Natural land carbon dioxide exchanges in the ECMWF integrated forecasting system: Implementation and offline validation



[1] The European Centre for Medium-Range Weather Forecasts land surface model has been extended to include a carbon dioxide module. This relates photosynthesis to radiation, atmospheric carbon dioxide (CO2) concentration, soil moisture, and temperature. Furthermore, it has the option of deriving a canopy resistance from photosynthesis and providing it as a stomatal control to the transpiration formulation. Ecosystem respiration is based on empirical relations dependent on temperature, soil moisture, snow depth, and land use. The CO2 model is designed for the numerical weather prediction (NWP) environment where it benefits from good quality meteorological input (i.e., radiation, temperature, and soil moisture). This paper describes the CO2 model formulation and the way it is optimized making use of off-line simulations for a full year of tower observations at 34 sites. The model is then evaluated against the same observations for a different year. A correlation coefficient of 0.65 is obtained between model simulations and observations based on 10 day averaged CO2 fluxes. For sensible and latent heat fluxes there is a correlation coefficient of 0.80. To study the impact on atmospheric CO2, coupled integrations are performed for the 2003 to 2008 period. The global atmospheric growth is well reproduced. The simulated interannual variability is shown to reproduce the observationally based estimates with a correlation coefficient of 0.70. The main conclusions are (i) the simple carbon dioxide model is highly suitable for the numerical weather prediction environment where environmental factors are controlled by data assimilation, (ii) the use of a carbon dioxide model for stomatal control has a positive impact on evapotranspiration, and (iii) even using a climatological leaf area index, the interannual variability of the global atmospheric CO2 budget is well reproduced due to the interannual variability in the meteorological forcing (i.e., radiation, precipitation, temperature, humidity, and soil moisture) despite the simplified or missing processes. This highlights the importance of meteorological forcing but also cautions the use of such a simple model for process attribution.

1 Introduction

[2] Earth system models are aimed at describing the behavior of the global energy, water, and carbon cycles and enhancing our capacity to monitor and predict natural resources and their evolution in time. These models often stem from numerical weather prediction (NWP) research, which involves introducing key climate processes and variables into models to perform realistic long-term simulations. Extending such models with natural land CO2 fluxes is a natural step because it allows a coupling between energy, water, and carbon cycles and the modeling of atmospheric CO2 concentrations [Baker et al., 2003; Krinner et al., 2005; Sitch et al., 2003; Bonan and Levis, 2010; Clark et al., 2011]. In addition, modeling the CO2 variability provides benefits for atmospheric data assimilation [Engelen et al., 2009] and monitoring of atmospheric composition [Peuch and Engelen, 2012].

[3] Land CO2 exchange is not only the result of complex vegetation processes but also known to be influenced by hydrological and meteorological state variables [Goudriaan et al., 1985; Farquhar and Sharkey, 1982; Ball et al., 1987; Bonan et al., 2003], such as soil moisture and soil temperature, and near-surface atmospheric conditions. CO2 modeling may therefore benefit from taking advantage of the accuracy of weather and land parameters in the NWP environment.

[4] Improving land surface representation of soil moisture increases the skill of weather forecasts [Betts et al., 1996; Beljaars and Viterbo, 1999] up to the subseasonal time scale [Koster et al., 2010] and has involved a continuous effort in the NWP context [e.g., Boussetta et al., 2008]. More recently, the vegetation layer, via its impact on radiation and rainfall partitioning (e.g., interception and transpiration processes), is also shown to play an important role in the surface-atmosphere exchanges for NWP forecasts [Boussetta et al., 2011] up to seasonal-to-interannual time scales [Cox et al., 2000; Van den Hurk et al., 2003].

[5] In this study, the European Centre for Medium-Range Weather Forecasts (ECMWF) operational land surface scheme, HTESSEL [Balsamo et al., 2011], is coupled with a photosynthesis-conductance (A-gs) model. This model is a modified version of the Jacobs [1994] scheme and takes into account the effects of soil water stress on photosynthesis and canopy resistance [Calvet et al., 1998; Calvet, 2000; Lafont et al., 2012]. The A-gs model links the leaf photosynthesis rate, and thereby the stomatal conductance, to external, surface and atmospheric factors. It describes the main responses of ecosystem-scale CO2 fluxes to atmospheric forcing parameters (e.g., temperature, humidity, radiation, and soil moisture) and to a fixed vegetation state via the climatology of the leaf area index (LAI). The land surface carbon pool is parameterized in a fairly simple way with a dependency on prescribed vegetation types/cover, soil moisture, soil temperature, and snow depth.

[6] The lack of a prognostic carbon pool limits the model's suitability for very long integrations (e.g., for the Intergovernmental Panel on Climate Change-IPCC multidecadal simulations); nevertheless, it would be suitable for monitoring applications within the NWP environment where relevant prognostic variables are kept under control by data assimilation or climatological data sets. This kind of system is, for example, applied to global reanalysis such as ERA-Interim [Dee et al., 2011]. Also, it provides an advanced analysis of atmospheric variables [Simmons et al., 2010], soil moisture and surface temperature using a wide range of observations.

[7] The main objective of the present study is to verify whether this parameterization albeit simple, especially its ecosystem respiration component, is capable of (i) producing the CO2 fluxes with reasonable accuracy compared to flux-tower observations and (ii) providing a meaningful evolution of CO2 over a period of a few years when integrated into atmospheric CO2 simulations. Furthermore, previous studies disagreed on the impact of having a photosynthesis-dependent canopy conductance for the surface fluxes. For instance, Lawrence et al. [2011] found that introducing a vegetation phenology had a detrimental effect on the global evapotranspiration while Kumar et al. [2011] found that a physically based canopy conductance improves the turbulent fluxes. The current study also aims to investigate the impact of such development on the surface fluxes within the ECMWF model. The CO2 model is evaluated at the local scale using flux-tower observations, and at the global scale using land CO2 flux products and the observed annual growth rate of CO2 in the atmosphere. Additionally, consideration is given to the ability of the model to reasonably simulate the CO2 fluxes based on a monthly LAI climatology and on the relative importance of the atmospheric forcing and vegetation state. These are important issues because the answers will give guidance for research priorities. For instance, it might be important to have observational information about interannual variability of vegetation state.

[8] The modeling components are briefly presented in section 2 (with details of equations in the appendices), while the vegetation state is described in section 3. The flux tower field sites from which observations were obtained and the model parameter were calibrated are presented in section 4. Model evaluation is carried out in an offline mode for various flux tower sites, with results presented in section 5 for both the turbulent energy and CO2 fluxes using different versions of the model. The impact of the CO2 fluxes on atmospheric CO2 simulations is also shown. Section 6 presents the conclusions of this study and discusses the implications of introducing vegetation dynamics into a global model.

2 The Modeling Components

[9] The backbone of this study is the land surface model, HTESSEL, which incorporates several years of research and development [Viterbo and Beljaars, 1995; Viterbo et al., 1999; Van den Hurk and Viterbo, 2003; Balsamo et al., 2009; Dutra et al., 2010; Boussetta et al., 2011; Balsamo et al., 2011]. This scheme has been extended to include modules dealing with photosynthesis and CO2 emission. A detailed technical description of HTESSEL is provided in the full scientific and technical documentation of the ECMWF Integrated Forecasting System (IFS) [ECMWF, 2011].

[10] For the computation of the plant transpiration, most of the operational land surface models use either the Jarvis approach [Jarvis, 1976] or a plant physiological approach [Farquhar and Sharkey, 1982; Goudriaan et al., 1985; Ball et al., 1987; Collatz et al., 1992] to estimate the stomatal conductance (gs). The basic assumption in the Jarvis approach is that the various environmental factors (i.e., soil moisture, temperature, humidity, solar radiation, CO2) have a mutually independent impact on gs, which can be parameterized as a simple product of functions representing these factors. On the other hand, the plant-physiological approach is based on the fact that CO2 taken up for photosynthesis (An) largely uses the same pathway as water transpired from the plants: the stomata. By linking the plant transpiration flux to An through gs, account can be taken of the interactions between the environmental factors that affect gs.

[11] Two new land model versions are presented here: (i) a fully coupled version (labeled CTESSEL) in which the plant-physiological approach (A-gs) is used to compute the stomatal conductance for water vapor transpiration and (ii) an uncoupled version (labeled CHTESSEL) which includes the CO2 module but calculates evaporation using the stomatal conductance formulation based on the original Jarvis approach.

[12] In order to obtain the CO2 balance at the ecosystem scale, the A-gs model is coupled to an ecosystem respiration module with a dependency on surface temperature [Norman et al., 1992] and soil moisture. This module is modified to take into account the effect on soil respiration of cold regions and snow packs (see section B).

2.1 The Jarvis Approach

[13] The Jarvis-type approach is used in many land surface models for NWP due to its straightforward formulation. Empirical stress functions (with values ranging between 0 and 1) depend on environmental conditions and are used to modulate a preset maximum stomatal conductance, which regulates the water vapor flux. One of the hypotheses behind this formulation is that the stress functions are independent of each other. In HTESSEL, the following formulation of the canopy conductance, gc [m s−1], is adopted:

display math(1)

with gs,max [m s−1] the vegetation type-dependant maximum stomatal conductance, LAI [m2 m−2] the leaf area index, and f1, f2, f3, three dimensionless stress functions expressing the effects on gc of shortwave radiation, soil moisture stress, and atmospheric humidity deficit, as described in section A. The original Jarvis formulation also includes a temperature dependency, but it was not used in HTESSEL. Temperature, shortwave radiation, soil moisture, and atmospheric humidity tend to be highly correlated and therefore the additional temperature dependency in the stress function has very little impact.

