Effects of scale on predictive power of two bald eagle habitat models

Abstract

We examined the role scale plays in determining the predictive power of bald eagle (Haliaeetus leucocephalus) habitat models. We used a bald eagle roost habitat database that included 35 roost sites and 123 random sites located and characterized on the Chesapeake Bay from 1985-1988. A micro-habitat model, based on 6 micro-scale variables, correctly classified 80% of the roost sites. A macro-habitat model, based on 10 macro-scale variables, correctly classified only 63% of the roost sites. A mixed model, incorporating the significant micro- and macro-scale variables, correctly classified 89% of the roost sites. Our results suggest there is a tradeoff between model performance (predictive power), model development costs, and model application.

Full text

Effects
of
Scale on Predictive Power
of
Two Bald
Eagle Habitat Models
David A. Buehler,1 Department of Fisheries and Wildlife
Sciences, Virginia Polytechnic Institute and State University,
Blacksburg, VA 24061
James D. Fraser, Department of Fisheries and Wildlife, Virginia
Polytechnic Institute and State University, Blacksburg, VA
24061
Janis K. D. Seegar, Chemical Research, Development, and
Engineering
Center,
U.S.
Army,
Aberdeen Proving Ground,
MD 21010
Abstract: We examined the role scale plays
in
determining the predictive power
of
bald
eagle (Haliaeetus leucocephalus) habitat models.
We
used
a
bald eagle roost habitat
database that included 35 roost sites and 123 random sites located and characterized on
the Chesapeake Bay from 1985-1988.
A
micro-habitat model, based
on 6
micro-scale
variables correctly classified 80%
of
the roost sites.
A
macro-habitat model, based
on
10 macro-scale variables, correctly classified only 63%
of the
roost sites.
A
mixed
model, incorporating the significant micro- and macro-scale variables, correctly classi-
fied
89%
of the
roost sites.
Our
results suggest there
is a
tradeoff between model
performance (predictive power), model development costs,
and
model application.
Proc. Annu. Conf. Southeast. Assoc. Fish
and
Wildl. Agencies 46:266-273
Wildlife species select habitat through
a
hierarchy
of
decisions that start
at a
geographic scale and continue to finer scales until an individual decides to perch on
a
particular branch
on a
tree
or
rest
in a
given thicket (Johnson 1980).
One
goal
of
wildlife-habitat models
is to
document
the
relationships that exist
at
these various
scales.
A
second goal
is to
develop the model
in
such
a
way that makes
it
useful
for
management (Salwasser 1986), either
by
predicting where suitable habitat exists
or
by predicting
how
management actions will affect habitat suitability. Wildlife-
habitat models
can be
developed
by
researchers
at any
scale
to
meet
the
first goal.
Model utility
(the
second goal), however, favors model development
at the
same
1
Present address: Department
of
Forestry, Wildlife, and Fisheries, University of Tennessee, Knox-
ville,
TN
37901-1071.
1992 Proc. Annu.
Conf.
SEAFWA

