Building a bridge: characterizing major anthropogenic point sources in the South African Highveld region using OCO-3 carbon dioxide snapshot area maps and Sentinel-5P/TROPOMI nitrogen dioxide columns

In this paper, we characterize major anthropogenic point sources in the South African Highveld region using Orbiting Carbon Observatory-3 (OCO-3) Snapshot Area Map (SAM) carbon dioxide (CO2) and Sentinel-5 Precursor (S5P) TROPOspheric Monitoring Instrument (TROPOMI) nitrogen dioxide (NO2) observations. Altogether we analyze six OCO-3 SAMs. We estimate the emissions of six power stations (Kendal, Kriel, Matla, Majuba, Tutuka and Grootvlei) and the largest single emitter of greenhouse gas (GHG) in the world, Secunda CTL synthetic fuel plant. We apply the cross-sectional flux method for the emission estimation and we extend the method to fit several plumes at the same time. Overall, the satellite-based emission estimates are in good agreement (within the uncertainties) as compared to emission inventories, even for the cases where several plumes are mixed. We also discuss the advantages and challenges of the current measurement systems for GHG emission monitoring and reporting, and the applicability of different emission estimation approaches to future satellite missions such as the Copernicus CO2 Monitoring Mission (CO2M) and the Global Observing SATellite for GHGs and Water cycle (GOSAT-GW), including the joint analysis of CO2 and NO2 observations.


Introduction
The Paris Agreement, adopted in 2015, requires monitoring of anthropogenic greenhouse gas (GHG) emissions and assessment of collective climate mitigation efforts. While GHG emission information can be derived using a wide variety of methods (e.g. NASEM 2022), the atmospheric-based approach has a potential of providing GHG information over the areas with lower inventory compilation capacity and/or for key sub-national sources (state/province, cities, and facilities). Recent GHG missions have demonstrated that space-based observations offer new opportunities by providing denser global data. Several space agencies have further responded to this call by designing new CO 2 monitoring missions with a focus on anthropogenic emissions, e.g. the Copernicus CO 2 Monitoring (CO2M) mission in Europe (Janssens-Maenhout et al 2020, Meijer et al 2020), Global Observing SATellite for Greenhouse gases and Water cycle (GOSAT-GW, Kasahara et al 2020) mission in Japan and TanSat-2 mission in China. Concerning fugitive methane emissions, several commercial small satellite instruments (e.g. GHGSat) have demonstrated their capability of targeting intense methane plumes from space (e.g. Varon et al 2019Varon et al , 2020. The SCIAMACHY instrument (2002-2012 was the first space-based sensor to return observations of reflected sunlight that could be analyzed to yield estimates of the column-averaged CO 2 dry air mole fraction (XCO 2 ) that were sensitive to CO 2 variations near the surface (Reuter et al 2014). Since then, Japan's Greenhouse gas Observing SATellite (GOSAT), launched in 2009, has also produced observations that enable the detection of anthropogenic CO 2 emission signatures (Kort et al 2012, Janardanan et al 2016. Recently, Kuze et al (2022) developed a new algorithm to retrieve lower-tropospheric CO 2 concentrations from GOSAT target observations over global megacities. Since the launch of the Orbiting Carbon Observatory-2 (OCO-2) mission in 2014 (Eldering et al 2017), several methods have been used for inferring CO 2 emissions from space, e.g. the Gaussian plume model (Nassar et al 2017) and cross-sectional flux method (Reuter et al 2019). Anthropogenic emission enhancements have also been observed by the Chinese TanSat mission (Yang et al 2022).
Analyzing anthropogenic emissions of short-lived trace gases, such as nitrogen dioxide (NO 2 ), from space has been a topic of numerous academic studies since the late 1990s (e.g. Beirle et al 2003). In recent years, satellite-based atmospheric observations have been used to assess the trends in air polluting emissions (Krotkov et al 2016, Georgoulias et al 2019, to verify the success of environmental policy measures (Castellanos and Boersma 2012) and the efficacy of new clean technologies in reducing emissions (Ialongo et al 2018). Atmospheric satellite data have also shown potential in monitoring the progress towards the sustainability development goals.
