Winter microbial respiration in soils drive vertical fluxes in GHGs across the soil-atmosphere interface by the production or consumption of CO₂ and CH₄ within the soil column. Snow acts as a porous intermediary medium between soil and atmosphere for gases to traverse, creating a vertical concentration gradient driven by molecular diffusion. Assuming steady state and concentration gradients across the snowpack are linear, then the rate of diffusion of gas (Fgas, mol m⁻² s⁻¹) can be calculated using Fick's 1st law of diffusion as: 1Fgas=-φτDdCdzNa where D (m⁻² s⁻¹) is the standard air diffusion coefficient of each gas species (here, DCO2 and DCH4 for CO₂ and CH₄, respectively) through the snowpack, dC / dz (mol m⁻¹) is the vertical gas concentration gradient across the snowpack of height z, φ (m³ m⁻³) is snowpack porosity across dz, τ (m m⁻¹) is snowpack tortuosity across dz, and Na (mol m⁻³) is the molar volume used to convert units of concentration (from ppm to mol m⁻³). According to the ideal gas law (PV = nRT, where R=8.314 J K⁻¹ mol⁻¹), Na (n/V) is calculated as: 2Na=PRT where P (Pa) is ambient air pressure and T (K) is the average temperature of the snowpack. Snow porosity is derived from measured snow density along its vertical profile following Kinar and Pomeroy (2015): 3φ=1-ρsnowρice where no liquid content is assumed to be present in the snowpack, ρsnow is measured snow density and ρice is pure ice density as a function of measured temperature at standard pressure (ρice=-0.0001 × T+0.9168 g cm⁻³; Haynes, 2016), where T is temperature in °C. Tortuosity through a porous medium is derived from Du Plessis and Masliyah (1991): 4τ=1-1-φ2/3φ

Standard CO₂ and CH₄ diffusion coefficients must be corrected for air temperature and pressure (Massman, 1998): 5D=D0P0PTT01.81 where P0 and T0 are standard atmosphere pressure and temperature respectively (P0=101.325 kPa and T0=273.15 K), and D0 is gas diffusion in air at those standards where: D0,CO2= 0.1381 cm² s⁻¹ (1.381 × 10⁻⁵ m² s⁻¹); D0,CH4= 0.1952 cm² s⁻¹ (1.952 × 10⁻⁵ m² s⁻¹).

Non-diffusive transport driven by wind pumping is represented through an effective diffusion coefficient altering Eq. (1) to: 6Fgas=-1+ϵφτDdCdzNa where ϵ is a dimensionless enhancement factor that is zero for pure molecular diffusion and increases with snow-surface wind speed, snowpack air permeability, and barometric variance (Massman et al., 1997; Bowling and Massman, 2011).

Turbulent and barometric pressure pumping is oscillatory with near-zero mean flow, so the net effect is a dispersive enhancement of diffusion rather than a mean advective flux (Massman et al., 1997; Bowling and Massman, 2011), whereas quasi-static (Venturi) advection depends on the wavelength and height of surface relief, such as sastrugi or drifts, and on the resulting surface pressure perturbation (Colbeck, 1997; Massman et al., 1997). Bowling and Massman (2011) directly measured ϵ with values reaching zero in calm air and up to 0.33 at a wind speed of 10 m s⁻¹. Modelling of a layered snowpack driven by such fluctuations at mean wind speeds of 1 to 4 m s⁻¹ shows that pressure pumping does not appreciably enhance through-snowpack diffusion unless it alters the boundary concentrations (Massman et al., 1997). Furthermore, a lower-permeability surface layer such as a wind crust further attenuates the penetration of pressure-driven airflow into the snowpack (Colbeck, 1997). Advective enhancement is therefore bounded near zero (ϵ≈ 0) under which conditions the flux reduces to its diffusive form as Eq. (1) when wind speed is low, the snow surface lacks pronounced relief, and a wind crust is present. As antecedent wind-generated pressure gradients were not directly measured within snowpacks over periods of hours to days before measurement of concentration gradients, their impact on basal snowpack CO₂ concentrations (e.g. Massman and Frank, 2006) cannot be excluded completely. Nevertheless, the contribution to barometric pressure pumping from small atmospheric pressure changes during the measurement period are expected to be negligible.

Measurements were undertaken in a 3 km² area surrounding the Wilfrid Laurier University Trail Valley Creek Research Station, (TVC; 68°74^′ N, 133°50^′ W), approximately 45 km north of Inuvik, in the Inuvialuit Settlement Region of western Inuit Nunangat, Northwest Territories, Canada (Fig. 1). TVC has a mean annual precipitation of 210 ± 9 mm for the 1991–2020 period, with approximately two thirds of annual precipitation falling as snow. Mean annual air temperature for that same period is -8.1 ± 0.7 °C (Environment and Climate Change Canada, 2024). Dominant westerly wintertime winds produce stable snow redistribution patterns across years trapping snow in areas with taller shrubs, and into drifts on eastward facing slopes where large eddies form forcing snow deposition (Wilcox et al., 2019). Typically, snow covers persist from mid-October to mid-May depending on accumulation patterns (Wilcox et al., 2019; Pomeroy et al., 1993), with seasonal maximum snow depths ranging from 45 ± 10 cm in riparian shrub and sparsely wooded areas to 71 ± 15 cm in upland tundra and tussock sites (Tutton et al., 2025).

