The research site was located in a floodplain of the Kolyma river near the city of Chersky (68.75∘ N, 161.33∘ E), in the northeastern part of Siberia, Russia. Situated south of the Arctic tree line within the floodplain of the Kolyma river, this area was subject to annual flooding in spring, leading to a common vegetation structure that was characterized by sedges, shrubs, and sphagnum mosses (Merbold et al., 2009) and defined as a wet tussock tundra. The mineral soil of the permafrost and active layer was influenced by the sediment transport of the river and was classified as silty loam (Corradi et al., 2005). The transition period between the snow-covered and snow-free seasons was strongly influenced by the flooding in late spring, which in most years led to standing water on the site as a result of snowmelt and rising water levels in the nearby river. Development of the vegetation in the flooded sections was therefore usually delayed compared to dry areas where snow melt is followed by rapid greening. Continuous daylight and continental air masses (which are warm, dry, and from the S to SE) characterized the summer (i.e., June–August). Overall, in 2014, the mean annual temperature was -10 ∘C, with daily means ranging from -49 ∘C in January to 22 ∘C in July. The mean annual precipitation was about 172 mm, with nearly 50 % of this precipitation occurring during the summer. As a result of limited light and low temperatures, the growing season was mainly restricted to 3 months of the year, from mid-June to the end of August.

The experimental concept of the disturbance was designed to modify the local water regime and soil moisture conditions. The manipulation consisted of a 200 m diameter drainage ditch that connected to the nearby river through a channel of ∼ 300 m length (Merbold et al., 2009). This drainage system was installed in autumn 2004 and is still in place today, altering the water table depth and thus significantly modifying biogeochemical and biogeophysical site characteristics during the last decade. For the recent observation program, two observation sites were chosen: one within the drainage ring (thus covering the areas affected by lowered groundwater levels) and another serving as a control site (close by but far enough away to represent natural conditions unaffected by the disturbance). Since both sites had to be close enough to share a power supply, the two observation sites were ∼ 600 m apart (Fig. 1).

To monitor ecosystem-scale CO2 fluxes, two eddy-covariance systems were installed in July 2013. One tower (Tower 1, 68.61∘ N and 161.34∘ E) was placed within the drainage area (Fig. 1, blue shaded area), so that its footprint primarily covered areas affected by reduced groundwater levels. The control tower (Tower 2, 68.62∘ N and 161.35∘ E) represented natural conditions (Fig. 1). Data acquisition started 13 July 2013 for Tower 1 and 16 July 2013 for Tower 2.

The ground under both towers had an elevation of approximately 6 m a.s.l., with an average vegetation height of 0.7 m in the surrounding area during peak growing season. Each tower was equipped with a heated sonic anemometer (uSonic-3 Scientific, by Metek GmbH, Elmshorn, DE) that monitored the three-dimensional wind field and the sonic temperature at the top of the towers (at heights of 4.91 and 5.11 m for Tower 1 and 2, respectively). During the first summer season of the experiment, the CO2 and H2O flux densities were measured exclusively using open-path sensors (LI-7500 by LI-COR Biosciences Inc., NE, USA). These gas analyzers were placed next to the sonic anemometer, with a sensor separation of 0.38 m at both towers. In 2014, new closed-path gas analyzers (FGGA by Los Gatos Research Inc., CA, USA) for monitoring GHG mixing ratios were added to the setup. These systems consisted of an inlet placed next to the sonic anemometer (vertical sensor separation: 0.30 m), a sampling line (heated and insulated Eaton Synflex decabon with 6.2 mm inner diameter and a length of 16.0 and 12.8 m for Tower 1 and 2, respectively), and an external vacuum pump (N940, KNF). The flow rate of 13 L min-1 (ambient pressure) translated into a replacement of sample air in the measurement cell at a frequency of ∼ 2–2.5 Hz. Both open- and closed-path analyzers had been running in parallel at Tower 1 since installation, while the open-path analyzer on Tower 2 was disassembled in July 2014. An overview of the active times for the different gas analyzing systems is given in Fig. 2. Additional instrumentation to monitor environmental variables is listed in Table 1.

