As shown in Table 1, the total runoff predicted by AnnAGNPS compared well with the observed data (RE = 0.01) during the calibration period (1998–1999).

In addition, monthly runoff predicted by AnnAGNPS also compared well with the observed data (NSE = 0.94), and the regression of the monthly predicted runoff with the observed runoff on the line of equal values was satisfactory, with an R-square value of 0.94 (p < 0.05) and a slope of 1.04. The mean and standard deviation of simulated runoff were close to the corresponding observed runoff as shown in Table 1. Thus, the model can predict monthly runoff well after calibration (Figure 1).

Comparisons of predicted and observed total runoff during the validation period (2003–2005) produced a RE of −0.06; and comparisons of predicted and observed monthly runoff during validation period produced a NSE of 0.93 and an R-square value of 0.93 (p < 0.05), which again demonstrates a satisfactory fit with field data (Figure 2).

Studies performed by Licciardello et al. [22] in a Mediterranean watershed produced a NSE of 0.76 and RE of −0.4 for runoff comparison without calibration; after calibration, the NSE was 0.84 and RE was zero, and the NSE was 0.83 and RE was 0.32 during validation period. A NSE of 0.89 and RE of −0.16 for runoff were reported by Zema et al. [23] on the Ganspoel watershed (Belgium) without calibration. A NSE of 0.82 was achieved by Baginska et al. [24] on the Currency Creek experimental catchment of the Sydney Region (Australia) after calibration. Based on a thorough review of model evaluation methods done by Moriasi et al. [25], model simulation can be judged as satisfactory if NSE is greater than 0.50 and very good if NSE is greater than 0.75 for runoff. Thus, the overall model performance was good. Generally, the monthly runoff events were slightly underpredicted by AnnAGNPS, although a few monthly runoffs were over-predicted (Figure 2). As pointed out by Yuan et al. [12], and also demonstrated by Baginska et al. [24], Licciardello et al. [22] and Yuan et al. [14], AnnAGNPS tended to under-predict runoff due to the assumption of a triangular hydrograph and some approximations of model parameters.

Annual average sediment yields monitored at the Wuxi station during 2000–2002 and 2003–2004 were 1.44 × 10⁶ t·y⁻¹ and 1.09 × 10⁶ t·y⁻¹, respectively [26]. Therefore, average sediment transport per square kilometer was calculated as 722 t·km⁻²·y⁻¹ for the period of 2000-2002 and 546 t·km⁻²·y⁻¹ for the period of 2003–2004, with the contribution area of 2000 km² above the Wuxi station. Annual average sediment yields predicted by AnnAGNPS were 0.95 × 10⁶ t·y⁻¹ for the period of 2000-2002 and 1.29 × 10⁶ t·y⁻¹ for the period of 2003-2004 with the sediment transport rate of 477 t·km⁻²·y⁻¹ and 645 t·km⁻²·y⁻¹respectively, which are approximately 34% lower than the observed value during 2000–2002 and 18% higher than the observed value during 2003–2004.

The use of RUSLE technology and the parameters associated with determining soil loss in AnnAGNPS are meant to be used as long term estimates. Yuan et al. [12] showed that AnnAGNPS adequately predicted average annual sediment load without calibration. Studies performed by Licciardello et al. [22] had a NSE of 0.51 and RE of 0.53 for sediment yield at event scale without calibration. A RE of 0.59 was achieved by Shrestha et al. [19] on a watershed in the Siwalik Hills of Nepal. Although no detailed observed sediment data were available for comparisons of the sediment simulation in this study, the RE of −0.34 and 0.18 without calibration indicates a satisfactory performance of AnnAGNPS application in simulation sediment.

Based on terrain, soil, land use, and crop management information and eight-year climatic data from eleven stations in the watershed, results predicted by AnnAGNPS were used in evaluating characteristics of soil erosion, sediment and sediment delivery of the watershed.

Soil erosion from individual cells was highly spatially variable (Figure 3). The soil loss tolerance in earth-rock mountainous areas of southwest China is 500 t·km⁻² per year by the Ministry of Water Resources of China.

