www.biogeosciences-discuss.net/8/7475/2011/ doi:10.5194/bgd-8-7475-2011 © Author(s) 2011. CC Attribution 3.0 License.

Abstract. New observations of the vertically integrated CO2 mixing ratio, ⟨CO2⟩, from ground-based remote sensing show that variations in CO2⟩ are primarily determined by large-scale flux patterns. They therefore provide fundamentally different information than observations made within the boundary layer, which reflect the combined influence of large-scale and local fluxes. Observations of both ⟨CO2⟩ and CO2 concentrations in the free troposphere show that large-scale spatial gradients induce synoptic-scale temporal variations in ⟨CO2⟩ in the Northern Hemisphere midlatitudes through horizontal advection. Rather than obscure the signature of surface fluxes on atmospheric CO2, these synoptic-scale variations provide useful information that can be used to reveal the meridional flux distribution. We estimate the meridional gradient in ⟨CO2⟩ from covariations in ⟨CO2⟩ and potential temperature, θ, a dynamical tracer, on synoptic timescales to evaluate surface flux estimates commonly used in carbon cycle models. We find that simulations using Carnegie Ames Stanford Approach (CASA) biospheric fluxes underestimate both the ⟨CO2⟩ seasonal cycle amplitude throughout the Northern Hemisphere midlatitudes and the meridional gradient during the growing season. Simulations using CASA net ecosystem exchange (NEE) with increased and phase-shifted boreal fluxes better fit the observations. Our simulations suggest that climatological mean CASA fluxes underestimate boreal growing season NEE (between 45–65° N) by ~40%. We describe the implications for this large seasonal exchange on inference of the net Northern Hemisphere terrestrial carbon sink.


Introduction
Accurate determination of surface fluxes of carbon dioxide is important both to predict future climate and increasingly to support climate and energy policy.The long-term trend in atmospheric CO 2 depends not only on anthropogenic emissions; natural fluxes of CO 2 to the ocean and terrestrial biosphere control the fraction of anthropogenic CO 2 , presently ∼45 %, that remains in the atmosphere (Gloor et al., 2010).These natural fluxes respond to changes in atmospheric composition and climate on interannual and longer timescales.
Partitioning uptake and release of CO 2 among its terrestrial and oceanic components is necessary to gain insight into the processes controlling these fluxes.
Ideally, self-consistent flux estimates over spatial scales ranging from local to global could be determined from a single set of observations.In reality, no single type of measurement is capable of providing such a constraint.For example, observations of atmospheric O 2 mixing ratio (Keeling et al., 1996) provide an estimate of net ocean and biospheric uptake globally, but provide little or no information about the spatial distribution of fluxes.At local scales, terrestrial biospheric carbon fluxes can be measured quite accurately via inventory (e.g., changes in carbon stocks, Saatchi et al., 2007) and by atmospheric observations (e.g., eddy-covariance measurements, Baldocchi, 2008).Such measurements are, however, representative only of small spatial scales (<10 km 2 ); other data sets such as measures of land color or chlorophyll fluorescence from satellites are needed to scale the local flux estimates to regional and global scales (e.g., Beer et al., 2010;Frankenberg et al., 2011).Likewise, estimation of ocean fluxes from surface pCO 2 measurements requires substantial upscaling of sparse observations.
To infer surface fluxes at intermediate spatial scales (e.g., continental scales), inverse modeling of atmospheric carbon dioxide observations has been used in a variety of approaches.For example, the TransCom experiment estimated continental fluxes at scales of 10 7 km 2 using a suite of transport models, each constrained by measurements of boundary layer CO 2 mixing ratio, [CO 2 ] ( Gurney et al., 2002).Carbon-Tracker optimizes regional fluxes via CO 2 data assimilation (Peters et al., 2007).Like other methods, inverse modeling has its limitations.Although the global network to monitor CO 2 is growing, there are only limited observations of atmospheric CO 2 , most of which are obtained within the planetary boundary layer.The relative contribution of local versus large-scale fluxes to variations in CO 2 varies based on location, time of day, and season, complicating the use of boundary layer CO 2 observations to constrain fluxes over a range of spatial scales.
Inverse modeling is also complicated by a more fundamental limitation.The ability to use surface concentration data to infer fluxes requires accurate modeling of boundary layer height and rates of entrainment between the boundary layer and the free troposphere.Gradients in atmospheric carbon dioxide are influenced by the covariance between the strength of surface fluxes and the strength of vertical mixing (Denning et al., 1995).Primarily as a result of this covariance, studies have shown that inversion results are very sensitive to how vertical mixing is represented in transport models (Stephens et al., 2007).
Total column carbon dioxide, denoted in this paper as CO 2 but denoted as X CO 2 elsewhere, is a new measurement that will complement surface data in inverse studies by bridging the spatial gap between the small-scale flux constraints provided by direct flux observations or boundary layer con-centration data and the global constraints provided by atmospheric O 2 .Because the column integrates not only boundary layer CO 2 but also free tropospheric CO 2 , its footprint is much larger.A companion paper (Keppel-Aleks et al., 2011), hereafter KWS11, shows that simulated CO 2 has a large footprint and is largely insensitive to assumptions about small-scale mixing.In the KWS11 simulations, variations in CO 2 at all timescales are primarily tied to large-scale rather than local fluxes.Diurnal changes in the column are only weakly influenced by local fluxes, while synoptic-scale variations in the column are due primarily to advection across the mean hemispheric-scale north-south gradient.
Here, we use data from the Total Carbon Column Observing Network (TCCON) to test the simulations of KWS11.TCCON is a network of high resolution ground-based Fourier transform spectrometers (FTS) that obtain direct solar absorption spectra (Wunch et al., 2011).We focus our analysis on two diagnostics for variations in the column, described in Sect.2.1.The first is the seasonal cycle amplitude (SCA), which quantifies the seasonal variation in CO 2 at a given location.The second is the north-south gradient in CO 2 , which can be estimated even from sparse ground-based observations using covariations in observed CO 2 and a dynamical tracer such as potential temperature, θ (KWS11).
In this paper, we evaluate the information content in CO 2 observations.Specifically, we compare diurnal variations in TCCON CO 2 data with eddy covariance flux data to determine the contribution of local surface fluxes to variations in CO 2 .We compare temporal variations in CO 2 at synoptic timescales with spatial gradients in free tropospheric CO 2 data from aircraft to understand the imprint of horizontal advection on CO 2 .A comparison of variations in CO 2 to variations in local fluxes (eddy covariance observations) and gradients in the free troposphere (aircraft observations) both confirm that CO 2 variations primarily reflect large-scale flux information.We illustrate how CO 2 observations can be used to constrain global carbon fluxes and address how they can be used in future studies to complement more established observations.

