Making footprints move: Temporal disaggregation of building footprint data using Sentinel-2 imagery and Bayesian deep learning

Abstract

High-resolution building footprints from Overture, Google, Meta, and OpenStreetMap are essential for urban and environmental studies. However, these datasets often lack temporal metadata, limiting their utility for applications, that require spatio-temporal information, such as risk assessment, population estimation, and urbanization monitoring. This study presents a novel method for temporally disaggregating static building footprints by leveraging Sentinel-2 satellite imagery and a Bayesian U-Net segmentation model. The approach allows the assignment of time labels to individual building footprints and probabilistic uncertainty estimation. Beyond temporal labeling, disaggregation greatly improves label quality, boosting from −0.10 to 0.80 for building count and from 0.74 to 0.89 for built-up area accuracy. Overall, the method robustly generalizes, enabling flexible temporal disaggregation of high-resolution building footprints with uncertainty estimates that indicate prediction trustworthiness.