Volume 124, Issue 2 p. 1170-1193
Research Article
Free Access

Quantifying Emissions of CO and NOx Using Observations From MOPITT, OMI, TES, and OSIRIS

X. Zhang

Corresponding Author

X. Zhang

Department of Physics, University of Toronto, Toronto, Ontario, Canada

Correspondence to: X. Zhang,

[email protected]

Search for more papers by this author
D. B. A. Jones

D. B. A. Jones

Department of Physics, University of Toronto, Toronto, Ontario, Canada

Search for more papers by this author
M. Keller

M. Keller

Department of Physics, University of Toronto, Toronto, Ontario, Canada

Search for more papers by this author
T. W. Walker

T. W. Walker

Department of Civil and Environmental Engineering, Carleton University, Ottawa, Ontario, Canada

Search for more papers by this author
Z. Jiang

Z. Jiang

School of Earth and Space Sciences, University of Science and Technology of China, Hefei, Anhui, China

Search for more papers by this author
D. K. Henze

D. K. Henze

Department of Mechanical Engineering, University of Colorado at Boulder, Boulder, CO, USA

Search for more papers by this author
H. M. Worden

H. M. Worden

Atmospheric Chemistry Observations and Modeling Laboratory, National Center for Atmospheric Research, Boulder, CO, USA

Search for more papers by this author
A. E. Bourassa

A. E. Bourassa

Department of Physics & Engineering Physics, University of Saskatchewan, Saskatoon, Saskatchewan, Canada

Search for more papers by this author
D. A. Degenstein

D. A. Degenstein

Department of Physics & Engineering Physics, University of Saskatchewan, Saskatoon, Saskatchewan, Canada

Search for more papers by this author
Y. J. Rochon

Y. J. Rochon

Air Quality Research Division, Environment and Climate Change Canada, Toronto, Ontario, Canada

Search for more papers by this author
First published: 07 January 2019
Citations: 9
The copyright line for this article was changed on 18 APR 2019 after original online publication.


We use the GEOS-Chem four-dimensional variational data assimilation system to estimate emissions of carbon monoxide (CO) and nitrogen oxides (NOx) in November 2009 and July 2010. We assimilated CO retrievals from The Measurements of Pollution In The Troposphere (MOPITT), ozone (O3) retrievals from the Tropospheric Emission Spectrometer (TES) and the Optical Spectrograph and InfraRed Imager System (OSIRIS), and NO2 columns from the Ozone Monitoring Instrument (OMI). By assimilating data for multiple chemical species with the four-dimensional variational scheme we obtain a consistent chemical state over the assimilation period. The modeled O3 bias was reduced to less than 3.5 ppbv everywhere, relative to independent aircraft O3 data, except in the high-latitude upper troposphere and lower stratosphere. We found that in November, the CO and NOx emission estimates obtained with the multiple species assimilation were generally comparable to those inferred from assimilating only MOPITT or OMI data, respectively, whereas in July the differences were larger. For the main anthropogenic source regions, the inferred North American and European emissions responded most strongly to the multispecies information in the assimilation. In July, North American and European CO emission estimates differed by 31% and 65% relative to the MOPITT-only estimates, respectively. We also found large differences for the North American and European NOx emission estimates, which differed by 27% and 16% relative to the OMI-only estimates, respectively. Our results highlight the potential benefit of exploiting the additional constraints offered by multispecies chemical data assimilation.

Key Points

  • Multispecies data assimilation provides a means of indirectly estimating tropospheric OH
  • Multispecies DA better fits assimilated MOPITT CO, whereas OMI-only DA better fits OMI NO2 data
  • European emission estimates are most sensitive to the multispecies information in GEOS-Chem

1 Introduction

Surface pollutants such as carbon monoxide (CO) and nitrogen oxides (NOx = NO + NO2) have a significant impact on air quality as well as the evolution of the chemistry of the troposphere (L. Zhang, Jacob, et al., 2008). Atmospheric CO is emitted from incomplete combustion and produced from the oxidation of atmospheric hydrocarbons. It is both a major sink for the atmospheric hydroxyl radical (OH) and a precursor of tropospheric O3. Since OH is the major oxidant in the atmosphere, CO emissions will indirectly impact the oxidative capacity of the atmosphere. Atmospheric NOx is produced from combustion, lightning, and microbial processes in soils and is a major precursor of tropospheric O3. As a consequence, changes in the abundance of both CO and NOx will impact the tropospheric ozone (O3) budget. Since preindustrial times, CO sources have increased by a factor of two or three (Haan et al., 1996). The current estimates of NOx emissions from fuel combustion are approximately 25 TgN/year, three times greater than the preindustrial NOx emission estimates (Müller & Stavrakou, 2005). However, recent satellite observations of CO and NO2 have suggested that there have been reductions in emissions of CO across much of the northern hemisphere in the past decade, and more heterogeneous changes in emissions of NOx, with decreases in much of North America and Europe and increases in parts of Asia (e.g., Duncan et al., 2016; Gu et al., 2014; Jiang et al., 2017; Worden et al., 2013).

Estimates of the changing regional emissions of CO and NOx are uncertain, and a variety of inverse modeling studies conducted so far have focused on better quantifying these emissions employing surface observations (e.g., Hooghiemstra et al., 2012; Pétron et al., 2002; Pison et al., 2009), aircraft measurements (e.g., Heald et al., 2004; Palmer et al., 2003), or satellite observations (e.g., Arellano et al., 2004; Fortems-Cheiney et al., 2011; Gaubert et al., 2017; Jones et al., 2009; Jiang et al., 2013, 2017; Kopacz et al., 2010; Müller et al., 2018; Stavrakou & Müller, 2006; Yin et al., 2015). Kopacz et al. (2010), for example, based on satellite observations, suggested that wintertime CO emissions in Europe, North America, and Asia were underestimated by as much as 60%. Hooghiemstra et al. (2012) compared their inferred estimates of regional CO emissions using surface and satellite observations with previously published values and found that discrepancies in the regional emission estimates among the different studies could be more than 100% for some regions (such as North America), despite the fact that the inferred total global emission were consistent across the studies. These differences in the regional emission estimates in the literature reflect the influence of model errors, differences in the employed data assimilation approaches, and the different observations assimilated in the studies.

Similar to CO, inverse modeling studies of NOx have used in situ aircraft measurements (e.g., Heland et al., 2002) and satellite observations (e.g., Boersma et al., 2008; Martin et al., 2003; Müller & Stavrakou, 2005; Miyazaki, Eskes, & Sudo, 2012; van der et al., 2008) to estimate regional NOx emissions. However, until recently, these inverse modeling studies did not account for the impact of possible discrepancies in the model chemistry on the inferred NOx emission estimates. For example, the relationship between the NOx emissions and the observed NO2 column abundance requires knowledge of the NOx/NO2 ratio, which could be influenced by biases in the modeled O3 abundances. Furthermore, for those studies using space-based column observations of NO2, discrepancies in the partitioning of total reactive nitrogen in the free troposphere could impact the modeled NO2 column abundance. Using a regional chemical transport model (CTM), Chai et al. (2006) demonstrated that assimilating aircraft observations of O3, NOx, nitric acid (HNO3), lumped organic nitrates (RO3), and peroxyacetyl nitrate (PAN) improved their model prediction of O3 and total reactive nitrogen. L. Zhang, Constantinescu, et al. (2008) also improved their model prediction of NO2 by simultaneously assimilating satellite observations of O3 and formaldehyde (HCHO) together with surface observations of O3 in their regional CTM. Miyazaki, Eskes, Sudo, et al. (2012) were the first to assimilate a suite of global space-based observations of atmospheric trace gases to constrain the chemical mechanism in their CTM in the context of their inversion analysis to quantify global NOx emissions. Using the CHASER-DAS CTM, with a local ensemble transform Kalman Filter data assimilation scheme, they simultaneously optimized 35 tropospheric tracers as well as the emissions of CO and NOx (from the surface and lightning). They assimilated observations of CO from the Measurements of Pollution in The Troposphere (MOPITT) instrument, O3 from the Tropospheric Emission Spectrometer (TES) and the Microwave Limb Sounder, NO2 from the Ozone Monitoring Instrument (OMI), and HNO3 from Microwave Limb Sounder (Miyazaki, Eskes, Sudo, et al., 2012).

