estimating atmospheric co from advanced infrared satellite ...citation: engelen, r. j., and a. p....

Estimating atmospheric CO2 from advanced infrared

satellite radiances within an operational

four-dimensional variational (4D-Var) data

assimilation system: Results and validation

Richard J. Engelen and Anthony P. McNallyEuropean Centre for Medium-Range Weather Forecasts, Reading, United Kingdom

Received 16 March 2005; revised 4 May 2005; accepted 7 June 2005; published 23 September 2005.

[1] More than a year of Atmospheric Infrared Sounder (AIRS) radiance observationshave been assimilated in the European Centre for Medium-Range Weather Forecastsfour-dimensional variational (4D-Var) data assimilation system to estimate troposphericCO2. The assimilation of a set of 18 spectral channels provides a mean troposphericmixing ratio representing a layer between �700 hPa and the tropopause. Analysiserrors for a 5-day mean on a 6� by 6� averaging grid box are on the order of 1%.Comparisons with independent flight data from Japanese Airlines and National Oceanicand Atmospheric Administration Climate Monitoring and Diagnostics Laboratory arefavorable. Differences between the averaged assimilation estimates and the onboardflask observations are generally within the 1-s error bars. Currently, this work is beingextended by introducing CO2 as a full assimilation model tracer variable, whichwill allow the operational monitoring of atmospheric CO2 using AIRS observations andobservations from upcoming instruments.

Citation: Engelen, R. J., and A. P. McNally (2005), Estimating atmospheric CO2 from advanced infrared satellite radiances within an

operational four-dimensional variational (4D-Var) data assimilation system: Results and validation, J. Geophys. Res., 110, D18305,

doi:10.1029/2005JD005982.

1. Introduction

[2] A proper understanding of the global carbon cycle iscritical for understanding the environmental history of ourplanet and its human inhabitants, and for predicting andguiding their joint future [Global Carbon Project, 2003].[3] One path to achieve the above goal is by inferring

surface fluxes of CO2 from atmospheric CO2 observationsusing an inverse transport model. Significant progress hasbeen made in the last decade using observations fromthe surface flask networks [e.g., GLOBALVIEW-CO2,2003] (available on the Internet via anonymous ftpto ftp.cmdl.noaa.gov, Path: ccg/co2/GLOBALVIEW) byimproving transport models and inversion techniques [e.g.,Gurney et al., 2002, 2004; Peters et al., 2004; Rodenbeck etal., 2003]. However, these surface flask networks, althoughbeing extended, are limited in number and geographicalarea, which limits the inversion approach. Rayner andO’Brien [2001] showed that the greater coverage in timeand space provided by satellite data can improve existingsurface flux estimates even though the precision ofindividual measurements may be an order of magnitudelower than those estimated from the air samplingnetwork.

[4] Engelen et al. [2001] performed a simulation study tolook at the capabilities of the Atmospheric Infrared Sounder(AIRS), Chedin et al. [2003] did similar simulations for theInfrared Atmospheric Sounding Interferometer (IASI), andO’Brien and Rayner [2002] studied the near-infrared option,which might be realized by the Orbiting Carbon Observa-tory (OCO) mission. AIRS and IASI (will) observe emis-sion in the infrared part of the spectrum and are thereforenot very sensitive to lower tropospheric CO2 (below800 hPa), but they can observe every location two timesper day. OCO will observe reflection in the near-infraredand will therefore be most sensitive to lower troposphericCO2, but can only observe CO2 during daytime. All threestudies showed that the required accuracy of 2.5 ppmv formonthly mean column-integrated data on a 8� � 10�footprint [Rayner and O’Brien, 2001] is in principle achiev-able. Chedin et al. [2002] used real data from the TirosOperational Vertical Sounder (TOVS) to infer atmosphericCO2 concentrations in the tropics, and Engelen et al. [2004]described the use of AIRS observations in the EuropeanCentre for Medium-Range Weather Forecasts (ECMWF)four-dimensional variational (4D-Var) data assimilation sys-tem to infer tropospheric CO2 concentrations.[5] In this paper, we present the current status of the CO2

data assimilation system as described by Engelen et al.[2004]. In section 2 we give a brief summary of thedata assimilation system and describe the main changescompared to system described by Engelen et al. [2004].

