Drivers of long-term variability in CO 2 net ecosystem exchange in a temperate peatland

. Land–atmosphere exchange of carbon dioxide (CO 2 ) in peatlands exhibits marked seasonal and interannual variability, which subsequently affects the carbon (C) sink strength of catchments across multiple temporal scales. Long-term studies are needed to fully capture the natural variability and therefore identify the key hydrometeorological drivers in the net ecosystem exchange (NEE) of CO 2 . Since 2002, NEE has been measured continuously by eddy-covariance at Auchencorth Moss, a temperate lowland peatland in central Scotland. Hence this is one of the longest peatland NEE studies to date. For 11 years, the site was a consistent, yet variable, atmospheric CO 2 sink ranging from − 5.2 to − 135.9 g CO 2 -C m − 2 yr − 1 (mean of − 64.1 ± 33.6 g CO 2 C m − 2 yr − 1 ). Inter-annual variability in NEE was positively correlated to the length of the growing season. Mean winter air temperature explained 87 % of the inter-annual variability in the sink strength of


Introduction
Northern peatlands are one of the most important global sinks of atmospheric CO 2 ; with their ability to sequester C controlled by hydrometeorological variables such as precipitation, temperature, length of growing season and period of snow cover, they also potentially represent an important climatic feedback mechanism (Aurela et al., 2001;Frolking et al., 2001;Lafleur et al., 2003). Peatland carbon models generally suggest a decline in net sink strength in a warming climate, although the magnitude of the decline predicted by individual models is variable . UK peatlands are predicted to become a net source of carbon in response to climate change (Worrall et al., 2007), with climate models predicting a rise in global temperature of ca. 3 • C between 1980-1999and 2100(IPCC, 2007; scenario A1B which considers a balanced distribution between fossil fuel intensive and non-fossil fuel energy sources). A greater understanding of drivers and feedback mechanisms, across a range of temporal scales, is therefore a current research priority.

