See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/23304213
Fine-scale habitat modeling of a top marine predator: Do prey data improve
predictive capacity?
Article in Ecological Applications · November 2008
DOI: 10.1890/07-1455.1 · Source: PubMed
CITATIONS
READS
185
1,097
3 authors:
Leigh G. Torres
Andrew Read
Oregon State University
Duke University
114 PUBLICATIONS 3,444 CITATIONS
251 PUBLICATIONS 15,429 CITATIONS
SEE PROFILE
Patrick Neal Halpin
Duke University
178 PUBLICATIONS 11,305 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Whale baleen View project
OPAL: Overlap Predictions About Large whales View project
All content following this page was uploaded by Leigh G. Torres on 26 March 2019.
The user has requested enhancement of the downloaded file.
SEE PROFILE
Ecological Applications, 18(7), 2008, pp. 1702–1717
Ó 2008 by the Ecological Society of America
FINE-SCALE HABITAT MODELING OF A TOP MARINE PREDATOR:
DO PREY DATA IMPROVE PREDICTIVE CAPACITY?
LEIGH G. TORRES,1,3 ANDREW J. READ,1
1
AND
PATRICK HALPIN2
Duke University Marine Laboratory, Nicholas School of the Environment and Earth Sciences, 135 Duke Marine Lab Road, Beaufort,
North Carolina 28516 USA
2
Duke University Marine Geospatial Ecology Laboratory, Nicholas School of the Environment and Earth Sciences, Box 90328,
Durham, North Carolina 27708 USA
Abstract. Predators and prey assort themselves relative to each other, the availability of
resources and refuges, and the temporal and spatial scale of their interaction. Predictive
models of predator distributions often rely on these relationships by incorporating data on
environmental variability and prey availability to determine predator habitat selection
patterns. This approach to predictive modeling holds true in marine systems where
observations of predators are logistically difficult, emphasizing the need for accurate models.
In this paper, we ask whether including prey distribution data in fine-scale predictive models of
bottlenose dolphin (Tursiops truncatus) habitat selection in Florida Bay, Florida, USA,
improves predictive capacity. Environmental characteristics are often used as predictor
variables in habitat models of top marine predators with the assumption that they act as
proxies of prey distribution. We examine the validity of this assumption by comparing the
response of dolphin distribution and fish catch rates to the same environmental variables.
Next, the predictive capacities of four models, with and without prey distribution data, are
tested to determine whether dolphin habitat selection can be predicted without recourse to
describing the distribution of their prey. The final analysis determines the accuracy of
predictive maps of dolphin distribution produced by modeling areas of high fish catch based
on significant environmental characteristics. We use spatial analysis and independent data sets
to train and test the models. Our results indicate that, due to high habitat heterogeneity and
the spatial variability of prey patches, fine-scale models of dolphin habitat selection in coastal
habitats will be more successful if environmental variables are used as predictor variables of
predator distributions rather than relying on prey data as explanatory variables. However,
predictive modeling of prey distribution as the response variable based on environmental
variability did produce high predictive performance of dolphin habitat selection, particularly
foraging habitat.
Key words: bottlenose dolphin; environmental variability; Florida Bay, Florida, USA; generalized
additive models; habitat selection; Mantel’s tests; predator–prey dynamics; predictive distribution models;
spatial scale; Tursiops truncatus.
INTRODUCTION
Modeling species distribution is a valuable tool of
biological conservation efforts, especially predictive
models of marine predators due to the logistical
difficulties of monitoring their distributions at sea. For
instance, managers of whale and dolphin populations
can benefit from accurate model-derived predictions of
cetacean habitat to mitigate anthropogenic effects, such
as fisheries by-catch (Torres et al. 2003), sonar (Cox et
al. 2006), and the impacts of habitat alterations on
ecosystem function (Baumgartner et al. 2000, D’Amico
et al. 2003), in order to protect critical habitat (Hooker
Manuscript received 5 September 2007; revised 23 January
2008; accepted 14 February 2008. Corresponding Editor: K.
Stokesbury.
3 Present address: National Institute of Water and
Atmospheric Research (NIWA), 301 Evans Bay Parade,
Greta Point, Private Bag 14-901, Kilbirnie, Wellington, New
Zealand. E-mail: l.torres@niwa.co.nz
et al. 1999, Gregr and Trites 2001) and understand the
ecology of these animals (Hamazaki 2002). By assuming
that the distribution of cetaceans is nonrandom relative
to environmental variability, predictive models of
cetacean distribution typically identify the ecological
relationships between the environment and species
habitat selection. With the goal of improving conservation applications of modeling efforts, our study examines the potential for increased predictive capacity by
models of dolphin distribution that include direct prey
data as an explanatory variable.
Abiotic variables may be correlated with the distribution of dolphins (i.e., temperature, salinity, depth,
dissolved oxygen, distance from shore). However, these
metrics often have little direct influence on the actual
selection of habitats by dolphins. In reality, these abiotic
variables are frequently used as proxies for prey
distribution. Predictive models of dolphin distribution
rarely include direct data on prey distribution because
prey sampling is more difficult than sampling of abiotic
1702
October 2008
MODELING MARINE PREDATOR HABITAT
variables. Therefore, as top marine predators, dolphins
are removed from the direct influence of the environmental variability that is commonly used to characterize
their habitat.
The ability to predict dolphin distribution requires an
understanding of the ecological factors relevant to
dolphin habitat selection. Environmental characteristics
influence fish distribution by spatial structuring of
phytoplankton and zooplankton populations and the
physiological limits of fish. The distribution of dolphin
predators is contingent upon this spatial structuring of
their prey. Therefore, if dolphins track their prey,
dolphin habitat selection should mimic the environmental variability that structures fish distribution. The
relationships between environment and prey and between prey and predators are well understood (Hugie
and Dill 1994, Sih 1998). However, despite established
predator–prey theory on the response of predators to
their environment (see Lima 2002), the extent to which
the environment can be used to predict predator
distribution without any information on the intermediary step (prey) remains unevaluated.
In a recent paper reviewing the techniques for
cetacean habitat modeling (Redfern et al. 2006), the
authors note that (1) environmental variables are used as
proxies of prey distribution and (2) cetacean habitat
models should ideally incorporate data on prey populations and understand the ecological relationships
between predators, prey, and their environment. It is
generally believed that predictive models of cetacean
distribution would improve with the inclusion of direct
data on prey distribution and densities (Davis et al.
1998, Ferguson et al. 2006a, Redfern et al. 2006), but
this assertion has never been tested. Few studies have
actually sampled prey and monitored dolphin distributions synoptically to describe correlations with dolphin
habitat use (Allen et al. 2001, Heithaus and Dill 2002,
Benoit-Bird and Au 2003). In this paper, we ask whether
including prey distribution data in fine-scale predictive
models of bottlenose dolphin (Tursiops truncatus)
habitat selection in Florida Bay, Florida, USA, improves predictive capacity.
We begin with an exploratory exercise to compare the
response of dolphin distribution and fish catch to a suite
of environmental variables. We assess the pragmatism of
predictive models’ reliance on proxy relationships
between environment and dolphin distribution under
the assumption that dolphin predators mimic the spatial
distribution of their prey. If this assumption is true,
dolphins and fish should display similar responses to the
same environmental variables. This phase of analysis
also evaluates the optimal descriptive metric of fish
sampling data (i.e., catch per unit effort, diversity,
richness) for use in models of dolphin habitat selection.
Next, in hypothesis 1 (H0 1), we assert that dolphin
habitat selection can be predicted without recourse to
describing the distribution of their prey. We assess four
types of models, with and without fish sampling data as
1703
explanatory variables, according to their capacity to
predict dolphin habitat selection. Hypothesis 2 (H0 2)
builds on the results of H0 1 to determine whether
accurate predictive maps of dolphin distribution can be
produced by modeling areas of high fish catch based on
significant environmental characteristics. The model
tested in H0 2 predicts habitat with high prey abundance
and assumes that dolphins will distribute themselves
accordingly.
We collected data in Florida Bay during four summer
field seasons (2002–2005) to test these hypotheses. These
data include standardize dolphin surveys and prey
sampling by bottom trawls, both with associated habitat
quality measurements. Florida Bay is a remarkably
heterogeneous ecosystem (Sogard et al. 1989, Zieman et
al. 1989, Fourqurean and Robblee 1999, Durako et al.
2002) that lends itself to a fine-scale study of dolphin
habitat selection due to large habitat variation over
small spatial distances. Ample data are available from
these four years of fieldwork to use independent data
sets to train and test models. Shark sampling conducted
in Florida Bay found few sharks of the species and sizes
that would pose a threat to dolphins or cause behavior
modifications (Heithaus 2001, Torres et al. 2006, Wiley
and Simpfendorfer 2007). Thus, we assume a minimal
predation effect on dolphin habitat selection in Florida
Bay. Commercial fishing is prohibited and recreational
boating is minimal in Florida Bay, so we also assume
that fine-scale dolphin distribution is not heavily
influenced by boat traffic, fishing activities, or other
proximate anthropogenic activities.