The link between the co-emitted nitrogen oxides (NO x ) and carbon dioxide (CO 2 ) has been studied since joint CO 2 and NO 2 observations were provided by SCIAMACHY (Reuter et al 2014). Since 2018, Copernicus Sentinel-5 Precursor (S5P), with its payload the TROPOspheric Monitoring Instrument (TROPOMI, Veefkind et al 2012), has been providing satellite-based NO 2 observations with improved spatial resolution (currently 5.5 km by 3.5 km at nadir), which allows the instrument to detect individual NO 2 plumes from point sources. These observations have been very useful for detecting the CO 2 plumes and approximating their shape (Reuter et al 2019, Hakkarainen et al 2021. The first concurrent airborne observations of NO 2 and CO 2 plumes from power plants was provided by (Fujinawa et al 2021). The improved spatial resolution has also fostered new methodological developments for the emission estimation, such as the divergence method (Beirle et al 2019(Beirle et al , 2021, which enabled the compilation of a global catalogue of NO 2 emissions from point sources. The divergence method can be potentially applied to CO 2 observations from the upcoming CO2M mission . Recent advances in space-based CO 2 observations have been provided by the NASA's OCO-3 mission  on the International Space Station (ISS). To support the quantification and monitoring of anthropogenic CO 2 emissions (Kiel et al 2021), OCO-3 incorporates a new key capability that provides observations in Snapshot Area Maps (SAMs), providing contiguous images over regions as large as 80 km by 80 km. These data have been recently used for assessing CO 2 emission reduction of the Bełchatów power station in Poland (Nassar et al 2022). NASA plans to extend further the OCO-3 mission on the ISS.
In this paper, we aim to characterize the emissions from the major point sources in the South African Highveld region using OCO-3 SAM CO 2 and TRO-POMI NO 2 observations. This area includes several strong point sources, i.e. multiple power plants and the Secunda CTL synthetic fuel plant. Hakkarainen et al (2021) presented the analysis of individual plumes generated from the isolated Matimba power station based on OCO-2 observations, which practically enables the analysis of only one cross-section of the plume due to the narrow swath. In this paper, we use OCO-3 SAM data, which can cover a larger fraction of the emission plumes and allow the analysis of multiple cross-sections for each plume. We extend the cross-sectional flux method for estimating CO 2 emissions to fit several plumes at the same time. We discuss the advantages and challenges of a setup where several cross-sections of the plumes are available and where plumes from multiple sources overlap. Finally, we discuss the applicability of similar emission estimation approaches for future satellite missions such as CO2M, GOSAT-GW and MicroCarb (Wu et al 2022), including the joint analysis of CO 2 and NO 2 plumes.

Space-based CO 2 and NO 2 observations
Orbiting Carbon Observatory-3 is NASA's CO 2 observing instrument operating on the International Space Station. It incorporates OCO-2 flight spare components along with a fast pointing mirror assembly that provides new capabilities. OCO-3 was launched on 04 May 2019 with focus on monitoring anthropogenic emission signatures. While producing observations similar to the OCO-2 mission in nadir and glint measurement modes, OCO-3 also provides observations in Snapshot Area Map mode that covers 80 km by 80 km target areas in 2 min. The along-track footprint size is about 2.2 km at nadir and the crosstrack footprint size is smaller than 1.6 km, which results in a footprint area of about 3.5 km 2 , slightly larger than that of OCO-2 (Kiel et al 2021). In this study, we use bias-corrected XCO 2 values with quality flag screening from the latest available data version V10.4r. The OCO-3 data can be accessed from NASA's GES DISC https://disc.gsfc.nasa.gov. More information about the OCO-3 mission operations is provided by Taylor et al (2020). The bias in the OCO-3 SAM observations is discussed in detail by Bell et al (2022). The bias correction for two cases (4 July 2021 and 2 August 2020) were further optimized for pointing errors (by C W O'Dell, Kiel et al 2019). The effect of this update on the mean CO 2 emission estimates was less than 4%.
