Carbon accumulation in a drained boreal bog was decreased but not stopped by seasonal drought

Drainage of peatlands is expected to turn these ecosystems to carbon sources to the atmosphere. We measured carbon dynamics of a drained forested peatland in southern Finland over four years, including one with severe drought during 10 growing season. Net ecosystem exchange (NEE) of carbon dioxide (CO2) was measured with eddy covariance method from a mast above the forest. Soil and forest floor CO2 and methane (CH4) fluxes were measured from the strips and from ditches with closed chambers. Biomasses and litter production were sampled, and soil subsidence was measured by repeated levellings of the soil surface. The drained peatland ecosystem was a strong sink of carbon dioxide in all studied years. Soil CO2 balance was estimated by subtracting the carbon sink of the growing tree stand from NEE, and it showed that also the soil was a sink 15 of carbon. A drought period in one summer significantly decreased the sink through decreased gross primary production. Drought also decreased ecosystem respiration. The site was a small sink for CH4, even when emissions from ditches were taken into account. Despite the continuous carbon sink, peat surface subsided slightly during the 10-year measurement period, which was probably mainly due to compaction of peat. It is concluded that even fifty years after drainage this peatland site acted as a soil C sink due to relatively small changes in water table and in plant community structure compared to similar 20 undrained sites, and the significantly increased tree stand growth and litter production. Although the site is currently a soil C sink, simulation studies with process models are needed to test whether such sites could remain C sinks when managed for forestry over several tree-stand rotations.

The mean air temperature in 2005-2008 was 15.3 °C during summer months (June-August) and -3.8 °C during winters (December-March). The annual mean air temperature was 5.1 °C and temperature sum (> 5 °C) 1356 d.d. Annual average 5 precipitation was 722 mm and maximum wintertime snow depth 20-60 cm.

Measurement setup
The micrometeorological measurements were conducted in the centre of the peatland at a 200-250 m distance to an upland forest in the north-west and to a small lake in the south-west. To the north-east the homogenous fetch was longer, about 600 m. The EC footprint thus covered the fairly homogenous peatland pine forest with at least 200 meter radius (Lohila et al. 2011). 10 The chamber measurements of CO2 and CH4 fluxes were conducted at four plots, located 50-100 m from the mast. As every plot consisted of 16 measurement points (collars), the whole setup contained 4×16, i.e. 64 measurement points. In addition, CH4 fluxes from ditches were measured in 2011 at four points on two parallel ditches located on both sides of the mast. The depth of the water table (WT) was manually measured from two perforated plastic pipes at each plot, along with chamber measurements. WT was also continuously recorded close to the EC mast by a PDCR 830 (Druck Messtechnik GmbH, in 2004-15 -2006, and a Hobo U20-001-01 (Onset Computer Corporation, MA, USA, in 2007--2009. Soil temperatures were recorded with temperature loggers (i-Button DS1921G, Maxim Integrated Products) from the depths of 5 cm (T5) and 30 cm (T30) below soil surface at intervals of 1-3 h. In 2005-2006, T5 was recorded from every measurement point and T30 from 16 points (four/plot). In 2007-2008, T5 recordings were taken from two points/plot and T30 recordings from two points in total.
The tree stand, ground vegetation and soil properties were measured on 33 plots located evenly along eight radial transects 20 extending 160 m from the mast (the centre plot). Four transects with plots spaced at 20, 60, 100 and 140 m distances from the mast were alternated with four other transects with plots spaced at 40, 80, 120 and 160 m distances from the mast. The area of each plot was 200 m 2 .

Ecosystem-atmosphere exchange of CO2 25
The Here we report results for the full years 2005-2008. 30 We used an SATI-3SX (Applied Technologies, Inc.) sonic anemometer/thermometer from 2004 to November 2006, after which a METEK USA-1 (METEK GmbH, Elmshorn, Germany) was used. The atmospheric concentrations of CO2 and H2O were measured with an LI-7000 (LI-COR, Inc.) analyzer. This instrument was calibrated bimonthly to monthly with two known CO2 concentrations [CO2] (0 and 421 ppm). CO2-free synthetic dry air was used as a reference gas. The heated inlet tube (3.1 mm Bevaline IV) for the LI-7000 was 17 m long, and a flow rate of 6 l min −1 was used. 35 Biogeosciences Discuss., https://doi.org /10.5194/bg-2017-530 Manuscript under review for journal Biogeosciences Discussion started: 2 January 2018 c Author(s) 2018. CC BY 4.0 License.
The signals were sampled at a frequency of 10 Hz, and the turbulent fluxes were calculated on-line as 30-min averages applying standard EC procedures. The effect of density fluctuations related to the water vapour flux (Webb et al. 1980) was included in the calculations, and the fluxes were corrected for systematic losses using the transfer function method of Moore (1986), including the losses due to autoregressive running mean filtering and the imperfect high-frequency response of the measurement system. Details of the flux calculation and correction procedures can be found in Pihlatie et al. (2010) and Lohila 5 et al. (2011).
To estimate the storage fluxes of CO2, the mean [CO2] observed at a height of 4 m with a LI-820 CO2 analyzer and the [CO2] measured at the top of the mast were used. The storage term was calculated with the central difference method from the mean concentration during the subsequent and preceding 30 min periods and added to the measured turbulent flux. Hereafter NEE refers to the sum of turbulent and storage fluxes. In this paper, we use the convention that a positive value of NEE indicates a 10 flux from the ecosystem to the atmosphere.

