It is a frontier in the field of soil erosion control and soil-water conservation to establish a model of extensive applicability for prediction of soil erosion and sediment yield on a watershed scale. With Chabagou Watershed of the Loess Plateau as a case for study, an artificial neural network model was established for sediment yield; analysis was done of various factors for their sensitivities to sediment yielding with the default factor method; and then based upon sensitiveness factors and fractal information dimension, a piecewise prediction model for erosion and sediment yield of individual rainfall events was established and verified. Results show that the artificial neural network model demonstrated capability of describing quantitatively the coupling relationship between runoff and sediment with sufficient high accuracy. The sensitivity of sediment yielding of individual rainfall events to runoff depth and runoff erosion power is related to complexity of the landform. With fractal information dimension as boundary, runoff depth and runoff erosion power were introduced piecewise into the prediction model. The prediction based on runoff erosion power was higher than that based on runoff depth in accuracy, when D_i＞0.8308, and it was the other way round when D_i＜0.8140. The findings demonstrate that the establishment of the piecewise prediction model and the proposition of the method for prediction of erosion and sediment yield are reasonable and reliable, and of some value as reference for establishment of other models for prediction of erosion and sediment yield.