3D residual attention network for hyperspectral image classification

Abstract

Convolutional neural network (CNN) has been widely applied to hyperspectral image (HSI) classification and shows good performance. In this paper, a novel 3D residual attention network (3D-RAN) based on CNN is proposed for HSI classification. The main contributions and novelties of the method are as follows: (1) the 3D-RAN can directly process the 3D HSI data and effectively preserve the data structure of HSI; (2) the constructed residual module can supplement the loss of some important information during the process of information transmission, so that it makes the extracted features contain richer information; (3) the embedded attention module can strengthen the important features and suppress the unimportant features by adjusting weights, which effectively improves the classification performance. Moreover, weight sharing makes the attention module require fewer parameters. The experiments on three public HSI data sets demonstrate that the proposed 3D-RAN outperforms several state-of-the-art methods.