SGR-Net: Learning Multimodal Embeddings for Traffic Accident Prediction via Geometry–State Attentive Fusion

Traffic accident prediction is a key challenge in road safety, and it is necessary to accurately identify high-risk sections from different data sources. Although graphical neural networks (GNNs) simulate the road network topology well, they ignore the visual and environmental clues from physical road conditions. This paper addresses this gap by proposing a Sequential Geometric Reasoning Network (SGR-Net), a deep learning framework for multimodal accident prediction. Unlike prior GNN-based approaches, SGR-Net introduces a Geometry–State Attentive Fusion (GSAF) module—its main novelty—which dynamically integrates visual features from satellite imagery with structural graph contexts. The framework also includes a stability-aware training objective and meta-learning for cross-region generalization. We evaluate on a large-scale dataset covering six U.S. states with over nine million accidents and one million satellite images. SGR-Net achieves strong results, with AUROC up to 96.8% and MAE as low as 0.08 in Delaware. Ablations confirm the GSAF module is essential: removing it reduces AUROC by 2.7% and increases MAE by over 40%. The framework establishes a new state-of-the-art for multimodal traffic accident prediction.