[Objective] To found a new approach to estimate soil erosion modulus, and achieve predictions of spatial distribution of soil erosion based on GIS. [Methods] Taking soil erosion modulus as discriminant conditions, each applicability of soil erosion prediction model built based on Logistic regression and RBF neural network was validated, and then the improved model(soil erosion prediction model) based on LOG-RBF neural network was built and validated. [Results] (1) There was obvious advantage for Logistic regression model to discriminant the occurrence of soil erosion, and the accuracy of prediction for un-occurring and occurring was 77.4% and 97.9%, respectively, the total predictive accuracy was 94.9%. (2) RBF neural network model had the stronger ability to estimate soil erosion modulus, the relative error and error sum of squares of the simulation results was 0.612% and 13.292, respectively, and R² was 0.57. (3) Relative error and error sum of squares of the simulation results was decreased by 0.157% and 2.601, respectively based on LOG-RBF neural network model than RBF neural network model, and R² was 0.82, so LOG-RBF neural network model had a better fitting degree, and with the soil erosion modulus increase, misjudge phenomenon showed a trend of gradual reduction. Determined by receiver operating characteristic curve, the value of area under curve based on LOG-RBF neural network model was 0.063 larger than RBF neural network model, and the accuracy was higher. [Conclusion] LOG-RBF neural network model could be used to estimate soil erosion modulus, and predict spatial distribution of soil erosion based on GIS.