Predictive performance of wildfire structure exposure modeling in the western US

Tyler Hoecker, Matthew Patrick Thompson , Bryce Young, Joe Scott, Chris Moran, Kayleigh Wilson, Jeremy Arkin, Michael Callahan

As development expands and extreme wildfire events worsen, people and property are increasingly exposed to wildfire hazards, necessitating comprehensive and accurate community and structure risk assessment tools. A variety of modeling frameworks and indices have been applied to support community wildfire planning, but few have been evaluated against subsequent fire events, which is a key gap for advancing the science and practice of risk management. Here we evaluate the performance of the Structure Exposure Score (SES) at predicting subsequent structure exposure. Specifically we analyze a SES layer current to 2017 landscape conditions and evaluate predictive performance over the western US across 2017-2022. We describe SES and evaluate its ability to discriminate wildfire-exposed from unexposed structures relative to a random classifier using an AUC-ROC approach. Results show that the mean SES value of exposed structures was 78.95% higher than unexposed structures. We report a mean per-fire AUC of 0.83, and an overall AUC of 0.76; the lower overall AUC is attributed to challenges predicting exposure from wildlands when events transition to urban conflagrations. Burn probability alone performed similarly to SES, suggesting it is sufficient where general exposure patterns are the focus, and SES more appropriate where the magnitude of exposure matters. Ember load was also an important predictor, especially on a per-fire basis (AUC = 0.76), consistent with the known role of ember transport in driving structure exposure in WUI fire environments. This work provides a foundational benchmark for future exposure prediction and assessment.