Only 48% of the area of the total watershed, total area of 1,996.8 km², is less than the tolerance value. The annual average soil erosion for the entire watershed is 1,551.7 t·km⁻²·y⁻¹ (Table 2), which is much larger than the tolerance value. A small portion (8%) of the watershed had soil erosion in excess of 5,000 t·km⁻²·y⁻¹ as shown in Table 2. Therefore, additional soil erosion control measures must be implemented to control soil erosion.

The soil erosion amount varied greatly with different land use types (Table 3). The forestland, shrub forestland and higher coverage grassland had low erosion amounts. The sloping land and lower coverage grassland covered 22% and 5% of the total watershed area, respectively, but their erosion amount accounted for 56% and 11% of the watershed total. Table 3 indicates that sloping land and lower coverage grassland are the main source of soil loss. Thus, the key of soil loss control is to control erosion from those lands and utilize those lands more rationally.

Soil erosion is the first step in the sedimentation processes which consist of erosion, transportation and deposition of sediments. A fraction of the eroded soil passes through the channel system and contributes to sediment yield, while much of soil deposits in landscape and water channels before reaching the outlet. Defining a sediment budget can help to better understand the sources, pathways and sinks (deposits) of sediment within a watershed system. The sediment budget of the watershed (Figure 4) shows that the average annual erosion amount of the Daning River watershed was 6.49 million tons, 21% deposited in the cells, 49% stored in the stream channels, and the rest was the output of the watershed. To reduce sediment loss at the watershed outlet, it is important to have more soil deposited in the cells. Thus, more conservation practices such as buffers should be adopted at the landscape scale to reduce sediment loss.

Table 4 shows annual average runoff and sediment characteristics from the main stem and six important tributaries in the watershed, which were also discretized by AGNPS 2001 data preparation tools using a CSA value of 8,000 ha and an MSCL value of 2,000 m. The contribution areas above Wuxi and Dachang stations are 2,000 and 3,050 km², respectively.

Among the six tributaries, Dongxi River, Xixi River, Houxi River, Boyang River, Bayanzhi River and Pingding River which altogether are about 73% of the entire watershed, Xixi River and Boyang River are the largest and smallest subwatersheds respectively (Table 4).

The entire watershed can be classified into upper (above the Wuxing station), middle (between Wuxi and Dachang station) and lower (from Dachang station to the entire watershed outlet) watershed. Average annual sediment yield is 1.25 × 10⁶ t·y⁻¹ at the Wuxing station of the upper watershed, rapidly adds up to 1.81 × 10⁶ t·y⁻¹ at the Dachang station at the middle watershed, whereas it increases slowly in the lower watershed, with only 1.95 × 10⁶ t·y⁻¹ at the watershed outlet. This indicates that the total 1.95 × 10⁶ t·y⁻¹ sediment yield, 64% comes from the upper, 29% from the middle and 7% from the downstream watershed, respectively. Among the six tributaries, the Dongxi River, with the average annual sediment yield of 5.23 × 10⁵ t, about 27% of the total sediment yield, had the largest sediment contribution to the entire watershed. Dongxi River watershed, with a large amount of deep sloping lands and lower coverage grassland, will need more conservation measures implemented to reduce sediment loss. In addition, strategies of controlling bank erosion should also be sought to reduce sediment loss from this mechanism. Boyang River with the largest runoff, 565.2 mm·y⁻¹ in six tributaries, yielded 3.16 × 10⁵ tons sediment per year, which was about 16% of the entire watershed. Bayangzhi River watershed had the least soil erosion of 766.6 t·km⁻²·y⁻¹, followed by Pingding River watershed (1,054.9 t·km⁻²·y⁻¹). The reason that these two subwatersheds had low soil erosion is that most of the watershed area was covered by forestland and shrub forestland.

Sediment yield is a critical factor in identifying non-point source pollution as well as in design and construction of dams and reservoirs. However, sediment yield is usually not available as a direct measurement but estimated by using a sediment delivery ratio (SDR) [27]. An accurate prediction of SDR is important in controlling sediments for sustainable natural resources development and environmental protection.