Here we use the GEOS-Chem four-dimensional variational (4D-Var) data assimilation system to conduct a multispecies data assimilation (MSA) to quantify NOx and CO emissions, following the approach of Miyazaki, Eskes, Sudo, et al. (2012). We assimilate O3 data from TES and the Optical Spectrograph and InfraRed Imager System (OSIRIS), NO2 column abundances measured by OMI, and retrieved CO profiles from MOPITT. Müller and Stavrakou (2005) were the first to consider the benefit of jointly assimilating NO2 and CO observations. They found that assimilating NO2 and CO observations to estimate NOx and CO emissions resulted in a posteriori CO and NO2 fields that were in better agreement with independent observations than those based on only the assimilation of CO observations. Moreover, including the NO2 in the inversion resulted in OH changes and CO precursor changes such as volatile organic compounds (VOCs), which, in turn, impact the atmospheric CO abundance through the coupling in the CO-OH-NOx-VOC chemistry. Discrepancies in OH is a particular challenge for inverse modeling of CO emissions. For example, Jiang et al. (2015) used different OH fields from two versions of the GEOS-Chem model (versions v8-02-01 and v5-07-08) to estimate summertime CO emissions. They found that the differences in the regional emission estimates could differ by a factor of two. More recently, Müller et al. (2018) showed the resulting CO columns in their inversion analysis best match independent data sets when the OH abundance in the northern hemisphere was 18% lower than the OH estimates derived from methylchloroform observations. In our study, we are interested in using the MSA to understand the impact of the tropospheric chemical coupling on the inferred CO and NOx emissions. We are particularly interested in investigating the impact on the inversion analysis of improving the modeled tropospheric O3 field, since tropospheric O3 is a key precursor of OH.

A key difference between the work here and that of Miyazaki, Eskes, Sudo, et al. (2012) is our use of the 4D-Var assimilation scheme. With the local ensemble transform Kalman Filter approach, Miyazaki, Eskes, Sudo, et al. (2012) had to employ localization to minimize the impact of long-range correlations, which could limit the chemical coupling among species. Furthermore, they used a short assimilation window of 90 min, which limits the model's transport impact on optimizing surface emissions of CO and NOx. In contrast, we use a long assimilation window of 2 weeks. This was chosen to ensure that sufficient observations are ingested to constrain the model states. In the 4D-Var scheme, the state vector is optimized to produce a smoothed model trajectory, which would allow the model forecast to best reproduce the assimilated observations over the assimilation period (Trémolet, 2006). As a result of this time dependence, the 4D-Var scheme inherently produces a consistent chemical state across the assimilation period (Eibern & Schmidt, 1999). Due to expensive computational costs of running MSA, we restrict the study periods to November 2009 and July 2010. We focus on November 2009 to take advantage of the observations from the HIAPER Pole-to-Pole Observations (HIPPO) aircraft campaign to evaluate the model. We also chose July 2010 to contrast the seasonal response of the assimilation.

The structure of the paper is as follows. Section 2 presents the details of the employed data sets, and introduces the inversion methodology. The GEOS-Chem model and our experiment setup is also described in section 15. Section 3 discusses the data assimilation and the validation results, followed by the uncertainties of our studies. Section 4 summarizes the results and provides implications for future MSA studies.

2 Data and Methods

2.1 Data

2.1.1 MOPITT CO Observations

MOPITT is an instrument on the EOS-Terra satellite, which was launched in December 1999. The satellite is in a Sun-synchronous orbit that crosses the equator at 10:30/22:30 local time. The MOPITT observational strategy consists of a 612 km cross-track scan that provides global coverage every 3 days. Atmospheric CO is retrieved on 10 pressure levels (the surface, 900, 800, 700, 600, 500, 400, 300, 200, and 100 hPa) using an optimal estimation approach (Rodgers, 2000) from measurements at 2.3 μm in the near infrared (NIR) and at 4.7 μm in the thermal infrared (TIR). The retrieval is conducted with respect to the log10 of the CO volume mixing ratio, and can be written mathematically as
where AMOP is the averaging kernel of the MOPITT CO retrievals, which represents the vertical sensitivity of the retrieval to the true CO state in the atmosphere; z is the true CO state; urn:x-wiley:jgrd:media:jgrd55198:jgrd55198-math-0002 is the a priori MOPITT profile, which was obtained from simulations from the MOZART-4 model; and ϵ is the measurement error. Details of the retrieval are described in Worden et al. (2010) and Deeter et al. (2011). In this study, we use version V7J TIR/NIR retrievals and assimilate only the daytime data. We use MOPITT profiles because the multispectral retrievals provide sensitivity to CO in the lower and upper troposphere (Worden et al., 2010). Deeter et al. (2017) validated the retrievals with aircraft measurements and found that the biases were small, 2.8% at the surface, −3.4% at 600 hPa, and −1.9% at 400 hPa. The largest biases were 4.2% at 200 hPa. Here we assimilate only the CO retrievals at altitudes below 200 hPa.

2.1.2 TES O3 Observations

TES is an infrared Fourier-transform spectrometer that was launched on the EOS Aura satellite in July 2004. It is in a Sun-synchronous orbit at an altitude of 705 km, with an equatorial crossing time of 13:43 local time, and with a 16-day repeat cycle. The nadir observations in the global survey mode have a footprint of 5.3 × 8.3 km and are spaced about 2° along the orbit track. Although the observational coverage from TES is sparse, compared to that from OMI, Parrington et al. (2008) showed that the TES O3 profiles provide sufficient information to significantly improve modeled O3 between 800 and 400 hPa when assimilated into the GEOS-Chem model. The O3 profiles are retrieved from measurements around 9.6 μm using an optimal estimation approach. The retrievals are conducted with respect to the natural logarithm (ln) of the volume mixing ratio on a 67-level vertical grid. Similar to the MOPITT CO retrievals, the TES O3 profile retrievals can be expressed mathematically as
where ATES is the TES averaging kernel, urn:x-wiley:jgrd:media:jgrd55198:jgrd55198-math-0004 is the TES a priori O3 profile, and z is the true O3 state. Details of the TES O3 retrievals are described in Bowman et al. (2006). Here we use version 4 of the TES O3 data product, which has good agreement in the troposphere with the later versions of the retrievals (Herman & Osterman, 2014). The spatially dependent biases reported by Nassar et al. (2008) were also removed to improve the quality of the observations.

2.1.3 Odin OSIRIS O3 Observations

OSIRIS was launched onboard the Odin Satellite in February 2001. The satellite is in a polar, sun-synchronous, near terminator orbit that provides observational coverage limited to the summer hemisphere and year-round in the tropics. The OSIRIS spectrograph scans the limb from 7 to 70 km at the tangent altitude and measures scattered sunlight from 280 to 820 nm. We use the O3 profile retrievals from version 5.07 of the SaskMART O3 product (Degenstein et al., 2009). The retrievals have a vertical resolution of about 2 km and have been validated by Adams et al. (2014), who reported a bias of less than 5% between 13.5 and 54.5 km relative to O3 data from the Stratospheric Aerosol and Gas Experiment II. The precision increases from 5% to 15% between 10 and 20 km to better than 5% between 25 and 50 km(Bourassa et al., 2012). In our analysis, we assimilate OSIRIS data between 18 and 53 km. Our focus is to reduce model biases in the upper troposphere and lower stratosphere (UTLS).

2.1.4 OMI Tropospheric NO2 Column Observations

The OMI instrument is a nadir imaging spectrograph onboard the Aura satellite. The instrument measures scattered sunlight between 270 and 550 nm. The instrument has a nadir footprint size of 13 × 24 km, with a cross-track swath of 2,600 km that enables global observational coverage daily. We use NO2 retrievals from version 3.0 of the standard product from the NASA Goddard Space Flight Center (Krotkov et al., 2017). Details of the retrieval process, which is based on fitting earthshine radiances in the spectral range 405 to 465 nm, is originally described in Bucsela et al. (2006). For the analysis presented here, we assimilate the tropospheric slant columns and transform the GEOS-Chem tropospheric vertical columns into slant columns using air mass factors computed from GEOS-Chem following the approach of Palmer et al. (2001). Under clear-sky conditions, the uncertainty in the tropospheric vertical columns are about 1 × 1015 cm−2 and increase with increasing cloud fraction. Also, the errors can become as large as 100% over polluted regions (Bucsela et al., 2013). The methods and impacts of assimilating slant columns were also discussed in Qu et al. (2017). Similar to these studies, we select only the observations with cloud fraction less than 0.2, and with quality flag = 0. Also, only the data south of 60°N latitude, with a viewing zenith angle less than 65° and a solar zenith angle less than 75° are ingested in the assimilation.

2.1.5 HIPPO Aircraft Campaign Data

The HIPPO aircraft campaign conducted five aircraft transects across the Pacific Ocean, from the Arctic to the Southern Ocean (Wofsy et al., 2011). Here we use data from the HIPPO-2 campaign, which was carried out in November 2009. The observations extend from 83°N to 66°S, and from the ground up to an altitude of 14 km, with individual vertical profiles being 2.2° latitude apart on average (Wofsy et al., 2011). The data consist of high frequency (up to 1 Hz) observations of a number of trace gases by different instruments onboard the Gulfstream V aircraft. We use the 10-s merged O3 data, which were made by the ultraviolet ozone photometer. The observations are used as an independent data set to evaluate the GEOS-Chem a priori and a posteriori O3 fields over the ocean.

