Soft-Label Denoising Contrastive Learning

Hyperspectral images (HSI) exhibit complex neighborhood contextual relationships and severe spectral uncertainty in ground objects, such as spectral diversity in local homogeneous regions and high inter-class spectral similarities. Denoising contrastive learning aims to construct advanced discriminative representations while preserving rich and specific contextual information. Although this approach enhances neighborhood discriminative feature learning, it faces inherent conflicts between spatial consistency and localized spectral variations in homogeneous regions due to the risk of misclassifying negative pairs in neighborhood contrastive learning. To address these challenges, we propose a Soft-Label Denoising Contrastive Learning (SDCL) method for HSI, integrating denoising contrastive learning with adaptive similarity calibration. Building upon probabilistic soft assignment, the framework dynamically reweights intra-neighborhood relationships through manifold-aware feature distribution matching, effectively balancing local consistency with discriminative feature preservation. Experimental results on two public datasets demonstrate that the proposed method exhibits strong competitiveness compared to other self-supervised approaches.