A coupled hydrology and carbon model was used to investigate the sensitivity of the soil thermal regime and soil carbon decomposition to changes in surface air temperature and snow cover conditions. The hydrology model accounts for the effects of soil organic layers, changes in surface snow cover properties and soil water phase change on the soil freeze–thaw process in permafrost landscapes (Rawlins et al., 2013). These factors represent important controls on soil thermal dynamics within the active layer (Nicolsky et al., 2007; Lawrence and Slater, 2008, 2010), enabling an improved estimation of subsurface soil temperature and moisture profiles, particularly in permafrost areas, and a representation of essential environmental constraints on soil carbon decomposition.

The hydrology model used for this investigation is an extension of previous efforts regarding large-scale pan-Arctic water balance modeling (PWBM; Rawlins et al., 2003, 2013). Recent updates to the model include an improved simulation of snow or ground and subsurface temperature dynamics using a 1-D heat transfer equation (Rawlins et al., 2013) instead of the empirical thaw depth estimation based on the Stefan solutions used in Rawlins et al. (2003). The updated PWBM model has 23 soil layers down to 60 m below surface, with increasing layer thickness at depth. Up to five snow layers are used to account for the effects of seasonal snow cover evolution on the ground thermal regime, and changes in seasonal snow density and thermal conductivities are also considered. Other model improvements include accounting for the impact of soil organic carbon content on soil thermal and hydraulic properties (Appendix Sect. A1, Eq. A3); this impact is an important feature of boreal and Arctic soils (Lawrence and Slater, 2008). Further details on the updated hydrology model are provided in Appendix Sect. A1.

A satellite-based terrestrial carbon flux (TCF) model (Yi et al., 2013) was coupled to the hydrology model for this investigation. The TCF model uses a light use efficiency algorithm driven by satellite estimates of FPAR (fraction of vegetation canopy intercepted photosynthetically active radiation) to estimate vegetation productivity and litterfall inputs to a soil decomposition model. In the original TCF model, soil carbon stocks and respiration fluxes were estimated using a simplified three-pool soil organic carbon (SOC) decomposition framework with environmental constraints on soil decomposition rates derived from either satellite-estimated surface soil moisture and temperature fields (Kimball et al., 2009) or reanalysis data (Yi et al., 2013). This approach assumes that the major source of soil heterotrophic respiration (Rh) comes from surface (≤ 10 cm) litter and surface organic layers. However, the contribution of deeper soils to total Rh may be non-negligible, especially in high-latitude boreal and Arctic tundra landscapes with characteristic carbon-rich soils (Koven et al., 2011; Schuur et al., 2015). Therefore, in this study, we incorporated a more detailed soil decomposition model representing SOC stocks, extending to 3 m below the surface and representing differences in litterfall and soil organic matter substrate quality within the soil profile (Thornton et al., 2002). The resulting soil decomposition model used for this study includes three litterfall pools, three SOC pools with relatively fast turnover rates and a deep SOC pool with a slow turnover rate (Fig. S1 in the Supplement). The three litterfall pools were distributed within the top 20 cm of the soil layers; the three fast SOC pools were distributed within the top 50 cm of the soil layers, and the deep SOC pool extended from 50 cm to 3 m below the surface. Substantial SOC may be stored in permafrost soils below 3 m depth (Hugelius et al., 2014) and may potentially undergo mobilization with continued warming. However, this contribution to total land–atmosphere carbon (CO2) exchange was assumed negligible for the recent historical period examined (Schaefer et al., 2011) and was not considered in this study. Further details on the carbon model used in this study are provided in Appendix Sect. A2.

The modeling domain for this investigation encompasses the pan-Arctic drainage basin and Alaska, representing a total land area extent of approximately 24.95 million km2. The model was run at a 25 km Northern Hemisphere Equal-Area Scalable Earth Grid (EASE-Grid) spatial resolution and daily time step from 1979 to 2010. Further details on the model validation data sets and inputs used for this study are provided below.