The TROPOspheric Monitoring Instrument (Veefkind et al 2012) was launched onboard the Copernicus Sentinel-5P satellite on 13 October 2017. The satellite has the equatorial crossing time of 13:30 LT. The current spatial resolution is 5.5 km by 3.5 km at nadir and it covers about 2600 km wide swath. TROPOMI observes atmospheric parameters related to air quality, ozone and UV radiation, and climate monitoring. In this study, we use the tropospheric NO 2 vertical columns with quality flag screening. Due to changes in the operational algorithm, we use the version v02.03.01 intermediate reprocessing on the S5P-PAL system https://data-portal. s5p-pal.com/, which provides a seamless connection with the operational version 2.3.1 data product. Technical details can be found from the readme file: https://data-portal.s5p-pal.com/product-docs/no2/ PAL_reprocessing_NO2_v02.03.01_20211215.pdf. The TROPOMI NO 2 products are operationally validated by the S5P-MPC-VDAF (S5P-Mission Performance Centre-Validation Data Analysis Facility) using ground-based observations. The operational validation results are periodically reported at the S5P-MPC-VDAF website: http://mpc-vdaf.tropomi.eu/.

FLEXPART simulations
We use the FLEXible PARTicle (FLEXPART) model to simulate CO 2 and NO 2 enhancements to support the interpretation of satellite observations and to analyze the NO x -to-CO 2 emission ratio. FLEXPART is a multi-scale Lagrangian particle dispersion model created to simulate the transport, diffusion, dry/wet deposition, radioactive decay, and 1st-order chemical reactions of tracers released from different sources (Stohl et al 2005). We utilise FLEXPART version 10.4 (Pisso et al 2019) in forward mode driven by ERA5 meteorological fields from ECMWF Integrated Forecast System (IFS). ERA5 data were used at 0.5 • × 0.5 • horizontal resolution, 1 h time resolution and 137 vertical levels. Concentrations of the particles released from the eight considered sources were determined on a regular latitude-longitude grid. The domain of the gridded output was 19 • -34 • E, 19 • -29 • S with 0.05 • × 0.05 • resolution. The number of particles released at each source was 10 000 and the release height was 150 m above ground. NO 2 simulations were carried out with chemical decay with the lifetime of 4 h (as in, e.g. Beirle et al 2019, Hakkarainen et al 2021.

Cross-sectional flux method
To derive the CO 2 emissions from the observed plumes, we use the cross-sectional flux method (Varon et al 2018, Reuter et al 2019, where the emission, E, is obtained by integrating the plume along the axis perpendicular to the wind direction and then multiplying this integral, C, by the effective wind speed: E = U eff × C. If the shape of the plume is assumed to be Gaussian, the integral can be calculated as: where A is the amplitude of the Gaussian function and FWHM is the full width at half maximum. Reuter et al (2019) applied this method to OCO-2 data, with the FWHM calculated from collocated TROPOMI NO 2 slant-column observations. The effective wind speed is calculated from the external data. Here, we extend this approach to fit several plumes at the same time (see section 3 for examples).
In the cross-sectional flux method, the plume is defined downwind of the emission source and it is 'sliced' along the principal wind direction. The crosssectional integral is then calculated from these slices, e.g. by fitting the Gaussian function. Due to the narrow swath, with OCO-2 observations, we essentially obtain one cross-section per plume (Reuter et al 2019, Hakkarainen et al 2021, which is not necessarily perpendicular to the plume direction. With OCO-3 SAM observations, we can observe a larger portion of the plume and we can select several transects along the plume. In this paper, we consider the transects every 0.025 • intervals along the direction of the plume in the latitude-longitude space. In practice, we calculate the mean from three transects for every second center point, i.e. running mean. We then fit the satellite data using Gaussian function to obtain the the amplitude A and the FWHM to calculate the cross-sectional flux. Here, we fit the plumes separately for NO 2 and CO 2 . Once all the cross-sectional fluxes have been calculated for each transect, we calculate the mean and the standard deviation of the fluxes for each plume. The calculation of cross-sectional fluxes for all transects of each plume is illustrated in the supplementary material. Concerning the wind speed information, we follow the approach of Varon et al (2018) and Reuter et al (2019) and calculate the effective wind speed, U eff , from the European Centre for Medium-Range Weather Forecasts (ECMWF) next-generation reanalysis ERA5 dataset (Hoffmann et al 2019) 10 m wind speed multiplied by the empirical correction factor, 1.4. If necessary, we manually correct the wind direction to match the observed plume direction, as also noted by Nassar et al (2017) and Reuter et al (2019).
