multi-model analysis

Abstract. We are comparing spatially explicit process-model based estimates of the terrestrial carbon balance and its components over Africa and confront them with remote sensing based proxies of vegetation productivity and atmospheric inversions of land-atmosphere net carbon exchange. Particular emphasis is on characterizing the patterns of interannual variability of carbon fluxes and analyzing the factors and processes responsible for it. For this purpose simulations with the terrestrial biosphere models ORCHIDEE, LPJ-DGVM, LPJ-Guess and JULES have been performed using a standardized modeling protocol and a uniform set of corrected climate forcing data. While the models differ concerning the absolute magnitude of carbon fluxes, we find several robust patterns of interannual variability among the models. Models exhibit largest interannual variability in southern and eastern Africa, regions which are primarily covered by herbaceous vegetation. Interannual variability of the net carbon balance appears to be more strongly influenced by gross primary production than by ecosystem respiration. A principal component analysis indicates that moisture is the main driving factor of interannual gross primary production variability for those regions. On the contrary in a large part of the inner tropics radiation appears to be limiting in two models. These patterns are partly corroborated by remotely sensed vegetation properties from the SeaWiFS satellite sensor. Inverse atmospheric modeling estimates of surface carbon fluxes are less conclusive at this point, implying the need for a denser network of observation stations over Africa.


Introduction
Understanding terrestrial sources and sinks of CO 2 and its variability is important for understanding the carbon cycleclimate feedback.Extensive research in this field has concentrated on the highly developed parts of the world, in particular North America and Europe with a strong research infrastructure.In a recent review, Williams et al. (2007) identified Africa as "one of the weakest links in our understanding of the global carbon cycle."Africa is the second largest continent of the world occupying about 20% of global land mass and inhabits a large variety of ecosystems ranging from perhumid tropical forest to semi-arid and arid grass and shrub communities.Although Africa's decadal scale mean carbon balance appears to be neutral, the continent contributes about half of the interannual variability of the carbon balance on global scale (Williams et al., 2007).This large interannual variability results primarily from climatic perturbations related to the El Nino phenomenon that directly affects the ecosystems' productivity and due to concomitant biomass burning (e.g.Le Page et al., 2007;Anyamba et al., 2002Anyamba et al., , 2003;;Myneni et al., 1996;Kogan 2000).
Given the scarcity of observation sites for atmospheric CO 2 concentrations and land -atmosphere CO 2 exchange the uncertainties of African carbon cycle research remain U. Weber et al.: Interannual variability of Africas ecosystem productivity high in particular regarding the spatial localization of hotspot regions of variability and the underlying driving forces.Simulations of terrestrial ecosystem models can provide insights here; however, these models are also associated with large uncertainties (e.g.McGuire et al., 2001;Friedlingstein et al., 2006) in particular for water limited conditions (e.g.Morales et al., 2005;Jung et al., 2007), and in addition have generally not been tested and parameterized specifically for Africa.Thus, confidence of a single model analysis is limited and a multi-model study is warranted to identify coherent and dissimilar behavior between different biosphere models.
In this study we assess the interannual variability of Africa's carbon cycle using four different terrestrial carbon cycle models in conjunction with remotely sensed indicators for the state of the vegetation as well as carbon balance estimates from global atmospheric inversions.We aim to identify (1) regions of largest carbon balance interannual variability, (2) the primary process (photosynthesis or respiration) dominating the carbon balance variability, and (3) which climate variables are driving the ecosystem's carbon cycle in the models.

