The Trail Valley Creek research site is located at the tree line in the tundra–taiga transition zone 45 km north of Inuvik, Northwest Territories, Canada, east of the Mackenzie Delta (68.742∘ N, 133.499∘ W, Fig. a).

The mean annual air temperature in the 1999–2018 period was -7.9 ∘C . During this 20-year period, it rose by 1.1 ∘C per decade, and the strongest warming trend was observed in May with an increase of 2.8 ∘C per decade . Although the site is only 70 km south of the Arctic Ocean, the climate is continental and summers can be quite warm (Fig.  in Appendix B). The study site is underlain by continuous permafrost 100–150 m thick and is characterised by an active layer 25–100 cm deep at the end of summer. The catchment is dominated by mineral–earth hummocks which have a 5 cm thick organic layer and are underlain by fine-grained material composed of roughly one-third sand, one-third silt, and one-third clay . Between the hummocks is a several-decimetre-thick peat layer, known as the inter-hummock zone . The permafrost is ice-rich and thus susceptible to warming and thawing .

Gently rolling hills structure the lowland landscape and provide habitats for different tundra vegetation types and, in favourable locations, forest patches. The vegetation can be divided into six main types (Fig. c–h) based on the classification by : 1. Trees

can be found in river channels and on adjacent slopes as well as in isolated patches. While trees growing in the forest patches can reach 10 m height, trees in isolated patches are usually 0.5–2.5 m tall . The tree species are white and black spruce (Picea glauca and Picea mariana) . Forest patches cover roughly 2 % of the landscape .

Topsoil temperature was measured for 2 years at 68 locations distributed below different vegetation types at a maximum distance of 1200 m between the sensors in the study area (Fig. b). We used iButton^® temperature loggers (DS1922L) at 11 bit (0.0625 ∘C) resolution with an accuracy of 0.5 ∘C provided by the manufacturer. The sensors are 17 mm in diameter and 6 mm thick. Before deployment, we coated them with plastic to ensure water resistance. We tested all sensors in an ice bath and verified the temperatures measured at the zero curtain. As all sensors had zero-curtain temperatures between -0.12 and 0.30 ∘C, we assume that the accuracy provided by the manufacturer is realistic.

We installed all temperature loggers as close as possible to the surface (soil, moss, or lichen) but deep enough to be protected from solar radiation (typically 1–3 cm deep). The measurement interval was every 3 h; the measurement periods were 27 August 2016, 17:00 (local time) to 3 September 2017, 08:00 and 4 September 2017, 05:00 to 22 August 2018, 14:00. In between the two measurement periods, the sensors had to be removed in order to obtain data records and were reinstalled at the same positions. We peeled off the coating, read out the data, and used tape as the new coating, as the plastic could not be reapplied in the field setting. This was done in the field to assure that each sensor was reinstalled as close as possible to the original location. However, the coating was different and possibly the micro-location and the contact to the surrounding moss, lichen, or soil also differed slightly. Thus, we treated the two periods separately in our statistical analysis.

We classified the vegetation type into one of the six categories described above at each topsoil temperature location based on observations in the field and pictures taken while installing and removing the sensors. Maximum vegetation height was measured from 4 to 6 August 2019 within 30 cm around the sensor position.

Depth to the frozen soil was measured on 22 and 23 August 2018 three times around each topsoil temperature location within a 30 cm radius. We averaged the three values before continuing with the statistical analysis. Thawing of the soil potentially continues until freeze-back starts in October; thus, maximum active layer thicknesses were not recorded. Snow depth was measured on 30 April 2017 approximately at the topsoil temperature locations based on a handheld GPS because the flags marking each location were buried below the snow. In 2018, snow depth was measured by calculating the difference between a digital elevation model (DEM) of the snow and a DEM of the bare ground. The snow DEM was created using structure-from-motion photogrammetry in the software Pix4Dmapper , using images taken by an eBee Plus RTK at 3.5 cm per pixel resolution on 22 April 2018 . The 1 m resolution bare-ground DEM was created using airborne lidar acquired in 2008 . The resulting 1 m resolution snow depth raster was calibrated using 1370 Magnaprobe snow depth measurements taken between 19 and 25 April 2018.