To calculate the NO x -to-CO 2 emission ratio, we use the approach introduced in our earlier work (Hakkarainen et al 2021). We calculate the crosssectional NO x and CO 2 flux at distance, x, from the source, using both observed and simulated data and then calculate the NO x -to-CO 2 emission ratio: We note that if we assume that NO x and CO 2 have the same cross-sectional form (in the Gaussian case, the same FWHM) and the same effective wind speed, U eff , these terms cancel out and we are only left with the ratios of the amplitudes A of the plumes (equation (1)).

Results
Anthropogenic emission signatures on regional scales over the globe can be identified by calculating the XCO 2 anomalies from satellite observations (Hakkarainen et al 2016(Hakkarainen et al , 2019. Figure 1 shows the OCO-3 XCO 2 anomalies using the method described by Hakkarainen et al (2019) and previously applied to OCO-2 observations. The anomaly is calculated with respect to the daily medians for each 10-degree latitude band and linearly interpolated to each OCO-3 data point. The median better represents the typical value in each latitude band, and it is not skewed towards extreme values. The XCO 2 anomalies are averaged over the period between August 2019 and the end of January 2022. Positive anomalies are seen over most industrialized areas such as China, India, Europe, USA, Middle East and South Africa. They are also seen in regions of Africa and South America with frequent and intense forest fires. Overall, the features in the map are quite similar to those obtained with OCO-2 (Hakkarainen et al 2019) although some of the point sources are more highlighted with OCO-3. The anomaly map in figure 1 does not include OCO-3 SAM data, but observations obtained while observing the local nadir (operation mode 1) or in the direction of the apparent glint spot, where sunlight is specularly reflected from the Earth's surface (called 'glint mode' or operation mode 2). One of the anthropogenic emissions areas highlighted in the anomaly map is the South African Highveld region (blue rectangle in figure 1), where several power plants and industrial facilities are located. This area will be the subject of further study based on the analysis of OCO-3 SAM observations and TROPOMI NO 2 retrievals. Figure 2 shows the six OCO-3 SAMs collected between August 2019 and January 2022 over the Highveld area and analyzed in this study. The information about the SAMs included in the analysis are summarized in table 1. We estimate the emissions by six power stations (Kendal, Kriel, Matla, Majuba, Tutuka and Grootvlei) and those from the Secunda CTL synthetic fuel plant (table 2). XCO 2 enhancements originating from emission point sources in the area were also visible during different days, however, it was only possible to clearly identify the emission plumes in these six cases. To better identify the XCO 2 plumes, we use spatio-temporally colocated TRO-POMI NO 2 observations. One challenge related to the joint analysis of OCO-3 and TROPOMI is related to the overpass time, which can vary significantly from case to case for the ISS whereas the S5P overpass time is always around noon UTC (±30 min) over South Africa. This particular challenge will not be present with future simultaneous CO 2 and NO x observations (e.g. CO2M, GOSAT-GW).
To support the satellite data analysis, we simulate the plumes from selected point sources using the FLEXPART dispersion model. For CO 2 emissions (table 2), we use the values from the ODIAC fossil fuel emission dataset (1 km × 1 km resolution, version 2020b, Oda et al 2018) except for Secunda CTL. The CO 2 emissions for Kendal and Grootvlei power stations were not available in ODIAC, and we use instead emissions reported by Eskom, the company that operates the power stations. Secunda CTL (owned by Sasol) is considered the largest single emitter of GHG in the world with emission of 155 kt d −1 (Hallowes andMunnik 2017, Sguazzin 2020). For the Secunda CTL synthetic fuel plant, ODIAC reports the emission of 67 kt d −1 . To estimate the NO x emission, we simply scale the CO 2 emissions with a constant NO x -to-CO 2 emission ratio of 2 × 10 −3 , which is the same order of magnitude of the value estimated for Matimba power station (Hakkarainen et al 2021). Table 2 also includes the CO 2 emissions from the Emissions Database for Global Atmospheric Research (EDGAR) v6.0 dataset (Crippa et al 2021). EDGAR does not include the emissions for the Secunda CTL, and the emissions of Matla and Kriel are in the same 0.1 • × 0.1 • grid cell, but are otherwise in good agreement with the emissions used in the FLEXPART simulations.