Topography is characterised by gently rolling hills of glacial till, overlain by silty clay and a thin organic top layer. The area is underlain by 350 to 500 m thick ice rich continuous permafrost (i.e., > 90 % in areal extent; Gruber, 2012), generally with active layer thickness varying between 0.5 and 0.8 m (Burn and Kokelj, 2009) and up to 1.5 m on some southward facing white and black spruce (Picea glauca, Picea mariana) populated slopes and below stream channels. Landscape microtopography is dominated by mineral soil hummocks of up to a metre in diameter and peaty inter-hummock hollows with some periglacial tundra landforms and associated cover types including ice-wedge polygons and barren ground (Quinton and Marsh, 1998).

Vegetation is composed of low stature shrub tundra composed of dwarf birch and willow (Betula nana, Betula glandulosa; S. arctica), taller alders surrounding drainage channels (Alnus sp.), graminoids and cotton grass tussocks (Eriophorum spp.), patches of lichen and mosses, with some areas of sparsely populated black spruce (Graveline et al., 2024). Shrub-dominated areas are seen as greener areas, and low-lying lichen and graminoid-dominated vegetation in lighter tones (Fig. 1).

All data were collected late winter between 20–26 March 2024 near maximum annual snow water equivalent, but before any snow surface melt had taken place and air temperatures remained below -10 °C. We identified five sampling transects, at length scales of approximately 100 m (T1 to T5; Fig. 1), within different vegetation types representative of the Trail Valley Creek site (Grünberg and Boike, 2019), and snow depth distribution patterns reported in Walker et al. (2021). Topography and vegetation types were used as a visual proxy for spatial delineation of the combined influence of snow depth and soil properties (temperature, moisture and organic content) on potential for carbon fluxes. To account for any preferential orientation of snow resulting from wind redistribution, sampling transects were split into two sections of 100 m, orthogonal to each other. Snow free images of transect vegetative cover are shown in Fig. A1.

Vertical snowpack structure profiles and CO₂ and CH₄ concentration profiles were sampled along five 200 m transects with 10 m spacing between each microsite along each transects (Fig. 1, described in Table 1). Three 1 m spaced samples also taken at the 50 and 150 m mark of each transect (at the midpoint of each section) to further identify highly localised variance at metre-scales. In total, 119 (Ntot) co-located snowpack structure profiles and CO₂ and CH₄concentration profiles were collected across the study site.

Vertical profiles of snow density, depth, and specific surface area were measured at sub-millimetre resolution using a Snow Micropenetrometer (SMP; Proksch, Löwe, and Schneebeli 2015). Raw SMP measurements were processed using the snowmicropyn (v. 1.2.1) python package with penetration force to snow density relationship coefficients adapted by Dutch et al. (2022) to the arctic snowpacks found in TVC.

Snowpack CO₂ concentration profiles were measured using a temperature controlled portable IRGA (LGR U-GGA-915 Ultraportable, Los Gatos Inc, Los Gatos, CA) deployed in a toboggan. Snow depths at each profile were calculated as the average of three measurements taken within 100 mm of each microsite profile (coloured icons, Fig. 1). Gases within the snowpack were sampled at increments of 100 or 200 mm depending on total thickness of the snowpack, with a minimum of 6 measurement depths per profile. Snowpack air was drawn to the analyser through a metal tube on a graduated avalanche probe by the instrument's fixed internal pump at approximately 0.8 standard L min⁻¹, and each depth was measured for approximately 60 s until CO₂ and CH₄ concentrations reached a stable plateau. The analyser is a flow-through instrument; sampled air was vented to ambient air rather than returned to the snowpack. The volume drawn per depth (∼ 0.8 L over 60 s) is a few percent of the pore-air volume of the surrounding snow, of the order of tens of litres within a radius of at most 0.3 m from the intake at the measured porosities (φ≈ 0.65 to 0.78). Cumulative depletion was therefore limited to the volume within each vertical layer measured. Ambient air concentrations were also recorded at the start of each profile within the transect drawn from approximately 1.5 m height to avoid contamination.

Gas concentration profiles were recorded as one continuous IRGA time series per transect accompanied by manually recorded total profile depths, creating a timeseries of step changes in concentration for each depth where air was sampled in the snowpack. IRGA time series were then partitioned into individual profiles with defined concentrations per depth in a MATLAB environment, using the Signal Processing Toolbox findchangepts.m function (The MathWorks, 2024), forcing known number of step changes in mean gas concentration between the start and end times of each profile. Step changes were first identified on the CO₂ concentration time series as signal noise to change in concentration per step change was lower and then used to partition the CH₄ time series. Concentrations at each step were then associated to independently recorded depths for each profile. Snowpack CO₂ and CH₄ concentration gradients used to calculate soil-atmosphere CO₂ and CH₄ fluxes through the snowpack (FCO2 and FCH4, g C m⁻² d⁻¹, respectively were then calculated by fitting linear models at each profile and testing for significance of gradient (T-test) and model coefficient of determination (R2).

Flux estimates assume quasi-steady-state diffusive transport and linear concentration gradients through the snowpack, computed from the diffusive form of Eq. (1) only. To ensure the validity of this assumption and satisfy the conditions under which the wind-pumping enhancement is negligible (Sect. 2.1), all measurements were conducted during stable meteorological conditions, characterized by wind speeds below 3.5 m s⁻¹, and were restricted to microsites free of pronounced sastrugi or drift relief and capped wind crust free of ice lenses. Barometric pressure was measured at the Environment and Climate Change Canada weather station (Trail Valley Creek) located near transect T2-Mixed, and varied by at most 10 hPa (1.0 %) within each sampling day and by 37 hPa (3.7 %) across the campaign. Pressure changes exchanged no more than the upper 1 % of snowpack pore air with the atmosphere per day (≤ 13 mm for the deepest snowpack), constituting a negligible advective contribution under low-frequency barometric forcing.

