We utilize the adjoint of the GEOS-Chem model (version 35n) with extended support for MERRA-2 meteorological data and HEMCO emission inventories. The analysis is conducted at a horizontal resolution of 2°×2.5° with 47 vertical levels (MERRA-2) up to 0.01 hPa and employs a CO-only simulation (tagged-CO mode), in which the chemical sink of CO is linearized with archived monthly mean OH fields. Two types of archived OH fields are used in this study: fixed monthly OH fields for 2013 from the GEOS-Chem full chemistry simulation (Fisher et al., 2017), and variable monthly OH fields for 2005–2020 from the Tropospheric Chemistry Reanalysis version 2 (TCR-2, Miyazaki et al., 2020). The TCR-2 OH fields have been validated against various aircraft observations and show generally good agreement (Miyazaki et al., 2020). Figure S2 shows global mean tropospheric OH concentrations from TCR-2, demonstrating a slight increasing trend (1.0 ± 0.6×103 molec.cm-3yr-1) in 2005–2020.

The global default anthropogenic emission inventory is the CEDS-CMIP6 (Community Emissions Data System) (Hoesly et al., 2018). Regional emissions are replaced as follows: MIX (Li et al., 2017) over Asia, NEI 2016 (National Emissions Inventory) over the United States, DICE_AFRICA and EDGARv4.3 over Africa, and APEI over Canada. The contribution of co-emitted anthropogenic VOC sources is considered by scaling up anthropogenic CO emissions by 11 %. Biogenic sources are simulated using the Model of Emissions of Gases and Aerosols from Nature, version 2.0 (MEGANv2.0, Guenther et al., 2006). CH4 oxidation source is considered by using a prescribed, spatially varying CH4 field following the default tagged-CO configuration (Fisher et al., 2017). The CO sources from both biogenic VOCs and CH4 oxidation are calculated online based on the assumption of instantaneous oxidation by OH radicals. Biomass burning emissions are based on the Global Fire Emissions Database version 4 (GFED4, van der Werf et al., 2010). For years beyond the end year of a specific inventory, emissions from the last available year within that inventory's coverage were used to fill the subsequent years. The distribution of the annual mean CO emissions from 2003 to 2022 is shown in Fig. 1a–c.

The annual global sources are 536.3 Tgyr-1 from anthropogenic emissions, 312.5 Tgyr-1 from biomass burning, 623.2 Tgyr-1 from the oxidation of biogenic VOCs, 922.5 Tgyr-1 from the oxidation of CH4, with a total sink (through the reaction with OH radicals) of approximately 2395.0 Tgyr-1 in a priori inventories in this work. For comparison, Zheng et al. (2019) reported inversion-based global CO budget estimates for 2005–2017 of approximately 700 Tgyr-1 for anthropogenic emissions, 500 Tgyr-1 for biomass burning, 300 Tgyr-1 for biogenic VOC oxidation, and 900 Tgyr-1 for CH4 oxidation, with a total sink of approximately 2600 Tgyr-1. Regarding anthropogenic emissions, the CEDS-CMIP6 inventory estimates an average of 607.6 Tgyr-1 for 2003–2014, while the updated CEDS-CMIP7 inventory yields lower values, averaging 480.2 Tgyr-1 over 2003–2023. For biomass burning, the GFED5 inventory estimates an average of 518.3 Tgyr-1 for 2003–2022.

The objective of the 4D-Var approach is to minimize the difference between simulations and observations by minimizing the cost function (Henze et al., 2007): 1J(x)=∑i=1N(Fi(x)-zi)TSΣ-1(Fi(x)-zi)+γ(x-xa)TSa-1(x-xa) where x is the state vector of CO emissions, N is the number of observations distributed in time over the assimilation period, zi are the MOPITT CO retrievals, and F(x) is the forward model. Error estimates are assumed to be Gaussian: SΣ is the observational error covariance, which combines a 10 % uniform error and the MOPITT CO retrieval error covariance; and Sa is a priori error covariance. Here the combustion-related CO sources (fossil fuel, biofuel, and biomass burning) and the oxidation source from biogenic VOCs are combined, with a uniform a priori error of 50 % assumed following previous studies (Jiang et al., 2013, 2017). The CO source from CH4 oxidation is optimized separately as an aggregated global source, with a priori uncertainty of 25 %. The cost function is minimized by iteratively adjusting the CO emissions using the quasi-Newton gradient-based optimization L-BFGS-B algorithm (Zhu et al., 1997) and the adjoint gradients: 2∇xJ(x)=∑k=1N2∂Fi∂xTSΣ-1(Fi(x)-zi)+2γ(x-xa)TSa-1. The LOGX2 method (Jiang et al., 2015a, 2017) is employed to improve the reduction of negative gradients.