2.1.6 EPA AQS Surface Data

Aside from HIPPO, measurements from the Air Quality System of the United States Environmental Protection Agency (EPA AQS) network (http://www.epa.gov/aqs) are also used for O3 validation. The controlled measurements of air pollutants are collected by EPA, state, local, and tribal air pollution control agencies over more than 1,000 sites within the United States. In our validation, modeled surface O3 time series are examined against hourly EPA AQS O3 data. To reduce bias with respect to the model, measurements under extreme O3 events and severe weather conditions are filtered to reduce the impact of meteorology (Chai et al., 2013). As the EPA AQS measurements are mostly distributed over emission source regions, the data set is valuable to examine the impact of optimized CO and NOx emissions on modeled O3 distribution.

2.2 Methods

2.2.1 GEOS-Chem Model

GEOS-Chem (www.geos-chem.org) is a global CTM driven by the assimilated meteorological data from the Goddard Earth Observing System (GEOS-5) of the NASA Global Modeling and Assimilation Office. We use version v34 of the GEOS-Chem adjoint model, which is based on version v8-02-01 of the forward model, with updates to the chemical mechanism. The native resolution of the GEOS-5 meteorological fields is 0.5° × 0.67° with 72 vertical levels from the surface to 0.01 hPa. However, to reduce the computational costs of running the assimilation, we degrade the resolution to 4° × 5° with 47 vertical pressure levels. Here we use the full-chemistry version of GEOS-Chem, with 43 chemical tracers that simulate the HOx-O3-NOx-VOC chemistry in the troposphere. In the stratosphere, O3 is simulated with the linearized O3 parameterization of McLinden et al. (2000).

Anthropogenic emission are taken from EDGAR 3.2FT2000 (Olivier & Berdowski, 2001), which is overwritten by a number of regional inventories: the US Environmental Protection Agency National Emission Inventory for 2005 (NEI05) for the United States, the Big Bend Regional Aerosol and Visibility Observational Study Emissions Inventory (BRAVO) for Mexico (Kuhn & Etyemezian, 2003), the Criteria Air Contaminants from the National Pollutant Release Inventory for Canada (www.ec.gc.ca/inrp-npri/), the Cooperative Programme for Monitoring and Evaluation of Long-Range Transmission of Air Pollutants in Europe (EMEP) inventory in Europe, and the anthropogenic emissions from Streets et al. (2006) for Asia. Biomass burning emissions are taken from version 3 of the Global Fire Emissions Database (GFED) inventory (van der Werf et al., 2006) with a monthly resolution. Biogenic emissions are from version 2.0 of the Model of Emissions of Gases and Aerosols from Nature (Guenther et al., 2006). Soil emissions of NOx are according to Wang et al. (1998), and the parameterization of lightning NOx (LNOx) follows Price and Rind (1992), with a global scaling, as described by Murray et al. (2012), to match climatological observations from the Optical Transient Detector and the Lightning Imaging Sensor.

2.2.2 Assimilation Approach

In the 4D-var data assimilation system of GEOS-Chem, the cost function is given by
where yn,r is the observation vector that contains the four different observation types (r) (MOPITT CO, TES O3, OSIRIS O3, and OMI NO2), at a given time n over the assimilation period; p represents the model parameters, which refers to surface CO and NOx emissions in our case; and x0 is the initial condition state vector which optimizes the O3 initial conditions. In minimizing J, the model is compared to the observations at the observation time n, and the GEOS-Chem adjoint is used to project the model-data mismatch backward onto the initial conditions and model parameters (Henze et al., 2007). pa and xa are the a priori estimates of the model parameters and the initial conditions, Bp and Bx are the a priori error covariance matrices for the model parameters and the initial conditions, Rr is the observation error covariance for observations of species r, and M(x0,p) represents the GEOS-Chem forward model that simulates the CO, O3, and NO2 concentrations. γr is a normalization factor used for balancing the contribution of the various observations to the cost function, which will be discussed in section 12. H is the observation operator that projects the model state to the observation space. For MOPITT and TES data, the observation operators are equivalent to equations 1 and 2, respectively, and are given by
where xn is the GEOS-Chem modeled profile of the appropriate species r. Because of the high vertical resolution of the OSIRIS data, the observation operator for OSIRIS consists only of interpolation of the GEOS-Chem O3 profile onto the OSIRIS vertical grid. For OMI NO2 data, the observation operator is given by
where h is the column operator that linearly aggregates the GEOS-Chem NO2 profiles into vertical column densities and AMF is the air mass factor. As described in section 5, AMF converts the vertical column densities into slant column densities, and captures the vertical distribution of NO2 assumed in the retrieval and the vertical sensitivity of the measurement.

In the assimilation we set the a priori error covariances as diagonal matrices with relative error standard deviations of 50%. Although we impose the diagonal covariance matrices, the use of the adjoint to propagate information across the assimilation window allows observations of a particular species to influence other species in the model. Our assumption of a uniform a priori error of 50% is an ad hoc simplification, but still consistent with previous studies (e.g., Jiang et al., 2013, 2015, 2017; Stavrakou et al., 2008). The assumption of a spatially uniform a priori error enables us to better identify the impact of the multispecies observation and their chemical coupling on the regional emission estimates without the confounding influence of some regions being more tightly constrained to match the a priori emissions than others. Our use of a uniform 50% error is similar to the approach of Miyazaki, Eskes, Sudo, et al. (2012). They assumed an initial a priori error of 40% for the emission estimates and then inflated the emission errors after each analysis step to ensure a minimum emission error of 30%. A notable difference between our assimilation configuration and Miyazaki, Eskes, Sudo, et al. (2012) is that they assumed a 10% error for the tracer concentrations, whereas our 50% a priori error is assumed for the emission and concentration estimates. We believe that a larger a priori error for the concentrations is appropriate in our configuration because of the much longer assimilation window used here compared to that in Miyazaki, Eskes, Sudo, et al. (2012) (2 weeks compared to 90 min).

For the observation error covariances, we implement a superobservation approach for MOPITT CO and OMI NO2 data. Following Jiang et al. (2013), we aggregate the MOPITT retrievals onto the model's 4° × 5° grid and assume a uniform observation error of 20% to capture the influence of representativeness errors. For OMI, which has the greatest observational coverage of the data sets assimilated, we generate superobservations following Miyazaki, Eskes, and Sudo (2012), with the exception that we use our own representativeness errors according to the cost function scaling discussed in section 12. In general, the superobservation error for NO2 is 25% of the mean measurement error.

In all experiments described below, we optimize surface emissions of CO and NOx, and O3 initial conditions (including lightning NOx emissions), depending on the observations assimilated. The quantities optimized using different combinations of the observations are listed in Table 1. We optimize only CO emissions when we assimilate only MOPITT data, only NOx emissions when we assimilate only OMI data, and ozone initial conditions when we assimilate only ozone observations (from TES and OSIRIS, separately or together). When we combine MOPITT and OMI data or MOPITT and TES, we optimize CO and NOx emissions or CO emissions and O3 initial conditions, respectively. Similarly, if we combine OMI and TES data, we optimize NOx emissions and O3 initial conditions.

Table 1. Optimized Quantities With the Corresponding Observations Used in This Study
Optimized quantities Assimilated observations
CO emissions MOPITT CO profiles
O3 initial conditions TES or OSIRIS O3 profiles
Surface NOx emissions OMI NO2 columns
Surface CO and NOx emissions, O3 initial All instruments (MOPITT, TES, OSIRIS,
conditions, LNOx emissions and OMI data)
  • Note. CO = carbon monoxide; O3 = ozone; NOx = nitrogen oxides; LNOx = lightning NOx; MOPITT = Measurements of Pollution In The Troposphere; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System; OMI = Ozone Monitoring Instrument.

The analyses were conducted for the first 2-week period in November 2009 and in July 2010. As mentioned above, the November 2009 period was selected because of the overlap with the HIPPO-2 aircraft campaign, which allows us to use the HIPPO data to evaluate the assimilation. The experiments in July 2010 are used to examine the seasonal differences in the response of the assimilation to incorporating the multiple observations. The initial conditions of all the transported tracers were obtained at the beginning of November 2009 by spinning up the forward model from July 2005 to November 2009. The run was then extended to July 2010 to get the initial conditions for the July 2010 analyses. We did not include a “spin-up” assimilation period (such as that used by Jiang et al., 2013) before each analysis period, which could introduce an initial condition bias, because our focus here is on the use of the MSA to adjust the overall chemical state and reduce the initial condition bias.

2.2.3 Cost Function Weighting Observing System Simulation Experiment

One major consideration in our MSA is the choice of the cost function normalization factor γ. If each observation is considered equally (γ = 1 for all observations), the model will favor observations with high accuracy, and the observations of the chemical species with maximum adjoint sensitivity. Our state vector consists of CO and NOx emissions and the initial O3 abundances. Changes in the CO emissions in the state vector will have a limited impact on the modeled NO2 mismatch with the OMI observations. Changes in the NOx emissions will have a larger, but still limited, impact on the modeled CO mismatch with the MOPITT observations. However, on the long time scales considered in the assimilation, changes in the CO and NOx emissions will influence the model mismatch with the O3 data. So the challenge in the MSA is to ensure that the MOPITT and OMI data provide a sufficient contribution to the cost function minimization to constrain the CO and NOx emissions, respectively.

