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

A major shortcoming of air pollution and road traffic noise models is a lack of traffic data, especially for local roads. We aimed to increase the coverage and accuracy of data on annual average daily traffic (AADT) flows on all road types to support the development of emission inventories. We developed statistical machine learning approaches (Random Forest - RF; extreme gradient boosting - XGB; generalised linear mixed model - GLMM) to model AADT, using open-data on measured traffic flows from 4744 sites, covering 12 areas in 7 European Countries. Traffic flow predictor variables (N = 51) described the types of roads, road network connectivity, population, drive-time, land use and topography. A 1000-fold cross-validation procedure was used in model training and testing, where a randomly stratified (by road type and area) selection of 80% of traffic counting sites were used to train models, with the remaining 20% of sites reserved for model testing. The predictions for all reserved sites from the 1000 iterations were pooled to provide an 'ensemble' prediction for each site. For performance evaluation, we compared 'single' models developed using the whole dataset with models 'combined' from road-type sub-models (highways, local, residential). RF and XGB performed similarly overall (R²: 0.85-0.86) and outperformed GLMM (R²: 0.69-0.73). The 'combined' method slightly outperformed 'single' models overall, with notably better performance for residential roads by using sub-models. For transferability to other areas, we recommend testing the need for local calibration of the models (i.e., regression) using measurements of AADT covering different road types.