Study species
The bobcat is a wide-ranging carnivore in North America, occupies relatively large home ranges, and exhibits low reproductive rates, territoriality, and male-biased dispersal (Knowles 1985; Hansen 2007). Bobcats move long distances within their home ranges and travel mostly by night, with males traveling farther and faster than females (Chamberlain et al. 2003). Bobcats use a variety of land cover types for travel, including stream valleys and associated ridgelines (Woolf and Nielsen 2002), thick understories (Litvaitis et al. 1986), and rocky ledges (Anderson 1990; Hansen 2007). Despite tolerance of habitat fragmentation, it has been suggested that maintaining corridors for travel between habitat fragments is necessary for bobcat population persistence (Tigas et al. 2002; Litvaitis et al. 2015). Although habitat use in this species has been well-documented, habitat selection by individuals as they move and how landscape context affects habitat selection are topics that have received little attention.
Field methods
To obtain data on the trajectories of individuals through space and time, 41 bobcats were captured and GPS-collared between 2005 and 2007 in areas representing the Champlain Valley lowlands and higher elevation regions in the Green Mountains of Vermont, USA (University of Vermont IACUC 05-036). Bobcat home ranges were distributed in Chittenden, Washington, Lamoille and Addison counties. This large study area was comprised of the following land cover types based on National Land Cover Database (NLCD, Homer et al. 2007) categories: forest (62.4%), agriculture (24.5%), scrub and shrub/rock (5.4%), wetland (4.2%), and development (3.5%). These five land cover categories were created by combining relatively similar land cover classes from the original NLCD dataset, which is comprised of 16 land cover classes. For example, the forest category used in this analysis was created by combining the deciduous forest, evergreen forest and mixed forest categories delineated in the original NLCD dataset. Within forested land cover, 10.0% was forest edge and 22.6% was deep forest core (300 m from a forest edge). The remaining forested land cover was comprised of superficial forest core (30–300 m from a forest edge). Forest edge was defined as the transition between two habitat types by the scale of the data available. The NLCD data set is available in 30 m resolution, allowing the delineation of core and edge habitat in 30 m increments. Forest edge is defined as the interface between forest habitat and a different ecosystem type (Harper et al. 2005), and has long been perceived to have higher abundance and diversity of game species (Leopold 1933), including bobcat prey. In contrast, forest interior is defined as habitat that show no detectable edge influence, defined by alterations in biophysical processes and ecosystem composition and structure. Harper et al. (2005) suggest that edge effects on primary forest processes (e.g. tree mortality) and structure can extend up to ~ 300 m. The topographic landscape contained ridges (5.3%), slopes (56.9%), flat areas (31.3%) and valleys (6.5%). The average slope was 8.3˚.
Due to the seasonal variation in environmental conditions and bobcat behavior, two capture methods were used (Donovan et al. 2011, University of Vermont Animal Care and Use Protocol 05-036). During warmer months, bobcats were captured using padded foot-hold traps (Victor Soft Catch no. 3, Woodstream Corporation, Lititz, Pennsylvania, USA) and sedated via pole injection. During cooler months, as well as in areas with a high probability of human presence, bobcats were captured using cage traps (91 × 28 × 30 cm wire mesh, Safeguard, New Holland, Pennsylvania, USA) that were baited with meat. Traps were distributed widely across the study area in locations that would increase the likelihood of capturing the species of interest. All captured individuals were anesthetized according to body weight with ketamine and xylazine at a ratio of 5 to 1. Both foot-hold traps and cage traps were checked once every 24-h (Donovan et al. 2011).
Bobcats were then outfitted with a GPS collar (model 3300S, Lotek Wireless, Newmarket, Ontario, Canada or model G2400, Advanced Telemetry Systems, Isanti, Minnesota, USA). To avoid effects that might negatively alter natural survival, movement and reproduction, collars did not exceed 2% of the total body weight of a captured individual (Withey and Boloxton 2001). Collars collected temporal data and spatial data in the form of x,y location points. All data were stored on-board the GPS collars.
Bobcats are most active during dawn, dusk and evening hours (Anderson and Lovallo 2003). GPS collars recorded location data every 20 min during these periods (1600–930 h) on alternate days. During off days and all diurnal periods, location fixes were attempted every 5 h in a 24-h period for home range analyses. Data were collected at 20-min intervals periodically during daytime hours. Collars were programmed to actively collect data for 130 days. When this time period expired, collars self-released and were retrieved by technicians (Donovan et al. 2011).
