Towards monitoring localized CO 2 emissions from space : co-located regional CO 2 and NO 2 enhancements observed by the OCO-2 and S 5 P satellites

Despite its key role for climate change, large uncertainties persist in our knowledge of the anthropogenic emissions of carbon dioxide (CO2) and no global observing system exists allowing to monitor emissions from localized CO2 sources with sufficient accuracy. The Orbiting Carbon Observatory-2 (OCO-2) satellite allows retrievals of the column-average dry-air mole fractions of CO2 (XCO2). However, regional column-average enhancements of individual point sources are usually small compared to the background concentration and its natural variability and often not much larger than the satellite’s measurement 5 noise. This makes the unambiguous identification and quantification of anthropogenic emission plume signals challenging. NO2 is co-emitted with CO2 when fossil fuels are combusted at high temperatures. It has a short lifetime of the order of hours so that NO2 columns often greatly exceed background and noise levels of modern satellite sensors near sources which makes it a suitable tracer of recently emitted CO2. Based on six case studies (Moscow, Russia; Lipetsk, Russia; Baghdad, Iraq; Medupi and Matimba power plants, South Africa; Australian wildfires; and Nanjing, China), we demonstrate the usefulness 10 of simultaneous satellite observations of NO2 and XCO2. For this purpose, we analyze co-located regional enhancements of XCO2 observed by OCO-2 and NO2 from the Sentinel-5 Precursor (S5P) satellite and estimate the CO2 plume’s cross-sectional fluxes. We take advantage of the nearly simultaneous NO2 measurements with S5P’s wide swath and small measurement noise by identifying the source of the observed XCO2 enhancements, excluding interference with remote upwind sources, allowing to adjust the wind direction, and by constraining the shape of the CO2 plumes. We compare the inferred cross-sectional 15 fluxes with the Emissions Database for Global Atmospheric Research (EDGAR), the Open-Data Inventory for Anthropogenic Carbon dioxide (ODIAC), and, in the case of the Australian wildfires, with the Global Fire Emissions Database (GFED). The inferred cross-sectional fluxes range from 31 MtCO2/a to 153 MtCO2/a with uncertainties (1σ) between 23% and 72%. For the majority of analyzed emission sources, the estimated cross-sectional fluxes agree within their uncertainty with either EDGAR or ODIAC or lie in between them. We assess the contribution of multiple sources of uncertainty and find that the 20 dominating contributions are related to the computation of the effective wind speed normal to the plume’s cross-section. The planned European Copernicus anthropogenic CO2 monitoring mission (CO2M) will not only provide precise measurements with high spatial resolution but also imaging capabilities with a wider swath of simultaneous XCO2 and NO2 observations.


Introduction
Carbon dioxide (CO 2 ) is the most important anthropogenic greenhouse gas and driver for climate change.By September 2018, 195 member states of the UNFCCC (United Nations Framework Convention on Climate Change) have signed the Paris agreement with the long-term goal to keep the increase in global average temperatures relative to preindustrial levels well below 2 • C. Actions need to be taken to halve anthropogenic greenhouse gas emissions (including CO 2 ) each decade after reaching peak emissions in 2020 (Rockström et al., 2017).However, there are still large uncertainties in the anthropogenic emissions and no global observing system exists that allows us to monitor country emissions and their changes with sufficient accuracy (e.g., Ciais et al., 2014;Pinty et al., 2017).
CO 2 is long-lived and well-mixed in the atmosphere and its largest gross fluxes are of natural origin (photosynthesis and respiration).As a result, regional column-average enhancements of individual anthropogenic point sources are usually small, compared with the background concentration and its natural variability, and often not much larger than the satellite's measurement noise (Bovensmann et al., 2010).This makes the identification of anthropogenic plume signals with past (SCIAMACHY, SCanning Imaging Absorption SpectroMeter for Atmospheric CHartographY, Burrows et al., 1995;Bovensmann et al., 1999) and current (GOSAT, Greenhouse Gases Observing Satellite, Kuze et al., 2009;OCO-2, Orbiting Carbon Observatory-2, Crisp et al., 2004) satellite sensors difficult and the quantification of anthropogenic emissions a challenging task.Usually, the latter requires knowledge of the source position and assumptions on plume formation (e.g., Nassar et al., 2017;Heymann et al., 2017) or statistical approaches applied to larger areas and/or time periods (e.g., Schneising et al., 2013;Buchwitz et al., 2017).Reuter et al. (2014) followed an alternative approach to identify anthropogenic regional CO 2 enhancements by analyzing simultaneous satellite observations of tropospheric nitrogen dioxide (NO 2 ) vertical columns and column-average dry-air mole fractions of CO 2 (XCO 2 ).Nitrogen monoxide (NO) is formed and emitted to the atmosphere when fossil fuels are combusted at high temperatures.In the atmosphere, it reacts rapidly with ozone (O 3 ) and at a much slower rate via a termolecular reaction with oxygen (O 2 ) to form NO 2 .The tropospheric daytime concentrations of NO 2 are coupled with the concentrations of NO and O 3 by the Leighton photostationary state.NO 2 has a short lifetime on the order of hours so that its vertical column densities often greatly exceed background and noise levels of modern satellite sensors near sources (Richter et al., 2005) making it a suitable tracer of recently emitted CO 2 .
In contrast to SCIAMACHY, which was used by Reuter et al. (2014), OCO-2 has no NO 2 sensor aboard.However, with the launch of the S5P satellite (Sentinel-5 Precursor, Veefkind et al., 2012) in October 2017, NO 2 observations with unprecedented spatial resolution and global daily coverage became available.Here we use these data to identify OCO-2 XCO 2 enhancements, which can be attributed to localized (up to city-scale) emissions for which we estimate the plume's cross-sectional CO 2 fluxes.
In the next section, we describe the used OCO-2 XCO 2 and S5P NO 2 datasets and the developed co-location method.Also in Sect.2, we describe the used plume detection and scenario selection method as well as the cross-sectional flux estimation method.The results of our case study analyses are presented and discussed in Sects.3 and 4, respectively.
2 Datasets and methods

