INTRODUCTION

Stand-replacing fire is a fundamental process in high-latitude boreal forests and understanding the factors that determine tree regeneration following these fires is an increasingly pressing scientific need. Arctic regions are experiencing twice the mean global rate of temperature increase associated with climate change (IPCC, 2021) and this creates strong potential for ecosystem transformations that influence the global climate system (Schuur et al., 2015). For example, in Interior Alaska, increased fire severity can transition forests from black spruce (Picea mariana) to deciduous hardwood dominance, leading to changes in carbon and nitrogen pools (Alexander, Mack, Goetz, Beck, et al., 2012;Alexander, Mack, Goetz, Loranty, et al., 2012) and surface energy (Beck et al., 2011). The effects of climate warming on fire and forest dynamics in systems that overlay Yedoma permafrost have particularly strong potential to influence the global carbon cycle due to its vast extent (approximately 1 million km 2 across Siberia and central Alaska) and high carbon content (approximately 500 Gt of carbon; Alexeyev et al., 1995;Zimov et al., 2006). Forest regeneration following wildfire is a driver of long-term ecosystem structure and function in many boreal forests and is regulated by tree seedling recruitment in the postfire environment (Brown et al., 2015;Furyaev et al., 2001;Hollingsworth et al., 2013;Kasischke et al., 2010;Ponomarev et al., 2016;Turner et al., 2002).

Recruitment limitation is an important constraint that dictates forest regeneration following wildfires and is linked to propagule availability and seedling establishment opportunity. Fire is a keystone ecological process in many coniferous forests and a requirement for germination and establishment of some species (Brown, 2006;Christensen, 1993;Moore et al., 1999;Pyne, 1982). By consuming organic matter, fire can create favorable mineral soil microsites for seed germination and seedling establishment, and future forest composition and structure are strongly influenced by patterns of postfire tree recruitment (Chu et al., 2017;Green, 1984;Johnstone et al., 2010;Johnstone & Chapin, 2006;Moore et al., 1999). Propagule limitation can be a strong determinant of population dynamics, which ultimately manifest in ecosystem structure and function as the stand matures (Chapin et al., 2004;Urza & Sibold, 2017). In addition to propagule availability, establishment may also be limited by availability of "safe sites" (sensu Harper et al., 1961), which are microsites in the landscape that are favorable to seedling establishment and development. In boreal forests, the dominant tree species can influence fire intensity (Rogers et al., 2015), and in larch forests, stand-replacing fires have been shown to create local cooling due to increased albedo (Chen et al., 2018). Stand-replacing fires represent 50%-60% of fires in northeast Siberia (Krylov et al., 2014); the frequency, extent, and prevalence of stand-replacing fires are increasing in the region in recent decades (Ponomarev et al., 2016;Shvidenko & Schepaschenko, 2013).

Larix cajanderi (Mayr; Cajander larch) is the northernmost occurring tree in the world and is an important species across vast forests in northeastern Siberia (Abaimov, 2010). The ecological feedbacks connecting climate warming and potential carbon release from Yedoma permafrost will be strongly influenced by long-term regeneration dynamics in forests in which L. cajanderi is monospecific or a canopy-dominant tree. In some Larix forests, postfire ecosystem structure and forest development are driven by establishment opportunities for new seedlings, and variation in seedling establishment can alter long-term stand density (Alexander et al., 2018;Alexander, Mack, Goetz, Beck, et al., 2012;Alexander, Mack, Goetz, Loranty, et al., 2012;Chu et al., 2017;Goldammer & Furyaev, 1996;Johnstone & Chapin, 2006). In particular, L. cajanderi regeneration occurs predominantly through seedling establishment from wind-dispersed seeds originating from adjacent stands following stand-replacing fire, which is distinct from the process in some coniferous systems where large trees persist following fire and disperse seeds into the postfire landscape through serotinous cones or other mechanisms (Lavoie & Sirois, 1998;Lotan, 1976). The tree densities established during postfire recruitment of L. cajanderi set the stage for long-term forest density. Species dominance shifts associated with forest succession shifts do not occur due to a lack of other trees species in the system. The opportunity for L. cajanderi seedling establishment is largely relegated to the immediate postfire environment after which seedling establishment may be limited over the short term by competition from quickly establishing grass and shrubs, and over the long term by the formation of a deep soil organic layer (SOL), which acts as a barrier to seedling establishment (Alexander et al., 2018;Shorohova et al., 2009). This regeneration pattern generally leads to a single cohort of saplings and relatively even-aged stands (Shorohova et al., 2009), and some research suggests safe sites may play an important role in Larix regeneration after a fire (Shirota et al., 2006). Regeneration of L. cajanderi forests has important implications for future patterns of productivity, albedo, and belowground conditions that effect carbon storage, nutrient cycling, and water cycling across vast regions of the Siberian Arctic (Alexander et al., 2018;Alexander, Mack, Goetz, Loranty, et al., 2012;Hewitt, Alexander, Izbicki, et al., 2022;Hewitt, Alexander, Miller, et al., 2022;Kropp et al., 2019;Loranty et al., 2018;Paulson et al., 2021;Rogers et al., 2015;Walker et al., 2021;Webb et al., 2017).