Forest floor CO2 efflux
CO2 efflux from forest floor was measured with a closed steady state chamber (diameter 31.5 cm, height 14.9 cm) attached to a portable infrared gas analyzer (EGM-4, PP-Systems, Hitchin, U.K.; NSF11 in Pumpanen et al., 2004). Chamber closure time was 81 s. Measuring points were delimited with permanent collars and had four different treatments including the following 15 respiration components: A) peat soil (including cut roots), B) A + above ground litter, C) B + living roots, and D) C + ground vegetation.
In order to exclude autotrophic respiration, treatments A and B involved trenching with 30 cm deep collars and removing aboveground parts of living vegetation by repeated clippings. From treatment A, the above ground litter was also removed every time before measurements. From treatment C, only the above ground parts of plants were removed and treatment D was 20 left intact. Collar depth in treatments C and D was only 2-3 cm to minimise disturbance to roots. Treatment D (RD) thus includes all respiration components of forest floor respiration (RFF) and treatment A respiration from peat soil only (RPEAT).

Forest floor and ditch CH4 fluxes
Soil CH4 fluxes from the strips between ditches were measured with static chambers from the D points and reported by Lohila et al. (2011). To complement the CH4 flux estimate for the whole area, fluxes from ditches were measured with the same 30 equipment and methods as earlier. Fluxes were measured from four points on two parallel ditches on the both sides of the mast, altogether 7 times between June 28th and December 8th, 2011. The annual flux was estimated as 365 × daily mean flux.

Organic carbon pools and fluxes
The carbon stock in peat, and biomasses and litter production of the tree stand and ground vegetation, were measured to estimate organic carbon pools and fluxes in the peatland. Peat C stock was estimated based on average peat layer thickness on 35 the tree stand transects (Lohila et al. 2011) and average carbon density in peat (Mathijssen et al. 2017 Repola (2008, 2009) and Laiho and Finér (1996 for pine belowground biomasses (root d>1 cm), as described in detail by Ojanen et al. (2012). In all biomass C stock and flux calculations, C content of 50% was assumed.
Above-ground biomass of ground vegetation vascular plants was sampled along the tree stand transects (n plots = 39), from an area of 0.25 m 2 /plot. Moss samples (n = 64) were collected from the same sites using corers with a diameter of 93 or 125 mm. In the lab, the dead part of the moss was cut and removed, based on ocular assessment (color change of the moss). The 5 samples were separated by species and dry mass (105 °C) was determined for each sample.
The biomass of roots (and rhizomes of shrubs) were determined by taking a soil sample of 15 × 15 × 20 cm (width × length × depth) along the tree stand transects, adjacent to the mid points of the tree sample plots (n = 32). In the laboratory all roots were carefully separated from peat, divided according to species/functional groups (pine, spruce, birch, shrubs, grasses and herbs) and diameter (below and over 2 mm), dried in 105 °C and weighed. According to Bhuiyan et al. (2016), 15 % of the 10 fine roots in Kalevansuo are located deeper than 20 cm. The biomasses estimated here were corrected accordingly. C flux in above ground litter was estimated with 14 litter traps (20 × 20 cm) per chamber plot (i.e. altogether 56 traps). Litter was collected 2-3 times per year, separated by species, dried in 105 °C and weighed. As moss litter is not captured by litter traps, moss litter production was estimated by harvesting moss biomass production over 2 and 5 years (Ojanen et al. 2012). As the whole moss biomass eventually dies and forms litter on site, annual moss biomass growth equals annual litter production. 15 Coarse root (>2 mm) litter production was estimated as biomass × turnover rate (0.12 for pine, 0.08 for shrub rhizomes; Finer and Laine 1998). Fine root litter production was estimated with root-ingrowth-cores by Bhuiyan et al. (2016). Sixty cores (diameter 3 cm, length 50 cm) filled with Sphagnum peat were installed into soil in October 2009, and 20 cores were collected every year for three years. The fine root production rate was calculated as the average fine root mass (live+dead) in the cores divided by incubation years (average for 2 nd and 3 rd years). 20

Change in peat layer thickness
To survey the changes in peat layer thickness, caused by compaction and decomposition of soil organic matter, litter production and moss height growth, soil surface around the mast was levelled in 2004, 2011 and 2014. In the beginning of measurements in 2004, a 20 mm thick steel rod was hammered through the peat layer firmly to the subsoil, serving as a stable benchmark.
The soil surface at the undisturbed chamber measurement points (D-collars), was repeatedly levelled in relation to the 25 benchmark. A manual levelling instrument with a levelling rod was used and the readings were recorded with the precision of ±0.5 cm.

