Deep Metric Learning for Disaster Damage Classification in Remote Sensing Images

Using space-borne images and their time series for disaster-induced damage detection and mapping has been a classic topic since the advent of remote sensing (RS) technology. The traditional methods usually start with a change detection step that extracts structural changes followed by a pattern classification model. By exploiting recent advances in deep learning, the authors developed a novel metric learning-based damage classification model, termed Enhanced Triplet Network for Damage Detection (EnTriNet-DD). The model enables a dual-phase training and learning of inter- and intra-class variances of damage levels in bi-temporal image pair for application in remote sensing-based disaster damage mapping. Compared with other existing damage-learning methods in the literature using remote sensing images, this model first model inter- and intra-class variances explicitly, significantly improving learning efficiency, damage-classification accuracy, and robustness to uncertainties. With this preliminary research, we argue that in the advent of remote sensing big data, deep-learning and RS-based damage mapping can contribute to the lifecycle management of disasters both for lifeline systems and communities.