A dynamic litterfall allocation scheme based on satellite NDVI data (Appendix Sect. A2) was used to prescribe the daily litterfall fraction through each annual cycle to account for litterfall seasonality, particularly for deciduous vegetation types (Randerson et al., 1996; White et al., 2000). The GIMMS3g NDVI bimonthly data were first aggregated to a monthly time step and then used to characterize monthly leaf loss and turnover rates of fine roots during the active growth period based on Eq. (A7). The monthly litterfall fraction was then evenly distributed at a daily time step within each month. This approach generally allocates more litterfall during the latter half of the growing season, while the model simulations show generally more soil heterotrophic respiration during the latter portion of the year (Fig. S3). A comparison of model simulations against tower measurements shows an overall improved net ecosystem exchange (NEE) seasonality relative to a previous TCF model application where litterfall was distributed evenly over the annual cycle (Yi et al., 2013).

We conducted a model sensitivity analysis to examine how the estimated soil thermal regime and SOC decomposition respond to changes in surface air temperature and snow conditions over the most recent 3 decades. Three sets of daily model simulations were run by (1) varying air temperature (T) and precipitation (P) inputs; (2) varying T inputs alone (temperature sensitivity analysis), and (3) varying P inputs alone (precipitation sensitivity analysis). Daily mean T (including daily mean and minimum temperature) and P climatology was first derived from the initial 3-year (1979–1981) WFDEI meteorological record and used in the model sensitivity runs. The daily climatology, based on 3-year (1979–1981) meteorological records rather than a single year (i.e., 1979), was used to minimize effects from characteristically large climate fluctuations in the northern high latitudes. For precipitation, we first created a monthly climatology from the daily record (1979–1981) and then scaled the daily WFDEI precipitation by maintaining the monthly climatology value (Lawrence and Slater, 2010): P′(y,m,d)=P(m)‾P(y,m)P(y,m,d), where y, m and d represent a particular year, month and day; P(m)‾ is the precipitation monthly climatology averaged from 1979 to 1981 and P(y,m) is the monthly total precipitation for a particular year and month; P(y,m,d) and P′(y,m,d) are the original and scaled daily precipitation, respectively, for a particular year, month and day. Due to a relatively short record (i.e., 1979–1981) and large variability in northern latitude precipitation, the ratio of P(m)‾P(y,m) may be too large for a particular month with very low precipitation rates. In this case, the daily precipitation was not adjusted to avoid unreasonable estimates. We then ran the model with different configurations of the daily surface meteorology data sets. Model simulations derived using the dynamic WFDEI daily surface meteorology from 1979 to 2010 (i.e., varying T and P) were used as the model baseline simulation. For the temperature sensitivity analysis, we ran the model using the dynamic daily WFDEI temperature records from 1979 to 2010 but holding P as the climatology value from 1979 to 1981. For the precipitation sensitivity analysis, we ran the model using the dynamic daily WFDEI precipitation records but with the T daily climatology. Since VPD is dependent on air temperature, we also created a daily VPD climatology (1979–1981) as an additional input to the carbon model, assuming negligible changes in relative humidity during the study period for the precipitation sensitivity analysis. There was no significant trend in solar radiation during the study period; we therefore used the historical (i.e., 1979–2010) solar radiation data for the three sets of simulations.

The model was initialized using a two-step process prior to the three sets of simulations. The model was first spun-up using the daily surface climatology (1979–1981) including T, VPD, and P for 50 years to bring the top 3 m soil temperature into dynamic equilibrium; the model was then run using the same climatology and simulated soil temperature and moisture fields over several thousand years to bring the SOC pools to equilibrium.

