EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS 1

Transcription

1 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION Vol. 45, No. 3 AMERICAN WATER RESOURCES ASSOCIATION June 2009 EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS 1 Huidae Cho and Francisco Olivera 2 ABSTRACT: The spatial variability of the data used in models includes the spatial discretization of the system into subsystems, the data resolution, and the spatial distribution of hydrologic features and parameters. In this study, we investigate the effect of the spatial distribution of land use, soil type, and precipitation on the simulated flows at the outlet of small watersheds (i.e., watersheds with times of concentration shorter than the model computational time step). The Soil and Water Assessment Tool model was used to estimate runoff and hydrographs. Different representations of the spatial data resulted in comparable model performances and even the use of uniform land use and soil type maps, instead of spatially distributed, was not noticeable. It was found that, although spatially distributed data help understand the characteristics of the watershed and provide valuable information to distributed hydrologic models, when the watershed is small, realistic representations of the spatial data do not necessarily improve the model performance. The results obtained from this study provide insights on the relevance of taking into account the spatial distribution of land use, soil type, and precipitation when modeling small watersheds. (KEY TERMS: geospatial analysis; surface water hydrology; precipitation; runoff; watersheds; simulation.) Cho, Huidae and Francisco Olivera, Effect of the Spatial Variability of Land Use, Soil Type, and Precipitation on Streamflows in Small Watersheds. Journal of the American Water Resources Association (JAWRA) 45(3): DOI: /j x INTRODUCTION A number of researchers have studied the effect of the spatial variability of the watershed data used in models on simulated flows and constituent loads (Cotter et al., 2003; Kalin et al., 2003; Chen and Mackay, 2004; Tripathi et al., 2006). The spatial variability of the data used in models includes the spatial discretization of the system into subsystems, the data resolution, and the spatial distribution of hydrologic features and parameters. Much has been discussed about system discretization and data resolution (Faurès et al., 1995; Bingner et al., 1997; Manguerra and Engel, 1998; FitzHugh and Mackay, 2000; Andréassian et al., 2001; Di Luzio et al., 2002; Muttiah and Wurbs, 2002; Cotter et al., 2003; Chen and Mackay, 2004; Jha et al., 2004; Chaplot, 2005; Chaplot et al., 2005; Haverkamp et al., 2005; Olivera et al., 2006); however, the effect of the spatial distribution of the hydrologic features and parameters is less well known. Although distributed hydrologic models allow the use of spatially distributed information of the watershed, the complexity of the models does not 1 Paper No. JAWRA P of the Journal of the American Water Resources Association (JAWRA). Received January 24, 2008; accepted November 10, ª 2009 American Water Resources Association. Discussions are open until six months from print publication. 2 Respectively, Graduate Research Assistant and Associate Professor, Department of Civil Engineering, Texas A&M University, 3136 TAMU, College Station, Texas ( Olivera: folivera@civil.tamu.edu). JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 673 JAWRA

2 CHO AND OLIVERA necessarily imply more preferable results (Perrin et al., 2001). The distributed hydrologic model Soil and Water Assessment Tool (SWAT) (Arnold et al., 1998) was used to assess the effect of the spatial distribution of land use, soil type, and precipitation on simulated streamflows. It is important to stress, though, that the results could be different when addressing water quality variables, but that discussion is beyond the scope of this paper. The SWAT geographic information system (GIS) interface available in BASINS 3.1 (Di Luzio et al., 2002) was used to develop the SWAT input files of the case studies. SWAT was developed to assess the long-term impact of land use and land management changes on hydrologic responses. In SWAT, the watershed is subdivided into subwatersheds, and unique combinations of land use and soil type in each subwatershed are referred to as hydrologic response units (HRUs). In GIS, HRUs are obtained by intersecting land use, soil type, and subwatershed georeferenced polygons, so that each resulting polygon has only one land use, one soil type and is part of only one subwatershed. HRUs are not georeferenced for modeling purposes; that is, their location and spatial distribution within the subwatershed is not taken into account. In fact, spatially disconnected identical combinations of land use and soil type within a subwatershed constitute a single HRU. Runoff generated at the HRUs is calculated either with the Soil Conservation Service (1972) curve number method or the Green and Ampt (1911) infiltration method, aggregated by subwatershed, and routed through the stream network to the watershed outlet using the variable storage routing method (Williams, 1969) or the Muskingum routing method (Lawler, 1964). In SWAT, one stream segment is defined for each subwatershed and is connected dendritically to up and down streams to construct the stream network. Subwatershed size depends on the threshold value used for stream initiation, which, in turn, affects the stream network density. The stream network, which is the flow routing structure, may play an important role in large watersheds but is not expected to be a factor for flow estimation in small watersheds, in which the time of concentration is likely to be shorter than the model computational time step of one day (i.e., 24 h), and most raindrops travel to the outlet within a single time step. Studies of the effect of the spatial discretization (i.e., subwatershed size) on flow estimations have been conducted by several researchers. Mamillapalli et al. (1996) found that finer discretization schemes and increased number