Abstract: The mountain hazards in the river courses such as the debris flow, landslide, mountain flood and soil erosion seriously threaten the safety of the important infrastructures such as roads, railways, bridges and large-scale hydroprojects around the river. More than 90% of the traffic interruption in the Sichuan-Tibet highway are caused by the mountain hazards, which restrict the development of economy and society in Tibet. It is of great significance to quickly identify the mountain hazards in the river courses for timely adoption of appropriate emergency disaster plans and release evacuation information. The observation of the mountain hazards in the river courses by professional inspectors in a traditional way is of great danger and obvious lag. Therefore, it is urgent to study a new method as an alternative. With the arrival of the big data era, a deep learning technology represented by the convolution neural network has many characteristics such as local connectivity, parameter sharing and pooling, and has more powerful features and learning skills compared with the traditional machine learning method. This study has completed the training of Caffenet and other depth models by using a large number of image data of the mountain hazards in the river courses with the deep learning open source framework. By use of the transfer learning, the recognition accuracy eventually reached more than 90 percent. The analysis results show that the recognition method provides a new idea for the fast recognition of the typical mountain hazards in the river courses and the improvement of the group measuring system.