A two-step machine learning method for casualty prediction under emergencies

Abstract

Casualty prediction is meaningful to the emergency management of natural hazards and human-induced disasters. In this study, a two-step machine learning method, including classification step and regression step, is proposed to predict the number of casualties under emergencies. In the classification step, whether there are casualties under an incident is firstly predicted, then in the regression step, samples predicted to have casualties are used to further predict the exact number of the casualties. Using an open-source dataset, this two-step method is validated. The results show that the two-step model performs better than the original regression models. Back propagation(BP) neural network combined with Random Forest performs the best in terms of the death toll and the number of injuries. Among all the two-step models, the lowest mean absolute error (MAE) for the death toll is 1.67 while that for the number of injuries is 4.13, which indicates that this method can accurately predict the number of casualties under emergencies. This study's results are expected to provide support for decision-making on rapid resource allocation and other emergency responses.