You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/301623629

Estimating landscape resistance from habitat suitability: effects of data


source and nonlinearities

Article  in  Landscape Ecology · November 2016


DOI: 10.1007/s10980-016-0387-5

CITATIONS READS

79 1,091

3 authors, including:

Annika Keeley Paul Beier


Delta Stewardship Council Center for Large Landscape Conservation
17 PUBLICATIONS   571 CITATIONS    164 PUBLICATIONS   8,996 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Essential Geodiversity Variables View project

Conserving Nature's Stage View project

All content following this page was uploaded by Annika Keeley on 09 April 2018.

The user has requested enhancement of the downloaded file.


Landscape Ecol
DOI 10.1007/s10980-016-0387-5

RESEARCH ARTICLE

Estimating landscape resistance from habitat suitability:


effects of data source and nonlinearities
Annika T. H. Keeley . Paul Beier .
Jeffrey W. Gagnon

Received: 31 August 2015 / Accepted: 18 April 2016


Ó Springer Science+Business Media Dordrecht 2016

Abstract resistance maps on an independent set of observed


Context Conservation corridors must facilitate long- long-distance, prospecting movements.
distance dispersal movements to promote gene flow, Results A negative exponential function best
prevent inbreeding, and allow animals to shift ranges described the relationship between resistance values
with climate change. Least-cost models are used to and habitat suitability for desert bighorn sheep indi-
identify areas that support long-distance movement. cating long-distance movers readily travel through
These models rely on estimates of landscape resistance, moderately-suitable areas and avoid only the least
which are typically derived from habitat suitability. suitable habitat. For desert bighorn sheep, all three
Objectives We examine two key steps in estimating suitability estimates performed better than chance, and
resistance from habitat suitability: choosing a procedure to resource and step selection functions outperformed
estimate habitat suitability, and choosing a transformation expert opinion. For elk, all three suitability estimates
function to translate habitat suitability into resistance. performed the same as chance.
Methods We used linear and nonlinear functions to Conclusions When designing corridors to facilitate
convert three types of habitat suitability estimates long-distance movements of mobile animals, we
(from expert opinion, resource selection functions, and recommend transforming habitat suitability into resis-
step selection functions) into resistances for elk tance with a negative exponential function. Use of an
(Cervus canadensis) and desert bighorn sheep (Ovis exponential transformation means that larger fractions
canadensis nelsoni). We evaluated the resulting of the landscape offer low resistance, allowing greater
flexibility in where a corridor is located.

Keywords Cervus canadensis  Desert bighorn


Electronic supplementary material The online version of
sheep  Elk  Expert opinion  Exploratory movements 
this article (doi:10.1007/s10980-016-0387-5) contains supple- Least-cost models  Ovis canadensis  Resource
mentary material, which is available to authorized users. selection function  Step selection function  Wildlife
corridors
A. T. H. Keeley (&)  P. Beier
School of Forestry, Northern Arizona University, 200 E
Pine Knoll Drive, Flagstaff, AZ 86011, USA Introduction
e-mail: annika.keeley@yahoo.com
Wildlife corridors are intended to allow animals to
J. W. Gagnon
Arizona Game and Fish Department, 5000 W Carefree move through the landscape, maintain genetic diver-
Highway, Phoenix, AZ 85023, USA sity, and shift ranges with climate change (Hilty et al.

123
Landscape Ecol

2012). Corridors must meet all life-history needs for Adriaensen et al. 2007; Beier et al. 2008a). Alterna-
relatively sedentary species (corridor dwellers—Beier tively, landscape variables at sites used by an animal
et al. 2008a). But for relatively mobile species, are compared to equivalent sites that are either unused
including most large mammals, corridors are intended or potentially available (Manly et al. 2002) to estimate
to support migration and dispersal movements. Dis- habitat suitability. When the used sites are animal
persal is defined as movement of an animal from the point locations, resource selection functions (RSF) are
area where it is born (or, rarely, from a breeding home employed to model habitat suitability (e.g., Johnson
range) to another area where it joins, or attempts to et al. 2000). However, when used sites are movement
join, the local breeding population (Baguette et al. steps (straight lines connecting consecutive locations
2013). Dispersal movements, although rare, are the separated by no more than a few hours), step selection
most important movements underlying gene flow, functions (SSF) are used to model habitat suitability
demographic viability of metapopulations, recolo- (e.g., Fortin et al. 2005). We hypothesize that all three
nization, and range shifts (Trakhtenbrot et al. 2005). procedures yield habitat suitability estimates that can
Prospecting movements, during which animals inform useful least-cost models. We predict that
become familiar with the habitat surrounding the resistance derived from expert opinion is least reliable
home range, frequently precede dispersal movements because expert opinion is subjective, and the resis-
(Stamps and Krishnan 1999; Stamps 2001). Here, we tance derived from SSF is most reliable because SSF
assume that habitat use during long-distance, are based on the data set that contains the most
prospecting movements is similar to habitat use during information about the animals’ habitat use.
long-distance, dispersal movements. When resistance is estimated from habitat suitabil-
Least-cost models (Adriaensen et al. 2003) are ity, it is usually assumed to be a negative linear
commonly applied to identify wildlife corridors in a function of suitability (Walker and Craighead 1997;
conservation context (Beier et al. 2008a, 2011; Sawyer Hoctor et al. 2000; Ferreras 2001; Singleton et al.
et al. 2011). These models use estimates of landscape 2002; Hunter et al. 2003; Larkin et al. 2004; Wikra-
resistance (the degree to which a grid cell facilitates or manayake et al. 2004; Richards-Zawacki 2009; Pull-
impedes movement of the study organism; Spear et al. inger and Johnson 2010) such that resistance decreases
2010) to identify the path with the least accumulated at a constant rate as suitability increases, as in the
resistance between two locations. Thus resistance upper line in Fig. 1. Beier et al. (2008a) called
estimates are critical to least cost corridor models attention to the lack of empirical support for this
(Beier et al. 2008a; Sawyer et al. 2011; Zeller et al. crucial assumption, labelled it the ‘‘subjective trans-
2012). If estimates are wrong, then land-use planners, lation problem,’’ and argued that it was a crucial
transportation agencies, and other decision-makers uncertainty in all corridor modeling. It seems plausible
may be implementing poor corridor designs. Resis- that during long-distance dispersal or pre-dispersal
tance estimates are commonly derived from habitat prospecting movements, animals might readily move
suitability values because habitat suitability is easier to through moderately suitable areas such that resistance
study than landscape use during rare prospecting and increases only modestly as suitability decreases from
dispersal movements. Here, we examine two key steps its maximum value, and then increases dramatically at
in estimating resistance from habitat suitability: lower suitability values (lower curves in Fig. 1). Such
choosing a procedure to estimate habitat suitability, a relationship was reported by Trainor et al. (2013) for
and choosing a transformation function to transform red-cockaded woodpeckers (Picoides borealis) and by
habitat suitability into resistance. Mateo-Sánchez et al. (2015) for brown bears (Ursus
Habitat suitability is typically modeled as a func- arctos). We hypothesized that this negative exponen-
tion of land cover, elevation, and other attributes. To tial relationship applies to most animals.
estimate habitat suitability, expert opinion models are In this study, we test whether estimates of habitat
popular because they are less expensive and do not suitability can predict habitat use during long-dis-
require intensive field data collection (Walker and tance, prospecting movements. We compare resis-
Craighead 1997; Hoctor et al. 2000; Singleton et al. tance estimates from linear and nonlinear
2002; Hunter et al. 2003; Larkin et al. 2004; Wikra- transformations of three types of suitability estimates
manayake et al. 2004; Beier and Majka 2006; (from expert opinion, RSF, and SSF) in terms of their

