Nitrogen dioxide (NO₂) column density measurements from satellites have been widely used in constraining emissions of nitrogen oxides (NO_x = NO + NO₂). However, the utility of these measurements is impacted by reduced observational coverage due to cloud cover and their reduced sensitivity toward the surface. Combining the information from satellites with surface observations of NO₂ will provide greater constraints on emission estimates of NO_x. We have developed a deep-learning (DL) model to integrate satellite data and in situ observations of surface NO₂ to estimate NO_x emissions in China. A priori information for the DL model was obtained from satellite-derived emissions from the Tropospheric Chemistry Reanalysis (TCR-2). A two-stage training strategy was used to integrate in situ measurements from the China Ministry of Ecology and Environment (MEE) observation network with the TCR-2 data. The DL model is trained from 2005 to 2018 and evaluated for 2019 and 2020. The DL model estimated a source of 19.4 Tg NO for total Chinese NO_x emissions in 2019, which is consistent with the TCR-2 estimate of 18.5 ± 3.9 Tg NO and the 20.9 Tg NO suggested by the Multi-resolution Emission Inventory for China (MEIC). Combining the MEE data with TCR-2, the DL model suggested higher NO_x emissions in some of the less-densely populated provinces, such as Shaanxi and Sichuan, where the MEE data indicated higher surface NO₂ concentrations than TCR-2. The DL model also suggested a faster recovery of NO_x emissions than TCR-2 after the Chinese New Year (CNY) holiday in 2019, with a recovery time scale that is consistent with Baidu “Qianxi” mobility data. In 2020, the DL-based analysis estimated about a 30 % reduction in NO_x emissions in eastern China during the COVID-19 lockdown period, relative to pre-lockdown levels. In particular, the maximum emission reductions were 42 % and 30 % for the Jing-Jin-Ji (JJJ) and the Yangtze River Delta (YRD) mega-regions, respectively. Our results illustrate the potential utility of the DL model as a complementary tool for conventional data-assimilation approaches for air quality applications.

Surface ozone (O₃) concentrations in China have increased largely in the past decade. An accurate understanding of O₃ pollution evolution is critical for making effective regulatory policies. Here we integrate data- and process-based models to explore the drivers of the observed summertime surface O₃ change in the North China Plain (NCP) over 2015-2021. The data-based model by the deep learning (DL) suggests the reverse of meteorological contributions to the observed O₃ change, i.e., 0.14 ppb/y in 2015-2019 and -1.74 ppb/y in 2019-2021. This is mainly resulted from the reversed changes in meteorological variables in surface air temperature and relative humidity. The simulations from a global chemical transport model, GEOS-Chem, also support those results, i.e., the meteorological contribution to O₃ changes are 0.26 ppb/y in 2015-2019 and -0.74 ppb/y in 2019-2021. Furthermore, our analysis exhibits possible weakened anthropogenic contributions to surface O₃ rise, for example, 1.53 and 0.54 ppb/y by DL in 2015-2019 and 2019-2021, respectively. Similarly, GEOS-Chem simulations suggest an accelerated decrease in surface O₃ concentrations driven by the decline in nitrogen dioxide (NO₂) concentrations, i.e., approximately 0.4 and 1.2 ppb in 2015-2019 and 2019-2021, respectively. The combined effects of meteorological and anthropogenic contributions led to a significant decrease in surface O₃ concentrations by -1.20 ppb/y in 2019-2021. The findings in this work offer valuable insights to mitigate O₃ pollution in China.

