jo ur nal ho me page: www.elsev ier .com/ locate /eco lmodel
dentifying critical source areas of nonpoint source pollution withWAT and GWLF
ewati Niraulaa, Latif Kalinb,∗, Puneet Srivastavac, Christopher J. Andersonb
Hydrology and Water Resources, University of Arizona, Tucson, AZ 85721, United StatesSchool of Forestry and Wildlife Sciences, Auburn University, Auburn, AL 36849, United StatesDepartment of Biosystems Engineering, Auburn University, Auburn, AL 36849, United States
r t i c l e i n f o
rticle history:eceived 15 June 2013eceived in revised form 13 August 2013ccepted 14 August 2013
Identification of critical source areas (CSAs) (areas contributing most of the pollutants in a watershed) isimportant for cost-effective implementation of best management practices. Identification of such areasis often done through watershed modeling. Various watershed models are available for this purpose,however it is not clear if the choice (and complexity) of a model would lead to differences in locationsof CSAs. The objective of this study was to use two models of different complexity for identifying CSAs.The relatively complex Soil and Water Assessment Tool (SWAT) and the simpler Generalized WatershedLoading Function (GWLF) were used to identify CSAs of sediment and nutrients in the Saugahatchee Creekwatershed in east central Alabama. Models were calibrated and validated for streamflow, sediment, totalnitrogen (TN) and total phosphorus (TP) at a monthly time scale. While both models performed wellfor streamflow, SWAT performed slightly better than GWLF for sediment, TN and TP. Sub-watershedsdominated by urban land use were among those producing the highest amount of sediment, TN and TPloads, and thus identified as CSAs. Sub-watersheds with some amount of agricultural crops were also
identified as CSAs of TP and TN. A few hay/pasture dominated sub-watersheds were identified as CSAsof TN. The identified land use source areas were also supported by field collected water quality data.A combined index was used to identify the sub-watersheds (CSAs) that need to be targeted for overallreduction of sediment, TN and TP. While many CSAs identified by SWAT and GWLF were the same, someCSAs were different. Therefore, this study concludes that model choice will affect the location of someCSAs.
. Introduction
Approximately 67% of lakes, reservoirs and ponds, and 53% ofivers and streams in the U.S. are classified as impaired, needingmmediate attention (USEPA, 2013). Impairment of water bodiesue to elevated levels of nutrients and sediments originating frompland areas (i.e. watersheds) is a serious problem around theorld. High level of nutrients can cause problems such as toxic
lgal blooms, oxygen deficiency, fish kills, and loss of biodiversity.hese problems can also make the water unsuitable for drinking,
ndustrial, agricultural and recreational use (Carpenter et al., 1998).
Watershed management offers a strong basis for developingnd implementing effective management strategies (such as
riparian zones, vegetation strips, retention ponds, etc.) to protectwater resources (USEPA, 2003). Past efforts in reducing pollutantloads from watersheds have mainly focused on point sources andhave failed to adequately address the impact of nonpoint sources.If nonpoint sources of pollutants are not addressed, water bodiescan continue to be impaired (USEPA, 2003). However, nonpointsources of nutrients and sediments are difficult to identify andcontrol because they originate from spatially and temporallyvarying areas (Carpenter et al., 1998).
The level of sediment and nutrient contribution from differentparts of a watershed can vary substantially. Some typically smalland well defined areas contribute much of the sediment, P, andN into the watershed outflow (Walter et al., 2000; Pionke et al.,2000) and over relatively short periods (Dillon and Molot, 1997;Heathwaite et al., 2005). But in many situations source areas arenot well defined but diffused. Certain areas with a particular type of
soil, land use/cover and slope are more vulnerable than the others interms of nutrient and sediment loss. These areas are known as crit-ical source areas (CSAs). It is extremely important to identify thesesources of pollutants for cost-effective management practices.
dentifying nutrient and sediment loss prone areas in a watershednd concentrating management efforts in those areas have beenecommended in numerous studies (e.g., Nonpoint Source Taskorce, 1984; Tim et al., 1992; Zhou and Goa, 2008). Such areasan be identified through sub-watershed level water monitoring,imulation modeling, or both (Sharpley et al., 2002). Direct wateronitoring and field studies are usually costly and labor inten-
ive, and require a number of years of monitoring to sufficientlyccount for climatic fluctuations. The use of watershed models,uch as Soil and Water Assessment Tool (SWAT) (Arnold et al., 1998)nd Generalized Watershed Loading Function (GWLF) (Evans et al.,002), can avoid most limitations associated with field studies andan help in identifying and prioritizing sub-watersheds for cost-ffective implementation of management practices (Tripathi et al.,005; Ouyang et al., 2008; Georgas et al., 2009).
GWLF has widely been used for estimating streamflow, nitrogenN) and phosphorus (P) loadings (Swaney et al., 1996; Lee et al.,000), hydrochemistry (Schneiderman et al., 2002) and also forssessing changes in streamflow (Chang, 2003; Wu et al., 2007) andater quality (Tu, 2009) under different land use scenarios. GWLFas also been used for identification of CSAs at sub-watershed levelMarkel et al., 2006; Georgas et al., 2009). Similarly, SWAT has beensed around the world for predicting streamflow, and sediment andutrient loads from watersheds (Spruill et al., 2000; Kirsh et al.,002; Veith et al., 2005; Srivastava et al., 2006; Jha et al., 2007;iraula et al., 2012a,b). SWAT has also been used in several studies
or identification and prioritization of CSAs of sediments and nutri-nts (Tripathi et al., 2005; Ouyang et al., 2008; White et al., 2009;hebremichael et al., 2009; Panagopoulos et al., 2011; Shang et al.,012; Niraula et al., 2012a).
Wanger et al. (2007) studied the impact of alternative wateruality models, by comparing GWLF and SWAT, on pollutant load-
ng for Total Maximum Daily Loadings (TMDL) development. These of alternative water quality models resulted in differences inequired sediment reduction. The SWAT model load estimates wereonsistently larger than loads from GWLF. However, it is not clearf the use of different models would lead to different CSAs that areignificant enough for practical implementation of Best Manage-ent Practices (BMPs). In a related study, Niraula et al. (2012b)
ound that calibration of the SWAT model had very little effect onocations of nutrients and sediment CSAs. Therefore, the objectivef this study was to assess the effect of model choice on CSA loca-ions. Two models of different complexity, SWAT and GWLF, weresed for this purpose. Both models were utilized to identify sedi-ent and nutrients CSAs in the Saugahatchee Creek watershed in
ast central Alabama.
