Introduction
Lakes and wetlands are linked to the atmospheric carbon pool via the exchange
of methane (CH4) and carbon dioxide (CO2), which are two important
climate-forcing gases (Myhre et al., 2013). While wetlands have been a focus
of study due to their high CH4 source function (Christensen et al.,
2003; Crill et al., 1988; Olefeldt et al., 2013) and carbon sequestration
capacity (Kayranli et al., 2009; Whiting and Chanton, 2001), lakes have only
recently been incorporated together with streams as a separate source into
global CH4 budgets with an uncertain global emission rate of
8–73 Tg CH4 yr-1 (Kirschke et al., 2013). The low number of
experimental studies and the variability in the magnitude of emissions across
lake types (Wik et al., 2016) explain some of this large uncertainty. Carbon
emissions from lakes outweigh part of the land carbon sink, because they emit
CH4 (Bastviken et al., 2011) and because they respire as CO2 a portion
of the carbon that is transported laterally from terrestrial soils to lakes
(Algesten et al., 2004; Battin et al., 2009; Cole et al., 2007; Tranvik et
al., 2009). Hence, lakes play an important role within the terrestrial carbon
budget.
Wetlands and lakes are particularly abundant in the subarctic and boreal
regions (Smith et al., 2007; Verpoorter et al., 2014), where climate warming
is occurring at a faster pace than in the rest of the world (Serreze and
Barry, 2011). In this context, freshwaters have received increasing attention
over the past decade, due to the potential for lakes and particularly Arctic
thermokarst lakes to exert a feedback on climate warming through large
CH4 emissions (Walter Anthony et al., 2016; Walter et al., 2006). While
non-thermokarst, post-glacial lakes emit less CH4 per unit area
(Sepulveda-Jauregui et al., 2015; Wik et al., 2016), they cover a larger area
and may as a whole emit half of the total CH4 emissions
(16.5 ± 9.2 Tg CH4 yr-1) recently attributed to northern
(> 50∘ N) lakes and ponds (Wik et al., 2016).
Biogenic CH4 production (methanogenesis) occurs in anoxic environments
such as lake sediments and water-saturated peat (Cicerone and Oremland,
1988). The process is controlled by the interplay between temperature and the
input of organic matter (Kelly and Chynoweth, 1981; Yvon-Durocher et al.,
2014; Zeikus and Winfrey, 1976). In lakes and waterlogged wetlands, CH4
reaches the atmosphere from its production zone via direct bubble release up
to the surface (ebullition), through emergent vascular plants, or via
turbulence-driven diffusion through the water column (Bastviken et al., 2004;
Lai, 2009; Rudd and Hamilton, 1978). The net flux of CH4 at the lake
surface is a balance between the production in the sediments and the
oxidation of CH4 into CO2 at oxic–anoxic boundaries within the
water column (Casper, 1992). In shallow lakes, ebullition is a main pathway
for CH4 to reach the atmosphere while bypassing the oxidation zones
(Bastviken et al., 2004). In wetlands, transport from the peat soil to the
surface through vascular plants by passive diffusion or by pressurization
effects depending on the plant species (Brix et al., 1992) is an efficient
pathway for CH4 to avoid oxidation in the soil and water before reaching
the atmosphere (Joabsson and Christensen, 2001).
Dissolved CO2 in lakes is produced throughout the water column and
sediments (Casper et al., 2000) or is directly imported from the catchment
(Maberly et al., 2013; Weyhenmeyer et al., 2015). In situ production comes
from the mineralization or the photochemical oxidation of carbon (C) input
from the surrounding catchment (Cory et al., 2014; Dillon and Molot, 1997;
Duarte and Prairie, 2005) and from the degradation of locally produced
organic carbon. CO2 exchange across the air–water interface is
primarily via diffusive release rather than ebullition (e.g., Casper et al.,
2000). Lake waters are generally observed to be supersaturated in CO2
with respect to atmospheric values due to in-lake respiration processes
outweighing rates of primary production (Duarte and Prairie, 2005; Sobek et
al., 2005). Hence they are generally CO2 sources to the atmosphere,
albeit nutrient-rich lakes and ponds can be small CO2 sinks during
summer months (Huotari et al., 2011) or an entire summer season (Laurion et
al., 2010; Pacheco et al., 2013; Shao et al., 2015; Striegl and
Michmerhuizen, 1998; Tank et al., 2009).
Near-surface atmospheric forcing is a key driver for the transport and net
emissions of gases from a lake to the atmosphere. Ebullition of CH4 in
lakes is partly triggered by water level changes or drops in atmospheric
pressure (Casper et al., 2000; Mattson and Likens, 1990), as a decrease in
the hydrostatic pressure of the overlying water column on gas saturated
sediments favors the release of bubbles (Varadharajan and Hemond, 2012).
Wind-driven turbulence is a recognized driver of diffusion-limited exchange
of CO2 and CH4 across the lake-water interface (Sebacher et al.,
1983; Wanninkhof et al., 1985). Convective mixing due to the cooling of the
lake surface or following the breakdown of thermal stratification in the
water column can increase advection of gas-rich water from the lake bottom,
thus enhancing the diffusion-limited release of gases to the atmosphere
(Eugster, 2003; MacIntyre et al., 2010; Podgrajsek et al., 2015), especially
if the main source of those gases is the sediments, as is the case for
CH4.
Northern ecosystems have strong seasonal contrasts, with short growing
seasons and long snow-covered winters. A snow cover insulates the soil from
very cold air temperatures (Bubier et al., 2002), while an ice lid on a lake
temporarily inhibits the exchanges of gas and heat between the water and the
atmosphere (e.g., Greenbank, 1945) and greatly dampens wind-driven turbulence
in the water column. In lakes, gases can accumulate at the bottom during long
stratification periods, or under lake ice. Lake overturn events most often
occur in spring for seasonally ice-covered lakes and/or in fall if lakes
thermally stratify during summer (Kirillin et al., 2012; Wetzel, 2001).
Overturn events can thus lead to the fast release of accumulated gas to the
atmosphere (Jammet et al., 2015; Michmerhuizen et al., 1996; Phelps et al.,
1998), but also to the input of atmospheric oxygen, thus increasing the
potential for methanotrophy (Kankaala et al., 2006; Schubert et al., 2012).
Extension of observations across all seasons of the year is rare at high
northern latitudes, particularly for lakes, yet is indispensable for reducing
the uncertainty in the magnitude of annual carbon exchange and improving
understanding of the processes driving them.