To better understand the need for balancing the cost function, we have conducted an Observing System Simulation Experiment (OSSE) using pseudo-observations during the first week of November 2009. Pseudo data for MOPITT, TES and OSIRIS, and OMI were generated by sampling the GEOS-Chem a priori CO, O3, and NO2 fields, respectively, at the observation locations and times for each instrument. In generating the TES and MOPITT pseudo data we applied the real averaging kernels and a priori profiles. For OMI, we produced pseudo NO2 column densities using the reported scattering weights from the real OMI NO2 data set. We also produced NO2, O3, and CO columns densities from an OMI-like instrument, for which we used the OMI observation locations and times to sample the GEOS-Chem CO, O3, and NO2 fields and generated column densities by vertically integrating the modeled profiles (without averaging kernels or scattering weights). We then assimilated these data in the five OSSE experiments described in Table 2, assuming biased CO and NOx emissions and biased O3 initial conditions. For the MOPITT OSSE, we reduced CO emissions by 50%, whereas for the OMI OSSE we increased NOx emissions by 50%. For both OSSEs, the model was run with perturbed emissions from early September to the beginning of November, followed by the assimilation of the pseudo data during the first week of November. For the MSA OSSEs, in addition to perturbing the CO and NOx emissions, we also increased the O3 initial conditions by 30% throughout the troposphere. Since we conducted the OSSEs for only 1 week in November 2009, we increased the precision of the pseudo data by assuming a measurement error that was 50% of the errors reported with the real data. We did not apply additional noise to the pseudo data since random errors would average out given the large amount of data assimilated, and because our focus was mainly on the impact of the weighting of the different terms in the cost function on the inferred sources.

Table 2. Configurations of the OSSEs for γ
OSSE CO bias (%) NOx bias (%) O3 bias (%) Pseudo data
OMI 0 +50 0 OMI
MSA3 −50 +50 +30 OMI-like CO, O3, NO2
  • Note. Shown are the imposed bias in the CO and NOx emissions and in the O3 initial conditions. Also listed are the pseudo data sets assimilated in each OSSE. OSSE = Observing System Simulation Experiment; CO = carbon monoxide; NOx = nitrogen oxides; O3 = ozone; MOPITT = Measurements of Pollution In The Troposphere; OMI = Ozone Monitoring Instrument; MSA = multispecies data assimilation; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System.

Figure 1 shows the ratio of estimated to true emissions for Northern Africa, East Asia, Europe, North America, Australia, and the globe. As expected, the OMI-only and MOPITT-only OSSEs recovered the global emissions, but there were regional differences. The OMI-only inversion successfully recovered the regional emissions, but the MOPITT-only inversion did not have sufficient information with only 1 week's worth of MOPITT pseudo data to recover the true emission estimates for Europe and North America to within 10%. For the MSA1 OSSE, when no scaling was applied to the cost function (γ=1), the global mean CO and NOx emission estimates reverted closer to the prior. The regional CO and NOx emission estimates were more biased than for the single instrument case for most regions. This is because the OMI and MOPITT pseudo data contributed less toward the cost function minimization as the inversion primarily focused on optimizing the O3 state. For the MSA2, we scaled the MOPITT and OMI terms with γOMI = 16 and γMOP = 64, respectively. The motivation for this particular scaling will be discussed below in section 12. As seen in Figure 1, with the increased weighting of the MOPITT and OMI pseudo data in MSA2, the assimilation was able to recover the global mean CO and NOx emissions. Furthermore, the regional NOx emission estimates were consistent with the estimates from the OMI-only assimilation. For the CO emissions, the estimates for North America and Australia were closer to the truth, whereas the North African and East Asian estimates were biased high. The European estimate was the same as the result for MSA1. In MSA3, there was no scaling applied to the cost function. However, unlike MSA1, the assimilation was able to recover the global mean CO and NOx emissions. MSA3 also recovered the regional NOx emissions. For CO, the assimilation recovered the emission estimates to within 10% for all regions expect for Northern Africa and East Asia. As a result of the OMI-like observational coverage, MSA3 also produced the best estimate of European emissions. Similarly, we believed that because the OMI-sampled O3 observations produced much greater observational coverage than TES, the assimilation was better able to adjust the O3 state using the O3 data, and consequently, better able to utilize the CO and NO2 to optimize the CO and NOx emissions, respectively. Overall, the choice of γ will depend on the type of data assimilated and how well they constrain the different quantities within the state vector. It should be noted that although the OMI-sampled tropospheric O3 column data worked well in the OSSE, we believe that real tropospheric O3 column data from OMI would be challenging to use in our MSA because of the potential impact of errors associated with the removal of the stratospheric contribution to the total column.

Details are in the caption following the image
Emission SF for different experiments in major emission sources when comparing with the unperturbed “true” state. MSA1 and MSA2 OSSE assimilate the pseudo-observations of CO, NO2, and O3 sampled from MOPITT, OMI and TES, and OSIRIS observations, respectively. γ=1 was applied in MSA1, Scaling following Case G in section 12 was applied in MSA2. In MSA3, the pseudo-observations of CO, NO2, and O3 are all sampled from OMI observations as tropospheric columns with γ=1. NOx = nitrogen oxides; SF = scaling factor; CO = carbon monoxide; MOPITT = Measurements of Pollution In The Troposphere; OMI = Ozone Monitoring Instrument; OSSE = Observing System Simulation Experiment; MSA = multispecies data assimilation.

2.2.4 Estimating the Cost Function Weighting

To determine the appropriate values for γ for the MOPITT and OMI data, we conducted a series of sensitivity tests in which we sequentially increased γMOP and γOMI from one to 64 by a factor of four. We also considered a more extreme case with γMOP = 400 and γOMI = 400. Table 3 shows the MOPITT and OMI data contributions to the observation term in the a posteriori cost function for eight of the sensitivity test conducted. When no scaling is applied, O3 observations from TES and OSIRIS provide the dominant contributions to the total observation term in the cost function (60% and 39%, respectively). The total contribution to the observation term from MOPITT CO is about 1% and that from OMI NO2 is 0.12%. The dominating cost function contribution by O3 observations would lead to all O3 precursor fields such as modeled NO2 being overadjusted to drive the assimilated O3 changes. Consequently, the assimilated NOx emissions will not strongly reflect the influence of the OMI NO2 observations, which is consistent with the results of the OSSE discussed above. As expected, as γMOP and γOMI become larger, the contributions from MOPITT CO and OMI NO2 increase, and the contributions from the two sets of O3 observations decrease. When γOMI = 400, the cost function share of OMI NO2 reaches 18%. Similarly, increasing γMOP to 16 (with γOMI = 16), the MOPITT CO a posteriori cost function contribution increases to 22%.

Table 3. Relative a Posteriori Contribution of the Various Observations to the Total Observation Term in the Cost Function Using Different Normalization Factors γMOP and γOMI
Case Cost function share (%) Total cost function
A 1 1 1.04 59 0.12 39 5.4
B 4 1 1.04 59 0.27 40 5.5
C 16 1 1.02 57 1.3 40 5.9
D 400 1 0.72 43 18 38 11
E 16 4 3.7 67 1.3 28 5.5
F 16 16 22 48 1.2 29 7.2
G 16 64 37 43 1.1 18 9.0
H 16 400 78 16 0.42 5.7 17
  • Note. γTES=1, γOSIRIS=1 for all experiments. MOPITT = Measurements of Pollution In The Troposphere; TES = Tropospheric Emission Spectrometer; OMI = Ozone Monitoring Instrument; OSIRIS = Optical Spectrograph and InfraRed Imager System.

Figure 2 shows the a posteriori global CO and NOx emission estimates for the different values of γMOP and γOMI in the MSA. With γMOP = 1 or γOMI = 1, the estimated surface CO or NOx emissions in the MSA are much smaller than that inferred using only MOPITT CO or OMI NO2, respectively. Instead, with γMOP = 64 and γOMI = 16, we recovered the global CO emissions to within 10% of the estimate obtained from the MOPITT-only assimilation and the global NOx emissions to within 2% of the estimate inferred in the OMI-only assimilation. These are the cost function weightings applied when data from all the instruments are assimilated together (referred to as the all instrument run) for November 2009 and July 2010. We believe that the γMOP = 64 is acceptable because global mean OH has an uncertainty of 10–15% (e.g., Krol et al., 1998; Naik et al., 2013; Spivakovsky et al., 2000; Wang et al., 1998), which leads to a similar uncertainty on the global mean CO emissions. The resulting global CO emissions are also consistent with both MOPITT CO assimilation and many other inverse modeling studies summarized in Hooghiemstra et al. (2012). Moreover, our choice of γOMI = 16 and γMOP = 64 ensures that no single data set will dominate the optimization throughout the assimilation. The scaling applied here can be considered as an observation error reduction for MOPITT and OMI, which is plausible since superobservations are considered for the two observations. The importance of scaling each observation type was also discussed in Müller and Stavrakou (2005). In their joint assimilation of ground-based CO observations with GOME NO2 data, they apply a different scaling on the CO and NO2 observations in order to produce a posteriori results that best match the CO and NO2 in situ measurements.

