Machine Learning for Estimating Daily Traffic on Local Roads

This technical report presents a methodology that employs machine learning (ML) techniques to estimate annual average daily traffic (AADT) on local roads in England and Wales. The study addresses the challenges associated with collecting AADT data, particularly the underrepresentation of minor roads in traditional data collection methods. The proposed approach integrates ML with spatial statistics and utilizes extensive geospatial data to enhance AADT estimation accuracy. A lightGBM model is applied, incorporating over 900 spatial features and accounting for spatial autocorrelation. The report details the feature selection process using the Boruta algorithm and describes the model evaluation through a tailored cross-validation method designed for spatial data. Results indicate that separate models for major and minor roads yield improved predictive performance compared to a universal model. The findings support the potential for ML methods to provide reliable AADT estimates, which can be beneficial for pollution and carbon emissions assessments.