GPS-Driven evacuation analysis and trajectory synthesis using Large Language Models

MSc Thesis Title: GPS-Driven evacuation analysis and trajectory synthesis using Large Language Models

Author’s name: Apratim Sen

Supervisors: Dr. Xin Wang, Dr. Jeong-Woo Kim (Co-supervisor)

Summary of the thesis: Wildfires pose escalating threats to communities worldwide, driven by climate change and its effects on environmental conditions. Effective evacuation planning requires a deep understanding of human mobility patterns, yet obtaining reliable, large-scale movement data remains a persistent challenge. This thesis presents GPS-driven methods for wildfire evacuation analysis and Large Language Model based trajectory generation framework, addressing both the behavioral understanding of evacuees and the data scarcity challenges that limit research and planning applications.

The first contribution examines evacuation dynamics during three Canadian wildfires: Upper Tantallon wildfires, Nova Scotia (2023), which took place a suburban community on the outskirts of Halifax; Jasper wildfires, Alberta (2024), which took place near Jasper, a geographically isolated tourist destination in the Canadian Rockies; and the McDougall Creek wildfires of Okanagan, British Columbia (2023) which primarily affected West Kelowna, a small city serving as a major hub in the Okanagan region. Through spatial and temporal analysis of GPS records, the study reveals significant differences in evacuation behavior between the regions. The findings underscore that evacuation dynamics are heavily context-dependent, and that region-specific strategies are essential for effective emergency response. 

The second contribution addresses a fundamental limitation of GPS-based research: real-world trajectory datasets are often scarce, inconsistently sampled, and carry significant privacy risks that restrict data sharing and reuse. To address this, RouteGenLLM, a novel fine-tuning framework for large language models is introduced. Raw GPS trajectories are discretized into tokens containing H3 cell IDs fused with road identifiers, enabling topologically valid autoregressive generation, guided by graph constraints, temporal features, and destination-aware prompting. Through GRPO-based reinforcement fine-tuning, the framework achieves over 85\% next-step prediction accuracy and generates trajectories that closely match real spatial distributions, providing a scalable and privacy-preserving tool for evacuation planning and urban mobility analysis.