The NASA Orbiting Carbon Observatory 2 (OCO-2) launched in 2014 and began collecting data in September of that year. Daily averages of XCO2 are calculated for each zone using observations from OCO-2 Lite files version 9, “B9”;. Ongoing improvements in the ACOS retrieval algorithm and previous efforts by to develop quality control thresholds tailored to OCO-2 B9 retrievals over boreal forest regions (boreal QC) have allowed sufficient data over our 5∘ latitude by 20∘ longitude zones to construct XCO2 time series that yield robust seasonal cycle parameterization. The boreal QC was evaluated for use with terrestrial OCO-2 B9 retrievals north of 50∘ N , and the zones considered here cover the majority of land north of 50∘ N. The southern boundaries of the southernmost zones of North America are at 47∘ N, but the 3∘ of latitude is not expected to significantly impact the effectiveness of the boreal QC filtering. Instead of the standard B9 bias correction, we use a modified bias correction that includes temperature at 700 hPa (T700), as discussed by , because it was found in previous results to reduce the seasonality of OCO-2 bias relative to ground-based TCCON and EM27/SUN measurements. Seasonal cycle fits to OCO-2 retrievals of XCO2 with the standard B9 bias correction, as well as fits to OCO-2 B10 retrievals, were also calculated and compared to model-derived seasonal cycle fits in the Supplement (see Sect. S2). The spatial distribution of seasonal cycle parameters across northern high-latitude regions is similar for all three types of OCO-2 retrievals, but the alternative bias correction yields improved agreement with model-derived seasonal cycle parameters.

The Total Carbon Column Observing Network (TCCON) is a ground-based network of sites observing XCO2 using high-spectral-resolution solar-viewing Fourier transform infrared spectrometers (FTSs). Data are included from four TCCON sites: East Trout Lake, Canada, in North American boreal zone 3 ; Sodankylä, Finland, in European boreal zone 6 ; Białystok, Poland, in European temperate zone 2 ; and Bremen, Germany, in European temperate zone 3 (see site details in Table  and locations mapped in Fig. ). The Collaborative Carbon Column Observing Network (COCCON) is a network of sites observing with the Bruker EM27/SUN FTS , which are lower-resolution mobile solar-viewing spectrometers that serve as complement to TCCON measurements. EM27/SUN observations have been compared to TCCON observations in multiple studies, most notably , , , , and . In most of these comparisons EM27/SUN and TCCON observations agree with biases less than 0.25 ppm on average. In some cases offsets between EM27/SUN and TCCON observations are reported to be as large as 2 ppm, but the proven stability of the EM27/SUN should allow for a bias correction that would yield good agreement between TCCON and EM27/SUN retrievals. The EM27/SUN instruments have measured XCO2 in a number of campaigns to validate OCO-2 and other satellite-based observations, including work by , , and , suggesting good agreement between EM27/SUN observations and satellite-based observations. In this analysis, observations with an EM27/SUN FTS in Fairbanks, Alaska, USA (65.859∘ N, 147.850∘ W; ), are used as a fifth ground-based comparison in the boreal forest. Fairbanks is an established COCCON site as of 2018, so the instrument participates regularly in performance and calibration checks at the central facility operated by KIT, and data processed in compliance with COCCON recommendations are available. In this study, we use the GGG2014 retrieval algorithm coupled with the EM27/SUN GGG interferogram processing suite (EGI; ) instead of the standard COCCON retrieval methods for consistency with TCCON retrievals and because this data product has already been bias corrected to TCCON using side-by-side observations at Caltech, as described in detail by . Seasonal cycle fits for ground-based observations at these five sites use daily averages of retrievals collected within 2 h of local solar noon, weighted by retrieval error, as described by . We refer to these daily averages as near-noon ground-based (NNG) observations, and this time frame is chosen because OCO-2 overpasses tend to occur within approximately 1 h of local solar noon over these northern high-latitude regions.

