SS-CAN: Semi-Supervised Contour Alignment Network for Building Footprint Correction in Off-Nadir Satellite Imagery

Accurate building footprint extraction from high resolution remote sensing imagery is essential for urban planning and GIS data production. However, off-nadir satellite images introduce significant challenges: high-rise buildings exhibit substantial relief displacement, where the visible rooftop appears as a “false” footprint that deviates from the actual ground base. Existing segmentation methods typically fail to correct this geometric offset, while supervised correction approaches are constrained by the difficulty of acquiring strictly paired annotations linking rooftops to bases. To address this challenge, we propose a Semi-Supervised Contour Alignment Network (SS CAN) designed for unpaired data scenarios. Our framework comprises three synergistic modules: (1) a Semi-Supervised Base Extraction module that transfers knowledge from sparse labels to large-scale unlabeled data using Consistent Teacher and Mean Teacher networks; (2) a Combinatorial Matching mechanism that associates predicted rooftops with potential base targets; and (3) a Shape-Similarity Alignment module that rectifies geometric offsets by maximizing morphological overlap. Experimental results on the self-constructed Wenzhou Building Dataset demonstrate that our method achieves an IoU of 59.68% and a Precision of 78.19%, significantly outperforming state-of-the-art baselines such as Swin Transformer by effectively decoupling visual segmentation from geometric correction.