The evolution of atmospheric CO concentrations with time (∂[CO]/∂t) in a 3-D atmospheric model grid box is expressed as the sum of multiple CO emission sources (SourceCO) minus the CO sink (SinkCO), which can be represented by the following Eq. (1). 1∂CO∂t=∑SourceCO-SinkCO=-ν⋅∇[CO]+∑sectorECO+PCH4→CO+PNMVOCs→CO-kCO+OH(T)[CO][OH]-DepCO The flux divergence term (υ⋅∇[CO]) represents the transport of CO into and out of each atmospheric model grid box, whose sum is equal to zero at the whole globe. ECO is the surface CO emission flux from different source sectors (i.e., anthropogenic, biomass burning, biogenic, and oceanic). PCH4→CO and PNMVOCs→CO represent the CO chemical production from CH4 and non-methane volatile organic compounds (NMVOCs), respectively, oxidized by OH in the atmosphere. The CO chemical sink (kCO+OH(T)[CO][OH]) is calculated on the basis of CO ([CO]), OH ([OH]), and a temperature (T)-dependent rate (kCO+OH), and DepCO is the dry deposition of CO that contributes about 7 % of the CO total sink (Stein et al., 2014).

We use the global 3-D transport model of the Laboratoire de Météorologie Dynamique (LMDz) coupled with a simplified chemistry module, Simplified Atmospheric Chemistry assimilation System (SACS) (Pison et al., 2009), to simulate the atmospheric physical and chemical processes described in Eq. (1) except the dry deposition not represented by this model. An atmospheric Bayesian inversion framework is built upon the LMDz-SACS model (Chevallier et al., 2005, 2009), and satellite observations of the relevant trace gas species (CO, HCHO, and CH4) are assimilated to constrain the inversion system (Zheng et al., 2018a, b) given some prior information on the initial model state, surface emissions, CO chemical production, and OH field. Section 2.2 provides details of the atmospheric inversion approach and the model evaluation protocol, and Sect. 2.3 describes how we analyze the global CO budget using inversion results.

The core of atmospheric Bayesian inversion is the minimization of the following cost function: 2J(x)=(x-xb)TB-1(x-xb)+(H(x)-y)TR-1(H(x)-y). x is the control vector that gathers the target variables we seek to optimize, and xb is a prior guess of these variables assuming unbiased Gaussian error statistics represented by a covariance matrix B. y is the observation vector containing all the observation data assimilated to constrain the inverse problem; their error statistics are assumed to be unbiased and Gaussian with a covariance matrix R. H is the forward model (the combination of the LMDz-SACS model, a sampling operator, and an averaging kernel operator) that calculates the equivalent of the observation data in y based on the control vector x. The forward model error and the representation error caused by the mismatch between model and observation resolutions are also included in R, making R represent a combination of measurement, forward model, and representation errors. Configurations of the variables and vectors in Eq. (2) are summarized in Table 1 and Table S1 in the Supplement, most of which have already been described in our previous papers (Zheng et al., 2018a, b). To solve the inverse problem, forward and adjoint codes are iteratively run until sufficient convergence of the cost function (Eq. 2), and the last iteration with optimized model states gives us the best estimate that matches all available information within their uncertainties.

The inversion system has several updates compared to the version developed by the same research team a few years ago to study atmospheric CO trends (Yin et al., 2015). First, we use a higher-resolution transport model with finer horizontal grid cells (3.75∘×1.9∘ compared to 3.75∘×2.5∘) and more vertical layers (39 layers compared to 19 layers) than Yin et al. (2015) used. Second, we assimilate the MOPITT version 7 data as an observation constraint, which is improved with respect to retrieval biases and bias drift compared to the previously used version 6 data (Deeter et al., 2017). Third, we use the latest global bottom-up emission inventories as prior, including the Community Emissions Data System (CEDS, Hoesly et al., 2018) for the anthropogenic source and the Global Fire Emissions Database (GFED) 4.1s for the biomass burning source. The regional studies of Zheng et al. (2018a, b) used the same configuration than here, except that they assimilated in situ measurement for CH4 rather than satellite retrievals.

