A Deep Learning Approach to Predict Severity Levels of Work Zone Crashes

The need for maintaining aging infrastructure and/or expanding road capacity leads to an increasing number of work zones presented on highways. Consequently, many crashes were observed in work zone areas. In order to help transportation agencies select more appropriate countermeasures to reduce crash risk in work zones, more comprehensive studies on work zone crash occurrences and the corresponding outcomes are needed. Existing research mainly sought to use parametric models such as logistic regression to examine the causal relationship between potential contributing risk factors and accordingly work zone crash severity. However, the predictive performances of the deployed models have been rarely examined. This paper introduces a non-parametric model building upon deep learning approach to predict the severity levels of work zone crashes. A numerical study that uses massive work zone crash data obtained from a State crash database was conducted to test the predictive capability of the proposed approach. Compared with two baseline approaches, logistic regression and support vector machine, the proposed method achieved improved performance. Key findings, recommendations, and the sensitivity analysis are also provided that can help users consider suitable parameters when using the proposed approach to predict the severity of work zone crashes.