Characterizing and Predicting Wildfire Evacuation Behavior: A Dual-Stage ML Approach

A research team led by Sazzad Bin Bashar Polock has published a novel machine learning framework for analyzing wildfire evacuation patterns, accepted for presentation at SoutheastConn 2026. Their paper, 'Characterizing and Predicting Wildfire Evacuation Behavior: A Dual-Stage ML Approach,' leverages a substantial survey of residents across three high-risk western U.S. states. The study addresses the critical challenge of understanding the complex, variable human decisions during fire emergencies, moving beyond traditional surveys by applying advanced ML techniques to uncover latent behavioral typologies and build predictive models for key outcomes like transportation mode choice.

The methodology combines unsupervised learning—including Multiple Correspondence Analysis, K-Modes clustering, and Latent Class Analysis—to identify consistent population subgroups differentiated by factors like vehicle access, disaster planning, pet ownership, and technological resources. A complementary supervised modeling stage demonstrates that household characteristics can predict an evacuee's likely transportation mode with high reliability, a finding with direct implications for emergency traffic management. However, the research confirms that predicting the precise timing of evacuation remains difficult, as it heavily depends on dynamic, real-time fire conditions and situational cues rather than static household attributes. This data-driven approach provides emergency managers with a new tool for developing more targeted preparedness campaigns, optimizing resource allocation for vulnerable groups, and promoting more equitable emergency planning strategies.