Deep Learning-Based Convolutional Neural Network for Crash Severity Prediction

Abstract

Predicting traffic crash severity is vital for enabling data-driven policies and interventions to improve road safety. This study aims to evaluate customized Convolutional Neural Network (CNN) architectures for classifying single-motorcycle crash severity in Thailand using police reports from 2018–2020. Systematic experiments reveal substantial variability in model performance across different CNN layouts and layer depths. The peak-performing architecture proves to be a 4-layer dropout-regularized CNN which improves the F1-score by over 2 percentage points compared to a standard CNN baseline. Additionally, this optimized model achieves an aggregate trade-off score (between prediction accuracy rate and false positive rate) of 72.23%–over 5 points ahead of other variants (including the logistic regression and multilayer-perceptron neural network models). It demonstrates resilient precision and reliability in classifying both severe and fatal crashes, even with increasing depth. However, the dataset encompasses just over 2,900 motorcycle crash cases, constraining feasible model complexity. Significantly larger datasets could enable further performance gains from depth and regularization as shown through initial experiments. Overall, this study highlights the promise of applying customized deep learning techniques to unlock essential insights from traffic injury data for guiding impactful road safety policies.