Hyperspectral Image Classification with Graph Attention Network

Convolutional neural networks have achieved great advantages in the hyperspectral image (HSI) classification. Nevertheless, their ability to model topological relations among samples is limited, especially in the class boundary regions. To alleviate this problem, a novel graph attention classification network (GANet) is proposed, which is mainly composed of a graph branch, a spectral branch, and a spatial branch. The graph branch is designed to extract topological relation attention features via a dense graph convolutional block and a multiway graph attention module. It can emphasize useful features and suppress useless features, thus expanding the distance between different classes and shortening the distance between the same classes. The spectral and spatial branch can extract the spectral attention features and the spatial attention features respectively. Furthermore, different from traditional concatenate operation, a simple yet effective parallel attention module is built to integrate different types of features, which can further improve the classification accuracy. Extensive experimental results demonstrate that on the Indiana Pines dataset, the proposed GANet obtains higher classification performance than state-of-the-art methods.