Ground based total column data
We use CO 2 data from TCCON observatories.The first dedicated TCCON observatories became operational in 2004, and the network has subsequently grown.Table 1 lists the sites used in this study, limited to midlatitude sites that have obtained data over at least one complete annual cycle.We summarize the data acquisition and retrieval method; more details are provided in Wunch et al. (2011) and references therein.
Briefly, TCCON observatories use FTS to record high resolution spectra (0.02 cm −1 ) in the near infrared (3800-15 500 cm −1 ).A profile scaling retrieval is used to determine the vertical CO 2 column from observed spectra.A priori CO 2 profiles are generated using an analytical function that varies seasonally and meridionally based on GLOBALVIEW CO 2 in the troposphere and applies the age-of-air measurements of Andrews et al. (2001) in the stratosphere.Temperature, water vapor profiles, and tropopause pressure from NCEP reanalysis (Kalnay et al., 1996) are used to map the assumed profile onto 71 vertical layers at 1 km resolution.Gases, including CO 2 , O 2 , and H 2 O are retrieved simultaneously in selected spectral windows.The CO 2 retrieval scales the a priori profile to best match the recorded spectrum in two spectral windows centered at 6220 cm −1 (80 cm −1 wide) and 6339 cm −1 (85 cm −1 wide).The retrieved column abundance [mol m −2 ] is normalized by the retrieved column abundance of O 2 to yield CO 2 [ppm].
TCCON data have been rigorously calibrated against aircraft profiles of CO 2 and are traceable to World Meteorological Organization standards (Wunch et al., 2010).After calibration, the accuracy and precision are both better than 0.2 % (0.8 ppm in CO 2 ).In the analysis presented here, we have used simultaneous measurements of CO to remove the local fossil component of observed CO 2 for data obtained at the Pasadena site, in the Los Angeles urban area (Wunch et al., 2009).Such a localized source will not be captured by the simulations described in this paper, which were run at moderate resolution (2 • by 2.5 • ).More details are provided in Appendix A. At the other sites, CO 2 and CO are not well correlated, so no empirical correction was applied.
Figure 1 shows the daily median CO 2 at the sites used in this study.We include all observations for which the retrieval uncertainty is <1.5 ppm.These timeseries show variability at several timescales: a long-term secular increase, seasonal variations, particularly in the Northern Hemisphere, and day-to-day variations.There is also substantial interannual variability in the SCA, which is discussed in Sect.3.4.
Because the column data are sparse and not all the observatories have operated simultaneously, we focus on a statistical comparison of the data with simulations.For each site, we subtract the linear trend in CO 2 at Lauder, New Zealand, a Southern Hemisphere site with small seasonal and interannual variability (Fig. 1).Thus, the detrended data retain the absolute CO 2 offset among sites.We then compare the seasonal variation in CO 2 by computing the peak-totrough seasonal cycle amplitude as the difference between maximum monthly median CO 2 and the minimum monthly median CO 2 in the preceding growing season.We estimate the north-south gradient across TCCON sites using the covariation between CO 2 and θ at 700 hPa, a dynamical tracer, as described in KWS11 Sect.3.2.Here we provide a brief rationale for this method, refering readers to KWS11 for a more complete description.The vertically integrated CO 2 and θ at 700 hPa observed at TCCON stations covary on synoptic timescales because large-scale eddy activity in the atmosphere advects both tracer fields.We can use θ to estimate the meridional distance by which an airmass was displaced, because θ is conserved following adiabatic motion, and then infer the north-south gradient in CO 2 from the estimated meridional displacement and temporal fluctuations in CO 2 .Although diabatic effects will modify θ and local fluxes will affect variations in CO 2 , it is possible to infer the large-scale meridional gradient in CO 2 given adequate data using this method (KWS11).
The daily mean CO 2 timeseries from TCCON sites are gap-filled using a cubic spline before we apply a 3 to 21 day bandpass filter to isolate synoptic-scale variations in CO 2 .This removes seasonal variations at frequencies lower than 21 days and mesocale variations at frequency higher than three days and so, among the dynamically generated CO 2 variations, isolates those generated by large-scale eddies.Our results are not sensitive to the exact filter cutoffs; using a 3-14 day filter changes the results by only 1-2 %, as does changing the high frequency cutoff by one day.We denote the bandpass-filtered variations as CO 2 .We use NCEP reanalysis data to calculate bandpass filtered variations and the meridional θ gradient at each site to estimate the meridional displacement, L, of observed air parcels from a reference latitude with monthly mean potential temperature θ 0 : We then use the meridional displacement and variations in CO 2 to calculate the meridional gradient, ∂ y CO 2 [ppb km −1 ]: Here, we use the overhat to indicate that the gradient is an estimated, rather than directly observed, quantity.We calculate ∂ y CO 2 only where there is more than two years of data from a single site.Additionally, direct calculation of ∂ y CO 2 using this method only works well in the midlatitudes where eddy activity is substantial and significant CO 2 and θ gradients exist (KWS11).We can, however, use θ to estimate meridional displacements everywhere outside the tropics.Because the mean westerly wind in the midlatitudes is 8.5 m s −1 during the summer, during one day, the column is influenced by airmasses originating more than 700 km upwind.Because the zonal winds are so strong, it is difficult not to alias transport error when comparing observations to a model.The use of a dynamically informed, rather than simply a geographic coordinate, facilitates intercomparison between data and models (KWS11).Therefore, we compare CO 2 at various TCCON sites as a function of the estimated meridional displacement.However, we only calculate ∂ y CO 2 directly from the observations at Park Falls and Lamont because of these methodological and data constraints.

