A machine learning model to predict wildfire burn severity for pre-fire risk assessments, Utah, USA

Abstract Background High-severity burned areas can have lasting impacts on vegetation regeneration, carbon dynamics, hydrology, and erosion. While landscape models can predict erosion from burned areas using the differenced normalized burn ratio (dNBR), post-fire erosion modeling has predominantly focused on areas that have recently burned. Here, we developed and validated a predictive burn severity model that produces continuous dNBR predictions for recently unburned forest land in Utah. Results Vegetation productivity, elevation, and canopy fuels were the most important predictor variables in the model, highlighting the strong control of fuels and vegetation on burn severity in Utah. Final model out-of-bag R 2 was 67.1%, residuals showed a correlation coefficient of 0.89 and classification accuracy into three classes was 85%. We demonstrated that dNBR can be empirically modeled relative to fuels and topography and found burn severity was highest in productive vegetation and at relatively cooler sites. Conclusions We found that prediction accuracy was higher when fuel moisture was lower, suggesting drier weather conditions drive more consistent and predictable burn severity patterns across a range of burn severity, vegetation types, and geographic locations. Moreover, burn severity predictions from this model can be used to inform hydro-erosion models and subsequent management actions aimed at reducing burn severity and post-wildfire erosion risks.