Use of Machine Learning to Characterize and Predict Wildfire Activity in Alaska

Abstract:

While a mostly natural phenomenon, recent climate change in Alaska has increased the frequency and severity of wildfires across the state. This study uses random forest machine learning models to better understand the variables that can be used to predict fire activity. Remote sensing records over the past 20 years have allowed for the collection of detailed records of daily fire activity. While in-situ weather observations remain sparse, increasing station density and model accuracy provides the opportunity for data-driven methods to provide fresh insight on large-scale wildfire behavior. Given access to a wide variety of weather and fuel data sources, the methods explored in this study show moderate skill when predicting remote sensing-based fire activity on any given day. The model reveals the critical importance of including a measure of previous day fire activity alongside general weather variables. When provided with a measure of previous fire activity, instantaneous measures of weather variables outperform longer term metrics of fuel conditions. Weather variables obtained from an atmospheric reanalysis (ERA5) generally make greater contributions to forecast skill than do the sparse local observations. The models explored in this presentation include forecast variables from weather prediction models to aid in medium range fire prediction. Data categorization may also be used to enhance the operational usefulness and accuracy of these methods.