The MOPITT instrument onboard the Terra Satellite provides the longest satellite-based record of atmospheric CO, with observations available from March 2000 to present. It provides nearly global coverage every 3 days (Edwards et al., 2004). We use the monthly Level 3 daytime column data from the Version 5 TIR product, which has negligible drift in the bias over time (Deeter et al., 2013). The Level 3 data are a gridded product and include the a priori and averaging kernel for each grid box. Supplemental Fig. S1 shows the MOPITT column averaging kernels averaged over four regions. The column averaging kernels depend on the observed scene, and vary year to year as well as seasonally. The dependence of the column averaging kernels on the CO mixing ratio profile (Deeter, 2009) explains the high values in the lower troposphere over eastern China in winter.

We calculate trends and deseasonalized anomalies for the eastern United States, Europe, and eastern China regions described by Worden et al. (2013). Trends that differ from zero by more than the 2σ uncertainty on the trend are considered statistically significant. We account for autocorrelation of the data for a 1-month lag when calculating the uncertainty on the trends. We calculate the annual cycle by fitting the data with a series of sines and cosines as well as the linear trend, and then remove the annual cycle to obtain the deseasonalized anomalies. Months with no MOPITT data or only a few days of MOPITT data are excluded from the trend analysis. This includes May–August 2001 and August–September 2009. We report the MOPITT trends for 2000–2010 for comparison with model simulations, and for 2000–2014 to give a longer-term view of the observed trends.

We use a suite of chemistry climate model (CCM) and chemical transport model (CTM) simulations to interpret the observed trends. The Global Modeling Initiative (GMI) CTM includes both tropospheric (Duncan et al., 2007) and stratospheric (Strahan et al., 2007) chemistry, including over 400 reactions and 124 chemical species. Meteorology for the GMI simulations comes from the Modern-Era Retrospective Analysis for Research and Applications (MERRA) (Rienecker et al., 2011). The GEOS-5 Chemistry Climate Model (GEOSCCM) (Oman et al., 2011) incorporates the GMI chemical mechanism into the GEOS-5 atmospheric general circulation model (AGCM). The GEOSCCM simulations are forced by observed sea surface temperatures (SSTs) from Reynolds et al. (2002).

The Community Earth System Model, CESM1 CAM4-chem, includes 191 chemical tracers and over 400 reactions for both troposphere and stratosphere (Tilmes et al., 2016). The model can be run fully coupled to a free-running ocean, with prescribed SSTs, or with nudged meteorology from GEOS-5 or MERRA analysis. CESM1 CAM4-chem is further coupled to the land model, providing biogenic emissions from the Model of Emissions and Aerosols from Nature (MEGAN), version 2.1 (Guenther et al., 2012).

Several simulations were conducted as part of the Chemistry-Climate Model Initiative (CCMI) project (Eyring et al., 2013). These include the Ref-C1 simulation of the GEOSCCM and a Ref-C1 CESM1 CAM4-Chem simulation, hereafter called G-Ref-C1 and C-Ref-C1, respectively, and the Ref-C1-SD simulation of the GMI CTM. Both the Ref-C1 and the Ref-C1-SD simulations use time-dependent anthropogenic and biomass burning emissions from the MACCity inventory (Granier et al., 2011), but the Ref-C1-SD simulations use specified meteorology while the Ref-C1 simulations run with prescribed SSTs. The MACCity inventory linearly interpolates the decadal anthropogenic emissions from the ACCMIP inventory (Lamarque et al., 2010) for 2000, and the RCP8.5 emissions for 2005 and 2010, to each year in between. The MACCity biomass burning emissions have year-to-year variability based on the GFED-v2 (van der Werf et al., 2006) inventory. From 2000 to 2010, CO emissions in the MACCity inventory decreased from 31 to 11 Tg yr-1 over the eastern United States, from 97 to 59 Tg yr-1 over Europe, and increased from 56 Tg to 72 Tg yr-1 over eastern China.

