Content uploaded by Allan M. Strong
Author content
All content in this area was uploaded by Allan M. Strong on Sep 21, 2018
Content may be subject to copyright.
Early View (EV): 1-EV
The relative contribution of local habitat and landscape context
to metapopulation processes: a dynamic occupancy modeling
approach
Sarah J. K. Frey , Allan M. Strong and Kent P. McFarland
S. J. K. Frey (sarah.frey@oregonstate.edu), Dept of Forest Ecosystems and Society, Oregon State Univ., Corvallis, OR 97331, USA. – A. M. Strong,
Rubenstein School of Environment and Natural Resources, Univ. of Vermont, Burlington, VT 05405, USA. – K. P. McFarland, Vermont Center
for Ecostudies, PO Box 420, Norwich, VT 05055, USA.
Changes in site occupancy across habitat patches have often been attributed to landscape features in fragmented systems,
particularly when considering metapopulations. However, failure to include habitat quality of individual patches can mask
the relative importance of local scale features in determining distributional changes. We employed dynamic occupancy
modeling to compare the strength of local habitat variables and metrics of landscape patterns as drivers of metapopulation
dynamics for a vulnerable, high-elevation species in a naturally fragmented landscape. Repeat surveys of Bicknell ’ s thrush
Catharus bicknelli presence/non-detection were conducted at 88 sites across Vermont, USA in 2006 and 2007. We used an
organism-based approach, such that at each site we measured important local-scale habitat characteristics and quantifi ed
landscape-scale features using a predictive habitat model for this species. We performed a principal component analysis on
both the local and landscape features to reduce dimensionality. We estimated site occupancy, colonization, and extinction
probabilities while accounting for imperfect detection. Univariate, additive, and interaction models of local habitat and
landscape context were ranked using AICc scores. Both local and landscape scales were important in determining changes
in occupancy patterns. An interaction between scales was detected for occupancy dynamics indicating that the relation-
ship of the parameters to local-scale habitat conditions can change depending on the landscape context and vice versa.
An increase in both landscape- and local-scale habitat quality increased occupancy and colonization probability while
decreasing extinction risk. Colonization and extinction were both more strongly infl uenced by local habitat quality relative
to landscape patterns. We also identifi ed clear, qualitative thresholds for landscape-scale features. Conservation of large
habitat patches in high-cover landscapes will help ensure persistence of Bicknell ’ s thrushes, but only if local scale habitat
quality is maintained. Our results highlight the importance of incorporating information beyond landscape characteristics
when investigating patch occupancy patterns in metapopulations.
Understanding distributional patterns of organisms in space
and time is a fundamental question in ecology. Variations in
species occurrence patterns can elucidate drivers of important
population processes such as site occupancy, colonization,
and local extinction (Gaston 1990). By linking these pro-
cesses to environmental features we can begin to understand
the ecological factors that motivate habitat selection and
drive changes in species distributions. Landscape structure
(MacArthur and Wilson 1967, Hanski 1998) and composi-
tion (With et al. 1997), local patch characteristics (Mortelliti
et al. 2010), and species ’ dispersal capabilities ( omas 2000)
have all been implicated as important drivers of distribution
dynamics.
Incorporation of spatial structure into population
dynamics is a central concept of metapopulation models
(Hanski 1998). A metapopulation is defi ned as a network
of sub-populations that are linked by migration. Changes
in occupancy state, through subpopulation extinction and
colonization, depends on the size and isolation of the habi-
tat patch (Hanski 1998). However, most landscape studies
only consider features at the landscape-scale (i.e. patch size
and isolation), while ignoring local habitat quality within
patches (Mortelliti et al. 2010). is can be an oversimpli-
fi cation (Hastings and Harrison 1994), especially in hetero-
geneous ecosystems where distributions are likely driven by
factors at multiple scales. ere is increasing evidence that
this variation in local habitat quality is an important fac-
tor in population dynamics and should be incorporated into
metapopulation models ( omas et al. 2001, Fleishman
et al. 2002, Armstrong 2005), however, there have been few
empirical tests (Mortelliti et al. 2010). Further, investigations
of the infl uence of landscape structure on metapopulation
processes are generally conducted in anthropogenically frag-
mented forest surrounded by an agricultural matrix (Opdam
1991), as opposed to naturally fragmented systems.
D y n a m i c ( o r m u l t i - s e a s o n ) o c c u p a n c y m o d e l s ( M a c K e n z i e
et al. 2003) can be used to assess distributional patterns at a
variety of scales. Dynamic occupancy models allow estimation
of the probability that a site will be occupied, as well as colo-
nization and extinction probabilities. Occupancy modeling
Ecography 34: 001–009, 2011
doi: 10.1111/j.1600-0587.2011.06936.x
© 2011 e Authors. Journal compilation © 2011 Nordic Society Oikos
Subject Editor: Michel Baguette. Accepted 8 July 2011
2-EV
can be used to address an array of questions about distribu-
tional patterns through inclusion of survey- and site-specifi c
covariates to calculate unbiased colonization and extinction
rates through the incorporation of detection probabilities
(MacKenzie et al. 2003). Including detection probabilities
into dynamic occupancy models avoids issues associated
with false absences thought to be the largest source of bias in
traditional approaches to the estimation of metapopulation
parameters (Moilanen 2002).
