Deep Learning-Based Damage Mapping With InSAR Coherence Time Series

Satellite remote sensing is playing an increasing role in the rapid mapping of damage after natural disasters. In particular, synthetic aperture radar (SAR) can image Earth's surface and map all weather conditions, day night. However, current SAR methods struggle to separate from other changes surface. this study, we propose a novel approach mapping, combining deep learning with full time history observations impacted region order detect anomalous variations properties due disaster. We quantify Earth change using series Interferometric coherence, then use recurrent neural network (RNN) as probabilistic anomaly detector on these coherence series. The RNN first trained pre-event series, forecasts probability distribution between pre- post-event images. difference forecast observed co-event provides measure confidence identification damage. method allows user choose detection threshold that customized for each location, based local behavior through before event. apply calculate estimates three earthquakes multi-year Sentinel-1 acquisitions. Our shows good agreement quantitative improvement compared loss proxy.