A deep-learning-based classification of European air quality monitoring sites for model verification and data assimilation
Abstract. Assessing the spatial representativeness of observational data is a key challenge in atmospheric composition modelling, for model validation and data assimilation. In addition to the valuable metadata provided by network operators, the European CAMS (Copernicus Atmosphere Monitoring Service) regional consortium (Colette et al., 2025) also uses the results of an objective classification method originally based on the work of Joly and Peuch (2012). The method has been significantly improved in 2017, and Section 2 provides the first full published description of the LDA method that has produced site lists for the CAMS Regional Air Quality service each year since 2017. However, the method still had limitations, in particular a high sensitivity to the quality and quantity of data, which led to the rejection of certain time series. This paper presents a new method that aims to be more effective using recent machine learning and deep learning techniques. The new procedure now comprises three stages: i) data pre-processing including gap filling, ii) unsupervised outlier detection based on the statistical features of each annual time series, and iii) a 10-fold supervised classification using deep learning. For each species, the choice of models and hyperparameters has been fine-tuned separately. When applied to the eight years of data from 2018 to 2025, the new method yields improvements for all species and all performance metrics compared to the previous operational version. Case studies show that time series data provide a useful source of information for scrutinising and supplementing the network operator’s metadata. This new method is now being used operationally to provide information to the CAMS-regional project and has contributed to a substantial reduction in the amount of data discarded before data assimilation and model validation.