Introduction

In the Arctic, the rate of climate change is now faster than natural ecosystem adaptation (Duarte and others 2012). Mean annual air temperature is currently 1.5°C higher than the 1971–2000 mean, with the greatest increases having occurred in the autumn and winter (Jeffries and others 2013). This warming in the Arctic is more than double the warming at lower latitudes (Overland and others 2013). In a changing arctic climate, one of the most significant potential feedbacks from terrestrial ecosystems to the atmosphere is carbon release from decomposition of soil organic matter (SOM) that was, until recently, frozen in permafrost (Schuur and others 2015).

Arctic tundra ecosystems underlain by permafrost store large amounts of soil organic C in the first several meters of the soil because decomposition rates are low in the seasonally unfrozen active layer of these cold, often wet soils (Hugelius and others 2014). Under a warming climate, thawing of the upper permafrost will add soil organic C to the active layer, thereby increasing the pool of SOM potentially available for decomposition. In northern Alaska, this formerly frozen SOM is labile and could become available for microbial degradation as the active layer continues to deepen (Mueller and others 2015). However, the net C balance of these ecosystems will be controlled by the balance between potential gains in plant productivity, which may also increase under warming (Shaver and Chapin 1991; Natali and others 2012) and losses through SOM decomposition and lateral export of carbon.

Recent data syntheses and modeling studies of arctic tundra net CO2 balance have suggested that the tundra is either presented a CO2 sink, a source, or approximately in balance (McGuire and others 2012; Belshe and others 2013; Fisher and others 2014). In a data synthesis, Belshe and others (2013) found that tundra has been a source of CO2 over the last 40 years, despite increases in summer gross primary productivity since the 1990s. In a synthesis of ground-based C flux observations, bottom-up process-based models, and top-down atmospheric inverse models, McGuire and others (2012) found that tundra has been a sink to neutral for CO2 in recent decades. In a large cross comparison among 40 terrestrial biosphere models simulated over the Alaskan Arctic, Fisher and others (2014) found no consensus, with different models showing the Alaskan Arctic as a sink, a source, or in balance. These studies agree that there is insufficient continuous CO2 flux data from tundra ecosystems to understand the current C balance in arctic tundra, and to parameterize and validate ecosystem models, particularly during the winter and shoulder seasons. It has long been recognized that winter CO2 efflux is a key component of the annual C budget in arctic ecosystems because microbes can continue to respire under the snowpack in temperatures close to and even below freezing (Coyne and Kelley 1971; Panikov and others 2006; Drotz and others 2010), but the magnitude and interannual variability of this winter flux remains uncertain (Björkman and others 2010; Webb and others 2016). Due to a harsh, remote environment and the difficulties of collecting measurements in areas without line power, continuous measurements of arctic tundra CO2 fluxes over the full annual cycle across numerous years have not existed until recently (Euskirchen and others 2012; Lüers and others 2014; Oechel and others 2014).

Arctic wetlands have been implicated as a major source of CH4 (McGuire and others 2009), a powerful greenhouse gas. But again, much uncertainty exists with regard to arctic CH4 flux (Christensen and others 2014). Some field studies indicate substantial current Arctic emissions and suggest that future Arctic emissions may also be large (O’Conner and others 2010; Nauta and others 2015), and that these emissions may explain the recent increase in global atmospheric CH4 (Dlugokencky and others 2009). Although CH4 has rarely been measured during the snow season, a recent study indicates that winter CH4 emissions are greater than previously estimated, with highest emissions from dry, upland tundra (Zona and others 2016). However, other data from airborne studies suggest that sources of CH4 in Arctic Alaska may not be particularly large (Chang and others 2014).

We measured net ecosystem exchange of CO2 (NEE, where a negative value denotes a terrestrial sink) over eight years (2008–2015) in the Imnavait Creek Watershed in northern Alaska (68º37′N, 149°18′W) at two ecosystems, heath and wet sedge tundra. These continuing eddy covariance measurements comprise the longest continuous record of Alaskan arctic tundra NEE that is currently available. We also measured NEE over eight growing seasons and two full years at another ecosystem in this watershed, tussock tundra. In 2012, we initiated growing and shoulder season CH4 eddy covariance measurements at the wet sedge tundra site. Beginning in 2006, we measured permafrost temperatures at a nearby borehole in the same watershed, thereby permitting us to relate long-term change in soil temperatures at various depths to C flux. We also quantified the amount of soil carbon in the annually thawed active layer at our study sites, allowing us to estimate the amount of soil carbon that is lost or gained from the system over the period of measurement.