Next, we analyze the OCO-3 SAMs case-by-case. All the figures for each day are provided in the supplementary material.
21 January 2022. Figure 3 shows the OCO-3 SAM and the TROPOMI NO 2 data collected on this day, which was characterized by strong winds (10 m wind speed about 6 m s −1 ). From the OCO-3 SAM, we observe two distinct CO 2 plumes, the southern plume originating from the Secunda CTL synthetic fuel plant and the northern plume from the Kriel and Matla power stations, with the Kendal power station downwind. The underlying TROPOMI NO 2 observations show that emissions plumes originating from the Tutuka and Majuba power stations also contribute to the southern plume. We also note a small enhancement originating from the Duvha power station.
We calculate the cross-sectional fluxes for both the NO 2 and CO 2 plumes along several sets of three transects (white lines in figure 3, left panel) perpendicular to the wind direction. Figure 3 illustrates an example of the calculation for one set of three transects, and all the other cases are given in the supplementary material. We fit two Gaussian functions to match the two plumes (figure 3, right panels). For NO 2 , it might be possible to fit more plumes, especially closer to the source, but for CO 2 data fitting, more than two Gaussians is impractical. For the southern plume, we obtain CO 2 emissions from 141 to 419 kt d −1 , with a mean of 265 kt d −1 and a standard deviation of 79 kt d −1 . The input values used for the Majuba, Tutuka and Secunda CTL simulations combined are 282 kt d −1 (table 2). The cross-sectional NO x emission estimates range from 194 t d −1 to 268 t d −1 , the values being generally lower downwind of the sources (as expected due to the photochemical decay of NO x ). We note that the plume width is generally larger for NO 2 than for CO 2 . If we adopt the NO 2 plume width for the CO 2 emission estimation (as done, e.g. We also observe an increase in the cross-sectional NO x emission estimates after Kendal power station. In general, however, the cross-sectional NO x emission estimates seem to increase rather than decrease as function of the distance from the sources, which would indicate that NO is not yet fully oxidized to NO 2 . This is a general feature that we observe throughout this study. The effect can also be observed from the NO 2 concentration map, i.e. the values are higher downwind than near the source. In this case, using the NO 2 plume width for the CO 2 emission estimation results in about 5% larger emission estimates. Figure 4 shows the FLEXPART model simulations for 21 January 2022. The emissions used as input in the simulations for each source are given in table 2. Figure 5 shows the cross-sectional NO x -to-CO 2 emission ratio at different transects across the southern plume. Transect number one corresponds to the location of the Secunda CTL. The transects are considered every 0.025 • along the direction of the plume. The blue line indicates the cross-sectional NO x -to-CO 2 emission ratio calculated from the FLEXPART model simulations at each transect using linear fit between the simulated NO 2 and CO 2 . The crosssectional NO x -to-CO 2 ratio decreases exponentially along the plume due to the decay of NO x (a lifetime of 4 h is assumed in the FLEXPART simulation). The assumed NO x -to-CO 2 emission ratio for all sources is 2 × 10 −3 , which is the same order of magnitude     Red circles indicate the NO x -to-CO 2 ratios calculated from TROPOMI and OCO-3 columns. We note that these ratios show a much higher variability compared to the FLEXPART simulations. The mean ratio between simulated and observed ratio is 0.9, which gives the average NO x -to-CO 2 emission ratio of 1.8 × 10 −3 with standard deviation of 0.6 × 10 −3 . In this case, using the NO 2 width for the CO 2 emission estimation would result in about 40% higher CO 2 emission and therefore in a lower NO x -to-CO 2 ratio, i.e. (1.3 ± 0.3) × 10 −3 . In Hakkarainen et al (2021), we obtained the NO x -to-CO 2 emission ratio of (2.6 ± 0.6) × 10 −3 for the Matimba power station in South Africa, which is somewhat higher than that for Secunda CTL, even though the approaches are not directly comparable. 4 July 2021. From the OCO-3 SAM, we observe a strong CO 2 plume originating from Secunda CTL and a fainter plume from Tutuka power station. We note the difference between the CO 2 and NO 2 plume direction and shape, especially in the second part of the Secunda CTL plume due to the two-hour time difference between OCO-3 and S5P overpasses. The change in the plume direction can also be seen in the FLEXPART simulations. For this reason, we do not use the NO 2 plume width for the CO 2 Gaussian fitting, but instead estimate the width directly from the CO 2 data. For the Secunda CTL plume, in the first three sets of transects, we obtain CO 2 emissions of 94, 99 and 102 kt d −1 , which are lower than the expected value of 155 kt d −1 . For the other two sets of transects, we obtain worse fitting results (the observed plume is twisted and less Gaussian) and lower CO 2 emission values of 71 kt d −1 . For the NO x emissions, we observe increasing cross-sectional emission from 44 t d −1 near the emission source and 165 t d −1 near the end of the OCO-3 swath. For the Tutuka power station plume, we obtain CO 2 emissions ranging from 29 to 8 kt d −1 with a mean of 16 kt d −1 and a standard deviation of 8 kt d −1 . These are relatively low compared to the expected value of 60 kt d −1 . Again, we note that the cross-sectional NO x emission estimates increase from 10 t d −1 to 30 t d −1 with increasing distance from the source.
