Multi-Gaussian Prototype Metric With Dual Optimization for Cross-Domain Few-Shot Hyperspectral Image Classification

Metric-based few-shot learning (FSL) methods have shown great potential in cross-domain hyperspectral image (HSI) classification. However, existing approaches face three key challenges: First, HSIs exhibit complex spatial–spectral structures, and cross-domain distribution shifts further exacerbate feature discrepancies among same-class samples, increasing the difficulty of extracting class-consistent features. Second, in cross-domain few-shot scenarios, although some methods incorporate pseudo-label estimation to achieve class-level alignment, their effectiveness is limited by the scarcity of labeled samples in the target domain and the uncertainty of pseudo-labels, and existing domain-adversarial strategies still primarily focus on global distribution alignment, making it difficult to effectively model class-level shared features across domains. Third, intraclass feature modeling typically relies on a single class prototype, which fails to capture the complex and long-tailed distribution of intraclass features in few-shot learning scenarios. To address these challenges, a multi-Gaussian prototype metric with dual optimization (MGPDO) is proposed for cross-domain few-shot HSI classification. Specifically, for consistent feature extraction from samples with complex structures, a multiscale feature extractor is designed to obtain features at multiple scales via parallel convolutions and pyramid pooling structures. For class-level alignment, the prototype-based cyclic domain triplet loss is proposed, integrating domain-adversarial learning with cross-domain prototype contrast. The module iteratively optimizes prototype pairs to increase similarity between positive pairs and reduce correlation between negative pairs, promoting class-level alignment and shared features extraction across domains. Furthermore, for intraclass feature modeling, a multi-Gaussian prototype metric is introduced, which assigns multiple Gaussian prototypes to each class to construct a one-to-many mapping, effectively modeling complex long-tailed feature distribution and improving classification performance. Experimental results on three standard hyperspectral datasets validate the effectiveness of the proposed method, offering a novel perspective on deep learning-based cross-domain few-shot HSI classification.