Materials and Methods

Study Site

The study site is located in the Imnavait Creek watershed (2.2 km2), in the northern foothills of the Brooks Range, Alaska (68°37′N, 149°18′W). The watershed is underlain by continuous permafrost, to a maximum thickness of 250–300 m (Osterkamp and others 2005). The predominant soils are 15–20 cm of organic peat underlain by silt and glacial till (Hinzman and others 1991). The mean annual air temperature for the years 1988–2007 was −7.4°C and the mean annual precipitation was 318 mm, with about 40% occurring as rain and 60% as snow. The landscape is treeless, located approximately 100 km north from latitudinal treeline.

We examined three different tundra types, including heath, moist acidic tussock, and wet sedge tundra. The moist acidic tussock tundra ecosystem is dominated by the tussock-forming sedge Eriophorum vaginatum, Sphagnum spp., deciduous dwarf shrubs such as Betula nana and Salix spp., and evergreen dwarf shrubs such as Rhododendron subarcticum (formerly known as Ledum palustre) and Vaccinium vitis-idaea. The dry heath tundra ecosystem is dominated by Dryas integrifolia, lichen, Carex spp., dwarf evergreen, and deciduous shrubs. The wet sedge ecosystem includes a variety of Carex species, Eriophorum angustifolium, and dwarf deciduous shrubs such as Betula nana, Salix spp, and mosses. Further information on the vegetation at these sites can be found in Euskirchen and others (2012) and Kade and others (2012).

For scaling the vegetation and flux measurements, we relied on the descriptions and estimated area from Jorgensen and Heiner’s (2004) vegetation map, which covers the 300,000 km2 of the North Slope of Alaska, from the Brooks Range to the Arctic Ocean. In this map, our tussock tundra site corresponds to tussock tundra on non-sandy substrates; our heath tundra site corresponds to prostrate dwarf-shrub, graminoid, sedge, forb, and lichen communities, excluding that found in the higher elevations of the Brooks Range, and our wet sedge tundra site corresponds to wet sedge, moss communities on the North Slope (excluding those in the far north of the Arctic Coastal Plain, which are assumed to differ from those found on the North Slope; Table 1).

Table 1 CO2 Exchange Budget Across Heath, Tussock and Wet Sedge Tundra Ecosystems of Arctic Alaska Based on Data Collected from January 2013 to December 2014
Full size table

Eddy Covariance Data Collection

Our time series of eddy covariance data builds on that described in Euskirchen and others (2012). Due to the remote location of the site and the absence of line power, electrical power for the equipment at the site was provided by solar panels, wind turbines, and batteries. For the wet sedge and heath tundra sites, the power supply at each site consisted of 32 6-V deep cycle marine batteries connected to five 130-W solar panels and a wind turbine. The solar panels were the predominant source of power from April to October. The wind turbine was the main source from November to March. From 2008 to 2012, the power supply at the tussock tundra site consisted of three 12-V deep cycle marine batteries connected to six solar panels totaling 300 W and consequently only operated from approximately May to early November. In 2013, a wind turbine and additional batteries were added to the tussock tundra site, thereby permitting year-round operations.

The eddy covariance system for measuring the fluxes of CO2, water, and energy was placed on a 3 m high tripod in the center of each site. The instrumentation consisted of a 3-D sonic anemometer (CSAT-3; Campbell Scientific Instruments, Logan, Utah, USA) mounted at a height of 2.5 m at all three sites. An open-path infrared gas analyzer (LI-7500 IRGA; LI-COR, 2004; Lincoln, Nebraska, USA) was used at all three sites until 2012. At this time, the LI-7500A IRGA (LI-COR, 2009), which is not subject to sensor heating, as can be the case with the LI-7500 IRGA, (discussed below), was installed at the wet sedge site. The LI-7500 IRGA remained at the heath tundra site, but with the addition in 2013 of the enclosed path LI-7200 IRGA (LI-COR, 2010a, b). The main axes of the IRGAs were tilted by 20º with respect to the horizontal to aid in draining condensation and precipitation from the optical windows. The IRGAs and the CSAT-3 sonic anemometers were both mounted on a shared horizontal bar and were laterally separated by 20 cm to reduce flux loss and flow distortion. The differing time delays in signals were taken into account by shifting the CSAT-3 data by one scan (at 10 Hz) to match the fixed 302.369 ms delay (or 3 scans at 10 Hz) that is programmed into the LI-7500 or LI-7500A. This instrumentation was connected to a digital datalogging system (either a CR3000 or CR5000; Campbell Scientific Instruments) to log data at 10 Hz intervals. Raw data were collected once a month from a CompactFlash card located in the datalogger. The IRGAs were calibrated following the instructions in the manuals (LI-COR 2004, 2009, 2010a, b). Calibration was checked during each site visit. Gas analyzers were calibrated monthly at first, although this frequency was reduced to every several (3–4 months) since inspections indicated that the instruments remained stable over a several month period.