C. Helfter et al.: Drivers of long-term variability in CO 2 net ecosystem exchange
Eddy covariance measurements using fixed flux towers provide the best method for assessing inter-annual changes in land-atmosphere exchange of CO 2 at the catchment scale Nilsson et al., 2008;Roulet et al., 2007). In most years and in most peatlands, net ecosystem exchange (NEE) is the largest and most variable of the C flux terms (Roulet et al., 2007). In combination with aquatic fluxes (downstream and evasive losses) and CH 4 emissions, it is a key component of C and greenhouse gas (GHG) budgets for peatland systems Dinsmore et al., 2010). Although more sites are now being established globally, there are still relatively few peatland sites (< 10) with published NEE measurements for periods of 3 years or more.
Including the Auchencorth Moss site, there are to our knowledge only six peatland sites in the Northern Hemisphere for which long-term (≥ 3 years) data sets of NEE are now available and all show that peatlands operate as a sink for atmospheric CO 2 , albeit with different annual sink strengths. The 6-year mean NEE for Mer Bleue peatland (Ontario, Canada) was −40.2 g C m −2 yr −1 (negative values signify uptake), varying year-to-year from a minor (−2) to a major (−112) CO 2 sink (Roulet et al., 2007). Similarly, McVeigh et al. (2014) found that a blanket bog in southwest Ireland had a mean 9-year NEE of −55.7 g C m −2 yr −1 and exhibited significant inter-annual variability (−32.1 to −79.2 g C m −2 yr −1 ). Degerö Stormyr in northern Sweden showed consistent yet variable CO 2 uptake over 12 consecutive years (12-year mean −58 ± 21 g C m −2 yr −1 , range −18 to −105 g C m −2 yr −1 ) (Peichl et al., 2014). Eddy covariance measurements at Lompolojänkkä, a nutrient-rich fen in northern Finland, again showed that the site operated as a weak (−3 g C m −2 yr −1 ) to strong (−59 g C m −2 yr −1 ) CO 2 sink over a 3-year period (Aurela et al., 2009). In contrast to the variability exhibited by these sites, a sub-arctic permafrost mire in Northern Sweden was relatively stable over the period 2001-2008 (−46 g C m −2 yr −1 ) .
Quantifying inter-annual variability in NEE is a prerequisite for detecting longer term trends or step changes in flux magnitude in response to climatic or anthropogenic influences and identifying its drivers. In the UK, there have been significant regional changes in precipitation and temperature since the beginning on the 20th century, with the most rapid changes occurring over the last 50 years (Jenkins et al., 2009). During the period 1961-2006 annual precipitation increased by 2.5-23.2 %, with the largest increases occurring in the winter (particularly in Scotland and northern England); summer months were typically characterised by a decrease in precipitation. Mean annual temperature during the same period increased in parts of the UK by 1.05-1.64 • C (Jenkins et al., 2009), with winter months (January-February) warming much faster than the other months of the year in some parts (Holden and Rose, 2011). These data show that significant changes are taking place in seasonal climatic patterns, which are likely to have a major impact on annual net CO 2 uptake by peatland systems.
Meteorological conditions such as rainfall, temperature and levels of photosynthetic active radiation (PAR) control NEE and its components, total ecosystem respiration (R eco ) and gross primary productivity (GPP). R eco is composed of a plant respiration term (autotrophic respiration, R A ), which quantifies metabolic respiration from both above-and belowground biomass, and a heterotrophic respiration term (R H ) resulting from microbial decomposition of organic matter in the soil. Autotrophic respiration can account for up to 60 % or R eco whilst total belowground respiration can account for up to 70 % (van der Molen et al., 2011). R eco and GPP have been shown to be tightly linked in a range of ecosystems on both short-term and annual timescales (Irvine et al., 2008;Ryan and Law, 2005) and respond similarly, although not necessarily with the same magnitude, to extreme events such as drought. For example, short-term dynamics of R eco can be more sensitive to the availability of labile C compounds produced by photosynthesis than to the effects of varying soil moisture on soil microbial activity (Irvine et al., 2008). On a global scale, drought is the main cause of decreased GPP alongside continent-specific secondary drivers such as cold spells and precipitation (Zscheischler et al., 2014a, b). Although less well understood and modelled than GPP, R eco plays a major role in ecosystem C exchange dynamics, and increases in R eco have been shown to turn a sink of C into a source (Lund et al., 2012). In order to interpret inter-annual variability in NEE, it is therefore crucial to partition NEE into GPP and R eco and study their dynamics with respect to meteorology. We have done this on Auchencorth Moss, an ombrotrophic peatland in south-east Scotland.
The first eddy covariance measurements of CO 2 exchange at Auchencorth Moss took place in 1995-1996(Hargreaves et al., 2003, with continuous measurements starting in 2002. Previous measurements of NEE have been published for specific 2-3 year time periods and suggest that inter-annual variability is high. Dinsmore et al. (2010) and  reported that over a 3-year period (2006)(2007)(2008) the peatland acted as a very strong CO 2 sink (−88 to −136 g C m −2 yr −1 ), whereas Billett et al. (2004) reported that between 1995-1996 it was acting as a weaker CO 2 sink (−36 and −8 g C m −2 yr −1 ). In comparison to NEE, CH 4 emissions at Auchencorth Moss are small (average of 0.32 g CH 4 -C m −2 yr −1 in 2007 and 2008; Dinsmore et al., 2010). Although these individual studies highlight significant inter-annual variability at Auchencorth Moss, they are for relatively short periods of time and are insufficient to investigate the drivers of inter-annual variability in NEE. Here we present the first complete analysis of the 2002-2013 data set in terms of monthly, seasonal and annual fluxes and explore the drivers of temporal variability in NEE. We use our data to test the following hypotheses: -Ecosystem respiration is related to water table depth and the peatland releases more CO 2 to the atmosphere during dry spells.
-Annual NEE is positively correlated with the length of the growing season. where the soils comprise peats (85 %), Gleysols (9 %), Humic Gleysols (3 %) and Cambisols (3 %). The open moorland site has an extensive uniform fetch over blanket bog to the south, west and north with a dominant wind direction from the south-west; winds from the north-east are the second most important wind direction. The terrain is relatively flat with a complex micro-topography consisting of hummocks and hollows. Hummocks are relatively small in size (typically 40 cm in diameter and ∼ 30 cm in height) and covered by either a mix of Deschampsia flexuosa and Eriophorum vaginatum, or Juncus effusus. In contrast, hollows are dominated by mosses (Sphagnum papillosum and Polytrichum commune) and a layer of grasses (Dinsmore et al., 2009). The site was drained more than 100 years ago (Leith et al., 2014); the drains have become progressively less effective and re-vegetated over time, leading to slow and progressive rewetting of the site. Over the last 20 years the site has been used for seasonal low intensity sheep grazing; areas of peat extraction occur at the margins of the catchment outside the footprint of the flux tower measurements.

