A spatiotemporal deep learning approach for citywide short-term crash risk prediction with multi-source data

Area of Science:

• Transportation Engineering
• Artificial Intelligence
• Urban Planning

Background:

• Accurate short-term traffic crash risk prediction is crucial for urban safety.
• Existing methods often struggle to capture complex spatiotemporal dynamics.
• Multi-source data integration presents a challenge for traditional models.

Purpose of the Study:

• To evaluate the effectiveness of a deep learning approach for citywide short-term crash risk prediction.
• To leverage multi-source datasets including crash, taxi GPS, road network, land use, population, and weather data.
• To compare the proposed deep learning model against econometric and machine-learning benchmarks.

Main Methods:

• Development and application of a spatiotemporal convolutional long short-term memory network (STCL-Net).
• Utilized diverse datasets from Manhattan, New York City.
• Conducted nine prediction tasks across different spatiotemporal resolutions (weekly, daily, hourly).

Main Results:

• STCL-Net demonstrated superior performance compared to benchmark models in most crash risk prediction tasks.
• Prediction accuracy increased and false alarm rates decreased with STCL-Net.
• Model performance declined with higher spatiotemporal resolution; econometric models outperformed ML models for weekly predictions, while ML models were better for daily predictions.

Conclusions:

• The proposed spatiotemporal deep learning approach effectively captures spatiotemporal characteristics for citywide short-term crash risk prediction.
• STCL-Net offers a promising solution for enhancing transportation safety.
• Findings can guide the selection of appropriate prediction methods for different transportation safety engineering tasks.