A convolutional neural network (CNN) can extract hierarchical features in an image, and the extracted images include a large amount of detailed information contained in the image. However, CNN-based methods do not take full advantage of all hierarchical features extracted by the network. To make full use of all hierarchical features and enhance feature reuse and information flow, we design a residual dense network suitable for hyperspectral image classification. The residual dense network combines residual and dense networks, including shallow feature extraction, residual dense units, and dense feature fusion. Firstly, shallow features of the original image are extracted using a convolution operation, which is input to the residual dense unit. Secondly, the output of the residual dense unit establishes a shortcut connection with each convolution layer and output layer in the next unit, thereby realizing continuous information transmission. Subsequently, dense features extracted from the two units are added to the shallow features to form global residual learning, which realizes the fusion of all hierarchical features. The fused features are then used for hyperspectral image classification. Experimental results demonstrate that the proposed method can obtain 98.71%, 99.31%, and 97.91% classification accuracies on the Indian Pines, University of Pavia, and Salinas data, respectively, which effectively improves the classification accuracy of hyperspectral images and enhances the stability of classification methods.