AI assessment of structural fuel load and fire risk via street house images in wildland-urban interface

Abstract

Wildfires directly threaten human lives and properties in the wildland-urban interface (WUI). Combustible materials of house structure are easily ignited by sustained thermal radiation or fire spotting, so the estimation of structural fuel load is critical for accurate WUI fire spread modelling and risk assessment. This work presents a proof-of-concept framework for evaluating structural fuel load that integrates deep learning, remote sensing, and geographic datasets (house images of Google Street View and OpenStreetMap). A convolutional neural network (CNN) model was developed to automatically identify from images the construction materials (e.g., wood, vinyl, metal, concrete, brick) for the exterior walls and roof of residential structures. Remote sensing data was employed to estimate structural dimensions and structural fuel load via combining material properties of density, combustibility, and heat of combustion. The CNN model was trained and validated using real-world data of more than 6,000 structure-material pairs from the 2025 Palisades fire in California, sourced from the CAL FIRE Damage Inspection (DINS) database. The recognition accuracy of structure attributes exceeds 0.7. Additionally, a case study of the 2025 Eaton fire in California was implemented, where the AI-enabled fuel load estimations showed methodological and feasibility and consistency with those derived from the DINS database under consistent assumptions. Overall, our approach advances a proof-of-concept approach for pre-fire hazard assessment by efficiently providing approximate fuel load information under simplified assumptions, which is crucial for fire spread modelling, disaster prevention, and fire emergency response in WUI communities.