Cross-domain few-shot hyperspectral image classification with frequency-aware augmentation

Cross-domain few-shot hyperspectral image classification remains challenging due to limited labeled samples and domain shifts between source and target domains. This paper proposes a Frequency-Aware Augmentation Cross-Domain Few-Shot Learning framework (FAA-CDFSL) to enhance domain generalization by incorporating frequency perception. A frequency-aware reconstruction strategy is introduced to generate augmented tasks by manipulating high-frequency components, while a mutual attention module facilitates cross-frequency interaction, emphasizing transferable low-frequency semantics. Extensive experiments on three benchmark hyperspectral datasets demonstrate that our method achieves superior performance and stability under 1–5 shot settings, consistently outperforming state-of-the-art cross-domain few-shot methods. These results confirm that frequency-aware augmentation provides an effective way to improve robustness and adaptability in hyperspectral image classification.