Large uncertainties in land surface models (LSMs) simulations still arise
from inaccurate forcing, poor description of land surface heterogeneity (soil
and vegetation properties), incorrect model parameter values and incomplete
representation of biogeochemical processes. The recent increase in the number
and type of carbon cycle-related observations, including both in situ and
remote sensing measurements, has opened a new road to optimize model
parameters via robust statistical model–data integration techniques, in order
to reduce the uncertainties of simulated carbon fluxes and stocks. In this
study we present a carbon cycle data assimilation system that assimilates
three major data streams, namely the Moderate Resolution Imaging Spectroradiometer (MODIS)-Normalized Difference Vegetation Index (NDVI) observations of vegetation
activity, net ecosystem exchange (NEE) and latent heat (LE) flux measurements
at more than 70 sites (FLUXNET), as well as atmospheric CO
Overall, the ORCHIDEE model is able to achieve a consistent fit to all three
data streams, which suggests that current LSMs have reached the level of
development to assimilate these observations. The assimilation of MODIS-NDVI
(step 1) reduced the growing season length in ORCHIDEE for temperate and
boreal ecosystems, thus decreasing the global mean annual gross primary
production (GPP). Using FLUXNET data (step 2) led to large improvements in the
seasonal cycle of the NEE and LE fluxes for all ecosystems (i.e., increased
amplitude for temperate ecosystems). The assimilation of atmospheric
CO
Atmospheric CO
Uncertainties in model simulations arise from inaccurate forcing, incorrect model parameter values and/or an inadequate or incomplete representation of biogeochemical processes in the model (for example the impact of nutrient limitation on C fluxes, or C release related to permafrost thawing). Arguably the best way to improve model predictions is to confront simulations with multiple sources of data within an appropriate and rigorous framework (Prentice et al., 2015). In the last 2 decades significant efforts by the site and satellite observation communities have resulted in a large increase in the number and type of C cycle-related observations. These data contain some information at various spatial and temporal scales and should be combined together to robustly address different aspects of the models. One way in which these data can be used to better quantify and reduce model uncertainty is by optimizing or calibrating the model parameters via robust statistical model–data fusion (or data assimilation – DA) techniques. In particular a Bayesian inference framework allows us to update our prior knowledge of the parameters based on new information contained in the observations.
There is a long history of using DA techniques for parameter optimization,
particularly in geophysics (Tarantola, 1987), but the initial studies in the
field of global terrestrial C cycle data assimilation started with the
initial study of Fung et al. (1987) and a pioneering work by Knorr and
Heimann (1995), who used atmospheric CO
When assimilating multiple different data streams we have two options: (i) to optimize the model with each data stream in turn, and to propagate the information gained on the parameter values from one step to the next (hereafter referred to as “stepwise” assimilation), or (ii) to include all data streams together in the same optimization (hereafter referred to as “simultaneous” assimilation). Kaminski et al. (2012) suggested that it is essential to perform a consistent, simultaneous assimilation that includes all data streams in the same optimization. It is important to note that this is an implementation question. Tarantola (2005) recast the fundamentals of the approach as the conjunction or multiplication of probability densities. This multiplication is associative so it makes no difference whether it is performed in one step or several (and whether the system is linear or not). In complex problems such as these, one cannot carry or even describe the full structure of the relevant probability densities, so which approach will work best in each case is unclear. In particular, technical difficulties associated with the different number of observations for each data stream and the characterization of error correlations between them, in addition to computational constraints to run global LSMs, might result in the preference for a stepwise assimilation framework. Additionally, it may be more straightforward, to expose a restricted set of parameters (following a global sensitivity analysis) to each observation type in a stepwise approach to ensure that each data stream constrains only the most relevant parts of the model. This reduces biases from other poorly represented processes caused by inadequate model structure. Note finally that more complex approaches based on random generation of parameter sets, such as the multi-objective approach using the Pareto ranking of several cost functions (e.g., Yapo et al., 1998), are not yet affordable for global LSMs from a computational point of view. For these reasons we follow the stepwise approach in this paper.
We present the first global-scale CCDAS that assimilates three of the main
global data streams that have been used to date to understand the terrestrial
carbon cycle – atmospheric CO
In this context, the main questions that we aim to answer in this paper are
as follows:
How and to which extend the optimization of the ORCHIDEE model allows one to
fit the three data streams that are considered? Does the stepwise optimization result in a degradation of the fit to
other data streams used in the previous steps? What are the main changes in the optimized parameters when using
sequentially these three data streams in a global CCDAS and which processes
are constrained? What are the improvements for the land C cycle in terms of net/gross
fluxes and stocks as a result of multi-data stream optimization? What
preliminary perspectives can we draw that may help us in improving model
predictions of trends, variability and the location of terrestrial C sources
and sinks?
