A post-hurricane building debris estimation workflow enabled by uncertainty-aware AI and crowdsourcing

Abstract

Climate disasters often result in large amounts of debris that need to be cleaned up in the event's aftermath. Effective post-disaster debris management poses unique challenges due to limited resources, high expenses, and the nonhomogeneous and hazardous debris content. Current practices of debris cleanup rely on time-consuming and error-prone field-based estimation, which impedes immediate response and accurate analysis. While new approaches such as artificial intelligence (AI) and crowdsourcing have shown promise in various domains, their potential for debris estimation remains underexplored. This study leverages a human-AI teaming workflow that estimates post-disaster debris volume and composition. The workflow begins with the utilization of drones to capture aerial views for efficient disaster infrastructure reconnaissance. Subsequently, an AI model and a crowdsourcing module are calibrated and work together to rapidly and reliably detect damaged buildings and classify the levels of damage based on aerial imagery. By establishing a connection between building damage states and the resulting generated debris, these damage level predictions serve as a foundation for the efficient and accurate estimation of debris. Overall, the proposed approach uses drones for rapid data collection and AI to automate damage detection and classification, which reduces the time required for debris estimation from days to a few hours as well as minimizes predictive uncertainty with crowdsourcing by up to 40%. A case study on Hurricane Laura in 2020 was conducted to validate the proposed approach. Findings of this study lead to predicting debris volume and composition with minimal errors, particularly for partially damaged buildings.