We define regions in North America, Europe, and Asia, which are further subdivided into 5∘ latitude by 20∘ longitude zones, and zones are designated as temperate, boreal, or tundra. For the purposes of this analysis, the classification of zones as temperate, boreal, or tundra, as well as the longitudinal division between the European and Asian regions, are guided by maps of ecoregions from and (see Fig. ). The boreal forest zones considered cover a narrower range of latitudes than the boreal regions defined for the Transcom 3 ecoregions , which include all high Arctic tundra and portions of temperate Asia as part of the boreal regions. Otherwise, the North American boreal region and Eurasian boreal region defined by Transcom 3 are very similar to the North American boreal and Asian boreal regions defined in this analysis. We differ markedly from Transcom 3 in defining separate European boreal and European temperate regions, while Transcom 3 combines all of Europe into a single European region. In North America, the zones are shifted by 3∘ latitude relative to the zones in Eurasia – starting at 47∘ N and extending in to 72∘ N in 5∘ increments. This was done to bring ground sites in the North American boreal region closer to the center latitude of their encompassing zones and to more accurately fit the boundaries of temperate, boreal, and tundra biomes.

Also shown in Fig.  are the locations of five ground sites where long-term observations of XCO2 have been collected (see Table  for details). These ground sites include two sites in the European temperate region (Białystok and Bremen), one site in the European boreal region (Sodankylä), and two sites in the North American boreal region (East Trout Lake and Fairbanks). Ground-based data are explained further in Sect. , and seasonal cycles of ground-based data are compared to satellite and model-derived seasonal cycles in Sect. .

The primary focus of our analysis is characterizing seasonality in XCO2 and exploring how and why this seasonality differs across northern high-latitude regions, with particular emphasis on the boreal forest. To this end, time series are constructed using daily averaged XCO2 from satellite retrievals, ground-based solar-viewing Fourier transform infrared (FTIR) spectrometers, and model estimates (see previous methods sections for details). Seasonal cycles are characterized following methods used by , in which daily XCO2 values are fit to a skewed sine wave of the form 1f(t)=a0+a1t+a2sin⁡(ω[t-a3]+cos⁡-1[a4cos⁡(ω[t-a5])]), where t is days, ω=2π365.25, the interannual trend is defined by a0+a1t, and the seasonal cycle amplitude (SCA) is defined by 2|a2|. As a metric for seasonal timing we define half drawdown day (HDD) as the day of year when the detrended seasonal cycle fit, f(t)-a0-a1t, crosses zero from positive to negative. The fit is calculated using nonlinear least squares optimization with the standard error, σ, defined as the mean of daily standard deviations in XCO2 over all days in the time series. If there is no daily standard deviation, as is the case with single point model estimates near ground sites from the CT2019B and GC-CT2019 models, the standard error is assumed to be 0.25 ppm. Specifically, we implement the Levenberg–Marquardt algorithm for nonlinear least squares fitting, which seeks to minimize the quantity, 2χ2=∑i=1nyi-f(xi,ao,a1,…aj)2σ2, and yields variance for each parameter in the fit equation given by 3σaj2=([JTWJ]-1)jj, where J is the Jacobian and W=Vxi-1, the inverse of the covariance matrix. In this case uncertainties are scaled by χ2, such that Vxi is defined as χ2σ2I. The variance in fit parameters is taken to be a direct estimate of fit uncertainty and translate to uncertainty in SCA, defined as 2σa2 and depicted in figures as error bars, where appropriate.

The seasonal cycle fitting approach used by was found to be more numerically stable than fitting to a truncated Fourier series, as has been employed in previous studies , because periods of missing data can produce unrealistic oscillations in a fit to a truncated Fourier series. Even in cases with continuous data, the fit to a truncated Fourier series can yield unrealistic oscillations within the overall seasonal cycle that do not appear to accurately depict variability in the data. These unrealistic oscillations are more pronounced when there are gaps in the time series of data. The fits to Eq. () also show a degradation of the seasonal cycle shape with larger gaps in the time series but yield more stability in fitting time series with some data gaps than the fits to a truncated Fourier series. This is an important consideration for high-latitude regions, which have winter gaps in observations for most passive atmospheric remote sensing measurements, due to lower solar elevation angles or night at satellite overpass time (near solar noon).

