Convolutional neural networks have been applied in the field of remote sensing image classification. For convolutional neural networks with shallower layers and simpler structures, the accuracy of the recognition of debris flow gully images is not ideal, while the number of layers is deeper and the structure is relatively simple. More complex neural networks often consume a lot of system resources and are difficult to deploy on the user side. Aiming at this problem, an optimized convolutional neural network method is proposed. First, through multiple sets of comparative experiments, select Resnet101 and Resnet18 models with good image classification performance; then, use the characteristics of debris flow gully images to pre-train the deeper and more complex Resnet101 model; finally, through the method of knowledge distillation, The trained “knowledge” is extracted into the Resnet18 model to achieve the effect of improving accuracy while reducing system resource occupation. Experimental data shows that after using knowledge distillation, the accuracy and sensitivity of the Resnet18 model are increased by 2.36 and 1.72 percentage points, respectively, and the image processor occupancy is reduced by 37 percentage points compared with the Resnet101 model.