Remote sensing image classification is an important method to distinguish different types of targets according to the different features reflected in the image information. When traditional convolutional networks solve this kind of problem, it could easily cause network degradation and increase the amount of calculation due to the deepening of depth. In view of this, the paper proposes an improved structural model of residual network to enhance the classification performance of the network. Firstly, a 3D neural network and a residual neural network are used to extract features from the Salinas scene dataset to reduce the data volume. Then a channel attention mechanism is added to the residual neural network to improve the efficient recognition of hyperspectral image feature weights. Finally, a comparison experiment between the 3D neural network and the residual neural network is conducted, and the results show that the optimized residual network is more efficient.