We mainly used correlation analysis to evaluate the climatic controls on simulated soil temperature and carbon fluxes. The outputs from the model baseline simulations (i.e., varying T and P) from 1982 to 2010 were used for this analysis. The period from 1979 to 1981 was excluded in order to reduce the impact of the spin-up process on model simulations. We first calculated the correlation coefficients between the time series of each climate variable and modeled soil temperature or carbon fluxes at each grid cell from 1982 to 2010. The resulting correlation coefficients were then averaged for each climate zone classified using the annual mean air temperature (1982–2010) and binned into 2.5 ∘C intervals. The climate variables used in the correlation analysis included air temperature, snow water equivalent (SWE) and snow cover extent (SCE). The model did not simulate SCE directly, and the SCE was estimated using the following equation: SCE=SNOWD0.1+SNOWD, where SNOWD is the simulated snow depth (m), and the surface roughness was set as 0.1 m (Lawrence and Slater, 2010).

The model simulations were generally consistent with observed daily carbon fluxes from the 20 EC tower sites across the pan-Arctic domain (Table 2), with mean R values of 0.84 ± 0.11 (SD) for gross primary productivity (GPP) and 0.63 ± 0.17 for NEE, and mean RMSE differences of 1.44 ± 0.50 g C m-2 d-1 for GPP and 1.04 ± 0.36 g C m-2 d-1 for NEE. The model results showed relatively large discrepancies with the tower-based carbon fluxes for tundra sites; however, large uncertainties are associated with the tower measurements in tundra areas due to the characteristically harsh environment and extensive missing data. The simulated temperature and moisture fields also capture the seasonality of the in situ surface (≤ 15 cm) soil measurements representing variable soil depths (not shown), despite large uncertainties in the surface meteorology inputs (particularly precipitation or snowfall) and soil parameters, including the definition of texture and peat fraction within the soil profile. Additional assessment of the model simulations was conducted using detailed in situ measurements at selected tundra and boreal forest validation sites (Table 1) as summarized below.

The model simulations compared favorably with in situ measurements at the tundra validation sites for surface soil temperature (R= 0.93, RMSE = 3.12 ∘C) and carbon fluxes, including GPP (R= 0.72, RMSE = 0.76 g C m-2 d-1) and NEE (R= 0.79, RMSE = 0.50 g C m-2 d-1) but had a relatively larger discrepancy during the winter when the model showed lower values of NEE (e.g., less CO2 emissions) than the measurements (December to February, DJF; Fig. 1). The simulated maximum soil thaw depth (∼ 50 cm averaged from 2008 to 2011) was also consistent with site measurements, ranging from 40 to 70 cm at three locations within the tundra validation site (Euskirchen et al., 2012). An apparent cold bias ranging from -2 to -5 ∘C in the simulated soil temperature during the fall and winter period of 2009 and 2010 (Fig. 1a) reflects lower model-simulated snow depth and associated reductions in thermal buffering between the atmosphere and underlying soil layers. This cold bias in the simulated soil temperatures results in early freezing of simulated soil water content (Fig. S4). Compared with the tower observations, the simulated daily surface soil temperatures generally show large temporal variations, particularly during the summer (June to August, JJA). There were also considerable differences among in situ soil temperatures at the different tundra sites. Summer (JJA) soil temperature at the wet sedge tundra location was generally lower than for the other tundra vegetation types, which may reflect higher soil water content and specific heat capacity and greater latent heat loss from evapotranspiration, leading to slower soil warming at this site. Overall, the model simulations compare well with the tower-observed carbon fluxes during the growing season but significantly underestimate NEE and soil respiration during the dormant season. Model underestimation of soil respiration during the dormant season may reflect less liquid soil water represented by the model under frozen (< 0 ∘C) temperatures than the tower measurements (Fig. S4) as well as a lack of model representation of wind-induced CO2 exchange between the atmosphere and surface snowpack (Lüers et al., 2014). The model generally shows earlier seasonal onset and offset of photosynthesis relative to the in situ measurements, while partitioning of the tower NEE measurements during the shoulder season may be subject to large uncertainties under partial snow cover conditions (Euskirchen et al., 2012).

