The study area is located in the Kolyma Lowland region in northeastern Siberia (68.51∘ N, 161.50∘ E), close to the town of Chersky, Sakha Republic, Russia, which is situated around 100 km south of the Arctic Ocean. The weather patterns frequently switch between maritime air masses from the north and continental air masses from the southeast, with the former dominating the wintertime conditions and the latter dominating in the summer. The mean daily air temperature can remain at or below -40 ∘C for several days during December to February, while daily means of more than 20 ∘C can be reached around the peak of summer. The mean annual temperature (averaged for 1960–2009) is approximately -11.0 ∘C. The total amount of annual precipitation (averaged for 1950–1999) is between 200 and 215 mm, with 80 to 110 mm falling as rain . Snowmelt leads to an annual flooding event in the Kolyma River and its tributaries, usually inundating large parts of the study area between middle to late May and late June .

In the context of this study, we compared carbon fluxes and ecosystem characteristics between two measurement sites in the Chersky region. Both sites are wetland ecosystems situated within the Kolyma Lowland region that are waterlogged for the largest part of the growing season, affected by the flooding regime of the Kolyma River. Our undisturbed reference ecosystem is a tussock tundra site situated about 15 km south of Chersky on the floodplain of the Kolyma River. Due to the very low natural abundance of grazing herbivores in the region, the influence of grazing disturbance on this dataset can be considered negligible. Situated about 15 km further south at the margins of the floodplain, the Pleistocene Park site, hosting a variety of herbivores (sheep, yaks, cows, horses, bison, muskoxen, reindeer), was used to study the effects of grazing on permafrost ecosystems.

The Pleistocene Park area (https://pleistocenepark.ru, last access: 14 March 2022) was established in 1996 on 144 km2 of permafrost territory mainly consisting of ecotonal upland forest–tundra, wet lowland tundra, and small lakes and rivers. About 2000 ha has been fenced in to concentrate animals on the core domain of the park; accordingly grazing disturbance has influenced these sections for about 25 years. Inside the lowland section on a former wet tussock tundra ecosystem, we selected one of the longest and most intensively grazed areas as our grazing study site, hereafter labeled as “GR”, with data being collected at three plots (GR-1 to GR-3) placed within a radius of approximately 6 m. This site, which similarly to the reference site is flooded every year in spring during snowmelt, is a moist or wet meadow almost without shrubs and also featuring decaying tussocks. The vegetation at this site primarily consists of grasses and sedges, including Calamagrostis langsdorfii, Carex appendiculata and Eriophorum spp. . Before the introduction of grazing herbivores, this site used to be dominated by tussocks and saturated with water during the whole year (Sergey Zimov, personal communication, 2019). Prior to disturbance, the study site therefore represented the typical lowland wet tussock tundra ecosystem dominating large parts of the Kolyma Lowland region.

The reference site, hereafter labeled as ungrazed site “UGR”, is located outside the Pleistocene Park domain on a wet tussock tundra floodplain along the Ambolikha River, a small tributary of the Kolyma River. Here, a long-term monitoring site was installed to investigate drainage effects on wet tundra ecosystems . The reference site for our study is the non-drained control area of this experiment. The dominant vegetation species are tussock-forming Carex appendiculata and C. lugens and Eriophorum angustifolium, with Betula nana and Salix spp. growing on elevated areas with a lower water table. An organic peat layer (15–20 cm deep) has accumulated on top of alluvial material soils (silty clay) . In the context of this study, we assume this site reflects the status that the grazed site would have if herbivores had not been introduced there since both sites showed a similar ecosystem structure in the early 1990s (Sergey Zimov, personal communication, 2019). Observations were collected at two plots (UGR-1, UGR-2) to capture some variability concerning vegetation structure and soil properties. Based on results of a previous study evaluating small-scale flux variability across a transect of 10 quasi-randomly selected locations , flux rates and environmental conditions within these two plots were shown to be close to the average at this site. In terms of spatial flux variability, UGR-1 (Control-0 site in Kwon et al., 2016) featured a relatively high productivity, while UGR-2 (Control-2) showed comparatively reduced flux rates.