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 110, D18305, doi:10.1029/2005JD005982, 2005

Copyright 2005 by the American Geophysical Union.0148-0227/05/2005JD005982$09.00

D18305 1 of 9

Section 3 shows monthly mean results, while section 4describes comparisons with independent CO2 observations.

2. Short Description of the Assimilation Systemand Recent Changes

[6] Engelen et al. [2004] described the setup of thecurrent CO2 data assimilation system at ECMWF. In short,radiance observations from the Atmospheric InfraredSounder (AIRS) [Aumann et al., 2003] are operationallyassimilated in the ECMWF 4D-Var data assimilation systemtogether with many other observations to constrain thedynamics and thermodynamics of the assimilation model[McNally et al., 2005]. For the CO2 assimilation experi-ments, a CO2 column variable is added to the minimizationstate vector of the analysis system at all available AIRSobservation locations within the 6 hour time window ofeach analysis. This means that an analysis does not providea full three-dimensional CO2 field for each analysis cycle,as is the case for most other variables, but individual CO2

estimates for all AIRS observation locations. There istherefore no CO2 transport within the 6 hour analysiswindow and there are no a priori horizontal correlationsapplied to constrain the CO2 estimation problem. The onlyexplicit prior CO2 constraint is in the vertical by assuming awell-mixed profile. Because the CO2 mixing ratios areestimated within the regular minimization, they fully usethe information about the other meteorological variables,such as temperature and water vapor, which are constrainedby various data sources (e.g., AMSU-A, AMSU-B, SSMI,radiosondes) as well as the assimilating model and back-ground state that is produced with a 3 hour forecast from theprevious analysis. This is a significant difference withstand-alone satellite retrievals that generally use only obser-vations from the same satellite platform. For the CO2

experiments we ran the assimilation model with 60 levelsat resolution T159, which is approximately 1.125� by1.125�.

2.1. CO2 Column Variable

[7] Engelen et al. [2004] estimated two column amounts:a tropospheric and a stratospheric column separated by avariable tropopause. However, careful analysis of the resultsshowed that there were problems with channels that hadsignificant sensitivity both in the troposphere and thestratosphere. Because there was no correlation betweenthe two layers applied by the background constraint, dipolestructures could arise between the troposphere and strato-sphere (because the temperature lapse rate changes sign).When the bulk of the sensitivity was in the stratosphere, astrong increment in the tropospheric CO2 value could becompensated by a small increment in the stratosphere. thefact that stratospheric temperatures in our assimilationmodel often suffer from biases increased this problem.Therefore the channel selection was changed to includeonly 18 channels that are mainly sensitive to troposphericCO2 as is shown by the weighting functions in Figure 1.This also implied that the stratospheric column could nolonger be estimated. Furthermore, we changed the values ofthe background estimate to a single global mean value of376 ppmv. This made the interpretation of the results morestraightforward compared to the previously used zonal

mean monthly mean background, because any horizontalstructure cannot be provided by the background. The376 ppmv is based on the approximate annual mean of2003 for the Mauna Loa flask station on Hawaii. Althoughthis value is not necessarily the same as the global annualmean for 2003, it is close enough to minimize thebackground bias on the timescale of a year. For compar-ison, the variation due to the seasonal cycle is on the orderof 10 ppmv.[8] An extra benefit of the small channel selection was

that it became computationally affordable to estimate theanalysis error using the full Bayesian equation instead ofrelying on the neural network approach as described byEngelen et al. [2004]. Individual analysis error estimates(sa) are calculated using

s2a ¼ s�2b þHTR�1H

� ��1 ð1Þ

where sb is the background error, R is the observation errorcovariance containing both the measurement error covar-iance (O and the observation operator error covariance Fmatrices, and H are the CO2 Jacobians (sensitivity of theobserved brightness temperatures to a change in the CO2

column concentration). The observation error covariancematrix is currently a diagonal matrix containing estimates ofthe combined instrument and observation operator errorswith variances set to (0.6 K)2. The observation operator is afast radiative transfer model (Radiative Transfer forthe TIROS Operational Vertical Sounder (RTTOV8)[Matricardi et al., 2004]) that links the atmospheric profilevariables to a simulated radiance. The error estimatetherefore contains uncertainties in spectroscopy and modelapproximations.[9] The analysis errors estimated using equation (1) do

