Abstract: The risk evaluation of the valley debris flow is a basic and important part in preventing debris flow. To make the evaluation, we combine the stereoscopic convolution and the residual structure, and propose a neural network model for feature learning on DEM and multi-spectral data simultaneously. This article takes Nujiang Prefecture as an example and investigates the susceptibility of the entire-valley debris flow. After training on the data of historical debris flow hazard valley, the network can give a reasonable evaluation of the hazard of the debris flow based on the debris flow similarities, and then the risk assessment map of debris flow disaster in Nujiang Prefecture is drawn. Among all 214 valleys, there are 114 high risk valleys, 40 medium risk valleys, and 60 low risk valleys. The experimental results show that the model can achieve the highest accuracy rate of 80%, the recall rate of 88% and Kappa coefficient of 0.81 on the task of classification of valley debris flow. In addition, in experiments that use less training data and experiments that compare different models, the designed models perform better. The degree of risk given by the model is consistent with the disaster facts and the result of field investigations.