XCO 2
The Orbiting Carbon Observatory-2 (OCO-2, Crisp et al., 2004) was launched in 2014, aiming to continue and improve XCO 2 observations from space.OCO-2 is part of the A-train satellite constellation and flies in a sun-synchronous orbit whose ascending node crosses the Equator at 13:36 LT.It measures the solar backscattered radiance in three independent wavelength bands in the spectral regions of the near infrared (NIR) and shortwave infrared (SWIR): the O 2 -A band at around 760 nm, the weak CO 2 band at around 1610 nm, and the strong CO 2 band at around 2060 nm.OCO-2 is operated in a near-push-broom fashion and has eight parallelogram-shaped footprints across track with a spatial resolution at ground of ≤ 1.29 km × 2.25 km.
We use NASA's operational bias-corrected OCO-2 L2 Lite XCO 2 product v9 (Kiel et al., 2019;see Fig. 1a for an example), which we obtained from https://daac.gsfc.nasa.gov(last access: 17 July 2019).The product is rigorously prefiltered and post-filtered for potentially unreliable soundings including, e.g., cloud and aerosol contaminated scenes.Additionally, the OCO-2 retrieval algorithm accounts for light scattering at optically thin aerosol layers by fitting the optical depth and height of two lower-atmosphere aerosol layers and the optical depth of a stratospheric aerosol layer (O'Dell et al., 2018).The OCO-2 v9 dataset has an improved bias correction approach that results in reduced biases, particularly over areas of rough topography.
The OCO-2 XCO 2 product includes an uncertainty estimate which we use for our study.For the selected scenarios, the reported single sounding uncertainty lies typically in the range of 0.4 to 0.7 ppm, which is similar to estimates based on the standard deviation of the difference of succeeding soundings.The validation study of Reuter et al. (2017) estimated that the single sounding precision relative to groundbased Total Carbon Column Observing Network (TCCON) data is about 1.3 ppm.However, this includes, e.g., the noise of the validation dataset and a larger pseudo-noise component, due to spatial and temporal representation errors when co-locating OCO-2 with the validation data and it should be noted that the study of Reuter et al. (2017) analyzed a predecessor NASA OCO-2 XCO 2 dataset (v7 instead of v9).

NO 2
The TROPOspheric Monitoring Instrument (TROPOMI) on Sentinel-5 Precursor was launched in October 2017 into a sun-synchronous orbit with an ascending-node Equator crossing time of 13:30 LT (Veefkind et al., 2012).TROPOMI is a nadir-viewing grating imaging spectrometer for the UV-visible spectral region with additional channels in the NIR and SWIR, extending the existing data records of the GOME (Global Ozone Monitoring Experiment), SCIA-MACHY, OMI (Ozone Monitoring Instrument), and the GOME-2 missions.It has a swath width of about 2600 km and in comparison to previous instruments a much better spatial resolution of 3.5 km × 7 km at nadir at a similar signal-to-noise ratio per measurement.Here we use radiances in the spectral region 425-465 nm to retrieve NO 2 slant columns with a standard Differential Optical Absorption Spectroscopy (DOAS) retrieval developed for previous satellite instruments (Richter et al., 2011), followed by a destriping step, as described by Boersma et al. (2007).Slant columns are defined as the absorber concentration integrated along the light path, and thus depend on both the atmospheric NO 2 profile and the light path of the individual measurement.
The random noise of our S5P slant columns has been estimated from the scatter of observations over a clean Pacific region (10 • S-10 • N, 160-230 • E).In order to account for the viewing-angle dependency of the slant columns, a geometric air mass factor has been computed using only the instrument's viewing zenith angle.The evaluation suggests that the random noise (1σ ) of our S5P slant column product is typically 5 × 10 14 molec.cm −2 , while enhancements near sources often exceed 10 16 molec.cm −2 .For individual soundings, the uncertainty can differ depending on viewing geometry and surface reflectance.
Usually, in order to extract the tropospheric vertical columns, first the stratospheric contribution to the retrieved slant columns needs to be removed and then the light path dependency of the remaining tropospheric slant columns is corrected for by dividing through a scene dependent air mass factor.In this study, another approach is taken as only localized enhancements are evaluated.By subtracting the surrounding background values (Sect.2.5), both the stratospheric contribution and any tropospheric background are removed from the signal as they are both smooth on the scale of a few tens of kilometers discussed here.What remains is the slant column plume signal of the lower troposphere from which we derive information on the CO 2 plume.