NEE
The NEE data obtained from the EC measurements were screened as described by Lohila et al. (2011). In short, screening 30 criteria were applied to remove spikes in the 10-Hz anemometer data and to discard poor-quality 30-min data. For the latter, the criteria were based on the expected range of the mean [CO2] and air temperature (from the sonic anemometer), and of the variances of [CO2], vertical wind speed and air temperature. In addition, a cumulative flux footprint of 70% was required, and a threshold of 0.1 m s -1 was set to the friction velocity (Lohila et al. 2011). The procedures of gap-filling of the EC flux data and partitioning of NEE to the GPP and RECO components are described in Appendix 1. The estimation of uncertainties in 35 annual NEE is described in Appendix 2.

Forest floor CO2 efflux
CO2 efflux from forest floor is a result of heterotrophic and autotrophic processes from different layers (vegetation and soil), which have different temperature dynamics. Therefore an additive, layerwise model was used, in which soil temperatures T5 and T30 predict fluxes from different layers, with different temperature dynamics. An Arrhenius type function (Lloyd & Taylor 1994), was fitted to the measured CO2 efflux (g CO2 m -2 h -1 ) from forest floor: 5 ( 1 ) where RREF5 and RREF30 are respirations at reference temperatures (TREF = 10 °C) and E05 and E030 describe temperature sensitivities of respiration in 5 cm and 30 cm peat depths, respectively. T0 = -46.02 °C is a constant.
Parameter values were estimated separately for different treatments (A-D) representing different components of RFF, the four gas measurement plots, and two groups of years (2005-2006 and 2007-2008; Appendix 3), as the decomposability of soil 10 organic matter changes in time at A collars. WT was also tested as an explanatory variable, but as it predicted the temporal flux variation poorly, it was not included in the final models. The models were used with measured soil temperature data to simulate the temporal dynamics and annual fluxes of different flux components.

Modeling of the tree stand CO2 fluxes
To analyse the contribution of the tree stand (aboveground) to the ecosystem CO2 exchange, we used the GPP and shoot 15 respiration (R) models in Stand Photosynthesis Program (SPP). SPP predicts canopy light interception, photosynthesis and shoot respiration in half-hourly time steps (Mäkelä et al., 2006). PPFD, air CO2 concentration, air temperature, and relative air humidity measured at the site were used as inputs for SPP. The photosynthesis model used was OPAC (Mäkelä et al. 2006).
Tree stand was described as three size classes (Ojanen et al. 2012), foliar masses for each class were estimated using the models of Repola (2009), and these were converted to leaf area index with specific leaf area of 11 m 2 kg −1 (Luoma, 1997). Stem 20 respiration was estimated with the model of Zha et al. (2004).

Meteorological conditions
Of the studied years, 2008 was the warmest, especially during the winter months January-March, which were almost snowless.
It was also the rainiest year. The summer (June to August) of 2008 was significantly cooler, but otherwise similar to the other 25 summers. In contrast, the year 2006 was exceptionally dry from January until the end of September, including a severe drought during the growing season. In summer 2006, air temperature and PPFD were higher than on other years, whereas relative humidity and water table were lower (Table 1, Figs. 1 and 2). The dry and warm growing season 2006 was preceded by a cold winter, which is why soil surface temperatures (T5) were much below average in the spring, and down in 30 cm stayed below average until September, i.e. for almost the whole growing season (Fig. 2). In September-October the deeper peat layers finally 30 warmed up and stayed warmer than average for the rest of the year.