The overarching goal of this project was to assess factors influencing tree regeneration in L. cajanderi forests in northeastern Siberia. Recruitment in postfire stands is typically from adjacent forests; thus, we first sought to measure seed production across a gradient of tree density within mature stands of similar age. We used cone abundance as a proxy measure for seed availability and surveyed forest stands across a range of densities to examine the relationship between tree density and reproductive potential in these forests. We also assessed seedling establishment in recently burned stands (<20 years old) with respect to available safe sites. We measured the presence of safe sites for seedling establishment and combined the availability of safe sites with patterns of seedling establishment to estimate the importance of safe sites for tree regeneration in these forests. Finally, we examined the role of safe sites in dictating seedling establishment of L. cajanderi across varying conditions in this endangered landscape.

METHODS

Study area

This research was conducted near the Northeast Science Station in Cherskiy, Sakha Republic,Russia (68.74 N,161.40 E). The boreal forests of this region are dominated by Cajander larch (L. cajanderi), a needle-leaf deciduous conifer, which is the only prominent large tree (Figure 1), with an understory of tall shrubs, short shrubs, forbs, grasses, mosses, and lichens. L. cajanderi has mast cycles of 2-3 years and very little seed dormancy (Abaimov, 2010). The research sites have been described by Alexander, Mack, Goetz, Loranty, et al. (2012), Paulson et al. (2021), andWalker et al. (2021): understory vegetation consists of deciduous shrubs (e.g., Alnus fruticosa, Betula divaricata, B. exilis, Salix pulchra, S. alaxensis, S. glauca, and other Salix species), evergreen shrubs (e.g., Arctous alpine, A. erytrocarpa, Empetrum androgynum, Ledum decumbens, Pinus pumila, Pyrola grandiflora, and Vaccinium vitisidaea), herbs (e.g., Artemisia tilesii, Carex appendiculata, Epilobium angustifolium, Luzula multiflora, Pedicularis lapponica) and grasses (Calamagrostis neglecta), mosses (e.g., Aulacomnium turgidum and Sphagnum spp.), and lichens (e.g., Cetraria cuculata, Cladonia rangiferina) (Paulson et al. 2021;Petrovsky & Koroleva, 1979). Historical fire return interval for Siberian larch forest is ca. 110 years, with intervals approaching ca. 200-300 years in more northern altitudes (Kharuk et al., 2016;Ponomarev et al., 2016).