We analysed the topsoil temperature data separately for the four seasons. The definition of seasons used in our statistical analysis is based on mean daily air temperatures of the last 20 years (1999–2018) at Trail Valley Creek . To obtain a complete 20-year record, we gap-filled the Trail Valley Creek data using a piecewise linear regression with data from the climate stations in Inuvik, 45 km further south. Furthermore, we smoothed the average annual cycle using a 7 d moving window (Fig.  in Appendix A). We defined winter as all days with average air temperature below -15 ∘C (6 November–10 April) while summer was defined as air temperature above 8 ∘C on average (10 June–25 August). The periods in between were defined as spring and autumn (Fig. ).

Before analysis, we checked the quality of all time series data during the two measurement periods and removed the series if (a) the data record had more than 10 % data gaps (four series); (b) the average summer topsoil temperature was more than 5 ∘C colder than air temperature, indicating that the sensor was either buried too deep and affected by the permafrost or affected by running water (three series); or (c) more than 5 % of the single summer measurements were more than 7 ∘C above air temperature, indicating additional sensor warming by direct solar radiation (eight series) (Fig. ). We used air temperature measurements by at Trail Valley Creek for this comparison. The number of remaining data series in each measurement period per vegetation type is listed in Table .

We calculated a variety of different characteristics for each topsoil temperature series, namely (i) the mean value for the whole year, each season, and each month; (ii) the range of all three hourly values between the 10th and the 90th percentiles for the whole year, each season, and each month; (iii) the slope of a linear regression of all daily average values within each season, i.e. the rate of warming or cooling; (iv) the cumulative sum of positive degree days from the beginning of the melt season until the end of August; (v) the day of the year when the soil temperature first rose above -0.5 ∘C, which indicates the beginning of the thawing period; (vi) the day of the year when the soil temperature first rose above 0.5 ∘C, which indicates the end of the thawing period; and (vii) the day of the year when the soil temperature first drops below -0.5 ∘C, which indicates the end of the freezing period in autumn. We used -0.5 and 0.5 ∘C as thresholds instead of 0 ∘C to account for sensor uncertainty. We used Python 3.6 to analyse the topsoil temperature series. In box plots, we always show the absolute minimum and maximum as whiskers, the 25th and 75th percentiles as the box, and the median as a line within the box.

In terms of statistical analysis, we used linear models (lm in R) to calculate Pearson's correlation coefficients and linear regressions for all numerical variables and to determine whether the variables were significantly related at α=0.05. In a first step, we used data of all vegetation types. In the second step, we split the data into two subsets, one containing the upland vegetation types lichen tundra and dwarf shrub tundra and one containing all other vegetation types. To assess the relationship of vegetation type to topsoil temperature characteristics, active layer thickness, and snow depth, we used a linear model (lm in R) to estimate the adjusted R2 and the fraction of variance explained by vegetation type. Prior to the analysis, we excluded the vegetation types tree and riparian shrub, as we do not have enough measurement series for these types. We assessed the impact of vegetation type on topsoil temperature characteristics based on all data series from both measurement periods. By analysing the two periods separately in the statistical model, we accounted for the systematic differences between the years. For example, the summer of 2017 was warmer and started earlier than the summer of 2018. When we were interested in the impact of vegetation type on active layer thickness and snow depth, we used all sensor locations, even those for which we removed the temperature record, for example due to missing data. We used R version 3.4 for all statistical analyses.

The two years 2016 and 2017 had different meteorological conditions in autumn. While air temperature dropped gradually to -7 ∘C before the first snowfall in 2016 (Fig. a), daily mean air temperature had only dropped to 0 ∘C by the date of the start of the snow accumulation in autumn 2017. Thus, the topsoil was significantly cooler in October 2016 compared to 2017 (Fig. d).

