Abstract: Aiming at the problem that convolutional neural network (CNN) could only be applied to Euclidean data and could not effectively 
obtain global relationship features between pixels and long-distance contextual information, we constructed a superpixel segmentation-based graph attention network (SSGAT). The network treated the segmented superpixel blocks as graph nodes in the graph structure, effectively reducing the complexity of the graph structure and reducing the noise of the classification graph.  
The classification accuracy of SSGAT and the comparison algorithm were tested on three datasets, and overall classification accuracy of 94.11%, 95.22%, and 96.37% were obtained, respectively. The results show that the method has excellent performance and significant advantages in dealing with classification problems in large-scale regions.

Key words: hyperspectral image, graph attention network, residual mechanism, superpixel segmentation