Model estimates of XCO2 from the Copernicus Atmospheric Monitoring Services (CAMS) global inversion-optimized greenhouse gas flux model v19r1 are used here as a model comparison to OCO-2 and NNG data. The modeling framework for CAMS CO2 flux inversions is described in detail by . Quality assessments for the Northern Hemisphere by report that nearly all biases in both CAMS estimates of in situ CO2 relative to unassimilated aircraft observations and CAMS estimates of XCO2 relative to TCCON observations are within 1 ppm, with standard deviation in bias around 2 ppm. CAMS estimates of XCO2 are available as 3-hourly estimates with 1.9∘ latitude by 3.75∘ longitude spatial resolution, which is sufficient for providing multiple grid points within each zone and coincidence with most ground sites within approximately 100 km (see Table S1 in the Supplement for exact coordinates of grid points nearest to the ground sites). Daily averages and standard deviation in CAMS XCO2, used to calculate seasonal cycle fits, are taken from combined spatial aggregations within zones or coincidence regions and temporal aggregations for each 24 h date in UTC. We use CAMS model estimates with data assimilation from a global network of surface in situ observations at 119 locations but without any satellite data assimilation. In addition to the XCO2 estimates, the CAMS model output includes surface flux estimates, which will be considered further in the Discussion, Sect. . Both CAMS flux estimates and CAMS XCO2 estimates use the same atmospheric transport modeling framework.

The CarbonTracker2019 (CT2019) model is an inverse model that provides estimates of global CO2 fluxes with a 1∘ by 1∘ spatial resolution and estimates of global XCO2 fields with a atmospheric transport simulated by the TM5 model with a spatial resolution of 2∘ latitude by 3∘ longitude . The model assimilates in situ measurements of atmospheric CO2 concentrations from aircraft, AirCore , tall tower, and surface measurement platforms at 460 sites around the world. For CT2019 fluxes considered in this analysis and used in GEOS-Chem simulations, we use the original CT2019 release from May 2020 , but for the posterior estimates of XCO2 we use results from an updated release, CT2019B . This was a matter of practical necessity because of the timing of the updated CT2019B release and the fact that XCO2 estimates from the previous CT2019 version were unavailable after this release. The purpose of the release of CT2019B was reported as “Correction of a minor bug in CT2019” . CT2019 uses the Global Fire Emissions Database, version 4.1 (GFED4.1s; ; ), as one part of the fire module that estimates emissions from biomass burning, while fossil fuel emissions are driven by a combination of the Miller emissions inventory (see , for more details) and the Open-Data Inventory for Anthropogenic Carbon dioxide (ODIAC; ) emissions datasets. Biospheric exchange is driven by the Carnegie-Ames-Stanford Approach (CASA) biogeochemical model introduced by . In this analysis we calculate seasonal cycle fits using daily CT2019B posterior XCO2 and compare these to corresponding seasonal cycle fits from ground-based and satellite-based observations. The CT2019 estimates of biospheric CO2 exchange are also used within a GEOS-Chem model framework, described in Sect. , and considered in an assessment of the role of fluxes in shaping XCO2 seasonality.