Soil temperatures at 5, 15, 25 and 35 cm depth were recorded at 1 Hz during the time when flux measurements were conducted (Th3-s Soil Temperature Profile Probe, UMS GmbH, Munich, Germany). While the two UGR sites were equipped with one probe each, a single probe was used for the three GR sites due to their proximity. To create a gapless dataset for the full experiment period, TS, 25 cm and TS, 35 cm were interpolated linearly (GR sites) or modeled based on a moving average of Tair (UGR sites). TS, 5 cm and TS, 15 cm have been modeled based on air temperature, and in the case of TS, 5 cm incoming shortwave radiation and corresponding TS, 35 cm were used additionally in the course of stepwise linear regressions to improve the prediction.

For soil moisture (SM) measurements, we installed a set of three TDR (time-domain reflectometry) probes (models CS640, CS630 and CS605; Campbell Scientific, Logan, UT, USA) combined for all three GR sites at depths of 7.5, 15 and 30 cm close to the soil temperature sensor and recorded at 1 Hz intervals. At the UGR site, due to the waterlogged conditions, soil moisture was permanently saturated, and no measurements were used within the context of this study. SM values at GR were flagged using plausibility limits, and systematic offsets were corrected. Since the soil moisture data were also only collected during times of flux measurements, values were subsequently linearly interpolated to create continuous time series. For both soil temperature and soil moisture sensors, the surface height between tussocks was used as the reference (0 cm).

Thaw depths were measured with a metal pole that was stuck into the ground, directly next to the chamber plot, to avoid disturbance, until it hit the frozen bottom of the active layer. The measured thaw depth was defined as the distance between this point and the ground surface, which included the organic layer on top, and loose material was softly pressed down until more or less stable ground was touched.

Radiation budget components for the GR sites in Pleistocene Park were measured using a net radiometer (model CNR1, Campbell Scientific, Logan, UT, USA) installed at 4 m height on a pole approximately 15 m away from the flux sampling sites. At the UGR sites located at the Ambolikha site, a net radiometer (model CNR4, Campbell Scientific, Logan, UT, USA) permanently installed on a 5 m tall flux tower located approximately 50 m away from the sampling sites provided radiation observations.

Measurements were stored as 10 min averaged intervals. Albedo was derived by dividing the average upwelling shortwave radiation by the average downwelling component for each single day. Then, these daily averages were averaged over the observation period. While photosynthetically active radiation (PAR) was measured directly at the UGR site, at GR it was converted from incoming shortwave radiation data using a single conversion factor provided as a function within the R package “LakeMetabolizer”  following an approach described by, e.g., . Calculations with this statistical package, as for the other R packages listed below, were conducted using R Version 4.1.1 .

Directly prior to measurements at GR, wooden fences were constructed to protect the sites from grazing animals during chamber operation. At both GR and UGR sites, walking boards were placed around the setup to prevent damaging plants and to minimize influences on measurements by disturbing the soil. Our flux chamber approach closely followed and is therefore only briefly outlined in the following paragraphs.

Carbon dioxide, CO2 and methane (CH4) fluxes were determined with a non-steady-state flow-through method using an Ultra-Portable Greenhouse Gas Analyzer (UGGA, Los Gatos Research, USA) for in situ measurements of gas concentrations at 1 Hz. PVC collars of 60 cm × 60 cm, which have a socket at the top for the chamber, were installed in the ground at each plot to prevent leaking of air during chamber measurements. The cubic chamber hoods with a 60 cm side length, made of 4 mm thick plexiglass, were placed on these collars to capture gases exchanged with the surface. They featured an opening valve on the top to avoid pressure effects when the chamber is placed onto the collars. Inside the chamber, three electric fans were installed to ensure well-mixed conditions. Air was pumped from the chamber to the gas analyzer through three tubes installed at different heights inside the chamber hood. Sensors for air temperature (Tair), relative air humidity (RH), air pressure (pair) and photosynthetically active radiation (PAR) were attached to one side of the chamber. For measurements, the chamber was oriented in a way to minimize shading the vegetation with the instruments.

Each flux measurement was conducted for a period of about 2 min. This period was chosen in order to minimize disturbance effects such as temperature increases or moisture saturation within the chamber e.g., while at the same time collecting enough information for the derivation of a stable flux rate. After completing one measurement, the chamber was lifted and tilted for ventilation until ambient CO2 concentrations were reached. Ecosystem respiration (Reco) was determined by covering the hood with a white polyethylene tarp that completely blocked incoming radiation. For each plot, one measurement iteration consisted of three NEE measurements and two Reco measurements.