The observed meteorological data (Table 1) were collected at 10 s intervals and stored on a data logger (CR3000, Campbell Scientific) as averages over 10 min periods. The post-processing quality control scheme of the meteorological data included a test for failure of the power supply, a test that checks the range of parameters within a certain time window, a flat lining test, and a spike test, as well as tests for identifying malfunctioning sensors and consistency limits. The final dataset was averaged to 30 min intervals.

For the eddy-covariance towers, the data were collected at a rate of 20 Hz with analogue output for the gas analyzers. All analog signals were first transmitted to the sonic anemometer, where they were digitized and sent to the site computer jointly with the wind signal. Data acquisition was handled through the software package EDDYMEAS (Kolle and Rebmann, 2007) on a local computer at the field site. The processing of the high-frequency eddy-covariance data was based on the software tool TK3 (Foken and Mauder, 2015), which implements the following major components: for both open- and closed-path gas analyzing systems, the flux processing includes (1) a 2-D coordinate rotation of the wind field, (2) a cross-wind correction (Liu et al., 2001), and (3) a correction for loss in the high-frequency range (Moore, 1986). For the closed-path systems, the greenhouse gas mixing ratios originally recorded as wet mole fractions were converted to dry mole fractions before processing by the TK3 software. To correct losses in the high-frequency range that occur when gases are transferred to the closed-path analyzers through inlet tubes at flow rates much lower than the sampling frequency, cutoff frequencies of 2.5 and 2 Hz for CO2 and H2O, respectively, were applied. For data collected by the open-path gas analyzers, the WPL correction (Webb et al., 1980) was applied. To account for the self-heating of the open-path analyzer, a correction was applied following Burba et al.'s (2006) approach for inclined positions (∼ 15∘), using an optimized fractionation factor of 0.07 based on high-quality data. This fractionation factor was subsequently used for the data post-processing of all open-path measurements used within the context of the present study.

Our post-processing quality control scheme was based on the quality flagging system initially proposed by Foken and Wichura (1996), which tests data for effects such as stationarity and well-developed turbulence. Afterwards, a set of additional tests was applied to flag implausible data points in the resulting flux time series. These tests included a check for absolute limits for net ecosystem exchange (NEE; 5 µmol m-2 s-1 < NEE < -15 µmol m-2 s-1), the open-path gas analyzing system status information (gain control maximum > 75), and overall errors in the log file recorded by the sonic anemometer, as well as a comparison of the absolute concentrations of CO2 for the two towers for specific wind directions. This comparison of concentrations was performed to detect potential contamination by the exhaust plume of the generator (based on a more than 5 % difference criterion).

Gap filling and flux partitioning were both based on the so-called marginal distribution sampling method (Reichstein et al., 2005), implemented through the R package REddyProc (https://r-forge.r-project.org/projects/reddyproc/). As there was a pronounced variability of environmental conditions during the growing season, the dataset was split into four different sub-seasons (see Sect. 2.6 for details), with individual gap filling and flux partitioning data pools for each sub-season. The final NEE time series used within the context of the present study were merged from open-path (until April 2014) and closed-path (since April 2014) fluxes.

Data coverage ranged between 60 and 80 % over the different years of eddy-covariance flux measurement using closed-path observations. Gaps were commonly less than a day and distributed equally over all sub-seasons. Longer gaps were caused by flooding events, when the entire system needed to be shut down to avoid damage, or resulted from a laser offset in the closed-path gas analyzer at Tower 1 in 2015.