Instrumentation and data processing
Fluxes of carbon dioxide (CO 2 ) have been measured continuously by eddy covariance (EC) at Auchencorth Moss since May 2002. The principles of operation and flux calculation methods using the eddy-covariance technique have been extensively discussed elsewhere (Aubinet et al., 2000;Baldocchi et al., 2001). The EC system at Auchencorth Moss consists of a LI-COR 7000 closed-path infrared gas analyser operating at 10 Hz for the simultaneous measurement of carbon dioxide and water vapour. Turbulence measurements were made with an ultrasonic anemometer (initially model Solent R1012A R2 operating at 20.8 Hz; from 2009 Gill Windmaster Pro operating at 20 Hz; both Gill Instruments, Lymington, UK), mounted atop a 3 m mast. The effective measurement height is 3.5 m with a vertical separation of 20 cm between the anemometer and the inlet of the sampling line. Air is sampled at 20 L min −1 through a 40 m long Dekabon line (internal diameter 4 mm). In addition to eddy-covariance measurements, the site is equipped with a Campbell Scientific 23X data logger for the automated acquisition of a comprehensive suite of meteorological parameters which include net radiation (Skye instruments SKS1110), PAR (Skye instruments SKP215), air temperature (fine wire type-E thermocouple), air pressure (Vaisala PTB101C), wind speed and direction (Gill Instruments WindSonic), soil water content (Campbell Scientific CS616 TDR probes), soil temperature (Campbell Scientific 107 thermistors at 10, 20, 30, and 40 cm), rainfall (tipping bucket rain gauge) and, since April 2007, water table depth (Druck PDCR 1830).
High-frequency eddy-covariance data is acquired by inhouse software written in LabView (National Instruments) and processed offline into half-hourly fluxes.
Half-hourly data points were excluded from further analysis if any of the criteria listed below were not met: -The total number of "raw" (high-frequency) data points per notional half-hour period was less than 90 % of the maximum possible number of points (36 000), i.e. below a minimum averaging period of 27 min.
-The number of spikes in raw w (vertical wind velocity component), CO 2 (CO 2 mole fraction) and H 2 O mole fraction exceeded 1 % of the total number of points per half-hour period.
-The stationarity test devised by Foken and Wichura (1996), which compares half-hourly fluxes to the average of six 5 min averaging periods within the half hour, did not fulfil the quality criterion.
After quality control, the number of good data points ranged from 45 % (in 2005) to 78 % (in both 2004 and 2008), with an annual mean of 65 ± 11 %. Due to technical difficulties with the sampling pump (gradual decline in pumping performance), which were not detected immediately, most of the flux data for the summer period of 2011 were discarded as a precautionary measure.

Calculations of ecosystem respiration, Q 10 and GPP
Gap filling of net ecosystem exchange (NEE) measured by eddy-covariance and partitioning of the gap-filled halfhourly fluxes into ecosystem respiration (R eco ) and gross primary production (GPP) was achieved using an online tool developed at the Max Planck Institute for Biogeochemistry, Jena, Germany 1 . In this flux partitioning approach, daytime R eco is obtained by extrapolation of the night time parameterisation of NEE on air temperature (using an exponential relationship of the form given in Eq. 1).
where T is air temperature and a and b are fitting coefficients. GPP was then calculated as the difference between R eco and measured NEE. The growth rate (Q 10 ) for ecosystem respiration for a change of 10 • C was determined using the relationship: (2) T 1 and T 2 are reference temperatures (5 and 15 • C, respectively), and R 1 and R 2 are the corresponding respiration rates. R 1 and R 2 for each calendar year of the study were calculated from Eq. (1) using 24 h averages of measured night time T air and NEE (see Supplement for non-linear regression statistics).
GPP was parameterised with respect to PAR using the following rectangular hyperbolic regression function: where GPP sat (GPP at light saturation) and α (quantum efficiency) are fitting parameters.