Following Jiang et al. (2017), we applied a two-step approach to mitigate the influence of systematic biases in the model simulations. First, a sequential Kalman filter (Todling and Cohn, 1994; Tang et al., 2022) was used to assimilate MOPITT CO retrievals from 1 October 2002 to 31 December 2022, providing optimized CO concentration fields with lower bias. As illustrated in Fig. 2a, the GEOS-Chem model driven by the original monthly CO initial conditions and a priori emission inventories (referred to as GC-original) substantially underestimated column CO concentrations by approximately 30 %–40 % (mean bias=-39.4×1016 molec.cm-2; Table 1). In contrast, simulations using the monthly CO initial conditions derived from the sequential Kalman filter together with a priori emissions (GC-a priori) showed markedly improved agreement with MOPITT CO retrievals (Fig. 2b), reducing the mean bias to about 10 % (mean bias=-9.7×1016 molec.cm-2). Similarly, the use of optimized monthly CO initial conditions led to considerable improvement in model performance against independent surface and aircraft measurements (Table 1). The mean bias decreased from -18.3 ppb (GC-original) to -1.4 ppb (GC-a priori) for World Data Centre for Greenhouse Gases (WDCGG) surface observations; from -18.9 to -3.8 ppb for HIAPER Pole-to-Pole Observations (HIPPO) aircraft data; and from -16.2 to -3.4 ppb for Atmospheric Tomography Mission (ATom) aircraft measurements. These results suggest that the substantial negative biases seen in Fig. 2a largely originate from the accumulation of biases over preceding months.

Furthermore, ocean scenes (pink grids in Fig. S3) were defined as land boundary conditions. The optimized CO fields from the Kalman filter were used to update CO concentrations over the ocean at hourly intervals during the forward simulation within the 4D-Var process. Meanwhile, the 4D-Var system constrained CO emissions over land without modifying oceanic CO distributions. As demonstrated by Jiang et al. (2017), the use of optimized CO land boundary conditions in 4D-Var assimilation effectively reduces systematic biases associated with long-range transport. By adopting this two-step assimilation framework, the inversion focuses on optimizing fresh continental CO emissions, while reducing the influence of uncertainties arising from transport and chemical processes, which tend to exhibit larger systematic biases. Consequently, a posteriori CO emissions estimated in this study are expected to be lower than those derived without adjustments to the initial and boundary CO conditions. This reflects both the specific inverse modeling setup and a possible underestimation in our a posteriori emission estimates, attributable to the emphasis on constraining fresh continental CO sources.

Based on this assimilation framework, three sets of CO emission inversion experiments are designed:

Column-FixOH: uses MOPITT CO column concentration data with default OH fields fixed in 2013.

By comparing the results of Column-FixOH and Profile-FixOH, the influence of different MOPITT CO data types on CO source estimates can be assessed. Similarly, comparing Column-FixOH and Column-VarOH allows for evaluation of the impact of different OH fields on CO source estimates.

The MOPITT instrument was launched on 18 December 1999, aboard the NASA Terra spacecraft. The satellite follows a sun-synchronous polar orbit at 705 km altitude, crossing the equator at 10:30 LT. The instrument made measurements over a 612 km cross-track scan, with a footprint of 22km×22km. The MOPITT data used in this study are from the joint retrieval (version 9J) of CO, which combines thermal infrared (TIR, 4.7 µm) and near-infrared (NIR, 2.3 µm) radiances using an optimal estimation approach (Worden et al., 2010; Deeter et al., 2022). The retrieved volume mixing ratios are reported as layer averages across 10 pressure levels (surface, 900, 800, 700, 600, 500, 400, 300, 200, and 100 hPa). The relationship between the retrieved CO profile and the true atmospheric state is expressed as: 3z^=za+A(z-za)+Gϵ where za is the MOPITT a priori CO profile, z is the true atmospheric state, Gε represents the retrieval error, and A=∂z^/∂z is the MOPITT averaging kernel matrix, indicating the sensitivity of the retrieval to the actual atmospheric CO. We only consider data with Cloud Description=2 (cloud free) and exclude MOPITT data with CO column amounts less than 5×1017 molec.cm-2. The threshold (5×1017 molec.cm-2) was selected to prevent the influence of certain potentially inaccurate, extremely low-concentration observations, which may also have low observation errors in the cost function, on the 4D-Var assimilation (Jiang et al., 2013, 2017). Since the NIR channel relies on reflected solar radiation, only daytime data are considered (Worden et al., 2010; Tang et al., 2024).