Autumn topsoil temperatures and cooling rates were not significantly related to vegetation type (Fig. a–d, Table ). The same was true for the start of the frozen period, which was almost identical at most sensor locations independent of vegetation type (Fig. e). However, we observed considerable variability in the mean temperature within each vegetation type, at the landscape level, and between the different years. Considering the complete autumn season (26 August–5 November), the mean autumn temperatures varied between -0.9 and 1.2 ∘C in 2016 and between 0.3 and 1.7 ∘C in 2017. For October only, the range of mean topsoil temperatures was even higher, between -5.3 and -1.0 ∘C for the coldest and warmest locations in the first period and between -2.8 and 0.2 ∘C in the second period (Fig. d). Similar differences could be observed in the cooling rates, which varied most strongly within tussock tundra and tall shrubs. Some locations within these two types cooled down more than twice as fast as other locations (Fig. c).

There were significant differences in topsoil temperature between the two winter seasons (6 November–10 April). Average topsoil temperatures were roughly 2 ∘C colder beneath all vegetation types in the winter 2016/2017 compared to the following year. This observation agrees well with the difference in mean winter air temperature of -20.3 and -18.2 ∘C in the first and second periods, respectively (Fig. a). The extremely cold December of 2016 (6.5 ∘C colder than in the second period) led to a much faster soil cooling in the first winter compared to the second (Fig. a).

There was a strong relationship between vegetation type and topsoil temperature in winter. Topsoil under tall shrub tundra stayed warmest, followed by topsoil beneath tussock, dwarf shrub, and lichen tundra (Fig. ). Vegetation type explained 36 % of the topsoil temperature variations in the coldest month (March) in a single year and 43 % of the cooling rate during winter (Table ). The coldest mean temperatures were associated with the highest temperature variations below lichen tundra (Fig. ).

Snow depth strongly mediated the association between vegetation type and winter soil temperatures. Different vegetation types showed characteristic snow depth values (Fig. e). The deepest snow cover was associated with tall shrubs, followed by tussock tundra and dwarf shrubs while lichen tundra was characterised by the shallowest snow cover. Vegetation explained 70 % and 58 % of the observed snow depth variability in 2017 and 2018, respectively (Table ). The predictability of snow depth from vegetation type is limited by the differences in snow cover in different years. For example, the snow depth at the end of April 2017 reached higher maximum and lower minimum values compared to 2018, indicating a stronger snow redistribution. The correlation between snow depth values of the different years was R2=0.49 (Fig. ).

Several winter and spring topsoil temperature characteristics were also very strongly related to snow cover, in particular the day of the year when the topsoil warmed above 0 ∘C, the topsoil temperature slope during winter cooling, and the mean temperature of the winter and spring months (Fig. ). In 2018, snow depth was less strongly related to vegetation or topsoil temperature characteristics than in 2017 (Fig. , right column, Table ).

Starting from the beginning of May, the relationship between vegetation and topsoil temperature was opposite to that found in winter (Fig. a, b). In general, topsoil temperatures beneath all vegetation types were very similar in the last few days of April. The soil below lichen tundra warmed up first and most strongly and showed the most pronounced diel variation. Dwarf shrub tundra was still cooler in May and warmed up a bit more slowly than lichen tundra. Tussock tundra topsoil temperatures rose above 0 ∘C even later; tall shrub temperatures were the last to reach positive temperatures in 2017 and were similar to tussock tundra in 2018. This order became apparent in the slope of spring temperatures, in the mean May temperatures, and in the day of the year when the thawing period ended (Fig. c–e). Using a statistical model, we found that vegetation type explained 55 % of the observed variability in the end of the thaw date and the mean May temperature and 33 % of the variability in the spring warming rate (Table ).