The percentage of data collected was near 95% during the months from April to October. Data collection from the months November to March was near 85%, except for some larger gaps in the data in the beginning months of data collection, when data loss was greater (Table 2). Data coverage during the entire period was approximately 68% after accounting for data loss from power outages, automatic gain control (which represents optical impedance by precipitation or aerial contaminants), and filtering for periods of low turbulence (friction velocity, u* ≤ 0.1). The fast-response (10 Hz) eddy covariance was reduced to half-hourly means.

Table 2 Periods of Data Loss >1 Week at the Heath, Tussock, and Wet Sedge Tundra Sites from January 1, 2008 to December 31, 2014
Full size table

To account for nocturnal CO2 advection, we calculated a storage term (Papale and others 2006) and then gap-filled the data removed when u* ≤ 0.1. The storage term had a negligible, no more than 2% effect, on the net annual fluxes. We use the ReddyProc tool for gap-filling of the eddy covariance data (Reichstein and others 2005). The gap-filling of the eddy covariance and meteorological data are not only performed through methods that are similar to Falge and others (2001), but also consider both the co-variation of fluxes with meteorological variables and the temporal autocorrelation of the fluxes (Reichstein and others 2005). The Webb–Pearman–Leuning (‘WPL’) terms were applied during post-processing to the CO2 and latent heat fluxes to account for changes in mass flow caused by changes in air density (Webb and others 1980). In addition, corrections were applied to account for frequency attenuation of the eddy covariance fluxes (Massman 2000, 2001). We used the EddyPro 4.2 software (LI-COR, 2013) to apply the WPL and frequency corrections, and to reduce the fast-response (10 Hz) eddy covariance data to 30-minute means. Our choice of ReddyProc and EddyPro software for use in data processing complies with Fratini and Mauduer (2014), who recommend relying only on data processing software that has been extensively documented and validated.

To evaluate the influence of surface heating on the open-path LI-COR 7500 IRGA (Burba and others 2008), we collected data using both an open- and enclosed path analyzer during 2013 and 2014 at the heath tundra site. As noted above, in addition to the LI-COR 7500 open-path instrument that had been in use since the beginning of the study, we installed an enclosed path LI-COR-7200 in 2013. The enclosed path LI-7200 is not subject to surface heating issues. We ran the LI-COR 7200 instrument nearly continuously, except for gaps in 2013 between September 14–25 and October 10–December 1, and in 2014 between May 5–June 6 and September 29–October 10. These data gaps occurred due to instrument malfunction when it had to be brought in from the field site for repair. The LI-7500 ran continuously except for September 15–26, 2013 (Table 2).

Data collected under normal operating conditions may still need to be discarded from both instruments, although the two instruments are subject to data loss for different reasons. The enclosed path LI-7200 is subject to some data loss due to the frequency dampening (but less than some closed-path models due to a shorter intake tube) and may also be subject to incomplete temperature attenuation even in the short tube. In arctic environments, icing can occur on the inside of the intake tube of the instrument. The LI-7500 is subject to data loss due to precipitation on the analyzer. The eddy covariance computation from both analyzers is subject to data loss due to precipitation on the sonic anemometer.

We compared the LI-7200 and LI-7500 datasets using common half-hourly periods of data acquisition in 2013 and 2014. We found that overall the LI-7500 showed slight uptake of CO2 during the winter, whereas the LI-7200 showed release (Figure 1A). Upon application of the heating correction (Burba and others 2008), the LI-7500 and LI-7200 analyzers generally agreed well in the winter (Figure 1A). The correction was negligible in the summer. We note that our magnitude of correction (~0.45 µmol m−2 s−1; Figure 1B) is in line with, or slightly smaller than, that discussed in Burba (2008), Burba and Anderson (2010), and Oechel and others (2014). The effect of increasing air temperatures at our study sites on the sensor heating correction would seemingly result in a smaller correction to the original open-path LI-7500 data since the sensor and the ambient air temperature would be closer together, as was seen in the warmer months (Figure 1A). In summary, based on our comparison between the open- and enclosed path analyzers, we determined that the sensor heating correction was a warranted, and necessary, adjustment to the open-path LI-7500 data.