Statistical tests
Statistical dependence between ecosystem response and hydro-meteorological variables was tested using Spearman's rank correlation. This allows testing for monotony between pairs of variables without making assumptions as to the nature of the function linking them. The independent variables winter air temperature, length of growing season (LGS) and annual water table depth (WTD), were tested for rank correlation against the dependent variables summertime NEE, R eco , GPP, α and GPP sat , annual NEE and annual GPP sat . The Spearman's correlation coefficient (ρ) is calculated using Eq. (4): In Eq. (4), d i is the difference between ranked variables and n the sample size.
Potential dependence between daily growing season (March-September) water table depth and ecosystem response (R eco , GPP and NEE) was further investigated using one-way analysis of variance (ANOVA). The assumptions made were that (a) the 10 WTD classes (> 0 cm to < −45 cm in increments of 5 cm) constitute different treatments and (b) that the plant community has reached a steady state in terms of growth. The null hypothesis tested using this ANOVA is that WTD has no influence on ecosystem response.

Site meteorology
During the study period (2002-2013) the site received a mean annual precipitation of 1018 ± 166 mm (± values denote standard deviation). Autumn (September-November) was the wettest season with 96 ± 11 mm of rain per month, and spring (March-May) was the driest with 64 ± 17 mm per month. Rainfall is highly variable year on year but records from a weather station of the UK Met Office (UK Meteorological Office, 2013) located 3.5 km north of the study site indicated a slight upward trend since the early 1970s (average annual precipitation 899 ± 166 mm for the period 1961-2001).
Mean annual air temperatures were 8.3 ± 4.6 • C for the study period 2002-2013 compared to 7.7 ± 4.5 • C for . Despite year-on-year variability there are indications of a steady increase of the order of 0.019 • C yr −1 since records began in 1961 at the nearby Met Office station, which is consistent with UK and global trends (Jenkins et al., 2009). All seasons were warmer in 2002-2013 than in 1961-2001, albeit not significantly. Summer (June-August) was the warmest season with an average temperature of 13.6 ± 1.1 • C, and winter (December-February) the coldest with 3.7 ± 1.0 • C (Fig. 1).
Over the period April 2007-December 2013, water table depth (WTD) was within 4 cm of the peat surface for 50 out of 81 months (62 %). During dry periods, however, the water table could fall quickly to depths < −35 cm (Table 2).

Seasonal and inter-annual variability of R eco , GPP and NEE
Ecosystem respiration typically peaked in July/August and was distributed asymmetrically around its peak value (Fig. 2), following the annual cycle of temperature. Plotting monthly GPP as a function of photosynthetically active radiation (PAR) reveals two separate plant productivity regimes culminating around mid-summer (Fig. 3). The hysteresis of GPP vs. PAR is characterised by an exponential growth phase from March to June/July followed by a logarithmic decline in photosynthetic efficiency. The ratio of GPP to R eco showed  that on average carbon uptake by vegetative growth exceeded losses to the atmosphere through respiration for six months of the year, from April to September (Fig. 2, inset). A negative correlation was established between mean annual values of GPP sat (GPP at light saturation, Eq. 3) and WT (Spearman ρ = −0.63, p < 0.05, Table 1) indicating that the photosynthetic capacity of the plant community tended to decrease as WT deepened. Furthermore, GPP sat was positively correlated to the average temperature during the preceding winter (ρ = 0.73, p < 0.01, Table 1).
Both GPP and R eco exhibited significant inter-annual variability with peak summer time values ranging from 96 to 245 g CO 2 -C m −2 month −1 for GPP and 76 to 198 g CO 2 -C m −2 month −1 for R eco (August 2010 and July 2006, for minima and maxima, respectively). The site was consistently a sink for CO 2 , however inter-annual variability was large. NEE (mean −64.1 ± 33.6 g CO 2 -C m −2 ) ranged from −5.2 to −135.9 g CO 2 -C m −2 yr −1 with minimum and maximum CO 2 uptake in 2013 and 2007, respectively (Fig. 4). As observed at other sites , annual values of NEE were well correlated to the length of the growing seasons (LGS from here onward; R 2 = 0.64; Fig. 5). Furthermore, whilst mean spring/summer (April-September) NEE (NEE SS ) at Auchencorth Moss was not significantly corre-  lated to summer temperature, a strong negative correlation (i.e. net uptake increased with increasing winter T air ) was observed between mean NEE SS and the mean air temperature of the preceding winter (December-March) (R 2 = 0.87, Fig. 6; p < 0.01). Comparable correlations to winter T air were observed for GPP SS and R ecoSS (R 2 = 0.43, p < 0.01; R 2 = 0.28, p = 0.02, respectively, Fig. 6).

