The influence of vegetation and soil characteristics on active‐layer thickness of permafrost soils in boreal forest

Abstract Carbon release from thawing permafrost soils could significantly exacerbate global warming as the active‐layer deepens, exposing more carbon to decay. Plant community and soil properties provide a major control on this by influencing the maximum depth of thaw each summer (active‐layer thickness; ALT), but a quantitative understanding of the relative importance of plant and soil characteristics, and their interactions in determine ALTs, is currently lacking. To address this, we undertook an extensive survey of multiple vegetation and edaphic characteristics and ALTs across multiple plots in four field sites within boreal forest in the discontinuous permafrost zone (NWT, Canada). Our sites included mature black spruce, burned black spruce and paper birch, allowing us to determine vegetation and edaphic drivers that emerge as the most important and broadly applicable across these key vegetation and disturbance gradients, as well as providing insight into site‐specific differences. Across sites, the most important vegetation characteristics limiting thaw (shallower ALTs) were tree leaf area index (LAI), moss layer thickness and understory LAI in that order. Thicker soil organic layers also reduced ALTs, though were less influential than moss thickness. Surface moisture (0–6 cm) promoted increased ALTs, whereas deeper soil moisture (11–16 cm) acted to modify the impact of the vegetation, in particular increasing the importance of understory or tree canopy shading in reducing thaw. These direct and indirect effects of moisture indicate that future changes in precipitation and evapotranspiration may have large influences on ALTs. Our work also suggests that forest fires cause greater ALTs by simultaneously decreasing multiple ecosystem characteristics which otherwise protect permafrost. Given that vegetation and edaphic characteristics have such clear and large influences on ALTs, our data provide a key benchmark against which to evaluate process models used to predict future impacts of climate warming on permafrost degradation and subsequent feedback to climate.


Introduction
The mass of the permafrost carbon (C) stock is estimated to be almost twice that of the atmosphere, totalling ca. 1300 Pg (Hugelius et al., 2014). As permafrost thaws an increasing amount of previously frozen C is exposed to microbial decomposition and hence can be transferred to the atmosphere and hydrosphere (Zimov et al., 2006;Schuur et al., 2009;Schaefer et al., 2011). This transfer is of major concern given that high latitudes are predicted to experience the fastest rate of warming compared to the rest of the globe (IPCC, 2013). Observations over recent decades demonstrate that permafrost is warming, thinning and shrinking in area . Therefore, to accurately predict the release of carbon from thawing permafrost and its feedback to climate, it is essential to fully understand the controls on permafrost thaw.
In the early stages of permafrost degradation, thickening of the active layer (the seasonally thawed soil layer above permafrost in which biological activity takes place) is thought to be the dominant process (Schuur et al., 2008). Although climatic warming is important in increasing active-layer thickness (ALT), the strength of the relationship between air temperature and ALT varies substantially between different regions and may be strongly influenced by factors such as vegetation cover and edaphic properties (Jorgenson et al., 2010;Shiklomanov et al., 2010;Shiklomanov & Nelson, 2013). As a result of the surface offset (the difference between air temperature and near-surface ground tem-perature) provided by ground cover and surface conditions, permafrost can persist in areas where the mean annual air temperature (MAAT) is as high as +2°C, or degrade in areas where MAAT is À20°C (Jorgenson et al., 2010). Therefore, along latitudinal gradients, increasing vegetation cover southward may compensate for greater summer warmth, weakening the relationship between MAAT and ALT (Walker et al., 2003). At finer scales, within catchments or hill slopes, ecosystem characteristics may play the dominant role in driving ALT (Jorgenson et al., 2010).
Several vegetation characteristics can influence soil temperature and hence ALT. Increasing leaf area reduces the amount of radiation reaching the soil, which should act to reduce ALT and hence protect permafrost (Marsh et al., 2010). However, with increasing stem density, particularly in shrubby species, vegetation can trap more snow, which insulates the ground and reduces heat loss in winter, potentially increasing ALT in the subsequent thaw season (Sturm et al., 2001). Experimental removal of shrub or dwarf shrub and nontussock sedge cover in Siberian and Alaskan tundra has been shown to increase ALT considerably (Kade & Walker, 2008;Blok et al., 2010). In the boreal region, the tree canopy leaf area performs a similar shading role to that of the understory, but evergreen canopies also trap snow aloft and reduce snow cover on the ground. This trapping may increase conductive heat loss from the ground in winter, which may decrease ALTs and so protect permafrost (Yi et al., 2007).
Mosses are another important component of high latitude vegetation (Street et al., 2012(Street et al., , 2013, and have been largely neglected in coupled C-climate models (Turetsky et al., 2007(Turetsky et al., , 2012. Mosses strongly dampen temperature fluctuations in the soil, largely because their open structure makes them effective insulators. However, their thermal conductivity is strongly influenced by their moisture content (Gornall et al., 2007;O'Donnell et al., 2009). In summer, a dry moss layer minimizes downward heat conduction, whereas when wet during the shoulder seasons, and when frozen in winter, the higher thermal conductivity increases upward heat conduction (Burn & Smith, 1988). Both processes keep the ground cool, thus reducing ALTs. Because the thermal properties of mosses can be explained solely by their physical properties (such as mat thickness and moisture content), this should simplify their inclusion in processed-based models (Soudzilovskaia et al., 2013).
The thickness of the soil organic layer beneath the moss layer performs a similarly important role in determining ALT (Johnson et al., 2013). The low bulk density of organic relative to mineral soils means organic soils can present more varied and extreme air and water contents, leading to a much greater range of thermal conductivities and specific heat capacities. Moisture content plays a major modifying role in the thermal properties of the soil organic layer, as it does for moss (O'Donnell et al., 2009), and can also influence ALT by nonconductive heat transfer through movement of liquid water and vapour (Hinkel & Outcalt, 1994;Kane et al., 2001). ALT monitoring in Canada has revealed that within-site variation is much reduced at sites with homogeneous thin organic layers, but where large variations in organic layer thickness or its water content exist, ALT is much more variable (Smith et al., 2009).
Our goal here is to better understand the magnitude of effects and relative importance of the multiple vegetation and edaphic characteristics that influence ALT. Such an understanding is particularly important given that climate change is likely to have contrasting impacts on different ecosystem characteristics. For instance, the tree line will (overall) move poleward with climate warming while tundra shrub cover is also predicted to increase (Grace, 2002;Jia et al., 2009;Forbes et al., 2010). Greater shrub cover, however, is likely to reduce moss cover and may over time result in thinner organic layers (Walker et al., 2006). Furthermore, fire activity in boreal forest and tundra is likely to increase in the future, causing further changes to vegetation structure and organic layer thickness and strongly influencing soil moisture (Stocks et al., 1998Kelly et al., 2013. Additionally, these factors affecting ALT are likely to be of particular importance within the discontinuous and sporadic permafrost regions, which are typically dominated by boreal forests, as these areas have relatively warm (À0.2°C) and thin permafrost, which may be particularly vulnerable to thaw (Smith et al., 2005;Baltzer et al., 2014).
Here, we aim to quantify the influence and importance of vegetation and soil characteristics in driving ALT in boreal forests, which cover over 50% of permafrost regions globally (Osterkamp et al., 2000). Our study includes four field sites within the discontinuous permafrost zone in the Northwest Territories, Canada. These incorporate different fire histories, substrates and tree canopies (deciduous or evergreen) that capture three representative and contrasting boreal forest cover types [black spruce (Picea mariana) at two sites of differing canopy density, a burned black spruce site and a paper birch (Betula papyrifera) site]. We employed a stratified sampling strategy to encompass the full range of variation in vegetation and edaphic characteristics within each site to produce the most detailed fine scale survey of the links between ALT, vegetation and soil characteristics to date. Specifically, we hypothesized that (i) increasing canopy and understory LAI would decrease ALT; (ii) taller understory vegetation would increase ALT, (iii) increasingly thick moss and soil organic layers would decrease ALT and (iv) increasing soil moisture would increase ALT. In addition to addressing these hypotheses, our approach allowed us to determine the relative importance of these different drivers of ALT, and how they interact to determine ALTs.