Net ecosystem exchange estimated from column drawdown
We examine diurnal variations in CO 2 by computing the mass change in the total column over the course of a single day.Because we only obtain measurements during daylight, we define diurnal variations as the daytime drawdown in CO 2 .The vertical column of any gas, VC G [mol m −2 ], is the product of the dry air, vertically averaged mixing ratio of the gas, G , and the mass of the atmospheric column as expressed in Eq. ( 3): where VC air is the vertical column of dry air, P s is the surface pressure, P q is the vertically integrated contribution of water to the surface pressure, g the gravitational acceleration, and M air the molecular mass of air.We differentiate Eq. ( 3), assuming the vertical column mass of air is constant, to determine the time rate of change in VC CO 2 attributable to changes in the vertically averaged mixing ratio, CO 2 : In Eq. ( 4), dVC CO 2 dt has units of flux [mol m −2 s −1 ].The column-averaged mixing ratio, CO 2 , may change due either to local surface fluxes, to changes in the tropopause height, or to advection of CO 2 .If we neglect the influence of advection (an assumption that will turn out to be untenable on short timescales), we can assess the extent to which variations in CO 2 capture local flux signals.Substituting VC air from Eq. ( 3) with G = O 2 into Eq.( 4), we calculate NEE from TCCON observations as: We calculate CO 2 as the four hour difference in CO 2 centered around solar noon.This method of calculating CO 2 t maximizes the number of days with adequate temporal coverage; it yields similar results to fitting a line through the data obtained on a single day.VC O 2 is averaged from retrievals within two hours on either side of solar noon, and we use a constant value of 0.20939 for O 2 (Tohjima et al., 2005).Although the atmospheric concentration of O 2 is decreasing by ∼4 ppm yr −1 (Manning and Keeling, 2006), this change has a negligible impact on our estimate of NEE.Equation ( 5) neglects the change in VC O 2 , which is at most 0.5 % on four hour timescales.
We test how well NEE inferred from column drawdown represents local biospheric fluxes by comparing Park Falls NEE to eddy covariance flux observations at the same location (discussed in Sect.2.3).Because the column retrieval is not equally sensitive to all levels of the atmosphere, we must incorporate the averaging kernel of the FTS into the comparison (see Appendix B for more details).Local drawdown impacts the CO 2 mixing ratio within the boundary layer.The column observations are less sensitive to CO 2 within the boundary layer than to CO 2 in the free troposphere (i.e., the surface averaging kernel at ±2 h of solar noon ranges from 0.7-0.9 over a seasonal cycle, while the averaging kernel within the free troposphere is close to unity).To compare column and flux observations, we damp the observed eddy covariance flux by the surface value of the column averaging kernel, as this is the signature of local fluxes that would be imparted on the observed column.This method is similar to the approach used by Wunch et al. (2009) to analyze diurnal variations in CO 2 and other greenhouse gases in the Los Angeles basin.

Eddy covariance flux data
Fluxes of CO 2 , water vapor, virtual temperature, and momentum are measured continuously at 30, 122, and 396 m on the WLEF tower in Park Falls.CO 2 and water vapor mixing ratios are measured using infrared gas analyzers (Li-COR Inc., Lincoln, Nebraska, model LI-6262).Virtual temperature and momentum are measured using sonic anemometers (Applied Technologies Inc., Boulder, Colorado, model SATI/3K or Campbell Scientific Inc., Logan, Utah, model CSAT3).High-precision, high-accuracy CO 2 mixing ratio measurements are simultaneously recorded at 30, 122, and 396 m using Li-COR model LI-6251and LI-7000 infrared gas analyzers (Bakwin et al., 1998;Zhao et al., 1997).These measurements are used to calibrate the fast-response infrared gas analyzers and to calculate the rate of change of storage for CO 2 (Davis et al., 2003).The eddy covariance instrumentation and methodology is described in detail in Berger et al. (2001) and Davis et al. (2003).Davis et al. (2003) developed an algorithm for preferred NEE that combines measurements at the three levels to maintain a large footprint while avoiding the influence of the clearing around the tower.The footprint of the midday flux measurements within the convective boundary layer is discussed by Wang et al. (2006), and surface layer flux footprints are discussed by many authors (e.g., Horst and Weil, 1992).Flux uncertainties due to changes in the level of the flux measurement and random errors were discussed by Ricciuto et al. (2008).We determine daily average NEE within two hours of solar noon to compare with drawdown in the column.We use data obtained between 2004 and 2009, only on days with both column and eddy covariance data.

Aircraft data
Free tropospheric [CO 2 ] data from two field campaigns that had extensive geographic coverage are used in this analysis.Free tropospheric [CO 2 ] data were obtained in situ during the Intercontinental Chemical Transport Experiment (INTEX-NA), a NASA-sponsored aircraft mission in July-August 2004 (Choi et al., 2008).Data were obtained between 28 • and 54 • N along the flight tracks shown in Fig. 2. Carbon dioxide mixing ratios were determined using a modified LI-COR model 6252 non-dispersive infrared gas analyzer.This dual-cell instrument achieves high precision by measuring the differential absorption between sample air and a calibrated reference gas that is traceable to the WMO primary CO 2 standards.The LI-COR-based CO 2 sampling system was operated at constant pressure (333 hPa) and temperature (40 • C), and had a precision of 0.07 ppm and accuracy of 0.20 ppm (1 s/1-σ precision, 1-σ accuracy).Experimental procedures are described in detail by Vay et al. (2003).
Free tropospheric [CO 2 ] was measured during the first HI-APER Pole-to-Pole Observations (HIPPO 1) campaign in January 2009 over the Pacific (Wofsy et al., 2011).[CO 2 ] was measured between 80 • N and 67 • S using three instruments, two Li-COR non-dispersive infrared gas analyzers as well as a quantum cascade laser system (QCLS) and two flask samplers.In this study, we use only QCLS data, which has precision of 0.02 ppm and accuracy of 0.1 ppm (1 s/1-σ precision, 1-σ accuracy).During HIPPO 1, [CO 2 ] was sampled in the free troposphere over a series of vertical profiles.We limit our analysis to those profiles that exceed 500 hPa in vertical extent and have a pressure floor of at least 900 hPa.We append stratospheric CO 2 profiles to the observed vertical profiles based on the TCCON a priori CO 2 profile for the sampled locations, and then integrate these profiles to determine the column mixing ratio.The locations of the profiles obtained in the Northern Hemisphere are shown in Fig. 3.
For both campaigns, the cited accuracy is relative to the WMO standards.It is worth noting that actual limits to accuracy are in the form of non-random, systematic biases.Expression of accuracy as a 1-σ value represents a judgment based on intercomparison activities that these biases are less than or equal to the given values 68 % of the time.During both INTEX-NA and HIPPO 1, profile measurements were obtained at TCCON sites.These profiles are among the data used to calibrate TCCON CO 2 data (Washenfelder et al., 2006;Wunch et al., 2010).