. Methodology
.1. Study area
The 570 km2 Saugahatchee Creek watershed (Fig. 1), selected forhis study, is a sub-watershed of the Lower Tallapoosa sub-basin inast central Alabama. The watershed, as determined using Nationaland Cover Data (NLCD, 2001), was comprised of 67.8% forest, 10.0%rassland, 11.7% agricultural land (hay/pasture and row crops) and.4% urban area (Fig. 1). Although most of the watershed lies inhe Piedmont physiographic province, a small portion lies in theoastal Plains. The Piedmont covers a transitional area betweenhe mostly mountainous Appalachians in the northeast and theelatively flat Coastal Plains in the southeast Alabama. While the
oils in the Piedmont are dominated by loam and sandy loam, soilsn remaining coastal plains are sandy loam based on the STATSGOoil database. Elevation in the watershed varies from 103 to 255 m.he study area is characterized by hot summers and mild winters
lling 268 (2013) 123– 133
with average temperatures of 26 ◦C and 7 ◦C, respectively. The longterm annual average rainfall in the watershed is 1336 mm. AlabamaDepartment of Environmental Management (ADEM) has identifiedtwo segments within the Saugahatchee Creek watershed as beingimpaired for nutrients and organic enrichment/dissolved oxygen(ADEM, 2008). The nutrient of concern in both of the tributaries isphosphorus. ADEM also recommended development of TMDLs foraddressing water quality problems in this watershed.
2.2. Watershed models
2.2.1. Soil and Water Assessment Tool (SWAT)The SWAT is a semi-distributed model that was primarily devel-
oped to predict the impact of land management practices on water,sediment, and agricultural chemical yields in large complex water-sheds over long periods of time (Neitsch et al., 2005). The modelinputs consist of topography, soil properties, land use/cover type,weather/climate data, and land management practices. The water-shed is sub-divided into sub-watersheds and each sub-watershedis further divided into hydrological response units (HRU) based ontopography, land use, and soil (Neitsch et al., 2005).
Surface runoff in each HRU was estimated using a modificationof the Soil Conservation Service Curve Number (SCS-CN) method(USDA, 1972). In the curve number method, daily precipitationis partitioned between surface runoff and initial and continuedabstractions as a function of antecedent soil moisture condition.The total sub-watershed discharge computed by SWAT includesrunoff from its HRUs and subsurface flow including lateral flowand return flow. Flow in SWAT is routed through channels usingeither Muskingum routing method or variable storage coefficientmethod (Neitsch et al., 2005). The latter was used in this study. Ero-sion and sediment yield from each HRU are estimated based on theModified Universal Soil Loss Equation (MUSLE) (Williams, 1975).Sediment is routed through channels using a modification of Bag-nold’s sediment transport equation (Bagnold, 1977). This equationestimates sediment transport capacity as a function of flow veloc-ity. The model either deposits or erodes sediment, depending on thesediment load entering the channel and the capacity of the flow.
SWAT models nitrogen and phosphorus cycles through five dif-ferent pools of nitrogen (two inorganic forms: NH4
+ and NO3−;
three organic forms: fresh, stable and active) and six different poolsof phosphorus (three inorganic forms: solution, active and sta-ble; three organic forms: fresh, stable and active) in soil (Neitschet al., 2005). Mineralization, decomposition, and immobilizationare important processes in both N and P cycles. Organic N andP transport with sediment is estimated using a loading functiondeveloped by McElroy et al. (1976) and later modified by Williamsand Hann (1978). Daily organic N and P runoff losses are calculatedby loading functions based on the concentrations of these elementsin the top soil layer, the sediment yield, and an enrichment ratio.Nitrate concentration in mobile water is calculated and multipliedwith mobile water volume to estimate total nitrate lost from thesoil layer. Mobile water is the sum of runoff, lateral flow and per-colation. The soluble P removed in runoff is estimated using the Pconcentration in the top soil layer, runoff volume and a P soil par-titioning coefficient. Further details can be found in Neitsch et al.(2005).
2.2.2. Generalized Watershed Loading Function (GWLF)The GWLF model is a combined distributed/lumped parameter,
continuous watershed model (Evans et al., 2002), which has theability to simulate runoff, sediment, and nutrient (N and P) loads
from various source areas, each of which is considered uniformwith respect to soil and cover. GWLF uses land use, soil, and dailyweather data for calculation of water balance. For estimation ofsediment and nutrient loads, monthly calculations are made based
R. Niraula et al. / Ecological Modelling 268 (2013) 123– 133 125
own
osmenlma1
E
Fig. 1. Saugahatchee Creek watershed in east central Alabama. Also sh
n the daily water balance aggregated to monthly values. It is con-idered a distributed model for surface loading because it allowsultiple land uses within an area, but for each area other param-
ters are considered to be uniform by the model. This model doesot consider spatial location of source areas, but simply adds the
oads from different source areas within each sub-watershed. Theodel works as a lumped parameter model using a water balance
pproach for modeling sub-surface loading (Haith and Shoemaker,987; Haith et al., 1992).
In GWLF, surface runoff is simulated using the SCS-CN method.rosion and sediment yield is modeled using the Universal Soil
are land use/cover and USGS flow and NOAA weather gaging stations.
Loss Equation (USLE). GWLF simulates soil erosion by consideringi) soil detachment by rainfall and ii) runoff transport relation-ships developed by Meyer and Wischeier (1969). A sedimentdelivery ratio, based on watershed size, and a transport capacity,based on average daily runoff, is then used to estimate sedimentyield from each source area. In USLE equation while soil erodi-bility factor (K) depends on soil properties, other factors such
as cover factor (C) and practice factor (P) depend on land usetype. The daily runoff volume which transport sediment is cal-culated based on CN which is also a function of soil and landuse/cover.
126 R. Niraula et al. / Ecological Modelling 268 (2013) 123– 133
Table 1Calibrated parameters of SWAT model with their default and calibrated values.
