Explainability Feature Bands Adaptive Selection for Hyperspectral Image Classification

Hyperspectral remote sensing images are widely used in resource exploration, urban planning, natural disaster assessment, and feature classification. Aiming at the problems of poor interpretability of feature classification algorithms for hyperspectral images, multiple feature dimensions, and difficulty in effectively improving classification accuracy, this paper proposes a feature band adaptive selection method for hyperspectral images. The proposed feature band adaptive selection model focuses on the joint salient feature regions of the hyperspectral image, visualizes the feature contribution of the bands, more intuitively reveals the selection basis of the feature bands of the hyperspectral features in the process of deep learning, and selects the feature bands with high contribution to carry out classification experiments for verification. Quantitative evaluations on four hyperspectral benchmarks (Pavia University/Centre, Washington DC, GF-5) demonstrate that EFBASN achieves state-of-the-art classification accuracy, with an overall accuracy (OA) of 97.68% on Pavia U, surpassing 12 recent methods including SSCFA (94.48%) and CNCMN (93.12%). Crucially, the attention weights of critical bands (e.g., Band 26 at 672 nm for iron oxide detection) are 3.2 times higher than redundant bands, providing physically interpretable selection criteria.