Introduction

The Earth’s climate is warming as a result of rapidly increasing CO2 emissions and the global mean temperature is expected to increase by 1.1–6.4°C by the end of this century (IPCC 2007). Although nearly 30% of carbon (C) released by anthropogenic activities is sequestered by terrestrial ecosystems (Canadell and others 2007), whether this natural sink will be sustainable into the future is a major concern (Luo and Weng 2011). The capacity of ecosystem C sequestration depends on the magnitude of net ecosystem exchange of CO2 (NEE), which usually varies among years. Large interannual variability (IAV) in NEE has been observed at almost all eddy-flux sites over the world (Baldocchi 2008).

Climatic variables, such as solar radiation, temperature, and water conditions (Barr and others 2007; Pintér and others 2008; Yuan and others 2009) as well as cloud cover, drought, snow cover, and El Niño-Southern Oscillation (ENSO, Baldocchi and others 2001; Goulden and others 1996; Weber and others 2009), are believed to be the direct drivers of the variation in NEE. However, climatic variables can also indirectly drive IAV in NEE by regulating ecological and physiological processes such as photosynthetic and respiratory traits (Humphreys and Lafleur 2011) and phenological features (for example, growing season length, transition dates, and phenological lags, Dragoni and others 2011; Richardson and others 2009; Wu and others 2013). In addition, stand age and nutrient conditions may also change ecophysiological properties (Buchmann and Schulze 1999; de Beeck and others 2010). In this paper, we defined the effects of ecological and physiological changes on IAV in C fluxes caused by either climate or other factors such as biotic effects. The direct effects of climatic variations were treated as climatic effects. Due to the complex interactions between the climate and ecophysiological processes, few studies have explicitly quantified the two types of effects on IAV in NEE, separately.

Over the past decade, two statistical approaches, the homogeneity-of-slopes method (Hui and others 2003) and the crossed model (Richardson and others 2007) have been developed to address the issue. Both methods apply models that simulate C fluxes with yearly varying parameters, and both consider the variation of NEE from changes of model parameters as biotic effects and those directly from changes of climatic variables as climatic effects. Analysis of variance (ANOVA) is then used to obtain the relative importance of the biotic and climatic effects in both methods. However, the multiple linear regression model used in Hui and others (2003) produced considerable model-data mismatch in some ecosystems (Polley and others 2008; Teklemariam and others 2010). Richardson and others (2007) used a process-based model to estimate the relative importance of climatic and biotic effects as well as the magnitude of the two effects each year, which has been successful for Howland forest, USA, but may not be suitable for other ecosystems. Thus, more flexible methods need to be developed with site-specific data sets.

Previous studies with these approaches have found that biotic effects became more important as the temporal scale increased from days to years (Hui and others 2003; Polley and others 2008; Richardson and others 2007), whereas climatic effects, mainly at diel to seasonal scales, had the opposite trend (Baldocchi 2008; Luo and Weng 2011). Moreover, the importance of biotic effects to IAV in NEE varied among ecosystems (for example, grassland > forest > peatland, Hui and others 2003; Polley and others 2008; Teklemariam and others 2010), and between deciduous and evergreen forests (deciduous > evergreen, Richardson and others 2007; Wu and others 2012).

Forest plantations, occupying about 200 million ha over the world (FAO 2007), are a large wood production and efficient C sink due to fast growth (Carle and Holmgren 2008; Pan and others 2011). In terms of area, China has the largest portion of global plantations (about one-third), 40% of which are planted in Southern and Eastern China with a subtropical monsoon climate (Huang and others 2012). However, the potential C sequestration in plantations was questioned by a recent meta-analysis, suggesting that soil C concentration in plantation forests was lower than that in natural forests (for all trees and the Genus Pinus, Liao and others 2012). Understanding this inconsistency requires more detailed studies on mechanisms of C cycling in plantations and their biotic responses to climate change, to which the partitioning of climatic and biotic effects would provide useful information.

Therefore, in this study, we aimed to investigate the relative importance of biotic and climatic effects on IAV in NEE and their potential interactions at a subtropical evergreen plantation in China. Instead of using the specific-scale functions in Richardson and others (2007), we applied the Bayesian information criterion (BIC) to optimize the relationships of climatic variables with maximum photosynthetic rate (A m) and temperature sensitivity (Q 10) using multiple regression models. An empirical model with the optimized relationships and yearly varying parameters was then constructed to simulate seasonality and IAV in NEE. A modified statistical approach from Richardson and others (2007) was used to partition the climatic and biotic effects on IAV in NEE (and its component fluxes GPP and RE) and their relative importance. Because the response magnitude of C fluxes to climate variability is the sum of biotic and climatic effects, the correlation between biotic and climatic effects can potentially reveal the possible responses of the ecosystem C cycle to future climate change and help clarify whether a positive feedback exists between the C cycle and climatic change (Cox and others 2000; Luo and others 2009). If the biotic effects are negatively correlated with climatic effects with similar magnitudes, ecosystem C fluxes may not fluctuate with climatic change and can be a robust C sink/source. To evaluate the performance of our modeling method, we also applied our approach to another two ecosystems, an oak forest and a C3 grassland in the US. The three sites not only represent different ecosystem types (evergreen and deciduous forests, and grasslands), but also cover a wide range of geography and climate. Therefore, the comparison of the results across the three ecosystems will strengthen the application of our approach.