We perform four inversion simulations with our inversion system (Table 2). Inversion no. 1 (2000–2017) is constrained by CO total columns derived from the MOPITT version 7 TIR-NIR retrievals; Inversion no. 2 (2005–2017) is constrained by both the MOPITT CO column and the Ozone Monitoring Instrument (OMI) version 3 HCHO column; and Inversion no. 3 (2010–2017) is additionally constrained by Greenhouse gases Observing SATellite (GOSAT) column-averaged dry air mole fractions of CH4 (XCH4). All three inversions also assimilate in situ measurement of methyl chloroform (MCF) to help constrain OH (Yin et al., 2015), but the rapidly declining levels of MCF in the atmosphere make this constraint progressively ineffective (e.g., Liang et al., 2017). Factorial simulations constrained by MOPITT CO, OMI HCHO, and GOSAT XCH4 allow us not only to better constrain the photochemical reaction sequence of CO, but also to facilitate a quantitative assessment of potential uncertainties in the inversion CO budget relating to the use of different observation constraints. We also do a sensitivity Inversion no. 4 (2000–2017) with the same observation constraints as Inversion no. 1 but flat prior surface CO emissions without any inter-annual variability to check whether the derived CO budget is robust to the inter-annual variation of the prior CO fluxes.

Inversion results of the optimized CO concentration in the atmosphere are evaluated against independent ground-based observations from the World Data Centre for Greenhouse Gases (WDCGG, https://gaw.kishou.go.jp/, last access: 10 September 2019) and the Total Carbon Column Observing Network (TCCON, Wunch et al., 2011). The WDCGG provides measurements of surface CO concentrations through in situ and flask sample measurements, and the TCCON provides retrievals of the column-averaged dry air mole fraction of CO (XCO). We collect observation data from 110 sites in WDCGG (Table S2) and from 32 sites in TCCON (Table S3; station names and references are shown in Appendix Fig. A1), which cover the whole globe (Fig. A1). To do the evaluation, we first sample the model at the location and time of the observation data and then calculate the average values and annual trends for both model and observation. The annual trends are estimated on the basis of monthly time series using a curve fitting method (https://www.esrl.noaa.gov/gmd/ccgg/mbl/crvfit/crvfit.html, last access: 10 September 2019), which is also used in Zheng et al. (2018a). p values and 95 % confidence intervals are given to assess the robustness of the estimated trends. The metrics used for evaluation include normalized mean bias (NMB), root mean square error (RMSE), Pearson's correlation (R), and the regression slope between model and observation among all surface sites.

The picture of the atmospheric CO budget derived from our inversions includes surface fluxes (the sum of direct emissions from different source sectors and of dry deposition), CO chemical production, and CO chemical sink. Given the marginal role played by dry deposition (about 20 % of the direct emissions, Stein et al., 2014), the inferred surface fluxes will be assumed to be made of direct emissions only in the following. The CO chemical production and chemical sink are direct outputs from the inverse system, which calculates the CO yield from the oxidation of CH4 and of NMVOCs and the CO oxidation sink with the linearized chemistry scheme in each model grid box at each time step of the model simulation.

To obtain sectoral surface emission fluxes, we multiply the optimized 8-daily surface total fluxes by the proportion of each sector in each model grid cell as given by the prior (Jiang et al., 2017; Yin et al., 2016; Zheng et al., 2018b). We distinguish between four source sectors, anthropogenic, biomass burning, biogenic, and oceanic, which have rather distinct spatial–seasonal patterns in CO emission distributions. Without considering the oceanic emission, 96 % of the CO emissions on land are distributed on grid cells with a dominant emission source (i.e., a sector that contributes more than 50 % of CO flux in that grid). Further, 85 % of CO emissions on land are distributed in grid cells where such a dominant source accounts for more than 65 % of the CO total flux. The distinct seasonal evolutions of different sectors also help attribute emissions to one specific source sector. For example, the biomass burning in Africa typically accounts for 80 %–90 % of surface CO emissions in the dry season (Zheng et al., 2018b). This local source homogeneity reduces the attribution bias of sectoral CO fluxes, although it cannot eliminate all biases. To minimize remaining biases, we focus on annual emission anomalies by removing the multiannual average or calculating linear trends to reduce the systematic errors. This makes it possible to directly compare the inversion-based emissions with bottom-up inventories to analyze the underlying emission drivers (Sect. 4.2).

We also collect previous top-down inversion estimates of the global CO budget from the scientific literature (Table S4). These studies used older versions of MOPITT retrievals (or other satellite data) and coarser-resolution transport models, and they did not have the capability of multi-species constraints. Despite a different observation data quality and inversion model setup, it is still meaningful to do such a review to visualize the evolution of the inversion-based global atmospheric CO budget.