The flux-partitioning algorithm that splits NEE into GPP and Reco signals is based on the estimation of the nighttime respiration as a function of temperature and the extrapolation of this relationship for the daytime (Reichstein et al., 2005). In a subsequent processing step, GPP is derived from the difference between NEE and Reco (for details see Reichstein et al., 2005). The results for the component fluxes GPP and Reco presented here are therefore not based on direct measurements but rather represent a model–data hybrid product. The performance of this procedure is dependent on a number of user-defined settings, such as the definition of nighttime conditions by setting a radiation threshold (present study: 20 W m-2, e.g., Parmentier et al., 2011; Runkle et al., 2013). To address the temperature dependence of Reco, we used air temperature instead of soil temperature. This decision was based on two factors: first, the patterns in near-surface temperatures were implausible at our site; secondly, higher-quality soil temperatures at 0.08 m below the surface cannot be correlated to short-term changes in respiration patterns because of a damped diurnal cycle. This decision was based on our observance of implausible patterns in near-surface temperature measurements at our site, and, because of a damped diurnal cycle, higher-quality soil temperatures at 0.08 m below the surface cannot be correlated to short-term changes in respiration patterns. During the growing season, air and near-surface soil temperatures are closely linked, and both represent the temperature regime within the soil, driving respiration processes (Lloyd and Taylor, 1994). However, this flux-partitioning approach using air temperature data is limited during the thawing and refreezing periods when energy is not used for heating, but instead for the phase change of ice water, and when soil temperatures do not change relative to air temperatures.

For the statistical analysis of the whole dataset, a linear regression with a moving window of 14 days was used to analyze the peak uptake. An hourly shift was chosen to account for short-term variabilities. Maximum negative slopes were identified as maximum uptake rates while the center of the window represents the day of year (DOY) with the maximum uptake. To determine statistical differences between drained and control site a one-way analysis of variance (ANOVA) was performed to compare the means of the time series. Measurements and thus errors are independent, the data are normally distributed and the variances are equal in most cases. All statistical designs and analysis were performed with the software R (R Core Team, 2014).

Eddy-covariance measurements were conducted during the 2002–2005 growing seasons in the context of the TCOS Siberia project (Corradi et al., 2005; Merbold et al., 2009, Fig. 2), at the same location that was used for the drainage tower in the new experiment starting 2013. Flux measurements until the end of October 2004 represented an undisturbed tussock tundra (Merbold et al., 2009). After the installation of the drainage channel in autumn 2004, flux measurements at the same location in the subsequent year indicated that this modification in the local water regime resulted in no significant change to the seasonal budgets of net carbon dioxide fluxes (Merbold et al., 2009). However, these trends were based on a single year of observations only and thus did not reflect the long-term effects of such a disturbance.

To facilitate a direct comparison between recent (2013–2015) and historic (2002–2005) observations, the raw datasets from the period of 2002–2005 were re-processed using the same data processing and quality assessment procedure outlined in Sect. 2.4. This re-processing caused minor changes in the flux rates compared to the values reported in Corradi et al. (2005) and Merbold et al. (2009) because of changed assumptions and inputs when using a different flux processing software. However, general patterns in interannual variability of fluxes were not affected. For the data collected in 2004, a parallel run of open- and closed-path gas analyzers resulted in systematic differences in the seasonal dynamics and final budgets (see also Merbold et al., 2009) for CO2 and H2O fluxes. These differences remained (R2= 0.75) even after re-processing and applying additional corrections that have since become available (Burba et al., 2006), as well as analyzer-specific corrections. Since the seasonal patterns in the open-path data displayed implausible patterns (e.g., near-zero fluxes at the beginning of the growing season, when dominating positive fluxes are expected) we decided to use only the closed-path data for 2004 in this study. More frequent gaps in the historic dataset compared to the recent dataset were caused by failures in the power supply system, as there was no continuous or stable power supply during this time (Merbold et al., 2009). To compare differences between historic and recent datasets the same statistical approaches as mentioned in Sect. 2.4 were used.