The data were screened such that only those trajectories containing consecutive location points less than 20–25 min apart were used in movement analysis. Using a relatively short time interval reduced the risk that trajectories recreated from discrete point data differed from the true path of the individual. This strict screening resulted in 13 collared bobcats suitable for movement analysis (Table 1).
Table 1 Summary of movement path data for bobcats tracked in the Champlain Valley of Vermont, USA
Movement metrics analysis (objective 1)
We used the x,y location data in the Universal Transverse Mercator projected coordinated system (Zone 18 N, North American Datum 1983) to generate a point feature class for each individual bobcat in a GIS (Geographic Information System) environment (ArcGIS 9.3, ESRI, Redlands, California, USA). We used the Hawth’s Tools analysis package (Beyer 2004) to create continuous movement paths from the input point data. The distance between one point to the next created a movement segment. An example of a single movement path consisting of multiple segments is shown in Fig. 1; each bobcat had multiple movement paths.
Hawth’s Tools movement tools were used to calculate movement metrics, including step length and speed of travel. Step length was the Euclidean distance from the current point to the next point (Fig. 1). As points were, on average, 20 min apart, step length was the distance (m) traveled in a 20-min period. We calculated average step length and average speed for each bobcat. Then, we used two sample t-tests with unequal variances (α = 0.1) to analyze the effect of sex and time of day on step length (the distance traveled in a 20-min period) and speed of travel (meters/minute), where the individual was the sample unit. We used a paired t test to test for differences in mean step length and speed by time of day, where day and night designations were determined by using sunrise and sunset times for each 24-h period.
Resource selection via compositional analysis (objective 2)
To determine habitat selection and preference as it related to movement paths, we compared the habitat compositions of “used” movement space to those of “available” movement space. Used movement space was defined as the space through which an animal chose to travel from one point to the next. Available movement space was defined as the space through which an animal could have chosen to travel (see Used Space Buffer and Available Space Buffer in Fig. 1). The available movement space included used movement space.
In this analysis, recreated trajectories, or the movement paths, were encapsulated by a 60 m buffer to provide an estimate of used movement space. We chose this buffer size based on the activity sensor embedded within GPS collars, and because this buffer liberally encompassed GPS error and uncertainty of the exact movement path between discrete points (Dickson et al. 2005). Only movement segments > 60 m were used in this analysis. Location data were buffered by the Euclidean distance traveled between consecutive points to create “available” movement space for each observation (Fig. 1). This method assumes that individuals make choices based upon the space it currently occupies (Aebischer et al. 1993).
We used Python 2.4.1 (Van Rossum 2005) in conjunction with ArcGIS 9.3 to extract statistics on land cover proportions (agriculture, forest, development, scrub-rock, wetlands), vegetation (forest edge, wetland edge, and deep forest core), topography (ridge, slope, flat, valley), stream density, degree slope, and road density within used and available space. The original NLCD (Homer et al. 2007) was comprised of 16 land cover classes in the state of Vermont. Based on the land cover classes believed to be of significance to bobcats (Berg 1979; Sunquist and Sunquist 2002; Hansen 2007; Donovan et al. 2011), and to facilitate clear land cover comparisons (Dickson et al. 2005), we reclassified the original NLCD 2001 into the following five categories: forest, wetlands, developed, agriculture, scrub-rock. By combining classes in this way, we were able to maximize our sample size because individuals that did not have a habitat type available were eliminated from the analysis as they provided no information on use or availability (Aebischer et al. 1993).
We used compositional analysis to test for habitat preference among the five different land cover types (Aebischer et al. 1993). In this movement study, the composition of land cover types in a given space, be it used or available, summed to 1.0. However, being a proportion, the proportion of landcover type 1 (x1) is dependent on the proportion of landcover type 2 (x2). Additionally, proportions make it difficult to determine individual preference and avoidance, as the total habitat is constrained such that a preference for landcover type “1” will automatically lead to an avoidance of landcover type “2.” Aebischer et al. (1993) solved this challenge by using log-ratio transformations, which ensures linear independence of the proportions and allows the use of multivariate compositional analysis to determine preference for one landcover type over another (Aitchison 1986; Aebischer et al. 1993; Zar 2010).
