Multiscale Fusion Graph Convolutional Networks for Hyperspectral Image Classification

Graph neural network (GNN) has shown remarkable performance in many fields, such as edge prediction, node classification, etc. However, the traditional graph neural network can only accept the full graph to train the model, which leads to a very high computational complexity. Besides, the lack of multi-scale mechanism limits the actual classification performance. In this paper, we propose a multiscale fusion model based on graph convolutional network (GCN) and convolutional neural network (CNN) for hyperspectral image (HSI) classification tasks. This model can first sample randomly from the full graph to subgraphs of different scales, then trains GCN on sub-graphs of different scales and extract features, and finally integrates the spatial-spectral features extracted by multi-scale CNN classifier and the structural features extracted by multi-scale GCN for end-to-end mini-batch joint training. Experiments on two real HSI data s demonstrate that the proposed method significantly improves the classification performance compared with other comparison methods.