Species with spatially isolated subpopulations, especially
those that occupy naturally restricted ranges, are well suited
for this type of approach (Hanski 1998). Bicknell ’ s thrush
Catharus bicknelli is a montane fi r-forest specialist that
inhabits a naturally fragmented breeding range in the north-
eastern United States, southeastern Qu é bec, and Maritime
Canada (Rimmer et al. 2001, Lambert et al. 2005). It
occupies ephemeral, disturbance-driven, mid-successional,
fi r-dominated forests within montane or highland regions.
Bicknell ’ s thrush is ranked as a top conservation prior-
ity among Nearctic-Neotropical migrants in the northeast
(Rich et al. 2004) with a global status of vulnerable (BirdLife
International 2000). A habitat model for Bicknell ’ s thrush
(Lambert et al. 2005) allowed us to defi ne landscape ele-
ments from a species-specifi c perspective (Betts et al. 2006).
Here we used dynamic occupancy modeling (MacKenzie
et al. 2003) and Akaike ’ s information criterion (AIC) model
selection techniques (Burnham and Anderson 2002) to test
the relative importance of local-scale habitat characteristics
versus landscape-scale features in determining Bicknell ’ s
thrush site occupancy patterns over time. We used this anal-
ysis to assess metapopulation processes within an existing
potential habitat model (Lambert et al. 2005) in Vermont
and determine how metapopulation processes relate to
habitat features at multiple scales.
Methods
Field surveys
Detection/non-detection data were collected over a two-
year period from 2006 to 2007 within the Bicknell ’ s thrush
breeding range across the state of Vermont, USA. We
focused on the metapopulation of the Green and Taconic
Mountains, and Northeastern Highlands of Vermont, where
sub-populations were defi ned by high-elevation habitat
islands delineated using an existing habitat model for this
species (Lambert et al. 2005). A total of 88 sites between 733
and 1236 m elevation were surveyed (Fig. 1). Twenty-nine
sites (hereafter, SF sites) were added to 59 sites surveyed in
Vermont through a citizen-science program called Mountain
Birdwatch (MBW, Hart and Lambert 2007). Each site con-
sisted of a 1-km transect of fi ve points separated by 200 –
250 m (seven sites contained 3 – 4 points due to patch size
constraints). Bird sample locations covered a rectangular area
of roughly ∼ 25 ha (1 km by ∼ 250 m), which approximates
the size of fi ve to 10 Bicknell ’ s thrush breeding home ranges
(2.33 – 4.53 ha, Rimmer et al. 2001).
MBW sites were selected through random selection of
high-elevation forests (montane areas ! 823 m, Hart and
Lambert 2007). SF sites were chosen from the remaining
un-surveyed MBW sites in Vermont and fi lled gaps in MBW
sampling by surveying marginal habitat and sites without
hiking trails. e 1-km transects were fi t into the habitat
polygons defi ned by Lambert et al. (2005), following a
straight line wherever possible, often along ridgelines.
Surveys were conducted during the peak of the breeding
season (late May – mid-July) at optimal activity times (dawn
and dusk) under favorable weather conditions. A maxi-
mum of three surveys were conducted at each site each year
(mean " SD # 2.1 " 0.9) with slight diff erences between
MBW and SF sites (see following two paragraphs for details).
In the MBW survey protocol, the fi rst survey period occurred
between 04:30 and 06:30 h EST and consisted of a 10-min
point count at each point along the transect. If no Bicknell ’ s
thrush were detected during the fi rst survey period, up to
two additional surveys were conducted to increase oppor-
tunity for detection. e second survey period directly fol-
lowed the fi rst survey and consisted of a 1-min playback of
Bicknell ’ s thrush songs and calls followed by a 2-min silent
listening period at each point. If no Bicknell ’ s thrush were
detected on either the fi rst or second surveys, a third survey
was conducted within two weeks following the initial sur-
veys (or before 15 July). e third survey occurred between
either 04:30 and 06:30 h or 20:00 and 21:00 h and was done
by broadcasting the 1-min playback and listening for 2 min
every 100 m along the transect.
For SF sites, the three surveys were almost always con-
ducted during a single visit to the site, weather permitting.
is was achieved by conducting an evening survey followed
by two morning surveys. During the evening survey a 5-min
point count was conducted followed by broadcasting a 1-min
playback and a 2-min listening period. Both morning sur-
veys followed the same protocol as MBW. Detections were
categorized as within or outside a 50-m radius around the
survey point, although all observations were counted assum-
ing an approximate detection limit of 125 m. For each sur-
vey at a given site (MBW and SF), either 1 (detection) or 0
(non-detection) was recorded based on whether a Bicknell ’ s
thrush was heard or seen anywhere along the transect. We
tested for an eff ect of survey technique on detection prob-
ability during the modeling process and found little support
for survey type to infl uence detection probability (Table 1).