Figure 1
figure 1

Monthly mean daily NEE (μmol CO2 m−2 s−1) in 2013 and 2014 at the heath tundra site based on common half-hourly data collected with the LI-7500 and LI-7200 analyzers, including LI-7500 data adjusted for sensor heating and not adjusted for sensor heating (A). Error bars represent the standard deviation of the mean. In B, a comparison between LI-7500 half-hourly NEE (μmol CO2 m−2 s−1) adjusted for sensor heating and not adjusted for sensor heating

Full size image

Data pertaining to CH4 fluxes were collected seasonally at the wet sedge site, beginning in July 2012. The site was instrumented with a with a fast-response open-path methane analyzer (LI-7700; LI-COR, 2010b), which used a LI-7550 interface unit to control mirror heating and cleaning cycles and to route the high-frequency data to the datalogger. The calculated methane fluxes exclude data during heating and cleaning, and when the available optical power was less than 5%33. The raw fluxes are also filtered to remove all periods where u* < 0.1 and momentum flux > 0. Gaps in the CH4 data, generally due to brief power outages or filtering, were filled by calculating the mean diurnal variation, where a missing observation is replaced by the mean for that time period (half hour) based on adjacent days. The CH4 data were converted to CO2 equivalents (CO2 e) by multiplying the CH4 flux by the 100-year global warming potential of methane, estimated at 28 (Myhre and others 2013), taking into account that this conversion is performed on the masses of the gases, not the masses of C.

Borehole Measurements

The 70 m borehole was drilled in spring 2006, at the top of the Imnavait Creek valley west-facing slope. The temperature is measured every 5 minutes at 0.34, 0.5, 0.9, 3, 4, 5, 7, 10, 15, 20, 30, 40, 50, and 60 m depths, and the hourly averaged data were stored by the datalogger (Campbell Scientific Instruments, Logan, UT). Temperatures at 0.34, 0.5, and 0.9 m were measured in a separate shallow borehole. The precision of temperature measurements is better than 0.05º C. Soil temperatures at depths of 0.34, 0.5, and 0.9 m were averaged in determining complete soil freeze or thaw. The date of complete soil freeze was determined by the mean day after which soil temperatures remained below −0.1° C at depths to 0.9 m. The date of soil thaw was determined to be the day after which soil temperatures remained above +0.1° C at depths to 0.9 m.

Partitioning of Net Ecosystem Exchange into Gross Primary Productivity and Ecosystem Respiration

NEE represents the balance between gross CO2 assimilation (gross primary productivity, GPP, where GPP ≤ 0, since CO2 uptake is denoted as a negative value) and ecosystem respiration (ER). Although we do not directly measure GPP and ER, NEE based on eddy covariance data can be partitioned into these counterparts to provide an approximation of GPP and ER, and consequently a broader understanding of the photosynthetic versus respiratory controls over NEE. This partitioning is calculated from the relation between NEE during the nighttime (defined as photosynthetically active radiation <50 µmol m−2 s−1) and air temperature, fit to an equation:

$$ ER = R_{0} *Q_{10}^{{T_{\text{a}} /10}} , $$

where T a is the air temperature; R 0 is a scale parameter; and Q 10 is the temperature sensitivity coefficient of ER. R 0 and Q 10 were estimated each day using a 29-day moving window and least squares method (Ueyama and others 2013). GPP is then calculated as GPP = NEE − ER. We calculated ER and GPP for the June–August growing season.

Error Estimate

Bootstrapping was used to estimate the error (95% confidence interval) about the total NEE, GPP, and ER values over the measurement periods for both CO2 and CH4 using SAS software version 9.4. The bootstrap calculated the confidence interval by (1) constructing 2,000 bootstrapped sample series by randomly sampling with replacement the observed total daily time series, (2) calculating an average from each constructed data series, and (3) calculating the grand mean (±95% CI) from the distribution of means calculated from the bootstrapped data series (Efron and Tibshirani 1998).

Trend Analysis

We assessed trends in metrics related to our soil temperature time series, including day of freeze in the autumn, annual minimum soil temperature, the day of thaw in the spring, and the duration of the zero curtain. The zero curtain is the period during soil freeze-up where the phase transition of water to ice is retarded due to latent heat release. Here, we examined the zero curtain period where soil temperatures up to 0.9 m depth remained between −0.1 and +0.1 °C.

We also assessed trends in the summer fluxes of NEE, GPP, and ER. To perform this trend analysis, we first tested for autocorrelation in the time series using the autocorrelation and partial autocorrelation functions (acf, pacf in R software version 3.2.3). We detected a positive autocorrelation of lag-1 in the metrics related to soil temperature, while the summer fluxes of NEE, GPP, and ER did not show autocorrelation. We then pre-whitened the soil temperature metrics to remove the lag-1 autocorrelation using the Yue and Pilon method (Yue and others 2002) in the zyp package of R. We applied the non-parametric Mann-Kendall test for trend at the 5% significance level on the data with the lag-1 autocorrelation removed, as well as the flux data without the autocorrelation. We also computed the magnitude of the trend (for example, a trendline with a slope) based on the Theil-Sen approach (Sen 1968).

