include small subbasin sizing 2% of watershed area, HRUs that reflect actual land cover composition, a check to evaluate surface runoff and ground water contributions and modification of these as needed to reflect site conditions, sensitivity analysis, and uncertainty analysis that includes sensitive parameters having significant impact on selected watershed response predictions. Future investigation is needed to determine if HRU spatial characterization of SWAT will influence the user s ability to calibrate. Model predictions only at annual scales were considered in this study; future work should consider stochastic validation of the model at smaller temporal scales e.g., monthly to evaluate its usefulness for making response predictions for ungauged watersheds. Fig. 6. Effect of interactions among CN, ESCO, and GW_REVAP on predicted flow for D2-07 SWAT Fig. 7. Effect of interactions among CN, ESCO, and AMP on predicted sediment for D2-07 SWAT 268 / JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008

12 Acknowledgments The writers would like to acknowledge the University of Arkansas Division of Agriculture and the University of Florida Institute of Food and Agricultural Sciences for their support in conducting this research. The writers would like to thank the Arkansas Soil and Water Conservation Commission for funding the Watershed Modeling Laboratory and Steven Cole who assisted in development of the Monte Carlo VB programming. Excellent comments provided by three anonymous reviewers greatly improved the earlier version of this manuscript. References Arabi, M., Govindadraju, R. S., Hantush, M. M., and Engel, B. A Role of watershed subdivision on modeling the effectiveness of best management practices with SWAT. J. Am. Water Resour. Assoc., 42 2, Arnold, J. G., and Fohrer, N SWAT 2000: Current capabilities and research opportunities in applied watershed modeling. Hydrolog. Process., 19 3, Arnold, J. G., Srinivasan, R., Muttiah, R. S., and Williams, J. R Large area hydrologic modeling and assessment. I: Model development. J. Am. Water Resour. Assoc., 34 1, Bingner, R. L., Garbrecht, J., Arnold, J. G., and Srinivasan, R Effect of watershed subdivision on simulation runoff and fine sediment yield. Trans. ASAE, 40 5, Center for Advanced Spatial Technologies CAST land use/land cover data. May, 12, Chaubey, I., Costello, T. A., White, K. L., and Cotter, A. S Stochastic validation of SWAT model. Proc., Total Maximum Daily Load: Environmental Regulations II, ASAE, St. Joseph, Mich., Chaubey, I., and White, K. L Influence of hydrologic response unit HRU distribution on SWAT flow and sediment predictions. Proc., Watershed Management to Meet Water Quality Standards and Emerging TMDL (Total Maximum Daily Load), ASAE, Atlanta, Cotter, A. S., Chaubey, I., Costello, T. A., Soerens, T. S., and Nelson, M. A Water quality model output uncertainty as affected by spatial resolution of input data. J. Am. Water Resour. Assoc., 39 4, Crawford, C. O Estimation of suspended-sediment rating curves and mean suspended-sediment loads. J. Hydrol., , Crawford, C. O Estimating mean constituent loads in rivers by the rating-curve and flow-duration, rating-curve methods. Ph.D. dissertation, Univ. of Bloomington, Bloomington, Ind. Fitzhugh, T. W., and Mackay, D. S Impacts of input parameter spatial aggregation on an agricultural nonpoint source pollution model. J. Hydrol., 236 1, Grayson, R., and Bloschl, G Spatial patterns in catchment hydrology: Observations and modelling, Cambridge University Press, Cambridge, U.K. Haan, C. T Statistical methods in hydrology, 2nd Ed., Iowa State Press, Ames, Iowa. Haan, C. T., Allred, B., Storm, D. E., Sabbagh, G. J., and Prabhu, S Statistical procedure for evaluating hydrologic/water quality models. Trans. ASAE, 38 3, Haan, C. T., Barfield, B. J., and Hayes, J. C Design hydrology and sedimentology for small catchments, Academic, San Diego. Haan, C. T., and Schulze, R. E Return period flow prediction and uncertain parameters. Trans. ASAE, 30 3, Hantush, M. M., and Kalin, L Uncertainty and sensitivity analysis of runoff and sediment yield in a small agricultural watershed with KINEROS2. Hydrol. Sci. J., 50 6, Haverkamp, K., Srinivasan, R., Frede, H. G., and Santhi, C Subwatershed spatial analysis tool: Discretization of a distributed hydrologic model by statistical criteria. J. Am. Water Resour. Assoc., 38 6, James, L. D., and Burges, S. J Hydrologic modeling of small watersheds, ASAE Monograph, St. Joseph, Mich. Jha, M., Gassman, P. W., Secchi, S., Gu, R., and Arnold, J Effect of watershed subdivision on SWAT flow, sediment, and nutrient predictions. J. Am. Water Resour. Assoc., 40 3, Kalin, L., Govindaraju, R. S., and Hantush, M. M Effect of geomorphologic resolution on modeling of runoff hydrograph and sedimentograph over small watersheds. J. Hydrol., , Lopes, V. L., and Canfield, H. E Effects of watershed representation on runoff and sediment yield modeling. J. Am. Water Resour. Assoc., 40 2, Neitsch, S. L., Arnold, J., Kiniry, J. R., and Williams, J. R. 2001a. Soil and water assessment tool theoretical documentation version Nov. 10, Neitsch, S. L., Arnold, J., Kiniry, J. R., and Williams, J. R. 2001b. Soil and water assessment tool user s manual version Oct. 12, Santhi, C., Arnold, J., Williams, J. R., Dugas, W. A., Srinivasan, R., and Hauck, R Validation of the SWAT model on a large river basin with point and nonpoint sources. J. Am. Water Resour. Assoc., 37 5, Sharpley, A. N., and Williams, J. R EPIC Erosion Productivity Impact Calculator, model documentation. Technical Bulletin No. 1768, U.S. Department of Agriculture, Agricultural Research Service, Washington, D.C. Sheskin, D. J Handbook of parametric and nonparametric statistical procedures, 2nd Ed., CRC, Boca Raton, Fla. Srinivasan, R., Ramanarayanan, I. S., Arnold, J. G., and Bednarz, S. T Large area hydrologic modeling and assessment. Part II: Model application. J. Am. Water Resour. Assoc., 34 1, U.S. Environmental Protection Agency USEPA Better assessment science integrating point and nonpoint sources. May 13, U.S. Geological Survey USGS Water resources of Arkansas. Aug. 9, Vieux, B., and Needham, S Nonpoint-pollution model sensitivity to grid cell size. J. Water Resour. Plann. Manage., 119 2, White, K. L., and Chaubey, I Sensitivity analysis, calibration, and validations for a multisite and multivariable SWAT model. J. Am. Water Resour. Assoc., 41 5, JOURNAL OF HYDROLOGIC ENGINEERING ASCE / APRIL 2008 / 269