Semantic-Guided Cross-Modal Feature Alignment for Cross-Domain Few-Shot Hyperspectral Image Classification

When facing the challenge of limited samples, existing hyperspectral image (HSI) classification methods typically assume that source domain samples (with prior knowledge) and target task samples share identical data distributions. However, real-world applications frequently encounter cross-domain scenarios where HSIs acquired from distinct environments exhibit significant distribution discrepancies. This domain shift phenomenon severely degrades the performance of models trained on source domains when deployed in target domains. To deal with this problem, this paper presents a semantic-guided cross-domain HSI classification method, namely Semantic-Guided Cross-Modal Feature Alignment (SG-CMFA). Our architecture employs a dualchannel spatial-spectral Mamba module to hierarchically extract discriminative spatial and spectral features from HSIs, enhancing feature representation capacity. Simultaneously, we integrate a BERT-based semantic encoder to distill label-specific textual semantics, which are subsequently fused with visual features to generate semantic-enhanced class prototypes. This cross-modal alignment strategy effectively mitigates domain shift by establishing semantic consistency between source and target domains. Comprehensive evaluations across three benchmark HSI datasets validate that our approach outperforms some state-of-the-art approaches in few-shot HSI classification scenarios.