The model simulations also compared favorably against observations at the boreal forest validation sites (Fig. 2), capturing observed seasonality in soil temperatures (R > 0.95, RMSE < 2.00 ∘C) at different soil depths and daily variations in tower-observed carbon fluxes for GPP (R= 0.89, RMSE = 1.24 g C m-2 d-1) and NEE (R= 0.73, RMSE = 0.65 g C m-2 d-1). Similar to the tundra sites, snow depth also has a large impact on simulated soil temperatures at the boreal forest sites but is subject to large uncertainties from both model snowfall inputs and forest canopy snow interception processes. The timing of simulated thaw and freeze of soil water at different depths is generally consistent with the tower measurements, with later seasonal thawing and freezing occurring in deeper soils (Fig. S5). The tower site soil moisture measurements show larger variability than the model simulations during the growing season and likely reflect differences in the model parameterization of surface moss or peat and mineral soil hydraulic conductivities relative to local site conditions. The model-simulated NEE fluxes during the non-growing season stem mainly from soil heterotrophic respiration and are largely consistent with the in situ tower observations, generally diminishing towards the end of the year and then gradually recovering with soil warming toward the onset of the growing season. Both the model and in situ tower NEE fluxes show large temporal variations during the growing season, largely due to GPP reductions caused by high vapor pressure deficits or water stress.

The model-simulated SCE was generally consistent with satellite-observation-based global climate data records documenting weekly SCE changes (Brown and Robinson, 2011; Fig. 3). The model simulations show a similar mean seasonal cycle as the satellite observations, with spring snowmelt mostly occurring from April to May and fall onset of seasonal snow cover occurring in October over the 1982 to 2010 record (Fig. 3a). The model-simulated SCE shows consistent changes with the satellite observations in spring, indicating realistic simulation of the snow melting process. However, the model generally underestimates SCE in the fall and winter. The model did not directly simulate SCE, which was calculated from simulated snow depth using an empirical equation (Eq. 2). Based on Eq. (2), the modeled SCE will never approach 100 %, while the satellite data indicates nearly complete winter snow cover over the study domain. Larger model SCE differences from the satellite observations are expected when the snow cover is relatively shallow and patchy owing to the relatively coarse spatial resolution of both model simulations and satellite observations. Moreover, the satellite SCE data set is presented as a binary classification at a weekly time step, which may not adequately depict transient SCE fluctuations under active surface melting and freezing processes in the fall (Kim et al., 2015).

The simulated permafrost area is generally consistent with reported estimates from previous studies. The simulated mean permafrost area from 1982 to 2010 is approximately 11.3 million km2, which is within the range of observation-based estimates (11.2–13.5 million km2) of the combined area for continuous (90–100 %) and discontinuous (50–90 %) permafrost extent over the northern polar region (≥ 45∘ N) (Zhang et al., 2000).

The simulated active layer depth (ALD) shows an overall increasing trend across the pan-Arctic domain over the 1982 to 2010 record (Fig. 4a, b). No strong bias was observed for the model ALD simulations compared to in situ observations for 53 pan-Arctic sites from the Circumpolar Active Layer Monitoring (CALM) program (Brown et al., 2000); these results showed a mean model bias of -9.48 cm, representing approximately 16.5 % of the estimated ALD but with low model correspondence (R= 0.31, p < 0.1) relative to in situ observations (Fig. S6). The discrepancy between model-simulated ALD results and in situ observations may be partly due to a spatial scale mismatch between the coarse-resolution model simulations and the local CALM site measurements, as well as uncertainties in the reanalysis surface meteorology data used as model forcings (Rawlins et al., 2013). Previous studies have shown large local spatial variations in ALD due to strong surface heterogeneity including microtopography, vegetation and soil moisture conditions (Romanovsky et al., 2010a, b; Mishra and Riley, 2014). Simulated widespread ALD deepening is consistent with generally decreasing snow cover extent in the pan-Arctic region (Fig. 4c). Simulated ALD trends over the 1982–2010 record range from -4.32 to 8.05 cm yr-1, with a mean value of 0.66 cm yr-1. A notable model ALD deepening trend occurs in discontinuous permafrost areas with relatively large mean ALD values. However, in portions of Alaska, the model simulations indicate slightly decreasing ALD trends across the study period (Fig. 4b), despite a strong reduction in the local snow cover extent (Fig. 4c). This mainly reflects a large decrease in the simulated snowpack (Fig. 4d) due to a decreasing trend in WFDEI precipitation or snowfall data, resulting in less thermal insulation of underlying soil, which may offset warming effects from decreasing snow cover extent.