Parameter Descriptions Range Default value Calibrated value
CN2 Initial SCS Runoff curve number for moisture condition II 25–92 Variesa −5b
ESCO Soil evaporation compensation factor 0.01–1 0.95 0.2GWQMN Threshold depth of water in shallow aquifer required for the return flow to occur 0–5000 0 1200SPEXP Exponent parameter for calculating sediment reentrained in channel sediment routing 1–1.5 1 1.5PRF Peak rate adjustment factor for sediment routing in the main channel 0–2 1 1.97SURLAG Surface runoff lag time 1–24 4 1ADJ PKR Peak rate adjustment factor for sediment routing in the subbasins(tributary) 0.5–2 0.5 2PPERCO Phosphorus percolation coefficient 10–17.5 10 17.5PHOSKD Phosphorus soil partitioning coefficient 100–200 175 200P-UPDIS Phosphorus uptake distribution factor 0–100 20 80PSP Phosphorus sorption coefficient 0.01–0.7 0.4 0.7SOL LABP Initial soluble P concentration in surface soil layer (mg/kg) 0–100 0 11NPERCO Nitrogen percolation coefficient 0–1 0.2 1SOL NO Initial NO concentration in the soil(mg/kg) 0–100 0 21
acSsApfci(
2
ehDcpdwdwtsam1so(waApnrmasr
2
T
3 3
a Varies with land use and soil type.b Value reduced by 5.
Nutrient loads from rural areas are transported in runoff waternd eroded soil from numerous source areas. Dissolved loads arealculated by multiplying dissolved N and P coefficients with runoff.olid phase nutrient loads are calculated by multiplying averageediment nutrient concentrations with monthly sediment yield.ll N and P inputs from urban areas are assumed to be in solidhase. The model uses an exponential accumulation and wash-offunction for estimating urban loadings. Sub-surface losses are cal-ulated using dissolved N and P concentrations (which are modelnputs) considering a single lumped-parameter contributing areaEvans et al., 2002).
.3. Model inputs
ArcSWAT 2.1 (Winchell et al., 2008) and AVGWLF 7.1 (Evanst al., 2008) were used to set up and develop models for the Sauga-atchee Creek watershed. Data required in this study includedigital Elevation Model (DEM), soil properties, land use/cover,limate data such as precipitation, and minimum/maximum tem-erature, and point sources. A 10-m resolution DEM was used toelineate the watershed and sub-watershed boundaries, whichere used in both models. State Soil Geographic (STATSGO)atabase was used to derive soil parameters. Land use/cover dataere obtained from the National Land Cover Dataset (NLCD) for
he year 2001 (Fig. 1). Point source discharges from three pointources (North Auburn and Opelika waste water treatment plants,nd West Point Stevens) were included. Daily precipitation andinimum and maximum air temperature data between January
995 and December 2008 were collected from three NOAA weathertations (Fig. 1). Flow data for the period from 2000 to 2008 werebtained from a USGS gauging station located within the watershedFig. 1). Water quality data for the same location (USGS station)as obtained from Alabama Department of Environmental Man-
gement (ADEM, 2002) and Department of Fisheries and Alliedquaculture at Auburn University (DFAA, 2001). At least 3 dataoints per month for total suspended solids (TSS), nitrate (NO3),itrite (NO2), Ammonia (NH3 + NH4), total nitrogen (TN), solubleeactive phosphorus (SRP) and total phosphorus (TP) includingeasurements immediately after some storm events, were avail-
ble for the period 2000–2003. Organic P and N were calculated byubtracting the sum of mineral components from the TP and TN,espectively.
.4. Calibration and validation of models
Both SWAT and GWLF models were run from 1995 to 2008.he periods 2000–2004 and 2005–2008 were selected as the
calibration and validation periods, respectively, for flow. The firstfive years (1995–1999) were used as a warm up period to mini-mize uncertain initial conditions (e.g., soil moisture, groundwaterlevel, ground residue, nutrient pool, etc.). A manual calibrationtechnique was adopted. To identify the most sensitive parametersfor calibration, a manual one-at-a-time (OAT) relative sensitivityanalysis was carried out for the list of parameters synthetizedfrom review of the current literature. Models were first calibratedfor streamflow using data from the USGS gauge station (Fig. 1).For this purpose, observed streamflow was separated into surfacerunoff and baseflow components with a baseflow separation filterprogram (Arnold et al., 1995; Arnold and Allen, 1999). In additionto matching hydrographs of model simulated baseflow and excessrunoff with observed counterparts, quantitative measures (percentbias, correlation coefficient, and Nash-Sutcliffe efficiency) werealso used during calibration.
Once the models were calibrated for flow, they were subse-quently calibrated for sediment, TN and TP at the same USGSstation. Due to lack of sufficient water quality data, monthly sed-iment was calibrated for year 2000 and validated for year 2002.High quality data were available only for those years. TN and TPwere calibrated for the year 2000 (to be consistent with sedimentcalibration, which affects TN and TP) and validated for the period2001–2002. The Load Estimator (LOADEST) software (USGS, 2013)developed by USGS was used to estimate monthly loads from spon-taneous grab samples of sediment and nutrients, instantaneousstreamflow, and daily average streamflow. The degree of fit (R2)between the observed instantaneous flow data and TSS, TN and TPused for estimating loads were 0.77, 0.72 and 0.76, respectively.
Various hydrologic and water quality parameters(Tables 1 and 2) were changed within their ranges to get thebest fit with the observed data. Four evaluation criteria: percentbias (PBIAS; Moriasi et al., 2007), Nash–Sutcliffe efficiency (NSE;Moriasi et al., 2007), coefficient of determination (R2; Niraulaet al., 2012b) and line graphs were used to assess streamflow,sediment, TN and TP loads simulated by SWAT and GWLF. Itshould be noted that the calibrated model parameters are likelyaway from their true values because of the uncertainties comingfrom measurement errors in water quality data and the additionaluncertainty exerted by the LOADEST model.
2.5. Identification of critical source areas (CSAs)
CSAs were identified at the sub-watershed level. The sedimentand nutrient yields from each sub-watershed were analyzed basedon loads per unit area to identify the CSAs. Maps were created basedon these loadings to depict the CSAs separately for sediment, TN
R. Niraula et al. / Ecological Mode
Table 2Calibrated parameters of GWLF model with their default and calibrated values.