Soil Carbon Determination

From July 25 to 28, 2013, soils were sampled at 20 points in a stratified random manner within the major plant community types in the eddy covariance footprint of each tower. To determine soil bulk density and element concentrations, organic soil horizons were sampled volumetrically with a serrated knife. Soil monoliths (approximately 5 × 5 cm and extending from surface of the green moss to the organic-mineral interface) were cut from the side of a soil pit. Cores (5.4 cm diameter) were taken through the mineral soil horizon below these monoliths to the ice surface at the bottom of the thawed layer. Monoliths and cores were wrapped in tin foil to preserve structure, returned to the lab, and refrigerated prior to analysis. Within 24 hours of collection, monoliths, and cores were processed to determine bulk density, soil moisture, and C content.

Soil samples from monoliths and cores were homogenized by hand, and coarse organic materials (twigs and roots more than 2.5 cm in diameter) and rocks were removed. The soil C stocks were corrected for rocks by calculating the volume and mass of the removed rocks. Coarse fractions and a 10 g subsample of the fine soil fractions were weighed wet, dried at 60°C for 48 h, reweighed for dry matter content, and then ground using a ball mill (organic samples; SPEX Certiprep 8000-D, SPEX, Inc. Metuchen, NJ) or mortar and pestle (mineral soils). Dry matter content of fine mineral soil was determined on subsamples dried at 105 °C for 48 h. For all samples dried at 60°C, C content was measured using a Leco TruSpec CHN Analyzer (LECO, Inc., St. Joseph, MI). The volume of each organic monolith layer or mineral core was calculated as depth multiplied by area minus rock volume. Bulk density and C pools were calculated separately for mineral and organic horizons, and C pools were then summed for each sampling point. C pools for each tower were averages weighted by the areas of the major plant community types sampled within each tower footprint.

Results

Long-Term Net Ecosystem Exchange, 2008–2015

Both the wet sedge and heath tundra were CO2 sources over the eight years, from January 2008 to December 2015 (Figure 2A; 158 ± 53 g C m−2 lost from heath tundra and 668 ± 83 g C m−2 from wet sedge tundra, with uncertainty surrounding the fluxes estimated by bootstrapping, as described in the Methods), due to CO2 loss during the cold seasons. The tussock tundra, measured continuously from January 2013 to December 2014, also acted as a source of CO2 during these two years (30 ± 17 g C m−2), with the seasonal patterns of uptake and emissions more closely following that of the heath tundra than the wet sedge tundra (Figure 3a).

Figure 2
figure 2

Cumulative net ecosystem exchange (NEE g C m−2) between January 1, 2008 and December 31, 2014 for the wet sedge and heath tundra (A). Inset B cumulative NEE between September and December for the wet sedge and heath tundra versus the duration of the zero curtain period during each measurement year

Full size image
Figure 3
figure 3

Cumulative NEE between January 1, 2013 and December 31, 2014 for the wet sedge, heath, and tussock tundra, corresponding to the common period of full annual measurements across the three ecosystems (A). Cumulative emissions of CO2 from both NEE and NEE + CH4, expressed as CO2 equivalents (CO2 e) for the wet sedge tundra (B)

Full size image

Soil Temperature Trends

Concurrent with these releases of CO2, the soils in our study area have warmed at all depths since the beginning of soil temperature measurements in 2006 (Figure 4A). Notably, soils at 3 m depth are nearing thaw (0°C), where they had only previously reached a maximum of −2 °C (Figure 4A). The soils in the active layer at a depth to 0.9 m are also freezing at a later date in the fall (Figure 4B). The annual minimum soil temperatures have continued to increase since the beginning of the measurements in 2006 (Figure 4C). In particular, late fall and early winter in 2014 were marked by extremely high soil temperatures, with the soils to 0.9 m depth not fully freezing until late December, about 8 weeks later than in the beginning of measurements (Figure 4B). We did not detect a trend in soil moisture during the time when surface soils to 7 cm depth were unfrozen in June, July, and August from 2008 to 2015 at either the wet sedge, heath, or tussock tundra sites (Supplementary Figure 1).

Figure 4
figure 4

In A, time series of soil temperature from September 2006 to March 2015 from the borehole monitoring station at Imnavait. Depths are from 0.34 to 10 m. Based on the time series in A, we show the day of complete soil freeze in the fall (B), the annual minimum soil temperature (C), the duration of the zero curtain period (D), and the day of thaw (E). Error bars represent the standard deviation based on computing the mean across soil temperatures recorded at depths to 0.9 m, as described in the Methods. The slope of the line is based on the Theil-Sen estimate, and the significance is assessed using a Mann–Kendall test, also described in the Methods. There is no trend line shown in the day of thaw (E) since there was not a statistically significant trend

Full size image

In conjunction with soil warming, there was also a trend in the lengthening of the zero curtain period in the fall (Figure 4D). The length of this zero curtain period was correlated with losses of NEE in both the wet sedge and heath tundra, with large emissions of CO2 between September and December in years with a long zero curtain period (Figure 2B). The wet sedge tundra appears to be more sensitive to changes in the length of the zero curtain period, with an estimated loss of 1.34 g C m−2 for each day that the zero curtain period increases, compared to the heath tundra with a loss of 0.65 g C m−2 for each day that the zero curtain period increases (Figure 2B).

