Modeling Wildfire Behavior with Forest Machine Learning Models Using the RAVE Dataset

Smoke-filled skies can ruin day. More and more people are turning to smoke models to check if activities, events, and more will be affected. One way to improve smoke forecasts is to improve the fire radiative power (FRP) modeling during the smoke model forecast period. Previous work has shown that by combining numerical weather model outputs with satellite FRP information (from the day before) as inputs into a random forest model can model, hourly FRP can be modeled as well as the current persistence model.

This work expands on that which was previously done by utilizing a new merged polar-geostationary satellite product called RAVE [https://doi.org/10.1016/j.rse.2022.113237] as the satellite input and training/testing over three years, 2019-2021. Models are trained using several more numerical weather model inputs as well, and explainable AI methods (XAI) are applied to some of the best performing models to gain insight into the model performance. Additionally, case studies are examined on how differently trained models perform regionally as well as on wildfire that undergo certain conditions, such as precipitation.