We use generalized additive models (GAMs) to model
dolphin and fish catch distribution because they are
nonparametric and describe nonlinear relationships.
GAMs have been used in previous work to detect
significant nonlinear relationships between cetacean
distribution and environmental variables (Forney 1999,
2000, Hedley et al. 1999, Ferguson et al. 2006b).
Moreover, Segurado and Araujo (2004) found that
GAMs are an appropriate technique to model species
with complex distribution and behavior patterns relative
to environmental variables.
Our spatial analysis is performed on presence/absence
data, incorporating each sighting equally regardless of
group size to minimize the effect of population density
on habitat selection patterns. A fine-grain resolution of
50 m is applied to all analyses because of the extreme
habitat heterogeneity in Florida Bay. The scale of study
is an important and, often confounding, variable related
to distribution matching between predator and prey in
marine environments (Schneider and Piatt 1986, Logwerwell and Hargreaves 1996, Heithaus and Dill 2006).
Working at a small scale fosters the incorporation of
small, yet important, refuges for prey and the ability of
predators to use patchy habitats with high prey
abundance. Moreover, fine-scale studies of species
biogeography coupled with behavioral data can reveal
considerable insight about the biological mechanisms
1704
Ecological Applications
Vol. 18, No. 7
LEIGH G. TORRES ET AL.
FIG. 1. Bottom types and zones of Florida Bay, USA. Inset: Florida Bay lies at the southern end of the Florida peninsula (area
in box enlarged). Black lines designate boundaries between the six environmentally homogeneous zones. The bottom types map is
based on USGS map OFR 97-526.
underlying associations between distribution, environment, and space. Therefore, we recorded the behavior
state of observed dolphins at each sighting to denote
information on the functional mechanisms of habitat
selection (Heithaus and Dill 2002, Hastie et al. 2004).
Dolphins may socialize or travel through marginal
habitat, but forage in areas where critical resources
exist; such areas may deserve increased management
attention and protection. Therefore, due to the importance of identifying feeding areas in conservation
modeling applications, we assess each model’s ability
to predict dolphin foraging habitat, as well as the
model’s overall predictive capacity.
METHODS
Study area
Florida Bay lies at the southern tip of the Florida
Peninsula and is ;1800 km2. Florida Bay is composed
of a heterogeneous mosaic of benthic habitat types
including seagrass, mud, sand, and hard-bottom areas
composed of sponge and coral structures (Fig. 1).
Previous research in Florida Bay subdivided the bay
into environmentally distinct zones based on fish
composition (Sogard et al. 1989, Thayer and Chester
1989, Thayer et al. 1999), water quality (McIvor et al.
1994, Boyer et al. 1999), and seagrass distribution
(Zieman et al. 1989, Hall et al. 1999). We used these
definitions in our study to divide the bay into six
relatively homogeneous zones: Atlantic, Central, Eastern, Gulf, Flamingo, and Western (see Plate 1).
Field methods
We conducted a fine-scale study of dolphin habitat
selection in Florida Bay during the summers of 2002
through 2005. This fieldwork consisted of track line
surveys for dolphin presence and absence. We drove a
5.5-m fiberglass skiff at ;30 km/h with three trained
observers searching for dolphins. When a dolphin
sighting occurred, we stopped the research vessel so
that the radial distance to the sighting and angle from
the track line could be estimated. We used these data to
determine the perpendicular distance from the track line
(Buckland et al. 2001) and to generate pseudo-absences
(see Pseudo-absence generation below). The field crew
then slowly approached the dolphin(s) to record data
including GPS location, depth, benthic habitat type,
water quality metrics (see Benthic habitat and water
October 2008
MODELING MARINE PREDATOR HABITAT
quality sampling below), group size, and behavior state
(forage, travel, socialize, rest, or unknown). For these
analyses, we classified dolphin behavior at each sighting
as foraging or non-foraging. We determined foraging
behavior if dolphins were observed catching fish, chasing
fish, or surfacing erratically and quickly within one area
using fluke-out dives to promote deep diving profiles.
Additionally, we recorded water quality and bottom
type at the start and end of each survey, as well as every
30 min if no dolphins were sighted. These methods
allowed us to assess environmental conditions at
locations of both dolphin presence (sightings) and
absence.
Due to limited knowledge on dolphin habitat use in
Florida Bay prior to fieldwork in 2002, we used a
random approach to dolphin surveys during the 2002
and 2003 field seasons. During each of these field
seasons we surveyed the entirety of Florida Bay twice.
This allowed us to identify areas of high use by dolphins.
From these surveys in 2002 and 2003, we established
track lines in areas of high dolphin use within three of
the environmentally distinct zones: Atlantic, Central,
and Gulf (Fig. 2). We completed replicate surveys along
each track line four times in 2004 and twice in 2005.
Benthic habitat and water quality sampling
The USGS produced a bottom types map of Florida
Bay in 1997 (USGS, OFR 97-526). We used this map as
the foundation for benthic habitat classification
throughout Florida Bay. To refine this map, we assessed
and recorded benthic habitat type at all water and
habitat quality sampling locations. These 1092 sampling
locations were used to update and improve the
resolution of the original USGS map (see Fig. 1). We
classified habitat types by visual inspection (through the
water column) or, when turbidity did not allow the
former, a small bottom grab sampler (;5 cm diameter,
;7.7 cm deep) was used. We categorized bottom types
with the same classification system used by the USGS
map with the addition of two classes: (1) patchy
seagrass, i.e., areas composed of dense patches of
seagrass interspersed in barren landscapes and, (2) hard
bottom with seagrass, i.e., areas composed on equal
amounts of sponge and coral habitat as seagrass habitat.
Thus, we used nine bottom type classifications in this
study: sparse seagrass, intermediate seagrass, dense
seagrass, patchy seagrass, hard bottom, hard bottom
with seagrass, mud, sand, and mud bank.
In 2002 and 2003, a YSI 30 conductivity instrument
(Yellow Springs Instruments, Yellow Springs, Ohio,
USA) measured salinity and temperature at the midpoint of the water column. To estimate turbidity,
percentage of dissolved oxygen, and chlorophyll a
values during these years, we acquired data through
the Southeast Environmental Research Center’s Water
Quality Monitoring Network at Florida International
University, Miami, Florida, USA. This program has 24
water quality stations with monthly sampling periods
1705
FIG. 2. Survey areas and track lines during 2004 and 2005
field seasons. Gray areas represent surveyed regions within
larger zones. Black lines designate the track line followed within
each zone for replicate surveys.
placed throughout Florida Bay. Using ArcGIS (version
9.1; Environmental Systems Research Institute, Redlands, California, USA), we interpolated these data
points with an inverse distance weighting (IDW)
technique that preserves local variation between sample
points. We chose an IDW interpolation technique as the
most appropriate method to create raster grids for
preservation of the water quality variation that exist in
Florida Bay across spatial scales less than or equal to the
distances between the 24 water quality stations. In total,
we created 18 separate rasters: one grid for each variable
(turbidity, percentage of dissolved oxygen, chlorophyll
a) for each month of the study period (June, July,
August of 2002 and of 2003). Finally, we used each
location (sightings, absences, trawls, and 30-min survey
interval water quality sites) to sample the appropriate
temporal set of three grids. We used these values of
turbidity, percentage of dissolved oxygen, and chlorophyll a in all further analyses regarding 2002 and 2003
data points.
During the 2004 and 2005 field seasons, we used a YSI
6600 Sonde monitor (Yellow Springs Instruments) to
sample the following water quality data in the upper and
lower portions of the water column: temperature (in
degrees Celsius), salinity (ppt), turbidity (nephelometric
turbidity units [NTU]), percentage of dissolved oxygen,
and chlorophyll a. We applied the average value of the
upper and lower water column measurements for each
location in all analyses because no discernable difference
was detected between water samples from the wellmixed, shallow waters of Florida Bay. (Methods on the
calibration of YSI-based chlorophyll a readings can be
found in Appendix A.)
Fish sampling
We sampled the fish community throughout Florida
Bay using a 3-m demersal research otter trawl towed at
1706
Ecological Applications
Vol. 18, No. 7
LEIGH G. TORRES ET AL.
;4 km/h for 3 min. We randomly generated the
locations of trawl sampling sites and stratified them by
benthic habitat type to sample the different bottom
types within each zone. We conducted a minimum of
three trawls on each survey day. All captured fish were
identified, measured, and then released alive. We
recorded all water and habitat quality metrics prior to
each trawl and the GPS positions at the start and end of
each trawl to calculate the exact distance trawled.
We converted the catch from each trawl into four
descriptive metrics: catch per unit effort (CPUE),
dolphin prey per unit effort (DPPUE), Simpson’s
diversity index per unit effort (SPUE), and Margalef’s
species richness index per unit effort (MPUE). Catch per
unit effort is the total number of fish captured per meter
of trawling. Dolphin prey per unit effort is a subset of
CPUE consisting of catch that may be potential dolphin
prey items. We determined this subset by fish species and
size. Due to the expansive, uninhabited nature of
Florida Bay, stranded dolphins are rarely recovered,
and to date no data exist from stomach contents in this
region. Therefore, we assumed the diet of dolphins in
Florida Bay to be similar to the diets of bottlenose
dolphins documented in adjacent areas (Barros 1987,
1993, Barros and Odell 1990, Barros and Wells 1998).
