Predicting daily firefighting personnel deployment trends in the western United States

Abstract

Projected increases in wildfire frequency, size, and severity may further stress already scarce firefighting resources in the western United States that are in high demand. Machine learning is a promising field with the ability to model firefighting resource usage without compromising dataset size or complexity. In this study, the Categorical Boosting (CatBoost) model was used with historical (2012–2020) wildfire data to train three models that calculate predicted daily counts of 1) total assigned personnel (total personnel), 2) assigned personnel that are at the fire (ground personnel), and 3) assigned personnel that either work with aircraft or in management (air/overhead personnel) based on daily wildfire characteristics. The main drivers behind personnel assignment under current management practices included structures threatened, acres burned, point of fire origin, and fire priority. While contextual variables such as preparedness level and the presence of other large fires were among the least important, the importance of fire priority reveals that factors beyond the features of the fire itself are influential in personnel assignment. CatBoost model predictions provide an historical context to firefighting resource assignment and could also be used to inform decision-makers and managers about future issues facing firefighting resources in the western United States given projected changes in climate.