Using CO2:CO correlations to improve inverse analyses of carbon fluxes



[1] Observed correlations between atmospheric concentrations of CO2 and CO represent potentially powerful information for improving CO2 surface flux estimates through coupled CO2-CO inverse analyses. We explore the value of these correlations in improving estimates of regional CO2 fluxes in east Asia by using aircraft observations of CO2 and CO from the TRACE-P campaign over the NW Pacific in March 2001. Our inverse model uses regional CO2 and CO surface fluxes as the state vector, separating biospheric and combustion contributions to CO2. CO2-CO error correlation coefficients are included in the inversion as off-diagonal entries in the a priori and observation error covariance matrices. We derive error correlations in a priori combustion source estimates of CO2 and CO by propagating error estimates of fuel consumption rates and emission factors. However, we find that these correlations are weak because CO source uncertainties are mostly determined by emission factors. Observed correlations between atmospheric CO2 and CO concentrations imply corresponding error correlations in the chemical transport model used as the forward model for the inversion. These error correlations in excess of 0.7, as derived from the TRACE-P data, enable a coupled CO2-CO inversion to achieve significant improvement over a CO2-only inversion for quantifying regional fluxes of CO2.

1. Introduction

[2] Measurements of atmospheric composition provide powerful constraints to improve understanding of surface fluxes of trace compounds and their subsequent fate in the atmosphere. Recent studies have exploited these data by employing an inverse model approach [Bousquet et al., 1999; Gilliland et al., 2003; Palmer et al., 2003], in which a chemical transport model (forward model) is used to relate the sensitivity of an observed chemical concentration measurement to changes in particular model parameters (e.g., surface fluxes). An optimal estimation inverse model then combines this sensitivity information with the observations to yield the solution that is most consistent with the observations, the a priori information about the model parameters, and their relative errors. Estimated error variance and covariance information is included in the inversion via the a priori and observation error covariance matrices. Previous studies that analyzed correlations between different trace gases [e.g., Ciais et al., 1995; Enting et al., 1995] did not take advantage of these additional correlations in inverse model analyses. These correlations can arise from having similar spatial and temporal flux distributions, from chemical mass balance, or from atmospheric transport processes. Exploitation of these correlations in an inverse model involves quantifying the associated error correlation coefficients to include in the off-diagonal entries to the a priori and observation error covariance matrices.

[3] We present in this paper a method to objectively quantify error correlations between CO and CO2 as constraints in inverse model calculations for the budgets of both gases. Carbon dioxide over continents is released by combustion and biospheric respiration, and is taken up by photosynthesis. Over many regions of the world these combustion and biogenic fluxes overlap significantly. Separation of these fluxes in inverse analyses of CO2 observations can therefore be difficult. Combustion is a source of CO and CO2, with the CO emission ratio decreasing with combustion efficiency. Biospheric fluxes of CO are small. Here we use the error correlation coefficients between CO and CO2 in an inverse model analysis to estimate simultaneously CO and CO2 fluxes with the principal purpose of reducing uncertainty of combustion and biospheric CO2 flux estimates. We apply this method to CO and CO2 concentration data from the NASA TRACE-P aircraft campaign [Jacob et al., 2003], conducted over the western Pacific during February–April 2001. Two aircraft (DC8 and P3B) were used during TRACE-P to sample Asian outflow between 10° and 45°N and 0 and 12 km altitude. TRACE-P was conducted during the season of strongest outflow from mainland Asia to the Pacific, driven by frequent midlatitude cyclones and associated cold fronts and warm conveyor belts [Liu et al., 2003]. The February–April period is also the biomass burning season in Southeast Asia [Heald et al., 2003].

[4] Recent studies have used observed CO2:CO slopes derived from linear regression of atmospheric composition data to identify the source origins of air masses sampled in TRACE-P [Suntharalingam et al., 2004; Takegawa et al., 2004]. Suntharalingam et al. [2004] showed strong correlations between CO2 and CO concentrations during TRACE-P with distinct CO2:CO slopes that they used to constrain regional flux estimates of CO2. They interpreted the CO2:CO correlations using the GEOS-CHEM chemical transport model (Appendix A) and identified a large underestimate in the Chinese biospheric CO2 flux. Other TRACE-P studies have used CO observations with success to constrain east Asian source estimates of CO [Carmichael et al., 2003; Palmer et al., 2003; Allen et al., 2004; Heald et al., 2004; Tan et al., 2004], but without consideration of the implications for CO2 surface fluxes.