On each sampling day at GR, we rotated between chamber locations, with about 1 full hour needed to complete the five individual measurements at each of the three plots. Since the two plots at UGR were located further apart, we conducted several individual measurements at one location before switching to the other to minimize the time needed for relocating instrumentation. On each day, only one of the study sites (UGR vs. GR) was sampled.

Each chamber measurement resulted in a 1 Hz time series of CO2 and CH4 concentrations. After excluding the equilibration period of at least 20 s, periods with a minimum length of 40 s and a linear concentration change in greenhouse gases were identified based on the change point detection algorithm provided by the R package “changepoint” . Subsequently, using the R package “boot”  we generated 1000 combinations of start and end times within this period and calculated an ensemble of stationarities (slopes) using a bootstrapping approach. The final slope used for flux computations was identified as the frequency distribution's median. Implausible or disturbed signals were manually flagged and excluded from further analysis. Such cases included for example unstable signals without a distinctly discernible, steady slope, which are not clearly interpretable, or signals obviously disturbed by leakage. The total quantity of measurements is shown in Table 1.

The median slope (a^) of greenhouse gas concentration change over time was transformed to a flux using the following formula: 1flux=a^VchAchpairRTair. Vch and Ach are the volume and surface area of the chamber, respectively. R is the ideal gas constant (8.3144 J K-1 mol-1); Tair and pair are the mean air temperature (K) and pressure (Pa) inside the chosen interval. Fluxes are derived in units of µmol C–CO2 m-2 s-1 and µmol C–CH4 m-2 s-1, respectively.

While fluxes of Reco and NEE could be calculated based on dark and light chamber measurements, respectively, the photosynthetic uptake portion of the flux (GPP) was calculated as the difference between measured NEE and the mean of measured Reco for one measurement. The root mean square error (RMSE) of each flux measurement was calculated using all bootstrapped slopes, distinguishing between Reco and NEE measurements. The slope error for GPP was taken as the summed errors in Reco and NEE measurements. Error values are given in Table B1 in the Appendix.

To analyze the implications of the grazing disturbance on net CO2 and CH4 exchange across the sites, the time series of flux estimates for each plot was interpolated across the entire measurement period of 17 d at a resolution of 10 min. Both Reco and CH4 fluxes were interpolated using empirical plot-specific linear models linking flux rates to environmental parameters (TS, Tair and SM). Parameters minimizing the fit uncertainty were not uniform across plots even at one site (see Appendix B for an exact description of this approach, including the derivation of formulas, and Sect. 4.2 for evaluation). The following formulas were derived to interpolate Reco and CH4 fluxes: 2Reco(GR-1,GR-2)=exp(a0TSoil, 5 cm+b0)+a1SM7.5 cm+b1, 3Reco(GR-3,UGR-1,UGR-2)=exp(a0Tair+b0),4FCH4(UGR)=exp(a0TSoil, 15 cm+b0),5FCH4(GR)=a0TSoil, 25 cm+b0+a1SM15 cm+b1. In each formula, a0 is the slope of the first applied model, and a1 is that of the second. b0 and b1 are corresponding intercepts. Total errors for all fluxes were derived considering the standard error from the final model compared to observed values (linear regression), further considering the standard error from the bootstrapping approach used to transform measured concentration slopes into fluxes and the standard error from modeled TS (for Reco and CH4 fluxes). A detailed error calculation is shown in Appendix B3.

GPP was modeled as a function of PAR, using a rectangular hyperbolic function . 6GPP=-PmaxαPARPmax+αPAR The fit parameters α and Pmax represent the initial canopy quantum efficiency (the initial slope of the GPP–PAR curve at PAR = 0) and the maximum canopy photosynthetic potential, respectively, the latter of which is the hypothetical maximum of GPP at infinite PAR. Both α and Pmax are assumed to have positive values, necessitating the negative sign on the equation's right-hand side to allow GPP to fit the NEE sign convention. In this context, positive fluxes imply carbon losses from the ecosystem into the atmosphere. This model contains the explicit assumption that GPP is insensitive to light stress or temperature effects . For each site, α and Pmax were determined by fitting PAR against the GPP fluxes from chamber measurements and applying a non-linear least-squares (nls) optimization. Implausible PAR values from chamber measurements were replaced by PAR derived from the net radiometer measurements. With these parameters, a continuous GPP time series could be modeled for the entire observation period.

