A rapid-updating method for anthropogenic NO x emissions based on convolutional neural networks and TROPOMI NO₂ observations

This is the authors' abstract. We don't add key points for this paper.

Nitrogen oxides (NO x = NO + NO2) are air pollutants primarily emitted from anthropogenic sources, however, bottom-up anthropogenic NO x emission inventories often suffer from update delays. The TROPOspheric Monitoring Instrument (TROPOMI) provides near-real-time, high-resolution NO2 column densities, offering new opportunities for timely NO x emission prediction. Given the short lifetime of NO x in the lower troposphere, local emissions can influence observations over distances of ~100 km. In this study, we develop a convolutional neural network to predict anthropogenic NO x emissions from TROPOMI NO₂ observations. The model operates at a monthly temporal resolution and 0.1 ° × 0.1 ° spatial resolution over Europe and the continental United States, using TROPOMI NO2 columns and ERA5 wind fields as inputs, with EDGARv8.1 emissions as training targets. Trained on 2019–2020 data, the model achieves an R2 of 0.958 and an RMSE of 8.194 Mg/month/cell on the test set. It demonstrates strong temporal generalisation, with an average R2 of 0.927 at the grid scale and 0.966 at the national scale during 2021–2022, although spatial generalisation remains limited. The model is applied to extend gridded NO x emissions to December 2025. Overall, the proposed satellite-driven deep learning approach enables accurate, high-resolution, and near-real-time updates of anthropogenic NO x emissions.