Content uploaded by Peter Corkeron
Author content
All content in this area was uploaded by Peter Corkeron
Content may be subject to copyright.
ENDANGERED SPECIES RESEARCH
Endang Species Res
Vol. 15: 39– 52, 2011
doi: 10.3354/esr00367 Published online October 21
INTRODUCTION
Habitat models, spatial models, and conservation
planning
Habitat models have evolved from Hutchinson’s
(1957) concept of niche as environmental hyperspace
(Basille et al. 2008). In general, habitat models are
used to inform cetacean management by attempting
to understand the relationships between cetaceans
and their environment, from which inference is then
drawn on space use (e.g. Johnston et al. 2005,
Cañadas & Hammond 2008, Redfern et al. 2008,
Stafford et al. 2009). In order to develop spatially-
based approaches to cetacean conservation, is this
the only way forward?
© Inter-Research 2011 · www.int-res.com*Email: peter.corkeron@noaa.gov
Spatial models of sparse data to inform cetacean
conservation planning: an example from Oman
Peter J. Corkeron1, 2, 3,8,*, Gianna Minton,4,5,Tim Collins4, 6, Ken Findlay7,
Andrew Willson4, Robert Baldwin4
1Integrated Statistics, Woods Hole, Massachusetts 02543, USA
2Bioacoustics Research Program, Cornell Lab of Ornithology, Ithaca, New York 14850, USA
3The New England Aquarium, Central Wharf, Boston, Massachusetts 02110-3399, USA
4Environment Society of Oman, Ruwi, Sultanate of Oman
5Sarawak Dolphin Project, Institute of Biodiversity and Environmental Conservation, Universiti Malaysia Sarawak,
94300 Kota Samarahan, Sarawak, Malaysia
6Ocean Giants Program, Wildlife Conservation Society, Bronx, New York 10460-1099, USA
7MaRe, Oceanography Department, University of Cape Town, Rondebosch 7701, South Africa
8Present address: NOAA Northeast Fisheries Science Center, Woods Hole, Massachusetts 02543, USA
ABSTRACT: Habitat models are tools for understanding the relationship between cetaceans and their
environment, from which patterns of the animals’ space use can be inferred and management strate-
gies developed. Can working with space use alone be sufficient for management, when habitat can-
not be modeled? Here, we analyzed cetacean sightings data collected from small boat surveys off the
coast of Oman between 2000 and 2003. The waters off Oman are used by the Endangered Arabian
Sea population of humpback whales. Our data were collected primarily for photo-identification, using
a haphazard sampling regime, either in areas where humpback whales were thought to be relatively
abundant, or in areas that were logistically easy to survey. This leads to spatially autocorrelated data
that are not amenable to analysis using standard approaches. We used quasi-Poisson generalized lin-
ear models and semi-parametric spatial filtering to assess the distribution of humpback and Bryde’s
whales in 3 areas off Oman relative to 3 simple physiographic variables in a survey grid. Our analysis
focused on the spatial eigenvector filtering of models, coupled with the spatial distribution of model
residuals, rather than just on model predictions. Spatial eigenvector filtering accounts for spatial
autocorrelation in models, allowing inference to be made regarding the relative importance of partic-
ular areas. As an exemplar of this approach, we demonstrate that the Dhofar coast of southern Oman
is important habitat for the Arabian Sea population of humpback whales. We also suggest how con-
servation planning for mitigating impacts on humpback whales off the Dhofar coast could start.
KEY WORDS: Spatial eigenvector models · Spatial planning · Marine Protected Area · Generalized
linear models · Oman · Whales
Resale or republication not permitted without written consent of the publisher
Contribution to the Theme Section: ‘Beyond marine mammal habitat modeling’
O
PEN
PEN
A
CCESS
CCESS
Endang Species Res 15: 39– 52, 2011
Science is generally used to mitigate inadvertent
anthropogenic mortality of cetaceans in a series of
steps: estimating the abundance of the population
of interest; determining population structure and
boundaries (e.g. Taylor et al. 2000); estimating an -
thropogenic mortality of the population; then model-
ing the likely sustainability of this mortality (e.g.
