Estimating Anthropogenic NOx Emissions Using Convolutional Neural Networks with Horizontal Transport Considerations

Nitrogen oxides (NO_x = NO + NO₂) are significant air pollutants, mainly emitted from anthropogenic sources. Bottom-up methods for the estimation of anthropogenic NO_x emissions are based on energy consumption data, which, if outdated, result in a delayed response in the produced emission inventories. The TROPOspheric Monitoring Instrument (TROPOMI) provides high-resolution NO₂ column densities, offering valuable data for estimating NO_x emissions. Given the short atmospheric lifetime of NO_x, horizontal transport influences over distances within a few tens to a few hundred kilometers must be taken into account. To address this, we developed a convolutional neural network (CNN) which incorporates the NO₂ divergence and horizontal transport features to estimate the anthropogenic NO_x emissions. Our model operates on a monthly timescale with a spatial resolution of 0.1°, utilizing TROPOMI NO₂ column densities and ERA5 wind field data as inputs, and the EDGARv8.1 0.1° gridded NO_x inventories as targets. The training set comprised data from 2019 and 2020, of which 70 % were used for training, and the remaining 30 % for testing of the model. The model achieved an R² of 0.922 and an RMSE of 11.214 Mg/month on the test set when estimating NO_x emissions in Europe and the USA. Additionally, the model demonstrated temporal generalization capabilities, achieving an average R² of 0.853 (±0.066) and an average RMSE of 16.545 (±3.804) Mg/month in monthly estimations for Europe and the USA during 2021-2022. The proposed method integrates satellite observations with emission inventories, employing CNNs to facilitate rapid updates of anthropogenic NO_x emissions.