2.2 The Photosynthesis Scheme (A-gs)

[14] The A-gs approach is based on plant-physiological considerations and describes the plant photosynthesis process and its dependence on CO2, temperature, and soil moisture [Jacobs, 1994; Jacobs et al., 1996; Calvet et al., 1998, 2004]. The stomatal conductance is active for regulating both water vapor and CO2 fluxes. The gross CO2 assimilation by the canopy Ag [kg CO2 m−2 s−1] is calculated using a photosynthesis module following Goudriaan et al. [1985]. The net assimilation An [kg CO2 m−2 s−1] (i.e., the net flow of CO2 through the stomata) is Ag minus the dark respiration Rd [kg CO2 m−2 s−1]. Once the net assimilation is known, the stomatal conductance for CO2, gsc [m s−1], can be derived by Kirchhoff's resistance/conductance analogy. It is defined as the net flow of CO2 through the stomata divided by the difference between the CO2 concentration outside the leaves Cs [kg CO2 m−3] and the concentration in the intercellular cavities, Ci [kg CO2 m−3] (see Figure 1)

display math(2)

where the functions (Env) indicate the dependence of An and Ci on the various environment factors. The description of the gsc formulation is detailed in section B. In the CTESSEL version, the stomatal conductance from the CO2 module is also used for water vapor transpiration, which leads to a coupling between CO2 and water vapor fluxes.

Figure 1.

Schematic representation of a leaf with the resistance analogies for carbon and water vapor, where gcu is the cuticular conductance, gsc the stomatal conductance to CO2, gs the stomatal conductance to water vapor, gm the mesophyll conductance, Cs and Ci the CO2 concentration at the canopy surface and inside the leaf cavity, respectively, and qs and qsat(Ts) the humidity at the canopy surface and the saturated humidity at the temperature Ts of the canopy surface, respectively.

2.3 The Ecosystem Exchanges Parameterization

[15] In order to obtain the net exchange of CO2 between the land surface and atmosphere by a NWP model, ecosystem respiration needs to be represented. Schemes relying on prognostic land carbon pools (BEPS-InTEC [Chen et al., 2000]; CLM-DGVM [Levis et al., 2004]; JULES [Clark et al., 2011]; LPJ [Sitch et al., 2003]; NCAR-DGVM [Bonan et al., 2003]; ORCHIDEE [Krinner et al., 2005]) are not really practical for NWP purposes because they are difficult to initialize without a very long spin-up. In CTESSEL/CHTESSEL the CO2 biomass assimilation gross primary production (GPP) [kg CO2 m−2 s−1] and the ecosystem respiration (Reco) [kg CO2 m−2 s−1] are parameterized as described in section B. The net ecosystem exchange (NEE) of CO2 [kg CO2 m−2 s−1] between the surface and atmosphere is given by

display math(3)

[16] This quantity is the main observed variable in field experiments.

3 Vegetation Description

[17] The state of vegetation is given by the LAI; it is crucial for deriving the plant assimilation and transpiration activity. In this study a climatology based on satellite observations (also used operationally in the IFS) was considered for the representation of LAI. The satellite product (MOD15A2) is derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument on board Terra satellite. It is produced daily for the land surfaces at 1 km spatial resolution from the MODIS spectral reflectance with a global coverage, and synthesized on an 8 day interval based on simultaneously retrieved maximum Fraction of Absorbed Photosynthetically Active Radiation to remove the atmospheric noise [Myneni et al., 2002].

[18] The collection 5 of the product (released in 2008 available from February 2000 to present) is used in this study. To derive the climatological time series, 9 years of data (2000−2008) were reprojected from the sinusoidal to a geographic regular lat/long projection, spatially averaged to 1/12th degree resolution, then temporally smoothed and monthly averaged [Jarlan et al., 2008], and finally interpolated to the IFS reduced Gaussian grid. The MODIS LAI products have been analyzed and validated in previous studies [Garrigues et al., 2008; MODIS Land team; Weiss et al., 2012]. After a positive assessment of this product within the IFS [Boussetta et al., 2011], it was adopted by ECMWF for operational use. As a first approach, this climatology is tested within CTESSEL/CHTESSEL to drive the CO2 flux module.

[19] The land use classification follows from the global land cover characteristics (GLCC) data [Loveland et al., 2000]. Use is made of the biosphere-atmosphere transfer scheme classification to assign dominant high and low vegetation types and associated parameters within each grid box (as detailed in Tables 1 and 2).

Table 1. Parameter Values as Specified in the Optimized CTESSEL Modela
Type CodeVegetation TypeR0[mgCO2 m−2 s−1]gm*(25)[mm s−1]gc [mm s−1]Dmax* [kg kg−1]Am,max(25) [mgCO2 m−2 s−1]fo* [-]Γ(25) [ppm]
  1. aThe vegetation types are from the GLCC database and used in the same way as in TESSEL [Van den Hurk et al., 2000].
1Crops, mixed farming0.1001.30.15Equation (B18)2.200.8542
2Short grass0.0801.30.20Equation (B18)3.000.6542
3Evergreen needle-leaf0.3600.80.200.1242.20Equation (B22)42
4Deciduous needle-leaf0.3300.80.200.1242.20Equation (B22)42
5Deciduous broadleaf0.2801.40.000.1091.83Equation (B22)42
6Evergreen broadleaf0.2701.10.250.1241.83Equation (B22)42
7Tall grass0.1502.30.20Equation (B18)1.830.702.6
9Tundra0.3602.00.25Equation (B18)3.000.9542
10Irrigated crops0.0961.40.25Equation (B18)1.830.9242
11Semidesert0.0191.00.25Equation (B18)1.830.8042
12Ice caps and glaciers-------
13Bogs and marshes0.2700.50.25Equation (B18)1.830.9642
14Inland water-------
16Evergreen shrubs0.1100.90.15Equation (B18)1.830.722.6
17Deciduous shrubs0.0801.90.20Equation (B18)1.830.9642
18Mixed forest- Wood0.4201.00.000.1242.20Equation (B22)42
19Interrupted forest0.1600.80.100.1242.20Equation (B22)42
20Water -land mixtures0.2701.00.25Equation (B18)1.830.9542
Table 2. CTESSEL Temperature Response, Quantum Use Efficiency, and Soil Moisture Stress Parameters
T1gm (°C)5(13 for veg. types 7 and 16)
T1Am,max (°C)8(13 for veg. types 7 and 16)
T2gm (°C)36 
T2Am,max (°C)38 
a2.381(5.323 for veg.types 7 and 16)
b−0.6103(−0.8923 for veg.types 7 and 16)
inline image300 
εo (mg CO2/J PAR)0.0142(0.0117 for vegetation types 7 and 16)

4 Verification Data and Parameter Calibration

4.1 Data From Observations

[20] Observational data for the 2004 and 2006 from the Boreal Ecosystem Research and Monitoring Sites (BERMS) [Betts et al., 2006], FLUXNET eddy-covariance network [Baldocchi et al., 2001; Baldocchi, 2008], and Coordinated Energy and water cycle Observations Project (CEOP) were used in this study.

[21] For many years, BERMS has been providing high-quality data, which is especially useful for model evaluation and parameter optimization. The BERMS sites used in this study consist of observations from two contrasting locations less than 100 km apart in Saskatchewan at the southern edge of the Canadian boreal forest (at about 54°N/105°W). The two sites are the Old Aspen (deciduous, open canopy, hazel under-story) and Old Black Spruce (evergreen needle-leaf and boggy, moss understory).

[22] As part of the CEOP program, reference site observations from the Amazonian region, also belonging to the Large Scale Biosphere-Atmosphere Experiment in Amazonia experiments, are available for scientific use. In this study, observations are taken from flux towers located within an evergreen broadleaf forest (Manaus) and a woody savannah region (Brasilia).

[23] Furthermore, 30 sites from LaThuile in situ observations (, which are available to the research community and belong to the FLUXNET eddy covariance network, are also used in this study. This data set provides latent heat flux, sensible heat flux, and NEE at high temporal resolution (30 to 60 min). Additionally, the LaThuile data set provides ecosystem respiration (Reco) and GPP, which are derived from NEE using a partitioning scheme. To find the daytime Reco, the nighttime values are extrapolated using an empirical function that includes a temperature dependency. GPP is then computed by subtracting Reco from NEE. This algorithm also considers the seasonal to diurnal variability of Reco, which has been subject of extensive research [Reichstein et al., 2005]. Although the derived GPP and Reco data are affected by uncertainties owing to the partitioning algorithm, they are considered as state-of-the-art estimates of these quantities. In this study, only observations flagged as high-quality data are used. Also, as stated above, the chosen observation sites cover a range of climates and ecosystem zones mostly located at the Northern Hemisphere midlatitudes (North America and Europe), with the addition of two tropical stations. Sites from high-latitude regions were not available. The sites used for optimization and subsequent evaluation are listed in Table 3.

