TopoFormer: A Topology-Aware UNetFormer for Surface-Structure Segmentation in Remote Sensing Imagery toward Traffic Perception

Leveraging structural information to guide segmentation improves performance and enhances applicability in intelligent transportation systems (ITSs). Existing methods often treat structural information as a post-processing constraint and lack differentiated modeling of structure-sensitive regions. To this end, we propose TopoFormer, a topology-aware semantic segmentation model. Within this model, a lightweight auxiliary branch, jointly trained with the UNetFormer backbone, generates four online structural representations: road skeletons, connectivity, road–building adjacency, and building boundaries. These representations are then injected into each decoder stage via channel-spatial attention, enabling topological information to directly guide feature learning. Building on this guidance, reliability weights derived from the auxiliary branch’s prediction confidence drive a differentiated loss that imposes stronger supervision on structurally critical yet uncertain regions, such as skeleton endpoints, narrow road segments, and low-confidence areas. Comprehensive experiments and analyses demonstrate TopoFormer’s strong performance.