2. Correlations Between CO and CO2 Errors

[5] There are two types of CO2:CO correlations that can serve as error covariance constraints for a joint CO2:CO inverse analysis. Correlations between CO2 and CO combustion sources provide a constraint on the a priori fluxes used for the joint inversion. Observed correlations between atmospheric CO2 and CO concentrations imply corresponding correlations in the source-receptor relationships determining the concentrations at the observation points, i.e., in the forward model for the inversion. Therefore they provide information on model error correlations for the observation error covariance matrix. Here we derive these two types of error correlation for the Asian outflow conditions sampled by the TRACE-P mission.

2.1. Observation Error Correlations

[6] Atmospheric transport processes during TRACE-P moved air masses laden with CO and CO2 from east Asia over the Pacific [Liu et al., 2003; Fuelberg et al., 2003], introducing correlation in the CO and CO2 concentrations. TRACE-P took place before the onset of the growing season, hence the correlations are positive. Figure 1 illustrates the spatial distribution of TRACE-P CO and CO2 concentration data in the boundary layer. Examination of the spatial distributions reveal strong regional correlations for individual flights (correlation coefficients >0.8) with CO2:CO slopes ranging from 10 mol/mol in biomass burning outflow from Southeast Asia, to 20–30 mol/mol in Chinese pollution plumes and 50 mol/mol in outflow from Japan [Suntharalingam et al., 2004; Takegawa et al., 2004]. TRACE-P sampled relatively fresh air masses (1–2 days old) so we can assume that loss of CO by reaction with OH does not affect the CO2:CO correlation driven by Asian outflow.

Figure 1.

Mean CO and CO2 boundary layer (0–2 km) concentrations during the TRACE-P aircraft mission in March–April 2001 [Jacob et al., 2003]. The data are averaged on the GEOS-CHEM 2° × 2.5° model grid.

[7] Figure 2 shows correlation coefficients between CO and CO2 concentrations measured during TRACE-P as a function of altitude and latitude. These correlation coefficients are typically >0.7, and vary only by 5–10% as a function or altitude, with smaller values at higher altitudes where aged air masses originating from regions outside Asia play a relatively large contribution to the total CO abundances. It is useful to split the TRACE-P data into two distinct latitudinal regions, characterized by differences in sampled air masses [Blake et al., 2003]. North of 30°N, air masses were heavily influenced by fossil fuel and biofuel emissions from China, Korea and Japan, with essentially no influence from biomass burning [Liu et al., 2003]. South of 30°N, and particularly in the free troposphere, air masses were strongly influenced by biomass burning from Southeast Asia. Correlation coefficients are larger at latitudes <30°N (Figure 2) because background concentrations are lower and less variable (Figure 1).

Figure 2.

Correlation coefficients (r) between CO and CO2 mixing ratios measured during TRACE-P as a function of altitude (2-km altitude bins). Numbers inset in each plot refer to the number of observations used to compute the correlation at each altitude bin. Correlation coefficients calculated from 10 or more data points are shown.

2.2. Source Error Correlations

[8] Combustion is a source of both CO and CO2, and consequently introduces a correlation in emissions. We take advantage of a detailed fuel emission inventory for east Asia [Streets et al., 2003] that offers an opportunity to quantify the source error correlation by providing uncertainty estimates for the sources from individual energy sectors. We do not consider error correlation for the biomass burning source, and as we will see this is justified because of large uncertainty in the CO emission factor.

