The full annual carbon balance of a subtropical coniferous plantation is highly sensitive to autumn precipitation

Deep understanding of the effects of precipitation on carbon budgets is essential to assess the carbon balance accurately and can help predict potential variation within the global change context. Therefore, we addressed this issue by analyzing twelve years (2003–2014) of observations of carbon fluxes and their corresponding temperature and precipitation data in a subtropical coniferous plantation at the Qianyanzhou (QYZ) site, southern China. During the observation years, this coniferous ecosystem experienced four cold springs whose effects on the carbon budgets were relatively clear based on previous studies. To unravel the effects of temperature and precipitation, the effects of autumn precipitation were examined by grouping the data into two pools based on whether the years experienced cold springs. The results indicated that precipitation in autumn can accelerate the gross primary productivity (GPP) of the following year. Meanwhile, divergent effects of precipitation on ecosystem respiration (Re) were found. Autumn precipitation was found to enhance Re in normal years but the same regulation was not found in the cold-spring years. These results suggested that for long-term predictions of carbon balance in global climate change projections, the effects of precipitation must be considered to better constrain the uncertainties associated with the estimation.


Results
Seasonal and inter-annual variation of environmental factors. DR (direct solar radiation), Ta (air temperature), and VPD (vapor pressure deficit) values had a single peak and reached their maximums in July (Fig. 1a,b and d). The PP (precipitation) showed strong seasonal and inter-annual variation and was usually more in the first half year than that in the second half of the year (Fig. 1c). The trends in VPD and SWC (soil water contents) generally followed the seasonal course of Ta and precipitation ( Fig. 1d and e).
The annual mean Ta from 2003 to 2014 ranged from 16.7 °C to 18.7 °C, with the minimum in 2012 and maximum in 2003, whereas the precipitation ranged from 944.9 mm in 2003 to 1918.3 mm in 2012 (Fig. 2a). In general, the precipitation decreased as the temperature increased at the annual scale. The Ta was lower in 2005, 2008, 2011 and 2012 than that in other years from January to March because cold early springs appeared in those years (Figs 1b and 2b). In normal years, the monthly mean temperatures were always above the active temperature threshold of 5 °C throughout the whole year. However, in the cold years of 2005, 2008, 2011 and 2012, the monthly average temperatures in January were 3.4 °C, 4.2 °C, 1.7 °C and 3.7 °C, respectively. In addition, because of the variation in the monsoon retreat time, the inter-annual variation in precipitation in autumn (from September to November) was large (Figs 1c and 2b). The coefficients of variation (CV) were 56.6%, 105.4% and 75.5% in September, October and November, respectively. The precipitation in the autumn of 2003,2004,2007,2013 and 2014 was respectively 56.4%, 46.3%, 40.6%, 6.6% and 48.3% lower than multi-year average. Seasonal and inter-annual variation of carbon fluxes. The GPP and Re exhibited clear seasonal variation, with peaks in July in most observed years, whereas the seasonal variation in NEP was more complex because of its dependence on both GPP and Re (Fig. 3). The multi-year monthly average of NEP also peaked in July with a value of 55.6 ± 16.4 gC m −2 month −1 .
Over the twelve years, the average annual GPP was 1724.0 ± 63.8 gC m −2 yr −1 (CV = 3.7%), with a maximum of 1835.5 gC m −2 yr −1 in 2006 and a minimum of 1621.5 gC m −2 yr −1 in 2005 (Fig. 4). The annual Re ranged from 1184.6 gC m −2 yr −1 in 2004 to 1457.5 gC m −2 yr −1 in 2006, with an average of 1262.2 ± 78.2 gC m −2 yr −1 (CV = 6.2%). As the difference between GPP and Re, the multi-year average NEP was 461.8 ± 62.2 gC m −2 yr −1 . The CV of NEP was 13.5%, which was a bit larger than that of GPP and Re.
The mean annual GPP in the years with cold early springs was significantly lower than that of the other years (P < 0.01). The average annual Re and NEP in the years with cold early springs were also lower than those of the other years, but the differences were statistical insignificant (P = 0.33 for Re, P = 0.23 for NEP) (Fig. 4). The anomalies in GPP in the years with cold springs were negative when the temperature anomalies were negative (Fig. 3b), but the same regulation was not found in the variation in Re.
To exclude the effects of cold springs, the annual carbon fluxes were separated into two groups (Fig. 4). In the normal years (2006, 2013, 2010, 2007 and 2009), the GPP was higher than the average. As shown in Fig. 2b, the autumn precipitation in the previous years, 2005,2012,2009,2006 and 2008, respectively, was also above average. Among the years with cold springs, the GPP was higher only in 2012, and the precipitation was higher in 2011.
