Residual forest structure influences behaviour of Pacific marten (Martes caurina) on post-fire landscapes

Keywords: carnivores, fire ecology, forest structure, habitat use, landscape management, montane forest, salvage logging, wildfire.

Wildfires are increasingly prevalent on North American landscapes. Large fires (>1000 ha) have become more common in conifer forests of Canada and the United States (Hessburg et al. 2005; Miller et al. 2009), and climate change models forecast substantial fire-driven shifts in forest composition and connectivity in the coming decades (McKenzie et al. 2004; Westerling et al. 2011). Fires have long-term consequences for forest-associated birds (Schieck and Song 2006), amphibians (Pilliod et al. 2003), and mammals (Fisher and Wilkinson 2005). Although some wildlife benefit from these disturbances (Fontaine and Kennedy 2012), the resilience of many species has not been determined (Hutchen et al. 2017; Volkmann et al. 2020).

Landscape structure matters to wildlife post-fire (Saab and Powell 2005; Fontaine and Kennedy 2012). Fire initially alters forests by killing trees, opening the canopy, and removing understory features such as groundcover and coarse woody debris (hereafter ‘deadfall’), which often creates a mosaic of lightly-to-severely-burned habitats (Agee 1993, 1998). Burns are therefore heterogeneous in terms of small-scale structures, such as green trees, that wildlife use for foraging (Covert-Bratland et al. 2006; Wood et al. 2007), shelter (Magoun and Vernam 1986; Simanonok and Burkle 2019), and reproduction (Meehan and George 2003; Cunningham and Ballard 2004). Residual structures, such as trees within lightly-burned stands or snags in more heavily-burned areas, influence wildlife distributions on burned landscapes (Barton et al. 2014; Swan et al. 2015). For species that select mature stands (‘forest specialists’), post-fire residuals can provide islands of suitable resources amid low-quality habitat (Robinson et al. 2014; Vanbianchi et al. 2017).

Further disturbances from post-fire salvage logging – the harvest of fire-damaged trees – often removes or reconfigures these residual habitats. Salvage logging is a major component of forest harvest in regions such as British Columbia (BC), where mountain pine beetles (Dendroctonous ponderosae) are projected to kill >50% of merchantable lodgepole pine (Pinus contorta) by the mid-2020s (Dhar et al. 2016), and where record-setting fires burned ~2.5 million ha in 2017–2018 (BC Wildfire Service 2020). However, fire and salvage logging create fundamentally different conditions for wildlife (Franklin et al. 2002; Lindenmayer et al. 2008). Conventional timber harvest influences animal distributions (Hargis et al. 1999; Nelson et al. 2019), home range structure (Lambert and Hannon 2000; Potvin et al. 2000), and habitat connectivity (Cushman et al. 2011; Popescu and Hunter 2011). These effects may be more pronounced on salvage-logged landscapes, where resources were already altered from burning. However, long-term impacts of salvage logging on wildlife remain uncertain (Nappi et al. 2004; Thorn et al. 2018).

Forest specialists such as marten (M. americana and M caurina; Hargis et al. 1999) are sensitive to disturbance and may respond negatively to fire and post-fire salvage logging. These carnivores select specific habitat features at multiple spatial scales (Buskirk and Powell 1994; Thompson et al. 2012), and are indicators of forest health in Canada and the United States (Watt et al. 1996; Guppy 2008). Understanding marten behaviour post-fire would help managers maintain marten populations in fire-prone regions.

Marten occupy burned landscapes (Magoun and Vernam 1986; Latour et al. 1994; Paragi et al. 1996; Volkmann and Hodges 2021), but many details of their post-fire habitat use are unclear. Like other carnivores, marten select habitats based on vegetation (Porter et al. 2005; Seip et al. 2018), prey (Vigeant-Langlois and Desrochers 2011), and conspecifics (Hutchings and White 2000). Identifying spatial changes in marten behaviour can thus provide insight into habitat quality (Powell 2004). For example, marten movements vary with small-scale changes in overhead cover (Nams and Bourgeois 2004; Moriarty et al. 2016). The spatial distribution of marten scent-marks (urine and scat) reflects the distribution of high-quality resources encountered while foraging (Hargis and McCullough 1984; Porter et al. 2005).