Effects of dry periods on CO 2 exchange
Throughout most years and most seasons Auchencorth Moss can be considered a wet site, with mean water table depth (WTD) −3.5 ± 6.8 cm and monthly range +3.8 cm (flooded; positive values denote water table levels above the peat surface) to −36 cm (WTD measurements commenced in  April 2007). The site was generally waterlogged during the autumn and winter months. During dry spells, which we arbitrarily define as any period lasting 1 week or longer with WTD < −5 cm, the water table can drop quickly at rates up to 3 cm day −1 (Table 2).
Three notable dry spells occurred during the summer of 2010 and two during the summer of 2008, characterised by cycles of rapid fall and rise of the water table. Meanwhile, air temperatures exhibited little variation. Details of the drainage rates (water table drawdown) and minimum water table depths are given in Table 2. Under normal hydrological conditions (water table typically within 3-5 cm of the peat surface), R eco at Auchencorth Moss did not exhibit a significant correlation with WTD. In contrast, during the dry spells of 2008 and 2010, daily R eco was non-linearly correlated to WTD (Fig. 7). The response of R eco to changes in WTD occurred with time lags ranging from 0 to 5 days (Table 2). The time lag is the number of days elapsed between the start of the dry period and the onset of a response from the ecosystem respiration (R eco ); the time lag was determined by optimising the polynomial fit between R eco and WTD. The minimum value of R eco for each dry spell and the water table depth corresponding to each minimum value of R eco were calculated using a second degree polynomial regression function between R eco and WTD. No parabolic relationship was observed in 2013 between R eco and WTD; for this reason, time lag, minimum R eco and WTD for minimum R eco could not be calculated.

Period
Drainage During the first two dry spells of 2010 the relationship between R eco and WTD was of clear parabolic form, with R eco reaching a minimum a few days after the onset of the dry period. Except for the second dry period of 2010, the residuals of the regressions between R eco and WTD were not correlated with air nor soil temperature. The two dry spells of 2008 exhibited similar parabolic relationships between R eco and WTD but differed in magnitude.
Such parabolic relationships between R eco and WTD were not observed during the summer of 2013, which was the second driest in the 2002-2013 study period (the driest was 2003 with 346 mm rain between April and September compared to 361 mm in 2013); 2013 also had the longest winter of the study period (start of the growing season at day 103 in 2013 compared to day 77 ± 21 for the entire study period) as well as the lowest soil temperatures. Soil temperature at −5 cm increased by 3 • C in the 10 days prior to the start of the thermal growing season; T soil increased steadily until mid-July and reached 15 • C, the highest value of the 11-year study period, on 26 July. The dry period began on 25 May, culminating on 22 July (WTD = −48 cm), and WTD was ≤ −5 cm until early September. In 2013, the relationship between R eco and WTD was linear across the 6 temperature classes considered (Fig. 8). Above 16 • C, the positive correlation between R eco and WTD was less pronounced and was even found to be negative for the 16-18 • C temperature class. Above 18 • C, the positive linear correlation was no longer statistically significant.
For all years for which WTD data was available, the sensitivity of R eco to air temperature (Q 10 ) decreased with a drop in water table; in contrast, the theoretical values of R eco at T air = 0 • C (obtained by extrapolation to the origin of the temperature-dependent functions fitted to monthly R eco and averaged to annual values) were found to increase with WTD. One-way analysis of variance (ANOVA) on GPP, NEE and R eco with respect to 10 WTD classes (Table 3)

Discussion
The following sections discuss the effects of winter meteorology and water table depth on ecosystem response during the growing season and place the Auchencorth Moss peatland into a broader context by comparing it to other sites in the Northern Hemisphere with published NEE records ≥ 3 years.