Given the uncertainty in CO emissions, we conduct a GMI CTM simulation using an alternative time-dependent emissions scenario, called AltEmis. This simulation is described in detail in Strode et al. (2015b). Briefly, anthropogenic emissions include time dependence based on EPA (https://www.epa.gov/air-emissions-inventories/air-pollutant-emissions-trends-data), the REAS inventory (Ohara et al., 2007), and EMEP (http://www.ceip.at/ms/ceip_home1/ceip_home/webdab_emepdatabase/reported_emissiondata/), and annual scalings from van Donkelaar et al. (2008). Biomass burning emissions are based on the GFED3 inventory (van der Werf et al., 2010). While the regional emission trends in this simulation are of the same sign as in the Ref-C1 case, the magnitude of the negative trends over the US and Europe are smaller and the positive trend over China is larger, leading to a positive global trend (Fig. 1). We also conduct a sensitivity study called EmFix with anthropogenic and biomass burning emissions held constant at year-2000 levels. Table 1 summarizes the simulations used in this study.

We regrid the model output to the MOPITT grid and convolve the simulated CO with the MOPITT averaging kernels and a priori in order to compare the simulated and observed CO columns. The averaging kernels are space- and time-dependent. We use the following equation from Deeter et al. (2013): Csim=C0+a(xmod-x0), where Csim and C0 are the simulated and a priori CO total columns, respectively, a is the total column averaging kernel, and xmod and x0 are the modeled and a priori CO profiles, respectively. The column averaging kernel is calculated from the standard averaging kernel matrix, which is based on the log of the CO concentration profile, following the method of Deeter (2009): aj=(K/log⁡10e)∑Δpivrtv,iAij, where Δpi and vrtv,i are the pressure thickness and retrieved CO concentration, respectively, of level i, A is the standard averaging kernel matrix, and K=2.12×1013 molec cm-2 hPa-1 ppb-1.

We deseasonalize the simulated CO columns and calculate their linear trend following the same procedure that we applied to the MOPITT CO. Months that do not have MOPITT data (June–July 2001 and August–September 2009) are excluded from the analysis of the model trends as well.

The Ref-C1 and Ref-C1-SD simulations requested by CCMI extend until 2010. However, the MACCity biomass burning emissions extend only until 2008. CAM4-Chem therefore repeated the biomass burning emissions for 2008 for years 2009–2010. In contrast, the GEOSCCM Ref-C1 and GMI Ref-C1-SD simulations used emissions from GFED3 (van der Werf et al., 2010) for years after 2008. Some simulations were available through 2011, while others ended in 2010. We therefore report results for 2000–2010, but note that extending the analysis through 2011 does not alter the conclusions.

The hindcast simulations driven by MACCity emissions (G-Ref-C1, Ref-C1-SD, and C-Ref-C1) show negative trends in CO over the US and Europe that agree with the observed slope from MOPITT within the uncertainty (Fig. 2, Table 2). The MOPITT trends for both regions are statistically significant for both regions, as shown by Worden et al. (2013). These results are consistent with the findings of Yin et al. (2015), whose inversion of MOPITT data showed a posteriori trends in CO emissions over the US and western Europe that were consistent with but slightly larger than the a priori trends. The EmFix hindcast shows a positive, though non-significant, trend for both regions, indicating that the decrease in CO emissions is necessary for reproducing the downward trend in the CO column. The AltEmis simulation fails to produce the negative trends, despite including negative trends in regional emissions for both the US and Europe. The impact of these negative regional trends is insufficient to overcome the positive global emission trend in the AltEmis scenario (Fig. 1), leading to positive trends in CO.

Figure 2 also reveals a negative bias in the simulated CO column between the models and MOPITT. A low bias in simulated CO at northern latitudes is often present in global models (Naik et al., 2013) and may indicate a high bias in northern hemispheric OH (Strode et al., 2015a) or CO dry deposition (Stein et al., 2014), as well as an underestimate of CO emissions.

The deseasonalized anomalies in the MOPITT and simulated CO columns are shown in Fig. 2b and d; the correlation coefficients between the observed and simulated monthly anomalies are presented in Table 2b. The highest correlations are for the AltEmis and Ref-C1-SD simulations of the GMI CTM. This result is consistent with the use of year-specific meteorology, which we expect to better match the transport of particular years. The lowest correlations are for the EmFix simulation. This is expected since the EmFix simulation does not include inter-annual variability (IAV) in biomass burning. The IAV in biomass burning makes a large contribution to the IAV of CO (Voulgarakis et al., 2015).