Common methods to measure lake–atmosphere fluxes include floating chambers
(e.g., Bastviken et al., 2004), gas transfer models (e.g., Cole and Caraco,
1998),
and bubble traps (e.g., Walter et al., 2008; Wik et al., 2013;
Sepulveda-Jauregui et al., 2015). If these methods are not integrated and
combined, they inherently omit part of the total surface flux. The
application of the eddy covariance (EC) method (Aubinet et al., 2012;
Moncrieff et al., 1997) to lake environments (e.g., Anderson et al., 1999;
Eugster, 2003; Huotari et al., 2011; Mammarella et al., 2015; Podgrajsek et
al., 2014; Shao et al., 2015) offers long-term flux monitoring and is a
potential methodological solution for solving the spatial and temporal issues
of measuring total gas exchange in lakes. Few studies, so far, have used eddy
covariance to quantify long-term CO2 emissions from boreal lakes
(Huttunen et al., 2011) as well as CH4 emissions from boreal (Podgrajsek
et al., 2014) or subarctic lakes (Jammet et al., 2015). We report here one of
the first year-round eddy covariance measurements of both CH4 and
CO2 fluxes from a seasonally ice-covered lake.
Location of the study site (a) and flux footprint of the
flux tower in summer averaged over all years (b). The color scale
indicates the extent of the fractional contribution from the source area to
the fluxes measured at the tower. The location of the flux tower is indicated
along with the location of the main environmental data sources.
![]()
Surface fluxes were monitored using the eddy covariance method in a subarctic
permafrost peatland undergoing thaw, a landscape with a high percentage of
pond and lake coverage. The location of the flux tower allowed for
measurements to alternate between surface fluxes from a shallow lake and from
a permafrost-free, waterlogged fen-type wetland. The overall aim of this work
was to quantify year-round carbon fluxes in a post-glacial lake, a type
widely present around the subarctic, as compared to the adjacent fen.
Specifically, the objectives were (1) to compare the seasonality of CH4
and CO2 fluxes from two contrasting subarctic ecosystems (lake and fen),
(2) to explore the possibility of identifying short-term environmental
controls on the surface–atmosphere exchange of CH4 and CO2 in a
lake, using high, sub-daily temporal resolution measurements covering all
seasons of the year, and (3) to assess and compare the annual atmospheric
carbon budget of a lake and a wetland.
Materials and methods
Study site
Stordalen Mire is a subarctic peatland complex (68∘20′ N,
19∘03′ E) with a high lake and pond coverage, located near Abisko
in northern Sweden. Mean annual temperature in the Abisko region has been
increasing and fluctuating around 0 ∘C since the 1990s as part of an
accelerated warming trend (Callaghan et al., 2010). Permafrost, which is
discontinuously present in the local mires, has been thawing at an increased
rate since the 1990s in the peatlands of the region, sometimes disappearing
completely (Åkerman and Johansson, 2008). In Stordalen Mire permafrost
thawing has led to changes in microtopography, which controls local
hydrology, which in turn leads to vegetation shifts (Christensen, 2004;
Malmer et al., 2005). This accentuates the heterogeneity of a landscape
comprised of elevated palsas with permafrost, thawing lawns with thermokarst
ponds, and permafrost-free, water-saturated fens. The mire is bordered by
post-glacial lakes on its western, northern, and eastern edges. This study
focuses on the lake Villasjön (Fig. 1) and the adjacent fen to the west
of the tower. According to Olefeldt and Roulet (2012), the two ecosystems are
hydrologically connected with a directional flow from the lake to the fen.
The wetter fen areas have expanded as a result of permafrost thaw over the
past decades (Johansson et al., 2006). The water table of the fen is at or
above the surface throughout the year. The dominant vegetation species are
vascular plants Carex rostrata and Eriophorium angustifolium. Villasjön is the largest (0.17 km2) lake of the
15 km2 wide Stordalen catchment containing 27 lakes (Lundin et al.,
2013). It has mean and maximum depths of 0.7 and 1.3 m, respectively
(Jackowicz-Korczyński et al., 2010; Wik et al., 2013). The upstream
catchment of Villasjön is dominated by birch forest (Olefeldt and Roulet,
2012), while its western shore is bordered by thawing palsas. During
snowmelt, there is a small surface inflow feeding Villasjön in the east
(Wik et al., 2013). The lake usually freezes to the bottom in winter. The DOC
concentration in the lake water was measured to be 8.1 mg L-1 in 2008
(Olefeldt and Roulet, 2012). Aquatic vegetation is present in the lake, on
its bottom as algae, as submerged plants within its southern arm, and in a
low density of emergent macrophytes at its shores.
Eddy covariance measurements
Measurement setup
Between June 2012 and December 2014, the surface–atmosphere exchange of
CH4, CO2, and latent heat and sensible heat was monitored nearly
continuously with an eddy covariance setup located at the shore of
Villasjön (Fig. 1). Data were logged on a CR1000 (Campbell Scientific,
Inc., UT, USA) until May 2013; from June 2013 it was replaced with a CR3000
(Campbell Scientific, Inc., UT, USA). The 2.92 m high mast was equipped with
a 3-D sonic anemometer (R3-50, Gill Instruments Ltd.) sampling wind
components and sonic temperature at 10 Hz. Throughout the study period,
ambient molar densities of CO2 and H2O were sampled at 10 Hz with
an open path infrared gas analyzer (IRGA), model LI7500 (LICOR Environment,
NE, USA), mounted on the mast at 2.50 m height. Following lightning that hit
the electric grid in Stordalen Mire on 27 July 2013, the initial LI7500
ceased functioning and was replaced on 1 October 2013 by a different
instrument of the same model.
From June 2012 to May 2013, the ambient CH4 mole fraction was sampled at
10 Hz with a closed path Fast Greenhouse Gas Analyzer (FGGA, Los Gatos
Research, CA, USA) in air that was taken from a gas inlet located at 2.50 m
height on the mast through a 6 mm inner diameter polyethylene (PE) tube
using a dry scroll pump (Varian TriScroll 300). The efficient flow rate in
the 95 m long sampling line was 16 L min-1, ensuring the maintenance
of turbulent conditions (Reynolds number ca. 4025). On 5 June 2013, the tube
was replaced with an 8 mm inner diameter PE tube, which changed the flow
rate in the sampling line to 23.88 L min-1 (Reynolds number ca. 4099).