of HRUs improved runoff flow estimations for a 4,297-km 2 watershed in Texas. However, they also found that there is a threshold beyond which increased model complexity does not lead to better model results and makes the model computationally more expensive. Haverkamp et al. (2005) demonstrated that the model efficiency, based on the Nash-Sutcliffe (1970) coefficient (NS), improves as more spatial heterogeneity of the watershed is taken into account. They used an entropy function to quantify the heterogeneity of the spatial input data and assumed that model efficiency is maximized when the entropy of the model parameters becomes equal to that of the spatial modeling units, such as subwatersheds or HRUs. In contrast, for a 21.3-km 2 watershed in northern Mississippi, Bingner et al. (1997) showed that the prediction of annual runoff volume was not greatly affected by the number and size of subwatersheds. They attributed these results to the aggregation effects of subwatersheds and HRUs. Using a 47.3-km 2 watershed in Wisconsin, Chen and Mackay (2004) also found that the level of watershed subdivision does not have a significant influence on annual streamflows, although it does on annual sediment yields. They recommended defining the maximum number of subwatersheds that the model interface allows and one HRU per subwatershed. Mamillapalli (1997), cited in Manguerra and Engel (1998), found that, for eight watersheds ranging in area from 2,000 to 5,000 km 2, different discretization schemes did not make a significant difference in flow estimation when there were enough HRUs to represent the spatial variability of the watersheds. Manguerra and Engel (1998) also evaluated the effect of discretization schemes on the estimation of monthly streamflow for 3.28-km 2 and km 2 watersheds in west central Indiana and a km 2 watershed in Mississippi, and showed that the adoption of a single subwatershed further subdivided into HRUs was sufficient to take into account the spatial variability of the watersheds. They suggest that a detailed discretization method, such as subwatersheds or grid elements, should be applied only when there are site-specific water impoundments or when there is the need for visualizing distributed output. According to FitzHugh and Mackay (2000) and Jha et al. (2004), subwatershed size does not significantly affect monthly and annual streamflow predictions because the overall precipitation abstractions are the same regardless of the subwatershed sizes. For 2,000- km 2 and 18,000-km 2 watersheds in Iowa, Jha et al. (2004) found a slight increase in annual streamflow due to increasing transmission gains in subsurface flow and a decrease in transmission losses with decreasing subwatershed size. Olivera et al. (2006), for a 116-km 2 watershed in Texas, found that simulated daily and monthly flows did not show much difference whether parameter values were assigned to each HRU based on their individual land use and soil type or averaged over the subwatersheds. They also JAWRA 674 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

3 EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS noted that longer times of concentration or shorter computational time steps could affect the estimation of HRU parameters during calibration and, consequently, also the flow estimation. Research on the effect of the resolution of the spatial data (i.e., topography, land use, or soil type) on model outputs has shown that model predictions are affected by the aggregation of model parameters. Using a 131-km 2 watershed and six large watersheds ranging in area from 2,807 to 7,812 km 2 in Texas, Muttiah and Wurbs (2002) found that flow was not sensitive to the resolution of soil type and precipitation data, except for one of the watersheds in which high soil type variability and wet climate were dominant. Di Luzio et al. (2005) showed that, for a small 21.3-km 2 watershed in Mississippi, coarser digital elevation models (DEM) resulted in a lower simulated monthly runoff volume due to the reduced accuracy of subwatershed delineation, coarser land use data increased monthly runoff volume, and coarser soil data decreased monthly runoff volume. They highlighted that the coarsest DEM data are not adequate for delineating small watersheds because of the inaccurate identification of drainage divides. DEM resolution has been found to have more impact on simulated flow than land use and soil type data resolution (Cotter et al., 2003); however, in small watersheds, increasing the DEM resolution does not improve the estimation of mean monthly flow because topography does not affect the estimation of rainfall abstractions by SWAT (Chaplot, 2005). Even though the conclusions of Di Luzio et al. (2005) appear to contradict those of Cotter et al. (2003), one refers to the minimum DEM resolution needed to correctly identify watershed boundaries while the other to the maximum DEM resolution beyond which no further improvements in runoff estimations are achieved. Andréassian et al. (2001) observed that decreasing the number of rain gauges resulted in large uncertainties in runoff prediction when modeling runoff in 71-km 2, 1,120-km 2, and 10,700-km 2 watersheds. However, for a 51-km 2 watershed in Iowa and a 918-km 2 watershed in central Texas, Chaplot et al. (2005) found that decreasing the rain gauge concentration did not significantly affect the estimation of monthly flows because of the averaging caused by considering such long time intervals. In this study, we investigate the effect of the spatial distribution of land use, soil type, and precipitation for a given subwatershed configuration on the simulated flows. Our hypothesis is that in small watersheds (i.e., watersheds with times of concentration shorter than SWATs computational time step of one day) spatial variability does not significantly affect the simulated flows because the entire watershed drains within one day, making the location where each raindrop fell immaterial. To evaluate our hypothesis, lumped and randomly distributed land use and soil type maps were used to contrast the estimated