Methods

Site Information

In this study, we mainly focused on a subtropical plantation in China using a statistical modeling approach to examine climatic and biotic effects on IAV in C fluxes. Another two ecosystems in the USA were used to validate the approach. Thus, the three sites included Qianyanzhou, Missouri Ozark, and Duke Forest Open Field. The Qianyanzhou flux site (QYZ, 26°44′29″N, 115°03′29″E, a.s.l. 100 m) is located in Jiangxi Province, China. The site is controlled by a subtropical monsoon climate, whereas other areas at a similar latitude are occupied by arid steppes and deserts (Huang and others 2007). The mean annual temperature and precipitation are 17.9°C and 1,475 mm, respectively, based on the meteorological record from 1985 to 2007 (Wen and others 2010). High temperature and drought often occur in summer, suppressing plant physiological activities. The vegetation is a needle-leaved forest plantation that is 25 years old and approximately 13 m tall. The dominant species are Pinus massoniana, Pinus elliottii, and Cunninghamia lanceolate.

The Missouri Ozark flux site (MOZ, 38°44′39″N, 92°12′00″W, a.s.l. 219 m) is located at the Baskett Research and Education Area (BREA) in Missouri, USA. The climate of the area is warm, humid, and continental. The mean annual temperature and precipitation are 13.6°C and 1,023 mm over the period of 1971–2000, respectively. The vegetation is a deciduous forest that is 77 years old and approximately 13 m tall. The dominant species are white oak (Ouercus alba) with other oak species and hickories (Yang and others 2010).

The Duke Forest Open Field (DK1, 35°58′16″N, 79°05′36″W, a.s.l. 168 m) is located at the Blackwood Division of the Duke forest in North Carolina, USA. The regional climate is characterized by warm and humid summers and mild winters with mean annual temperature of 15.5°C and mean annual precipitation of 1,145 mm based on the long-term record (111 years). The dominant species is the C3 grass, Festuca arundinacea (Stoy and others 2008).

Data Sources

Half hourly data for NEE and corresponding climatic variables at the QYZ flux tower (2003–2009) were obtained from CERN (Chinese Ecosystem Research Network, www.cern.ac.cn/), whereas the data at MOZ (2005–2009) and DK1 (2003–2007) were from AmeriFlux (public.ornl.gov/ameriflux/index.html). These variables included F c (CO2 flux), u* (friction velocity), PAR (photosynthetically active radiation), T a (air temperature), T s (soil temperature), PPT (precipitation), SWC (soil water content), and VPD (vapor pressure deficit).

Gap-Filling

The raw data were preprocessed by spike screening and nighttime filtering based on the methods in Papale and others (2006). Then a Q 10 function was used to model the nighttime flux, which represents the nighttime ecosystem respiration (RE, Richardson and Hollinger 2005) and fill the gaps:

$$ RE = R_{10} Q_{10}^{{\frac{{T_{\text{s}} - 10}}{10}}} , $$
(1)

where R 10 is the respiration rate at 10°C, Q 10 is the temperature sensitivity of RE. The parameterization of equation (1) follows the method from Reichstein and others (2005) with slight modifications. First, we set 1-month moving windows with a step of 1 day. If the valid nighttime data in a certain moving window were less than 100 points, the window was expanded to get enough data points. Then the fitted parameters R 10 and Q 10 were obtained for each moving window. Only the R 10 values within a range of 0–0.15 mg C m−2 s−1 and the coefficient of variance (CV) less than 50% were accepted (Reichstein and others 2005). These valid R 10 values were averaged with the inverse of standard error as a weighting factor. The weighted R 10 was considered as the long-term reference respiration rate through the whole period. Then we put the weighted R 10, a constant, into equation (1), and obtained the estimates of daily Q 10 using 1-month moving window with a step of 1 day. Once the parameters in equation (1) were estimated, we applied it to fill the gaps at nighttime. Daytime RE was also estimated by this method to extend the equation to daytime.

GPP was estimated as the difference between estimated daytime RE and observed NEE. A Michaelis–Menten equation (Falge and others 2001) was used to describe the GPP for the daytime:

$$ GPP = \frac{{PAR \cdot A_{\text{m}} }}{{PAR + K_{\text{m}} }}, $$
(2)

where A m is the maximum photosynthetic rate and K m is the Michaelis constant, which is PAR at which the photosynthetic rate is half of A m. We fitted A m and K m using a 1-month moving window as described above to get the daily parameters of a photosynthesis model. We then used these parameters to estimate daytime GPP. The daytime flux gaps for NEE were filled by using the equation: NEE = RE − GPP.

