Remote sensing semantic change detection via semantic editing-driven change synthesis

Semantic change detection (SCD) typically demands extensive semantic labels and laborious annotations, while models trained on synthetic data often generalize poorly to real scenes due to domain shift. To address these issues, we propose the Remote Sensing Semantic Change Detection via Semantic Editing-driven Change Synthesis (SECS-SCD). Change events are modeled as Markov-driven edits to semantic masks, and pseudo-change pairs are generated by a mask-conditioned latent diffusion module with spatially aligned cross-attention and coordinate-based positional encodings. A mask-sampling prior preserves unchanged regions via reference injection. Joint training on synthetic and limited real data, regularized by multi-kernel MMD, aligns latent representations and mitigates multi-temporal bias, thereby reducing dependence on strictly registered dual-temporal labels and achieving robust performance on real SCD datasets.