Prediction of expressway accident locations using contribution information from Grad-CAM

Abstract

This study proposes a method for predicting accident locations on expressways using convolutional neural networks (CNNs) and Grad-CAM-based interpretability. A CNN model is trained using vehicle trajectory images generated from ETC 2.0 probe data. Grad-CAM is applied to extract spatiotemporal contribution information associated with accident risk, enabling the identification of expressway sections where risk-prone traffic patterns are spatially concentrated for the subsequent 30-min period. Based on the spatial aggregation of these contribution scores, a framework is developed to estimate the most likely accident section within a target route. The model outputs are intended to support short-term, location-specific prioritization of safety interventions by traffic operators, rather than forecasting individual crash events. Validation on a 15.34 km section of the Tomei Expressway in Japan demonstrates that the proposed method reduces location prediction error by approximately 19% compared to a naïve prediction, and further outperforms representative tree-based ensemble methods, including Random Forest, XGBoost, and LightGBM. The method shows higher accuracy for multi-vehicle accidents and for high-traffic conditions. Furthermore, the model exhibits superior performance over a statistical baseline across various risk thresholds, indicating its robustness and practical applicability in real-world implementation scenarios where resource allocation must be prioritized. These results highlight the utility of interpretable deep learning in supporting targeted and effective traffic safety interventions.