CNN -Enhanced Multi-Scale Graph Attention Network for Hyperspectral Image Classification

In recent years, the utilization of both Graph Neural Network (GNN) and Convolutional Neural Network (CNN) in hyperspectral image (HSI) classification has gained significant attention. GNN, in particular, have proven effective in modelling irregular image regions. However, the limitations of single-scale graph structures and the focus on super-pixel nodes instead of pixel nodes within GNN hinder the extraction of pixel-level spectral-spatial features. To address these challenges and leverage the strengths of both CNN and GNN, we propose a novel heterogeneous deep network called CNN-Enhanced Multi-Scale Graph Attention Network (CEMSGAT). In CEMSGAT, we employ semi-supervised Local Fisher Discriminant Analysis (SELF) for dimensionality reduction and spectral-spatial convolution to extract surface features. Furthermore, our utilize super-pixel segmentation to create multi-scale graphs and implementing an improved graph attention algorithm at each scale to process the features obtained from the spectral-spatial convolutions. A spatial transformation operation is designed to enable seamless integration between the different scales of the graphs. Simultaneously, the features obtained from the previous spectral-spatial convolution are fed into a multilayer convolutional network for deep feature extraction and enhance the accuracy of the classification of connected areas between different land cover types calculated by the graph attention algorithm to achieve a clearer classification. Finally, the super-pixel level features derived from the multi-scale graph attention network are fused with the pixel level features obtained from the multilayer convolutional network for precise hyperspectral image classification. Experimental results on three hyperspectral datasets demonstrate the superiority of CEMSGAT over numerous state-of-the-art methods.