[9] The emission of a gas from a combustion process (g gas emitted yr−1) can be described as a product of the activity rate A (g fuel burned yr−1) and the emission factor F (g gas emitted per g fuel burned). Activity rates are obtained from energy statistics. Emission factors are determined from laboratory test burns or field measurements. The activity rates and emission factors are both subject to uncertainties. Emission factors for CO2 and CO vary with the combustion process. For example, Chinese CO emission factors range from 0.7 Gg/PJ for industrial coal burning to 7.0 Gg/PJ for gasoline-powered cars [Streets et al., 2003]. Chinese CO2 emission factors for the same processes are 24.7 Gg/PJ and 20 Gg/PJ, respectively [Streets et al., 2003]. Even for inefficient forms of combustion, e.g., open biomass burning, values of F for CO2 are an order of magnitude greater than for CO [Andreae and Merlet, 2001]. The sum of CO and CO2 typically represents >95% of the total emitted carbon, with CH4, and nonmethane volatile organic compounds (NMVOC) contributing the remaining few percent [Streets et al., 2003].

[10] Correlation between CO and CO2 emissions can be introduced via A or F. We assume activity rates are common to both gases, while CO and CO2 emission factors are inversely correlated. We use the east Asian emission inventory for 2000 from Streets et al. [2003], which includes information about A, F, and associated 1σ uncertainties (Δ) for different energy sectors and for individual regions. The inventory includes sector emissions from industry (coal and oil), domestic biofuel, and transportation (domestic and commercial, gas and diesel powered, vehicles). We derive error correlations for the national inventories, summing over all sectors and over all regions within a country.

[11] Table 1 presents values of ΔA and ΔF, expressed as a percentage of A and F, from Streets et al. [2003] for individual east Asian countries. CO emission estimates are the most uncertain, due to the uncertainty of the emission factors. The most uncertain emission estimates are from biofuels, reflecting large uncertainties in both F and A.

Table 1. Percentage Uncertainties of Activity Rates (ΔA/A) and Emission Factors (ΔF/F) Associated With East Asian CO and CO2 Emissions From Fuel Consumption
RegionBiofuelFossil FuelaCO2:CO Source Error Correlation
  • a

    CO and CO2 sectors include domestic coal, domestic oil, industrial coal, and transport [Streets et al., 2003]. Values of ΔA and ΔF given here are national averages (Figure 4) weighted by activity rates over different sectors and regions resolved by Streets et al. [2003].

  • b

    SEA in Figure 4.

Southeast Asiab36240306311211090.17

[12] We use a Monte Carlo approach to estimate error correlations between CO and CO2 sources for individual countries, in which we generate a large ensemble of CO and CO2 emissions Ei for each energy sector i within each country by perturbing the mean values for Ai and Fi by their estimated 1σ uncertainties, ΔAi and ΔFi, and a population of normally distributed random numbers εi with unit variance (equation set (1)):

equation image

[13] Regional CO2:CO error correlations are calculated by first summing the ensembles of CO and CO2 emissions generated using the Monte Carlo approach from different energy sectors within a particular country. For each energy sector i within a particular country we generate an ensemble of size 104, ensuring a statistical significant correlation. In the absence of a regional and sector breakdown of CH4 and other VOC emission factors, we assume that CO2 and CO represents 97.5% of the total carbon emitted [Streets et al., 2003], and that CH4 and other VOCs contribute the remaining 2.5%. In perturbations to this best estimate we require that the sum of the CO2 and CO emission factors is between 95% and 100%, ensuring that we only consider physically realistic realizations, e.g., the total carbon emitted does not exceed the carbon content of the fuel burned.

[14] The value of the CO2:CO error correlation for a particular sector i depends on the relative uncertainties associated with Ai and Fi. In the limiting case, where values of Fi are known perfectly the correlation will be +1 because ΔA is common to both CO and CO2. Conversely, if Ai is known perfectly, uncertainties in Fi will lead to correlation coefficients between −1 and +1. In practice, there are significant uncertainties associated with both Ai and Fi, and error correlations between CO2 and CO sources represent a balance between the relative magnitudes and uncertainties of Ai and Fi, e.g., image is a smaller factor of variability for CO2 emission than the activity rate, while the converse holds for CO.

[15] We find that the error correlations on the a priori sources of CO2 and CO, corresponding to values of ΔAi and ΔFi prescribed by Streets et al. [2003], are negligible or weakly negative (Table 1). This is due primarily to large values of ΔFCOi. However, analyses of the TRACE-P CO data indicate that the Streets et al. [2003] inventory underestimated A and F [Tan et al., 2004], so we explored the sensitivity of the CO2:CO source error correlation to the range of possible values of ΔAi and ΔFi (Figure 3) using the same percentage change in ΔAi and ΔFi over all sectors i and over all regions. As we will show later, source error correlation coefficients need to be greater than 0.5 before we get a significant improvement in the inverse analysis for CO2 fluxes. Such correlations are unrealistic, given constraints on A from the CO2 concentrations [Suntharalingam et al., 2004].

