6 February 2023 · Kahramanmaraş earthquakes
A model compares each building's satellite image from before and after the earthquakes and estimates how badly it was damaged. It was trained on five cities and is tested here on Antakya, a city it never saw during training or tuning.
Each building is cut out of a pre-event image (BRIGHT, Maxar) and the earliest usable post-event Maxar image, shifted onto the same grid and co-registered. Drag the line to compare. The outline is the building footprint the model is asked about.
Every dot is a building in the held-out test city, coloured by the class the model predicts. Switch to the expert labels to see where it agrees and where it does not.
The same model, run the way it would run on the day: building footprints made before the earthquake (Microsoft's machine-made footprints), the Maxar scene released five days after it, no labels. The result is a ranking. Below are its first 500 buildings, coloured by what the experts later found there.
Share of all destroyed buildings found by checking the buildings in ranked order, for the response pipeline, the evaluation pipeline and a random order.
Macro-F1 averages the F1 score of the three classes, so the rare damaged and destroyed buildings count as much as the many intact ones. "Top 10 %": share of destroyed buildings found by inspecting the 10 % of buildings the method ranks highest. ECE: calibration error of the probabilities after temperature scaling (lower is better).