Details are in the caption following the image
Changes (in percent) in the global a posteriori emissions of CO and NO2 as a function of the normalization factors γ shown in Table 3. The changes are shown relative to the global emission estimates obtained from assimilating only MOPITT CO or only OMI NO2 data to separately infer the CO or NOx emissions, respectively. OMI = Ozone Monitoring Instrument; MOPITT = Measurements of Pollution In The Troposphere; CO = carbon monoxide; MSA = multispecies data assimilation; NOx = nitrogen oxides.

3 Results

3.1 Assimilated O3 Distribution

The O3 distribution plays a key role in the HOx-NOx-O3-VOC chemistry and, as discussed above, has a strong influence on the cost function minimization, therefore, we begin with evaluating the assimilated O3 distribution. The a priori O3 distribution at 500 hPa and the changes in O3 that result from assimilating TES or OSIRIS O3 observations are shown in Figure 3. The a priori O3 concentrations show both a distinct minimum along the Intertropical Convergence Zone in the tropics and along the South Pacific Convergence Zone, located near 20°S, and a maximum between 20° and 40° in both hemispheres. The changes in O3 in the assimilation using only TES data have a strong zonal structure. Over the tropics and subtropics there are negative corrections exceeding 10 ppbv, whereas over the northern high latitudes there are positive corrections also exceeding 10 ppbv. In contrast, in the OSIRIS assimilation, the changes are negligible, with a maximum decrease in O3 of 3 ppbv over northern Africa. This is because OSIRIS does not make observations at these altitudes. The changes in the OSIRIS assimilation at these altitudes are caused by downward transport of the changes in O3 in the UTLS.

Details are in the caption following the image
The a priori Ox distribution and the difference (ppbv) between the a posteriori and the a priori Ox at 500 hPa for the TES-only, the OSIRIS-only, and the TES + OSIRIS assimilation. TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System.

The a posteriori O3 changes at 300 hPa (shown in Figure 4) for the experiment assimilating TES O3 data are similar in structure to the O3 changes at 500 hPa. There are negative changes over the tropics and subtropics, and positive changes over the northern high latitudes. The averaged magnitude of changes in the UTLS varies between −15 and 8 ppbv. As expected, the O3 changes at 300 hPa when using OSIRIS O3 are larger than the changes obtained in the lower troposphere, with negative corrections everywhere. For the experiment combining both TES and OSIRIS data, the assimilation results are dominated by the influence of the OSIRIS data. The modeled O3 biases are larger within the UTLS than at lower altitudes, and OSIRIS provides more O3 data with a smaller retrieval error than TES. As a result, the optimization responds more strongly to the OSIRIS data in the UTLS, which mitigates the impact of the TES data, as shown in Figures 3 and 4.

Details are in the caption following the image
As in Figure 3, but for 300 hPa.

For the experiment combining all instruments, the assimilated O3 distribution (not shown) in both the middle troposphere and UTLS is similar to that obtained in the TES + OSIRIS assimilation. However, at the surface, the all instrument assimilation produces increases of up to 5 ppbv in O3 over the main anthropogenic emission regions in the northern hemisphere (see Figure 5), whereas the changes in surface O3 in the TES + OSIRIS assimilation are negligible (not shown). In the TES-only assimilation there are large changes in high latitude O3, which we believe is due to the downward transport of the changes in O3 in the middle troposphere (see Figure 3). In the tropics and southern subtropics, the all instrument assimilation produce modest reductions in surface O3. The increase in O3 in the northern extratropics as well as the decrease in the tropics in the all instrument case are due to the reduced NOx emissions in the assimilation. These results show that optimizing the O3 initial conditions is a less effective means of adjusting surface O3 concentrations than optimizing the O3 precursor emissions. This is because the O3 lifetime is short near the surface, so the perturbations in the initial conditions are quickly destroyed. The lifetime is expected to be longer at high latitude in November and, thus, we see the persistence at the high-latitude surface of the signal of the midtroposphere O3 changes in the TES assimilation in Figure 5.

Details are in the caption following the image
As in Figure 3, but for the surface.

3.1.1 Evaluation Against HIPPO O3 data

The assimilated O3 in November 2009 is compared with HIPPO data to evaluate O3 distribution both in the lower troposphere and midtroposphere as well as in the UTLS. To map the modeled state to the observations, we sampled the assimilated O3 at the time and locations closest to the HIPPO observations. The differences between the observations and the a priori and a posteriori simulations were then averaged into three latitude bins: 90–45°N, 45°N–45°S, and 45–90°S. As shown in Table 4, within the tropics and subtropics, the observed O3 concentrations are lower than the a priori O3 concentrations for both lower troposphere (by −4.4 ppbv) and UTLS (by −15.5 ppbv). When TES O3 is added to the assimilation, the a posteriori O3 concentrations are closer to the observed values over this region. However, the positive biases in the model are not removed completely. The experiment using TES O3 also provides a posteriori O3 concentrations that are closer to the HIPPO values over the southern hemisphere. However, the assimilated O3 using TES degrades the model agreement with the observations in the northern high latitudes. This could be due to the influence of the TES averaging kernels smearing O3 from the stratosphere down into the troposphere. In late fall and early winter, the tropopause is lower and O3 abundances are greater in the high-latitude lower stratosphere, which exacerbates the smearing effects of the TES averaging kernels. The experiment combining TES and OSIRIS observations compensates for the adverse impact of the TES assimilation in the northern high latitudes and also significantly reduces the a priori model bias in the southern high latitudes. However, at low latitudes in both hemisphere, the TES-only assimilation provides the best agreement with the HIPPO data. The all instrument assimilation produces an O3 distribution that is similar to that of the TES + OSIRIS assimilation, but with slightly larger absolute biases. Overall, despite the adverse impact of the TES assimilation in the northern high latitudes, both the TES + OSIRIS and the all instrument assimilation experiments effectively reduced the a priori bias.

Table 4. Comparisons Between the HIPPO O3 Observations and the a Priori and a Posteriori O3 Fields in November 2009
Region A priori TES O3 OSIRIS O3 TES and OSIRIS All instrument
Global mean −7.3 (36) −3.3 (24) −1.7 (23) −1.4 (23) −1.9 (24)
90–45°N LT & MT −1.9 (11) −5.8 (15) −1.0 (7.3) −1.6 (7.6) −1.2 (7.2)
90–45°N UTLS 26 (52) 34 (50) 27 (41) 27 (40) 28 (42)
45°N–45°S LT & MT −4.4 (19) −1.2 (9.0) −3.7 (9.5) −3.3 (9.2) −3.4 (9.9)
45°N–45°S UTLS −16 (37) 0.9 (25) −4.7 (29) −1.7 (27) −0.8 (25)
45–90°S LT & MT −4.3 (17) −3.8 (6.8) −1.0 (7.3) −1.6 (7.3) −2.1 (7.1)
45–90°S UTLS −36 (50) −33 (47) 29 (55) 12 (53) 14 (51)
  • Note. Shown are the mean difference (HIPPO − model) and the mean absolute difference |(HIPPO − model)|, in parentheses. LT & MT refers to 0–7 km in altitude, whereas UTLS refers to 7–13 km. HIPPO = HIAPER Pole-to-Pole Observations; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System; LT & MT = Lower and Middle Troposphere; UTLS = Upper Troposphere and Lower Stratosphere.

3.1.2 Evaluation Against EPA AQS O3 Data

The assimilated surface O3 in July 2010 is evaluated using EPA AQS O3 data. We focus on O3 in July because models tend to overestimate summertime surface O3 abundances in the eastern United States (Reidmiller et al., 2009). Figure 6 shows the differences between the observations and the a priori model, the TES-OSIRIS assimilation, the OMI NO2 assimilation, and the All Instrument assimilation. The AQS data were binned onto the model grid using the weighted-average algorithm suggested by Klonecki et al. (2012). As expected the a priori model overestimates surface ozone across much for the United States. Canty et al. (2015) suggested that these biases on modeled surface ozone could be due to discrepancies in vertical mixing as well as errors in the O3 precursors. In the TES-OSIRIS assimilation in which we adjusted only the O3 initial conditions there is no significant improvement on surface O3, which is expected since the O3 lifetime is short in the boundary layer. The All instrument run, however, shows significant surface O3 improvements. In the eastern and western United States (east and west of 95°W) the mean ozone bias decreased from 13.1 to 6.2 ppbv and from 5.4 to 0.3 ppbv, respectively. In fact, OMI NO2 assimilation alone would correct surface O3 in similar patterns as O3 changes assimilated by all instruments. The surface O3 bias were corrected to 7.9 and 1.9 ppbv for eastern and western United States, respectively, driven by the chemical response to the adjusted NOx emissions. This confirms the utility of the MSA approach for providing additional corrections to the O3 distribution through the emission optimization.