Wade 1998). With these scientific inputs, managers
and stakeholders can devise measures that should
reduce mortality to a level that will allow adequate
mitigation within the social and cultural norms of the
people using the area over which the cetacean popu-
lation ranges.
The view of oceans underpinning this process is
one in which the marine environment was viewed
as generally undisturbed, with patches of impact,
which has recently been questioned (e.g. Crowder
et al. 2006). This has led, in some nations, to the
view that marine zoning, i.e. the marine equivalent
of terrestrial conservation planning (Margules &
Pressey 2000), is a more appropriate paradigm to
adopt (e.g. Fernandes et al. 2005). When providing
data on habitat use by cetaceans to inform spa-
tially-explicit conservation planning (e.g. Parra et
al. 2006a), knowing the absolute abundance of
cetaceans becomes less crucial than in the tradi-
tional approach. Note also that knowledge of the
ecological (or social) processes driving spatial dis-
tribution of whales, although very useful, is not
necessarily more important than simply having a
well-quantified understanding of what the spatial
distribution is. At this point, the interaction between
model outputs of space use, and the management
and policy milieu of the area of interest, is what
matters (e.g. Grech & Marsh 2008).
Spatial autocorrelation
Spatial autocorrelation (i.e. the closer samples are,
the more similar they are) introduces challenges
when making inference from models, as standard
errors of fixed effects from linear models are likely
underestimated (Dormann 2009). Recently, several
modeling techniques that can account for spatial
autocorrelation have been brought to the attention of
ecologists (Dormann et al. 2007). Even when working
with data from surveys specifically designed to pro-
duce distance-sampled estimates of abundance (e.g.
Gómez de Segura et al. 2007, Redfern et al. 2008),
spatial autocorrelation arising from niche-related or
social factors must be considered. However, when
working from vessels of opportunity, or from field
data where the principal aims did not require sys-
tematic or random sampling for survey coverage,
another source of autocorrelation needs considera-
tion, viz. that introduced by the haphazard sampling
regime.
Note that here we are using ‘haphazard’ techni-
cally to refer to sampling that is not explicitly ran-
domized, nor a fixed sampling regime starting from a
randomized point. It is generally assumed in these
instances that simply accounting for effort is suffi-
cient (e.g. Macleod et al. 2004). However, papers
analyzing haphazardly-collected data that then
account for effort rarely analyze model residuals to
demonstrate that spatial autocorrelation has been
handled satisfactorily by the model used (e.g.
Macleod et al. 2004).
Arabian Sea humpback whales
Here we address this problem by modeling the dis-
tribution of the Arabian Sea population of humpback
whales Megaptera novaeangliae off the coast of the
Sultanate of Oman (hereafter, Oman). The Arabian
Sea humpback whales are the only known non-
migratory population of humpback whales, and were
designated as an Endangered subpopulation in the
2008 revision of the IUCN Red List for cetacean spe-
cies (Minton et al. 2008). The current population,
estimated to number less than 100 individuals, does
not appear to be recovering from depletion due to
whaling in the 1960s (Minton et al. in press). Data
from photo-identified individuals (Minton et al. 2010)
and genetics (Rosenbaum et al. 2009) demonstrate
that this population is isolated from the nearest
neighboring Indian Ocean populations.
The distribution of the Arabian Sea population of
humpback whales is assumed to include the waters
of other nations (Fig. 1), particularly the Islamic
Republic of Pakistan, India, the Islamic Republic of
Iran, and the Republic of Yemen (Minton et al.
2010), but dedicated survey effort in these nations’
waters is either absent or very limited (e.g. Braulik
et al. 2010). Thanks to the combination of historical
records (Mikhalev 2000), continued research efforts
in recent years, a history of attendance at Interna-
tional Whaling Commission meetings, and the
establishment of a well networked non-government
organization as a platform to support conservation-
based research, Oman has become the range state
of primary importance for the research and protec-
tion of the Arabian Sea population of humpback
whales.
