Earthquake damage mapping from Sentinel-1 modeling against coherence damage proxy

Mohamad El Moussawi 

This paper tests whether a supervised model that uses the same Sentinel-1 signal as the coherence damage proxy can rank earthquake
damage better than the proxy itself. Damage grades from Copernicus EMS for four earthquakes are joined to 783 635 OpenStreet-
Map footprints over 28 assessed areas, and the coherence drop is aggregated to buildings, morphological blocks and cells of 50 m
to 400 m. A ladder of models, from a logistic regression to gradient boosting, a random forest and an RBF SVM, is compared with
the proxy on identical folds, using the held-out area as the replicate. Under leave-one-area-out validation, modeling improves on
the proxy for blocks, where the two boosting models beat it in 13 and 15 of 15 areas, significant after correction for multiple testing.
For buildings the best model gains in 12 of 15 areas, which is not significant after correction, and for cells no model improves on
the proxy, with the flexible models tending to perform worse. The logistic model is the most reliable one. It is never significantly
worse than the proxy and its gain changes little between validation schemes. Only 2 % to 7 % of its weight is on the mean coherence
drop used by the proxy, while 79 % to 87 % is on the distribution of the drop inside the unit and on its neighbourhood. These results
hold within one event. The cross-event gains depend on assumed negatives and should be read as upper bounds.