Parameter Default value Calibrated value
Curve number Variesa +5b
Recess coefficient 0.1 0.01Seepage coefficient 0 0.006Nitrogen in sediments (mg/kg) 3000 7500Phosphorus in sediments (mg/kg) 750 1500Sediment A factor 1.0997 × 10−4 1.0997 × 10−3
C factor: low/high intensity dev 0.02 0.4P factor: low/high intensity dev 0.02 0.6Nitrogen runoff coefficient
Low intensity dev (kg/ha/day) 0.012 0.02High density dev (kg/ha/day) 0.101 0.15
Phosphorus runoff coefficientCropland(mg/l) 0.079 0.6Forest(mg/l) 0.006 0.036Low intensity dev (kg/ha/day) 0.002 0.007High density dev (kg/ha/day) 0.011 0.015
N (mg/l) in groundwater 1 0.4P (mg/l) in groundwater 0.01 0.02
aeowMtcopoit
wfpiatNchib
I
wfi
Y
wRi0reon
a Varies with land use and soil type.b Value increased by 5.
nd TP using simulation results from both SWAT and GWLF mod-ls. In order to do this, sub-watersheds were ranked in descendingrder based on yields. The sub-watershed with the highest yieldas ranked first and the one with the lowest yield was ranked last.oving from the highest ranking to the lowest, sub-watersheds
hat collectively contribute to 20% of the sediment, TP or TN wereonsidered as CSAs. Different CSAs for sediment, TN and TP werebtained. The 20% threshold was chosen to provide results for com-arison across two models. Actual threshold is a function of costf implementing management practices. For instance, if budgets limited for implementation of management practices a lowerhreshold is needed.
A combined index was also defined to identify the sub-atersheds that can be considered as CSAs and need to be targeted
or overall reduction of sediment, TN and TP. Although TMDLrograms target reduction of individual stressors and, therefore
dentification of CSAs of individual stressors seems more reason-ble, sometimes a water body can be on the 303(d) list for morehan one pollutant. Since many BMPs can simultaneously reduce, P and TSS loads (e.g. vegetative filter strips), with a single BMPombined load reduction can be achieved. The combined index canelp identify those areas that are critical for multiple stressors. Hav-
ng BMPs in those areas will be more economical. This index is giveny
j =∑
(ωiYi,j) (1)
here, Ij is the combined index for sub-watershed j, Yi,j is an indexor sediment (i = 1), TN (i = 2) and TP (i = 3) for sub-watershed j, ands defined as
i,j = Ri,min − Ri,j
Ri,min − Ri,max(2)
here, Ri,j is the rank of watershed j for constituent i, and Ri,min andi,max are respectively the lowest and highest ranks for constituent
. Note that Ri,max = 1, Ri,min = total number of sub-watersheds and ≤ Yi,j ≤ 1. For our study watershed Ri,min and Ri,max are 106 and 1,
espectively. In Eq. (1), ωi is a subjectively chosen weight given toach Yi based on their importance, where
∑ωi = 1. For the purpose
f our study we assigned equal weights (1/3) to all three compo-ents (sediment, TN and TP). Theoretically Ij varies between 0 and
lling 268 (2013) 123– 133 127
1. If the same sub-watershed is ranked last in all three constituentsthen Ij = 0. Similarly, if the same sub-watershed is ranked highestin all three, then Ij = 1.
3. Results and discussion
3.1. Calibration/validation of SWAT and GWLF (monthly timescale)
Both SWAT and GWLF models were able to predict the monthlystreamflow with good accuracy (Fig. 2a and Table 3). Accordingto the performance statistics, SWAT and GWLF performed equallywell with respect to all three measures. GWLF had slightly betterNSE values and SWAT had marginally better PBIAS values, but forpractical purposes the differences are insignificant. Both modelswere able to capture the months with high- and low-flows (Fig. 2a).
Although both SWAT and GWLF models were able to predict themonthly sediment loadings with sufficient accuracy (Fig. 2b andTable 3), SWAT performed better than GWLF during both calibra-tion and validation periods on the basis of NSE and R2. However,SWAT and GWLF performed equally well based on PBIAS duringboth calibration and validation periods. SWAT overestimated sedi-ment by 0.9% and 3.7% during the calibration and validation periods,respectively, while GWLF underestimated the monthly sedimentloadings by 0.5% during the calibration period and overestimatedby 4.0% during the validation period.
Both models predicted the monthly TN loadings with sufficientaccuracy (Fig. 2c and Table 3). SWAT performed slightly better thanGWLF on the basis of NSE and R2 during both calibration and vali-dation periods. However, SWAT performed markedly better basedon PBIAS. SWAT underestimated TN loading by 2.4% during the cal-ibration period and underestimated by 0.3% during the validationperiod. GWLF, on the other hand, overestimated TN loading by 7.8%during the calibration period and underestimated by 9.0% duringthe validation period.
Even though both SWAT and GWLF models predicted monthlyTP loadings well (Fig. 2d and Table 3), SWAT performed better thanGWLF on the basis of performance statistics (NSE, R2 and PBIAS)during both calibration and validation periods. SWAT overesti-mated phosphorus loading by 7.0% during the calibration periodand underestimated by only 0.1% during the validation period.GWLF overestimated TP by 10.2% and underestimated by 7.8% dur-ing the calibration and validation periods, respectively.
3.2. Critical source areas (CSAs)
The calibrated SWAT and GWLF models were used to identifyand compare the CSAs in the Saugahatchee Creek watershed at thesub-watershed level. Different CSAs were identified with respectto sediment, TN and TP loadings because the factors driving eachare likely to be different, but not mutually exclusive. To eliminatethe effect of differences in areas of the sub-watersheds, average ofannual loadings per unit area (i.e. yield) for the 2000–2008 periodwere used to identify the CSAs.
3.2.1. CSAs of sedimentBoth models produced somewhat similar areas as CSAs of sedi-
ment (Fig. 3a and d). Based on 20% contribution, 5 sub-watersheds,covering 4.2% of the watershed, were identified as CSAs by GWLF.Similarly, 6 sub-watersheds, covering 4.6% of the watershed, wereidentified as CSAs by SWAT. Although the ranks are not exactly inthe same order, sub-watersheds 25, 53, 56, 69 and 73 were iden-
tified as CSAs by both SWAT and GWLF models. Sub-watershed 23was also identified as a CSA of sediment by SWAT. While averagesediment yield ranged from as high as 9.77 tons/ha/yr (t/ha/y) toas low as 0.06 t/ha/y based on the SWAT results (Fig. 3a), it ranged
128 R. Niraula et al. / Ecological Modelling 268 (2013) 123– 133
F b) sed
bTvst
3
eC7GawowT
TP
ig. 2. Observed, SWAT-simulated, and GWLF-simulated monthly (a) streamflow, (
etween 0.07 t/ha/y and 6.15 t/ha/y based on GWLF results (Fig. 3d).his showed that sediment yields obtained by SWAT showed largerariation than GWLF generated sediment yields. The average basinediment yield was found to be 0.94 t/ha/y with GWLF comparedo 1.44 t/ha/y with SWAT.