Tropospheric CO columns observed by MOPITT v7 declined at an average rate of -0.32±0.05 % yr-1 (p<0.01; numbers with ± signs are 95 % confidence limits from a linear fit) from 2000 to 2017 over the whole globe (Figs. 1a, S1). This trend, equivalent to -5.6×1015±0.9×1015 molec. cm-2 yr-1, is much larger than the retrieval bias drift of 1.0×1015±1.0×1015 molec. cm-2 yr-1 in the MOPITT v7 TIR-NIR product (Deeter et al., 2017), indicating that the observed downward trend is robust to satellite retrieval errors (Worden et al., 2013). The distributed ground-based sites of the WDCGG network (dots in Fig. 1a) also measure rapidly decreasing surface CO concentrations during 2000–2017, broadly consistent with the negative trends of the MOPITT CO columns. The sites located on or near continents, more affected by land sources upwind, generally show faster CO concentration declines than the MOPITT CO columns, while the background sites, especially those over islands, agree better with MOPITT observations. The TCCON-observed XCO also presents consistent declining trends with MOPITT (Fig. S2).

The trends in MOPITT CO columns reveal heterogeneous spatial patterns (Fig. 1b and c). The largest decrease is seen in the northern mid-latitudes (30–60∘ N), where Canada (CAN), the USA, Europe (EU), Russia (RUS), and China (CHN) all present statistically significant declining trends (Fig. 1c). The tropical region (30∘ S–30∘ N) has a smaller decrease in CO columns due to some increasing trends existing over South Asia (SAS) and over a large part of Africa. South Asia is the only region that has a statistically significant trend of rising CO columns since 2000 according to our region splitting (Fig. A1). The African continent presents both increasing and decreasing CO columns that compensate each other and therefore lead to no statistically significant trends over equatorial Africa (EQAF) and southern Africa (SAF) (Fig. 1c). This is consistent with ground-based observations at Ascension Island, UK (ASC, Table S2), located downwind of the West African coast that sees CO plumes from Africa. The southern mid-latitudes (30–60∘ S) primarily consist of the ocean with very few lands, where the MOPITT and WDCGG observations both show consistently moderately decreased CO abundance.

Modeled CO columns from Inversion nos. 1, 2, and 3 all match MOPITT observations within their assigned errors and also in terms of trends (Inversion no. 1 is shown in Fig. S6). The posterior simulation corrects the underestimates of prior modeled CO columns, especially over Europe, Africa, and South America, where the CO source is increased by inversion. For trends, the model with prior fluxes can only simulate slightly declining CO columns over the USA and EU (Fig. S6g), where anthropogenic CO emissions decrease in prior, but present increasing CO columns over all the other regions. The inverse system reverses the upward trend in prior CO sources that is inconsistent with atmospheric CO observations (e.g., in Asia), which consequently reproduces the global decline in CO total columns (Figs. S6c, S6e).

The independent ground-based observations from WDCGG and TCCON confirm an improvement of modeled CO and XCO in Inversion nos. 1, 2, and 3 with respect to both annual averages and trends (Figs. B1–B3, S7–S9). Compared with the WDCGG data, all three inversions correct the underestimates of the prior surface CO concentrations, which reduces the NMB and RMSE and increases the slope (closer to one) and R-squared of the linear regressions. For trends, most of the WDCGG sites, especially those located in the USA, EU, and CHN, present statistically significant downward trends between 2000 and 2017, while the prior results tend to underestimate the declining trends. These biases are reduced by the inversions, though uncertainties still exist in view of the scattered dots. The WDCGG sites that show large disagreements are mostly located in coastal terrain areas, where our coarse-resolution model simplifies the coastline and thus cannot resolve the associated meteorology well (e.g., land–sea breeze circulation) (Palau et al., 2005; Ahmadov et al., 2007) and possible local emission sources. Several sites at high northern latitudes also suggest relatively large modeling bias due to the lack of high-quality satellite data as an observational constraint. The modeled XCO with posterior fluxes agrees better than the prior results with the TCCON observations, especially for the rapidly declining trends between 2000 and 2017.