Table 3. List of Sites Used for the Verification of the Simulated Fluxes, Where the Biome Types Are Deciduous Broadleaf Forest (DBF), Evergreen Broadleaf Forest (EBF), Deciduous Needle-Leaf Forest (DNF), Evergreen Needle-Leaf Forest (ENF), Mixed Forest (MF), Woody Savannah (WSA), Grasslands (GRA), Crops (CRO), and Wetlands (WET)
NumberSiteNetworkLat [°N]Lon [°E]Vegetation TypeReference/PI
1sk-oaberms53.63−106.20DBFT. Andrew Black
2sk-obsberms53.99−105.12ENF/WETT. Andrew Black
3brasiliaceop−15.93−47.92WSA/GRA/SHAntonio Ocimar Manzi
4manausceop−2.61−60.21EBFAntonio Ocimar Manzi
5at-neufluxnet47.1211.32GRAWohlfahrt et al. [2008a, 2008b]
6ca-merfluxnet45.41−75.52WETLafleur et al. [2003]
7ca-qfofluxnet49.69−74.34ENFBergeron et al. [2007]
8ca-sf1fluxnet54.49−105.82ENFM.S. Mkhabela et al. [2009]
9ca-sf2fluxnet54.25−105.88ENFM.S. Mkhabela et al. [2009]
10ch-oe1fluxnet47.297.73GRAAmmann et al. [2007]
11fi-hyyfluxnet61.8524.29ENFSuni et al. [2003]
12fr-hesfluxnet48.677.06DBFGranier et al. [2000]
13fr-lbrfluxnet44.72−0.77ENFBerbigier et al. [2001]
14il-yatfluxnet31.3435.05ENFGrünzweig et al. [2003]
15it-ampfluxnet41.9013.61GRAGilmanov et al. [2007]
16it-cpzfluxnet41.7112.38EBFGarbulsky et al. [2008]
17it-mbofluxnet46.0211.05GRAMarcolla and Cescatti [2005]
18it-ro1fluxnet42.4111.93DBFRey et al. [2002]
19it-ro2fluxnet42.3911.92DBFTedeschi et al. [2006]
20nl-ca1fluxnet51.974.93GRAGilmanov et al. [2007]
21nl-haafluxnet52.004.81GRAEddy Moore
22nl-horfluxnet52.035.07GRAJacobs et al. [2007]
23nl-loofluxnet52.175.74ENFDolman et al. [2002]
24ru-fyofluxnet56.4632.92ENFKurbatova et al. [2008]
25ru-ha1fluxnet54.7390.00GRABelelli et al. [2007]
26ru-ha3fluxnet54.7089.08GRABelelli et al. [2007]
27se-sk2fluxnet60.1317.84ENFAnders Lindroth
28us-armfluxnet36.61−97.49CROFischer et al. [2007]
29us-barfluxnet44.06−71.29DBFJenkins et al. [2007]
30us-ha1fluxnet42.54−72.17DBFUrbanski et al. [2007]
31us-mmsfluxnet39.32−86.41DBFSchmid et al. [2000]
32us-syvfluxnet46.24−89.35MFDesai et al. [2005]
33us-tonfluxnet38.43−120.97MF/WSAMa et al. [2007]
34us-varfluxnet38.41−120.95GRAMa et al. [2007]

4.2 Offline Simulations

[24] The offline (or stand-alone) simulations offer a convenient framework for studying the benefits and deficiencies of a given land surface parameterization without having to consider complex surface/atmosphere interactions as is necessary in a coupled mode. Additionally, the computational costs of offline simulations are much lower than those of coupled integrations. Offline simulations are also much faster to run.

[25] In this study, offline runs were performed both at the global scale and for specific sites, and all the land simulations were forced with 3-hourly meteorological data extracted from the ECMWF ERA-Interim (ERA-I) reanalysis [Dee et al., 2011], which covers the period from 1979 to present. These forcing data are gridded on a reduced Gaussian grid (N128) corresponding to a resolution of about 80 km. The temperature, surface pressure, humidity, and wind fields are instantaneous values and representative of the lowest level in the atmospheric model corresponding to a height of 10 m above the surface. The incoming surface radiation (in its longwave and shortwave components), rainfall, and snowfall are provided as 3-hourly accumulations. The instantaneous fields are linearly interpolated in time to the 30 min time-step of the land surface model. They are from the 3, 6, 9, and 12 h forecasts starting from the daily analyses at 00 and 12 UTC. As a compromise between spin-up effects (mainly in radiation) and forecast errors, the fluxes are averages from the forecast intervals 9−12, 12−15, 15−18, and 18−21 h starting from the daily analyses at 00 and 12 UTC (see Kallberg [2011] for a discussion on the spin-up characteristics of ERA-I). Fluxes and instantaneous fields are matched by verification time. Precipitation is kept constant over the 3-hourly interval, long-wave downward radiation is linearly interpolated and downward solar radiation is disaggregated in time making use of the solar angle, but conserving the 3-hourly integral.

[26] For the global simulation (section 5.2), the land-use information has been derived from the GLCC data set and aggregated to the same resolution as the forcing data. When performing the optimization procedure for site simulations (section 4.3), the land-use information was set to correspond to the specific status of the site to ensure that the optimization is applied to the correct vegetation type. However, for validation (section 5.1), the vegetation type of the GLCC data was taken from the 80 km grid-box and no attempt was made to tune the derived vegetation characteristics to specific site conditions. Therefore, the evaluation includes both the model and the supporting vegetation data, and should be indicative of the expected accuracy of a global model.

4.3 Parameter Optimization

[27] During its development, the A-gs model was initially designed and tested for a single field location [Jacobs et al., 1996; Calvet et al., 1998], followed by an extension to regional and global domains [Gibelin et al., 2006; Brut et al., 2009] using the ECOCLIMAP vegetation database [Masson et al., 2003]. This uses seven vegetation types including the C3/C4 types, which are linked to the plant functional behavior under various climate conditions.

[28] To accommodate the incorporation of this land CO2 module in the operational HTESSEL, the model parameters have to be compatible with the 20 vegetation types of the biosphere-atmosphere transfer scheme classification used in the GLCC data [Loveland et al., 2000]. This classification does not distinguish between C3 and C4, but with the much larger number of biomes globally distributed over different climate regions, it is believed that functional classes are reasonably well represented. The model parameters for each vegetation type were optimized making use of observations representative for that vegetation type. The eddy covariance sites with similar vegetation type were grouped together and the parameter optimization was performed separately on each of those groups for 2006. The largest observation coverage was in 2004, but these data are kept for verification only to ensure reasonable independence from the calibration. Default parameters extracted from previous studies [White et al., 2000; Calvet, 2000; Calvet et al., 2004] were assigned to the vegetation types that are not represented by the available eddy covariance sites (Tundra, interrupted forest, and evergreen shrubs).

[29] The parameters for which the CO2 fluxes showed a higher sensitivity were chosen for optimization. These parameters are the unstressed mesophyll conductance, inline image, and the reference respiration, R0. For the remaining model parameters, values from the literature [White et al., 2000; Calvet, 2000; Calvet et al., 2004] were assigned.

[30] Because GPP is computed independently from the ecosystem respiration (Reco) (equation (B39)), the parameter optimization procedure was performed in two steps. First, there is an estimate of inline image for which the GPP calculation showed a high sensitivity. Subsequently, the optimization is applied to Reco; this is highly sensitive to the reference respiration R0, which is derived by minimizing the root mean square error (RMSE) between observed and simulated CO2 fluxes (GPP for inline image and Reco for R0). For each site within the same vegetation type group, the optimized parameters were allowed to vary within a fixed range chosen from the literature. Using this procedure, the optimized parameters converged toward values yielding minimum errors with respect to the observed fluxes for each vegetation type. An illustrative example is shown in Figure 2 for the needle-leaf forest type where a minimum RMSE is obtained between observed and simulated ecosystem respiration for R0 = 0.36 [mgCO2 m2 s−1], while the errors between observed and simulated GPP converge to an optimal value of 0.80 [mm s−1] for inline image. The sites used in this optimization procedure are listed in Table 3 and the results of the optimized value for R0 and inline image are presented in Table 1 together with the other model parameters extracted from previous studies described in Calvet [2000], Calvet et al. [2004], and White et al. [2000].

Figure 2.

RMSE of Reco as a function of the reference respiration R0 (left) and the RMSE of GPP as a function of the mesophyll conductance inline image (right). These curves apply to the evergreen needle-leaf forest type. The minimum values of these curves have been selected as optimal parameters, i.e., R0 = 3.6 10−7 [kg CO2 m−2 s−1] and inline image = 0.0008 [m s−1].

5 Results

[31] To assess the quality of the land surface scheme, CTESSEL and its variant, CHTESSEL, in situ and global integrations of the model were performed. The in situ simulations consisted of single-point off-line simulations for the available FLUXNET, BERMS, and CEOP sites covering various types of vegetation for the year 2004, which had the largest observation coverage of the available records (34 stations). Both the energy and the carbon cycles are evaluated. The global offline runs are compared for the period 2003 to 2008 with the Carnegie-Ames-Stanford-Approach (CASA) driven Global Fire Emissions Database (GFED3.0) product (CASA-GFED3) [Potter et al., 1993; Van der Werf et al., 2010], which is the result of a CO2 model to be described later in this section. The purpose is twofold. First, the relative performance of CASA-GFED3 versus CHTESSEL is of interest, because it is the intention to replace CASA-GFED3 fluxes by CHTESSEL in the ECMWF system. Second, it is of interest to assess the differences between results from CASA-GFED3 and CHTESSEL in view of the differences in forcing of the two systems. The forcing in CASA-GFED3 is monthly and relies heavily on vegetation state and radiation, whereas for CHTESSEL, the quality of soil moisture, meteorology, and radiation at the 3-hourly time scale is a key feature. Finally, the CO2 fluxes from CHTESSEL are used as boundary conditions in the global transport model of the Monitoring of Atmospheric Composition and Climate project to evaluate the impact of simulated NEE on the atmospheric CO2 evolution.

