Technical note - Improving the European air quality forecast of the Copernicus Atmosphere Monitoring Service using machine learning techniques

Aug 5, 2026

Upload by Himbo

Model output statistics based on machine learning are applied to downscale European air quality forecasts from CAMS (Copernicus Atmosphere Monitoring Service) across hundreds of monitoring sites. Two strategies are compared: a “local” MOS training per station and a “global” model trained on the full geographical domain. The study evaluates performance for PM10, PM2.5, O3, and NO2 using predictive-error and exceedance detection scores. Results show both approaches improve the raw ensemble forecast, with the global method often matching or exceeding local performance and giving the best RMSE with random forests, while gradient boosting better detects EU-threshold exceedances.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

Introduction

Model output statistics (MOS) and motivation

Two MOS strategies: local vs global

Pollutants and evaluation metrics

Machine learning algorithms and key findings

Atmos. Chem. Phys., 23, 5317–5333, 2023 [https://doi.org/10.5194/acp-23-5317-2023](https://doi.org/10.5194/acp-23-5317-2023)[ ](https://doi.org/10.5194/acp-23-5317-2023)© Author(s) 2023 . This work is distributed under the Creative Commons Attribution 4 .0 License. Technical note: Improving the European air quality forecast of the Copernicus Atmosphere Monitoring Service using machine learning techniques Jean-Maxime Bertrand, Frédérik Meleux, Anthony Ung, Gaël Descombes, and Augustin Colette Institut National de l'Environnement Industriel et des Risques (INERIS), Parc Alata, BP2, 60550 Verneuil-en-Halatte, France Correspondence: Frédérik Meleux (frederik.meleux@ineris.fr) Received: 14 November 2022 – Discussion started: 18 November 2022 Revised: 17 March 2023 – Accepted: 20 March 2023 – Published: 11 May 2023 Abstract. Model output statistics (MOS) approaches relying on machine learning algorithms were applied to downscale regional air quality forecasts produced by CAMS (Copernicus Atmosphere Monitoring Service) at hundreds of monitoring sites across Europe. Besides the CAMS forecast, the predictors in the MOS typically include meteorological variables but also ancillary data. We explored ﬁrst a “local” approach where speciﬁc models are trained at each site. An alternative “global” approach where a single model is trained with data from the whole geographical domain was also investigated. In both cases, local predictors are used for a given station in predictive mode. Because of its global nature, the latter approach can capture a variety of meteorological situations within a very short training period and is thereby more suited to cope with operational constraints in relation to the training of the MOS (frequent upgrades ofthe modelling system, addition of new monitoring sites) . Both approaches have been implemented using a variety of machine learning algorithms: random forest, gradient boosting, and standard and regularized multi-linear models. The quality of the MOS predictions is evaluated in this work for four key pollutants, namely particulate matter (PM 10 and PM2:5), ozone (O 3) and nitrogen dioxide (NO 2), according to scores based on the predictive errors and on the detection of pollution peaks (exceedances of the regulatory thresholds) . Both the local and the global approaches signiﬁcantly improve the performances of the raw ensemble forecast. The most important result of this study is that the global approach competes with and can even outperform the local approach in some cases. This global approach gives the best RMSE scores when relying on a random forest model for the prediction of daily mean, daily max and hourly concentrations. By contrast, it is the gradient boosting model which is better suited for the detection of exceedances of the European Union regulated threshold values for O 3 and PM 10. 1 Introduction Outdoor air pollution induced by natural sources and human activities remains a major environmental and health issue worldwide. Producing reliable short-term forecasts of pollutant concentrations is a key challenge to support national authorities in their duties regarding the European Air Quality Directive, like planning and communications about the air quality status towards the general public in order to limit the exposure of populations. Progress in computing technologies during the last decades has allowed for the rise of large-scale chemistry transport models (CTMs) which provide a comprehensive view of the air quality on a given time period and geographical domain by solving the differential equations that govern the transport and transformation of pollutants in the atmosphere. An overview of such deterministic air quality forecasting systems operating in Europe was provided by Zhang et al. (2012) . Ensembles of several CTMs have also been used in order to improve single-model forecasts (Delle Monache and Stull, 2003; Wilczak et al., 2006) . Such an ensemble approach is cur

FAQs

What is the goal of the technical note?

To improve European short-term air quality forecasts from CAMS by applying machine learning-based model output statistics for downscaling and bias/error reduction.

How do the “local” and “global” approaches differ?

The local approach trains a separate model at each monitoring site, while the global approach trains a single model using data from the entire geographic domain and then applies it to stations.

Which machine learning models perform best for accuracy versus exceedance detection?

Random forest provides the best RMSE scores for daily mean, daily max, and hourly concentrations, while gradient boosting is better at detecting exceedances of EU regulatory thresholds for O3 and PM10.

Aug 5, 2026

Upload by Himbo