Figure 3.

Regional error correlation coefficients r between bottom-up emission estimates for CO2 and CO from fuel consumption as a function of relative uncertainties in activity rates A and emission factors F. These illustrative calculations assume the same relative uncertainties for all energy sectors within the region.

3. Inverse Model

[16] Our inverse model aims to achieve the best estimate of CO and CO2 fluxes that is consistent with the TRACE-P concentration measurements of CO and CO2, the a priori CO and CO2 flux estimates, and their respective uncertainties. It builds on our previous CO-only inverse model that used the TRACE-P observations to estimate east Asian CO fluxes [Palmer et al., 2003].

3.1. Model Description

[17] Anthropogenic sources of CO and CO2 include burning of fossil fuels and biofuels, and biomass burning (Tables 2 and 3). We use a priori anthropogenic emission estimates of CO and CO2 from Streets et al. [2003] for east Asia and from Palmer et al. [2003] and Suntharalingam et al. [2004] for the rest of the world (ROW). We use the same biomass burning emission estimates as Palmer et al. [2003], based on daily satellite fire count data for the TRACE-P period [Heald et al., 2003]. We account for the secondary source CO from oxidation of anthropogenic VOCs coemitted with CO by scaling up fossil fuel and biofuel (19%) and biomass burning (16%) [Heald et al., 2004]; these scaling factors are different from those used by Palmer et al. [2003] and are reflected in the a priori emissions presented in Table 2. The ROW represents a background of CO and CO2, including sources outside of Asia and chemical production of CO from CH4. ROW also includes the chemical production of CO from biogenic NMVOCs, representing a small diffuse source over the TRACE-P period (which was mainly outside the growing season). Daily mean terrestrial biosphere fluxes of CO2 are provided by the CASA balanced biosphere model [Randerson et al., 1997]. TRACE-P data are a few days downwind of continental sources and at this offshore distance the influence of diurnal variation from the terrestrial biosphere is weak [Suntharalingam et al., 2004]. For the TRACE-P season (March) there is a net positive flux of CO2 from the east Asian biosphere [Suntharalingam et al., 2004]. Ocean CO2 fluxes, prescribed by Takahashi [1999], are assumed not to contribute significantly to the magnitude and variability of TRACE-P data [Suntharalingam et al., 2004]. We aggregate the CO2 flux from the global ocean with ROW.

Table 2. A Priori Sources of CO for the Inverse Model Analysis
RegionBiofuels (BF),a Tg CO yr−1Fossil Fuels (FF),a Tg CO yr−1Biomass Burning (BB),a Tg CO yr−1Global CH4 and Biogenic NMVOCs, Tg CO yr−1
  • a

    Sources from BF, BB, and FF include the secondary source of CO from the oxidation of NMVOCs coemitted with CO.

China (CH)49 ± 3859 ± 4618 ± 9-
Korea (KR)3 ± 13 ± 10.08 ± 0.04-
Japan (JP)3 ± 16 ± 10.3 ± 0.1-
Southeast Asia (SEA)23 ± 2310 ± 1081 ± 40-
Boreal Asia (BA)--12 ± 6-
Rest of world (ROW)124 ± 26301 ± 78396 ± 1981205 ± 301
Table 3. A Priori Surface Fluxes of CO2 for the Inverse Model Analysis
RegionBiofuels, Tg CO2 yr−1Fossil Fuels, Tg CO2 yr−1Biomass Burning, Tg CO2 yr−1Terrestrial Biosphere,a Tg CO2/March 2001Global Oceans, Tg CO2 yr−1
  • a

    Net flux for March 2001.

