Short-Horizon Wildfire Growth Forecasting Using Satellite-Derived Fire-Behavior Features

Abstract

Predicting which active wildfires will grow substantially over the next day can help fire managers direct limited crews and resources. This paper tests whether weather and terrain improve that prediction beyond a fire's own recent behavior. A dataset of 10,076 fire-days from 320 western United States fires (2018-2021) was built from the NASA Fire Event Data Suite (FEDS) and joined to ERA5 weather, terrain, and gridMET fire-weather indices. Substantial growth was defined in advance as more than 25% relative growth and at least 2km^2 of absolute growth in 24 hours. Tree-based models were trained with a fire-level temporal split and compared with a paired, fire-clustered bootstrap. A random forest using only fire-history features reached a test ROC-AUC of 0.911, compared with 0.694 for persistence. Adding weather changed ROC-AUC by -0.002(95% CI [-0.007, 0.003]), and even the actual weather during the forecast window did not help. On quiet days, when history gives the least warning, neither ERA5 nor gridMET gave a
significant improvement. These results suggest that satellite-observed fire behavior carries most of the usable signal for next-day growth.