streamflows with those obtained with the original spatially distributed data for three watersheds ranging from 277 to 1,005 km 2. Similarly, the estimated streamflows obtained with multiple and single rain gauge models were compared. The findings from this study provide insights on the relevance of accounting for the spatial distribution of land use, soil type, and precipitation when estimating streamflows in small watersheds with a daily computational time step. METHODOLOGY To evaluate the effect of the watershed spatial distribution of land use and soil type on the simulated flows, alternative land use and soil type maps were created as explained below. The National Land Cover Data (NLCD) (USGS, 2006b) was used to create two alternative land use maps: a single land use map and a randomly distributed land use map. The single land use map had the land use category that was most frequent in the watershed; while, in the randomly distributed land use map, the total area of each land use was kept, but the location of the different areas of each land use were shuffled randomly within the watershed. Similarly, the State Soil Geographic database (STATSGO) (USDA-NRCS, 2006) was used to create a single soil type map and a randomly distributed soil type map. The parameter values estimated from the NLCD (USGS, 2006b) and the STATSGO (USDA-NRCS, 2006) data were recalculated in the calibration process. Likewise, because of the spatial variability of the precipitation and relief, even with a single land use and soil type, not the same amount of runoff was generated in the different subwatersheds. To address the effect of the spatial variability of the precipitation, additional models with a single rain gauge were developed for comparison purposes. The single rain gauges were selected based on their proximity to the centroid of each watershed. The calibration of the model parameters was performed with the Shuffled Complex Evolution algorithm (SCE-UA) (Duan et al., 1993). The SCE-UA method is widely used in hydrology because it is a robust and efficient global search optimization algorithm (Eckhardt and Arnold, 2001; Muttil and Liong, 2004). In calibration, the spatially distributed parameter values were changed according to a JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 675 JAWRA

4 CHO AND OLIVERA predefined one-parameter rule, which decreased the number of decision variables and kept the relative parameter values between spatial modeling units. The rule was p new ¼ p 0 þ ajp b p 0 j; ð1þ where p new is the new parameter value, p 0 is the initial parameter value, p b is either the upper or lower boundary of the parameter value, and a is a real number that varies from )1 to +1 and is the decision variable evolved by the SCE-UA algorithm. The value of p b is defined as p b ¼ 0:5p u ð1 þ sgn aþþ0:5p l ð1 sgn aþ; where p u and p l are the upper and lower limits of the parameter, respectively, and sgn a is +1 if a is positive and )1 ifa is negative. That is, for a = )1, p b = p l ; and for a = +1, p b = p u. For streamflow calibration, Neitsch et al. (2002a) recommends adjusting nine of the parameters listed in Table 1 plus the average slope steepness and the average slope length (SLOPE and SLSUBBSN, respectively, in the SWAT documentation). In this study, SLOPE and SLSUBBSN were not included in calibration because they could be estimated more accurately from the topographic data. In addition to the nine parameters recommended by Neitsch et al. (2002a), the ground-water delay time and the deep aquifer percolation fraction (GW_DELAY and RCHRG_ DP, respectively, in the SWAT documentation), which affect ground-water flow and baseflow, were adjusted. Thus, the calibration process had 11 decision variables. The decision variables and the range in which they were allowed to vary in calibration ð2þ are listed in Table 1. The objective function used in calibration was the root mean squared error (RMSE) of the streamflows: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 X n RMSE ¼ ðq obs;i Q sim;i Þ 2 ; ð3þ n i¼1 where Q obs,i and Q sim,i are the observed and simulated streamflows on the i th day, respectively, and n is the number of days of the calibration period. It is to be noted that the objective function of Equation (3) tends to give more weight to high flows when compared with low flows because errors in high flows are usually greater in absolute value than errors in low flows (Gan and Burges, 1990; Gan and Biftu, 1996; Eckhardt and Arnold, 2001; van Griensven and Bauwens, 2001; Huisman et al., 2003). Therefore, low flows might not be accurately estimated, which could have large impacts on the stream s water quality. This problem, however, is beyond the scope of this study. After calibration, three types of model validation were conducted: spatial validation, temporal validation, and spatiotemporal validation. In spatial validation, the models were validated at a different location within the watershed for the same period as in calibration; in temporal validation, the models were validated at the same location but for a different period; and, in spatiotemporal validation, the models were validated at a different location within the watershed and for a different period. These validations assessed the applicability of the model parameters obtained in calibration to different scenarios. The NS coefficient (Nash and Sutcliffe, 1970) was used to assess the performance of SWAT models. The use of this coefficient for the assessment of hydrologic model performances is recommended by the American TABLE 1. Model Parameters for Streamflow Calibration (Neitsch et al., 2002a). Parameter Description Range CN2 Initial NRCS runoff curve number for moisture condition II SOL_AWC Available water capacity of the soil layer (mm H 2 O mm soil) ESCO Soil evaporation compensation factor GWQMN Threshold depth of water in the shallow aquifer required for return flow to occur (mm H 2 O) GW_REVAP Ground-water revap coefficient REVAPMN Threshold depth of water in the shallow aquifer for revap or percolation to the deep aquifer to occur (mm H 2 O) GW_DELAY Ground-water delay time (days) RCHRG_DP Deep aquifer percolation fraction CH_K2 Effective hydraulic conductivity in main channel alluvium (mm h) ALPHA_BF Base flow alpha factor (days) OV_N Manning s n value for overland flow JAWRA 676 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

