关键词: 遥感影像, 建筑物足迹, 语义分割, 域适应, segment anything model(SAM)

Abstract: To address the substantial decline in generalization performance of deep learning models for building footprint extraction across different platforms and spatial resolutions caused by inter-domain distribution inconsistency,this study introduces a cross-scale geometric-refined network (CSG-Net) aimed at enhancing both adaptability and extraction accuracy in unlabeled target domains.The proposed approach establishes a cascaded probabilistic-geometric pseudo-label refinement framework.Firstly,pseudo-label uncertainty is quantified by computing the Jensen-Shannon divergence (JSD) between the dual prediction branches of the model,and a probabilistic weighting scheme is applied to suppress noise in unreliable regions.Subsequently,geometric priors derived from the segmentation results of the segment anything model (SAM) are incorporated to perform explicit geometry-constrained refinement of pseudo-label boundaries based on overlap ratio analysis,thereby generating high-quality training targets.Extensive experiments on challenging cross-scale building extraction tasks demonstrate that CSG-Net achieves an intersection over union (IoU) of 73.05%,markedly surpassing the Baseline (52.49%) and outperforming other state-of-the-art domain adaptation approaches.These findings confirm the effectiveness of the proposed framework in improving cross-domain robustness and extraction accuracy.

Key words: remote sensing imagery, building footprint, semantic segmentation, domain adaptation, segment anything model (SAM)