Study sites
All study sites were located on gently sloping topography, avoiding low points in the landscape where permafrost may impede drainage and create wetlands (median slope angle about 6-8°; Table S1). Thus, we selected sites where the permafrost dynamics were likely to be controlled by ecosystem properties rather than conversely the permafrost controlling ecosystem properties (see Discussion).
Two study sites in the Great Slave Lowland High Boreal Ecoregion (Fig. S1) were adjacent burned and unburned areas of black spruce (Picea mariana) forest located near Mosquito Creek, NWT (62°42 0 2.3″N, 116°8 0 8.8″W), subsequently referred to as Mosquito Spruce Burned (MSB) and Mosquito Spruce Unburned (MSU) (Fig. 1). The effects of a large fire at this site in 2008 (Canadian Forest Service, 2014) were still clearly visible in our survey year (2014). The burned site was characterized by charred snags and ground scorching, bare ground coverage and associated heterogeneous losses of moss and organic soil horizon thickness. In some areas, taller shrub birch (Betula glandulosa) had begun to establish, while other patches were covered with shrubby species such as Rhododendron groenlandicum and Vaccinium vitis-idaea (Fig. S2). The density of dead trees was 2720 stems ha À1 , with a mean diameter at breast height (DBH) of 6.8 AE 0.3 cm. A neighbouring (ca. 50 m) study site was established in an unburned area (ca. 500 m 2 of unburned forest) dominated by mature black spruce trees interspersed with tamarack (Larix laricina), with a varying degree of canopy closure (stem density 4161 stems ha À1 , DBH 7.8 AE 0.2 cm, Fig. 2 , Table S1) and a carpet of feather mosses, predominantly Hylocomium splendens. Again, shrub and herbaceous species, mainly Arctostaphylos rubra and Geo-caulon lividum, formed a patchy understory (Fig. S2). The soil profile at these sites transitioned sharply from organic to mineral soil, and the mineral horizon was generally poorly poreice cemented and, down to 1 m depth, dominated by grey sand with occasional rounded to subangular pebbles <2 cm in diameter. Texturally, the <2 mm fraction of mineral soil at MSU had mean values (n = 6) of 82 AE 8% sand, 11 AE 6% silt and 7 AE 4% clay, similar to those at MSB (n = 6) of 79 AE 5% sand, 15 AE 4% silt and 6 AE 2% clay.
Two other sites were established ca. 63 km east in the Great Slave Lowland within adjacent black spruce and paper birch (Betula papyrifera) forest near Boundary Creek (62°31 0 36.3″N, 114°57 0 41.3″W and 62°31 0 37.7″N, 114°57 0 38.9″W respectively), subsequently referred to as Boundary Creek Spruce (BS) and Boundary Creek Birch (BB) (Figs 1 and S1). Paper birch stem density was 2980 stems ha À1 and mean DBH was 9.6 AE 0.3 cm ( Fig. 2 and Table S1 for LAI Tree ). There was no moss layer in the birch understory; instead the ground was covered with leaf litter with sparse patches of fireweed (Chamerion angustifolium), low stature shrubs such as Ribes glandulosum and Rubus chamaemorus or denser stands of Rosa acicularis or emergent P. mariana saplings (Fig. S2). Nearby, the spruce site (BS) had a stem density of 6620 stems ha À1 and the trees had a mean DBH of 5.7 AE 0.1 cm. The understory was similar to that of the Mosquito Creek unburned site though with some fruticose lichen dominated patches. The organic soil horizon at these sites was underlain, to at least 1 m depth, by grey silty clay mineral soil with ice lenses <1 cm thick and disseminated ice crystals a few millimetres in diameter. Texturally, the <2 mm fraction of mineral soil at BS had mean values (n = 6) of 13 AE 11% sand, 24 AE 5% silt and 63 AE 15% clay, and at BB (n = 4) 26 AE 10% sand, 29 AE 10% silt and 45 AE 20% clay. Abundant segregated ice was commonly observed in the BS soil profiles at depths between 60 and 90 cm, which coincides with the top of the permafrost.
During the growing season, the average mean daily air temperature at both the Mosquito Creek and Boundary Creek sites, measured using screened TinyTag probes (Gemini, Chichester, UK), was 15.6°C, the average daily maximum and minimum temperatures were 21.6 and 8.2°C respectively (Fig. S3a). Total rainfall through the growing season, measured at the nearby Environment Canada Yellowknife-Henderson station (62°27 0 00.0″N, 114°22 0 48.0″W) was 77 mm ( Fig. S3b). At all sites, bottom-sealed access tubes filled with antifreeze were installed after soil coring to monitor soil temperature profiles by means of sealed thermistors connected to a digital multimetre. Soil temperatures, at 1 m depth obtained from access tubes installed after soil coring, at the end of the growing season (29 August 2014) were À0.2 AE 0.2°C in MSU and 1.6 AE 1.6°C in MSB. Soil temperatures at 1 m depth at the end of the growing season (2nd September 2014) were À0.5 AE 0.1°C in BS and À0.4 AE 0.1°C in BB.