5 EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS Society of Civil Engineers (1993). The NS coefficient is defined as P n i¼1 NS ¼ 1 ðq obs;i Q sim;i Þ 2 P n i¼1 ðq obs;i Q obs Þ 2 ; ð4þ where Q obs is the mean value of the observed flows for the period for which records are available. Q obs is defined as the no-model condition because no hydrologic concept is used to estimate it. The no-model is the simplest possible flow prediction model (i.e., the long-term average flow). Note that the NS coefficient assesses how the hydrologic model compares to the no-model. That is, high coefficients do not necessarily imply a good match to observed values, but could indicate poor performance of the no-model and vice versa. NS values of one correspond to a perfect model, zero to a model that does as well as the no-model, and negative to a model that performs poorer than the no-model. Because the no-model for this study considered all the available historical data, for each watershed, the same no-model was used to calculate model efficiencies for calibration and validation. Each of the 18 models of each watershed (i.e., three land use distributions, three soil type distributions, and two precipitation distributions, giving = 18 models) was calibrated independently and, for evaluation purposes, their performance during validation was compared with the performance of the other models. FIGURE 1. East Fork of the San Jacinto River Watershed in Texas. APPLICATION Study Area and Hydrologic Data The three watersheds shown in Figures 1 and 2 were selected for this study. The three watersheds were selected because each had a U.S. Geological Survey (USGS) flow gauge at the outlet and another within it, and a period of record of at least 12 years. The areas of the East Fork of the San Jacinto River ( ), Barton Creek ( ), and Onion Creek ( ) watersheds are 1,005, 277, and 831 km 2, respectively. A summary of the land use and soil type areas according to the NLCD (USGS, 2006b) and STATSGO (USDA-NRCS, 2006) data is presented in Tables 2 and 3, respectively. Forest is dominant in the East Fork of the San Jacinto River watershed, while both forest and rangeland are major land uses in the Barton Creek and Onion Creek watersheds. Sand is the dominant soil type in the FIGURE 2. Barton Creek and Onion Creek Watersheds in Texas. East Fork of the San Jacinto River watershed, silt and sand in the Barton Creek watershed, and clay and silt in the Onion Creek watershed. Additionally, the recharge zone of the Edwards aquifer underlies part of the Onion Creek watershed in Hays and Travis counties, with flow gauges and located upstream and downstream of it, JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 677 JAWRA

6 CHO AND OLIVERA TABLE 2. Land Use Distribution According to NLCD Land Use East Fork of the San Jacinto River (%) Barton Creek (%) Onion Creek (%) Urban Agriculture Pasture Forest Rangeland Water Wetland Note: NLCD, National Land Cover Data. Soil Type TABLE 3. Soil Type Distribution According to STATSGO (first layer only). East Fork of the San Jacinto River (%) Barton Creek (%) Onion Creek (%) Clay Silt Sand respectively. In the Barton Creek case, both gauges are located upstream of the Edwards aquifer. The aquifer s outcrop area is known to be heavily fractured with major faults (Khorzad, 2003; Lindgren et al., 2004); therefore, the highly permeable soils in the recharge zone were assigned a value of 950 mm h for hydraulic conductivity and for specific yield as reported by Lindgren et al. (2004) for median values. Watersheds and subwatersheds were delineated using the 30-meter resolution National Elevation Dataset (USGS, 2006a) with a threshold area of 10 km 2 for the Barton Creek watershed and 20 km 2 for the other two watersheds. The number of subwatersheds delineated in the East Fork of the San Jacinto River, Barton Creek, and Onion Creek watersheds was 20, 16, and 61, respectively. The significantly higher number of subwatersheds in Onion Creek was caused by the addition of outlet points at the aquifer s boundary. The model calibration period ranged from January 1, 1989, to December 31, 1994, and the validation period from January 1, 1995, to December 31, In both cases, the first two years of simulation were used for the initial stabilization of the model and only four years were used to evaluate the model performance. Daily precipitation and temperature data were obtained from the National Climatic Data Center (NCDC) website (NOAA-NCDC, 2006). Flow data were obtained from the USGS National Water Information System website (USGS, 2006c). Rain and flow gauges are shown in Figures 1 and 2 for the three watersheds studied. Summaries of the precipitation and streamflow data are presented in Tables 4 and 5, respectively. Rain gauges in the same watershed recorded similar amounts of annual rainfall in both the calibration and validation periods. For each of the three watersheds, the rain gauge closest to its centroid was selected for the single-gauge models (i.e., rain gauge for the East Fork of the San Jacinto River watershed and for the Barton Creek and Onion Creek watersheds), and Pearson s correlation coefficients between them and the other rain gauges in their watershed were calculated. Higher correlation coefficients were observed in the East Fork of the San Jacinto River watershed than in the other two watersheds. Although the amounts of annual rainfall were similar, low correlation coefficients suggest that their patterns of rising and declining do not match well and that the precipitation fields are non-uniform, and vice versa. Table 5 shows that the standard deviations of the daily flows were three or more times greater than the mean values, which is caused by a small number of extremely high flows. TABLE 4. Summary of Annual Rainfall Data. Calibration Period Validation Period NCDC Gauge Watershed Mean (mm year) SD (mm year) Pearson s Correlation Coefficient Mean (mm year) SD (mm year) Pearson s Correlation Coefficient East Fork of the 1, , * San Jacinto River 1, , , , , , Barton Creek and 1, * Onion Creek 1, , Onion Creek 1, , , *Rain gauge against which Pearson s correlation coefficients were calculated. JAWRA 678 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