On 4 August 2013, the closed path CH4 system was renewed: the IRGA was
changed to a FGGA model 911-010 (Los Gatos Research, CA, USA). Due to
instrumental maintenance, the IRGA was offline between February and
March 2014 and replaced on 24 March 2014 by the previous FGGA. From
August 2013 to December 2014, due to a failure in the electronic connection
between the gas analyzer and the data logger, the raw CH4 data measured
by the FGGA were pre-processed in order to align the time stamp and frequency
of the gas analyzer recordings with the time stamp of the logger sampling the
wind data. The synchronization procedure was quality-checked after flux
computation (Supplement S2). Each change in the CH4 measurement setup
was taken into account in the flux calculation.
Flux calculation and quality check
CO2, CH4, sensible and latent heat fluxes, as well as atmospheric
turbulence quantities were calculated and output as 30 min averages using
the EddyPro version 5.2 open-source software (hosted by LICOR Environment,
USA). Processing of the 10 Hz raw data followed standard eddy covariance
procedures (Aubinet et al., 2012; Lee et al., 2005); methodological choices
differed partly for CH4 and CO2 flux processing, as detailed in
Appendix A.
The 30 min averaged fluxes were quality-checked to detect measurement errors
and to ensure the fulfillment of theoretical assumptions for the application
of the eddy covariance method. Spikes were present in the CO2 flux time
series, which can be due to weather conditions, fast changes in the
atmosphere's turbulent conditions, or faulty instrumentation. Outliers in the
CO2 flux dataset were detected using the median of absolute deviation
from the median (MAD) as described in Papale et al. (2006) using a threshold
z=4 and with no distinction between day and night. Additionally, CO2
flux averaging periods were rejected when the number of spikes per half hour
was > 100 (Mammarella et al., 2015) and when skewness and
kurtosis were outside the [-2,2] and [1,8] ranges (Vickers and Mahrt,
1997), respectively. Required flux stationarity (FST) within the 30 min flux
averaging period was ensured by rejecting each flux value when the FST
criterion as defined by Foken and Wichura (1996) was above 0.3, a strict
criterion that was chosen, given the challenging footprint on the lake side,
to ensure that fluxes represented the surface of interest. The fen part of
the CO2 flux dataset was additionally filtered for poorly developed
turbulence (Mauder and Foken, 2006) and u*-filtered (Papale et al.,
2006) with a threshold determined to be 0.1 m s-1 using the online
tool available at
https://www.bgc-jena.mpg.de/bgi/index.php/Services/REddyProcWeb
(Reichstein et al., 2005). The impact of self-heating on the open path
CO2 flux measurements (Burba et al., 2008) is a potential issue when
using the LI7500 model, especially for low-flux environments. It could not be
correctly quantified in this study and was thus not applied to avoid a
potentially large systematic error. A rough estimate of the correction
indicated that it could increase the magnitude of the CO2 fluxes by
0.28 µmol m-2 s-1 on average.
Quality check and screening of the CH4 flux time series included
rejection of averaging periods when the number of spikes per half hour was
> 100. Fluxes were rejected when skewness and kurtosis were
outside the pre-cited range, in winter only, due to the pulse character of
CH4 emissions during non-winter periods that could be misinterpreted as
faulty raw data due to a high skewness (Jammet et al., 2015). Flux
stationarity within the averaging time period was also ensured by rejecting
values when the FST criterion was above 0.3. Additionally, flux values for
which a time lag was not found within the plausibility time lag window were
rejected (Eugster et al., 2011; Wille et al., 2008). Gaps were initially
present in the flux dataset due to instrument malfunctioning, particularly in
winter, and power cuts (Fig. S6 in the Supplement). After quality check and
filtering, in total 13 474 CH4 fluxes and 11 629 CO2 fluxes were
available for further analysis (Table S1, Fig. S6 in the Supplement). The
data rejection rate was high; eddy covariance studies in lake environments
usually report high rejection rates from quality-check routines (Jonsson et
al., 2008; Mammarella et al., 2015; Nordbo et al., 2011).
Flux footprint and partitioning between lake and wetland
As in Jammet et al. (2015), the bidirectional wind pattern at the Stordalen
Mire was used to partition the flux dataset into two main wind sectors which
crossed the two different studied ecosystems. To ensure a clear distinction
between the lake and the fen fluxes and the homogeneity of each surface, the
lake sector was conservatively defined as 20–135∘, and the fen
sector as 210–330∘ (0∘ = 360∘ is true north).
Over the study period, 46.2 % of the measured fluxes originated from the
lake sector, 45.7 % of the measured fluxes originated from the fen
sector, and the residual 8.1 % of the measured flux data was excluded,
for being of lower wind speed and originating from mixed sources. Each season
was nevertheless well represented in the flux dataset of each ecosystem (fen
and lake) thanks to regular shifts in wind direction.
The flux footprint was calculated using a 2-D model developed by Kljun et
al. (2015). Inputs to the model include turbulence quantities measured at the
tower (friction velocity (u*), standard deviation of cross-wind velocity
(σv), Obukhov length, horizontal wind speed), height of the
boundary layer (derived from the Era-Interim reanalysis product; Dee et al.,
2011), the measurement height and the surface roughness length which was
separately estimated for the lake (0.001 m), and the fen sectors (0.002 m
in winter to a maximum of 0.10 m in summer, accounting for vegetation
growth). Footprints were calculated for each 30 min time step where the
required data were available, and averaged for the periods of interest
(summer and winter).
According to the footprint model, most of the flux measured at the tower
(peak fetch) originated from a distance of 38 m on the fen sector and 73 m
on the lake sector, on average during the ice-free season. The lake surface
was within the cumulative 80 % of flux footprint during the ice-free
seasons (Fig. 1). In winter, the footprint model revealed that the 80 %
cumulative footprint, on the lake side, included part of the land in the
middle of the lake (Fig. S1 in the Supplement). This is likely due to the
lower roughness length (snow cover), but also to more stable atmospheric
conditions: the sensible heat flux H is negative during most of the winter
and the atmospheric stability parameter indicated more stable conditions in
winter as compared to summer conditions. Thus, to avoid large contamination
of the land respiration in the winter lake fluxes due to an extended
footprint over the lake shores, the winter fluxes from the lake sector were
only kept for further analysis when the standard deviation of lateral wind
velocity σv ≤ 1 m s-1 (Forbrich et al., 2011)
in order to limit lateral contamination of CO2 fluxes into the footprint
area of interest. This removed an additional 3.7 % of the lake CO2
flux dataset.