Empirical Model

The seasonality of the parameters in the gap-filling model (for example A m, R 10 and Q 10) has been proven to be related to climatic variables (for example, temperature, soil moisture, and VPD) (Ricciuto and others 2008; Richardson and others 2006). At the QYZ site, SWC has been proven to be a good indicator for maximum carboxylation rate (Ju and others 2010). In addition, Yu and others (2008) have used multiple linear regression with T a and SWC to simulate Q 10. In this study, we applied multiple linear regression of meteorological variables to simulate the seasonality of A m and Q 10. Quadratic and interactive terms were included in the regression to capture the potential nonlinear relationships between climatic variables and the parameters. Regression variable selection was based on Bayesian information criterion (BIC, Burnham and Anderson 2002).

First, we considered daily T a, VPD, and SWC as the potential explanatory variables for A m, and T s, and VPD and SWC for Q 10, according to univariate linear correlation between climatic variables and model parameters. We then constructed a series of candidate models with different combinations of potential explanatory variables and their quadratic and two-order interactive terms. The BIC of each candidate model was calculated by

$$ BIC = - 2 \times \log \,(Likelihood) + k \times \log \,(n), $$
(3)

where k is the number of parameters, n is the length of data, and Likelihood is the likelihood function. The likelihood function is the probability density function (PDF) of the joint distribution of the model parameters when the data are given. The model with the lowest BIC was selected and then the explanatory variables in this model were used in the next steps. After model selection, the best model for Q 10 at QYZ was identified as:

$$ {Q_{10}} = {a_{0}} + {a_{1}} \times {T_{\text{s}}} + {a_{2}} \times SWC + {a_{3}} \times {T_{\text{s}}} \times SWC $$
(4)

and the best model for A m was

$$ A_{\text{m}} = b_{0} + b_{1} \times T_{\text{a}} + b_{2} \times VPD + b_{3} \times SWC , $$
(5)

where a i (i = 0, 1, 2, 3) and b i (i = 0, 1, 2, 3) are parameters. Therefore, we constructed an empirical model for NEE, which is expressed as

$$ NEE_{\text{predict}} = R_{10} Q_{10}^{{\frac{{T_{\text{s}} - 10}}{10}}} - \frac{{A_{\text{m}} \cdot PAR}}{{PAR + K_{\text{m}} }}, $$
(6)

where Q 10 and A m are functions of climate variables and R 10 and K m are constant parameters.

The parameters of the empirical model were estimated by Bayesian parameterization with the Markov Chain Monte Carlo (MCMC) method (McCarthy 2007) using non-gap filled half-hour data. The posterior probability of a parameter set θ i could be obtained from

$$ \Pr (\theta_{i} |Data) = \frac{{\Pr \,(\theta_{i} ) \times \Pr (Data|\theta_{i} )}}{{\sum\nolimits_{j} {(\Pr \,(\theta_{i} ) \times \Pr (Data|\theta_{i} ))} }}, $$
(7)

where Pr(θ i |Data) is the posterior probability of θ i , which means the joint probability of parameter set θ i when given the data. Pr(θ i ) is the prior probability of θ i , and Pr(Data|θ i ) is the likelihood function. For a normal distribution, the likelihood function is expressed as

$$ Likelihood = \prod\limits_{i = 1}^{n} {\left( {\frac{1}{{\sqrt {2\pi } \sigma }}e^{{ - \frac{1}{2}\left( {\frac{{\varepsilon_{i} }}{\sigma }} \right)^{2} }} } \right)} , $$
(8)

where \(\varepsilon_i\) is the model residual. When the maximum likelihood method is used to estimate model parameters, the statistics σ can be calculated from the following equation:

$$ \sigma^{2} = \frac{{\sum\nolimits_{i = 1}^{n} {\varepsilon_{i}^{2} } }}{n} $$
(9)

There are a total of ten parameters to be optimized with uniform prior distributions.

The same approach was applied to MOZ and DK1, but the best models for Q 10 and A m at these sites were

$$Q_{10} = a_{0} + a_{1} \times T_{\text{s}} + a_{2} \times SWC + a_{3} \times T_{\text{s}} \times SWC A_{\text{m}} = b_{0} + b_{1} \times T_{\text{a}} + b_{2} \times VPD + b_{3} \times VPD^{2}$$
(10)

and

$$\begin{aligned} Q_{10} &= a_{0} + a_{1} \times T_{\text{s}} + a_{2} \times SWC + a_{3} \times T_{\text{s}} \times SWC \\ A_{\text{m}} &= b_{0} + b_{1} \times T_{\text{a}} + b_{2} \times VPD + b_{3} \times VPD^{2} \end{aligned}$$
(11)

respectively.

