Progressive Contrastive Learning With Confidence-Weighted Sample Contribution for Cross-Scene Hyperspectral Image Classification
Abstract
Cross-scene hyperspectral image classification (HSIC) under unsupervised domain adaptation (UDA) remains challenging due to substantial distribution discrepancies caused by variations in acquisition conditions. In this setting, labeled samples are only available in the source domain, while the target domain lacks supervision. Despite recent progress, existing hyperspectral image domain adaptation (HSIDA) methods are still vulnerable to severe domain shifts, particularly when unreliable pseudo-labels introduce erroneous supervision. Moreover, unconstrained cross-domain feature interaction may further amplify distribution discrepancies and degrade representation quality. To address these issues, we propose a progressive confidence-weighted contrastive learning network (PC2L-Net) for HSIDA. At its core, a progressive confidence-weighted contrastive learning (PCCL) strategy is developed to regulate target-domain sample contributions based on pseudo-label reliability, mitigating error propagation during contrastive learning. In addition, an LMMD gradient-guided reweighting (LGGR) module performs feature dimension-level discrepancy-aware alignment by adaptively reweighting spectral bands. Furthermore, an interdomain sparse cross-attention (ISCA) mechanism enables controlled cross-domain interaction while suppressing discrepancy amplification. Extensive experiments on three hyperspectral benchmark datasets demonstrate that the proposed method consistently outperforms representative HSIDA methods, yielding more discriminative and robust cross-domain representations.