SHAP-based convolutional neural network modeling for intersection crash severity on Thailand's highways

Abstract

Intersection-related crashes on Thailand's highways pose a significant risk to road users, particularly motorcyclists. This study develops customized Convolutional Neural Network (CNN) models to classify the severity of intersection crashes and utilizes SHapley Additive exPlanations (SHAP) to interpret the models. The methodology involves using three years of crash data from Thailand's highways, covering the period from 2018 to 2020. Additionally, three CNN model variations were developed: a basic CNN, a CNN with dropout (CNN-D), and a CNN with both dropout and L2 regularization (CNN-DR). The results demonstrate the superior performance of the CNN-DR model in classifying crash severity for both motorcycle-related and nonmotorcycle-related intersection crashes. SHAP analysis reveals key factors influencing crash severity, including the year of the crash, with a clear distinction between pre-COVID-19 years (2018–2019) and the pandemic year (2020). Crash mechanisms, such as impacts with vehicles from adjacent approaches and rear-end collisions, are significant factors that increase the likelihood of serious crashes. The study also identifies the type of intersection, specifically curved intersections, T-intersections, and Y-intersections, as major determinants of crash severity, particularly for motorcycle-related crashes. Time-of-day analysis reveals early morning hours (00:00 to 5:59) as high-risk periods for nonmotorcycle-related crashes. Furthermore, the influence of highway types and vehicle involvement, such as regional secondary highways and the presence of trucks, is linked to the increased severity of motorcycle-related crashes. The insights derived from this study can guide road safety managers in implementing targeted interventions to reduce intersection crash severity on Thailand's highways.