Co-location of OCO-2 and S5P data
OCO-2 and S5P both fly in sun-synchronous orbits with similar Equator crossing times of their ascending nodes and with orbit times of about 100 min.S5P has a swath width of about 2600 km, which provides nearly global coverage each day.For these reasons, each scene observed by OCO-2 is also observed by S5P within a maximum time difference of about 50 min.We project the S5P and OCO-2 data of the same day in a surroundings of a potential target on a high-resolution (0.001 • × 0.001 • ) grid to compute NO 2 averages representative for the footprints of the CO 2 soundings (see Fig. 1c for an example).

Geophysical databases
As input for the computation of the cross-sectional fluxes (Sect.2.5), we compute the number of dry air particles in the atmospheric column from meteorological profiles which we read at the same time with the wind information from the ECMWF (European Centre for Medium range Weather Forecast) ERA5 (fifth-generation of ECMWF atmospheric reanalyses) data archive at 0.25 • × 0.25 • hourly resolution.This data archive also provides an uncertainty estimate of the wind information from an ensemble statistic but at a reduced resolution of about 0.5 • × 0.5 • over 3 h.We compare the inferred cross-sectional CO 2 fluxes with the following emission databases.The Emissions Database for Global Atmospheric Research (EDGAR v4.3.2, https: //edgar.jrc.ec.europa.eu,last access: 17 July 2019) provides information on anthropogenic CO 2 emissions at 0.1 • × 0.1 • annual resolution.EDGAR v4.3.2 ends in 2012 and we use the data of that year for our comparisons.The Open-Data Inventory for Anthropogenic Carbon dioxide (ODIAC v2017, http://db.cger.nies.go.jp/dataset/ODIAC (last access: 17 July 2019), Oda et al., 2018) also provides information on annual anthropogenic CO 2 emissions but at a finer resolution (1 km × 1 km monthly), and the database ends in 2016.For the reason of comparability, we re-gridded the ODIAC emissions to the EDGAR resolution (0.1 • × 0.1 • annually) and use 2012 data as baseline.Additionally, we use ODIAC v2017 data re-gridded to 0.1 • × 0.1 • monthly resolution.The Global Fire Emissions Database (GFED v4.1s, https://www.globalfiredata.org,last access: 17 July 2019) provides information on CO 2 emissions from wildfires at a resolution of 0.25 • × 0.25 • over 3 h, which we re-gridded to 0.1 • × 0.1 • resolution for a 6 h average, ending approximately at the time of the overpass.

