Jeremy Kauwe
February 5, 2025
This abstract summarizes a long-run evaluation of the accuracy, classification performance, and threshold validity of Oregon’s Wildfire Hazard Risk Map using 21 years of wildfire data. The findings reveal that the model systematically overestimates burn probability, inflating long-term fire likelihood by approximately 78%. Additionally, within its highest risk category, the model overestimates extreme fire risk by 74%, indicating a persistent pattern of overprediction across risk levels.
Most critically, the model fails to distinguish between fire-prone and non-fire zones over a long history of data. A robust wildfire risk model should indicate that areas classified as high hazard have historically experienced more fire than low-risk areas. However, this model does not meet that expectation. Instead, its classification performance is statistically indistinguishable from that of a random number generator, meaning it offers no meaningful ability to separate areas that have burned from those that have not, undermining its utility as a predictive tool.
Compounding this issue, the model’s hazard classification thresholds lead to extreme overclassification. The high hazard threshold (0.13787), originally set at the 90th percentile within the wildland-urban interface (WUI), corresponds to the 40th percentile when applied to the full dataset. This means that nearly 60% of tax lots are classified as high hazard instead of the intended 10%, drastically inflating perceived wildfire risk across the state. Given the model’s inability to differentiate between fire and non-fire areas, this overclassification suggests that hazard values do not reflect actual fire history.
These findings collectively highlight fundamental flaws in the model’s burn probability estimates, classification performance, and threshold calibration. Given that this evaluation is based on long-term wildfire data spanning over two decades, the results underscore the necessity of validating wildfire risk models against extended historical datasets to ensure their predictive reliability and policy relevance over time.
A more detailed analysis of these findings, including methodology and supporting data, is available in the full paper.