The SPUE is Simpson’s diversity index ðRp2i Þ for each
trawl divided by the length of the trawl in meters. We
calculated Margalef’s richness index ((S 1)/log(N)) for
each trawl and converted it into the per meter of
trawling metric MPUE. We implemented these four
metrics in the exploratory exercise to determine the fish
catch metric most correlated with dolphin habitat
selection. Unfortunately, we were unable to include a
metric of fish biomass captured by each trawl in the
analysis due to a lack of length–mass parameters for a
number of frequently caught fish.
Pseudo-absence generation
While surveying for dolphins, we constantly collected
absence data. However, these data are in the form of
strip transect areas, as opposed to dolphin sightings,
which are point locations. Therefore, to test for dolphin
habitat selection with binomial presence/absence data,
we generated pseudo-absence points in the areas of
absence by accounting for detection probabilities
(MacKenzie et al. 2002, Brotons et al. 2004). The
quality of a model can be significantly influenced by the
location of absence points (Engler et al. 2004).
Therefore, we proportionally distributed pseudo-absence locations while accounting for survey effort and
conditions to avoid labeling areas of missed sightings as
absence locations. We generated 10 times the number of
sightings as pseudo-absences and randomly distributed
these pseudo-absence points in proportion to survey
effort, distance from track line, and sighting conditions.
(See Appendix B for more details on pseudo-absence
generation.)
Grid sampling
We created interpolated raster grids, with 50-m grid
cell size, to obtain water quality values for all absence
points. We recorded water quality and habitat metrics at
all dolphin sightings so it was unnecessary to sample
presence locations (except to acquire turbidity, percentage of dissolved oxygen, and chlorophyll a values for
2002 and 2003 sightings). We created daily water quality
grids for each survey day through interpolations of all
water quality sampling locations on each survey day
(sightings, trawls, 30-min survey intervals, and start and
end of surveys samples). We employed a kriging
interpolation method to generate daily water quality
grids because numerous data points occupied relatively
small spatial extents, allowing the kriging method to
accurately interpolate spatial trends between data
points.
In contrast to water quality, we created seasonal
interpolated surfaces of trawl data of the four fish
metrics. With only three trawls conducted on each
survey day we did not have enough spatial coverage to
adequately describe the daily spatial variability of fish
distribution. Therefore, we utilized all sampling locations conducted in each zone and summer field season to
interpolate fish grids. With this method we make the
assumption that the fish community structure within
each zone of Florida Bay does not change throughout a
summer (Matheson et al. 1999, Thayer et al. 1999,
Gaertner 2000). We created CPUE, DPPUE, SPUE, and
MPUE grids for each zone during each summer between
2002 and 2005. We interpolated these grids from trawl
data points using either a spline or kriging method,
determined by the accuracy of real data point representation.
Analysis
Generalized additive models.—Using S-plus 7.0 (Insightful Corporation, Seattle, Washington, USA), we
created models of dolphin habitat selection using
generalized additive models (Hastie and Tibshirani
1990) with a smoothing function and backwards
variable selection. We chose the optimal model,
composed of the combination of variables that best fit
the observed data, based on the lowest Akaike
Information Criterion (AIC; Burnham and Anderson
1998). The GAMs generate smoothed curves representing the relationship between the response and each
predictor variable in the model. The GAMs are
particularly good at identifying and describing nonlinear
relationships that are more typical than linear relationships in ecology (Oksanen and Minchin 2002). We
conducted GAMs with two types of data: (1) binomial
dolphin presence/absences data with a logit link function
and (2) continuous fish catch data using a Gaussian
family model with an identity link function.
Mantel’s tests.—Mantel’s tests (Mantel 1967) are able
to overcome many problems associated with examining
species–environment relationships. They are multivari-
October 2008
MODELING MARINE PREDATOR HABITAT
ate, explicitly test for the effect of space on the response
variable, account for multicolinearity between predictor
variables, and identify and account for spatial autocorrelation of explanatory variables (Schick and Urban
2000). Mantel’s tests are a more robust analysis than
GAMs, which have been criticized for their tendency to
overfit data and give artificially high P values (Guisan
and Zimmermann 2000, Guisan et al. 2002, Vaughan
and Ormerod 2005). Unlike GAMs, greater correlation
between predictor and response variables is required to
obtain a significant P value from a Mantel’s test.
Therefore, for the exploratory exercise, we used Mantel’s tests (run in S-Plus 7.0) to determine significant
predictor variables correlated with dolphin presence/absence in order to limit the number of variables included
in the GAMs. (See Appendix C for details on Mantel’s
tests.)
Predictive maps.—We used the zero line on each
GAM plot to divide the range of the explanatory
variable that had a positive effect from the range that
had a negative effect on the response variable. We
applied simple threshold cutoffs from these GAM plots
in a geographic information system (GIS) framework to
define dolphin habitat. Where the response curve was
above the zero line, the explanatory variable was used to
select habitat. This threshold approach is an extension
of envelope models that generate reproducible results
based on minimum and maximum values of explanatory
variables (Redfern et al. 2006). Using GIS, we selected
grid cells from interpolated surfaces representing various
variables (i.e., salinity, bottom type) based on the
thresholds determined by the GAM models. We refer
to this technique as ‘‘GAMvelopes.’’
Model evaluation.—Model evaluation is a critical step
in producing predictive species distribution models
(Rushton et al. 2004, Vaughan and Ormerod 2005).
Although evaluation standards can vary depending on
the goal of the model, most evaluation techniques must
reflect the model’s ability to correctly select habitat, as
well as exclude non-habitat. We assessed the predictive
capacity of each GAM model used to map dolphin
habitat using the true skill statistic (TSS; Allouche et al.
2006) calculated from traditional model accuracy
measures of sensitivity and specificity. The TSS is
defined as
TSS ¼ sensitivity þ specificity 1:
The TSS is insensitive to both prevalence and the size of
the validation set, combines both errors of commission
(sensitivity) and omission (specificity), and is simple to
calculate and interpret. The TSS ranges from 1 to þ1,
where þ1 indicates perfect model performance. A zero
value means the model performed no better than
random, and a negative value indicates that the model
performed worse than random guessing would have
predicted. The TSS assigns equal weight to sensitivity
and specificity, which makes false positive as unwanted
1707
as false negatives. (See Appendix D for details on model
evaluation.)
Hypothesis testing
We performed the following analyses by creating
dolphin habitat selection models with training data sets
and evaluating each models’ predictive capacity with
independent testing data sets. As explained above, two
different survey methods were employed in our fieldwork. Thus, we matched testing and training data sets
based on consistent field methods: we used data
collected from the Atlantic, Central, and Gulf zones in
2004 and 2005 in our exploratory exercise; for H0 1, we
used 2004 data to train models and 2005 data to test
each model; we explored H0 2 with models trained based
on 2002 data and tested by predicting data collected in
2003.
Exploratory exercise.—To examine the effect of the
same environmental variables on predator and prey, we
first determined the significant predictors of dolphin
presence/absence with binomial Mantel’s tests using the
following explanatory variables: temperature, salinity,
turbidity, chlorophyll a, percentage of dissolved oxygen,
distance from mud banks, CPUE, DPPUE, MPUE, and
SPUE. We transformed the last five variables to a log þ
1 scale. (Detail on the use of Mantel’s tests in the
exploratory exercise can be found in Appendix E.)
The next step of our exploratory exercise modeled fish
catch relative to environmental variability. We used the
fish metric identified as most correlated with dolphin
presence/absence by the Mantel’s tests (CPUE). Those
explanatory variables found to be correlated with CPUE
were then tested in a binomial GAM of dolphin
presence/absence for that same zone. To conclude this
exploratory exercise, we created and compared GAM
plots of CPUE and binomial dolphin presence/absence
to the same explanatory variables.
H0 1.—To evaluate the utility of fish distribution data
in predictive models of dolphin habitat selection we
created four types of models that used different sets of
explanatory variables. The first model predicted dolphin
presence/absence based on environmental factors (notation: DOLPHIN ; ENV). The next model predicted
dolphin distribution based only on fish catch data
(notation: DOLPHIN ; FISH). The third model
predicted dolphin habitat by modeling both environmental factors and fish catch data (notation: DOLPHIN
; ENV þ FISH). The final model predicted areas of
high fish catch (CPUE) based on environmental factors
(notation: FISH ; ENV). This model essentially
predicted habitat with high prey abundance and
assumed that hungry dolphins would distribute themselves appropriately. We created the first three types of
models using presence and absence points from four
surveys conducted in each zone (Atlantic, Central, and
Gulf) in 2004. We developed models of the fourth type
based on fish catch data from all trawls conducted in
each of the three zones during the 2004 summer.
1708
LEIGH G. TORRES ET AL.