123
Landscape Ecol

rugged mountainous terrain with steep talus slopes and


cliffs as well as dry washes among rolling hills.

Habitat data

To estimate habitat suitability, we first identified land


cover and geographic variables that are known to affect the
distribution of elk (Wallace and Krausman 1997; Boyce
et al. 2003; Wait and McNally 2004; Creel et al. 2005;
Mao et al. 2005; Coe et al. 2011; Beck et al. 2013) and
desert bighorn sheep (Turner et al. 2004; Sappington et al.
2007; Hoglander 2012). We reclassified vegetation types
from the Landfire Program (Landfire 2008) into broad
Fig. 1 Eight curves used to transform habitat suitability values land-cover variables (Tables S1, S2) to limit the number of
into resistance values. The curves are based on the transforma- variables. We obtained elevation and slope from 30-m
tion function R ¼ 100  99  ðð1  expðc  HÞÞ =
resolution digital elevation models (DEM, http://
ð1  expðcÞÞÞ, where R is resistance, H is suitability, and
the factor c determines the shape of the curves nationalmap.gov/elevation.html, last accessed August
21, 2015). We applied the DEM Surface Tools extension
(Jenness 2013) to generate aspect and slope rasters in
consistency with observed long-distance prospecting ArcGIS 10.2 (ESRI 2011). We transformed aspect, a cir-
movements for two species of large mobile mammals. cular variable, into two independent variables, northness
Our method of evaluating each resistance model (a (cos(aspect)) and eastness (sin(aspect)). We obtained road
combination of a particular suitability estimate and and highway maps from the topologically integrated
transformation function) assumes that during geographic encoding and referencing (TIGER) project of
prospecting animals choose paths of low cumulative the U.S. Census Bureau (http://www.census.gov/geo/
resistance (Adriaensen et al. 2003). If this assumption maps-data/data/tiger.html, last accessed August 21,
is true, long-distance paths will have a relatively low 2015). We calculated distances to roads or highways in
cost compared to random paths of identical topology ArcGIS 10.2 (ESRI 2011). All variables were resampled
in the same geographic area. to the North American geographic coordinate system from
1983 with a cell size of 30 9 30 m.
Following the commonly used threshold of r = 0.7
Methods (Booth et al. 1994; Dormann et al. 2013), correlations
among habitat covariates for elk were relatively low
Study area (maximum r = -0.57, Table S3), allowing us to
retain all predictor variables in our models. For
We studied elk on and near the Tonto, Coconino and bighorn sheep, we excluded the topographic position
South Kaibab National Forests in northern Arizona. index because it was strongly correlated with slope
Elevations range from 1100 to 3800 m, and vegetation (r = 0.72, Table S4).
types include pinyon (Pinus edulis)–juniper (Junipe-
rus spp.) woodlands and chaparral (Arctostaphylos Animal locations
spp.) at lower elevations, ponderosa pine (P. pon-
derosa) forest at medium elevations and mixed conifer The Arizona Game and Fish Department (AZGFD)
forest at upper elevations. radio collared elk from 2004 through 2012 with global
We studied desert bighorn sheep in the Black positioning system (GPS) collars (models TGW-3600,
Mountains in northwestern Arizona on land managed TGW-3690, TGW 4600; Telonics, Inc., Mesa, AZ;
by the Bureau of Land Management. Elevations range and NSG-LD2 Northstar Science and Technology,
from 150 m at the Colorado River to 1663 m. Mojave LLC, King George Virginia; Dodd et al. 2012a, b;
Desert scrub and creosote (Larrea tridentata) shrublands Gagnon et al. 2012, 2013). Each collar recorded the
are dominant vegetation types. Topography includes animal’s location every 2 h between 17:00 and 7:00 or

