Efficient Foundation Model Adaptation for Hyperspectral Image Classification Under Label Scarce Conditions

This is the authors' abstract. We don't add key points for this paper.

Hyperspectral image classification (HSI) classification under sample-scarce conditions remains challenging due to limited labeled samples, spectral–spatial distribution shifts, and overfitting during model adaptation. Existing methods usually rely on metalearning or metric-learning-based paradigms, where episodic tasks or support–query pairs are constructed from a narrow set of source scenes. Such limited spectral diversity may introduce dataset-specific priors and weaken generalization. To address these issues, we propose an efficient foundation-model-driven adaptation framework for sample-scarce HSI classification. A pretrained spectral foundation model is adopted to provide transferable spectral priors learned from large-scale HSI data. We design regional contextual spectral–spatial tokenization (RCSST) to aggregate locally homogeneous pixels into compact region-level tokens and model interregion contextual dependences. Within RCSST, learnable global aggregation selectively fuses informative bandwise cues into regional embeddings. To efficiently adapt the pretrained encoder under limited supervision, spectral gating and NormalLayer adaptation recalibrates spectral responses and feature distributions while keeping most backbone parameters frozen. Experiments on four public benchmarks show that our method consistently outperforms state-of-the-art approaches under standard label scarce settings and achieves robust source-only cross-scene generalization on unseen target scenes.