Local-scale habitat measurements
Local habitat conditions were quantifi ed once in either 2006
or 2007 and were assumed to be constant within this time
period. Because within-site variation was negligible, local
habitat measurements were averaged across all points to
obtain a single site value. We measured site variables rep-
resentative of habitat quality for Bicknell ’ s thrush based on
the species ’ natural history (Table 2, Rimmer et al. 2001),
assuming these variables are linked to resource availability
(Strong et al. 2004). To quantify coniferous shrub density
at each point, we used the point-centered quarter method
(Cottam and Curtis 1956). We measured basal area of snags
(using a wedge prism), as snags are a useful structural indica-
tor of the two main causes of natural disturbance in montane
ecosystems in the northeastern US: severe weather and fi r
waves (Sprugel 1976), both of which result in areas of forest
3-EV
Figure 1. Area of study and survey sites in Vermont located within a Bicknell ’ s thrush predicted habitat model (Lambert et al. 2005). Suit-
able habitat was identifi ed as conifer-dominated forest (white pixels) within high elevation land units (outlined in black) delineated by an
elevational-latitudinal threshold.
Table 1. AICc model selection results for detection probability ( p ).
Each model is ranked by its AICc score, which represents how well
the model fi t the data. A lower ∆ AICc value is indicative of a better
model. The probability that the model (of the models tested) would
best explain the data is indicated by the AICc weight ( ω
i ). K is the
number of parameters each model estimates. Initial occupancy ( y ),
colonization ( g ), and local extinction ( e ) probabilities are constant in
all models.
Model AICc ∆ AICc ω
i K
y $ g $ e $ p (survey % ps) 362.19 0 0.77 7
y $ g $ e $ p (survey ) 366.92 4.73 0.07 6
y $ g $ e $ p (ps) 367.58 5.39 0.05 5
y $ g $ e $ p
(type ) 367.93 5.74 0.04 6
y $ g $ e $ p (playback ) 368.02 5.83 0.04 5
y $ g $ e $ p ( morning
v. evening)
370.88 8.69 0.01 5
y $ g $ e $ p (time ) 371.64 9.45 0.01 5
y $ g $ e $ p 372.16 9.98 0.01 4
y $ g $ e $ p (date ) 374.28 12.09 0.00 5
ps # p a t c h s i z e , s u r v e y # s u r v e y n u m b e r ( 1 – 3), type # s u r v e y t y p e .
regeneration. Bicknell ’ s thrush tend to select dense regener-
ating patches for nesting sites (Rimmer et al. 2001). Canopy
species composition was classifi ed as coniferous, deciduous,
or mixed to determine the proportion of conifer-dominated
points along the transect. Average canopy height (m) was
also measured at each point.
Due to the complexity inherent in incorporating four
local habitat covariates (Table 2) in the modeling process, we
collapsed them into a single variable using a principal com-
ponent analysis. We used principal component one (PC1)
in the analysis. is value was termed local in the modeling
process and is the sum of the products of each variable ’ s fac-
tor coeffi cient and standardized value.
Landscape features
Landscape features were quantifi ed using GIS software
(ESRI 2005). Lambert et al. (2005) created a habitat model
for Bicknell ’ s thrush by delineating high-elevation land
units with an elevational-latitudinal threshold based on
breeding season presence-absence data. Within this range,
conifer-dominated forest (based on forest composition from
National Land Cover Data [Vogelmann et al. 2001]) was
considered potential habitat (Fig. 1). We included two of the
most common metrics of landscape structure that captured
both landscape confi guration and composition of Bicknell ’ s
thrush habitat (Table 2). We defi ned patch size as the total
area of conifer-dominated forest (30 $ 30 m pixels) within a
4-EV
is determined the occupancy status of a site the following
year. Within a primary sampling period are surveys, or second-
ary sampling periods. Between surveys (or within a season), the
population is assumed to be closed. Two additional assumptions
are no false detections and detections at one site are indepen-
dent of another (MacKenzie et al. 2003). ese assumptions
are reasonable for this species as adults display high breeding
site fi delity (Rimmer et al. 2001), Bicknell ’ s thrushes are eas-
ily identifi ed by sound (majority of detections), and sampling
locations were spaced a minimum of 8 km apart.
We conducted the modeling analyses using the program
Presence (Hines 2006). Covariates relating to site-specifi c char-
acteristics, at the local and landscape scale, were included to test
the strength of their relationship to probability of initial site
occupancy ( y ) , c o l o n i z a t i o n ( g ) , a n d e x t i n c t i o n ( e ). Although
not the primary parameter of interest, detection probability ( p )
was included to correct for imperfect detection. We tested for
the eff ect of survey type, number (fi rst, second, or third), time,
date, and patch size on detection probability. Maximum likeli-
hood techniques were used to estimate the four parameters ( p ,
y , g , a n d e ) based on site detection histories with the following
likelihood equation (MacKenzie et al. 2003):
LX
ni
N
yeg
1,,, Pr( )pX X
i
11
,...
()
==
∏
Here, y 1 refers to the initial occupancy in the fi rst pri-
mary period, where thereafter e and g determine occupancy
in the following seasons. is is a Marcovian process in
which occupancy in a particular time step depends on occu-
pancy in the previous time step. X i
are the data in the form
of detection histories.
