Technical Note: Accurate, reliable and high resolution air quality predictions by improving the Copernicus Atmosphere Monitoring Service using machine learning techniques

Abstract. Starting from the regional air quality forecasts produced by the Copernicus Atmosphere Monitoring Service (CAMS), we propose a novel post-processing approach to improve and downscale results on a finer scale. Our approach is based on the combination of Ensemble Model Output Statistics (EMOS) with a spatio-temporal interpolation process performed through the Stochastic Partial Differential Equation-Integrated Nested Laplace Approximation (SPDE-INLA). Our interpolation approach includes several spatial and spatio-temporal predictors, including meteorological variables. A use-case is provided, scaling down the CAMS forecasts on the Italian peninsula. The calibration is focused on the concentrations of several air quality pollutants (PM₁₀, PM_2.5, NO₂ and O₃) at daily resolution from a set of 750 monitoring sites, distributed throughout the Italian country. Our results show the key role played by conditioning variables to improve the forecast capabilities of ensemble predictions, thus allowing a net improvement of the calibration with respect to ordinary EMOS strategies. From a deterministic point of view, the predictive model performance shows a significant improvement of the performance of the raw ensemble forecast, with an almost zero bias, significantly reduced root mean square errors and correlations almost always higher than 0.9 for each pollutant; moreover, the post-processing approach is able to significantly improve the prediction of exceedances, even for very low thresholds, such as those recently recommended by the World Health Organisation. This is particularly significant if a forecasting approach is to be used to predict air quality conditions and plan adequate human health protection measures, even for low alert thresholds. From a probabilistic point of view, the forecast quality was verified in terms of reliability and credible intervals. After post-processing, the predictive probability density functions were sharp, and much better calibrated than the raw ensemble forecast. Finally, we present some additional outcomes based on a set of gridded (4 km x 4 km) daily maps covering the whole Italian country, for the detection of areas where pollution peaks forecasts (exceedances of the regulatory thresholds) occur.