not take horizontal and temporal error correlations intoaccount. However, these correlations are introduced in theassimilation system itself (e.g., through error correlations ofthe temperature field). We therefore previously defined anupper and lower estimate for the error of the gridded time-averaged product by assuming either full horizontal andtemporal correlation between the individual error estimatesor zero correlation, respectively. This wide range of possiblemean error was not ideal and we therefore have tried toestimate these error correlations empirically. A statisticalanalysis of a large ensemble of analysis error estimates wasused to infer an approximation of the error correlation. Theerror correlation is here defined as

r ¼S si � �sð Þ sj � �s

� �

S si � �sð Þ2S si � �sð Þ2h i1=2 ð2Þ

and Figure 2 shows this correlation as a function of distancebetween the error estimates in degrees as well as of timeseparation in days. The figure shows that for observations ofthe same day the correlation falls of to zero within a radiusof �12�. The time correlation is becomes very small within�5 days.[10] Table 1 shows the effect of using these estimated

error correlations on the mean error estimate in the tropicsfor various grid box sizes and averaging time periods. While

D18305 ENGELEN AND McNALLY: CO2 DATA ASSIMILATION

2 of 9

D18305

the maximum error estimate (using full error correlations)remains the same (4.6 ppmv) for all combinations, theminimum error (using zero error correlations) is clearly afunction of

ffiffiffiffiN

p, where N is the number of observations

used in the average. This lower estimate is far toooptimistic as can be seen from the mean errors using themore realistic estimated error correlations. These estimatesvary from 1.3 ppmv for a 10� � 10� grid box size over a30-day period to 3.7 ppmv for a 1� � 1� grid box size overa 5-day period. These estimates fall well within the targetvalue of 2.5 ppmv for a monthly mean as described byRayner and O’Brien [2001]. It is important to note thatthese estimates only reflect the random errors. Any sys-tematic errors that have not be corrected will deterioratethe results. Systematic errors can arise from errors in theradiative transfer, errors in the cloud detection, and sys-

tematic errors in the temperature and water vapor fields ofthe final analysis.

2.2. Cloud Detection

[11] In the first results described by Engelen et al. [2004],high CO2 values were seen in the western Pacific region, amajor area of tropical convection in February. There was nodirect explanation of these increased values and it wassuggested that the cloud detection could have had an effecton the CO2 estimates. The cloud detection scheme for AIRSis described by McNally and Watts [2003], and the detailswill not be reproduced here. In summary it is a noveltechnique for the identification of clear channels at aparticular location rather than the more traditional approachof identifying completely clear locations. Departures of theobserved spectrum from clear-sky background values arefirst reordered into a vertically ranked space (i.e., in order ofincreasing sensitivity to cloud), in which the characteristicsignature of cloud becomes monotonic and more readilyidentifiable as shown in Figure 3. A digital filter is thenapplied to the departures to reduce the instrument noise (andnoise due to errors in the background estimate of theatmospheric state). This essentially isolates the pure cloudsignal such that the level (or channel in the ranked space)where the cloud contamination first becomes significant canbe determined. Channels ranked above this level (i.e., lesssensitive to cloud) are retained for assimilation and channels

Table 1. Mean Errors as a Function of Spatial and Temporal

Averaging Using Full Error Correlation (r = 1), Zero Error

Correlation (r = 0), and Error Correlations as a Function of Spatial

and Temporal Distancea

Monthly 5-Day

r = 0 r = 1 r = f (x, t) r = 0 r = 1 r = f (x, t)

10� � 10� 0.1 4.6 1.3 0.3 4.6 2.55� � 5� 0.2 4.6 1.6 0.6 4.6 3.11� � 1� 1.0 4.6 2.1 1.9 4.6 3.7

aMean errors are given in ppmv.

Figure 3. Example of the cloud detection scheme. Solidcircles denote channels flagged as cloudy. See text forfurther explanation.

Figure 1. CO2 weighting functions for the 18 channelsused in the assimilation. The weighting functions representthe change in brightness temperature (dBT) for a 1% changein CO2 mass mixing ratio at every level.