Ecosystem CO2 exchange
According to the EC flux measurements, the site acted as a CO2 source typically during the winter months (October-March) and a sink during the growing season (April-September) (Fig. 3, 4a). The variation in NEE during winter was small, ranging from about -0.1 to 0.1 mg m -2 s -1 (Fig. 3). While there were occasional, warm days with net CO2 uptake during the winter, the actual spring recovery of photosynthesis seemed to occur typically in the beginning of April, the only exception being the 5 spring after the warm winter of 2006-2007, when the recovery started already in March. In summer (June-August), the highest night-time CO2 emission values, representing RECO, were on average 0.35 mg m -2 s -1 , and the highest day-time CO2 uptake typically fluctuated around -0.75 mg m -2 s -1 . Only in summer 2006 the amplitude in the diurnal dynamics was smaller.
The site was a sink of CO2 in all years, NEE varying between -520 and -990 g CO2 m -2 a -1 ( Table 2). The average value for the four years was -860 g CO2 (i.e. -234 g C m -2 a -1 ). With the exception of the dry year of 2006, the annual NEE was 10 surprisingly similar in other years, varying from -950 to -990 g CO2 m -2 a -1 .
The drought during the spring and the growing season of 2006 was clearly reflected in the CO2 exchange. The (gap-filled) NEE and GPP were markedly less negative in June and July 2006, indicating lower CO2 uptake by photosynthesis as compared to the other years ( Fig. 4c). However, in July and August the RECO was also clearly suppressed ( Fig. 4d), thus decreasing the net loss of CO2 from the peatland (NEE). In September 2006, the GPP had fully recovered to the level of other years, but the 15 RECO stayed at slightly higher level during the rest of the year, leading to clearly higher NEE during the last months of the year.
After the first week of June until the end of July 2006, there were only a few days with accepted NEE observations (Fig. 3), so the results shown for these months (Fig. 4) largely depend on gap-filling. However, the data coverage was considerably better in August, making it possible to study the impact of drought on NEE. By plotting the daytime NEE against PPFD ( which suggests that the ecosystem was slowly recovering from the drought in the autumn. Thus, the distinct decrease in the annual net CO2 uptake in 2006 (Table 2) seemed to be caused by the GPP decrease during 25 the summertime, although RECO decreased during the drought as well. The higher RECO in autumn months after the drought and heavy rains in October (Fig. 2) furthermore increased the difference to other years: the cumulative NEE in October-December in 2006 was 320 g CO2 m -2 , while in other years it varied from 130 to 190 g CO2 m -2 .
The regression models with T5 and T30 as explanatory variables (Eq. 2) explained 70% (46%-90%) of the variation in the 35 fluxes of the entire dataset (Appendix 3). Respiration rates at 10 °C (RREF5 and RREF30) increased from A to D collars, i.e. as respiration components were added, and decreased at A collars with time since the beginning of the study (05-06 to 07-08).
The modelled annual respiration ranged during the first two years from 1233 g CO2 m -2 a -1 in A collars (RPEAT) to 2312 g CO2 m -2 a -1 in D collars (RFF, Table 3). During 2007-2008, RPEAT clearly decreased from the previous years, to ca. 830 g CO2 m -2 a -1 , whereas RFF varied little between the studied years. In A collars the decomposability of organic matter is likely gradually decreased when the labile components are decomposed and the recalcitrant ones are enriched. Also, as we had to remove a newly grown moss layer from A-collars in in spring 2007 (inevitably with some soil organic matter attached), this procedure 5 probably decreased the proportion of labile components on the soil surface.
Based on the modelled fluxes of the first two years, RHET contributed 75% and RAUT 25% to the mean annual RFF (Table 3).
RPEAT comprised 53% of the flux, RLITTER 22%, RROOT 16% and RGV 8%. The four-year mean of RFF was 2197 g CO2 m -2 a -1 , i.e. ca 600 g C m -2 a -1 . Using this mean value with the proportions from 2005-2006, we get an estimate for RHET of 450 g C m -2 a -1 and RAUT 150 g C m -2 a -1 . 10 In 2006, the main part of summertime (June 15th to September 12th) measurements were lost due to instrument failure. Thus, we cannot reliably analyse the impact of 2006 summer drought on forest floor respiration. The existing soil CO2 efflux data from September 2006, when WT was extremely low, do show higher effluxes than those in early June 2006, although soil surface temperatures (T5) were lower in September. However, at the same time T30 was much higher (10.7 °C) than in June (5.4 °C), explaining the increased efflux. Compared to the other years, soil temperatures in September were at their highest in 15 2006 (Fig. 2), and the temperature response models thus predicted higher fluxes for September 2006 than for the other years.
Following the heavy rains in the beginning of October, respiration decreased at the same time with the the rise of WT -and the decrease in T5.
The impact of WT on forest floor respiration was ambiguous. Correlations between WT and CO2 efflux were weak and variable by year and treatment. The residuals of the model (Eq. 1) estimates vs. WT indicated a positive response especially in D collars 20 (lower RFF with lower WT). However, this effect was caused mostly by spatial variation, as measurement points in hummocks generally had lower WT and lower respiration than the points in the lawn-level. Since the models were used for predicting temporal dynamics, WT was not included in the models.

Simulated tree stand CO2 flux
The SPP-model simulated the tree stand GPP and respiration well. For the year 2008 with the most complete NEE data, the 25 RECO, derived from the gap-filling and partitioning of the EC measurements, matched very well (0.8% difference) the modelderived sum of RFF and above-ground tree respiration (RTREE, Fig. 7a, Table 2). Not surprisingly, the model was not able to simulate the suppression of respiration in 2006 (Fig. 7b), apparently since it does not have linkages to soil moisture. The simulated four year average was 9% higher than the EC-derived RECO ( Table 2).
The simulated four-year average GPP of the tree stand was 2473 g CO2 m -2 a -1 (675 g C). The GPP for the ground vegetation, 30 measured by manual flux chambers in another campaign, was 1040 g CO2 m -2 a -1 (Badorek et al. 2011). Altogether the tree stand and the ground vegetation GPP sum up to 3513 g CO2 m -2 a -1 , which is relatively close (92%) to the ecosystem GPP obtained from the partitioning of the EC fluxes (3805 g CO2 m -2 a -1 ). These independent findings suggest that the tree stand contributes about 70% and the ground vegetation 30% of the GPP at Kalevansuo.

CH4 fluxes 35
CH4 flux from ditches was very variable, especially spatially but also temporally. The instantaneous fluxes varied between -0.098 and 1.757 mg CH4 m -2 h -1 . The wettest plot, with cottongrass (Eriophorum vaginatum) emitted on average 0.936 mg CH4 m -2 h -1 , significantly more (p <0.001) than the other three, slightly drier plots, (mainly Sphagnum riparium), with mean Biogeosciences Discuss., https://doi.org /10.5194/bg-2017-530 Manuscript under review for journal Biogeosciences Discussion started: 2 January 2018 c Author(s) 2018. CC BY 4.0 License. fluxes of 0.006, 0.056 and -0.006 mg m -2 h -1 . Temporal variation was high but no clear seasonality was observed. At the wettest plot, fluxes had similar temporal pattern with WT, i.e. the highest flux took place in September during the highest WT.

