A comprehensive evaluation of input data-induced uncertainty in nonpoint source pollution modeling

Introduction Conclusions References Tables Figures


6
More information about the study area are referred to Shen et al., (2012aShen et al., ( , 2013a. 115 Based on the characteristics of the river system, the studied watershed was broken 116 into six drainage regions: Dongxi river, Xixi river, Baiyang river, upper region of the 117 Wuxi hydrological gauge, Houxi river, and upper region of the county boundary 118 (watershed outlet). As illustrated in Fig. 1, the corresponding outlets of are referred to 119 as DX, XX, BY, WX, HX, and CF, respectively. In this study, TP was evaluated as P 120 was recognized as the key limiting factor of eutrophication in this region. validation. The measured water quality and flow data were obtained from the 7 2.3.1 Spatial data 1: Rainfall data 143 In this study, rainfall datasets were collected from twelve rain gauges located within 144 the watershed boundary and two outside stations that were within approximately 10 145 km of the watershed boundary were also used (Fig. 1). The rain gauge falling within a 146 given sub-catchment is identified using the GIS software. The annual mean rainfall 147 recorded by these rain gauges is listed in Table 1. Previous studies have demonstrated 148 rainfall uncertainty comes from the lack of representative rain gauges and then the 149 need to interpolate the rainfall data between rain gauges (André assian et al., 2001;150 McMillan et al., 2011). Our previous study (Shen et al., 2012a) has already focused on 151 the impact of interpolation methods on the spatial rainfall heterogeneity so we focused 152 on the representativeness of rainfall stations. In this sense, rainfall data-induced 153 uncertainty was analyzed in two steps: 1) the dataset of each rain gauge was used as 154 inputs for the SWAT model separately, and the model performances were ranked 155 based on the Nash-Sutcliffe efficiency coefficient (ENS) values for single gauge 156 simulations; 2) random combinations of m rain gauges (m ranged from 2 to 12) were 157 generated and used as SWAT inputs. The expected rainfall spatial distributions were 158 only generated by the centroid method was selected because it was the current 159 approach incorporated into the current version of SWAT model and the easiest to  As discussed above, land use data available for the modeling effort will likely come 174 from numerous sources; therefore, an assessment of available land use data and the 175 time period covered by these data should be made. In this study, land use data were 176 obtained from the 1980s (1980-1989), 1995, 2000, and 2007. Specifically, maps from  Table 2. Second, these four land use maps were used as model inputs 183 and their impacts were estimated respectively using the calibrated SWAT model. In 184 our previous study (Shen et al., 2013a), the resolution of land use data was shown to 185 have only a slight influence on simulated NPS-P for the study region; therefore, the 186 land use map was not resampled in this study.

216
This study focused on error-propagation from input data to NPS-TP predictions (the 217 sum of organic P and mineral P) at the WX for the period from 2000 to 2007. First, 218 the sensitivity of simulated TP to each input data was quantified in the form of 219 summary statistics, such as the SD and the coefficient of variation (CV). Specifically, 220 the CV, which is a normalized measure of dispersion of a probability distribution, is 221 defined as a dimensionless number by quantifying the ratio of the SD to the MV. 222 Compared to SD, the CV is more appropriate for comparing different data sets; 223 therefore, it was used as the main approach for expressing uncertainty in this study. of simulated data.

229
The ENS was used as the goodness-of-fit indicator to evaluate the model x is the simulated and measured data for the ith pair, 233 respectively, mea x represents the mean value of the measured values, and n is the 234 total number of paired values.

235
In this study, model structure was fixed and model uncertainty will stems 236 predominantly from input errors. Based on the performance ratings by Moriasi et al.

237
(2007), 0.5 was judged as a reasonable ENS value for TP simulation so a threshold of 238 ENS≥ 0.5 was defined to select acceptable SWAT runs (Liu and Gupta, 2007). In the 239 next step, behavior input data (ENS≥0.5), which refer to the phenomenon of 240 equifinality and can be representative of a watershed system (ENS≥0.5), were grouped 241 to express the prediction uncertainty by using a multi-input ensemble method. Finally, 242 input-induced model uncertainty was generated via sampling from the output 243 distributions that are generated from these effective input datasets.

Calibration and validation 246
As shown in  (2007), the accuracy of flow prediction could be judged as 252 very good, while the sediment and TP simulations were judged to be satisfactory.    there might not be many dense rain gauge networks similar to those used for this 329 study; therefore, the fact that spatial rainfall variation is a function of key gauges 330 rather than all gauges would indicate a wider range of applicability. For this study 331 area (2,421 km 2 ), the optimal number of gauges were identified as 6 beyond which 332 improvements to the model predictions would not be found.

333
As illustrated in Fig. 2c and 2d, the second highest uncertainty was caused by DEMs, 334 and the ASTER GDEM-induced uncertainty was higher than by uncertainty induced 335 by NFGIS DEM. These higher values could be due to the following two reasons: first,

336
NFGIS DEM was already validated in many places in China, which was not the case 337 for ASTER GDEM (Wu et al., 2007;Dixon and Earls, 2009 to the nonlinearity of erosion processes and its subsequent effect on P processes ground cover. However, these low values in our study could be due to minor land use 365 changes during the period from the 1980s to 2007. As shown in Table 2, the fraction 366 of forest area decreased gradually from 61.75% to 54.76%, whereas agricultural land 367 increased from 25.68% to 33.47%. Fig. 2f indicates that the fertilizer input has only a 15 slight impact on in-stream TP loads. This was because P application was low in this 369 watershed with the inorganic N being applied in greater amounts and more widely.

370
Additionally, the major forms of P in mineral soils are plant-available soluble P, 371 insoluble forms of mineral P and organic P. According to the mechanism of the 372 SWAT model, P would be taken up firstly by plant uptake and then by erosion, and 373 these processes would govern the turnover rates and transport of P (Arnold et al.,

374
1998). Therefore, only a small proportion of P will finally flow into the water body as

381
As shown in Fig. 3, this demonstrated that input-induced uncertainty may be highly     Table 4 The sensitivity of simulated TP (CV values) to different input dataset