The HIPPO (Wofsy and HIPPO Science Team, 2011) were conducted using the Gulfstream V aircraft from 2009 to 2011. The flights primarily covered the Pacific Ocean, spanning latitudes from 67° S to 87° N, with continuous sampling from 0.2 to 12 km altitude. The ATom (Wofsy and Atom Science Team, 2018) used the DC-8 aircraft from 2016 to 2018. ATom covered similar altitude and latitudinal ranges as HIPPO but with broader spatial coverage, particularly over the Atlantic Ocean. For HIPPO, a total of 687 CO profiles from five missions were used directly. For ATom, CO measurements during continuous ascents and descents were used to construct 523 CO profiles from four missions. Surface CO measurements from the WDCGG are also included in this analysis. The WDCGG, operated by the Japan Meteorological Agency under the World Meteorological Organization's Global Atmosphere Watch (GAW) program, collects, archives, and distributes atmospheric greenhouse gas data, including CO, contributed by various institutions worldwide.

Before presenting the estimated emission trends, we first evaluate the performance of our assimilation system. The evaluation involves comparing modeled CO concentrations from the GC-original, GC-a priori, and a posteriori simulations (Column-FixOH, Profile-FixOH, Column-VarOH) over the period 2003–2022 against MOPITT satellite measurements, as well as independent surface observations from WDCGG and aircraft measurements from HIPPO and ATom. As summarized in Table 1, a posteriori simulations exhibit mean biases relative to MOPITT retrievals ranging from -5.1 to -7.3×1016 molec.cm-2. These values are notably smaller than the biases in the GC-a priori simulation (-9.7×1016 molec.cm-2) and the GC-original simulation (-39.4×1016 molec.cm-2). Similarly, for the HIPPO aircraft observations, a posteriori simulations show mean biases between -2.5 and -2.1 ppb, improved compared to the GC-a priori (-3.8 ppb) and GC-original (-18.9 ppb) simulations. For ATom aircraft data, a posteriori mean biases range from -2.9 to -1.6 ppb, also lower than those from the GC-a priori (-3.4 ppb) and GC-original (-16.2 ppb) simulations. In the case of surface CO concentrations, a posteriori simulations yield mean biases between 0.3 and 1.9 ppb relative to WDCGG observations (Table 1), which are reduced compared to GC-original (-18.3 ppb) simulations, and comparable with the GC-a priori (-1.4 ppb) simulations. A posteriori simulations slightly overestimate surface concentrations relative to WDCGG data, while underestimating CO in the free troposphere according to MOPITT and aircraft measurements. This systematic discrepancy may be attributable to uncertainties in convective transport parameterizations within the model.

Overall, the good consistency between a posteriori simulations and multiple independent observation platforms demonstrates the capability of our assimilation system to effectively constrain CO emissions. Given this confidence in the system's performance, we now present the central findings of this study: the long-term evolution of CO emissions. As mentioned in Sect. 2.1, the combustion-related CO sources and the oxidation source from biogenic VOCs are combined, and thus, the inverse system optimizes total CO emissions within each model grid cell. The subsequent attribution of emissions to specific source types (e.g., anthropogenic, biomass burning) in an individual grid cell is based on the relative contribution of each source category from a priori emission inventories. Specifically, a posteriori emission for a given source type in a grid cell is calculated by applying the grid-scale scaling factor (the ratio of a posteriori to a priori total emissions) to the corresponding a priori emission of that source type. Different sources can finally be calculated because each source category possesses distinct spatial patterns and seasonal variations.