5.1 In Situ Simulations

[32] Blyth et al. [2010] suggests that even with as few as 10 observation sites over different climate regimes a reasonable impression of the performance of a land surface model can be obtained. Nevertheless, a bigger number of in situ observations covering all climate regions and vegetation types is always beneficial to minimize the representativity issues related to site observation. In this study, results and performance metrics of 34 stations covering various biomes and climate zones (section 3) are presented. A selection of three stations representing crops, grass, and forest (which are dominant vegetation types in terms of global coverage), is used to show time series for illustrative purposes.

5.1.1 Energy Fluxes

[33] A comparison has been made between observed and simulated energy fluxes for all the stations for the two model versions, i.e., using the Jarvis approach CHTESSEL or the A-gs approach CTESSEL. Figure 3 shows the correlation of the two models against the eddy-covariance measurements of the 10 day averaged energy fluxes. Overall, both model versions show high correlations (>0.80 on average) over all biomes except for the Manaus tropical station (Figure 3c). Here the radiative forcing suffers from a known cloudiness bias over the Intertropical Convergence Zone (ITCZ) region [Dee et al., 2011]. Overall, an important improvement for the energy fluxes (Table 4) was achieved with CTESSEL using the photosynthetic-based A-gs formulation compared to CHTESSEL using the Jarvis approach. The latent heat flux (LE) with CTESSEL has an RMSE of 22.7 [W m−2], and a bias of 13.4 [W m−2] against an RMSE of 28.4 [W m−2], and a bias of 19.5 [W m−2] when using CHTESSEL. For the sensible heat flux (H), CTESSEL is also better mainly with a smaller a bias of −2.9 [W m−2] against −9 [W m−2] for CHTESSEL.

Figure 3.

Correlation for 2004 of the 10 day averaged simulated energy fluxes with the eddy-covariance observations over the 34 sites using N = 36 data points for each site. Black bars are for CHTESSEL runs and dark grey bars are for the CTESSEL results. (a) Latent heat flux, (b) sensible heat flux, and (c) net radiation. Blank space in the graph refers to no or incomplete observations at the station for that parameter.

Table 4. Average Performance Metrics for 2004 of the 10 day Averaged Simulated Energy Fluxes of CTESSEL and CHTESSEL Over the 34 Sites
ModelLE RMSE [W m−2]LE Bias [W m−2]LE Corr -H RMSE [W m−2]H Bias [W m−2]H Corr -NR RMSE [W m−2]NR Bias [W m−2]NR Corr -

[34] The annual cycle of latent and sensible heat fluxes has been compared with in situ observations for both model versions for three sites with different vegetation types: crops at the Us-Arm site, grassland at the It-Mbo site, and evergreen needle-leaf forest at the Berms.Sk-Obs site. Overall, as shown in Figure 4, both models perform reasonably well at the seasonal scale, with relatively small differences between the three sites. This is an encouraging result given the fact that the Jarvis approach has benefited from a long history of model evaluation and optimization because it has been used in ECMWF's operational model for many years, whereas the A-gs did not benefit from such operational evaluation. Latent and sensible heat fluxes have biases with opposite sign and are both reduced with CTESSEL compared to CHTESSEL (Table 4). On the other hand, there is a slight deterioration in terms of bias and RMSE in both models for some stations; these are identified as managed sites like the Italian Monte Bondone site (It-Mbo), which has a grazing season in summer, or the Austrian Neustift/Stubai Valley site (At-Neu), which has two grass clipping events. Mediterranean forest sites also showed a deterioration in the skill; this problem was mentioned in previous studies [Migliavacca et al., 2011] and attributed to a specific plant functional type of the trees within the arid area of the Mediterranean and which is not present in the GLCC vegetation type classification.

Figure 4.

Seasonal cycle (2004) of 10 day averaged simulated (lines) and observed (dots) latent heat flux [W m−2] (left) and sensible heat flux [W m−2] (right) for CTESSEL (with A-gs, grey line) and CHTESSEL (with Jarvis-type evaporation, black line) at different observation sites with different biomes: (a) crops (Us-Arm), (b) grassland (It-Mbo), and (c) evergreen needle-leaf forest (Berms-Obs).

[35] The average diurnal cycle in UTC time (not local) of the latent and sensible heat fluxes in July for the same selected sites has been investigated. It is found that, although, the latent heat flux for crops and grass sites shows reasonable agreement with the observations, the forest site tended to overestimate the midday value by about 50 [W m−2]; these tendencies were reversed for the sensible heat flux, which was underestimated at the forest site and overestimated at the crops and grass sites. This behavior is illustrated for July 2004 by Figure 5. Nevertheless, results for all the 34 stations confirm that CTESSEL and CHTESSEL have reasonable skills over the vegetation types and regions that have been considered. The correlation between observed and modeled mean diurnal cycle exceeds 0.80 for both the latent and sensible heat fluxes.

Figure 5.

July 2004 average diurnal cycle of simulated (line) and observed (dots) latent heat flux [W m−2] (left) and sensible heat flux [W m−2] (right) for CTESSEL (with A-gs, grey line) and CHTESSEL (with Jarvis-type evaporation, black line) at different observation sites with different biomes: (a) crops (Us-Arm), (b) grassland (It-Mbo), and (c) evergreen needle-leaf forest (Berms-Obs).

[36] The number of sites where CTESSEL is better than CHTESSEL, based on RMSE and correlation, indicates that CTESSEL is always better, although the difference can be small (see Table 5).

Table 5. Number of Sites for Which One Scheme Is Better Than the Other in Terms of RMSE or Correlation Based on 10 Day Averages for the 34 Sites
  Number of Sites With Better Performance
ScoreSchemes That are ComparedLatent Heat FluxSensible Heat FluxNet Ecosystem Exchange

5.1.2 Natural CO2 Fluxes Gross Primary Production and Ecosystem Respiration

[37] As for the energy fluxes, the GPP and Reco fluxes calculated by CTESSEL and CHTESSEL are evaluated with in situ observations. Although both models carry the same CO2 flux parameterization, different treatments of transpiration may lead to differences in CO2 fluxes. Both GPP and Reco show a generally good performance for 10 day mean values with an average correlation of 0.80 for GPP and 0.70 for Reco (see Figure 6 and Table 6). However, for some stations both models have a rather poor performance. The correspondence between the performance metrics of the two models indicates that the problem is not due to difference between CTESSEL and CHTESSEL (i.e., it is not related to the use of the A-gs canopy conductance in the evaporation formulation).

Figure 6.

Performance metrics for 2004 of the 10 day averaged simulated carbon fluxes compared with eddy-covariance observation over the 34 sites using N = 36 data points for each site (no data were available for Brasilia and Manaus). Black bars are for CHTESSEL runs, dark grey bars are for the CTESSEL results. (a) Gross primary production correlation, (b) Gross primary production RMSE [µmol m−2 s−1], (c) Ecosystem respiration correlation, and (d) Ecosystem respiration RMSE [µmol m−2 s−1].

Table 6. Average Performance Metrics for 2004 of the 10 Day Averaged Carbon Fluxes Simulated With CTESSEL and CHTESSEL for the 34 Sites
ModelGPP RMSE [µmol m−2 s−1]GPP Bias [µmol m−2 s−1]GPP CorrNEE RMSE[µmol m−2 s−1]NEE Bias [µmol m−2 s−1]NEE Corr [µmol m−2 s−1]Reco RMSE [µmol m−2 s−1]Reco Bias [µmol m−2 s−1]Reco Corr

[38] For the three selected sites, the 2004 seasonal cycles of GPP and Reco show that the observed seasonal cycles are reasonably well simulated by both models, which is also partly due to the parameter optimization performed on these sites even though for a different year. From Figure 7 it can be seen that the grass (It-Mbo) and forest (Berms.Sk-Obs) sites show almost a perfect match between the CTESSEL and CHTESSEL CO2 fluxes, but the US-Arm site (crops) showed larger differences in GPP between the two model versions, while the differences in the Reco were small. Among the three shown sites, the soil moisture of the US-Arm site showed the biggest difference between CTESSEL and CHTESSEL; while the surface temperatures were quite similar; this indicates the high sensitivity of GPP to the soil moisture. Nevertheless, some stations show less skill. As stated above, the problem is not always due to the canopy conductance formulation, but rather to other error sources such us an inadequate representation of the coefficients dependent on vegetation type. For instance the Il-Yat site with a Mediterranean needle-leaf type is represented by parameters for boreal needle-leaf forest in the current model set-up. Also, the management practices, which are not represented by the model, are sources of error at the At-Neu and It-Mbo sites. The error in the radiative forcing over the tropical region (Figure 3c) is another likely source of error.

Figure 7.