Statistical Analysis

At the QYZ site, we ran the models with separate year’s data, and obtained seven parameter sets. Following Richardson and others (2007) procedure, we then ran the empirical model by crossing each “driver year” with each “parameters year” and obtained a 7 × 7 matrix of model predictions. This “cross model” was run for 500 iterations. The analysis of variance (ANOVA) with factors of “driver year,” “parameter year,” “driver year” × “parameter year” interaction, and “parameter year” × “model run” interaction was used to partition the total variance of the model predictions. The majority (>95%) of the interannual variation came from the factors “driver year” and “parameter year,” the contributions to total variation were regarded as the contributions of climatic and biotic effects to the interannual variation in C fluxes, respectively. This approach was applied to model predictions aggregated at daily, weekly, monthly, seasonal, and annual periods, resulting in the contributions of biotic and climatic effects to the IAV of C fluxes at the corresponding periods. To investigate the possible relationship between biotic and climatic effects, the main effects of the factor “driver year” and “parameter year” were defined as the magnitude of biotic and climatic effects, respectively. The same approach was applied to MOZ and DK1. Following Yuan and others (2009) approach, we used standard deviation to represent the absolute interannual variability (AIAV) and the coefficient of variation (CV) to represent the relative interannual variability (RIAV).

All analyses but Bayesian parameter estimation were applied in R (R Development Core Team 2011), which is a free, open source software environment for statistical computing and graphics with thousands of packages. WinBUGS (version 1.4.3, http://www.mrc-bsu.cam.ac.uk/bugs/) was used to conduct the Bayesian analysis.

Results

Interannual Variability in Climate and Carbon (C) Fluxes

The climatic variables (PAR, T a, T s, PPT, SWC, and VPD) showed large interannual variability (IAV) at both annual and monthly scales (Figure S1, S2, S3). The relative IAV (RIAV, represented by the coefficient of variation (CV)) in annual PAR, PPT, and VPD was greater than that in T a, T s, and SWC at QYZ (Figures S1B, D, F), whereas the RIAV in PPT, SWC, and VPD was greater than that in PAR, T a, and T s at MOZ and DK1 (Figures S2, S3). At the monthly scale, the RIAV in T a and T s was lower than that in other climatic variables at all the sites. The variables associated with water conditions (PPT, SWC, and VPD) showed larger IAV in the growing season than other periods (Figures S1, S2, S3). Note that T a and T s did not show the same patterns at the annual scale at QYZ (r 2 = 0.22, P > 0.05) and DK1 (r 2 = 0.46, P > 0.05), although the seasonality and IAV at the monthly scale were close (Figures S1C, S3C). For the period considered in this study, two extreme climate events occurred at QYZ, a severe summer drought with high temperatures in 2003 and an ice storm event from late January to early February 2008. At MOZ, the precipitation in 2006 and 2007 was lower than that in other years. At DK1, the precipitation in 2007 was the lowest over the period 2003–2007.

During the study period, the annual NEE was −333 ± 47 (mean ± SD), −479 ± 65 and 34 ± 92 g C m−2 y−1 at QYZ (2003–2009), MOZ (2005–2009), and DK1 (2003–2007), respectively (Figure 1B, D, F). At the monthly scale, AIAV of NEE was 4–17, 6–58, and 5–41 g C m−2 mon−1 in QYZ, MOZ, and DK1, respectively. The RIAV in the monthly NEE was much larger than that at the annual scale. Although the monthly NEE was strongly correlated with the climatic variables, significant correlations were only found between NEE and T s at QYZ (r 2 = 0.61, P < 0.05) and between NEE and PPT at DK1 (r 2 = 0.65, P < 0.05) at the annual scale.

Figure 1
figure 1

Seasonal (A, C, E) and interannual variability (B, D, F) in NEE for the study period at QYZ, MOZ, and DK1. The scaled NEE was calculated by (X i  − mean)/SD, where X i is NEE in a specific month during the study period. B, D, F represent observed and modeled annual NEE. Error bars in modeled values are the range of the 95% credible interval predicted by the empirical model. The gray dashed horizontal lines are the means of the observed annual NEE.

Full size image

Model Parameters and Performance

The modeled NEE at the half-hour, daily, weekly, and monthly scales had good agreement with observed values for all three sites (Figure 2). At the annual scale, the modeled NEE was −324 ± 39 (mean ± SD), −475 ± 51 and 34 ± 95 g C m−2 y−1 at QYZ, MOZ, and DK1, respectively, closely matching the corresponding observed values in terms of magnitude, RIAV, and interannual patterns (r 2 = 0.89, P < 0.01; r 2 = 0.88, P < 0.05; and r 2 = 0.96, P < 0.01, respectively, Figure 1B, D, F). At QYZ, the model caught the suppression of summer drought on GPP and RE in 2003, but failed to capture the weak summer suppression in 2004 and 2005 (Figures S6A, S7A). However, the model underestimated the RE on cold days with high Q 10 (Figures 5B, S7A). At the annual scale, the model overestimated RE in 2009 and GPP in 2008 and 2009 (Figures S4B, S5B). At MOZ, the NEE had a clear seasonality, which was well captured by the model. However, the model overestimated both GPP and RE in 2006 (Figures S4D, S5D) when the temperature was higher and the precipitation was lower than those in the normal years (Figure S2B, D). At DK1, some large model-data mismatching of NEE and GPP occurred on some days in warmer and drier years (2006 and 2007, Figures 2C, S6C).

Figure 2
figure 2

Comparison of modeled and observed NEE at QYZ (A), MOZ (B), and DK1 (C) for the study period. The values represent the weekly mean NEE.