40
Corkeron et al.: Spatial models for cetacean conservation planning 41
Fig. 1. Study area, showing bottom topography. (A) Arabian Sea and environs, showing countries named in the text. White box
delineates area shown in detail in (B). (B) Study area in the waters off the Sultanate of Oman in detail. Black boxes delineate
the 3 study areas described in the text
A
B
Endang Species Res 15: 39– 52, 2011
The available data present several challenges for
modeling. As the Arabian Sea population of hump-
backs is small, there are few sightings of individuals,
and those sightings are clustered. The data were col-
lected by small boat, primarily for photo-identifica-
tion and genetic sampling, so surveys were haphaz-
ard, with coverage affected by logistical constraints,
and with a concentration on areas where whales
were likely to be encountered. Furthermore, as sur-
veys were conducted by researchers on a volunteer
basis, survey timing had to fit around researchers’
normal occupations, and were spread out over 4 yr
(although they were timed to coincide with the likely
presence of humpback whales as indicated by his-
toric whaling records; Mikhalev 2000).
For these reasons, previous publications from these
surveys have either been descriptive (e.g. Minton et
al. 2011), or provided results from photo-identifica-
tion (e.g. Minton et al. 2010, in press) or genetic stud-
ies (e.g. Rosenbaum et al. 2009). Our aim in this
paper was to develop a spatial model from the Oman
sightings data in order to identify the areas of great-
est relative abundance of humpback whales off the
Oman coast. We could not use biological oceano-
graphic predictors for our model, given the timing
over which data were collected and the relatively few
sightings in each year (Minton et al. 2011), and
because oceanographic processes off
Oman are driven largely by mon-
soonal conditions. The timing of the
monsoon, and its strength, varies
between years (Burkill 1999). That
being so, we chose simple physio-
graphic predictor variables for our
model.
This means that although we are
developing spatial models, we are
not constructing habitat models from
our data. Instead, we show how rel-
atively simple spatial models, based
on data that violate most models’
assumptions of spatial indepen-
dence, can still provide the scientific
foundation for management action.
In doing so, we aim to demonstrate
how others with similar sightings
data and issues of spatial autocorre-
lation can extract statistically robust,
meaningful results that can be used
to inform conservation measures. As
we have survey data for several
cetacean species, we used the dif-
ferences between humpback and
Bryde’s whales Balaenoptera sp. in model results to
begin to differentiate between spatial autocorrela-
tion caused by species-specific ecological factors
from those due to haphazard sampling.
MATERIALS AND METHODS
Study area and field techniques
Table 1 shows a summary of survey dates and dis-
tances covered on effort. Fig. 2 shows effort within
the study grid (see below). Full details of the survey
design and field methods are given by Minton et al.
(2011). Small boat surveys (most frequently a 6.5 m
rigid-hulled inflatable boat) were run between Janu-
ary 2000 and October 2003 in 3 areas: off Muscat, the
Gulf of Masirah, and the Dhofar coast (Fig. 1). Sur-
veys were generally conducted on a monthly basis in
the Muscat region through most of the study period;
the Gulf of Masirah was surveyed in October and
November, and the Dhofar coast was surveyed in
February and March. As the research focus was on
humpback whales, areas of known or suspected
humpback whale distribution were targeted, based
on historical data (e.g. Wray & Martin 1983,
Mikhalev 2000) and anecdotal reports. The excep-
42
Survey area Survey dates Effort hours
Muscat
Monthly surveys 15 Mar 2001 − 15 Jul 2003 104.21
Dhofar
Hallaniyat Islands 15−24 Jan 2000, 8−21 Feb 2000 63.5
Dhofar 9−22 Feb 2001 34.26
Dhofar 10 Feb − 2 Mar 2002 62.37
Hasik Bay 24−26 Jun 2002 4.32
Sharbitat and Hallaniyats 17−20 Nov 2002 36.83
Dhofar 24 Feb − 19 Mar 2003 116.31
Dhofar (Hasik only) 15−17 May 2003 2.17
Total 319.76
Gulf of Masirah
N. Gulf of Masirah 15−17 Oct 2000 11
Gulf of Masirah 4−27 Oct 2001 83.15
Gulf of Masirah 24 Oct − 16 Nov 2002 58.2
Total 152.35
Other areas
Ras al Hadd 30 Mar − 2 Apr 2001 8.13
Shore-based observations
Duqm 10−13 Jun 2001 25
Table 1. Dates and locations of small boat surveys in Oman. Effort indicates
time spent actively searching for whales and excludes time spent working
with whales, in transit, or on breaks
Corkeron et al.: Spatial models for cetacean conservation planning
tion to this was the area around Muscat, as authors
who ran the field surveys lived there. Within each of
the 3 survey areas, tracks were designed to provide
as much coverage of the area as possible within the
logistic and safety limitations of daily excursion small
boat surveys.
