Fig. 5: Overview of the machine learning framework.
From: An open-source tool for mapping war destruction at scale in Ukraine using Sentinel-1 time series
For training (left), we use per-pixel Sentinel-1 time series extracted at the location of UNOSAT point annotations. The model is fed with a pair of time series from the same location. The first one spans a fixed 12-month time interval T0 from 2020, and the second one spans one of the 3-month time intervals Tn between 2021 and 2023. Both time series are encoded with a custom features extractor, and damage labels are dynamically assigned according to Tn. At inference time (center), the model generates a damage probability heatmap valid at Tn and spanning the entire country, aggregating the predictions of different Sentinel-1 orbits. The raw damage probabilities are intersected with building footprints. For the final map the estimates for different time intervals Tn are thresholded and aggregated.