To compare the effects of autumn precipitation, we divided the observations into two groups, as shown in Fig. 4. In combination with Figs 2 and 3, the variation in the regulation of Re was found to follow the inter-annual variation in precipitation in normal years, with the opposite pattern in cold-spring years.    The response of GPP to autumn precipitation in the previous year. During the observed years, the annual GPP was found to be positively related to the autumn precipitation of the previous year (Fig. 5). When the data were placed into one group, the linear-fit relation seemed to be comparatively weak (Fig. 5a). The observed values were somewhat scattered around the fit line. In addition, the residual sum of squares was large.
When the data were classified by early spring cold, the autumn precipitation more clearly explained the annual GPP: 63% and 69% of GPP can be explained by the precipitation in the previous autumn in the years without and with cold early springs, respectively (Fig. 5b). The response slope of GPP to the precipitation in the previous autumn in the cold-spring years was greater than that of the ordinary years, which indicated a more sensitive response.
The effects of autumn precipitation on Re. Autumn precipitation had divergent effects on annual Re.
When all of the annual Re values from 2003 to 2014 were studied, no clear relation was observed between Re and autumn precipitation (Fig. 6a). When the data were put into two pools, the effects of autumn precipitation emerged.
In the years without cold early springs, Re increased with increasing autumn precipitation. The liner relation was significant, and the R 2 was 0.66. In the years with cold springs, Re showed decrease trend as autumn precipitation increased. The relation was not significant because of the limited samples, but autumn precipitation explained 43% of the variation in Re.

Discussion
Effects of temperature and precipitation on carbon fluxes. Over a long time, an ecosystem carbon budget will reach a dynamic equilibrium through the process of photosynthesis and respiration. As the variation in NEP depends on GPP and Re 42 , the controlling mechanisms of both GPP and Re should be examined to gain a comprehensive understanding of carbon budgets.
In addition, long-term observation is necessary to investigate the effects of climatic factors to avoid sampling in the trough or crest of a multi-annual pattern 28 . Therefore, we used 12-year observations to examine the mechanisms through which Ta and precipitation control GPP and Re in our study. During the observation years, the dataset represented much of the typical inter-annual variability at this site (Figs 3 and 4), including both typical and extreme climatic events; for example, the cold-spring years 8 , the typical El Niño event, which resulted in an extreme summer drought in 2003 17 , and a strong La Niña event that happened in 2009-2010 41 .
Previous studies have confirmed that variations in GPP and Re were driven by environmental forcing, especially temperature and precipitation 43, 44 . In the climate change context, warming temperatures and an altered precipitation regime would largely impact carbon budgets 45,46 . Hence, many studies have focused on the effects of Ta and Precipitation. However, to date, no consensus has been reached regarding whether Ta or precipitation was the dominant factor 8,47 . Especially at large temporal and spatial scales, the controlling mechanisms of Ta and precipitation remain controversial. Therefore, the existing models, which were constructed based on the driving effects of climatic factors, were inadequate to assess the inter-annual variability of carbon budgets 48 . Fortunately, when studies focused on a single site or a region, some instructive results were obtained 49 . These studies were conducted based on annual carbon budgets and their corresponding Ta and precipitation and reached general conclusions that temperate forests were temperature limited 12 . However, for subtropical and tropical forests, precipitation was more important than temperature 50 . At our study site, previous studies had determined that GPP, Re and NEP was depressed by summer-autumn drought 17 , whereas Zhang et al. indicated that the low temperature in spring played a more important role in regulating the inter-annual variability of annual net carbon uptake at this subtropical site 8 .
These results revealed that the full annual precipitation or annual mean Ta played limited roles in interpreting the variability in carbon budgets and their controlling mechanisms. Based on previous studies, the Ta and precipitation may be more crucial in transitional periods in the growing season, such as spring and autumn 6,51,52 . In other words, carbon budgets would be sensitive to Ta and/or precipitation within a given period 30 . In this study area, Ta in spring, especially low temperatures, could regulate phenology and depress carbon sequestration 8,41 . Another key factor might be autumn precipitation because the carbon budget is easily affected by seasonal droughts 17,31 . In addition, the autumn precipitation in this region is clearly affected by El Niño-Southern Oscillation (ENSO) 53 , with relatively great variability. Therefore, spring temperatures and autumn precipitation should be considered key factors in our study (Fig. 2b).