Simulations with a transport model
We compare variations in observed CO 2 against simulations from KWS11.Briefly, these simulations use the AM2 general circulation model (GCM) of NOAA's Geophysical Fluid Dynamics Laboratory, a free-running GCM with prescribed sea surface temperatures (Anderson et al., 2004).We run the model at 2 Fig. 3. HIPPO partial CO 2 columns overlaid on January monthly mean potential temperature, θ .The inner circle is colored by CO 2 , while the outer square is colored by θ at 700 hPa at the time the profile was acquired.
25 vertical levels.As land surface fluxes in the GCM, we use monthly NEE (balanced to give zero annual mean) from the climatological mean Carnegie Ames Stanford Approach (CASA) simulation (Randerson et al., 1997).These are the same CASA fluxes as in Olsen and Randerson (2004) with NEE representing the residual of monthly net primary production (NPP) and respiration fluxes.NEE has been redistributed at 3 h resolution based on the year 2000 meteorology to capture diurnal and synoptic flux patterns, but the monthly mean flux at any grid box is equal to the climatological mean.These fluxes provide a good estimate of the expected patterns in atmospheric CO 2 in an average year, even though our simulations do not correspond to any given year.CASA biospheric fluxes are commonly used by the carbon cycle community and are the basis for prior fluxes in many inversion simulations including TransCom (Gurney et al., 2002) and CarbonTracker (Peters et al., 2007), which uses CASA-GFEDv2 for its prior (van der Werf et al., 2006).
Ocean surface fluxes are prescribed as monthly-mean fluxes derived from surface ocean pCO 2 data (Takahashi et al., 2002).These ocean fluxes represent an annual and global mean sink of atmospheric CO 2 of ∼1.4 Pg C yr −1 .A more recent version of these fluxes is available, but our results are not sensitive to variations in oceanic fluxes.We use annual mean fossil emissions with net global emissions of ∼5.5 Pg C yr −1 (Andres et al., 1996).These emissions are determined from self-reported fuel consumption at the national level (Marland and Rotty, 1984) and are converted to regional fluxes proportional to local population density.We also use updated fossil fluxes determined at monthly time resolution from 2004-2007(Gregg and Andres, 2008;Gregg et al., 2009;Andres et al., 2011) to test the sensitivity of the column to these larger and temporally varying parameterization of the fossil fluxes.
We compute the SCA and the estimated gradient ∂ y CO 2 for the model gridboxes corresponding to TCCON sites to compare with the observations, incorporating the FTS averaging kernel (Appendix B).
As shown in KWS11, biospheric fluxes impart the largest variations in CO 2 .We therefore focus our comparisons of observed CO 2 against three simulations with variations in biospheric fluxes: CASA exchange, boreal amplified CASA exchange, and boreal amplified CASA exchange with an early growing season onset as described in KWS11.In the boreal amplified simulations, NEE is amplified by a Gaussian weighting function centered at 55 • N with a peak value of 1.5 that decays to unity over a standard deviation of 10 • .Integrated over 45-65 • N, NEE is increased by 40 %; integrated over the Northern Hemisphere, NEE is increased by 30 %.In the boreal amplified simulations with early growing season onset, we use NEE with the boreal amplification described above, but add July NEE to May NEE between 50-60 • N, which changes the sign of May NEE, thus converting a net source to a net sink and moving the onset of the growing season earlier by one month.We focus on sensitivity simulations with perturbed boreal fluxes because the large seasonality of boreal NEE has a large imprint on the seasonal cycle in CO 2 (KWS11, Randerson et al., 1997); the KWS simulations show that changing NEE in the subtropics or midlatitudes has a small effect on the SCA of CO 2 . To

Diurnal variations
Diurnal variations in observed CO 2 are small and influenced by a large spatial footprint.At Park Falls, the change in CO 2 during the summer growing season is ∼2 ppm within a given day (Fig. 4).Although this change is small in comparison to diurnal variations at the surface, it is larger than the change predicted in the total column predicted by the simulations in Olsen and Randerson (2004).NEE inferred from column drawdown (Eq.5) shows little correlation with NEE inferred from eddy covariance on daily timescales (Fig. 5a).A regression of NEE from column drawdown against NEE from eddy covariance on individual days yields a slope of 0.9 ± 0.3, an intercept of −5 ± 2 µmol m −2 s −1 , and an R 2 of 0.11.Consistent with this large scatter, KWS11 show that  Although the regional flux signal is not apparent in the total column variations on daily timescales, it does become more apparent when we aggregate NEE on weekly and monthly timescales (Fig. 5b, c).We use a fit that accounts for the standard error of the mean in both eddy covariance NEE and column NEE over the aggregated time period (York et al., 2004).The slopes are 1.5 ± 0.2 for the weekly mean data and 1.5 ± 0.2 for monthly mean data, and the fits have a negative intercept of −4.9 ± 0.8 µmol m −2 s −1 for weekly mean data and −4.3 ± 0.7 µmol m −2 s −1 for monthly mean data.The weighted R 2 values of 0.94 and 0.86 for weekly and monthly data account for the standard error of the mean in the datasets and are much improved relative to the daily comparison.The larger NEE inferred from the FTS column may be attributable to column drawdown having a larger footprint.Regional studies have shown divergent estimates for NEE in the area around Park Falls using several bottom-up and topdown methods (Desai et al., 2010).We have confirmed that the offset between NEE from column drawdown and from eddy covariance is not due to seasonal uptake by removing the seasonal trend from CO 2 before determining the diurnal drawdown in the column.Neither the slope nor the intercept are significantly affected.
While horizontal and vertical advection make a negligible contribution to daytime NEE determined from eddy covariance (Wang et al., 2005;Yi et al., 2000), our results confirm that horizontal advection does have a large influence on variations in the column even over four hours.We attribute the poor comparison of column drawdown and local NEE on daily timescales to the neglect of advection in Eq. ( 5). on 15 July 2004 during INTEX-NA.Although the two profiles were obtained only three hours apart, a frontal system moved through the region and the free tropospheric [CO 2 ] above 5 km increased by ∼5 ppm during this time.The drawdown of boundary layer [CO 2 ] over the same three hours is much smaller.Only when we aggregate column drawdown and eddy covariance fluxes over several synoptic cycles is the influence of the large synoptic variability in free tropospheric [CO 2 ] reduced sufficiently to reveal a robust relationship between column drawdown and local eddy covariance fluxes (Fig. 5b-c).We have investigated whether there is a relationship between the mismatch in NEE derived from the two measurements and meteorological variables such as wind speed, wind direction, and diurnal changes in pressure and temperature.We do not find correlations between diagnostics for advection and mismatch in NEE.Modelling tools such as back trajectory analysis originating from multiple levels in the column may prove helpful in diagnosing changes in CO 2 on sub-diurnal timescales, but they require accurate background CO 2 distributions to be useful.In summary, the comparison between drawdown in CO 2 and eddy covariance flux confirms that while regional information is contained in column observations, these regional flux signals are obscured by larger-scale variations in CO 2 even on short timescales.