.2.2. CSAs of TPIn spite of the fact that most of the CSAs of TP identified by
ach model overlap, several sub-watersheds were identified asSAs by one model and not by the other. While SWAT identified
sub-watersheds (5.8% of total watershed area) as CSAs for TP,WLF identified 11 sub-watersheds (9.6% of total watershed area)s CSAs. Although the ranks are not exactly in the same order, sub-
atersheds 25, 51, 53, 56, 59, 69, and 73 were identified as CSAs
f TP by both the SWAT and the GWLF models. Most of these sub-atersheds were also identified as sediment CSAs, indicating that
P load is highly correlated with sediment load. Sub-watersheds
able 3erformance measures of the SWAT and the GWLF during calibration and validation peri
Parameter Model Calibration
NSE R2
Streamflow SWAT 0.90 0.90
GWLF 0.91 0.92
Sediment SWAT 0.72 0.72
GWLF 0.68 0.68
TN SWAT 0.75 0.81
GWLF 0.70 0.78
TP SWAT 0.85 0.89
GWLF 0.65 0.79
iment load, (c) TN load, and (d) TP load, for the calibration and validation periods.
64, 79, 88 and 97 were also identified as CSAs of TP by GWLF, butnot by SWAT. TP yields ranged from 0.02 kg/ha/y to 0.87 kg/ha/yaccording to SWAT results (Fig. 3b), and between 0.06 kg/ha/y and0.71 kg/ha/y according to GWLF results (Fig. 3e). This again showedthat SWAT generated outputs showed slightly wider range thanGWLF generated outputs. The basin-average TP yield was found tobe 0.22 kg/ha/y with GWLF, which is slightly higher than the SWATgenerated average of 0.19 kg/ha/y. Although the average TP loadfrom SWAT was slightly smaller than the GWLF average, the factthat highest TP load from all the sub-watersheds was obtained withSWAT points to the strong correlation between P and sediment.
3.2.3. CSAs of TN
In the case of TN, while 10 sub-watersheds (10.5% of total water-
shed area) were identified as CSAs by SWAT, 13 sub-watersheds(13% of total watershed area) were considered as CSAs by GWLF.Both models identified sub-watersheds 4, 56, 69, 79, 88, 97 and
ods for flow, sediment, TN and TP.
Validation
PBIAS (%) NSE R2 PBIAS (%)
+3.1 0.74 0.78 −1.4−6.7 0.79 0.82 +3.5
+0.9 0.86 0.87 +3.7−0.5 0.78 0.81 +4.0
−2.4 0.90 0.92 −0.3+7.8 0.87 0.91 −9.0
+7.0 0.88 0.91 −0.1+10.2 0.87 0.88 −7.8
R. Niraula et al. / Ecological Modelling 268 (2013) 123– 133 129
Fig. 3. Sediment, TP and TN yields from each sub-watershed as estimated by SWAT (a, b, c, respectively) and GWLF (d, e, f, respectively). Numbered sub-watersheds arei
1C1n50ST2aiwaww
dentified as CSAs.
03 as CSAs. Sub-watersheds 25, 73 and 102 were also identified asSAs based on SWAT, but not by GWLF. Similarly, sub-watersheds, 11, 14, 53, 59 and 85 were identified as CSAs by GWLF, butot by SWAT. While TN yields ranged between 0.57 kg/ha/y and.31 kg/ha/y based on SWAT results (Fig. 3c), it ranged between.85 kg/ha/y and 4.73 kg/ha/y according to GWLF (Fig. 3f). Again,WAT had slightly more variation in TN yields. The average basinN yield was found to be 1.93 kg/ha/y with GWLF compared to.1 kg/ha/y with SWAT. The differences in values obtained by SWATnd GWLF can be related to the way NO3 pool is handled. NO3s mostly transported through subsurface flow in watersheds as
as the case with SWAT in this study too. However, with GWLF, dissolved N coefficient is simply used to simulate subsurface N,hich might not be sufficient to capture the NO3-N loads from theatershed.
3.3. Combined CSA Index
Combined CSA index Ij was determined for each sub-watershedj by assigning the same weights to sediment, TN, and TP (ωi = 1/3in Eq. (2)). Sub-watersheds with Ij ≥ 0.9 were subjectively identi-fied as CSAs collectively for all parameters (sediment, TN and TP).Again, the purpose here was to present the idea. The threshold 0.9was chosen so that we deal with a manageable number of CSAs.A map (Fig. 4) was produced to show these CSAs based on resultsfrom both SWAT and GWLF models to visualize the differences.Sub-watersheds 25, 53, 56, 59, 69 and 73 had Ij ≥ 0.9 based on both
models and thus were identified as CSAs by both models. How-ever, sub-watersheds 23 and 51 had Ij ≥ 0.9 only based on SWATresults and thus were also identified as CSAs based on SWAT. Like-wise, sub-watershed 88 had Ij ≥ 0.9 only based on GWLF results
130 R. Niraula et al. / Ecological Modelling 268 (2013) 123– 133
F b-was
ab2iaT
atlstaaACp
3
wsoICtMwsalcmehmtmS(iiwg
ig. 4. Critical source areas based on the combined index (Eq. (1)). The numbered suediment, TP and TN loads.
nd thus was identified as CSA by the GWLF model. CSAs identifiedy SWAT covered only 6.5% of the watershed area and contributed6.5% of the sediment, 23.1% of TP and 13.9% of TN loadings. Sim-
larly, CSAs determined by GWLF covered 5.6% of the watershednd were responsible for contributing 23.1% of sediment, 16.5% ofP and 12.7% of TN.