Plot establishment
In 2014, plots of 1 m 2 were located to ensure that a full range of ground cover types, tree canopy cover and moss and organic layer thicknesses were sampled at each site. Additionally, we only selected plots where the ground cover was homogeneous over an area of at least 2 9 2 m such that the 1 m 2 study plot was representative of the larger area. Initially, 30 plots were established at each of the four study sites. However, as the summer thaw season progressed it became apparent that, for a small number of these plots, thaw depth could not be accurately determined due to the resistance of unfrozen, clay-rich soils or rocks. Hence the final dataset had 30 plots from the MSB site, 30 from the MSU site, 24 from the BS site and 24 from the BB site, giving a total of 108 plots.

Vegetation and edaphic characteristics
Vegetation characteristic survey work was carried out between 27th July 2014 and 25th August 2014, with thaw depths recorded on 28th August -1st September 2014 to capture near maximum ALT. Tree canopy leaf area index (LAI Tree ) was determined using a Nikon D5000 DSLR camera with a Sigma EX 4.5 1:2.8 DC HSM hemispherical lens. Nine photographs were taken at dif-  ferent locations 1 m above each plot and processed with CAN-EYE software (Weiss & Baret, 2010). The 'LAI2000, 5 rings.Eff' output was used for maximum comparability with understory LAI. To control for influence of sunlight conditions, images were thresholded individually by the same researcher, allowing the threshold to be set appropriately for the sky conditions. Whenever possible all LAI measurements were taken as late as possible in the day to reduce influence of directly incident sunlight. Understory LAI (LAI U ) was measured using a LI-COR LAI2000 Plant Canopy Analyser (LI-COR, Lincoln, USA). One measurement was taken above the understory canopy with a 90°field of view cap and four measurements were taken below the understory canopy at the four corners of a 20 9 20 cm square at the centre of each plot. This protocol allowed LAI U to be determined independently of LAI Tree (White et al., 1997). LAI U therefore comprises the vascular plant component but not the moss ground layer. To control for sunlight conditions, the LAI2000 has a relatively narrow (90°) field of view cap that minimizes the impact of uneven cloud conditions on days when the sky was cloudy or nonuniformly overcast (as recommended in the user manual). When the sky was clear a tarpaulin was used to cast shade on the part of the understory canopy that was being measured (again following the manual).
The maximum understory vegetation height was measured from the moss or soil surface (where moss was not present) at the four corners and in the centre of a 50 9 50 cm quadrat in the centre of the 1 m 2 plot. The mean of these five values was used in subsequent analyses.
Moss thickness was determined by carefully removing a section of moss and organic material from the ground with a serrated knife while avoiding compression. Moss thickness was measured as the distance from the surface of the living moss to the point at which dead moss (fibric material) became decomposed to a state that its structure was no longer discernible. Moss thickness therefore includes living and dead moss layers.
Surface moisture was measured in the upper 6 cm of the moss/soil layer using an ML3 ThetaKit (Delta-T Devices Ltd, Cambridge, UK) with an accuracy of 1% for 0-50% range volumetric moisture content, and precision of 1 mV. Measurements were taken by placing the probe gently into the surface of the moss carpet or soil surface (in the case of bare ground plots), and in most cases readings reflected the moisture in the moss layer. Given the similar properties of soil organic matter and moss in insulating permafrost, we deemed it more appropriate to measure the depth of the moisture measurement from the surface of the soil or moss when present, rather than always measuring from the soil surface (even when there was a moss layer above this). A deeper soil moisture measurement was taken by parting the moss/organic layer so that the probe was inserted to a depth of about 11 cm, hence the measurement volume was between ca. 11-16 cm below the surface of the moss or soil. For clarity, we refer to this reading as deeper soil moisture, though in the thickest moss areas, this 11-16 cm volume included some moss. Four measurements were taken at each plot throughout the growing season and the mean of these readings was used for data analysis. In previous work, we found that millivolt readings varied over small scales even within the same plot, so building new calibration curves for the moisture probe was problematic (and ideally would have needed a calibration curve for each plot). Hence, we used the standard factory calibration settings for organic soils and concentrated effort in obtaining multiple readings from each plot at the multiple sites over the growing season.
Soil organic matter (OM) thickness, determined using a soil corer (1.8 cm internal diam.), was measured to the depth of the base of the O horizon. The thickness of the moss layer was then subtracted from this measurement. Slope was measured, as it influences surface water runoff, snowpack depth and solar radiation interception. A 50 cm wooden plank was laid along the steepest gradient through the centre of each plot, and the angle below the horizon was measured using a digital angle meter with a bubble level.