Survey tracklines generally followed an irregular
saw-tooth pattern along the coast, and were tra-
versed at speeds of 12 to 15 knots (22 to 28 km h−1).
Search effort was suspended when cetaceans were
sighted and groups were approached to confirm spe-
cies identity and collect data (e.g. photographs to
identify individual animals, biopsy samples). Search-
ing stopped in Beaufort states of 4 or higher. Sighting
positions and other positional data were recorded
using Garmin 12 or 12XL GPS units. Tracks were
logged, with the vessel’s position recorded every 30
to 45 s. Georeferenced data were imported into
ArcView®3.2a (ESRI: www.esri.com) and checked at
the end of each day. Sightings data were stored in an
MS Access®database.
Geoprocessing
Sightings data were overlaid onto a 0.1 × 0.1°
lat./long. grid (at these latitudes, approximately 11 ×
11 km). Grid cell size was determined as a compro-
mise between accuracy in classifying habitat charac-
teristics within grid cells and the need for sufficient
encounters within each cell to yield usable results
(e.g. Hamazaki 2002). On-effort portions of survey
tracks were imported into ArcGIS (ESRI, WGS84 pro-
jection) and converted into shape files, one for each
day’s effort. The geo-processing ‘intersect’ and ‘dis-
solve’ functions of ArcGIS were then used to calcu-
late the total distance (in decimal degrees) surveyed
on-effort in each cell. The ‘spatial join’ function of
ArcGIS was used to calculate the total number of
cetacean groups, by species, in each cell, from the
MS Access®database.
Digitized depth files were generated for each sur-
vey area using rasterized nautical charts (British
Admiralty Raster Chart Series, British Admiralty
chart nos. 2851, 2828, 2896, 3519, 3522, 3784, and
3785). Depth files were interpolated using ArcGIS
Spatial Analyst to generate depth rasters for the grid,
with a mask applied to exclude terrestrial surfaces
from grid cells overlapping the coast. Minimum and
maximum values for slope and depth were calculated
from the rasters for each grid cell. All geoprocessing
was conducted by G. Minton.
The ArcGIS shape file of the Oman coast
(WGS84 projection) was imported into R (R Devel-
opment Core Team 2010) using the maptools v0.7-
34 (package Lewin-Koh et al. 2010), and converted
into a SpatialLines object. The center point of
each grid square was also read into R as a Spa-
tialPoints object, using the same projection. Both
objects were then projected to UTM (zone 40Q).
In order to calculate the distance to shore for the
center of each grid square, the SpatialPoints and
SpatialLines objects were transformed into spatial
point patterns and line segment patterns, respec-
tively, using the spatstat v1.21-2 (package Badde-
ley & Turner 2005). The ‘nncross’ command was
used to calculate distances.