While sub-watersheds 23, 25, 51, 53, 56, 59, 69 and 73 werell dominated by urban land use with considerable amount of pas-ure land, sub-watershed 88 was comprised of some cropland anday predominantly on the coastal plain soils. It was clear from thistudy that, sub-watersheds dominated by urban area were amonghose producing highest amount of sediment and nutrient loadsnd thus were identified as CSAs. Also sub-watersheds with somemount of agricultural crops were identified as CSAs of TP and TN.
few hay/pasture dominated sub-watersheds were identified asSAs of TN (e.g. sub-watershed 4, which is 46% hay/pasture). Table 4rovides the land use composition of the combined CSAs only.
.4. Differences in model performances and identified CSAs
As mentioned earlier, both SWAT and GWLF performed equallyell for streamflow. This could be because both models use the
ame method (SCS-CN) for estimating runoff. However, in the casef sediment, TN and TP, SWAT performed slightly better than GWLF.n the case of sediment, there was also a difference in one of theSAs identified by the two models. While GWLF uses the conven-ional USLE method for estimating soil erosion, SWAT uses the
USLE equation. USLE uses rainfall intensity as the erosive energy,hereas MUSLE uses runoff volume and peak flow rate to simulate
ediment erosion and yield. This improves the prediction accuracynd also eliminates the need for delivery ratio. In GWLF, nutrientoads from rural areas are calculated based on dissolved N and Poefficients in surface runoff, and a sediment coefficient in sedi-ent load. Nutrient loads form urban areas are estimated through
xponential accumulation and wash-off function. On the otherand, SWAT models N and P cycles comprehensively. Processes likeineralization, decomposition, and immobilization are allowed to
ake place in the soil in each HRU. Thus, SWAT provides a moreechanistic and process-based approach than GWLF. As a result,
WAT predicted sediment, TN and TP loads better than GWLF didsee Table 3). Further, because SWAT and GWLF conceptualizes sed-
ment, TN and TP processes differently, there were some variationsn the locations of identified CSAs. Also, since SWAT divides the sub-
atersheds into smaller computational units, i.e. HRUs, there is areater chance that it can distinguish sub-watersheds with higher
tersheds should be first targeted for management practices for overall reduction of
loadings than the others. The wide range of output values fromdifferent sub-watersheds for sediment, TN and TP obtained withSWAT further support this.
3.5. Implications of not choosing the right model
Models are simplified mathematical representations of realwatershed systems, and thus are never perfect (Gupta et al., 1998).Model structural adequacy can vary with “engineering” view point(focused on functional adequacy and usually take a decision mak-ing perspective); “physical science” viewpoint (consistency withthe physical system) and “system science” viewpoint (data-basedhybrid of the first two which stresses on both consistency andprinciple of parsimony) (Gupta et al., 2012). Thus, choice of theright model depends on number of factors. Water quality mod-els are typically chosen for a specific application based on themodel’s capability to simulate the dominant processes and desiredoutput, professional and personal preference, software availabil-ity, prior experience in using the model, and computational costthat include both time and financial resources (Wanger et al.,2007).
Results from this study showed that not choosing the rightmodel may have important implications. For instance, GWLF iden-tified 4 extra sub-watersheds as CSAs for TP compared to SWAT.They had 7 sub-watersheds in common as CSAs. So, if GWLF isused as the base in deciding where to implement BMPs, about65% more area should be targeted compared to SWAT. This mighthave extremely important economic implications. The total areastargeted for BMPs implementation in conservation programs aregenerally the limiting factor due to limited funding. The differ-ences in locations of CSAs were more evident with TN. Out ofthe 13 CSAs identified by GWLF only 7 were also recognized asCSA by the SWAT model. That means GWLF has 6 sub-watershedsnot identified as CSA by SWAT. Note that we are not promot-ing one model over another, rather pointing out the differencesin CSAs due to the use of two different models. Since no mea-sured data were available at sub-watershed level (we had 106sub-watersheds), we cannot establish which model was identify-ing the locations of CSAs more properly, which is a limitation of thisstudy.
Although we did not have the water quality data to verify the
CSAs identified by the models, some unpublished water quality dataexists from several first and second order creeks around the Cityof Auburn (Fig. 1) to support most of the land use source areasidentified by the models. Table 5 summarizes these water quality
R. Niraula et al. / Ecological Modelling 268 (2013) 123– 133 131
Table 4Land use composition of combined CSA’s in the Saugahatchee Creek watershed.
Table 5Mean (±SE) total suspended solid (TSS), TN, and TP concentrations (mg L−1) for urban and reference creeks in the Saugahatchee Creek basin sampled during 2009–2010(unpublished data).
ata and generally supports that urban areas represented CSAs ofutrients and sediment into the Saugahatchee Creek. Water quality
rom urban areas around the City of Auburn showed mixed resultsor TSS but consistently higher concentrations of TN and TP com-ared to reference streams (streams draining forested watersheds).urprisingly, some of the highest TSS concentrations are found inhe reference streams but only during the rising limb of storm-ow. During the baseflow and falling limb, the average urban creekad 3–5 times greater TSS concentration than the reference creeks.utrient concentrations (N and P) were consistently higher in urbanreeks compared to the reference creeks (often near twice the con-entration) for baseflow, falling limb, and rising limb. These dataupported indications provided by the two models that urban landenerated higher levels of nutrients and possibly sediment.
. Summary and conclusions
Two watershed models differing in their conceptual structuresnd complexities, SWAT and GWLF, were set up, calibrated, and
validated in an east-central Alabama watershed. The models werethen utilized to identify critical source areas (CSAs) of sediment, TNand TP for implementation of cost effective management practicesin the watershed.
Based on the overall model performance statistics, it wasobserved that SWAT performed slightly better than GWLF.Although performance of GWLF was similar to SWAT for stream-flow during both calibration and validation periods, SWATperformed better for sediment, TN, and TP. The calibrated and vali-dated models were used to identify the CSAs in the study watershedat sub-watershed level based on loadings per unit area (yield). Ingeneral, sub-watersheds dominated by urban areas were amongthose producing the highest amount of sediment, TN and TP yields,and thus were identified as CSAs. However, sub-watersheds withsome amount of agricultural crops were also identified as CSAs
of TP and TN. A few hay/pasture dominated sub-watersheds wereidentified as CSAs of TN.