7 EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS TABLE 5. Summary of Daily Average Streamflow. Watershed USGS Gauge Drainage Area (km 2 ) Calibration Period Validation Period Mean (m 3 s) SD (m 3 s) Mean (m 3 s) SD (m 3 s) East Fork of the San Jacinto River , Barton Creek Onion Creek RESULTS The NS coefficients for calibration and validation based on daily flows are presented in Tables 6-9. Figure 3 shows plots of simulated vs. observed flows at the outlet of the East Fork of the San Jacinto River watershed for calibration and the three validations conducted for the distributed, lumped, and random models. The distributed model considered the original land use and soil type maps, and multiple rain gauges; the lumped model considered the single land use and soil type maps, and a single rain gauge; and the random model considered the randomly distributed land use and soil type maps, and multiple rain gauges. Similar plots were prepared for the TABLE 6. Nash-Sutcliffe Coefficients for Calibration. Soil Type Multiple Rain Gauges Single Rain Gauge Calibration Original Single Random Original Single Random East Fork of the San Jacinto River Land use: Original Land use: Single Land use: Random Barton Creek Land use: Original Land use: Single Land use: Random Onion Creek Land use: Original Land use: Single Land use: Random TABLE 7. Nash-Sutcliffe Coefficients for Spatial Validation. Soil Type Multiple Rain Gauges Single Rain Gauge Spatial Validation Original Single Random Original Single Random East Fork of the San Jacinto River Land use: Original Land use: Single Land use: Random Barton Creek Land use: Original Land use: Single Land use: Random Onion Creek Land use: Original Land use: Single Land use: Random JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 679 JAWRA

8 CHO AND OLIVERA TABLE 8. Nash-Sutcliffe Coefficients for Temporal Validation. Soil Type Multiple Rain Gauges Single Rain Gauge Temporal Validation Original Single Random Original Single Random East Fork of the San Jacinto River Land use: Original Land use: Single Land use: Random Barton Creek Land use: Original )0.22 )0.20 )0.22 )0.22 )0.28 )0.01 Land use: Single )0.44 )0.20 )0.19 )0.39 )0.10 )0.35 Land use: Random )0.28 )0.32 )0.15 )0.22 )0.10 )0.20 Onion Creek Land use: Original Land use: Single Land use: Random TABLE 9. Nash-Sutcliffe Coefficients for Spatiotemporal Validation. Soil Type Multiple Rain Gauges Single Rain Gauge Spatiotemporal Validation Original Single Random Original Single Random East Fork of the San Jacinto River Land use: Original Land use: Single Land use: Random Barton Creek Land use: Original )1.05 )1.00 )1.04 )0.54 )0.61 )0.32 Land use: Single )1.37 )0.97 )0.95 )0.65 )0.49 )0.69 Land use: Random )1.15 )1.25 )0.93 )0.53 )0.46 )0.58 Onion Creek Land use: Original )0.71 )0.48 )0.47 )0.26 )0.28 )0.16 Land use: Single )0.75 )0.78 )0.02 )0.43 )0.49 )0.40 Land use: Random )1.21 )1.02 )1.09 )0.11 )0.66 )0.01 Barton Creek and Onion Creek watersheds and for the other models, but were not included here because of space limitations. These plots showed overall similar patterns as those of Figure 3. Likewise, Figure 4 shows annual values of different hydrologic components for the East Fork of the San Jacinto River watershed during the calibration period. Again, similar maps were developed for the Barton Creek and Onion Creek watersheds, for the other models and for validation, but were not included here because of space limitations. Overall, all sets of maps showed similar patterns. Only the water yield, surface runoff, lateral flow, ground-water flow and evapotranspiration (WYLD, SURQ, LATQ, GW_Q and ET in the SWAT documentation, respectively) are shown in the figure because other components such as the transmission losses in tributaries (TLOSS in the SWAT documentation) were negligible in comparison. For the subbasins of the models presented in Figure 4, the water yield varied from 260 to 580 mm year, the surface runoff from 0 to 150 mm year, the lateral flow from 0 to 260 mm year, ground-water flow from 160 to 490 mm year, the evapotranspiration from 890 to 1,020 mm year, and transmission losses from 0 to 2 mm year. Calibration In calibration, as shown in Table 6, no clear advantage of using one land use or soil type map over the others was identifiable. Likewise, the single rain gauge models performed almost as well as their corresponding multiple rain gauge models, except for the case of the Onion Creek watershed with random land use and original soil type, which performed slightly better. JAWRA 680 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

9 EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS FIGURE 3. Daily Simulated vs. Observed Streamflows at Flow Gauge for the Distributed, Lumped, and Random Models for Calibration, Spatial Validation, Temporal Validation, and Spatiotemporal Validation. The dotted lines represent the long-term mean flow. The relatively low NS coefficients in the case of the East Fork of the San Jacinto River compared with those of Barton Creek and Onion Creek can be explained by the presence of a few extremely high flows that could not be accurately simulated by SWAT. For example, there was a severe flood in the East Fork of the San Jacinto River watershed during the calibration period that conveyed 1,320 m 3 s at gauge on October 19, To assess the effect of this single value on the model efficiency, another set of calibrations were performed excluding this unusually high value. In these new models, the sum of the square of the residuals (SSR) of the simulated flows significantly decreased and, consequently, the NS coefficients increased. For the model with original land use, original soil