A t o t a l o f 6 6 m o d e l s w e r e c o m p a r e d u s i n g A I C m o d e l
selection procedures (Burnham and Anderson 2002). e
models consisted of all possible combinations of 1) local,
2) landscape, 3) local % l a n d s c a p e , a n d 4 ) l o c a l $ landscape
for occupancy, colonization, and extinction. Models were
ranked based on their AICc score (a small sample size adjust-
ment of AIC) and models with ∆ AICc of & 2 w e r e c o n -
sidered plausible. A test for model goodness-of-fi t was not
performed because no such test exists for dynamic occupancy
models or models that have missing surveys (D. MacKenzie
pers. comm.). We calculated relative variable importance
for 1) landscape, 2) local, 3) landscape % lo c a l , a n d 4 ) la n d -
scape $ local for occupancy, colonization, and extinction by
summing the weights (AICc ω i ) o f t h e m o d e l s i n w h i c h t h e y
appeared for the particular parameter in question (Burnham
and Anderson 2002).
Results
We found both landscape context and local habitat qual-
ity to be important in determining Bicknell ’ s thrush occu-
pancy patterns with both being included in the top eight
most plausible models (Table 3). e additive relationship of
landscape and local was the most supported model and had
the highest relative importance for all population parameters
(Table 3, Fig. 2). Landscape-scale characteristics alone had
little support (Table 3, Fig. 2). We detected a cross-scale
interaction and potential thresholds for both local and land-
scape variables (see resholds below).
given high-elevation land unit and used this as our measure
of confi guration (Betts et al. 2006). For landscape compo-
sition, we used patch isolation measured as the amount of
potential habitat surrounding the survey point (Fahrig 2003)
at three spatial extents (2-, 5-, and 10-km radii). While habi-
tat amount is not a direct measure of connectivity, such as
distance to nearest patch, habitat isolation is thought to be
best predicted by amount of surrounding habitat, which is
also more meaningful for conservation (Fahrig 2003). To
create one variable that captured the landscape-scale fea-
tures around each site (hereafter termed landscape) we used
PC1 from a second principal component analysis of the four
landscape-scale variables (as described above for local).
Dynamic occupancy modeling and parameter
estimation
Site encounter histories were created by compiling detec-
tions (1) and non-detections (0) from surveys conducted in
2006 and 2007. Missed surveys were not used in parameter
estimation (MacKenzie et al. 2003). Model input consisted
of encounter histories and the two covariates describing local
and landscape habitat characteristics.
A multi-season dynamic occupancy model framework
was used following MacKenzie et al. (2003). e dynamic
model estimates four parameters: 1) probability of detection
( p ), 2) probability of initial site occupancy ( y ), 3) probabil-
ity of site colonization ( g ), and 4) probability of local site
extinction (e ). Sites that were vacant in year t could become
colonized in year t % 1 or remain vacant. Sites that were
occupied in year t could become locally extinct in year t % 1
or remain occupied.
D y n a m i c o c c u p a n c y m o d e l s i n c o r p o r a t e p r i m a r y a n d s e c -
ondary time periods. In this study, the primary sampling period
was defi ned as a Bicknell ’ s thrush breeding season. Between
primary sampling periods movement could occur in and out of
the populations causing local extinction or colonization events.
Table 2. Local- and landscape-scale habitat variables used in this
study. The factor coeffi cients for each variable are listed for principal
component one (PC1) for the local and landscape variables. The
proportion of variance explained by PC1 for each scale is shown in
parentheses. Sites scoring high for local were conifer-dominated,
had a high coniferous shrub density, greater basal area of dead trees,
and shorter canopies. Low local scores were associated with open
understories, more deciduous vegetation, and taller canopies. Sites
with high landscape scores were located within larger patches sur-
rounded by higher amounts of habitat (conifer-dominated forest). A
low landscape score represents a site situated in a small patch in a
low-cover landscape.
Variable PC1 factor coeffi cient
Local (0.51)
Coniferous shrub density (stems m
– 2 ) 0.500
Dead basal area (m
2 ha
– 1 ) 0.400
Proportion of coniferous dominated forest
points along the transect
0.614
Average canopy height (m) '0.462
Landscape (0.82)
Patch size (ha) 0.511
Amount of habitat within 2 km (ha) 0.500
Amount of habitat within 5 km (ha) 0.531
Amount of habitat within 10 km (ha) 0.456
5-EV
structure in driving changes in patch occupancy (Table 4,
Fig. 4). Both landscape- and local-scale habitat features posi-
tively infl uenced colonization and had negative eff ects on
extinction (Fig. 4, Table 4). Compared to occupancy and
colonization, the infl uence both of landscape- and local-
scale habitat features on extinction was greater (Table 4).
is resulted in a narrow range of covariate values in which
extinction went from zero to one (Fig. 4c–d). Average colo-
nization and extinction probabilities ( " 1SD) across all sites
were 0.52 ( " 0.4) and 0.27 ( " 0.33), respectively. Note
the diff erence in rates compared to those based on the raw
detections (see Raw occurrences below) not accounting for
variation in detection probability (i.e. assuming p # 1).