Seasonal cycle (2004) of 10 day averaged simulated (lines) and observed (dots) gross primary production GPP [µmol m−2 s−1](left) and ecosystem respiration Reco [µmol m−2 s−1] (right), for CTESSEL (with A-gs, grey line) and CHTESSEL (with Jarvis-type evaporation, black line) at different observation sites with different biomes: (a) crops (Us-Arm), (b) grassland (It-Mbo), and (c) evergreen needle-leaf forest (Berms-Obs). Net Ecosystem Exchange

[39] For CTESSEL and CHTESSEL the NEE results are compared with the CASA-GFED3 NEE outputs, which are also driven by the MODIS collection 5 data. The CASA-GFED3 system simulates NEE with a monthly time step, based on input from net primary productivity (NPP) and CO2 emissions through heterotrophic respiration (Rh). It uses satellite-derived estimates of fraction of absorbed photosynthetically active radiation and a light use efficiency function that depends on temperature and moisture conditions to calculate NPP, while the soil respiration component is simulated by a set of carbon “pools”, which also depend on soil moisture and temperature. The atmospheric forcing used to derive the CASA-GFED3 NEE are the Global Precipitation Climatology Project version 1.1 precipitation, the International Institute for Applied System Analyses temperature climatology, and the International Satellite Cloud Climatology Project solar radiation (International Satellite Cloud Climatology Project FD) (for further details, see Van der Werf et al. [2010] and Potter et al. [1993]). To obtain the same high temporal resolution as ERA-I, the CASA-GFED3 data are disaggregated to a 3-hourly time-step based on the Olsen and Randerson [2004] method. The monthly CASA NPP is separated into GPP and autotrophic respiration (Ra) assuming that NPP and Ra are equal (i.e., GPP = 2NPP). Then, GPP is assumed to be proportional to solar radiation from ERA-I with a constant of proportionality that guarantees the same monthly average as CASA-GFED3. The CASA-GFED3 autotrophic respiration is disaggregated in a similar way making use of a Q10 temperature function [Olsen and Randerson, 2004]. Because of disaggregation with ERA-I, CASA-GFED3 and CTESSEL/CHTESSEL cannot be considered fully independent for the short time scales. However, the longer time scales in CASA-GFED3 (from month up to multiyear) are fully independent of ERA-I.

[40] Similar to the energy fluxes, both model versions are comparable over the 34 sites in 2004 (Table 6) with slightly better performance for CTESSEL (Table 5). The comparison with in situ observations yields similar results for CHTESSEL and CTESSEL with an average correlation of 0.65 and an RMSE of 1.6 µmol m−2 s−1, which are better than those of the CASA-GFED3 results (correlation of 0.37 and RMSE of 1.8 µmol m−2 s−1 ) (Figure 8).

Figure 8.

Performance metrics for 2004 of the 10 day averaged simulated carbon net ecosystem exchange with eddy-covariance observations over the 34 sites using N = 36 data points for each site. Black bars are for CHTESSEL runs, dark grey bars are for the CTESSEL results and light grey bars are for the CASA-GFED3 data. (a) Correlation and (b) RMSE [µmol m−2 s−1] .

[41] The NEE seasonal cycle for 2004 at the three selected sites is shown in Figure 9. Most of the sites show that CTESSEL and CHTESSEL have a larger sink (average bias = −0.2 [µmol m−2 s−1]) than the CASA-GFED3 inventory (average bias = +0.7 [µmol m−2 s−1]), and the overall seasonal cycle of NEE is better represented by CTESSEL, and to less extent by CHTESSEL. A dominant feature at most sites is that NEE in CASA-GFED3 shows smaller amplitude of the seasonal cycle than observed, sometimes up to a factor of 2. It is likely that CTESSEL and CHTESSEL benefit from the optimization on the basis of the 34 flux towers. As discussed before, a few sites show particular features that could not be simulated by the model, namely management practices in summer at the It-Mbo and At-Neu sites, the biased radiation forcing for the tropical sites, and the mismatch between optimized model parameters and the site vegetation type for Mediterranean forest sites.

Figure 9.

Seasonal cycle (2004) of 10 day averaged simulated (lines) and observed (dots) net ecosystem exchange NEE [µmol m−2 s−1], for CTESSEL (with A-gs, grey line) and CHTESSEL (with Jarvis-type evaporation, black line) and CASA-GFED3 (dashed line) at different observation sites with different biomes as indicated in Figure 7.

[42] With regard to the average diurnal cycle, CTESSEL and CHTESSEL show similar results. This is illustrated by the results for July shown in Figure 10. There is an underestimation of the midday NEE for the low vegetation (the grass and crops sites) and an overestimation for the forest site, the nighttime behavior is reversed marking a lower source for both model versions. Compared to the CASA-GFED3 NEE, both model versions show smaller overall bias, but the midday NEE at the crops site is better represented by CASA-GFED3.

Figure 10.

Simulated and observed mean July diurnal cycle of the net ecosystem exchange NEE [µmol m−2 s−1] for CTESSEL (with A-gs, grey line) and CHTESSEL (with Jarvis-type evaporation, black line) and CASA-GFED3 (dashed line) at different observation sites with different biomes as indicated in Figure 7.

[43] The results obtained with the Jarvis-based evaporation in CHTESSEL indicate that the simulation of the photosynthesis process, although tightly linked to the transpiration, could be performed in a modular way and lead to comparable results to those from the photosynthesis-based evaporation in CTESSEL.

5.2 Multiannual Global Offline Simulations

[44] In this section multiyear offline simulations with CHTESSEL forced by the ERA-I are considered for 2003 to 2008. As demonstrated in the previous section, CHTESSEL and CTESSEL have quite similar skills in predicting NEE. For this reason, CHTESSEL, which would preserve the current weather prediction performance, was introduced into the operational IFS, and results from this configuration are now presented.

[45] Figure 11 shows global patterns of January and June NEE averaged over 2003 to 2008 as simulated by CHTESSEL compared to CASA-GFED3 with positive NEE indicating a CO2 sink. In January, although the two models predict an emission of CO2 from the Northern Hemisphere, the CO2 source is less pronounced in CHTESSEL partly owing to the snow attenuation function introduced in the ecosystem respiration formulation (section B.7). In the tropics, the two models disagree: CHTESSEL simulates an uptake of CO2 whereas CASA-GFED3 predicts a release, particularly in the Amazonian region. In June, CASA-GFED3 and CHTESSEL predict an uptake of CO2 over the Northern Hemisphere with similar spatial patterns, whereas CHTESSEL shows more spatial variability. Over the Amazonian tropics, the models also disagree in June.

Figure 11.

The 2003–2008 monthly average of the NEE [µmol m−2 s−1] simulated by CHTESSEL (upper panels) and CASA-GFED3 (lower panels) for January (left panels) and June (right panels).

[46] Based on the tower observations discussed above, it is concluded that CHTESSEL is better in the extra-tropics, mainly due to the stronger and more realistic seasonal cycle in CHTESSEL than in CASA-GFED3. The source of these differences is not always clear. Differences in the model formulation and differences associated with the meteorological forcing play a role, particularly in the tropics. Zhao et al. [2006] showed that using the MODIS GPP algorithm, the annual NPP over the Amazonian region can vary by more than 50% depending on the forcing used.

[47] The evolution of the global atmospheric CO2 budget is assessed by comparing the annual global means of NEE from CHTESSEL and CASA-GFED3 for 2003 to 2008 with the global atmospheric growth obtained from observations [GLOBALVIEW-CO2, 2011, ] as shown in Figure 12. The global CO2 budget is obtained by integrating all the instantaneous model fluxes in time and space, taking into account the area weights for each model grid point. These fluxes are NEE from either CHTESSEL or CASA-GFED3, surface CO2 fluxes from the ocean [Takahashi et al., 2009], biomass burning GFED3.0 [Van der Werf et al., 2010], and anthropogenic emissions (EDGARv4.2) [Olivier and Berdowski, 2001]. Note that the global annual ocean flux is constant because it is based on climatology, as its interannual variability is small [Le Quéré et al., 2000]. The contribution of biomass burning to the interannual variability is also small for 2003 to 2008 [Van der Werf et al., 2010, Table 7]. Therefore, the interannual variability in the global atmospheric growth is mainly linked to the interannual variability of NEE.

Figure 12.

Time series of surface flux contributions to the annual atmospheric CO2 mass budget: NEE from CHTESSEL (green); NEE from CASA-GFED3 (orange); ocean sink from the Takahashi climatology (blue); biomass burning from GFED3.0 (red); anthropogenic emissions from the EDGARv4.2 inventory (cyan); the total surface emission with NEE from CHTESSEL (solid grey) and the total surface emission with CASA-GFED3 (dash grey). The latter two can be compared with the CO2 atmospheric growth (black) from GLOBALVIEW-CO2/ 2011 which is used as a proxy for observations as it is heavily constrained by flask observations. To be consistent with a positive sign for an increase of CO2 in the atmosphere, the surface fluxes are positive in this figure for upward fluxes.

[48] The net surface contribution to the atmospheric budget with CHTESSEL NEE is of similar magnitude to that of CASA-GFED3. However, both the interannual variability of CHTESSEL NEE and the resulting total CO2 flux obtained by combining the NEE with the other prescribed CO2 surface fluxes in the model have a consistently better correlation with the interannual variability of the atmospheric growth based on observations (0.70 for the total flux based on CHTESSEL NEE and 0.20 for the total flux based on CASA-GFED3 NEE). Because of the important role of the meteorological and land surface conditions in the interannual variability of NEE, and given that NEE in both CHTESSEL and CASA-GFED3 are driven by the same MODIS collection 5 data, it is likely that one of the main factors contributing to differences in the interannual variability between CHTESSEL and CASA-GFED3 is the different meteorological forcing and land surface conditions.

