Deep Metric Learning for Damage Detection Using Bitemporal Satellite Images

In this article, we propose a novel deep metric learning approach to multi-hazard damage detection in remote-sensing (RS) images. The model is realized through a triplet network, which addresses the challenge of detecting disaster-induced damage with high accuracy for built-objects using pre- and post-event RS data. Previous efforts using a machine learning methodology do not sufficiently address the significant inter- and intra-class variations amid the disaster-scene complexities in built-environment RS images. Through this model, numerical evaluation is conducted based on wind storm disasters, and high performance is observed.