Sky islands, ecosystems found on geographically isolated mountain peaks, are among the most biodiverse ecosystems in the world but face a disproportionately high threat from climate change. High‐elevation, montane ecosystems, which are already at their upper altitudinal limits, are predicted to severely contract in response to climate change. The identification and conservation of refugia is an increasingly important approach for protecting biodiversity associated with imperiled ecosystems. We explored the spruce‐fir–northern hardwood ecotone as a possible refugium for mammals in the Southern Appalachian red spruce (Picea rubens)‐Fraser fir (Abies fraseri) sky islands. We conducted livetrapping, camera trapping, and ultrasonic acoustic surveys to characterize mammal diversity across the spruce‐fir–northern hardwood forest gradient on Grandfather Mountain and Roan Mountain Highlands in western North Carolina, USA. We detected four out of the five spruce‐fir‐associated small mammal species in both spruce‐fir and ecotone habitats. Mammal species richness, alpha diversity, and bat activity tended to be higher in the ecotone than in the other forest types on both mountains. Next, the abundance of small mammals associated with spruce‐fir was higher in the spruce‐fir and ecotone forests for one of the three species we were able to estimate. Together, our results suggest that the spruce‐fir–northern hardwood ecotone might serve as refugium for mammal species that are associated with spruce‐fir sky islands in the Southern Appalachian Mountains and mammalian conservation efforts in this biodiversity hotspot should consider focusing on the ecotone in addition to the adjacent spruce‐fir ecosystem.

Study Area

This study was conducted in the Southern Appalachian Mountains, in western North Carolina, USA. We surveyed mammal communities on two mountains: Grandfather Mountain (36°6′ N and 81°48′ W; 1812 m above sea level [asl]), and Roan Mountain Highlands (36°6′ N and 82°7′ W; 1916 m asl). On each mountain, we conducted surveys at three sites along an elevational gradient, each located within one of the following forest types: spruce‐fir, spruce‐fir–northern hardwood ecotone, and northern hardwood. We surveyed six sites in total.

Mammal Surveys

We surveyed mammals along an elevational gradient at each mountain, from 1324 to 1635 m asl on Grandfather Mountain and 1547 to 1872 m asl in the Roan Mountain Highlands (Table S1 in Appendix S1). The peaks of Grandfather and Roan Mountains were approximately 27 km straight‐line distance from one another. We used three field survey techniques to document volant and non‐volant mammal species within each forest type on each mountain: live traps, remote camera traps, and ultrasonic acoustic recorders (Figure 2). All surveys occurred in June and July 2023 on the southern aspect of both mountains (Table S1 in Appendix S1). We surveyed a total of six sites (three forest types on each mountain) that were spaced by 0.6–1.9 km. Our protocol conformed to the guidelines outlined by the American Society of Mammalogists (Sikes et al. 2016) and was approved by the Appalachian State University Institutional Animal Care and Use Committee (permit #22‐12) with permissions from relevant management authorities.

Livetrapping

We live‐trapped non‐volant small mammal species using Sherman traps (3″ × 3.5″ × 9″; H.B. Sherman Traps, Tallahassee, FL, USA). Each survey site consisted of 36 traps arranged in a 6 × 6 grid with 20 m spacing between traps (Figure 2). We set traps at sunset and checked and closed them at sunrise. Each trapping session consisted of four consecutive trap nights for a total of 144 trap nights per each 100 m² survey site. We baited traps with sunflower seeds, oats, and mealworms, and placed a cotton ball into each trap to help animals thermoregulate during colder nights. We placed traps at grid intersections in proximity to woody debris, trees, or rocks (Converse et al. 2006).

For each captured rodent, we measured weight, hind foot length, body length, and tail length for species identification (Stephens et al. 2014; Berl et al. 2017), and determined sex, age class (adult or juvenile), and external reproductive status (Steele and Powell 1999; Polyakov et al. 2021). We uniquely marked each adult rodent in the right ear with an ear tag (Style 1005‐1L1, National Band & Tag Company, Newport, KY, USA; McCain 2004) and used surgical scissors to take ear tissue samples from Peromyscus for molecular species assignment (see “Molecular species assignment and discriminant function analysis” section below). Shrews were weighed, sexed, and identified to species. We used a permanent ink pen to mark juvenile rodents and all shrews with a unique symbol on the stomach or tail. All animals were released at their capture location.

Molecular Species Assignment and Discriminant Function Analysis

Morphology‐based identification is often unreliable for the two Peromyscus species in the region, the cloudland deer mouse (P. maniculatus nubiterrae) and the white‐footed mouse (P. leucopus) (Choate 1973; Stephens et al. 2014; Berl et al. 2017). We extracted DNA from the collected Peromyscus ear tissue samples via Qiagen DNeasy blood and tissue kits (Qiagen, Germantown, MD, USA). A random subset of the DNA extracts (n = 48) was analyzed at the North Carolina Museum of Natural Sciences for identification to species using the cytochrome b mitochondrial gene (primers: MTCB‐F and R; Naidu et al. 2012).

