Abstract: In order to improve the classification accuracy of hyperspectral images, a novel hyperspectral image classification method based on generative adversarial network is proposed. The proposed generative adversarial network consists of generator, discriminator and classifier, in which the generator is used to approximate the data distribution of hyperspectral samples and generate specific categories of samples. The discriminator is a binary classifier to determine whether the input samples are real data. And the classifier is used to classify the input samples. The parameters of generator, discriminator and classifier are updated sequentially by back propagation algorithm to minimize the loss function, so as to achieve the goal of training the network. Generators and discriminators can approximate the sample distribution of hyperspectral images to assist the training of the classifier. Therefore, they can effectively improve the classification accuracy of hyperspectral images. Pavia University and Salinas hyperspectral data sets are used for classification experiments respectively. The experimental results show that the proposed classification method can effectively improve the classification accuracy of hyperspectral images with only small samples.

Key words: hyperspectral image classification, small samples, generative adversarial networks, deep learning, generative models