Details are in the caption following the image
The absolute differences between the AQS and the modeled O3 during July 2010 for the a priori (top left), the TES + OSIRIS assimilation (top right), OMI assimilation (bottom left), and the All Instrument assimilation (bottom right). AQS = Air Quality System; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System; OMI = Ozone Monitoring Instrument; O3 = ozone.

3.2 Tropospheric OH Concentrations

Figure 7 shows the mass weighted global mean OH concentrations for the a priori and the changes in OH from MOPITT assimilation as well as MSA. The global mean OH produced by a priori is 11.9 × 105 molecules per cubic centimeter for November 2009, and 13.9 × 105 molecules per cubic centimeter for July 2010. The distribution of tropospheric mean OH is consistent the tropospheric water vapor abundances. In November, the highest OH concentrations are found in the tropics and southern subtropics due to higher actinic fluxes as well as higher water vapor concentrations. In July 2010, the OH maximum shifts into the northern hemisphere. For the assimilated results in comparison with the a priori, the global mean OH constrained by MSA is reduced to 10.5 × 105 molecules per cubic centimeter (by 11.7%) and 12.8 × 105 molecules per cubic centimeter (by 7.9%) in November and July, respectively. In contrast, with the MOPITT-only assimilation we obtained reductions in the global mean OH of 0.9% and 1.8% for November and July, respectively. Our results show that the MSA is conducive to constraining the tropical and subtropical OH, which could reduce the uncertainties of CO emission estimates due to biased tropospheric OH (Müller et al., 2018). There are also significant OH changes due to CO and NOx emission adjustments over the extratropics. In November 2009, the a posterior OH is reduced by 6.5% between 30 and 60°N. In July 2010, as the tropospheric chemical cycling of CO-HOx-NOx-O3-OH becomes more active over the northern extratropics, the resulting OH is reduced by 10.5%. These OH changes inferred by MSA suggest that tropospheric OH could be constrained indirectly, relying on the assimilation of the OH-related species in the 4D-var assimilation system. We note, however, that improving the OH abundance in the assimilation should also account for the influence of VOCs on OH, which was not done in our analysis.

Details are in the caption following the image
The a priori mass weighted tropospheric mean OH distribution and the relative difference ( urn:x-wiley:jgrd:media:jgrd55198:jgrd55198-math-0009 × 100%) between the a posteriori and the a priori OH for different experiments. The mass weighted tropospheric mean was defined the same as Spivakovsky et al. (2000). OH = hydroxyl radical; MOPITT = Measurements of Pollution In The Troposphere.

3.3 Regional CO and NOx Emissions in November 2009

The a priori CO combustion emissions and the estimated change in the emissions relative to the a priori for the MOPITT-only assimilation and for MOPITT combined with the different data sets are shown in Figure 8. The a posteriori CO emissions based only on MOPITT data and on MOPITT combined with OMI data (MOPITT + OMI) indicate much higher emissions in East Asia, Europe, and North America. In contrast, the biomass burning emissions in northern equatorial Africa are decreased. There are also increases in the emissions in southern Africa and northern Australia. With MOPITT and TES combined, the large emission increases estimated in the extratropical northern hemisphere are more muted, and with MOPITT and OSIRIS the assimilation largely reverts to the a priori CO emissions, except for East Asia and northern Africa. Two reasons might explain this significant damping effect when OSIRIS is combined with MOPITT. The first is that the total cost function is dominated by OSIRIS O3 data so the model does not effectively optimize the CO emissions to match the MOPITT data. The other reason is since OSIRIS does not produce observations in the middle and lower troposphere, the assimilated O3 concentration does not change in the middle and lower troposphere other than O3 changes transported from stratosphere. Consequently, we would expect the a posterior O3 in those layers of the atmosphere to induce small changes in CO emissions.

Details are in the caption following the image
The a priori CO emissions and the a posteriori CO emission changes relative to the a priori for experiments assimilating only MOPITT data, MOPITT + OMI data, MOPITT + TES data, MOPITT + OSIRIS data, and the data from all four instruments. Shown are the emissions and absolute changes in the emissions in units of 1010 molecules per square centimeter per second. CO = carbon monoxide; MOPITT = Measurements of Pollution In The Troposphere; OMI = Ozone Monitoring Instrument; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System.

To assess the a posteriori emissions quantitatively, the CO emissions have been aggregated over the different continental regions shown in the left corner of Figure 9. We find the MOPITT + OMI and the all instruments cases are consistent and are generally in agreement with the MOPITT-only assimilation. For example, for North America the MOPITT, MOPITT + OMI, and the all instrument estimates are 21.5, 18.4, and 18.5 Tg/month, respectively. For Europe, the corresponding estimates are 12.6, 12.0, and 10.5 Tg/month, respectively. For the biomass burning emissions in northern equatorial Africa (N. Africa), we find that combining MOPITT with the O3 observations results in larger departures from the a priori. The MOPITT + TES and MOPITT + OSIRIS emission estimates are 21.8% and 22.1%, respectively, lower than the a priori. In contrast, the all instrument estimate is only 4.3% lower than the a priori. Given the large biases in O3 in the UTLS and the importance of convective transport in the tropics, it is possible that the larger changes relative to the a priori for MOPITT + TES/OSIRIS runs are due to the assimilation trying to reduce the overestimate in O3 in the model by reducing the emissions that are convectively transported throughout the free troposphere, counteracting with CO emission increase brought by MOPITT CO assimilation. For all instrument run, such ozone adjustment was brought by optimized NOx emissions, which leaves assimilated CO emissions primarily responding to MOPITT CO observations.

Details are in the caption following the image
Surface CO emissions (Tg CO/month) for November 2009 aggregated for the seven continental regions shown in the top left. CO = carbon monoxide; MOPITT = Measurements of Pollution In The Troposphere; OMI = Ozone Monitoring Instrument; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System.

The a priori NOx combustion emissions and the estimated changes in the emissions relative to the a priori for the OMI-only assimilation and for OMI combined with the different data sets are shown in Figure 10. For the OMI-only assimilation, there are reductions of NOx emissions over the eastern United States and Europe, but the changes are heterogeneous over East Asia, with increased and decreased emissions across the region. There is also a modest increase in emissions in the southern United States and in the biomass burning region in northern equatorial Africa. Adding MOPITT CO observations to the assimilation does not change the spatial patterns of the optimized emissions significantly. In contrast, ingesting O3 observations has a large impact on the NOx emission estimates. The impact is more clearly seen in the regional emission estimates in Figure 11. With TES data included, East Asian emissions are 24% lower than the a priori, whereas the OMI-only estimate is 11% lower. For northern equatorial Africa, the OMI-only case suggested a 5% increase in emissions, whereas the OMI + TES assimilation produced a 58% decrease in the emissions. These changes are occurring because the TES O3 data override the information coming from the OMI tropospheric NO2 observations. Consequently, the NOx emission are overadjusted to reduce the bias in the modeled O3 to give a better fit to the TES retrievals. In the all instrument case, in which we balance the contributions to the cost function, the estimated regional emissions are 20% lower than the OMI-only estimate. For the major anthropogenic emission regions, North America and East Asia, we find that the all instrument estimate is consistent with that inferred in the OMI-only assimilation. Europe is an exception, for which we obtain a large difference (18%) in emissions between the OMI-only and the all instrument cases. The situation with the European emission estimates will be discussed further in section 21.

Details are in the caption following the image
The a priori NOx emissions and the a posteriori NOx emission changes relative to the a priori for experiments assimilating only OMI data, OMI + MOPITT data, OMI + TES data, OMI + OSIRIS data, and the data from all four instruments. Shown are the emissions and absolute changes in the emissions in units of 109 molecules per square centimeter per second. NOx = nitrogen oxides; OMI = Ozone Monitoring Instrument; MOPITT = Measurements of Pollution In The Troposphere; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System.
Details are in the caption following the image
Surface NOx emissions (Tg N/month) for November 2009 aggregated for the seven continental regions shown in Figure 9. NOx = nitrogen oxides; OMI = Ozone Monitoring Instrument; MOPITT = Measurements of Pollution In The Troposphere; TES = Tropospheric Emission Spectrometer; OSIRIS = Optical Spectrograph and InfraRed Imager System.

