Data-driven optimization for efficient post-hazard housing reoccupation

Abstract

Flooding poses significant challenges to community resilience, highlighting the need for accurate, building-level economic loss estimates and equitable recovery planning. This thesis addresses limitations of macro-scale flood loss models by developing integrated, data-driven frameworks for predicting flood impacts and allocating post-hazard recovery resources. The first component presents a machine learning framework to estimate direct building-level flood losses, using a two-stage XGBoost model to identify damaged structures and predict associated economic losses. Applied to Lumberton, North Carolina, following Hurricanes Matthew and Florence, results show that direct physical and locational variables—including first-floor elevation, building value, and location—yield accurate and interpretable loss estimates. The second component develops a stochastic, multi-objective optimization model for post-hazard housing recovery that incorporates household-level social vulnerability and uncertainty in resource effectiveness. Using Markov chain transition dynamics and empirical data from 1,729 households, optimized allocations reduce emergency shelter reliance and accelerate permanent housing reoccupation while revealing clear efficiency–equity trade-offs. Together, these frameworks form a compatible decision-support pipeline that enables timely, equitable, and effective flood recovery planning under uncertainty.