Bald Eagle Habitat Models 267
scale as management operation. Although it is possible to develop models that
incorporate several different scales simultaneously, economics may limit the num-
ber of scales that can be developed.
In general, models based on micro-scale habitat variables that require field
mensuration tend to be more expensive to develop and apply than models based on
macro-scale habitat variables measured in a remotely-sensed fashion (e.g., from
aerial photos or satellite imagery). Macro-scale models also are becoming more
economical to develop and apply as national and regional databases become more
widely available. Therefore, model builders tend to opt for development of
the
more
economical macro-scale wildlife-habitat models.
The selection of a scale for model development must in part be determined by
the biology of the wildlife species under study. Habitat selection by some species,
such as small mammals, is determined primarily by micro-scale habitat features
(Dueser and Shugart 1979, Healy and Brooks 1988), such that macro-scale habitat
models are unlikely to result in high predictive power.
Selection of scale is influenced also by the planned management application.
Management may be conducted at a fairly broad, geographic scale. Model develop-
ment at this scale is desirable so that model results can be directly incorporated into
the management scheme.
The goal of
this
study was to compare the predictive power of
a
bald eagle roost
habitat model based on micro-scale variables with a roost habitat model based on
macro-scale variables. We also discuss the tradeoffs in final model selection for
management application.
The U.S. Army Chemical Research, Development, and Engineering Center
funded this study. A. B. Jones assisted with ARC/INFO analyses. We thank E. M.
Adams, L. L. Arnold, A. K. DeLong, D. C. DeLong, Jr., D. W. Leidlich, M.
Roeder, J. M. Seegar, and T. L. Weller for assisting with fieldwork. We thank E. A.
Kascenska, C. P. Campbell, L. Serlin, K. K. Stout, and J. M. T. Cazell for
computerizing the data and digitizing maps. We thank H. Galbreath for allowing
access to Remington Farms. We thank R. B. Owen and an anonymous person for
reviews of the manuscript.
Methods
The study area was the northern Chesapeake Bay, extending from the Bay
Bridge at Annapolis, Maryland, northward to the Conowingo Dam on the Sus-
quehanna River, a distance of 3,426 km. The area included 2,472 km of bay, river,
and creek shoreline and much of the Baltimore metropolitan area. Forested habitat
ranged from bottomland hardwood forests on the western shore to mixed pine (Pinus
spp.)-hardwood forests on the eastern shore. The Baltimore area on the western
shore was highly developed, whereas the eastern shore consisted of agricultural land
with an interspersion of small woodlots.
We located 35 roost sites from 1985-1988 by tracking radio-tagged bald eagles
from the late afternoon until they roosted in the evening (Buehler et al. 1991a). We
1992 Proc. Annu.
Conf.
SEAFWA

268 Buehler et al.
also randomly selected 123 trees and sites from throughout the study area for com-
parison with the roost-site habitat.
A roost site was defined as the area enclosed by a minimum convex polygon
connecting all perimeter trees in which we observed eagles roosting. Roost sites
varied from a single-tree site used on only
1
occasion to communal sites involving
many roost trees used traditionally year after year (Buehler et al. 1991a).
We defined micro-scale variables as habitat measurements that had to be made in
the field, whereas macro-scale variables were habitat measurements made on aerial
photos, standard
U.S.
Geological Survey (USGS) topographic maps, or derived from
the USGS Land Use and Land Cover (LULC) database (Anderson et al. 1976).
We measured 6 micro-scale variables for each roost site. We measured roost
tree diameter at breast height (dbh), and we measured roost tree height and surround-
ing canopy height. We estimated roost tree accessibility to an eagle as the total arc
(0°-360°) that was unobstructed by other tree canopies for a distance of 10 m out
from the trunk and 3 m below the tree's crown. We noted the presence of snags in
11.3 m-radius circular plots centered on each roost tree. We defined snag presence at
each roost site as the percent of roost tree plots that contained at least 1 snag. We
counted the number of trees >10-cm dbh on each circular plot and calculated tree
density as the number of
trees
per plot divided by the plot area (0.04
ha).
Because we
considered each roost site the basic sampling unit, we averaged values for tree dbh,
tree height, canopy height, access, and tree density across all roost trees at a roost
site to generate a mean value for each variable and each roost site. These mean
values were used to develop the logistic-regression models. Similar measurements
were made on each randomly-selected tree.
We measured 10 macro-scale variables for each roost site located. We digitized
all shoreline boundaries using USGS 7.5-min topographic maps and we digitized all
building using 1985
1:12,000
color aerial photos. We used ARC/INFO (Environ.
Systems Res. Inst., Inc., Redlands, Calif.) to measure the distance on the digitized
maps from each roost site to the nearest water of any kind, the Chesapeake Bay,
rivers, creeks, ponds, buildings, and roads. To calculate building density at each
site,
we used ARC/INFO to count all buildings within 500 m of each site and divided
by the area (78.5 ha). We used ARC/INFO to overlay the site location onto the
USGS LULC database (Anderson et al. 1976) to identify land cover at each site, and
to measure the distance from the site to the nearset USGS LULC habitat edge.
We defined a randomly-selected site as an 11.3 m-radius circular plot, centered
on each randomly-selected tree. We made similar measurements on these sites as
were made on the roost sites.
We used the LOGIST procedure (SAS Inst., Inc. 1986) to develop logistic
regression models, with the micro- and macro-scale habitat variables serving as the
independent variables and the roost-site classification (roost or random) serving as
the dependent variable in the models.
To determine the classification accuracy of each model, we used a modified-
jackknife procedure (Meyer et al. 1986). We wanted to compare the classification
accuracy for roost sites with the classification accuracy for random sites. Because
1992 Proc. Annu.
Conf.
SEAFWA