The effects of Ta and precipitation are always entangled. Therefore, to isolate the effects of autumn precipitation, we grouped the observed data into two pools according to whether the years experienced cold springs (Figs 4,5 and 6). The GPP and Re in cold-spring years were less than in normal years, which was consistent with the results of Zhang et al. 8 . Based on this result, we examined the contributions of autumn precipitation to GPP and Re in the two groups. The results underscored the prominent role that precipitation plays in the inter-annual variability of GPP and Re (Figs 4 and 5).
Previous studies have indicated that seasonal droughts occurred in this region in 2003 and 2007 and have observed the different responses of GPP and Re to water condition. At short-time scales, increased precipitation would be accompanied by a decrease in temperature and radiation, which would depress the photosynthesis of trees and thereby reduce GPP. However, at annual or longer time scales, the temperature and radiation are both generally high in this subtropical region. By contrast, the asynchronism of precipitation and temperature causes water to be a limiting factor, thereby reduced GPP and Re to different extents 17 . Therefore, a comprehensive understanding of the controlling mechanisms of precipitation was needed.
Recently, time-lag effects of precipitation have been reported in many studies [54][55][56] . Our study also found that GPP increased as autumn precipitation increased in the previous year for either normal or cold-spring years, and thus a time-lag effect of precipitation on GPP emerged. In contrast, for Re, previous studies mainly focused on the accelerating effects of precipitation on soil respiration and devoted less attention to the trade-off between depressed plant respiration and increased soil respiration. According to our results, precipitation showed varying effects on Re in normal and cold-spring years. Below, we discuss the lag effects of precipitation on GPP and the effects of precipitation on Re.
The lagged effects of precipitation on variation in GPP. Any individual climatic factor is far from adequate to explain carbon flux variability, especially precipitation, which has different controlling effects on GPP and Re. Our analysis shows that the precipitation in autumn will affect the Re of the given year and the GPP of the following year. Additionally, we used a statistical method to reveal and quantify the time lag effects of precipitation on ecosystem GPP and Re (Fig. 7). The statistical results strongly supported our analysis. Precipitation exhibited a significant 3-5 month lag effect on GPP (R = 0.51, P < 0.01) and a weak lag effect on Re in this subtropical coniferous plantation (Fig. 7).
Some studies have previously captured the lag effects of precipitation 57 , which are highly variable across different ecosystems and regions. Similar to our observation that annual GPP was sensitive to the precipitation of the previous autumn, Vasconcelos et al. 56 found that aboveground net primary productivity in tropical forest regrowth increased following wetter dry-seasons. In contrast, Zhang et al. found weak lag effects of precipitation on GPP and Re in grassland ecosystems 31 . These different responses of ecosystems may be due to the different water use strategies of grasses, crops and trees. Yang et al. have reported that crops only use shallow soil water 58 , whereas trees can use deep soil water, especially in the dry season 59 .
The time-lag effects of precipitation in forests can also be ascribed to the altered water use strategies of plants under soil water stress 18 or to the long turnover time of deep soil moisture 35 . Autumn precipitation is important to the recharge of soil water. Inadequate autumn precipitation can occasionally result in severe drought, in which decreases in GPP, Re and energy reserves would constrain the bud preformation in succeeding years 60,61 . Similar results have been reported by Breda et al. 62 , where reduced net primary productivity was caused by reduced storage of carbohydrate, lipid and protein reserves during the previous drought year. Additionally, the altered water use strategy will increase deep soil water consumption, thereby leaving less water available for twig shooting in the following spring 41,59 .

Divergent responses of Re to autumn precipitation in normal and cold-spring years. Ecosystem
respiration is a complex process that can be separated into two major components: plant respiration (root respiration was excluded) and soil respiration (Rs). Plant respiration includes leaf respiration, stem respiration, and respiration of the understory. This component is closely related to the GPP of the ecosystem 63, 64 because they are coupled processes that are joined at the stomata. Therefore, the ratio between plant respiration and GPP would be a relatively constant value. The other important component of Re is soil respiration, which has been widely studied [65][66][67] . Soil respiration is generally considered to be controlled by temperature and sometimes by soil moisture and was thus deemed to be sensitive to precipitation 68,69 . At our study site, soil respiration was also positively related to soil moisture and precipitation 70 .
To conduct a comprehensive study of how precipitation affects Re, we should understand how precipitation affects these two components. Plant respiration was found to be positively related to air temperature, with obvious seasonal variation 71 . Precipitation events were always accompanied by relatively lower temperature and radiation, likely decreasing plant respiration. In contrast, soil respiration would be promoted by precipitation, as discussed by Wang et al. 70 . Therefore, to predict how precipitation would affect Re, we should consider the ratio between Rs and Re. At our study site, the Rs/Re ratio was negatively related to the enhanced vegetation index (EVI) (Fig. 8), with slopes of −0.95 in normal years (Fig. 8a) and −1.42 in cold-spring years (Fig. 8b). Based on these statistical results, Rs dominated Re when the EVI was less than 0.48 in the normal years, whereas in the cold-spring years, the Rs constituted a larger proportion when the EVI was less than 0.41.