Ecological Applications
Vol. 18, No. 7
FIG. 3. Example ‘‘GAMvelope’’ predictive map for the Atlantic region on 23 July 2005. Habitats selected by thresholds
generated from generalized additive model (GAM) plots are shown modeling 2004 dolphin presence/absence based on
environmental variables (DOLPHIN ; ENV). The dashed horizontal black lines represent the zero line in each plot. The vertical
dotted lines delineate the range of the explanatory variable above the zero line used as thresholds. The dotted lines bracketing the
response curves are twice the standard error and function as confidence limits of the model. Thresholds from GAM plots applied in
GIS to select habitats are: bottom type (hard bottom with seagrass, intermediate seagrass, or sparse seagrass); salinity (,38.3 or
.41.5; P ¼ 0.01); and turbidity (,2.2 or .2.5 and ,2.9; P ¼ 0.02). Each tick mark on the x-axis represents one sample at that value;
i.e., sampling intensity. Note the different scales on the y-axis between plots. The habitat of six out of six sightings (open squares)
was selected including three out of three foraging sightings (demarcated with an F). The locations of 45 of 61 absences (open circles)
were not selected as habitat. This predictive map yielded a sensitivity of 1, a specificity of 0.7377, and a true skill statistic (TSS) of
0.7377. Abbreviations are: D, dense seagrass; HS, hard bottom with seagrass; H, hard bottom; I, intermediate seagrass; MB, mud
bank; SP, sparse seagrass. NTU is nephelometric turbidity units.
For each model we applied the ‘‘GAMvelope’’
technique to any explanatory variable with a significant
parameter estimate (P , 0.05). We tested the predictive
capacity of each 2004 model with presence and absence
points from two independent surveys conducted in 2005
in each zone. The GAMvelope thresholds were applied
to daily water quality grids for each 2005 survey day,
seasonal CPUE grids for each zone in 2005, and a static
bottom type grid of each zone. We subsequently
calculated the sensitivity, specificity, and TSS for each
2005 predictive map. Additionally, we tallied the ratio of
foraging sighting locations predicted to foraging sightings observed. Using six 2005 surveys as test data sets,
we compared the TSS and percentage of foraging
sightings correctly predicted between the four model
types. Fig. 3 depicts an example GAMvelope predictive
map based the DOLPHIN ; ENV model from the
Atlantic zone in 2004 and tested by a dolphin survey in
the Atlantic zone on 23 July 2005. The GAMs and
threshold ranges applied to create the predictive maps of
dolphin distribution for each model type tested in H0 1
are described in Table 1.
H0 2.—This hypothesis explicitly tests the predictive
capacity of dolphin distribution by the FISH ; ENV
model type using large training (2002) and testing (2003)
data sets. We created one GAM of CPUE for each zone
based on trawls conducted in 2002 in each zone:
Atlantic, Central, Gulf, Eastern, and the combined
region of the Flamingo and Western zones. The
Flamingo and Western zones are geographically adjacent and have similar environments, which allowed us to
combine the zones to increase the sample size of trawls
used to create the 2002 fish catch GAM. The predictor
variables tested for a significant relationship with CPUE
were bottom type, salinity, temperature, chlorophyll a,
turbidity, percentage of dissolved oxygen, and distance
October 2008
MODELING MARINE PREDATOR HABITAT
1709
TABLE 1. Description of the four types of generalized additive models (GAMs) created for each zone and tested in hypothesis 1
(H0 1; see Methods: Hypothesis testing: H0 1): DOLPHIN ; ENV þ FISH, DOLPHIN ; ENV, DOLPHIN ; FISH, and FISH
; ENV.
Models based on
data from zone/year
Model type
GAM model
R2
Atlantic 2004 (9 presences,
137 absences)
DOLPHIN ; ENV
þ FISH
Dolphins ; Bottom þ
s(DPPUE) þ s(TEMP) þ
s(MPUE)
Dolphins ; Bottom þ
s(SALINITY) þ
s(Dist.MB) þ s(TURB)
0.9762
DOLPHIN ; ENV
DOLPHIN ; FISH
Dolphins ; s(SPUE) þ
s(DPPUE)
CPUE ; Bottom
0.7188
0.3761
Bottom ¼ HBSG, INT,
Sparse; DPPUE , 0.005
or .0.022 (0.05)
Bottom ¼ HBSG, INT,
Sparse; SALINITY ,
38.4 (0.01); TURB , 2.2
or .2.5 and ,3.0 (0.02)
SPUE . 0.0185 and
,0.0225 (0.05)
Bottom ¼ INT
12 trawls
FISH ; ENV
Central 2004 (21 presences,
206 absences)
DOLPHIN ; ENV
þ FISH
P.A ; Bottom þ s(DPPUE)
þ s(SALINITY) þ
s(TEMP) þ s(CHLA) þ
s(Dist.MB) þ s(TURB)
0.7846
DOLPHIN ; ENV
P.A ; Bottom þ s(TEMP)
þ s(CHLA) þ s(Dist.MB)
þ s(TURB)
0.6676
DOLPHIN ; FISH
P.A ; s(CPUE) þ
s(DPPUE)
CPUE ; Bottom þ
s(SALINITY)
0.2274
Bottom ¼ Dense, Mud
Bank; SALINITY , 42
or .50 (0.002); DPPUE
. 0.011 and ,0.0225
(0.004); TEMP , 28.7 or
.31.3 (0.015); TURB ,
4.1 (0.016)
Bottom ¼ Dense, Mud
Bank; TEMP , 28.5 or
.31 (0.0009); TURB ,
4.5 (0.002)
CPUE . 0.05 (,0.0001)
0.9807
Bottom ¼ Mud Bank
P.A ; Bottom þ s(CPUE)
þ s(TEMP) þ s(CHLA)
þ s(TURB) þ s(DO)
P.A ; Bottom þ
s(Dist.MB) þ
s(SALINITY) þ s(TEMP)
þ s(CHLA) þ s(TURB)
þ s(DO)
P.A ; s(SPUE) þ s(CPUE)
0.5827
Bottom ¼ INT, MUD,
SAND; DO . 102 (0.001);
TURB , 9 (0.04)
Bottom ¼ INT, MUD,
SAND; Dist.MB , 4.1 or
.7.2 (0.004); SALINITY
, 34.7 or .38.1 (0.008)
CPUE ; Bottom þ
s(TEMP)
0.9234
13 trawls
FISH ; ENV
Gulf 2004 (21 presences,
145 absences)
DOLPHIN ; ENV
þ FISH
DOLPHIN ; ENV
DOLPHIN ; FISH
12 trawls
FISH ; ENV
0.3610
‘‘GAMvelope’’ thresholds
applied in GIS
0.5134
0.1750
SPUE . 0.013 and ,0.0275
(0.003); CPUE , 0.08 or
.0.42 (0.005)
Bottom ¼ INT, MUD
BANK
Notes: Column 1 describes the number of presence and absence locations used to fit the first three types of GAMs for each zone.
The number of trawls used to fit the last GAM (FISH ; ENV) is described in the bottom row for each zone. The selected model,
based on the minimum Akaike Information Criterion (AIC) produced by the backward selection process, is described under ‘‘GAM
model’’ with the corresponding R2 value. Thresholds to be applied in GIS to produce predictive maps were only generated for those
explanatory variables that were significant at the 0.05 level. Values in parentheses are P values from GAM. Abbreviations:
DOLPHIN, dolphin presence or absence; Bottom, bottom type; TEMP, temperature; CHLA, chlorophyll a; TURB, turbidity; DO,
percentage of dissolved oxygen; CPUE, catch per unit effort; DPPUE, dolphin prey per unit effort; MPUE, Margalef’s richness
index per unit effort; SPUE, Simpson’s diversity index per unit effort; Dist.MB, distance from mud banks (log þ 1); Dense, dense
seagrass; INT, intermediate seagrass; Sparse, sparse seagrass; HBSG, hard bottom with seagrass.
from mud banks (log þ 1). To select habitat in GIS we
applied derived GAMvelope thresholds of significant
explanatory variables to daily grids of the 2003 water
quality metrics and a static bottom type grid for each
zone (see Table 2 for model specifications). We used
these habitat maps of fish abundance to predict dolphin
sightings and pseudo-absence points from each 2003
survey in each zone. We then assessed sensitivity,
specificity, the TSS, and the percentage of foraging
sightings correctly predicted for each predictive map. In
total, we created five zonal models from 99 trawls
conducted in 2002. We validated these CPUE ; ENV
models with 36 independent dolphin surveys from 2003.
RESULTS
Exploratory exercise
The global Mantel’s tests for each zone found
significant correlation between the explanatory variables
and dolphin presence/absence, even when the effect of
space was accounted for (Table 3). Many of the
explanatory variables had significant spatial autocorrelation but the pure partial Mantel’s tests were able to
compensate for these spatial trends to identify those
explanatory variables with direct correlation to dolphin
distribution. In the Atlantic zone during 2004 and 2005,
CPUE and DPPUE were the only variables significantly
1710
Ecological Applications
Vol. 18, No. 7
LEIGH G. TORRES ET AL.