Model construction
Our data were counts, and although the humpback
data were approximately Poisson distributed, the
Bryde’s whale data were not. As we needed a model-
ing approach that would be consistent across both
species, we used quasi-Poisson generalized linear
models (GLMs) with log-link (Venables & Ripley
2002). To account for survey effort differing across
grid cells, the natural log of on-effort distance for
each cell was included as an offset. Mapping the
43
Fig. 2. On-effort survey (in decimal degrees) for cetaceans
surveyed off Oman coastal waters, 2000 to 2003. Grid is that
used for all analyses
Endang Species Res 15: 39– 52, 2011
residuals of GLMs (see below) showed spatial pat-
terning, so further analysis was undertaken. We used
spatial eigenvector mapping (SEVM; Dormann et al.
2007) to account for residual spatial autocorrelation
(SAC), as SEVM builds on GLM results. Also, as Dor-
mann et al. (2007, p. 612) noted, it is a method that
‘could thus be very useful for data with SAC stem-
ming from larger scale observation bias,’ and we
know that biases due to haphazard design confound
our data.
SEVM works by ‘whitening out’ residual spatial
autocorrelation in a model, rather
than incorporating it into the model.
We used the ‘ME’ command from the
spdep 0.5-16 (package Bivand et al.
2010), which takes a brute force
approach to finding the smallest pos-
sible subset of eigenvectors that
removes residual spatial autocorrela-
tion from a GLM. The residual auto-
correlation is then accounted for by
refitting the original model with the
eigenvectors included as covariates
(for further details, see Dormann et al.
2007, Bivand et al. 2008). Hereafter,
we refer to these as SEVM-GLMs. A
flow diagram outlining the process of
model construction and listing com-
mands used is provided in Fig. 3.
Analyses were run using R 2.11.1
(R Development Core Team 2010)
through rgedit 0.7.0.1 on an x86 com-
puter running Ubuntu 9.04.
RESULTS
Of the slope and depth values, max-
imum depth and minimum slope
were the least correlated and so were
selected for inclusion in the model.
Distance from shore was not strongly
correlated with any other physio-
graphic variable. Of the 3 separate
study areas, the Muscat and Dhofar
coasts are relatively similar in topo -
graphy, with deep water within 1 grid
cell of shore. The Gulf of Masirah, on
the other hand, includes one of the
largest areas of shallow waters any-
where off the Arabian Sea coast of
Oman (approximately 80 km at the
widest, see Fig. 1B), with a gently
sloping shelf extending to the outer edge of the sur-
vey area. Coefficients for the GLMs for both species
are shown in Table 2.
To calculate spatial eigenvectors, first we con-
structed a neighborhood, then calculated eigenvec-
tors (see Fig. 3 for details). The distribution of Bryde’s
whales showed little spatial autocorrelation, so we
used an alpha value of 0.25 as a stopping rule for
eigenvector calculation. All neighborhoods are, by
definition, confined within one of the 3 separate study
areas, i.e. the Dhofar coast, the Gulf of Masirah, or off
44
Estimate SE tp r(>|t|)
Megaptera
Intercept −2.42 × 10−1 2.04 × 10−1 −1.182 0.238
DepthMax 5.83 × 10−4 2.62 × 10−4 2.228 0.0268*
DistShore −1.46 × 10−5 1.69 × 10−5 −0.862 0.390
SlopeMin −3.74 × 10−6 1.95 × 10−6 −1.915 0.0566
Balaenoptera
Intercept −2.59 4.13 × 10−1 −6.256 < 0.001***
DepthMax −2.55 × 10−4 6.21 × 10−4 −0.410 0.682
DistShore 4.04 × 10−5 2.19 × 10−5 1.842 0.0667
SlopeMin 1.93 × 10−6 1.36 × 10−6 1.418 0.157
Table 2. Coefficients from the quasi-Poisson generalized linear models of
humpback whales Megaptera novaeangliae and Bryde’s whales Balaenoptera
sp. surveyed off Oman coastal waters, 2000 to 2003, with results
of significance testing. *p < 0.05, ***p < 0.001
Fig. 3. Process by which the choice to use a generalized linear model (GLM)
or a spatial eigenvector mapping GLM (SEVM-GLM) is made. Text in italics
indicates commands used from the R packages indicated in bold
Corkeron et al.: Spatial models for cetacean conservation planning
Muscat. Coefficients for the SEVM-GLMs for both
species are shown in Table 3. Tests comparing the fits
of GLMs and SEVM-GLMs are given in Table 4.