The GEOS-Chem atmospheric transport model version 12.7.2 (more detailed information at https://geos-chem.seas.harvard.edu/, last access: 31 July 2020) has 2∘ latitude by 2.5∘ longitude spatial resolution, using MERRA-2 meteorology . We use the GEOS-Chem CO2 simulation and GEOS-Chem surface contact tracers, for 2014–2016, to examine the relationships between seasonal cycle parameters and atmospheric transport patterns and speculate on the role of atmospheric transport in determining spatial distributions of XCO2 seasonality across northern high latitudes. The GEOS-Chem CO2 simulation provides daily XCO2 estimates and source attributions for total column CO2, with biospheric fluxes for land and ocean taken from the CT2019 model , so we refer to this combination of GEOS-Chem atmospheric transport and CT2019 biospheric exchange as GC-CT2019 throughout this paper. Similar to CT2019, this GC-CT2019 simulation uses GFED4.1 to estimate emissions from biomass burning, but unlike CT2019, it uses the Community Emissions Data System (CEDS; ) for fossil fuel emissions and results from for the chemical production of CO2 in the atmosphere. For the fossil fuel emissions, the CEDS inventory ended in 2014, so 2015 and 2016 emissions were scaled by the CEDS 2014 emissions to match the global total in those later years, as reported by ODIAC. We used this approach, rather than using ODIAC alone, because the CEDS inventory includes anthropogenic biofuel emissions that are not in ODIAC. The GC-CT2019 simulation is initialized on 1 January 2007 with observed global ocean surface mean CO2 (Dlugokencky and Tans, NOAA/GML, http://www.esrl.noaa.gov/gmd/ccgg/trends/, last access: 15 December 2020), giving the model 7 years of spin-up before generating the output for 2014–2016 used here. GC-CT2019 estimates, like CT2019B posterior XCO2 estimates, are obtained as daily values for grid points with daily averages and standard deviation calculated spatially for each 5∘ latitude by 20∘ longitude zone or 5∘ latitude by 10∘ longitude satellite coincidence region; there is no temporal averaging performed on these estimates for this analysis. Unlike the CAMS model and CT2019B XCO2 estimates, which provide optimized CO2 flux, and XCO2 estimates using internally consistent atmospheric transport models, GC-CT2019 uses CT2019 fluxes that are estimated using TM5 to simulate atmospheric transport rather than the GEOS-Chem transport model. In addition, the fossil fuel and biomass burning flux estimates used in GC-CT2019 are based on slightly different source datasets. GEOS-Chem simulations are run for 2014–2016 rather than for 2014–2019, like the OCO-2 observations and other model estimates. The analysis shown in the Supplement (see Fig. S39) suggests that the spatial distributions of SCA and HDD across northern high-latitude zones are not likely to significantly change for CAMS, CT2019B, or OCO-2 when calculated for 2014–2016 instead of 2014–2019. In particular, changes in SCA for the different time periods are less than approximately 0.5 ppm, which represents less than 10 % of the ∼5 ppm variability in SCA across these northern high-latitude regions. There are some larger discrepancies between OCO-2 HDD for 2014–2016 and OCO-2 HDD for 2014–2019, with a couple of zones yielding a difference of around 8 d, but this may be partially attributed to the fact that some zones lack sufficient data points in the 2014–2016 time period for a stable and accurate seasonal phase determination.

To examine the role of atmospheric transport in shaping the seasonal cycle of XCO2, we define new tracers of air mass surface contact in the GEOS-Chem atmospheric transport model. These surface contact tracers are emitted uniformly and continuously over specific surface types at a rate of 1 molecule m-2 s-1, transported like CO2 and other constituents, and decay with a prescribed e-fold lifetime. One set of these surface contact tracers is emitted uniformly over the ocean and another set is emitted uniformly over land. For both surfaces, we release multiple tracers with lifetimes of 5, 15, 30, and 90 d, making eight surface contact tracers in total. After spinning up the simulation for several e-fold lifetimes, the tracer concentration at a point in the model indicates the integrated upwind contact with land or ocean over the timescale of the tracer lifetime. For example, high concentrations of the land surface contact tracers reveal locations where atmospheric circulations have confined air over continents. These surface contact tracers are like e90 , except that e90 is emitted uniformly from all surfaces and therefore indicates upwind contact with any surface rather than particular surface types as we have done. The sum of our land and ocean tracers with 90 d lifetimes is equal to e90. The surface contact tracers are initialized and run for the same periods as the GC-CT2019 simulation, 2014–2016. The surface contact tracers for a given zone were found to vary minimally in time, so a total zonal average of surface contact tracer contributions was taken spatially and temporally within each zone (see map in Fig. ) and over all days in 2014–2016.

