Discrepancy-Conditioned Residual Feature Refinement for Multi-Source Hyperspectral Classification

Multi-source cross-domain hyperspectral image (HSI) classification is challenged by heterogeneous sensor configurations, scene-dependent distribution shifts, and limited labeled target data, which hinder effective knowledge transfer across multiple scenes. Motivated by progressive feature correction, we propose a three-stage Residual Feature Discrepancy Refinement (RFD) framework for collaborative representation learning across heterogeneous HSI domains. RFD formulates this correction as a deterministic, discrepancy-conditioned residual refinement process. First, domain-specific encoders project four source domains and the target domain, which may have unequal spectral dimensions and label spaces, into a common-dimensional feature space. Adaptive severity and domain weighting uses first- and second-order feature discrepancies to estimate source-specific conditioning coordinates and collaborative contribution weights. A shared discrepancy-conditioned residual refiner then performs multi-step feature refinement to reduce domain-dependent statistical deviations. Finally, an exponential-moving-average historical prototype memory stabilizes target adaptation, followed by cosine 1-nearest-neighbor classification. Across ten randomized runs, RFD achieves mean overall accuracies of 94.65%, 94.87%, and 96.95% on NC12, Salinas, and WHU-Hi-LongKou, respectively, and obtains the highest mean overall accuracy, average accuracy, and κ among the evaluated unified-protocol methods.