Bald Eagle Habitat Models 269
the number of roost sites sampled (N = 35) was much less than that for random sites
(N = 123), we used all roost sites in determining model classification accuracy for
roosts, whereas we randomly selected with replacement 35 random sites to test the
model classification accuracy for random sites. In both cases, we used the jackknife
approach of withholding 1 observation, developing the model, and using the
withheld observation to test the model. We used the number of actual roost sites that
were correctly classified as roost sites by each model as a measure of the predictive
power of the model. We selected the significant variables in each model (P
=£
0.05)
to develop a combined model to determine the upper limit on the predictive capa-
bility of our modeling effort.
Results
The micro-scale variables were significant predictors of the roost-site classifica-
tion (P < 0.001) (Table 1). Three variables (tree height, the presence of snags, and
tree access) were significantly related to the roost classification in the logistic regres-
sion (P < 0.05), whereas tree dbh, canopy height, and tree density were not (P >
0.05). Based on this model, 28 of 35 roost sites (80%) and 32 of 35 random sites
(91%),
respectively, were correctly classified.
The macro-scale variables also were significant predictors of the roost-site
classification (P < 0.001) (Table 2). Five variables (land cover type, distance to the
Chesapeake Bay, distance to ponds, distance to water of any type, and building
density) were significantly related to the roost classification in the logistic regression
model (P < 0.05). Based on this model, 22 of 35 roost sites (63%) and 32 of 35
random sites (91%), respectively, were correctly classified.
Table 1. Logistic regression model based on micro-scale
explanatory variables of bald eagle roost habitat, Chesapeake
Bay, Maryland, 1985-1988.
Explanatory
variable
Tree height
Snags present
Tree access
Tree dbh
Canopy height
Trees/ha
Observed
Parameter
estimate
0.344
0.089
0.015
0.041
0.122
0.001
Parameter
SE
0.109
0.030
0.005
0.022
0.078
0.001
Classification Table
Roost
Random
Total
Roost
28
3
31
10.03
8.64
7.70
3.55
2.44
0.73
Predicted
Random
7
32
39
P
0.001
0.003
0.006
0.059
0.119
0.392
Total
35
35
70
1992 Proc. Annu.
Conf.
SEAFWA

Table 2. Logistic regression model based on macro-scale
explanatory variables of bald eagle roost habitat, Chesapeake Bay,
Maryland, 1985-1988.
Explanatory
variable
Land cover type
Distance
to Bay
Distance
to
pond
Distance
to
water
Building density
Distance
to
creek
Distance
to
building
Distance
to
river
Distance
to
edge
Distance
to
road
Observed
Parameter
estimate
2.848
-0.171
-1.122
-2.655
-0.188
0.174
-0.582
0.051
-3.119
0.247
Parameter
SE
0.847
0.076
0.513
1.246
0.093
0.109
0.629
0.073
5.028
0.801
Classification Table
Roost
Random
Total
Roost
22
3
25
X*
11.30
5.04
4.79
4.54
4.08
2.56
0.86
0.49
0.38
0.10
Predicted
Random
13
32
45
P
0.001
0.025
0.029
0.033
0.043
0.110
0.355
0.485
0.535
0.757
Total
35
35
70
Table 3. Logistic regression model based on micro-scale and
macro-scale explanatory variables of bald eagle roost habitat,
Chesapeake Bay, Maryland, 1985-1988.
Explanatory
variable
Tree height
Tree access
Land cover type
Distance
to
water
Distance
to
pond
Snags present
Building density
Distance
to Bay
Observed
Parameter
estimate
0.421
0.013
2.828
-3.749
-1.326
0.044
-0.088
-0.055
Parameter
SE
0.107
0.005
1.334
1.902
0.698
0.034
0.727
0.114
Classification Table
Roost
Random
Total
Roost
31
1
32
15.61
5.64
4.50
3.89
3.61
1.66
1.47
0.23
Predicted
Random
4
34
38
P
0.001
0.018
0.034
0.049
0.058
0.198
0.225
0.632
Total
35
35
70
1992 Proc. Annu.
Conf.
SEAFWA

