Fire Intensity and spRead forecAst (FIRA): A Machine Learning Based Fire Spread Prediction Model for Air Quality Forecasting Application

Abstract

Fire activities introduce hazardous impacts on the environment and public health by emitting various chemical species into the atmosphere. Most operational air quality forecast (AQF) models estimate smoke emissions based on the latest available satellite fire products, which may not represent real-time fire behaviors without considering fire spread. Hence, a novel machine learning (ML) based fire spread forecast model, the Fire Intensity and spRead forecAst (FIRA), is developed for AQF model applications. FIRA aims to improve the performance of AQF models by providing realistic, dynamic fire characteristics including the spatial distribution and intensity of fire radiative power (FRP). In this study, data sets in 2020 over the continental United States (CONUS) and a historical California fire in 2024 are used for model training and evaluation. For application assessment, FIRA FRP predictions are applied to the Unified Forecast System coupled with smoke (UFS-Smoke) model as inputs, focusing on a California fire case in September 2020. Results show that FIRA captures fire spread with R-squared (R²) near 0.7 and good spatial similarity (∼95%). The comparison between UFS-Smoke simulations using near-real-time fire products and FIRA FRP predictions show good agreements, indicating that FIRA can accurately represent future fire activities. Although FIRA generally underestimates fire intensity, the uncertainties can be mitigated by applying scaling factors to FRP values. Use of the scaled FIRA largely outperforms the experimental UFS-Smoke model in predicting aerosol optical depth and the three-dimensional smoke contents, while also demonstrating the ability to improve surface fine particulate matter (PM_2.5) concentrations affected by fires.