The role of fluxes in determining XCO2 seasonality at northern high latitudes is assessed using flux estimates from CAMS, as well as the sum of CO2 flux estimates from the CEDS and GFED4.1s inventories and CarbonTracker2019 biospheric fluxes from land and ocean used to generate the GC-CT2019 XCO2 estimates. While the CAMS v19r1 and CT2019B model frameworks have internally consistent atmospheric transport because both CO2 flux and XCO2 estimates are generated using the same atmospheric transport model , GC-CT2019 includes biospheric fluxes from CarbonTracker2019 using the TM5 transport model but then applies GEOS-Chem atmospheric transport modeling to estimate XCO2. In addition, the fossil fuel fluxes from CEDS and biomass burning fluxes from GFED4.1s may include other assumptions about atmospheric transport.

First, CO2 fluxes from each model are averaged spatially within each 5∘ latitude by 20∘ longitude zone (see Fig. ) for each 3-hourly time step. A total average annual flux is calculated for each zone by summing all 3-hourly CO2 fluxes in each year and taking an average over the 6 years in 2014–2019. To calculate the flux seasonal cycle amplitude (flux SCA), the 3-hourly, spatially averaged fluxes within each zone are summed for each 24 h period in UTC and used to derive a 15 d rolling mean, which is then averaged by day of year to yield an average annual cycle for 2014–2019. The difference between the maximum and minimum of the average annual cycle is then taken to be the flux SCA. The annual cycles of fluxes are plotted in Figs. S32 through S49 of the Supplement.

Before attempting to interpret spatial distributions of seasonal cycle parameters on continental scales, it is of value to get a better idea of how seasonal cycle parameters from observations at a single location compare to those from spatially averaged data. To this end, five high-latitude sites are considered with long-term ground-based observations, as described in Sect. . There are three spatial scales considered with seasonal cycle fits to near-noon ground-based (NNG) observations and seasonal cycle fits to spatially averaged data over a commonly used satellite coincidence region of 5∘ latitude by 10∘ longitude centered on the location of the ground site, as well as over the 5∘ latitude by 20∘ longitude zone in which the ground site is located (reference zones in Fig. ). In Fig. , observed SCAs and HDDs from NNG and OCO-2 observations are correlated against model-derived SCAs and HDDs from CAMS, CT2019B, and GC-CT2019 at each of the three spatial scales, and the corresponding linear regression equations and correlation coefficients are reported in Table . These correlations exhibit tight linearity for SCA (with most R2>0.7) and reasonable linearity for HDD when comparing observed and model-derived parameters at all scales. SCAs from CAMS or CT2019B are in better agreement with observations than those from GC-CT2019 in every case, as demonstrated by the fact that the CAMS and CT2019B linear regressions fall closer to the y=x line than the GC-CT2019 linear regression in panels (a), (b), and (c) of Fig. . Agreement between model-derived and observed HDD is better for the single-point model estimates nearest the ground site versus NNG results in panel (d) of Fig. , and the scatter increases in the comparisons of HDD from spatially averaged model estimates versus spatially averaged OCO-2 observations in panels (e) and (f) of Fig. . All three models tend to yield slightly larger SCAs than observations at all sites, as well as later HDD than observations with the exception of CAMS HDD at Bremen and Sodankylä. These results stand in contrast to earlier work by , who found that the Transcom model underestimated SCA of CO2 mixing ratios relative to aircraft observations at nearly every altitude. Details of the full time series, plots of seasonal cycle fits, and seasonal cycle fit parameters, with uncertainties, for these ground sites, coincidence regions, and encompassing zones are reported in the Supplement.

