Machine Learning to Predict Vehicular Crash Severity from Weather Conditions

Adverse weather conditions are responsible for thousands of vehicular deaths, millions of vehicular crashes, and billions of dollars in economic and congestion costs each year. To date, the meteorology community has provided climatology and relative risk assessments of vehicular crashes with respect to weather conditions that fall short of providing predictive capability of crash frequency and severity to transportation network operators. The transportation community has used machine learning to predict crash severity; however, weather conditions are often overlooked in favor of other contributing factors such as driver age, driver behavior, vehicle type, and the presence of driver distractions and/or inebriation.

This research focuses on using machine learning algorithms, such as random forests, to predict vehicular crash severity from meteorological variables using the jointly created Governors Highway Safety Association’s and National Highway Traffic Safety Administration’s Model Minimum Uniform Crash Criteria injury severity scale known as KABCO. Statewide reported vehicular crashes were obtained from Nebraska (2009–2013) and Colorado (2013–2018). Meteorological variables obtained from Automated Surface Observing System stations via the National Centers for Environmental Information are used as predictors for vehicular injury crash severity. As part of the random forest model implementation, the relative importance of these predictors was computed and compared between the two states and across diverse weather conditions to better understand the predictive value of each variable. Additionally, this is done across diverse weather conditions such as winter season versus warm season precipitation events and non-precipitation events like wind and visibility. To provide a tool for diverse end-users, a winter severity index metric representing the combined influence of individual meteorological predictors in a single metric was computed for both states to associate with vehicular crash severity. Finally, a winter season case study provides an application context for the final random forest model to compare crash severity predicted from meteorological conditions with road closure frequency and duration information. The output from this analysis will inform transportation policy and incident management with respect to weather-related challenges and underscore the need for continued and expanded partnerships between the meteorology and transportation communities.