Soil-snow interface temperatures (basal temperatures) and snowpack temperatures sampled at 10 cm vertical intervals were measured using a handheld type K thermocouple probe. Missing basal temperature observations due instrument malfunction were inferred using a linear relationship between snow depth and soil temperature (missing = 53, R2=0.76, RMSE = 1.47 °C).

Measurement uncertainties were propagated to FCO2 and FCH4 according to the generalised error propagation method for nonlinear systems (Taylor, 2022). Gas concentration profile gradient uncertainty is defined as the linear fit 95 % confidence interval using the Wald method for binomial series, assuming that measurements used to calculate gradients are exact as the IRGA concentration measurement error was negligible compared to the error on profile gradients. Error on average profile snow density was defined as the standard deviation of binned densities in 50 mm vertical increments, where error-per-binned densities were propagated to the total average. Soil temperature, snow temperature, and air pressure measurements errors were also fully propagated to FCO2 and FCH4 using the generalised error propagation method.

Snowpack concentration gradients derived from all vertically spaced measurements in each microsite profile were compared to sub-sampled vertical gradients calculated from; (a) two concentration measurements: one within the air-snow interface layer of snow and one at the bottom of the snowpack (two measurements per microsite), and (b) one measurement from the bottom of the snowpack and the average concentration from ambient air above snow surface for the whole transect (one measurement per microsite and an average of ambient near-surface air).

To assess gas flux spatial variability across transects (homogenous surface covers), robust median estimates with acceptable variances are required to calibrate and test TBMs. However, the ideal number of samples per surface cover cannot be know in advance. Here, we used a bootstrap method where median flux values (Md_full) of the total number of flux measurements (N) from each surface cover transect were compared to median flux values from incrementally smaller subsamples per transect. Subsample size decreased from a total number of flux measurements from each transect n=N-1 to n=2 (T1 N=19; T2 to T5 N=25). Subsamples with n-values less than the maximum in each transect were randomly selected from the total possible. Random selection was repeated 10 000 times, the median calculated for each selection, and then the median of all 10 000 selections was calculated (Md_n). Differences between the medians of subsamples and all measurements in a transect are expressed as a ratio (Md_n / Md_full). Subsampled Md_n were considered different to the median of total number of flux measurements from each transect (Md_full) when the ratio Md_n / Md_full fell within the 19.4 % relative uncertainty reported for the CO₂ concentration-gradient (d[CO₂]/dz) regression of manually sampled snowpack fluxes at Trail Valley Creek by Mavrovic et al. (2023). This tolerance is a method-based precision band rather than a statistical significance level, adopted so that the subsampling result is directly comparable with the accuracy achievable by the gradient method. Profiles exhibiting non-significant concentration gradients (p>0.05) were excluded from subsampling and scaling analyses to avoid introducing noise from profiles indicating negligible net exchange. Profiles exhibiting non-significant concentration gradients were interpreted as representing near-zero flux conditions for the corresponding sampling period.

Snow density profiles exhibited typical Arctic tundra characteristics of high-density near-surface wind crust and lower density basal depth hoar layers (Fig. 2, a to f). Depth hoar layers were most pronounced in T1-Shrub and T2-Mixed, and T5-Inlet, with median wind crust densities of 297 ±70 kg m⁻³ and depth hoar densities of 223 ± 46 kg m⁻³. Greater spread in T3-Tussock mid-snowpack densities may indicate several harder snow layers forming during consecutive high wind events as the snowpack built over winter. Snowpacks were dry and free of discrete ice lenses or basal ice layers at all microsites throughout transects.

Pronounced microtopography and shallow snow in T3-Tussock created more variable density profiles, corroborating with observed variable wind crust presence and profile depth. T4 forest snowpack were deeper than for other transects but had similar profiles of high-density surface layers overlying lower density depth hoar (wind crust Md: 314 ± 84 kg m⁻³, depth hoar Md: 219 ± 40 kg m⁻³).

Snow depths and basal (soil-snow interface) temperatures (Fig. 2f) were strongly correlated (Pearson's Linear Correlation Coefficient, r=0.90, α< 0.01). Transect scale basal temperature and snow depth medians were significantly different (alpha < 1 %) to each other, except for T1-Shrub and T5-Inlet that had similar distributions. T3-Tussock basal temperatures were lowest with lowest snow depths (Md: -9.5 ± 0.8 °C, 440 ± 118 mm), contrasting with T4-Forest with highest basal temperatures and snow depths (Md: -2.7 ± 2.6 °C, 1280 ± 343 mm). Basal temperature and snow depth distributions in other transects are all found between T3-Tussock and T4-Forest ranges, with the T2-Mixed significantly colder (α< 0.05) than T1 and T5. Transect density, basal temperature, and snow depths are summarised in Table 2.

Individual snowpack CO₂ and CH₄ concentration profile analysis revealed that most profiles across transects T1-Shrub, T2-Mixed, T4-Forest, and T5-Inlet were explained with high certainty by simple linear models. Transect T3-Tussock microsites contained weak GHG gradients, where the majority of snowpack CO₂ and CH₄ concentrations were not significantly different to atmospheric concentrations with 80 % (CO₂) and 56 % (CH₄) of gradient p-values > 0.05. Although well explained by linear models (gradient p-value < 0.05), some profiles display nonlinearity, suggesting the influence of snowpack vertical structure for some locations.