ALT was measured using a graduated stainless steel rod (1.5 cm diameter) inserted to the point at which it was impeded by frozen soil (Nelson & Hinkel, 2003). At this late stage in summer, thaw depth approaches its maximum and is therefore close to that of the ALT (Walker et al., 2003). A temperature probe was used to confirm that the soil was frozen (0°C) at the point of refusal. The probe was custom built (British Rototherm Co. Ltd. Port Talbot, UK), and consisted of a robust 1.3 m long tube of stainless steel (11 mm outer diameter, 7 mm inner diameter) fitted with a 300 mm wide 'T' handle for inserting and extracting the probe from the soil. The sensing tip was 7 mm in outer diameter, sharpened to a point and contained a platinum resistor (100 Ω at 0°C, 4 wire, Class B, made to IEC 751 Standard; manufacturer's stated tolerance AE0.3°C). Temperature measurements were made by connecting the probe to a hand-held digital thermometer. Where temperatures were >0°C, these plots were excluded from analysis.

Statistical analyses
The influence of the measured vegetation and edaphic factors on thaw depth was assessed using both tobit multiple regression models and structural equation modelling (SEM). These approaches are complementary, with SEM giving a more mechanistic insight into the controls on ALT whereas multiple regression modelling provides a clearer insight into the direct influence of the measured variables on ALT. In addition to their complementary strengths, the combined approach reinforces confidence in the conclusions drawn where there is agreement between them.
Tobit regression was used in place of standard multiple regression because 17 of the plots in the burned sites had ALTs greater than 150 cm, which were beyond the limit of the probe, resulting in a censored response variable. Tobit models were developed specifically to deal with this kind of censored data (Tobin, 1958) and were implemented in the VGAM R package (Yee, 2014). Prior to tobit analysis predictor variables were mean centred to aid interpretation of the results. Curvature in the relationship between explanatory and response variables was tested by fitting all explanatory variables and their squared terms then retaining only those terms which were significant in the full model. A series of models each containing main effects and a subset of all possible two-way interactions were used to identify potentially significant interactions. A full model was then constructed using all main effect terms plus the identified potentially significant interactions and quadratic terms. This model was simplified by sequentially removing nonsignificant terms in order to obtain the model with minimum Akaike information criterion (AIC) (Crawley, 2012). a levels for testing the significance of the terms remaining in the model were determined using the false discovery rate control method described by Benjamini & Hochberg (1995). Interactions were interpreted using the methods of Aiken & West (1991). Model fits were checked visually to ensure that they conformed to model assumptions.
This process was repeated at the individual site level to determine whether the same factors that emerged as important drivers of ALT across all sites were also significant within each land cover type (paper birch, black spruce and burned black spruce). However, interaction terms were not fitted as the sample size within each land cover type was too small, and further data collection within each individual site to the level needed to elucidate interactions would have been impossible in the time available. Data are back transformed for presentation in figures to ease visualization of relationships, but our interpretations are based on the nonback transformed data in the statistical analysis. These analyses were carried out using R 3.1.2 (R Core Team, 2014).
The data were also interrogated with SEM, which allows links between measured variables to be analysed with direct paths implying causality and indirect paths occurring where the impact of one factor is modified by another. Using SEM, direct and indirect impacts of exogenous and endogenous variables can be estimated and compared. Additionally SEM is well suited to data where there may be colinearity among predictor variables, since SEM can be used to build meaningful models of ecological systems where this is present (Graham, 2003). Bayesian SEM was used as it allowed the incorporation of the censored ALT measurement (burned site plots with ALT >150 cm) by restricting the posterior distribution of those ALT estimates which could not be measured directly. Diffuse priors were set for all parameter estimates except where an admissibility test determined that the lower bound of the prior needed to be set to zero in order to generate a proper solution. The rationale behind the structure of the model is outlined in Supporting Information. Ninety-five percentage highest density intervals (HDIs) were used to assess whether parameter values and differences between parameter values were credibly different from zero (i.e. zero did not lie within the 95% HDI). SEM was carried out using IBM SPSS Amos 22 (Arbuckle, 2013).