The evaluation with measurement from WDCGG suggests that Inversion no. 3 gives a fair estimate of surface CO trends during 2010–2017 (NMB =-8 %, RMSE =1.4 % yr-1, Fig. B3c), while Inversion no. 2 (NMB =-34 %, RMSE =2.0 % yr-1, Fig. B2c) and Inversion no. 1 (NMB =-47 %, RMSE =1.8 % yr-1, Fig. B1c) still present moderate biases in their study period. During the overlap period of 2010–2017 with Inversion no. 3, Inversion no. 2 and Inversion no. 1 both present a slightly larger RMSE of 1.5 % yr-1 in the trend estimates. Compared with TCCON observations, we also see a slight improvement of the modeled XCO trends in Inversion no. 3 (NMB =21 %, RMSE =0.3 % yr-1, Fig. B3d) and in Inversion no. 2 (NMB =10 %, RMSE =0.3 % yr-1, Fig. B2d) than those estimated by Inversion no. 1 (NMB =9 %, RMSE =0.4 % yr-1, Fig. B1d).

Inversion no. 4 has the same inversion setup as Inversion no. 1 except that it used the flat prior CO fluxes (seasonal climatology) without inter-annual variation. The prior CO fluxes used in Inversion no. 1 (Table S6) are composed of increasing anthropogenic emissions (0.33±0.14 % yr-1, p<0.01) and decreasing but highly time-variable biomass burning emissions (-0.43±1.43 % yr-1, p=0.53), which lead to a slightly increasing global total source (0.16±0.18 % yr-1, p=0.07). Without this prior inter-annual variability, Inversion no. 4 estimates that the global CO source declined by -0.32±0.23 % yr-1 (p<0.01) and the global CO sink declined by -0.32±0.20 % yr-1 (p<0.01) during 2000–2017 (Table 3). The declining CO source can be decomposed into a relatively flat CO chemical production (0.03±0.09 % yr-1) and a rapidly declining surface emission (-0.62±0.36 % yr-1) that is primarily due to the decreasing anthropogenic (-0.70±0.26 % yr-1) and biomass burning (-0.91±0.81 % yr-1) emissions. Overall, Inversion no. 4 is highly consistent with Inversion no. 1 in regard to the relative trends in the anthropogenic (Table S7) and biomass burning (Table S10) sources globally (Fig. 6) and regionally (Fig. 7), especially for the top five emitters (Figs. 8, 9). This consistency suggests that the estimated declining trends in the global CO source and CO sink are not affected by the inter-annual variations of prior CO fluxes.

We compare the inversion-based anthropogenic CO emissions to regional bottom-up inventories that are not used in our inversions (Fig. 8). Here we use the bottom-up inventory data from EPA for the USA (USEPA, 2018, Fig. 8a), from MEIC v1.3 for CHN (Zheng et al., 2018c, Fig. 8b), from TNO-MACC III for EU (Kuenen et al., 2014, Fig. 8c), from REAS v2.1 for SAS (Kurokawa et al., 2013, Fig. 8d), and from EDGAR v4.3.2 for EQAF (Crippa et al., 2018, Fig. 8e), while our prior inventory is the CEDS global inventory (see Sect. 2.2). These bottom-up emissions data are derived from detailed regional activity maps and sector-specific emission factors, except for EQAF, where we use the EDGAR v4.3.2 global inventory (up to 2012) due to the lack of regional inventories there. Figure 8 shows consistent anthropogenic CO emission trends between bottom-up and top-down estimates in all five regions, indicating that the inversion results capture the inter-annual variation of anthropogenic CO emissions well. The sectoral detail of bottom-up data allows us to identify the driving source sector in each region. The transport sector dominates the rapidly decreasing emissions in the USA and EU, while the industrial and residential sectors have driven down the emissions in CHN after 2005. The road transport, industrial, and residential sources all pushed up emissions in SAS during 2000–2010, while the growing residential source is mainly responsible for the continuous rising emissions in EQAF.

Distinct emission driver sectors reflect different stages of socio-economic development, fuel use, and emission regulation policies in different regions. Developed economies such as the USA and EU have improved their industrial and residential combustion facilities that now burn relatively clean energy at high combustion efficiencies. The transport sector accounts for the majority of the current emissions, and the progressive pollution control on vehicles has successfully cut CO emissions in the USA (Jiang et al., 2018) and EU (Crippa et al., 2016). CHN and SAS are both in the process of rapid industrialization and urbanization, which made their anthropogenic CO emissions rise up rapidly since 2000, driven by all the source sectors, especially the industrial and residential sources. To save energy and reduce air pollution, China has improved the combustion efficiency and strengthened the end-of-pipe pollution control since 2005, which has successfully cut industrial and residential CO emissions (Zheng et al., 2018a, c). China has also implemented stringent vehicle emission standards; however, the explosive growth in vehicle sales and oil use partly offsets the impact of vehicle pollution control, making the transport sector contribute less to emission reductions. The emissions from SAS are estimated to have increased up to 2010 and have remained flat since then, while the cause of flattening emissions is not very clear due to lack of bottom-up inventories for recent years. The underdeveloped economies in the EQAF have few emissions from the industrial and transport sources, and most of the emissions increase is driven by the growing residential source mainly for cooking and heating.