Thresholds
All parameters, (with the exception of occupancy for local),
indicated a potential threshold value beyond which an
increase in local or landscape quality did not translate to
an increased probability of a site becoming or remaining
occupied (Fig. 3, 4). Although we did not explicitly test for
threshold values, there are clear qualitative delineations seen
in the relationships between the parameters and the indi-
vidual landscape-scale variables (Supplementary material
Appendix 1–3). Occupancy and colonization probabilities
are ∼ 1.0 and extinction rates are essentially 0 when patch
size approaches ∼ 600 ha. When the proportion of suitable
habitat in the surrounding landscape within 2 km reaches
0.35 (and reaches 0.10 within 5 km) occupancy and coloni-
zation rates are ∼ 1.0. Extinction probability drops to 0 when
the proportion of habitat reaches 0.25 and 0.08 within 2 and
5 km, respectively.
Detection probability
Detection probability ( p ) was best explained by survey
number and patch size (Table 1). Probability of detection
decreased with increasing number of surveys and increased
with patch size. Patch size had an equal positive eff ect on
detection probability across all surveys (Table 4). In the fi rst
two surveys, usually done during the same visit to the site,
detection probability varied little (0.77 to 0.95) and in large
patches ( ! 850 ha), probability of detection was ! 0.75,
Table 3. AICc model selection results for determining the effects of landscape and local-scale habitat covariates on initial occupancy ( y ),
colonization ( g ), and local extinction ( e ). K is the number of parameters estimated in the model. Each model is ranked by its AICc score, which
represents how well the model fi t the data. A lower ∆ AICc value is indicative of a better model. Only models within 2 AICc points of the top
model were considered plausible and are displayed. The probability that the model (of the models tested) would best explain the data is
indicated by the AICc weight ( ω
i ).
Model AICc ∆ AICc ω
i K
y (local % landscape) g (local % landscape ) e ( local % landscape)
p (survey % ps) 297.17 0 0.14 13
y (local $ landscape) g (local % landscape) e (local % landscape) p (survey % ps) 297.72 0.55 0.11 14
y (local % landscape) g (local % landscape) e (local ) p (survey % ps) 298.06 0.89 0.09 12
y (local % landscape) g (local) e (local % landscape) p (survey % ps) 298.40 1.23 0.08 12
y (local $ landscape) g (local % landscape
) e (local) p (survey % ps) 298.53 1.36 0.07 13
y ( local % landscape) g (local % landscape) e (local $ landscape) p (survey % ps) 298.55 1.38 0.07 14
y (local % landscape) g (local $ landscape) e (local % landscape) p (survey % ps) 299.14 1.97 0.05 14
y (local $ landscape) g (local % landscape) e (local $ landscape) p (survey % ps) 299.17 2.00 0.05 15
Figure 2. Relative variable importance (determined by summing
the AICc ω i for the models in which each covariate was present) for
1) local % landscape, 2) local $ landscape, 3) local, and 4) land-
scape by parameter (probability of initial occupancy, site coloniza-
tion and local site extinction).
Occupancy
ere was strong support for occupancy probability to
be driven by an interaction between scales. Specifi cally, if a
site contained moderate quality local habitat, an ideal land-
scape context was necessary (i.e. large patches within high-
cover landscapes, Fig. 3a) for it to be occupied. Conversely,
a site situated in a poor landscape context (i.e. small patches
within low-cover landscapes) would only be occupied if the
local habitat quality was high (Fig. 3b). e average site
occupancy probability ( " 1SD) was 0.58 ( " 0.31) in 2006
and 0.61 ( " 0.38) in 2007 based on estimates from the top
model (2007 occupancy estimate was calculated using Eq. 7
in MacKenzie et al. 2003).
Colonization and extinction
Local and landscape covariates were both important drivers
of site colonization and extinction (Table 3). However, the
eff ect of local-scale habitat was stronger relative to landscape
6-EV
Raw occurrences and principal
components analyses
Bicknell ’ s thrush overall patch occupancy remained relatively
consistent during the study period. Based on raw detection
data, 28 sites remained vacant (31.8%) and 38 remained
occupied (43.2%), whereas nine sites were colonized (raw
colonization rate # 24.3%) and six sites showed local extinc-
tion (raw extinction rate # 13.6%) between years (Fig. 5).
Landscape and local habitat scores were generally higher for
sites that remained occupied or became colonized than those
that stayed vacant or went extinct (Fig. 5).
Discussion
We found diff erential site occupancy by Bicknell ’ s thrush
due to an interaction between local and landscape scales. A
site with high-quality local habitat could be occupied even if
the landscape context was marginal (i.e. smaller patch and/or
low-cover landscape), suggesting that local site characteris-
tics can compensate for poor landscape context. Similarly, if
local habitat quality is low, it may be necessary for the site to
be in a large patch and/or surrounded by a high proportion
of suitable habitat to be occupied. ere is some evidence
that these cross-scale interactions may be common in birds
(Betts et al. 2006, Renfrew and Ribic 2008) and perhaps in
other taxa. A similar study in a naturally fragmented wet-
land system (Schooley and Branch 2007) identifi ed cross-
scale interactions between patch size, spatial connectivity,
and wetland quality as determinants of patch occupancy of
round-tailed muskrats Neofi ber alleni .