Following these objectives, the paper first describes the new ORCHIDEE-CCDAS
including the concept, the observations, the models and the optimization
approach. We then present the results, including the fit to the data,
consistency checks (question i, above) as well as the mean global and
regional C cycle budget for the period 2000–2009. The last section discusses
issues and perspectives associated with these results.
Schematic of the ORCHIDEE Carbon Cycle Data Assimilation System (ORCHIDAS).
We have designed a CCDAS around the ORCHIDEE land surface model
(ORCHIDEE-CCDAS, later also referred to as ORCHIDAS for simplicity) that
combines a state-of-the-art description of the driving biogeochemical
processes within the model with multiple observational constraints in a
robust statistical framework, in order to improve the simulation of land
carbon fluxes and stocks. The system allows us to retrieve the best
estimate, given the observations and prior information, of selected
parameters (see Sect. 2.3.3) as well as to evaluate their
uncertainty. It relies on a stepwise assimilation of a comprehensive set of
three C cycle-related observations that are representative of small (100 m)
to large (continental) scales (see Sect. 2.2):
step 1: satellite measurements of vegetation activity using the Normalized Difference Vegetation Index (NDVI) from the MODIS instrument over the
2000-2008 period for a randomly selected set of sites for boreal and
temperate deciduous vegetation types; step 2: in situ eddy covariance net CO step 3: in situ monthly atmospheric surface CO the ORCHIDEE global LSM, whose main C cycle parameters are optimized (see
Sect. 2.3); the GCM of the Laboratoire de Météorologie
Dynamique, LMDz (see Sect. 2.3), to relate the surface carbon fluxes
to atmospheric CO
The system relies on two models:
The framework combines the different observational data streams within
ORCHIDAS in order to optimize selected model parameters using a variational
data assimilation system, described in Sect. 2.4. Figure 1 illustrates the
structure of the CCDAS and the different components that are involved. Such
a framework distinguishes (i) the assimilated observations, (ii) an ensemble
of forcing and input data streams, (iii) the models and optimization
framework, as well as (iv) an evaluation step, where independent data sets are
compared to the optimized model stocks and fluxes. As explained in the
introduction, a major feature of the current system is the stepwise
approach, in which all data streams are assimilated sequentially (i.e., one
after the other). The information retrieved at a given step (retrieved
optimal parameter values and associated uncertainty) is propagated to the
next step (see Fig. 2 and Sect. 2.4). Note that for simplicity we did
not propagate the error correlations in this first implementation of the
system, a simplification that appeared sufficient (see the consistency
analysis in Sect. 3.2); Sect. 4 also discusses the potential impact of
this simplification.
Illustration of the stepwise data assimilation approach used for the assimilation of multiple data streams in the ORCHIDEE-CCDAS. The list of parameters for each step is summarized in Table 1.