Flux estimation
S5P's spatial resolution is considerably coarser than that of OCO-2.Consequently for our case studies, we concentrate on plumes that are significantly larger than the swath width of OCO-2.This means that for the selected scenarios, OCO- 2 actually sees only a cross section of a plume (see Fig. 1c for an example).
We model the cross-sectional NO 2 columns along the OCO-2 orbit via a linear polynomial, accounting for largescale variations in the background values, overlaid by a Gaussian function describing the enhancement within the plume.Simultaneously, the cross-sectional CO 2 concentrations are modeled in a similar manner.However, the width of the CO 2 Gaussian function is constrained to equal the width of the NO 2 Gaussian function.This means that the plume shape is determined from the NO 2 measurements, but we allow for a shifted position of the maximum in order to account for potential plume displacements resulting from different overpass times.Additionally, it should be noted that the CO 2 and NO 2 plumes may have small differences, e.g., due to different decay rates of NO 2 in different altitudes.These differences, however, are considered minor compared with the precision of the XCO 2 soundings.Specifically, the co-located NO 2 and XCO 2 values along the distance in OCO-2's flight direction x are fitted with the maximum likelihood method by the following vector function: The free fit parameters a 0-8 correspond to the polynomial coefficients of the background values (a 0, 1, 5, 6 ), the amplitudes (a 2, 7 ), shifts (a 3, 8 ), and the full width at half maximum (FWHM, a 4 ) of the Gaussian functions.We force the FWHM to be constrained entirely by the NO 2 measurements by setting the CO 2 part of the corresponding Jacobian artificially to zero.However, we expect only little differences with a combined FWHM fit because of the lower relative noise of the NO 2 measurements.
Integration over the Gaussian enhancement results in the cross-sectional CO 2 flux F CO 2 (mass of CO 2 over time) of the plume depending on the FWHM a 4 , the amplitude of the XCO 2 enhancement a 7 , the effective wind speed v e within the plume normal to the OCO-2 orbit, and the number of dry air particles in the atmospheric column n e : Here, M CO 2 is the molar mass of CO 2 (44.01 g mol −1 ) and N A is the Avogadro constant (6.02214076 × 10 23 mol −1 ).We approximate the number of dry air particles n e and the effective wind speed's normal v e from ECMWF ERA5 meteorological profiles at the position of the maximum of the fitted Gaussian XCO 2 function.In regions with large variations in surface elevation or wind conditions within the plume's cross section, it might be appropriate to account for variations in the number of dry air particles and/or the wind conditions when integrating over the Gaussian enhancement.
We manually adjust the ECMWF wind direction (not the wind speed) to subjectively fit the plume direction observed in the NO 2 fields (e.g., Fig. 1a).The manual adjustment to wind direction but not wind speed is similar to the approaches of, e.g., Krings et al. (2011) or Nassar et al. (2017).
For a hydrostatic atmosphere with a standard surface pressure of 1013 hPa, n e is about 2.16 × 10 25 cm −2 and the crosssectional CO 2 flux F CO 2 (Eq.2) in units of MtCO 2 a −1 becomes approximately given that the FWHM a 4 , the amplitude of the XCO 2 enhancement a 7 , and the effective wind speed v e are provided in the units km, ppm, and m s −1 , respectively.As n e approximately scales with the surface pressure, Eq. ( 3) may be easily adapted to other meteorological conditions.As discussed by Brunner et al. (2019), the plume height (and subsequently the wind speed in plume height) depends on many aspects like emission height, stack geometry, flue gas exit velocity and temperature, meteorological conditions, etc.Some of these parameters are not known for many sources and their explicit consideration would go beyond the scope of this study focusing on demonstrating the benefits of simultaneous NO 2 and XCO 2 measurements rather than on most accurate flux estimates.Varon et al. (2018) proposed approximating the effective wind speed within the plume from the 10 m wind by applying a multiplier in the range of 1.3-1.5.Therefore, we decided to use a multiplier of 1.4 for convenience.This empirical relationship accounts for plume rise and mixing into altitudes with larger wind speeds, for example.For the present, we consider this approximation adequate for this first study, but we recognize that uncertainties (see next section) resulting from this estimate of the effective wind speed's normal may be reduced in the future by improved wind knowledge.
Additionally, it should be noted that the plume crosssectional flux (Eq.2) is only a good approximation for the actual source emission under steady-state (temporally invariant) conditions for wind speeds greater than about 2 m s −1 (Varon et al., 2018), when advection dominates over diffusion (Sharan et al., 1996).Changes in wind direction, wind speed, or atmospheric stability in the time span between emission and observation may result in differences between the plume cross-sectional flux and the source flux.Temporal variations in the source emissions of course also result in (temporally delayed) variations in the plume cross-sectional flux, which is always only a snapshot and must not be confused with, e.g., the annual average, even though it is given in the same units.In case of chemically active species (such as NO 2 ), chemical processes along the plume path would also have to be considered in order to compute source emissions from plume cross-sectional fluxes.

Uncertainty propagation
In order to estimate the uncertainty of the CO 2 plume crosssectional flux (F CO 2 , Eq. 2), we propagate the uncertainties of the FWHM (a 4 ), the amplitude (a 7 ), and the wind speed normal (v e ) by assuming uncorrelated errors.The uncertainties of the FWHM and the amplitude result from the maximum likelihood fitting method propagating the uncertainties of the individual XCO 2 and NO 2 soundings as reported in the data products.The uncertainties of the wind components are read from the ECMWF ERA5 data archive resulting in total wind speed uncertainties ranging from 0.18 to 0.33 m s −1 for the analyzed scenarios.Additionally, we assume that the manual adjustment of the wind direction is accurate by ±10 • .These uncertainties propagate into the uncertainty of the wind speed normal.Varon et al. (2018) estimated that computing the effective wind speed from the 10 m wind introduces an additional uncertainty of 8 %-12 %.However, we analyze scenarios with larger plume structures and probably also larger variations in the injection heights, which we consider by enhancing this error component to 20 % for convenience.Uncertainties in the number of dry air particles are neglected as they are much smaller compared to the wind speed uncertainty, for example.As mentioned earlier, the assumption of constant meteorological conditions might not be valid in regions with large variations in surface elevation or wind conditions within the plume's cross section, which may result in an underestimation of the total cross-sectional flux uncertainty in such cases.

