Adaptive Band Selection for Hyperspectral Classification with Spatially Disjoint Evaluation
Researchers have developed a new two-stage method called SGBR-HC for hyperspectral band selection, which aims to improve classification accuracy. This approach uses a supervised spectral ranking to initialize trainable sparse gates, allowing the number of selected bands to be determined during training rather than being fixed beforehand. When evaluated on standard datasets with spatially disjoint validation, SGBR-HC achieved high accuracy with approximately twenty bands, highlighting the importance of its ranking prior and careful evaluation to avoid spatial leakage. AI