AI Meets the Eye of the Storm: Machine Learning-Driven Insights for Hurricane Damage Risk Assessment in Florida

Abstract

Abstract Due to Florida’s exposure to hurricanes originating from both the Atlantic Ocean and the Gulf of Mexico, hurricane risk assessments serve as a critical tool for mitigating potential impacts. This is the first novel study to develop a machine learning based risk assessment for hurricane induced flood damage, which demonstrates the potential of granular building level insurance data from 1985 to 2024, enriched with remote sensing derived variables. The stacked ensemble machine learning model predicted hurricane flood damage with an MAE of 11.3% at a granular ZIP Code Tabulation Area level (ZCTA). The model’s explainability tools determined that building property value was a significant predictor of hurricane damage, as it correlated with property size, complex architectural design, and proximity to waterfront locations, all of which affect potential repair costs. Other predictive factors include construction year, occupancy type, and flood zone designation. Partial dependency plots (PDPs) identified that northwest Florida is particularly susceptible to hurricane damage, attributed to the Gulf of Mexico’s warm and shallow waters than eastern Florida’s cooler Atlantic conditions and steep ocean floor. Miami’s significant coastal urbanisation, rendered it a hotspot despite southeast Florida’s overall low hurricane risk. Similarly Jacksonville in north-eastern Florida was identified as a hotspot due to compounded flooding from storm surge and nearby river systems. Partial dependency plots also quantified the significant positive impact of 1970s building code regulation. Future studies should examine coastal morphology, landfall angle, and proximity to barrier islands. A study limitation is that insurance data may be an imperfect representation of Florida, due to underinsurance and inability to afford insurance.

Topics & Concepts