Predictors of Evacuation Rates During Hurricane Laura: Weather Forecasts, Twitter, and COVID-19

Abstract

Machine learning was applied to predict evacuation rates for all Census tracts affected by Hurricane Laura. The evacuation ground truthwas derived from cellphone-based mobility data. Twitter data, Census data, geographical data, COVID-19 case rates, the CDC/ATSDR social vulnerability index, and relevant weather and physical data were used to do the prediction. Random forests were found to perform well, with a MAPE of 4.9% on testing data. Feature importance for prediction was analyzed using Shapley additive explanations and it was found that previous evacuation, rainfall forecasts, COVID-19 case rates, and Twitter data rank highly in terms of importance. Social vulnerability indices were also found to show a very consistent relationship with evacuation rates, such that higher vulnerability consistently implies lower evacuation rates. These findings can help with hurricane evacuation preparedness and planning as well as real-time assessment.