Transect scale concentration profile linearity were also examined by aggregating normalised CO₂ and CH₄ concentrations as a function of depth (Fig. 3). Linear models were fitted to each normalised gas species (CO₂ and CH₄) with model 95 % confidence intervals illustrating the level of certainty about the model. Normalised T1-Shrub concentration gradients show positive linear trends for both CO₂ and CH₄, with variance well explained by the model (R2 > 0.75), indicating a strongly linear trend across all profiles. T2-Mixed has similar relationship to T1-Shrub but shows more spread (R2> 0.55), suggesting either more variance in flux magnitudes or less linearity in concentration profiles. T3-Tussock shows positive gradients for both gas species, with CO₂ variance being poorly explained (R2 < 0.1) while CH₄ shows a tighter albeit poor fit (R2=0.37). T4-Forest concentration profiles are the most linear for both CO₂ and CH₄ with very little spread around the linear model (R2> 0.85), indicating the least dependence to vertical snow structure. Contrarily to all other transects, All T4-Forest CH₄ vs. depth gradients are negative, indicating a net CH₄ uptake in the forested area. T5-Inlet CO₂ profiles are also strongly linear (R2=0.75), while CH₄ show more spread (R2=0.35), with 76 % of profile gradients significant (p-value < 0.05). Similarly to T3-Tussock, several profiles across T5 had CH₄ concentrations near atmospheric levels, with a few profiles with high basal concentrations. T5-Inlet was the most varied in terms of land cover type, i.e., dry shallow snowpack slope, and wet high snowpack drainage channel.

Instantaneous CO₂ and CH₄ fluxes (FCO2 and FCH4, respectively) through the snowpack for each vertical profile were calculated using paired snow gas concentrations and snow density profiles, aggregated by transect (Fig. 4). Distributions of CO₂ fluxes closely correspond with snow depth distributions, which strongly control soil surface temperatures. Distributions of fluxes between transects were all significantly different to one another except for T1-Shrub and T5-Inlet. Very low fluxes at T3 coincide with the shallowest snowpacks and coldest basal temperatures. Median CO₂ and CH₄ fluxes at T4-Forest were an order of magnitude higher than other transects, coinciding with the deepest snowpacks and warmest basal temperatures. Five high CO₂ emitting outliers were found in T2-Mixed, and T3-Tussock, T4-Forest.

Methane fluxes through the snowpack did not follow the same pattern as CO₂ fluxes Instead, significant net CH₄ uptake (-FCH4) was found throughout T4-Forest where snowpacks are thickest and basal temperatures in frozen soils are warmer than -3 °C. Low positive CH₄ fluxes were found across T2-Mixed with T1-Shrub and T3-Tussock, including some high emitting profiles where net fluxes were one order of magnitude higher than profile medians. T5-Inlet CH₄ fluxes had most variability with weak net uptakes and strongly emitting profiles which show little correlation to snowpack depth and basal temperatures (R2=0.06 and -0.02, respectively), contrary to patterns in CO₂ fluxes (Table 2).

We found a significant relationship between T4-Forest microsite rates of net CH₄ uptake (negative FCH4) and FCO2 (Fig. 5). Although significantly related (p-value < 0.01), T4-Forest FCO2 to FCH4 ratios are still poorly explained by magnitudes alone with an R2=0.14. Spread around the linear model is partially explained by basal temperatures, FCO2 increasing proportionally more with warmer soil surfaces than CH₄ uptake rates.

Measured CO₂ fluxes followed an exponential relationship to basal temperatures, with R2 of 0.79 and RMSE of 0.020, most situated within the model's 95 % regression parameter confidence interval (Fig. 6). A two-parameter exponential regression curve was fitted on FCO2 and basal temperature, and compared to relationships found elsewhere (Natali et al., 2019; Mavrovic et al., 2023). While within this study's 95 % regression parameter confidence interval, Mavrovic et al. (2023) overestimated fluxes with basal temperatures below -4 °C relative to our study, whereas Natali et al. (2019) overestimated low temperature CO₂ fluxes (as noted in Mavrovic et al., 2023). This may be due to our study including a substantial number of fluxes measured at temperatures of -6 to -0.5 °C, which were insufficiently represented in previous datasets.

Methane fluxes have less distinct relationship with soil surface temperature than CO₂ fluxes. T1-Shrub, T2-Mixed and T3-Tussock microsites were all below -6 °C and showed very little CH₄ emissions, with two high emitting microsites (up to 0.58 mgC m⁻² d⁻¹ or 0.56 nmol m⁻² s⁻¹) in the coldest transect T3. All fluxes measured in T4-Forest and had basal temperatures were above -5 °C and showed weak net CH₄ sinks between 0 and -0.1 mgC m⁻² d⁻¹ (-0.10 nmol m⁻² s⁻¹). Although within the same temperature range as T1-Shrub and T2-Mixed (-8 °C < T < -4.5 °C), T5-Inlet had the highest CH₄ emissions with three profiles displaying uptake of the same magnitude as T4-Tussock. T5-Inlet had the most diverse surface vegetation and snow depths across its sampling transect, with the first 100 m of the transect situated on a well-drained and wind scoured eastward facing slope covered with dwarf birch, and the second 100 m along the lake inlet with taller alder shrubs, tussocks and several micro drainage channels which are generally saturated with water across the shoulder and growing seasons.

