Post-Disaster Damage Assessment with Artificial Intelligence and Uncertainty Quantification

Abstract

Climate disasters have severely impacted our built environment, infrastructure, and way of life, causing significant economic and fatal losses. In the immediate aftermath of a disaster, preliminary damage assessment (PDA) is one of the critical first steps in estimating the scale of damage and seeking a major disaster declaration by providing pivotal information and narratives of disaster impacts. The process of PDA involves assigning damage measurements to buildings, facilities, and infrastructures based on functionality and visible damage to the structure. Traditionally, human teams are deployed to the field to thoroughly assess damages to critical structural components and evaluate the extent of damage to structures. However, the human-based PDA process is time-consuming, labor-intensive, unsafe, and in many cases, subject to bias and error. Besides, a successful ground-based PDA requires uninterrupted access to the affected disaster site, which may not be possible due to the presence of hazardous materials and debris. To address this challenge, the tradeoff between speed, accuracy, and safety motivates new practices that leverage remote sensing and artificial intelligence (AI) technologies. To date, however, there is only a limited body of work on PDAs using AI and remote sensing data. Moreover, data-driven AI can introduce new sources of bias since training data is often limited and labeled by a small group of human experts whose damage annotations may not be entirely reliable or in agreement with one another. Inaccurate prediction of disaster damage by AI models can mislead the allocation of recovery and rebuild resources and government assistance funds. In such scenarios, crowdsourcing approaches using social networking systems are usually expected to perform more reliably by scaling up the data collection and annotation process, thus expanding the scope and certainty of AI inferences. However, fully crowd-based damage assessment is also quite expensive and slow due to the large number and diversity of input information. To this end, this dissertation aims at developing a human-AI partnership system for automating large-scale damage assessment using remote sensing and crowdsourcing technologies. The dissertation leads to three significant contributions: (a) it creates a novel vision-based deep learning model for building damage assessment in accordance with Federal Emergency Management Agency (FEMA) guidelines and for debris estimation; (b) it quantifies the uncertainty for AI-assisted post-disaster damage assessment for more trustworthy and explainable AI-assisted damage assessment, and (c) it reduces the uncertainty and enhances the consistency of damage assessment through a proposed probabilistic human-AI partnership framework leveraging the strength of both humans and machines.