Uncertainty-aware downscaling of NO2 surface levels in urban environments

Air pollution is recognised as one of the leading environmental risks to global health, contributing to severe respiratory and cardiovascular diseases. In urban environments, air pollutant concentrations exhibit strong spatial variability at very fine scales, which cannot be adequately resolved by regional air quality models. The recently adopted Directive (EU) 2024/2881 establishes stricter regulatory standards, including a new annual mean limit of 20 μg/m³ for NO₂, which is frequently exceeded at specific urban locations, highlighting the need for high-resolution data fusion approaches that integrate air quality models with observational data to support exposure assessment and regulatory compliance.

Current state-of-the-art methods increasingly fuse Sentinel-5P (TROPOMI) tropospheric NO₂ columns with high-resolution geospatial proxies and modelled data to refine regional air quality outputs for urban-scale applications. However, high-resolution uncertainty quantification is largely absent from these products, limiting their interpretability. Establishing a clear methodology to provide this information is essential for effective air quality management, as stakeholders require reliable confidence metrics alongside best estimates to design robust mitigation strategies.

In this work, we present a statistical downscaling framework applied to CALIOPE, an operational air quality system integrating meteorological, emission, and photochemical models within a three-level nested configuration covering Europe, the Iberian Peninsula, and Catalonia. In the innermost domain, CALIOPE provides hourly forecasts at a 1 km × 1 km resolution. The proposed methodology bridges the gap between regional and urban scales by integrating high-resolution geospatial covariates (including traffic intensity networks, CORINE land-use data, terrain elevation, and distances to industrial sources) to produce concentration maps at the target high resolution. In particular, the proposed downscaling procedure operates on a non-uniform spatial mesh, achieving resolutions of up to 25 m near emission sources, with dense sampling along the road network and progressively coarser resolution towards the regional background. Sentinel-5P (TROPOMI) tropospheric NO₂ column data are interpolated to the final grid and incorporated as a spatially continuous covariate, ensuring regional consistency, particularly in areas lacking ground-based monitoring. The modelling strategy follows a source-oriented stratified approach inspired by area-oriented Kriging principles, separating traffic-influenced and background environments. Deterministic concentration trends are estimated using non-linear machine learning algorithms, including Random Forest and Gradient Boosting, while spatially correlated residuals are interpolated using ordinary kriging. Crucially, uncertainty quantification is explicitly integrated by propagating both model and spatial interpolation uncertainties, resulting in an uncertainty-aware product that provides local confidence estimates alongside predicted concentrations.

The framework is applied to estimate annual mean NO₂ concentrations for 2024 in Catalonia, Spain, serving as a high-resolution diagnosis for the regional government. Beyond standard concentration maps, the system provides stakeholders with probability of exceedance maps, relative to the regulatory thresholds, and pixel-level uncertainty metrics. Statistical performance evaluated through Leave-One-Out Cross-Validation demonstrates significant improvements over raw regional outputs, achieving an R² of 0.87 and reducing the Root Mean Square Error by 35% to 3.4 μg/m³. These results highlight the potential of the proposed approach to resolve complex urban patterns at regional scales for multiple cities, while supporting targeted public health interventions and evidence-based policy-making.