Classification Research Based on Residual Network for Hyperspectral Image

Classification is one of the main ways to mining the rich information in hyperspectral image (HSIs). At present, the deep learning methods have been applied to image processing and achieved good classification results. But overfitting is still a problem in the image classification by the deep neural networks. Residual network (ResNet) can help to overcome the overfitting phenomenon. It can optimize parameters and extract characteristic information conveniently by adding identity mapping between input and output. In this paper, ResNet is introduced to the HSI classification. In order to reduce the dependence of network on parameter initialization and to improve the generalization ability of the model, batch normalization method is adopted to optimize the network. Then, some virtual samples are generated to reduce the impact of limited number of training samples on the classification accuracy. Two different HSIs have been classified by the proposed method and the experimental results show that this method has promising prospects in the HSI classification.