Figure 2. Estimated error correlations as a function ofdistance and time separation.


3 of 9

D18305

ranked lower (i.e., more sensitive to cloud) are discarded.The result of such an approach is that the data coverage inthe analysis is different for each channel. Channels withweighting functions peaking higher in the atmosphere areused extensively whereas channels with lower peakingweighting functions (e.g., window channels) are used sig-nificantly less often. The cloud detection scheme has anumber of tunable parameters, such as the window width ofthe digital filter and the gradient threshold of the smoothedcurve that stops the detection, which have initially been setto rather stringent values. This possibly results in thewrongful rejection of some clear data, but ensures thaterrors due to undetected residual cloud contamination inchannels passed as clear are very small (estimated to betypically less that 0.2K for a midtropospheric temperaturesounding channel).[12] Although the cloud detection works very well in

most cases [McNally and Watts, 2003], the assumption thatthere are no systematic errors proved to be a problem in thetropical convective areas. The detection of thin cirrus (thinenough to still show the atmosphere and/or surface under-neath it) was compromised by large systematic errors in thebackground water vapor profiles that affect the lowestpeaking channels in the long-wave band (sensitive to watervapor). Figure 4 (left) shows an example of an undetectedthin cirrus cloud. A large error in the water vapor back-ground profile has lifted these lowest peaking channels inthe long-wave band up toward the zero departure line. Thecloud detection scheme therefore detects a cloud at channel123, which is a very low peaking channel. It fails to detectthe real cloud effect that is causing the smoothed curve todrop again outside the threshold range. The simple fix thatwas implemented consisted of removing the 30 lowestpeaking channels from the cloud detection to remove thebiased water vapor effect. Figure 4(right) shows the result.Because the channels used are now not sensitive to watervapor anymore, the cloud detection is able to detect thecirrus cloud, which affects all the channels with an indexhigher than 79. This fix improved the cloud detection in thetropics and therefore removed some (but not all, as will beshown in section 4) anomalous CO2 estimates around the

edges of convective clouds. The channels that wereremoved from the cloud detection and therefore flagged ascloudy for all observation locations were not used for theCO2 estimation itself. The above described cloud detec-tion problem affected other convective areas as well, butwe have focused on the Pacific, because we haveindependent CO2 data in that area and also because thewestern Pacific is the most important area for large-scaletropical convection.

3. Results

[13] As done by Engelen et al. [2004], monthly meanresults are presented by averaging on a 1� by 1� latitude-longitude grid. Within a grid box the data were averagedusing the individual analysis error estimates as weights.This 1� by 1� grid was then smoothed with a 15� by 15�moving boxcar average for clarity. Monthly mean errorswere averaged on the same grid taking the estimated errorcorrelations into account. AIRS data in the period from 1January 2003 until 31 March 2004 have been processed,and Figure 5 shows monthly mean CO2 analysis results forMarch 2003, September 2003, and March 2004. Alsoshown is the monthly mean analysis error for March2003. The mean errors for the other two months are verysimilar and therefore not shown. The largest signal inatmospheric CO2 concentrations comes from the terrestrialbiosphere [e.g., Erickson et al., 1996]. A strong seasonalcycle is produced, although the annual net biosphere flux isvery close to zero. The terrestrial biosphere also creates alatitudinal gradient in the atmospheric concentrations due tothe large amount of land in the Northern Hemispherecompared to the Southern Hemisphere. This latitudinalgradient is amplified by anthropogenic emissions that mainlyoriginate from the Northern Hemisphere [Denning et al.,1995]. Both the seasonal cycle and the latitudinal gradientare visible in the results of Figure 5. It is encouraging to seethat the assimilation is capable of producing these spatialand temporal variations without having that information inthe background, as was the case in the previous results ofEngelen et al. [2004]. Another marked difference with the

Figure 4. Example (left) of undetected thin cirrus with the old cloud detection scheme and (right) of thedetection of the thin cirrus with the new cloud detection scheme.


