Temporal Disaggregation of Building Footprints using Sentinel-2

Kurzfassung

High-resolution building footprint datasets from sources like Overture Maps (OV), Google, Meta and OSM, lack temporal information, limiting their use for dynamic monitoring tasks. This study addresses this gap by proposing a method to temporally disaggregate static building footprints using Sentinel-2 imagery. A Bayesian U-Net (BCNN) segmentation model is employed to assign year-specific labels while also estimating prediction uncertainty. The approach enables time-aware, uncertainty-informed building maps at scale.