Full size image

Parameterization of the empirical model was conducted yearly during the study period with seven parameter sets for QYZ and five for MOZ and DK1. The PDFs of parameters K m and R 10 varied greatly among years at all three sites with larger variation at MOZ and DK1 (Figure 3). Parameter K m was significantly correlated with annual PAR (r 2 = 0.60, P < 0.05, Figure 4A) and SWC (r 2 = 0.62, P < 0.05, Figure 4B) at QYZ, and with annual PAR (r 2 = 0.80, P < 0.05, Figure 4C) and PPT (r 2 = 0.89, P < 0.05, Figure 4D) at DK1 but not at MOZ as well as R 10.

Figure 3
figure 3

Posterior probability distribution functions (PPDF) of parameters K m (A, C, E) and R 10 (B, D, F) at QYZ, MOZ, and DK1. The area below each curve is 1. At QYZ, the distribution of K m in 2003 and 2007 was not significantly different, the distribution of R 10 in 2008 was not significantly different from those in 2003 and 2004. At MOZ, the distribution of R 10 in 2008 and 2009 was not significantly different. Others are all significantly different for K m and R 10 at the three sites.

Full size image
Figure 4
figure 4

Linear relationships between parameter K m and annual mean climatic variables at QYZ (A, B) and DK1 (C, D). There were no other significant relationships of K m and R 10 with other climatic variables for all three sites.

Full size image

At QYZ, the A m increased by 0.035 ± 0.0058 mg C m−2 s−1 (mean ± SD) with the increment of 1°C in T a based on the yearly relationship between A m and T a during the study period and decreased by 0.48 ± 0.11 mg C m−2 s−1 with 1 kPa in VPD, but the effects of SWC on A m were minor. The maximum A m was 0.4920 mg C m−2 s−1 in 2006, whereas severe summer drought resulted in a low A m in both 2003 and 2007 (Figure 5A). The Q 10 values decreased with T s (0.037 ± 0.029 for every 1°C when SWC was kept at 30%) and increased with SWC (0.042 ± 0.038 for every percent of SWC when T s was kept at 18°C). The mean Q 10 in a calendar year ranged from 1.9 to 2.7, and the Q 10 pattern was different among years (Figure 5B). For example, the Q 10 values declined from 3.5 to 2.0, and then increased to the former level in 2005 and 2007, whereas it declined from 4.5 to 1.2 during the whole year in 2003 (Figure 5B), which might have resulted mainly from the continuously declining SWC due to severe drought (r 2 = 0.41, P < 0.001). In 2009, the Q 10 values remained relatively unchanged within 1.6–2.1 (Figure 5B), probably due to the relative low variations of SWC.

Figure 5
figure 5

Daily A m (A) and Q 10 (B) from 2003 to 2009 at QYZ. The values were obtained from the equations A m = b 0 + b 1 × T a + b 2 × VPD + b 3 × SWC and Q 10 = a 0 + a 1 × T s + a 2 × SWC + a 3 × T s × SWC using the values of parameters from the parameterization of the empirical model. Daily A m and Q 10 in the other two ecosystems are shown in Supplementary material.

Full size image

The relationships between daily A m and climatic variables T a and VPD at MOZ were nonlinear and varied among years. However, the A m reached the maximum when T a was about 24°C and VPD about 1 kPa. The Q 10 values deceased by 0.052 ± 0.049 with the increment of 1°C in T a and increased by 0.038 ± 0.028 with 1% in SWC. At DK1, the daily A m increased with T a (0.0062 ± 0.0032 mg C m−2 s−1 for 1°C), and reached the maximum when VPD was about 1 kPa. However, the relationships between Q 10 and climatic variables T s (−0.056 to 0.049 for 1°C increment when SWC kept at 30%) and SWC (−0.0056 to 0.049 for 1% increment when T s kept at 16°C) were not consistent among years.

Climatic and Biotic Effects on C Fluxes

The contributions of climatic effects to the IAV in C fluxes decreased when the temporal scale increased from days to years for all three sites (Figure 6A, C, E) with a larger contribution in the forests (QYZ and MOZ) than in the grassland (DK1). At QYZ, the contribution of climatic effects to the IAV in NEE was closer to that in GPP than RE (Figure 6A), whereas DK1 had the opposite pattern (Figure 6E). At MOZ, the contributions of climatic effects were similar for NEE, GPP, and RE (Figure 6C). At all three ecosystems, the contributions of climatic effects to the IAV in NEE were lower in the growing season than the non-growing season (Figures 7, S10, S11). The magnitude of climatic effects on annual NEE at QYZ, MOZ, and DK1 ranged from −131 to 127, from −73 to 107, and from −60 to 89 g C m−2 y−1, respectively (Figure 8A–C). Annual PAR, SWC, and VPD all were correlated with climatic effects on NEE at QYZ (Table 1). The climatic effects on NEE were correlated with annual PAR in MOZ and annual VPD in DK1 (Table 1).