There has also been a trend toward later soil thaw in the spring (May–June), although this trend is not as strong as the delay in the freeze-up seen in the late fall and early winter (Figure 4E). While the wet sedge tundra and heath tundra generally show greater loss of CO2 in years with a late spring (Figure 5), this loss is not as large as that observed in the fall. Furthermore, in the heath and wet sedge tundra, the relation between spring soil thaw and CO2 loss is weaker (Figure 5A, C; P = 0.01 to 0.04) than the relation between the length of the zero curtain and CO2 release in the fall and early winter (Figure 1B, P < 0.0001). In the tussock tundra, there was not a statistically significant relation between soil thaw and spring CO2 loss (Figure 5B; P = 0.4). Consequently, it appears as though increases in soil temperatures in the fall and early winter, and the corresponding length of the zero curtain period (Figures 2b, 4d), are a key predictor of CO2 release at these sites.

Figure 5
figure 5

Day of soil thaw versus total May–June net ecosystem exchange (NEE, g C m−2) at the heath (A), tussock (B), and wet sedge tundra. Each point for soil thaw represents the mean and standard deviation across depths of 0.32, 0.50 m, and 0.90 for the day after which soil temperature measurements remained above 0.1°C. The NEE error bars represent standard error based on bootstrapping, as described in the Methods. There is no trend line shown in the tussock tundra (B) since there was not a statistically significant trend

Full size image

Long-term data from three other borehole sites in northern Alaska show that since 1986 the day of freeze-up is significantly later (Supplementary Figure 2A; slope = 0.90 to 1.36 days later per year; P = 0.001 to <0.0001). Although these sites show a trend towards an earlier day of soil thaw (Supplementary Figure 2B), the trend is weaker (slope = 0.28 to 0.57 days earlier per year; P = 0.006 to 0.03).

Summer Carbon Fluxes

All three tundra ecosystems were CO2 sinks during the summer, when more uptake of C via gross primary productivity (GPP = NEE − ER; Methods) was found compared to C respired in ecosystem respiration (ER; June–August; Figure 6). However, the 2008–2015 summer trends in NEE, GPP, and ER differed among the ecosystems. The heath and tussock tundra ecosystems showed significantly (P = 0.001) greater C uptake (NEE was more negative) with time (heath: −5.0 g C m−2 y−1, Figure 7A; tussock: −6.5 g C m−2 y−1, Figure 6B), whereas the wet sedge tundra did not show a clear trend (Figure 6C). GPP was more significantly negative (that is, greater productivity) only for tussock tundra (−23.3 g C m−2 y−1, P < 0.001; Figure 6E). Consequently, it is not likely that at these sites gains in productivity during the growing season will outpace annual increases in respiration.

Figure 6
figure 6

Trends in total growing season (June, July, and August) net ecosystem exchange (NEE, g C m−2), gross primary productivity (GPP, g C m−2), and ecosystem respiration (ER, g C m−2) from 2008 to 2015 for the wet sedge, heath, and tussock tundra. Theil-Sen trend lines are only shown for significant trends (P ≤ 0.001) based on Mann–Kendall test, with the trend analysis described in the Methods. The error bars represent standard error based on bootstrapping, also described in the Methods

Full size image
Figure 7
figure 7

Soil C in the active layer for the wet sedge, tussock, and heath tundra (mean ± standard error)

Full size image

Soil Carbon

The wet sedge tundra at our site has greater C stocks in the annually thawed active layer than the tussock or heath tundra sites (16.4 ± 2.4 kg m−2 in the wet sedge versus 11.3 ± 1.0 kg m−2 in the tussock and 11.7 ± 1.9 kg m−2 in the heath tundra; Figure 7). Given the loss of 668 g C m−2 in the wet sedge tundra and 154 g C m−2 in the heath tundra (Figure 2A), this translates to approximately 4.1% of the total active layer soil C stocks (0.5% per year) in the wet sedge and 1.4% of the total active layer C stocks (0.2% per year) in the heath tundra from 2008 to 2015.