Furthermore, snowmelt timing was different between the years. Our observations of topsoil temperature indicated that the snowmelt period of the entire landscape was 20 and 40 d long in 2017 and 2018, respectively (Fig. e). This observation was in contrast to the more spatially variable snow distribution observed in 2017 (Fig. ). The long melt period in 2018 was associated with a 3.4 ∘C colder mean May air temperature and more cold spells compared to May 2017 (Fig. ). However, due to the differences in snowmelt timing, the cold May air temperatures in 2018 did not necessarily translate into cold topsoil temperatures. While lichen tundra measurements indeed revealed colder May topsoil temperatures in 2018, both years were similar for dwarf shrub soil, and May 2018 was associated with warmer topsoil temperatures below tussock tundra and tall shrubs.

Topsoil temperatures in summer were more similar for all vegetation types than spring or winter temperatures (Fig. a, b).

In summer 2017, dwarf shrub tundra had the coolest summer temperatures. The other three vegetation types (excluding tree/riparian shrub) had almost identical mean values, and the distributions largely overlapped (Fig. c, d). In 2018, the difference between dwarf shrubs and the other vegetation types was much smaller and the mean summer slope was almost equal for lichen, dwarf shrub, and tall shrub tundra, while tussock tundra warmed at a slightly lower rate. Some locations had negative summer slopes in 2018 as the end of summer was relatively cool. The temperature difference between the two years was much larger than the variability between vegetation types. The 2018 mean summer air temperature was 2.7 ∘C cooler than in 2017. This difference translated into topsoil temperatures which were, on average per vegetation type, 1.3–2.2 ∘C colder in 2018, with the smallest difference for dwarf shrubs. The small difference between vegetation types was also reflected in the low fraction of July mean temperature variance (12 %) that could be explained by the vegetation type (Table ). Lichen tundra featured slightly higher cumulative degree days than the other three vegetation types, which all showed similar values on average (Fig. e). Even though the influence of vegetation type on cumulative degree days was statistically significant, vegetation only explained 7 % of the observed variability (Table ).

There was a weak association between active layer thickness and vegetation type. On average, we found deeper active layers below lichen and dwarf shrub tundra compared to tussock and tall shrub tundra (Fig. i). The variability in active layer thickness within each vegetation type was substantial. Correspondingly, vegetation was only weakly related to active layer thickness at the end of summer as it explained only 34 % of the variability between different locations (Table ).

The relationship between active layer thickness and topsoil temperature characteristics depended on vegetation type. If we considered all vegetation types jointly, active layer thickness was most strongly related to the date when the topsoil temperature first rose above 0 ∘C in spring (Fig. l). While a moderate association existed with other spring, autumn, and winter characteristics, such as the cooling rate in winter (Fig. k), summer characteristics did not show a significant correlation with active layer thickness when all vegetation types were considered (Fig. m). If we split the data in two subsets, lichen and dwarf shrub tundra revealed very different winter characteristics compared to the other vegetation types. While cold winter temperatures were associated with deeper active layers in the following summer for lichen and dwarf shrub tundra (Fig. a), tall shrub and tussock locations with cold winter temperatures developed shallow active layers (Fig. e). The spring temperature characteristics were not significantly related to active layer thickness for tall shrub and tussock locations (Fig. g). Conversely, the date of the first T>0 ∘C also apparently explained a moderate fraction of the active layer thickness variance within lichen and dwarf shrub tundra (Fig. c). Higher positive degree days were associated with deeper active layers below dwarf shrub and lichen tundra (Fig. d); however, this relationship was not statistically significant (p=0.08). We observed significantly higher topsoil temperatures in October at locations with deep active layers (Figs. , ).

Vegetation type did not play a significant part in the autumn topsoil temperature variation (Fig. ), which agrees well with , who observed very similar dates of the start of ground freezing across three sites up to 63 km apart. On the other hand, found that tall shrubs experienced a delayed start of freezing by almost 1 month compared to other vegetation types. We found that at all locations the topsoil temperature signal was dominated by air temperature and snowfall timing. The importance of snow cover onset for soil temperature was highlighted by many authors, including and .