Horizontal gradients as sources of synoptic variations in CO 2
Advection leaves a notable signature on CO 2 at seasonal and shorter timescales.Figure 7 illustrates the relationship between synoptic-scale variability and north-south gradients in CO 2 at Park Falls and Lamont, two North American TCCON sites that are separated by 9 • latitude.Outside the growing season, the two timeseries essentially overlap, indicative of a small meridional CO 2 contrast between the two sites.At the onset of the growing season (May), the two timeseries diverge; the larger decrease in CO 2 at Park Falls results in a significant mean contrast (∼3 ppm) between the two sites.During the period of divergence, the variability at each site increases noticeably.This variability is a direct re- Park Falls Lamont Fig. 7. Fifteen months of daily median CO 2 at Park Falls and Lamont.We see the impact of large-scale gradients in these two timeseries: the records are similar outside the growing season (when north-south gradient in CO 2 is small) and the records diverge during the growing season (June-August) when CO 2 is drawn down in the north.When the north-south gradient is larger, synoptic scale variations in daily median CO 2 are larger, reflecting the fact that variations arise from advection of gradients.
sult of the increased meridional gradient and weather systems advecting air to each site from the north and south.Although the median CO 2 is quite different at the two sites during the growing season, there is overlap within the observed variability: higher CO 2 values at Park Falls are similar the lower CO 2 values at Lamont.We exploit the variability in daily median CO 2 and free tropospheric potential temperature, θ, to infer a north-south gradient representative of large spatial scales.Figure 8 shows daily median CO 2 at Park Falls, Wisconsin for one month in summer 2006.Day-to-day variations are large (∼5 ppm within one week), approximately half the amplitude of the seasonal cycle.These variations correlate with changes in θ at 700 hPa.Cold air (low θ) from the north has lower CO 2 ; warm air (high θ) from the south has higher CO 2 .During the growing season, approximately one-third of the variations in CO 2 at Park Falls can be accounted for by variations in θ (R 2 = 0.34).We estimate Northern Hemisphere summer ∂ y CO 2 at Park Falls to be 1.1 ppb km −1 and at Lamont to be 1.3 ppb km −1 (Table 2).In Fig. 9, we show TCCON CO 2 in July and August as a function of meridional displacement relative to Park Falls.A negative meridional displacement signifies that an airmass originated south of Park Falls, and a positive displacement that the airmass originated north of Park Falls.Variations in CO 2 , CO 2 , are determined relative to annual mean CO 2 at Lauder.Negative CO 2 values, such as those at Park Falls and Bialystok, indicate that during Northern Hemisphere summer, CO 2 is consistently lower than at Lauder.The CO 2 values at Lamont are centered around zero, indicating that summer mean CO 2 is similar at Lamont and Lauder.The range of CO 2 sampled at sites between 34 • and 53 • N is ∼9 ppm corresponding to meridional displacements between 3000 km south and 1500 km north of Park Falls at 46 • N.

Free tropospheric data from aircraft
Free tropospheric [CO 2 ] between 350-800 hPa measured during the INTEX-NA mission show a similar correlation with θ.We obtained coincident 700 hPa θ from the NCEP North American Regional Reanalysis (Mesinger et al., 2006) by interpolation along the flight tracks, which yields the same results as using θ sampled on the aircraft interpolated to 700 hPa.The in situ data show a positive correlation of CO 2 with θ throughout the free troposphere (Fig. 10a).The correlation between [CO 2 ] and θ is much stronger than that between CO 2 and latitude because synoptic systems passing through the sampled regions obscure the mean north-south gradient (Fig. 10b).
We can quantitatively estimate the north-south gradient in CO 2 from INTEX-NA data using the same approach we use for column data.We designate three free tropospheric layers between 350-800 hPa and a fourth bin for the boundary layer (between the surface and 800 hPa).We recenter [CO 2 ] and θ by the mean value within each pressure bin and fit a line to relate variations in [CO 2 ] to variations in θ.The resulting slopes are weighted by the atmospheric mass in each pressure bin and summed.We then normalize by the meridional gradient in θ at 45 • N to determine the meridional gradient in CO 2 .The resulting estimate of north-south gradient agrees very well with the gradient estimated at Park Falls and Lamont (Table 2).
Analysis of Northern Hemisphere HIPPO data obtained in January 2009, primarily over the north Pacific, show a northsouth gradient that has reversed sign relative to the summer gradient.In Fig. 3, we map January 2009 mean 700 hPa θ contours from NCEP reanalysis, and overlay HIPPO partial column CO 2 with θ at the time of the observation.The observed CO 2 and θ are generally anti-correlated, as expected for winter when respiration dominates photosynthesis at high northern latitudes, setting up a positive gradient in biospheric CO 2 .The calculated gradient is −0.2 ± 0.1 ppb km  In summary, these aircraft data demonstrate that the covariation between TCCON CO 2 and θ reflects large-scale spatial gradients.Free tropospheric CO 2 observed during aircraft campaigns can be used to directly calculate spatial gradients, but such campaigns are of limited duration and spatial extent.

