Predicting fire severity in Montana using a random forest classification scheme

Fire managers often make decisions about wildfire incidents on a landscape scale. While several well developed models can predict fire behaviour at these scales, the paucity of data they draw upon can limit their range of validity. Other models explicitly represent the physical complexities of a fire environment, but at increased computational costs and increased sensitivity to boundary conditions. In this paper, we explore a middle ground between landscape level, data-driven fire behaviour predictions and physics-based, computationally expensive models. As a first step, we employ machine learning to predict fire severity from a set of well recognized covariates such as weather, fuel, and topography. A gradient boosted regression tree is used for the classification task, and sets of model hyperparameters tested to determine their role in the classification. The model demonstrates skill in prediction of the burn severity, but is limited by the dynamically evolving set of features; especially the weather. Results were found to be biassed towards under-predicting the severity of wildfires. We conclude with a discussion of approaches available to improve model performance and future applications of the predicted burn severity.