Deep learning to evaluate US NOx emissions using surface ozone predictions

Emissions of nitrogen oxides (NOx = NO + NO2) in the United States have declined significantly during the past three decades. However, satellite observations since 2009 indicate total column NO2 is no longer declining even as bottom-up inventories suggest continued decline in emissions. Multiple explanations have been proposed for this discrepancy including 1) the increasing relative importance of non-urban NOx to total column NO2, 2) differences between background and urban NOx lifetimes, and 3) that the actual NOx emissions are declining more slower after 2009. Here we use a deep learning model trained by NOx emissions and surface observations of ozone to assess consistency between the reported NOx trends between 2005-2014 and observations of surface ozone. We find that the 2005-2014 trend from older satellite-derived emission estimates produced at low spatial resolution best reproduce ozone in low NOx emission (background) regions, reflecting the blending of urban and background NOx in these low-resolution top-down analyses. The trend from higher resolution satellite-based estimates, which are more capable of capturing the urban emission signature, is in better agreement with ozone in high NOx emission regions, and is consistent with the trend based on surface observations of NO2. In contrast, the 2005-2014 trend from the US Environmental Protection Agency (EPA) National Emission Inventory (NEI) results in an underestimate of ozone. Our results confirm that the satellite-derived trends reflect anthropogenic and background influences and that the 2005-2014 trend in the NEI inventory is overestimating recent reductions in NOx emissions.