TABLE 2. Generalized additive models (GAMs) specifications of models used to test hypothesis 2 (H0 2): catch per unit effort as a
function of environmental characteristics (CPUE ; ENV) from trawls conducted in each zone in 2002.
Model
R2
No. trawls
used to
create GAM
Atlantic 2002
Central 2002
CPUE ; Bottom
CPUE ; Bottom þ
s(CHLA) þ s(Dist.MB)
0.4805
0.9662
22
23
Eastern 2002
Gulf 2002
Western and Flamingo 2002
CPUE ; Bottom
CPUE ; s(Dist.MB)
CPUE ; Bottom
0.2073
0.5827
0.4483
27
14
13
CPUE ; ENV GAM
based on data
from zone/year
‘‘GAMvelope’’ thresholds
applied in GIS
Bottom ¼ Dense, INT
Bottom ¼ Dense, INT, Mud
Bank; Dist.MB ¼ 0 or . 5.6
(0.002)
Bottom ¼ Dense, INT, Mud
Dist.MB . 5.75 and , 7.7 (0.04)
Bottom ¼ Dense, Mud Bank
Notes: Values in parentheses are P values from GAM. Key to abbreviations: Bottom, bottom type; CHLA, chlorophyll a;
Dist.MB, distance from mud banks (log þ 1); Dense, dense seagrass; INT, intermediate seagrass.
related to dolphin presence/absence once the effect of
spatial location was removed. Salinity, chlorophyll a,
percentage of dissolved oxygen, and CPUE were
significantly correlated with dolphin distribution in the
Central zone. The predictor variables correlated to
dolphin presence/absence in the Gulf zone in 2004 and
2005 were salinity, chlorophyll a, and CPUE.
With the effects of autocorrelation and multicollinearity removed in the pure partial Mantel’s tests,
CPUE was identified as the fish metric most correlated
with dolphin presence/absence in all three zones. No
other fish metric was found significant, with the
exception of DPPUE in the Atlantic zone. Based on
this result, CPUE was used as the response variable
TABLE 3. Mantel P value results from binomial Mantel’s tests on dolphin presence/absence data
from 2004 and 2005 grouped by zone.
Atlantic
Central
Gulf
Test type and response variable
P
P
P
Global Mantel tests
Explanatory on dolphins
Space on dolphins
Space on explanatory
Explanatory on dolphins accounting for space
Space on dolphins accounting for explanatory
0.023
0.511
0.001
0.016
0.714
0.007
0.967
0.001
0.001
1
0.088
0.509
0.001
0.071
0.847
Spatial autocorrelation of explanatory variables
Temperature
Salinity
Turbidity
Chlorophyll a
Percentage of dissolved oxygen
Distance from mud banks (log þ 1)
CPUE (log þ 1)
DPPUE (log þ 1)
MPUE (log þ 1)
SPUE (log þ 1)
0.002
1
0.144
1
NA
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.95
0.001
0.001
0.001
0.001
0.001
0.001
NA
0.575
NA
1
0.001
0.001
0.001
0.001
Pure partial Mantel’s tests
Temperature
Salinity
Turbidity
Chlorophyll a
Percentage of dissolved oxygen
Distance from mud banks (log þ 1)
CPUE (log þ 1)
DPPUE (log þ 1)
MPUE (log þ 1)
SPUE (log þ 1)
Space
0.56
0.83
0.326
0.842
NA
0.959
0.001
0.064
0.473
0.506
0.817
0.993
0.069
0.515
0.002
0.001
0.179
0.003
0.998
0.573
0.786
0.962
0.497
0.041
NA
0.01
NA
0.4
0.041
0.959
0.447
0.842
0.188
Notes: Values in boldface type are significant using the 0.09 cutoff. NA values for percentage of
dissolved oxygen are a result of a broken sensor during fieldwork. ‘‘Explanatory’’ signifies
explanatory variables; ‘‘dolphins’’ signifies dolphin presence/absence; ‘‘Space’’ signifies the
geographic distance between data points. See Table 1 for an explanation of abbreviations. See
Methods: Analysis: Mantel’s tests for explanation of tests.
NA values are due to a strong correlation between turbidity and chlorophyll a, making it
necessary to remove turbidity from the analysis (turbidity was not significant without chlorophyll a
in the data set).
October 2008
MODELING MARINE PREDATOR HABITAT
1711
PLATE 1. Bottlenose dolphins swimming in shallow, seagrass habitat within the Central zone of Florida Bay, Florida (USA).
Photo credit: L. G. Torres.
tested in the second step of this exploratory exercise
(GAMs of trawl catch) and in the FISH ; ENV models
of H0 1 and H0 2.
Despite using different data sets to create GAM
models of CPUE and dolphin distribution, results from
this exploratory exercise showed strong evidence that
dolphins and fish catch responded similarly to the same
environmental variables. Fig. 4 compares GAM plots
between significant variables related to CPUE with
GAM plots of binomial dolphin distribution vs. the
same variables in each zone. Each zonal GAM of CPUE
; ENV returned bottom type and one water quality
variable as significant explanatory variables.
Salinity and bottom type were the significant predictor variables of trawl catch in the Atlantic zone. The
GAM plots comparing the effects of salinity on fish
catch and dolphin distribution showed that a similar
range of salinity had a positive effect on both response
variables (Fig. 4a). More fish we caught at salinity levels
,37 ppt and in the GAM plot of dolphin presence/
absence a similar range of salinity, ,38 ppt, had a
positive effect on dolphin presence. Moreover, the same
bottom types had similar effects on CPUE and dolphin
presence in the Atlantic zone: intermediate seagrass was
positively associated with both dolphin presence and
CPUE, while hard bottom habitats were negatively
correlated with both dolphin presence and CPUE.
In the Central zone, more fish were caught in habitats
with salinity levels .47 ppt and in mud or mud bank
bottom types (Fig. 4b). The GAM plots of dolphin
presence/absence depict that the same habitats had a
positive impact on dolphin presence: salinity . 47 ppt
and in mud and mud bank bottom types. Sparse
seagrass habitats had negative effects on both CPUE
and dolphin presence. These GAM plots indicate that
both dolphins and fish responded positively to the
hypersaline conditions of the Central zone.
Chlorophyll a was significantly correlated with both
CPUE and dolphin distribution in the Gulf zone. Plots
from the Gulf zone GAM models show that chlorophyll
a values between ;1.5 and 4 lg/L had a positive
influence on dolphin presence and CPUE (Fig. 4c). The
effects of bottom type were less congruent in the Gulf
zone; only intermediate seagrass and mud bank habitats
had positive effects on both dolphin presence and
CPUE.
H0 1.—The predictive capacity of dolphin habitat by
the four model types created from 2004 data and tested
by each 2005 dolphin survey is described in Appendix F.
The TSS for all prediction maps ranged from a low of
0.3087, poor predictive capacity, to a high of 0.7377,
very good predictive capacity. For every 2005 survey day
used to validate the four model types, the resulting TSS
values were used to rank each set of four models. The
1712
LEIGH G. TORRES ET AL.
Ecological Applications
Vol. 18, No. 7
October 2008
MODELING MARINE PREDATOR HABITAT
TABLE 4. Overall model performance by the four tested model
types in hypothesis 1 (H0 1; see Methods: Hypothesis testing:
H0 1).
Model
Dolphin
Dolphin
Dolphin
FISH ;
; ENV þ FISH
; ENV
; FISH
ENV
Sum of
TSS ranks
14
10
18
11
Mean
TSS
value
0.0055
0.2644
0.0807
0.2407
Percentage
of foraging
sightings
predicted
12.50
62.50
0
87.50
Notes: The generalized additive models (GAMs) were trained
with 2004 data and tested by dolphin presence and absence
from 2005 surveys. Predictive capacity is assessed in three ways:
the sum of true skill statistic (TSS) ranks (lower rank indicates a
higher TSS, which means better performance), the mean TSS
for all survey dates, and the percentage of observed foraging
sightings on each survey date correctly predicted by the model.
highest TSS value received a rank of one and the lowest
TSS value received a rank of four.
The overall evaluation of predictive performance by
the four model types was determined using three ranking
methods: sum of TSS ranks, mean TSS value, and the
percentage of foraging sightings correctly predicted
(Table 4). The order of predictive performance by the
four model types was identical between the sum of TSS
ranks and the average TSS value. These two evaluation
methods determined that the DOLPHIN ; ENV
models performed best, but only slightly better than
the FISH ; ENV models. However, the ability of the
DOLPHIN ; ENV models to predict foraging habitat
was 25% worse than the ability of the FISH ; ENV
models. The FISH ; ENV models were able to correctly
predict dolphin foraging locations 87.5% of the time.
The predictive capacity of the DOLPHIN ; ENV þ
FISH models placed third, with an average TSS value
only marginally better than random. Those models that
attempted to predict dolphin distribution based on fish
distribution alone, DOLPHIN ; FISH, performed the
worst and were able to predict 0% of dolphin foraging
sightings.