In unburned forests, high-quality marten habitats have abundant vertical structure (trees and snags), horizontal structure (uneven-aged stands), understory cover (saplings, shrubs, and deadfall), and ground vegetation (Thompson et al. 2012). Marten avoids meadows and cut-blocks, and choose direct routes across low-quality habitats (Cushman et al. 2011). Trees and snags provide cover for resting and denning (Buskirk et al. 1989; Ruggiero et al. 1998), hunting prey (Coffin et al. 1997; Andruskiw et al. 2008), and avoiding predators (Herman and Fuller 1974). Deadfall creates overwinter habitat for prey such as southern red-backed voles (Myodes gapperi; Sullivan et al. 2011) and shelter for marten (Taylor and Buskirk 1994). Saplings, shrubs, and ground cover offer favourable habitat for snowshoe hares (Lepus americanus; Lewis et al. 2011), northern red-backed voles (Myodes rutilus; Vanderwel et al. 2010), and other small mammals. These habitat features likely matter to marten on post-fire landscapes as well.

We studied marten habitat use in Washington, USA, and BC, Canada, where wildfires have caused substantial landscape changes over the past 15 years (2005–2020). We developed habitat selection models to identify small-scale structural features influencing marten behaviour post-fire, and to assess further impacts of post-fire salvage logging on marten habitat quality. Given the strong links between marten and structurally-complex habitats, we hypothesised that marten would respond to residual structure on burned landscapes. We expected foraging marten to select sites with relatively abundant residual trees, deadfall, understory cover, and ground vegetation. Likewise, we expected marten to move more quickly through sparsely-covered (low-quality) habitats. Finally, we expected marten to scent-mark at sites with relatively abundant deadfall and rocks – structures suitable for broadcasting scent (Hargis and McCullough 1984; Porter et al. 2005). Given the importance of scent for marking resources, we anticipated strong selection for high-quality sites along foraging paths. For both behaviours, we expected marten to avoid salvage-logged areas due to loss of forest structure.

We examined marten populations on post-fire landscapes in north-central Washington (48.790°N, −119.953°W) and central BC (52.071°N, −122.436°W), approximately 400 km apart. We chose study areas representative of regional forest types, management strategies, and disturbance events. Because numerous large, recent fires have occurred in Washington and BC (BC Wildfire Service 2020; National Interagency Fire Center 2020), we also selected landscapes based on overall accessibility.

The 2006 burn (Tripod Complex wildfire, 70 575 ha), is situated within the Okanogan-Wenatchee National Forest (Fig. 1, Table 1). This landscape has rugged topography typical of the northeastern Cascade Range, with alpine parkland above ~2100 m and shrub-steppe communities below ~1000 m (Koehler et al. 2008). Pre-fire, much of the study area was mature, highland conifer forest composed of lodgepole pine, Engelmann spruce (Picea engelmannii), and subalpine fir (Abies lasiocarpa), with Douglas fir (Pseudotsuga menziesii), ponderosa pine (Pinus ponderosa), and quaking aspen (Populus tremuloides) at lower elevations. The regenerating landscape is dominated by snags, deadfall, conifer saplings, and willows (Salix spp.). We sampled elevations of 1500–2000 m.

Fig. 1.  Post-fire landscapes examined in this study: (a) the 2006 Tripod Complex wildfire in north-central Washington, USA; (b) the 2010 Meldrum Creek and 2017 Hanceville-Riske Creek wildfires in central British Columbia, Canada. White borders denote burn perimeters and map colours are burn severities derived from the differenced Normalised Burn Ratio (2006 and 2010 burns) and Relativised Burn Ratio (2017 burn): gray = unchanged, yellow = low, orange = moderate, and red = high.

Table 1.  Characteristics of post-fire landscapes examined in this study.

The Okanogan-Wenatchee National Forest is not actively managed for timber harvest; a single major access road runs north-south through the study area. Aside from a ~10 m road buffer cleared of hazard trees, the 2006 burn has seen little human disturbance post-fire. State forests east of the burn contain cut-blocks of various ages, but were not sampled in this study.

The 2010 burn (Meldrum Creek wildfire, 15 553 ha) and the 2017 burn (Hanceville-Riske Creek wildfire, 239 340 ha) are situated on the Chilcotin Military Reserve and multi-use (Crown) land near Williams Lake (Fig. 1, Table 1). This landscape has gentle topography at elevations of 900–1000 m, and straddles the Interior Douglas-fir and Sub-Boreal Pine-Spruce biogeoclimatic zones (Hope et al. 1991). Pre-fire, the study area contained mature lodgepole pine, Douglas fir, hybrid Engelmann-white spruce (Picea glauca), and quaking aspen. The regenerating landscape is composed of cut-blocks, natural meadows, snags, and residual trees with a patchy understory of conifer saplings, aspen, and willows. Approximately 15% of the 2010 burn re-burned in 2017.

The BC study area has an extensive road network and a complex history of timber management. Approximately 6000 ha (26%) of the 2010 burn was salvage-logged in large blocks in 2010–2012. Some selective harvest also occurred following post-fire insect outbreaks. Salvage logging has occurred at similar intensity in the 2017 burn, beginning in winter 2017–2018.

