Inferring near-surface NO2 concentrations for Belgium using multiple machine learning models and TROPOMI data

Near-surface nitrogen dioxide (NO₂) is of great concern due to its impact on air quality and human health. Inferring the high-resolution spatiotemporal distribution of surface NO₂ is necessary to assess the NO₂ effects on human society and the ecosystem. Machine learning (ML) is an efficient approach to establishing a data-driven nonlinear mapping between targets and predictors. Various ML models have been used in past studies to estimate ground NO₂ distributions from satellite observations (e.g., TROPOMI and OMI) and ancillary predictors (e.g., meteorology, land cover, and anthropogenic emissions) with good resolution, efficiency, and accuracy. In spite of these successes, the application of ML to infer near-surface NO₂ remains challenging due to model stability issues and missing uncertainty estimations. In this research, we compare different ML models with respect to their predictive accuracy and spatiotemporal patterns. Moreover, we analyze the impact of the satellite remote sensing dataset and different predictors on model predictions using ML interpretation techniques. Based on these, we further investigate the possibility of assembling various ML models to provide a reliable ground NO₂ estimation with uncertainty assessments. Overall, this study explores how ML models can be used to produce surface NO₂ products, offering a perspective on practical applications for ML methods in atmospheric science.