AI for Disaster Rapid Damage Assessment from Microblogs

Formal response organizations perform rapid damage assessments after natural and human-induced disasters to measure the extent of infrastructures such as roads, bridges, buildings. This time-critical task, when performed using traditional approaches experts surveying disaster areas, poses serious challenges delays response. paper presents an AI-based system that leverages citizen science collect images reported on social media assessment in real-time. Several image processing models tackle non-trivial posed by a data source, high-volume redundant irrelevant content. The determines severity state-of-the-art computer vision model. Together with organization US, we deployed identify reports during major real-world disaster. We observe almost 42% are unique, 28% relevant, more importantly, only 10% them contain either mild or severe damage. Experts from our partner provided feedback system's mistakes, which used additional experiments retrain models. Consequently, retrained based expert target domain helped us achieve significant performance improvements.