Ancillary measurements
Supporting environmental variables were measured at 1 Hz and averaged and
logged every half hour. Temperature probes (T107, Campbell Scientific, Inc.,
UT, USA) were installed in the peat 5 m south-west of the tower at 5, 10,
25, and 50 cm depths. Net radiation at the fen surface was recorded with an
REBS Net Radiometer, model Q7.1 (Campbell Scientific, Inc., UT, USA). Air
temperature used in this study was measured at 2 m height on a mast located
in the middle of the lake (Fig. 1) with a CS215 probe (Campbell Scientific,
Inc., UT, USA). At the same mast, radiation components were measured with a
CNR4 Net Radiometer (Kipp & Zonen, the Netherlands) from which net
radiation at the lake surface was computed. Water temperature in the center
of the lake (Fig. 1) was measured at 10, 30, 50, and 100 cm depths with
intercalibrated HOBO Water Temp Pro v2 loggers (Onset Computer Corporation,
MA, USA). The loggers are suspended on a nylon line from a mooring float,
which stays at the surface throughout the year. The string assembly was
designed so that the bottom sensor at 100 cm depth is in the surface
sediment. Lake surface albedo was computed daily as the midday ratio (10:00
to 13:00 UTC+1) of the shortwave
radiation components, and used in combination with temperature data to
delimit the ice-cover seasons. Air pressure and precipitation were measured
at a weather station 640 m south of the eddy covariance mast.
Gap filling of CH<inline-formula><mml:math id="M122" display="inline"><mml:msub><mml:mi/><mml:mn mathvariant="normal">4</mml:mn></mml:msub></mml:math></inline-formula> and CO<inline-formula><mml:math id="M123" display="inline"><mml:msub><mml:mi/><mml:mn mathvariant="normal">2</mml:mn></mml:msub></mml:math></inline-formula> flux time series
The estimation of annual carbon exchange budgets required filling of the gaps
in the CH4 and CO2 flux time series. When working with highly
skewed flux datasets such as CH4 emissions from lakes (Fig. C1),
integrating the mean flux over the whole time period may lead to important
overestimation. There is as yet no published account of gap filling CH4
fluxes measured with eddy covariance. In the present study, gap filling of
the CH4 flux time series was performed separately on the lake and fen
flux datasets, using artificial neural networks (ANNs) (Moffat et al., 2010;
Papale and Valentini, 2003). ANNs are multivariate, nonlinear regression
models that are fully empirical: the observational data are used to constrain
the model's numerical relationship between the inputs (independent variables,
i.e., environmental drivers) and outputs (dependent variables, i.e., fluxes)
(Moffat et al., 2010). ANNs have been tested and successfully used to
estimate missing values and gap fill CO2 flux time series measured with
eddy covariance in forests (Moffat et al., 2007; Papale and Valentini, 2003)
or in urban terrain (Järvi et al., 2012) and to gap fill CH4
emissions in wetlands (Dengel et al., 2013). To our knowledge, we present the
first attempt at using ANNs to gap fill eddy covariance measurements of
CH4 emissions from a lake.
Three ANN models were built separately on CH4 emissions from the fen,
CH4 emissions from the lake, and CO2 emissions from the fen, on the
hourly scale. Model development followed workflows introduced by Papale and
Valentini (2003), Moffat et al. (2007), and Dengel et al. (2013).
Environmental variables to be included as inputs were selected according to
their physiological relevance to the production and transport of CH4 and
CO2 from the surface to the atmosphere, as reported in the literature.
The relevance of the drivers was confirmed by correlation analysis. Input
variables used for each model are reported in Table B1. A further detailed
description of ANN model development can be found in Supplement S1. Gaps in
the measured CH4 and CO2 flux time series were replaced by
predicted values, and the annual sums were computed by integrating the hourly
flux values over time. The performance of the ANN models was assessed by
comparing the predicted values with original observed values over the whole
dataset (Appendix B).
The ANN method did not perform well on the lake CO2 flux time series,
which comprised a large number of gaps and a high signal-to-noise ratio. ANN
results were thus excluded for the lake CO2 fluxes to avoid introducing
a high and unnecessary uncertainty. Instead, the seasonal lake CO2
exchange was computed by multiplying the mean flux rate during each season by
the number of days. Considering the normal distribution of the CO2 lake
fluxes (Fig. C1), this method was considered an acceptable way of filling
missing values.
Uncertainty analysis
The total random error is a composite of errors associated with instrument
noise, the stochastic nature of turbulence, the instrument precision, and the
variation of the flux footprint (Moncrieff et al., 1996). The relative random
error increases with the magnitude of the flux (Richardson et al., 2006),
while over time it decreases with increasing size of the dataset because of
its random nature (Moncrieff et al., 1996). It is therefore negligible when
propagated over annual sums, but can be important for single flux values. The
random error was calculated as the sampling error for each flux value in the
flux calculation software using the method of Finkelstein and Sims (2001).
The mean random error of the quality-checked measured CH4 fluxes was
2.9 ± 4.3 (mean ± SD) nmol m-2 s-1 at the lake and
4.7 ± 3.8 nmol m-2 s-1 at the fen, which corresponds to
7.6 and 6 % of the overall mean measured fluxes, respectively. This is in
the lower end of ranges reported for CO2 and energy fluxes. The mean
random error of the individual measured lake CH4 fluxes was highest in
winter (24 %) and lowest during the thaw season (11 %) when the
largest fluxes were measured. The mean random error of the quality-checked
measured CO2 fluxes was
0.20 ± 0.57 µmol m-2 s-1 at the lake and
0.32 ± 0.48 µmol m-2 s-1 at the fen, equivalent
to 31 and 13 %, respectively, of the overall mean absolute measured flux,
indicating that our setup was able to measure both CH4 and CO2
fluxes observable at our site. These values are comparable to what has been
reported in vegetated (Finkelstein and Sims, 2001), urban (Järvi et al.,
2012), boreal lake (Mammarella et al., 2015), and other typical eddy
covariance sites (Rannik et al., 2016). In spring the random error of
CO2 fluxes at the lake was 18 % of the measured flux; however,
during the ice-free season it was above the absolute measured flux in
20 % of the cases. Thus air–lake CO2 exchange rates during the
summer were low and sometimes close to the detection limit.
