Development and Evaluation of a Machine Learning Based Wildfire Spread Prediction Model for Regional Air Quality Forecasting

Wildfire spread, the spatiotemporal variations of wildfires, is critical to regional-scale air quality forecasting due to the large abundance of trace gas and aerosol emissions from the fires. Wildfire spread is a complex and highly variable process attributed to wildfire behavior, including various physical and dynamical phenomena. Substantial time and computational resources are required to parameterize and simulate the interactions between these phenomena at regional scales. Hence, this study proposes a novel machine learning wildfire spread model for regional-scale air quality forecast model applications. The proposed model predicts both fire propagation and intensity by predicting the spatial distribution and magnitude of fire radiative power (FRP) for the next hour. Numerous input variables are considered for the model, including fuel type (land use), canopy characteristics, wind fields, air temperature, humidity, precipitation, and topographic characteristics. The contribution of a more appropriate midflame wind speed (i.e., the wind speed at the midpoint of the flame) for wildfire spread is also evaluated. The model is examined using hourly datasets during the prolific wildfire period of September 1^st to October 7^th, 2020 over CONUS with 3-km spatial resolution. Preliminary results show that the machine learning model well captures the propagation of wildfires with R-squared (R²) near 0.6 and good image similarity. The model performs better for big wildfire events with high FRP values in terms of fire propagation. The application of midflame wind speed improves fire propagation prediction to some extent. However, the machine learning model generally underestimates fire intensity by 30 – 50 %, probably due to the limitation of machine learning algorithms.