We found little support for eff ects of landscape-scale fea-
tures in driving metapopulation processes independent of
local habitat quality; both local habitat characteristics and
landscape-scale features played important roles in explain-
ing Bicknell ’ s thrush occupancy patterns in Vermont. In our
system, local habitat is of equal or greater importance than
landscape patterns for colonization and extinction rates.
F i g u r e 3 . E s t i m a t e d p r o b a b i l i t y o f initial site occupancy (adjusted
for imperfect detection) as a function of (a) local habitat PC1
score and (b) landscape PC1 score at each site from the top
ranked model: y (local % landscape ) g (local % landscape) e
(local % landscape) p ( survey % ps) . Parameter estimates were
categorized as low, moderate, and high quality landscape con-
text (a) or local habitat quality (b) to facilitate visualization of
the interaction between the local and landscape scales in deter-
mining site occupancy by Bicknell ’ s thrush in Vermont. Quality
categories for both local and landscape were delineated by divid-
ing the data range into thirds. is fi gure shows that in order for
a site with moderate quality local habitat to be occupied, the
landscape context must be ideal (i.e. large patches within high-
cover landscapes). On the other hand, for sites located in poor
landscape contexts to be occupied, they must contain very high
quality local habitat.
Table 4. Parameter estimates ( β ) and 95% confi dence intervals
(lower # LCI and upper # UCI) for probability of initial site occu-
pancy ( y ), site colonization ( g ), local site extinction ( e ) and detection
probability ( p ) for the most supported model ( y ( local % landscap e)
g
(local % landscape) e
(local % landscap e) p ( survey % p s )). Bicknell ’ s
thrush occupancy patterns were estimated as a function of local
habitat and landscape context. Detection probability was modeled
as a function of survey number and patch size.
Variable β LCI UCI
y intercept 0.822 0.057 1.587
local 0.797 0.270 1.324
landscape 0.776 0.164 1.388
g intercept 0.798 ' 1.289 2.885
local 1.693 0.322 3.063
landscape 1.057 ' 0.270 2.385
e intercept ' 3.519 ' 6.307 ' 0.732
local ' 1.979 ' 3.791 ' 0.167
landscape ' 1.440 ' 3.345 0.465
p intercept survey 1 1.874 1.234 2.514
intercept survey 2 1.459 0.733 2.185
intercept survey 3 0.177 ' 0.618 0.972
patch size 0.375 ' 0.131 0.880
regardless of the survey number. e third survey was gener-
ally done either at a later date in the breeding season or later
in the morning (after an earlier survey), which may explain
the lower detection probability. Essentially, if a Bicknell ’ s
thrush occupied a site, the chances of detecting it on the fi rst
survey were 83% or greater, regardless of patch size.
7-EV
Figure 4. Predicted probability of site colonization (a and b) and local extinction (c and d) as a function of a site ’ s local habitat PC1 score
(a and c, fi lled circles) and landscape PC1 score (b and d, open circles) for Bicknell ’ s thrush in Vermont. All estimates were derived from the top
ranked model ( y ( l o c a l % l a n d s c a p e ) g (local % l a n d s c a p e ) e ( l o c a l % l a n d s c a p e ) p ( s u r v e y % ps) ) and were adjusted for imperfect detection.
Figure 5. Comparison of means ( " 1 SE) for local and landscape
principal component one (PC1) scores for sites that remained occu-
pied (11), remained vacant (00), were colonized (01), or went
extinct (10) during the 2006 and 2007 Bicknell ’ s thrush breeding
seasons. Sample sizes are displayed in brackets above mean values
for each detection history. e landscape and local habitat PC1
scores were signifi cantly, positively correlated (r # 0.51, p ( 0.001).
However, we considered this correlation to be moderate and there-
fore acceptable (Ott and Longnecker 2001) for including both in
the model set to test for eff ects of scale and to explore interactions.
Patch size and habitat amount play a key role in changes
in patch occupancy, but persistence at a site requires that
the proper local habitat conditions exist. ese results sug-
gest that local and landscape scales act together to infl uence
occupancy patterns and motivate habitat selection for this
vulnerable species.
eoretical work states that landscape structure is of pri-
mary importance in driving metapopulation dynamics in
patchy environments (MacArthur and Wilson 1967, Hanski
and Ovaskainen 2003), although it has also been suggested
that habitat quality of a patch likely infl uences extinction
probability through birth and death rates of a subpopulation
(Opdam 1991). However, incorporating local habitat qual-
ity does not always improve model fi t (Moilanen and Hanski
1998) and landscape features may be more important in
some systems (Vogeli et al. 2010). Our study contributes to
the growing body of literature highlighting the importance
of incorporating both the local patch scale and landscape
context when modeling species distributions (Mortelliti
et al. 2010). Collectively, both local and landscape scales have
been found to be important in metapopulation dynamics
for a variety of taxa including insects ( omas et al. 2001),
mammals (Schooley and Branch 2009), and birds (Verboom
et al. 1991).
Identifying thresholds in landscape context at which
immigration rates are no longer suffi cient to maintain con-
nectivity is useful when making conservation decisions.