We separately calculated FCO2 and FCH4 from only basal and near surface, and from snow basal and average ambient air samples to assess advantages and need for snow gas profiles. Both subsampling strategies tightly predicted fluxes similar to those fluxes calculated using multiple evenly spaced snowpack samples (Fig. 7). Both subsampling methods showed near 1:1 ratio for FCO2 and FCH4 with R2> 0.99 and RMSE two orders of magnitude smaller than FCO2 and one order of magnitude smaller than FCH4. As two-sample and one-sample methods show similar linear relationships, concentration gradient and consequently vertical fluxes have been shown to be almost entirely driven by the snowpack basal concentration.

Transect subsampling comparisons show that FCO2 Md_n variability remains within 90 % of the fully sampled dataset Md_full with n>9 in transects T1-Shrub, T4-Forest, and T5-Inlet, and n > 16 for T3-Tussock (Fig. 8). T3-Tussock subsampled FCO2 Md_n did not stabilise within subsamples suggesting surface cover variability within T3 may require a larger sampling size than we conducted (n > 25). FCH4 transect Md_n required more subsamples to capture surface cover scale variability in Md_full, with n> 12 for T2-Mixed and T4-Forest, while other transects required n > 20 when comparing means but did not reach a stable plateau within N samples when comparing medians. TVC site scale analysis (black lines Fig. 8a and b) shows that FCO2 were within 90 % of Md_full with transect n> 12 for medians and n> 13 for means, while CH₄ failed to be adequately reproduced with subsamples below N.

We consistently observed positive (soil-to-atmosphere) CO₂ emissions (FCO2) at all microsites with significantly different rates between surface cover transects. Instantaneous FCO2 were similar to those derived using syringe extraction methods by Mavrovic et al. (2023) at TVC during the late 2022 and 2023 winters, and comparable to pan-Arctic measurements using comparable in-situ concentration measurements or a wide range of other techniques including eddy covariance and flux chambers (Natali et al., 2019; Pirk et al., 2016; Webb et al., 2016). However, our study resolved a paucity of Arctic FCO2 at soil temperatures between -6 and 0 °C. Filling this measurement gap reduced error in exponential model fits by 0.01 and 0.004 gC m⁻² d⁻¹ for Natali et al. (2019) and Mavrovic et al. (2023), respectively (Fig. 6a). A wide basal temperature range and lower FCO2 uncertainty at temperatures below -6 °C, constrains exponential FCO2 response to soil temperature, indicating FCO2 tends to zero at lower temperatures observed at TVC (<-9 °C).

Measured winter CO₂ flux magnitudes at TVC are consistent with early observations from Arctic Alaska showing sustained cold-season emissions under snow cover (Fahnestock et al., 1998, 1999; Jones et al., 1999b; Welker et al., 2000). Subsequent work has demonstrated that snow depth, thermal insulation, and active-layer dynamics exert strong controls on winter respiration and the age of respired carbon (Lupascu et al., 2018; Pedron et al., 2023; Schimel et al., 2006; Sullivan et al., 2008). Measurements at TVC extend these findings by resolving fine-scale spatial variability in winter fluxes and by evaluating how simplified sampling strategies capture this variability under dry snowpack conditions.

Significant differences in the magnitudes of FCO2 between transects coincide with spatial patterns of basal temperatures, indicating CO₂ emissions from each surface cover are dominated by microbial respiration in frozen soils (Pedron et al., 2022). Consistently high emitting microsites in T4-Forest are typical of fluxes observed in permafrost free northern boreal forests with similar snow depths (Mavrovic et al., 2023). Near zero basal temperatures at T4-Forest suggest a portion of the vertical soil column might remain at or near zero-curtain conditions, a narrow temperature range (0.05 °C) in which changes in heat energy alters the phase of water without changing soil temperature (Zona et al., 2016; Lupascu et al., 2018; Devoie et al., 2022). FCO2 outliers (Fig. 4) at low surface temperatures (basal temperature <-15 °C), may have resulted from wintertime microbial activity in localised warmer soil pockets, or slow release of trapped CO₂ produced during autumn freeze-up similarly to those found in other tundra sites at similar latitudes (Pedron et al., 2022). Talik formations have been reported in dry tundra environments by Kim et al. (2021) at similar latitudes (Cambridge Bay, NU Canada), but have yet to be verified in TVC.

Consistent net uptake measured in the T4-Forest transect between -6 and 0 °C reveals that high-latitude forested soils can function as significant net sinks even under frozen conditions, representing an integrated balance of diffusive transport and microbial methanogenic and methanotrophic turnover within soils (Mast et al., 1998). Few measurements of wintertime CH₄ uptake specific to land surface types have been reported in high-latitude tundra environments, mainly due to the historical difficulty of resolving low-magnitude fluxes through Arctic snowpacks. Recent studies have begun to address this gap, with winter CH₄ uptake rates in northern closed-crown boreal forest uplands recently reported at -0.43 ± 0.34 to -0.47 ± 0.26 mg C m⁻² d⁻¹ (Mavrovic et al., 2025). For comparison, our observed winter forest uptake (-0.04 to 0.08 mg C m⁻² d⁻¹) represents a persistent, albeit diminished sink relative to peak growing season consumption. At TVC, automated chambers have recorded summer CH₄ uptake rates for lichen and shrub communities between -0.37 and -0.44 mg C m⁻² d⁻¹ (converted from -0.49 to -0.59 mg CH₄ m⁻² d⁻¹), suggesting winter forest uptake magnitudes are roughly one order of magnitude lower than peak summer rates in the same region (Voigt et al., 2023).