Figure 6
figure 6

The relative contribution of climatic (A, C, E) and biotic effects (B, D, F) to IAV in fluxes of QYZ, MOZ, and DK1 at the temporal scales from daily to yearly. The error bars represent the SD of a certain contribution at the specific scale across the whole period of year.

Full size image
Figure 7
figure 7

Contributions of climatic and biotic effects to interannual variability in NEE (A–D), GPP (E–H), and RE (I–L) at daily, weekly, monthly, and seasonal scales at QYZ. The values were derived by partitioning the variance by analysis of variance (ANOVA) in crossed model predictions at the specific day (or week, month, season) of the year. Those for the other two ecosystems are shown in Supplementary material.

Full size image
Figure 8
figure 8

Climatic and biotic effects on NEE (A–C), GPP (D–F), and RE (G–I) at the annual scale, and their relationship between the two effects (J–L) in QYZ, MOZ, and DK1. The values were calculated as the difference between the estimated value of climatic (or biotic) effect in the specific year and the mean across 7 years. The error bars represent standard deviation for the study period. Linear relationships between climatic and biotic effects in J were significant for NEE (r 2 = 0.80, P < 0.01) and GPP (r 2 = 0.61, P < 0.05).

Full size image
Table 1 Correlation Coefficients (r) Between Annual Climatic Variables and Climatic (or Biotic) Effects on IAV in Annual NEE at QYZ, MOZ, and DK1
Full size table

On the contrary, the contributions of biotic effects increased when the temporal scale increased from days to years (Figure 6B, D, F). The magnitude of biotic effects on annual NEE at QYZ, MOZ, and DK1 were from −95 to 170, from −99 to 78, and from −204 to 318 g C m−2 y−1, respectively (Figure 8A–C). Interestingly, we found a significant negative correlation between climatic and biotic effects on NEE (r 2 = 0.80, P < 0.01) and GPP (r 2 = 0.61, P < 0.05) at the annual scale at QYZ (Figure 8J) but not at the other two sites.

Discussion

Climatic and Biotic Drivers of the IAV in NEE

Quantifying net ecosystem exchange of CO2 (NEE) and its components (GPP and RE) and identifying its controlling drivers is critical for understanding ecosystem functioning and global C cycling. In this study, we partitioned the IAV in C fluxes into climatic and biotic effects, and quantified the relative contributions of these two effects. At QYZ, the relative contribution of climatic and biotic effects to the IAV in NEE was closer to that in GPP compared to RE, whereas the pattern was the opposite at DK1. At MOZ, GPP and RE were similar to NEE in terms of climatic and biotic effects (Figure 6). Yuan and others (2009) defined the C uptake period as the continuous period when a 5-day running average of NEE was negative and suggested that the longer the C sink period, the more important GPP (and less importance of RE) was to NEE, because GPP was the dominant flux in the C sink period whereas RE drove the variation of NEE during the dormancy period of plants. Our results confirmed this idea, which showed that the C sink periods were 298, 254, and 183 days at QYZ, MOZ, and DK1, respectively.

Annual NEE was usually linearly related to annual solar radiation, air or soil temperature, and water conditions (Archibald and others 2009; Pintér and others 2008; Wen and others 2010) as well as climatic factors in specific month(s) (Allard and others 2008; Zhang and others 2011a) and climatic events such as severe drought and El Niños (Aires and others 2008; Wharton and others 2009). However, the simple correlations between annual NEE and climatic variables above might confound the underlying mechanisms of IAV in NEE, because climatic effects on IAV in NEE were different from annual NEE. For example, although annual NEE was significantly correlated with T s at QYZ (r 2 = 0.61, P < 0.05), the modeling results showed that radiation and water conditions, rather than temperature, were the main climatic drivers on IAV (Table 1). At MOZ and DK1, the primary climatic drivers of IAV in NEE were radiation and water conditions, respectively (Table 1). The warm climate of the three sites indicated that the temperature did not constrain the C fluxes, whereas the relatively dry summer made water conditions a primary controller of NEE at QYZ and DK1. Radiation might affect NEE at QYZ via light supply, change in evapotranspiration, and photoinhibition. At MOZ, there was no strong constraint of water conditions on NEE due to the year-round humid climate, resulting in the dominant effect of radiation.

IAV in NEE was also affected by biotic drivers such as maximum photosynthetic rate (A m) and reference respiration rate (R 10, Humphreys and Lafleur 2011), which contributed to biotic effects. At all three sites, we found significant changes in A m, the Michaelis constant (K m), R 10 and Q 10 (Figures 3, 5, S8, S9) among the years as well as in other studies (Richardson and others 2007; Adkinson and others 2011). IAV in biotic drivers may mainly result in biotic effects on IAV in both annual GPP and RE. Because IAV in NEE was controlled by different fluxes (GPP or/and RE) among the three ecosystems, the main biotic drivers of annual NEE also differed from photosynthetic (A m and K m) to respiratory capacities (Q 10 and R 10). However, it is difficult to identify one or two primary biotic drivers due to complex interactions and the limited data.