The regional differences in snow cover effects on model-simulated ALD can be explained by different climatic controls on warm-season (May to October) soil temperatures. The correlation analysis between climate variables and warm-season soil temperatures (Fig. 5) indicates that surface warming has a dominant control on upper (< 0.5 m) soil temperatures in all climate zones, and also on deeper (≥ 0.5 m) soil temperatures in warmer climate zones (mean annual Tair > -4 ∘C). A deep snowpack has a strong warming effect on simulated deeper (≥ 0.5 m) soil temperatures in colder climate zones (mean annual Tair ≤ -4 ∘C) but with limited warming effects on surface soil temperatures across all pan-Arctic climate zones. Correspondingly, the effects of seasonal snow cover duration on model soil temperatures vary across different climate zones and soil depths. In colder climate areas, a longer snow cover duration has a relatively strong warming effect on deeper (≥ 0.5 m) soil temperatures but with negligible warming effects on surface soil layers. In warmer areas, a shorter snow cover season promotes warmer soils, particularly within surface soil layers, due to stronger air and soil thermal coupling. Additional analysis also indicates that earlier snow cover seasonal onset in the fall has a stronger warming effect on modeled soil temperatures in colder climate areas, while earlier offset of seasonal snow cover in the spring has a stronger warming effect on modeled soil temperatures in warmer climate areas.

The model simulations indicated that air temperature has an overall dominant control on the two main components of the NEE flux (i.e., net primary productivity, NPP, and Rh) across all pan-Arctic climate zones, while snow has a larger control on estimated annual NEE fluxes in colder climate areas (Fig. 6). These results indicate that warming generally promotes vegetation photosynthesis and soil heterotrophic respiration in the pan-Arctic region. However, a reduced positive correlation between NPP and air temperature in warmer climate zones (mean Tair > 0 ∘C) also indicates that warming-induced drought may reduce vegetation productivity to some extent (Kim et al., 2012; Yi et al., 2014). No significant correlation (p > 0.1) between NEE and air temperature was observed for most areas (mean Tair ≤ 5 ∘C) due to NEE being a residual between two large fluxes (i.e., NPP and Rh) with similar temperature responses. A predominantly positive correlation (mean R= 0.32; p < 0.1) between annual NEE and SWE in colder regions (mean Tair < -4 ∘C) is mainly due to a strong positive correlation (R > 0.60, p < 0.01) between SWE and NEE fluxes during the cold season (November to April; Fig. S7). A deeper snowpack promotes warmer soil conditions (Fig. 5b) and associated SOC decomposition and heterotrophic respiration, which contributes significantly to annual NEE, especially in colder climate areas (Zimov et al., 1996). No significant correlation (p > 0.1) between annual SCE or SWE and warm-season (MJJASO) carbon fluxes was observed.

While snow cover has a negligible effect on total estimated carbon fluxes during the warm season, it has a strong control on the composition of soil Rh (Fig. 7). An overall, deeper snowpack promotes soil decomposition and respiration from deeper (≥ 0.5 m) soil layers while inhibiting contributions from surface (≤ 0.2 m) soil layers, especially in colder climate areas. This response is due to a stronger warming effect of snow cover on deeper soil layers in colder areas (Fig. 5). Comparatively, even though air temperature has a strong control on total warm-season Rh fluxes, it has a limited effect on the contribution of different soil depths to total soil decomposition and respiration except in the warmer climate areas (mean annual Tair > 0 ∘C). In the cold season, a deeper snowpack also promotes soil decomposition in deeper (> 0.2 m) soil layers more than in surface (0–0.2 m) soil layers.

