VegeSSL : A Multistage Contrastive Learning Framework Combining Self‐Supervised Representation and Label‐Guided Refinement for Segment‐Level Noisy Label Detection in Remote Sensing Data

ABSTRACT

Research on semantic segmentation has long been hindered by the presence of noisy labels in large‐scale datasets, which significantly degrade the performance of deep learning (DL) models. Although semi‐supervised and self‐supervised learning (SSL) approaches have alleviated data scarcity, the automatic identification of mislabeled segments in dense vegetation mapping remains a major challenge. To address this, we propose VegeSSL, a multistage contrastive learning framework that combines self‐supervised pixel‐level representation learning with label‐guided hard negative refinement and segment‐level anomaly detection. First, pixel‐level contrastive learning is employed using a ResNet‐18 backbone to extract robust feature representations. Second, a hard negative mining (HNM) strategy is applied to enhance the encoder's sensitivity to class boundaries. Third, a multi‐threshold isolation score (IS) is introduced, the core of VegeSSL which identifies anomalous labels based on neighborhood consistency in the learned embedding space. Among all thresholds, achieved an overall accuracy (OA) of 0.892 and a F 1‐score of 0.81. This OA exceeded that of the other thresholds, which yielded OA values of 0.72, 0.80, and 0.84 for , respectively. Consequently, was selected as the optimal threshold, and the error candidates generated under, were finalized as the corrected segments. The proposed approach was validated against three benchmark methods on two publicly available datasets: (a) LoveDA and (b) DeepGlobe, where VegeSSL demonstrated superior performance in identifying segment‐level mislabeled samples. By reducing reliance on extensive manual verification and providing a ranked list of potential labeling errors, VegeSSL offers a scalable and practical solution for improving labeled data quality in terrestrial monitoring and environmental management application.