Predicting Wildfire Spread with Machine Learning and a Socio-Ecological Approach to Identify Areas of High Vulnerability

Wildfires, exacerbated by climate change, cause significant devastation to human livelihoods in the USA. The impacts of wildfires are felt more adversely by socially vulnerable communities, such as those with high poverty levels or high percentage of disabled population. Hence, it is critical not only to accurately model wildfire spread, but also to identify and prioritize these at-risk communities. Such targeted efforts are essential to mitigate the impacts of wildfires and create more equitable communities. This study leverages a comprehensive dataset of approximately 1.6 million data points, recorded between 2012 and 2017, to develop machine learning models that predict wildfire intensity based on meteorological, topographical, and vegetation features. Among the models tested, Random Forest demonstrated the highest performance, achieving 86.4% accuracy, 85.0% precision, and 88.0% recall in predicting wildfire intensity. Additionally, this research introduces a new nationwide socio-ecological framework by integrating wildfire risk with the Social Vulnerability Index provided by the Centers for Disease Control and Prevention to develop a Combined Vulnerability Score. This score identifies census tracts that have high social vulnerability coupled with high wildfire risk. This study marks the first comprehensive initiative across the USA to merge machine learning-driven predictions of wildfire intensity with social vulnerability assessments, determining areas in urgent need of support during wildfire disasters.