4 of 9

D18305

earlier results is that the high values in the tropical convec-tive areas are significantly reduced and fit better within theCO2 patterns of the surrounding areas. Furthermore, March2004 shows generally higher CO2 concentrations thanMarch 2003, probably because of the upward trend inglobal atmospheric CO2. The difference between March2003 and March 2004 at the location of Hawaii is 1.6 ppmvcompared to the 2.3 ppmv observed at the Mauna Loa flaskstation. This is reasonably close, especially considering thatthe observations represent different parts of the troposphere.The monthly mean error shows the clear dependence of theanalysis error on the temperature lapse rate as well asthe thickness of the tropospheric layer. Errors are smallestin the tropics were the tropopause is high and the temper-ature lapse rate is large, while they increase at higherlatitudes where the tropopause is lower. The relatively lowerrors over Europe are caused by a higher tropopause(deeper tropospheric layer) in the subtropical air mass.The data density does not significantly affect the monthlymean error, because of the applied error correlations (seesection 2.1). Monthly mean errors range from about 1 ppmvto 6 ppmv. These are the random errors; uncorrectedsystematic errors are not included in the estimate. Thepatterns in the CO2 distribution are quite reasonable and

compare well with independent CO2 simulations over oceanfrom the LMDz model [Sadourny and Laval, 1984;F. Chevallier, personal communication, 2005].

4. Validation

[14] The presentation of monthly mean results and thecomparison with model simulations is interesting by itself,but an important check of the validity of our analysis resultsis by comparing these results to independent observations ofatmospheric CO2. There are only very few data sources for2003. We cannot use the surface flask data as our estimatesrepresent a layer between �700 hPa and the tropopause,while the surface flasks are sampled in the boundary layer.Only if we are sure that the full tropospheric CO2 profile iswell mixed, a comparison would be useful. However, twodata sets of middle to upper tropospheric CO2 do exist andcomparisons are presented here.

4.1. Comparisons With Japanese AirlinesMeasurements

[15] The first data set consists of CO2 data sampled fromJapanese Airlines (JAL) commercial airliners flyingbetween Australia and Japan [Matsueda et al., 2002]. These

Figure 5. Monthly mean analysis results for March 2003, September 2003, and March 2004 as wellas the monthly mean analysis error for March 2003. See color version of this figure at back of thisissue.


5 of 9

D18305

observations consist of automatic flask samples gathered ataltitudes between 8 and 13 km on biweekly commercialflights. For 2003, 21 flights were available for our compar-isons. Figure 6 shows the CO2 annual cycle for both theflight observations and the assimilation estimates. For thefull processed period (1 January 2003 to 31 March 2004),CO2 estimates were sampled in a 6� � 6� box around thelocation of the flight observation over a period of 5 daysaround the date of the flight observation. A minimum of100 analysis estimates for the average was required toproduce a mean CO2 value for each location. This resultedin annual cycle plots for 12 different latitudes, because eachJAL flight has 12 flask observations at approximately thesame latitudes.[16] For clarity, we generated three plots from these 12

annual cycles, shown in Figure 6, that represent the North-ern Hemisphere region, the equatorial region, and theSouthern Hemisphere region, by averaging three latitudeplots together for each region. The CO2 analysis estimatesrepresent a thick layer, while flight observations are taken ata certain height level, but Sawa et al. [2004] showed that (atleast) in February 2000 the CO2 mixing ratios in the zone13�–24� N are very similar at various altitudes and agreewell with the JAL observations. Figure 6 shows that theanalysis estimates follow the JAL observed annual cyclequite well. All differences fall within the 1-s error bars andare of the order of 1 ppmv in most cases. There is a clearimprovement compared to the used background, which is376 ppmv throughout the year. The main anomaly can be

seen in both the Northern Hemisphere and the SouthernHemisphere in January and February. The analysis estimatesare consistently higher than the JAL observations for thisperiod. The 12 individual annual cycles show the samebehavior, although the results have more random errorbecause of less averaging.[17] Figure 7 shows geographical comparisons for Janu-