We plotted the results of GLMs and the SEVM-
GLMs side by side in order to display their differ-
ences. Plots of the predicted values generated by
GLMs and SEVM-GLMs are shown in Fig. 4, with
residuals in Fig. 5. Note that predicted values are for
counts of groups of cetaceans (the unit of sighting)
per grid square over the entire period of the study, so
the maps show spatial patterns of relative abun-
dance. For both species, the standard errors esti-
mated for the SEVM-GLMs were greater than those
for the GLMs (as expected), and so these results
are not mapped. Fig. 6 shows maps of eigenvector
values.
For humpback whales, the most important habitat
variable identified in the GLM was depth. The 3
eigenvectors extracted fell clearly into the 3 survey
areas: off Muscat, the Gulf of Masirah, and the Dho-
far coast, respectively (Fig. 6A). The SEVM-GLMs
fitted the data much better than the GLM, with slope
and depth appearing important. As expected, the
SEVM-GLM residuals are substantially smaller and
less spatially clustered than the GLM residuals
(Fig. 5A).
Bryde’s whales produced a very different result
from humpbacks. The distribution of sightings across
the study area was more even than Poisson (disper-
sion parameter for the GLM was ~0.6). There was
relatively little spatial autocorrelation in the initial
GLM, and the SEVM extracted only 1 eigenvector
(Fig. 6B), in which autocorrelation in the Muscat
study area predominated. Habitat variables iden -
tified as being important in the SEVM-GLMs are
distance from shore and slope. Both models’ predic-
tions give relatively similar patterns, with the great-
est relative abundance of Bryde’s whales off Muscat
and, to a lesser extent, off the Dhofar coast (Fig. 4B).
Mapping residuals (Fig. 5B) suggested that the
SEVM-GLM predicts the relative abundance of
Bryde’s whales somewhat better than does the
GLM, although there is a small area
in the Gulf of Masirah where neither
model predicted their relative abun-
dance well.
DISCUSSION
Context
The mark−recapture estimate of
abundance for the Arabian Sea pop-
ulation of humpback whales (82;
95% CI: 60−111; Minton et al. in
press) is from data that are now
almost a decade old. Relying on a
time series of mark−recapture esti-
mates of a cetacean population of
around 100 animals to determine
trends in abundance is futile (e.g.
Thompson et al. 2000, Parra et al.
2006b). Analysis of scarring on the
caudal peduncle region of photo-
graphically identified humpback
whales in Oman in 2003 indicated
that 30 to 40% of all whales exam-
ined were likely to have been
involved in entanglements with fish-
ing gear (Minton et al. in press).