To visualize and compare carbon fluxes between plots and study sites, daily means for GPP, Reco, NEE and CH4 fluxes were calculated. Prior to post hoc pairwise t tests and a correction of p values with Holm's method , a repeated measures ANOVA was conducted using the package “ez” in R ; herein the approach is based on and . In this way, we can provide evidence that the measurements were independent between days and can correct the results of the pairwise comparisons for multiplicity problems. Statistics and further explanations are given in Appendix B3. This approach was used to analyze fluxes, soil temperatures and radiation (i.e., upwelling and downwelling shortwave radiation) measurements.

All interpolation models described in Sect. 2.4 yielded a significant linear regression fit between observed and calculated values (for details, see Table 2).

The application of flux chambers may lead to biases in the ecosystem fluxes themselves . Installing collars in the ground before starting the experiment, while necessary to prevent leaking of air, can disturb the soil and plant roots. At Pleistocene Park, we had to cut a shallow slit in the ground to be able to tightly fit in collars. In a long-term study assessing both chamber and eddy-covariance fluxes conducted in the Alaskan tundra, disturbing roots and soil when the chamber was set up was shown to have a depressing effect on both Reco and, to a stronger extent, GPP. The chamber fluxes only caught up with the eddy-covariance fluxes a few years after installing the chambers, most likely linked to a regeneration of belowground structures after the initial disturbance . In the case of our study, the collars at UGR had already been installed 6 years before the start of the experiment, and accordingly the disturbance effect should have been negligible. In contrast, installation artifacts are likely at GR, where collars were only installed days before starting the 2019 measurement campaign, potentially leading to underestimated fluxes, mainly considering GPP. The observed enhancement in both GPP and Reco following drainage may therefore be a conservative finding. A net effect on the derived shifts in NEE is also possible, since GPP and Reco can be affected in different ways, but a quantification of the potential bias cannot be performed without a longer-term observational dataset.

Another important aspect concerning representativeness of chamber-based carbon flux measurements is small-scale heterogeneity in the ecosystem, which may exist even within plots that are seemingly homogeneous. This heterogeneity is observable at sub-meter scales and can be a result of disturbances by soil fauna, pockets of fine-root proliferation, moisture gradients, freeze–thaw dynamics, spatial structures in vegetation distribution such as those introduced by tussock-forming plants or remnants of decaying organic matter . In the Arctic tundra, this small-scale heterogeneity is common and is, e.g., reflected by variations in soil temperature, soil moisture and thaw depths. To account for it in flux uncertainties, more than one chamber is needed to adequately assess the mean and variance of surface–atmosphere exchange fluxes . At UGR, spatial variability in carbon fluxes and environmental conditions was analyzed in detail along a transect of 10 flux chamber plots in a previous study , including a description of the vegetation community structure and hydrologic status at each plot. Based on these findings, we selected two positions for the presented study (IDs 2-0, 2-2) which best represent the vegetation composition of a waterlogged tussock tundra that dominated the grazed site before the Pleistocene Park experiment was started in the 1990s. At GR, the relatively low number of plots available for this study, combined with the lack of previous studies within the area, implies that a larger-scale representativeness of this part of the dataset cannot be validated. In combination with the short temporal coverage, the GR dataset should therefore be regarded as a snapshot in both space and time, demonstrating that there is the potential for significant changes in carbon cycle processes following grazing disturbance in permafrost wetlands while not necessarily providing representative flux quantification for a larger domain. As a guideline to evaluate the captured variability across plots, coefficients of variance (CVs) can be computed. CVs of Reco measurements in seemingly homogeneous ecosystems typically range around 30 % and reach higher values for CH4 fluxes, which tend to be more location-specific . For our study, the CV for Reco at GR was 11 % but reached 20 % at UGR. For CH4, CVs were 77 % and 32 %, respectively. The characteristics found in our dataset are therefore generally in line with findings presented by , with the low CVs for Reco potentially linked to the relatively low number of plots observed. However, one has to keep in mind that the GR sites feature more plots than UGR and also a higher number of observations, both of which may influence a comparison of derived CVs.

