3.2.1. Critical Periods in Measurements
 Figure 5 shows the critical periods as detected in measured time series at each site (black lines). The critical periods for GPP (CPGPP) were detected at the beginning of the growing season (DoY 90–140) in the LOO pine forest (Netherlands). CPGPP were evenly distributed during the entire growing season in the THA spruce forest (Germany) and in the HYY pine forest (Finland). The three temperate beech forests showed a distinct CPGPP distribution throughout the season. CPGPP at beech sites were detected in the summer and early autumn (HES, eastern France, SOR, Denmark, HAI, Germany) and in the spring (SOR). At the Mediterranean holm oak forest site (PUE) in southern France, CPGPP occurred between DoY 150 and 260. At nearly all sites, the correlations between CPGPP and annual GPP anomalies were positive. An exceptional negative correlation was observed near the end of the growing season at HYY. The monthly variance of this period represented 7% of the annual variance. A potential mechanism for explaining the negative correlation at HYY in autumn was the coincidence of low Ta and high Rg (anticyclonic conditions). After mid-October, Pinus sylvestris was actively assimilating and responding positively to Rg, albeit at a much lower yield than during the peak growing season. Autumn Ta (DoY 270–335) was found to be positively correlated to annual Ta (data not shown). Therefore, the combination of low Ta and high Rg, which drove high autumnal canopy photosynthesis, was generally associated with low annual Ta (low annual GPP anomaly).
Figure 5. Moving correlation coefficients between monthly and annual NEE, GPP, and TER at seven CarboEurope sites. Flux tower measurements (black) and ORCHIDEE model simulations (red) are calculated from 6 to 8 years of data (see Table 1). Thick lines identify critical periods.
Download figure to PowerPoint
 CPTER were more evenly distributed than CPGPP throughout the growing season because the seasonal duration of ecosystem respiration activity was longer than the duration of canopy photosynthesis. Several periods of the year therefore played an important role in controlling TER interannual variability. Finally, CPNEE indicated the most important month for the annual carbon balance variability (Figure 5). At four out of seven forest sites (THA, HAI, HES and PUE), CPNEE were found only during the growing season. At three other sites (HYY, SOR and LOO) CPNEE were found near the margins of the growing season.
3.2.2. Comparison of Simulated and Observed Critical Periods
 Figure 5 further illustrates the seasonal distribution of critical periods modeled by ORCHIDEE for each site. A comparison of modeled with observed critical periods allowed an assessment of whether critical periods could be captured. Overall, the model was able to simulate observed CPGPP with κ > 0.2 (fair agreement) for four sites out of seven (Table 3). For the THA and HYY sites, CPGPP occurring at the middle and end of the growing season were correctly located by the model. However, the model did not capture the early growing season CPGPP at THA, HYY, or LOO due to a bias in phenology (as discussed earlier, see Figure 2). At LOO, this failure could also be due to the fact that GPP of the herbaceous understory was not simulated. The CPGPP simulation was poor at HYY (κ = 0.09), and the model was not able to reproduce observed negative autumnal correlations. On average, the model accurately located the CPGPP in the deciduous forests (SOR: κ = 0.57; HES: κ = 0.19; HAI: κ = 0.12, but for this last site κ was very sensitive to the significance threshold: κ = 0.3 for a significance threshold of 0.11 instead of 0.1 in equation (1)). In the early season at SOR, the error probably came from the nonmodeled herbaceous strata. At the Mediterranean holm oak forest (PUE), the simulated CPGPP also matched (κ = 0.32) observations except in midsummer (DoY 200–230). This shortcoming in the model reflected excessive water consumption, which overdepleted the soil water content each year during this period (Figures 2 and 3). Excessive depletions could have been caused by the use of a generic rather than a site-specific formulation of stomatal conductance as a function of SWC. Keenan et al.  recently showed that fitting the response of the dependency of photosynthetic capacity on soil water content variations helped to improve the ORCHIDEE simulation of canopy photosynthesis at PUE. Here, we used the standard version of the ORCHIDEE drought stress simulation. The overestimated water consumption caused only a slight bias in monthly summer canopy photosynthesis. Therefore, the summertime GPP interannual variability was dampened. Compared to GPP, NEE critical periods were more sensitive to the length of the drought stress period (DoY before 200 and after 230).
Table 3. Values of Cohen's Kappa (κ) With Significance Indexa
 ORCHIDEE was not as effective for locating CPTER as it was for CPGPP (3 sites with κ > 0.19, Table 3), due to the multiplicity of processes controlling ecosystem respiration (i.e., autotrophic and heterotrophic processes), their different responses to varying meteorology, and probably the difficulty in obtaining reliable estimates of ecosystem respiration from flux-tower data [see, e.g., van Gorsel et al., 2009]. The simulated CPTER and CPGPP showed similar patterns because of the growth respiration response, which closely paralleled canopy photosynthesis in ORCHIDEE, and in the modeled direct response of respiration to temperature. On the other hand, CPTER in winter were due only to variations in heterotrophic and maintenance respiration processes (see HES in Figure 5).
 As for the field data, CPNEE reflected the superimposed responses of TER and GPP. Any model error in simulating the gross flux response to climate thus worsened the model's ability to capture CPNEE. This was seen at HYY, for example, in Figure 5 with κ < 0. However, the κ statistics showed that, for most of the sites, the CPNEE model-data agreement was fair (3 sites, Table 3) or moderate (2 sites).