H0 2.—Predictive maps of 2003 dolphin distribution
produced by GAM models of 2002 CPUE ; ENV were
evaluated by the TSS and percentage of foraging
sightings correctly predicted (Appendix G). Only 7 of
36 predictive maps produced a TSS score of zero or less.
1713
TABLE 5. Overall predictive capacity results of dolphin
distribution by zone for hypothesis 2 (H0 2).
Zone
Atlantic
Central
Eastern
Gulf
Western and
Flamingo
Total, all zones
No.
surveys
tested
Mean TSS
for all
models
7
8
5
6
10
0.5251
0.6557
0.3145
0.2247
0.2326
36
0.3905
Percentage
of foraging
sightings
correctly
predicted (ratio)
60
86
76
80
100
(3/5)
(6/7)
(2/3)
(4/5)
(10/10)
83 (25/30)
Notes: Models of dolphin presence/absence were trained by
data from 2002 trawls (CPUE ; ENV) and tested by 2003
dolphin surveys. The ratio for percentage of foraging sightings
correctly predicted is the number of foraging sightings correctly
predicted divided by the number of foraging sightings observed.
TSS, true skill statistic.
Two maps had perfect predictive capacity and three
other maps produced a TSS of 0.80 or higher.
The five 2002 models of CPUE ; ENV produced
positive mean TSS values ranging from 0.2247 to 0.6557
when validated by 2003 observed presence and absence
locations (Table 5). The mean TSS value for all 36
predictive maps was 0.3905, denoting that the maps
performed on average 40% better at predicting dolphin
habitat than random guessing would have produced.
Additionally, each zone model was able to predict 60%
or more of all dolphin foraging locations, including
100% accuracy in the combined region of the Western
and Flamingo zones. In total, habitat models of high fish
abundance generated by 2002 trawl data accurately
predicted 83% of all dolphin foraging sightings observed
in 2003 in Florida Bay.
DISCUSSION
Our results indicate that incorporating prey data as an
explanatory variable into fine-scale models of dolphin
habitat selection within a heterogeneous environment
does not improve predictive capacity. However, predictive modeling of prey distribution based on environmental variability did produce high predictive
performance of dolphin habitat selection, particularly
foraging habitat. This relationship between dolphin
habitat and the habitat preferences of their prey is
FIG. 4. Generalized additive model (GAM) plot comparison between significant explanatory variables for dolphin distribution
(upper plots in each group of panels) and catch per unit effort (CPUE; lower plots in each group of panels) in the (a) Atlantic, (b)
Central, and (c) Gulf zones based on data from 2004 and 2005. The data sets used to fit each GAM and the resulting models are
described in the right column adjacent to the plots from that model. Bottom type (plots in central column) was a significant
explanatory variable in all GAMs. The left column for each zone compares plots of the same explanatory variable related to
dolphin presence/absence and CPUE. The dashed horizontal black lines represent the zero line in each plot. The vertical dotted
lines delineate the points at which the response curves move above or below the zero line. The dotted lines bracketing the response
curves are twice the standard error and function as confidence limits of the model. Tick marks on the x-axis represent sampling
intensity. Note the different scales on the y-axis between plots. Abbreviations: DOLPHIN, dolphin presence/absence; D, dense
seagrass; H, hard bottom; HS, hard bottom with seagrass; I, intermediate seagrass; M, mud; MB, mud bank; P, patchy seagrass; S,
sand; SP, sparse seagrass.
1714
LEIGH G. TORRES ET AL.
corroborated by the exploratory exercise that depicted
an overall congruency of dolphin distribution and fish
catch response to the same environmental characteristics. While it is difficult to distinguish whether dolphin
predators are tracking their prey, or the resources of
their prey, the FISH ; ENV approach to dolphin
habitat modeling successfully used environmental characteristics as proxies of prey distribution to predict
dolphin distribution.
We conclude from model testing of H0 1 that the
DOLPHIN ; ENV or FISH ; ENV models were the
most appropriate to predict fine-scale dolphin habitat
selection. Why did those models of dolphin distribution
that included fish catch data (DOLPHIN ; ENV þ
FISH and DOLPHIN ; FISH) not perform as well? It
is likely that the scale of fish sampling was inappropriate
for such a fine-scale study. Prey items are able move
between habitats within a grain size of 50 m on a weekly
or possibly daily basis, and our fish sampling scheme
was unable to capture such variability. The assumption
that the spatial distribution of the fish community does
not change during a three-month summer field season
was incorrect, leading to imprecise seasonal fish catch
grids. Conversely, the more frequent habitat quality
sampling enabled greater spatial and temporal resolution, allowing us to capture this fine-scale variability.
Therefore, those models of dolphin distribution that
used only environmental factors as explanatory variables (DOLPHIN ; ENV) were more successful
because of the increased spatial and temporal resolution
at which these parameters were sampled. The scale at
which data are collected and analyzed has important
consequences to model output (Levin 1992, Guisan et al.
2005, Redfern et al. 2006), and previous studies have
also encountered scale-dependent relationships between
marine predators and their prey (Schneider and Piatt
1986, Piatt and Methven 1992, Fauchald et al. 2000,
Guinet et al. 2001).
The FISH ; ENV models performed well because
there is no issue of spatial or temporal scale associated
with these models. At every trawl location, environmental quality was sampled in situ. No spatial interpolations
or temporal assumptions were necessary to associate fish
catch rates with environmental variability. Therefore,
when these models of fish catch based on environmental
characteristics were produced, there were no residual
effects of scale to misrepresent the factors influencing
fish catch distribution. The FISH ; ENV models in
H0 1 and H0 2 predicted fish catch well and the dolphin
predators distributed themselves appropriately, with
83% of 2003 and almost 90% of 2005 foraging sightings
correctly predicted. These results reveal a strong spatial
coupling between foraging habitat and areas of
increased fish catch. Therefore, due to the importance
of identifying feeding areas in conservation modeling
applications and the lack of scaling issues associated
with the FISH ; ENV models, we believe that fine-scale
predictions of top marine predator distribution can be
Ecological Applications
Vol. 18, No. 7
reliably produced by understanding and modeling the
environmental factors that determine prey distribution.
Mismatches between the distributions of predator and
prey may be due to their mutual responses to each other
at various temporal and spatial scales. For instance, the
divergence of bottom type influence in the Gulf zone
between dolphin presence and areas of high fish catch
(Fig. 4c) may be due to an increased encounter rate in
this zone of Florida Bay with sharks of the size and
species that may modify the distribution of dolphins
(Heithaus 2001, Torres et al. 2006, Wiley and Simpfendorfer 2007). Therefore, predictive models of marine
predator habitat use can either attempt to incorporate
all variables relevant to predator–prey interactions (i.e.,
anti-predator tactics, resource availability, predation
and competition effects, density-dependent relationships), which most models cannot do (Redfern et al.
2006), or focus on those variables with weaker or no
response to predator and prey distribution (ENV). The
latter option provides a simpler and effective approach
to modeling marine predator habitat based on the
distribution of the resources of their prey, as demonstrated in this study. The link between dolphins and
environment is easier to model than the relationship
between dolphins and their prey because both predator
and prey are mobile and biotic. At small spatial scales,
in variable coastal ecosystems, it appears ineffective to
predict the distribution of one complex biotic predator
based on the complex distribution patterns of multiple
biotic prey items, especially because these two interact at
multiple spatial and temporal scales.
Despite conducting almost 400 trawls in Florida Bay,
our fish sampling methods were too coarse grained in
time and space to accurately describe overlap between
dolphins and their prey at the fine scale at which these
data were applied. To successfully incorporate prey
distribution data into our predictive models in Florida
Bay we could have increased the spatial scale, thus
leading to a loss of predictive resolution. Alternatively,
we could have trawled at a frequency equal to that of
our environmental quality sampling scheme, achieving
an increased resolution of fish distribution but decreasing time allotted to dolphin surveys. Another option
would have been to employ a prey sampling method
with increased sampling resolution. Benoit-Bird and Au
(2003) accredit their ability to identify congruent
distribution patterns between spinner dolphins and their
prey at multiple spatial and temporal scales to an echo
sounder that simultaneously measured the abundance of
dolphins and prey in the water column. In lieu of such a
synchronous sampling protocol, our results suggest that,
due to high habitat heterogeneity and the spatial
variability of prey patches, fine-scale models of marine
predator habitat selection in coastal habitats will be
more successful if environmental variables are used as
proxies of both prey and predator distributions rather
than relying on direct prey distribution data. Although
we do not believe our fish sampling gear captured all
October 2008
MODELING MARINE PREDATOR HABITAT
potential prey items of dolphins in Florida Bay, we are
confident that the bottom trawl sampled a large majority
of dolphin prey, revealing relative abundance of prey
across the heterogeneous landscape. The issue of
appropriate sampling technique highlights another
reason why relying on prey data alone to model predator
distribution is challenging. It is unlikely that one
sampling method will be able to capture all possible
prey items of opportunistic foragers, such as dolphins,
within a dynamic ecosystem. Moreover, it is difficult to
merge fish catch data sets collected with different gear
types due to effort biases.