123
Landscape Ecol

between 16:00 and 10:00, or, for a few individuals, For testing resistance estimates from expert opin-
every 1.5 or 4 h throughout the day. GPS error was ion, RSF, and SSF in terms of their consistency with
±10 m (Dodd et al. 2006). To minimize heterogeneity observed long-distance prospecting movements, we
within models, we restricted our analyses to data of visually identified long-distance movements by indi-
yearling bull elk in April, May, and June, because vidual elk and bighorn sheep outside their established
yearling bulls made more prospecting movements home range (Fig. 2). The beginning and ending points
than any other age-sex class, and because elk disperse were located at the edge of the home ranges; all the
at about this age (Petersburg et al. 2000). To avoid other points of the prospecting movements were
locations with large errors, we removed two-dimen- located outside the home ranges. To avoid circularity,
sional fixes with a PDOP [5 (Lewis et al. 2007). we used data for stay-at-home individuals to estimate
AZGFD equipped desert bighorn sheep in the Black resistance and data for animals that undertook these
Mountains from 2004 through 2012 with remote- prospecting movements to evaluate the models.
download, self-dropping, store-on-board GPS collars
(model TTW-3590; Telonics, Inc., Mesa, AZ; McKin- Estimating habitat suitability
ney and Smith 2007; Bristow and Crabb 2008) and
satellite uplink collars (NSG-D2 Northstar Science To build habitat suitability models based on expert
and Technology, LLC, King George Virginia; Gagnon opinion, we used previously published expert opinions
et al. 2011). Most collars recorded the animal’s of resistance for elk (Table S5; Beier et al. 2008b) and
location every 5 h for 6–24 months, some collars bighorn sheep (Table S6; Beier and Majka 2006). To
recorded the location every 2 h (Gagnon et al. 2011). generate these opinions, biologists familiar with the
Because precipitation amounts can affect desert big- region reviewed the literature on habitat selection by
horn sheep space and resource use (Hoglander 2012), the focal species. They then assigned a score between
we selected bighorn sheep location data from July to 1 and 10 (1 being most suitable) to vegetation classes,
March, but omitted data from April, May and June topographic positions, elevation classes, and distance-
when mean precipitation over the last 30 years was to-road classes, and weighted these four factors
below 1 cm per month (http://www.usclimatedata. according to their estimated importance.
com/climate/henderson/nevada/united-states/usnv0122). To estimate habitat suitability with a resource
We removed two-dimensional fixes with a PDOP [ 5. selection function (RSF), we conducted a two-step
For the RSF, we reduced the GPS collar data sets conditional logistic regression to quantify selection
to resemble VHF telemetry study data sets by for each habitat attribute (Compton et al. 2002;
randomly selecting one location between 17:00 and Manly et al. 2002; Zeller et al. 2014). First, we
19:00 h per day for each elk, and one location determined the optimum spatial scale of each habitat
between 9:00 and 11:00 h per day for each bighorn covariate to consider as available. Second, we created
sheep. The time windows reflected times of peak a multi-scale model by using the best spatial scale for
activity for elk (Green and Bear 1990) and bighorn each habitat covariate. In a conditional logistic
sheep (Alderman et al. 1989). Short time intervals regression, used habitat is compared to available
would likely have greater dependence between habitat. We characterized used habitat by values of
successive locations. We only used the data for habitat variables in grid cells with GPS fixes. To
animals that maintained coherent home ranges characterize available habitat, we first fit a general-
throughout the 12- to 18-month life of their collars ized Pareto distribution to the straight-line distances
(i.e., ‘‘stay-at-home’’ individuals). between GPS points separated by a fixed time interval
For the SSF, we limited the data sets to include only (Zeller et al. 2014) using the POT package (Ribatet
consecutive GPS fixes with the minimum GPS fix 2012). At each location, we then placed a Pareto
acquisition intervals (elk: 2 h; bighorn sheep: 5 h). We kernel and truncated this kernel at the smaller of the
defined a movement step as the straight-line path 97.5 percentile or the maximum observed movement
between two consecutive GPS fixes. To focus on distance. We calculated the mean value of each
movement behavior, we excluded steps \200 m for habitat variable in the kernel, inversely weighted by
elk and\100 m for bighorn sheep. We only used data the distance to the observed location (Zeller et al.
of stay-at-home individuals. 2014). We calculated the difference between used

123
Landscape Ecol

Fig. 2 Example of a long-


distance path (diamonds) of
a yearling male elk in
relation to the home range
(circles)

and available for each variable, and set the response To estimate habitat suitability with a step selection
variable equal to one for each GPS location. function (SSF), we characterized used habitat as the
To determine the optimum spatial scale of each proportion of each land-cover type, and the mean of the
habitat covariate to consider as available in the first continuous variables, within a 30-m buffer on each side
step, we calculated four Pareto kernels for elk, of the segment to account for GPS error (Rettie and
corresponding to 2-, 4-, 8-, and 16-h intervals between McLoughlin 1999) and to include the immediate
consecutive fixes, and 3 Pareto kernels for bighorn environment around each step. We characterized avail-
sheep, corresponding to 5-, 10-, and 20-h acquisition able habitat as these same proportions and means within
intervals. We then used a conditional mixed-effects a Pareto kernel centered on the starting point of each
logistic regression (lmer function in the lme4 package, step. We generated the Pareto kernel the same way as in
v.0.999999-2, Bates et al. 2014), with individual the RSF calculations, but only for one scale (Zeller et al.
animals as a random effect, and identified the scale 2014, 2015). We used the differences between used and
with the lowest AICc ranking (AIC corrected for small available habitat values as predictor variables. In this
sample size; Burnham and Anderson 2002) as the best specification, the response variable is 1 and there is no
scale at which to assess availability. model intercept (Agresti 2002; Zeller et al. 2015). We
Our top models were generally clearly better than created a full model and conducted an automated model
alternative models. Therefore, to create the multi-scale selection with the glm function in the glmulti package
model we used the best scale for each habitat (Calcagno and de Mazancourt 2010). We selected the
covariate, as identified in step one. Because many of best model based on the AICc value.
our paired mixed-effects models using individual
animals as the random effect failed to converge, we Generating resistance surfaces
used generalized linear models with a binomial error
distribution (glm function, R Core Team 2014). We We used the expert habitat suitability estimates and
conducted an automated model selection using the the coefficients of the best RSF and SSF models to
glmulti package in R (Calcagno and de Mazancourt create habitat suitability surfaces. We parameterized
2010). This package uses a genetic algorithm to find resistances per cell, rx, across the landscape as:
the best models for large candidate sets. We selected
r x ¼  1  z 1 þ 2  z 2 þ    þ i  z i ; ð1Þ
the best model based on the AICc value.

