Three-dimensional time series building reconstruction framework in global south based on openly available satellite data

Rapid urbanization necessitates accurate 3D building data for effective urban planning and analysis. Building height provides critical vertical information reflecting urban morphology, land use intensity, and energy demand. However, high-resolution large-scale 3D datasets remain limited, particularly in the Global South, due to the high cost and complexity of traditional methods. Although machine learning approaches have been widely explored, they often underperform in dense urban areas and tend to underestimate tall buildings due to limited training data and generalizability. In this study, we propose a novel framework for large-scale building height estimation using only free remote sensing data. Leveraging 0.5 m resolution open-access satellite imagery and ICESat-2 photons, we construct an accurate parallel projection model for each image. It enables the generation of dense height points via triangulation across image pairs without additional geometric parameters. The height points are then integrated with 2D building footprints to reconstruct building height maps. Validation results of the full urban area in Nairobi, achieving a root mean square error (RMSE) of 3.338 m, demonstrated the feasibility of our framework. The method also exhibits strong temporal consistency, with a maximum mean deviation of only 1.93 m across multi-temporal height maps. Experiment results in the three additional Global South cities (Medellín, Salvador and Jakarta), achieving mean absolute errors (MAE) of 3.867 m, 3.642 m, and 2.484 m, respectively, further confirmed the transferability of our framework. These results highlight our framework’s capability to deliver low-cost, accurate, and high-resolution 3D urban reconstruction, particularly in resource-constrained cities, providing a scalable tool for urban analysis, planning, and policy support.