Human Mobility Prediction and Simulation After Natural Disasters: A Big Data Approach Using GPS Tracking and Machine Learning
This peer-reviewed journal article presents a computational framework for predicting and simulating population movements following large-scale natural disasters, developed through analysis of big and heterogeneous data from Japan. Using GPS records from 1.6 million anonymized users collected over three years, alongside earthquake intensity data, government declarations, news reports, and urban transportation network data, the authors conduct a detailed empirical analysis of human behavior following the 2011 Great East Japan Earthquake and Fukushima nuclear accident. The study finds that post-disaster mobility patterns are shaped by a combination of factors including disaster intensity, damage levels, proximity to the hazard source, social relationships, government-designated shelter locations, and media reporting — and that these patterns are significantly more predictable than previously assumed.
Building on these empirical findings, the study develops a two-level predictive model. The first level uses a Hidden Markov Model (HMM) to predict individual behavioral states — such as staying home, evacuating to a shelter, moving toward social relations, or relocating to an unknown safe place — based on observed movements and current disaster conditions. The second level applies a Markov Decision Process (MDP) framework to predict specific evacuation routes and transportation modes, drawing on a population mobility graph constructed from large-scale trajectory data. Together, these components allow the model to generate likely movement paths for individuals given their known important places and the unfolding disaster situation, with transportation mode (walking, cycling, car, train) incorporated to improve prediction accuracy.
A key contribution of the study is a knowledge transfer mechanism that extends the model beyond the specific context of the 2011 earthquake. By extracting generalizable patterns from disaster mobility data, the authors develop a simulation capability that can generate plausible population movements for any person, any location, and any disaster scenario — including smaller-scale events for which individual historical GPS data may not be available. For small-scale earthquakes, the simulator uses multimodal route planning on standard transportation networks; for large-scale disasters where public transport is disrupted, it relies on the pretrained urban mobility graph.
Experimental validation demonstrates that the model outperforms existing baseline approaches across multiple metrics, including predictive accuracy, route distance matching, and log-likelihood scores. The authors note limitations including potential demographic bias toward younger, mobile-device-owning populations. The study has direct relevance for disaster risk reduction practitioners and emergency managers working on anticipatory action, evacuation planning, population movement modeling, and humanitarian logistics — particularly in contexts where understanding where people will go before, during, and after a disaster can inform resource prepositioning and response coordination.