In winter (December–March, 2016–2019), we located marten trails weekly along ~75 km of snowmobile-accessible forestry roads chosen ad hoc to cover burned, unburned, and salvage-logged habitats representative of each study area (Volkmann 2021). Suitable trails were 1–4 days old with clearly-identifiable tracks for >50 m. We recorded each trail in straight-line 5 m segments (‘steps’) marked with pin flags and georeferenced at every sixth step (30 m). We also georeferenced all scent-marking (urine or scat) unambiguously associated with these trails. Whenever possible, we back-tracked trails to avoid influencing marten behaviour; forward-tracking was limited to trails >2 days old. We prioritised back-tracking >1 km from previously-sampled trails to maximise our coverage of habitats and marten home ranges.

Marten leave easily-identifiable footprints in snow (see Supplementary material; Elbroch 2003). We counted footprints per 5-m step to index relative movement speed (‘track count’), where a lower track count implies faster movement (Hodges et al. 2014). In some cases only a minimum track count was possible due to poor tracking conditions or subnivean travel. We applied linear interpolation to these segments to determine a most likely track count; missing entries at trail endpoints were assigned a nearest-neighbour value.

In summer (July 2017 and July–August 2018 for Washington; April–July 2018 and March 2019 for BC), we documented habitats along marten trails and at randomly-located sites available to marten. For habitats used by marten (‘foraging sites’), we sampled a random subset of 3–6 sites from each trail, excluding sites <20 m from one previously sampled (if a trail doubled back). We also sampled all sites where marten scent-marked (‘marking sites’). To characterise habitats available to marten but not necessarily used, we randomly selected sites on each landscape via ArcMap where no post-fire salvage logging had occurred (‘burned sites’), sites with salvage logging in the 2010 burn (‘salvaged sites’), and sites unaffected by fire for >50 years (‘unburned sites’). To mimic the spatial autocorrelation of marten trails, we sampled available habitats in sets of 3–6 sites separated by a randomly-generated number of 10 m increments (30–200 m) and a randomised bearing; we excluded sites <20 m from previously sampled sites.

At each site, we established a 10 m radius circular plot containing a randomly-oriented 20 m transect for measuring deadfall. For vertical structure, we recorded diameter at breast height (DBH) to the nearest 0.1 cm for all trees and snags within the plot; we counted trees and snags <7.5 cm DBH as saplings, and any snags with a lean angle >45° from vertical as deadfall. For deadfall, we recorded diameter to the nearest 0.1 cm where each piece crossed the transect; we excluded pieces <7.5 cm in diameter at the transect or <1 m long. We determined decay class for each snag and piece of deadfall: (1) dead foliage present, (2) foliage absent but bark intact, (3) bark absent but heartwood intact, or (4) heartwood soft and crumbling (Resources Information Standards Committee (RISC) 2007). We recorded height to the nearest 50 cm (0–200 cm) for each sapling and shrub within a 2 m radius circle at the centre of each site. For groundcover, we visually assessed total grass/forb, moss/lichen, bare soil, and bare rock to the nearest 5%, using three 1 m radius circles centred at 5, 10 and 15 m along the deadfall transect. Finally, we measured canopy closure with a spherical densiometer from the centre of each 10 m plot, averaging four readings.

We mapped burn perimeters in ArcMap using shapefiles from the USGS Geosciences and Environmental Change Science Centre (http://rmgsc.cr.usgs.gov/outgoing/GeoMAC) and DataBC (https://catalogue.data.gov.bc.ca/dataset). Burn severity layers were derived from Landsat mapping of near-infrared landscape reflectance at ~12 months post-fire; these datasets came from the Monitoring Trends in Burn Severity project and the BC Ministry of Forests, Lands, Natural Resource Operations and Rural Development. Mapping of the 2006 and 2010 burns used the differenced Normalised Burn Ratio (Key and Benson 2006), and the 2017 burn used the more recently developed Relativised Burn Ratio; in practice the two metrics have a ~2% difference in accuracy (Parks et al. 2014). We analysed burn severity layers as rasters at a 30 × 30 m (0.09 ha) resolution.

To estimate deadfall volume, we used the formula:

where V is deadfall volume in m²/ha, k is a unit conversion constant, a is the lean angle correction factor, c is the slope correction factor, L is the survey transect length, and d is deadfall diameter at the transect. We used k = 1.234 and set both a and c at 1 based on recommended methods for North American montane forests (Woodall and Williams 2007).