Effect of winter meteorology on ecosystem response
Mean winter T air explained 87 % of inter-annual variability in NEE during the following summer (NEE SS ), 43 % of GPP SS and 28 % of R ecoSS (Fig. 6), which is consistent with observations over a 12-year period at a boreal fen in northern Sweden (Peichl et al., 2014). A number of studies have reported correlations between winter meteorological conditions and the development of plant populations later in the year. Weltzin (2000) reported increased total net primary productivity (TNPP) in shrubs, a decrease in graminoids and no effect on bryophytes exposed to a gradient of infrared loading (i.e. continuous heating by infrared lamps). Individual species of bryophytes at a temperate UK site have been shown to respond to winter warming and/or summer drought in opposite ways, but this was not reflected at the community level, whose mean cover did not exhibit significant differences between treatments (Bates et al., 2005). Kreyling (2008) demonstrated enhancement of aboveground net primary productivity (ANPP) in grasses as a result of freeze-thaw cycles the preceding winter, whilst belowground net primary productivity (BNPP) was adversely affected. Eddy-covariance measurements provide spatially integrated fluxes representative of the entire plant community within the footprint of the flux tower. The contributions of individual species, whose productivity can vary from year to year (Bates et al., 2005;Kreyling et al., 2010Kreyling et al., , 2008Weltzin et al., 2000) cannot be assessed by EC. However, based on the knowledge that Sphagnum mosses are capable of photosynthesising as soon as the snow cover disappears and daily air temperature reaches > 0 • C (Loisel et al., 2012), we speculate that the sensitivity of GPP, GPP sat and α to winter air temperature is predominantly caused by graminoids and other non-moss species.

