Manuel Huber (German Aerospace Center, DLR)
Accurate information about where buildings are located is a key input for many disaster risk applications. It helps estimate the number of exposed people and assets and supports planning before, during, and after hazardous events. While several global building footprint datasets are now available, most represent a single, often unknown, point in time. As settlements continue to grow, these static datasets become increasingly outdated, making it difficult to reconstruct historical exposure or analyse how the built environment has changed over time.
Our recent study, Making Footprints Move: Temporal Disaggregation of Building Footprint Data Using Sentinel-2 Imagery and Bayesian Deep Learning, addresses this limitation by combining freely available Sentinel-2 satellite imagery with deep learning to estimate when buildings were constructed and generate annual high-resolution building footprint datasets. This is relevant within the context of the PARATUS project, where understanding how exposure evolves over time is an important component of improving dynamic disaster risk assessment. By adding a temporal dimension to existing building footprint data, this work supports urban development analysis and provides more temporally consistent exposure information for risk assessments.
Updating building footprint data using open satellite imagery
Our approach relies entirely on openly available datasets, combining multi-temporal Sentinel-2 imagery with building footprints from the Overture Maps Foundation. A Bayesian deep learning model is trained to predict whether a building was already present in a given year based on the corresponding satellite observations. These annual predictions are then used to temporally disaggregate the static building footprint dataset into yearly or sub-yearly building footprint maps.
The framework is based on a Bayesian U-Net architecture. In addition to producing annual footprint estimates, the model also provides an estimate of prediction uncertainty. Prediction quality is not uniform across all locations, as factors such as cloud cover, seasonal vegetation, or land-cover changes unrelated to urban development can reduce confidence. The uncertainty estimates therefore provide additional information that can support quality assessment, identify areas where manual validation may be beneficial, and help users interpret the generated datasets.
Looking ahead
This work demonstrates one possible approach to reconstruct historical building development using freely available earth observation data. Future work will investigate the integration of additional satellite data sources, improvements in uncertainty estimation, and methods that better adapt to regional differences in urban development. Within PARATUS, this contributes to ongoing efforts to improve the information available for multi-hazard risk assessment and disaster management
