Explainable deep learning powered building risk assessment model for proactive hurricane response.

Area of Science:

• Environmental Science
• Computer Science
• Civil Engineering

Background:

• Climate change and urban development exacerbate hurricane impacts on U.S. Southeastern coasts.
• Existing hurricane risk assessments lack building-level specificity and actionable insights for residents.
• Traditional methods struggle to model complex building-specific risk factors.

Purpose of the Study:

• To develop and validate a building-level hurricane risk assessment model using deep feedforward neural networks (DFNN).
• To integrate diverse data sources for accurate prediction of building damages from wind and surge hazards.
• To provide timely and interpretable risk information for coastal property owners.

Main Methods:

• Utilized deep feedforward neural networks (DFNN) for risk assessment.
• Incorporated building characteristics, meteorological, and hydrological data as input features.
• Employed Local Interpretable Model-agnostic Explanations (LIME) for model interpretability.

Main Results:

• The DFNN model accurately assessed building-level hurricane risk, including probability and intensity of wind and surge damage.
• Demonstrated timely risk assessment for a case building in Cameron County, Louisiana, using updating weather forecasts.
• Showcased the potential of AI in integrating multi-sourced data for predicting weather extreme risks.

Conclusions:

• Deep learning models offer a powerful approach for granular hurricane risk assessment.
• AI-driven insights can enhance coastal populations' understanding and response to extreme weather.
• This technology supports proactive risk mitigation and long-term climate adaptation strategies for property owners.