Model descriptions
The four ecosystem models ORCHIDEE (Krinner et al., 2005), LPJ-DGVM (Sitch et al., 2003), LPJ-GUESS (Smith et al., 2001), and JULES/TRIFFID (Cox et al., 2001;Essery et al., 2001;Hughes et al., 2006) applied in this study are coupled biogeography-biogeochemistry models, i.e. they combine representations of both vegetation dynamics and land-atmosphere carbon and water exchanges (dynamic global vegetation models -DGVMs).The concept of plant functional types (PFT) (Smith et al., 1997) is used to discretize differences in physiology and allometry of species including adaptations to climatic conditions and disturbance regime.PFTs compete for resources like light, and water.Representations of vegetation structure like allometry, and function like phenology, allocation, mortality, and establishment are essential for this.Gross primary production (GPP) is calculated based on a coupled photosynthesis-water balance scheme after (Farquhar et al., 1980;Collatz et al., 1991Collatz et al., , 1992)), where simplifications have been incorporated in the different models as described in detail in the above references.Heterotrophic respiration is assumed to be represented by first-order decay of organic material with decay rates for a few pools (Foley, 1995) which depend on temperature and moisture following (Lloyd and Talor, 1994) or a Q 10 formula.Fire, the main disturbance in the region is simulated following (Thonicke et al., 2001) within LPJ-DGVM, ORCHIDEE, and LPJ-GUESS.
The applied models differ in the temporal resolution and in the resolution of vegetation structure representations.While LPJ-DGVM and LPJ-GUESS are designed as stand-alone models running on a daily time step, ORCHIDEE and JULES can be coupled to General Circulation Models functioning as land-surface schemes.Therefore, latter models resolve the diurnal cycle with time steps of 30 to 60 min.Usually, the LPJ family of models is driven by monthly climatic data which are interpolated to pseudo-daily using a whether generator for precipitation.In this study, the LPJ-DGVM is driven by daily climatic inputs.In doing so, net radiation is approximated from global radiation after Linacre (1969).The LPJ-GUESS has the most advanced representation of vegetation structure including age with features of a forestgap model (Shugart, 1984) resolving forest succession.LPJ-DGVM and ORCHIDEE have intermediate complexity applying the concept of "average individuals" for a whole grid cell (Sitch et al., 2003) while JULES/TRIFFID employs a heuristic approach to determine the vegetation coverage and carbon allocation to each PFT.

Experimental setup
The simulation of the terrestrial carbon and water budgets are carried out at a spatial resolution of 1 • ×1 • for entire continental Africa for the target period 1982-2006.Spin-up calculations were performed by repeating the years 1982-1992 using the meteorological data sets of the appropriate years and a fixed CO 2 concentration of 341.13 ppm (Mauna Loa value in 1982) until carbon pools reach equilibrium.After the spin-up, model simulations start in 1982 with rising CO 2 concentration per annum.Potential vegetation distribution was dynamically simulated by the models.

Definition of regions
We define six major regions of sub-saharan Africa based on the broad distribution of ecosystem types (available from global land cover maps), and thus implicitly according to bioclimatic conditions: Northern Savannah Belt, Central African tropical forest, Horn of Africa, Southern raingreen woodlands, South African grasslands, and Madagascar (Fig. 1).Given that Madagascar is small in comparison to the other regions and very heterogeneous we do not discuss this region extensively.

Quantification of interannual variability
Interannual variability (IAV) is calculated for each pixel and model as the standard deviation of the respective quantity (e.g.GPP) for a yearly time step across all years, resulting in a grid of IAV for each model and variable.For identifying spatial patterns of relatively high and low interannual variability, the IAV grid of each model was z-transformed. z Where z(IAV I ) is the standardized IAV of grid cell i, IAV and σ IAV are the spatial mean and standard deviation of the IAV, respectively.Hence z(IAV I ) measures the degree of variability for each pixel in units of standard deviations, i.e. values larger than 0 refer to above average interannual variability and for example a value of two indicates variability of 2 standard deviations above the mean variability of the continent.Consistent spatial patterns between all models are defined as model agreement and are implemented as the sum of models per grid cell which show variability larger than one standard deviation above the mean variability of the continent.

Principal component analysis
To analyze the relationship between meteorological conditions and simulated carbon fluxes on annual scale we follow the approach of Jung et al. (2007).First we reduce the array of several meteorological driver variables to their principal components and then calculate the correlation between meteorological principal components and relative flux variations for each grid cell.The principal component analysis (PCA) of the meteorological data reduces the dimensionality of the data set and often extracts major weather patterns or gradients.We used four annual data fields as input to the PCA: precipitation, air temperature, radiation, specific air humidity.

Corroboration against satellite data
We use the fraction of absorbed photosynthetic active radiation (FAPAR) product of Gobron et al. (2006) based on the SeaWiFS satellite sensor as a proxy for vegetation productivity (available at: www.fapar.jrc.it).It has been designed as an optimized indicator for the state and health of the vegetation and overcomes several limitations of the classically used normalized difference vegetation index (NDVI) (Gobron et al., 2000).Jung et al. (2008) have shown that the annual sum of FAPAR growing season values correlates strongly with annual gross primary production from eddy covariance measurement sites in Europe.At first glance it appears to be more consistent to compare simulated FAPAR by the models with the FAPAR satellite retrievals instead of using the FAPAR as a proxy for GPP.However, there are several reasons why it makes more sense to interpret the remotely sensed FAPAR as an indicator for productivity: (1) ecosystem models tend to capture interannual variability of GPP primarily via interannual variations of radiation use efficiency and not via changes of leaf area (Jung et al., 2007) while the interannual anomaly patterns of the SeaWiFS-FAPAR provide a realistic picture of the GPP anomalies (Jung et al., 2008;Gobron et al., 2005), and (2) there are conceptual mismatches between the FAPAR from the satellite and the FAPAR simulated by models.For example the response of herbaceous understorey, which is very sensitive to e.g.water stress, plays likely an important role in the anomaly patterns of the satellite FAPAR by indicating the direction of change of the ecosystem (see in Jung et al., 2008 for more discussion).In addition, changes of the remotely sensed FAPAR may originate from changes of leaf colour (e.g.leaf darkening or yellowing), which indicates changes of chlorophyll content and thus radiation use efficiency.In contrast, the leaves in the models do not change their reflective properties; effects of leaf aging on photosynthesis are in some cases captured as Vcmax being a function of leaf age as in ORCHIDEE.
We calculate mean annual FAPAR as a proxy for GPP after filling gaps of the FAPAR time series as described in Jung et al. (2008).In addition we compare the mean seasonal cycles of simulated GPP and remotely sensed FAPAR for the defined regions.