Comparison with models
The SCA in CO 2 is underestimated throughout the Northern Hemisphere in AM2 with CASA NEE as a boundary condition.Table 1 shows that the SCA at all four Northern Hemisphere TCCON sites is underestimated in CASA simulations by 25-35 %, a finding consistent with Yang et al. (2007) and Nakatsuka and Maksyutov (2009).There is substantial interannual variability in the SCA at TCCON sites with multiple years of data.We report the median SCA and standard deviation for Park Falls, Lamont, and Lauder.For Bialystok and Pasadena, the data record contains only one complete seasonal cycle.We have averaged five seasonal cycles from AM2 to estimate a climatological mean SCA for each site.Our simulations ignore interannual variability in fluxes, but they show that for an average year, CASA underestimates column SCA in Northern Hemisphere midlatitudes.Figure 12a shows the seasonal cycle in observed CO 2 at four Northern Hemisphere TCCON sites compared to CO 2 from AM2 with underlying CASA biospheric and standard ocean and fossil fluxes.At all Northern Hemisphere sites, the amplitude is smaller in AM2 than in the observations; the y-axis reflects the offset from annual mean Lauder CO 2 .We have investigated whether sampling biases could account for the discrepancy between the observed and simulated seasonal cycle amplitudes.Although the TCCON data have been calibrated, we expect that any bias between the sites, or at a single site over a seasonal cycle, would be primarily attributable to retrieval bias associated with the solar zenith angle at which the observations are obtained.To estimate the potential error due to varying solar zenith angle, we average TCCON data obtained only between 60 • -70 • solar zenith angle.The SCA values change by at most 0.1 ppm at all sites.This suggests that biases in the estimated SCA due to observational biases are small.
We have also investigated the potential bias in the comparison due to the fact that we use daily mean AM2 output to calculate SCA.If instead we analyze hourly CO 2 output for one model year, SCA changes only by −0.1 to 0.2 ppm using daylight-only CO 2 , a small difference relative to the disagreement between the simulated and observed seasonal cycle amplitude at TCCON sites.We also tested the potential bias in the comparison due to the fact that TCCON data are only obtained when the sun is not obscured by clouds.We limit sampling in AM2 to days in which the cloud fraction in the 2 • by 2.5 • gridbox corresponding to each TCCON site is less than 70 %.This cloud filter removes ∼80 % of days in winter and ∼25 % of days in the summer at Park Falls.Removing cloud-contaminated days decreases the SCA by ∼5 % at Park Falls and Lamont, and has little effect at Bialystok and Pasadena, suggesting that sampling the model only under cloud-free conditions would increase the disagreement between AM2 with CASA fluxes and the observations.
The north-south gradient in CO 2 during summer is underestimated by ∼45 % with CASA fluxes.This is evident from ∂ y CO 2 estimated at Park Falls or along INTEX tracks (Table 2), or simply by inspecting the CO 2 contrast between TCCON sites.At the gridbox corresponding to Park Falls, ∂ y CO 2 in AM2 is only 0.6 ppb km −1 , compared with 1.1 ppb km −1 determined from the data.For Lamont, ∂ y CO 2 is only 0.4 ppb km −1 , in contrast to 1.3 ppb km −1 determined from the data.When sampling in AM2 is limited to days with less than 70 % cloud cover, the estimated gradients decrease by ∼10 %.Likewise, the gradient estimated from free tropospheric [CO 2 ] sampled in the model along weaker than observed by 35 %.The north-south contrast across TCCON sites is also too small, as is evident in Fig. 9.
The large circles represent summer mean CO 2 sampled in AM2.While the mean Pasadena and Lamont CO 2 generally agree with the observations, the simulated Northern Hemisphere summer mean CO 2 is too high farther north.
We tested whether the use of 1990 annual mean fossil fuel CO 2 emissions could account for the discrepancy by running AM2 with 2004-2007 monthly mean fossil fuel emissions (Andres et al., 2011).At all four Northern Hemisphere TCCON sites, including urban Pasadena, the SCA changes by less than 0.1 ppm when accounting for seasonality in the fluxes.Fossil fuel emissions in 2004-2007 show a more pronounced subtropical source than do 1990 emissions estimates, but the change in the north-south gradient in the column due to updated fluxes is small (<0.04 ppb km −1 at Park Falls).
Simulations with enhanced boreal fluxes better match CO 2 observations.Figure 12b-c show weekly median timeseries for TCCON data compared to AM2 simulations using boreal-amplified CASA fluxes.The two simulations with amplified boreal fluxes have increased SCAs that bracket the observed column SCAs (Table 1).The simulation with an early onset to the growing season has a growing season net flux somewhat greater than 40 % of the CASA net flux in boreal regions, so overestimation of the SCA is expected.Nevertheless, this simulation better matches the phasing of the observations (Fig. 12b-c), particularly at higher latitude sites.Even sites that are not affected locally by the flux enhancement (i.e., Lamont at 34 • N) show an increase in the modeled SCA with the boreal amplified fluxes.
The dominant role of the high latitude biosphere in determining the SCA in the subtropics has been noted before (Randerson et al., 1997).There is still a mismatch between simulated and observed CO 2 at Pasadena due to AM2 having a different mean CO 2 contrast between Lauder and Pasadena, indicating that the model may be placing Pasadena further into the subtropics dynamically than is realistic.
As the seasonal cycle amplitude increases due to greater summer drawdown in boreal regions in the enhanced simulations, the north-south gradient across TCCON sites grows correspondingly larger (Fig. 9).Increasing boreal uptake during summer has little effect on the mean CO 2 at Pasadena but reduces mean CO 2 at Bialystok such that the contrast is in better agreement with the data.At Park Falls, ∂ y CO 2 increases in the boreal-enhanced and borealextended simulations to 0.9 and 1.1 ppb km −1 respectively, in much better agreement with the observations.At Lamont ∂ y CO 2 doubles to 0.7 ppb km −1 , but is still too small by 50 % compared to observations.We caution that Lamont may lie too far in the dynamical subtropics in AM2 for accurate calculation of ∂ y CO 2 .

