2.1. Study Area
Our study took place on a 1242 hectare tract of upland oak-hickory forest located on the surface of the Cumberland Plateau on the campus of The University of the South (UoS) in Sewanee, TN, USA. The upland topography gently varies from convex to concave surfaces with only subtle changes in vegetation that map these gradients across the landscape. The study site is bounded on three sides by a steep bluff transition to a cove community perforated in places by stream drainages. On the fourth side, the forest rapidly transitions to agricultural and residential land use. Elevations of the study area vary from 580–625 m above mean sea level [
33].
The forest canopy is generally dominated by oaks (
Quercus alba L.,
Q. coccinea Muenchh.,
Q. montana Willd., and
Q. velutina Lam.) and hickories (
Carya glabra Miller,
C. pallida (Ashe) Engl. & Graebn., and
C. tomentosa (Lam.) Nutt.) along with
Acer rubrum L.,
Nyssa sylvatica Marshall,
Oxydendrum arboreum (L.) DC., and
Prunus serotina Ehrh. with a distinct understory layer of shrubs and small trees including:
Sassafras albidum L.,
Vaccinium corymbosum L.,
Vaccinium stamineum L., and
Viburnum acerifolium L. White-tailed deer (
Odocoileus virginianus) were largely eliminated from the Sewanee area due to severe hunting pressure during the 1930s, but were reintroduced in the 1950s, along with a predator trapping program which targeted bobcats and feral dogs [
34]. By the late 1990s, the decline in sapling densities within forested areas of the UoS was attributed to deer herbivory [
35]. Deer culls have been performed in selected areas of various sizes within the study area since 2010 for the purpose of management of deer populations in the vicinity of agricultural and residential land use.
2.2. Field Methods
To examine sapling density across the landscape, we established 46 stratified random 20 × 50 m box transects between June and August 2012. Within each transect, there were 10 randomly distributed circular plots (5 m diameter). Within each circular plot, saplings (0.25–3 m height) were recorded by species and counted. The number of saplings per circular plot was added together, so a single value that represented the total number of saplings was generated for each transect. Plot locations were revisited in July 2015, and tree saplings were resampled within a box transect with the same area as the initial survey. The entire 20 × 50 m box transect was surveyed for sapling species and number. While sampling methodology differed between the two years, the area sampled and plot location remained the same.
Four fenced exclosures within the study area, established by the UoS prior to the study, were used to assess the plant community condition in the absence of deer. Two 16 m
2 exclosures were established in 1996, and two 24 m
2 exclosures were established in 1998 (
Figure 1). Exclosures were fully sampled in 2012 and 2015 after excluding a 0.5 m buffer zone near the fence.
2.4. A Model of Sapling Density Based on Landscape Predictors
We evaluated several predictors of sapling density on the landscape that were grouped into three classes: edge predictors (7), deer cull predictors (3), and plateau accessibility predictors (4) (
Table 1).
Edge predictors included distance (m) of a sampling point from the nearest edge in each forest edge category. We defined forest edges as belonging to one of the following categories: residential development, road, cleared road, recent (present in 2012; there was no change in the landscape between 2012 and 2015) clearcut, livestock pasture/agriculture, and lake. The nearest edge predictor denotes the distance (m) to the nearest type of edge in these categories.
Deer cull predictors were generated using deer cull area locations and the number of deer culled (per km2) within each area. In 2014, 14 plots fell within deer cull areas, and 31 were not within deer cull areas. For all plots outside of the deer cull areas, we assumed that the deer cull was equivalent to zero deer. Deer cull area predictors include the number of deer culled per km2 in 2014, 5-year deer cull, and 3-year deer cull.
Plateau accessibility predictors related to topography include latitude and derived measures which describe the steepness and inaccessibility of movement from the adjacent coves to the top of the Cumberland Plateau. Using the digital elevation model (DEM), we derived potential plateau access zones by determining the minimum elevation around each individual pixel (10 m cell size) along the bluff edge within a 50 m radius and then subtracting this minimum from the actual elevation of the pixel. This result gave a measure of potential plateau access locations along the bluff edge which had low differences between the neighborhood minimum and the local elevation. If the range of difference between the bluff highs and local lows in elevation were 4 m or greater, then the location was defined as a plateau access zone, and its width was evaluated. To characterize the number of nearby access zones in relation to sampling points, we identified the closest bluff edge to each sampling point and evaluated the number and average width of potential access zones 500 m of either side of the closest bluff edge. We evaluated the following predictors of plateau accessibility: distance to plateau access zone, number of nearby (within 500 m of bluff edge) plateau access zones, and average nearby (within 500 m of bluff edge) plateau access zone width.
We used an information theoretic approach [
38] to evaluate the association of each of these parameters to our observed landscape patterns of sapling density. To avoid multicollinearity, we evaluated correlation using Pearson’s Correlation Coefficient and
p-values to determine the likelihood of correlation. If the
p-value was 0.05 or less, we did not include predictors within the same candidate model. Using a global model with uncorrelated predictors representing each major category, we evaluated model fit assuming an underlying negative binomial distribution using a chi-square test. We chose the following predictors for the global model that were highly correlated with the other predictors in their group: latitude (correlated with number of nearby plateau access zones (
p-value < 0.0001), average nearby plateau access zone width (
p-value = 0.0007)), distance to nearest edge (correlated with cleared road (
p-value < 0.0001) and clearcut (
p-value = 0.027)), 5-year deer cull (correlated with 3-year deer cull (
p-value < 0.0001) and average deer culled per km
2 in 2014 (
p-value < 0.0001)), and sampling year (
Table A1). We also evaluated if variation in sampling protocols in 2012 and 2015 altered model likelihood and found that the global model including time did not perform as well as the global model excluding time (ΔAIC = 1.6; RMSE = 1.057). To reflect different sampling methodologies, we included time as a random variable in all subsequent models. To separately examine the effect of treatment (exclosure or sampling plot) and year, we used the Kruskal-Wallis test [
39].
Models were fit in R using the MASS package [
40]. To determine the relative plausibility of each model given our data, we used Akaike’s Information Criteria (AIC) [
41] adjusted for small sample sizes using the small-sample bias adjustment (AICc) [
42]. Model weights were calculated for each candidate model and were ranked from most likely to least likely, based on their weight [
38]. Models were only included in the confidence set if they had Akaike weights greater than 10% of the best fitting model’s Akaike weight [
43]. Model fit was also evaluated using the root mean square error (RMSE). We performed a step-wise protocol by assessing the best-fitting predictor within each category (edge predictors, deer cull predictors, plateau accessibility predictors) before assessing the fit of the best predictors relative to one another. To evaluate the importance and influence of each predictive parameter, we calculated importance weights and scaled estimates to assess the biological relevance of each parameter.