SHINE: Using Machine Learning to Incorporate Social Vulnerability Indicators into Wildfire Risk Management

Abstract

Eve Myadze-Pike and Alper Yilmaz

This study introduces SHINE, a novel tool designed for wildfire risk mitigation and adaptation planning. The name is derived from the tool’s primary function in assessing the Susceptibility and Human Impact of Natural Emergencies. Leveraging 22 years of historical data from California, the tool employs a random forest initialized AdaBoost classifier to analyze the intersection of social vulnerability and wildfire risk. Six categories were derived by considering subsets of two from the four thematic areas identified by the Social Vulnerabil- ity Index from the U.S. Centers for Disease Control and Prevention. These categories were cross-referenced with spatial data on square kilometers burned by wildfires. The classifier, validated using 20 years of California data and tested on an additional 2 years of Califor- nia data, demonstrated remarkable effectiveness. It achieved high accuracy and precision, with a mean accuracy of 99.22%, a precision of 99.40%, a recall of 98.87%, and an F1 score of 99.13% on the California test set. In a groundbreaking extension of our research, the classifier is further tested against 22 years of data from Oregon and Washington. On the Oregon test set, the model achieved a mean accuracy of 97.66%, a mean precision of 96.98%, a mean recall of 95.10%, and a mean F1 score of 95.99%. The model performed with a mean accuracy of 97.37%, mean precision of 96.96%, mean recall of 95.25%, and a mean F1 score of 96.06% on the Washington test set. The successful application of the classifier, trained on California data, to Oregon and Washington data demonstrates its adaptability and effectiveness across varied geographic and environmental contexts. This research showcases the potential of machine learning models in enhancing disaster risk reduction strategies and enabling targeted community-specific interventions.

PDF