Large-Scale Polygonal Building Mapping From 3–5 m Medium-Resolution Satellite Imagery

Abstract

Mapping buildings in vectorized form from remote sensing imagery is a key task for digitalizing and monitoring our planet. While most existing methods rely on sub-meter-resolution aerial or satellite imagery to extract fine-grained building polygons, these sources suffer from long revisit intervals, high cost, and incomplete coverage, limiting their scalability. This raises a fundamental question: can we enable global-scale building polygon mapping using only medium-resolution (3–5 m) satellite imagery, while maintaining acceptable geometric quality? In this article, we explore the feasibility of large-scale polygonal building mapping using medium-resolution (approximately 4.77 m at the equator) PlanetScope imagery. Our goal is not to surpass high-resolution methods in geometric precision, but to achieve substantially improved coverage and update frequency with a favorable accuracy–scalability trade-off. To this end, we present a complete and scalable pipeline encompassing global data acquisition, automated annotation, novel algorithmic components, and an efficient inference framework tailored for real-world deployment. To address the common issue of small, adjacent buildings merging during binarization, we introduce peak-based clustering (PBC) for robust instance separation. In addition, we propose a point feature sampling (PFS) strategy that extends the global collinearity-aware polygonization (GCP) method to efficiently generate regularized building polygons from large-scale satellite mosaics. Extensive experiments show that our approach achieves substantially higher completeness than existing high-resolution-based products, with only minimal accuracy trade-offs. These results highlight the potential of our method to significantly enhance global building footprint coverage. The code and data are publicly released under: https://github.com/zhu-xlab/MRPolyBuild.