A machine learning approach for predicting hurricane evacuee destination location using smartphone location data

Abstract

Evacuation destination choice modeling is an integral aspect of evacuation planning. Outputs from such models are required to estimate the clearance times on which evacuation orders are based. The number of evacuees arriving at each destination also informs allocation of resources and shelter planning. Despite its importance, evacuee destination modeling has not received as much attention as identifying who evacuates and when. In this study, we present a new approach to identify evacuees and determine where they go and when using privacy-enhanced smartphone location data. We demonstrate the method using data from four recent U.S. hurricanes affecting multiple geographies (Florence 2018, Michael 2018, Dorian 2019, and Ida 2021). We then build on those results to develop a new machine learning model that predicts the number of evacuees that move between pairs of metropolitan statistical areas. The machine learning model incorporates hurricane characteristics, which have not been thoroughly exploited by existing methods. The model’s predictive power is comprehensively evaluated through a tenfold cross validation, holdout validation using Hurricane Ida (2021), and comparison with the traditional gravity model. Results suggest that the new model substantially outperforms the traditional gravity model across all performance indicators. Analysis of feature importance in the machine learning model indicates that in addition to distance and population, hurricane characteristics are important in evacuee destination choices.

Funding

National Science Foundation,(CMMI-1331269)