Mixture of experts with virtual class for unsupervised domain adaptation in hyperspectral image classification

Pseudo-label-based methods have gained significant attention in unsupervised domain adaptation for cross-scene hyperspectral image (HSI) classification. However, existing approaches often treat spatial and spectral information uniformly under domain shift and typically rely on confidence thresholds to filter pseudo-labels, which may discard informative samples or amplify errors under severe shift. To address these issues, we propose a mixture-of-experts framework with virtual class supervision for HSI domain adaptation. The framework comprises a spatial expert and a spectral expert to capture modality-specific class cues, together with a reconstruction expert regularized by a spatial–spectral mask reconstruction task to enhance cross-modal interaction modeling. Instead of thresholding, an expert-meeting mechanism categorizes target samples by expert consensus and disagreement, and applies risk-aware pseudo-label supervision, where conflicting samples contribute through virtual-class-based gradient suppression rather than unreliable positive reinforcement. Experiments on three cross-scene benchmarks demonstrate consistent improvements over state-of-the-art methods, validating the effectiveness and robustness of the proposed framework.