14 March 2021. From the OCO-3 SAM, we observe a clear CO 2 plume signal from the Secunda CTL, even if a part of the observations is screened out by the quality flags of the retrievals. The plumes from Kriel and Matla power station are sampled poorly, but the enhancements are still visible. The area near the Kendal power station is completely flagged and the underlying NO 2 observations show only a faint plume. On this day, the NO 2 and CO 2 plumes are very similar, in the observations and in the FLEXPART simulations, although, the time difference between the OCO-3 SAM and TROPOMI is about 3 h. We however correct the plume direction by 15 degrees. Since a large part of the scene does not include good quality observations, we also test the data rejected by the quality flags (see the figures in the supplementary material for 14 March 2021). For Secunda CTL CO 2 emissions, we obtain quite variable results, with values ranging from 38 to 130 kt d −1 with a mean of 75 kt d −1 and standard deviation of 28 kt d −1 . Again, we note that the cross-sectional NO x emission estimate increases with the distance from the source. The effect can be also seen in the NO 2 columns as the NO 2 concentration increases from about 2.5 × 10 16 near the source to about 5 × 10 16 near the end of the area covered by the OCO-3 SAM.
2 August 2020. From the OCO-3 SAM, we observe a clear CO 2 plume generated from the Tutuka power station. Also, enhancements from the Majuba power station can be observed, but are not sampled very well. On this day, the time difference between OCO-3 SAM and TROPOMI is about 4 h, which makes the difference in NO 2 and CO 2 plume direction and shape very large. This is also visible in the FLEXPART simulations. The CO 2 emission estimates for Tutuka power station range from 10 to 33 kt d −1 with the mean 18 kt d −1 and the standard deviation of 8 kt d −1 . These are essentially the same values as obtained on 4 July 2021 and lower than the expected value of 60 kt d −1 .
25 July 2020. We observe several overlapping plumes generated from multiple point sources in both OCO-3 and TROPOMI observations. We analyze the plume from the Kendal power plant and correct the wind direction by 30 degrees for the emission estimation. We obtain CO 2 emissions values in the range 30-45 kt d −1 . These values are lower than the expected emission of 70 kt d −1 . We are also able to analyze the isolated plume generated from the Grootvlei power station that can be observed in a narrow part of the OCO-3 SAM. We fit the data with two set of transects and obtain CO 2 emissions of 32 and 22 kt d −1 . These are just slightly higher than the expected CO 2 emissions of 18 kt d −1 .
28 May 2020. We observe strong enhancements in both CO 2 and NO 2 datasets originating from Tutuka and Majuba power stations. For Tutuka power station, we observe nearly constant CO 2 emission between 33 and 37 kt d −1 . For Majuba power station, the CO 2 emissions range from 44 and 60 kt d −1 , which is slightly lower than the expected CO 2 emission of 67 kt d −1 . This case is characterized by strong wind speed (10 m wind speed of about 6 m s −1 ) and a clear difference between the CO 2 and NO 2 plume direction, even if the time difference between OCO-3 and TROPOMI overpass is only 2 h. Figure 6 summarizes the CO 2 emission estimates for all of the cases analyzed. For the case on 21 January 2022, we have combined the emissions from several point sources that contributed to the northern and southern plumes. Overall, we observe good agreement between expected and estimated emissions. On the other hand, our results generally slightly underestimate the ODIAC CO 2 emission estimates. We point out that ODIAC emissions for Secunda CTL were significantly lower (67 kt d −1 ) than the reported estimates (155 kt d −1 ) and also that these emissions are missing from the EDGAR v6.0 dataset. For the expected CO 2 emissions, we have included 30% error bar derived from the uncertainty estimate available for the high-resolution gridded inventory (Oda et al 2019). This value is quite conservative compared to typical uncertainties assumed for reported emissions since the emission estimates are derived from a global gridded inventory.