While all vegetation types had similar average autumn temperatures and cooling rates, we observed considerable variability within single types as well as between years. The temporal variability of the start of freezing was described by , who found up to 18 d difference in the start of freezing date at a single location between 1987 and 1992. In their study, the temporal variability was higher than the spatial variability, although the stations were up to 63 km apart. On the other hand, we observed considerable local variability of up to 39 d difference in the timing of the first below-zero topsoil temperatures in a single year (Fig. e). This variability is potentially driven by the soil cover of mosses or lichen, soil moisture, and microtopography rather than vascular plants. The importance of soil moisture in the freezing period was highlighted by . We observed the largest range of mean October topsoil temperatures in tussock tundra (Fig. d), which may be caused by a large range of moisture conditions depending on the microtopography in this vegetation type. However, it should be noted that variations in the sensor depth will play a major role in autumn when steep temperature gradients can be expected.

In winter, warm soil temperatures can be expected at locations with more snow such as below tall shrubs and in poorly drained and somewhat sheltered tussock-dominated areas . The snow observed at our measurement locations was indeed deepest in tall-shrub areas, where soils were also the warmest in winter (Fig. d, e). Across all vegetation types, the differences in snow depth influenced the winter topsoil temperatures, temperature variability, and cooling rates. Winter temperatures were coldest and most variable at vegetation types associated with low snow, in particular below lichen tundra and, to a lesser extent, below dwarf shrubs (Fig. a, b, d).

The variability of snow depth within vegetation types (Fig. e) is likely due to (i) topography and microtopography influencing the deposition of blowing snow , (ii) vegetation height and density differences within one vegetation type , and (iii) factors other than snow limiting shrub growth such as poor soil, too much or too little soil moisture, disturbances, and slow colonisation .

We also observed variability in the relationship between snow depth and winter topsoil temperature characteristics such as the soil cooling rate and the mean and range of temperature values in the winter months (Fig. a–d). The topography of the site is gentle and cannot account for major differences in cold-season insolation; therefore, the variability is likely largely due to differences in autumn soil moisture, snow density, and snow texture. Differences in snow density and texture across the landscape can be caused by snow compaction at wind-exposed sites, by loose snow accumulating within shrub canopies, and by the formation of depth hoar. We do not have detailed snow observations at our measurement locations to analyse such differences in detail. However, our results suggest that, for the same snow depth, topsoil temperatures in March are colder at lichen sites compared to dwarf shrub sites (Fig. c, d). This agrees very well with the observation that lichen tundra can be found at the most wind-exposed ridges. While historical surveys of snow density did not find differences between lichen and dwarf shrub tundra at the site (0.22 g cm-3; ), snow texture and depth hoar formation likely differed between shrub and shrub-free areas , leading to differences in the heat conductivity of each snowpack.

In spring, a thick snow cover delays soil warming and thus results in a reversal of the topsoil temperature–vegetation type relationship. We observed similar topsoil temperatures beneath all vegetation types in the last few days of April. The soil below tall shrubs was coldest in May, followed by the temperatures below tussocks and dwarf shrubs, while lichen tundra topsoil was already the warmest (Fig. d).

We found that the date when the topsoil warmed above 0 ∘C was strongly related to vegetation type in 2017, and less so in 2018 (Fig. e). A strong relationship can be expected due to the influence of vegetation on snow depth, snow density, and snowmelt energetics . The weaker relationship in 2018 is likely due to the more complex spring weather patterns including multiple periods of warming and subsequent freezing periods. The colder May air temperature in 2018 was associated with colder topsoil temperatures below lichen tundra and warmer topsoil temperatures below tall shrubs, where the snow melted later compared to May 2017 (Fig. d). This indicates a first-order control of snow depth as a buffer between air and topsoil temperature.