The model sensitivity analysis using different surface meteorology inputs indicated that warming and reduced snow cover extent promoted widespread ALD deepening across the pan-Arctic domain over the 1982 to 2010 record (Fig. 8). In Eurasia, strong winter warming reduced model-simulated SWE and SCE, while increasing winter precipitation generally increased SWE and SCE. In North America, regional trends in winter snowpack and SCE were more variable due to variable trends in winter air temperature and precipitation. Therefore, the resulting model-simulated trends in SWE and SCE based on varying temperature and precipitation inputs showed strong spatial heterogeneity across the pan-Arctic domain. The model sensitivity analysis based on varying temperature inputs alone indicated overall ALD deepening in permafrost areas, corresponding with widespread warming and reduced SCE. However, the sensitivity analysis based on varying precipitation alone showed more variable trends in the simulated ALD results. Areas with strong decreasing winter precipitation and snowpack trends, such as interior Alaska and eastern Siberia, showed a decreasing ALD trend, attributed to reduced snow insulation effects. The results also indicated that changing air temperature had an overall dominant effect on the simulated ALD trends, though changing precipitation also contributed to ALD changes in some areas.

The model sensitivity analysis indicated that varying precipitation accounts for more of the change in the simulated Rh contribution from different soil depths (i.e., soil Rh fraction; Figs. 9 and 10, and Fig. S8), which is consistent with the above results indicating strong control of snow cover on the soil Rh fraction at different soil depths. The model sensitivity results also indicated that changing air temperature has minimal impact on the simulated soil Rh fraction, while increasing (decreasing) winter snowpack in permafrost areas generally corresponded to increasing (decreasing) soil Rh fraction from deeper (> 0.5 m) soil layers and decreasing (increasing) soil Rh contributions from surface (0–0.2 m) soil layers (Fig. 9). This is particularly true in cold climate regions (mean annual Tair < -10 ∘C; Fig. 10). The simulated Rh fraction from the deeper soil layers (0.5–3.0 m) based on model runs using varying precipitation alone did not show significant differences (p > 0.1) from model simulations based on varying air temperature and precipitation. However, the simulated soil Rh fraction from both surface (0–0.2 m) and deeper (0.5–3.0 m) soil layers based on model runs using varying temperature alone was significantly (p < 0.01) different from model simulation results based on varying air temperature and precipitation. Moreover, cold regions (mean Tair < -10 ∘C) showed stronger decreasing trends in the Rh fraction from surface soil layers and increasing soil Rh contributions from deeper soil layers, likely due to increasing winter precipitation and snow cover (Figs. 8 and 9) and consistent with field studies involving snow cover manipulations and associated impacts on soil respiration (e.g., Nowinski et al., 2010).

Our results show that recent strong surface warming trends in the pan-Arctic region have promoted widespread soil thawing and ALD deepening (Fig. 8), while changing precipitation and snow depth have had a relatively smaller impact on ALD trends (Figs. 4 and 8). We find a mean increasing ALD trend of 0.66 ± 1.20 cm yr-1 across the pan-Arctic region over the past 3 decades, which is similar to values reported in previous studies (Zhang et al., 2005; Romanovsky et al., 2010a), albeit representing different time periods. This overall ALD deepening trend across the pan-Arctic domain corresponds with widespread warming and warming-induced decreases in SCE (Fig. 4c) and increasing non-frozen-season duration (Kim et al., 2012). Our analysis indicates that air temperature has a dominant control on upper (< 0.5 m) soil layer temperatures during the warm-season, with an increasing control in warmer climate zones (Fig. 5a). The model simulations also suggest that most pan-Arctic permafrost areas, especially continuous permafrost areas, have a relatively shallow (< 1 m) active layer (e.g., Fig. 4a). Therefore, rapid warming of the upper soil layers corresponds with general ALD deepening.

