DISGT: Dynamic-Interactive Subgraph Transformer for Unsupervised Hyperspectral Band Selection

Band selection (BS) is one of the effective ways to reduce data dimensionality while retaining important information in hyperspectral imagery (HSI). The ultimate goal is to represent the original HSI using as few informative and discriminative bands as possible. However, existing BS methods usually divide all bands into independent spaces while neglecting the relationships between these spaces. Moreover, the evaluation of band importance through its neighbors requires more time to capture global information from the spectral perspective. Therefore, a dynamic-interactive subgraph Transformer (DISGT) is proposed for unsupervised hyperspectral BS. Specifically, a dynamic distance norm is designed to measure the correlation between bands and characterize several subspaces with better separability. Bands with uncertain subspace membership are shared with other subspaces through information interaction to enhance the diversity of band combinations, compensate for the effect of subspaces on global information processing, and construct subspaces into subgraphs. Furthermore, the subgraph Transformer (SGT) captures local and global spectral information using a sparse multihead attention (SparseMHA) mechanism to accurately evaluate band importance. Finally, the original HSI is reconstructed using the crucial bands to continuously adjust and obtain the optimal band subset. The superior performance of the DISGT is demonstrated through extensive comparisons with other state-of-the-art methods on three public datasets.