Deep Learning-based Damage Mapping with InSAR Coherence Time Series

Fast response in the aftermath of natural disasters is essential to minimize casualties. In order to assist effective use of limited response resources, rapid and accurate mapping of the extent and intensity of disaster-induced damage over tens to thousands of square kilometers is necessary. Satellite remote sensing is playing an increasingly important role in damage mapping due to improvements in data availability and spatiotemporal coverage. In particular, synthetic aperture radar (SAR) can image the Earth's surface in all-weather conditions, day and night, and has been used to map damage. Current SAR damage mapping methods rely on using changes in interferometric SAR (InSAR) coherence between two pairs of images before and spanning an earthquake but they depend on the particular choice of pre-event SAR image pair that sometimes suffers from temporal decorrelation from other changes in the Earth's surface.

Here, we propose a new method for damage mapping using a history of InSAR sequential coherence images of a region from a single satellite constellation. We train a recurrent neural network (RNN) on the pre-seismic coherence time series, then forecast the coherence image spanning the event. The difference between the RNN-based forecast and calculated coherence from the coseismic interferometric SAR image pair is used to map anomalous changes, inferred to be due to damage. We apply this method to calculate damage proxies for two earthquakes in Iran and Italy using multi-year time series of Sentinel-1 SAR acquisitions. We demonstrate that this method shows good qualitative and quantitative agreement with independent analyses based on visual interpretation of before-and-after high-resolution optical images and reduces the false detections compared to conventional damage mapping methods.