Deep Learning or Machine Learning or Statistical Models for Weather-related Crash Severity Prediction?

Road traffic crashes are a significant public health concern. Rain, snow, ice, and poor visibility often create unsafe road conditions for drivers. Recent reports published by the United States Department of Transportation (USDOT) indicate that nearly 5,000 people are killed and more than 418,000 people are injured in weather-related crashes each year. The drivers may not have enough time to react and stop on a wet pavement surface or in poor visibility conditions, often resulting in a severe crash. Understanding the contributing factors to weather-related crashes will help to select and implement suitable solutions to minimize the number or severity of these crashes.

Researchers in the past adopted statistical methods such as binary logistic regression models, ordered probit models, and random parameter logit models to evaluate crash severity. A few researchers have explored machine learning methods to predict crash severity in the past.

Recent advances in deep learning resulted in powerful and robust tools like convolutional neural networks (CNN) to train the deep architecture. These models are proven to be superior in many transportation engineering applications. The capabilities of CNN-based deep learning algorithms to predict crash severity have not been explored in the past. This research aims to propose a novel deep learning-based algorithm to predict the severity of weather-related traffic crashes and compare it with machine learning and statistical models. This will be complemented with efforts to model crash severity by specific weather condition like rain, poor visibility, and snow.