Dual-Level Domain Alignment Meets Fine-Grained Contrast: Advancing Cross-Scene Few-Shot Hyperspectral Image Classification

Abstract

In hyperspectral image classification, existing cross-domain few-shot learning (CDFSL) approaches primarily focus on feature-level adaptation yet often neglect distribution-level shifts, which restricts the model’s transferability across different domains. This limitation is further exacerbated in challenging environments like coastal wetlands, where highly mixed vegetation, significant intra-class spectral variability, and strong inter-class similarity result in severe inter-class overlap and intra-class dispersion. To address these challenges, we propose a cross-domain few-shot HSI classification framework incorporating dual-level domain adaptation and fine-grained supervised contrastive learning (CDFS- $\mathrm {D}^{2}$ AFC), aiming to capture class-discriminative domain-invariant features with few labeled target samples. First, we put forward a global–local adaptive interaction transformer network as the backbone feature extractor, which dynamically fuses global self-attention and local convolutional features via an adaptive interaction strategy to enrich feature representations. Subsequently, a dual-level domain adaptation module is designed to align intra-domain features and minimize inter-domain distribution discrepancies, alleviating feature and distribution shifts. Furthermore, we propose a mask-based fine-grained supervised contrastive learning approach to enhance the model’s discriminative capability and robustness against noise. This method applies contrastive constraints on mask-augmented samples at two granularity levels, self-positive and context-positive, effectively mitigating the adverse effects of ambiguous class boundaries in complex environments. Extensive experiments on five cross-domain tasks, including three coastal wetland datasets and two public benchmark datasets, show that CDFS- $\mathrm {D}^{2}$ AFC outcomes state-of-the-art approaches, demonstrating strong robustness in challenging scenarios and consistently achieving high performances on the public benchmark.