Our final dataset contained 28 habitat attributes (Table 2). To maximise the dataset available for models, we used predictive mean matching and polytomous regression with the R package ‘mice’ to impute missing values due to recording errors (van Buuren and Groothuis-Oudshoorn 2011). These entries represented <3% of the original dataset and were deemed to be missing at random. We converted continuous variables representing DBH, diameter, decay class, and height class into categorical variables to capture biologically meaningful missing data – for example, a treeless site would have no entry for tree DBH on a continuous scale, but this result is still relevant for marten habitat use. For each attribute, we used the R package ‘boot’ to bootstrap means and 95% confidence intervals with the bias-corrected and accelerated method (Canty and Ripley 2019). Each bootstrap procedure took 10 000 replicates. We pooled habitat data for all years.

Table 2.  Characteristics of burned and unburned landscapes measured in this study.

We built generalised linear mixed models with the R package ‘MASS’ (Venables and Ripley 2002) to determine whether heterogeneous cover affected marten movements. Marten trail characteristics vary in response to local habitat quality, particularly small-scale differences in overhead cover (Moriarty et al. 2016; Slauson et al. 2017). Thus, we expected faster movement (fewer tracks per 5 m) in open areas relative to closed-canopy areas along marten trails. For our models, canopy closure was a fixed-effect predictor of movement speed, and year/trail ID were nested random predictors. We used a Penalised Quasi-Likelihood model fit with an identity link based on inspection of quantile-comparison plots, since track count distributions showed a strong positive skew (Wolfinger and O’Connell 1993). We tested for differences in track counts among burns using an analysis of deviance against a null model of burn ID ~ 1.

For each burn, we built generalised linear models to separately examine marten selection of foraging and marking (‘used’) sites compared to burned (‘available’) sites. We a priori excluded salvaged and unburned sites from models because the 2006 and 2017 burns lacked salvage-logged areas, and because marten in the 2010 burn avoided salvage-blocks (Volkmann and Hodges 2021). We pooled habitat data for all years. To reduce the likelihood of nonsense models, we first screened for individually insignificant habitat attributes using single-factor logistic regressions for continuous variables and Fisher’s exact test for categorical variables. We then used the R package ‘Hmisc’ to calculate Pearson correlations for habitat attributes in each dataset (Harrell and Dupont 2020). From correlograms, we identified a subset of individually significant habitat attributes with pairwise correlations of r² < 0.25; this relatively strict cutoff ensured that only <10 strongly independent variables were retained for each model run (Table 2; Schober et al. 2018).

We applied multi-model inference with the R package ‘glmulti’ to build and rank multivariate logistic regressions linking habitat attributes to marten foraging and marking sites, and to assess the individual importance of each attribute across candidate models (Calcagno 2019). We used differences in Akaike’s Information Criterion (ΔAIC), IC weight, and model diagnostics to judge model performance, retaining models with ΔAIC < 2 and individual variables occurring in >50% of models (Anderson and Burnham 2002). We estimated effect sizes using Cohen’s d. Finally, to assess the quality of salvage-logged areas for marten, we compared habitat attributes of foraging and marking sites in the 2010 burn to those of salvaged and unburned sites, using two-tailed Welch’s t-tests.

We back-tracked 100 marten trails (65.5 km total): 50 trails (24.3 km) in the 2006 burn, 33 trails (25.2 km) in the 2010 burn, and 19 trails (16.0 km) in the 2017 burn. Individual trails were 70–2200 m long with 42–440 5-m steps. We recorded 70 scent-marks: 46 sites with urine, 21 with scat, and 3 with both.

We measured track counts from 2317 trail segments: 871 in the 2006 burn, 888 in the 2010 burn, and 558 in the 2017 burn (Fig. 2). Mean track counts were 10.2 ± 0.2 per 5-m step in the 2006 burn, 9.9 ± 0.2 in 2010 burn, and 10.5 ± 0.3 in the 2017 burn; track counts did not differ among burns (residual deviance = 229.52, d.f. = 2314, P = 0.517) or across years within each burn. Track counts varied with canopy closure in the 2006 burn (est. = 0.020 ± 0.005, t₈₂₀ = 4.22, P < 0.001) and the 2010 burn (est. = 0.015 ± 0.004, t₈₅₄ = 3.48, P < 0.001). Marten movement speed increased 15–20% overall as canopy closure decreased. We did not see this behavioural shift in the 2017 burn (est. = 0.009 ± 0.006, t₅₃₈ = 1.59, P = 0.119).

Fig. 2.  Influence of canopy closure on marten movement speed. Data are from burned landscapes in north-central Washington and central British Columbia: (a) 2006 burn, (b) 2010 burn, and (c) 2017 burn. Lower track counts indicate faster movement and imply lower-quality sites for marten. Solid lines denote model fit (black), confidence intervals (blue), and prediction intervals (red), based on generalised linear mixed-models. Horizontal dashed lines denote mean track counts; vertical dashed lines indicate where the model fit crosses the mean. We have omitted three outliers with track counts >30 (one in the 2006 burn; two in the 2010 burn).