Climatic and Biotic Effects on IAV in NEE

Variations in climatic variables (diel and seasonality) drive short-term variability in C fluxes, whereas biotic drivers may be more important at the long-term scale (Baldocchi 2008; Luo and Weng 2011). Our results showed that the contributions of biotic effects to variations in C fluxes become more important with the increasing temporal scale from days to years in spite of different vegetation types (Figure 6, Richardson and others 2007; Wu and others 2012). Temporally, the contributions of biotic effects were larger in the growing season than the non-growing season within a year (Figures 7, S10, S11), which was consistent with our definition of biotic effects as the result of IAV in plant physiological activities. Spatially, it was suggested that the biotic effects were more important in deciduous than evergreen vegetation for both forests and wetlands (Adkinson and others 2011; Richardson and others 2007; Wu and others 2012) and more important in grasslands than forests (Polley and others 2008), because the former was more sensitive to climate variation (Adkinson and others 2011; Wu and others 2012). In this study, we also found that the contribution of biotic effect to IAV in NEE at the annual scale was the largest in the grassland (DK1, 77%), followed by the deciduous (MOZ, 69%) and evergreen forests (QYZ, 47%). However, whether the vegetation type is the primary factor controlling the relative contribution of biotic effects is still unclear.

Interestingly, we found a strong negative correlation between the climatic and biotic effects on both annual NEE and GPP (Figure 8J) at QYZ, compared to a weak non-significant correlation in Howland Forest (Richardson and others 2007), MOZ and DK1 (Figure 8K, L). Because the responses of ecosystem C fluxes to climatic change can be partitioned into climatic and biotic effects with the opposite directions, the combined effect (climatic + biotic effects) at QYZ may not fluctuate dramatically. Therefore, the QYZ plantation has robust ecosystem functioning in terms of C cycling, in spite of the variations of climatic variables. For example, the climatic effect caused the plantation to absorb less C from 2003 to 2009 with a decrease of 36.2 g C m−2 y−1, whereas the biotic effect offset the climatic one to a large degree (Figure 8A, J).

The relatively stable C sink may result from ecological resistance to climatic variability for several reasons. First, the ecological and biological properties of the dominant species P. massoniana and P. elliottii, with the waxy structures on the leaves and associating mycorrhizal fungi, may enhance water availability and water use efficiency and then reduce evapotranspiration (Wang and Ding 2013; Zhang and others 1999). The subtropical evergreen plantation may thus be drought resistant and has the ability to reduce the stress of water deficit. Second, although the summer drought suppressed both GPP and RE, there was usually sufficient water during the pre-summer period (data not shown), which might stimulate microorganism activity, increase nutrient availability (Brooker 2006), and thus ameliorate environmental limitations to the biotic effects.

At all three sites, the IAV in radiation and water conditions was greater than that in temperature. However, how these climatic variables directly and indirectly affect C fluxes are crucial to understand the underlying mechanisms of response of ecosystem C cycling to climatic change. At QYZ, the model parameter K m was strongly positively correlated with annual PAR and negatively with SWC (r 2 > 0.60, P < 0.05, Figure 4A, B). Because the annual mean A m was not correlated with annual PAR or SWC (both P > 0.9) at QYZ, the maximum photosynthetic rate was more difficult to reach if the K m was the larger. Thus, the positive correlation between PAR and K m indicated that the direct and indirect effects of PAR on GPP were opposite based on the light compensation curve (Falge and others 2001), whereas the direct and indirect effects of SWC were consistent due to the negative correlation between SWC and K m. On the contrary, K m was strongly negatively correlated with annual PAR and positively to PPT (r 2 > 0.80, P < 0.05, Figure 4C, D) at DK1. The annual mean A m was positively correlated with K m and PPT (r 2 = 0.80, P < 0.05) and negatively to PAR (r 2 = 0.46, P = 0.21), resulting in the opposite direct and indirect effects of PAR through K m and A m on GPP and similar effects of PPT. However, it is difficult to partition the effects of PAR and water conditions on photosynthetic capacity due to the high correlation between them at the annual scale, and the ultimate relationships between the indirect and indirect effects of climate depend on the relative importance of these climatic variables to GPP.

The previous and current studies found that the seasonality of A m and Q 10 was correlated with temperature and water conditions (Ju and others 2010; Ricciuto and others 2008; Yu and others 2008). At QYZ, the positive effect of T a on A m was usually greater than the negative effect of VPD, whereas in the severe drought summer, the negative effect of VPD dominated and then suppressed A m, especially in 2003 (Figure 5A). At MOZ and DK1, the effects of T a and VPD on A m were similar to those at QYZ, except that there was no strong summer suppression on A m (Figures S8A, S9A). At QYZ and MOZ, the temperature sensitivity of RE (Q 10) decreased with T s and increased with SWC. When SWC was kept constant at an average level of 30%, Q 10 values decreased 0.037 ± 0.029 and 0.052 ± 0.049 with every 1°C in T s at QYZ and MOZ, respectively, which was similar to the average value of 0.046 ± 0.0033 for foliage respiration from 56 species in arctic, boreal, temperate, and tropical biomes (Tjoelker and others 2001). The dependence of Q 10 on temperature was usually attributed to the acclimatization of respiration (Luo and others 2001). However, a recent study has found that Q 10 values of RE did not differ among biomes at a global scale, being independent of temperature (Mahecha and others 2010). The results from DK1 showed that the effects of 1°C increment in T s on Q 10 varied from −0.056 to 0.049 due to a complex interactions between temperature and water conditions. The different effects of temperature on Q 10 values might have resulted from the influence of the confounding factors as well as methods and scales.

