A novel deep learning approach to predict crash severity in adverse weather on rural mountainous freeway

Abstract

Abstract The main focus of this study was to develop a robust prediction model based on deep learning capable of providing timely predictions of injury and fatal crashes in adverse weather on rural mountainous freeways. This study leveraged a promising deep learning technique named ResNet18. To apply the proposed deep learning model, the numeric crash data were converted to images utilizing a cutting-edge method, called DeepInsight. In addition, considering the imbalanced nature of the crash data, this study leveraged two data balancing techniques, namely Random Under Sampling (RUS) and Synthetic Minority Oversampling Technique (SMOTE); and experimented with several data sampling ratios. The best prediction performance was found using a ratio of 1:2:2 (Fatal:Injury:PDO) coupled with both RUS and SMOTE, which produced an overall prediction accuracy of 99.3% and 80.5% for fatal and injury crashes, respectively. This study also investigated the importance of variables on crash severity, which revealed that driver residency, vehicle damage extent, airbag deployment, driver conditions, weather, and road surface conditions were the most important variables contributing to the severity of crashes. The proposed deep learning framework can provide an accurate prediction of fatal and injury crashes, which is crucial to ensuring effective traffic collision management.