We conclude that methods and scales of prey data
collection and incorporation into models must be
appropriate and considered with forethought. For
instance, depending on the spatial coupling between
predators and their prey, direct prey data may be more
effective in models of pelagic predator distribution.
Models for pelagic ecosystems are typically applied to
larger spatial extents but provide decreased spatial
resolution. The resulting larger grid cells allow greater
predictability of prey, which, in conjunction with
minimal prey refuges, may allow increased predictability
of predators based on direct prey data. Again, due to the
patchy nature of fish distribution and abundance, fish
sampling intensity must be such that the applied grain
size is able to realistically represent the spatial and
temporal variability of the ecosystem.
As apex predators, bottlenose dolphins are frequently
considered an indicator of healthy habitats (Caro and
O’Doherty 1999, Torres and Urban 2005). However, the
mere presence of dolphins does not necessarily imply a
healthier habitat, but a sighting of foraging dolphins
does indicate a habitat that provides an important
feeding opportunity. Therefore, because marine predators such as dolphins preferentially select different
habitat characteristics when foraging as compared to
socializing, traveling, or resting, behavior data should be
incorporated when using predators as an indicator of
relative habitat health.
CONCLUSIONS
The results of our study suggest that high predictive
capacity of fine-scale marine predator habitat selection
can be achieved without including prey distribution data
as a predictor variable in the model. Fish aggregations
can be ephemeral, leading to poor performance as a
spatial proxy for predator distribution unless prey is
sampled at an appropriate intensity. Unfortunately, prey
sampling demands more time, skill, and money than
measuring environmental conditions, and managers of
marine predator populations typically face limits on
funding, time, and expertise. Using the example
predator in this study, managers are left with two
options to model dolphin distribution at fine scales.
Both options use environmental characteristics as
explanatory variables but have different response
variables: DOLPHIN ; ENV or FISH ; ENV.
1715
Predictive maps of dolphin distribution based on FISH
; ENV models performed with similar success as
DOLPHIN ; ENV models, but were appreciably better
at predicting dolphin foraging habitat, proving tight
spatial links between foraging dolphins, their prey
community, and their environment. Due to inadequate
prey sampling, our models including fish catch data as
an explanatory variable (DOLPHIN ; ENV þ FISH
and DOLPHIN ; FISH) were unable to account for the
spatial variation of fish distribution, and subsequently
the spatial relationship between dolphin presence and
high prey abundance decoupled.
In a dynamic coastal environment, high sampling
intensity is necessary to successfully incorporate explanatory variables of prey into fine-scale predictive models
of predator distribution due to the complex habitat
selection patterns of both predator and prey. However,
because the environmental dynamics that aggregate prey
are more consistent, more limited prey sampling can be
conducted to develop accurate models of prey distribution based on environmental predictors. These models
can then be applied to predict marine predator
distribution based on the proxy relationships between
predator, prey, and environment.
ACKNOWLEDGMENTS
Data were collected under Everglades National Park permits
EVER-2002-SCI-0049, EVER-2003-SCI-0042, EVER-2004SCI-0062, and EVER-2005-SCI-0062, General Authorizations
911-1466 and 572-1639, and approved by the Duke University
Institutional Animal Care and Use Committee (IACUC).
Funding and support was provided by the NOAA/Coastal
Ocean Program, the Dolphin Ecology Project, the Florida Keys
National Marine Sanctuary, and the Duke University Marine
Lab. Data were provided by the SERC-FIU Water Quality
Monitoring Network, which is supported by SFWMD/SERC
Cooperative Agreement #C-15397 as well as EPA Agreement
#X994621-94-0. Todd Chandler, Kate Freeman, Gretchen
Lovewell, Erica Morehouse, Liz Tuoy-Sheen, Anne Starling,
Danielle Waples, and Margaret Worthington assisted with
fieldwork. We also thank two anonymous reviews for helpful
comments on this manuscript.
LITERATURE CITED
Allen, M. C., A. J. Read, J. Gaudet, and L. S. Sayigh. 2001.
Fine-scale habitat selection of foraging bottlenose dolphins
Tursiops truncatus near Clearwater, Florida. Marine Ecology
Progress Series 222:253–264.
Allouche, O., A. Tsoar, and R. Kadmon. 2006. Assessing the
accuracy of species distribution models: prevalence, kappa
and the true skill statistic (TSS). Journal of Applied Ecology
43:1223–1232.
Barros, N. B. 1987. Food habits of bottlenose dolphins
(Tursiops truncatus) in the southeastern United States, with
special reference to Florida waters. Thesis. University of
Miami, Coral Gables, Florida, USA.
Barros, N. B. 1993. Feeding ecology and foraging strategies of
bottlenose dolphins on the central east coast of Florida.
Dissertation. University of Miami, Coral Gables, Florida,
USA.
Barros, N. B., and D. K. Odell. 1990. Food habits of bottlenose
dolphins in the southeastern United States. Page 653 in S.
Leatherwood and R. R. Reeves, editors. The bottlenose
dolphin. Academic Press, San Diego, California, USA.
1716
LEIGH G. TORRES ET AL.
Barros, N. B., and R. S. Wells. 1998. Prey and feeding patterns
of resident bottlenose dolphins (Tursiops truncatus) in
Sarasota Bay, Florida. Journal of Mammalogy 79:1045–
1059.
Baumgartner, M. F., K. Mullin, L. N. May, and T. D. Leming.
2000. Cetacean habitats in the northern Gulf of Mexico.
Fisheries Bulletin 99:219–239.
Benoit-Bird, K. J., and W. W. L. Au. 2003. Prey dynamics
affect foraging by a pelagic predator (Stenella longirostris)
over a range of spatial and temporal scales. Behavioral
Ecology and Sociobiology 53:364–373.
Boyer, J. N., J. W. Fourqurean, and R. D. Jones. 1999.
Seasonal and long-term trends in the water quality of Florida
Bay (1989–1997). Estuaries 22:417–430.
Brotons, L., W. Thuiller, M. B. Araujo, and A. H. Hirzel. 2004.
Presence–absence versus presence-only modelling methods
for predicting bird habitat suitability. Ecography 27:437–448.
Buckland, S. T., D. R. Anderson, K. P. Burnham, J. L. Laake,
D. L. Borchers, and L. Thomas. 2001. Introduction to
distance sampling: estimating abundance of biological
populations. Oxford University Press, New York, New
York, USA.
Burnham, K. P., and D. R. Anderson. 1998. Model selection
and inference: a practical information-theoretic approach.
Springer-Verlag, New York, New York, USA.
Caro, T. M., and G. O’Doherty. 1999. On the use of surrogate
species in conservation biology. Conservation Biology 13:
805–814.
Cox, T., et al. 2006. Understanding the impacts of anthropogenic sound on beaked whales. Journal of Cetacean Research
and Management 7:177–187.
D’Amico, A., A. Bergamasco, P. Zanasca, S. Carniel, E.
Nacini, N. Portunato, V. Teloni, C. Mori, and R. Barbanti.
2003. Qualitative correlation of marine mammals with
physical and biological parameters in the Ligurian Sea.
IEEE Journal of Oceanic Engineering 28:29–43.
Davis, R., G. S. Fargion, N. May, T. D. Leming,
M. Brumgartner, W. E. Evans, L. J. Hansen, and K. Mullin.
1998. Physical habitat of cetaceans along the continental
slope in the north-central and western Gulf of Mexico.
Marine Mammal Science 14:490–507.
Durako, M., M. O. Hall, and M. Merello. 2002. Patterns of
change in the seagrass dominated Florida Bay hydroscape.
Pages 523–537 in J. W. Porter and K. G. Porter, editors. The
Everglades, Florida Bay, and coral reefs of the Florida Keys:
an ecosystem sourcebook. CRC Press, Boca Raton, Florida,
USA.
Engler, R., A. Guisan, and L. Rechsteiner. 2004. An improved
approach for predicting the distribution of rare and
endangered species from occurrence and pseudo-absence
data. Journal of Applied Ecology 41:263–274.
Fauchald, P. K., E. Erikstad, and H. Skarsfjord. 2000. Scaledependent predator–prey interactions: the hierarchical spatial
distribution of seabirds and prey. Ecology 81:773–783.
Ferguson, M. C., J. Barlow, P. Fiedler, S. B. Reilly, and
T. Gerrodette. 2006a. Spatial models of delphinid (family
Delphinidae) encounter rate and group size in the eastern
tropical Pacific Ocean. Ecological Modelling 193:645–662.
Ferguson, M. C., J. Barlow, S. B. Reilly, and T. Gerrodette.
2006b. Predictive Cuvier’s (Ziphius cavirostris) and Mesoplodon beaked whale densities as functions of the environment in the eastern tropical Pacific Ocean. Journal of
Cetacean Research Management 7:287–299.
Forney, K. A. 1999. Trends in harbour porpoise abundance off
central California, 1986–1995: Evidence for interannual
changes in distributions? Journal of Cetacean Research
Management 1:73–80.
Forney, K. A. 2000. Environmental models of cetacean
abundance: reducing uncertainty in populations trends.