Biomass burning emissions are determined by burned area, fuel combustion rate per unit area, and emission factors per unit mass of fuel burned (van der Werf et al., 2017). Our inversion-based biomass burning CO emissions broadly follow the trends of GFED 4.1s burned area that is derived from satellite observations during 2000–2017 (Fig. 9), which suggests that burned area is the primary driver of biomass burning emissions variation. The global burned area is observed to have declined since 2000 (Fig. 9a) with the largest decline in the grassland and savanna ecosystems over EQAF (Fig. 9b). This declining trend is primarily driven by agricultural expansion and intensification, with more fire management and suppression due to population increase and socioeconomic development (Andela and van der Werf, 2014; Andela et al., 2017). Although extreme drought years could shortly expand the burned area, the overall downward trends in global burned area are robust due to the strong inverse relationship between fire activities and economic development (Andela et al., 2017). We have also noticed that the variations in burned area and biomass burning emissions did not have perfect matches in some years (e.g., the year of 2015). The mismatch can be explained by the inter-annual variation of fuel combustion rate and emission factors that mostly depend on burning conditions, land cover, and fuel type. For example, the 2015–2016 El Niño event caused severe fires on the drained peatlands in Indonesia that are not burned in normal years. This caused only a moderate increase in burned area but released a disproportionately large amount of CO because the peat fire emission factor is 210 g CO kg-1 dry matter (van der Werf et al., 2017), 3.3 times higher than that of savanna fires (63 g CO kg-1 dry matter) that regularly burn every year and contribute roughly half of the global biomass burning CO emissions.

The global CO budget is affected by the interactions between OH and CO. A lower OH level translates into a proportionally smaller CO sink and thus estimates a smaller CO total source to achieve the source–sink balance (Müller et al., 2018). As such, the inter-annual variability of OH (if any) perturbs the long-term trends of the global CO budget. There are similar discussions for CH4 that a combination of declining OH and slightly growing CH4 source may explain the renewed growth of CH4 concentrations since 2007 (Rigby et al., 2017; Turner et al., 2017); otherwise, an abrupt increase in CH4 emissions since 2007 has to be assumed to match the CH4 observations (Turner et al., 2017). On the other hand, the declining CO burden in the atmosphere can leave more OH available to oxidize CH4 through the coupling of CO–OH–CH4, which means that the declining CO can stimulate the growth of CO chemical production (Gaubert et al., 2017). These interactions between OH and CO help understand the uncertainties induced by OH in our inversion results.

The tropospheric OH mean derived from Inversion no. 1 is 9.9×105 molec. cm-3 and Inversion nos. 2 and 3 both give a mean OH of 10.0×105 molec. cm-3. These values are close to the TransCom prior with less than 1 % difference (the prior uncertainty is 5 %). This is a medium level compared to the modeled OH of 6.5×105 to 13.4×105 molec. cm-3 in the Atmospheric Chemistry and Climate Model Intercomparison Project (ACCMIP) simulations (Voulgarakis et al., 2013). The three inversions all estimate a small inter-annual variation in OH, less than 2 % (Fig. S10). Inversion no. 1 that constrains OH through CO and MCF gives a small positive trend in OH during 2000–2008 (0.21±0.14 % yr-1, P<0.01) and a small negative trend during 2008–2017 (-0.07±0.06 % yr-1, P=0.02). Inversion nos. 2 and 3 both estimate slightly increasing OH with the trends of 0.26±0.25 % yr-1 (P=0.05) and 0.19±0.21 % yr-1 (P=0.06), respectively, during their overlapping period 2010–2017. The estimated growing OH trends are larger than the downward trends estimated by Inversion no. 1. As Inversion nos. 2 and 3 additionally assimilate HCHO and CH4 that also react with OH, these results suggest that HCHO and CH4 tend to have a stronger constraint on the OH level than CO and MCF assimilated in Inversion no. 1. It should be noted that the mixing ratios of MCF in the atmosphere are approaching zero and therefore hardly constrain OH any more (Liang et al., 2017).

