Hyperspectral image classification with multi-scale graph convolution network

Hyperspectral images with high intra-class variability and dimensionality, which greatly weakens the performance of supervised classification methods, especially with the limited number of samples. The graph convolution networks (GCNs) have attracted much attention in the field of hyperspectral classification as it is able to formulate robust spectral-spatial features simultaneously. To this end, we propose a multiscale graph convolutional network based on the spectral-spatial features called MSGCN for hyperspectral image classification. The framework allows us to fuse spectral-spatial features within the graph for node and edge feature generation. Based on that, we further aggregate information from multi-order adjacent nodes to obtain multiscale spectral-spatial features, which proves to be beneficial for hyperspectral image classification. To demonstrate the robustness of the proposed method, three widely used hyperspectral data sets were selected, with the overall accuracy of 99.52%, 99.34% and 98.31%, respectively.