Generally, Reco in wet tundra and peatlands is enhanced by warmer temperatures due to increased microbial activity , and it may also increase due to drying, with increased potential for aerobic respiration . Identifying and comparing the controls for Reco resulted in different equations across the GR and UGR sites, linked to the fact that their vegetation and soil structures differed, as described in detail below.

The GR site is a flat meadow with dense grass tillers, some decaying tussocks, and relatively dry and warm soils. At UGR, soils are waterlogged and interspersed with tussocks raised above the water level, and the relatively cool, wet soils are covered by a thick organic layer. These differences justify the use of different sets of controls to explain temporal variability in Reco. At UGR, we used only Tair as a driver for interpolating Reco, since the water table was constantly above the ground, and accordingly fluctuations in soil moisture were ruled out. At GR, we used Tsoil at 5 cm and SM at 7.5 cm instead to explain variability in Reco. Our choice is in line with other studies correlating Tsoil and SM with Reco . Here it is important to remember that we had only one set of instruments to measure Tsoil and SM at GR. However, thaw depths and CH4 fluxes (an indicator for wet/anaerobic conditions; ) at the grazed plots varied strongly at this small scale. Therefore, we concluded that the measured soil conditions are not fully representative across the three sub-plots. For that reason, at GR-3 Tair was also used to interpolate Reco, since the coefficient of determination was much higher compared to that of Tsoil. No statistical relationship between Reco and SM could be found for this plot.

Also CH4 fluxes varied strongly between GR plots, being highest at GR-1 and lowest at GR-3. Principally, high Tsoil and water-saturated conditions promote a high CH4 release . We observed an almost continuous cooling trend from warm to cold across the whole observation period. At the same time, soil moisture also continuously decreased. Therefore, high SM was always accompanied by high Tsoil and low SM by low Tsoil. This made it difficult to disentangle their effects on CH4 fluxes.

Estimates of carbon fluxes obtained by chamber measurements in this study cover a similar range compared to recent reference studies related to grazing in the Arctic tundra (see Table 4). Still, when directly comparing the results between studies, one has to keep in mind inter- and intra-annual variability in environmental conditions and therefore also in carbon fluxes, which are highly pronounced in tundra ecosystems . Differences in experimental approaches regarding how to assess and display the fluxes further aggravate the comparison of fluxes across studies. Furthermore, many studies cover longer timescales compared to this study.

Values presented by , obtained during flux chamber campaigns from the period 2013–2015 at the Ambolikha site that served as the UGR reference within the context of the presented study, largely agree with our findings. A wet Carex meadow (grazed by geese) in a subarctic coastal tundra showed very similar values compared to values at UGR . Most of the other studies displayed in Table 3 were conducted in the high Arctic; therefore flux magnitudes are expected to be lower compared to our study . observed similar patterns concerning the amplification of both GPP and Reco as a response to grazing compared to our study for a wet tundra ecosystem. However, we are not aware of studies examining carbon fluxes on a similar type of grazed ecosystem to that at Pleistocene Park, in corresponding climatic conditions.

Grazing by large herbivores had a number of obvious impacts on the vegetation in Pleistocene Park. However, one issue that complicates the attribution of the herbivore influence on the vegetation is the year-long human disturbance by the park operations, such as operation of vehicles or feeding animals in winter. While this influence is mostly restricted to selected areas and transects across the park and no major direct impact was apparent for the study plots used within the context of this study, one cannot completely disentangle the impacts of anthropogenic and grazing disturbances. Another issue that needs to be considered when interpreting the presented intercomparison of GR and UGR sites is potential differences in site characteristics that were already present before grazing management in the Pleistocene Park areas started. Lacking pre-treatment flux datasets for both sites, the only reference that is available is the similar appearance of both GR and UGR sites in photographs from the early 2000s, i.e., a time when ecosystem transformation due to grazing was still in an initial stage. Accordingly, while the differences in aboveground ecosystem characteristics described below can clearly be attributed to grazing pressure over the past 2 decades, the highlighted differences in carbon fluxes may at least partially have been present in the pre-treatment state. While unlikely to affect the qualitative tendencies of higher carbon turnover and reduced methane emissions, this potential systematic bias clearly needs to be considered when evaluating and interpreting the flux numbers.

