Estimating annual average daily traffic on local roads: Integrating spatial insights with machine learning

Abstract

Obtaining street-level annual average daily traffic (AADT) is crucial for environmental assessments, infrastructure planning, and sustainable transport policy. However, comprehensive traffic data collection remains resource intensive and spatially sparse. This study presents a scalable machine learning (ML) framework to predict AADT of road links across large networks. The framework integrates over 900 contextual features (e.g. demographics) with over 100 engineered spatial statistical features (e.g. eigenvector spatial filtering) to capture spatial dependencies and heterogeneity. Model performance is evaluated at over 19,000 count point locations in England and Wales with two spatially informed cross-validation (CV) strategies, i.e. sampling-intensity-weighted and spatial block CV, and achieves test R² of 0.80 and 0.72, respectively. Residual diagnostics further confirm the model's effectiveness in capturing spatial dependencies. This framework is then applied to estimate street-level AADT and vehicle kilometres travelled (VKT) by vehicle and fuel type across all road links in Greater London in 2021. Aggregated VKT estimates at the local authority level broadly align with official statistics, with median absolute relative differences of 8–10%. Agreement is stronger in data-rich central areas, while larger discrepancies occur in west and south outer London, reflecting limited observational coverage where spatial extrapolation is needed. To inform future traffic monitoring, we further classify local authorities by model disagreement and data density. By integrating spatial statistical theory with computationally efficient ML, the framework offers a bottom-up approach for large-scale spatial prediction in urban analytics, environmental modelling, and regional science. The disaggregated outputs also support high-resolution emissions modelling and inform sustainable transport planning.