Methane Fluxes

Emissions of CH4 generally peaked between July and early August, with smaller emissions observed during late spring, early summer and fall (Figure 8). Total CH4 emitted across the common measurement period over the four years of collection (July 11–September 16) was 1399 ± 222 mg CH4 m−2 in 2012; 1726 ± 286 mg CH4 m−2 in 2013; 1621 ± 278 mg CH4 m−2 in 2014; and 946 ± 169 mg CH4 in 2015 (Figure 8). Converting to CO2 equivalents (CO2 e; Methods) and adding them to measured NEE resulted in increased emissions of only 190 g CO2 equivalents m−2 (2,642 g CO2 e m−2 taking into account CH4 + CO2 versus 2,452 g CO2 e m−2 taking only CO2 into account; Figure 3B).

Figure 8
figure 8

Total daily methane (CH4) flux in the wet sedge tundra from 2012 to 2015. The vertical lines indicate the beginning and end of the ‘common measurement period’ over the four years of data collection (July 11–September 16). The error bars represent daily standard error based on bootstrapping, as described in the Methods

Full size image

Regional scaling

The vegetated tundra land area of the Alaska Arctic ecoregion extending from the Brooks Range to the Arctic Ocean is approximately 300,000 km2, comprised various types of heath, shrub, tussock, and wet sedge tundra (Table 1). Of this tundra area, the vegetation within the footprints of our eddy covariance towers is generally representative of 42,552 km2 of the heath, 74,716 km2 of the tussock, and 30,027 km2 of the wet sedge tundra, or in total approximately 50% of the area of tundra in northern Arctic Alaska. By scaling our two years of continuous measurements from January 2013 to December 2014 to the three tundra types by area, we estimate a release over the whole of this Alaska Arctic ecoregion of 14.27 ± 3.84 Tg C from these three ecosystems during this time, if all of this area responded in the same way as the ecosystems that we measured in this study. In this calculation, the C release from the wet sedge tundra dominates the carbon signal (8.80 ± 1.17 Tg C loss over two years). The area of tussock tundra would have to be extended to that of nearly the entire region, to 290,000 km2, to result in the same spatially scaled CO2 emissions as the wet sedge tundra (Table 1).

Discussion

Overview

Our study illustrates the importance of measuring long-term arctic tundra carbon fluxes over the full annual cycle. We have captured the release of CO2 from these ecosystems over multiple years, and how they respond to warming soils, particularly during the most recent years (2013–2015) of this study. Moreover, we demonstrate the importance of considering different types of tundra ecosystems. Many studies target tussock tundra (for example, Hicks Pries and others 2013; Sistla and others 2013), because it is the most dominant type of tundra vegetation in northern Alaska. However, here we find that wet sedge tundra, with a relatively smaller area than tussock tundra, can have disproportionate impacts if it responds strongly (Table 1). Below, we further discuss our findings in the context of other studies of tundra C fluxes and soil carbon, warming soil temperatures in northern Alaska, and biogeochemical modeling of these fluxes across the region.

Net Ecosystem Exchange and Soil Carbon Loss

We are aware of only a few published studies of year-round arctic tundra eddy covariance measurements of NEE (Oechel and others 2014; Lüers and others 2014), whereas other NEE eddy covariance measurements take into account a subset of the winter months in alpine, subarctic tundra (for example, Webb and others 2016). Over one year of full annual measurements of Alaskan tussock tundra, Oechel and others (2014) found a source of 13.6 g C m−2 y−1, which agrees well with the two full annual measurements of tussock tundra NEE we present here, estimated as a source of 15 g C m−2 y−1 in 2013 and 7 g C m−2 y−1 (Figure 3A). Annual NEE over a full year at a high arctic site on Svalbard is measured at 0 g C m−2 y−1 at a site with low productivity and only moderate emissions in the winter, possibly due to little carbon stock available for microbial decomposition (Lüers and others 2014). Lüers and others (2014) also note that nearly two-thirds of growing season CO2 uptake is compensated by these moderate winter emissions. Webb and others (2016) report a large source of CO2 based on eddy covariance NEE measurements in subarctic tussock tundra, 130 g C m−2 y−1, with the overall conclusion that subarctic and arctic tundra are shifting from an overall C sink to a C source. Out of 1,400 years of flux site measurements across terrestrial ecosystems worldwide, the mean annual flux of NEE is a sink of -156 ± 284 g C m−2 y−1, with the greatest sinks generally found in evergreen plantation forests, and the sources associated with disturbed sites (Baldocchi 2014). The CO2 release from the undisturbed, intact tundra ecosystems presented here is similar to that found in boreal forests in the first 10 years following disturbance, which range from a loss of about 20–200 g C m−2 y−1 (Amiro and others 2010). Overall, our results agree with the conclusion of Webb and others (2016) that the arctic is shifting or has shifted to a source of C.