Around our chamber site at GR, almost all sedge tussocks were in a state of decay or had disappeared nearly completely. In place of them or between their remnants, many single plant tillers (mainly Carex spp. and Calamagrostis langsdorfii) grew. These apparent changes in soil and vegetation properties in Pleistocene Park are in accordance with previously reported observations documenting grazing impacts. For example, the transformation from tussocks to grass mats by grazing, accompanied by a strong increase in belowground biomass, has already been observed for montane biomes . Some sedges found in Arctic environments, such as Carex aquatilis, were shown to benefit from muskox grazing, since they feature strong root production and the ability to produce dense grass tillers and therefore more easily recover from grazing . This ability gives them an advantage over other species , leading to a more turf-like vegetation structure that gradually replaces the original plant community.

Fertilization of tundra ecosystems through available nutrients from urine and faeces also influences vegetation communities under grazing pressure . Accelerated urea-nutrient uptake by living plants has been reported for upland tundra , where graminoids were more efficient in using these resources compared to shrubs. At ungrazed sites such as UGR, the aboveground parts of the plant die off, wither and accumulate on the topsoil, where they rot slowly, leading to a thick organic layer . In comparison, at grazed sites such as GR, the plant shoots were regularly removed by the animals, resulting in reduced litter accumulation. Linked to the same effect, valleys in the Canadian Arctic which are regularly grazed by muskoxen give the impression of a productive meadow, while ungrazed sites in the same region appear overgrown and rather nutrient starved . Similar effects were observed by , where excluding muskoxen from an Arctic mire decreased the number of plant tillers and increased litter and moss cover. In both upland and lowland tundra ecosystems, herbivores, mostly reindeer or muskoxen, have been shown to reduce moss cover and decrease shrub cover by trampling and browsing, promoting the expansion of graminoids . Landscapes covered by graminoids usually have a higher albedo compared to shrub-covered ones . Accordingly, herbivore grazing can systematically increase the surface reflectivity and therefore reduce ecosystem energy input. This observation agrees with our results, which show a significantly higher albedo at GR compared to UGR. Within Pleistocene Park, we are confident that in more shrubby, heavily browsed upland tundra and taiga areas, the increase in albedo following the grazing impact is even more pronounced due to the expansion of graminoids. Since the grazing history of about 22 years at Pleistocene Park is still relatively short, the ecosystem is most probably still in a transition state. We therefore expect further changes in vegetation community structure over the coming decades, given a persistent grazing pressure, i.e., a further condensing of the grass mat and accordingly increased tiller formation and living belowground biomass. Denser grass cover is likely to further enhance the albedo and will also more effectively shade the soil surface. In the long term, both effects should therefore contribute to altering the energy budget, potentially leading to a cooling of shallow soil layers (see also below).

We found overall Tsoil, especially in the topsoil, to be higher at the grazed sites, as well as temperature gradients to be steeper and thaw depths to be greater. At the same time, GR was relatively dry compared to UGR, where the water table was above the ground throughout our observation period. Ultimately, TS, 5 cm at GR reacted 1 order of magnitude faster to changes in Tair compared to UGR-1 and UGR-2. Overall, we suggest there were two processes dominating how grazing pressure transformed the soil at GR: compacting the soil by trampling the organic peat layer (plus related effects on vegetation) and the reduction in soil moisture. A long-term drainage experiment conducted at UGR demonstrated that the topsoil peat layer can strongly influence the soil thermal regime within tundra ecosystems . When drying out the organic layer, it warms up faster but poorly conducts the heat to deeper soil layers . When this peat layer is trampled by herbivores, as observed at GR, the soil thermal regime may be significantly modified, with effects on both thermal conductivity and diffusivity.

