Engineering Department, University of Georgia, Tifton, Georgia. Corresponding author: Jaepil Cho, USDAâARS Southeast Watershed. Research Laboratory ...
EFFECT OF SPATIAL DISTRIBUTION OF RAINFALL ON TEMPORAL AND SPATIAL UNCERTAINTY OF SWAT OUTPUT J. Cho, D. Bosch, R. Lowrance, T. Strickland, G. Vellidis
ABSTRACT. Accurate rainfall data are critical for accurate representation of temporal and spatial uncertainties of simulated watershed‐scale hydrology and water quality from models. In addition, the methods used to incorporate the rainfall data into the simulation model can significantly impact the results. The objectives of this study were to (1) assess the hydrologic impacts of different methods for incorporating spatially variable rainfall input into the Soil and Water Assessment Tool (SWAT) in conjunction with subwatershed delineation level and (2) assess seasonal and spatial uncertainty in hydrologic and water quality simulations of SWAT with respect to rain gauge density. The study uses three different methods to incorporate spatially variable rainfall into the SWAT model and three levels of subwatershed delineation. The impacts of ten different gauge‐density scenarios on hydrology and water quality were subsequently evaluated by using the highest gauge‐density scenario as a baseline for comparison. Through the centroid method, which is currently used by the AVSWAT‐X interface, variations in the representation of measured annual rainfall as model input and corresponding simulated streamflow increased as subwatershed delineation level decreased from high‐density to low‐density. The rainfall input by the Thiessen averaging method for each subwatershed (Thiessen method) and the inverse‐distance‐weighted averaging method for the entire watershed (average method) were not sensitive to subwatershed delineation. The impacts of delineation on streamflow were also less with these two methods. The Thiessen method is recommended for SWAT simulation of a watershed with high spatial variability of rainfall. The currently used AVSWAT‐X centroid method will also accurately represent spatially variable rainfall if a subwatershed delineation is used that sufficiently incorporates the density of observed rainfall stations. As the number of rain gauges used for the simulation decreased, the uncertainty in the hydrologic and water quality model output increased exponentially. Total phosphorus was most sensitive to the changes in rain gauge density, with an average coefficient of variation of root mean square difference (CVRMSD ) of 0.30 from three watersheds, followed by sediment, total nitrogen, and streamflow, showing CVRMSD values of 0.24, 0.18, and 0.17, respectively. Seasonal variations in simulated streamflow and water quality were higher during summer and fall seasons compared to spring and winter seasons. These seasonal and temporal variations can be attributed to the rainfall patterns within the watershed. Keywords. Hydrology, Rainfall, Sediment, SWAT, Temporal and spatial uncertainty, Total nitrogen, Total phosphorus.
A
ccurate modeling of runoff and transport of diffu‐ sive nonpoint‐source pollutants at the watershed scale requires accurate representation of spatially distributed input parameters. Spatially distrib‐ uted model input parameters, representative of topography, soils, and land uses within a selected watershed, can be easily obtained through readily available national GIS databases. Several studies have been conducted to assess the impact of the spatial resolution of land use, soils, and topographic cov‐ erages on the accuracy of the simulated model output (Bosch et al., 2004; Chaplot, 2005; Chaubey et al., 2005; Wang and Melesse, 2006). In addition, accurate representation of spa‐
Submitted for review in September 2008 as manuscript number SW 7690; approved for publication by the Soil & Water Division of ASABE in August 2009. The authors are Jaepil Cho, ASABE Member Engineer, Postdoctoral Research Associate, David Bosch, ASABE Member Engineer, Research Hydraulic Engineer, Richard Lowrance, Research Ecologist, and Timothy Strickland, Supervisory Soil Scientist, USDA‐ARS Southeast Watershed Research Laboratory, Tifton, Georgia; and George Vellidis, ASABE Member Engineer, Professor, Biological and Agricultural Engineering Department, University of Georgia, Tifton, Georgia. Corresponding author: Jaepil Cho, USDA‐ARS Southeast Watershed Research Laboratory, Tifton, 2329 Rainwater Rd., GA 31793‐0748; phone: 229‐391‐6854; fax: 229‐386‐3958; e‐mail: Jaepil.Cho@ars. usda.gov.