China (CH)856 ± 692553 ± 638294 ± 147693 ± 520-
Korea (KR)43 ± 3335 ± 841 ± 0.720 ± 15-
Japan (JP)30 ± 1813 ± 2036 ± 341 ± 31-
Southeast Asia (SEA)419 ± 189608 ± 2741320 ± 66023 ± 17-
Boreal Asia (BA)--197 ± 99360 ± 270-
Rest of world (ROW)1800 ± 90017137 ± 47416467 ± 3331698 ± 524−8050 ± −8050

[18] Observed CO and CO2 mixing ratios (measurement vector y) are related to CO and CO2 surface fluxes (state vector x) by

equation image

where K is the Jacobian matrix describing the forward model sensitivity. We use the GEOS-CHEM CTM as the forward model. Loss of CO is computed with fixed 3-D OH fields from GEOS-CHEM to enforce linearity as described by equation (2). The linearized CO and linear CO2 models allow us to describe total CO and CO2 model concentrations as a linear sum of contributions from different countries and different sources.

[19] The state vector x comprises annual mean CO and CO2 anthropogenic (including biomass burning) source estimates assuming seasonal variations [Duncan et al., 2003; Streets et al., 2003]. It also includes CO2 biospheric surface fluxes for March 2001 from the different geopolitical regions of Figure 4.

Figure 4.

Source regions for tagged CO and CO2 simulations. See Tables 2 and 3 for flux estimates.

[20] The state vector components were chosen on the ability of the TRACE-P data to independently resolve source regions (Figure 4) and processes (fuel FL and biomass burning BB), as derived formally by Palmer et al. [2003] and Heald et al. [2004]. For CO we use six components: CHFL, KRJP, SEA, CHBB, BABB, and ROW. For CO2 we add regional biospheric terms (BS) and use eight components: CHFL, KRJP, CHBB, BABB, ROW, CHBS, KRJPBS, BABS. Observed CO2 signatures for biomass burning from China (CHBB), boreal Asia (BABB), and for SEA are too weak to be resolved independently.

[21] The measurement vector y consists of the TRACE-P CO and CO2 mixing ratio data. CO and CO2 mixing ratios are measured using a differential absorption tunable diode laser system [Sachse et al., 1987] and a nondispersive infrared analyzer [Vay et al., 2003], respectively. The data used for the inversion includes 299 hours of measurements from the two aircraft over 28 flights between 27 February and 3 April, 2001. We use the data for individual flights averaged on the GEOS-CHEM 2° × 2.5° model grid.

[22] Errors associated with K and y (the sum of which is described by equation image), include measurement accuracy, subgrid variability of observations (representation error), and model error. The ensemble characteristics of these errors are described by the observation error covariance, SΣ, representing a sum of the covariance matrices from individual sources of error. This will be discussed further in the next section.

[23] The maximum a posteriori solution [Rodgers, 1976] is given by

equation image
equation image

where Sa is the a priori error covariance matrix, equation image is the optimized a posteriori state vector, equation image is the a posteriori error covariance matrix describing the error associated with equation image, and other variables are as defined previously. Source and observation CO2:CO error correlations, as derived in section 2, are implemented in the inversion as off-diagonal terms in Sa and SΣ, respectively.

[24] Inverse modeling using Bayesian synthesis generates artificially small a posteriori errors, due to the assumption that observational errors are random, Gaussian, representatively sampled by the observations, and described accurately by its covariance [Heald et al., 2004]. Here we use the a posteriori errors as a qualitative measure of the constraint provided by the CO2:CO error correlations.

3.2. Error Variance and Covariance Specification

[25] Error variance and covariance information is included in the inversion as diagonal and off-diagonal terms in Sa and SΣ, respectively. We specify a priori emission error variances for Asia following Streets et al. [2003]. We assign an uncertainty of 25% on the global source of CO from the oxidation of CH4 and biogenic NMVOCs [Palmer et al., 2003], and further assume flux uncertainties for CO2 from the ocean and terrestrial biosphere of 100% and 75%, respectively.

[26] Observation error variances are quantified using the mean statistics of the relative difference between the observed and model concentrations as a function of altitude and latitude [Palmer et al., 2003; Heald et al., 2004]. The mean bias between the model and observed concentration is assumed to reflect errors in model fluxes to be retrieved by the inversion, while the standard deviation about the mean bias (residual relative error, RRE) is assumed to reflect total observation error. RRE values for CO are typically 20–30% [Palmer et al., 2003]. Similar analysis for CO2 shows RRE values of 1–2%, with little variability as a function of altitude or latitude. The observation error variance for an individual observation yi is given by (RRE × yi)2.