At each step, the parameter optimization relies on a Bayesian framework that
explicitly minimizes the difference between the simulated and observed
quantities in addition to minimizing the difference between the optimized
model parameters and “a priori” values (see Sect. 2.4.2). The
dependence of the simulated quantities on the optimized variables is
nonlinear, which thus necessitates the use of an iterative algorithm. Note
that all components of the surface C budget need also to be included in the
ORCHIDAS, particularly when using atmospheric CO
MODIS collection 5 obtained from surface reflectance data (from 2000 to 2008)
in the red (R) and near-infrared (NIR) bands at 5 km resolution (CMG) are
used to optimize the phenology-related parameters of ORCHIDEE in the first
step. The R and NIR data were processed to correct for directional effects
following Vermote et al. (2009) and then used to calculate the NDVI, which
is assumed to be linearly related to the model FAPAR. The NDVI are then (i) aggregated to the 0.72
Eddy covariance flux measurements of net surface CO
Atmospheric CO
Location of the different observations used for each data stream
assimilated in the system: MODIS-NDVI measurements, FLUXNET sites with NEE
and LE measurements and atmospheric CO
In this study we use the ORCHIDEE process-oriented land surface model
(Krinner et al., 2005), which computes water, carbon and energy balances at
the land surface on a half-hourly time step, using a mechanistic description
of the physical and biogeochemical processes (see,
ORCHIDEE uses the concept of the plant functional type (PFT) to describe the
vegetation distribution, with 13 PFTs (including bare soil) that can co-exist
in each grid cell. Except for the phenology (see a recent description in
MacBean et al., 2015), the equations governing the different processes are
generic, but with specific parameter values for each PFT. Detailed
descriptions of model equations can be found in numerous publications (see
for instance Krinner et al., 2005). ORCHIDEE can be run at either global
scale on a grid, or at site level using point-scale surface meteorological
forcing variables. It is the land surface component of the Institut Pierre
Simon Laplace (IPSL) Earth system model, and the version that we used
corresponds to Coupled Model Intercomparison Project Phase 5
(CMIP5) simulations in the IPCC
5th Assessment Report (Dufresne et al., 2013). However, in this study the
model is run offline using the ERA-Interim 3-hourly near-surface
meteorological forcing fields (Dee et al., 2011) aggregated at the spatial
resolution of the atmospheric transport model for the global simulations
(2.5
The transport model used in this study is version 3 of the GCM, LMDz (Hourdin
and Armengaud, 1999) with a horizontal resolution of 3.75
Parameters description, generality (PFT dependent, global, specific to FLUXNET sites or for a set of regions) and data stream(s) that were used to constrain them.
The optimized parameters are described in Table 1, and their prior values,
uncertainty and range are given in Table 2. In the most recent studies using
ORCHIDAS at site scales, a large set of ORCHIDEE parameters has been
optimized (Kuppel et al., 2014; Santaren et al., 2014; Bacour et al., 2015).
In this study a smaller set was chosen, based on a Morris sensitivity
analysis (Morris, 1991; results not shown) that determines the sensitivity of
the NEE and LE to all model parameters at various FLUXNET sites (for each
PFT), in order to reduce the computational cost of the global optimization in
step 3 (see Sect. 2.5). We considered nine PFT-dependent and four “global”
(i.e., non-PFT-dependent) parameters that control mostly the fast carbon
processes (diurnal to seasonal). In addition, we introduced a new parameter,
Overall (including all PFT-dependent parameters), we optimize 16 parameters related to phenology, 36 to photosynthesis, 3 to respiration, 1 to the energy budget, 78 soil C pool scalars (one for each FLUXNET site) and 30 regional soil C pool scalars for the global simulations – a total of 184 parameters (16, 134 and 86 in step 1, 2 and 3, respectively). Note that the soil C pool multipliers at the FLUXNET sites are independent from the regional C pool multipliers, as the history of soil carbon over large eco-regions of several millions square kilometers is rather heterogeneous (as it is mainly related to previous land use changes) and, most likely, the FLUXNET sites are not representative of larger regions in terms of the soil carbon disequilibrium. The prior standard deviation for each parameter is equal to 40 % of the parameter range (lower and higher boundaries) prescribed for each parameter following Kuppel et al. (2012). The parameter ranges were specified following expert judgment of their meaning in the ORCHIDEE equations and based on literature reviews or databases (such as the global database of plant traits, TRY; Kattge et al., 2011).
Prior information for all parameters except initial soil C pool
multipliers: prior value, uncertainty and range of variation for the
different plant functional types (tropical broadleaf evergreen/rain-green (TrBE/TrBR)
forests, temperate needleleaf/broadleaf evergreen (TeNE, TeBE) forests, temperate broadleaf deciduous (TeBD) forest,
boreal needleleaf evergreen (BoNE) forests, boreal broadleaf/needleleaf deciduous (BoBD/BoND) forests and C
The ORCHIDAS system relies on a stepwise assimilation of the three data streams described in Sect. 2.2. Figure 2 illustrates the flow of information in this sequential approach:
In each step the statistically optimal parameter values are derived with an
optimization procedure following the principle of the 4-D variational
assimilation systems (developed for numerical weather prediction), using a
tangent linear operator (and finite differences for a few parameters, Bacour
et al., 2015). Assuming that the errors associated with the parameters, the
observations and the model outputs follow Gaussian distributions, the
optimal parameter set corresponds to the minimum of a cost function,
The determination of the optimal parameter vector that minimizes
For steps 1 and 2, the model “
For step 3, the model “
For improved minimization efficiency, the inversion is
preconditioned (following Chevallier et al., 2005), which means that
L-BFGS-B is fed with the control variable
The posterior parameter error covariance matrix,
In order to analyze the fit to the atmospheric CO
In this section we describe the other components of the carbon cycle (apart from the surface C exchange with terrestrial vegetation) that are imposed in step 3 of the optimization process as fixed fluxes.