tially distributed rainfall is essential in hydrologic and water quality modeling because rainfall is the major driving force of runoff and contaminant transport. Using multiple rain gauges has advantages in considering spatially variable hydrologic processes within a rainfall‐runoff model. Even though spatially uniform rainfall is generally assumed in modeling the hydrological behavior of small watersheds, the assumption of spatially uniform rainfall may not be valid for larger watersheds or watersheds that experience convective rainfall (Faures et al., 1995; Goodrich et al., 1995). As a re‐ sult, many studies have been conducted to evaluate the im‐ pact of spatial variability of rainfall on the accuracy of model predictions (Andreassian et al., 2001; Chaplot et al., 2005; Chaubey et al., 1999; Duncan et al., 1993; Hamlin, 1983; Ma‐ millapalli, 1998; Moon et al., 2004; Shah et al., 1996). Uncertainties in modeling output as a function of spatial variability of rainfall can be caused by two different sources of error (Troutman, 1983): (1) inappropriate input parameters or representation (input error), and (2) error in the model structure and algorithms (structural error). Many studies have focused on input error. In the intensive review of previous SWAT applications by Gassman et al. (2007), inadequate rep‐ resentation of rainfall input due to a lack of rain gauges was indicated to be one cause of weak simulation results. Mamil‐ lapalli (1998) compared four different algorithms for repre‐
Transactions of the ASABE Vol. 52(5): 1545-1555
2009 American Society of Agricultural and Biological Engineers ISSN 0001-2351
1545
senting multiple rain gauge input within the SWAT model. In the study, subwatershed delineation levels were not consid‐ ered in conjunction with the rainfall input algorithms even though spatial distribution of rainfall can be influenced by subwatershed delineation level due to the model structure. SWAT structure allows only one rainfall input for each delin‐ eated subwatershed (Neitsch et al., 2002). Spatial variability of rainfall can consequently be disregarded by selecting a coarse subwatershed delineation despite multiple rain gauges being available throughout the watershed. If multiple rain gauges exist within a delineated subwatershed, the model se‐ lects the rain gauge nearest the centroid of the watershed and uses data only from it. As a result, a method using single Thiessen average input over the entire watershed was recom‐ mended by Mamillapalli (1998). While these and other studies have focused on simulation error as a function of the representation of rainfall input, the combined impacts of temporal variability of rainfall and structural error have rarely been discussed. Because of the methods used to incorporate rainfall data into the SWAT model, watershed delineations can be selected that incorpo‐ rate the level of rainfall data available in a given watershed. However, little guidance is available for selecting the ap‐ propriate watershed delineation as a function of rainfall den‐ sity. In addition, other methods for incorporating rainfall variability into the model that are less sensitive to watershed delineation need to be explored. Few studies have been performed to extend the analysis to water quality (Chaplot et al., 2005; Chaubey et al., 1999). In long‐term simulation using SWAT, runoff and nitrogen (NO3‐N) output varied slightly while the accuracy of sedi‐ ment loads greatly improved with increasing rain gauge den‐ sity (Chaplot et al., 2005). A study using the AGNPS model showed large uncertainty in the output of runoff, total sedi‐ ment, and sediment‐attached nitrogen and phosphorus ac‐ cording to the spatial variability of rainfall (Chaubey et al., 1999). Generalizations regarding the impacts of rain gauge density on the uncertainties of water quality results are diffi‐ cult because the degree of spatial variability of rainfall and representation of watershed response using models depends on the characteristics of the selected watershed and model. SWAT was used in the Little River experimental wa‐ tershed (LREW), one of 14 ARS benchmark watersheds for the Conservation Effects Assessment Project (CEAP) (USDA, 2007) to evaluate the impacts of conservation prac‐ tices on hydrology and water quality by considering spatial distribution of input parameters (Bosch et al., 2004; Feyerei‐ sen et al., 2007; Van Liew et al., 2005; Van Liew et al., 2007). Bosch et al. (2004) evaluated the accuracy of hydrologic sim‐ ulation of SWAT according to different spatial resolutions of soil and land use data within the LREW. SWAT predicted the total water yield and general streamflow trends more accu‐ rately using high spatial resolution input data with modified initial conditions (Bosch et al., 2004). Rainfall patterns with‐ in the LREW showed the characteristics of convective thun‐ derstorms with higher intensity and more frequent occurrence during summer compared to other seasons (Bosch et al., 1999). However, the impact of rainfall input for SWAT on hydrologic and water quality simulations has not been evaluated by considering the structure/algorithm of the mod‐ el and spatial/temporal variability of rainfall within the LREW.
