Multihazard Disaster Damage Classification for Buildings in Optical Satellite Images

While deep learning methods have significantly advanced remote sensing-based disaster damage detection, the challenge of reliably classifying structural damage levels across multiple hazard types using optical satellite imagery remains unresolved. This paper addresses this gap by proposing a triplet deep metric network for bitemporal remote sensing damage detection. The model, which combines a triplet network structure, ResNet50-based feature extraction, and an enhanced triplet loss, is designed to effectively distinguish building damage levels under bidirectional uncertainties. A systematic experimental approach is employed, utilizing the publicly available xBD data set and hazard-specific disaster subsets (wind, flooding, and fire). Our experimental results confirm the viability of a unified vision-based damage detection framework, particularly in distinguishing between structures at opposite ends of the damage spectrum (i.e., no damage and destroyed). These findings provide insights into advancing remote sensing–based methodologies, enabling the scalable and rapid assessment of damage across multihazard natural disasters.