Field methods

Seed sources

Cone counts and stand surveys were conducted in the summers of 2017 and 2018 in 17 mature L. cajanderi stands of similar age, all of which had regrown from a

large fire that occurred in 1940. The burned area is approximately 2 km from the Northeast Science Station in Cherskiy,Sakha Republic,Russia (68.76 N,161.46 E) and is just east of the Kolyma River approximately 130 km south of where the river enters the Arctic Ocean. These stands were a subset of a broader set of 25 L. cajanderi stands that have been described in detail by Paulson et al. (2021). Fieldwork complications created the need to select a subset of those original 25 stands, and the 17 stands used in our study were selected because they were broadly representative of the tree density gradient present in this forest type. In our 17 stands, densities range from approximately 300-33,000 trees per ha. Within each stand, we established three plots located ~30 m apart. Plot area varied between 30 and 120 m 2 , with greater area sampled in low-density stands. Tree density was measured for each stand by counting every L. cajanderi tree within each plot. Trees selected for cone counts were chosen haphazardly by moving 10 m perpendicular to the plot, spinning a pencil, and following the point to the nearest tree. Trees were felled and separate cone counts were recorded for the "new" cones that were produced in the year of sampling as distinguished morphologically, and for "old" cones that had remained on the tree from the previous years (i.e., new cones were reddish in color with a flexible texture, while old cones were gray in color with a brittle texture; Appendix S1: Figure S1A). Basal diameter and dbh were measured for each sampled tree using calipers or a dbh tape. Tree height was measured using a meter tape after trees were felled. The average cone production per tree in each stand was multiplied by the average number of trees per hectare in that stand to estimate the average number of cones produced by a given hectare of forest.

Cones were collected for the analysis of seed production per cone across the density gradient. In each stand, cones were collected haphazardly from trees, with no more than two cones taken from any individual tree and with one taken from the lower branches and one from near the top of the crown. Mature scales on these cones were counted and multiplied by two in order to estimate seed production, as each scale has the potential to hold two seeds. A total of 156 cones were counted for seed production, 48 from low-density stands, 33 from mediumdensity stands, and 75 from high-density stands as identified from visual observation following Alexander, Mack, Goetz, Loranty, et al. (2012).

We found evidence of a masting cycle on yearly cone production for L. cajanderi in these stands (Appendix S1: Figures S1A and S2). Field observations indicated that 2016 was a mast year, and our own results show a near absence of new cone production for summer 2017 (Appendix S1: Figure S2), with only 62 total new cones counted, 52 of which were found on a single tree. Due to this lack of data for new cone production, the analyses presented here were based on counts of old cones.

Safe sites

L. cajanderi recruitment patterns in relationship to safe site availability were assessed in 15 stands in an area spanning approximately 800 km 2 along the northern portion of the Kolyma River (69.5477 N, 161.3641 E) over the past 18 years. These sites were selected opportunistically from visual observations of burned areas during an expedition by boat on the Kolyma River (the area is largely roadless). In each burned area, three 30-m long transects were established perpendicular to the unburned forest edge, extending into the burned area. The starting point of the first transect was chosen haphazardly 25-35 m from the unburned forest edge, and the other transects were set up at 50-m intervals parallel to the first transect. These sites were also sorted into two broad landscape position categories: upland sites were relatively drier sites (n = 21 transects), whereas lowland sites (n = 23) were "marshier," near rivers, and exhibited conditions and the presence of tussocks. We measured the presence of safe sites along each transect using a point-intercept method by recording whether or not the transect intersected a safe site at 2-m intervals. In this study, we defined safe sites as tussocks, hummocks, tip up mounds, edges of coarse woody debris, and the bases of standing L. cajanderi trees (live or dead) (Appendix S1: Figure S1B-D). These formations (1) can be identified for years after the fire, (2) are known to limit inundation and provide the roots of new seedlings with more ready access to mineral soil, and (3) have been previously identified as potential microtopographic safe sites for larch (Shirota et al., 2006). We also considered any area of bare mineral soil as a safe site for seedling establishment given research for both L. cajanderi (Alexander et al., 2018) and other fire-dependent boreal conifers (Johnstone & Chapin, 2006), indicating that bare mineral soil is associated with seedling establishment. When points along the transect intersected a safe site, the type of safe site was recorded. Seedlings were counted in a 2 × 30 m plot along the transect and categorized by whether the seedling occurred on a safe site, along with the type of safe site. The width of this belt was narrowed to 1 or 0.5 m when the sampling of a transect took more than 45 min due to seedling density. We also collected baseline data regarding other factors that can affect L. cajanderi recruitment. In three 0.5 × 0.5 m plots along each transect, we measured soil moisture, SOL depth, permafrost thaw depth, and percent cover of other vegetation. Soil moisture was measured in the middle of two quadrants of each plot using a moisture sensor (GS-3 and ProCheck, Decagon Devices, Pullman, WA). SOL was measured in two quadrants of each plot by carefully cutting into the soil and measuring the depth with a ruler from the bottom of the moss to the top of the mineral soil. Thaw depths were measured in the center of two quadrants of each plot using a hand-held metal probe that was inserted into the soil until the resistance of frozen soil or ice was reached. Vegetation cover was estimated visually in a single 0.25 m 2 plot on each transect.