The vertical error bars in figure 6 express onesigma uncertainties calculated from individual sets of transects. This approach does not therefore consider the systematic errors. For example, the effective wind speed appears as a multiplicative factor in the cross-sectional flux emission estimate, so that an error in the effective wind speed, will directly transfer to the emission estimates. Systematic errors can also come from the observations themselves, or for example from systematically estimating the plume width incorrectly. The error bars in figure 6 thus represent the variation of the emission estimates within each SAM, i.e. the emissions derived from different sets of transects along the plume. The use of sample statistics is not possible with other satellite instruments, such as OCO-2, which essentially yields emissions along only one cross-section for each plume. The OCO-3 SAM observations quantify a larger portion of the plume and therefore improve the estimates of the emissions and their uncertainties.

Discussion
In this paper, we characterized the CO 2 emissions from several anthropogenic point sources in the South African Highveld region by analyzing six OCO-3 Snapshot Area Maps together with TRO-POMI NO 2 observations. We find that analyzing CO 2 and NO 2 data collected at different times can be challenging if the time difference is too large. In this study, the time difference between the OCO-3 and TROPOMI observations was up to 4 h, which can result in very different NO 2 and CO 2 plumes due to the changing wind patterns, and possibly also to temporally changing emissions. This makes the often-taken assumption that we can use NO 2 observations as a proxy for CO 2 observations invalid when combining OCO-3 and TROPOMI data. This largely affects the calculation of the NO x -to-CO 2 emission ratio, but can cause large biases also when using the NO 2 plume width for CO 2 cross-sectional flux calculations, as the width appears as a multiplicative factor. Other factors that cause the plume widths to be different are related to the different CO 2 and NO 2 background, detection limit and different pixel size of the instruments. In this study, we tested the NO 2 plume width only for the case on 21 January 2022, when the time difference between OCO-3 and TROPOMI overpasses is less than 1 h, while we fit the plume width directly from the CO 2 observations in all other cases. Luckily, this problem will be mostly solved with future missions such as GOSAT-GW and CO2M, which will provide simultaneous CO 2 and NO 2 observations. Despite these limitations, we find that even with temporally distant observations, analyzing OCO-3 together with TROPOMI NO 2 observations can still help in the identification of the emission sources.
We noted that in many cases the cross-sectional NO x flux estimates do not decrease with increasing distance from the source (as it was expected from the model simulations), which suggests that NO might be not yet fully oxidized into NO 2 or that some other chemistry-related issues need to be taken into account. This can substantially affect the calculation of the NO x -to-CO 2 ratio, especially near the emission source. Correcting for this effect may require a more complex modeling approach to be considered also for future observation systems including simultaneous CO 2 and NO 2 observations. The effect on the plume shape is expected to be less crucial.
We applied the cross-sectional flux method to obtain CO 2 emissions from OCO-3 SAMs along several transects along the plume and calculated the mean and the standard deviation. We demonstrated how the method can be extended to multiple and overlapping plumes within one scene. Overall, the emission estimates were comparable to the expected values from emission inventories (e.g. ODIAC), also for the cases where several plumes mixed. A similar approach can be applied to observations from future satellite missions such as CO2M, although some effort is needed to clearly identify and analyze overlapping plumes especially in a potentially automatized framework. Other methods for inferring point source emissions from column measurements (such as Gaussian plume inversion method, source pixel method and integrated mass enhancement method) can be also considered (Varon et al 2018). Concerning NO x , the divergence method (Beirle et al 2019) may provide a solution to infer emissions when there are several strong nearby point sources, but it requires averaging over long periods of time (e.g. one year). Recently, Hakkarainen et al (2022) proposed an extension of the divergence method to CO 2 based on synthetic CO 2 observations that could also be applied to future satellite observations over areas with multiple point sources close to each other.