The debate on OH variation is still ongoing (Turner et al., 2019; Nisbet et al., 2019). OH was thought to be well buffered in the atmosphere, which means that OH is not sensitive to the variations of anthropogenic and natural emissions. This is reflected in the results of global chemistry and climate models (Naik et al., 2013; Voulgarakis et al., 2013). A recent 3-D inverse modeling by McNorton et al. (2018) also gave a small inter-annual variation of 1.8±0.4 % in the tropospheric OH from 2007 to 2015, broadly consistent with our inversion results (Fig. S10). In contrast, some two-box model inversions give larger inter-annual variations of about 5 %–8 % in the global OH mean (Rigby et al., 2017; Turner et al., 2017; Naus et al., 2019, Fig. S10). Rigby et al. (2017) and Turner et al. (2017) primarily used MCF to infer OH, and they both estimated declining OH from 2004/2005 to 2014/2013. This declining trend is not revealed by 3-D modeling studies, including our inversions here. Turner et al. (2017) also suggested slightly increasing OH since 2013, which agrees with our estimates of Inversion nos. 2 and 3. However, it should be noted that the uncertainty ranges assessed by those two-box model studies are larger than their estimated OH variations. Our 3-D inversions further suggest that the OH trend is sensitive to the observation constraints used. All these features suggest that the OH inter-annual variation is still underdetermined (Rigby et al., 2017; Turner et al., 2017).

The ability to simulate the nonlinear chemistry of OH is still weak in global models, which is another challenge to understand the OH variation. The LMDz-SACS model adopts a linearized chemical scheme to simulate the hydrocarbon reactions including CH4+OH, HCHO+OH, CO+OH, and MCF+OH. The nonlinear dynamics of the OH chemistry, such as the secondary OH production (Lelieveld et al., 2016) and the interaction of OH with the NOx chemistry (Miyazaki et al., 2017), is not represented. The negligible computational cost of this configuration for OH motivates it, but we also expect the optimization of OH through the joint assimilation of CH4, HCHO, CO, and MCF observations to counterbalance the simplicity of the scheme. Alternatively, it would be interesting to sophisticate the scheme by introducing key tracers in the OH chemistry (e.g., tropospheric ozone, NO, and NMVOCs) in the scheme together with prescribed (though uncertain) reaction rates, but we currently lack enough observations to constrain this additional complexity.

The OH trends derived from our inversions may be ambiguous, but we can still speculate that our main conclusion is robust to the possibly larger variation of OH. If a strong downward OH trend existed as two-box model studies suggested, we could see faster decreases in both global atmospheric CO sink and global atmospheric CO source than our current estimates. Although decreasing rates may be varied, the overall trends and drivers of the estimated global CO budget are not changed.

We compile the top-down estimated global CO budget from 12 papers (Table S4) to compare with our inversion results. Compared to the nine studies using different inversion systems (Fig. 10), our inversion system is the only one with the capability of multi-species constraints and also the only one using MOPITT v7 data to constrain CO, while the previous studies had to use older versions of MOPITT data.

All these top-down studies converge on the estimates of global CO sources (Fig. 10a) but disagree on the split between surface emissions and chemical production, and also on the proportions of CH4-based and NMVOC-based CO productions. For example, Jiang et al. (2017) estimated consistent declining trends in anthropogenic (Fig. 10d) and biomass burning (Fig. 10e) CO emissions as our estimates, but their annual average emissions are 20 %–37 % lower than our results. The CO production from NMVOCs shows a large spread (Fig. 10h) among different inversion studies. Gaubert et al. (2017) estimated a CO chemical production quite close to our estimates (Fig. 10f), while they gave lower CH4 oxidation and higher NMVOC oxidation. The biogenic and oceanic CO emissions derived from different inversions are 100–200 and 20 Tg CO yr-1 (Table S4), respectively, which is consistent with our estimates (Table 3). Few studies estimated the global CO sink (Fig. 10b), except Gaubert et al. (2017), who gave a declining CO sink from 2002 to 2013 that agrees with our estimates, but the Gaubert et al. (2017) values are 15 % lower on average.

The other three inversions in the comparison are all derived from previous versions of our inversion system with multi-species constraints (Fortems-Cheiney et al., 2011, 2012; Yin et al., 2015). Compared to these previous estimates (Fig. S11), we have increased the spatial resolution of our transport model, used improved prior data, and assimilated the new MOPITT v7 observations. Therefore, the updated inversion results in this paper are expected to have better quality, leading to updated trend estimates of the global CO source (Fig. S11a), CO sink (Fig. S11b), surface CO emissions (Fig. S11c), and CO production from NMVOCs (Fig. S11h).