Predicting human mobility flows in response to extreme urban floods: A hybrid deep learning model considering spatial heterogeneity

Resilient post-disaster recovery is crucial for the long-term sustainable development of modern cities, and in this regard, predicting the unusual flows of human mobility when disasters hit, could offer insights into how emergency responses could be managed to cope with such unexpected shocks more efficiently. For years, many studies have been dedicated to developing various models to predict human movement; however, abnormal human flows caused by large-scale urban disasters, such as urban floods, remain difficult to capture accurately using existing models. In this paper, we propose a spatiotemporal hybrid deep learning model based on a graph convolutional network and long short-term memory with a spatial heterogeneity component. Using 1.32 billion movement records from smartphone users, we applied the model to predict total hourly flows of human mobility in the “7.20” extreme urban flood in Zhengzhou, China. We found that the proposed model can significantly improve the prediction accuracy (i.e., $R^2$ from 0.887 to 0.951) for during-disaster flows while maintaining high accuracy for before- and after-disaster flows. We also show that our model outperforms selected mainstream machine learning models in every disaster stage in a set of sensitivity tests, which verifies not only its better performance for predicting both usual and unusual flows but also its robustness. The results underscore the effective role of spatial heterogeneity in predicting human mobility flow in a disaster context. This study offers a novel tool for better depicting human mobility under the impact of urban floods and provides useful insights for decision-makers managing how people move in large-scale disaster emergencies.