In Fig.  and Table  the relationships between seasonal cycle parameters from spatially averaged data versus those from a single point, at or nearest the ground sites, are explored. In this case, seasonal cycle parameters derived from NNG observations are correlated against seasonal cycle parameters derived from spatially averaged OCO-2 retrievals, while model-derived seasonal cycle parameters for a single point near the ground site are correlated against model-derived seasonal cycle parameters from spatially averaged model estimates. have shown that an alternative bias correction, parameterized for temperature at 700 hPa, resulted in reduced seasonality in OCO-2 bias within the 5∘ latitude by 10∘ longitude coincidence region relative to NNG observations at East Trout Lake, Sodankylä, and Fairbanks. Results in the Supplement show that the alternative bias correction improved agreement in both SCA and HDD between NNG seasonal cycles and coincident OCO-2 seasonal cycles. The results in Fig.  and Table  indicate that HDD correlations across scales tend to be slightly weaker and markedly different depending on whether one considers observed or model-derived seasonal cycles. For the coincidence region and the encompassing zone, OCO-2 data consistently yield later HDD than NNG data, while spatially averaged model estimates tend to yield HDD that is in good agreement or slightly earlier than the point nearest to the ground site. SCAs for ground sites are well correlated and mostly in close agreement with both SCAs for 5∘ latitude by 10∘ longitude coincidence regions and SCAs for encompassing 5∘ latitude by 20∘ longitude zones, demonstrating that SCA scales well with spatial averaging and is a credible metric to consider in the context of this analysis. The relatively weaker correlations in HDD across spatial scales in panels (c) and (d) of Fig.  suggest greater spatial heterogeneity in HDD within zones from both observed and model-derived seasonal cycles. This result, in combination with the scatter in panels (d), (e), and (f) of Fig. , implies that HDD may display more random variability than SCA and may therefore be a less useful metric in this context.

The full set of seasonal cycle fit parameters, corresponding uncertainty estimates, and plots of time series and seasonal cycle fits for each zone and ground site in Fig.  are reported in the Supplement. While the fits to model estimates are generally similar in shape with fits to observations, there are some zones that yield an unrealistic drop in wintertime values in the OCO-2 seasonal cycle fits, which is more pronounced for zones with fewer satellite-based XCO2 observations near the peak and trough of the seasonal cycle. This wintertime drop in the shape of the seasonal cycle is evidence that should motivate further efforts to increase satellite-based observations over high-latitude regions outside of the summer months. Only a small number of zones, particularly in the Asian tundra, have seasonal cycle fits that are obviously compromised by insufficient data in spring and autumn, while the majority of zones have seasonal fits that look reasonable and are similar to seasonal cycle fits to continuous model estimates, with the exception of the noted wintertime drop. An analysis of changes in CAMS seasonal cycle fits when selecting only days with OCO-2 observations available rather than the full continuous time series of model estimates is presented in the Supplement, which indicates that shifts in CAMS SCA due to imposed data gaps are -0.044±0.197 ppm, with shifts in SCA for all zones less than 0.5 ppm, while biases in CAMS SCA for the full time series relative to OCO-2 SCA are 0.547±0.720 ppm and range from -1.0 to 3.0 ppm. These shifts in CAMS SCA are also small relative to the approximately 5 ppm variability in SCA seen across the northern high-latitude regions. In addition, the Lindqvist method of fitting still provides a more constrained shape than a fit to a truncated Fourier series, which can yield highly unrealistic oscillations in time series with only minimal gaps in data coverage. The close similarities between spatial distributions of seasonal cycle parameters from OCO-2, CAMS, CT2019B, and GC-CT2019 are apparent in Fig. .

Seasonal cycle amplitudes (SCA) and half drawdown day (HDD) are mapped in Fig. , showing results from OCO-2 observations, CAMS model estimates, CT2019B model estimates, and GC-CT2019 model estimates of XCO2. Figure  demonstrates that both OCO-2 observations and model estimates from CAMS, CT2019B, and GC-CT2019 yield larger SCAs and earlier HDDs in the Asian boreal forest than any other region. The earlier seasonal timing of the Asian boreal forest zones is consistent with results from and , and these studies also linked earlier drawdown in Asia to larger SCA. Although one would not necessarily expect SCA of surface in situ measurements to match SCA of XCO2, this finding also aligns with the study by , who found that Siberia had the largest SCA in surface CO2 concentrations when normalized for area. The CAMS, CT2019B, and GC-CT2019 results in Fig.  exhibit more apparent latitudinal gradients, whereas the results from OCO-2 show more spatial heterogeneity, and this is particularly true for HDD. Seasonal cycles of direct observations are expected to display more heterogeneity than seasonal cycles of model estimates, which depend on mathematical modeling of atmospheric transport to calculate XCO2, even if the underlying fluxes are based on in situ data assimilation. Overall, the spatial distributions in SCA and HDD from OCO-2 agree more with those from CAMS and CT2019B than those from GC-CT2019. GC-CT2019 yields larger magnitudes of SCA for many regions, as well as SCAs and HDDs spread across a larger range for northern high-latitude regions, relative to SCAs from OCO-2, CAMS, or CT2019B. In addition, OCO-2 observations yield notably smaller SCA than CAMS, CT2019B, or GC-CT2019 in the western zones of the Asian boreal forest, in the Asian tundra, and in the eastern zones of North America.