Statistical methods

Modeling methods

To assess the influence of safe sites on L. cajanderi seedling establishment, we developed a predictive model. This model is based on the assumption that the total L. cajanderi establishment in a given area is a function of the total number of seeds landing in the area, the proportion of seeds that are viable, the proportion of the area that is made up of safe sites, and the establishment probabilities for seeds that land on safe and unsafe sites. For this model, we assumed that seed viability is uniform and that seeds stochastically landed in an area; thus, we assume that seed rain and seed viability are identical across the landscape. Future work that may include seed baskets, assessment of the spatial arrangement of seed viability, or other experiments would be needed to test this assumption. Even so, the safe and unsafe sites in this study were literally adjacent to one another over very small spatial scales (30-m-long transects) supporting the assumption of uniform seed rain (and viability). Under this assumption, we use seedling presence and arrangement as a post hoc measure of regeneration success as follows:

where S is the total number of seeds entering the area; A s is the proportion of the area occupied by safe sites; A u = 1 -A s , the proportion of the area not occupied by safe sites; E s is the establishment probability of seeds that land on safe sites; E u is the establishment probability of seeds that land on unsafe sites.

Because the total number of seeds entering the area is unknown for these study areas, we used the ratio of seedlings that were found on safe sites as a way to understand the relative establishment on safe versus unsafe sites.

Safe establishment ratio ¼ safe establishment total establishment :

In this comparison, using the safe establishment ratio, the total number of seeds entering the area can be assumed to be equal and can be mathematically factored out of the equation. This equation can then be simplified to the following relationship between the establishment probability of the seeds that land on safe sites (E s ) and the establishment probability of seedlings that land on unsafe sites E u ð Þ:

This relationship E s /E u , which we have deemed "safety power," estimates how much greater the chance of establishment of a seed that lands on a safe site has versus seeds that land elsewhere. For instance, if the safety power equals 3, then seeds that land on safe sites are three times more likely to establish than those that did not land on a safe site. The closer safety power is to 1, the less important safe sites are for understanding L. cajanderi establishment patterns in a given area. In any transect where all the seedlings were found on safe sites, the safety power is infinite, as the equation for safety power has a vertical asymptote when the proportion of safe seedlings equals one. For the sake of display, these infinite values have been changed to 30, larger than the next highest value of 25. While this value is arbitrary, we believe it adequately describes the observed disparity in the establishment rates of safe sites and unsafe sites. For example, with a safety power of 30, an area where the actual establishment probability of a seed that lands on a safe site is 0.1 (or 10%), the establishment probability on unsafe sites is 0.0033 (or 0.33%).

Because we do not have data on seed rain for these sites, nor seed viability, we cannot estimate actual establishment probability; however, our safety power model does allow for an estimate of the magnitude of the difference between these probabilities. For instance, we can estimate that a seed that landed on a safe site was three times more likely to establish than a seed that landed on a nearby unsafe site (safety power equals 3), but we cannot say whether the seed had a 15% chance of establishing whether it landed on a safe site and a 5% chance whether it landed elsewhere versus a 9% chance on safe sites and 3% chance elsewhere.

Statistical testing

Data were scanned for normality using the Shapiro-Wilk normality test and scanned for homogeneity of variance using Bartlett's test. The tree density values were base 2 logarithmically transformed to account for the leftward skew of the data. Cone production per tree and per hectare of forest was also base 2 logarithmically transformed to limit the influence of extreme values and bring the data into a normal distribution. Seed count and cone count data were analyzed using linear and polynomial regression models. Safe site data were analyzed using linear regression models and 95% CIs. Upland and lowland site differences were not normally distributed and were analyzed using the Wilcoxon rank sum test. All analyses and data visualization were carried out using base functions in the R statistical software package (R Core Team, 2018).