ary and May 2003. The figure for 20 January 2003 is anexample of a bad match between the JAL observations andthe analysis results, while the figure for 20 May 2003 is anexample of a good match. The main problem area in theJanuary plot is an area affected by clouds, although this isnot immediately visible in the smoothed average. There is asmall area that is rightly detected as cloudy, but observa-tions around the edges of this area, detected as cloud-free,provide high CO2 values. This causes the average CO2 fieldto be high compared to the surrounding area. These obser-vations that are detected as cloud-free, still suffer fromproblems in the cloud detection related to the backgroundwater vapor field. In contrast, the May plot shows nicecorrespondence in the north-south gradient. This differencein the cloud detection between January 2003 and May 2003is most likely caused by the effect of the GOES-9 satelliteon the moisture fields of the ECMWF analysis. The normalgeostationary satellite over the Pacific area was GMS-5, butcontinuing problems with the onboard imager required areplacement by the GOES-9 satellite in May 2003. There-fore, during the critical (in terms of amount of convectionover the Pacific) months of January, February and March

Figure 6. Comparison of CO2 estimates with JAL observations for three different latitude zonesfrom January 2003 to March 2004. Missing ECMWF data are caused by extensive cloud cover in thearea.


6 of 9

D18305

the geostationary constraint on the humidity field waslacking. After the middle of May, this was corrected withthe GOES-9. This is a cautionary example showing theimpact of seemingly unrelated satellite instrument changeson the CO2 estimates. Continuous validation of the results istherefore of great importance. Currently, efforts are underway to make the cloud detection more robust.

4.2. Comparisons With Climate Monitoring andDiagnostics Laboratory Measurements

[18] Another source of data to validate our analysisresults are the vertical profiling sites managed by the

National Oceanic and Atmospheric Administration ClimateMonitoring and Diagnostics Laboratory (NOAA/CMDL).Air samples are collected using an Automated Air SamplingSystem aboard chartered aircraft at a frequency of aboutonce a month at several sites. These samples are thenanalyzed at CMDL to provide concentrations of CO2,CH4, CO, H2, N2O, and SF6 as a function of altitude(CMDL Summary Report 27, 2004; available from http://www.cmdl.noaa.gov/publications/annrpt27/). Most sites arelocated in the continental United States and thereforeoutside the tropical area where our analysis estimates aremost reliable. However, three sites are located elsewhere:

Figure 7. Comparison of CO2 estimates with JAL observations for 20 January 2003 and 20 May 2003.See color version of this figure at back of this issue.

Figure 8. ECMWF tropospheric mean mixing ratio (solid lines) compared to CMDL flight profiles(solid circles) for 2 April, 5 June, 17 August, and 9 October 2003 at Molokai Island, Hawaii. The analysisbackground value is represented by the dashed lines, and the mean analysis error is represented by thehorizontal solid line.


7 of 9

D18305

Molokai Island (Hawaii; 21.23�N, 158.95�W), Rarotonga(Cook Islands; 21.25�S, 159.83�W), and Santarem (Brazil;2.85�S, 54.95�W). Figure 8 shows 4 profiles measured atthe Hawaii site for 2 April 2003, 5 June 2003, 17 August2003, and 9 October 2003. The solid vertical line representsthe CO2 analysis estimate with the 1-s error bar shown asthe horizontal line. The vertical dashed line is the back-ground value used in the analysis. Only four profiles areshown for clarity, but they were chosen to be representativeof the other available profiles in 2003. The CO2 analysisvalues were calculated by averaging within a 6� � 6� box ofthe CMDL site and over an period of 5 days around the dateof the CMDL observation. The average error was calculatedthe same way as for the comparisons with the Japanese data.The figure shows that the analysis estimates are closer to theCMDL observed profiles than the background and fit wellwithin the 1-s error. The changes due to the seasonal cycleare clearly captured. Figure 9 shows scatter diagrams withall available data for Hawaii, Rarotonga, Santarem, andHarvard Forest. Each plot also provides the linear correla-tion coefficient (r). The horizontal dashed line represents thebackground value used in the assimilation. The analysisvalues correlate well with the CMDL data, although thescatter is not negligible. This means that the annual cycle is

generally well captured as with the Japanese data, butindividual comparisons can have differences up to 3 ppmv.The error bars seem large in this kind of scatter diagram, butare well within 1%. The Harvard Forest plot was added toshow an example for higher latitudes.

