Enhancing Label Noise Robustness for Hyperspectral Image Classification by Neighborhood Contrastive Learning

Recent advancements in hyperspectral images classification (HIC) rely on high-quality annotations and thus inevitably suffer from noisy labels. To address the negative effects of noisy labels, some methods employ neighborhood samples to select clean samples and demonstrate promising results. However, they typically rely on robust feature extraction and remain limited under high noise ratios. To overcome the limitations, we propose a novel robust sample selection and correction method based on robust contrastive learning and neighborhood feature modeling. The proposed RSC adopts a dual-branch spectral-spatial network combining spatial and channel-based residual attention modules to extract robust feature. Furthermore, unsupervised contrastive learning at both feature and logit-level are introduced to bolster the feature extractor. Finally, a clean sample selection strategy based on neighborhood consistency in feature space and relabelling scheme by the maximum confidence are integrated to resist the noisy labels. Extensive experiments conducted on publicly available hyperspectral datasets, including Houston and Indian Pines, demonstrate the superior performance of the proposed method, particularly in high noise ratios, where substantial improvements in classification accuracy are observed. The code is available at https://github.com/kovelxyz/RSC.