RESULTS

Seed sources

There were strong relationships among L. cajanderi density, tree size, and cone production per tree in our research sites. We found a strong negative relationship between tree size and tree density within a given stand (Figure 2A;

p < 0.001, R 2 = 0.713). The number of cones produced by each tree was strongly related to tree size, with larger trees producing more cones (Figure 2B; p < 0.001, R 2 = 0.703).

In stands where the mean tree dbh was <5 cm, those trees tended to have <20 cones per tree and, in contrast, stands with the largest average dbh contained trees that produced >1000 cones per tree (Figure 2B). The number of cones that the average tree produced was negatively related to the density of trees in a stand (Figure 3A; p < 0.001, R 2 = 0.444). When cone production was standardized by stand area, the relationship to tree density was best described by a third-order polynomial model (Figure 3B; p < 0.001, R 2 = 0.467). At ≤500 trees per ha, the total stand cone production was quite low. Cone production per hectare increased rapidly as the tree density increased from low to medium density, and then declined again in high-density stands (Figure 3B). The results of seed counts in cones collected from various stands showed a weak negative trend (p < 0.001, R 2 = 0.074) in seeds per cone as tree density increased (Figure 4).

Safe sites

The percentage of seedlings found on safe sites increased as safe site availability increased (Figure 5A). The relationship between the percentage of the landscape that F I G U R E 2 Relationships among Larix cajanderi size, tree density, and cone production in northeastern Siberia. (A) The strong linear relationship between base 2 logarithm of larch tree density and mean dbh within each stand. (B) The strong linear relationship between mean larch dbh and base 2 logarithm of mean cone production per tree in each stand. Stand density designations reflect transect establishment in the field based on visual observation. Linear units are displayed on the logarithmically transformed axes: the smaller tick marks represent an increase of the smallest unit label on that axis (250 trees per ha and 4 cones per tree for panels A and B, respectively). Gray shaded area represents the 95% CI.

was taken up by safe sites (safe area percent) and the percentage of seedlings that occurred on safe sites (safe seedling percent) was linear (Figure 5A). The 95% CI of this linear relationship line was nearly entirely outside of the dotted, null, 1:1 line, which represents equivalent establishment on safe sites versus other terrain (Figure 5A). Where safe sites occupied <10% of the area, the CI did include the null line, indicating the relative importance of safe sites in seedling establishment was minimal. In three quarters of our sampled transects, the safety power was >1, indicating increased establishment success on safe sites (horizontal dashed line, Figure 5B). The median safety power value across all sites was 2.81, indicating that seeds that landed on safe sites were nearly three times as likely to establish as those that landed elsewhere (Figure 5B).

We found evidence that the importance of safe sites was greater in lowland forests than in upland. The relationship between safe seedling percent and safe area percent was generally more positive in lowland sites (with the 95% CI farther from the null line) than that for upland sites (Appendix S1: Figure S3A,B). The median safety power value in lowland sites was more than 2× higher than that of upland sites, and many values in the lowland site were above 10, indicating 10× greater F I G U R E 3 Relationship between Larix cajanderi tree density and cone production in northeastern Siberia. (A) The relationship between L. cajanderi density and the mean number of cones produced per tree. (B) The relationship between tree density and the estimated number of cones produced per hectare at each stand. Stand density designations reflect transect establishment in the field based on visual observation. All axes have been base 2 logarithmically transformed. Linear units are displayed on the logarithmically transformed axes: the smaller tick marks represent an incremental increase of the smallest unit on that axis (250 trees per ha on both horizontal axes; 4 cones per tree and 25,000 cones per ha for the vertical axes of panels A and B, respectively). Gray shaded area represents the 95% CI.

