The Loess Plateau suffers from severe soil erosion in China. The accurate estimation of soil erosion has important guiding significance for the implementation of local projects such as land management, water and soil conservation. In this study, we selected Yanhe River Basin as the study area, and four years Landsat TM remote sensing data (in 2000, 2005, 2010, and 2015), rainfall, land use, DEM and soil properties as the source data. We compare the RUSLE, InVEST and USPED models in the studied area. We analyze the prediction accuracy under different terrains and vegetation distributions. The results showed that (1) soil erosion in the Yanhe River Basin increased first and then decreased from 2000. The reduction gradually increased after 2005. It indicates that the project of Grain for Green had a significant effect; (2) compared with the sediment yield data, the amount of soil erosion predicted by the RUSLE model are relatively large, and the USPED and InVEST models are more accurate than RUSLE model. Therefore, the InVEST and USPED models are suggested to predict soil erosion in the Yanhe River Basin; (3) soil erosion increased with the increase of slope, and the RUSLE model increased the most; (4) when NDVI > 0, soil erosion decreased with the increase of NDVI; and when NDVI was between 0 and 0.1, soil erosion was the largest.