Previous studies have also shown that summer air temperature is a primary control on ALD trends, while the relationship between snow cover and ALD is more variable (Zhang et al., 2005; Romanovsky et al., 2010a, b). Our results demonstrate that deeper snowpack conditions promote warming of deep (> 0.5 m) soil layers, especially in colder climate areas (Fig. 5b), and this effect exceeds the impact of surface warming on deeper soil layers (e.g., > 1 m). Previous studies indicate that changes in snow depth can influence borehole (10–20 m) permafrost temperatures as much as changes in air temperature (Stieglitz et al., 2003; Romanovsky et al., 2010a, b). Regional simulations from the improved Community Land Model (CLM) also indicate that snow state changes can explain 50 % or more of soil temperature trends at 1m depth over the recent 50-year record (Lawrence and Slater, 2010). On the other hand, the impact of changing snow cover duration on soil temperatures may vary across different climate zones (Fig. 5c) due to the influence of both air temperature and precipitation or snowfall on snow cover duration. A shorter snow cover season may cool the soil in colder climate zones due to less insulation from cold temperatures but may warm the soil in warmer climate zones by promoting greater atmospheric heat transfer into soils (Lawrence and Slater, 2010; Euskirchen et al., 2007). Our results indicate that recent regional trends toward continued warming, earlier spring snowmelt onset and a shorter snow cover season will likely enhance soil warming and permafrost degradation in relatively warmer (mean annual Tair > -5 ∘C) regions of the pan-Arctic domain.

Snow cover is an important control on the annual carbon budget in cold regions (annual mean Tair < -4 ∘C; Fig. 6b–c), even though air temperature has a dominant control on both annual NPP and Rh fluxes across all climate zones (Fig. 6a). Strong snow cover buffering of underlying soil temperatures sustains soil respiration even under very cold winter air temperatures, and the resulting winter soil respiration can be a large component of the annual NEE budget (Sullivan et al., 2010). Field experiments have shown that winter soil respiration in tundra areas can offset total net carbon uptake during the growing season and thus switch the ecosystem from a net carbon sink to a carbon source (Zimov et al., 1996; Euskirchen et al., 2012; Lüers et al., 2014). Our results also indicate that cold-season (November–April) Rh accounts for ∼ 25 % of total annual Rh over the entire pan-Arctic domain, while this estimate may be conservative since our model may underestimate soil respiration in tundra areas (Fig. 1b). The model simulations indicate very low (< 5 %) unfrozen water below ∼ -3 ∘C at the tundra sites, while previous studies and the tower measurements (Fig. S4) indicate that substantial unfrozen water may remain even under very low soil temperatures (e.g., ∼ -10 ∘C), sustaining soil microbial activities (Romanovsky and Osterkamp, 2000). On the other hand, winter warming may change the depth and structure of insulating snow cover, affecting underlying soil temperatures, which could alter soil N mineralization rates and soil microbial activities that influence ecological processes during the growing season (Schimel et al., 2004; Sturm et al., 2005; Monson et al., 2006).

Even though air temperature has a dominant control on Rh during the warm season (from May to October), snow cover strongly influences the contribution of different soil depths to total soil decomposition and Rh (Fig. 7). This nonlinear response is due to different controls of surface air temperature and snow cover on soil temperatures at different soil depths (Zhang, 2005; Romanovsky et al., 2010a, b; Lawrence and Slater, 2010). Surface warming during the summer has a dominant control on upper soil layer temperatures (< 0.5 m; Fig. 5a), while a deeper winter snowpack has a persistent warming effect on deeper soil temperatures in colder climate areas (Fig. 5b; Gouttevin et al., 2012). Therefore, surface warming likely promotes more heterotrophic respiration from surface litter and soil layers, while a deeper snowpack promotes soil respiration from deeper soil layers. This is particularly important for soil carbon dynamics in permafrost areas, where a large amount of soil carbon occurs in deep perennial frozen soils (Hugelius et al., 2014). Previous studies including field experiments have primarily focused on the effects of surface warming on permafrost soil carbon decomposition (e.g., Schuur et al., 2007; Koven et al., 2011; Schaefer et al., 2011), while our results show that snow cover may play a larger role than air temperature in influencing deeper soil temperatures and permafrost stability. This is also supported by a recent snow addition experiment in Alaskan tundra areas (Nowinski et al., 2010), which showed that a deeper snow treatment resulted in a larger contribution of deep and old soil carbon decomposition to total soil heterotrophic respiration.