Plume detection and scenario selection
We use a semiautomatic method to select potentially interesting targets.In a first step, all co-locations of OCO-2 and S5P are computed similarly to those described in Sect.2.3 but based on a coarser high-resolution grid (0.01 further considered that have at least 100 co-locations without data gaps exceeding 3 s (∼ 20 km) within the search window.In the next step, we perform a least-squares fit of the co-located XCO 2 and NO 2 data with a Gaussian vector function.This fitting function corresponds to Eq. ( 1) but with independent FWHM for XCO 2 and NO 2 and centered within the search window (a 3 and a 8 set to zero), which improves the convergence rate.Only those time steps that fulfill the following criteria are considered further: the fit converges, the NO 2 amplitude exceeds 10 15 molec.cm −2 , the XCO 2 and NO 2 FWHM (a c and a n , respectively) do not exceed the half width of the search window (a c , a n ≤ 15 s) and do not differ by more than their average (|a c − a n | ≤ (a c + a n )/2), and the XCO 2 and NO 2 amplitudes are at least 2 times larger than their uncertainties and larger than the maximum variations in the backgrounds.In the last step, we decided by manual inspection of the XCO 2 and NO 2 co-locations plus the surrounding NO 2 fields and ECMWF wind information if the scenario is a promising candidate for further flux analyses.Potential reasons to reject an automatically preselected scenario are, e.g., too low wind speed, wind direction nearly parallel to OCO-2 orbit, unclear source attribution, or poor fit quality.In total, we manually identified about 20 promising scenarios in the time period January to August 2018 of which we selected and analyzed six examples for this study.

Results
From the time period of January to August 2018, we selected the following scenarios as examples for flux analyses based on co-located XCO 2 and NO 2 observations.

Moscow
Figure 1a shows the NO 2 enhancement in the city plume of Moscow (approx.12.4 million inhabitants) as retrieved from S5P overlaid by OCO-2's XCO 2 measurements.The NO 2 enhancement is clearly also visible in the plume's cross section along OCO-2's ground track (Fig. 1c).Due to the larger relative noise of the XCO 2 retrievals, the XCO 2 enhancement is less obvious but still visible (Fig. 1c).The Gaussian fit of the enhancements is excellent for NO 2 and reasonable (χ 2 = 2.2) for XCO 2 .There was nearly no adjustment needed (1 • ) to bring the ECMWF 10 m wind in good agreement with the NO 2 plume (Fig. 1a).The effective wind speed normal to the OCO-2 orbit amounts to 1.6±0.6 m s −1 , which is a bit lower than optimal for reasonable flux estimates (Varon et al., 2018).The cross-sectional CO 2 flux amounts to 76 ± 33 MtCO 2 a −1 .This compares to 2012 average upwind emissions (white boxes in Fig. 1a) of 195 MtCO 2 a −1 (EDGAR) and 102 MtCO 2 a −1 (ODIAC).ODIAC's emission estimate for August 2016 amounts to 88 MtCO 2 a −1 .The NO 2 far field shows no indications of overlaid CO 2 plumes from other sources (Fig. 1b).The total flux uncertainty is dominated by the uncertainty of the wind direction followed by the uncertainty of the effective wind speed.

Lipetsk
Figure 2a shows the surroundings of Lipetsk (approx.0.5 million inhabitants) with, among other industries, the Novolipetsk steel plant and the Lipetskaya TEC-2 gas-fired power plant (515 MW) only 1 min (∼ 400 km) away from Moscow along OCO-2's flight track (see also Fig. 1b).The cross-sectional NO 2 and XCO 2 enhancements clearly stand out from the noise in the data (Fig. 2c) and the Gaussian function fits the XCO 2 data reasonably well (χ 2 = 2.4).We applied a small correction of 5 • to the ECMWF wind direction.However, as the wind direction is similar to OCO-2's flight direction, the normal effective wind speed is unfavorably low (0.9 ± 0.7 m s −1 ), which makes the cross-sectional flux estimates (69 ± 50 MtCO 2 a −1 ) less reliable and highly uncertain.The largest uncertainty contribution by far comes from the uncertainty of the wind direction.The 2012 average EDGAR and ODIAC upwind emissions (white marked boxes in Fig. 2a) are 23 and 4 MtCO 2 a −1 (same for August 2016), respectively, but the NO 2 far field shows no indications of overlaid CO 2 plumes from other sources (Fig. 2b).

Baghdad
Figure 3a shows the S5P NO 2 slant columns overlaid by OCO-2 XCO 2 data in a surroundings of Baghdad (approx.5.4 million inhabitants).Enhanced values are clearly visible in the cross section of the NO 2 plume and less obviously also visible in the XCO 2 data (Fig. 3c).The XCO 2 enhancement is well-fitted (χ 2 = 1.0) by the Gaussian fitting function.The manually adjusted wind direction deviates by 17 • from the ECMWF wind direction and the normal wind speed amounts to 4.4 ± 1.7 m s −1 .From the XCO 2 enhancement and the normal wind speed, we compute the crosssectional CO 2 flux to be 95 ± 36 MtCO 2 a −1 .This compares to an upwind source of 22 or 13 MtCO 2 a −1 (12 MtCO 2 a −1 for July 2016) of EDGAR or ODIAC, respectively.The flux uncertainty is dominated by the uncertainty of the wind direction and the uncertainty of the effective wind speed.The NO 2 far field shows no indications of overlaid CO 2 plumes from other sources (Fig. 3b).