Following Aebischer et al. (1993), the log-ratios of the average used space (xi,u) to available space (xi,a) for each land cover type were computed for each bobcat. Then, a “difference matrix” was created for each landcover type separately, in which the difference between the log-ratios of each landcover type was compared to the reference type on a bobcat-by-bobcat basis (Aebischer et al. 1993). For example, with ‘forest’ as a reference type, the difference (d) between the log-ratios of forest and the log-ratios of every other landtype (agriculture, development, wetland, shrub-scrub) were computed for each bobcat. In each of the 5 d matrices, positive d scores indicated preference, and negative scores indicated avoidance for a given land cover type relative to other land cover types. We used a MANOVA on the forest-reference d score matrix to test the omnibus null hypothesis of no difference in d scores between the alternative land types. Additionally, the scores for each d matrix were then averaged across individuals in the population, and were used to rank land cover types from most preferred to least preferred. Due to the large variance in the locations contributed by each bobcat, it was necessary to implement a weighting structure to account for the variation in available data per individual.
Similarly, we used compositional analysis to evaluate the effects of topography on movement. Elevation data were extracted from a 1:24,000-scale Digital Elevation Model (DEM) for the Vermont portion of the United States Geological Survey National Elevation Dataset (raster with 30 m × 30 m pixel size). We created a topographic position index map, which categorized the state of Vermont into valley, flat, sloped and ridge topographies based on change in elevation. This resulted in 4 categories whose proportions summed to 1 within a used or available space. Using this topographic position index, we performed a compositional analysis to test for preference as described for the landcover analysis.
Resource selection via paired tests (objective 3)
In addition to the compositional analysis of landcover types and topographic position, we also tested for selection in the form of slope, forest edge, forest core, and wetland edge vegetation (variables that are not proportions). Average slope encountered for used and available space was calculated using ArcGIS 9.3. Edge habitat existed where two habitats of contrasting composition met in the landscape. Forest edge (30 m) and wetland edge (30 m) consisted of forested and wetland habitats within 30 m of a contrasting habitat type. Core habitats were those habitats that were surrounded by habitat of like composition. Forest core (30 m) represented superficial core forest that was 30 m or more in distance from a habitat of contrasting type. Forest core (300 m), a measure of deep core, represented forest that was 300 m or more in distance away from a habitat of contrasting type. These edge and core categorizations were imposed by the use of a 30 m pixel landcover raster. We used the svyratio function in the R package, survey (Lumley 2004) to determine if the ratio of used to available significantly differed from 1 for slope, core habitats, and edge habitats, where the statistical sample was an individual’s average used and available score. Ratios > 1 indicate resource selection, and ratios < 1 indicate non-selection.
The effects of linear features, including streams and roads, on bobcat movement were similarly evaluated. The most accurate road data available for Vermont was the Emergency 911 roads centerline feature class data (McMullen 2008). This GIS layer was created from 1:5000 scale orthophotos. Both paved and dirt road densities were calculated using ArcGIS 9.3 for available and used movement space. Stream density was calculated and used as an indicator of riparian zones, which are thought to be of importance to bobcats in Vermont (Hansen 2007). We used the svyratio function in the R package, survey (Lumley 2004) to determine if the used to available ratio for road density and stream density differed from 1, where the average used and available scores for each individual constituted a data point for statistical analysis.
Effects of landscape composition on travel speed (objective 4)
We assessed travel speed as a function of proportion of used forest and as a function of used scrub-rock landcover types, with the expectation that speed would decline as these proportions increased if bobcats preferred these habitats. However, because speed through these habitats may be shaped by the landscape composition of the “available” habitat, our analysis included the main effect of used habitat (proportion of forest or scrub-rock), the main effect of available habitat (proportion of forest, scrub-rock, agriculture, and development), and the interaction between used and available habitat proportions. As such, we evaluated eight GEE models with an autoregressive correlation structure, and used QIC to rank models. The general model framework was speed = usedi + available habitatj + usedi * availablej, where i was either the proportion of used forest or the proportion of used scrub rock, and j was the proportion of available forest, scrub-rock, development, or agriculture. The model included a repeated measurement by individuals to account for multiple observations per individual.