Bald Eagle Habitat Models 271
The 8 significant micro- and macro-variables combined also were significant
predictors of the roost-site classification (P < 0.001) (Table
3).
Although all variables
included in this model were significant in the original models, only 4 variables (tree
height, tree access, land cover type, and distance to water), were significantly
related to the roost classification in the combined logistic regression model (P <
0.05). Based on this combined model, 31 of 35 roost sites (89%) and 34 of 35
random sites (97%), respectively, were correctly classified.
Discussion
Our results suggest that bald eagle roost habitat models may lose predictive
power as the scale becomes larger and that a combination of
micro-
and macro-scale
variables ultimately lead to the most parsimonious model (fewest parameters) with
the greatest predictive power.
Bald eagles may be more discriminating in roost habitat selection at a micro
scale. Bald eagle roosts on the Chesapeake Bay and elsewhere provide protected
environments where eagles can avoid buffeting by prevailing winds (Stalmaster and
Gessaman 1984, Keister et al. 1985, Buehler et al. 1991fc). These habitat characteris-
tics may be recognizable only to an eagle at a micro scale. If true, there should be
greater difference between described eagle habitat at this scale and what is available
at random, such that micro-habitat models would be able to discriminate between
roost and random sites more easily.
Livingston et al. (1990), however, developed macro-scale bald eagle nesting
habitat models in Maine with correct classification rates ranging from
75%
to 100%.
Although inclusion of micro-scale variables may have improved classification accu-
racy in some of Livingston et al.'s models, overall accuracy of their macro-scale
models appears to be very good. The inclusion of 39 variables in the original
modeling effort by Livingston et al. (1990), suggests selection of an appropriate
array of variables may be as important as scale in determining model performance
and development and application efficiency. Macro-scale variable measurement
may be more economical on a per variable basis. The point where time spent
measuring a large number of macro-scale variables on aerial photos or maps is more
efficient than time spent in the field measuring a more limited set of micro-scale
variables is yet to be determined. Care needs to be taken in the variable selection
phase of model development to ensure that variables strongly related to actual
habitat selection are included.
Our results suggest that a combined micro-macro approach to habitat modeling
that may be consistent with the species' actual hierarchical habitat selection process,
may lead to the best wildlife-habitat models. We envision eagles making gross
habitat selection decisions consistent with the scale at which the USGS LULC data-
base was developed (1:250,000), before focusing in on individual tree characteris-
tics to select an actual roost site. Using the USGS LULC database to identify gross
land cover and measuring 2 simple tree characteristics (height and access) appears to
mimic this habitat selection process and lead to accurate classification of roost sites.
1992 Proc. Annu.
Conf.
SEAFWA