Building on the emission estimates evaluated above, this section investigates their ultimate influence in the atmosphere by analyzing the spatiotemporal patterns and trends of CO concentrations. We first present the mean state and long-term changes in CO concentrations, and then quantitatively attribute these changes to their underlying drivers: emissions and meteorology. Figure 8a–c show the mean surface CO concentrations in 2003–2022 from a posteriori simulations and WDCGG surface observations. Higher CO concentrations are evident in regions with strong anthropogenic emissions, such as East Asia, India, and Southeast Asia, as well as in areas with significant biomass burning, i.e., Central Africa and South America. The long-term trends in surface CO (Fig. 8d–f) reveal declining concentrations over North America, Europe, East Asia, and South America, which contrast with rising trends over India, Boreal Asia, Central Africa, and Australia. The 20 year mean CO columns (Fig. 9a–c) show a consistent spatial pattern, with the highest column concentrations over East Asia and Central Africa, followed by South America, India, and Southeast Asia. In contrast, the long-term trend of CO columns (Fig. 9d–f) exhibits a more uniform decrease across the Northern Hemisphere, lacking the distinct regional hotspots observed in the surface trends. This suggests that changes in CO are more thoroughly mixed within the column.

To quantitatively attribute the concentration trends to specific drivers, we conducted a series of sensitivity experiments. The experimental design isolates the influence of individual emission sectors by building a baseline scenario in which all emissions are fixed at 2003 levels to reflect the impact of meteorological condition changes in 2003–2022. Three more sensitivity experiments were then conducted in 2003–2022 in which only one emission category, i.e., anthropogenic, biomass burning, or biogenic VOC sources, was allowed to vary over time, respectively. The time-varying sources in these sensitivity experiments were prescribed from the Column-FixOH a posteriori inversion.

The results indicate that meteorological influences induced positive trends in surface CO concentrations in regions such as central Africa, Southeast Asia, and the Tibetan Plateau (0.6 %–1.8 % yr-1), along with slight negative trends in areas such as South America (Fig. 10a). The meteorological impact on CO column concentrations was comparatively weaker (Fig. 10b), showing positive trends of 0.45 % yr-1 over central Africa and the Tibetan Plateau. This vertical differentiation implies that meteorological influences may primarily alter the vertical distribution of CO through changes in convective transport, with a more limited effect on larger horizontal scales. The derived meteorological impact is noticeably weaker than that reported by Jiang et al. (2017), a discrepancy likely attributable to our use of consistent MERRA-2 meteorological fields, which enhances the reliability of the long-term trend analysis. Similarly, the impact of biogenic VOC changes on CO concentrations (Fig. 10g and h) was markedly weaker than in Jiang et al. (2017).

Anthropogenic emission changes were identified as the principal driver behind declining CO levels, inducing strong negative trends in industrial regions of the Northern Hemisphere, such as eastern North America, Europe, and eastern China. This signal is consistent across both surface and column concentrations (Fig. 10c and d). Globally, anthropogenic emission changes led to an average annual decrease of 0.27 % yr-1 in CO column concentrations, with a more pronounced decline rate of 0.51 % yr-1 in the Northern Hemisphere (Table 4). Regionally, the US, Europe, and eastern China exhibited the most substantial decreases, at -0.57 % yr-1, -0.69 % yr-1 and -0.69 % yr-1, respectively. In contrast, India experienced a slight concentration increase (0.03 % yr-1) due to rising emissions, while Southeast Asia showed a more moderate decline (-0.19 % yr-1) compared to other major industrial regions.

Conversely, changes in biomass burning emissions generally contributed to positive trends in CO, particularly at high latitudes (Fig. 10e and f). At global and Northern Hemispheric scales, this positive trend was largely attributable to extreme wildfire activity in 2021. When 2021 is excluded, the long-term trend in CO columns due to biomass burning becomes statistically insignificant at these broad scales (Fig. 7g and h), with only regionally and seasonally confined increases remaining apparent, notably over Boreal Asia (July–August, Fig. 7b) and Africa (January–April, Fig. 7d). In the full record (including 2021), biomass burning emissions led to an average annual increase of 0.10 % yr-1 in global CO columns, and 0.24 % yr-1 in the Northern Hemisphere (Table 4). It is noteworthy that the CO concentration response lagged behind emission pulses by about one month and persisted longer. In the Northern Hemisphere, for instance, enhanced emissions occurred mainly from July to September, whereas the significant concentration response extended from August to December (Fig. 7g). This lag and prolonged influence were primarily attributable to the delayed response over Boreal North America (Fig. 7a). At the regional scale, increases occurred in Boreal North America (0.43 % yr-1) and Boreal Asia (0.48 % yr-1). In contrast, South America, Australia, and Southeast Asia experienced declining trends ranging from -0.13 % yr-1 to -0.22 % yr-1, while Africa showed a slight increase of 0.09 % yr-1.

