Predicting Wildfire Strength and Duration Over the United States with Machine Learning Methods

Abstract

Due to the recent increase in global warming, the magnitude and frequency of wildfires have grown. Through the application of machine learning methods, we can better predict the growth and duration of these fires to improve the planning of safety measures and the forecast of air quality. The employment of machine learning in predicting natural occurrences and disasters is relatively new but with the growing quality and availability of data, more machine learning methods are being applied and tested. In this study, we use three widely used machine learning methods, including random forest (RF), artificial neural network (ANN), and BRT, to predict wildfire strength and duration with the FIRMS fire radiation power (FRP) data, NARR meteorological data, IGBP land use, Bluesky fuel loading, and NLDAS soil moisture data from 2002-2019. We chose.2002-2017 data for training and 2018-2019 for verification. We initially tested the effectiveness of multilinear regression and we got an RMSE of 182.594 MW for predicting the FRP and an RMSE of 2.087 days for predicting duration, and through variable manipulation, along with using RF and BRT models, we brought the prediction errors of the FRP and duration down to 77.982 MW and 1.925 days, respectively. Seen by the drastic improvement of our results, it is plausible that we will be able to predict the factors of wildfires. This is consequential, as the ability to predict wildfires will help in a multitude of aspects, such as ecosystem management, climate adaptation, and safeguard planning.