The role of biomass burning in driving the CO variability is even more evident at the hemispheric scale. Figure 2g and h show the anomalies in MOPITT and the simulations for the Northern Hemisphere (0–60∘ N). The EmFix simulation shows almost no correlation, while the other simulations have correlation coefficients exceeding 0.6 (Table 2). The role of changing anthropogenic emissions is also evident, as the Ref-C1-SD simulation captures the 2008–2009 dip in the CO column while the EmFix simulation does not. Gratz et al. (2015) found decreasing CO concentrations at Mount Bachelor Observatory in Oregon during spring for 2004–2013, which they attribute to reductions in emissions leading to a lower hemispheric background. We also note that Ref-C1-SD and G-Ref-C1 have similar correlations with the observed variability for the Northern Hemisphere (Table 2), indicating that transport differences are less important for variability at the hemispheric scale.

Observations from MOPITT show a negative trend in the CO column over eastern China for 2002–2012 (Worden et al., 2013). The negative trend for the years 2000–2014 exceeds that for 2000–2010 (Table 2), showing that it is not driven solely by temporary emission reductions in 2008. Our simulations do not reproduce this trend, and instead show increases in the CO column (Fig. 2e), which is expected given that CO emissions from China increase in four of the five simulations. The anomalies (Fig. 2f) show that the discrepancy in the simulated versus observed trends is driven largely by the failure of the simulations to capture the 2008 dip in the CO column, leading to an overestimate that continues through 2010. This suggests emission reductions in China during this time period are not adequately captured by the emission inventories. However, the good agreement between the observed and simulated decreases in CO for the Northern Hemisphere as a whole (Fig. 2g, h) suggest that on a global scale, the emission time series is reasonable. Consequently, we examine several other factors that may contribute to the difference in sign between the MOPITT and simulated CO trends.

Regional trends in CO are expected to vary with altitude, with surface concentrations most heavily influenced by local emissions. MOPITT TIR retrievals have higher sensitivity to CO in the mid-troposphere than at the surface (Deeter et al., 2004), so the trend in the MOPITT CO column will be weighted towards the trends in free tropospheric CO rather than near-surface CO. We quantify this impact on our Ref-C1-SD CO column trends by comparing the trend in the pure-model CO column with that of the simulated column convolved with the MOPITT averaging kernels.

The simulated CO trend over eastern China for 2000–2010 is positive (but not significant) both with and without the averaging kernels, but application of the MOPITT kernels increases the positive trend from 1.3×1016 to 1.4×1016 molec cm-2 yr-1. This result is initially surprising since we expect trends in the mid-troposphere to be more strongly influenced by the decrease in the hemispheric CO background. Indeed, the trends in CO concentration over eastern China simulated in Ref-C1-SD switch from positive in the lower troposphere to negative in the middle and upper troposphere. However, the application of the kernels results in more positive (or less negative) trends in all regions.

Yoon et al. (2013) show that since the averaging kernels vary over time, a bias between the true atmosphere and the a priori assumed by MOPITT can lead to an artificial trend in the retrieved CO. Similarly, the bias between the average simulated CO concentrations and the MOPITT a priori, evident in Fig. 2, can lead to an artifact in the simulated CO trend when the simulation is convolved with the MOPITT averaging kernels. This is due to the changing contribution of the a priori when the vertical sensitivity (averaging kernel) is varying in time. MOPITT vertical sensitivity varies with time due to instrument degradation as well as the change in CO abundance. The bias in CO varies with altitude, so if the vertical sensitivity described by the averaging kernel changes, this will change the value of the convolved CO column even if there were no changes in the CO profile. Furthermore, changes in the averaging kernel result in more or less weight placed on the a priori versus the CO simulated by the model. Thus, a difference between the a priori and the model means that placing more (or less) weight on the a priori will change the resulting value of Csim. Since the a priori profiles and columns are constant in time, taking the time derivative of Eq. (1) yields ∂Csim/∂t=a(∂xmod/∂t)+∂a/∂t(xmod-x0).

The second term on the right-hand side shows that the larger the bias between the modeled CO and the a priori, the larger the impact of the changing averaging kernel.

