Semantic-centric change detection framework: considering spatiotemporal heterogeneity and spatiotemporal correlation of land cover

The semantic change detection (SCD) task aims to detect ‘from-to’ change types of land cover occurring over time in registered bitemporal remote sensing images. The spatiotemporal heterogeneity of land cover introduces significant intra-class variance, whereas spatiotemporal correlation provides a reference for land cover change. Current multi-task SCD methods achieve encouraging results by enforcing consistency between semantic segmentation and binary change detection; however, they fall short in exploring the geological priors of land cover. In this study, we propose a Semantic-centric Change Detection framework (S-cCDNet) to model ‘from-to’ semantic changes by considering two geographical priors: spatiotemporal heterogeneity and spatiotemporal correlation. To overcome spatiotemporal heterogeneity, we design an intra-temporal Pixel-to-Prototype (Pix2Pro) and inter-temporal Prototype-to-Prototype (Pro2Pro) contrast strategy to promote intra-class compactness and inter-class separability. To enhance spatiotemporal correlation, we introduce a Spatiotemporal Correlation Transformer (STCformer) to progressively model the spatiotemporal correlation of semantic changes from the scene to pixel levels. In addition, a direction-agnostic SCD loss is proposed as a regularization term to prevent overfitting of semantic change directions. Comprehensive experiments conducted on two large-scale SCD datasets with different scales and scenes demonstrated the superiority of the proposed S-cCDNet.