We surveyed habitats at 476 sites in Washington and 995 sites in BC. Marten responded to both overstory and understory habitat attributes when foraging (Table 3) and scent-marking (Table 4), and had stronger patterns of selection in the 2006 and 2017 burns compared to the 2010 burn (see Supplementary material). Foraging marten consistently selected mesic sites with greater moss/lichen cover, less bare soil, and less-decayed snags. Marten scent-marks were associated with sheltered sites with greater canopy closure and a greater basal area of trees. Neither behaviour was strongly influenced by the presence of large trees or snags.

Table 3.  Marten foraging sites in relation to habitat features on post-fire landscapes.

Table 4.  Marten scent-marking sites in relation to habitat features on post-fire landscapes.

In the 2006 and 2017 burns, foraging marten selected relatively intact sites with lower burn severity, greater canopy closure, and more trees compared to availability; selection was weaker in the 2010 burn (Table 3). Ten habitat attributes best explained marten foraging post-fire (Table 5). In the 2006 burn, marten chose foraging sites with lower burn severity (X² = 27.51, d.f. = 3, P < 0.001), a greater basal area of snags (21.9 [19.7–24.1] vs 16.9 [14.9–19.0] m²/ha), more vertical structures (1035 [918–1160] vs 622 [528–750] per ha), and greater moss/lichen cover (22.5 [19.7–25.7] vs 9.5 [7.5–12.1]%; Figs 3, 4; see Supplementary material). Marten in the 2010 burn similarly chose foraging sites with greater moss/lichen cover (14.5 [12.8–16.3] vs 10.1 [8.6–11.9]%), along with more saplings (9937 [8146–12 775] vs 7042 [5771–8966] per ha) and fewer shrubs (15 389 [13 215–17 847] vs 19 397 [16 293–23 351] per ha). Marten in the 2017 burn chose foraging sites with greater canopy closure (52.1 [48.4–55.9] vs 39.3 [35.5–43.5]%), fewer snags (300 [266–338] vs 463 [402–530] per ha), more saplings (4726 [3551–6642] vs 1870 [1366–2575] per ha), and less bare rock (1.9 [1.5–2.5] vs 4.1 [3.3–5.1]%).

Table 5.  Top models predicting marten foraging sites on burned landscapes.

Fig. 3.  Marten foraging (‘F’) and scent-marking (‘M’) sites on burned landscapes in north-central Washington and central British Columbia, compared to overstory habitat features at burned (‘B’), unburned (‘U’, green), and post-fire salvage-logged (‘S’, red) sites. Horizontal bars denote bootstrapped means and 95% confidence intervals, and asterisks denote statistically significant differences from burned (available) sites. Data points have horizontal jitter.

Fig. 4.  Marten foraging (‘F’) and scent-marking (‘M’) sites on burned landscapes in north-central Washington and central British Columbia, compared to understory habitat features at burned (‘B’), unburned (‘U’, green), and post-fire salvage-logged (‘S’, red) sites. Horizontal bars denote bootstrapped means and 95% confidence intervals, and asterisks denote statistically significant differences from burned (available) sites. Data points have horizontal jitter, and we have omitted 14 outliers with sapling densities >60 000/ha (four in the 2006 burn, seven in the 2010 burn, and two in the 2017 burn).

Marten foraging sites in burns differed from unburned areas in terms of habitat structure. Compared to unburned sites, marten in the 2006 burn selected foraging sites with a greater basal area of snags (21.9 [19.7–24.1] vs 3.9 [2.8–5.6] m²/ha), more saplings (12 027 [9834–15 337] vs 3354 [2520–4647] per ha), and greater moss/lichen cover (22.5 [19.7–25.7] vs 7.1 [5.7–8.9]%; Fig. 3, 4; see Supplementary material). Marten in the 2010 burn similarly chose foraging sites with more saplings (9937 [8146–12 775] vs 3842 [3048–5487] per ha) than at unburned sites (but less moss/lichen cover – 14.5 [12.8–16.3] vs 27.7 [24.3–31.2]%), while marten in the 2017 burn chose foraging sites with a greater basal area of snags (5.0 [4.1–6.3] vs 1.5 [0.9–2.4] m²/ha) and less moss lichen cover (6.7 [5.7–8.0] vs 27.7 [24.3–31.2]%). Foraging sites in all three burns had significantly less canopy closure compared to unburned sites.