Comparison with atmospheric inversions
We compare the simulated variations of the African carbon balance in terms of Net Ecosystem Productivity (NEP), and Net Biome Production (NBP) with results from global atmospheric inversions from Rödenbeck (2005) (version s96 v3.1) covering the period of 1997 to 2006 based on globally 51 stations of atmospheric CO 2 records.This way we also determine the role of fire on the interannual variability of the carbon balance.Simulated NEP is calculated as GPP minus Terrestrial Ecosystem Respiration (TER).Modeled NBP is based on the difference of NEP and model specific fire emissions, except JULES which does not include fire.The inversions detect carbon emissions from fire.To facilitate comparability between NEP simulations and inversions, we used the Global Fire Emission Database (GFED version 2.1, available at: ftp://daac.ornl.gov/data/globalvegetation/ fire emissions v2.1) from Randerson et al. (2007) and Van der Werf et al. (2006) to correct for carbon fire emissions in the atmospheric inversions.1).Previous modeling studies indicated mean annual NPP between 7 and 13 Pg C y-1 (Cramer et al., 1999;Cao et al., 2001;Potter, 2003).Possibly, the too extensive forest cover simulated by LPJ-DGVM (17.28 Pg C y-1) and OR-CHIDEE (15.38 Pg C y-1) may have caused an overestimation of continental scale NPP.A realistic representation of savannah ecosystems is still a challenge for dynamic vegetation models.
The comparison of simulated interannual variations of the African carbon balance with atmospheric inversions reveals a consistent pattern but some discrepancies remain (Fig. 2).There is agreement of above average net uptake for 1997, 2000, 2001, and 2006 and a below average carbon balance for 1998, 2002, 2003 and 2005.The general pattern of continental scale IAV is consistent between NEP and NBP, which indicates a small contribution of fire emissions on the variability of the African carbon balance in the considered period.Relatively low interannual variability of fire emissions in Africa is consistent with van der Werf et al. (2006).Some of the deviations between the models and the inversions is certainly also related to uncertainties of the latter given that the density of atmospheric CO 2 measurement stations is low.
Differences between models are even more pronounced at regional level (Fig. 3).GPP numbers are similar between JULES, ORCHIDEE, and LPJ-DGVM for the Northern Savannah Belt, the Central Tropical Forest, and the Southern Raingreen Woodland, where LPJ-Guess represents only 50 percent of other modeled GPP numbers.GPP estimates for  (1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006).Due to the standardization method, 0 represents the mean of each individual distribution, but no neutral balance.NEP for models was calculated as the difference of GPP and TER.NEP from atmospheric inversions was calculated by removing fire emissions using GFED (van der Werf et al., 2006).NBP refers to the originally inversion results, and NEP minus model specific fire emissions (except JULES, which does not simulate fire).
the Horn of Africa and the South African Grassland are consistent between JULES and LPJ-DGVM, but significantly lower for ORCHIDEE but in the same range as LPJ-Guess.Annual NPP estimates are more consistent between all models.Exceptions are lower LPJ-Guess NPP at the Northern Savannah Belt, twice as much higher ORCHIDEE NPP than inter-model average at the Central Tropical Forest, as well as a high and low plateau situation in the South African Grassland region as represented by JULES/LPJ-DGVM and ORCHIDEE/LPJ-GUESS.
Despite the discrepancies among models regarding the absolute flux magnitudes, consistent patterns emerge between models and with satellite observations regarding seasonal changes of photosynthesis in the different regions (Fig. 4).Both, remotely sensed FAPAR and simulated GPP show that the seasonality of photosynthesis varies in concert with rainfall in all regions except for the inner tropical forest, where low intra-annual variation of rainfall creates climatic conditions without water limitation, reducing the effect of precipitation on photosynthesis.The heterogeneous vegetation of Madagascar together with known extensive anthropogenic transformations (Green and Sussman, 1990) are likely a major reason for the disagreement between the models and between models and remotely sensed FAPAR.