This study revealed that CSAs can vary based on the parametersof interest (sediment, TN or TP). Based on a combined index, while 8
1 Mode
sw6stsSfidaGcsesctolmv
stiw
R
A
A
A
A
A
B
C
C
D
D
E
E
G
G
G
G
H
32 R. Niraula et al. / Ecological
ub-watersheds were identified as CSAs by the SWAT model, 7 sub-atersheds were identified as CSAs by the GWLF, among which
sub-watersheds were identical. Therefore, although there wereimilarities in many of the identified CSAs, not all the CSAs iden-ified by the models were mutual. GWLF did not recognize twoub-watersheds, which were identified as CSAs by the SWAT model.imilarly, SWAT did not detect one of the sub-watersheds identi-ed as CSA by GWLF. Dissimilarities in CSAs are attributed to theifferences in model conceptualizations implemented in the SWATnd GWLF models. Although SWAT performed slightly better thanWLF for predicting flow, sediment, TN, and TP, we cannot con-lude that SWAT was better for identifying and locating the CSAs,ince there was no measured data to verify the CSAs identified byach model. Unpublished water quality data from several first andeconds order tributaries generally agreed with the modeled indi-ations of CSAs in the Saugahatchee Creek watershed. We foundhat urban creeks in the basin normally had higher concentrationsf sediment, N, and P although considerable variability (particu-arly regarding sediment) was noted. Targeted and more intensive
onitoring of creek sediment may shed light on the dynamics andariability found among creeks with different land uses.
This study was conducted in a forest dominated watershed withignificant urban portion concentrated at the upstream headwa-ers. It would be interesting to carry out similar studies in futuren intensely cultivated watersheds and/or more heterogeneous
atersheds representative of the majority of NPS pollution sources.
eferences
rnold, J.G., Allen, P.M., Muttiah, R., Bernhardt, G., 1995. Automated base flow sep-aration and recession analysis techniques. Ground Water 33 (6), 1010–1018.
rnold, J.G., Srinivasan, R., Muttiah, R.S., Williams, J.R., 1998. Large area hydrologicmodeling and assessment, Part I: model development. J. Am. Water Resour.Assoc. 34 (1), 73–89.
rnold, J.G., Allen, P.M., 1999. Automated methods for estimating baseflow andgroundwater recharge from streamflow records. J. Am. Water Resour. Assoc.35 (2), 411–424.
DEM, 2002. Surface water Quality Screening Assessment of the Tallapoosa RiverBasin-2000. Aquatic Assessment Unit, ADEM. September 2002.
DEM, 2008. Final Saugahatchee Creek Watershed Total Maximum Daily Load(TMDL) – Nutrients & OE/DO. Alabama Department of Environmental Manage-ment, Water Quality Branch, EPA Region 4.
agnold, R.A., 1977. Bedload transport in natural rivers. Water Resour. Res. 13,303–312.
arpenter, S.R., Caraco, N.F., Correll, D.L., Howarth, R.W., Sharpley, A.N., Smith, V.H.,1998. Nonpoint pollution of surface waters with phosphorus and nitrogen. Ecol.Appl. 8 (3), 559–568.
hang, H., 2003. Basin hydrological response to changes in climate and landuse: theConstoga River Basin, Pennsylvania. Phys. Geogr. 24 (3), 222–247.
FAA, 2001. Water quality, loading and bio-assessment of Saughatchee Creek,2000–2001. Department of Fisheries and Allied Aquaculture, Auburn University,December 2001.
illon, P.J., Molot, L.A., 1997. Effect of landscape form on export of dissolved organiccarbon, iron, and phosphorus from forested stream catchments. Water Resour.Res. 33 (11), 2591–2600.
vans, B.M., Lehning, D.W., Corradini, K.J., Petersen, G.W., Nizeyimana, E., Ham-lett, J.M., Robillard, P.D., Day, R.L., 2002. A comprehensive GIS-based modelingapproach for predicting nutrient loads in watersheds. J. Spat. Hydrol. 2 (2),1–18.
vans, B.M., Lehning, D.W., Corradini, K.J., 2008. AVGWLF Version 7.1. User’s Guide.Penn State Institutes of Energy and the Environment, The Pennsylvania StateUniversity, University Park, PA.
hebremichael, L.T., Watzin, M.T., Veith, T.L., 2009. SWAT modeling of critical sourceareas of runoff and phosphorus loss: Lake Champlain Basin, VT. In: 2009 Interna-tional SWAT Conference Proceedings, Texas Water Resources Institute TechnicalReport No. 356, pp. 385–393.
eorgas, N., Rangarajan, S., Farley, K.J., Jugupilla, S.R.K., 2009. AVGWLF-based esti-mation of the nonpoint sources nitrogen loads generated within long islandsound sub-watersheds. J. Am. Water Resour. Assoc. 45 (3), 715–733.
upta, G.V., Clack, M.P., Vrugrt, J.A., Abramowitz, G., Ye, M., 2012. Towards a com-prehensive assessment of model structural accuracy. Water Resour. Res. 48,W08301, http://dx.doi.org/10.1029/2011WR011044.
Haith, D.R., Mande, l, R., Wu, R.S., 1992. GWLF: Generalized Watershed LoadingFunctions User’s Manual, Version 2.0. Cornell University, Ithaca, NY.
Heathwaite, A.L., Quinn, P.F., Hewett, C.J.M., 2005. Modeling and managing criticalsource areas of diffuse pollution from agricultural land using flow connectivitysimulation. J. Hydrol. 304 (1–4), 446–461.
Jha, M.K., Gassman, P.W., Arnold, J.G., 2007. Water quality modeling for the RaccoonRiver watershed using SWAT. Trans. ASAE 50 (2), 479–493.
Kirsh, J., Kirsh, A., Arnold, J.G., 2002. Predicting sediment and phosphorus loads inthe Rock River basin using SWAT. Trans. ASAE 45 (6), 1757–1769.
Lee, K.Y., Fisher, T.R., Jordan, T.E., Correll, D.L., Weller, D.E., 2000. Modeling the hydro-chemistry of the Choptank River basin using GWLF and GIS. Biogeochemistry 49,143–173.
Markel, D., Somma, F., Evans, B.M., Tyson, J., 2006. Using a GIS transfer model toevaluate pollutant loads in Lake Kinneret watershed, Israel. Water Sci. Technol.53 (10), 75–82.
McElroy, A.D., Chui, S.Y., Nebgen, J.W., Aleti, A., Bennet, F.W., 1976. Loading Functionsfor Assessment of Water Pollution from Non Point Sources. EPA Document EPA600/2-76-151. USEPA, Athens, USA.
Meyer, L.D., Wischeier, W.H., 1969. Mathematical simulation of the process of soilerosion by water. Trans. ASAE 12, 754–762.
