Introducing spatio-temporal graph networks to displacement predictions in the context of the 2023 Turkey-Syria earthquake

Abstract

A devastating earthquake impacted southern Turkey and northern Syria in 2023, laying waste to infrastructure alongside causing massive displacement of affected people. Utilizing mobile-phone records of people in the effected area surrounding the disaster, we construct mobility flow records to model the flows and to assess the predictive power of such data. We use graph networks to explore the influence of spatial and temporal data granularity, how inter-regional relations can be represented as graphs, and whether a disaster-related model can enhance prediction performance during such events. Results indicate that proper inter-regional relation representations, such as centrality measures based on road networks between regions, provide a significant performance enhancement. While we see opportunities for our proposed model in context of less-sparse and more-geographically dependent mobility data to improve performance, our results do not indicate that the graph network and disaster-components are effective solutions, where a gated recurrent unit (GRU) and Historic Average model, considering only temporal dependencies, outperform these models.