What is it about?
Traffic accidents are shaped by both road layout and the physical environment around a road, but most prediction systems mainly use road network structure. We built a dataset covering six US states with nine million accident records, one million satellite images, weather information, road types, and traffic volume. Our model combines these data sources to predict accident risk for each road segment. It achieved an average AUROC of 90.1%, improving by 3.7% over models that use road structure alone. We also used the learned representations to compare similar road segments and estimate how specific conditions affect accident rates. After accounting for other observed factors, accident rates increased with greater precipitation, on higher-speed roads such as motorways, and during higher-risk seasons. The results show that satellite images provide useful information for both accident prediction and the analysis of contributing factors.
Featured Image
Photo by NASA on Unsplash
Why is it important?
Traffic accident prediction is often based mainly on road maps, leaving out visual information about the road and its surroundings. Our study shows that satellite images provide useful information that road structure alone cannot capture. Combining satellite imagery with road networks, weather, road type, and traffic data raises the average AUROC to 90.1%, a 3.7% improvement over graph models that use only road structure. The learned representations also allow comparisons between similar road segments, helping estimate how precipitation, road speed, and seasonal conditions affect accident rates. These results can help transport agencies identify high-risk locations and decide where safety reviews and preventive measures may be most useful. The released dataset and code also give researchers a shared resource for developing and testing road safety models across different regions.
Perspectives
This work brings multimodal analysis into transportation research by combining satellite imagery, road networks, weather, traffic, and accident records within one framework. It shows how transportation studies can move beyond a single data source to support both more accurate accident prediction and a clearer analysis of contributing factors. For me, this project was an exciting opportunity to connect machine learning with a real public safety problem. I especially enjoyed collaborating with faculty members from outside my research area. It was rewarding to see how visual information about roads could improve our understanding of accident risk. I hope this work encourages closer collaboration between machine learning and transportation researchers and contributes to safer roads.
Ziniu Zhang
Northeastern University
Read the Original
This page is a summary of: Learning Multimodal Embeddings for Traffic Accident Prediction and Causal Estimation, April 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3770854.3785677.
You can read the full text:
Resources
Contributors
The following have contributed to this page