Model Performance, Limitations, and Implications

In our study, the empirical model captured both the seasonality and IAV in NEE relatively well, with good agreement between the observations and modeled results both in terms of AIAV and RIAV (Figure 1; “Interannual Variability in Climate and Carbon (C) Fluxes” and “Model Parameters and Performance” sections). For example, the observed NEE was very close to the simulated annual NEE (−333 ± 47 vs. −324 ± 39 g C m−2 y−1) within the range of −197 to −430 g C m−2 y−1 in previous studies at QYZ (Yu and others 2008; Liu and others 2009; Wen and others 2010; Zhang and others 2011b). Nevertheless, systematic and random errors may challenge the reliability of estimated annual fluxes and their IAV (Mauder and others 2013). Fortunately, standardized data-processing approaches considerably reduce the systematic error to typically 5–10% and have little effect on the IAV (Baldocchi 2008). The aggregated random error on the annual scale is generally about 5% (Baldocchi 2008), and compared to the much larger IAV in NEE, this level of error may not significantly affect our results (Mauder and others 2013; Stoy and others 2008).

The plantation experienced a severe drought in 2003 at QYZ, becoming a C source in the summer, but it was still a strong C sink over the whole year comparable to normal years. This is because the ecosystem has higher C uptake before the drought due to the sufficient supply of soil water (Wen and others 2010). In early 2008, a strong ice storm caused great biomass losses (227 g C m−2), resulting in lower C uptake than that in normal years due to the large reduction in GPP (Zhang and others 2011b). Usually, such extreme climatic events increase uncertainties in simulating C fluxes (Ju and others 2010). The water deficiency and heat stress usually suppress both GPP and RE, and in most cases the response of GPP is more dramatic than that of RE (Schwalm and others 2010). At QYZ, the severe drought in 2003 and 2007 considerably decreased C fluxes in summer, which was difficult to simulate in the empirical and/or mechanistic models (Ju and others 2010). The extreme drought in the summer of 2003 thus resulted in a larger disagreement between modeled and observed daily NEE (r 2 = 0.54, P < 0.001) compared to the normal condition (r 2 = 0.70, P < 0.001, Figure 2A). Such an effect of drought on model performance also occurred for NEE in 2006 and 2007 at MOZ (Figure 2B). At QYZ, the severe ice storm in early 2008 caused both physical damage and physiological injury (Ma and others 2010; Zhang and others 2011b). The physical damage resulted in the loss of 200 and 2,206 g C m−2 in branch breakage area and severely damaged area, respectively (Ma and others 2010). Although the empirical model was able to capture the IAV by changing the parameters, it seemed not to be as flexible as the gap-filling model to capture the sudden large loss of photosynthetic tissues, resulting in overestimated annual GPP in 2008 and 2009 and overestimated annual RE in 2009.

The difference between modeled and observed NEE at QYZ mainly resulted from the difficulties in simulating RE, which might be the result of low correlations between nighttime F c and climatic variables (r 2 = 0.32, P < 0.001 for half-hour scale) and the extensive gaps due to unrealistic conditions. The poor model performance also resulted in a large contribution of the biotic effect on RE (Figure 7I–L). At MOZ, similar model errors were also found for RE in 2005 and 2006 when the precipitation was low and the temperature was high (Figure S7B). Although some studies suggested that those models with GPP as biotic drivers could better simulate the dynamics of RE (Kuzyakov and Gavarichkova 2010; Migliavacca and others 2011), the problem of simulating RE was still pervasive (Mitchell and others 2011). Time lag between GPP and RE (Kuzyakov and Gavarichkova 2010; Vargas and others 2011) and the lack of other biotic data (for example, soil C pool) might have influenced model performance. Moreover, we found that the effect of SWC on Q 10 decreased with temperature at QYZ, which may not be suitable for severe drought situations, in which high temperature accelerates the negative effect of water stress. This causes additional difficulties in simulating IAV in C fluxes.

Understanding the underlying mechanisms of IAV in C fluxes is important for developing ecological theories and projecting future ecosystem changes (Hui and others 2003). Our study highlights the necessity of partitioning IAV in C fluxes into climatic and biotic effects, because the current process-based models generally failed to reproduce the interannual dynamics of C fluxes (Keenan and others 2012). Process-based mechanistic models would be more robust for predicting IAV in C fluxes if the biotic effects are incorporated into the model at annual and longer time scales. Moreover, it is difficult for both statistical and mechanistic models to simulate the sudden and dramatic effects of extreme events on C fluxes, which need to receive more attention in climate-impact research (Reichstein and others 2013).