Large seasonal flux = small net land sink?
There is no geobiological reason that seasonality of ecosystem exchange will necessarily have any relationship to the net uptake of carbon by the terrestrial biosphere.Nevertheless, Gurney et al. (2004) have shown a strong negative correlation between the retrieved seasonality and net flux in the TransCom inversions: models that infer large flux seasonality typically infer weak net uptake by the terrestrial biosphere, while those with small flux seasonality infer strong net uptake.This correlation is tied to differences in the transport implementation among models, illustrating the challenge of inferring mass fluxes from concentration data obtained at the surface.In inversions, fluxes are determined by optimizing the mass flux necessary to reproduce the simulated mass gradient.The simulated mass gradient is, however, only constrained by observed variations in boundary layer [CO 2 ].When constrained by surface concentration measurements, models that have stronger vertical mixing will yield larger simulated column mass gradients because the observed anomaly at the surface extends deeper into the atmospheric interior whereas models that more slowly advect the concentration anomalies produced by surface fluxes will infer smaller column mass gradients.
So why in the TransCom inversions are the estimates of the seasonality and net flux correlated?In these inversions, the global net uptake of carbon is highly constrained by the observed global growth rate of CO 2 and the imposed fossil and land use-change fluxes of CO 2 to the atmosphere.The decomposition of the global net flux into oceanic and terrestrial components is ultimately tied to the inferred column gradient in CO 2 between the Northern and Southern Hemispheres: a small simulated north-south gradient implies that sinks in the Northern Hemisphere counteract fossil emissions within the Northern Hemisphere.This explains the correlation between the net land flux and the seasonal flux: the models with strong vertical mixing produce larger seasonal fluxes because the expression of these fluxes at the surface is minimized while the expression of these fluxes on the column mass is significant.Such models will also project a larger interhemispheric fossil fuel CO 2 column gradient from the observed variations at the surface because these variations extend into the free troposphere.Thus, the inferred Northern Hemisphere net sink is smaller in models with stronger vertical mixing, which pushes the inversion toward a more globally distributed net sink (e.g., in the oceans rather than on land).
The above discussion illustrates the importance of having improved mass constraints for inverse modeling.Densely sampled aircraft profiles provide one such constraint (Stephens et al., 2007;Nakatsuka and Maksyutov, 2009).We show here that measurements of CO 2 provide another such improvement.In this study, we show using column data that the seasonal fluxes are larger than previously estimated from surface data.In fact, the column derived land flux amplitude (35 Pg C yr −1 ) is larger than that from any of the the TransCom inversions (Median: ∼26 Pg C yr −1 ; Range: 21-31 Pg C yr −1 ).The prediction is that because the TransCom models that came closest to the true value (the models with strongest vertical mixing) yielded the smallest net land sink in the Northern Hemisphere, an inversion using additional constraints from the column will likewise produce small estimates for the net Northern Hemisphere land sink.This remains to be confirmed in a proper inversion of both data types.
Reanalysis fluxes from CarbonTracker also suggest greater growing season uptake in the boreal forests than CASA fluxes.Figure 13 shows the zonal mean June-August NEE from CASA and 2008 CarbonTracker biospheric fluxes (Peters et al., 2007).NOAA's CarbonTracker data assimilation system uses CASA-GFEDv2 fluxes as its prior NEE estimate (van der Werf et al., 2006).By assimilating surface and tower [CO 2 ] data, CarbonTracker increases growing season NEE, with most of the increase in boreal ecosystems.Compared to the CASA NEE used in AM2 simulations, CarbonTracker NEE is greater by ∼25 %, integrated between 45 and 65 • N for June through August (and greater by ∼15 % relative to the prior CASA-GFEDv2 NEE), with a seasonal flux amplitude of 30 Pg C yr −1 compared to 27 Pg C yr −1 .This high latitude increase to the growing season flux leads to better agreement in ∂ y CO 2 estimated in midlatitudes and SCAs closer but still less than the SCAs in observations and in AM2 simulations with our boreal flux amplification (Table 1).With respect to the partitioning of the net CO 2 sink, CarbonTracker yields a net Northern Hemisphere land sink since 2000 of ∼2 Pg C yr −1 , approximately one-third of the global (land + ocean) carbon sink.This represents a significant reduction in the contribution of Northern Hemisphere land compared to Gurney et al. (2004), where the mean Northern Hemisphere land sink was 2.4 Pg C yr −1 relative to a global sink of 2.8 Pg C yr −1 for the period from 1992-1996.

Conclusions
Consistent with results from KWS11, TCCON data from Park Falls show that even on hourly timescales, changes in the column are dominated by non-local effects.Diurnal variations in CO 2 show a relationship with the underlying flux inferred from eddy covariance observations only after averaging over several synoptic cycles.
Midlatitude column CO 2 data and its correlation with synoptic-scale variations in θ can be used to estimate the large-scale gradient in CO 2 .We estimate the north-south gradient during the growing season at Park Falls to be 1.1 ppb km −1 , equivalent to a ∼4 ppm contrast across 30 • of latitude.Free tropospheric [CO 2 ] data obtained during INTEX-NA are consistent with this meridional gradient.We use θ to define a meridional displacement scale and show that variations in CO 2 sampled at Northern Hemisphere midlatitude TCCON sites are primarily driven by advection of airmasses that originate north and south of the observatories.Rather than obscure flux signatures, these synoptic scale variations reveal the meridional gradient in surface fluxes.
While variations in CO 2 primarily reflect large scale gradients, variations in boundary layer [CO 2 ] reflect both local fluxes and entrainment of free tropospheric CO 2 into the boundary layer.Therefore, it is possible to alias the largescale component of boundary layer variability into local surface fluxes when attempting to optimize surface fluxes based only on boundary layer observations.Our findings suggest that by combining CO 2 and boundary layer CO 2 observations, we can properly attribute variability to local or largescale influences based on the correlation of CO 2 with θ.We expect that more robust flux estimates can be inferred using this approach.
TCCON data suggest that Northern Hemisphere biospheric exchange is underestimated by the climatologicalmean CASA net ecosystem exchange used in our simulations.From the SCA in CO 2 at the Park Falls, Wisconsin TCCON site, Yang et al. (2007) concluded that Northern Hemisphere growing season net flux is underestimated by ∼30 % in the mean of TransCom inversions.
Here, we use additional TCCON data from other sites to show that this NEE underestimation is traceable to boreal regions.Simulations presented in KWS11, in which CASA NEE was increased in the subtropics and midlatitudes, could not simultaneously rectify the SCA and the north-south gradient in CO 2 .In this paper, we show that the SCAs at four Northern Hemisphere TCCON sites are consistent with a boreal NEE integrated over 45-65 • N that is increased by 40 %.Because column SCA is influenced by hemisphericscale fluxes, we can simultaneously improve the SCA fit at all four sites through boreal flux amplification.We also find that CO 2 simulated using CASA NEE with amplified boreal fluxes better matches the contrast in CO 2 across the Northern Hemisphere TCCON sites.The estimated gradient ∂ y CO 2 at Park Falls is misrepresented by 45 % using CASA fluxes.Both ∂ y CO 2 at Park Falls and Lamont and from INTEX data are better represented using boreal amplified fluxes.
While CASA is only one biospheric flux model, and we have not examined its interannual flux variability in the simulations described in this paper, it is quite commonly used for biospheric and atmospheric science and is related to many other biospheric flux models.An underestimation of terrestrial exchange at high latitudes points to significant error in the description of boreal ecosystem fluxes and suggests further research is necessary to determine the representativeness of available flux data.In terms of future atmospheric CO 2 levels, understanding boreal ecosystem functioning is important, as climate change is amplified in this region (Holland and Bitz, 2003).
Results from atmospheric inversion studies using surface data show that the net Northern Hemisphere terrestrial carbon sink is sensitive to the gross seasonal fluxes (Gurney et al., 2004).The correlation between seasonal land flux amplitude and net terrestrial uptake in TransCom inversion results underscores the utility of column data as a constraint on the mass gradient in atmospheric CO 2 .The large seasonality in Northern Hemisphere NEE necessary to account for variations in CO 2 data suggest that the Northern Hemisphere terrestrial biosphere may provide only a small net sink for anthropogenic CO 2 .In this paper, we demonstrate that the statistics of variations in CO 2 can provide important insights into the distribution of seasonal fluxes, but we anticipate results from a full inversion of column and boundary layer CO 2 data will provide robust constraints on both seasonal and net fluxes.
weight the eddy covariance observations by the surface averaging kernel.For comparison with the model output, where we integrate the CO 2 mixing ratio at 25 levels to determine CO 2 , we must apply the averaging kernel at each level.We apply the FTS averaging kernels to AM2 data by calculating a monthly mean solar zenith angle (SZA) for spectra obtained at each site (Fig. B1).We likewise calculate the mean CO 2 profile used in GFIT.GFIT uses prior profiles based on GLOBALVIEW-CO2 and incorporates the seasonal cycle and secular increase of CO 2 .These prior profiles are recentered to the model CO 2 , maintaining the profile shape.The averaging kernel corresponding to the appropriate monthly mean SZA and the daily a priori profile are interpolated to the pressure profile over each site in AM2.The AM2 CO 2 incorporating the averaging kernels are then calculated as where [CO 2 ] p is the prior value and q is the water vapor mixing ratio.We use a simplified parameterization of the averaging kernel for our model (i.e., monthly mean profiles and averaging kernels) to reduce the computational intensity of the comparison and because we use daily mean CO 2 model output, therefore we cannot take into account actual solar zenith angles.We find that applying the FTS averaging kernel to AM2 output has only a small effect on large-scale diagnostics in AM2, such as the seasonal cycle amplitude (<0.2 ppm) or the estimated north-south gradient (<0.01 ppb km −1 ).Therefore, we are satisfied that for these bulk diagnostics, our method is sufficient.

