Taking Advantage of Machine Learning Methods to Better Represent Fire Radiative Power (FRP) for Smoke and Weather Forecasting Models

Currently, the High Resolution Rapid Refresh (HRRR) coupled with Smoke (HRRR-Smoke) and many other smoke forecasting models represent diurnal cycle of fire emissions by using climatological curves. In HRRR-Smoke the fire emissions are estimated from the measured value of FRP from satellite data taken by the Visible Infrared Imaging Radiometer Suite (VIIRS) and Moderate Resolution Imaging Spectroradiometer (MODIS) instruments. The forecasting of fire smoke emissions are based on the persistence approach. As is, it is quite challenging to accurately forecast fire emissions and smoke concentrations in rapidly changing fire behaviour and weather conditions.

We propose improving upon older methods by introducing a machine learning model that integrates available Geostationary Operational Environmental Satellites (GOES) R satellite irradiance data and measured FRP with existing satellite-measured FRP values in addition to modeled meteorological values to better represent the FRP curve of individual fires based on current weather conditions and geospatial location. With information such as more frequently available satellite imagery and temperature, relative humidity, and wind, we present an algorithm to improve FRP curve for each pixel point at hourly time scale. We will compare different machine learning methods, such as random forest and neural networks, during 2017 and 2018 to demonstrate the potential of better representing FRP curves of fires for smoke modeling applications. We will also investigate how to analyze the importance of the different inputs into these machine learning models to better understand what drives FRP variability according to the machine.