[27] Observed correlations between CO2 and CO concentrations imply corresponding observation error correlations (section 2 and Figure 2), which we use to specify the off-diagonal elements in SΣ. Correlations in CO and CO2 emission errors provide similar information to construct the off-diagonal elements of Sa (section 2). These CO2:CO error correlation coefficients, r, are included as off-diagonal elements in SΣ and Sa following

equation image

We neglect the spatial correlations of observations as previous work for CO during TRACE-P has shown that it decays on a length scale of 150 km [Jones et al., 2003; Heald et al., 2004], which is short relative to the 2° × 2.5° grid over which the observations are sampled here.

4. Inversion Results

4.1. Best-Case Estimate

[28] Figure 5 shows results for our best-case inversion of CO and CO2 fluxes that uses the state vector defined in section 3, and the CO2:CO error correlations derived in section 2. The results for CO are generally the same as those presented by Palmer et al. [2003], including a 72% increase in the anthropogenic source from China but the 45% decrease in Southeast Asia obtained here is less than the 60% decrease inferred from using only TRACE-P CO data, and is more consistent with the constraints from the MOPITT data [Heald et al., 2004]. The CO source from Chinese biomass burning is weakly negative but is not significantly different from zero, given that equation image underestimates actual a posteriori errors.

Figure 5.

A priori (red), uncorrelated a posteriori (green), and best-case a posteriori (black) CO and CO2 flux estimates. The vertical lines superimposed on each estimate denotes the 1σ uncertainty.

[29] A posteriori emissions of CO2 show a 18% decrease in Chinese anthropogenic sources and a 10% increase in sources from Korea and Japan. The a posteriori decrease (increase) in Chinese anthropogenic CO2 (CO) source estimates implies an overestimate of sources with high CO2:CO signatures, and a missing source with a low CO2:CO signature, e.g., which could be contributed by small heavily polluting power plants [Tan et al., 2004] or burning of domestic coal and biofuels [Carmichael et al., 2003]. There is also a 95% reduction in the net Chinese biospheric CO2 source for March 2001 relative to the CASA model, consistent with the results of Suntharalingam et al. [2004], and a 10% increase in CO2 from the boreal Asia biosphere.

4.2. Sensitivity Studies

[30] Here we assess the added value of CO2:CO error correlations by comparing our best-case estimates with sensitivity calculations using modified correlations, i.e., modified off-diagonal terms in SΣ and Sa. The uncorrelated inversion (best-case inversion but with off-diagonal terms set to zero) estimates independently the optimal flux estimates of CO and CO2 (Figure 5). A posteriori emissions of CO are within a few percent of the best case inversion analysis, with the exception of biomass burning from Southeast Asia, as described above. A posteriori CO estimates are tightly constrained by the CO concentration data and are largely insensitive to changes in Sa and SΣ introduced by the CO2:CO error covariances. In the case of the Southeast Asian source, transport to the TRACE-P sampling region involved lifting to the free troposphere by convection and warm conveyor belts [Liu et al., 2003]. Consideration of the CO2:CO correlations in other geopolitical regions appears to help reduce the model transport errors associated with the Southeast Asian source, resulting in better agreement with the CO inversions that use the much denser MOPITT data [Heald et al., 2004].

[31] Including CO2:CO correlations further reduces the Chinese biospheric flux by 80%, while KRJPBS and BABS estimates are not significantly different from the uncorrelated inverse analyses. Examination of the a posteriori correlation matrix (normalized equation image) shows a particularly strong anticorrelation (−0.8) between CO2 fluxes from anthropogenic and biospheric sources. Consideration of the CO2:CO error correlation in the inversion facilitates the separation of these two sources by reducing the correlation between biospheric and anthropogenic sources. Figure 6 illustrates for China the strong anticorrelation between CHFL and CHBS remains until source or observation CO2:CO error correlations are greater than 0.7–0.8, after which it drops rapidly to zero. The source CO2:CO error correlation is the more effective constraint for separating biospheric and anthropogenic sources but as discussed earlier this error correlation is weak. Our best estimate, which uses both source and observation CO2:CO error correlations, leads to a CHFL:CHBS correlation 10% less than the uncorrelated inversion.

