S3DG-Net: A Spectral-Spatial Dynamic Superpixel Graph Network for Hyperspectral Image Classification

Abstract

Hyperspectral image classification is challenging due to the existence of “same object with different spectra” and “different objects with the same spectra”, which makes the classification accuracy difficult to improve. Existing methods based on CNN have limitations in their receptive fields and thus are unable to effectively capture the long-range spatial dependencies of land features. To address these issues, this paper proposes a Spectral-Spatial Dynamic Superpixel Graph Network ($\mathrm{S}^{3}$ DG-Net). This network consists of three parallel feature extraction paths. The grouped spectral-spatial feature enhancement branch uses the grouped dual attention mechanism to highlight discriminative land feature characteristics and achieve spectral-spatial feature enhancement; the multi-scale feature extraction branch extracts multi-scale land spatial features through the construction of parallel convolutional structures with multiple convolution kernels, improving the land feature recognition ability; the graph convolution branch based on superpixel segmentation constructs the long-range spatial dependencies between land features. Finally, the features extracted by the three branches are adaptively fused using cross-attention methods. The experimental results on the Indian Pines dataset show that $\mathrm{S}^{3}$ DG-Net outperforms 9 existing comparison methods in terms of overall accuracy coefficient, verifying the effectiveness of this network in hyperspectral image classification.