123
Landscape Ecol

where ß1-i are the habitat suitability beta coefficients, using the R-package gdistance (van Etten 2014). We
and z1-i are the environmental covariates. then arrayed the 50 paths by cost, and calculated the
We rescaled the suitability values to a range of 0–1 cost rank of the observed path. The mean cost rank
with Eq. 2. (MCR) of several observed paths is a measure of
aptness of the underlying resistance model. An MCR
f ðxÞ ¼ ðx  minÞ=ðmax  minÞ; ð2Þ of 1 is best, MCR values near the median (25.5)
where x is the habitat suitability value of a grid cell, indicate a model that performs about the same as
and min and max are the minimum and maximum chance, and MCR of 50 is worst. Therefore, MCR
habitat suitability values of the habitat suitability evaluates how well the model predicts long-distance
surface, respectively. Values near one indicate the paths. If actual paths consistently have low costs under
most suitable habitat conditions and values near zero a model, then the model can be expected to identify
indicate the least suitable habitat. corridors that would support exploratory and dispersal
movements by that species.
Transforming habitat suitability into resistance We tested for differences between the eight trans-
formations of each suitability model (expert opinion,
To transform suitability values into resistance values, RSF, SSF) with linear mixed effects models (R-
we used eight curves based on the transformation packages lme4: Bates et al. 2014; and lmerTest:
function (Eq. 3) to define the relationship between Kuznetsova et al. 2015):
suitability and resistance (Trainor et al. 2013): cost difference  suitability model  transformation
100  99  ðð1  expðc  hÞÞ = ð1  expðcÞÞÞ; þ ð1jIDÞ þ ð1jID : rotated pathÞ;
ð3Þ ð4Þ
where h is the suitability matrix, and c = 0.25, 0.5, 1, where cost_difference is the difference in cost
2, 4, 8, 16 or 32. between the observed path and the rotated paths and
For each of the eight transformations, resis- ID refers to an observed path and its rotated versions.
tance = 1 when suitability = 1 and resistance =100 The last two terms specify random effects attributed to
when suitability = 0. At c = 0.25, the relationship is the effects of the ID, and to the interaction between the
nearly linear; as c increases, resistance values become ID and the individuals rotated paths, respectively.
an increasingly nonlinear negative exponential func- P-values based on Satterthwaite’s approximations
tion of suitability (Fig. 1). For each group of animals indicate significantly different models and transfor-
(yearling bull elk, bighorn sheep ewes, bighorn sheep mations. We also tested whether a resistance model
rams), we generated 24 resistance surfaces: eight was significantly better than random (rank of 25.5)
transformations each of expert habitat suitability with 1-tailed Student’s t-tests.
estimates (Tables S5, S6), RSF suitability values,
and SSF habitat suitability values.
Results
Evaluating resistance surfaces
Animal locations
To evaluate whether the observed long-distance
movement paths were low-cost paths as predicted by We analyzed 35,836 fixes across 33 stay-at-home
one or more resistance models, we generated 49 yearling bulls (range = 68–3092, median = 987);
‘‘random’’ paths per observed path by rotating the 44,390 GPS fixes of 29 stay-at-home ewes (range =
observed path around the center point (see Cushman 359–2418; median = 1711), and 22,308 fixes of 17
et al. 2010, 2011) using the R-package adehabitatLT stay-at-home rams (range = 627–2266, median =
(Calenge 2006; Fig. S1). For each resistance model, 1481). We identified nine long-distance movements
we calculated the cost of each of the 50 paths (1 made by six yearling bull elk, 11 long-distance
observed, 49 random) by summing the least cost movements by seven ewes, and 15 long-distance
distances between consecutive GPS fixes of paths movements by 13 rams.

123
Landscape Ecol

Habitat suitability estimates preferred desert scrub and riparian areas in the RSF
model, but avoided these landcover types in the SSF
The Pareto kernel distances for elk in the RSF varied model. Rams showed a slight preference for developed
from 1430 to 4705 m for elk and from 1000 to 2529 m areas in the RSF model, but strongly avoided them in
for bighorn sheep (Table 1). The characteristic scale the SSF model. The expert model matches the RSF
varied among habitat variables. For yearling bull elk, and SSF models closely: all three models show that
the strongest response occurred at the finest scale (2-h bighorn sheep prefer medium elevations and steep
and 1430-m) for elevation, slope, northness, distance slope but avoid roads.
to highways, and shrub; but at a coarse scale (16-h and
4705-m) for eastness, distance to road, open, riparian Transforming habitat suitability into resistance
and developed areas of medium to high intensity
(Table S7). Bighorn sheep ewes and rams responded The relationship between resistance and suitability
most strongly to most habitat variables at the 20-h and was better defined by negative exponential functions
[2400-m scale (Tables S8, S9). Rams selected for the resource and step selection models of bighorn
elevation and sparse vegetation at the 5-h and sheep rams and ewes (Fig. 3; Tables S12, S13). For the
1366-m scale and grassland at the 10-h and 1743-m expert models of bighorn sheep and for all elk models,
scale. Ewes also responded to elevation and sparse there were no significant differences among transfor-
vegetation, but also to solar insolation and open mation functions, but there was a trend towards better
developed areas at the finest scale. performance of moderately exponential functions
In both the best multivariate RSF and SSF (Tables S12, S13; Fig. 3).
models, yearling bull elk strongly avoided devel-
oped areas of medium–high intensity, but strongly Evaluating resistance surfaces
preferred developed open areas such as golf courses
(Tables S10, S11a) and north-facing areas. Although At the transformations that yielded the lowest MCR,
elk avoided roads and highways in the RSF model, none of the elk resistance models performed better
elk were indifferent to small paved roads and tended than chance (1-tailed t-tests: p [ 0.07; Fig. 3). For
to occur close to highways in the step selection bighorn sheep ewes and rams, all models performed
model. The expert opinion model differed most from better than chance (1-tailed t-tests: p \ 0.015), includ-
the RSF and SSF models in suitability of open ing models in which suitability was linearly trans-
developed areas, which were moderately avoided in formed to estimate resistance (Fig. 3).
the expert opinion model but strongly favored in For bighorn sheep ewes and rams, the resource and
RSF and SSF models. step selection models performed significantly better
As expected, bighorn sheep preferred medium than the expert model, and there was no difference in
levels of elevation and ruggedness and preferred to performance between the resource and step selection
be near permanent water (Table S11b, c). Ewes models (Fig. 3; Table S13). For elk, we did not detect

Table 1 GPS fix Species and sex GPS acquisition interval Pareto kernel distance (m)
acquisition intervals and the
associated Pareto kernel Elk yearling bulls 2-h steps 1430
distances
4-h steps 2173
8-h steps 2891
16-h steps 4705
Bighorn sheep rams 5-h steps 1366
10-h steps 1743
20-h steps 2529
Bighorn sheep ewes 5-h steps 1003
10-h steps 1679
20-h steps 2396

123
Landscape Ecol

(a) a difference in performance among the three models


(Fig. 3; Table S13).

