Multi-Scale AI for Wildfire Risk Assessment: Integrating Regional Spread Mapping with Computer Vision Hazard Recognition

Abstract

With the increasing frequency of wildfires in Wildland-Urban Interface (WUI) regions, current risk management tools require residents to manually cross-reference regional fire forecasts, evacuation maps, and property-level mitigation guidelines during an active threat. To bridge this gap, we developed an automated home risk assessment system that identifies wildfire risks using image recognition and is supported by spread prediction and evacuation routing features. The application integrates ForeFire’s Rothermel Simulation Engine trained on multi-source geospatial, meteorological, and historical fire data to forecast wildfire propagation in near real-time across the US and Canada. By integrating predictive fire perimeter outputs with Mapbox’s live-traffic routing, the platform provides ETA’s of the fastest routes that minimize exposure to high-risk zones. Across 23 perimeter constrained tests, where the tests were run along the edge of active fire zones, 87% of all generated evacuation routes had a median clearance of 17.5 km from active wildfires. While the macro spread modeling applies broadly across North American WUI regions, at the micro scale, the vision module evaluates user-uploaded property imagery using a prompt-engineered, zero-shot Qwen2.5-VL-3B-Instruct model. Our framework uses California Public Resources Code (PRC) 4291 and California Assembly Bill 3074 (Zone 0 Ember-Resistant Mandate) standards as a benchmark to assess property compliance across Home Ignition Zone hazard classifications; fire hazards were assessed under Title 14, California Code of Regulations (CCR) 1299.03. We added an RGB matrix color filtering layer to validate vegetation dryness and filter hallucinated detections. We show how AI can bridge the gap between fire modeling and local user safety, with the ultimate objective of reducing response times and improving evacuation outcomes.