3.4 Evaluation of the Emission Estimates in November 2009

To evaluate the inferred emissions, we compared the a posteriori CO and NO2 fields with the assimilated data. Figure 12 shows that the MOPITT-only assimilation reduces the mean bias, relative to the MOPITT data, by more than a factor of two, from 12% to 5%. The MSA reduces the bias further, to 4%. The distribution of the MSA difference is also less skewed, with a standard deviation of 14% (compared to 16% for a priori). For NO2, the OMI-only assimilation also reduces the mean bias by more than a factor of two, from 45% to 20%. However, with the MSA, the residual bias is reduced to only 32%. The larger residual bias in the MSA is expected since the assimilation is trying to adjust the NOx emissions to accommodate the O3 data as well as the NO2 data. This makes the adjustment on surface NOx emissions less effective on correcting the modeled NO2 bias compared to the OMI-only assimilation. The distribution is skewed due to the fact that the modeled NO2 columns are negatively biased throughout much of the background atmosphere, which the assimilation could not correct. The better fit to the MOPITT data with the MSA may reflect the impact of a reduced OH bias on the CO simulation mentioned in section 18. The improvement on the a posteriori distribution of CO and NO2 assimilated by all instrument was also found for July 2010 (not shown).

Details are in the caption following the image
Distribution of the modeled minus assimilated CO (left panel) and NO2 (right panel). Each color represents the difference for the a priori (blue), the single instrument assimilation a posteriori (MOPITT-only on the left and OMI-only on the right; green), and the All Instrument assimilation a posteriori (red). MOPITT = Measurements of Pollution In The Troposphere; MSA = multispecies data assimilation; CO = carbon monoxide; OMI = Ozone Monitoring Instrument.

3.5 Regional CO and NOx Emissions in July 2010

The regional CO emission estimates for July 2010 and November 2009 are compared in Table 5. In contrast to the results in November, we find that the differences between the emission estimates from the MSA and those from the MOPITT-only. For example, for North America and Europe, the MSA CO emission estimates are higher by 30% and lower by 65% than those in the MOPITT-only assimilation, respectively. In November, the MSA estimates for these regions were lower by 14% and 17%, respectively, compared to the MOPITT-only case. In East Asia, the other major anthropogenic source region, the source estimate in the MSA was lower than the MOPITT-only estimate by about 10% in both months. Globally, the total CO emission estimates between the two assimilations differ by 9% in both July and November. Regional wise, the European emission estimate from the MSA differed the most from the MOPITT-only assimilation. Our results are consistent with those of Jiang et al. (2015), who also used the GEOS-Chem adjoint model, and found that summertime emission estimates for North America and Europe were most sensitive to the different OH fields. They argued that relative to East Asia, for example, North American and European CO emissions in GEOS-Chem were more strongly confined to the lower troposphere. This suggests that the free tropospheric CO from these regions was more chemically aged. Thus, inferring emission estimates from observations that exhibit high retrieval sensitivity to CO in the free troposphere could result in emission estimates that are sensitive to biases in OH, since the observations will contain a mix of lower tropospheric air with fresh emissions and free tropospheric air that is more chemically aged. In contrast, emission estimates inferred from observations in the boundary layer will be less sensitive to discrepancies in OH since the measurements will capture fresh CO emissions, with little chemical aging.

Table 5. Regional Surface CO and NOx Emission Estimates for the a Priori, MOPITT/OMI only, and All Instruments Runs
CO regional emissions (TgCO/month) NOx regional emissions (TgN/month)
November 2009 July 2010 November 2009 July 2010
Regions A priori MOPITT MSA A priori MOPITT MSA A priori OMI MSA A priori OMI MSA
S-America 3.2 4.6 4.1 8.3 3.5 6.3 0.13 0.13 0.11 0.19 0.22 0.20
N-America 7.5 21.5 18.5 15.1 30.8 40.4 0.54 0.42 0.39 0.61 0.44 0.56
E-Asia 16.3 44.8 40.2 13.5 30.9 27.6 0.66 0.59 0.55 0.61 0.67 0.59
SE-Asia 1.8 3.4 3.0 1.7 1.5 1.6 0.06 0.09 0.04 0.06 0.07 0.08
Australia 5.1 19.8 18.1 1.5 2.9 1.5 0.08 0.09 0.07 0.06 0.07 0.06
Europe 4.4 12.6 10.5 5.0 27.4 9.7 0.45 0.28 0.33 0.41 0.45 0.38
N-Africa 9.2 9.8 8.8 1.3 5.0 2.3 0.19 0.20 0.16 0.07 0.11 0.07
S-Africa 1.2 6.5 5.1 17.9 23.4 24.1 0.05 0.05 0.02 0.28 0.44 0.40
Rest of the world 8.6 14.4 15.4 17.9 39.7 36.4 0.31 0.33 0.24 0.44 0.46 0.44
Total 57.3 137.4 124.7 82.2 165.1 149.9 2.47 2.14 1.91 2.73 2.93 2.78
  • Note. The emissions include both anthropogenic and biomass burning components. CO = carbon monoxide; NOx = nitrogen oxides; MOPITT = Measurements of Pollution In The Troposphere; OMI = Ozone Monitoring Instrument; MSA = multispecies data assimilation.

For the main anthropogenic NOx emission regions (North America, Europe, and East Asia), we find that the emission estimates from the MSA for Europe and East Asia are lower than those from the OMI-only assimilation by 16% and 12%, respectively, in July, whereas the North American estimate is 27% higher. In contrast, in November, the North American and East Asian MSA estimates are 5% and 7% lower, respectively, whereas the European estimate is 18% higher. As with CO, we find that the European NOx emission estimates in both seasons are sensitive to the MSA approach. As noted by Jiang et al. (2015), European emissions are strongly confined to the boundary layer. This could make estimating European NOx emission more sensitive to discrepancies in the vertical distribution of NOx when using NO2 column data. Additional work is needed to better understand the source of the differences in the European emission estimates between the MSA and OMI-only assimilation. The large difference between the MSA and OMI-only NOx emission estimate for North America in July could be due to the fact that we do not optimize isoprene emissions in the assimilation. The isoprene oxidation pathways are NOx dependent, and discrepancies in the isoprene emissions in the model could impact the local NOx abundances. The other reason is that the model might try to correct biases owing to LNOx, which is larger in summer, by over-adjusting surface emissions.

In July, the dominant biomass burning region in the tropics is southern Africa where the MSA suggested a 35% increase in CO emissions and a 43% increase in NOx emissions over the a priori. In contrast, in November 2009, which is the beginning of the biomass burning season in northern Africa, the MSA suggested reductions of 4% and 15% in the CO and NOx emissions relative to the a priori, respectively. It is interesting to note that the MOPITT-only and OMI-only assimilations produced increases of 7% and 5% in the CO and NOx emissions relative to the a priori in November 2009, respectively. Interannual variability in Africa biomass burning emissions is low except for 2010 reported by Giglio et al. (2013). In 2010, the emissions are associated with much higher than averaged burned area in southern Africa (Giglio et al., 2013). Our higher estimate is part of an increasing trend in burned area of +1.5 Mha/year for southern Africa present in both GFED-3 and GFED-4 (Giglio et al., 2010, 2013). Giglio et al. (2013) also argued that northern African emissions in 2009 and 2010 experienced significantly reduced burning, which is consistent with our estimated emission changes.

3.6 Comparison With Previous Inversion Analyses

We compared the assimilated CO emission estimates with the inversion results shown in Jiang et al. (2017) and Kopacz et al. (2010). For North America we estimated a source of 40 Tg CO with the MSA and 31 Tg CO for the MOPITT-only assimilation for July, whereas Jiang et al. (2017) estimated 15 Tg CO with the MOPITT V6 TIR/NIR profile. For Europe, our July estimates were 10 and 27 Tg CO from the MSA and MOPITT-only assimilation, respectively. In contrast, Jiang et al. (2017) estimated 5 Tg CO. So both our MSA and the MOPITT-only European estimate were larger than those in Jiang et al. (2015). For East Asia we inferred 28 and 31 Tg CO for July for the MSA and MOPITT-only assimilation. Jiang et al. (2017) estimated 24 Tg CO. Our East Asian emission estimates seem comparable to those of Jiang et al. (2017). For Europe and East Asia, Kopacz et al. (2010) estimated July 2004 combustion emissions of about 9 and 12 Tg CO, respectively, which was comparable to our European MSA estimate of 10 Tg CO, but much lower than our East Asian estimate of 28 Tg CO for July 2010. For November 2004, Kopacz et al. (2010) reported European and East Asian combustion emissions of about 11 and 21 Tg CO. For November 2009, our MSA emission estimates for Europe and East Asia are 11 and 40 Tg CO, respectively. So our European emission estimates are comparable to Kopacz et al. (2010). One issue when comparing our results to other studies is that all the inversion analyses were conducted at different time periods or with different assimilation configurations. For instance, both Jiang et al. (2017) and Kopacz et al. (2010) used v5-07-08 of GEOS-Chem OH fields, whereas the OH in our analysis is more similar to the v8-02-01 OH fields used by Jiang et al. (2015). Jiang et al. (2015) argued that CO emission estimates using v8-02-01 OH could be 40% higher than using v5-07-08 OH, which explains why our emission estimates are higher in many regions. Jiang et al. (2017) also included the biogenic source of CO in their state vector and optimized the VOC source together with the combustion emissions of CO at the model resolution. In comparison, Kopacz et al. (2010) separately optimized the combustion and VOC sources of CO. They optimized the combustion emissions at the model resolution and aggregated the VOC source with the methane source to produce a global mean chemical source of CO. For this study, we did not include biogenic emissions in our state vector, so our emission estimate accounts only for combustion-related emissions, which may also explain some differences on emission estimates.