We used MODIS NDVI (MOD09GQ, MOD13Q1) with both daily and 16-day temporal resolution to analyze the interannual variability in phenology at the observation site. Raw data for the 250 m daily NDVI products were filtered for outliers based on the 16-day composite NDVI time series, and smoothed using a moving average approach (window size of 10 days). Based on the GreenBrown R package (Forkel et al., 2013), the following phenological key dates were derived from trends and patterns in the resulting seasonal course of NDVI: start of season (SOS), peak of season (POS), and end of season (EOS), for each year individual. From the temporal trends in MODIS time series of NDVI, the POS is determined by maximum value. For SOS and EOS a threshold of 46 and 59 % of the corresponding POS was taken, respectively. These key dates allowed the growing season to be separated into four distinct sub-seasons: (1) pre-season (until SOS), (2) early season (SOS to POS), (3) late season (POS to EOS), and (4) post-season (after EOS). To mark the beginning of the growing season (start of the pre-season), the first date when the daily mean temperature remained above 4 ∘C for 4 consecutive days was used. The end of the growing season (end of the post-season) was set at when there was a MODIS fractional snow cover > 50 %. Mean values for SOS, POS, and EOS, as well as year-to-year variability of these key dates within the period 2000–2015, are provided in Fig. 3.

A webcam (CC5MPXWD, Campbell Scientific) provided an additional data source for vegetation characterization. Pictures were taken daily at noon over the course of the growing season. Figure 4 provides an overview of the vegetation dynamics from pre-season (Fig. 4a) towards post-season (Fig. 4d).

Net ecosystem exchange and component fluxes varied considerably between 2013 and 2015 (Figs. 5 and 6). Splitting the time series into sub-seasons allows for an in-depth analysis of the changing impact of the drainage effects on the CO2 flux over the course of the growing season:

Data coverage between the historic and recent observation periods varied considerably, with measurement starting dates ranging from the beginning of May to mid-July and end dates differing by up to 2 months (Fig. 2). To facilitate a direct comparison of the carbon budgets between all years, the starting dates for cumulative time series (Figs. 7 and 8) were set to 1 May. For time series that started later in the season, we used the available budget from the nearest year (either prior to or after the year of observation). Accordingly, not all displayed time series start with a value of zero, and final values therefore do not always represent the cumulative budget based on the available observations. This approach facilitates a qualitative comparison of seasonal trends in CO2 exchange between different data years, while absolute numbers of the seasonal budgets are subject to significant uncertainty for years with data gaps.

The observed differences in cumulative uptake for the historic dataset reflect the interannual variability in CO2 fluxes that can be expected in this type of ecosystem (Fig. 7a, 2002–2004). Such changes are primarily caused by varying climate conditions from year to year (Merbold et al., 2009). In contrast, 2005 – the year immediately following the installation of the drainage system – shows a systematic dampening in the seasonal course of net carbon exchange (Fig. 7b), even though the cumulative NEE budget at the end of the growing season is comparable to previous years' levels.

A direct comparison of the recent and historic carbon flux patterns in Fig. 7 shows that the seasonal amplitude of the cumulative CO2 budget has increased over the past decade. Maximum CO2 uptake rates are higher at peak growing season in July, and initial carbon release immediately following snowmelt (∼ May–June) seems to have increased as well. In contrast, net CO2 emission rates in autumn appear to be similar between recent and historic datasets. As a consequence, the average cumulative growing season uptake of CO2 by the ecosystem has grown over the past ∼ 10 years.

While the range of the cumulative CO2 budget has partial overlap between the recent and historic data years for control conditions, the fluxes observed immediately after the disturbance (2005) and recent fluxes from the drained area display systematic differences, both in absolute values as well as in seasonal patterns. This finding suggests that the disturbed area may have rebounded from the immediate drainage effects and the resulting dampened seasonality that occurred shortly after the disturbance. Instead, recent CO2 uptake rates from the drained section of the tundra partly exceed the rates observed for the control area before installation of the drainage (Fig. 7, 2002–2004).

