Toward Accurate, Efficient, and Interpretable Hyperspectral Image Classification: A detachable network with adaptive spectral–spatial scale selection

Ideal hyperspectral image (HSI) classification should be accurate, efficient and interpretable. HSIs exhibit multi-scale characteristics in both spectral and spatial dimensions, which existing work fails to model comprehensively, limiting classification performance. Moreover, since the optimal interpretation scales vary across land cover classes, features extracted by deep networks are often redundant, influencing the inference efficiency and interpretability. Therefore, in this paper, a detachable network is proposed for HSI classification. Parallel branches with different spectral and spatial resolutions are progressively constructed for feature extraction of HSIs. To improve the inference efficiency, multi-scale gating modules are designed at each stage to adaptively select branches for different inputs. A two-step training strategy with hierarchical clustering further embeds semantic information into the gating process, yielding more reliable branch selection. Experimental results on WHU-OHS and Luojia-HSSR datasets indicate that the proposed network performs better than existing ones, with the overall accuracy ranging from 0.821 to 0.837 and 0.827 to 0.851 respectively. The detachable network improves efficiency by 35% to 58% relative to the corresponding dense counterpart, and exhibits scale interpretability for different land cover classes.