Although soil temperature and moisture are the two major environmental controls on soil carbon decomposition, other factors may also influence soil decomposition rates and permafrost carbon feedback potential but are not represented by our modeling study (Hobbie et al., 2000). A number of chemical and biological factors can affect the temperature sensitivity of soil carbon decomposition in northern soils, including enzyme abundance, microbial population size and oxygen availability (Waldrop et al., 2010). Previous studies also show that soil carbon decomposition rates may be depth-dependent. Accounting for vertical changes in soil biogeochemical properties and processes (including the size and substrate quality of the soil active layer and permafrost carbon pool, and the degree of N mineralization with decomposing permafrost carbon) may have significant impacts on the sign and magnitude of the projected high-latitude carbon response to future warming (Koven et al., 2011, 2015). Finally, changing wintertime soil microclimate will alter the amount and timing of plant-available nutrients (N) in tundra ecosystems and may drive a positive feedback between snow, soil temperature, microbial activity, and plant community composition (Schimel et al., 2004; Sturm et al., 2005).

A number of processes, notably fire disturbance, shrub expansion and thermokarst, are not included in this study but may be important factors affecting regional permafrost and soil carbon dynamics (Grosse et al., 2011; Schuur et al., 2015). A warming climate has been linked with increasing boreal–arctic fire activity and severity (Grosse et al., 2011). Fire can change the surface vegetation composition and consume a large portion of the soil organic layer, which can dramatically alter the surface energy balance and soil thermal properties, and cause rapid permafrost degradation (Harden et al. 2006; Jafarov et al., 2013). Both field experiments and satellite measurements indicate a “greening” Arctic with increasing shrub abundance due to climate warming (Tape et al., 2006). Shrub expansion in Arctic tundra can change the snow distribution and surface albedo, affecting the surface energy balance and underlying active layer and permafrost conditions (Sturm et al., 2005). The development of surface water ponding with thermokarst in ice-rich permafrost areas can alter the local surface hydrology, affecting permafrost and soil carbon decomposition (Schuur et al., 2007; Grosse et al., 2011).

Another important feature of the Arctic is strong surface heterogeneity, characterized by widespread lakes, ponds, wetlands and waterlogged soils as a result of both topography and restricted surface drainage due to underlying permafrost. Changes in both surface and subsurface hydrology are tightly coupled with local permafrost conditions and potential carbon and climate feedbacks (Smith et al., 2005; Watts et al., 2012; Yi et al., 2014; Schuur et al., 2015). Current large-scale model simulations, including this study, generally operate on the order of tens of kilometers or even larger, and may not adequately represent the effects of surface heterogeneity on simulated permafrost hydrologic processes and soil carbon decomposition processes (Koven et al., 2011; Rawlins et al., 2013; Schuur et al., 2015). For example, most models prescribe a dominant vegetation type or a single value for the organic layer thickness commensurate with the model spatial resolution, which likely introduces large uncertainties to the model-simulated moisture and heat fluxes and thus the permafrost properties. Next generation satellites, including the NASA SMAP (Soil Moisture Active Passive) mission provide for finer-scale (i.e., 3–9 km resolution) monitoring and enhanced (L-band) microwave sensitivity to surface (∼ < 5 cm) soil freeze–thaw and moisture conditions (Entekhabi et al., 2010) and may enable improved regional hydrological and ecological model parameterizations and simulations that more accurately represent active layer conditions. Finer-spatial-scale observations using lower-frequency (such as P-band) synthetic aperture radar (SAR) measurements from airborne sensors such as AirMOSS (Airborne Microwave Observatory of Subcanopy and Subsurface instrument; Tabatabaeenejad et al., 2015) may also provide improved information on sub-grid-scale processes and subsurface soil thermal and moisture profiles, providing critical constraints on model predictions of soil active layer changes and soil carbon and permafrost vulnerability.