Semi-Supervised Semantic Change Detection Based on Weak-to-Strong Consistency

Deep supervised semantic change detection (SCD) methods require extensive annotated data, which is challenging and expensive, while the abundance of unlabeled multi-temporal remote sensing images offers a potential alternative. In this paper, we attempt to transfer semi-supervised methods for semantic segmentation to the task of remote sensing semantic change detection, leveraging information from unlabeled dual-temporal images to explore the feasibility of semi-supervised detection methods for SCD tasks. In this work, we conduct semi-supervised semantic change detection guided by the strong-weak consistency principle. Our experiments conducted on publicly available SCD datasets reveal that a more extensive perturbation space exhibits limited generalization capabilities in our scenario. More future work can be inspired by our experimental results and the transfer of UniMatch.