Marten in the 2006 and 2010 burns chose structurally-complex sites for scent-marking, with greater canopy closure and more vertical structures; selection was weaker in the 2017 burn, likely because of low sample sizes (Table 4). Marten chose marking sites with abundant deadfall in the 2006 burn, but avoided bare rock in the 2006 and 2010 burns. Eight habitat attributes best explained marten scent-marking post-fire (Table 6). Marten in the 2006 burn chose marking sites with lower burn severity (X² = 11.25, d.f. = 3, P = 0.01), a greater basal area of vertical structures (37.4 [30.6–43.4] vs 20.0 [17.6–22.7] m²/ha), and greater deadfall volume (263.5 [177.6–375.2] vs 153.4 [132.8–178.4] m³/ha; Fig. 3, 4; see Supplementary material). In the 2010 burn, marten chose marking sites with greater canopy closure (54.8 [46.7–62.7] vs 29.3 [25.3–33.6]%), more snags (224 [173–292] vs 119 [98–145] per ha) and saplings (16 374 [10 186–27 477] vs 7042 [5771–8966] per ha), and less bare rock (2.5 [1.3–4.2] vs 6.8 [5.4–8.4]%). Similarly, marten in the 2017 burn chose marking sites with greater canopy closure (71.8 [56.6–80.0] vs 39.3 [35.5–43.5]%), more saplings (11 597 [4078–32 428] vs 1870 [1366–2575] per ha), and less bare soil (30.3 [20.0–54.0] vs 54.2 [50.1–58.2]%).

Table 6.  Top models predicting marten scent-marking sites on burned landscapes.

Marking sites in burns also differed from unburned areas in terms of habitat structure. Compared to unburned sites, marten in the 2006 burn selected marking sites with a greater snag basal area (28.2 [21.2–35.4] vs 3.9 [2.8–5.6] m²/ha) and greater deadfall volume (263.5 [177.6–375.2] vs 109.2 [90.1–134.8] m³/ha; Fig. 3, 4; see Supplementary material). In the 2010 burn, marten chose marking sites with a greater snag basal area (3.1 [2.2–4.6] vs 1.5 [0.9–2.4] m²/ha) and more saplings (16 374 [10 186–27 477] vs 3842 [3048–5487] per ha). Marking sites in the 2010 and 2017 burns had significantly less deadfall volume compared to unburned sites.

In the 2010 burn, marten selected habitats substantially different from post-fire salvage-logged areas (Fig. 3, 4). Marten foraged at sites with significantly greater canopy closure (33.7 [29.6–37.8] vs 1.9 [1.4–2.8]%), more trees (258 [221–301] vs 5 [2–10] per ha) and snags (132 [113–154] vs 26 [20–34] per ha), and greater grass/forb cover (65.8 [63.0–68.5] vs 52.6 [50.1–55.0]%) than what was available at salvaged sites (Table 7; see Supplementary material). Scent-marking sites had significantly higher sapling density (16 374 [10 186–27 477] vs 7535 [6414–9067] per ha) compared to salvaged sites. Conversely, marten foraged at sites with significantly less deadfall (60.1 [49.0–73.7] vs 146.2 [128.3–167.2] m³/ha) and less bare rock (4.9 [3.8–6.4] vs 11.5 [9.8–13.6]%) compared to salvaged sites.

Table 7.  Habitat features at post-fire salvage-logged sites versus sites selected by marten in burns.

Residual forest structure matters to marten post-fire. In Washington and BC, marten altered their movements in response to canopy closure. Marten selected structurally-complex habitats in burns, foraging and scent-marking at sites with closed canopies, abundant vertical structures, and dense regeneration. These features correspond to lightly-burned stands less altered by fire.

Previous work hinted at the importance of structure for marten post-fire. Marten evolved with wildfire and probably benefit from the shifting mosaic of resources on burned landscapes (Koehler and Hornocker 1977; Fisher and Wilkinson 2005). Trappers in Alaska associated marten with deadfall and shrubs <3 years post-fire (Stephenson 1984). Radio-collared marten used an Alaskan burn extensively at 7 years post-fire, selecting sites with abundant trees and deadfall (Magoun and Vernam 1986). Our data from landscapes burned in 2006 (10–13 years post-fire), 2010 (6–9 years post-fire) and 2017 (1–2 years post-fire) broadly agree with these earlier studies.