3.7 Sensitivity to Assimilation Window Selections

For the experiments mentioned above, we used a 2-week assimilation period. The length of the assimilation window was selected to match the lifetime of midtropospheric O3 in the tropics and the repeat time for TES observational coverage. We focused on O3 since it was a key component of tropospheric chemistry and because the MOPITT and OMI provided greater observational coverage. Here we vary the assimilation window of the MSA in November 2009 from 2 weeks to 1 week, and then to 2 days. In the experiments with the shorter windows, we are assimilating the data over for the same total 2-week period. For example, for the experiment with the 2-day window, after the optimized CO and NOx emissions and O3 concentrations are obtain at the end of the initial 2-day period, they are used as initial parameters for the next 2-day assimilation period. In this manner, we stepped through the original 2-week period, at 2-day intervals.

As shown in Table 6, the total regional CO and NOx emission estimates are robust. Although the total emissions retreat to the a priori with the shorter assimilation windows, the differences relative to the standard 2-week case are small. In our experiments, for the total CO emissions, the differences are −4% and −6% for the experiments with 1-week and 2-day window experiments, respectively. Similarly, for NOx we found that total emission estimates differ by 8% and 13% for the 1-week and 2-day experiments, respectively. For the regional anthropogenic CO emission estimates, the largest discrepancy between the 1-week window experiment and the standard 2-week case was for the East Asian emissions, which were 9% lower in the 1-week window experiment. This difference increased to −16% for the 2-day window experiment. For the NOx emission estimates, the absolute differences between 1-week window experiment and the standard 2-week case were less than about 5% for the major anthropogenic source regions. This agrees with L. Zhang, Constantinescu, et al. (2008), who also found consistent optimized NO2 despite varying assimilation windows. The maximum absolute difference increased to about 11% for the 2-day assimilation window case. Our results suggest that although we originally selected the 2-week assimilation window based on the TES observational coverage, we could reduce the window length to 1 week when we integrate the data from all of the instruments.

Table 6. Regional Surface CO and NOx Emission Estimates in November 2009 for the all Instrument Case Using Assimilation Windows of 2 Weeks, 1 Week, and 2 Days
CO regional emissions (TgCO/month) NOx regional emissions (TgN/month)
Regions AW = 2 weeks AW = 1 week AW = 2 days AW = 2 weeks AW = 1 week AW = 2 days
S-America 4.1 4.1 4.1 0.11 0.12 0.13
N-America 18.5 18.6 18.0 0.39 0.41 0.43
E-Asia 40.2 36.5 33.9 0.55 0.58 0.61
SE-Asia 3.0 2.9 3.0 0.04 0.06 0.05
Australia 18.1 15.5 18.9 0.07 0.07 0.07
Europe 10.5 10.9 10.4 0.33 0.32 0.35
N-Africa 8.8 9.3 9.1 0.16 0.20 0.17
S-Africa 5.1 5.4 5.2 0.02 0.03 0.03
Rest of the world 15.4 16.5 15.0 0.24 0.27 0.31
Total 124.7 119.7 117.6 1.91 2.06 2.15
  • Note. CO = carbon monoxide; NOx = nitrogen oxides.

4 Conclusions

We used the GEOS-Chem 4D-Var data assimilation system to examine the impact of assimilating multiple chemical species to quantify surface emissions of CO and NOx. The 4D-Var scheme seeks to obtain a model trajectory that best matches all the available observations over the assimilation period. Observations of O3 from TES and OSIRIS, NO2 from OMI, and CO from MOPITT were assimilated for a 2-week period in November 2009 and in July 2010. We found that the multispecies assimilation reduced the absolute mean bias in modeled O3, relative to HIPPO-2 data, to 10 ppbv or less in the middle and lower troposphere. The residual mean biases were less than 3.5 ppbv everywhere, with the model overestimating O3 concentrations, except in the high-latitude UTLS, where the model is biased low, with residual mean biases of 29 and 14 ppbv between 45–90°N and 45–90°S, respectively. The large residual biases in the UTLS might be due to the low (4° × 5°) horizontal resolution of the model (Deng et al., 2015). At the surface, the MSA reduced mean surface O3 over North America in July 2010 by 20.1% when validating against EPA AQS data.

For CO and NOx emissions, we found that in November 2009, the estimates obtained from MSA were generally comparable to those inferred from the MOPITT-only and OMI-only assimilation, respectively. For CO emissions from the main anthropogenic source regions, the largest discrepancies were obtained for the emission estimates for Europe and North America. In November, the European and North American emission estimates from the MSA differed from the MOPITT-only estimates by −17% and −14%, respectively. For the NOx emissions, the North American and European MSA estimates differed from the OMI-only estimates by −5% and 18%. In July the differences between the MSA estimates and those based on the MOPITT-only and OMI-only assimilation were larger. The differences for the North American and European CO emission estimates increased to −65% and 31%, respectively. For the North American and European NOx emission estimates, the July differences were −16% and 27%, respectively. The large difference between the MSA and MOPITT-only estimates for North American and European CO emissions are consistent with the findings of Jiang et al. (2015), who found that summertime emissions from these regions were particularly sensitive by the choice of OH distribution specified in their model. They attributed this sensitivity to the fact that emissions from these regions are relatively more slowly exported to the free troposphere, so the emission signals from these regions in the free troposphere reflect more chemically aged air. Further investigation is needed to better understand the source and seasonal dependence of this sensitivity of the North American and European emissions.

Comparison of the a posteriori CO and NO2 fields with the assimilated data showed that the MSA provided a better fit to the MOPITT data than the MOPITT-only assimilation. In contrast, the OMI-only assimilation produced a better fit to the OMI data than the MSA. The better agreement between the OMI-only assimilation and the OMI data was due to the fact that the assimilation adjusted the NOx emission to minimize the model-data mismatch, only by the assumed prior and observation errors. Whereas in the MSA, the NOx emissions were adjusted to improve the model agreement with the O3 data as well as with the OMI NO2 columns. For CO, the closer agreement between the MSA a posteriori fields and the MOPITT data could be due to the reduced chemical bias in the model, associated with the constrained OH fields in the MSA.

Our results highlight the importance of producing consistent chemical states for inverse modeling of CO and NOx emissions. However, this poses several challenges. A major issue is the biases in the modeled O3 distribution, which affects tropospheric OH abundances and the NOx-O3 coupling. The NOx inversion analysis responded strongly to the modeled O3 to mitigate discrepancies in between the modeled and observed O3. Consequently, it is important to effectively mitigate the O3 biases through assimilation of O3 data or through improving the tropospheric chemistry in the model. Given the short lifetime of O3 in the lower troposphere, it would be helpful to assimilate O3 data that can provide global observational coverage on at least daily time scales. This would help mitigate the O3 biases in the assimilation on short time scales and minimize the potential impact of the biases on the shorter-lived chemical tracers. Another issue is that only combustion-related sources were optimized in our inversion analyses. Tropospheric OH was also estimated relying on chemical feedback within the model. The global mean OH levels were decreased by 11.7% and 7.9% in November and July, respectively. However, the biogenic source of CO from VOC oxidation is a large component of the CO budget, particularly for North America in the summer (Hudman et al., 2008), and there are large uncertainties in the bottom up biogenic inventories (Marais et al., 2012). Studies (e.g., Kleinman et al., 1994; Travis et al., 2016; R. Zhang et al., 2004) shown that these VOC sources, as well as anthropogenic VOC emissions (which are also uncertain), have a strong impact on tropospheric O3 and OH abundances, and their oxidation is sensitive to local NOx concentrations. Several inverse modeling studies (e.g., Barkley et al., 2013; Kaiser et al., 2018; Marais et al., 2012) have utilized formaldehyde (HCHO) satellite observations to constrain isoprene emissions (the dominant biogenic CO source). Future multispecies assimilation studies should consider incorporating HCHO data to constrain the CO and VOC emissions as well as tropospheric O3 and OH concentrations.


This work was supported by funding from Environment and Climate Change Canada and the Natural Sciences and Engineering Research Council (NSERC). D.K.H. recognizes support from NASA ACMAP NNX17AF63G. We also thank the data source providers who archived MOPITT CO data (Ziskin, 2016), TES O3 data (available at https://tes.jpl.nasa.gov/data/), OMI NO2 SP data from NASA (Krotkov et al., 2018), OSIRIS O3 data from University of Saskatoon (available at odin-osiris.usask.ca). Furthermore, we would like to acknowledge the free use of validation data products including HIPPO products from NASA EOL (available at http://data.eol.ucar.edu/master_list/?project=HIPPO-2), and EPA AQS O3 data from EPA (available at https://aqs.epa.gov/aqsweb/airdata/download_files.html#Raw).