In addition, the EVI showed great differences in normal and cold-spring years (Fig. 9). Firstly, the EVI during January to April was much lower in clod-spring years than that in normal years. Secondly, the cold-spring postponed the peak time of EVI, which showed up in August and September instead of July in normal years. These phenomena indicated the great differences of ecosystem functions between normal and cold-spring years, which agreed well with some previous researches 12, 41 . In normal years, the EVI was lower than 0.48 in autumn (Fig. 9). Combing the results of Fig. 8 as mentioned above, the Rs was generally found to account for a larger proportion of Re in normal years. As the Rs is positively related to precipitation 70 , the increases in autumn precipitation would increase the total annual Re. In contrast, Re in cold-spring years showed a decrease trend with increasing autumn precipitation, possibly due to the decreasing effects of precipitation on plant respiration. In the cold-spring years, the plant respiration constituted over 50% of the Re in the autumn according to Figs 8 and 9. Therefore, autumn precipitation might have an opposite effect on Re in the years with cold springs.
However, the ecosystem responses to the precipitation are complicated and difficult to quantify. To demonstrate the underlying mechanisms, this study explored the lag effects of precipitation on GPP and the divergent effects on Re in cold-spring and normal years. Indeed, predicted climate change may include more frequent and severe dry seasons in response to global warming 72 and more frequent El Niño episodes 73 . The results reported here indicate that the effects of precipitation must be considered to better constrain the uncertainties associated with estimation under that projected scenario.

Materials and Methods
Site description. This study was conducted in a subtropical coniferous plantation (26°44′29″N, 115°03′29″E, at 102 m elevation above sea level) at Qianyanzhou Ecological Research Station (QYZ), located in a typical red

Flux correction and gap filling.
This study adopted the methods of calculating and correcting carbon dioxide fluxes in Wen et al. 17,31 . The CO 2 fluxes were calculated every 30 minutes from the 10 Hz raw data. Processing of the flux data was performed using routine methods, including three-dimensional rotation 74 , the Webb, Pearman and Leuning correction for the effects of air density fluctuations (WPL correction) 75 , storage calculations and spurious data removal 8,31 . Spurious data caused by rainfall, water condensation or system failure were removed from the dataset. To avoid the possible underestimation of the fluxes under stable conditions at night, nighttime data (solar elevation angle < 0) were excluded when the friction velocity (u * ) was less than the relevant thresholds, which were identified based on the researches of Reichstein et al. 76 . The threshold values of u * ranged from 0.16 to 0.22 m s −1 , with an average value of 0.20 m s −1 for the years from 2003 to 2014. Any data gaps in meteorological variables were filled using the mean diurnal variation method 77 . The linear fitting, nonlinear fitting and mean diurnal variations were used to fill missing data points and to replace spurious data 43 . Further details of data processing are presented in ChinaFLUX 17, 43, 74 . Data analysis. The carbon fluxes and their corresponding main climatic factors, including direct solar radiation (DR), air temperature (Ta), precipitation (PP), vapor pressure deficit (VPD) and soil water content (SWC), were integrated at the month and annual scales to represent seasonal and inter-annual variation. Their anomalies were also calculated to indicate their deviations from normal levels. In previous studies, this subtropical coniferous plantation was temperature sensitive in early spring 8 . Therefore, the averaged early spring (January to March) temperatures were calculated. In addition, for the purpose of our study, the cumulative autumn precipitation values (September to November) were calculated.
The linear relationships between annual carbon flux and precipitation were calculated to determine the sensitivity of carbon flux. In addition, the linear relation between annual carbon flux and autumn precipitation in the same year and in the previous year were estimated to examine the sensitivity of this ecosystem to autumn precipitation.
Lag effects of precipitation on GPP/Re. To identify the relationship between autumn precipitation and GPP/Re, we analyzed their correlation at the annual scale, placing emphasis on the time lag effects between the autumn precipitation and GPP/Re.
For this purpose, annual cumulative precipitation and GPP/Re were calculated. Multiple "yearly" statistics (approximately 130 values from the twelve years) were obtained for 12-month intervals, shifting them one month at a time 33,41 . To investigate the lag of the GPP/Re response to the climatic factors, we shifted the climatic series backward one month at a time (up to twelve months) to calculate the correlations between climate drivers and GPP/Re. Student's t-tests were applied to verify the statistical significance of the correlation coefficients 28,33 .