Regions of large interannual variability
Similar spatial patterns of interannual variability for GPP and NEP are estimated by all models (Fig. 5).Despite some differences of the patterns among the models, we can identify areas of large interannual variability of GPP and NEP in east and south Africa predicted by all models (Fig. 6).Both regions are dominated by herbaceous vegetation, including extensive agricultural land, and are known to be strongly influenced by El Nino conditions (Plisnier et al., 2003;Kogan, 2000).The independent remote sensing based FAPAR data confirms these two variability hotspot regions (Fig. 7).However, ORCHIDEE and JULES further simulate considerable variability in the northern savannahs and partly in the tropical forest, which is not evident in the satellite observations.

What drives NEP interannual variability -Photosynthesis or respiration?
The previous section indicated that regions of large interannual NEP variability are associated with also large GPP variability, suggesting that variations of photosynthesis are driving the variations of the net carbon balance.are strongly correlated with GPP anomalies in all models.In many cases anomalies of ecosystem respiration are also positively correlated with the NEP anomalies which may appear counterintuitive at first glance.If respiration would drive the carbon balance we would expect negative correlations; the positive correlations instead originate from the tight coupling of GPP and TER in the models, which is also evident in eddy-covariance based, estimates of GPP and TER (e.g.Reichstein et al., 2007;Baldocchi, 2008).In such case a positive anomaly of TER results in a positive anomaly of NEP because the GPP increase is even higher, than the (GPP induced) TER stimulation.
Although the overall pattern of GPP controlled NEP interannual variability is robust and consistent among models and regions, we find differences in the strength of this relationship depending on the model and region.JULES and LPJ-GUESS show strongest correlations between GPP and NEP anomalies (R GPP/NEP >0.92) in all regions, while LPJ-DGVM and ORCHIDEE simulate larger inter-regional variability and slightly weaker correlations.

The response of simulated gross primary production to meteorological conditions
Having identified GPP as a crucial driver of the carbon balance and the hotspot regions of largest interannual variability in east and South Africa, we now infer which meteorological conditions drive the variability of GPP.We use a principal component analysis to reduce the dimensionality of the meteorological dataset and to extract typical weather gradients (see Sect. 2.4).The first two modes of the PCA of the annual meteorological data explain 80% of its variability (Table 3) and are used to infer the primary driving factor of GPP interannual variability in the models.The first mode is most strongly associated with precipitation and specific humidity and thus represents a gradient of moisture availability.The factor loadings of the second mode show that it represents a gradient of increasing temperature and radiation with decreasing values of mode 2.
The correlation maps between the first principal component and GPP clearly show that moisture availability is the main limiting factor for photosynthesis (Fig. 8).This finding is in conjunction with Williams et al. (2008), who identified water stress as the primary governing factor of IAV  of photosynthesis.Correlation maps between the meteorological PCAs and the mean annual FAPAR from SeaW-iFS (Fig. 9) confirm that primary productivity responds to moisture in south and east Africa, which are the regions of largest interannual variability (see Sect. 3.2).The correlations between moisture and FAPAR on interannual scale in the Northern Savannah Belt are not as extensive as indicated by the models.This is possibly an artifact related to the relatively short time series (8 years) and coarse meteorological data.Camberlin et al. (2006) found more widespread strong correlations between integrated NDVI and rainfall for a 20 year time series in the northern savannahs.We also find no correlations between moisture and FAPAR in the Horn  of Africa region while the models suggest such relationship.For this region Camberlin et al. (2006) confirm no correlation between precipitation and vegetation productivity for the northern corner of the African Horn but indicate strong correlations with rainfall in the southern part of the region.The partly different spatial correlation patterns between PCA1 and GPP among the models is likely related to different parameterizations for water stress effects such as root profiles, soil depth, and the coupling between photosynthesis and transpiration via canopy conductance which largely determines soil water depletion.The simulation of water stress effects on photosynthesis has been previously identified as a major source for differences among models regarding interannual variability of GPP in Europe (Jung et al., 2007;Vetter et al., 2008).
The general water limitation of African vegetation's primary production is consistent with Churkina and Running (1998) based on Biome-BGC simulations and Jolly et al. (2005) based on the analysis of remote sensing data from MODIS.Using a production efficiency model Nemani et al. (2003) find that tropical Africa is primarily radiation limited, while both Churkina and Running (1998) as well as Jolly et al. (2005) masked the inner tropical regions and concluded that no climatic constrain limits productivity here.Interestingly, LPJ-DGVM and JULES also suggest that not rainfall but radiation limits GPP in some regions, especially in the inner tropics.This is represented by the negative correlations between PCA 2 and GPP in the tropical forest where simulated GPP increases with increasing radiation and temperature (Fig. 10).Given that temperature limitation does likely not play a role here, we can interpret the negative correlations between PCA2 and GPP as an indication for light limitation.The latter also makes sense since frequent cloud cover is present in the tropics, which controls incoming radiation with otherwise favorable climatic conditions for productivity.Light limitation for parts of the Amazon tropical forests has been inferred from studying wet to dry season transitions using in-situ measurements of CO 2 gas exchange (Saleska et al., 2003), manipulative experiments (Graham et al., 2002), and remote sensing based studies (Huete et al., 2006;Xiao et al., 2006;Nemani et al., 2003).Light limitation for parts of the Amazon result from access to groundwater during dry seasons when the vegetation capitalizes from the increases incoming radiation.From the correlation maps between the remote sensing based FA-PAR and PCA2 we also find patches in the inner tropics of Africa where productivity increases with increasing radiation but to a much smaller extent than suggested by some of the models.Thus, the role of light availability in the African tropics remains controversial.There are several possible reasons why we find only small areas with significant correlation between radiation and productivity from the remote sensing data: (1) the short time series (8 years) in combination with small variances make it difficult to achieve large correlations, (2) the uncertainty of the meteorological data that originate from global reanalysis, (3) possible errors in the satellite FAPAR retrievals due to subpixel cloud contamination, (4) a strong role of non-climatic limitations of productivity such as nutrients.If nutrients are not most limiting, ecosystems are more sensitive to climatic variations.Possibly, light limitation in the African tropics occurs where less phosphorous depleted soils are present since phosphorous is believed to be generally the main limiting factor for productivity in the tropics (Vitousek, 1984).The overriding effect of nutrient availability may explain why we find only small areas with light limitation from the analysis using the satellite data, in contrast to the extensive areas of light limitation indicated by LPJ-DGVM, JULES, and Nemani et al. (2003) because none of the latter considers explicitly nutrient cycles.