The random error of the fluxes modeled with ANN (Appendix B) was on average
4 nmol m-2 s-1 for fen CH4 fluxes,
0.23 µmol m-2 s-1 for fen CO2 fluxes, and
11 nmol m-2 s-1 for lake CH4 fluxes. These errors were
small when propagated onto seasonal and annual sums using the random error
propagation principle (Moncrieff et al., 1996). The systematic bias in the
annual flux due to the gap filling method, the bias error (BE), was
calculated on the seasonal and annual flux sums as in Moffat et al. (2007) as
the sum of the difference between the predicted values pi and the
observed values oi:
BE=1N∑(pi-oi),
where N is the number of gap filled values in the flux time series. The
bias error adds up over time. It was multiplied by the number of gap filled
values to obtain a total seasonal and annual offset (Moffat et al., 2007).
The offset can be the largest source of uncertainty in the computation of
annual budgets. The systematic offset due to gap filling was larger for
annual CO2 fen fluxes (Table 3), likely because of the higher noise in
the measurements and higher number of gaps during the second year. For total
lake CO2 fluxes, the annual BE was calculated in the same way, using the
seasonal mean as the predicted values. The resulting offset due to gap
filling was close to zero (Table 3), because flux values were normally
distributed, which confirms that using the mean did not induce a significant
bias for this particular dataset.
Daily mean of air, peat, and surface sediment
temperature (a); cumulative daily precipitation (b); daily
mean net radiation input at the fen and the lake surfaces (c);
albedo of the lake surface (d); daily temperature gradient in the
lake water column defined as (Tw,10cm -
Tw,100cm) divided by the depth difference Δz
(e); water temperature profile in the lake center derived from
continuous daily temperature measurements at depths 10 cm, 30 cm, 50 cm,
and 100 cm (= bottom) (f). The shaded areas indicate the periods
of ice cover at the lake. PN stands for polar night, i.e., near zero solar
input. The black arrows indicate the estimated time of complete lake overturn
after ice-out.
![]()
Definition of seasons
Observations started in June 2012 and were sustained nearly continuously
until December 2014. We adopted a lake-centric definition of seasons based on
the lake ice phenology. A full year was defined from 1 June to 31 May of the
next year, so that an entire ice-cover season was included in a given year.
This keeps the connected thaw period and previous ice-cover season within the
same year. The year was further divided into an ice-cover season (winter), an
ice-free season (summer and fall), and a thaw season (spring). The thaw
season represents a transitional period during which the snow cover and lake
ice melt. It is separated from summer, since the hydrological and
biogeochemical dynamics in seasonally ice-covered lakes differ from the rest
of the open-water season.
Climatic conditions per season: mean air temperature measured at the
Abisko Scientific Research Station, peat temperature at 10 cm depth in the
fen, surface sediment temperature in the lake (1 m depth), and total
incoming solar radiation (Si). Season delimitations (see text for a
definition) are reported for each year.
Season
Year
Dates
Length
Air T
Peat T at
Surf. sed.
Total Si
(days)
(∘C)
10 cm (∘C)
T (∘C)
(103 Wm-2)
Ice-free
2012
1 June–14 Oct.
136
7.8
8.9
10.2
808
2013
26 May–15 Oct
143
10.2
9.9
12.0
902
2014
2 Jun–9 Oct
130
10.4
10.5
12.8
1019
Ice-cover
2012–2013
15 Oct–14 Apr
182
-7.9
0.0
0.7
279
2013–2014
16 Oct–10 Apr
177
-6.5
0.2
0.8
239
Thaw
2013
15 Apr–25 May
41
4.3
0.3
1.9
250
2014
11 Apr–1 Jun
52
2.2
1
1.9
502
Annual
2012–2013
1 Jun–31 May
365
-0.3
3.5
4.7
1424
2013–2014
1 Jun–31 May
365
0.9
4
5.1
1543
Daily air temperature was used to define seasons. The start of the ice-cover
season was defined as the start of the lake freeze-up, i.e., the first day on
which daily mean air temperature is below zero for 3 consecutive days.
Further, this date coincided each year (2012 to 2014) with the formation of
thermal stratification in the water column (Fig. 2d) due to a rise in bottom
water temperature right after the first day of freezing (Fig. 2a, blue line).
This indicates the inhibition of direct heat exchange between the lake and
the atmosphere when ice forms at the lake surface. In the first year (2012),
the first day of the ice-cover season was confirmed by visual observation of
ice over the whole lake surface, while in the following years, these combined
temperature observations were used to define the start of ice cover.
The end of the ice-cover season was defined as the start of thaw, i.e., the
first date on which daily mean air temperature rose above 0 ∘C for
at least 3 consecutive days. This date preceded by 1 to 2 days the
temperature rise to 0 ∘C in the surface (10 cm) of the lake ice.
Further, the start of the thaw season could be confirmed by the increase in
mean daily net radiation to net positive values at the lake and the fen
surfaces (Fig. 2c) and by a decrease in lake albedo (Fig. 2d) until
open-water values were reached (mean = 0.05). We thereby defined two
complete ice-cover seasons during the study period: from 15 October 2012 to
14 April 2013 (182 days) and from 16 October 2013 to 10 April 2014
(177 days). In 2014, freeze-up of the lake occurred on 10 October. The
ice-cover seasons were systematically characterized by negative daily energy
input at the lake and at the fen (Fig. 2). Thus our season definition based
on daily air temperature was robust and reproducible each year.
The end of spring (i.e., the thaw season) and the beginning of the ice-free
season were defined as the first date with a daily temperature gradient in
the water column close to 0 ∘C m-1 after the spring overturn,
i.e., when the lake enters its isothermal conditions after complete ice thaw.
The dataset of this study covers 2 complete years plus an additional ice-free
season, i.e., three ice-free seasons, two ice-cover seasons, and two thaw
seasons. Season dates and
lengths are summarized in Table 1.
