Development and evaluation of statistical models for international-scale road traffic flow estimation.

Area of Science:

• Environmental Science
• Data Science
• Transportation Engineering

Background:

• Accurate traffic data is crucial for air pollution and noise modeling, yet current models lack sufficient data, particularly for local roads.
• Existing models struggle with comprehensive Annual Average Daily Traffic (AADT) data, hindering the development of precise emission inventories.

Purpose of the Study:

• To enhance the coverage and accuracy of AADT data across all road types.
• To support the development of more reliable emission inventories by improving traffic flow modeling.

Main Methods:

• Developed and compared statistical machine learning models: Random Forest (RF), extreme gradient boosting (XGB), and generalised linear mixed model (GLMM).
• Utilized open-data from 4744 traffic flow measurement sites across 7 European countries, incorporating 51 predictor variables.
• Employed a 1000-fold cross-validation with stratified sampling (by road type and area) for robust model training and testing, creating ensemble predictions.

Main Results:

• RF and XGB models demonstrated strong performance (R²: 0.85-0.86), significantly outperforming GLMM (R²: 0.69-0.73).
• A 'combined' modeling approach, using road-type specific sub-models, slightly improved overall performance compared to 'single' models.
• The 'combined' method showed particular improvement for residential roads, highlighting the benefit of tailored models.

Conclusions:

• Machine learning, specifically RF and XGB, offers a powerful solution for improving AADT data accuracy and coverage.
• The 'combined' modeling strategy, leveraging road-type specific models, enhances prediction performance, especially for complex road networks.
• Local calibration may be necessary to ensure model transferability and maintain accuracy in diverse geographical areas.