Accident severity prediction on arterial roads via multilayer perceptron neural network.

Abstract

Traffic accidents continue to be a major cause of death in urban areas. While recent research has demonstrated the utility of predictive modelling in rural, express and highway environments, a gap remains in understanding the factors that influence accidents in urban areas, particularly on arterial roads. This study developed multilayer perceptron (MLP), random forest (RF) and multinomial logistic regression (MLR) models to predict accident severity on urban arterial roads in Amman, Jordan's capital. The MLP demonstrates clear superiority over RF and MLR, achieving 97.3% training accuracy and 96.55% testing accuracy. Additionally, a Sobol Global Sensitivity Analysis (GSA) for the MLP model identified critical interactions between variables, especially between collision types and weather conditions. This study provides an in-depth understanding of the key factors influencing accident severity, which can be used to develop new safety regulations and countermeasures to prevent crashes.