F I G U R E 4

The number of Larix cajanderi seeds per cone as estimated by cone scales in nine forests that varied in tree density in northeastern Siberia. The line represents the negative linear relationship ( p < 0.001, R 2 = 0.074) and shaded area represents the 95% CI.

probability of success for seedlings to establish on safe sites (Appendix S1: Figure S3C). We did not detect any environmental factors (e.g., thaw depth, soil moisture, vegetative cover) that directly correlated with safety power; however, comparisons among upland and lowland sites indicated differences among these habitat types. Upland sites consistently had a thinner SOL (Appendix S1: Figure S4A). The types of safe sites found at upland and lowland sites also differed, with lowlands typically characterized by the presence of tussocks, which were absent in the upland sites (Appendix S1: Figure S4B). Upland sites had significantly more hummocks than lowland sites (Appendix S1: Figure S4C).

DISCUSSION

As high-latitude wildfires increase in frequency, extent, and severity due to climate warming (Kasischke et al., 2010;Ponomarev et al., 2016), the factors that dictate tree seedling establishment in postfire landscapes will be an increasingly important linkage between the ecology of Arctic forests and the global climate system (Ito, 2005;Kasischke, 2012). In particular, the dynamics of L. cajanderi forests that cover vast swaths of Yedoma permafrost in northern Siberia have strong potential to influence long-term climate feedbacks (Loranty et al., 2021). Regeneration of L. cajanderi in these forests is predominately from establishment following stand-replacing fire when soil conditions are favorable and competition is minimal (Alexander et al., 2018). Propagules to initiate forest regeneration are largely a result of a seed rain from adjacent unburned stands (Alexander et al., 2018;Shorohova et al., 2009) and our study focused on cones as a proxy for seed availability. Our data strongly indicate that seed availability of L. cajanderi is influenced by tree density in fire-adjacent forests and that maximum seed production occurs in stands of intermediate tree density. In very low-density stands, each tree produces many cones; however, there is unused space in these stands and so increasing density from low to medium density results in an increase in cones per hectare. Cone production per area and seed production per cone both decreased in the highest density stands, potentially due to competition among trees or to shading of the forest floor extending the period of frozen soil. These results suggest the potential for complex relationships between fire, forest stand establishment, and the global climate system. For example, the establishment of high-density stands may provide a faster accumulation of aboveground carbon through relatively higher productivity (Alexander, Mack, Goetz, Loranty, et al., 2012), but they may not provide as many seeds to adjacent areas following wildfire, resulting in lower density stands in regenerating forests. Future work is needed to verify our results in the field, including assessments of seed rain, and observations of establishment patterns following fires along gradients of L. cajanderi stand density.

There was strong evidence of annual variation in seed availability within L. cajanderi forests, suggesting that mast cycles will influence long-term forest dynamics. Masting is a reproductive pattern that is present in many forest ecosystems (e.g., Koenig et al., 2003;Minor & Kobe, 2017;Silvertown, 1980), and L. cajanderi forests tend to have heavy cone crops in 3-year cycles (Abaimov, 2010). The timing of a fire in relation to the mast cycle could be a driver of long-term patterns of regeneration, particularly, if there is a mismatch between seed availability and the ephemerally more conducive conditions in the postfire landscape. Seed viability is another factor that could influence propagule availability in the postfire landscape (Batkhuu et al., 2010;Moles & Westoby, 2004). In Larix gmelinii, a closely related larch species, seeds with an underdeveloped embryo (<40 of the embryonic channels filled) had much lower germination and were more commonly found in northern populations

The importance of safe sites in Larix cajanderi seedling establishment in the Siberian Arctic. (A) The relationship between the percentage of L. cajanderi seedlings observed growing on safe sites and the percentage of the area occupied by safe sites along transects. The dotted line represents a 1:1 ratio, which would indicate no difference in tree seedling establishment on safe sites compared with the rest of the landscape. The solid line represents the trend line of the statistically significant linear relationship ( p < 0.001, R 2 = 0.53), with the gray area showing the 95% CI. Where this CI is above the dotted line, there is 95% confidence that the seedlings are preferentially establishing on safe sites. (B) The distribution of the data points in terms of their safety power, which is the magnitude of the difference in establishment in safe sites versus unsafe sites. The horizontal dotted line represents a safety power of 1, which indicates no difference in safe site establishment success. (Barchenkov, 2011). L. cajanderi has relatively low germination rates (12%-16%; Alexander et al., 2018) and future studies are needed to assess the influence of stand density on seed viability. The factors affecting seed dispersal are another key area of future research given that relatively small changes in dispersal distance can have substantial effects on the species distributions (Clark et al., 1999). L. cajanderi is a wind-dispersed species and work is needed to assess the size and shape of the seed shadow of this species, and to examine how the spatial pattern of seed dispersal is influenced by stand characteristics and landscape features.