Landscape-scale tundra winter CH₄ sinks have also been observed in Svalbard, Greenland, and Norway, where soil organic layers are less spatially pervasive than in TVC (D'Imperio et al., 2023; Donateo et al., 2025; Treat et al., 2018). In Svalbard's dry upland tundra, median CH₄ fluxes remain negative throughout the winter, ranging from -0.18 to -0.37 mg C m⁻² d⁻¹ (reported as -0.17 to -0.36 nmol m⁻² s⁻¹), while notable net uptake events during the freezing shoulder season can reach -0.72 mg C m⁻² d⁻¹ (reported as -0.69 nmol m⁻² s⁻¹) (Donateo et al., 2025). Growing season and start-of-cold-season CH₄ uptake measured at TVC across drier landforms, including upland tundra and polygonal rims, further suggest that methanotrophic activity persists long after plant senescence (Ivanova et al., 2025; Voigt et al., 2023).The magnitude of this winter sink is particularly important when contrasted with comparatively wetter tundra sites in Northern Alaska, where summer emissions from even relatively dry polygon rims can range from 4.5 to 9.6 mg C m⁻² d⁻¹ (McEwing et al., 2015). Comparisons between different pan-Arctic land covers highlight that while mid-winter CH₄ uptake may be a small component of the annual budget, persistence of aerobic oxidation in well-drained forested and upland soils provides a non-negligible offset to winter CH₄ emissions from saturated hotspots.

Near-zero basal temperatures at T4-Forest suggest that much of the near-surface soil profile likely remained at zero-curtain temperature conditions, When soil temperatures are poised near 0 °C due to release of latent heat during the phase transition of water to ice, allowing liquid water content and ice to coexist in soil pore space (Devoie et al., 2022; Tucker, 2014; Outcalt et al., 1990). Remaining liquid soil water over winter sustains microbial metabolic activity in thin films surrounding soil particles even at temperatures as low as -18 °C (Dutch et al., 2024; Elberling and Brandt, 2003; Sullivan et al., 2008).

Northern boreal forest CH₄ uptake rates have been reported to be within the same order of magnitude as those found in T4-Forest, where soils also remained in zero-curtain conditions throughout winter (Lee et al., 2023; Mavrovic et al., 2025). Evidence for microbial CH₄ consumption within the T4-Forest is provided by the significant inverse relationship between FCO2 and FCH4 (Fig. 5). This inverse relationship aligns with global patterns where winter CO₂ emissions explain nearly 96 % of the variability in CH₄ oxidation, effectively serving as a proxy for the biological activation and oxygen diffusion required for substrate consumption (Voigt et al., 2023; Kim et al., 2024). High-affinity methanotrophs likely maintain this activity in the aerobic upper soil layers during mid-winter; however, the preferential coupling of basal temperatures with CO₂ rather than CH₄ flux magnitudes suggests that CH₄ consumption may also take place in deeper, relatively warmer soil layers (Donateo et al., 2025; Mikan et al., 2002). Further variability in CH₄ uptake magnitudes could also be explained by oxygen, and the liquid portion of soil moisture available in upper soil horizons have been identified as important regulators of uptake (D'Imperio et al., 2023; Lee et al., 2023; Rößger et al., 2022).

However, soil temperature measurements taken at the soil-snow interface may be insufficient to infer liquid water availability across the upper soil layers. As the active layer typically freezes faster from the surface downward due to large soil-atmosphere thermal gradients, interface temperatures may drop below ideal temperatures for microbial activity, while deeper carbon-rich layers remain in a zero-curtain state for weeks or months longer (Pedron et al., 2022; Yi et al., 2020). Relying solely on surface temperatures contributes to the cold-soil temperature biases in carbon fluxes observed in terrestrial biosphere models such as CLM5.0, which often fail to resolve the extended duration of these subsurface thermal refugia and resulting non-growing season carbon losses (Damseaux et al., 2025; Dutch et al., 2024; Tao et al., 2021).

We also measured CH₄ uptake in locations with notably cooler soils in T5-Inlet (temperature <-6 °C), which also contained the highest sitewide CH₄ emitting microsites. The T5-Inlet transect crosses a lake inlet area that is typically saturated or submerged during the autumn shoulder season, forming a patchwork of dry grassy hummocks, tall alder shrubs, and 10–40 cm deep drainage channels often permeating the active layer. While soil surface temperatures in T5-Inlet did not differ significantly from those in T1-Shrub, greater heterogeneity in snow depths and elevated spatial variability in soil water saturation in T5-Inlet may account for enhanced positive FCH4 and suppressed FCO2 typical of anoxic environments. Conditions for preferentially anoxic microbial activity are likely driven by prolonged thermal retention in saturated microsites during winter due to increased latent heat release during freeze-up, similarly documented in hydrologically comparable “wet” tundra systems in northern Alaska (Blanc-Betes et al., 2025).