5. Summary

[19] More than a year of AIRS data have been assimilatedin a CO2 configuration of the ECMWF 4D-Var dataassimilation system. Eighteen spectral channels sensitiveto tropospheric CO2 are used to estimate the mean CO2

concentration in layer between 700 hPa and the tropopause.Results look realistic and compare well with model simu-lations and independent CO2 observations.[20] The analysis estimates are currently not constrained

by the transport model and the background constraint isvery weak. Therefore individual estimates are quite noisyand require spatial and temporal averaging. However,5-day averages on a 6� by 6� grid compare well withobservations from the JAL commercial airliner flyingbetween Japan and Australia at altitude of �11 km.The estimated mean errors are of the order of 1%, whichshould be sufficient to have an impact on current flux

Figure 9. Scatter diagrams of analysis CO2 estimates versus CMDL flight data for Hawaii, CookIslands, Brazil, and; Harvard Forest.


8 of 9

D18305

inversions. Comparisons with profile data from theCMDL network shows similar agreement.[21] An important problem area of the assimilation esti-

mates is undetected systematic errors. Not only can biasedtemperature and water vapor fields directly bias the CO2

estimates through the radiative transfer modeling, but it wasalso shown that undetected clouds can have an impact onthe estimated CO2 values and the cloud detection algorithmwas adjusted to minimize this problem. Continuing valida-tion of the analysis results is required to remove anyremaining systematic errors.[22] On the basis of the results presented in this paper we

will introduce CO2 as a tracer in the ECMWF transportmodel, which will make it part of the full 4D-Var assimi-lation, including the constraint in the time dimensionprovided by the transport model. This will require a moresophisticated background constraint that realistically repre-sents the background errors.[23] At the same time first efforts will be made to use the

CO2 estimates in a flux inversion model. This will showwhere data from the flask network and the data presented inthis paper agree and disagree. It will also provide anestimate of the added information from the satellite esti-mates compared to just using flask observations.

[24] Acknowledgments. The work described in this paper was finan-cially supported by the EC project COCO (EVG1-CT-2001-00056). Theauthors are very grateful to Hidekazu Matsueda from the MeteorologicalResearch Institute in Japan for providing the JAL flight data. The authorsare also very grateful to Pieter Tans from NOAA/CMDL for providing theCMDL flight data. Furthermore, we would like to thank two anonymousreviewers for their very useful comments on the manuscript.

ReferencesAumann, H. H., et al. (2003), AIRS/AMSU/HSB on the Aqua mission:Design, science objectives, data products, and processing systems, IEEETrans. Geosci. Remote Sens., 41, 253–264.

Chedin, A., A. Hollingsworth, N. A. Scott, S. Serrar, C. Crevoisier, andR. Armante (2002), Annual and seasonal variations of atmosphericCO2, N2O and CO concentrations retrieved from NOAA/TOVSsatellite observations, Geophys. Res. Lett., 29(8), 1269, doi:10.1029/2001GL014082.

Chedin, A., R. Saunders, A. Hollingsworth, N. Scott, M. Matricardi,J. Etcheto, C. Clerbaux, R. Armante, and C. Crevoisier (2003), Thefeasibility of monitoring co2 from high-resolution infrared sounders,J. Geophys. Res., 108(D2), 4064, doi:10.1029/2001JD001443.

Denning, A. S., I. Y. Fung, and D. Randall (1995), Latitudinal gradient ofatmospheric CO2 due to seasonal exchange with land biota, Nature, 376,240–243.

Engelen, R. J., A. S. Denning, K. R. Gurney, and G. L. Stephens (2001),Global observations of the carbon budget: 1. Expected satellite capabil-

ities for emission spectroscopy in the EOS and NPOESS eras, J. Geo-phys. Res., 106, 20,055–20,068.

Engelen, R. J., E. Andersson, F. Chevallier, A. Hollingsworth, M.Matricardi,A. P. McNally, J.-N. Thepaut, and P. D. Watts (2004), Estimating atmo-spheric CO2 from advanced infrared satellite radiances within an opera-tional 4D-Var data assimilation system: Methodology and first results,J. Geophys. Res., 109, D19309, doi:10.1029/2004JD004777.