Site characteristics
Site characteristics contrasted as expected for these land cover units. Briefly, MSU had relatively low average surface and deep soil moisture contents, an intermedi-ate tree canopy LAI, and thick moss (dead + live moss) and organic layers (Fig. 2, Table S1). BS was similar, but had a more closed tree canopy, slightly greater deep soil moisture, and slightly lower moss and OM thicknesses (Fig. 2, Table S1). MSB had the greatest surface and deep soil moisture content, lowest tree canopy LAI, OM thickness, and a very thin moss layer ( Fig. 2; Table S1). The BB site almost completely lacked moss, instead having a layer of leaf litter, but the total organic layer thickness was similar to that of the BS site (Fig. 2b, c; Table S1). Surface moisture was intermediate at this site, but deep soil moisture content was greater than in the neighbouring BS site (Fig. 2d, h; Table S1). Tree canopy LAI was very similar across both of the Boundary Creek sites. ALTs were greatest at the MSB site, intermediate at BB and smallest at both MSU and BS (Fig. 2a, Table S1).

Cross-site tobit multiple regression analysis
The tobit multiple regression of all site data combined revealed significant effects of OM thickness, moss layer thickness, surface moisture and LAI Tree on ALT (detailed below, Table 1, Fig. 3). ALT was also influenced by significant interactions between LAI Tree and deeper soil moisture, between LAI U and deeper soil moisture and between OM thickness and ground slope angle (Table 1, Fig. 4). The combination of factors retained in the final model (OM thickness, moss layer thickness, LAI U , LAI Tree , slope, deeper soil moisture, surface moisture and the interactions detailed in Table 1) explained 73% of the variation in ALT across our four sites (Adjusted McFadden's pseudo R 2 = 0.734).
As moss layer thickness increased ALT decreased. This relationship took the form of an exponential decay, suggesting that increases in shallow moss layers had a greater impact on ALT than in deeper moss layers (Fig. 3a, Table 1).
Similarly, increasing OM thickness decreased ALT overall (Fig. 3b). However, there was also an interaction between OM thickness and slope (Table 1, Fig. 4a). This interaction arose from a decreasing influence of increasing organic layer thickness on ALT with steeper slopes. For plots on the steepest slopes, increasing OM thickness only weakly reduced ALT (Fig. 4a).
The relationship between surface moisture and ALT was straightforward, with greater surface moisture resulting in greater ALTs (Table 1, Fig. 3d). In contrast, deeper soil moisture influenced ALT by modifying the influence of other factors (interactions described below).
Overall, increasing LAI Tree caused a decrease in ALT (Table 1, Fig. 3c). This effect, however, was moderated by a significant interaction with deeper soil moisture (Table 1, Fig. 4b). As deeper soil moisture declined, the influence of LAI Tree on ALT was also decreased. Specifically, when soil moisture was high (one standard deviation above the mean, 0.55 m 3 m À3 ) the relationship between LAI Tree and ALT was strongly negative (Fig. 4b), but became less negative when soil moisture was at its mean (0.33 m 3 m À3 ). The relationship between LAI Tree and ALT was not significant when soil moisture was low (one standard deviation below the mean, 0.11 m 3 m À3 ; Fig. 4b).
ALT was also influenced by a similar interaction between LAI U and deeper soil moisture (Table 1). At the mean value of deeper soil moisture (0.33 m 3 m À3 ), increasing LAI U did not have a significant impact on ALT (Fig. 4c). However, at one standard deviation above the mean soil moisture (0.55 m 3 m À3 ), increasing LAI U decreased ALT, while at one standard deviation below (0.11 m 3 m À3 ) increasing LAI U increased ALT (Fig 4c).

Individual land cover type tobit regression analysis
Different combinations of vegetation and edaphic factors were revealed to be the most important determinants of ALT within each land cover type (for brevity, greater detail for the site-specific analyses, and discussion, are provided in supporting information, Table S2). Across the two black spruce sites (MSU and BS), increasing OM thickness and increasing LAI Tree both resulted in decreasing ALTs. Conversely, increasing surface moisture content promoted greater ALTs (Table S2).
At the paper birch site (BB), only LAI Tree had a significant effect on ALT, with more closed tree canopies associated with smaller ALTs (Table S2). In the burned black spruce site (MSB), LAI U was the only significant factor, with greater LAI U resulting in smaller ALTs (Table S2).