Moriasi, D.N., Arnold, J.G., Van Liew, M.W., Bingner, R.L., Harmel, R.D., Veith, T.L.,2007. Model evaluation guidelines for systematic quantification of accuracy inwatershed simulations. Trans. ASABE 50 (3), 885–900.
Niraula, R., Norman, L.M., Meixner, T., Callegary, J., 2012a. Multi-gaugecalibration for modeling the Semi-Arid Santa Cruz watershed inArizona–Mexico border area using SWAT. Air Soil Water Res. 2012, 41–57,http://dx.doi.org/10.4137/ASWR.S9410.
Niraula, R., Kalin, L., Wang, R., Srivastava, P., 2012b. Determining nutrient and sedi-ment critical source areas with SWAT model: effect of lumped calibration. Trans.ASABE 55 (1), 137–147.
NLCD, 2001. National Land Cover Data 2001. Downloaded from alabamaview.org onSeptember 2008.
Ouyang, W., Hao, F.H., Wang, X.L., 2008. Regional point sources organic pollutionmodeling and critical areas identification for watershed best environmentalmanagement. Water Air Soil Pollut. 187, 251–261.
Panagopoulos, Y., Makropoulos, C., Baltas, E., Mimkou, M., 2011. SWAT parameteri-zation for the identification of critical diffuse pollution source areas under datalimitations. Ecol. Model. 222 (19), 3500–3512.
Pionke, H.B., Gburek, W.J., Sharpley, A.N., 2000. Critical source area controls on waterquality in an agricultural watershed located in the Chesapeake basin. Ecol. Eng.14 (4), 325–335.
Schneiderman, E.M., Pierson, D.C., Lounsbury, D.G., Zion, M.S., 2002. Modeling thehydro-chemistry of the Cannonsville watershed with Generalized WatershedLoading Functions (GWLF). J. Am. Water Resour. Assoc. 38 (5), 1323–1347.
Sharpley, A.N., Kleinman, P.J.A., McDowell, R.W., Gitau, M., Bryant, R.B., 2002.Modeling phosphorus transport in agricultural watersheds: processes and pos-sibilities. J. Soil Water Conserv. 57 (6), 425–439.
Spruill, C.A., Workman, S.R., Taraba, J.L., 2000. Simulation of daily and monthlystream discharge from small watersheds using the SWAT model. Trans. ASAE43 (6), 1431–1439.
Srivastava, P., McNair, J.N., Johnson, T.E., 2006. Comparison of process-based andArtificial Neural Network approaches for streamflow modeling in an agriculturalwatershed. J. Am. Water Resour. Assoc. 42 (3), 545–563.
Swaney, D.P., Sherman, D., Howarth, R.W., 1996. Modeling water, sediment andorganic carbon discharges in the Hudson–Mohawk basin. Estuaries 19 (4),833–847.
Shang, X., Wang, X., Zhang, D., Chen, W., Chen, X., Kong, H., 2012. An improvedSWAT-based computational freamework for indtifying critical source areas foragriclutiurla pollution of the lake basin scale. Ecol. Model. 226, 1–10.
Tim, U.S., Mostaghimi, S., Shanholtz, V.O., 1992. Identification of critical nonpointsource areas using Geographic Information System and water quality modeling.Water Resour. Bull. 28 (5), 877–887.
Tripathi, M.P., Panda, R.K., Raghuwanshi, N.S., 2005. Development of effective man-agement plan for critical sub-watersheds using SWAT model. Hydrol. Process.19, 809–826.
Tu, J., 2009. Combined impact of climate and land use changes on stream-flow and water quality in eastern Massachusetts, USA. J. Hydrol. 379 (3–4),268–283.
USDA Soil Conservation Service, 1972. National Engineering Handbook. U.S. Gov-ernment Printing Office, Washington, DC, Hydrology Section 4 (chapters 4–10).
USEPA, 2003. National management measures for the control of non-point pollution from agriculture. EPA-841-B-03-004. Available at:http://water.epa.gov/polwaste/nps/agriculture/agmm index.cfm
USEPA, 2013, National summary of water quality assessments of each waterbodytype in US. Available at: http://ofmpub.epa.gov/waters10/attains nation cy.control#prob surv states
USGS, 2013. Load Estimator (LOADEST): A Program for Estimating Constituent Loadsin Streams and Rivers. http://water.usgs.gov/software/loadest
Veith, T.L., Sharpley, A.N., Weld, J.L., Gburek, W.J., 2005. Comparison of measuredand stimulated phosphorus with indexed site vulnerability. Trans. ASAE 48 (2),557–565.
alter, M.T., Walter, M.F., Brooks, E.S., Steenhuis, T.S., Boll, J., Weiler, K., 2000. Hydro-logically sensitive areas: variable source area hydrology implications for waterquality risk assessment. J. Soil Water Conserv. 55 (3), 277–284.
anger, R.C., Dillaha, T.A., Yagow, G., 2007. An assessment of the reference water-shed approach for TMDL with biological impairments. Water Air Soil Pollut. 181,341–354.
hite, M.J., Storm, D.E., Busteed, P.R., Stoodley, S.H., Phillips, S.J., 2009. Evaluatingnonpoint sources critical areas contributions at the watershed level. J. Environ.Qual. 38, 1654–1663.
illiams, J.R., 1975. Sediment routing for agricultural watersheds. Water Resour.Bull. 11 (5), 965–974.
lling 268 (2013) 123– 133 133
Williams, J.R., Hann, R.W., 1978. Optimal Operation of Large Agricultural Watershedswith Water Quality Constraints. Technical Report No. 96. Texas Water ResourcesInstitute, Texas A&M University.
Winchell, M., Srinivasan, R., Di Luzio, M., Arnold, J.G., 2005. 2008. ArcSWAT 2.1 inter-face for SWAT. User’s Guide. Grassland, Soil and Water Research Laboratory.USDA Agricultural Research Sevice Temple, Texas.
Wu, W., Hall, C.A.S., Scetena, F.N., 2007. Modeling the impact of recent landusechanges on the stream flows in northeastern Puerto Rico. Hydrol. Process. 21,2944–2956.
Zhou, H.P., Goa, C., 2008. Identifying critical source areas for non-point phosphorusloss in Chaohu watershed. Huan Jing Ke Xue 29 (10), 2696–2702.