Figure 6.

Sensitivity of the a posteriori correlation between Chinese anthropogenic (CHFL) and biospheric (CHBS) CO2 flux estimates to the source (Sa) and observation (SΣ) CO2:CO error correlations.

[32] Figures 7 and 8illustrate the sensitivity of a posteriori Chinese fluxes to the strengths of the source and observation CO2:CO error correlations. Source CO2:CO error correlation (Figure 7) provides little improvement in a posteriori CO source estimates, since CO atmospheric concentration signals are much stronger than for CO2 and the transport from China to the TRACE-P sampling region is principally by boundary layer advection [Liu et al., 2003] which has low transport error. Anthropogenic and biospheric CO2 flux estimates are statistically significantly different from the uncorrelated inversion where the source CO2:CO error correlation is greater than 0.5. However, with the emission uncertainty for CO contributed mostly by emission factor [Streets et al., 2003] the source CO2:CO error correlation is much less than 0.5 (Table 1). We conclude that the source CO2:CO error correlation is unlikely to offer much benefit for joint CO2:CO inversions. In contrast, accounting for forward model errorCO2:CO correlation (i.e., off-diagonal terms of SΣ) effectively reduces the uncertainty of the observations (Figure 8). The magnitude of a posteriori CO estimates increases slightly with increasing CO2:CO error correlation, with the associated uncertainties decreasing. This reduction in CO emission uncertainty is observed only with changes in SΣ because the observation uncertainties dominate the calculation of equation image (equation (4)). The magnitude of the Chinese biospheric CO2 flux reduces as the observation error CO2:CO correlation increases and is a CO2 sink at correlations greater than 0.8. The magnitude of the a posteriori Chinese anthropogenic CO2 estimate is not significantly affected by increasing the observation error CO2:CO correlation. Separation of the biospheric and anthropogenic CO2 sources, indicated by the a posteriori correlation matrix (Figure 6), occurs at correlations coefficients >0.7, similar to observed CO2:CO correlation (Figure 2). The importance of the CO2:CO error correlations in the TRACE-P inverse problem lies in their role of reducing model transport error and subsequently reducing uncertainty on a posteriori CO2 flux estimates.

Figure 7.

Sensitivity of a posteriori Chinese anthropogenic source estimates and their uncertainties to the source CO2:CO error correlation (Sa).

Figure 8.

Sensitivity of a posteriori Chinese anthropogenic source estimates and their uncertainties to the observation CO2:CO error correlation (SΣ).

5. Closing Remarks

[33] Accurately quantifying CO2 exchanges with the terrestrial biosphere represents an important challenge for understanding the global carbon cycle. A standard method has been to use inverse models constrained with atmospheric observations of CO2, and occasionally CO2 isotopes. Strong observed correlations between CO2 and CO in atmospheric data, as well as CO2:CO correlations in combustion sources, suggest that a joint CO2:CO inversion might provide information to improve the separation of combustion and biospheric CO2 surface flux estimates. We showed here that such an inversion, using concurrent observations of CO2 and CO in Asian outflow from the TRACE-P aircraft campaign in Spring 2001, does indeed provide useful additional constraints for improving a posteriori CO2 regional flux estimates.

[34] Observed correlations between CO2 and CO concentrations in atmospheric data imply a corresponding correlation in source-receptor relationships, and therefore in the forward model errors for the inversion. Mathematically, they can be used to estimate off-diagonal terms in the observation error covariance matrix (SΣ) of a joint CO2:CO inversion. Observed CO2:CO correlation coefficients during TRACE-P are generally larger than 0.7 [Suntharalingam et al., 2004; Takegawa et al., 2004] and enable significant improvement of a posteriori flux estimates. CO2:CO correlations may be different for other regions and during other seasons.

