Aiming at the problem that it is impossible to construct an effective deep learning classification model for hyperspectral images when there are few training samples, a “pyramid” residual block classification algorithm is proposed by improving the traditional residual block model. A deep learning network containing dozens of convolution layers is designed, which greatly reduces the number of parameters compared with the traditional residual model, and can fully extract the deep spatial-spectral features of hyperspectral images; Experiments are carried out on two open source datasets Indian Pines and University of Pavia, and three classical classification methods are selected for comparison. The experimental results show that the algorithm has the best performance and can effectively improve the classification accuracy of hyperspectral images.