From a landscape perspective, Bicknell ’ s thrush appears to be
able to persist in landscapes with relatively low proportions
of suitable habitat (0.10–0.35), depending on landscape
extent. ese fi ndings are similar to thresholds determined
for songbirds in anthropogenically-fragmented landscapes
8-EV
(Andren 1994, Betts et al. 2007). resholds in local habitat
conditions (Guenette and Villard 2005) also likely exist for
Bicknell ’ s thrush (Fig. 4, 5).
Many studies assessing the eff ect of forest patch size
and isolation on bird population processes have been con-
ducted within predominantly human-modifi ed landscapes
(Andren 1994, Prugh et al. 2008). In these landscapes,
connectivity may be much more signifi cant for population
persistence or occupancy of small patches (Donovan et al.
1995) because species that occupy these areas have been
recently separated by habitat fragmentation and may not
have the capabilities to move between fragmented habitat
patches or persist in small fragments. It is possible that the
degree of patch isolation may be less important to birds
with strong dispersal capabilities (With et al. 2006) and
naturally fragmented ranges such as the Bicknell ’ s thrush
(Rimmer et al. 2001). Consequently, this may explain why
our results show that landscape context is less important
than local habitat quality for colonization and extinction
dynamics for this species.
To ensure persistence of Bicknell ’ s thrush, our results
indicate that it is imperative that local-scale habitat char-
acteristics are considered in addition to landscape context.
Small, isolated patches are only likely to be occupied when
they contain optimal local conditions, while large, con-
nected patches without suffi cient local habitat quality may
be unsuitable.
Dynamic occupancy modeling provides an eff ective
approach to incorporate the quality of both local habitat
and landscape features into metapopulation dynamics for
species with populations in heterogeneous environments.
ese methods provide unbiased estimates of coloniza-
tion and extinction rates through the incorporation of
detection probabilities. Simple detection/non-detection
information can be easily gathered over large spatial scales
making this an effi cient and useful method for assessing
distributional changes with respect to land use or cli-
mate change. For example, Mustin et al. (2009) showed
through simulation that patch colonization and extinction
rates infl uenced the rate at which populations shift their
ranges in order to track changes in climate. Our study
also points out that range changes are unlikely to be con-
stant throughout the range due to uneven colonization
and extinction rates across gradients, such as in habitat
quality or landscape pattern. Information regarding the
responses of species to factors at multiple scales is useful
for parameterizing dynamic models aimed at predicting a
species response to environmental change (Midgley et al.
2010) and will likely result in more accurate predictions
of distributional changes.
Acknowledgements – We are grateful to the Mountain Birdwatch
coordinators, J. Hart and D. Lambert, and all of the MBW vol-
unteers who headed to the mountains in their free time to survey
for Bicknell ’ s thrush. We highly appreciate fi eld assistance pro-
vided by J. Juillerat, H. Slongo, J. Klavins, and K. Pindell and
recognize the demanding nature of the work. A. S. Hadley pro-
vided insightful and thoughtful comments that greatly improved
this manuscript. Financial support was provided by a grant from
the USDA Forest Service McIntire-Stennis Cooperative Forestry
Research Program.
References
Andren, H. 1994. Eff ects of habitat fragmentation on birds and
mammals in landscapes with diff erent proportions of suitable
habitat: a review. – Oikos 71: 355 – 366.
Armstrong, D. P. 2005. Integrating the metapopulation and habi-
tat paradigms for understanding broad-scale declines of species.
– Conserv. Biol. 19: 1402 – 1410.
Betts, M. G. et al. 2006. Independent eff ects of fragmentation on
forest songbirds: an organism-based approach. – Ecol. Appl.
16: 1076 – 1089.
B e t t s , M . G. et al. 2007. resholds in songbird occurrence in rela-
tion to landscape structure. – Conserv. Biol. 21: 1046 – 1058.
BirdLife International 2000. reatened birds of the world. – Cam-
bridge and Lynx Editcions.
Burnham, K. P. and Anderson, D. R. 2002. Model selection and
multimodel inference: a practical information-theoretic
approach. – Springer.
Cottam, G. and Curtis, J. 1956. e use of distance measures in
phytosociological sampling. – Ecology 37: 451 – 460.
Donovan, T. et al. 1995. Reproductive success of migratory birds
in habitat sources and sinks. – Conserv. Biol. 9: 1380 – 1395.
ESRI 2005. ArcGIS, ver. 9.1. – Environmental Systems Research
Inst., Redlands, CA.
Fahrig, L. 2003. Eff ects of habitat fragmentation on biodiversity.
– Annu. Rev. Ecol. Evol. Syst. 34: 487 – 515.
Fleishman, E. et al. 2002. Assessing the roles of patch quality, area,
and isolation in predicting metapopulation dynamics. – Con-
serv. Biol. 16: 706 – 716.
Gaston, K. J. 1990. Patterns in geographical ranges in species.
– Biol. Rev. Camb. Philos. Soc. 65: 105 – 129.
Guenette, J. S. and Villard, M. A. 2005. resholds in forest bird
response to habitat alteration as quantitative targets for conser-
vation. – Conserv. Biol. 19: 1168 – 1180.
Hanski, I. 1998. Metapopulation dynamics. – Nature 396: 41 – 49.