type, and multiple rain gauges, the SSR of the calibrated models with and without the 1,320 m 3 s flow value were 1,362,433 and 384,465 m 6 s 2, respectively; while the SSR of the no-model was 2,446,531 m 6 s 2. In terms of the NS coefficient, it increased from 0.44 to That is, the presence of an unusually high-flow value and the inability of SWAT to predict it significantly affected the NS coefficient. The calibrated models of the Barton Creek and Onion Creek watersheds had high NS coefficients, which resulted from successfully predicting high flows combined with the fact that the their no-models did not. In Figure 3, for the East Fork of the San Jacinto River watershed, it can be seen that the random model did similar to the distributed model except in low-flow days, which were mostly underpredicted. In the case of the lumped model, a bimodal distribution of flows was observed with values concentrated around 0.1 and 10 m 3 s but less frequent in the vicinity of 1 m 3 s. The lumped model underpredicted in low-flow and some average-flow days with respect to the distributed model. Note, however, that errors in low-flow days are small compared with those in highflow days and that their effect on the objective function and NS coefficient tends to be small. JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 681 JAWRA

10 CHO AND OLIVERA FIGURE 4. Annual Water Yield, Surface Runoff, Lateral Flow, Ground-Water Flow, and Evapotranspiration for the Distributed, Lumped, and Random Models in Calibration. Figure 4 shows dissimilar spatial distributions of water yield, surface runoff, lateral flow, ground-water flow, and evapotranspiration for the distributed, lumped, and random models of the East Fork of the San Jacinto River watershed. Note that the hydrologic components in the lumped model are somewhat uniformly distributed because the land use and soil type distributions are uniform. The reason why they are not equal is the variability of the topography. The dissimilar surface flow, lateral flow, ground-water flow, and evapotranspiration values lead to dissimilar water yield values in the subwatersheds; however, after routing the water yield through the stream network, it produced almost equal streamflows at the outlet and very similar model efficiencies. Spatial Validation Table 7 shows the model performance in spatial validation. The East Fork of the San Jacinto River models performed as well as in calibration; but the Barton Creek and Onion Creek models did not, although their NS coefficients are within the range of acceptable values found in the literature (Hanratty and Stefan, 1998; King et al., 1999; Rosenthal and Hoffman, 1999; Spruill et al., 2000; Eckhardt and Arnold, 2001; Weber et al., 2001; Fontaine et al., 2002; Neitsch et al., 2002b; Eckhardt et al., 2003; Tripathi et al., 2003; Olivera et al., 2006). In general, performances in spatial validation similar to those of calibration are expected in nested watersheds of comparable areas; however, the Barton Creek models did not do as well. For the Onion Creek case, note that flow gauge is located right upstream of the Edwards aquifer s recharge zone, while flow gauge lies on the aquifer s artesian zone. The high rate of ground-water recharge between the two gauges increases uncertainty in the parameter estimation in calibration because less accurate runoff generation from upstream can be compensated for by adjusting ground-water related parameters in the aquifer. JAWRA 682 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

11 EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS In the East Fork of the San Jacinto River models, the single rain gauge models did as well in spatial validation as the multi rain gauge models. The similarity of the performances in both cases is explained by the fact that the precipitation in the watershed is highly uniform, as indicated by the similar amounts of rainfall and the high correlation coefficients (see Table 4). In the case of the Barton Creek watershed, the similarity is not explained by uniformity of the precipitation field (see Table 4), but by the fact that the rain gauge used in the single-gauge models covers 83% of the watershed area when multiple rain gauges are used. Finally, in the case of the Onion Creek watershed, the dissimilarity was caused by both non-uniformity of the precipitation field and the fact that the rain gauge used in the single-gauge models covers only 58% of the watershed area when multiple rain gauges are used. With regard to the daily flows in the East Fork of the San Jacinto River watershed, Figure 3 shows similar patterns as those of calibration. That is, the random model underpredicted in low-flow days and the lumped model showed a bimodal distribution of flows in which low-flow and some average-flow days were underpredicted. Temporal Validation Table 8 presents the results of temporal validation. In this case, the multiple rain gauge models of the East Fork of the San Jacinto River performed similar to calibration, but the single rain gauge models showed lower NS coefficients than in calibration. These lower NS values were obtained despite the fact that the precipitation correlation coefficients were somewhat high. In the Barton Creek and Onion Creek watershed cases, rain gauge NCDC , which was used in the multiple and single rain gauge models, had three days of precipitation greater than 100 mm (i.e., June 11, 1997; January 7, 1998; and October 18, 1998) that did not generate high flows at the corresponding flow gauges. That is, in given days, high precipitation depths caused high simulated flows, when observed flows did not show a noteworthy change. In those days, simulated flows were appreciably different from the observed flows, while the no-model flows (i.e., long-term average flow) did not differ from them significantly. Consequently, the SSR of the model resulted greater than the SSR of the no-model, and the NS coefficient became negative. As can be seen in Figure 3, the daily flows in the East Fork of the San Jacinto River watershed were distributed as in calibration and spatial