Effect of water table level on GPP and R eco
WTD had a statistically significant negative effect on GPP indicating a decrease in plant productivity caused by the onset of drought stress. This has previously been shown to be important at other sites, particularly in moss species (Aurela et al., 2009;Lafleur et al., 2003;van der Molen et al., 2011); furthermore, a negative linear relationship between leaf area index (LAI) and WTD has been reported for a grassland established on drained organic soil in Ireland (Renou-Wilson et al., 2014), which illustrates the effect of water availability on graminoid productivity. It must however be noted that the WTD range in the Renou-Wilson (2014) study was significantly deeper (typically 20-60 cm below peat surface) than at our study site. Wet-adapted moss species growing in hollows are known to have large variability in growth rate directly linked to WTD (faster growth than hummock and lawn species under wet conditions, but susceptible to desiccation under dry conditions; Gunnarsson, 2005;Loisel et al., 2012). Weltzin et al. (2000) showed that, along a gradient of decreasing WTD of range consistent with our study site, TNPP increased in bryophytes, decreased in shrubs and was un- changed in graminoids. As graminoids and bryophytes were the dominant species in the EC footprint, the sensitivity of GPP to WTD observed at our study site was likely to be mainly due to mosses. The parabolic trend seen in the relationship between R eco and WTD during dry spells (Fig. 7) may help understand the mechanistic drivers of R eco at Auchencorth Moss. The parabolic trends were especially strong during the two first dry spells of 2010 (15 May-9 June 2010 and 10 June-10 July 2010) during which the prevailing wind direction was from the south. The WTD measurements might not be representative of the entire flux footprint, which could perhaps explain the markedly different trends observed in 2008 when wind was blowing from the east. We postulate that the initial decline in respiration was caused by a reduction in plant metabolic activity as water availability decreased (Lund, 2012). Drought has been shown to decrease C assimilation, slow the translocation of photosynthates between above-and belowground biomass, and reduce root-mediated respiration within days (Ruehr et al., 2009). Meanwhile, the lowering of the WT also favours aerobic processes, increases microbial activity and decomposition of organic matter (Hendriks et al., 2007;Moyano et al., 2013), and facilitates CO 2 diffusion within the peat profile (Moldrup et al., 1999;Tang et al., 2005) causing a net increase in CO 2 efflux from the soil. Minimum R eco could then correspond to equilibrium between declining autotrophic and increasing heterotrophic respiration. The decrease of the sensitivity of R eco with respect to T air (Q 10 ) at our site is consistent with findings at other hydric sites where soil respiration (in particular heterotrophic respiration) has been shown to be enhanced by drought . The subsequent net increase in R eco with deepening WTD could then be explained by a gradual increase in the ratio of heterotrophic to autotrophic respiration.
Based on these observations, we attribute the differences in respiration patterns during the dry spells to water table dynamics, which differs from drier sites where temperature (not WT) was found to be the dominant control of R eco (Lafleur et al., 2005;Updegraff et al., 2001). This is further supported by the result of one-way ANOVA, which demonstrates a statistically significant correlation between R eco and WTD for all growing seasons (except for 2012 which had a wetter than average growing season with WT near or above the peat surface for the entire growing season). The linear (rather than parabolic) response of R eco to WTD in 2013 could perhaps be linked to the long winter of 2013 (the thermal growing season began 69 days later than in 2008, and 10 days later than in 2010) and the fact that the dry spell, which lasted most of the summer, began less than a month after the start of the growing season. Under these conditions, the moss population could have switched from relatively low metabolic activity to desiccation while active growth had just begun in the graminoid community. Hence, the R H / R A ratio could have been smaller than in previous years. In contrast to other years, GPP during summer 2013 was positively correlated to WTD (p < 0.001), which suggests growth in species less susceptible to drought-stress than mosses.
Disentangling the effects of lower than average winter air temperature and summer dry spells on annual NEE is not straightforward, but the former seems to be the dominant driver based on our results (Table 1). The combined effects of a long, relatively cold winter and warm, dry summer which could have slowed plant growth, disturbed the normal phenological cycle and enhanced carbon losses from the peatland through enhanced heterotrophic respiration, were illustrated in 2013 when the sink strength of Auchencorth Moss was significantly weakened (−5.2 g C-CO 2 m −2 yr −1 ) compared to the long-term mean of −64.1 ± 33.6 g C-CO 2 m −2 yr −1 (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012).

NEE in Northern Hemisphere peatland C budgets
Compared to other peatlands in the Northern Hemisphere, annual values of NEE at Auchencorth Moss are at the high end of both the mean (−64.1 ± 33.6 g CO 2 -C m −2 yr −1 ) and inter-annual range (−5.2 to −135.9 g CO 2 -C m −2 yr −1 ). However, when the length of the growing season (LGS; the start of the growing season was defined as the first day of the year when mean diurnal air temperature exceeded 5 • C for 5 consecutive days. Conversely, the end of the growing season was defined as the first day of the year when mean diurnal air temperature fell below 5 • C for 5 consecutive days) was accounted for, the mean daily growing season NEE (NEE GS ) at Auchencorth Moss (−0.57 g CO 2 -C m −2 day −1 ) was remarkably similar to that found at both Mer Bleue (cool temperate bog; −0.58 g CO 2 -C m −2 day −1 ; Roulet et al., 2007) and Degerö Stormyr (boreal mire; −0.48 g CO 2 -C m −2 day −1 ; Peichl et al., 2014). By contrast, mean daily NEE GS at Glencar (maritime blanket bog; Koehler et al., 2011, McVeigh et al., 2014 was slightly lower (−0.39 g CO 2 -C m −2 day −1 ), whilst the two sub-arctic Scandinavian peatlands Lompolojänkä (nutrient-rich sedge fen; Aurela et al., 2009) and Stordalen (sub-arctic palsa mire; Christensen et al., 2012) stand out with mean daily growing season NEE rates 2 to 2.5 times higher than the values found for Auchencorth Moss, Degerö Stormyr and Mer Bleue, and over 3 times higher than the value found at Glencar (Table 4).
Despite the lower daily mean NEE, the long growing season at Auchencorth Moss made its total NEE GS comparable to that of Lompolojänkä and Stordalen. The vigorous net uptake at Lompolojänkä during the growing season was offset by relatively high carbon losses during the rest of the year. LGS and LDS are the length of growing and dormant season respectively, and subscripts GS and DS denote growing and dormant season. The length of the growing season for the study site Auchencorth Moss was bounded by the first and last day for which mean daily air temperatures exceeded 5 • C for 5 consecutive days. For the other sites, LGS was estimated from data available in the respective articles.

