Adversarial Constraints Guide Multiview Learning for Semi-Supervised Change Detection

Semi-supervised learning (SSL)has emerged as a powerful paradigm to mitigate the challenge of limited labeled data, and is gaining traction inremote sensing change detection (RSCD). Traditional SSL methods typically rely on pseudo-labeling coupled with consistency learning to exploit unlabeled data; however, performance is critically dependent on the quality of the pseudo-labels. To address this issue, we propose a novel pseudo-label learning framework for RSCD that introduces an error information adversarial constraint to enhance consistency learning. This constraint explicitly maximizes the discrepancy between unlabeled data and the downstream task while minimizing the divergence for labeled data, thereby dynamically correcting suboptimal pseudo-labels through adversarial error feedback. Building on this constraint, we further develop aMulti-View Joint Consistency Learning (MV-JCL)strategy to refine pseudo-label predictions. By aggregating perturbed data from multiple randomized complementary views, MV-JCL significantly enhances the model’s generalization capability. Extensive experiments validate the superiority of our method. Under the 5% labeled data setting, it achieves performance gains of 3.83%, 1.97%, and 1.85% on the LEVIR-CD, WHU-CD, and cross-dataset, respectively, surpassing state-of-the-art approaches. These results not only demonstrate the effectiveness of the framework, but also establish it as a promising solution for label-efficient change detection in remote sensing applications.