<p>Hyperspectral image (HSI) classification is the basis for the application of hyperspectral technology in many fields. Traditional HSI classification methods use manually designed features, and these models do not generalize well. Methods relying on deep convolutional neural networks are capable of autonomously deriving profound abstract characteristics from images for classification, but require a large number of labelled samples for supervised learning. However, in the actual application scenarios, it is often challenging to label the categories of ground objects in HSI. To address this problem, we propose a HSI classification method based on semi-supervised deep residual generative adversarial network, which improves the feature extraction ability of the discriminator by introducing residual module and long-short connection in the discriminator. At the same time, unlabelled samples and pseudo-samples generated by the generator are selected to expand the labelled sample dataset. The adversarial learning of generator and discriminator is utilized to enhance the ground object recognition ability and network generalization ability of the discriminator. Extensive experiments were carried out on three HSI datasets, Indian Pines, Pavia university and Salinas Valley. The classification results indicate an enhancement in the overall classification accuracy of the proposed method to 97.06%, 98.79%, and 98.61% respectively when random selections of 10%, 3%, and 3% of the samples in the three datasets are made for training. Compared with some existing deep learning-based HSI classification methods, the method can obtain higher classification accuracy.</p>