Site
Auchencorth  Auchencorth Moss, Lompolojänkä and Stordalen therefore had comparable NEE but for very different reasons: Auchencorth Moss had long growing seasons but also relatively high carbon losses the rest of the year, which could be due to milder winters with minimal snow cover. Lompolojänkä and Stordalen had vigorous carbon uptake rates, their LGS were comparable to one another, but were half that of Auchencorth Moss, whilst Lompolojänkä had high carbon losses during the dormant season which strongly reduced the site's sink strength.
Carbon uptake rates at Degerö Stormyr and Mer Bleue were very similar to Auchencorth Moss but their carbon loss rates, which were comparable to Stordalen's, were half or less than that of Auchencorth Moss. This could be explained by cooler climate and prolonged periods of snow cover compared to Auchencorth Moss.
Considering the differences in latitude, climate, hydrology and vegetation, these sites (with the exception of Stordalen and Lompolojänkä) are remarkably similar in terms of their daily mean NEE GS . NEE represents only one flux pathway within the full net ecosystem C budget (NECB). When terrestrial CH 4 emissions Dinsmore et al., 2010), downstream aquatic flux losses and water surface evasion (2007-2011; Dinsmore et al., 2013) are accounted for, the total longterm sink strength of Auchencorth Moss is reduced to approximately −28 g C m −2 yr −1 (whilst recognising un-certainty as the fluxes are not measured over the same time period). Using literature values of CH 4 (Roulet et al., 2007) and aquatic C losses for Mer Bleue (Billett and Moore, 2008) results in an approximate total C sink strength of −17 g C m −2 yr −1 ; for Degerö Stormyr the total C sink strength is −24 g C m −2 yr −1 (Nilsson et al., 2008), −30 g C m −2 yr −1 for Glencar (Koehler et al., 2011) and −34 g C m −2 yr −1 for Stordalen Lundin et al., 2013;Olefeldt et al., 2013); data for Lompolojänkä could not be found. Hence when all flux pathways are accounted for the C balances of the different peatlands appear to converge.

Summary
Eleven years of continuous monitoring of net ecosystem exchange of carbon dioxide at a temperate Scottish peatland revealed highly variable inter-annual dynamics despite little or no change in land management. Variation in climate and especially winter air temperature is thought to be the dominant control at the study site. The latter explained 87 % of inter-annual changes in NEE and a modest rise of 1 • C above average winter air temperature for the 2002-2013 study period was accompanied by a 20 % increase in CO 2 uptake. Colder winters appear to have an adverse effect on the peatland CO 2 sink strength possibly due to disturbances to the phenological cycle of the graminoid species at the site. Dry spells have been linked to enhanced ecosystem respiration and depressed GPP and it is thought that (a) heterotrophic respiration can become the dominant term as water availability decreases, and (b) mosses are more sensitive to WTD than other species at the site. Cold winters and dry summers both have negative effects on the CO 2 sink strength of the bog; these two factors converged in 2013 and led to a significant reduction in net CO 2 uptake (−90 % compared to the 11year mean). Auchencorth Moss, although always a sink of CO 2 during the study period, is highly sensitive to even modest changes in hydro-meteorological conditions at relatively short timescales. The large inter-annual variability of NEE observed to date makes future trends difficult to predict and quantify. Changes in seasonal hydro-meteorological conditions, especially changes in precipitation patterns and intensity, could however be pivotal for the CO 2 cycling of this peatland. Drier summers could lead to a reduction in net CO 2 uptake but this could be offset by milder temperatures, particularly in winter, and longer growing seasons. Mean annual temperatures at the study site have risen by 0.019 • C yr −1 since 1961, which could, in theory, benefit C uptake by the peatland in the long-term since NEE was found to be closely linked to the length of the growing season.
The Supplement related to this article is available online at doi:10.5194/bg-12-1799-2015-supplement.