WCD-30. Applying Machine Learning Techniques to Forecasting Fire Behavior

Abstract
Wildfires have become an increasing concern in the past few decades. Global warming has dried forests, allowed tree damaging invasive species to proliferate, increased the prevalence of thunderstorms, and lengthened fire seasons. All of these effects and more contribute to increases in wildfire frequency and severity. Emissions from forest fires, especially carbon dioxide (CO2), have an important role in the inter-annual variability of the atmospheric Greenhouse Gas (GHG) budgets. The satellite fire radiative power (FRP), defined as the instantaneous radiative energy output of a fire, is a key variable in understanding the behavior of a fire. Generating high temporal and spatial resolution FRP models allows not only the prediction of future fire behavior and emissions, but also allows the building of robust relationships between carbon monoxide (CO) and FRP to quantify CO emissions on large wildfires, and to improve our understanding of biomass-burning emissions in general. Here we present developments towards a tool (AIFire) that uses techniques from machine learning to predict hourly FRP on the contiguous US (CONUS) domain. We combine the outputs of satellite instruments such as the Visible Infrared Imaging Radiometer Suite (VIIRS), meteorological variables from the High Resolution Rapid Refresh (HRRR) forecast model to form our training dataset. AIFire can be used with meteorological analysis data (historical) to interpolate FRP behavior between satellite passes, or it can be used with forecast data to predict future FRP values. AIFire has the potential to be used not only in the air quality forecast models but also in applications such as plume rise parametrization, FRP-based emissions estimations and in helping to derive FRP diurnal climatologies when the number of FRP samples from satellite sources are insufficient for a given land cover type.