Using open data to derive parsimonious data-driven models for uncovering the influence of local traffic and meteorology on air quality: The case of Madrid

Air pollution remains a critical public health and environmental challenge, particularly in urban areas where traffic emissions and meteorological conditions strongly influence air quality. While Machine Learning (ML) techniques have been increasingly used to model pollutant concentrations, many existing studies rely on complex architectures that often integrate multiple heterogeneous data sources. In contrast, this study presents a parsimonious, data-driven ML model that predicts local hourly concentrations of key pollutants-NO₂, O₃, PM_2.5, and PM₁₀-in Madrid using only open data sources. A key factor of our approach is the incorporation of hourly road traffic data collected in the immediate vicinity of each pollutant monitoring station as a predictor. This localized traffic information, combined with local meteorological data, allows our model to outperform other existing solutions that often depend on historical and/or proprietary data. Our results clearly demonstrate that better data might surpass the benefits of more complex ML architectures. The model achieves strong predictive accuracy, with test R² scores ranging from 0.77 to 0.86 for NO₂, 0.8 to 0.85 for O₃, 0.63 to 0.82 for PM_2.5, and 0.68 to 0.95 for PM₁₀. This remarkable performance underscores the utility of dense networks of vehicle count sensors providing high-resolution spatiotemporal traffic data as a critical input for accurate urban air quality modeling. Additionally, we conducted a sensitivity analysis to assess the impact of reducing vehicle emissions on local NO₂ levels, offering actionable insights for policymakers. The findings highlight the potential of open-data-driven models in urban air quality management, providing a scalable, cost-effective, and interpretable tool to support evidence-based decision-making and environmental policy design.