A critical determinant of establishment success in postfire L. cajanderi forest will be microsite features that determine seedbed characteristics. Fire plays an important role in these systems by decreasing the organic layer depth to a level more conducive to seedlings establishment (Alexander et al., 2018;Beck et al., 2011;Chu et al., 2017;Wieczorek et al., 2017). Even in postfire conditions, establishment success varies due to microsite conditions, and long-term forest dynamics may be strongly regulated by the relative importance of safe sites for seedlings. Increased seedling establishment success on safe sites is related to factors such as more stable conditions in the presence of disturbance processes and better access to mineral soil for developing roots (Brown & Johnstone, 2012;Shirota et al., 2006). We found that L. cajanderi seedlings were positively associated with safe site availability in the landscape, and that upland and lowland sites varied in the relative importance of safe sites. One caveat of these findings is that the total number of seeds entering our study areas, and their spatial arrangement, was unknown, preventing us from being able to estimate the actual establishment probability of a seed that lands on a safe site versus an unsafe site. Measuring seed establishment probabilities could be a useful direction for future research that could be conducted using seed addition experimental designs (Brown et al., 2015). Even so, our evidence from seedling establishment patterns strongly indicate that safe sites facilitate L. cajanderi establishment and that the relative importance of safe sites varies across landscape conditions. Specifically, we found that in lowland areas, the importance of safe sites was much greater than in upland areas, suggesting that in uplands, ambient seedbed conditions were consistently more favorable than in lowlands. This difference may have been related to relatively lower organic layer depth in upland sites, which may provide better overall conditions for seedling establishment reducing the value of safe sites. Field observations suggest lowland sites may have consistently wetter soil conditions, creating the possibility for more consistent inundation that may increase the relative importance of safe sites for seedling establishment success. Testing this flooded-soil hypothesis and assessing other factors, including seed predation, that may differ across upland and lowland sites may help elucidate the factors limiting L. cajanderi seedling success.

In summary, we found strong evidence that tree density and safe site availability may play important roles in long-term dynamics of L. cajanderi forests in the Siberian Arctic. Our data indicated a unimodal relationship between L. cajanderi forest density and cone production that suggests competition or other factors are limiting per area cone production in the highest density stands. Additionally, we found a significant (albeit weak) negative relationship between tree density and per cone seed production. We also found strong evidence that L. cajanderi seedling establishment is favored on safe sites and demonstrated that the relative importance of safe sites is greater in stands classified as lowlands in comparison to upland areas. In L. cajanderi forests, uplands may have more favorable ambient seedbed conditions, whereas, in some lowland areas, regeneration may be strictly limited to safe sites. The biological reason for this pattern is presently unknown; however, this may be related to relatively thinner SOLs in uplands and that safe sites provide drier conditions in relatively marshy lowlands. Taken together, maximum reproductive potential for L. cajanderi following stand-replacing fire may be in postburn areas that have upland conditions and are adjacent to moderate-density forests. Higher or lower density forests may produce fewer seeds per area, and in lowland sites, regeneration may be relegated to the relatively small portion of the landscape occupied by safe sites. Climate-driven shortening of fire return intervals and increasing fire severity have the potential to initiate complex shifts in vegetation dynamics of Arctic forests with important implications for global carbon cycling (Schaphoff et al., 2016;Shuman et al., 2017). Our data suggest the potential for complex linkages between forest density, propagule availability, fire, safe sight colonization, and seedling establishment that may regulate forest regeneration and long-term dynamics in the understudied and endangered L. cajanderi forests of the Siberian Arctic.