Marten in the 2006 and 2017 burns selected lightly-burned sites (‘unchanged’ to ‘low’ burn severity) over severely-burned ones. Prior studies did not link marten behaviour to burn severity (Volkmann et al. 2020), but our results resemble observations of other forest specialists including Canada lynx (Lynx canadensis; Vanbianchi et al. 2017), squirrels (Sciurus arizonensis and S. aberti; Ketcham et al. 2017) three-toed woodpeckers (Picoides dorsalis; Murphy and Lenhausen 1998), spotted owls (Strix occidentalis; Eyes et al. 2017), and saproxylic insects (Nappi et al. 2010). Since burn severity relates to tree mortality (Key and Benson 2006), lightly-burned sites often resemble pre-fire conditions in terms of canopy closure, tree abundance, and understory structure.

Vertical structure influenced marten behaviour post-fire. Marten avoided open areas, consistent with these habitats having greater exposure to predators (Herman and Fuller 1974), fewer preferred prey (Hutchen and Hodges 2019; Kelly and Hodges 2020), and fewer subnivean access points (Corn and Raphael 1992). Relative importance of residual trees and snags differed among landscapes. Foraging marten selected for abundant snags in the 2006 burn, had no clear patterns of use in the 2010 burn, and chose densely treed sites with few snags in the 2017 burn. Snags provide less cover for marten, and prey such as red squirrels (Tamiasciurus hudsonicus) are more abundant in treed stands (Allard-Duchêne et al. 2014; Kelly and Hodges 2020). Marten in the 2006 burn may have used snag-rich stands as movement corridors, weakening the selection for trees that we expected (Volkmann and Hodges 2021). Alternatively, subnivean access from snags and a regenerating understory may have supported more prey in the 2006 burn.

Marten also responded to understory features, choosing sites with more saplings and moss/lichen cover. Marten select similar mesic features on other landscapes (Bull et al. 2005; Proulx et al. 2006). Mesic sites regenerate more rapidly post-fire (Dwire and Kaufmann 2003; Halofsky and Hibbs 2009), creating structural complexity sooner for marten and prey. Understory complexity benefits snowshoe hares (Hutchen and Hodges 2019) and meadow voles (Microtus pennsylvanicus; Sullivan et al. 2000), which marten eat when red-backed voles are scarce (Paragi et al. 1996).

Combined understory and overstory features influenced marten scent-marking. Marten chose marking sites with greater coverage from trees and saplings. Structurally-complex sites with abundant deadfall were important to marten in the 2006 burn, but use of deadfall in the 2010 and 2017 burns was ambiguous. Mustelids use scent to signal the presence of prey (Hutchings and White 2000). Although marten frequently investigated deadfall in subnivean forays, we could not determine prey abundance at these sites. Marten also avoided bare rock when scent-marking, contrary to expectations (Hargis and McCullough 1984). Elevated deadfall and boulders help broadcast scent (Pulliainen 1982; Porter et al. 2005), but these structures may have been too low or scarce for selection in BC, compared to densely interlocking deadfall in the 2006 burn (Magoun and Vernam 1986).

Marten showed weaker habitat selection in the 2010 burn compared to the 2006 and 2017 burns, implying greater similarity between used and available sites. We suspect that this similarity was due to salvage logging homogenising large areas of the burn. In effect, marten in the 2010 burn had fewer suitable post-fire habitats (Volkmann and Hodges 2021).

Surprisingly, we did not detect marten in unburned forests bordering either study area. Marten similarly selected a 6-year-old burn over mature forest in Alaska, likely due to differences in habitat structure and prey abundance (Paragi et al. 1996). Prey densities in unburned and lightly-burned areas of the 2010 burn were similar during our study (Kelly and Hodges 2020; Kelly 2021); this combined with greater snag density may have offered better habitat in the 2010 burn (Payer and Harrison 2003; Bull et al. 2005). Alternatively, marten may have chosen burns to avoid larger carnivores (fishers Pekania pennanti and lynx; Volkmann 2021). A broader survey of burned and unburned landscapes would help determine how these factors influence marten habitat use.

Post-fire salvage logging severely degraded marten habitat. Marten in the 2010 burn chose sites with substantially more structural complexity than salvaged sites. When crossing open areas, marten took direct routes between available trees and saplings, and rarely approached salvage-blocks (Volkmann and Hodges 2021). Although deadfall was abundant at salvaged sites, these structures were usually stripped of branches and flush with the ground, making subnivean access difficult (Wilbert et al. 2000). Notably, salvage-blocks in the 2010 burn also lacked debris piles to provide cover for marten and microtine prey (Sullivan and Sullivan 2021; Sullivan et al. 2021).