In general the presence of shrubs enhances snowmelt as soon as branches stick out of the snow, thus reducing surface albedo and increasing long-wave emissions . made extensive measurements of snow-free dates with a drone at the same study site in 2016 and showed that dwarf shrub areas become snow free earlier than non-shrub areas, regardless of snow depth and hillslope aspect. However, we did not observe this relationship at our locations, likely due to the small number of points we measured compared to . Instead, we found that topsoil warmed above 0 ∘C slightly earlier at the lichen tundra locations, even though they were similar in snow depth to the dwarf shrub locations (Fig. e, f).

Although shrubs may reduce summer soil warming through shading and evapotranspiration , in our study, the summer difference between all vegetation types is very small, and shrub tundra is comparable to tussock tundra in terms of topsoil temperature (Figs.  and ). Lichen tundra topsoil warms up slightly more, in particular during midday, but the average difference is less than 1 ∘C. These generally small differences in topsoil temperature likely contribute to the weak relation between vegetation type and summer temperature characteristics (Fig. i, m, Table ).

On the other hand, winter and spring temperature characteristics such as the date when the topsoil warmed above 0 ∘C are strongly related to active layer thickness (Fig. ). At first glance, our data suggest that the snow-free date and thus the length of the summer period are the most important drivers for active layer thickness (Fig. l). The strong influence of snow-free date on active layer thickness has been highlighted in several other studies . However, if we consider single vegetation types, the importance of snowmelt timing is strongly reduced. In particular, for tall vegetation (tall shrubs, riparian shrubs, trees) and tussock tundra we did not observe any correlation between snow-free date and active layer thickness at the end of summer (Fig. g).

While mean winter temperature is not significantly related to active layer thickness at the landscape scale, we found opposite effects for upland vegetation (lichen and dwarf shrubs) compared to tall vegetation and tussocks (Fig. a, e, j). Warm winter temperature is unexpectedly associated with a shallow active layer in upland areas, while warm winter temperature is related to deep active layers below tall vegetation and tussocks (Fig. a, e). This agrees well with results by from the outer Mackenzie Delta. They found that deep snow was associated with thick active layers in alluvial sites whereas it was associated with thin active layers in upland terrain. The opposite response of active layer thickness to mean winter temperature (and other winter and spring characteristics) below different vegetation types may be an artefact of the spatial correlation between snow depth and soil properties . For instance, both snow depth and organic layer thickness tend to be shallow on exposed lichen-covered ridges, and the latter favours deeper active layers. As organic and moss layer thicknesses are important controls of active layer thickness , the positive relationship between organic layer thickness and snow depth masks the actual effects of snow on the active layer thickness. Similarly, correlations between soil moisture, ice content, and snow depth may be different across the landscape, with variable effects on active layer development . It should also be noted that it was not recorded whether active layer thickness measurements were taken in a hummock or inter-hummock zone, which due to their vastly different soil properties have a stronger effect on active layer thickness than any other variable . This reduces our ability to draw inferences from our active layer thickness measurements.

Summer topsoil temperature is only weakly related with vegetation type, which agrees well with . Vegetation type explains 12 % of the variability of the summer topsoil temperature and 34 % of the variability of active layer thickness in our study (Table ), likely through its effects on snow depth. The generally weak relationships hinder upscaling approaches based on vegetation such as in and . We found that thawing degree days are not correlated with active layer thickness (Fig. m) at the landscape scale, echoing previous findings by . This is likely related to variable soil thermal properties and soil moisture. Within dwarf shrub and lichen tundra, where more uniform soil conditions may be expected, larger cumulative positive degree days are indeed associated with deeper active layers. In summary, our observations suggest that vegetation type is a better predictor of the near-surface thermal regime in winter than in summer in the Low Arctic. Furthermore, the soil temperature – active layer thickness relationship differs between dwarf shrubs and tall shrubs. Both vegetation types are currently expanding in the Arctic and our results indicate that their future distribution will govern the importance of summer versus winter processes for active layer thickness.