validation. In this case, however, the NS coefficient of the lumped model (i.e., 0.47) was higher than those of the distributed (i.e., 0.43) and random (i.e., 0.41) models (Table 8). The explanation for the better performance of the lumped model is that its errors considerable in a logarithmic scale plot are mostly in small flows, which do not significantly affect the SSR of the simulated flows and the NS coefficient. Spatiotemporal Validation As shown in Table 9, the East Fork of the San Jacinto River models show good performance in spatiotemporal validation while the other two watershed models perform poorly. As in the previous case, in the Barton Creek and Onion Creek cases, the extremely high precipitation depths of October 18, 1998, not associated with similarly high streamflows caused large errors in the simulated flows and low or even negative NS coefficients. Again, the daily flows in the East Fork of the San Jacinto River watershed were distributed as in calibration, and spatial and temporal validation (see Figure 3). However, in this case, the NS coefficient of the lumped model (i.e., 0.65) was clearly higher than those of the distributed (i.e., 0.56) and random (i.e., 0.52) models (Table 9). As in the previous case, the explanation for the better performance of the lumped model is that its errors are mostly in small flows. DISCUSSION It was observed that the watershed models based on single and random land use and soil type maps did as well as those based on the NLCD (USGS, 2006b) and STATSGO (USDA-NRCS, 2006) datasets. In other words, the location of specific land use and soil type areas within the watershed did not significantly affect the overall performance of the models. The limited effect of the spatial variability of land use and soil type on the model performance, as indicated by the NS coefficient, is explained by the fact that, because the watershed time of concentration is shorter than one day, the entire watershed drains within a single computational time step. The impact of time resolution on the streamflows of small watersheds is conceptually similar to the buffering effect that reduces the importance of daily rainfall variability by using monthly flow estimates (Chaplot et al., 2005). Thus, unless losses take place, daily runoff and flow at the outlet should be equal. It was noticed, though, that multiple rain gauge models performed better than single rain gauge models, both in temporal and spatial validation. JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 683 JAWRA

12 CHO AND OLIVERA Additionally, to assess the effect of flow velocities in the channel network on daily streamflows, the calibrated models were run with Manning s coefficients (CH_N2 in the SWAT documentation) of and 0.01 to 0.10 at 0.01 intervals. Manning s coefficient is a channel-roughness parameter that affects flow velocities, travel times, and residence times in the watershed (Hsu et al., 2006). It was observed that the effect of the different coefficient values on the hydrographs was negligible because the time taken in the stream network was shorter than one day. Figure 5 is a plot of daily flow at the outlet vs. daily runoff generated in the drainage area for the East Fork of the San Jacinto River and the Onion Creek watersheds. Points on the 1:1 line correspond to days in which the flow and the runoff were equal, implying that there were no losses and that the draining took place within a day. On the contrary, points below the 1:1 line reflect losses or residence times in the watershed longer than one day. Points on the 1:1 line were found for the East Fork of the San Jacinto River (see Figure 5a) and Barton Creek watershed models; while, for the Onion Creek watershed, flows were lower than runoff because part of the runoff was lost to the Edwards aquifer s recharge zone before reaching the watershed outlet (see Figure 5b). Finally, although the NS coefficient is a wellknown and commonly used metric of model performance, the results of this study reveal limitations of the coefficient because of its sensitivity to the nomodel, which is independent from the model being assessed. For this reason, the comparison of NS coefficients can only provide meaningful results if they have the same no-model (i.e., same watershed). It was also observed that single events can strongly affect NS coefficients. SUMMARY AND CONCLUSIONS (a) Streamflow (m 3 s) (b) Streamflow (m 3 s) Runoff (m 3 s) Runoff (m 3 s) FIGURE 5. Simulated Streamflow vs. Simulated Runoff for the Calibration Period. (a) East Fork of the San Jacinto River watershed model with NLCD 1992 land use, STATSGO soil type, and multiple precipitation stations; and (b) Onion Creek watershed model with lumped land use, lumped soil type, and single precipitation station. The plot for Barton Creek watershed was similar to that of the East Fork of the San Jacinto River watershed. The present study discusses the impact of the spatial distribution of land use, soil type, and precipitation in small watersheds (i.e., watersheds with times of concentration shorter than the model computational time step) on simulated streamflows. The SWAT model, with a computational time step of one day, was used to estimate hydrographs at the watershed outlets. Different synthetic representations of the spatial data, including uniform and randomly distributed land use and soil type maps, resulted in model performances comparable to those obtained using the NLCD (USGS, 2006b) and STATSGO (USDA-NRCS, 2006) datasets as indicated by the NS coefficient. It was observed that, in small watersheds, realistic representations of the spatial data do not necessarily improve the model performance. In such watersheds, the routing process and, therefore, its parameters does not affect the simulated streamflows because the watershed drains in less than a computational time step, and the model can only capture the lumped effects at the outlet. Based on the case studies presented here, spatial distribution of land use, soil type, and precipitation appears to matter more for the JAWRA 684 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