Figure 5 shows the spatial distribution of SDR of the watershed. Estimated SDRs range from 0.376 to 0.531 in six tributaries, which is consistent with the ranges from 0.15 to 0.61 for main watersheds in the Three-Gorge area of the Yangtze River [28]. SDR of the whole watershed is approximately 0.30. The maximum SDR is 0.531 in the Xixi River watershed, and the minimum SDR is 0.376 in the Houxi River watershed. The reason for the variation of SDR is that each subwatershed varies greatly not only in the drainage area, but also in slope, relief-length ratio, runoff-rainfall factors, land use/land cover and sediment particle size. This SDR information is useful for future conservation planning of the watershed for effective soil erosion control and sediment loss.

The Annualized Agricultural Nonpoint Source Pollution (AnnAGNPS) model is an advanced technological watershed evaluation tool that has been developed through a partnership between the U.S. Department of Agriculture (USDA)—Agriculture Research Service (ARS) and the USDA—Natural Resources Conservation Service (NRCS) to aid in the evaluation of watershed responses to agricultural management practices [3,5,12,29]. AnnAGNPS is a continuous-simulation, daily time-step, pollutant loading model designed to simulate long-term chemical and sediment movement from agricultural watersheds; and it includes significantly more advanced features than AGNPS [6]. The spatial variability of soils, land use, and topography within a watershed is accounted for by dividing the watershed into many user-specified, homogeneous, drainage-area-determined cells. Runoff, sediment, and chemicals are routed from each cell through a channel network to the outlet of the watershed. The model has the capability to identify the sources of pollutants at their origin and to track them as they move through the watershed system. The surface runoff from a field is determined using the Soil Conservation Service Curve Number (SCS-CN) technique [30] within AnnAGNPS. The peak flow is calculated by the extended TR-55 technique [31]. The lateral subsurface flow and tile drainage are also accounted for in the model. Runoff in channels is calculated using Manning’s equation. Soil erosion from each field is predicted by Revised Universal Soil Loss Equation (RUSLE) [32]. The sediment yield leaving each field is based upon the Hydro–Geomorphic Universal Soil Loss Equation (HUSLE) [33]. The sediment reach routing is based on a modified Einstein deposition equation [34] using the Bagnold suspended sediment formula for the transport capacity by particle size class. The model can be used to study the effects of alternative cropping and tillage systems including the effects of fertilizer, pesticide, and irrigation application rates as well as point source yields and feedlot management [12].

Required input parameters for application of the model include climate data, watershed physical information, and management information. Daily climate information, which includes daily precipitation, maximum and minimum temperatures, dew point temperature, cloud cover, and wind speed, is needed to account for temporal variation in the weather. Climate data for simulation can be historically measured, synthetically generated using the climate generator program [35], or a combination of the two. GIS data layers of a watershed are extremely helpful in characterizing the watershed physical information.

Using the GIS digital data layers of digital elevation model, soils, and land-use, a majority of the data input requirements of AnnAGNPS were developed by using a customized ArcView GIS interface. Inputs developed from the ArcView GIS interface include physical information of the watershed and subwatershed (AnnAGNPS cell), such as boundary and size, land slope and slope direction, and channel reach (AnnAGNPS reach) descriptions. The ArcView GIS interface also assigned a soil and land-use type to each cell by using the generated subwatershed and the soil and land-use GIS data layers. Additional steps to provide the model with the necessary inputs included developing the soil layer attributes to supplement the soil spatial layer, establishing the different crop operation and management data, and providing channel hydraulic characteristics. Those inputs can be organized using the AnnAGNPS Input Editor, a graphical user interface designed to aid users in selecting appropriate input parameters. Management information includes various field management operations such as planting, cultivation, fertilization, pesticides and harvesting, much of which can be obtained from RUSLE [31] databases or from actual activities implemented.

Output information produced by the model includes runoff, sediment, nutrient and pesticide loads at a temporal scale ranging from daily to yearly and at any desired location such as specific cells, stream reaches, feedlots, gullies, or point sources. The model also has the capability to provide source accounting information in terms of the fraction of a pollutant load passing through any reach location that originated from an upstream watershed pollutant source area. Further details on the theoretical background of AnnAGNPS can be found in the literature [3,5,30,36].