Marten have not been directly studied on post-fire salvage-logged landscapes. Indirectly, Steventon and Daust (2009) modelled marten habitat in response to post-insect salvage logging in BC and forecast substantial marten declines in the next 20–40 years, even with lower quotas for conventional timber harvest. More recent work suggests that marten use residual structures in salvage blocks for resting and denning, but these features are a poor long-term substitute for forest cover (Wilk and Raphael 2018). GPS-tracked marten in the 2010 burn avoided salvage-blocks (Volkmann and Hodges 2021), and the relative abundances of both marten (Volkmann 2021) and preferred prey (Kelly and Hodges 2020; Kelly 2021) were markedly lower in salvage-logged areas. Given that marten avoid meadows (Spencer et al. 1983) and conventional clear-cuts (Soutiere 1979), avoidance of post-fire salvage-logged areas is unsurprising.

Wildfire and post-fire salvage logging are not equivalent disturbances for marten, and create different landscapes (Fig. 5). Salvage logging homogenises burned forests, removing critical resources for marten and prey (Kelly and Hodges 2020; Kelly 2021). Loss of structures suitable for foraging, resting, and denning will likely hinder marten persistence on salvage-logged landscapes. Female marten are particularly intolerant of disturbance (Cheveau et al. 2013; Slauson et al. 2017) and require specific structures for reproduction (Wynne and Sherburne 1984; Ruggiero et al. 1998). For marten, post-fire salvage logging has secondary impacts more severe than high-severity wildfire.

Fig. 5.  Winter conditions on post-fire landscapes in north-central Washington, USA (2006 burn), and central British Columbia, Canada (2010 and 2017 burns), showing typical habitats available to marten. Images are from December 2018 to March 2019. Rows show (a) unburned forests in each study area, (b) areas burned at low severity, (c) areas burned at high severity. Panels (d₁) and (d₂) are natural meadows and (e) is a salvage-logged area of the 2010 burn. Marten on these landscapes selected areas with closed canopies and abundant residual trees. Photo credit: L. Volkmann.

Landscape alterations from salvage logging threaten marten and other forest specialists. Salvage-blocks support fewer snowshoe hares (Thomas et al. 2019; Kelly and Hodges 2020), an important prey for many carnivores including fishers (Powell 1979; Raine 1987). Like marten, female fishers require large trees and snags for denning (Weir et al. 2012). Individual snags can persist for >80 years post-fire (Chambers and Mast 2005; Russell et al. 2006), particularly in severely-burned areas (Angers et al. 2011), and provide both structural complexity and a long-term source of deadfall on regenerating landscapes. Sparse, evenly-spaced snags left by salvage logging are less suitable to black-backed woodpeckers (Picoides arcticus) compared to undisturbed stands (Hutto and Gallo 2006; Saab et al. 2009), prompting specific snag-retention guidelines for this species (Nappi and Drapeau 2009, 2011). Our results underscore the need for clearer retention guidelines to support other species on salvage-logged landscapes (Thorn et al. 2020).

Climate change is promoting larger, more severe fires in North American forests (McKenzie et al. 2004; Girardin and Mudelsee 2008). Many fire regimes are also shifting towards more frequent disturbance (Kasischke and Turetsky 2006; Westerling et al. 2006; Dennison et al. 2014), which may permanently alter forested landscapes (Adams 2013; Busby et al. 2020). Biodiversity losses from salvage logging are occurring in Australian forests (Lindenmayer et al. 2008; Blair et al. 2016), but fewer data exist for North American wildlife despite mounting pressure to harvest in burns. Amid rapid landscape change, it is critical to identify post-fire habitat features that wildlife need (Hutchen et al. 2017; Volkmann et al. 2020) and reassess forest management strategies accordingly (Hutto 2006; Millar et al. 2007; Lindenmayer et al. 2014).

For marten, we recommend that managers prioritise the retention of residual live trees, corresponding to areas burned at low to moderate severity. Future work should determine how snags and deadfall influence habitat quality and landscape connectivity in severely-burned areas (Seip et al. 2018; Sullivan and Sullivan 2021), and the extent to which debris piles improve salvage-logged areas for marten and prey (Wilk and Raphael 2018; Sullivan et al. 2021). Because post-fire landscapes are diverse and complex, we call on researchers to study a broader range of wildfire ages, severities, and sizes (Volkmann et al. 2020). We also recommend linking post-fire habitat selection with prey abundance (Coffin et al. 1997). More data on wildfire and salvage logging will improve our ability to maintain suitable landscapes for wildlife.

The data that support this study will be shared upon reasonable request to the corresponding author.

The authors declare no conflicts of interest.

This project was supported by grants from the Habitat Conservation Trust Foundation and the Forest Enhancement Society of British Columbia (Project #8-400), the Skagit Wildlife Research Grant Program (City Light Grant #2016-01), the Natural Sciences and Engineering Research Council of Canada (Discovery Grant #31222), and by fellowships from Killam Trusts and the University of British Columbia.

Supplementary material is available online.