Conclusions
Using four terrestrial ecosystem models in combination with remote sensing based information of vegetation productivity we were able to identify that (1) the largest interannual variability of gross primary production and net ecosystem productivity are concentrated in east and south Africa, (2) interannual variations of gross primary production is driving net ecosystem production in the models, and (3) the availability of moisture is the primary determinant of interannual variations of gross primary production and consequently the net carbon balance.
Nevertheless, our current simulations reveal substantial discrepancies among models regarding the actual flux magnitudes.Future simulations should be performed with improved forcings by prescribing the actual distribution of vegetation types provided by a remote sensing based land cover map and improved meteorological reanalysis.In addition it should be investigated to what extent the models are able to reproduce the ecological patterns found in the synthesis of insitu measurements of carbon and water fluxes across Africa (Merbold et al., 2009).Given the significance of moisture control on the carbon fluxes in most parts of Africa, particular attention should be dedicated to the models' ability to accurately simulate soil hydrological conditions, and the sensitivity of photosynthesis and soil respiration to soil moisture.

Fig. 2 .
Fig. 2. Comparison of Atmospheric Inversion and Models based on standardized NEP(a) and NBP(b)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006).Due to the standardization method, 0 represents the mean of each individual distribution, but no neutral balance.NEP for models was calculated as the difference of GPP and TER.NEP from atmospheric inversions was calculated by removing fire emissions using GFED (van derWerf et al., 2006).NBP refers to the originally inversion results, and NEP minus model specific fire emissions (except JULES, which does not simulate fire).

Table 1 .
Continental annual average GPP, NPP per model (temporal standard deviation), all numbers are in Pg C y-1We use annual sums of GPP and Net Primary Production (NPP) over the entire study period to investigate the mean annual carbon fluxes at the continental as well as on the regional scale, and compare them to global numbers based on the IPCC Fourth Assessment Report(IPCC 2007).Modelled absolute NEP estimates strongly depend on factors that cannot be taken into account (e.g.change in land-use history) and thus are relatively meaningless.Hence, we discuss NEP only in terms of interannual variability as previously done in other contexts (e.g.

Table 2 .
Correlation coefficient (R) between GPP and NEP anomalies (left), and TER and NEP anomalies 1 (right) per region and model

Table 3 .
Eigenvectors and percent variance per Principal Component Modes