Ecological compression of vegetation information in wildfire spread prediction: a case study of the 2020 Brattain Fire

Christopher Ko, Hannah Ko 

Broad fuel categories simplify mapped vegetation, but it is unclear whether the distinctions removed help predict wildfire spread. We examined whether vegetation composition improved prediction of the next observed fire-growth interval and measured the information lost by grouping fuel models. Using the 2020 Brattain Fire in Oregon, we compared baseline predictors describing fire state, terrain, fuels, canopy, and weather with models that included five additional vegetation fractions. We used chronological validation, logistic regression, and Random Forest, alongside a separate post hoc analysis of native and grouped fuel codes. Across 2,754 held-out cell-interval observations, adding vegetation reduced logistic-regression log loss from 0.363889 to 0.363117 and Brier score from 0.107189 to 0.106988. However, both 95% moving-block uncertainty intervals included zero, and the direction of improvement varied across evaluation periods. Grouping 27 fuel codes into five categories increased the conditional entropy of detailed vegetation classes by 0.919 bits, but retaining native fuel codes did not consistently improve prediction. These findings show that grouping fuel codes removes measurable vegetation-class information without a demonstrated consistent predictive penalty in this single fire. Further research across additional fires and with more direct measures of vegetation condition is needed to determine when greater ecological detail improves wildfire spread prediction.