 One explanation for the differences between the simulated and observed critical periods shown in Figure 5 could be that the model did not account for variations in woody biomass, soil carbon stocks or LAI following thinning or defoliation (i.e., nonclimatic or age-related changes). The HYY, THA, HES, and SOR forests were thinned or otherwise affected by severe storms during the observation period, which induced an increase in interannual flux variability that was not incorporated by the model. Vesala et al.  showed that a thinning event removing 25% of the basal area at the HYY Pine forest (Finland) in spring 2002 had no significant effects on NEE due to the compensating effects of ecosystem respiration and canopy photosynthesis and the enhancement of understory photosynthesis. Thinnings at HES during the winters of 1998–1999 and 2004–2005, in which 25% of the basal area was cut, did not result in any strong reduction in NEE or GPP at stand scale [Granier et al., 2008]. However, defoliation by Lymanthria dispar at Puechabon led to a 15% drop in GPP in 2004 [Allard et al., 2008]. Therefore, we could not conclude whether the differences between modeled and simulated critical periods can be attributed to thinning or defoliation episodes or to limitations in model parameterizations.
 Other model shortcomings for identifying the critical periods may be caused by the changing NEE footprints during the year [Chen et al., 2009a] or by unresolved seasonal errors in the partitioning between ecosystem respiration and canopy photosynthesis. The results presented here compare carbon flux simulations with either gap-filled measurements (in the case of NEE) or statistically separated [Reichstein et al., 2005] gap-filled time series (canopy photosynthesis, ecosystem respiration). In the latter case, we were therefore essentially comparing the results of two models (the ORCHIDEE process-based model versus statistically partitioned gross fluxes). Desai et al.  had shown on a range of sites and years that different flux-separation methods converged to similar monthly sums. This tended to confirm that the statistical separation of canopy photosynthesis and ecosystem respiration did not affect the detection of critical periods in the data. Further, Moffat et al.  showed that structurally different gap-filling schemes yielded similar results for annual NEE sums. However, one should keep in mind that the definition of canopy photosynthesis as the difference between ecosystem respiration and net carbon exchange introduces a spurious correlation between both elementary fluxes [Vickers et al., 2009b]. Any error in the estimation of ecosystem respiration yields a similar uncertainty in the estimate of canopy photosynthesis. The comparison of elementary modeled and partitioned fluxes can therefore best be considered an assessment of the plausibility of model estimation (see also the discussion by Ibrom et al. ).
 Only three out of seven sites had a significant κ for TER critical periods (p < 0.05), compared to four and five sites for GPP and NEE, respectively (Table 3). Based on equation (3), the probability of obtaining three or more successes out of seven sites by chance was 3.7 10−3. We concluded from this that the critical periods reproduced by ORCHIDEE were very unlikely to reflect only chance.
3.2.3. Comparison of Simulated and Observed Critical Periods With Meteorological Drivers
 The levels of agreement between model results and data based on kappa statistics are provided in Table 4. Crosses indicate if ORCHIDEE significantly identifies the meteorological drivers that determine the observed critical periods.
Table 4. Ability of the ORCHIDEE Model to Simulate Significant Correlation Between the Critical Period Fluxes (for GPP, TER, or NEE) and Meteorology (Ta, SWC, VPD, and Rg) Based on Kappa Statisticsa
| ||CP Flux||Hyy||Sor||Loo||Hai||Tha||Hes||Pue|
|Ta||GPP||X||X|| || ||X||X||X|
|TER||X||X|| ||X||X||X|| |
|NEE||X|| ||X|| ||X||X||X|
|SWC||GPP|| ||X||X||X|| ||X||X|
|TER||X||X|| || || ||X||X|
|NEE||X|| || ||X||X|| ||X|
|VPD||GPP|| ||X||X|| ||X||X||X|
|TER|| || ||X||X||X||X|| |
|NEE||X||X|| ||X||X|| || |
|Rg||GPP||X||X|| || ||X||X||X|
|TER||X|| ||X||X||X|| ||X|
|NEE|| || ||X||X|| ||X||X|
 For at least four of the seven sites, the model significantly captured the drivers of the observed critical periods (Table 4). According to equation (3), the probability that four or more of seven values would be significant by chance was very unlikely (1.9 10−4). Therefore, we concluded that, overall (though not for a particular site), the model was able to attribute critical periods to the correct climate drivers.
 The model correctly reproduced the r significance between CPGPP and Ta or SWC; five of the seven sites had significant r. The overall correlations between CPTER and Ta were also well simulated, although the CPTER for a given site was not always well captured (Table 3). The agreement between simulated and measured r for CPTER and SWC was lower (four significant sites out of seven). The agreement between modeled and measured correlations between CPNEE and SWC was low (four out of seven sites) but significant, and the correlation between CPNEE and Ta was significantly reproduced by the model for five out of seven sites. When we used ORCHIDEE gridded results to simulate critical periods at the European scale in section 3.3, these limitations became significant.
 After demonstrating relatively good agreement between measured and modeled fluxes and ancillary variability (section 3.1), we showed (section 3.2) that ORCHIDEE was also able to reproduce the basic features of the detection patterns for critical periods. ORCHIDEE, a PFT-parameterized DGVM, performed reasonably well at simulating fluxes and identifying critical periods at the site scale, although results are more significant for GPP and NEE than for TER. Agreement was within the expected boundaries that were widened due to the use of generic rather than site-specific parameters. Using site-specific parameters was not possible in a continental simulation. However, the comparisons enhanced our confidence in the overall ability of ORCHIDEE to simulate realistic spatiotemporal patterns in critical periods at the continental scale, as described below.