One issue is related to the interpretation of the emission estimation obtained from individual plumes and their relation to annual emission estimates (Hill and Nassar 2019). With current measurement systems, we obtain at maximum a handful set of plumes for individual point sources per year to infer the emissions. The emissions can change during the day and during the season, which makes the evaluation of the annual emission challenging. The situation will be improved with upcoming measurements systems, and a simulation-based study (Kuhlmann et al 2021) indicates that with a threesatellite CO2M constellation, about 15 plumes can be observed in Central Europe (Jänschwalde and Mělník power stations) during one year. However, there will be no observations available during wintertime, which may introduce a seasonal bias in the estimation of the annual emissions. Also, no plumes can be observed during persistently cloudy conditions or during night-time. This clear-sky bias will remain an issue for any emission estimation method based on passive remote sensing observations, including those methods based on temporal averaging .
We noted that the cross-sectional flux may change along the plume. Calculating the mean and the standard deviation partly solves this problem, but may introduce some additional errors, as some of the values within a plume maybe closer to the true emission value than others. On the other hand, the approach based on sample statistics does not consider the systematic errors. In the cross-sectional flux method, different terms (e.g. effective wind speed, U eff , amplitude and width of the plume) are multiplicative, which means that small (absolute) errors in each term can cause large errors in the emissions. Moreover, for example, systematic errors in estimating the effective wind speed do not cancel out when calculating the mean and the standard deviation, but will lead to systematically biased estimates. Challenges related to the accuracy of wind speed information are common to all available methods and will possibly remain an issue in the future.
Related to the random errors of the emissions, we note that having multiple transects per plume (as is the case with OCO-3) makes the interpretation of the results significantly easier than with just one crosssection (OCO-2). This also allows sample-based error analysis that represents the variation of the emission estimates within each plume. Obtaining multiple transects per plume will also be possible with future satellite missions such as CO2M, with its expected swath width of 250 km. Considering the random error part of the averaged emissions over the year, the standard error of the mean goes down with the square root of the number of plumes. This indicates that, for example, if 16 plumes are observed per year, the random error goes down with the factor of 4.
Finally, we note that many of the recent studies, this paper included, that analyze CO 2 emissions from individual plumes are based on case studies (e.g. Nassar et al 2017, Reuter et al 2019, Hakkarainen et al 2021, although some part of the analysis, e.g. the plume detection could have been (semi-)automatized. In the future, when the data volumes are larger, algorithms that can be fully automated and operate systematically across the globe are needed. A step towards this approach was presented by Chevallier et al (2022), although they compare their estimates to gridded emission inventories and do not attribute the observed plumes to specific emission sources. In addition, Finch et al (2022) propose a machine learning algorithm for automated NO 2 plume detection. This approach will be useful with simultaneous CO 2 and NO 2 observations. A global assessment of methane ultra-emitters based on hundreds of TROPOMI images was provided by Lauvaux et al (2022), although extensive human labeling was still needed. Alternative data sources have also come available for GHG emission monitoring. For example in recent applications like Climate TRACE (https:// climatetrace.org) satellite imagery has been used with machine learning to estimate GHG emissions, however the underlying data and code are not available to the public (NASEM 2022). The methods based on data averaging and inverse modeling, are quite automated almost by definition, although still the final source attribution may need some human supervision.

Data availability statement
The OCO-3 data can be accessed from NASA's GES DISC https://disc.gsfc.nasa.gov. These data were produced by the OCO-3 project at the Jet Propulsion Laboratory, California Institute of Technology. The Copernicus Sentinel-5P/TROPOMI data used in this study can be accessed from the S5P-PAL system https://data-portal.s5p-pal.com/. 958927), and Academy of Finland (Grant Numbers 336798, 337552 and 331829). T O acknowledge financial support from the NASA Grant #80NSSC18K1313. Part of the work presented here was performed at the Jet Propulsion Laboratory, California Institute of Technology under contract to the National Aeronautics and Space Administration (NASA). Government sponsorship acknowledged. The Copernicus Sentinel-5P/TROPOMI data used in this work are provided as part of the Sentinel-5P Product Algorithm Laboratory (S5P-PAL).