Discussion

Our results indicate habitat suitability estimates can


yield resistance estimates that reliably identify long-
distance movement paths for desert bighorn sheep, a
habitat specialist, but that do not identify long-
distance paths of elk, a habitat generalist. For bighorn
sheep ewes and rams, the expert opinion model did not
perform as well as the data-intensive RSF and SSF
(b) models, but still performed better than chance. Con-
sidering the cost of obtaining data for RSF or SSF
models, the combination of an expert opinion model
generated for the study area and a negative exponential
transformation of habitat suitability into resistance
may be, as in the case of the bighorn sheep, a reliable
foundation for corridor design. However, for little
studied species it may be safer to spend the time and
money needed to estimate habitat suitability to ensure
a sound underlying model for generating resistance
values.
Contrary to our hypothesis, resistance derived from
RSF (i.e., from telemetry locations obtained once per
day during a period of peak activity) was as reliable as
(c)
resistance derived from SSF (movement steps). We
observed a few differences between habitat suitability
models that might be due to incorporating different
spatial scales in the RSF models but only working with
the 2-h scale in the SSF models. We are not aware of
any studies comparing these two approaches, but if
this pattern holds for other species, resistance can be
estimated using data from relatively inexpensive
radio-transmitters.
For desert bighorn sheep, the resource and step
selection models had the lowest MCR when the
relationship between resistance values and habitat
suitability followed a strongly negative exponential
Fig. 3 Effect of the function used to transform habitat curve. This indicates that bighorn sheep, when under-
suitability (see Fig. 1) into resistance estimates on the mean taking long-distance movements, move through habi-
(±1 SE) cost rank of observed paths in comparison to 49 random tat of moderate and moderately low suitability as
paths. Lower mean ranks indicate a better transformation or
model (expert opinion, resource selection functions, step readily as through highly suitable habitat. Only the
selection functions). A rank of 25.5 (dashed line) indicates least suitable habitat poses high resistance to long-
performance of a random resistance model. a elk, b bighorn distance movements. Similarly, landscape resistance
sheep ewes, c bighorn sheep rams to exploratory movements of red-cockaded

123
Landscape Ecol

woodpeckers was a strongly negative exponential move their genes through these long corridors. For
function of habitat suitability (Trainor et al. 2013), these species, a long corridor can be conceptualized as
and landscape genetic patterns of brown bears and stepping stones of suitable habitat, separated by gaps
kinkajous (Potos flavus) were consistent with a negative crossed in discrete dispersal events (Beier et al.
exponential relationship between resistance and habitat 2008a). Research is needed on the minimum size of
suitability (Mateo-Sánchez et al. 2015; Keeley 2015). stepping stones, and on how resistance and habitat use
Similarly, habitat conditions that were avoided within are related for these less-mobile species.
the home range were readily traversed by dispersing For elk, none of the three models performed better
individuals in many species, including the butterflies than chance. We think the main reason for this
Junonia coenia and Euptoieta claudia (Haddad and outcome is that elk are generalist herbivores (Van
Tewksbury 2005), hispid cotton rats (Sigmodon Deelen et al. 1997), whereas bighorn sheep are habitat
hispidus, Bowne et al. 1999), Siberian flying squirrels specialists (Turner et al. 2004). The habitat niche of
(Pteromys volans, Selonen and Hanski 2006), Iberian elk was strikingly broad during the long distance
lynx (Lynx pardinus, Gastón et al. 2016), and lions movements that we observed. Similarly, McClure-
(Panthera leo, Elliot et al. 2014). Similar results are Rainey (2012) found that least-cost paths informed by
evident in landscape genetic studies that—instead of habitat suitability models using observed migration
transforming combined habitat suitability values into paths of elk predicted an independent set of migration
resistance with negative exponential functions—first paths no better than null models. In contrast, Poor et al.
conducted univariate scaling to determine the best scale (2012) examined how well least-cost models matched
and functional form of each landscape variables. In migratory movements of pronghorn (Antilocapra
these studies, initial parameter values were selected americana), a habitat specialist (Wood 1989). They
based on literature review and expert opinion. In the created least-cost paths based on Maxent habitat
optimized multivariate models that best explained how suitability models and buffered them by 10 % of the
landscape structure influences gene flow, only 1–2 of most traversable habitat. Because 71 % of migrating
the initial 4–5 landscape variables were retained, pronghorn GPS locations occurred in these least-cost
indicating that mountain goats (Oreamnos americanus, corridors, Poor et al. (2012) concluded the least-cost
Shirk et al. 2010), American marten (Martes ameri- method is a valid approach for conservation planning.
cana, Wasserman et al. 2010), and American pikas Collectively, these results support the conclusion of
(Ochotona princeps, Castillo et al. 2014) are con- Beier et al. (2009) that habitat specialists are the most
strained by fewer landscape variables during dispersal important focal species for design of multi-species
and mating movements than during daily movements. corridors because habitat generalists can easily move
Given this evidence, we propose that when through landscapes conserved for habitat specialists,
researchers and managers design corridors to facilitate but habitat specialists may not readily move through a
dispersal, migration, and other long-distance move- corridor designed for a generalist.
ments of mobile animals, a negative exponential This paper introduces MCR as a tool to evaluate
relationship should be the default assumption when aptness of resistance models. MCR can determine if a
transforming habitat suitability into resistance. Until model is statistically better than random with only a
now, the negative linear relationship has been the small number of observed paths (as few as nine paths
default, despite the lack of any empirical studies in our analyses). MCR can also be used to statistically
supporting a linear transformation (Beier et al. 2008a). test for differences between alternative resistance
Use of an exponential transformation has important models. Finally, this tool considers habitat use along
implications for corridor design because it means that the entire path, not only at the actual telemetry fix
larger fractions of the landscape offer low resistance, locations, and thus takes advantage of the rich
allowing greater flexibility in where a corridor is information yielded by GPS radio-tags.
located. We caution, however, that all species studied Some studies that compared resistance estimates
were relatively mobile animals that can pass through a from expert opinion and empirical methods found
long corridor designed to connect adjacent popula- expert opinion is not reliable (e.g., Seoane et al. 2005;
tions in a single movement lasting a few hours or days. Shirk et al. 2010; Charney 2012). However, our study
In contrast, corridor dwellers require generations to shows expert opinion can predict long-distance