Change in peat layer thickness
The soil surface on the undisturbed D collars had subsided on average by 1.4 cm in ten years from 2004 to 2014, i.e. 1.4 mm a -1 (Fig. 8). There was considerable variability between points from an increase in elevation by 2 cm to a subsidence of 5 cm, so that the change was not quite statistically significant (p=0.067). Also, some back and forth variation in peat thickness 10 between years was observed: in August 2011 all but four points had lower elevation than in 2014. This can be either a measurement error or real shrink-swell behaviour (breathing) of the peatland.

Carbon balance
The biggest carbon pool at Kalevansuo (Fig. 9.) was the 2.2 m thick peat layer making 95.3% of the total carbon pool. Tree stand (without fine roots) comprised 4.3% and ground vegetation only 0.4%. Fine roots comprised 0.2%. The total C pool in 15 vegetation in 2008 was 5.5 kg m -2 , which corresponds to about 10 cm layer of peat. Aboveground parts comprised 62% of the total biomass. Of the moss biomass, Sphagna comprised 20% and forest mosses 80%.
The tree stand volume increased from 90 m 3 ha -1 in 2000 to 130 m 3 ha -1 in 2008, i.e. on average by 5 m 3 ha -1 a -1 . The corresponding carbon pool was 4.6 kg m -2 in 2008 and 3.2 kg m -2 in 2000. The tree stand thus sequestered ca. 170 g C m -2 a -1 . This made 74% of the carbon accumulation at Kalevansuo, while the rest was attributed to peat soil (Fig. 9). 20 Total litter production was estimated at 437 g C m -2 a -1 . Of this mosses comprised 20% and vascular plants 80%. Of the litter production by vascular plants, trees comprised 79% (aboveground) and 66% (belowground). Fine root production was estimated at 120 g C m -2 a -1 (Bhuiyan et al. 2016), comprising 76% of the belowground litter.
As the average of the four years, the Kalevansuo peatland ecosystem fixed ca. 1040 g C m -2 a -1 through photosynthesis, 70% of which was attributed to the tree stand. Simultaneously it lost 810 g C m -2 a -1 through RECO. Ca. 50% of RECO resulted from 25 heterotrophic respiration and 50% from autotrophic respiration of trees and ground vegetation. RFF was comprised mainly of heterotrophic respiration of peat and litter (75%), and less by autotrophic respiration of tree roots and ground vegetation (25%). Some C may have been lost through leaching (not measured), but this is considered a minor component due to ineffective ditches and high transpiration. No C was lost as methane, as the site was a small CH4 sink (-0.06 g CH4 m -2 a -1 ), which is insignificant for the C balance. 30

Ecosystem CO2 fluxes -the effects of drought
The drained peatland forest Kalevansuo was a strong CO2 sink in all the four years studied (2005)(2006)(2007)(2008) reduced in summer, and the reduction in the GPP was larger. In addition, the higher-than-normal soil temperatures and consequently higher RECO in autumn 2006, explained partly the much lower annual CO2 net uptake.
Despite the long gap in the NEE data in June and July 2006, we were able to demonstrate that drought had an impact on the CO2 exchange. Based on the direct responses between the night-time NEE (respiration) and temperature, and the daytime NEE and PPFD both the ecosystem respiration and daytime net CO2 uptake were reduced in August 2006 as compared to the other 5 years. In September 2006, the difference in respiration and photosynthesis parameters to the other years got smaller, indicating recovery of the ecosystem from the drought-induced suppression of GPP and RECO.
Drought has been shown to strongly affect NEE through decreased GPP in pristine mires where vegetation is adapted to high water table (Alm et al. 1999, Bubier et al. 2003, Lafleur et al. 2003. Although Scots pine, the main tree species in Kalevansuo peatland, is a drought-tolerant species, summer droughts have been reported to decrease its radial growth in drained peatlands 10 (Huikari and Paarlahti 1967). The water table in Kalevansuo is usually rather high, which means that the roots of pines are located mainly in the top 40 cm (Bhuiyan et al. 2016), in the oxic layer above the average water table. During drought, when water table may drop down to 80 cm for several weeks, even pines will probably suffer from water deficit, and close their stomata.
In contrast to GPP, RECO and soil respiration have often been shown to increase in peatlands, when water-table is lowered and 15 more peat is exposed to oxidation (e.g. Silvola et al. 1996, Flanagan and Syed 2011, Ballantyne et al. 2014, Munir et al. 2014. However, many studies have shown only a weak or no impact of WT on RECO, whereas soil temperature has been driving the respiration fluxes (Lafleur et al. 2005, Nieveen et al. 2005, Juszczak et al. 2013, Olefeldt et al. 2017. In Kalevansuo, the latter seems to be the case. RECO was slightly lower during the drought in August 2006 compared to other years (Fig. 5).
RFF was strongly controlled by soil temperatures, whereas WT had only a weak and varying effect in different treatments and 20 years.
The decrease in RECO may be caused by decrease of both RAUT and RHET. As the drought decreases GPP, it will decrease also photosynthetically driven autotrophic respiration (Olefeldt et al. 2017), while heterotrophic respiration may well continue in deeper, still moist but now more oxic, peat layers. However, a large part of RHET is originated from the decomposition of the new organic matter (Chimner and Cooper 2003), i.e. above ground and root litter, deposited mainly in the very surface of the 25 peat soil. In drained peatlands the decomposition rate of this surface layer is hardly ever restricted by too high WT, but sometimes it can be restricted by too low moisture content (Mäkiranta et al. 2009).
If water levels were lowered for a longer period, e.g. through deeper ditching, the effect might be different than that of drought: a more efficient drainage would induce higher decomposition and heterotrophic respiration through changes in microbial communities (Mäkiranta et al. 2009) but also probably increased root growth into the deeper layers. 30

