Understanding and predicting population behavior in hurricane evacuations
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This dissertation makes three significant contributions towards understanding and predicting population behavior in hurricane evacuations through advancement in data and modelling. They are summarized as follows. ☐ To advance understanding and predicting who evacuates and the timing of their departures, we develop and compare five different evacuation demand models based on their practical utility for use in managing future hurricane evacuations. The models—participation rate (PR-S), logistic regression (LR-S), random parameter logit (RPL), time-dependent Cox (TD-Cox), and dynamic discrete choice (DDC)—were fitted using population survey and hurricane data collected in a consistent format across four different hurricanes (Florence 2018, Michael 2018, Dorian 2019, and Barry 2019). Out-of-sample predictive power was evaluated in terms of prediction of total evacuation rates, spatial distribution of evacuees, evacuation timing, and individual behavior. The final set of predictors can be obtained for a whole region and applied in the future for prediction. The results suggest that if only an estimate of the total evacuation rate for the whole region is required, the LR-S is easiest to implement and provides good predictive power. However, if spatial and/or timing predictions are required, the DDC is recommended. The results suggest that in general, for future hurricanes, the best models currently available can estimate total evacuation rate within one to nine percentage points; evacuation rate for each county within 10 to 15 percentage points; and departure curves within several hours. Results also indicate that errors become smaller as the geographic area increases. ☐ Along the dimension of destination choices/trip distribution, we present an approach to identify evacuees and characterize evacuation flows using smartphone location data collected across three different hurricanes affecting multiple geographies (Florence 2018, Michael 2018, and Dorian 2019). We further leverage this data to develop a new machine learning model, which predicts the number of evacuees between an origin-destination (O-D) pair at the metropolitan statistical area (MSA) level. The machine learning model incorporates hurricane characteristics which have not been thoroughly exploited by existing methods. The model’s predictive power is comprehensively measured through a ten-fold cross validation using data aggregated across the three hurricanes and compared with the traditional gravity model. Further evaluation of the model’s generalizability is conducted by applying a model trained on the three events to a fourth hurricane (Ida 2021). Results suggest that the new approach significantly outperforms the traditional gravity model across all performance indicators. Analysis of the feature importance of the machine learning model shows that in addition to distance and population, hurricane characteristics are also important in the destination choice decisions of evacuees. ☐ Finally, as the first of its kind, this study provides new knowledge on understanding evacuation behavior by exploring the relationship between visits to certain businesses and evacuation decisions. We use smartphone location and point of interest (POI) data collected across three hurricanes—Dorian (2019), Ida (2021), and Ian (2022)—for residents in voluntary and mandatory evacuation zones. Results suggest residents visit POIs as part of preparatory activities before a hurricane impacts land. Statistical tests suggest that POI visits can be used as precursor signals for predicting evacuations in real time. Specifically, people are more likely to evacuate if they visit a gas station and are more likely to stay if they visit a grocery store, hardware store, pet store, or a pharmacy prior to landfall. Additionally, they are even less likely to leave if they visit multiple places of interest. These results provide a foundation for using smartphone location data in real time to improve predictions of behavior as a hurricane approaches.
Description
Keywords
Hurricane evacuation, Machine learning, Point of interest data, Predictive models, Smartphone data, Statistical methods