BTSR-Net: Boundary-topology guided structural refinement for high-resolution remote sensing image segmentation

Semantic segmentation of high-resolution remote sensing imagery in complex urban scenes often suffers from insufficient modeling of boundary neighborhoods, causing ambiguity near class interfaces and weakening the continuity of elongated structures. To address this issue, we propose a Boundary-Topology-Guided Structural Refinement Network (BTSR-Net). Built upon HRNet-OCR, BTSR-Net introduces a Geometric Boundary-Distance Branch (GBD) with joint boundary-distance supervision to capture spatial transitions around class boundaries, a Multi-scale Topology-Guided Structural Fusion (MTSF) module to enhance connectivity representation of elongated structures, and a dual-stream structural refinement mechanism to perform local residual correction within a boundary-distance-defined refinement narrow-band. Experiments on the ISPRS Potsdam and Vaihingen datasets demonstrate improvements in overall segmentation accuracy, boundary quality, and structural connectivity. On Potsdam, BTSR-Net achieves 84.70% mIoU, 43.17% mBIoU, and 81.63% clDice, exceeding the baseline by 2.91, 5.66, and 3.74 percentage points, respectively. On Vaihingen, it achieves 93.94% mIoU, 73.90% mBIoU, and 95.46% clDice, with corresponding improvements of 2.27, 9.39, and 2.12 percentage points. These results demonstrate the effectiveness of BTSR-Net for structure-sensitive semantic segmentation in high-resolution remote sensing imagery.