1546
The overall goal of this study is to assess the impacts of spatial and temporal variability of rainfall on SWAT hydro‐ logic and water quality simulation results by considering the structural and input effects. The specific objectives are to: (1)assess the hydrologic impacts of different methods for in‐ corporating spatially variable rainfall input into SWAT in conjunction with subwatershed delineation level, and (2) as‐ sess seasonal and spatial uncertainty in hydrologic and water quality simulations of SWAT with respect to rain gauge den‐ sity.
MATERIALS DESCRIPTION OF THE STUDY WATERSHED The study area under consideration is the Little River ex‐ perimental watershed (LREW), located near Tifton, Georgia, in the South Atlantic Coastal Plain (fig. 1). The 334 km2 LREW is a mixed land‐use watershed that contains row crop agriculture, pasture and forage, upland forest, and riparian forest. The surface soil textures on the watershed are general‐ ly sands and sandy loams with high infiltration rates. Surface soils are underlain by the upper part of the shallow and rela‐ tively impermeable Hawthorne formation, which restricts downward movement of infiltrated precipitation and leads to lateral movement to the stream channels (Sheridan, 1997). The LREW landscape is dominated by a dense dendritic net‐ work of stream channels bordered by riparian forest wetlands (Asmussen et al., 1979). The climate of the LREW is humid subtropical with a long growing season. The LREW rain gauge network consists of 31 stations within and immediately outside the LREW boundaries. The spatial distribution of rainfall within the watershed has been monitored by the USDA‐ARS Southeast Watershed Research Laboratory (SEWRL) since 1967 (Bosch et al., 2007; Sheridan, 1997). Subwatersheds K (LRK), J (LRJ), and I (LRI) in the headwa‐ ters of the LREW were selected for this research because of the denser spatial distribution of rain gauges in this area (fig.1). Rainfall events during the summer yield less average depth, have shorter duration and higher intensity, and occur more frequently than events during other seasons of the year (Bosch et al., 1999). MODELING FRAMEWORK AND AVAILABLE DATA The AVSWAT‐X user interface was used to create primary inputs for SWAT 2005. The spatial data include a USGS digi‐ tal elevation model (DEM) (http://seamless.usgs.gov/web‐ site/seamless/), USDA‐NRCS soil survey geographic (SSURGO) coverage (http://soildatamart.nrcs.usda.gov/), and land‐use from a 1998 landsat image obtained through the Georgia GIS data clearinghouse (https://gis1.state.ga.us/). Six possible crop rotations based on a typical six‐year crop rotation (corn‐peanut‐corn‐peanut‐cotton‐peanut) were ran‐ domly assigned on classified crop areas. The typical crop rotation over the simulation period from 1985 to 1994 was de‐ fined according to the harvested crop area data obtained from the USDA National Agricultural Statistics Service for Turner County, Georgia (USDA, 2008). A two‐year equilibration pe‐ riod, 1983 and 1984, was simulated to allow the model to es‐ tablish initial conditions prior to the period examined. The model performance guidelines proposed by Moriasi et al. (2007) for monthly time‐steps were selected for evaluating the simulation performance in the study. Model simulation
TRANSACTIONS OF THE ASABE
Figure 1. Location of Little River experimental watershed and spatial distribution of selected subwatersheds and rain gauge stations.
was considered very good if the monthly Nash‐Sutcliffe effi‐ ciency index (NSE) was >0.75, monthly RMSE‐observations standard deviation ratio (RSR) was