Medupi and Matimba power plants
The Medupi (4764 MW) and Matimba (3990 MW) coal-fired power plants lie close to each other in South Africa, about 300 km north of Johannesburg.Their NO 2 plume is shown in Fig. 4a overlaid by OCO 2 XCO 2 measurements.NO 2 measurements in the larger surrounding do not suggest any additional nearby upwind sources (Fig. 4b).The cross-sectional NO 2 values show a clear elevation within the plume that is less obvious for XCO 2 , which has larger relative scatter, especially south of the plume.Nevertheless, the Gaussian  function fits the XCO 2 values reasonably well (χ 2 = 1.4).The wind direction (corrected by 13 • ) is nearly perpendicular to the OCO-2 orbit and the effective normal wind speed is 2.6 ± 0.6 m s −1 .The cross-sectional CO 2 flux amounts to 31±7 MtCO 2 a −1 , which is consistent with ODIAC 2012 emissions of 24 MtCO 2 a −1 and ODIAC July 2016 emissions of 26 MtCO 2 a −1 but EDGAR does not have significant emissions in this area.It should be noted that the Medupi power plant started operation in 2015 with limited capacity and that it still has not reached its nominal capacity.Therefore, it is no surprise that the Medupi power station is not included in either EDGAR or ODIAC 2012 data.The flux uncertainty is dominated by the uncertainty of the effective wind speed.

Australian wildfires
Figure 5a shows the NO 2 plumes of two Australian wildfires on 5 May 2018 overlaid by an OCO-2 orbit of XCO 2 measurements.Enhanced NO 2 and XCO 2 values are clearly visible within the plume's cross section (Fig. 5b).The NO 2 (and also less obviously the XCO 2 ) cross section has two maxima that cannot be accounted for by the Gaussian fitting function.However, this is not reflected in the good XCO 2 fit quality (χ 2 = 0.6) but should be taken into account when valuing the results.We applied a small manual correction of 7 • to the wind direction and the effective wind speed normal to the OCO-2 orbit is 6.7 ± 1.7 m s −1 .For the snapshot of the overpass, we computed a cross-sectional CO 2 flux of 153 ± 40 MtCO 2 a −1 .Its uncertainty is driven by the uncertainty of the effective wind speed and wind direction.As the shown plumes originate from wildfires, EDGAR and ODIAC do not include their emissions.However, GFED has average emissions of 52 MtCO 2 a −1 within the 6 h period 00:00-06:00 UTC including the time of the overpass (05:00 UTC).The maximum GFED emissions are approximately at the position of the largest NO 2 concentrations.Figure 5c shows no indications that additional upwind sources explain the discrepancy between our cross-sectional flux estimate and GFED.

Nanjing
Figure 6a shows the NO 2 slant columns in the surroundings of Nanjing (approx.5.8 million inhabitants) overlaid by OCO-2 XCO 2 measurements.The cross section along the OCO-2 orbit shows strong XCO 2 and NO 2 plume signals distinctively above the noise level that are well-fitted with the Gaussian fitting function (χ 2 = 0.6).The ECMWF wind direction is not far from being rectangular to the OCO-2 orbit, and we applied a moderate manual correction of 11 • .The effective normal wind speed is 2.2 ± 0.5 m s −1 .This results in a cross-sectional flux estimate of 120 ± 27 MtCO 2 a −1 , which lies in between the upwind emissions of EDGAR (163 MtCO 2 a −1 ) and ODIAC (89 MtCO 2 a −1 for 2012, 96 MtCO 2 a −1 for March 2016).Figure 6b does not indicate additional major remote upwind sources.The uncertainty of the cross-sectional flux estimate is dominated by the uncertainty of the effective wind speed.