Results
Environmental conditions and lake climatology
Mean annual (June to May) air temperature measured at the Abisko Scientific
Research Station was -0.3 ∘C in 2012–2013 and 0.9 ∘C in
2013–2014. The latter was significantly above the long-term average
(1913–2014) of -0.5 ± 1 ∘C (mean ± SD). The ice-free
season was warmest in 2014 and coolest in 2012 (Table 1). This difference
between years was reflected in both the mean daily peat temperature measured
at 10 cm depth and the mean daily lake water temperatures, which were
warmest in 2014 (Fig. 2; Table 1). The mean air temperature observed over our
study period correlated with total incoming solar radiation, both annually
and during the ice-free periods (Table 1). Winter 2013–2014 was on average
2.1 ∘C warmer than the previous one and the ice-cover season was 5
days shorter due to an earlier thaw start. The thaw period started earlier in
2014 but lasted longer (Fig. 2d, Table 1); complete lake overturn (isothermal
water column) following ice thaw occurred 7 days later than in 2013.
The development pattern of thermal stratification along lake depth at
freeze-up and its breakdown in spring was similar in both years (Fig. 2d). In
both spring 2013 and 2014, the lake overturn occurred after the development
of strong thermal stratification in the lake during thaw (Fig. 2e, f). The
temperature at the bottom of the lake was up to 4 ∘C warmer than the
surface in spring 2013 and up to 6 ∘C warmer in 2014, likely
indicating the penetration of solar radiation through thinning ice before
complete ice-out and full water mixing. The thermal structure at the
beginning of freeze-up, as well as during the period of ice thaw preceding
spring overturn in 2013, has been previously described in detail for this
site in a winter-focused study (Jammet et al., 2015).
During the ice-free, summer season, there was slight to no thermal
stratification in the water column (Fig. 2d). The shallow lake water column
reacted quickly to temperature changes (Fig. 2a, d) and the dark bottom
warmed up quickly since solar radiation can reach the sediment surface. Thus,
the lake had a polymictic behavior; i.e., it was regularly mixing to the
bottom during the ice-free season, which ensures isothermal conditions in the
water column throughout the summer. Water temperature reached a maximum of
23.8 ∘C in July 2014. Conversely, a strong temperature gradient
formed during winter (Fig. 2f). In both winters the surface sediment
temperature dropped below 0 ∘C by February, with minima of
-1.9 ∘C in March 2013 and -1.6 ∘C in February 2014.
This suggests that the water column froze to the bottom. The temperatures in
the peat soil and at the surface sediment in the lake were de-coupled from
air temperature (Fig. 2a), showing the hindrance of heat exchange with the
atmosphere due to the presence of ice and snow at the surface.
Statistical exploration of the lake flux dataset: Spearman's rank
correlation coefficient (ρ) shows the degree of association between
half-hourly lake CH4 flux and lake CO2 flux with potential drivers
of variability. All data were grouped per season. See Table 1 for the
limitation of the seasons. Lake fluxes are filtered in winter for high
standard deviation of lateral wind speed.
Lake CH4 flux (Spearman's ρ)
Lake CO2 flux (Spearman's ρ)
Annual
Ice-free
Thaw
Ice-cover
Annual
Ice-free
Thaw
Ice-cover
Air T
0.36***
0.40***
0.26***
0.14**
-0.21***
-0.22***
0.38***
0.16***
T water surface
0.27***
0.48***
0.37***
0.18*
-0.21***
-0.21***
0.27***
0.09***
T bottom
0.24***
0.49***
0.49***
0.10*
-0.26***
-0.21***
0.40***
-0.10***
Wind speed
0.36***
0.32***
0.40***
0.27***
0.26***
0.13***
0.38***
0.36***
H flux
-0.15***
0.12***
-0.18***
-0.12*
-0.62***
-0.67***
-0.38***
-0.50***
LE flux
0.24***
0.38***
0.16***
0.08
-0.29***
-0.49***
0.05
-0.14***
Solar radiation
0.27***
0.13***
-0.02
-0.11*
-0.09***
-0.52***
0.02
-0.22***
CO2 flux
0.41***
-0.13***
0.67***
0.21**
–
–
–
–
* p-value < 0.1;
** p-value < 0.01;
*** p-value < 0.001.
Measured CH4 and CO2 fluxes at the fen (a, b) and
the lake (c, d) over the full study period. Light grey dots are
half-hourly values and black dots show a 5-day running mean. Note on
panel (c): nine flux values above 800 nmol m-2 s-1 are
not displayed for visibility.
![]()
Seasonal and annual sums of CO2 and CH4 fluxes after gap
filling, in g C m-2 yr-1. Sum BE is the bias due to the gap
filling model (Eq. 1), scaled to annual flux units and multiplied by the
number of gaps in the flux dataset over the season or year. The sum of
CO2 fluxes from the lake is only reported for the first year.
Fen CH4 flux
Lake CH4 flux
Fen CO2 flux
Lake CO2 flux
Season
Year
Total flux
Sum BE
Total flux
Sum BE
Total flux
Sum BE
Total flux
Sum BE
Ice-free
2012
14.8
0.1
2.5
-0.1
-179.4
-14.8
-24.4
< 0.001
2013
16.2
-0.1
2.8
-0.1
-201.3
13.2
nc*
nc
2014
16.8
0.09
3.0
-0.5
-239.6
-5.1
nc
nc
Ice-cover
2012–2013
4.1
-0.2
0.3
-0.3
109.6
-8.8
12.4
< 0.001
2013–2014
4.2
0.7
0.4
-0.4
115.5
20.7
nc
nc
Thaw
2013
1.4
-0.1
2.5
-0.2
11.3
-0.4
33.3
< 0.001
2014
1.7
-0.07
1.3
0.5
11.7
1.1
nc
nc
Annual
2012–2013
20.3
0.2
5.3
-1.9
-58.5
-26.1
21.5
< 0.001
2013–2014
22.1
-0.1
4.4
0.6
-74.1
97.4
nc
nc
Annual
Average ± SD
21.2 ± 1.3
-0.2
4.9 ± 0.6
-0.7
-66.3 ± 11
35.7
21.5
< 0.001
∗ nc: not computed.
Measured flux rates of CH4 (a) and
CO2 (b) per season and per ecosystem. The central line of the
boxplots shows the median, box edges show the 25th and 75th percentiles, and
whiskers show the 5th and 95th percentiles. The black dots indicate the mean
flux rate. Outliers are not displayed.