Soil subsidence
Even though the flux and biomass data indicate a steady increase in soil C stock, a small (insignificant) subsidence of the soil surface was measured (0.14 cm/year). The value is considerably smaller than what has been reported for agricultural fields Subsidence of peat is caused by physical compaction and loss of organic matter through oxidation. In physical compaction, solid matter is compacted into a smaller space. The result is the increase in bulk density, which is evident in all drained peatlands (e.g. Minkkinen and Laine 1998b). In oxidation, organic matter is lost as CO2 from the peat to the atmosphere. In peat soil, both processes take place at the same time, and in forested sites especially, the C loss through oxidation is to varying extent compensated for by litter production. Thus, given the estimated positive soil C balance at Kalevansuo, we conclude that 5 the observed small subsidence is caused by compaction, not by loss of peat.

Carbon balance
Kalevansuo accumulated atmospheric C every year studied. Given that the average net carbon uptake of the site was 230 g m -2 a -1 and that 170 g m -2 a -1 was sequestered to the growing tree stand, the remaining 60 g C m -2 a -1 must have been accumulated in the other parts of the ecosystem. If the ground vegetation biomass is assumed constant, the surplus must be in the peat soil. 10 This assumption is based on ocular assessment at the site. It is reasonable to assume that the ground vegetation biomass is not increasing, since the tree stand is steadily growing bigger and the correlation between tree stand and ground vegetation biomass is negative (Reinikainen et al. 1984). Furthermore, an increase of 60 g C m -2 a -1 would equal the doubling of shrub biomass in 5 years, and that should be clearly visible. Thus the method should not be overestimating soil C pool increase, more likely underestimating it. However, as the C pool in ground vegetation is one-tenth of that in the tree stand, the change in C pool 15 would be irrelevant, assuming the same relative growth rate.
Despite the small biomass pool compared to the tree stand, ground vegetation was estimated to produce above ground litter at a rate of 130 g C m -2 a -1 , i.e. almost as much as the tree stand (Fig. 9). The majority of this litter originates from mosses, the coverage of which is almost 100% in Kalevansuo. Another rapidly renewing biomass pool was that of fine roots, which was composed almost totally of tree and shrub roots. About half of this pool is renewed annually, producing root litter at a rate of 20 120 g C m -2 a -1 (Bhuiyan et al. 2016). Thus, although being small C pools, both ground vegetation and fine roots have a large impact on the soil C balance.
In our estimation, the C in the below ground parts of trees (stumps and roots > 1 cm diameter) was considered as tree biomass, which increases as the stand grows. When trees die, either naturally or as they are harvested, the below-ground part of C becomes a part of the soil C pool. Considering this below-ground biomass as a part of the soil C pool, would increase the soil 25 C accumulation estimate to over 100 g C m -2 a -1 . The biomass of smaller roots could of course also change, but as the biomass pool of the 2-10 mm roots is only a small fraction of that of the bigger ones (Fig. 9), and as the fine root turnover is rapid (50% a -1 ), this is not considered a major uncertainty. Ojanen et al. (2012) evaluated different chamber-based methods for calculating the soil C balance, and compared these to the EC-based method described above. The "L-RHET -method" (litter production minus heterotrophic respiration) produced 30 varying results depending on the variable fine root turnover rates available from literature. Using the recent results of fine root production in Kalevansuo (Bhuiyan et al. 2016) we end up with L of 437 g C m -2 a -1 -and RHET of 450 g, which results in a loss of 13 g C m -2 a -1 . Thus there is still a difference of about 73 g C m -2 a -1 to the EC-based estimate. This difference is probably caused by uncertainties in estimating RHET (Ojanen et al. 2012). The cutting of roots causes an extra litter input (e.g. Subke et al. 2006) and on the other hand prevents further input. Roots may also reach under the 30-cm deep collars (Bhuyian 35 et al. 2016). Trenching also affects soil moisture that regulates respiration (Subke et al. 2006).
In addition to the "L-RHET -method", soil C balance can be estimated using data from transparent chamber and tree litter measurements, as follows: where GPPFF is chamber measured GPP of forest floor vegetation and LTREE total litter from trees. Since the chamber-measured RFF includes also tree stand root respiration this must be subtracted from RFF.
This gives an estimate for the soil C balance of -30 g C m -2 a -1 (sink), which is relatively close to the EC-based estimate of -60 g C m -2 a -1 , and supports our finding of the soil C sink. 5

