Machine Learning, Process-Based Modeling, and Observational Data: the Wildfire Modeling Triangle
Abstract
Between January 1 and July 27, 2021 wildfires consumed about 2.8 million acres across the United States compared with 1.9 million during the same period in 2020, continuing the trend in increased burn area the U.S. has experienced since 2004 (National Interagency Fire Center). As of July 13, 2021, 95% of the Western U.S. was expecting moderate to severe drought (U.S. Drought Monitor), and the heatwaves that much of the U.S. experienced in June compounded conditions ripe for more wildfire activity. Clearly, this is a critical time for wildfire scientists to engage all of our available knowledge and resources. This means coalescing observational data with process-based modeling and machine learning to further our understanding of wildfire behavior and improve forecasting. Here, we discuss three uses of machine learning methods that advance our understanding of wildfire. First we discuss an approach to determine regional-scale wildfire drivers using random forest models trained with the Fire Events Delineation (FIRED) dataset, coupled with meteorological, terrain, and fuel data. The combination of observation data with machine learning provides insights into the hierarchy of wildfire drivers by ecoregion, which can in turn inform process-based models. Second, we apply random forest models to update fuel inputs to improve Weather Research Forecasting-Fire (WRF-Fire) forecasts of the 2020 Colorado East Troublesome Fire using Sentinel-2 imagery and the USGS tree mortality survey. The burn area forecast improvements emphasize how machine learning may aid in adjusting inaccurate input data for use in process-based models. Finally, we conclude with a discussion of future uses of machine learning integrated with processed-based modeling for wildfire science, including the potential for machine learning to efficiently emulate computational and time-intensive process-based model forecasts.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMIN21A..03D