Structural equation model
The structural equation model supported our a priori interpretation of the importance of moisture mediation on the impact of the vegetation and edaphic factors measured, and explained 70.5% of the variance in ALT (Fig. 5, Table S3, and supporting information for SEM rationale).
The effects observed are causal relationships in the context of the SEM approach where the choice of model -and so the direction of causality-is determined from the modeller's existing understanding of the system. Deeper soil moisture, LAI Tree, , moss layer thickness, surface moisture and OM thickness all had direct effects on ALT (95% highest density interval, HDI), and the standardized path coefficients indicated that the importance of these direct effects was in this order (Fig. 5). Increasing LAI Tree , moss layer thickness and OM thickness all had negative direct effects on ALT, whereas increasing deeper soil moisture and surface moisture both acted to increase ALT (95% HDI, Fig. 5).
LAI Tree , moss layer thickness, OM thickness and LAI U had indirect effects on ALT, which were mediated via either surface or deeper soil moisture (zero not within 95% HDI for indirect effects). In the case of moss layer thickness, OM thickness and LAI Tree , the indirect effects via moisture acted to reinforce the direct effects. The indirect effects of LAI U as mediated by moisture were more complicated as LAI U simultaneously increased surface moisture and decreased deeper soil moisture. Moss layer thickness had a negative effect on both surface moisture and deeper soil moisture (95% HDI). LAI Tree and OM thickness each had a negative effect on deeper soil moisture, but not on surface moisture, and LAI U had a positive effect on surface moisture but not deeper soil moisture (zero not within 95% HDI). There was also a positive effect of deeper moisture on surface moisture. In the case of LAI Tree , moss layer thickness and OM thickness, the standardized direct effect of these variables was not credibly different to their standardized indirect (moisture mediated) effects (95% HDI of difference contains zero), implying that the indirect effect of these factors on ALT via soil moisture was of similar importance to their direct effects. However, the direct effect of deeper soil moisture on ALT was greater than the indirect effect via surface moisture.

Discussion
This study provides the most comprehensive assessment of how ALT varies with vegetation and soil characteristics in boreal forest, and is the first to separate data on moss, vascular vegetation and soil organic layer properties. Our main approach of seeking emergent drivers that operate irrespective of land cover type shows that moss layer thickness and LAI Tree , in combination with soil OM thickness, are the most important and broadly applicable factors influencing ALT. However, the influence of both LAI Tree and LAI U on ALT is highly dependent on soil moisture, which also has its own direct effect of increasing ALTs. Therefore, the important central role that moisture plays in influencing ALT should not be underestimated given future changes in precipitation regimes and evapotranspiration that may significantly alter soil moisture. Indeed, all measured variables that had a direct impact on ALT also had indirect effects that were mediated by soil moisture. Understanding the direct effects and interactions of these key vegetation and soil characteristics is crucial both for understanding the controls on current ALTs, and also future impacts of climate change on permafrost degradation due to climate-driven changes in vegetation and soil properties.

Direct and indirect effects of soil moisture
Our SEM showed that surface soil moisture had the greatest direct impact on ALT, and also highlighted deeper soil moisture content as an important factor influencing ALT indirectly by modifying the effect of other factors. This is supported by the regression analysis that also revealed a significant influence of surface moisture in increasing ALTs, whereas the importance of deeper soil moisture lay in modifying the impact of other factors on ALT. Taken together, these findings strongly indicate that (among the factors we measured), future changes in precipitation and evapotranspiration patterns and hence associated changes in soil moisture, will have a particularly strong influence on the rate of permafrost degradation, both directly by increasing ALTs and indirectly by modifying the influence of other drivers (Iijima et al., 2010).
In the ecosystems we studied, the critical role of soil moisture emphasizes the importance of moss and organic matter layers as critical insulators, and the role of greater wetness in increasing heat conductance and resulting in deeper active layers. However, we note that in very wet and waterlogged soils in other systems, it is possible that greater moisture can alter ALTs due to the greater latent heat effect delaying freezing (Morse et al., 2015). Similarly, for wetland ecosystems where the water table is at or near the surface during the growing season, we expect changes in precipitation regime to have less effect on ALTs. Clearly, the critical role of soil moisture warrants further attention, for both empirical research and for model development.

Moss layer thickness
Both regression analysis and SEM revealed that ALTs decreased with increasing moss layer thickness (live (c) Fig. 4 Interaction plots with partial residuals derived from the multiple regression model. (a) Interaction between organic layer thickness (OLT) and slope; points represent partial residuals for OLT; dotted, solid and dashed lines show the relationship between active-layer thickness (ALT) and OLT when slope is at its mean value (6.99°), one standard deviation (SD) above its mean value (10.5°) and one SD below its mean value (3.45°) respectively. (b) Interaction between tree canopy LAI (LAI Tree ) and deeper soil moisture (11-16 cm depth); points represent partial residuals for LAI Tree ; dotted, solid and dashed lines show the relationship between ALT and LAI Tree when deeper soil moisture is at its mean value (0.33 m 3 m À3 ), one SD above its mean value (0.55 m 3 m À3 ) and one SD below its mean value (0.11 m 3 m À3 ) respectively and (c) Interaction between understory canopy LAI (LAI Understory ) and deeper soil moisture; points represent partial residuals for LAI Understory ; dotted, solid and dashed lines show the relationship between ALT and LAI Understory when deeper soil moisture is at its mean value (0.33 m 3 m À3 ), one SD above its mean value and one SD below its mean value (0.11 m 3 m À3 ) respectively. Significant stars in figure legends for individual relationships are as follows (*P < 0.05, **P < 0.01, ***P < 0.001). and dead moss). This is consistent with dry moss being a poor thermal conductor which readily insulates permafrost from warm summer air temperatures (Turetsky et al., 2012). The exponential nature of the regression relationship shows that while initial increments in moss layer thickness may have a large impact in reducing ALT, subsequent increases are less effective as has been noted in 1-d modelling analysis (Riseborough et al., 2013). The SEM demonstrated that moss also affects ALT indirectly because increases in moss layer thickness resulted in drier soil. The tendency for the whole organic layer to dry in feathermoss covered soils has been demonstrated elsewhere (Harden et al., 1997) and a thicker moss layer may dry more quickly during hot, dry summers because its pore fraction is likely to be greater than that of the underlying OM. Overall, moss is the single most important component of the plant community in driving shallow ALTs and hence protecting permafrost.