The seasonal CO2 exchange for the overlapping data period (19 July to 23 September) for the years 2002–2004 summed up to -42, -21, and -16 gCO2 C m-2, respectively, and -46, -63, and -41 gCO2 C m-2, respectively, for the years 2013–2015 under control conditions. For the drained part of the tundra, we found -10 gCO2–C m-2 for 2005 and -20, -48, and -12 gCO2–C m-2 for 2013–2015, respectively. Negative budgets over most of the growing season in all observation years indicate that both the drained and control ecosystems act as net sinks for atmospheric CO2 during this part of the year. However, the drained ecosystem displays slightly different exchange patterns for CO2, with higher respiration rates throughout the growing season, making it a weaker CO2 sink in comparison to the control site.

Separation of NEE into its individual component fluxes – GPP and Reco – demonstrates that the interannual variability in both cumulative time series (recent: 2013–2015; historic: 2002–2005) increased strongly after mid-July. Net flux variability is dominated by changes in photosynthetic uptake (Fig. 8b). Absolute cumulative flux rates of GPP and Reco tend to be higher at the drained site, indicating that the carbon turnover rates have increased as a consequence of drainage. The long-term tendencies toward higher net CO2 uptake of the ecosystem during the summer are mainly driven by increases in GPP since overall increases in photosynthetic uptake outweigh higher respiration losses.

We found a strong interannual variability in timing for key phenological stages in the Chersky area (for the MODIS period, 2000 to present) with respect to the start (SOS), peak (POS), and end (EOS) of the season (Fig. 3). Maximum year-to-year variability ranged between 18, 26, and 6 days for SOS, POS, and EOS, respectively. Focusing exclusively on the observation periods 2002–2005 and 2013–2015, we found a maximum difference of 18 days between the shortest (118 days, in 2015) and longest (136 days, in 2005) growing season lengths. Neither the timing nor duration of seasons showed any long-term trends, resulting in no detected systematic differences in phenology between the historic and recent observation periods.

Our limited data availability precludes establishing significant links between phenology and NEE data. The analysis of the interannual variability of cumulative CO2 flux budgets was executed only for the control site, limiting the data pool to just 6 years (2002–2004 and 2013–2015). Moreover, observations made during the historic period did not cover the entire growing season in some years, further reducing the database for the purpose of statistical analyses. It was therefore not possible to identify meaningful links between individual key dates and NEE patterns. Similarly, no direct correlation between growing season length and NEE budgets in either the recent or historic datasets was found.

Normalizing CO2 time series to account for shifts in phenology (see Fig. 3 and Sect. 2.6) removes interannual variability in cumulative flux budgets and separates the cumulative CO2 budgets into two distinct groups, the recent (2013–2015) and the historic (2002–2004) datasets (Fig. 9). Differences in CO2 flux budgets become particularly obvious in the early and late seasons (Fig. 9b and c). A linear regression fit to identify the maximum gradient of the growing season CO2 uptake (Table 2) confirms this separation, indicating a significant increase in this peak NEE uptake (ANOVA, p < 0.05) between historic and recent datasets. However, the timing of the maximum uptake (as DOY of maximum gradient, Table 2) seems not to be influenced, as dates vary from mid- to late July without any clear pattern. For this gradient analysis we tested different sizes of the moving window, and found no sensitivity of the obtained results towards this parameter. During both the pre- and post-season the net CO2 release signal shows only minor differences between the recent and historic data (Fig. 9a and d).

In Arctic tundra ecosystems reduced soil moisture and the related decrease in water table depth increase the decomposition of soil carbon. This stems from the increased diffusion of oxygen into the soil (Billings et al., 1982; Oechel et al., 1998; Peterson et al., 1984), which induces higher CO2 exchange fluxes to the atmosphere. In contrast, partitioning of the NEE signal indicates a minor increase in GPP rates at the drained site, which can be attributed to a shift in the vegetation composition. Drainage has led to the establishment of more tussock-forming sedges (Carex species) and shrubs (Betula exilis and Salix species), while the undisturbed parts of this tundra ecosystem are instead dominated by cotton grasses (Eriophorum angustifolium, Kwon et al., 2016). Such shifts in vegetation community structure have been associated with differing CO2 uptake patterns (Christensen et al., 2000). Major differences – which were especially clear in the early and late season, when emissions from the drained site exceeded those found at the control site – are caused by higher respiration activity, triggered by higher soil temperatures (Kwon et al., 2016; see also Fig. A1c and d) in the upper part of the active layer (Lloyd and Taylor, 1994).