For DNA sequencing, polymerase chain reactions (PCR) occurred in 10 μL volumes consisting of 5–10 ng of extracted DNA, 10 nM of forward and reverse primer, 2.0 mM of MgCl₂, 5.0 μL of 1:5 mix of Takara Taq Polymerase and Promega GoTaq MasterMix, and 2.9 μL nuclease‐free water. PCRs began with initiation at 95°C for 2 min, then 35 cycles of denaturing at 95°C for 15 s, annealing at 52°C for 30 s, and extension at 72°C for 1 min with a final 10‐min extension at 72°C. We sequenced the forward primer for cytochrome b as it yields sufficient power to discern Peromyscus species (~600 bp of cytochrome b). Sequencing reactions initiated at 96°C for 3 min, then underwent 30 cycles at 96°C for 10 s, 50°C for 5 s, and 60°C for 2.5 min. We used the ethanol precipitation method to clean sequencing reactions (Latch and Rhodes 2005) and then sequenced them on a 3500 Genetic Analyzer (Applied Biosystems, Foster City, CA, USA). We aligned all sequences in Geneious Prime (Kearse et al. 2012) and uploaded them to GenBank (Accession Numbers PV487868–PV487919).

Finally, to identify the remaining live‐trapped Peromyscus individuals to species, we used a linear discriminant function analysis model in the “mda” 0.5‐4 package (Hastie et al. 2023) in R 4.3.2 (R Core Team 2023). The model function was built using four body measurements (weight, hind foot length, body length, and tail length). Analysis of variance tests confirmed that each body measurement differed significantly between the two species (p < 0.01). The model correctly assigned species identity for 93.9% of the 48 individuals previously identified by DNA analysis, and we used it to predict the species of the remaining captured individuals (n = 100).

Remote Camera Trapping

We used remote camera traps to detect small to large non‐volant mammal species. We installed four remote cameras (Spec Ops Elite HP5, Browning, Morgan, UT, USA) at each survey site within each forest type so that cameras were at a minimum of 150–200 m apart (Figure 2). We mounted each camera at 30–50 cm on the bole of a tree following standard methodology to survey for small to large mammals (Evans and Mortelliti 2022; Rooney et al. 2025) near animal trails, rock outcrops, logs, and other desirable habitat features to increase the probability of detection (Trolle et al. 2008; Gebert et al. 2019; Hofmeester et al. 2021). We set cameras to take five‐image photo bursts every 5 s when motion was detected. We considered all images taken of the same species at the same camera within an hour as one detection (Hegerl et al. 2017; Gebert et al. 2019). We set the cameras at each site for 31 days, for a total of 372 trap nights per mountain. We uploaded, sorted, and manually identified species in all camera images on Wildlife Insights (https://www.wildlifeinsights.org/; Ahumada et al. 2020).

Ultrasonic Acoustic Monitoring

We used ultrasonic acoustic detectors (SM4, Wildlife Acoustics Inc., Maynard, MA, USA) to record vocalizations and quantify bat activity and species richness, and the activity of American flying squirrels (Carolina northern flying squirrel and southern flying squirrel; Diggins et al. 2016, 2020) across the forest types. We installed two acoustic detectors at each trap grid in the same location as the camera traps (Figure 2). To detect bats and American flying squirrels while minimizing the detection of nontarget species, we set detectors to record sounds with a minimum frequency of 16 kHz and minimum length of 1.5 ms from 30 min before sunset to 30 min after sunrise for 10 days (Gilley et al. 2019; U.S. Fish and Wildlife Service 2024). We attached detectors to tree trunks 1.5 m off the ground, facing the direction with the least amount of clutter to improve recording quality (Diggins et al. 2016, 2020). We conducted acoustic recording for a total of 10 consecutive nights per survey site.

We used SonoBat software (SonoBat 4.4.5, DND Design, Arcata, CA, USA) to sort and identify flying squirrel calls. We confirmed all species identifications using two observers. A total of 845 calls were identified as American flying squirrels. We were able to identify four flying squirrel call types: chirps, trills, upsweeps, and crows (Gilley et al. 2019). We identified bat passes to species using Kaleidoscope Pro Analysis Software (Kaleidoscope 4.4, Wildlife Acoustics Inc., Maynard, MA, USA). Each bat pass had to match a species identification accuracy of at least 60% or the pass was classified as a “no ID” (Li and Kalcounis‐Rueppell 2018; Schimpp et al. 2018; Parker et al. 2019). We estimated bat species activity at each site as the mean number of bat passes recorded per night over the 10‐night survey period. We identified a total of 6436 bat passes to species.

Vegetation Surveys

We conducted vegetation surveys at every survey site to characterize each forest type. We established a circular 400 m² plot centered at each live‐trap grid. In every plot, all live trees (≥ 1.4 m height) were identified to species and measured at diameter at breast height. In addition, we counted the number of tree snags. Within each 400 m² plot, we established a 40 m² subplot where all live saplings (< 1.4 m height) were identified to species and dead saplings counted (Kalies et al. 2012). We used the line interception method (Canfield 1941) along four perpendicular transects within the subplot to identify shrubs and herbaceous plant species. We calculated plant species richness as the total number of species detected at each survey site, including tree, shrub, and herbaceous species. We also calculated the percentage of trees that consisted of Fraser fir, red spruce, and northern hardwood species. Finally, we measured slope and aspect at the center of each plot.