However, a key question is whether these annual increases (net losses to the atmosphere) in NEE in the wet sedge and heath tundra will be sustained over the longer term. The warm borehole temperatures suggest that permafrost thaw may have begun, thereby promoting decomposition of labile soil C in the active layer and releasing CO2 to the atmosphere, at the wet sedge tundra in particular. The rate of loss of soil C we find here would be difficult to sustain over decades if this C is simply coming from increased turnover of SOM in the current active layer, including surface litter C losses, or increased root respiration under warmer conditions. However, it is possible that as the active layer continues to deepen with permafrost thaw, more SOM is becoming available for decomposition (Hicks Pries and others 2013). The dynamics of C deeper in the soil than the current active layer are unknown at these sites, although research suggests that these deeper C stores can be large (Hugelius and others 2014) and that these deeper stores may be attributed to releases of older soil C (Hicks Pries and others 2013). Decomposition of added SOM due to thawing permafrost deeper than the current active layer would likely continue to result in increases in respiration that are greater than any increases in productivity, unless there are extreme shifts in plant community composition to much larger growth forms, such as shrubs or trees (for example, Myers-Smith and others 2015). Although we have documented increases in shrub abundance and biomass in this watershed, encompassing shrubs in the heath, tussock, and wet sedge tundra sites, since the 1980s (Bret-Harte and others unpublished data), it does not appear as though increases in productivity associated with this growth have compensated for soil C release (for example, Figure 5). However, it is possible that the soil C losses we detected may have been even greater without these documented increases in shrub biomass. Furthermore, it is important to note that much of the soil warming occurred during the fall and winter months (Figure 4b, c, d) and in conjunction with recent losses of late summer and early fall sea ice, particularly that off the coast of northern Alaska (Parmentier and others 2013). This fall and winter soil temperature warming followed plant senescence, which thereby precluded additional plant uptake of CO2, but enhanced CO2 respiratory losses.

Methane Fluxes

The wet sedge tundra is the site at which we would expect to see the greatest CH4 release given the saturated soil conditions and presence of sedges with aerenchyma, or gas conducting tissues (von Fischer and others 2010), but nevertheless, the emissions over our four growing seasons of measurements were low (Figures 3B, 8). Although recent work suggests that wintertime fluxes of arctic tundra CH4 can be significant (Zona and others 2016), we observed declining fluxes by mid to late September or October (Figure 8). However, during the period of measurement from July 11 to September 16 across our four years of CH4 measurements (for example, the ‘common period’ and when the flux rates were generally greatest, Figure 8) the mean total emission (1.4 ± 0.3 g CH4 m−2) was slightly less than that reported across five sites in Zona and others (2016) during the cold season (1.7 ± 0.2 g CH4 m−2). Extensions of the measurement systems here to other wetland tundra sites in this region would help clarify the amount and heterogeneity of CH4 emissions. Further, it would be interesting to determine if the CH4 from this wet sedge tundra site is being oxidized and then released as CO2.

Soil Temperatures

Other borehole monitoring stations on the Alaskan North Slope and in other parts of the Arctic have shown similar warming trends, particularly in the fall and early winter (Osterkamp 2005; Osterkamp and Jorgenson 2006; Romanovsky and others 2010). For example, Osterkamp (2005) finds that across a transect over the Alaskan Arctic, sites on the Arctic Coastal Plain have shown warming of 3 to 4°C, whereas those of the Brooks Range, including its northern and southern foothills, the warming was 1 to 2°C. Osterkamp and Jorgenson (2006) predict that the discontinuous permafrost boundary could extend north of the Brooks Range in the coming decades, if not sooner. These studies indicate that because the warming is not restricted to our site, increased amounts of CO2 could be released across a wider geographic area than we are currently measuring.

Conclusions

Our results suggest that warmer autumn and winter periods are already leading to substantial release of C from undisturbed, intact tundra permafrost ecosystems in northern Alaska. Further, they suggest that wet sedge ecosystems may contribute disproportionately to C loss. Our data illustrate the importance of measuring C fluxes over the full annual cycle across numerous years and representative types of tundra. They also illustrate the need to extend these types of long-term measurements in a coordinated manner using similar types of instrumentation, and effectively transfer this information into biogeochemical modeling frameworks.

Given the large amounts of C stored in permafrost soils worldwide (Hugelius and others 2014) and research suggesting that C stores in these soils may be more vulnerable to climate warming due to ‘enhancing’ microbial responses to temperature change (Karhu and others 2014), there is the potential for a strong positive feedback to global climate warming from arctic terrestrial ecosystems. With the continued increases in greenhouse gas emissions worldwide, and the current trend toward increased Arctic temperatures in fall and winter (Jeffries and others 2013), it is likely that elevated CO2 emissions from these ecosystems will continue. These ecosystems may be on a new trajectory of CO2 release, which could cause an enhanced positive feedback to climate warming, if our results are representative of larger areas of the Arctic.