Fig. 2 .
Fig. 2. Flight tracks in the free troposphere (350-800 hPa) during INTEX-NA during July-August 2004.Free tropospheric CO 2 is highly variable during these two months.
describe seasonally resolved TransCom flux inversion results, the land flux amplitude is defined as 12 •(NEE Jun−Aug − NEE Oct−Mar ), with optimized flux amplitudes ranging from 21 to 31 Pg C yr −1 (Gurney et al., 2004).Northern Hemisphere temperate and boreal terrestrial regions, roughly consistent with land area north of 20 • , are included in this calculation.When we apply this definition to the prescribed fluxes described above, CASA fluxes have a flux amplitude of 27 Pg C yr −1 , boreal amplified CASA fluxes have a flux amplitude of 35 Pg C yr −1 , and boreal amplified CASA fluxes with early growing season onset have a flux amplitude of 37 Pg C yr −1 .

Fig. 4 .
Fig. 4. Daily drawdown in CO 2 at Park Falls during one week in July 2006.Within one day, CO 2 changes by ∼2 ppm.

Fig. 5 .
Fig. 5. Comparison between NEE calculated from drawdown in the total column and NEE infered from eddy covariance measurements at Park Falls, Wisconsin.Drawdown in the column is calculated over a 4 h period symmetric about solar noon.Eddy covariance fluxes are averaged over the same time period.(a) NEE calculated each day for which data are available.(b) Data aggregated by week of year.(c) Data aggregated for each month.The dashed line represents the 1:1 line, while the solid line represents the best fit to the data.

Fig. 6 .
Fig. 6.Two profiles obtained above Park Falls on 15 July 2004 within three hours during the INTEX-NA campaign.A frontal system moved through the region between the profiles, leading to a ∼5 ppm increase in free tropospheric [CO 2 ] between 5-9 km.The decrease in boundary layer [CO 2 ] is notably smaller.

Fig. 8 .
Fig. 8. Daily median CO 2 at Park Falls for two months in summer2006 overlaid with daily mean potential temperature, θ , from NCEP reanalysis sampled at Park Falls at 700 hPa.The mean CO 2 and mean θ for the two-month period have been subtracted from the respective time series.Within a week, daily median CO 2 varies by up to 5 ppm.CO 2 and θ covary because both have strong northsouth gradients and variations arise from advection of these gradients.

Fig. 9 .
Fig.9.July-August CO 2 from four Northern Hemisphere TC-CON sites plotted against its meridional displacement relative to Park Falls, determined from variations in θ (small circles).A positive meridional displacement indicates a northerly airmass.Also plotted are the July-August mean CO 2 for TCCON sites from three AM2 simulations with modified NEE.The CO 2 values plotted are referenced to the CO 2 at Lauder, New Zealand.CO 2 from a simulation using CASA NEE (large circles) underestimates the north-south gradient among the TCCON sites.Amplified boreal fluxes (squares) and amplified boreal fluxes with an early growing season onset (diamonds) better capture the north-south gradient during summer.

Fig. 10 .
Fig. 10.Free tropopsheric [CO 2 ] measured during INTEX-NA plotted against (a) θ at 700 hPa determined from regional reanalysis data and (b) latitude.[CO 2 ] better correlates with potential temperature because synoptic activity obscures the north-south gradient while a dynamical coordinate reveals it.

Fig. 11 .
Fig. 11.Free tropopsheric[CO 2  ] measured during HIPPO-1 plotted against local θ and colored by latitude.The contrast across the entire Northern Hemisphere is smaller during winter than the contrast across only the midlatitudes during summer (Fig.9).Variations in CO 2 at Lamont and Park Falls during the campaign show a similar north-south gradient (black squares).

Fig. 12 .
Fig.12.Detrended weekly median CO 2 at four Northern Hemisphere TCCON sites (circles) plotted with simulated CO 2 from AM2 (solid lines).The simulated CO 2 has been smoothed using a two week running mean.The observed and simulated CO 2 values are referenced to annual mean CO 2 at Lauder, New Zealand.In each panel, the data are the same, while the simulated seasonal cycles result from (a) CASA net ecosystem exchange, (b) boreal amplified CASA net ecosystem exchange, (c) boreal amplified CASA net ecosystem exchange with an early growing season onset.CASA NEE underestimates the seasonal cycle amplitude in the column at all four sites, while simulations with enhanced boreal exchange better match the data.The colors are the same as in Fig.1.
Fig. B1.Monthly mean solar zenith angles for TCCON sites.

Table 1 .
TCCON sites and their median seasonal cycle amplitudes (SCA) in ppm.

Table 2 .
Estimated meridional gradient [ppb km −1 ] in CO 2 during the Northern Hemisphere summer growing season determined from TCCON sites and in free tropospheric [CO 2 ] from INTEX-NA data compared to AM2 simulations.