[49] The fact that CHTESSEL is picking up the interannual variability signal does not mean that it simulates the variability for the correct reasons (i.e., based on the observed ecosystem processes). Further examination shows that the tropical land regions are responsible for the global interannual variability in CHTESSEL. These are the regions that have the largest uncertainties and errors in the model due to missing processes and lack of observations. An illustrative example is the prolonged drought in the Amazon forest during 2005, which had a large impact on the global budget estimated between 1.2 and 1.6 PgC [Phillips et al., 2009]. In CHTESSEL, the 2005 Amazon drought also contributed significantly in the global CO2 anomaly (Figure 12) despite the model not being able to represent the actual tree morality due to water stress and fires, which were responsible for the observed CO2 release. In the model, the net CO2 release in the Amazon was associated with a large increase in heterotrophic respiration from top soil linked to increasing temperatures and a moderate decrease in GPP linked to soil moisture decrease; whereas in reality the main factor associated with the CO2 release was decomposition from dead trees as a result of the prolonged water stress and more biomass burning than usual [Phillips et al., 2009; Aragão et al., 2007]. CHTESSEL is able to reproduce the interannual variability because the fluxes are tuned to the resulting climate response, even though all the underlying observed processes cannot be represented in such a simplified model. Therefore, although CHTESSEL is suitable for providing boundary conditions to an atmospheric NWP model due to tuning of parameters based on flux observations, it has limitations when examining the causality of anomalous events.

[50] Besides the difference in their interannual variability, both models have the same range and seem to overestimate the CO2 source. An additional comparison was also performed against known land surface models and several state-of-the-art flux inversion systems used in the Carboscope project for the period 2002 to 2004 [ ts&param = co2_dgvm]. Despite the large differences in the formulation of the various models and flux inversion systems, the CHTESSEL budget is within the range of the listed estimates and reveals an interannual variability consistent with the other inventories.

[51] In spite of the simple parameterization of the respiration (meant to avoid long spin-up windows not suitable in NWP) and the absence of land use change (which is a reasonable assumption for 2003 to 2008), CHTESSEL in combination with an accurate atmospheric forcing lead to reasonable simulations of the yearly terrestrial CO2 sink and its interannual variability. However, as stated in the previous sections, additional investigations of modeling uncertainties would be needed for a better global CO2 estimate.

5.3 Impact of CO2 Fluxes on Atmospheric Concentrations

[52] In this section, terrestrial NEE fluxes are used as a surface boundary condition in a global atmospheric CO2 transport model that takes its meteorological fields from NWP analyses. Atmospheric CO2 is integrated in space and time, which allows the resulting concentrations to be evaluated against independent observations of atmospheric CO2 concentration.

[53] The impact of terrestrial NEE from CHTESSEL is compared to results obtained using NEE from CASA-GFED3 [Engelen et al., 2009]. As in the previous section, ocean surface fluxes are prescribed by climatology [Takahashi et al., 2009], biomass burning by GFED3.0 [Van der Werf et al., 2010], and anthropogenic emissions by EDGARv4.2 [Olivier and Berdowski, 2001]. The tracer transport model is based on the IFS model (2012 version of the operational system) with 60 model levels and a horizontal resolution of approximately 80 km.

[54] The atmospheric CO2 latitudinal distribution and seasonal cycle from the simulation using CHTESSEL NEE is consistent with that from CASA-GFED3 as shown in Figure 13 by comparing these with GLOBALVIEW-CO2 [2011]. Both CHTESSEL and CASA-GFED3 underestimate the CO2 sink in the Northern Hemisphere summer, but CHTESSEL shows a better match to GLOBALVIEW-CO2 in the corresponding winter. This is confirmed by comparing the modeled atmospheric CO2 with observations from various NOAA/ESRL (Earth System Research Laboratory) baseline stations [Thoning et al., 2012]. In particular, Barrow (Alaska, US) (Figure 14a) illustrates very clearly the shortcomings of the modeled CO2 regarding the seasonal cycle. In the CHTESSEL-based simulation there is an underestimation of both the CO2 growth in the winter and the CO2 sink in the summer, whereas in the CASA-GFED3-based forecast there is a large overestimation of the CO2 winter growth and a similar underestimation of the summer sink.

Figure 13.

Hovmöller diagram of weekly atmospheric CO2 concentrations [ppm] for the year 2003 across latitude from (a) the NOAA GLOBALVIEW-CO2 product based on observations and the model forecasts using (b) CHTESSEL NEE and (c) CASA-GFED3 NEE. The GLOBALVIEW-CO2 product is based on the processing of flask sites sampling the marine boundary layer (see Masarie and Tans [1995] for details). The model output has been sampled at the same sites following the same data processing procedure.

Figure 14.

(a) Time series of daily atmospheric CO2 [ppm] from simulations using CHTESSEL (red), CASA-GFED3 (orange) and observations from the NOAA/ESRL baseline station (grey) at Barrow, AK, US (71.32°N, 156.60°W, 11 m above sea level (asl) for the 2003–2005 period [Thoning et al., 2012]; (b) Hourly time series of atmospheric CO2 [ppm] for October 2003 from the NOAA/ESRL tower at Argyle, ME, US (45.03°N, 68.68°W, 53 m asl; Andrews et al. [2009, 2013]). The symbols are the same as in Figure 14a.

[55] Overall, the simulations forced with CHTESSEL fluxes have less bias with respect to observations than the simulations forced with CASA-GFED3 with annual biases in 2003 of −0.7 and +3.4 ppm, respectively. The standard error for CHTESSEL is 2.5 and 2.2 ppm for CASA-GFED3. The biases and standard errors in 2004 and 2005 are similar to those in 2003. However, in 2005 the underestimation of the summer sink is more pronounced and the annual bias is also larger in magnitude for CHTESSEL (+0.9 ppm). The annual negative bias in CHTESSEL for 2003 and 2004 is dominated by the underestimation of the winter growth and an early start of the growing season at high northern latitudes. This may be due to the model not being able to simulate accurately the CO2 flux attenuation by snow (in spite of the improvements) and the timing of the snow melt at the beginning of the growing season.

[56] In the tropics (e.g., Mauna Loa and Samoa) and the South Pole (not shown), the seasonal cycle of the baseline stations is weaker. Forecasts based on CHTESSEL and CASA-GFED3 have a similar overestimation of the atmospheric CO2 throughout the year. Although the magnitude of forecast atmospheric CO2 dry molar fraction is generally smaller than that observed, its short-term variability follows qualitatively the observations for both CHTESSEL and CASA-GFED3. This is shown by comparing the modeled and observed CO2 at continental sites (e.g., Argyle in Maine, US from the ESRL/NOAA tower network [Andrews et al., 2009, 2013] in Figure 14b). The diurnal cycle is reproduced by the forecasts based on CHTESSEL and CASA-GFED3, but there is an underestimation of its amplitude in both simulations. The high peaks of CO2 at nighttime are clearly underestimated. This is due to a combination of uncertainty about the respiration fluxes and transport/mixing within the atmospheric boundary layer [Yi et al., 2004]. The synoptic variability is well modeled by the IFS which has a high skill at that time scale [Patra et al., 2008].

6 Discussion and Conclusions

[57] To simulate the global CO2 cycle, a photosynthesis-based stomatal conductance and CO2 exchange parameterization has been added to the ECMWF model. The proposed parameterization simulates the diurnal and seasonal variations of CO2 fluxes, and allows for interaction with atmospheric CO2 in global monitoring and prediction applications. CO2 and water vapor fluxes share the same pathway through the leaf stomata and therefore the stomatal conductance computed from the photosynthesis module is used for evaporation resulting in a consistent treatment of the surface energy fluxes along with biospheric GPP and Reco. The latter two quantities are components of the biospheric NEE that can provide the surface boundary condition to atmospheric CO2 transport models.

[58] The scheme is calibrated by optimizing the parameters that actively control the CO2 fluxes, using tower observations of surface CO2 fluxes for a range of vegetation types. The model is then validated using results from a different time period for the same sites. The model is also compared to the well-established and widely used CASA-GFED3 product. It is shown that the two versions (CTESSEL/CHTESSEL with/without coupled net assimilation and evaporation) are very similar and that they show slightly smaller errors than CASA-GFED3. This result is consistent with Balzarolo et al. [2011] who found that CTESSEL and CHTESSEL are also slightly better than two other models (ISBA-Ags and ORCHIDEE) over a multiannual period at the field-site scale. It is important to note that most calibration/verification stations are in the Northern Hemisphere extra-tropics, so it is difficult to draw firm conclusions for the tropics for which only two stations were available. It is also likely that the slight advantage of CTESSEL/CHTESSEL over CASA-GFED3 is related to the calibration, which is not entirely independent of the verification. There is however, an important difference between the setup of the calibration and verification procedures (apart from using a different year). Calibration uses vegetation settings that are characteristic for the tower site, whereas the verification uses vegetation settings from climatology (GLCC and LAI) as in the global model. The idea is to verify the complete system including climatological fields.