This attribution analysis highlights the substantial impact of extreme wildfire years on the CO budget. Although anthropogenic emission reductions lowered Northern Hemisphere CO columns by approximately 0.51 % yr-1, the intense biomass burning emissions in 2021 introduced a positive perturbation of about 0.24 % yr-1 in the full-record trend, thereby offsetting a considerable fraction of the anthropogenic-driven decline. As a result, the net concentration decline was reduced to approximately 0.27 % yr-1 in the analysis including 2021. This implies that nearly half (47 %) of the potential air quality improvement from anthropogenic emission controls can be offset by wildfire emissions. This finding provides a clear mechanistic explanation for the decline in atmospheric CO concentrations in recent years, and underscores the growing role of extreme wildfire events in modulating regional to hemispheric air composition.

The MOPITT instrument provides retrievals for both CO total column and vertical profile. The degrees of freedom for signal (DFS) for MOPITT multi-spectral profile retrievals (TIR + NIR) is approximately 1.5–2.0 over land, reducing to about 1.0 when converted to a total column (Worden et al., 2010). The discrepancy between a posteriori emission estimates constrained by CO column (Column-FixOH) and profile (Profile-FixOH) data helps evaluate the influence of systematic errors associated with the vertical sensitivity of the satellite retrievals (Tang et al., 2024). Globally, a posteriori anthropogenic and biomass burning CO emissions from Profile-FixOH were both slightly lower than those from Column-FixOH, with average differences of -6.6 % and -5.5 %, respectively, over the period 2003–2022 (Table 2). The two configurations also showed broadly consistent long-term trends in inferred anthropogenic emissions, both indicating a global decline of approximately -0.9 % yr-1. Larger regional discrepancies were observed over eastern China (-2.1 % yr-1 for Profile-FixOH vs. -1.6 % yr-1 for Column-FixOH) and India (1.1 % yr-1 vs. 0.7 % yr-1). Similarly, the trends in global biomass burning CO emissions were consistent (0.3 % yr-1 for Column-FixOH and 0.5 % yr-1 for Profile-FixOH), though regional differences were more pronounced for boreal North America (3.1 % yr-1 vs. 4.9 % yr-1) and Australia (-1.5 % yr-1 vs. -0.7 % yr-1). The limited differences in inferred emissions between the two configurations resulted in a consistent declining trend in simulated CO columns (-0.5 % yr-1 for both).

OH concentrations in model simulations significantly influence the inverse analysis of CO emissions (Jiang et al., 2011; Müller et al., 2018). By assimilating MOPITT CO column data, we compared the inverted CO emission estimates driven by fixed (Column-FixOH) and variable (Column-VarOH) OH fields to investigate the potential influence. As shown in Fig. 11c, OH concentrations from the TCR-2 reanalysis are broadly 10 %–40 % lower than the fixed climatological OH concentrations over land (differences over the ocean are not considered here due to the use of CO land boundary conditions in the 4D-Var assimilation). Lower OH concentrations over land lead to reduced chemical loss, which is compensated by lower global anthropogenic CO emissions in Column-VarOH inversion (590.1 Tgyr-1) in 2003–2022, approximately 3.7 % lower than in Column-FixOH (612.8 Tgyr-1).

Variations in OH concentrations influence the oxidation of biogenic VOCs to CO and their subsequent chemical loss. These two counteracting processes establish a complex balance, ultimately reflected in the inverted estimates of biogenic CO sources. Specifically, the Column-VarOH inversion yields an average global biogenic CO sources of 391.4 Tgyr-1 in 2003–2022, approximately 3.9 % lower than the 407.6 Tgyr-1 in Column-FixOH inversion. The sensitivity experiments described above address the third objective of this study, which is to evaluate the robustness of our central findings against potential systematic errors associated with satellite retrieval vertical sensitivity and OH concentrations. This robustness can be attributed, in part, to our two-step inversion framework, which mitigates systematic biases through optimized initial and boundary CO conditions.