Can the carbon sink last?
Here we have shown that the Kalevansuo drained peatland ecosystem and even the soil is currently a carbon sink despite the drainage. It would be reasonable to assume that drainage would turn a peatland soil into a carbon source, because the decomposition of peat is typically increased after drainage. Drainage in Kalevansuo is, however, rather superficial, the average water table being at 35-40 cm, i.e. only about 15-20 cm lower than in natural dwarf-shrub pine bogs (Minkkinen et al. 1999). 10 The site is topographically rather even, as is typical for nutrient-poor pine bogs, so draining with open ditches is not efficient.
Thus the ditches are blocked by vegetation and drainage is mainly mediated by the transpiration through the tree stand (Sarkkola et al. 2010). The soil is also almost fully covered with vegetation, including mire species like Sphagnum mosses.
Such a small change in vegetation structure is typical for drained dwarf shrub pine bogs (Minkkinen et al. 1999). It thus appears that this peatland has not lost the ability to keep up the relatively high water table and surface moisture supporting the 15 continuous growth of mosses. Only very dry seasons, like summer 2006, may disturb the hydrology so much that C dynamics are seriously affected.
It is evident that most boreal and temperate peatland forest ecosystems, where drainage has been successful, act as contemporary C sinks (Ojanen et al. 2013, Meyer et al. 2013, Hommeltenberg et al. 2014, because the tree stand C sequestration exceeds the loss of C from soil. In peatlands used for forestry it is however the soil C storage that is important 20 in the long-term, given that the tree stock will eventually be harvested and the C in wood products will gradually be lost back to the atmosphere. Thus the most relevant question is: Will sites like Kalevansuo remain C sinks in the long-term if they are managed for forestry? After the site is harvested, as typical, by clear-cutting, soil decomposition processes will go on, whereas litter production from tree stand is ceased for several years. Logging residues will decompose rather fast, and may enhance the decomposition rate of the underlying peat soil (Mäkiranta et al. 2012, Ojanen et al. 2017). This will create a loss of soil C 25 through soil respiration, the magnitude of which is dependent on soil quality (von Arnold et al. 2005a, b, Minkkinen et al. 2007).
On the other hand, in typical stem-harvesting method, tree stumps and roots are left at the site, increasing the C stock in the soil significantly. This C pool of coarse woody debris is not easily decomposed (Laiho and Prescott 2004) especially when buried in peat soil, and its inclusion will compensate for the soil C losses for several years. Also, after clear-cut, the water table  30 will rise because of the removal of the transpiring tree stand, likely reducing peat decomposition rate (Mäkiranta et al. 2010).
This reduction is, however, probably quite small and the site is likely to be a strong C source at least for the first five years, after which the growing vegetation again starts to bind carbon to the system (Mäkiranta et al. 2010, Kolari et al. 2004).

Conclusions
The Kalevansuo drained peatland forest in southern Finland was a strong carbon sink despite the drainage during all the four years studied. The peat soil also continued to accumulate carbon, at an estimated mean rate of 60 g C m -2 a -1 . It was thus at least as strong a C sink as natural peatlands in general. In addition the site was a small CH4 sink, contrary to natural mires.
Based on earlier knowledge of similar sites on drained peatlands, Kalevansuo is not an exception, but rather represents a typical 5 drained pine bog, regarding the greenhouse gas fluxes. Modelling studies, in addition to further measurements focusing on young stands the first 20 years after cuttings would be necessary to show whether the sink is maintained under long-term production forestry.
Drought affected the CO2 fluxes and had a strong impact on the C balance of Kalevansuo mainly through the decrease in photosynthesis. Simultaneously suppressed respiration decreased the potential C loss from the system, however, and the site 10 remained a clear C sink even during the drought. Occasional droughts thus do not seem to threaten the sink capacity of such peatlands.  Table 1. Meteorological parameters for the full years and summer months, June-August. T = mean air temperature, P= precipitation sum, PPFD = mean daily sum of photosynthetic photon flux density, RH = mean relative humidity, VPD = mean vapour pressure deficit in the afternoon, 12:00-16:00 local time.

Acknowledgements
Year June-August Year  Table 2. EC-measured (and gap-filled and partitioned) annual net ecosystem exchange (NEE ± error; Appendix 2), gross primary production (GPP) and ecosystem respiration (RECO) of the Kalevansuo peatland, in comparison with the simulated tree stand GPP (GPPTREES), tree stand above ground respiration (RTREES_AG) and forest floor respiration (RFF). Unit: g CO2 m -2 a -1 .
-     . Measured carbon pools (rounded boxes; g C m -2 , and the changes in pools in italics; g C m -2 a -1 ) and fluxes (arrows and square boxes; g C m -2 a -1 ) in Kalevansuo drained peatland. Soil C accumulation is calculated as NEE (230 g C m -2 a -1 ) -C sequestration in tree stand biomass (170 g C m -2 a -1 ; above-126 g C m -2 a -1 and belowground 43 g C m -2 a -1 ). Other fluxes and pools

