Spectral Attention Graph Convolutional Network for Hyperspectral Image Classification

Graph Convolutional Network (GCN) can handle feature information efficiently via an end-to-end manner, which is popularly exploited for hyperspectral image classification. However, using only the normal graph convolution structure cannot deeply detect neighborhood data relationship. To further explore neighborhood data relationship, we propose a spectral information attention graph convolutional network (SIAGCN). We employ the graph convolutional network efficiently to acquire and gather spectral-spatial information. Then, spectral information attention (SIA) module is added to effectively explore the significance of spectral band and enhance the feature acquisition ability of GCN. Finally, we employ a broad learning network to enhance the width expansion. Experiments demonstrate that SIAGCN exceeds the comparison models.