In an earlier study, Merbold et al. (2009) could not show a significant change in the CO2 exchange patterns immediately following this drainage disturbance. As this result can be attributed to the shorter observation period of only a single data year, neither long-term effects nor interannual variability could be taken into account. In contrast, in the present study we demonstrate that long-term (> 10 years) drainage had a significant impact on the carbon cycle processes of this tundra ecosystem. Since recent CO2 uptake rates from the drained site at times exceed the rates observed for the control area before installation of the drainage, we conclude that the drained section has largely adapted to the new conditions, with features such as vegetation and microbial community structures that can tolerate lower average water table depths and stronger fluctuations in water availability over the course of the growing season (Kwon et al., 2016). Concerning interannual variability in the fluxes, ecosystem respiration seems to be relatively unaffected by annual shifts in either climate or water conditions, while assimilation can vary substantially from year to year. This variability in GPP is primarily due to changing temperatures and cloud cover and is further amplified under the influence of fluctuating water regimes at the drained site. Finally, despite the fact that long-term changes in site conditions (e.g., water availability, plant communities) can lead to systematic shifts in both absolute fluxes and flux components, we found interannual variability to remain at approximately the same level between historic and recent data.

To our knowledge, our study represents the only long-term experiment focusing on the net impact of lowered water table depths on summertime CO2 fluxes in Arctic tundra ecosystems. Existing water manipulation approaches have been conducted that can be used to evaluate our findings despite the fact that differences between wet and dry sites were obtained by wetting the natural tundra instead of drying as in case of the presented study. For example, in one study, long-term water table depth increase triggered a stronger CO2 sink during the growing season by promoting photosynthetic uptake during the summer (Lupascu et al., 2014). Other studies found lower CO2 emissions during autumn as a consequence of wetter conditions (Christiansen et al., 2012). Similar conclusions were drawn in the present study, where wetter conditions under undisturbed conditions led to lower respiration rates and a higher CO2 sink compared to the drained site. However, a direct comparison of our results with prior studies is compromised due to the different measurement techniques used, as well as the multi-factorial interaction of different treatments combined with the increased water table depth.

Similar findings to these long-term experiments have been described by studies that focus on short-term processes using an eddy-covariance method. Dry conditions are considered to be an important control factor for the terrestrial carbon cycle and have been linked to increased respiration rates during the growing seasons (Aurela et al., 2007; Huemmrich et al., 2010; Lund et al., 2012; Zona et al., 2010). Such findings are in line with the results we obtained shortly after the setup of the drainage channel (2005). Although Zimov et al. (1996) observed the highest effluxes during the wettest year in an area close to our site, these results may have been influenced by small-scale heterogeneity and differences in measurement technique (i.e., chamber vs. eddy covariance).

The methods and limits used in determining the key dates for seasonality have a high degree of uncertainty. This uncertainty can influence the key dates derived to indicate SOS, POS, and EOS, which can in turn lead to uncertainty in the growing season and sub-seasonal NEE budgets. Variable spring flooding can also affect the duration of the growing season through an indirect shift in key dates and can reduce CO2 uptake during the growing season (Zona et al., 2012). However, as we do not expect potential uncertainties in the key dates to cause systematic shifts in seasonal trends and budgets, the qualitative results presented in the present study should not be affected.

