Abstract: [Purposes] Semantic change detection aims to identify and distinguish different types of land surface changes from multi-temporal remote sensing images，holding significant application value in fields such as urban planning，environmental monitoring，and disaster assessment.However，traditional supervised methods heavily rely on pixel-level annotated data，which severely limits their practical deployment.In recent years，self-supervised learning has provided a promising approach to mitigate annotation dependency by leveraging large-scale unlabeled data for pre-training. [Methods] To this end，this paper proposes a self-supervised framework termed GLSNet for semantic change detection tasks.Built upon an encoder-decoder architecture，GLSNet incorporates joint global multi-scale contrastive learning and local feature matching contrastive learning during the pre-training stage to learn semantically consistent representations across temporal phases. [Findings] Experiments on the SECOND and Hi-UCD mini dataset demonstrate that GLSNet achieves superior performance in semantic change detection，particularly exhibiting significant advantages over existing methods in few-shot scenarios. [Conclusions] These results indicate that the proposed approach offers an efficient and practical solution for remote sensing semantic change detection．

Key words: semantic change detection, self-supervised learning, limited annotations, global multi-scale, local matching