123
Landscape Ecol

movements of desert bighorn sheep. Because many Beck JL, Smith KT, Flinders JT, Clyde CL (2013) Seasonal
corridor designs are based on resistances derived from habitat selection by elk in north central Utah. West N Am
Nat 73:442–456
expert opinion (e.g., Arizona Missing Linkages: http:// Beier P, Garding E, Majka D (2008b) Arizona Missing Link-
corridordesign.org/linkages/arizona, South Coast ages: Bradshaw Mountains to Agua Fria linkage design.
Missing Linkages in California: Beier et al. 2006), it is Report to Arizona Game and Fish Department, School of
useful to know that expert opinion paired with litera- Forestry, Northern Arizona University
Beier P, Majka DR (2006) Arizona Missing Linkages: US-93:
ture review of empirical studies can be a sound source Wickenburg to Santa Maria River linkage design. Report to
for resistance estimates for some focal species. We Arizona Game and Fish Department, School of Forestry,
expect expert opinion models of well-studied species Northern Arizona University
to be reliable, but recommend habitat suitability Beier P, Majka DR, Newell SL (2009) Uncertainty analysis of
least-cost modeling for designing wildlife linkages. Ecol
modeling for little studied species to ensure a good Appl 19:2067–2077
underlying model for generating resistance values. Beier P, Majka DR, Spencer WD (2008a) Forks in the road:
Studies of additional habitat specialists are needed to choices in procedures for designing wildland linkages.
verify our findings and to quantify the tradeoff between Conserv Biol 22:836–851
Beier P, Penrod KL, Luke C, Spencer W, Cabañero C (2006)
cost and reliability of different suitability estimates. South Coast Missing Linkages: restoring connectivity to
wildlands in the largest metropolitan area in the USA. In:
Acknowledgments We thank Ester Rubin, AZGFD, for Crooks K, Sanjayan M (eds) Connectivity conservation.
arranging use of data, Christopher Coffey for computer cluster Cambridge University Press, New York, pp 55–586
support, Jeff Jenness for ArcGIS support, and Kathy Zeller for Beier P, Spencer WD, Baldwin R, McRae B (2011) Best science
advice on the analyses. NAU School of Forestry, McIntire- practices for regional connectivity mapping. Conserv Biol
Stennis Cooperative Forestry Research Program, Arizona 25:879–892
Board of Forestry, and the Hafen, Krimminger, Prather, Czak, Booth GD, Niccolucci MJ, Schuster EG (1994) Identifying
Berry, and Forestry Faculty scholarships at Northern Arizona proxy sets in multiple linear regression: an aid to better
University supported A.T.H.K. during this work. Sam Cushman, coefficient interpretation. Research paper, US Department
Carol Chambers, and two anonymous reviewers helped to of Agriculture, Forest Service
significantly improve the manuscript. Ungulate data were Bowne DR, Peles JD, Barrett GW (1999) Effects of landscape
collected by AZGFD with funding from Wildlife Restoration spatial structure on movement patterns of the hispid cotton
Act, Special Big Game License Tag Funds raised by the Arizona rat (Sigmodon hispidus). Landscape Ecol 14:53–65
Desert Bighorn Sheep Society, Arizona Elk Society, Rocky Boyce MS, Mao JS, Merrill EH, Fortin D, Turner MG, Fryxell J,
Mountain Elk Foundation, Arizona Antelope Foundation, and Turchin P (2003) Scale and heterogeneity in habitat
Arizona Big Game Super Raffle. selection by elk in Yellowstone National Park. Ecoscience
10:421–431
Bristow K, Crabb M (2008) Evaluation of distribution and trans-
highway movement of Desert Bighorn Sheep: Arizona
References Highway 68, Arizona, USA. Final Report 588
(2005–2008). Arizona Department of Transportation
Adriaensen F, Chardon JP, deBlust G, Swinnen E, Villalba S, Research Center, Phoenix
Gulinck H, Matthysen E (2003) The application of ‘least- Burnham KP, Anderson DR (2002) Model selection and mul-
cost’ modeling as a functional landscape model. Landsc timodel inference: a practical information-theoretic
Urban Plan 64:233–247 approach. Springer, New York
Adriaensen F, Githiru M, Mwang’ombe J, Matthysen E, Lens L Calcagno V, de Mazancourt C (2010) glmulti: an R package for
(2007) Restoration and increase of connectivity among easy automated model selection with (generalized) linear
fragmented forest patches in the Taita Hills, Southeast models. J Stat Software 34:1–29
Kenya. Part II technical report, CEPF project 1095347968, Calenge C (2006) The package adehabitat for the R software: a
University of Gent, Gent tool for the analysis of space and habitat use by animals.
Agresti A (2002) Categorical data analysis, 2nd edn. Wiley, Ecol Model 197:516–519
Hoboken Castillo JA, Epps CW, Davis AR, Cushman SA (2014) Land-
Alderman JA, Krausman PR, Leopold BD (1989) Diel activity scape effects on gene flow for a climate-sensitive montane
of female desert bighorn sheep in western Arizona. J Wildl species, the American pika. Mol Ecol 23:843–856
Manag 53:264–271 Charney ND (2012) Evaluating expert opinion and spatial scale
Baguette M, Blanchet S, Legrand D, Stevens VM, Turlure C in an amphibian model. Ecol Model 242:37–45
(2013) Individual dispersal, landscape connectivity and Coe PK, Johnson BK, Wisdom MJ, Cook JG, Vavra M, Nielson
ecological networks. Biol Rev 88:310–326 RM (2011) Validation of elk resource selection models with
Bates D, Maechler M, Bolker B, Walker S (2014) Lme4: linear spatially independent data. J Wildl Manag 75:159–170
mixed-effects models using Eigen and S4 (R package Compton BW, Rhymer JM, McCollough M (2002) Habitat
version 11-7). http://www.CRANR-projectorg/package= selection by wood turtles (Clemmys insculpta): an appli-
lme4. Accessed 4 Mar 2016 cation of paired logistic regression. Ecology 83:833–843

