Cross-Domain Few-Shot Learning for Hyperspectral Image Classification Based on Global-to-Local Enhanced Channel Attention

Cross-domain few-shot learning (FSL) has garnered significant attention in hyperspectral image classification (HSIC). However, current transfer learning (TL) approaches often struggle to effectively capture both global and local spectral-spatial dependencies, particularly under substantial domain shifts. To address these challenges, we propose a novel approach incorporating a spectral–spatial enhanced channel attention (SECA), which dynamically extracts multiscale global-to-local feature relationships. In addition, we introduce correlation alignment (CORAL) loss to explicitly reduce distributional discrepancies between domains, thus enhancing the cross-domain transferability of the model. To achieve a balance between efficiency and accuracy, the proposed framework integrates a lightweight inverted residual (IR) module. Experimental results on multiple benchmark hyperspectral image (HSI) datasets demonstrate that our method outperforms state-of-the-art techniques, offering superior classification accuracy, robustness, and domain adaptability.