Improved global daily nitrogen dioxide concentrations from 2005 to 2023 derived using a deep learning approach

Abstract

Nitrogen dioxide (NO2) is a critical air pollutant with significant environmental and human health impacts, yet global and long-term NO2 datasets with daily continuity and fine spatial resolution remain limited. In this study, we construct a continuous global daily NO2 concentration (https://doi.org/10.5281/zenodo.13842191, Mu and Tao, 2025) spanning from 2005 to 2023 at a 0.1° resolution using the advanced Air Transformer deep learning framework that integrates satellite observations, ground-based measurements, meteorological reanalysis, land-use information, and auxiliary geophysical variables. The resulting dataset shows robust performance across diverse regions and pollution regimes, with improved spatial consistency and reduced biases relative to existing global products. Based on this dataset, we characterize the spatiotemporal evolution of global NO2 concentrations over the past two decades. Global annual mean NO2 increased from 2005 to 2015, followed by a moderate decline during 2016–2019, a pronounced decrease in 2020 associated with COVID-19-related reductions in economic activity and transportation, and a partial rebound thereafter, reaching 3.38 ppbv in 2023. The Northern Hemisphere and tropical regions largely followed the global trend, whereas the Southern Hemisphere exhibited distinct behaviour, with relatively stable or declining NO2 levels prior to 2015, a sharp decrease in 2020, and a stronger post-pandemic rebound during 2021–2023. As one of the global, multi-decadal NO2 datasets with daily resolution, this dataset provides a valuable resource for air quality assessment, exposure analysis, and atmospheric model evaluation.