Here, we compare two observation periods, the historic (2002–2005) and the recent dataset (2013–2015), both covering 3 years of control measurements. These time frames might not be sufficient to fully represent conditions within each observation period, since surface–atmosphere exchange processes are often subject to pronounced interannual variability. Long-term continuous observations would allow a more in-depth statistical analysis. However, we observed characteristic patterns (e.g., the maximum CO2 uptake rate in summer) that clearly showed a distinctive trend between historic and recent periods. Thus, even from this limited database, we can see that the ecosystem has undergone a systematic shift over the past ∼ 12 years.

Pronounced net CO2 emissions to the atmosphere at the start of the growing season, as observed in this study during the pre-season, have previously been described by other studies for Arctic tundra sites (Corradi et al., 2005; Euskirchen et al., 2012; Grøndahl et al., 2007; Kutzbach et al., 2007; Nordstroem et al., 2001; Oechel et al., 1997, 2014; Parmentier et al., 2011; van der Molen et al., 2007; Zimov et al., 1996). During the pre-season, respiration clearly dominates the CO2 flux signal, while the onset of photosynthetic uptake is delayed even though environmental conditions are adequate (e.g., mean temperature = 8.1 ∘C, mean shortwave incoming radiation = 170 W m-2). This observation, which differs markedly from the results of other studies focusing on Arctic tundra CO2 flux exchange (Harazono et al., 2003; Kutzbach et al., 2007), might be associated with the very low abundance of mosses in this specific ecosystem (Kwon et al., 2016).

The interannual variability of cumulative CO2 fluxes is weak during the pre-season, even if environmental conditions (i.e., air and soil temperatures, soil moisture) are highly diverse both within and before the individual pre-seasons. For example, 2004 showed exceptionally high air temperatures during this sub-season, while net CO2 emissions remained at the same levels as had been observed for years with more regular temperature courses (Fig. B2 in Appendix B). In contrast, in 2015 release rates were higher than average during a relatively cold pre-season. We were unable to correlate the early start of the growing season and warm springtime conditions with increased uptake patterns during the growing season as described by other studies such as Aurela et al. (2004) and Euskirchen et al. (2012). These findings suggest that, for our study site near Chersky, early-season emissions may be affected by both conditions during winter (i.e., snow cover and depth, onset of snowmelt) and transition period (i.e., start and intensity of the flooding event). A detailed effect cannot be quantified here since all relevant snow regime dynamics happen before the start of the pre-season. Based on the available data, no direct correlation between the years with the lowest uptake over the growing season and the strongest flooding event were observed; an isolated effect of the flooding event therefore cannot explain the observed interannual variability.

Both peak uptake rates and the timing of these peaks at the end of the early season (i.e., mid- to late July) agree with findings presented for tundra systems across the Arctic circle (Euskirchen et al., 2012; Grøndahl et al., 2007; Harazono et al., 2003; Kutzbach et al., 2007; Lafleur and Humphreys, 2007; Nordstroem et al., 2001; Oechel et al., 2014). During the late season the shift of the balance between respiration and assimilation towards a net release of CO2 is highly variable between years. Results from 2014 indicate that the length of the net uptake season is highly sensitive towards climate conditions in July to August. In 2014, warm temperatures (> 5 ∘C compared to the average monthly mean in August) led to an increased net CO2 uptake from mid-July to the end of October by ∼ 20 gCO2–C m-2.

The year-to-year variability in NEE during the early and late seasons is dominated by variability in GPP (Euskirchen et al., 2012; van der Molen et al., 2007). In our study, we found GPP also to be the main driver of shifts in summertime CO2 fluxes between the recent and historic datasets. The observed net increase in the CO2 sink strength in recent years can be attributed to significant increases in CO2 uptake rates around the peak of the growing season (Table 2). This observed separation in the cumulative CO2 patterns (Fig. 9) between historic and recent periods cannot be attributed to systematic changes in climate: There are no systematic trends in air temperatures over the last decade in the Chersky region, and, particularly during the early and late season, average conditions were found to be very stable over the years (Fig. B1). Small deviations in temperature during pre- and post-seasons are not correlated with the observed trends in CO2 uptake. However, we observed a minor increase in cumulative NDVI values (Fig. B2) in recent years (2013–2015), and these higher photosynthetic leaf areas might indicate an overall increase in light use efficiency within this ecosystem. Since the MODIS NDVI data result from a mixed product covering both the control and drained sites, values might have changed slightly due to changes in the vegetation composition at the drained site (Kwon et al., 2016). However, considering these small differences without a trend between recent and historic years, this data product alone cannot provide a final answer regarding overall shifts in vegetation structure.