[59] The relatively good performance of both CTESSEL and CHTESSEL compared to GFED3-CASA NEE is unlikely to come from LAI, because all three models use the MODIS collection 5 data [Van der Werf et al., 2010]. It is postulated that the coupling to a good quality NWP meteorological forcing with high temporal resolution brings a benefit because CO2 models are known to be sensitive to atmospheric forcing [Zhao et al., 2006]. NWP type reanalysis of relevant variables (e.g., temperature, atmospheric moisture, soil moisture, radiation, and precipitation) has been documented and evaluated [Simmons et al., 2010; Balsamo et al., 2010; Szczypta et al., 2011].

[60] In addition, the results presented here also indicate limitations of the parameter optimization by vegetation types; for some biomes the optimal parameters did not have the same skills over different climatic regions and thus a global extrapolation for those biomes is not always feasible. Mediterranean forest, for example, belongs to the same class as the boreal forest. Also, some model deficiencies were shown to be linked to anthropogenic actions such as management practices. Another problem is the lack of tropical stations in this study, which results in large uncertainty for the tropics. Skill in the tropics comes from optimization of parameters in the extra-tropics, which might be beneficial, but needs further investigation. The study however highlighted the importance of the site-level flux measurements in the context of model development and validation.

[61] The comparison between CTESSEL and CHTESSEL showed that the separation of the evapotranspiration and CO2 exchange processes does improve evaporation, but it does not seem to significantly affect the overall CO2 results. It is therefore possible to model terrestrial CO2 exchange in NWP models without impact on other land surface processes that might affect weather prediction; in other words, the full coupling of CO2 and evapotranspiration is not a precondition to model NEE with reasonable success. However, a more realistic variability of the turbulent energy fluxes at both seasonal and diurnal time scales is seen with the A-gs scheme, compared to fluxes with the Jarvis-based formulation. This is encouraging because it suggests that the basic plant physiological mechanism, where stomata control CO2 and water transport in the same way, is realistic. It also suggests that a fully photosynthesis-based parameterization has the potential to better capture the basic dependencies on environmental parameters and that it is suitable in an operational NWP such as the IFS. It is very difficult to establish which environmental factor is behind the improvement of the turbulent fluxes; nevertheless, it is believed that further testing of the two model versions during drought periods with enough available observations would better clarify the underlying carbon-water interactions. Currently, a temperature response is not included in the actual Jarvis-based formulation, although we know from earlier work that the temperature dependence does not add much. The photosynthesis approach by definition includes a physiological response of stomatal conductance to atmospheric CO2, which is known to be increasing at present. This model feature could be an advantage of the A-gs approach over the Jarvis-based approach.

[62] The evaluation shows that CHTESSEL can predict the spatial and seasonal to interannual variability of NEE reasonably well, especially in the Northern Hemisphere where most of the flux verification sites are located. The coupling of CHTESSEL NEE with an atmospheric transport model shows that there is also skill in simulating the seasonal variability by sampling the model at the GLOBALVIEW sites and at the NOAA/ESRL continuous baseline observatories. However, there are large systematic atmospheric CO2 errors in the Northern Hemisphere (i.e., an underestimation of CO2 in the winter at high latitudes, a premature decrease of CO2 at the onset of the growing season and overestimation of CO2 during summer). The fact that these atmospheric CO2 errors are not consistent with the NEE flux evaluation implies that more diverse flux sites are required to draw conclusions on a global scale regarding the seasonal cycle. The global NEE budget is within the range of uncertainty of other products (from state-of-the-art land vegetation models and flux inversion systems) and its interannual variability correlates well with the observed atmospheric growth. This does not imply that the processes explaining the variability are always correct. The optimization of parameters in CHTESSEL may have contributed to the tuning of the sensitivity of the model to the meteorological forcing without necessarily representing the right complex underlying processes. Therefore, caution should be exercised when attributing changes of CO2 or NEE to specific causes based on modeled fluxes only. These results support the hypothesis concerning the importance of both the LAI remote sensing data (for the climatology) and the meteorological forcing in simulating the CO2 fluxes.

[63] The current respiration does not account for land use change (which is a reasonable assumption for short time scales) and follows a simple parameterization. Nevertheless, the CHTESSEL model in combination with an accurate atmospheric forcing is able to simulate the yearly terrestrial CO2 sink and its interannual variability. These model features are particularly useful for monitoring applications within the NWP environment.

[64] Although the model described here performs reasonably well, it is believed that further improvement is possible. For instance, the leaf area index data used in this study are based on climatology extracted from satellite observations. Inclusion of deviations from climatological LAI through assimilation of real-time satellite observations would be a way to improve the seasonal cycle and the interannual variability. The too early drop of Northern Hemisphere CO2 in spring compared to observations of atmospheric CO2 is puzzling, because this shift in the seasonal cycle is not seen in the flux tower comparisons. It suggests that the signal is coming from areas that are not sampled by our set of flux towers. Improvements in the handling of cold processes will be necessary and support from high-latitude observations would be very helpful. Also, improvements to the vegetation classification (e.g., Mediterranean forest) could lead to a better seasonal cycle. All these model developments will be addressed in the future as part of the regular ongoing upgrades in the operational NWP system. Owing to its simple parameterization, and its strong sensitivities to meteorological forcing and satellite LAI, CTESSEL/CHTESSEL is well suited for monitoring applications within the NWP environment.

Appendix A

[65] The three stress functions f1, f2, f3 used in the Jarvis approach to describe the shortwave radiation, soil moisture stress, and atmospheric humidity deficit are given by

display math(A1)

where Rs is the downward shortwave radiation and a, b, c are empirical constants,

display math(A2)

with θpwp and θcap (dependent on soil type) representing soil moisture at permanent wilting point and at field capacity, respectively, and inline image soil moisture of a root-density-weighted average over the various soil layers for the unfrozen soil water, and

display math(A3)

[66] In (equation (A3)), Da is the atmospheric humidity deficit, and gD is a vegetation-type dependent coefficient. The details of HTESSEL and its parameter settings are given in Balsamo et al. [2009].

Appendix B

[67] The basics of the A-gs model are described by Jacobs [1994] and Jacobs et al. [1996]. Further details of the current extended formulation are given in Calvet et al. [1998, 2004]. For the description of the canopy conductance to CO2, gsc, a stepwise approach is followed with (i) the definition of the temperature-dependent parameters, (ii) the radiation response, (iii) the calculation of the ratio between internal and external CO2 concentration, (iv) the computation of stomatal conductance, (v) inclusion of the soil moisture response, and (vi) the vertical integration over the canopy.

B1. Temperature Responses

[68] There are several parameters in the photosynthesis model that are temperature dependent, namely the compensation point, the mesophyll conductance, and the maximum photosynthetic capacity. The compensation point Γ is defined as the CO2 concentration at which the net CO2 assimilation of a fully lit leaf becomes zero. It can be measured in a laboratory by exposing plants to a variable CO2 concentration. The mesophyll conductance gm describes the transport of CO2 from the substomatal cavities to the mesophyll cells where the CO2 is fixed. It includes the representation of physical and chemical processes. The maximum photosynthetic capacity Am,max is specified as an absolute upper limit to the photosynthesis rate in full sunlight and nonlimiting CO2 concentration.

[69] The temperature dependence is described with so-called Q10 functions, where Q10 represents the proportional increase of a parameter for a 10°C increase in temperature [Berry and Raison, 1982]. For the compensation point the formulation is

display math(B1)

where Γ(25°) is the compensation point at 25°C, Q10Γ is the Q10 -constant and Ts is the leaf surface temperature. For gm and Am,max, the temperature dependence is further adjusted by the inhibition functions after Collatz et al. [1992]

display math(B2)
display math(B3)

where Q10gm, Q10Am,max, T1gm, T2gm, T1Am,max, and T2Am,max, are constants affecting the sensitivity to the plant surface temperature Ts. Parameter gm(25°) depends on soil moisture stress and will be further described in section B5. Its unstressed value inline image is optimized here with the help of observations (see section 4). The constants in these functions are vegetation-type dependent and are listed in Tables 1 and 2.

B2. Radiation and CO2 Response

[70] For An, two regimes are distinguished: the radiation limiting regime and the CO2 limiting regime [Goudriaan et al., 1985; Jacobs, 1994]. In the radiation limiting regime with sufficient CO2, An is controlled by the amount of photosynthetically active radiation (PAR) Ia

display math(B4)

where Rd is the dark respiration and where ε is the quantum efficiency expressed as

display math(B5)

[71] Parameter εo is the maximum quantum use efficiency and Cs is the ambient CO2 concentration at the leaf surface. At high radiation intensities, net assimilation saturates at a level Am and becomes CO2 limited according to Am = (Ci − Γ) gm with Ci the CO2 concentration inside the leaf cavities (see below). An absolute limit to account for the maximum photosynthetic capacity of the leaves is further applied as follows:

display math(B6)

[72] The radiation and CO2 limiting regimes are combined via a smooth exponential transition function

display math(B7)

[73] The autotrophic dark respiration is simply parameterized according to Van Heemst [1986] and includes the respiration from the leaves only

display math(B8)

[74] The respiration from other parts of the vegetation is included in the heterotrophic respiration.

