Development of a Machine Learning Model to Predict Daily Wildfire Growth Across the Continental United States

Wildfires burn millions of acres across the United States each year threatening human lives and property. As rapid growth occurs in the wildland-urban interface, where houses and vegetation meet, wildfires are becoming more of a pressing issue. Additionally, climate change continues to result in a warmer and drier climate, promoting increasingly favorable conditions for wildfires. As a result, forecasting and messaging elevated wildfire potential to partners and the public is crucial. Several fire weather indices, such as the Hot-Dry-Windy index developed by Srock et al 2019, have been developed to aid forecasters. However, these indices are either tied to a particular topographical region or composed of different atmospheric variables. Given this disparity, it can be a headache for forecasters to know which will perform best in any given forecasting situation.

In order to consolidate these forecasting indices, a machine learning model has been developed to predict daily wildfire spread across the Continental United States should a fire develop in any given location. The model takes into account both atmospheric variables (3km WRF model output) (e.g, surface temperature, relative humidity) and fuel variables (gridMET archive) (e.g., vegetative moisture) by training on 14 years of fire data between 2005 and 2018. Fire data for the period is provided through the global satellite derived wildfire database developed by Artes et al 2019 and filtered using the USFS database produced by Short 2020 to remove controlled fires. Additionally, the model considers other variables such as regional fire climatology, population density and topographical slope. Ultimately, a significant fire parameter will be implemented using predicted wildfire size normalized to local fire climatology. The model can be implemented into an ensemble, allowing it to provide probabilistic information to forecasters.