[35] Correlation between CO2 and CO combustion sources in a priori estimates could also provide constraints to the joint inversion. We used a Monte Carlo approach to quantify the resulting CO2:CO source error correlations in individual Asian countries on the basis of uncertainties in activity rates and emission factors for individual energy sectors and source regions reported by the east Asia emission inventory of Streets et al. [2003]. We find that these correlation coefficients are weak (r < 0.2) because CO emission factors are the largest contribution to the overall uncertainty. Significant improvements in a posteriori CO2 and CO would require correlation coefficients >0.7 but these would require unrealistic uncertainties in activity rates.

[36] We find that the benefit of including CO2:CO error correlations in joint CO2:CO inversions is mainly for improving retrievals of CO2 surface fluxes, not CO sources. Atmospheric variability of CO is far greater than that of CO2, and separation of biospheric and combustion terms is not an issue, so that information on CO2 has little benefit for the CO source inversion.

[37] Satellite observations of CO2 have the potential to improve quantitative understanding of the global carbon cycle [e.g., Pak and Prather, 2001; Rayner and O'Brien, 2001]. Our work shows that this potential can be enhanced by exploiting correlations with concurrent satellite observations of CO, where CO information may be coretrieved from the measured spectra of one instrument (e.g., SCIAMACHY and AIRS), retrieved from another instrument aboard the same satellite platform, or taken from an instrument in the same orbit (e.g., OCO and TES).

Appendix A:: Forward Model

[38] The GEOS-CHEM version used here (6.01, has a horizontal resolution of 2° × 2.5°, with 48 vertical levels ranging from the surface to the mesosphere, 20 of which are below 12 km. The model is driven by GEOS-3 assimilated meteorology data from the Global Modeling and Assimilation Office GCM based at NASA Goddard. The 3-D meteorological data are updated every 6 hours, and the mixing depths and surface fields are updated every 3 hours.

[39] The CO and CO2 simulation are as described by Palmer et al. [2003] and Suntharalingam et al. [2004], respectively. The main sink for CO is oxidation by OH, leading to a lifetime of a few months. We prescribe global monthly mean OH fields calculated from a full chemistry model simulation (GEOS-CHEM v4.33) and consistent with observational constraints [Prinn et al., 2001]. The CO sink has little effect for the inversion of TRACE-P measurements, which are only a few days downwind of the Asian sources. The impact of the OH sink on the CO background is effectively taken into account by the adjustment of the ROW CO source [Palmer et al., 2003]. Suntharalingam et al. [2005] showed that not properly accounting for reduced carbon species (CO, CH4 and NMVOCs) in CO2 inversion analyses can significantly bias a posteriori flux estimates. This bias arises for two reasons: not accounting for the 3-D production of CO2 from the oxidation of reduced carbon species and neglecting that reduced carbon species are often included in bottom-up combustion and biospheric CO2 flux inventories. Air masses sampled during TRACE-P are relatively fresh and so the oxidation effect will only affect observations that sample the background atmosphere. The Streets et al. [2003] combustion emission inventory only accounts for direct emissions of CO2, however the terrestrial biospheric fluxes from CASA do account for reduced carbon species.

[40] The Jacobian matrix for the inversion is constructed by transporting separately CO and CO2 from individual sources describing the state vector. This required initialization of the corresponding “tagged tracers.” The CO simulation is initialized in January 2000 and integrated for 16 months (through April 2001). The 14 months prior to TRACE-P effectively remove the influence of initial conditions. In the case of CO2, we adopt the model initialization approach previously used by Xiao et al. [2004] for CH4 in GEOS-CHEM. We use the total CO2 simulation to define a background CO2 tracer as the CO2 present in the atmosphere on 1 January 2001. The background tracer has no surface fluxes after January 2001 and is allowed to mix. The other tagged tracers are initialized with zero concentration on 1 January 2001. Their concentrations evolve over the course of the simulation as a result of surface fluxes. We also included an additional initialization to the CO2 simulation to correct the model bias introduced by not accounting for the net uptake of CO2 by the land biosphere. We make this correction by comparing the 20th percentile (assumed to represent the CO2 background) in the model and observed CO2 concentration data (east of 150°W) as a function of latitude (5° increments), and remove it from the model values. The corresponding model bias ranges from 4 to 5 ppm and is consistent with the values calculated using GLOBALVIEW CO2 data [Suntharalingam et al., 2004].


[41] This work was supported by the NASA and NOAA carbon cycle programs.