Discussion and conclusions
Based on six case studies (Moscow, Russia; Lipetsk, Russia; Baghdad, Iraq; Medupi and Matimba power plants, South Africa; Australian wildfires; and Nanjing, China), we demonstrated the usefulness of simultaneous satellite observations of NO 2 and the column-average dry-air mole fraction of CO 2 (XCO 2 ).For this purpose, we analyzed colocated regional enhancements of XCO 2 observed by OCO-2 and NO 2 from S5P and estimated the CO 2 plume's crosssectional fluxes.For atmospheric standard conditions, we approximated as a rule of thumb that a Gaussian enhancement of 1 ppm with a width of 1 km at a wind speed (normal to the cross section) of 1 m s −1 corresponds to a plume crosssectional flux of roughly 0.53 MtCO 2 a −1 .
For Moscow, we derived a cross-sectional flux of 76 ± 33 MtCO 2 a −1 , which agrees (within its uncertainty) with ODIAC 2012 emissions of 102 MtCO 2 a −1 (88 MtCO 2 a −1 for August 2016) but not with EDGAR emissions of 195 MtCO 2 a −1 .The cross-sectional flux estimate of Lipetsk with the Novolipetsk steel plant and the Lipetskaya TEC-2 power plant is 69±50 MtCO 2 a −1 .Within its uncertainty, this estimate agrees with EDGAR emissions of 23 MtCO 2 a −1 but not with ODIAC emissions of 4 MtCO 2 a −1 .However, the uncertainty of the estimate is large due to a wind direction with an acute angle relative to the OCO-2 orbit, which also results in a low effective normal wind speed.This can serve as an example for low wind speeds being favorable for plume detection but not necessarily for flux quantification.In the case of Baghdad, we derived a cross-sectional flux of 95 ± 36 MtCO 2 a −1 for the time of the overpass, which is considerably larger than the annual average EDGAR (22 MtCO 2 a −1 ) and ODIAC (13 MtCO 2 a −1 for 2012, 12 MtCO 2 a −1 for July 2016) emissions of 2012.The wind conditions were relatively good and S5P NO 2 measurements do not suggest an overlaying significant upwind source.In this context, it is interesting to note that Georgoulias et al. (2019) found a strongly increasing trend (17.0 ± 0.8 % / a in the period April 1996-September 2017) for the tropospheric NO 2 concentrations in Baghdad (and a decreasing trend of −2.2 ± 0.7 % / a for Iraq) hinting at strongly increasing CO 2 emissions in Baghdad since 2012.The cross-sectional flux of the plume of the Medupi and Matimba power plants have been estimated to 31±7 MtCO 2 a −1 , which agrees (within its uncertainty) with ODIAC (24 MtCO 2 a −1 for 2012, 26 MtCO 2 a −1 for July 2016) but not with EDGAR (no significant emission).Nassar et al. (2017) also estimated the emissions from the Matimba power plant (but not Medupi) using  OCO-2 XCO 2 v7 data.For a direct overpass in 2014 and a close flyby (∼ 7 km away) in 2016, they found fluxes, converted to annual values, of 12.1 ± 3.9 MtCO 2 a −1 and 12.3 ± 1.2 MtCO 2 a −1 , respectively.For the Australian wildfires, we estimated a plume cross-sectional flux of 153 ± 40 MtCO 2 a −1 , which is about 3 times larger than the GFED estimate (52 MtCO 2 a −1 ) for a 6 h average ending approximately at the time of the OCO-2 overpass.Unfavorable wind conditions or a strong overlaying upwind source can be excluded as a reason for the discrepancy.The same is true for the fact that a double-plume structure has been fitted with a Gaussian function.However, it should be noted that GFED's emission estimate for the same time interval but 1 d before the OCO-2 overpass amounts to 252 MtCO 2 a −1 .
The total uncertainty of the derived plume cross-sectional fluxes ranges from 7 to 50 MtCO 2 a −1 or in relative measures from 23 % to 72 %.The total uncertainty is always dominated by an uncertainty contribution related to meteorology.Specifically, the (manually adjusted) wind direction or the computation of the effective wind speed from the 10 m wind contribute most to the total uncertainty.The noise of the XCO 2 retrievals contributes with only 1 to 8 MtCO 2 a −1 to the total error and the noise of the NO 2 retrievals contributes 3 times less on average.
It is unlikely that the observed XCO 2 enhancements are dominated by uncorrected enhancements due to co-emitted aerosols because the OCO-2 retrieval algorithm accounts for light scattering at optically thin aerosol layers and filters scenes with stronger aerosol contamination.Additionally, Bovensmann et al. (2010) estimated for the proposed CarbonSat (Carbon Monitoring Satellite) instrument that neglecting co-emitted aerosols in power plant plumes results in errors between 0.2 and 2.5 MtCO 2 a −1 , which is small compared with the derived cross-sectional fluxes and their total uncertainties (Table 1).Aerosols can also effect the S5P NO 2 slant columns which is, however, less important for our work because we derive only the plume width and direction from the NO 2 observations.
It should be noted that differences of the cross-sectional flux estimates and the emission databases are not necessarily coming from inaccuracies of the satellite retrievals or the emission databases.Our estimates are valid only for the time of the overpass, while the emission databases give annual or monthly averages.Velazco et al. (2011) illustrated that power plants can have substantial annual and day-to-day variations.Additionally, the cross-sectional flux is only a good approximation for the source emission under meteorological steadystate conditions with wind speeds greater than about 2 m s −1 (Varon et al., 2018).
Table 1.Summary of cross-sectional flux results including uncertainty contributions (1σ ) and comparison with emission databases EDGAR and ODIAC or GFED in the case of the Australian wildfires.The ODIAC values in brackets represent ODIAC emissions of 2016 and the month of the overpass in the same grid boxes as summed up for 2012.Note that the cross-sectional flux results correspond to the instantaneous time of the overpasses, while EDGAR and ODIAC emissions are annual or monthly averages; GFED emissions correspond to 6 h averages (see Sect. 2.4).The uncertainty estimate is comprised of the total uncertainty and the uncertainties introduced by the ECMWF wind uncertainty, the uncertainty of the wind direction (10 • ), use of the 10 m wind (20 %), the XCO 2 precision as reported in the data product, and the NO 2 precision as reported in the data product.All values are in units of MtCO 2 a −1 .For the analyzed scenarios, we observe rather large differences between the EDGAR and ODIAC emission inventories.However, note that only those grid boxes are shown (and summed up) in Figs.1a-6a for which either EDGAR or ODIAC emissions are larger than 0.5 MtCO 2 a −1 .This means a smoother distribution of emissions may be misinterpreted as fewer emissions if a significant fraction of the total emission is located in grid boxes not exceeding the 0.5 MtCO 2 a −1 threshold.Additionally, it should be noted that ODIAC emissions correspond to fossil fuel combustion and cement production only, while EDGAR also includes emissions from other sectors (e.g., agriculture, land use change, and waste).
NO 2 is co-emitted with CO 2 when fossil fuels are combusted at high temperatures and has a relatively short lifetime on the order of hours, which makes it a suitable tracer for recently emitted CO 2 .Despite less strict quality filtering being needed, plume enhancements of NO 2 columns near sources can be retrieved from satellites with much lower relative noise than is the case for XCO 2 .We take advantage of these points by using NO 2 measurements to (i) identify the source of the observed XCO 2 enhancements, (ii) to exclude interference with potential additional remote upwind sources, (iii) to manually adjust the wind direction, and (iv) to put a constraint on the shape of the observed CO 2 plumes.
In principle, it is also possible to fit only the XCO 2 values without constraining the plume shape by NO 2 .In this case, XCO 2 is used to derive the amplitude and FWHM of the enhancement.We repeated the flux estimation of all shown scenarios with such a setup and got fluxes of 61 ± 27 MtCO 2 a −1 , 63±46 MtCO 2 a −1 , 75±29 MtCO 2 a −1 , 35± 9 MtCO 2 a −1 , 166±44 MtCO 2 a −1 , and 119±28 MtCO 2 a −1 for the Moscow, Lipetsk, Baghdad, Medupi/Matimba, Australian wildfire, and Nanjing scenarios, respectively.The derived fluxes are consistent within their uncertainty with our main results shown in Table 1, but the uncertainty contribu-tion due to the noise in the XCO 2 data increased by 34 % from 4.7 to 6.3 MtCO 2 a −1 on average.Reuter et al. (2014) discussed that post-ENVISAT missions such as OCO-2 would benefit from co-located measurements of co-emitted species from other satellites or (ideally) multispecies measurements from the same instrument.We demonstrated that the analysis of small-scale emissions in OCO-2 XCO 2 data indeed profits from simultaneous NO 2 observations of S5P as they not only allow us to set the XCO 2 observations into context but also to constrain the plume structure.The uncertainties of the cross-sectional flux estimates due to meteorology and their agreement with the actual emissions might be improved in subsequent studies by making use of dedicated simulations with Lagrangian particle dispersion models with either known source positions (and injection heights) or source positions inferred from the NO 2 data.
However, we expect the largest room for improvement to be in satellite missions such as the planned European Copernicus anthropogenic CO 2 monitoring mission (CO2M), which will provide not only precise measurements with high spatial resolution but also imaging capabilities with a wider swath of simultaneous XCO 2 and NO 2 observations.Its imaging capabilities will reduce the uncertainty of the inferred emissions due to measurement noise simply because of the increased number of soundings.Additionally, simultaneous XCO 2 and NO 2 observations from the same platform will allow stricter constraints on the plume shape.More importantly, the meteorology related uncertainties will reduce (Varon et al., 2018) because deviations from steady-state conditions can average out and are, therefore, less critical if the entire plume structure is sampled rather than only a cross section.
Data availability.The research data used are available at the sources specified in the dataset section.