B3. The inline image Ratio

[75] The CO2 concentration inside the leaf cavities Ci needs to be known to derive the stomatal conductance from the net assimilation. Observations indicate that the ratio Ci/Cs is a rather conservative quantity for moist atmospheric conditions and that increasing humidity deficit exerts a strong stomatal control affecting this ratio. Therefore, Ci/Cs is specified as a function of atmospheric moisture deficit Ds at the leaf surface.

display math(B9)

where f is the coupling factor defined by

display math(B10)

and fo is the value of f at Ds = 0 kg/kg, Dmax is the maximum saturation deficit and

display math(B11)

[76] Some transport of CO2 is maintained in the situation where f = fmin through the leaf cuticle or because of imperfect closure of the stomata. This process is represented by the cuticle with conductance gcu.

B4. Stomatal Conductance

[77] The first computation of the stomatal conductance for CO2, inline image is achieved by dividing net assimilation by the difference between CO2 concentration in and outside the leaves. It is modified here to account for the limiting cases of very dry air and dark respiration

display math(B12)

where Amin represents the residual photosynthesis rate (at full light intensity) associated with cuticular transfers when the stomata are closed because of a high specific humidity deficit

display math(B13)

[78] In this equation, Cmin is the value of Ci at maximum specific humidity deficit

display math(B14)

[79] The diffusion of CO2 through the stomatal openings interacts with that of water vapor and therefore stomatal conductance to CO2 is corrected for this interaction by an iterative refinement

display math(B15)

where Mv and Ma are molecular masses of water vapor and air respectively, ρa is the air density and E is the leaf transpiration based on the previous guess of the stomatal conductance

display math(B16)

[80] Finally, the stomatal conductance to water vapor gs is given by

display math(B17)

[81] The total conductance used by the transpiration scheme is gs + gcu, where gcu is the vegetation-type dependent cuticular conductance (Table 1).

B5. Soil Moisture Stress Response

[82] Among other possible A-gs formulations for which the soil moisture stress response is directly applied to the gross assimilation, Ag [Ronda et al., 2001] or the net assimilation An [Sala and Tenhunen, 1996], Calvet [2000] found that the soil moisture stress response is driven in a complex way through the mesophyll conductance, gm, the maximum specific humidity deficit tolerated by the vegetation, Dmax, and the ratio Ci/Cs controlled by f. The soil moisture response behaves differently for high and low vegetation. In CTESSEL the adopted soil moisture stress response follows the function described in Calvet [2000], Calvet et al. [2004] and is based on a meta-analysis of several herbaceous and woody vegetation types.

B5.1. Low Vegetation Formulation

[83] Calvet [2000] found that the mesophyll conductance gm and the maximum atmospheric moisture deficit Dmax vary with soil moisture but that they remain correlated according to

display math(B18)

[84] Therefore, this equation is used to derive inline image (maximum saturation deficit without soil moisture stress) from the tabulated inline image, where superscript * indicates optimal soil moisture conditions and a and b are tabulated empirical coefficients (Table 2). Then the soil moisture stress index f 2 (see equation (A2)), is applied to find Dmax in stressed conditions according to a bilinear function with a breakpoint at a critical soil moisture stress index f2c

display math(B19a)
display math(B19b)

where inline image is the maximum value of Dmax corresponding to f2c. The resulting Dmax is substituted in (equation (B.18)) to find gm(25°).

B5.2. High Vegetation Formulation

[85] Observations show that gm(25°) is well correlated with the coupling factor f0 according to the following empirical expression [Calvet et al., 2004]:

display math(B20)

[86] In this case (equation (B20)) is used to derive inline image from the value inline image as tabulated according to vegetation type (Table 1). Subsequently a soil moisture stress function is applied to find gm(25°)

display math(B21a)
display math(B21b)

where inline image is the stressed value of gm derived from the Calvet et al. [2004] meta-analysis with the following empirical function:

display math(B22a)

[87] After computing gm(25°) according to equation (25), the stressed value for f0 is derived with

display math(B22b)

[88] Further details on the soil stress parameterization can be found in Calvet [2000], Calvet et al. [2004], and Voogt et al. [2006].

B6. Vertical Integration from Leaf to Canopy

[89] The net CO2 assimilation calculated at the leaf scale is up-scaled to the canopy scale assuming that (a) leaf parameters do not vary with height in the canopy, and (b) the attenuation of the incoming shortwave radiation in the canopy can be computed using a simple radiative extinction model. The incoming PAR above the vegetation (Ia(h), with h the canopy height) is assumed to be 48% of the incoming shortwave radiation and then further attenuated in the canopy. The dependence of PAR on height z within the canopy is described by Roujean [1996] according to

display math(B23)

where K is the extinction function given by

display math(B24)

Kdf (z) and Kdr(z) are the extinction coefficients of diffuse and direct light, respectively

display math(B25)
display math(B26)

where μs is the solar zenith angle and G is a parameter that describes the distribution of leaves (a spherical angular distribution is assumed with G = 0.5), δ is the ratio of diffuse to total downward shortwave radiation at the top of the canopy, LAI(h − z) is the cumulative leaf area index above height z, and b is the foliage scattering coefficient given by:

display math(B27)

based on the leaf single scattering albedo ω (=0.2) for the solar spectrum corresponding to the PAR. Parameter δ is given by

display math(B28)

[90] Assuming a homogeneous leaf vertical distribution, the integrated canopy net CO2 assimilation, dark respiration and conductance can be written as

display math(B29)
display math(B30)
display math(B31)

[91] In the above equations, LAI is defined as the ratio of leaf area covering a unit of ground area (m2 m−2).

[92] The integrations are parameterized with a three-point Gaussian quadrature method following Goudriaan [1986]:

display math(B32)
display math(B33)
display math(B34)

where Wi and zi are the Gauss weights and levels, respectively.

B7. Ecosystem Respiration and Gross Primary Production

[93] The ecosystem respiration Reco is given by two terms: the autotrophic dark respiration, RdI (equation (B30)), and Rsoilstr, which represents both heterotrophic respiration from the soil and autotrophic respiration from the above and below ground structural biomass. It is parameterized following a modified formulation of Norman et al. [1992] as a function of soil temperature, soil moisture, snow depth, and vegetation type

display math(B35)

[94] In this equation fsn and fsm are snow and soil moisture attenuation functions respectively defined as

display math(B36)

Cvs is the surface fraction covered by snow, α is a constant expressing the attenuation of the soil CO2 emission within the snow pack and zsnow is the snow depth. The soil moisture stress function for soil respiration is defined following a study by Albergel et al. [2010] as

display math(B37)

[95] Q10Ro represents the proportional increase of a parameter for a 10°C degree increase in temperature [Berry and Raison, 1982]. Given its variability with climate regimes, Q10Ro is defined as a function of soil temperature after McGuire et al. [1992]. The vegetation types are affecting the ecosystem respiration through a reference respiration at 25°C (R0 (25)) estimated by minimizing the root mean square errors between simulated and observed Reco for each vegetation type (see section 4).

[96] Finally, the gross primary production GPP and the ecosystem respiration are given by

display math(B38)
display math(B39)

[97] These two quantities are often derived from flux-tower NEE observations and used for evaluation of the two main CO2 processes.


[98] We thank Sebastien Massart for his support and comments on the CO2 concentration results, and Bob Riddaway for his careful review and improvements of the use of the English language. We would also like to thank the three anonymous reviewers who made many valuable suggestions that led to substantial improvement of the paper. 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), CarboItaly, CarboMont, Fluxnet-Canada (supported by The Canadian Foundation for Climate and Atmospheric Sciences (CFCAS), The Natural Sciences and Engineering Research Council of Canada (NSERC), BIOCAP Foundation, Environment Canada, and Natural Resources Canada (NRCan)), the Sources and Sinks of Greenhouse Gases from managed European Grasslands and Mitigation Strategies (GreenGrass) project, the Large-scale Biosphere-Atmosphere Experiment in Amazonia (LBA), the Nordic Centre for Studies of Ecosystem carbon exchange and its Interaction with the Climate system (NECC), and the Terrestrial Carbon Observation System Siberia (TCOS-Siberia). We thank in particular the PIs that decided to share their data freely within the scientific community. We acknowledge the support to the eddy covariance data harmonization provided by the Integrated Project Assessment of the European Terrestrial Carbon Balance (CarboEuropeIP), The Food and Agriculture Organization-The Global Terrestrial Observing System-The Terrestrial Carbon Observations project (FAO-GTOS-TCO), the Integrated Land Ecosystem-Atmosphere Processes Study (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.

[99] We also thank The Fluxnet-Canada Research Network for making available the BERMS data (, the PIs and researchers involved in the Coordinated Enhanced Observing Period (CEOP) project for provision and archiving of the data used in this study (, and the CARBOSCOPE project contributors for making their data freely available through the project website ( We further thank NOAA/ESRL, Pieter Tans, Kirk Thoning, and Arlyn Andrews, for providing the atmospheric CO2 observations to evaluate the atmospheric simulations. The research leading to the results on atmospheric CO2 simulations has received funding from the European Community's Seventh Framework Programme (FP7 THEME [SPA.2011.1.5-02]) under grant agreement n.283576 in the context of the MACC-II project (Monitoring Atmospheric Composition and Climate - Interim Implementation). This work is a contribution to the GEOLAND-2 project, funded by the European Commission 7th Framework Programme in preparation to the Global Monitoring for Environment and Security (GMES) initiative.