Alternatively observed net CH₄ uptake in saturated landcovers may be limited by physical encapsulation of soil layers by ice, where basal ice or thick frozen soil layers creates an impermeable barrier occluding efflux of CH₄ produced in deeper, unfrozen soil horizons (Elberling and Brandt, 2003; van Bochove et al., 2001). Such diffusion-limited transport, exacerbated by the occlusion of soil pores by ice, can decouple subsurface CH₄ production from soil-atmosphere exchange, potentially resulting in apparent uptake signals if atmospheric CH₄ is consumed in the uppermost aerated soil or if production is physically trapped at depth (Mavrovic et al., 2025; Treat et al., 2018; Schimel et al., 2006). Slower CH₄ diffusion through ice-filled pores in frozen soils can further mask the temperature dependence of winter CH₄ fluxes, suggesting that surface net uptake may reflect a lack of transport from deeper source layers rather than high methanotrophic potential (Mavrovic et al., 2025). Consequently, the fine-scale variability at T5-Inlet likely reflects a gradient where some microsites maintain active methanogenesis while others are governed by the passive trapping of greenhouse gases within the soil profile.

Further efforts to partition FCH4 into its autotrophic and heterotrophic respiration components through isotopic analysis (e.g., Pirk et al., 2016) or via metagenomic sequencing may help identify the relative importance of soil surface temperature, liquid moisture availability and nutrient availability on CH₄ flux, and allow estimation of CH₄ specific temperature response curves under methanogenic or methanotrophic prevailing conditions. Identifying sampling locations prior to snowfall and optimizing repeated flux measurements to capture more variability in vegetation, soils, and hydrology, throughout the growing and autumn shoulder seasons would also allow process level understanding of abiotic and biotic drivers of flux variability, providing a more comprehensive dataset to upscale fluxes to landscape and regional spatial extents.

Overall GHG concentrations throughout snowpacks were linear in nature (Fig. 3). Surprisingly, denser upper snowpack wind crust layers did not appear to limit CO₂ and CH₄ diffusion rates. Consistently significant linear fits across transects independent of snow depth (except T3-Tussock, where fluxes were too small) demonstrate that the snowpack acted as a uniform porous material (i.e., independent of snow structure). Concentrations at distinct vertical sampling points concentration profiles were reproducible across replicate profiles and consistent between full-profile and two-point flux estimates (Fig. 7), indicating that extraction did not draw down or homogenise air from different layers or ambient air. As GHG concentrations follow a linear trend from basal to air above snowpacks, microsite level concentration gradients are controlled by their respective basal concentrations. Future GHG emissions could therefore be estimated using basal concentration variability within landscapes.

Robust agreement between full-profile fluxes and estimates derived from basal and near-surface concentrations was evident under the predominantly dry, cold snowpack conditions encountered at TVC. Where snowpacks lack extensive ice layers and experience limited mid-winter melt-freeze cycles, molecular diffusion will remain the dominant transport mechanism. Melt-refreeze events promoting the formation of spatially coherent ice lenses within snowpacks may cause diffusive transport to deviate from linear assumptions and decouple subsurface microbial production from surface exchange (Elberling and Brandt, 2003; Mavrovic et al., 2023). Soil stratigraphic barriers or basal snowpack ice lenses can trap gases produced in deep soil layers, potentially leading to zero-gradient fluxes within the overlying snow despite ongoing soil metabolic activity

Non-linear departures in snowpack gas concentration profiles, particularly homogenization of gas concentrations observed in basal depth hoar layers, introduce a known potential for gradient underestimation (Dutch et al., 2022). Such concentration gradient (dC / dz) underestimations, particularly in low carbon flux environments such as the T3-Tussock transect, may reflect a reduced concentration change over the sampled depth, resulting in FCO2 and FCH4 magnitudes that are lower than the actual soil metabolic flux rates or mechanical release. Wind-induced advection (wind pumping) can further perturb gradients in upper snow layers, although these effects are often transient and confined to periods of high wind speeds (Jones et al., 1999a; Mavrovic et al., 2025), although those effects were negated in this study by sampling while wind speeds remained below 3.5 m s⁻¹.

Previous studies have also identified non-linear profiles in snowpack GHG distributions when impermeable soil layers were present (van Bochove et al., 2001). In such cases, using snowpack basal concentrations would likely overestimate instantaneous flux magnitudes, and would affect estimates in the timing of emissions rather than seasonal budget estimates. While profiles at TVC were predominantly linear, significant gas trapping below frozen soil barriers can artificially increase concentration gradients once released through thermal or mechanical cracking, or conversely, mask subsurface production entirely (Jones et al., 1999a). Such diffusion-limited transport decoupling suggests that measurements taken strictly within the snowpack may periodically reflect physical encapsulation rather than active biological oxidation (Elberling and Brandt, 2003; van Bochove et al., 2001).

The application of Fickian diffusion models to wintertime greenhouse gas exchange assumes linear concentration gradients through the snowpack. Uncertainty bounds (Figs. 5 and 6) incorporate the standard error associated with linear regression of concentration versus depth, accounting for variance introduced by non-linear yet significant vertical profile concentration gradients. Robustly quantification of uncertainty for FCO2 and FCH4 from both instrument measurement error and linear model uncertainty was then made possible. High resolution snow density profile measurements (SMP) and high precision in situ gas measurements using a portable IRGA allowed a tighter relationship between FCO2 magnitudes and basal temperatures than comparable studies and gave statistical confidence on reported FCH4 magnitudes in a low flux environment. Some profiles only became linear at 100 to 300 mm above the ground surface with concentrations homogenised below these depths. Profiles where CO₂ and CH₄ concentrations were more homogenous in the depth hoar layer likely caused gradient underestimation thus leading to lower FCO2 and FCH4magnitudes. Conversely, bulk snowpack density and temperature values derived from the entire snow column could lead to FCO2 and FCH4overestimation due to an underestimation of effective snow porosity and tortuosity, making the snowpack less resistant to vertical molecular diffusion. In either case, we recalculated FCO2 and FCH4 using per-microsite range of densities and air temperatures and found resulting instantaneous flux variability to be within reported propagated flux uncertainty.