Another possible factor influencing the differences observed between recent and historic datasets is the location of the control site. Since the control site from the historic dataset was converted into a drained site in autumn 2004 with the installation of the drainage ditch, a new control site had to be chosen for the recent measurements. Elevation measurements in 2014 indicated that parts of the drained site are slightly lower than the surrounding terrain (up to ∼ 0.5 m). Within the floodplain environment, such minor topographic depressions preferentially collect melt and flood water; accordingly, control observations made during the historic period may have been subject to more pronounced flooding as compared to the recent control observations at the new tower position. More standing water surrounding the tower site could in turn cause differences in photosynthetic activity, respiration rates, and NDVI. Flooding dynamics can also influence carbon exchange fluxes and increase CO2 losses (Zona et al., 2012). However, differences in elevation might have been less pronounced in the historic dataset, since subsidence may have been a secondary disturbance effect triggered by the lowered water table depth in this area.

Similar year-to-year patterns of pronounced CO2 efflux rates can be observed in the post-season, where temperatures remain above freezing in most soil layers with favorable conditions for the microbial decomposition, triggering enhanced CO2 emissions (Parmentier et al., 2011). However, unlike Oechel et al. (2014), respiration rates at our site are lower during post-season (∼ 35 %) than during late season. This is primarily due to biogeophysical differences in the site conditions in Chersky as compared to the Alaska North Slope sites near Barrow. While thawing of the active layer continues in autumn near Atqasuk (Alaska, US), soil temperatures are already starting to decrease by August to September (see Fig. A1c and Table A1) in Chersky.

Our observations show that the net CO2 sink strength of the drained ecosystem was reduced by 20–40 gCO2–C m-2 per year over the three observed growing seasons (2013–2015). Differences in NEE between the drained and control ecosystems are dominated by increased respiration rates at the drained site, primarily triggered by higher near-surface temperatures. In contrast, we found only minor effects on photosynthetic uptake, and the differences between the drained and control sites were less systematic across data years. Still, both ecosystems show negative flux budgets over the growing seasons, representing a net sink of CO2 from the atmosphere.

The comparison of recent flux data to historic conditions (2002–2005) indicates that this tundra ecosystem is now a stronger sink for CO2 over the summer months than it was about 10 years ago. This finding can be linked to an increase in seasonality, with higher net emissions after the snowmelt, as well as to strongly intensified uptake rates at the peak of the growing season in July. The observed long-term shifts are dominated by higher photosynthetic uptake and a resulting stronger carbon sink capacity for CO2 over the whole growing season, most likely linked to a trend towards higher plant biomass.

Our results demonstrate that the drained ecosystem did recover from the intermediate disturbance impact in 2005 (directly after the installation of the drainage), when net carbon exchange rates showed a systematically damped seasonal course. CO2 flux data from recent years indicate that the drained ecosystem has adapted to the disturbance, with changes in the biotic and abiotic conditions leading to a rebound to previous carbon flux levels.

Concerning recent effects of the drainage disturbance, 70 % of the observed differences in cumulative CO2 budgets can be attributed to conditions in the high-summer period (July–August), when the effect of reduced water availability is most pronounced. Concerning long-term trends, the major part of the differences between historic and recent flux budgets was found in the early season, when the long-term increasing trend in photosynthetic uptake has the strongest effect.