Panels (a), (c), (e), and (g) of Fig.  show a clear increase in SCA from west to east across the Eurasian continent in both model-derived and observational results. In North America, longitudinal gradients are more subtle. While OCO-2 observations exhibit a slight gradient in SCA across North America that increases from east to west, CAMS, CT2019, and GC-CT2019 yield SCAs that increase from west to east, in the same direction as gradients across Eurasia. This discrepancy in North America hinges primarily on the zones of boreal forest and tundra in eastern North America, which have the smallest SCAs for that continent when using OCO-2 data but have the largest SCAs for that continent when using CAMS, CT2019B, or GC-CT2019 model estimates. As expected, panels (a), (c), (e), and (g) in Fig.  show latitudinal gradients with increasing SCA from south to north for both observed and model-derived seasonal cycles. However, the Asian boreal forest zones stand apart from other regions in all panels of Fig. , particularly when plotted against latitude, with larger SCA than other data at similar latitudes or longitudes. Results in Fig.  demonstrate that HDDs are far more scattered and do not follow the distinct trends with latitude and longitude that SCAs do. HDD spatial gradients seem to be inverted relative to SCA with a vague tendency toward later HDD at more northern latitudes and more western longitudes, and HDD also exhibits similar discrepancies between observed and model-derived longitudinal gradients for North America.

Figure  and the calculated linear regressions in Table  show that there is a negative correlation between HDD and SCA for both observed and model-derived seasonal cycle fits, such that earlier HDD corresponds to larger SCA. CAMS model estimates yield correlations between HDD and SCA that are more similar to those from OCO-2 and NNG measurements, while GC-CT2019 yields stronger correlations with steeper slopes. The results in Fig.  emphasize the exceptionally early HDD and large SCA of the Asian boreal forest, such that many of the Asian boreal zones fall more in line with the tundra zones than with the other boreal zones. A latitudinal gradient is suggested by the fact that the linear regressions plotted in Fig.  are shifted up for the tundra zones and shifted down for the temperate zones, with the boreal zones in between the two. Furthermore, the separate linear regressions for the temperate, boreal, and tundra zones plotted in Fig.  have much larger R2 than the linear regressions for all the zones together (see Table ), suggesting that there are different dynamics in different biomes that affect relationships between extent and timing of apparent CO2 uptake. The strength of this correlation in observed and CAMS seasonal cycles was highest for tundra and lowest for temperate zones with the boreal zones falling in between.