Figure 1 .
Figure 1.Moscow on 25 August 2018.(a) S5P NO 2 slant column (background, pale colors) overlaid by OCO-2 XCO 2 (foreground, saturated colors).Gray and white 0.1 • boxes show EDGAR (bottom number in each box) and ODIAC (top number in each box) 2012 annual emissions with either EDGAR or ODIAC being larger than 0.5 MtCO 2 a −1 .The white arrows show the direction of the 10 m wind as read from ECMWF (dotted), manually corrected to (subjectively) best match the NO 2 plume (solid), and normal to the OCO-2 orbit (dashed).Effective wind speed normal to the OCO-2 orbit, estimated cross-sectional CO 2 flux, time of OCO-2 overpass, and time difference between OCO-2 and S5P overpass are also listed.The hatched area corresponds to the urban area (World Urban Areas dataset, Geoportal of the University of California, https://apps.gis.ucla.edu/geodata/dataset/world_urban_areas).(b) Larger section of the S5P NO 2 slant columns including the OCO-2 orbit and the bounding box of (a).(c) OCO-2 XCO 2 values (red) and co-located S5P NO 2 slant columns (black) within the plume's cross section in OCO-2 flight direction.

Figure 3 .
Figure 3.As in Fig. 1 but for Baghdad on 31 July 2018.

Figure 4 .
Figure 4.As in Fig. 1 but for the Medupi and Matimba power plants in South Africa on 11 July 2018.

Figure 5 .Figure 6 .
Figure 5.As in Fig. 1 but for the Australian wildfires on 5 May 2018.The ODIAC emission data (the top number in the white boxes) have been replaced by GFED emissions for the time of the OCO-2 overpass.