Observations of carbon monoxide (CO) from the Measurements Of Pollution In The Troposphere (MOPITT) instrument aboard the Terra spacecraft were expected to have an accuracy of 10 % prior to the launch in 1999. Here we evaluate MOPITT Version 7 joint (V7J) thermal-infrared and near-infrared (TIR–NIR) retrieval accuracy and precision and suggest ways to further improve the accuracy of the observations. We take five steps involving filtering or bias corrections to reduce scatter and bias in the data relative to other MOPITT soundings and ground-based measurements. (1) We apply a preliminary filtering scheme in which measurements over snow and ice are removed. (2) We find a systematic pairwise bias among the four MOPITT along-track detectors (pixels) on the order of 3–4 ppb with a small temporal trend, which we remove on a global scale using a temporally trended bias correction. (3) Using a small-region approximation (SRA), a new filtering scheme is developed and applied based on additional quality indicators such as the signal-to-noise ratio (SNR). After applying these new filters, the root-mean-squared error computed using the local median from the SRA over 16 years of global observations decreases from 3.84 to 2.55 ppb. (4) We also use the SRA to find variability in MOPITT retrieval anomalies that relates to retrieval parameters. We apply a bias correction to one parameter from this analysis. (5) After applying the previous bias corrections and filtering, we compare the MOPITT results with the GGG2014 ground-based Total Carbon Column Observing Network (TCCON) observations to obtain an overall global bias correction. These comparisons show that MOPITT V7J is biased high by about 6 %–8 %, which is similar to past studies using independent validation datasets on V6J. When using TCCON spectrometric column retrievals without the standard airmass correction or scaling to aircraft (WMO scale), the ground- and satellite-based observations overall agree to better than 0.5 %. GEOS-Chem data assimilations are used to estimate the influence of filtering and scaling to TCCON on global CO and tend to pull concentrations away from the prior fluxes and closer to the truth. We conclude with suggestions for further improving the MOPITT data products.

Emissions of nitrogen oxides (NO_x = 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 (a) the increasing relative importance of nonurban NO_x to total column NO₂, (b) differences between background and urban NO_x lifetimes, and (c) that the actual NO_x emissions are declining more slowly after 2009. Here, we use a deep learning model trained by NO_x emissions and surface observations of ozone to assess consistency between the reported NO_x trends between 2005 and 2014 and observations of surface ozone. We find that the satellite‐derived trends best reproduce ozone in low NO_x emission (background) regions. The 2010–2014 trend from older satellite‐derived emission estimates produced at low spatial resolution results in the largest bias in surface ozone in regions with high NO_x emissions, reflecting the blending of urban and background NO_x in these low‐resolution top‐down analyses. In contrast, 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 NO_x emission regions, and is consistent with the trend based on surface observations of NO₂. Our results confirm that the satellite‐derived trends reflect anthropogenic and background influences.

Uptake of atmospheric carbon by the ocean, especially at high latitudes, plays an important role in offsetting anthropogenic emissions. At the surface of the Southern Ocean south of 30^∘S, the ocean carbon uptake, which had been weakening in 1990s, strengthened in the 2000s. However, sparseness of in-situ measurements in the ocean interior make it difficult to compute changes in carbon storage below the surface. Here we develop a machine-learning model, which can estimate concentrations of dissolved inorganic carbon (DIC) in the Southern Ocean up to 4 km depth only using data available at the ocean surface. Our model is fast and computationally inexpensive. We apply it to calculate trends in DIC concentrations over the past three decades and find that DIC decreased in the 1990s and 2000s, but has increased, in particular in the upper ocean since the 2010s. However, the particular circulation dynamics that drove these changes may have differed across zonal sectors of the Southern Ocean. While the near-surface decrease in DIC concentrations would enhance atmospheric CO₂ uptake continuing the previously-found trends, weakened connectivity between surface and deep layers and build-up of DIC in deep waters could reduce the ocean’s carbon storage potential.

There has been a proliferation of dense observing systems to monitor greenhouse gas (GHG) concentrations over the past decade. Estimating emissions with these observations is often done using an atmospheric transport model to characterize the source–receptor relationship, which is commonly termed the measurement “footprint”. Computing and storing footprints using full-physics models is becoming expensive due to the requirement to simulate atmospheric transport at high resolution. We present the development of FootNet, a deep-learning emulator of footprints at the kilometer scale. We train and evaluate the emulator using footprints simulated with a Lagrangian particle dispersion model (LPDM). FootNet predicts the magnitudes and extents of footprints in near real time with high fidelity. We identify the relative importance of input variables of FootNet for improving the interpretability of the model. Surface winds and a precomputed Gaussian plume from the receptor are identified as the most important variables for footprint emulation. The FootNet emulator developed here may help address the computational bottleneck of flux inversions using dense observations.