123
Landscape Ecol

Creel S, Winnie J Jr, Maxwell B, Hamlin K, Creel M (2005) Elk agriculture by a woodland species depends on cover type
alter habitat selection as an antipredator response to and behavioural state: insights from resident and dispersing
wolves. Ecology 86:3387–3397 Iberian lynx. J Appl Ecol. doi:10.1111/1365-2664.12629
Cushman SA, Chase M, Griffin C (2010) Mapping landscape Green RA, Bear GD (1990) Seasonal cycles and daily activity
resistance to identify corridors and barriers for elephant patterns of Rocky Mountain elk. J Wildl Manag 54:272–279
movement in Southern Africa. In: Cushman SA, Huett- Haddad NM, Tewksbury JJ (2005) Low-quality habitat corri-
mann F (eds) Spatial complexity, informatics, and wildlife dors as movement conduits for two butterfly species. Ecol
conservation. Springer Verlag, Japan, pp 349–367 Appl 15:250–257
Cushman SA, Raphael MG, Ruggiero LF, Shirk AS, Wasserman Hilty JA, Lidicker WZ Jr, Merenlender A (2012) Corridor
TN, O’Doherty EC (2011) Limiting factors and landscape ecology: the science and practice of linking landscapes for
connectivity: the American marten in the Rocky Moun- biodiversity conservation. Island Press, Washington, D.C.
tains. Landscape Ecol 26:1137–1149 Hoctor TS, Carr MH, Zwick PD (2000) Identifying a linked
Dodd NL, Gagnon JW, Boe S, Ogren K, Schweinsburg RE reserve system using a regional landscape approach: the
(2012) Wildlife-vehicle collision mitigation for safer Florida Ecological Network. Conserv Biol 14:984–1000
wildlife movement across highways: State Route 260. Hoglander CJ (2012) Developed waters for wildlife conserva-
Final Project Report 603, Arizona Department of Trans- tion: collaborative approaches and landscape models for
portation Research Center, Phoenix desert bighorn sheep (Ovis canadensis mexicana) conser-
Dodd NL, Gagnon JW, Boe S, Schweinsburg RE (2006) Char- vation. Thesis, Northern Arizona University
acteristics of elk—vehicle collisions and comparison to Hunter RD, Fisher RN, Crooks KR (2003) Landscape level
GPS-determined highway crossing patterns. In: Proceed- connectivity in coastal southern California USA as asses-
ings of the 2005 international conference on ecology and sed through carnivore habitat suitability. Nat Area J
transportation, North Carolina State University, Raleigh, 23:302–314
pp 461–477, 29 Aug–2 Sept 2005 Jenness J (2013) DEM surface tools. Jenness Enterprises,
Dodd NL, Gagnon JW, Sprague S, Boe S, Schweinsburg RE Flagstaff
(2012) Wildlife accident reduction study and monitoring: Johnson BK, Kern JW, Wisdom MJ, Findholt SL, Kie JG (2000)
Arizona State Route 64. Final Project Report 626, Arizona Resource selection and spatial separation of mule deer and
Department of Transportation Research Center, Phoenix elk during spring. J Wildl Manag 64:685–697
Dormann CF, Elith J, Bacher S, Buchmann C, Carl G, Carré G, Keeley ATK (2015) Comparing estimates of landscape resis-
Marquéz JR, Gruber B, Lafourcade B, Leitão PJ, Münke- tance to animal movement. Dissertation, Northern Arizona
müller T (2013) Collinearity: a review of methods to deal University
with it and a simulation study evaluating their perfor- Kuznetsova A, Brockhoff PB, Christensen RHB (2015)
mance. Ecography 36:27–46 lmerTest: tests in linear mixed effects models (R package
Elliot NB, Cushman SA, Macdonald DW, Loveridge AJ (2014) version 20-25). http://www.CRANR-projectorg/package=
The devil is in the dispersers: predictions of landscape lmerTest. Accessed 4 Mar 2016
connectivity change with demography. J Appl Ecol LANDFIRE (2008) Existing vegetation type layer. US Depart-
51:1169–1178 ment of the Interior, Geological Survey. www.landfire.gov.
ESRI (2011) ArcGIS desktop: release 10. Environmental Sys- Accessed 4 Mar 2016
tems Research Institute, Redlands Larkin JL, Maehr DS, Hoctor TS, Orlando MA, Whitney K
Ferreras P (2001) Landscape structure and asymmetrical inter- (2004) Landscape linkages and conservation planning for
patch connectivity in a metapopulation of the endangered the black bear in west-central Florida. Anim Conserv
Iberian lynx. Biol Conserv 100:125–136 7:23–34
Fortin D, Morales JM, Boyce MS (2005) Elk winter foraging at Lewis JS, Rachlow JL, Garton EO, Vierling LA (2007) Effects
fine scale in Yellowstone National Park. Oecologia of habitat on GPS collar performance: using data screening
145:334–342 to reduce location error. J Appl Ecol 44:663–671
Gagnon JW, Dodd NL, Sprague S, Loberger C, Boe S, Sch- Manly BFJ, McDonald LL, Thomas DL, McDonald TL,
weinsburg RE (2011) Evaluation of measures to promote Erickson WP (2002) Resource selection by animals: sta-
desert bighorn sheep highway permeability: US Highway tistical analysis and design for field studies. Kluwer,
93. Final Project Report 677, Arizona Department of Dordrecht
Transportation Research Center, Phoenix Mao JS, Boyce MS, Smith DW, Singer FJ, Vales DJ, Vore JM,
Gagnon JW, Dodd NL, Sprague S, Nelson RE, Loberger C, Boe Merrill EH (2005) Habitat selection by elk before and after
S, Schweinsburg, RE (2013) Elk movements associated wolf reintroduction in Yellowstone National Park. J Wildl
with a high-traffic highway: Interstate 17. Final Project Manag 69:1691–1707
Report 647, Arizona Department of Transportation Mateo-Sánchez MC, Balkenhol N, Cushman S, Pérez T, Dom-
Research Center, Phoenix ı́nguez A, Saura S (2015) A comparative framework to
Gagnon JW, Sprague S, Dodd NL, Loberger C, Nelson RE, Boe infer landscape effects on population genetic structure: are
S, Schweinsburg RE (2012) Research report on elk move- habitat suitability models effective in explaining gene
ments associated with Interstate 40. Design concept study flow? Landscape Ecol 30:1405–1420
and environmental assessment. I-40, NH 040-C(211)S40 McClure-Rainey M (2012) Validating alternative methods of
CN 183 H7586 01L, Arizona Game and Fish Department modeling wildlife corridors using relocation data from
Gastón A, Blázquez-Cabrera S, Garrote G, Mateo-Sánchez MC, migrating elk and dispersing wolverines. Dissertation,
Beier P, Simón MA, Saura S (2016) Response to Montana State University

