Exploring the Use of Machine Learning to Develop a Predictive Model for Future Fire Seasons

Wildfires in the United States can be extremely costly, both economically and through the loss of life. The cost of fire suppression efforts alone surpassed $2B in 2017, which is greater than 50% of the U.S. Forest Service’s total budget. This leaves significantly less funds available for forest management and fire prevention efforts. With ballooning costs, predictive estimates of variables such as the start, severity and length of future fire seasons could be beneficial to management officials. This project looks to explore the use and feasibility of using machine learning to develop a predictive model based on a variety of remotely sensed, model and in situ datasets. More specifically, this project will take advantage of various NASA remote sensing assets, the NASA Short Term Prediction and Research Transition Center’s Land Information System (SPoRT-LIS) and gridded surface meteorological data. Example variables of interest may include LIS soil moisture, the ALEXI Evaporative Stress Index, MODIS green vegetation fraction (GVF) and leaf area index (LAI), fuel moisture, temperature, precipitation, and vapor pressure deficit. A database of over 10,000 fire cases from 2000-2015 will be used to perform analysis on all selected parameters at the start location of the fire to determine their use in a predictive model. In addition, antecedent land surface and meteorological conditions will be analyzed to gather any precursory fire signals.