Hanski, K. and Ovaskainen, O. 2003. Metapopulation theory for
fragmented landscapes. – eor. Popul. Biol. 64: 119 – 127.
Hart, J. A. and Lambert, J. D. 2007. Mountian birdwatch 2006:
fi nal report to the US Fish and Wildlife Service. – Vermont
Inst. of Natural Science Technical Report 07 - 03.
Hastings, A. and Harrison, S. 1994. Metapopulation dynamics and
genetics. – Annu. Rev. Ecol. Syst. 25: 167 – 188.
Hines, J. E. 2006. Presence2 – software to estimate patch occu-
pancy and related parameters. – USGS, Patuxent Wildlife
Research Center, Laurel, MD, USA.
Lambert, J. D. et al. 2005. A practical model of Bicknell ’ s thrush
distribution in the northeastern United States. – Wilson Bull.
117: 1 – 11.
MacArthur, R. H. and Wilson, E. O. 1967. e theory of island
biogeography. – Princeton Univ. Press.
MacKenzie, D. et al. 2003. Estimating site occupancy, coloniza-
tion, and local extinction when a species is detected imper-
fectly. – Ecology 84: 2200 – 2207.
Midgley, G. F. et al. 2010. BioMove – an integrated platform
simulating the dynamic response of species to environmental
change. – Ecography 33: 612 – 616.
Moilanen, A. 2002. Implications of empirical data quality to
metapopulation model parameter estimation and application.
– Oikos 96: 516 – 530.
Moilanen, A. and Hanski, I. 1998. Metapopulation dynamics:
eff ects of habitat quality and landscape structure. – Ecology 79:
2503 – 2515.
Mortelliti, A. et al. 2010. e role of habitat quality in fragmented
landscapes: a conceptual overview and prospectus for future
research. – Oecologia 163: 535 – 547.
Mustin, K. et al. 2009. e dynamics of climate-induced range
shifting; perspectives from simulation modelling. – Oikos 118:
131 – 137.
9-EV
Supplementary material (Appendix E6936 at ( www.
oikosoffi ce.lu.se/appendix ! ). Appendix 1 – 3.
Opdam, P. 1991. Metapopulation theory and habitat fragmenta-
tion: a review of holarctic breeding bird studies. – Landscape
Ecol. 5: 93–106.
Ott, R. L. and Longnecker, M. 2001. An introduction to statistical
methods and data analysis. – Duxbury.
Prugh, L. R. et al. 2008. Eff ect of habitat area and isolation on
fragmented animal populations. – Proc. Natl Acad. Sci. USA
105: 20770 – 20775.
Renfrew, R. B. and Ribic, C. A. 2008. Multi-scale models of grass-
land passerine abundance in a fragmented system in Wiscon-
sin. – Landscape Ecol. 23: 181 – 193.
Rich, T. D. et al. 2004. Partners in Flight North American Land-
bird Conservation Plan. – Cornell Lab of Ornithology.
Rimmer, C. C. et al. 2001. Bicknell ’ s thrush ( Catharus bicknelli ).
– In: Poole, A. and Gill, F. (eds), e Birds of North America.
e Birds of North America.
Schooley, R. L. and Branch, L. C. 2007. Spatial heterogeneity in
habitat quality and cross-scale interactions in metapopulations.
– Ecosystems 10: 846 – 853.
S c h o o l e y , R. L. and Branch, L. C. 2009. Enhancing the area-isolation
paradigm: habitat heterogeneity and metapopulation dynamics
of a rare wetland mammal. – Ecol. Appl. 19: 1708 – 1722.
Sprugel, D. G. 1976. Dynamic structure of wave-regenerated Abies
balsamea f o r e s t s i n t h e n o r t h - e a s t e r n U n i t e d S t a t e s . – J . E c o l .
64: 889 – 911.
Strong, A. M. et al. 2004. Eff ect of prey biomass on reproduc-
tive success and mating strategy of Bicknell ’ s thrush
( Catharus bicknelli ), a polygynandrous songbird. – Auk
121: 446 – 451.
omas, C. D. 2000. Dispersal and extinction in fragmented land-
scapes. – Proc. R. Soc. B 267: 139 – 145.
omas, J. A. et al. 2001. e quality and isolation of habitat
patches both determine where butterfl ies persist in fragmented
landscapes. – Proc. R. Soc. B 268: 1791 – 1796.
Verboom, J. et al. 1991. European nuthatch metapopulations
in a fragmented agricultural landscape. – Oikos 61:
149 – 156.
Vogeli, M. et al. 2010. e relative importance of patch habitat
quality and landscape attributes on a declining steppe-bird
metapopulation. – Biol. Conserv. 143: 1057 – 1067.
Vogelmann, J. E. et al. 2001. Completion of the 1990s National
Land Cover Data set for the conterminous United States from
Landsat ematic Mapper data and ancillary data sources.
– Photogramm. Eng. Remote Sens. 67: 650 – 662.
With, K. et al. 2006. e implications of metalandscape connectiv-
ity for population viability in migratory songbirds. – Landscape
Ecol. 21: 157 – 167.
With, K. A. et al. 1997. Landscape connectivity and population
distributions in heterogeneous environments. – Oikos 78:
151 – 169.