We quantify this effect by convolving the simulated CO for each year with the MOPITT averaging kernels for the year 2008, thus removing the effect of the time dependence of the averaging kernels. The resulting trend, 0.56×1016 molec cm-2 yr-1, is less positive than the pure model trend or the original simulated trend. Thus, accounting for the time dependence of the averaging kernels convolved with model bias reduces but does not eliminate the discrepancy with the observed trend. Comparing the trend for the constant averaging kernel case with the original simulated trend for Ref-C1-SD (1.4×1016 molec cm-2 yr-1) suggests that the changing averaging kernels combined with the model bias contribute 0.84×1016 molec cm-2 yr-1 to the simulated trend. Other regions also show a more negative trend when the same averaging kernel is applied to the model results for all years. The large bias in CO at middle and high northern latitudes commonly seen in modeling studies thus impacts the ability of models to reproduce and attribute observed trends in satellite data.

Figure 2 and Table 2 also show a positive trend in the GMI EmFix simulation for eastern China. This larger trend in the EmFix simulation than the Ref-C1-SD simulation indicates that the net decrease in emissions contributes to decreasing CO over eastern China, consistent with the observed negative trend, but other factors in the model cause an increase in CO over eastern China even when all emissions are constant. Subtracting the EmFix trend from the Ref-C1-SD trend shows that the changing emissions contribute a CO trend of -0.7 molec cm2 yr-1 over eastern China. The 2.1 molec cm2 yr-1 trend in the EmFix simulation, which reflects the impacts of the simulated chemistry and transport, thus contributes to the erroneous sign of the trend in the GMI simulations. The trends in the EmFix simulation for the northern hemispheric average and the eastern United States and Europe are positive as well (Table 2). We examine their cause in the next section.

Since the EmFix simulation shows a positive trend in the Northern Hemisphere, we next examine the variability in the CO sink, OH. We also examine variability in the total ozone column, since overhead ozone is a major driver of OH variability (Duncan and Logan, 2008). Figure 3 shows the variability in CO and OH in the EmFix simulation. The positive and negative anomalies in CO correspond with the negative and positive anomalies, respectively, in OH. The anomalies in OH are in turn inversely related to anomalies in the total ozone column. The correlation coefficient between OH and column ozone is -0.53 for the 15∘ S–15∘ N average, -0.72 for the 15–25∘ N average, and -0.75 for the 30–60∘ N average. The large NH ozone anomaly in 2010, in particular, leads to a large anomaly in OH and thus CO. This OH anomaly extends from the northern tropics to the midlatitudes. The large CO anomaly near the end of the time series contributes to the apparent 11-year trend. We note that since the lifetime of CO is several months, CO anomalies are not expected to have a one-to-one correspondence with the OH anomalies.

The large anomaly in the simulated total ozone column in 2010 is overestimated compared to observations. Figure 4 shows the time dependence of the total ozone column from 30 to 60∘ N in EmFix compared to SBUV data (Frith et al., 2014). While the observations show an anomaly in 2010, the magnitude is smaller than that produced by the simulation. Steinbrecht et al. (2011) attribute the 2010 anomaly in northern midlatitude ozone observations to a combination of an unusually strong negative Arctic Oscillation and North Atlantic Oscillation and the easterly phase of the quasi-biennial oscillation.

While the impact of OH interannual variability on the apparent trend in CO is clear in the EmFix simulation, this source of variability is partially masked by large interannual variability in CO emissions in the other simulations. We examine the correlation between the detrended and deseasonalized CO anomalies from 10∘ S–10∘ N in the Ref-C1-SD simulation and the CO emissions as well as the simulated OH and column ozone. Since the CO emitted in a given month can influence concentrations for several subsequent months, we use a 3-month smoothing of the emission time series. We find a high correlation (r=0.88) between the CO anomalies and the CO emissions. This correlation is also evident in the MOPITT data, as the MOPITT CO anomalies have a correlation of r=0.70 with the emissions. Figure 5 shows the strong relationship between the simulated CO anomalies and the CO emissions. However, the colors in Fig. 5 indicate that the scatter for a given level of emissions is often linked to the OH anomalies, with low/high OH anomalies leading to CO that is higher/lower than would be predicted just from the CO emissions. We find that the 10∘ S–10∘ N OH in the Ref-C1-SD simulation is anticorrelated with CO (r=-0.62) and with the total ozone column (r=-0.68). Consequently, the simulated ozone column plays a role in modulating tropical CO variability even when variable CO emissions are included, although the emissions still play the strongest role.