Data-Driven Assessment of Disaster Damage and Recovery Time

Abstract

Although natural hazards may sometimes be predictable, their occurrence is not preventable, especially in low-frequency-yet-high-impact events such as earthquakes and hurricanes. The catastrophic effects of natural hazards can vary vastly from year to year, depending on the seasons, locations, demographics, or resilience of the affected areas. Therefore, improving response system and recovery time is one of the most efficient ways to limit fatality and economic loss from a hazard event. Unfortunately, without being able to gain adequate situation awareness about the damage extent and the potential recovery, we cannot effectively improve these processes. This dissertation provides a suite of methodological frameworks utilizing statistical tools to aid in the damage assessment and estimation of various infrastructures' recovery to provide emergency managers and stakeholders with timely and extensive situation awareness after a hazard event. The initial step is to assess the actual damage extent immediately after a hazard event so that adequate planning and resources can be allocated. The first methodological framework aims to speed up the post-event damage assessment process. Instead of the more time-consuming and labor-intensive windshield survey method, machine learning algorithms are applied to automatically annotate the damaged and/or flooded buildings on satellite imagery. The annotation results can be used as a proxy for assessing how badly an area is affected. The machine learning algorithms require much less time and resources while still yielding results with reasonable accuracy. Secondly, to improve generalizability and accuracy of the previous damage assessment framework, a mixed data approach is adopted to combine satellite imagery and other geolocation features such as each building's elevation and proximity to water bodies. Finally, a recovery trajectory estimation framework is introduced to aid in recovery planning for critical infrastructures. The estimation will provide infrastructure management agencies with an idea of the most likely recovery pattern of various critical infrastructures (such as electricity, water, and gas), given different hazard scenarios. This will give them a quantitative assessment of how resilient their infrastructure systems are so that resources can be allocated and necessary investment can be informed effectively. Besides extensive results from numerical studies and empirical data, this dissertation research also contributes two curated datasets to open-access repositories so that others can reproduce and improve the proposed framework.