The ocean contributes to an uptake of about one-quarter to one-third of the
anthropogenic emissions with significant year-to-year variations (Sabine et
al., 2004). For this version of the ORCHIDAS, we developed a statistical
model to estimate the spatial and temporal variations (monthly) of the ocean
surface CO
The computation of
We have used a recently developed CO
Fire emissions data from the Global Fire Data (GFEDv3;
Mean seasonal cycle (2000–2008) of the normalized modeled FAPAR
before and after optimization, compared to that of the MODIS NDVI data, for
the temperate and boreal deciduous PFTs (TeBD, BoBD, BoND and NatC3).
Black
The optimization in step 1 resulted in an improved fit to the MODIS NDVI
observations for the four PFTs considered (TeBD, BoND, BoBD, NC
Following the improvement at the sites selected for the optimization, we
evaluated the impact for each PFT at the global scale using the global median
correlation between the MODIS-NDVI and the model FAPAR time series (from all
pixels where the fraction of a given PFT is above 60 %; see Maignan et
al., 2011). The global correlation increased for BoND trees and NC
The optimization in step 2 brings an improvement to the simulated NEE and LE
for all seven PFTs considered, with Fig. 5 showing the corresponding
PFT-averaged mean NEE seasonal cycles (mean across all sites/years). NEE is
overestimated by the prior model for all PFTs on average. This is partly due
to the model spin-up procedure, which brings each simulated site to a near
equilibrium state with a mean NEE close to zero (i.e., no net carbon sink,
see Sect. 2.1.1). This bias is significantly corrected by the
optimization to match the observed carbon uptake at most sites, notably via
the scaling of the initial soil carbon pool content at each site (parameters
Mean seasonal cycle of the net ecosystem exchange (NEE) for the different plant functional type optimized in step 2 of the assimilation. The mean across all sites for a given PFT is provided for the observations (black), the posterior of step 1 (green) and the posterior of step 2 (blue).
Monthly mean atmospheric CO
In addition, the optimization increases the NEE seasonal amplitude in
temperate evergreen (TeNE and TeBE) forests and temperate broadleaf deciduous (TeBD)
forests, and reduces the amplitude for boreal needleleaf evergreen (BoNE) forest
and natural C
The final optimization step with the atmospheric CO
Figure 6 illustrates the simulated concentrations for four stations
(representative of different conditions), over the assimilation period, with
the standard prior parameter vector (used in step 1), the posterior vector
from step 2 (used as prior in step 3) and the posterior vector from this last
step. The improvement in the fit to the observations can be quantified with
the reduction in RMSE (from the prior to the posterior of step 3) – the
largest reduction is at the South Pole station (73 %) and is on average
around 25 % across all sites. Note that for a few stations the fit is
slightly degraded (up to 10 %) except for one Pacific site (regular ship
measurements around the Equator, POCN00) for which there is a 40 %
degradation, possibly due to small biases in the simulation of the
ITCZ (Intertropical Convergence Zone) position in
LMDz. When calculated with respect to the standard prior (used in step 1) the
RMSE decrease is slightly larger on average, especially for the northern
mid-to-high latitude stations. For these stations the optimization performed
in step 2 with FLUXNET data led to a significant improvement of the mean
seasonal cycle amplitude of the atmospheric CO
Changes in the mean seasonal cycle of the atmospheric CO
We then investigated the fit to the observed CO
Finally, we verified that the improvement is valid not only at the
optimization sites but also at independent atmospheric CO
The main issue with a stepwise data assimilation system (vs. a
simultaneous approach) concerns the potential degradation of the model–data fit for the different data streams that are assimilated in previous
steps. We noted that CO
Figure 8 summarizes the performance of the model data fit for MODIS-NDVI and FLUXNET-NEE data streams for the prior and posterior of each step by evaluating the median RMSE between the model and the observations across all sites. The values are calculated for each PFT separately. In this section, we keep in mind the fact that we do not optimize the same PFTs with FLUXNET data and with MODIS-NDVI.