Organic matter thickness
Consistent with our hypotheses, thicker soil organic matter layers decreased ALTs, demonstrating the capacity of organic soils to protect permafrost from thaw by performing a similar role to that of the moss layer, with good insulating properties in summer when dry (O'Donnell et al., 2009). However, the actual relationship between OM thickness and ALT was more complicated due to a significant interaction with slope. Moisture may provide a mechanistic explanation for the OM thickness interaction with slope, where the influence of increasing OM thickness diminished on increasingly steep terrain. Because runoff and drainage can be improved on steeper terrain, as supported by our SEM analysis, the importance of increases in OM thickness is likely to be reduced because on a steeper, drier, slope less OM would be required to provide the same level of insulation compared to a shallower, wetter slope (Jorgenson et al., 2010). Steeper slopes may also have decreased snow pack thickness, reducing insulation from this and so furthering the importance of OM thickness on ALT (Nowinski et al., 2010;Johansson et al., 2013). This is consistent with OM thickness having a negative effect on deeper soil moisture in the SEM. The mechanism driving this is unclear but in thicker organic layers, the soil moisture reading at 11-16 cm may be situated within organic matter that is more 'surface-like' and so less compacted and more freely draining (the opposite direction of influence of drier soils creating deeper organic layers seems very unlikely). While the soil organic layer plays a similar role to the moss layer in insulating permafrost, a greater OM thickness may not compensate for a thinner moss layer; for instance a moss layer loss (e.g. from 15 to 0 cm) increases ALT by ca. 40 cm on average, whereas the same loss of OM thickness only increases ALTs by ca. 10 cm. Moss has a lower bulk density than the organic layer which will contribute to it draining/drying more readily in summer and having greater insulating properties (see also 'justification for SEM design' in supporting information). Additionally, although we did not measure the influence of litter, given the consistent emergence of moss as one of the most important factors influencing ALT, where moss is replaced by litter as the surface cover (as in deciduous vs. evergreen stands) it is highly unlikely that litter could maintain ALTs and hence protect permafrost to the extent that moss can.

Leaf area index
Overall, increasing LAI Tree caused a decrease in ALT, emphasizing the importance of a dense tree canopy in boreal forest in protecting permafrost. Three potential mechanisms could explain this observation; (i) increasingly closed evergreen canopies could intercept more snow, reducing the insulating snow pack and hence allowing more heat loss from the ground in winter (Lundberg & Koivusalo, 2003); (ii) larger tree canopies would transpire more, drying the soil and reducing thermal conductivity in summer, or (iii) a greater canopy would shade the ground more, reducing downward heat flux, as identified beneath shrubs in Siberian tundra (Blok et al., Heijmans, 2010). Our study suggests that the transpiration and shading mechanisms are both important in explaining the impact of LAI Tree on ALT. Furthermore, whereas evergreen black spruce canopies at our study sites may intercept snow in winter, deciduous paper birch leafless canopies trap snow much less effectively, yet LAI Tree still emerged as a significant factor for the birch site when analysed separately. Also, snow pack thickness, which typically exceeds 30 cm depth across this area has been suggested to be functionally homogeneous, and hence of limited importance in determining ALT (Morse et al., 2015). Nevertheless, moisture plays an important role in modifying LAI Tree influence: as deeper soil moisture decreased, the strength of the impact of tree canopy shading also decreased because drier deeper soil will conduct less heat downwards, rendering the cooling effect of shading less important. A similar interaction was also present between LAI U and deeper soil moisture. However, in this case increasing LAI U only decreases ALT in wetter soils. In drier soils, increasing LAI U increases ALT, an effect which is not easy to explain, but could result from the prevention of further evaporation of moisture in these drier soils or from increased snow trapping. Also, LAI U simultaneously increased surface moisture and decreased deeper soil moisture. That direction of causality (i.e. LAI U driving moisture) in the SEM is likely for deeper soil moisture since greater LAI U would dry soil more through greater transpiration, whereas the opposite direction of influence of more LAI U occurring in drier soils is harder to accept. The mechanism for greater LAI U increasing surface moisture is unclear, though greater shading and wind shelter may reduce surface evaporation from moss and litter layers.
Overall, LAI U may have a smaller impact on ALT than LAI Tree if they both influence ALT through shading. LAI U may be less important under more closed tree canopies, such as those in the unburned black spruce and paper birch sites because any further shading they provide will have a lesser impact on the amount of radiation reaching the ground. However in the burned site, where the soil moisture content is greater and the tree canopy is absent, the vascular understory vegetation plays an important role in preventing greater active-layer thickening. The rate of recovery of the ground layer vegetation postfire and prior to re-establishment of trees could therefore be important in determining the extent of ALT deepening from fire, both through its shading effect and its capacity to dry the soil (Viereck et al., 2008).