![]()
Year-round CH<inline-formula><mml:math id="M362" display="inline"><mml:msub><mml:mi/><mml:mn mathvariant="normal">4</mml:mn></mml:msub></mml:math></inline-formula> and CO<inline-formula><mml:math id="M363" display="inline"><mml:msub><mml:mi/><mml:mn mathvariant="normal">2</mml:mn></mml:msub></mml:math></inline-formula> fluxes
CH4 emissions from the lake showed a highly skewed distribution
(Fig. C1a); there was a large difference between the overall mean
(40 nmol m-2 s-1) and overall median
(12 nmol m-2 s-1). The seasonal flux pattern was characterized
by low background emissions and occasional, large degassing events (Fig. 3)
with 25 % of measured data above 111 nmol m-2 s-1 in the
thaw period, and 5 % of measured data above 75 nmol m-2 s-1
within the ice-free season (Fig. 4). Over the full measurement period, summer
emission rates averaged to 26 nmol m-2 s-1 (median
12 nmol m-2 s-1), spring emission rates to
84 nmol m-2 s-1 (median 33 nmol m-2 s-1), and
winter emissions to 2.8 nmol m-2 s-1 (median
3.0 nmol m-2 s-1). The mean and median of the winter emissions
were not significantly different from the mean random error of the fluxes.
CO2 fluxes were close to normally distributed at the lake (Fig. C1b) and
the overall mean rate was 0.22 µmol CO2 m-2 s-1
(median 0.18 µmol m-2 s-1). There was a distinctive
CO2 outgassing at the time of water overturn during the spring of both
2013 and 2014 (Fig. 3d). The mean measured CO2 exchange at the lake was
significantly negative during the ice-free seasons (one sample t-test,
p < 0.001), -0.14 µmol m-2 s-1,
indicating a low net uptake of CO2 (Fig. 4). Negative flux rates started
right after the spring CO2 outgassing in 2013 and 2014 (Fig. 3d). The
highest CO2 uptake rates were observed during the summer of 2014, which
was the warmest summer of the study period, with the highest solar radiation
input (Table 1). In fall 2014, a burst of CO2 was measured at the lake,
which was not present in previous years (Fig. 3d). This fall burst was not
observed in the CH4 flux measurements. Wintertime CO2 emissions
from the lake were significantly above the flux random error and
significantly positive (Figs. 3d, 4). There was an inter-annual variability
in the magnitude of the CH4 spring degassing from the lake between the 2
years (Fig. 3c). In contrast, mean CO2 degassing was higher during the
second thaw season, 0.78 µmol CO2 m-2 s-1 (median
0.47 µmol CO2 m-2 s-1) in 2013 and
0.99 µmol CO2 m-2 s-1 (median
0.75 µmol CO2 m-2 s-1) in 2014. Both CH4 and
CO2 emissions during spring were significantly higher than during the
following ice-free season, in both years (Fig. S2 in the Supplement).
The distribution of CH4 emissions from the fen was less skewed
(Fig. C1a) and the overall measured mean was 77 nmol m-2 s-1
(median 58 nmol m-2 s-1). The highest CH4 emissions at the
fen occurred during the ice-free season (Fig. 3), with a mean rate of
110 nmol m-2 s-1 (median 108 nmol m-2 s-1).
Sustained CH4 emissions were measured at the fen throughout the winter
(Fig. 3), with a mean rate of 25 nmol m-2 s-1 (median
25 nmol m-2 s-1). Flux rates during the snowmelt and thaw season
averaged to 35 nmol m-2 s-1 (median
33 nmol m-2 s-1). CO2 fluxes from the fen averaged to
-1.3 µmol m-2 s-1 (median
0.2 µmol m-2 s-1) over the full period, and the mean
rate was -2.6 µmol m-2 s-1 during the ice-free
seasons. CO2 respiration was sustained throughout winter (Fig. 3b) at a
mean rate of 0.8 µmol m-2 s-1, and the release of
CO2 during the melt season was low (Fig. 4).
Averaged seasonality of CO2 and CH4 fluxes in both
ecosystems. Dots show daily means across the whole measurement period; lines
are a smoothing filter of the daily mean with a 30-day window. Dashed lines
show the standard deviation around the daily means, smoothed with a 30-day
window.
![]()
CH4 emissions at the lake during the ice-free seasons of 2012,
2013, and 2014, from day 152 to day 290. Grey dots are half-hourly eddy
covariance observations, open black dots are daily means of the eddy
covariance fluxes, and red dots show spatially averaged daily CH4
ebullition measured in the lake with bubble traps. Atmospheric pressure
(black line), and water temperature at depths 10 cm (red line) and 100 cm
(sediment surface, blue line) are also shown.
![]()
The average annual seasonality of CH4 and CO2 fluxes in both
ecosystems is shown in Fig. 5. The lake dominated CH4 and CO2
emissions during spring. During the ice-free seasons, by contrast, the lake
was a lower emitter of CH4 than the fen per unit area, and its CO2
exchange was close to neutral with a small uptake. There was a slight
seasonality in CH4 lake emissions during summer but, annually, CH4
and CO2 lake fluxes peaked in spring (Fig. 5). CH4 emissions from
the fen peaked in August and net CO2 exchange peaked in July (Fig. 5).
The emission of both gases occurred at lower rates but continuously in
winter.
Variability of lake–air carbon exchange within seasons
The thaw and ice-free periods are different in terms of flux dynamics. The
dataset was therefore separated into seasons to explore controls on the lake
fluxes. During the ice-free seasons, half-hourly CH4 fluxes at the lake
were characterized by degassing events that coincided with drops in
atmospheric pressure (Fig. 6). Daily EC flux data were compared with
spatially averaged, daily ebullition fluxes measured in the lake with
ebullition traps located nearby or within the tower footprint. The degassing
events measured with the eddy covariance system coincided in timing and in
magnitude with the daily ebullition fluxes measured with the ebullition traps
(Fig. 6).
Due to the high skewness of the CH4 lake flux dataset (Fig. C1) and the
presence of outliers in both CO2 and CH4 flux datasets from the
lake, we used Spearman's ρ coefficient, which is a statistical measure
of association that is robust to outliers and applicable to skewed
distribution (Kowalski, 1972), to explore bivariate associations. Among
potential flux drivers, the highest correlation of ice-free CH4
emissions was found with surface sediment temperature (ρ=0.49,
Table 2). Degassing events occurred often after an increase in surface
sediment temperature (Fig. 6). When averaging all 30 min CH4 fluxes
during the three ice-free seasons per bins of 1 ∘C, an exponential
regression between lake CH4 fluxes and surface sediment temperature
could explain 82 % of the variability in CH4 emissions (Fig. 7).
