Discovering Crash Severity Factors of Grade Crossing With a Machine Learning Approach

Abstract

According to the Federal Railroad Administration (FRA) Highway-Rail Grade Crossing Accident/Incident database, more than 12,000 accidents occurred between 2012 and 2017 in the United States with casualties of around 3900. Despite repeated efforts to fully understand the risk factors that contribute to highway-rail grade crossing collisions, there still remain many uncertainties. A machine learning approach is proposed in this paper to find out significant factors, along with their individual impacts of crash severities at grade crossings. One of the most efficient and accurate machine learning algorithms, extreme gradient boosting (XGB or XGBoost), is applied to analyze 21 different accident and crossing -related characteristics per driver severities. The XGB model has been proven in previous studies across many research areas in transportation to outperform other machine learning-based methods and statistical classification methods, such as multinomial logit model, multiple additive regression trees, decision tree, and random forest, especially in prediction accuracy. Thereby, applying the algorithm is expected to provide highly reliable results to identify important factors that have impacts on injury severities at grade crossings. Such application will further aid the discovery of potential crossings with significant factors. The FRA’s Highway-Rail Grade Crossing Accident/Incident database from 2012 to 2017 is fused with the FRA Highway-Rail Crossing Inventory database for the analysis. Observations with missing information were removed from the original database. Crossing position under or over the railroad and pedestrian or other types of highway users were also not considered since they were not specifically of interest in this study. After the database cleaning process, it condensed to the total of 1,250 accidents out of the retrieved 12,630 from the combined database. The results show that adjacent highway traffic volume and train speed are the most significant factors causing accidents and injury severity. They are followed by the driver’s age and the estimated vehicle speed. It also indicated that truck-involved accidents and crossings with gates, flashing lights, and other types of warning devices combined, and highway user’s gender as a male also pertain to the higher injury rate. Through this study, it is possible to provide guidance to decision-makers in recognizing possible risks at-grade crossings that may cause driver casualties.