A reduction in soil moisture, which was observed at GR, generally leads to a decrease in thermal conductivity and heat capacity in organic soils . Since the warming of the soil was stronger overall at GR, the effects of lower water content on reducing the soil heat capacity along with the higher conductivity caused by topsoil compaction may have taken over as the dominant control. Accordingly, higher temperatures and temperature fluctuations are likely, since less energy is needed to warm up the soil and increase thaw depths. In previous studies, trampling and grazing in the tundra were already shown to lead to the diminishing of the organic litter/moss layer on the topsoil, a decrease in shading by shrubs, and consequently higher soil temperatures and more compact soils . The important role of soil moisture conditions is also highlighted in the results by Windirsch et al. (2021), who investigated places with different grazing pressure within the same thermokarst basin in Pleistocene Park as where our GR sites were located. The drier locations showed a deeper thaw depth in their study, even though the grazing pressure at these sites was actually lower. In accordance with , we therefore conclude that both grazing pressure and soil moisture differences have the potential to substantially influence soil properties, and their co-existing influence needs to be tested in further experiments.

The lower soil moisture observed at GR can be linked to an increase in evapotranspiration. The transpiration rate correlates with observed stronger photosynthetic activity (i.e., GPP) due to the adjustment of stomatal conductance to match the biochemical potential for photosynthesis . Additionally, evaporation increases due to a decrease in litter as a result of grazing . Accordingly, a decline in soil moisture is reported for tundra ecosystems under grazing influence  and also for other grazed biomes where soil structural (i.e., compaction) and subsequent hydrological changes have been reported. In summary, our results for the grazed site within Pleistocene Park, which show drier soils and a higher GPP compared to UGR, agree with previous findings, emphasizing that grazing can exert strong changes in the thermal and hydrological regimes.

Our findings confirm that grazing in tundra ecosystems can lead to higher Tsoil and lower soil moisture, which is usually related to an increase in Reco . This increase in Reco was similar in experiments studying warming and drainage effects on tundra ecosystems exclusively . In addition to triggering these biogeophysical shifts in soil properties, grazing was also shown to lead to faster nutrient cycling in both upland and lowland tundra ecosystems . Higher nutrient availability , along with soil thermal and hydrological changes, might therefore be an explanation for high GPP fluxes observed in our Pleistocene Park study plots. As discussed earlier, these increases may even be more pronounced after full recovery from the disturbance inflicted by installing the chamber collars.

Accordingly, despite the reported increases in GPP, our experiments did not yield an enhancement in net carbon uptake by grazing, i.e., a more negative NEE, during the growing season. In contrast, in other wet Arctic graminoid communities , NEE was greatly enhanced by grazing. In upland tundra ecosystems, however, net uptake was often lower at grazed compared to ungrazed or only lightly grazed sites. The main explanations for these shifts were either a simultaneous increase in Reco at grazed sites balancing GPP gains or a decrease in GPP linked to a decrease in plant biomass and leaf area index following grazing. However, at the same time grazed upland tundra exhibits stable or even increased belowground carbon storage , indicating that decreased Reco in the non-growing season possibly (over-)compensates for the decreased NEE in the growing season. It can be speculated that the reported differences in ecosystem properties between grazed and non-grazed areas have further implications for carbon fluxes in the non-growing season that have not yet been studied in detail.

GR is an ecosystem in transition from a tussock tundra topped by a thick organic layer to a grassland with dense tillers. At the same time, the mean water table during the growing season has been lowered. These shifts may lead to a change in which soil horizon carbon is primarily accumulated, with stronger accumulation in deeper soil layers, as observed by and . The decay of the remnants of tussocks, as well as the decomposition of large carbon pools in the now drier organic topsoil, may contribute to the high ecosystem respiration rates observed. These fluxes, however, have the potential to decrease in the future, leading to decreasing rates of Reco and therefore increased NEE.

In contrast to the current situation in most circum-Arctic ecosystems, Pleistocene Park exhibits a high diversity of herbivores that are absent from the landscape elsewhere. Previous studies have indicated that herbivore diversity leads to a more balanced use of food plants and correlates with increased carbon uptake and soil carbon storage . While such shifts have yet to be shown for permafrost ecosystems, the first observations in Pleistocene Park hint at positive long-term effects of big herbivore introduction on carbon sequestration. These insights and also questions emerging from ongoing studies call for future research concerning the influence of big herbivores on the carbon balance of permafrost ecosystems, with a focus on long-term, year-round monitoring.