Adaptive Structure-Aware Connectivity-Preserving Loss for Improved Road Segmentation in Remote Sensing Images

Accurate road extraction from remote sensing images is a critical step for many applications, including navigation, environmental monitoring, and disaster response. In remote sensing images, roads appear as thin curvilinear structures and are very prone to fragmentation due to challenges such as low-contrast materials and occlusions by surrounding vegetation or buildings. Although deep learning has greatly advanced image segmentation techniques, pixel-level loss functions are insufficient for preserving connectivity and topology of thin curvilinear structures such as roads. Small pixel-level errors can disrupt the extraction of accurate road graphs, complicating subsequent characterizations and analysis tasks. In this paper, we propose a novel, adaptive, structure-aware, connectivity-preserving loss function SAC-Loss. This loss function combines global and local processing to improve structure awareness and uses a proximity-based weighting scheme with asymmetric penalties to fill gaps in the road structures while limiting spurious detections elsewhere. Experimental results demonstrate improved road segmentation in terms of visual quality and quantitative performance.