GEOS-Chem surface contact tracers were used to simulate the release of tracers from land and ocean yielding relative concentrations of tracers with lifetimes of 5, 15, 30, and 90 d for a given grid point and day. In these results the relative contributions of surface contact tracers for each zone represent an overall average of surface contact tracer contributions for all days in 2014–2016 and aggregated spatially across the 5∘ latitude by 20∘ longitude zone. The surface contact tracers show that the largest XCO2 SCAs occur in areas with the greatest influence from air that contacted land surfaces 15 to 30 d prior. There are clear similarities in the spatial distributions of SCA in panels (a), (c), and (e) of Fig.  and those of the surface contact tracers from land with a 15 or 30 d lifetime, as shown in panels (c) and (e) of Fig. . There are also similarities in the spatial distributions of HDD in panels (b), (d), and (f) of Fig.  and those for the surface contact tracers from ocean with a 15 or 30 d lifetime, as shown in panels (d) and (f) of Fig. . The relative strength of linear relationships between seasonal cycle parameters from OCO-2 observations and surface contact tracers is quantified with correlation coefficients in Fig. . Figure  shows that the observed SCAs are most correlated with land-based surface contact tracers with 15 and 30 d atmospheric lifetimes, while observed HDDs are most correlated with ocean-based surface contact tracers with 15 and 30 d atmospheric lifetimes. Correlations between HDD and land tracers are weak (see Fig. S45) and seem to follow a curve rather than a line or may be representative of two different linear relationships for different groups of zones. The correlations with ocean tracers were always inverted relative to those with land tracers, such that reduced contribution from ocean tracers and increased contributions from land tracers consistently correspond with larger SCA and often correspond with earlier HDD.

and demonstrated the ability of GEOS-Chem to represent synoptic transport from midlatitudes to the Arctic. evaluated meridional transport of CO2 and SF6 in GEOS-Chem, using the same MERRA-2 meteorology used here, and suggested that the model may underestimate vertical and meridional mixing, which slightly increases CO2 seasonal cycle amplitude in GEOS-Chem in high northern latitudes compared to some other models. Despite some discrepancies in SCA and HDD from GC-CT2019, relative to observed SCA and HDD, which may arise from differences between the GEOS-Chem and TM5 atmospheric transport models, the GEOS-Chem transport model reproduces surface temperature and pressure patterns and thus is a reasonable representation of atmospheric transport . Therefore, comparing these GEOS-Chem surface contact tracers with observed SCA and HDD should provide useful insights into the influence of atmospheric transport patterns on observed XCO2 seasonality. The higher correlation coefficients obtained when comparing OCO-2 SCAs to land and ocean tracers with 15 and 30 d lifetimes suggest that accumulation of CO2 flux due to atmospheric transport on roughly monthly timescales plays an important role in affecting XCO2 SCA.

Figure  shows maps of average annual fluxes and flux SCAs for GC-CT2019 and CAMS, which show that flux SCAs in panels (c) and (d) are distributed without the apparent longitudinal gradient seen in XCO2 SCA in panels (a), (c), (e), and (g) of Fig. . The weak correlations (R2<0.2) between flux SCAs and XCO2 SCAs, shown in panels (a), (c), (e), and (g) of Fig.  and quantified in Fig. , combined with the relatively strong correlations (R2>0.6) between XCO2 and surface contract tracers with 15 and 30 d lifetimes, suggest that accumulation of CO2 flux due to atmospheric transport on roughly monthly timescales is more influential in determining XCO2 SCA than fluxes within a 5∘ latitude by 20∘ longitude zone. However, there are slightly stronger correlations between average annual flux and XCO2 SCA, shown in panels (b), (d), (f), and (h) of Fig.  and quantified in Fig. , which suggest some possible link between XCO2 SCA and the relative source or sink strength of a given zone. Panels (a) and (b) of Fig.  show that both models predict large negative average annual fluxes for Asian boreal zones, designating the Asian boreal region as a major sink for CO2 and suggesting that anomalously large XCO2 SCA in the Asian boreal region may be partially influenced by enhancements in CO2 uptake within that region. In another instance, the European temperate zone 3 and Bremen TCCON site both have exceptionally large positive average annual CO2 flux due to a significant contribution from fossil fuel emissions (as shown in Sect. S5), but they did not yield anomalously large flux SCA because the fossil fuel emissions do not exhibit significant seasonal variability. The European temperate zone 3 and Bremen also yielded smaller XCO2 SCAs and later HDDs than most of the other zones and ground sites for both observed and model-derived seasonal cycles. In this case, the large fossil fuel emissions may be indirectly influencing the XCO2 SCA despite the fact that these emissions are not directly contributing to the seasonal variability in atmospheric CO2 in this zone or at this site.