We found CO₂ and CH₄ concentration measurement error from the portable IRGA to have negligible effect on FCO2 and FCH4 uncertainties, which were mostly controlled by snowpack gas concentration gradient confidence intervals. Flux uncertainties are often reported but rarely well constrained, whether from microsite scale measurements due to low sample size, or from integrated landscape scale flux measurements such as eddy covariance, where systematic errors are difficult to separate from random measurement uncertainty (Levy et al., 2022). The importance of accurate uncertainty quantification in greenhouse gas flux measurements is underscored by (Fisher et al., 2014), who highlighted the significant uncertainties in carbon cycle models for the Alaskan Arctic. Similarly, (Björkman et al., 2010) demonstrated that different methods for estimating winter CO₂ effluxes can yield results that vary by up to two orders of magnitude, emphasizing the need for robust error analysis and uncertainty calculation. Measurement error propagation to FCO2 and FCH4 uncertainties presented here are nevertheless within similar magnitudes as snowpack gradient method derived fluxes reported by (Mavrovic et al., 2023, 2024).

Understanding the distribution of spatial variance in CO₂ and CH₄ soil-atmosphere fluxes across different land surface types is vital to enable up-scaling to the landscape scale, where flux uncertainties must account for surface cover aggregation error. Past studies reporting winter fluxes have either compromised spatial coverage for temporally continuous data (Zhu et al., 2014; Seok et al., 2009), or report spatially aggregated landscape scale fluxes that pose challenges to partitioning into their component surface covers (Levy et al., 2022; Pirk et al., 2017). Fast, instantaneous CO₂ and CH₄ concentration measurements using a roving, high-precision portable IRGA and SMP allowed robust quantification of spatial variability within surface cover types. This is particularly relevant to upscaling small-scale fluxes from tens of metres to scales typically used in Terrestrial Biosphere Models (TBMs), where finer-scaled data is aggregated by averaging discrete data points to much larger grid cells (Levy et al., 2022).

Measurements from TVC can directly inform TBMs by providing high-resolution empirical bounds for temperature sensitivity (Q10) of cold-season biogeochemical processes as a primary source of uncertainty in Arctic carbon simulations (Dutch et al., 2024). The relationship between late-winter soil temperature, snow depth, and FCO2 suggests that deterministic models using grid-cell mean snow depths likely underestimate winter activity by failing to capture the disproportionate contribution of FCO2 from warmer frozen soils under deeper snow. Our findings suggest that bulk snowpack diffusivity remains a robust approximation for models, provided that refined snow thermal conductivity parameterizations are implemented to correct the cold soil temperature biases that currently inhibit accurate flux simulations. A need for revised temperature-response curves in Arctic tundra extends to CH₄ dynamics; where persistent uptake observed in the forested transects indicates that current TBM methane production and oxidation schemes require better low temperature calibration to account for microbial oxidation occurring at soil temperatures below 0 °C. Following the process-based evaluation by Dutch et al. (2024), which highlighted how default model configurations often simulate near-zero winter fluxes at TVC due to overly restrictive soil moisture thresholds (Ψ_min) and respiration temperature responses, a transition is recommended toward sub-pixel distribution-based parameterizations (e.g., probability density functions) that can represent sub-grid heterogeneity in snow insulation and soil thermal regimes.

Errors in the estimation of spatial variability using methods such as Bayesian Kriging can be constrained by informing the number of measurements needed to quantify spatial variability within homogenous surface covers (Levy et al., 2022). A subsampling experiment (Figure 8), demonstrated that at TVC the homogenous surface cover scale FCO2 variance can generally be captured with n >=12 sampling sites, while larger sample sizes were needed for FCH4 (n > 25). Both median and average FCO2 and FCH4 at T3-Tussock failed to be reproduced with fewer subsamples because fluxes were either one order of magnitude smaller than those found in other transects or were not significantly different to zero. T5-Inlet covered both a well-drained low snow slope and a flat high snow and historically wet drainage area, exhibiting both high CH₄ emitting and low uptake microsites. Although closely grouped spatially (within 150 m), T5-Inlet's more heterogeneous snow depths and soil moisture (as well as associated nutrient availability), in contrast to other transects most likely explained why FCH4 average and median were not well reproduced with subsamples. This suggests future scaling attempts should consider a data informed method, such as multivariate spatial interpolation (and prior information) to divide up land surface for spatial flux upscaling.

Vertical snow structure had little effect on concentration profile linearity (and thus FCO2 and FCH4), and surface cover scale flux could confidently be derived from concentrations near the soil-snow and air-snow interfaces (Fig. 7). We show no significant difference between calculating soil-atmosphere fluxes from basal snowpack CO₂ and CH₄ concentrations and ambient air above the snowpack, compared to calculations from a larger number of equally spaced samples throughout the snowpack. Together these findings reduce the measurement burden needed for rapid in situ instantaneous flux measurement, opening up potential for landscape-scale FCO2 and FCH4 winter budgets at larger spatial scales using remotely sensed snow properties while expanding measurement spatial coverage improving quantification of variability over a wider set of surface covers (Dutch et al., 2024; Ivanova et al., 2026; Virkkala et al., 2022).