RMSE between model outputs and observations for two types of
observations: MODIS-NDVI on the left and FluxNet-NEE on the right, for
different plant functional types (PFT: TrBE, TeNE, TeBE, TeBD, BoBD, BoND,
NC
First, we notice again the significant RMSE reduction between the prior and
step 1, as discussed in Sect. 3.1. The fit to MODIS-NDVI (normalized data)
for steps 2 and 3 shows only a significant degradation (increased RMSE) for
temperate broadleaf deciduous (TeBD) forest, which decreases the improvement
achieved in step 1 (compared to the prior) by a factor of 2. A marginal
degradation for natural C
Figure 8 again reveals the significant reduction of the RMSEs for NEE in
step 2 compared to the standard prior or to the posterior of step 1 for most
PFTs, except BoNE. We see only small degradations (increases) in RMSE
between steps 2 and 3 for TeNE forests
(from 1.06 to 1.13 g C m
We now discuss the parameter values, focusing on the changes obtained though
the successive steps. Figure 9 presents the prior and posterior values for
each parameter together with their associated uncertainties (estimated
through Eq. 2) and the allowed range of variation. Note that nine parameters
are PFT dependent while four are global (non PFT dependent). For the global
non-PFT-dependent parameters included in the step 2 optimization, we took the
mean value and error variance (see Sect. 2.4) as the prior for step 3. Note
finally that the parameters linked to the initial soil carbon pools
(
If we first consider the phenology parameters optimized in step 1
(
Prior and posterior parameter values and uncertainties for a set of
optimized parameters (nine PFT dependent and four non-PFT dependent). The
prior value corresponds to the horizontal black line and the physical allowed
range of variation to the “
For the photosynthesis parameters (
The non-PFT-dependent respiration-related parameters
(HR
Left: net ecosystem exchange (NEE) for three regions (north of
35
The main objective of a carbon cycle data assimilation procedure is to improve the simulated land surface net and gross carbon fluxes as well as the simulated carbon stocks for both present and future conditions. Given the focus of the paper, i.e., to describe the potential of a stepwise global carbon cycle data assimilation system, we only discuss a few large-scale features of the optimized annual net and gross carbon fluxes, as well as one of the carbon stock variables (forest aboveground biomass). We thus do not discuss the interannual flux variability.
The mean annual carbon fluxes (NEE) for the globe, northern extra tropics,
tropics and southern extra tropics are reported in Fig. 10 for the
2000–2009 decade for the prior and posterior model simulations for all
steps. We ran the optimized model over the full decade in the 2000s in order to
compare with one other estimate of the land surface residual from the GCP (Le Quéré et al., 2015) over the same decade. The
prior NEE indicates a total sink of 0.5 PgC yr
Simulated annual net carbon exchange (NEE) for the land ecosystems
prior to any optimization (left column) and after step 3 of the optimization
process (right column). Upper figures correspond to the mean NEE (in
g C m
Figure 11 shows the spatial distribution of NEE averaged over 2002–2004 for
the standard prior and posterior after step 3. The large tropical net land
carbon sink that is inferred in step 3 is mainly explained by an increase of
the carbon uptake for the tropical forests of the Amazon basin and equatorial
Africa, as well as a decrease of the carbon release on the southern edge of
the Amazon basin (tropical rain-green forests and grasses). In the northern
mid-to-high latitudes only smaller regional changes from the prior occur. For
Europe, most of north Asia and Canada, the strength of the C sink slightly
decreased from the prior (up to 30 g C m
For the GPP the relative changes from the prior are smaller than for the NEE
(Fig. 10b). The mean annual total GPP is 172, 155, 156 and
157 PgC yr
We analyze the impact of the optimization on the forest aboveground biomass at equilibrium (i.e., after spin-up; see Fig. 12) as an example of the impact on model C stocks, and compare the simulated values, for the same three latitude bands than above, to the estimate based on field observations and remote sensing data. This product, which was produced in the GEOCARBON project (and thus is referred to by the same name), integrates a pan-tropical biomass map (Avitabile et al., 2016) with a boreal forest biomass product (Santoro et al., 2015).
Aboveground forest biomass data for the prior ORCHIDEE model and after step 1, step 2 and step 3 of the optimization process. Estimates from satellite observations (Santoro et al., 2015) and referred as “GEOCARBON” (following the EU-GEOCARBON project) are provided for comparison.