Conservation Biology 14:1271–1286.
Ecological Applications
Vol. 18, No. 7
Fourqurean, J. W., and M. B. Robblee. 1999. Florida Bay: a
history of recent ecological changes. Estuaries 22:345–357.
Gaertner, J. C. 2000. Seasonal organization patterns of
demersal assemblages in the Gulf of Lions (north-western
Mediterranean Sea). Journal of Marine Biological Association of the United Kingdom 80:777–783.
Gregr, E. J., and A. W. Trites. 2001. Predictions of critical
habitat for five whale species in the waters of coastal British
Columbia. Canadian Journal of Fish and Aquatic Science 58:
1265–1285.
Guinet, C., L. Dubroca, M.-A. Lea, S. Goldsworthy, Y. Cherel,
G. Duhamel, F. Bonadonna, and J.-P. Donnay. 2001. Spatial
distribution of foraging in female Antarctic fur seals
Arctocephalus gazella in relation to oceanographic variables:
a scale-dependent approach using geographic information
systems. Marine Ecology Progress Series 219:251–264.
Guisan, A., T. C. Edwards, and T. Hastie. 2002. Generalized
linear and generalized additive models in studies of species
distributions: setting the scene. Ecological Modelling 157:89–
100.
Guisan, A., W. Thuiller, and N. Gotelli. 2005. Predicting
species distribution: offering more than simple habitat
models. Ecology Letters 8:993–1009.
Guisan, A., and N. E. Zimmermann. 2000. Predictive habitat
distribution models in ecology. Ecological Modelling 135:
147–186.
Hall, M. O., M. Durako, J. W. Fourqurean, and J. C. Zieman.
1999. Decadal changes in seagrass distribution and abundance in Florida Bay. Estuaries 22:445–459.
Hamazaki, T. 2002. Spatiotemporal prediction models of
cetacean habitats in the mid-western north Atlantic Ocean
(from Cape Hatteras, North Carolina, U.S.A. to Nova
Scotia, Canada). Marine Mammal Science 18:920–939.
Hastie, G. D., B. Wilson, L. J. Wilson, K. M. Parsons, and
P. M. Thompson. 2004. Functional mechanisms underlying
cetacean distribution patterns: hotspots for bottlenose
dolphins are linked to foraging. Marine Biology 144:397–403.
Hastie, T. J., and R. J. Tibshirani. 1990. Generalized additive
models. Chapman and Hall/CRC, Boca Raton, Florida,
USA.
Hedley, S. L., S. T. Buckland, and D. L. Borchers. 1999. Spatial
modelling from line transect data. Journal of Cetacean
Research Management 1:255–264.
Heithaus, M. R. 2001. Predator–prey and competitive interactions between sharks (order Selachii) and dolphins (suborder
Odontoceti): a review. Journal of Zoology 253:53–68.
Heithaus, M. R., and L. M. Dill. 2002. Food availability and
tiger shark predation risk influence bottlenose dolphin
habitat use. Ecology 83:480–491.
Heithaus, M. R., and L. M. Dill. 2006. Does tiger shark
predation risk influence foraging habitat use by bottlenose
dolphins at multiple spatial scales? Oikos 114:257–264.
Hooker, S. K., H. Whitehead, and S. Gowans. 1999. Marine
protected area design and the spatial and temporal distribution of cetaceans in a submarine canyon. Conservation
Biology 13:592–602.
Hugie, D. M., and L. M. Dill. 1994. Fish and game: a game
theoretic approach to habitat selection by predators and
prey. Journal of Fish Biology 45:151–169.
Levin, S. A. 1992. The problem of pattern and scale in ecology.
Ecology 73:1943–1967.
Lima, S. L. 2002. Putting predators back into behavioral
predator–prey interactions. Trends in Ecology and Evolution
17:70–75.
Logwerwell, E. A., and N. B. Hargreaves. 1996. The
distribution of sea birds relative to their fish prey off
Vancouver Island: opposing results at large and small spatial
scales. Fisheries Oceanography 5:163–175.
MacKenzie, D. I., J. D. Nichols, G. B. Lachman, S. Droege,
J. A. Royal, and C. Langtimm. 2002. Estimating site
October 2008
MODELING MARINE PREDATOR HABITAT
occupancy rates when detection probabilities are less than
one. Ecology 83:2248–2255.
Mantel, N. 1967. The detection of disease clustering and a
generalized regression approach. Cancer Research 27:209–
220.
Matheson, R. E. J., D. K. Camp, S. M. Sogard, and K. A.
Bjorgo. 1999. Changes in seagrass-associated fish and
crustacean communities in Florida Bay mud banks: The
effects of recent ecosystem changes? Estuaries 22:534–551.
McIvor, C. C., J. A. Ley, and R. D. Bjork. 1994. Changes in
freshwater inflow from the Everglades to Florida Bay
including effects on biota and biotic processes: a review.
Pages 117–146 in S. M. Davis and J. C. Ogden, editors.
Everglades: the ecosystem and its restoration. St. Lucie Press,
Delrey Beach, Florida, USA.
Oksanen, J., and P. R. Minchin. 2002. Continuum theory
revisited: What shape are species responses along ecological
gradients? Ecological Modelling 157:119–129.
Piatt, J. F., and D. A. Methven. 1992. Threshold foraging
behavior of baleen whales. Marine Ecology Progress Series
84:205–210.
Redfern, J. V., et al. 2006. Techniques for cetacean-habitat
modeling. Marine Ecology Progress Series 310:271–295.
Rushton, S. P., S. J. Ormerod, and G. Kerby. 2004. New
paradigms for modelling species distributions? Journal of
Applied Ecology 41:193–200.
Schick, R. S., and D. Urban. 2000. Spatial components of
bowhead whale (Balaena mysticetus) distribution in the
Alaskan Beaufort Sea. Canadian Journal of Fish and
Aquatic Science 57:2193–2200.
Schneider, D., and J. F. Piatt. 1986. Scale-dependent correlation of seabirds with schooling fish in a coastal ecosystem.
Marine Ecology Progress Series 32:237–246.
Segurado, P., and M. B. Araujo. 2004. An evaluation of
methods for modelling species distributions. Journal of
Biogeography 31:1555–1568.
Sih, A. 1998. Game theory and predator–prey response races.
Pages 221–238 in L. A. Dugatkin and H. K. Reeve, editors.
Advances in game theory and the study of animal behavior.
Oxford University Press, New York, New York, USA.
Sogard, S. M., G. V. N. Powell, and J. G. Holmquist. 1989.
Spatial distribution and trends in abundance of fishes
residing in seagrass meadows on Florida Bay mudbanks.
Bulletin of Marine Science 44:179–199.
Thayer, G. W., and A. J. Chester. 1989. Distribution and
abundance of fishes among basin and channel habitats in
Florida Bay. Bulletin of Marine Science 44:200–219.
Thayer, G. W., G. V. N. Powell, and D. E. Hoss. 1999.
Composition of larval, juvenile, and small adult fishes
relative to changes in environmental conditions in Florida
Bay. Estuaries 22:518–533.
Torres, L. G., M. R. Heithaus, and B. Delius. 2006. Influence of
teleost abundance on the distribution and abundance of
sharks in Florida Bay, USA. Hydrobiologia 596:449–455.
Torres, L. G., P. E. Rosel, C. D’Agrosa, and A. J. Read. 2003.
Improving management of overlapping bottlenose dolphin
ecotypes through spatial analysis and genetics. Marine
Mammal Science 19:502–514.
Torres, L. G., and D. Urban. 2005. Using spatial analysis to
assess bottlenose dolphins as an indicator of healthy fish
habitat. Pages 423–435 in S. A. Bortone, editor. Estuarine
indicators. CRC Press, Boca Raton, Florida, USA.
Vaughan, I. P., and S. J. Ormerod. 2005. The continuing
challenges of testing species distribution models. Journal of
Applied Ecology 42:720–730.
Wiley, T. R., and C. A. Simpfendorfer. 2007. The ecology of
elasmobranchs occurring in the Everglades National Park,
Florida: implications for conservation and management.
Bulletin of Marine Science 80:171–189.
Zieman, J. C., J. W. Fourqurean, and R. L. Iverson. 1989.
Distribution, abundance and productivity of seagrasses and
macroalgae in Florida Bay. Bulletin of Marine Science 44:
292–311.
APPENDIX A
Methods on the calibration of YSI-based chlorophyll a readings (Ecological Archives A018-059-A1).
APPENDIX B
Pseudo-absences generation (Ecological Archives A018-059-A2).
APPENDIX C
Explanation of Mantel’s tests (Ecological Archives A018-059-A3).
APPENDIX D
Model evaluation (Ecological Archives A018-059-A4).
APPENDIX E
Exploratory exercise (Ecological Archives A018-059-A5).
APPENDIX F
Predictive results for hypothesis 1 (H0 1) (Ecological Archives A018-059-A6).
APPENDIX G
Predictive results for hypothesis 2 (H0 2) (Ecological Archives A018-059-A7).
View publication stats
1717