123
Landscape Ecol

McKinney T, Smith T (2007) US93 Bighorn Sheep Study: Stamps JA, Krishnan VV (1999) A learning-based model of
distribution and trans-highway movements of desert big- territory establishment. Q Rev Biol 74:291–318
horn sheep in Northwestern Arizona. Final Report 576, Trainor AM, Walters JR, Morris WF, Sexton J, Moody A (2013)
Arizona Game and Fish Department Empirical estimation of dispersal resistance surfaces: a
Petersburg ML, Alldredge AW, de Vergie WJ (2000) Emigra- case study with red-cockaded woodpeckers. Landscape
tion and survival of 2-year-old male elk in northwestern Ecol 28:755–767
Colorado. Wildl Soc Bull 28:708–716 Trakhtenbrot A, Nathan R, Perry G, Richardson DM (2005) The
Poor EE, Loucks C, Jakes A, Urban DL (2012) Comparing importance of long-distance dispersal in biodiversity con-
habitat suitability and connectivity modeling methods for servation. Divers Distrib 11:173–181
conserving pronghorn migrations. PLoS ONE 7:e49390 Turner JC, Douglas CL, Hallum CR, Krausman PR, Ramey RR
Pullinger MG, Johnson CJ (2010) Maintaining or restoring (2004) Determination of critical habitat for the endangered
connectivity of modified landscapes: evaluating the least- Nelson’s bighorn sheep in southern California. Wildl Soc
cost path model with multiple sources of ecological Bull 32:427–448
information. Landscape Ecol 25:1547–1560 Van Deelen TR, McKinney LB, Joselyn MG, Buhnerkempe JE
R Core Team (2014) R: a language and environment for sta- (1997) Can we restore elk to southern Illinois? The use of
tistical computing. R Foundation for Statistical Comput- existing digital land-cover data to evaluate potential habi-
ing, Vienna. http://www.R-project.org/. Accessed 4 Mar tat. Wildl Soc Bull 25:886–894
2016 Van Etten J (2014) gdistance: distances and routes on geo-
Rettie WJ, McLoughlin PD (1999) Overcoming radio telemetry graphical grids (R package version 1.1-5). http://CRAN.R-
bias is habitat-selection studies. Can J Zool 77:1175–1184 project.org/package=gdistance. Accessed 4 Mar 2016
Ribatet M (2012) Generalized Pareto distribution and peaks Wait S, McNally H (2004) Selection of habitats by wintering elk
over threshold (R package). http://cran.r-project.org/web/ in a rapidly subdividing area of La Plata County. In: Shaw
packages/POT/index.html. Accessed 15 Apr 2014 WW, Harris LK, Vandruff L (eds) Proceedings of the
Richards-Zawacki CL (2009) Effects of slope and riparian fourth international symposium on urban wildlife conser-
habitat connectivity on gene flow in an endangered Pana- vation, Tucson, 2004
manian frog, Atelopus varius. Divers Distrib 15:796–806 Walker R, Craighead L (1997) Analyzing wildlife movement
Sappington J, Longshore KM, Thompson DB (2007) Quanti- corridors in Montana using GIS. In: Proceedings of the
fying landscape ruggedness for animal habitat analysis: a 1997 international ESRI users conference, Environmental
case study using bighorn sheep in the Mojave Desert. Sciences Research Institute, San Diego, 7–11 July 1997
J Wildl Manag 71:1419–1426 Wallace MC, Krausman PR (1997) Movements and home-ran-
Sawyer SC, Epps CW, Brashares JS (2011) Placing linkages ges of elk in eastern Arizona. In: deVos JC Jr (ed) Pro-
among fragmented habitats: do least-cost models reflect ceedings of the 1997 deer/elk workshop, Rio Rico, 1997
how animals use landscapes? J Appl Ecol 48:668–678 Wasserman TN, Cushman SA, Schwartz MK, Wallin DO (2010)
Selonen V, Hanski IK (2006) Habitat exploration and use in dis- Spatial scaling and multi-model inference in landscape
persing juvenile flying squirrels. J Anim Ecol 75:1440–1449 genetics: Martes americana in northern Idaho. Landscape
Seoane J, Bustamante J, Dı́az-Delgado R (2005) Effect of expert Ecol 25:1601–1612
opinion on the predictive ability of environmental models Wikramanayake E, McKnight M, Dinerstein E, Joshi A, Gurung
of bird distribution. Conserv Biol 19:512–522 B, Smith D (2004) Designing a conservation landscape for
Shirk AJ, Wallin DO, Cushman SA, Rice CG, Warheit KI tigers in human-dominated ecosystems. Conserv Biol
(2010) Inferring landscape effects on gene flow: a new 18:839–844
model selection framework. Mol Ecol 19:3603–3619 Wood AK (1989) Comparative distribution and habitat use by
Singleton PH, Gaines WL, Lehmkuhl JF (2002) Landscape antelope and mule deer. J Mammal 70:335–340
permeability for large carnivores in Washington: a geo- Zeller KA, McGarigal K, Beier P, Cushman SA, Vickers TW,
graphic information system weighted-distance and least- Boyce WM (2014) Sensitivity of landscape resistance esti-
cost corridor assessment. Research Paper PNW-RP-549 mates based on point selection functions to scale and behav-
US Forest Service Pacific Northwest Research Station, ioral state: pumas as a case study. Landscape Ecol 29:541–557
Portland Zeller KA, McGarigal K, Cushman SA, Beier P, Vickers TW,
Spear SF, Balkenhol N, Fortin MJ, McRae BH, Scribner K Boyce WM (2015) Using step and path selection functions
(2010) Use of resistance surfaces for landscape genetic for estimating resistance to movement: pumas as a case
studies: considerations for parameterization and analysis. study. Landscape Ecol. doi:10.1007/s10980-015-0301-6
Mol Ecol 19:3576–3591 Zeller KA, McGarigal K, Whiteley AR (2012) Estimating
Stamps JA (2001) Habitat selection by dispersers: integrating landscape resistance to movement: a review. Landscape
proximate and ultimate approaches. In: Clobert J, Danchin Ecol 27:777–797
E, Dhondt AA (eds) Dispersal. Oxford University Press,
Oxford, pp 230–242

123

View publication stats

You might also like