Abstract: Abstract: Small watersheds are typically used as basic units for the comprehensive control of soil and water loss in the Loess Plateau. Study on the mechanism of the critical topography on erosion and sediment is very important to the establishment of prediction model for watershed-scale erosion and sediment. The research presented was conducted in the Chabagou watershed, which was located in the hill-gully area of the Loess Plateau, China. A back propagation artificial neural network (BPANN) model for watershed-scale erosion and sediment yield was established, whose accuracy was then compared to that of the multiple linear regression (MLR) model. The sensitivity degree of various factors to erosion and sediment yield was quantitatively analyzed. Based on the sensitive factors and the fractal information dimension, the critical geomorphic prediction model for erosion and sediment yield of individual rainfall event was established and further verified. The results revealed that the BPANN model performed better than the MLR model in terms of predicting the erosion modulus, and the former was able to effectively characterize dynamic changes in sediment yield under comprehensive condition of the factors. The sensitivity of runoff erosion power and runoff depth to the erosion and sediment yield associated with individual rainfall event was found to be related to the complexity of surface topography. The characteristics of such a hydrological response were thus closely related to topography. When surface topography was complex, the erosion modulus was more sensitive to runoff erosion power; conversely, when surface topography was simple, the erosion modulus was more sensitive to runoff depth. The developed sensitivity method based on BPANN was employed in order to select the main predictive (sensitive) factors for erosion and sediment yield; as the influence of these factors gradually increased, this quantitative method was increasingly helpful. Therefore, when the fractal information dimension was greater than the topographic threshold, the accuracy of prediction using runoff erosion power was higher than that using runoff depth. In contrast, when the fractal information dimension was smaller than the topographic threshold, the accuracy of prediction using runoff depth was higher than that using runoff erosion power. The characteristics of erosion and sediment transport, i.e. the hydrological response of a watershed, were closely related to topography. Sheet (inter-rill) erosion, which exhibited rainfall erosion characteristics, was more prone to occur when topography was simple. In contrast, rill and gully erosion, which exhibited the dual characteristics of rainfall and sediment transport, were more inclined to occur when topographic thresholds were exceeded. The fractal information dimension was used as a model boundary; when the value of the fractal information dimension was greater than the selected topographic threshold, the accuracy of predictions using runoff erosion power was higher than that using runoff depth. In contrast, when the value of the fractal information dimension was smaller than the topographic threshold, the accuracy of predictions using runoff depth was higher than that using runoff erosion power. Therefore, the piecewise prediction model for watershed-scale erosion and sediment yield of individual rainfall event, in which runoff erosion power and runoff depth are introduced using the fractal information dimension as a boundary, can be considered feasible and reliable, and has a high prediction accuracy. Considering the observed watershed differences and the relatively insufficient fractal dimension values, further comprehensive analysis and comparison should be carried out in order to establish piecewise erosion prediction models for other watersheds. However, as the present piecewise erosion prediction model takes account of watershed-scale surface topography, runoff depth and runoff erosion power, as well as the relationships between these factors, it can be used as a basis to establish and popularize other erosion models.