272 Buehler
et al.
Finally, about
3%
of
random sites were classified
as
roost sites
in the
combined
model. These sites could represent either true roost sites that
we did not
locate
or
potential roost habitat.
In
either case,
it
would
be
important
to
consider including
these sites when developing roost habitat management plans because they
are
struc-
turally similar
to
actual roosts.
Management Implications
Managers
are
faced with
a
series
of
decisions involving tradeoffs when identi-
fying
the
scale
and
variables
to be
used
in
developing wildlife-habitat models
for a
specific management area. Managers want optimal model accuracy
but
need
eco-
nomically efficient models that
are
applicable across
the
management area
(Sal-
wasser 1986).
A
priori selection
of
important variables
is
difficult, such that extra-
neous variables
may
have
to be
measured
to
identify
the
minimum
set of
variables
needed
for
desired model accuracy.
In
some cases, such
as in the
bald eagle roost
habitat model example, these desires
are in
conflict.
A
combination
of a
limited
number
of
micro-
and
macro-scale variables
can be
1 solution.
In addition, inclusion
of
micro-scale variables
in the
model development phase
may identify variables that could
be
estimated
by
macro-scale variables
in the
model
application phase.
For
bald eagle habitat,
it is
possible
to
estimate tree height
and
tree access from aerial photos (Howard 1970).
If
these parameters
are
estitmated
accurately,
a
macro-scale model could
be
developed that
has
excellent predictive
power
and
broad applicability.
Literature Cited
Anderson,
J. R., E. E.
Hardy,
J. T.
Roach,
and R. E.
Witmer,
1976. A
land
use and
land
cover classification system
for
use with remote sensor data. U.S. Geol. Surv.
Prof. Pap.
964.
28pp.
Buehler,
D. A., T. J.
Mersmann,
J. D.
Fraser, and
J. K. D.
Seegar. 1991a. Nonbreeding bald
eagle communal
and
solitary roost behavior
and
roost habitat
on the
northern Chesa-
peake
Bay. J.
Wildl. Manage. 55:273-281.
,
, , and —.
19916. Winter microclimate
of
bald eagle roosts
on
the northern Chesapeake
Bay. Auk
108:612-618.
Dueser,
R. D. and H. H.
Shugart,
Jr. 1979.
Niche pattern
in a
forest-floor small-mammal
fauna. Ecology 60:108-118.
Healy,
W. M. and R. T.
Brooks.
1988.
Small mammal abundance
in
northern hardwood
stands
in
West Virginia.
J.
Wildl. Manage. 52:491-496.
Howard,
J. A.
1970. Aerial photo-ecology.
Am.
ElsevierPubl.
Co., New
York, N.Y. 325pp.
Johnson,
D. H.
1980.
The
comparison
of
usage
and
availability measurements
for
evaluating
resource preference. Ecology
61:65-71.
Keister,
G. P., Jr., R. G.
Anthony,
and H. R.
Holbo. 1985.
A
model
of
energy consumption
in bald eagles:
an
evaluation
of
night communal roosting. Wilson Bull. 97:148-160.
Livingston,
S. A., C. S.
Todd,
W. B.
Krohn,
and R. B.
Owen,
Jr. 1990.
Habitat models
for
bald eagles nesting
in
Maine.
J.
Wildl. Manage. 54:644-653.
1992 Proc. Annu.
Conf.
SEAFWA

Bald Eagle Habitat Models 273
Meyer, J. S., C. G. Ingersoll, L. L. McDonald, and M. S. Boyce. 1986. Estimating uncer-
tainty in population growth models: jackknife versus bootstrap techniques. Ecology
67:1156-1166.
Salwasser, H. 1986. Modeling habitat relationships of terrestrial vertebrates—the manager's
viewpoint. Pages 419-424 in J. Verner, M. L. Morrison, and C. J. Ralph, eds. Wildlife
2000.
Modeling habitat relationships of terrestrial vertebrates. Univ. Wis. Press, Mad-
ison.
SAS Institute, Inc. 1986. SUGI supplemental user's guide. Version 5 ed. SAS Inst., Inc.,
Cary, N.C. 662pp.
Stalmaster, M. V. and J. A. Gessaman. 1984. Ecological energetics and foraging behavior of
overwintering bald eagles. Ecol. Monogr. 54:407^28.
1992 Proc. Annu.
Conf.
SEAFWA