The study area is the Daning River watershed (108°44' to 110°11' E, 31°04' to 31°44' N), which is located in the center of the Three-Gorge reservoir area of the Yangtze River, China (Figure 6). The entire watershed has a contributing drainage area of about 4,181 km² and drains into the Yangtze River. The terrain is undulating, ranging from 95 m to 2,793 m above sea-level, and the slope varies from 0 to 138%, which is derived from 1:50,000 DEM. The region has a subtropical humid-warm climate, with annual average temperatures of around 19.8 °C. Annual rainfall ranges from 1,000 mm in the valleys to 1,700 mm in the high mountain areas of the watershed with large variation from year to year. Over 87% of the rainfall occurs between April and October, which markedly affects runoff and soil erosion. The main crops grown in the Daning River watershed are rice, wheat, maize and rape. The soils in the study area vary from sandy loams to loamy sands with most of the soil types in the watershed classified as clay loam. The dominant soil in the watershed is yellow soil and yellow limestone soil, which represents almost 50% of the watershed. Major soil types in the watershed and their basic physical and chemical properties are listed in Table 5.

Detailed records of agricultural operations including tillage, planting, harvesting, and fertilization have been maintained since 1998. The operation management of four main crops related to this study is listed in Table 6.

Daily runoff data at the Wuxi monitoring station during 1998 and 2005 were obtained from the Ministry of Water Resources of China. Runoff was monitored using a velocity instrument method, i.e., the two point method and the eleven point method. The two point method, employed for the low water level in the station, is that velocities are sampled at two tenths and eight tenths of the water depth, and the results are averaged. The eleven point method, used for the high water level, is the same as the two point method, but velocities are sampled at eleven points of the water depth (top, one tenths, two tenths, three tenths, four tenths, five tenths, six tenths, seven tenths, eight tenths, nine tenths and bottom). Unfortunately, the data between 2000 and 2002 were missing. In addition, detailed sediment data were also unavailable.

Runoff data from 1998–1999 were used for calibration and data from 2003–2005 were used for validation. The purpose of calibration is to achieve a satisfactory simulation through an iterative procedure of parameter evaluation and refinement, as a result of comparing simulated and observed values of interest. The purpose of the model validation is to determine the quality of the model predictions for other time periods not considered in calibration. The following statistics were used to help evaluate model performance for calibration: the mean and standard deviation of both observed and simulated values, the relative error (RE), the gradient k, intercept b, and coefficient of determination (R²) of linear regression, the Nash-Sutcliffe coefficient of efficiency (NSE) [42], and root mean square error (RMSE).

The RE is the ratio of the total difference between simulated and observed values versus the total observed value. It ranges from minus one to ∞ while zero indicates that there is no difference between model simulation and field observation. The smaller the absolute value of a RE, the better performance of the model is.

The coefficient of determination, R², gives the proportion of the variance in the observed values explained by the simulated values. For a good agreement the gradient k should be close to one and the intercept b should be close to zero.

The NSE is a technique often employed to evaluate simulated results in hydrological modeling. It is calculated according to Equation (1):

where O_i, P_i and ō are observed, predicted, and the mean of the observed value respectively; and n is the total number of events. The NSE ranges from −∞ to 1 while one indicates that the model is perfect. A value larger than zero indicates that the model is minimally acceptable. Values between zero and one indicate that the model is a better indicator than the mean of the observed values; and negative values indicate that the mean of the observed values is a better indicator than the model [42].

The RMSE describes the difference between the observed and simulated values in the unit of the variable [23,43]. The RMSE range from 0 to ∞ while zero indicates there is no difference between model simulation and field observation. The RMSE can be expressed by Equation (2):

AnnAGNPS did not simulate base-flow, hence to compare the model predicted runoff to observed runoff, baseflow was separated from the observed runoff using the baseflow filter method [44].

SDR, a sediment transmission coefficient, is calculated as the ratio of sediment yield at the watershed outlet (point of interest) to the gross erosion in the entire watershed. It can be obtained from the following Equation:

where SDR is the sediment delivery ratio. SY is the sediment yield, and E is the gross erosion per unit area above a measuring point.