13 EFFECT OF THE SPATIAL VARIABILITY OF LAND USE, SOIL TYPE, AND PRECIPITATION ON STREAMFLOWS IN SMALL WATERSHEDS estimation runoff depth than for the determination of where it was generated and when it will reach the outlet. Although the three watersheds analyzed in this study presented different hydrologic characteristics, generalization of these conclusions should be made carefully. ACKNOWLEDGMENTS This study was supported in part by the Ministry of Science and Technology of the Korean Government through the Korea Science and Engineering Foundation grant M The authors would like to thank the manuscript reviewers for their input and comments. LITERATURE CITED American Society of Civil Engineers, Criteria for Evaluation of Watershed Models. ASCE Task Committee on the Definition of Criteria for Evaluation of Watershed Models of the Watershed Management Committee, Irrigation and Drainage Committee. Journal of Irrigation and Drainage Engineering 119(3): Andréassian, V., C. Perrin, C. Michel, I. Usart-Sanchez, and J. Lavabre, Impact of Imperfect Rainfall Knowledge on the Efficiency and the Parameters of Watershed Models. Journal of Hydrology 250: Arnold, J.G., R. Srinivasan, R.S. Muttiah, and J.R. Williams, Large Area Hydrologic Modelling and Assessment, Part I: Model Development. Journal of the American Water Resources Association 34(1): Bingner, R.L., J. Garbrecht, J.G. Arnold, and R. Srinivasan, Effect of Watershed Subdivision on Simulation Runoff and Fine Sediment Yield. Transactions of the ASAE 40(5): Chaplot, V., Impact of Dem Mesh Size and Soil Map Scale on SWAT Runoff, Sediment, and NO 3 -N Loads Predictions. Journal of Hydrology 312: Chaplot, V., A. Saleh, and D.B. Jaynes, Effect of the Accuracy of Spatial Rainfall Information on the Modeling of Water, Sediment, and NO 3 -N Loads at the Watershed Level. Journal of Hydrology 312: Chen, E. and D.S. Mackay, Effects of Distribution-Based Parameter Aggregation on a Spatially Distributed Agricultural Nonpoint Source Pollution Model. Journal of Hydrology 295: Cotter, A.S., I. Chaubey, T.A. Costello, T.S. Soerens, and M.A. Nelson, Water Quality Model Output Uncertainty as Affected by Spatial Resolution of Input Data. Journal of the American Water Resources Association 39(4): Di Luzio, M., J.G. Arnold, and R. Srinivasan, Effect of GIS Data Quality on Small Watershed Stream Flow and Sediment Simulations. Hydrological Processes 19: Di Luzio, M., R. Srinivasan, and J.G. Arnold, Integration of Watershed Tools and SWAT Model Into BASINS. Journal of the American Water Resources Association 38(4): Duan, Q.Y., V.K. Gupta, and S. Sorooshian, Shuffled Complex Evolution Approach for Effective and Efficient Global Minimization. Journal of Optimization Theory and Applications 76(3): Eckhardt, K. and J.G. Arnold, Automatic Calibration of a Distributed Catchment Model. Journal of Hydrology 251: Eckhardt, K., L. Breuer, and H.-G. Frede, Parameter Uncertainty and the Significance of Simulated Land Use Change Effects. Journal of Hydrology 273: Faurès, J.-M., D.C. Goodrich, D.A. Woolhiser, and S. Sorooshian, Impact of Small-Scale Spatial Rainfall Variability on Runoff Modeling. Journal of Hydrology 173: FitzHugh, T.W. and D.S. Mackay, Impacts of Input Parameter Spatial Aggregation on an Agricultural Nonpoint Source Pollution Model. Journal of Hydrology 236: Fontaine, T.A., T.S. Cruickshank, J.G. Arnold, and R.H. Hotchkiss, Development of a Snowfall-Snowmelt Routine for Mountainous Terrain for SWAT. Journal of Hydrology 262: Gan, T.Y. and G.F. Biftu, Automatic Calibration of Conceptual Rainfall-Runoff Models: Optimization Algorithms, Catchment Conditions and Model Structure. Water Resources Research 32(12): Gan, T.Y. and S.J. Burges, An Assessment of a Conceptual Rainfall-Runoff Model s Ability to Represent the Dynamics of Small Hypothetical Catchments, 1. Models, Model Properties, and Experimental Design. Water Resources Research 26(7): Green, W.H. and G.A. Ampt, Studies on Soil Physics, 1. The Flow of Air and Water Through Soils. Journal of Agricultural Sciences 4: van Griensven, A. and W. Bauwens, Integral Water Quality Modelling of Catchments. Water Science and Technology 43(7): Hanratty, M.P. and H.G. Stefan, Simulating Climate Change Effects in a Minnesota Agricultural Watershed. Journal of Environmental Quality 27: Haverkamp, S., N. Fohrer, and H.-G. Frede, Assessment of the Effect of Land Use Patterns on Hydrologic Landscape Functions: A Comprehensive GIS-Based Tool to Minimize Model Uncertainty Resulting From Spatial Aggregation. Hydrological Processes 19: Hsu, M.-H., J.-C. Fu, and W.-C. Liu, Dynamic Routing Model With Real-Time Roughness Updating for Flood Forecasting. Journal of Hydraulic Engineering 132: Huisman, J.A., L. Breuer, K. Eckhardt, and H.-G. Frede Spatial Consistency of Automatically Calibrated SWAT Simulations in the Dill Catchment and Three of Its Subcatchments. In: Proceedings of the 2003 International SWAT Conference, R. Srinivasan, J. Jacobs, and R. Jensen (Editors). Texas Water Resources Institute (TWRI) Technical Report 266, College Station, Texas, pp Jha, M., P.W. Gassman, S. Secchi, R. Gu, and J. Arnold, Effect of Watershed Subdivision on SWAT Flow, Sediment, and Nutrient Predictions. Journal of the American Water Resources Association 40(3): Kalin, L., R.S. Govindaraju, and M.M. Hantush, Effect of Geomorphologic Resolution on Modeling of Runoff Hydrograph and Sedimentograph Over Small Watersheds. Journal of Hydrology 276: Khorzad, K., Edwards Aquifer Evaluation: Kinney County, Texas. Journal of the American Water Resources Association 39(5): King, K.W., J.G. Arnold, and R.L. Bingner, Comparison of Green-Ampt and Curve Number Methods on Goodwin Creek Watershed Using SWAT. Transactions of the ASAE 42(4): Lawler, A.L., Chapter 25. Hydrology of Flood Control Part II. Flood Routing. In: Handbook of Applied Hydrology, V.T. Chow (Editor). McGraw-Hill Book Company, New York, pp to Lindgren, R.J., A.R. Dutton, S.D. Hovorka, S.R.H. Worthington, and S. Painter, Conceptualization and Simulation of the JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 685 JAWRA