FGAPA: Feature-Guided Adversarial Prototype Alignment For Cross-Domain Few-Shot Hyperspectral Classification

Cross-domain learning provides an effective way to the scarcity of labeled samples in hyperspectral image classification (HSIC). Existing single-domain adaptation methods often suffer from feature loss, disruption of inter-class structures, and limited ability to extract cross-domain invariant features. To address these challenges, this paper proposes a feature-guided adversarial prototype alignment (FGAPA) framework for cross-domain few-shot HSIC. FGAPA first constructs a feature fusion enhancement module to strengthen inter-class correlations, mine latent semantic information between samples and prototypes, and refine feature embedding representations. Subsequently, an adversarial cross-domain feature alignment strategy is introduced to achieve prototype alignment from a dual-domain perspective, effectively improving feature quality and alleviating domain shift. Experimental results on three public datasets demonstrate that the proposed FGAPA outperforms state-of-the-art methods.