Erickson, D. J., III, P. J. Rasch, P. P. Tans, P. Friedlingstein, P. Ciais,E. Maier-Reimer, K. Six, C. A. Fischer, and S. Walters (1996), The sea-sonal cycle of atmospheric CO2: A study based on the NCAR CommunityClimate Model (CCM2), J. Geophys. Res., 101, 15,079–15,097.

Global Carbon Project (2003), Science framework and implementation,Rep. 1, 69 pp., Earth Syst. Sci. Partnership (IGBP, IHDP, WCRP,DIVERSITAS), Canberra, A.C.T.

GLOBALVIEW-CO2 (2003), Cooperative Atmospheric Data IntegrationProject: Carbon dioxide [CD-ROM], NOAA Clim. Monit. andDiagnostics Lab., Boulder, Colo.

Gurney, K. R., et al. (2002), Towards robust regional estimates of CO2

sources and sinks using atmospheric transport models, Nature, 415,626–630.

Gurney, K. R., et al. (2004), Transcom 3 inversion intercomparison: Modelmean results for the estimation of seasonal carbon sources and sinks,Global Biogeochem. Cycles, 18, GB1010, doi:10.1029/2003GB002111.

Matricardi, M., F. Chevallier, G. Kelly, and J.-N. Thepaut (2004), Animproved general fast radiative transfer model for the assimilation ofradiance observations, Q. J. R. Meteorol. Soc., 130, 153 – 173,doi:10.1256/qj.02.181.

Matsueda, H., H. Y. Inoue, and M. Ishii (2002), Aircraft observation ofcarbon dioxide at 8-13 km altitude over the western Pacific from 1993 to1999, Tellus, Ser. B, 54, 1–21.

McNally, A. P., and P. D. Watts (2003), A cloud detection algorithm forhigh-spectral-resolution infrared sounders, Q. J. R. Meteorol. Soc., 129,3411–3423, doi:10.1256/qj.02.208.

McNally, A. P., P. D. Watts, J. A. Smith, R. J. Engelen, G. A. Kelly, J.-N.Thepaut, and M. Matricardi (2005), The assimilation of AIRS radiancedata at ECMWF, Q. J. R. Meteorol. Soc., 131, in press.

O’Brien, D. M., and P. J. Rayner (2002), Global observations of the carbonbudget: 2. CO2 column from differential absorption of reflected sunlightin the 1.61 mm band of CO2, J. Geophys. Res., 107(D18), 4354,doi:10.1029/2001JD000617.

Peters, W., et al. (2004), Toward regional-scale modeling using the two-waynested global model TM5: Characterization of transport using SF6,J. Geophys. Res., 109, D19314, doi:10.1029/2004JD005020.

Rayner, P., and D. O’Brien (2001), The utility of remotely sensed CO2

concentration data in surface source inversions, Geophys. Res. Lett.,28, 175–178.

Rodenbeck, C., S. Houweling, M. Gloor, and M. Heimann (2003), Time-dependent atmospheric CO2 inversions based on interannually varyingtracer transport, Tellus, Ser. B, 55, 488–497.

Sadourny, R., andK. Laval (1984), January and July performance of the LMDgeneral circulationmodel, inNewPerspectives inClimateModeling, editedby A. Berger and C. Nicolis, pp. 173–197, Elsevier, New York.

Sawa, Y., et al. (2004), Aircraft observations of CO2, CO, O3 and H2 overthe North Pacific during the PACE-7 campaign, Tellus, Ser. B, 56, 2–20.

��R. J. Engelen and A. P. McNally, European Centre for Medium-Range

Weather Forecasts, Shinfield Park, Reading RG2 9AX, UK. ([email protected])


9 of 9

D18305

5 of 9

Figure 5. Monthly mean analysis results for March 2003, September 2003, and March 2004 as wellas the monthly mean analysis error for March 2003.

D18305 ENGELEN AND McNALLY: CO2 DATA ASSIMILATION D18305

7 of 9

Figure 7. Comparison of CO2 estimates with JAL observations for 20 January 2003 and 20 May 2003.

D18305 ENGELEN AND McNALLY: CO2 DATA ASSIMILATION D18305

estimating atmospheric co from advanced infrared satellite ...citation: engelen, r. j., and a. p....

Documents