Appendix 1. Gap-filling and partitioning of net ecosystem exchange
The gap-filling of the net ecosystem exchange ( ) data obtained from the eddy covariance measurements was performed with the procedures incorporated into the FluxPartFill.py program developed at the Finnish Meteorological Institute. The gapfilling algorithm is based on empirical functions for total ecosystem respiration ( ) and gross primary production ( ) 5 and thus additionally provides the partitioning of into the and components ( 1 ) Ecosystem respiration was assumed to respond to temperature according to the Arrhenius-type relationship suggested by Lloyd 10 and Taylor (1994) (Eq. (2)): where is temperature, is the reference respiration ( at 283.15 K), describes the temperature sensitivity of 15 to , and 227.13 K is a constant. Eq.
(2) is fitted to nocturnal (photosynthetic photon flux density 5 μmol m -2 s -1 ) flux data by optimizing the parameters and .
Gross primary production was assumed to depend on according to a rectangular hyperbola that is multiplied by a function ( ) representing the reduction of with increasing water vapour pressure deficit ( (Eq. (3)

Appendix data 2. Uncertainty analysis of NEE
The uncertainty of the annual CO2 balance was estimated separately for each year. We followed here the approaches presented by Aurela et al. (2002), Lohila et al. (2011) andRäsänen et al. (2017). The random error arising from the stochastic variability of turbulent fluxes (EMEAS) was estimated, similarly to Räsänen et al. (2017), from the difference between the measurements and the corresponding values obtained from the gap-filling model fits (Eqs. 1-4 in Appendix A). This error varied between 5.8 5 and 13 g CO2 m -2 yr -1 in 2005-2008 (Table A1). The same approach was applied to the random error arising from the gapfilling of the data (EGAPS), which ranged from 6.0 to 7.3 g CO2 m -2 yr -1 in 2005-2008 (Table A1). The uncertainty associated with the corrections for the high-frequency flux loss (EHFL) was estimated at 3% of the annual balance (Lohila et al. 2011). For the uncertainty due to the gap-filling of the longer (>2 days) gaps in the flux data (ELONGGAPS), we adopted a new approach: as year 2008 had only few gaps, we simulated the impact of longer gaps in other years by assuming similar data gaps in the time 10 series of 2008 and then ran the gap-filling procedures for these compromised data. For each gap, a cumulative CO2 balance was estimated from two differently gap-filled data sets, i.e. the original and the simulated, and the difference of these was assumed to represent the error. The annual error was calculated by assuming that the errors obtained this way for separate gaps were independent of each other. For 2006, a similar simulation was also done with the data of 2005 and 2007, and the average of the three annual estimates obtained was taken as the total error related to the gap-filling of long gaps. In 2008, there was 15 only one longer (10 day) gap in December. The uncertainty due to this gap was calculated by adopting the daily errors estimated for December 2007, resulting in an ELONGGAPS of 13.3 g CO2 m -2 yr -1 for 2008. The total uncertainty in the annual balance was calculated by combining the different errors using the error propagation principle. The total error of ±37 g CO2 m -2 yr -1 in 2005 was significantly reduced from that reported by Lohila et al. (2011) (± 100 g).
This was mainly due to the different approach adopted for the error estimates for the compensation of long data gaps: for this, Lohila et al. (2011) shifted model parameters 2 weeks forward and backward, which resulted in a relative error of 10.7% of 25 the annual balance. We consider the present approach more realistic, as it is based on assessing the effect of realized gaps on actual measurement data. However, it is obvious that the uncertainty estimate for 2006 is limited by the fact that the summer of that year was exceptionally dry, and the changes in NEE induced by the drought cannot be accurately estimated based on the data of 2008. It is likely that the dynamics of photosynthesis and respiration during a dry summer are different from a normal year. This hypothesis gains support from the observation that the RECO and GPP parameters in August 2006 differed 30 markedly from those estimated for the other years ( Fig. xx in chapter 3.2). Biogeosciences Discuss., https://doi.org /10.5194/bg-2017-530 Manuscript under review for journal Biogeosciences Discussion started: 2 January 2018 c Author(s) 2018. CC BY 4.0 License.
Despite the large gaps during the growing season of 2006, and the large uncertainty resulting from these, the annual NEE balance of 2006 differed significantly from the other years. This difference between two annual balances (NEEi and NEEi+1) was considered significant, if the 95% confidence interval of the difference, defined as 2 ( E q . 1 A x ) 5 where SEi is the standard error of NEEi, did not cross zero.
Biogeosciences Discuss., https://doi.org /10.5194/bg-2017-530 Manuscript under review for journal Biogeosciences Discussion started: 2 January 2018 c Author(s) 2018. CC BY 4.0 License. Appendix 3. Parameter values of forest floor CO2 efflux models (Eq.1) for different collar treatments (A = peat, B = peat + litter, C = peat + litter + roots and D = peat + litter + roots + ground vegetation) and years. RREF5 and RREF30 are respirations at 10 °C for the 5 cm and 30 depths (g CO2 m -2 h -1 ), E05 and E030 are is temperature sensitivities of respiration for the same layers, r 2 is coefficient of determination of the model and n is the number of observations.