The analysis of accident risk factors based on the similarity between the contribution distribution of accident risk prediction models and traffic state contour maps

Numerous studies have examined the relationship between traffic conditions and accident risk on expressways. However, few have explored the interpretability of deep learning models in this context. This study aims to identify accident risk factors by analyzing the spatiotemporal features that contribute to predictions made by a CNN-based Accident Risk Prediction Model (ARPM). The ARPM is trained using vehicle trajectory diagrams generated from ETC 2.0 probe data and predicts accident occurrence within the subsequent 30 min. The proposed model achieved a recall of 0.233, a precision of 0.177, and an F1-score of 0.201 on the test dataset at a threshold of 0.43, indicating reasonable predictive performance under a highly imbalanced accident dataset. Grad-CAM is then integrated to visualize the contribution distributions that influence the model's outputs. To interpret these attention patterns, cosine similarity was calculated between Grad-CAM contribution maps and traffic state contour maps of flow, speed, and density. The results show that the similarity with density contours was significantly higher than that with flow or speed, suggesting that the model tends to focus on density-related traffic patterns when predicting accident risk. Furthermore, samples with higher density similarity showed higher future occurrence rates of rear-end and contact collisions, indicating that density-related traffic states are closely associated with congestion-related accident risk. These findings demonstrate that Grad-CAM–based similarity analysis can provide interpretable insights into the traffic conditions underlying CNN-based accident risk predictions.