Comparison of corridor designs (objective 5)
We used least cost path analysis (Adrianensen et al. 2003) and circuit theory analysis (McRae et al. 2008) to map the potential corridors between two state management areas in a region within the bobcat study area (Little Otter Creek Wildlife Management Area and Huntington Gap Wildlife Management Area; focal region = ~ 34 × 16 km). We chose these areas because bobcats have been recorded in both and the region between them was heterogeneous and contained several landcover types. Both corridor design approaches require a cost map (the inverse of habitat suitability) which identifies each pixel’s resistance to animal movement. We evaluated four corridor designs based on four alternative habitat suitability maps. For the first map alternative (‘Resource Selection Function’), we used a habitat suitability map inferred from analysis of all GPS locations, regardless of behavioral state (moving or non-moving). The suitability score for each pixel was obtained by calculating the proportion of scrub, deciduous forest, mixed forest, evergreen forest, wetlands, and road density within 1 km of each pixel, multiplying each factor by its corresponding resource selection linear coefficient (see Donovan et al. 2011, Table 5), and then summing the scores; scores were positively related to habitat suitability. This suitability map was based on habitat use with no regard for habitat preference.
For the second suitability map alternative (‘Compositional Analysis’), we created a suitability map based on the habitat preferences of moving bobcats revealed by compositional analysis in Objective 2 (Aebischer et al. 1993). In this approach, the suitability score for each pixel was obtained by calculating the proportion of agriculture, forest, development, scrub-rock, and wetlands within a 161 m radius of each pixel (the average step distance for a moving bobcat; see “Results”), multiplied by their corresponding average d score, and then summed; scores were positively related to habitat suitability. This suitability map was based on habitat preferences by moving animals (highly weighted forested pixels and strongly penalized developed pixels; see Results), but was restricted to the subset of landcover variables analyzed in the compositional analysis.
The third suitability map alternative (‘Use-to-Available Ratios’) was based on the used versus available ratio results from Objective 3. In this approach, we used the used-to-available ratios to determine which variables were selected (ratio significantly greater than 1) and which were non-selected (ratio significantly less than 1). For each pixel, we used a moving window analysis to calculate the average score of each variable within a 161 m radius. We then determined the minimum (min) and maximum (max) value for each variable in the focal region. For each pixel in the focal region, each variable was scaled from 0 to 1: variables that were selected were scaled with the equation scaled score = (score − min)/(max − min), whereas variables that were non-selected were scaled with the equation scaled scored = (score − max)/(min − max). Consequently, a variable score of 1 represented the “ideal” condition for a moving bobcat. For example, if forest cover was preferred, a score of 1 indicated the highest level of forest cover in the focal region. In contrast, if roads were not preferred, a score of 1 indicated the lowest level of road density in the focal region. The total suitability of each pixel was then calculated by summing the scaled scores. The resulting suitability map was based on habitats selected by moving animals and based on multiple variables; however, each variable contributed equally to the resulting map.
The final suitability map (‘Weighted Used-to-Available Ratios’) incorporated a slight modification to the third method, in which the scaled scores were weighted by how far the use:availability ratios deviated from 1.0. For example, variables with a ratios of 0.8 or 1.2 would be weighted by 1.2, while variables with ratios of 0.9 or 1.1 would be weighted by 1.1, and variables with ratios of 1 would be unweighted. This suitability map was based on habitat preferences by moving animals and based on multiple variables, with each variable weighted based on the preference of moving bobcats.
For each of the four alternative suitability inputs, we evaluated connectivity between the two management areas with a three-step process. The first step involved converting the habitat suitability map to a cost map between the two areas. For each design, we first converted the habitat suitability raster to a cost raster by dividing the raster by its maximum value, and then subtracted the result from 1.0. Values were then rescaled from 1 (pixel with minimal cost) to 100 (pixel with maximal cost). Thus, pixels with a low score had low travel cost.
The second step involved creating a cost-weighted distance map and least-cost corridor map between the two areas. In a cost-weighted distance map, the value of a pixel is function of its resistance value and distance to a core area. We created a cost-weighted distance map for each management area. We then created a least-cost corridor map by adding, normalizing, and mosaicking the cost-distance maps, and truncated this final map by removing all cost-distance values > 25 km. We chose this value based on observations of bobcats and information on movement and dispersal elsewhere. The remaining pixels represented corridors where bobcats could be expected to travel between management areas. We used Linkage Mapper v. 1.1.0 (McRae and Kavanagh 2011) to create cost-weighted distance and least-cost corridor maps, and calculate least-cost paths between areas.
The third step involved estimating movement flow through the least-cost corridor map. We estimated flow using a circuit theory approach, which treats movement like the flow electrical current through a circuit with varying resistances (McRae et al. 2008). We used Pinchpoint Mapper (McRae 2012) to pass current from one management area to another through the least-cost corridor map, which consisted of pixels with varying resistances (or cost-weighted distance values). The resulting map showed the distribution of movement flow within the corridor.