For the northern extra tropics, the prior aboveground C stock
(
For the tropics, while there is nearly no change with the assimilation of
MODIS-NDVI in step 1, the use of FLUXNET data leads to a significant increase
of the forest aboveground biomass (close to 25 %). Such an increase does
not correspond to an increase of the GPP (Fig. 10) but to changes in the
autotrophic respiration parameter (MR
In this paper we have described a first global carbon cycle data
assimilation system that assimilates three major carbon-cycle-related data
streams, namely MODIS-NDVI observations of vegetation activity at 60 sites,
FLUXNET NEE and LE measurements, at more than 70 sites, and atmospheric
CO
The assimilation of MODIS-NDVI (60 grid-cell points, step 1) improved the
phenology of ORCHIDEE with a significant reduction of the growing season
length and thus a direct impact on the GPP. The results are similar to those
presented in MacBean et al. (2015), who describe the impact of such
optimization on the global FAPAR simulations and the improvement in the bias
of the calculated leaf onset and senescence dates in more detail. The
optimization with FLUXNET data (78 sites, step 2) led to large improvements
in the seasonal cycle of the NEE and LE fluxes, constraining primarily the
photosynthetic processes. Some discrepancies remain due to site
heterogeneity (i.e., different species and edaphic conditions) that the model
does not fully capture, and due to missing processes in the model (see
Kuppel et al., 2014 for a more thorough discussion). However, without the
assimilation of atmospheric CO
The consistency of the stepwise approach has been evaluated with
back-compatibility checks after the final step (step 3: assimilation of
atmospheric CO
Most of the optimized parameter values have significantly changed compared
to their prior values, with a large error reduction for most (Fig. 9) that
results in a strong constraint on the simulated fluxes (Fig. 11). In the
last step, the assimilation of atmospheric CO
Nonetheless, several limitations, inherent to the optimization of model
parameters in a CCDAS, need to be called to mind when evaluating these
results (see also Rayner, 2010). First, the structure of the land surface
model (i.e., how biogeochemical processes are represented) is critical. Any
missing/misrepresented processes may have a direct impact and thus lead to
biases in the selected parameters. Note that this limitation could be even
more severe when using atmospheric CO
To conclude, this work is a step forward in terms of multiple
data streams assimilation that opens new
perspectives for a better understanding of the carbon cycle and better
predictions of the fate of the land carbon sink in the 21st century as a
consequence of anthropogenic changes. As ORCHIDEE is part of the IPSL Earth
system model the impact of the optimization on future climate change
predictions will be assessed in a future study. However, we first need to run
the ORCHIDAS with a longer atmospheric CO
The ORCHIDEE model code and the run environment are open source
(
Figure A1 displays the air–sea fluxes from the statistical model.
In order to account for fundamental differences between six fire flux
categories provided by the GFED product, we grouped these emissions into
three types with specific treatments. The first group includes C emissions
from deforestation and peat fires, which are considered to be permanent
carbon lost to the atmosphere, at least over the considered timescales. These
fluxes are rescaled to an annual emission of 1.1 PgC yr
Figure A2 provides the optimized values of the
CO
Map of the posterior values of the coefficient scaling the initial carbon pool sizes per regions.
This work was mainly funded by the EU FP7 CARBONES project (contracts
FP7-SPACE-2009-1-242316), with also a small contribution from GEOCARBON
project (ENV.2011.4.1.1-1-283080). This work used eddy covariance data
acquired by the FLUXNET community and in particular by the following
networks: AmeriFlux (U.S. Department of Energy, Biological and Environmental
Research, Terrestrial Carbon Program; DE-FG02-04ER63917 and
DE-FG02-04ER63911), AfriFlux, AsiaFlux, CarboAfrica, CarboEuropeIP,
CarboItaly, CarboMont, ChinaFlux, Fluxnet-Canada (supported by CFCAS, NSERC,
BIOCAP, Environment Canada, and NRCan), GreenGrass, KoFlux, LBA, NECC,
OzFlux, TCOS-Siberia, USCCC. We acknowledge the financial support to the eddy
covariance data harmonization provided by CarboEuropeIP, FAO-GTOS-TCO,
iLEAPS, Max Planck Institute for Biogeochemistry, National Science
Foundation, University of Tuscia, Université Laval and Environment Canada
and US Department of Energy and the database development and technical
support from Berkeley Water Center, Lawrence Berkeley National Laboratory,
Microsoft Research eScience, Oak Ridge National Laboratory, University of
California-Berkeley, University of Virginia. Philippe Ciais acknowledges
support from the European Research Council through Synergy grant
ERC-2013-SyG-610028 “IMBALANCE-P”. The authors wish to thank M. Jung for
providing access to the GPP MTE data, which were downloaded from the
GEOCARBON data portal
(