Fire impacts
The burned spruce site (MSB) had the greatest mean ALT of all our sites, consistent with it having the wettest soil, a very thin moss layer, and the thinnest organic layerall factors promoting a deeper ALT. Indeed, with most of the other factors that control ALT reduced or removed by fire (OM thickness, moss depth, LAI Tree ), only LAI U remains as the main controlling factor in ALT across the burned site (Table S2). The interaction between LAI Tree and soil moisture probably explains a large proportion of the considerable difference in ALT between our burned and unburned sites. The loss of the tree canopy postfire will increase soil moisture due to a lack of transpiration and may also allow the accumulation of more snow in winter. This is coupled with an increase in solar radiation reaching the ground, and a loss of insulation provided by reduced moss and organic layers, which will result in greater downward heat flux in summer and therefore greater ALT. Boreal forest fires are known to impact dramatically on ALT (Mackay, 1995;Burn, 1998;Yoshikawa et al., 2002), and our work shows that they have this major impact by concurrently altering several ecosystem characteristics that would otherwise provide shallow active layers. Fire severity ranges widely within boreal forests, and our study suggests that fires of differing intensity will have different levels of influence on ALT, with less severe impacts where only tree and understory canopies are removed compared to fires where insulating moss and OM layers are also lost (Turetsky et al., 2011). It is therefore of considerable concern that both forest fire frequency and intensity are predicted to increase with climate change; indeed clear increases in North American boreal forest affected by fire have been observed over recent years (Gillett et al., 2004;Kasischke & Turetsky, 2006). Increased fire intensities will lead to greater ALT increases, and shorter intervals between fires will lessen permafrost recovery, leading to loss of permafrost which may shift black spruce ecosystems from being a C source to a sink (Jorgenson et al., 2010;O'Donnell et al., 2011). However, in areas where permafrost is a driver of soil conditions, and thin active layers can impede drainage and increase soil moisture contents, the depth of burn may be reduced, as shown in lowlands in Alaska (Turetsky et al., 2011). In such ecosystems, fire may have less effect on permafrost-protecting ecosystem characteristics and thus there may be more potential for ecosystem recovery.
Finally, while the selected sites allow particular insight into how fire drives deepening of ALTs, the relationships seen might also be used to determine the extent that other processes altering vegetation structure (not directly assessed here) may result in deeper ALTs (e.g. stand damaging forest pests and disease, drought, storms; Gauthier et al., 2015).
Permafrost as a responder to, or driver of, ecosystem characteristics In our study, permafrost is considered to be a responder to ecosystem characteristics, rather than a driver of them. At our study sites, vegetation and soil properties drive permafrost because the permafrost here is climate-driven and ecosystem-protected sensu Shur & Jorgenson (2007) (Morse et al., 2015), and evapotranspiration is the main component of the surface energy balance (Burn, 1998). Hence, vegetation (through evapotranspiration) and summer rainfall determine surface and deeper soil moisture in our study, rather than the presence of permafrost itself. In other areas, especially where permafrost impedes drainage and promotes peat accumulation, permafrost can be a key driver of ecosystem dynamics. This occurs, for example in flat terrain with low-centred ice-wedge polygons in Arctic lowlands or where a bowl-shaped permafrost table drives cryoturbation in nonsorted (e.g. hummocky) patterned ground (Mackay, 1980).
Our approach of conducting detailed surveys across multiple land cover types has allowed us to determine the relative importance of the critical factors controlling ALT in boreal forest, and has also revealed how they interact to modify ALT. Moss layer thickness, tree canopy LAI and organic layer thickness are demonstrated to play critical roles in determining ALTs in boreal forest, which underlines the importance of including these components in processed-based models, and of testing models that include vegetation-soil interactions against data sets such as that presented here. Crucially though, the importance of these influences is highly dependent on soil moisture. This result suggests that changes in the magnitude and timing of precipitation, along with changes in evapotranspiration, could dramatically alter the interactions between vegetation, soil and permafrost. The impacts of future changes in precipitation and evapotranspiration on ALT and permafrost degradation require much more attention. We also demonstrate that forest fires influence ALT by simultaneously removing or reducing multiple ecosystem components that would otherwise reduce ALTs and protect permafrost. Again, this raises further concern in light of the already increasing fire frequency and intensity in boreal regions. Understanding the mechanisms through which vegetation and edaphic factors determine ALT and how they will interact with future changes in fire regime or precipitation patterns is vital in order to predict future rates and carbon cycle consequences of permafrost degradation in a changing climate.

Supporting Information
Additional Supporting Information may be found in the online version of this article: Data S1. Regional setting, Justification of SEM design, Discussion of individual site ALTs. Fig. S1. Map showing location of study sites. Fig. S2. Percentage cover of understory species within the four sites.  Table S1. Summary statistics for each of the parameters measured at the four field sites. Table S2. Parameter estimates for a multiple regression models of the effect of the vegetation and edaphic variables on thaw depth within each site. Table S3. SEM parameters.