Precise Building Boundary Extraction Using Deep Learning Method Lin Y. H.1, Tsai F.2*
1Student, Department of Civil Engineering, National Central University, Taiwan
2*Professor, Center for Space and Remote Sensing Research, National Central University, Taiwan
0123kan[at]gmail.com, *ftsai[at]csrsr.ncu.edu.tw
Abstract
Buildings stand out as the most important elements in urban areas, and their morphology and spatial characteristics are distinctly captured in high-resolution satellite imagery. Although LiDAR point clouds enable more precise 3D reconstructions of urban structures, their high cost and lack of real-time availability impose significant constraints. In contrast, satellite images offer shorter revisit intervals and more cost-effective coverage. Deep learning methods introduce an innovative, automated pipeline for interpreting remote sensing data. Among these approaches, Mask R-CNN has proven effective at extracting building boundaries from high-resolution satellite imagery. Recent studies have emphasized producing smoother, more consistent boundary geometries and strengthening the robustness of network across diverse urban scenes. In this study, we used Mask R-CNN as the base method for feature extraction. We experimented with different datasets offering varied spatial characteristics, aiming to fine-tune the model for high-resolution satellite imagery in Taiwan and integrate a boundary-regularization module to produce cleaner, sharper building boundaries. Future work will focus on developing a fully automated framework tailored for Taiwan, aiming to achieve higher accuracy in building footprint extraction and Level-of-Detail (LoD)-2 building modeling.
Keywords: Building boundary extraction, Deep learning, Mask R-CNN
