Two-Stage Machine Learning Framework for Predicting Building-Level Direct Economic Losses due to Flood Hazard
Conference
Schoolcraft, MA, Gupta, HS, González, AD et al. (2025). Two-Stage Machine Learning Framework for Predicting Building-Level Direct Economic Losses due to Flood Hazard
. 1236-1241. 10.21872/2025IISE_6768
Schoolcraft, MA, Gupta, HS, González, AD et al. (2025). Two-Stage Machine Learning Framework for Predicting Building-Level Direct Economic Losses due to Flood Hazard
. 1236-1241. 10.21872/2025IISE_6768
Flooding poses significant challenges to community resilience, necessitating better predictions of economic losses to support mitigation and recovery planning. Better building-level loss estimates are beneficial to enable effective mitigation and recovery planning. Macro-level flood loss predictions dominate many studies, often overlooking the granularity necessary for accurate building-specific loss assessments. The absence of reliable, micro-level damage predictions hinders resource allocation and policy decisions, especially for vulnerable communities. This study addresses the gaps by developing a machine learning approach to predict direct economic losses at the building level, focusing on Lumberton, North Carolina, a community severely impacted by consequent hurricanes: Hurricanes Matthew (2016) and Florence (2018). A two-stage approach was implemented: a classification model identified buildings with flood-related damages, followed by a regression model estimating the loss amount for affected structures. Both models utilized the XGBoost algorithm, with direct variable inputs outperforming principal component-based inputs in accuracy and interpretability. Key predictors included first-floor elevation, building value, and location. The study highlights limitations, including reliance on a single dataset and exclusion of social vulnerability metrics, which constrain model generalizability. Future research should integrate diverse datasets, social vulnerability factors, and uncertainty quantification to improve predictive capabilities. The findings underscore the potential of machine learning to enhance flood loss prediction and inform policy decisions, particularly for pre-hazard resource allocation and disaster resilience planning.
