Deep learning-driven semantic change detection in remote sensing: a systematic review

With the rapid advancement of high-resolution remote sensing technology, Remote Sensing Semantic Change Detection (RS-SCD) has emerged as a core technology for monitoring dynamic Earth surfaces. RS-SCD enables simultaneous localization of changed regions and inference of their ‘From-To’ class transitions, overcoming the semantic ambiguity limitations of traditional binary change detection. In this systematic review, we first formalize the RS-SCD task and present a conceptual probabilistic perspective to elucidate the structural coupling between change detection and semantic segmentation. Moving beyond conventional architecture-centric surveys, we propose a task-oriented taxonomy based on five semantic-change coupling strategies: sequential decoupling (primarily CNN-based), parallel coupling (Transformer and Mamba architectures), relational coupling (graph neural networks), cross-modal fusion (optical, SAR, and vision-language modalities), and partial-evidence inference (weakly supervised, self-supervised, and foundation-model fine-tuning). For each strategy, we analyse representative models, highlighting trade-offs among accuracy, efficiency, and robustness. We further discuss practical bottlenecks, including pseudo-change suppression, annotation scarcity, and limited generalization, and outline future trends towards large-model-driven RS-SCD that is generalized, all-weather, and trustworthy. This review provides a systematic, conceptually grounded reference for both research and operational deployment of RS-SCD.