Relationship between lake CH4 emissions and surface sediment
temperature during summer, using all half-hourly CH4 flux rates averaged
by bins of 1 ∘C across the three ice-free seasons. Error bars show
the standard deviation of the CH4 flux rates around the means within
each averaging bin. The red line is a regression fit with r2 = 0.82
(p < 0.001).
![]()
Diel medians, mean and 25th–75th percentiles of CH4 fluxes,
CO2 fluxes, sensible heat flux (H) and latent heat flux (LE) measured
at the lake (a) and at the fen (b) from June to August over
the three ice-free seasons.
![]()
Diel median of the net CO2 exchange at the fen (red) and at the
lake (blue) vs. diel median solar radiation, between June and August. Each
dot is the hourly median flux of the combined three ice-free seasons. Note
that the two y axes have different scales.
![]()
Wind speed correlated best with CH4 lake emissions during the thaw
period (Table 2, ρ=0.40). The relationship was weaker during the
open-water period. We observed a few CH4 degassing events in summer that
coincided with the likely mixing of the water column following a short period
of thermal stratification (0.8 to 1 ∘C gradient between 10 and
100 cm depths, Fig. S3a, b in the Supplement), but these were not systematic
(Fig. S3c, d in the Supplement). During the ice-free season, there was a weak
anti-correlation between CO2 exchange at the lake and air and water
temperature (Table 2). Wind speed correlated weakly with CO2 lake
exchange, while there was a strong anti-correlation with sensible heat flux
at the lake surface (Table 2), and with solar radiation input (Table 2).
During the thaw seasons, half-hourly CO2 fluxes strongly correlated with
CH4 emissions (ρ=0.67, p < 0.001, Table 2) and
followed the same emission pattern at the half-hourly scale, in both 2013 and
2014 (Fig. S4 in the Supplement). This correlation was not sustained during
the open-water seasons, when the two flux datasets had a very different
short-term variability. Both CH4 and CO2 degassing in spring
positively correlated with increasing air temperature.
Evaluation of the artificial neural network models: measured vs.
predicted hourly flux values for fen CH4 fluxes (a), lake
CH4 fluxes (b), and fen CO2 fluxes (c). In red is
the 1:1 line.
![]()
The diurnal course of CO2, CH4, and turbulent energy fluxes at the
lake and the fen was calculated on hourly fluxes. Only days with more than
75 % of hourly data coverage were selected. The median flux of each hour
was then computed across all days within each year during the open-water
season, and plotted along with the 25th and 75th percentiles (Fig. 8). There
was no diel cycle visible in the fen CH4 fluxes, while net CO2
exchange at the fen surface showed a clear peak uptake at noon. Lake methane
fluxes tended to be higher fluxes in the morning hours. A systematic diel
pattern was observed in the lake CO2 fluxes during each open-water
season (Fig. 8), with a slight peak in the mornings in 2012 and 2013, while
CO2 uptake peaked in the middle of the day in 2014. The 24 h cycle of
sensible heat flux (H) at the lake peaked in the late morning (Fig. 8) at
ca. 10:00 LT (median 20 W m-2), while
being < 10 W m-2 in the afternoon. The diel CO2 pattern
at the lake also coincided with daily variation in water surface and air
temperature and was in antiphase with the diel pattern of sensible heat flux.
Latent heat flux (LE) at the lake peaked in the afternoon. At the hourly
timescale, solar radiation could explain 88 % of the diel variability in
air–lake CO2 exchange during the summer months (Fig. 9). This light
response curve resembled the one measured at the fen, although less
pronounced.
Annual atmospheric carbon budget
Performance of the ANN modeling
The ANN gap filling method was most performant on the fen dataset, achieving
an r2 of 0.88 during the training phase, and an r2 of 0.85 between
measured and predicted values over the whole dataset (expressing the capacity
of generalization of the model) with a relative mean square error (RMSE) of
23 % (Fig. 10). The ANN gap filling of lake CH4 fluxes achieved an
r2 of 0.71 in the training phase and an r2 of 0.70 between
predicted and measured fluxes on the whole dataset, with an RMSE of 51 %.
The lake model was most accurate for periods with the best data coverage in
the measured dataset (spring seasons, r2=0.77). The lower accuracy of
the model during the ice-free seasons (r2=0.47) is also due to the
pulse character of lake ebullition, which was not always reproduced by the
model, while the background seasonal trend was present.
For the fen CH4 fluxes, the model was most accurate during the ice-free
seasons when fen CH4 emissions are tightly linked to peat temperature
and least performant during the thaw periods. Unsurprisingly, the prediction
performance of the models was dependent on data coverage in the original
dataset, but also on the accuracy of the choice of environmental drivers. On
the annual scale, both fen and lake models were most performant during the
first year (June–May), which had the least amount of data loss, with an
r2 of 0.88 (RMSE 23 %) in the first year for the fen model and of
0.82 (RMSE 40 %) for the lake model between predicted and measured
fluxes.
The ANN modeling was likewise performant on the fen CO2 flux dataset,
achieving an r2 of 0.86 (RMSE 35 %) over the whole dataset between
measured and predicted values (Fig. 10). The model was most performant during
the ice-free seasons. This can be explained by a better data coverage but
also by better constrained processes during the growing season and on the
annual scale than during the snow-cover season, when additional drivers than
the one selected as model inputs may play a role in the variability of the
fluxes at the hourly scale (e.g., sensitivity to footprint changes,
turbulence, variation in snow depth, partial melt).
Annual and seasonal sums
We report seasonal emissions for all available seasons, and for lake CO2
total annual emissions for the first year only (June 2012 to May 2013), due
to the absence of a robust gap filling model for the second year when data
coverage was lower. During the ice-free seasons, total lake CO2 flux was
negative and total lake CH4 flux was positive (Table 3). Annually, the
largest contribution to total carbon exchange at the lake was during the
spring season, whereas the ice-free season was quantitatively the most
important period for the annual carbon exchange at the fen (Table 3). On
average over both years, the ice-cover season accounted for 33 % of the
fen annual carbon (CH4 + CO2) exchange per m2. During the first
year, the lake C-emissions equaled 70 % of the total net fen C-exchange.
On a carbon mass basis, CO2 exchange dominated the total carbon budget.
Total net annual carbon exchange (CH4+CO2) at the fen was
-38.2 g C m-2 in the first year and -52 g C m-2 in the
second year (average -45.1 g C m-2 yr-1), while the lake
total carbon exchange was positive at 26.7 g C m-2 yr-1 (first
year only), of which 80 % was emitted as CO2.