Despite this apparently high level of
interaction with fisheries, there are
no estimates of fisheries-related mor-
45
Estimate SE t p r(>|t|)
Megaptera
Intercept −7.10 × 10−1 2.49 × 10−1 −2.853 0.005**
DepthMax 4.11 × 10−4 1.91 × 10−4 2.150 0.033*
DistShore −9.02 × 10−6 1.48 × 10−5 −0.611 0.542
SlopeMin −3.29 × 10−6 1.24 × 10−6 −2.644 0.009**
fitted(meg.ME.quasi)vec1 −16.6 6.06 −2.747 0.007**
fitted(meg.ME.quasi)vec5 8.84 1.58 5.605 <0.001***
fitted(meg.ME.quasi)vec4 8.67 2.06 4.205 <0.001***
Balaenoptera
Intercept −3.17 5.50 × 10−1 −5.767 <0.001***
DepthMax −9.69 × 10−4 8.83 × 10−4 −1.097 0.274
DistShore 6.03 × 10−5 2.42 × 10−5 2.489 0.015*
SlopeMin 3.62 × 10−6 1.71 × 10−6 2.117 0.035*
fitted(bal.ME.quasi) −8.44 2.55 −3.310 0.001**
Table 3. Coefficients from spatial eigenvector mapping of quasi-Poisson
generalized linear models of cetaceans surveyed off Oman coastal wa-
ters, 2000 to 2003. For species information see Table 2. *p < 0.05, **p < 0.01,
***p < 0.001
Residual df Residual dev. Δdf ΔDeviance p (>|χ|)
Megaptera
GLM 242 163.450
SEVM-GLM 239 100.810 3 62.637 <0.001***
Balaenoptera
GLM 242 48.284
SEVM-GLM 241 40.535 1 7.749 <0.001***
Table 4. Tests comparing fits of generalized linear models (GLMs) and spatial
eigenvector mapping of quasi-Poisson GLMs of cetaceans surveyed off Oman
coastal waters, 2000 to 2003. For species information see Table 2. Dev.:
deviance; ***p < 0.001
Endang Species Res 15: 39– 52, 2011
tality for this population. It is thus unrealistic to
expect that the now-traditional approach outlined
in the Introduction (estimate abundance, estimate
anthropogenic mortality and model sustainability of
anthropogenic mortality), can be implemented in a
timely manner for this population.
46
Fig. 4. Megaptera novaeangliae and Balaenoptera sp. Predicted numbers of whale groups for each grid square from the quasi-
Poisson generalized linear models (GLM Predicted), and spatial eigenvector mapping of quasi-Poisson generalized linear
models (SEVM Predicted) of cetaceans surveyed off Oman coastal waters, 2000 to 2003. (A) Humpback whales, (B) Bryde’s whales
A
B
Corkeron et al.: Spatial models for cetacean conservation planning
How then can scientific input inform plans to
manage anthropogenic activities impacting these
whales? Further, are there any results from our study
that can inform management more generally? In our
experience, these related problems, i.e. small popu-
lation, no reliable quantification of anthropogenic
47
Fig. 5. Megaptera novaeangliae and Balaenoptera sp. Model residuals for each grid square from the quasi-Poisson general-
ized linear models (GLM Predicted) and spatial eigenvector mapping of quasi-Poisson GLMs (SEVM Predicted) of cetaceans
surveyed off Oman coastal waters, 2000 to 2003. (A) Humpback whales, (B) Bryde’s whales
A
B
Endang Species Res 15: 39– 52, 2011
mortality or population trends, and limited resources,
are not uncommon in most of the developing world.
Further, throughout the world, there is likely to be
a large body of data collected using haphazard sam-
pling methods, e.g. cetacean sighting data with asso-
ciated effort (note that effort data are essential), but
collected on platforms of opportunity, or with addi-
tional survey aims that grossly violate the assump-
tions of line- or strip-transect sampling. Haphazard
sampling for photo-identification and genetic sam-
pling is common, as it is for coastal patrols in marine
protected areas. How can cetacean biologists make
best use of such data to inform conservation plan-
ning? Here we show why SEVMs may be the most
appropriate modern tool for analyzing this type of
spatially autocorrelated data.
48
Fig. 6. Megaptera novaeangliae and Balaenoptera sp. Eigenvalues for
each grid square from spatial eigenvector mapping of quasi-Poisson
GLMs (SEVM eigenvectors) of cetaceans surveyed off Oman coastal
waters, 2000 to 2003. (A) Humpback whales, (B) Bryde’s whales
A
B
Corkeron et al.: Spatial models for cetacean conservation planning
Other modeling approaches
Before discussing the results from our models, we
outline why we did not use other approaches to habi-
tat modeling.
Mixed models
We initially attempted to run generalized linear
mixed models (GLMMs) with spatial structure to the
random effect, as outlined by Dormann et al. (2007).
First, we note that a caveat with this technique is that
those authors refer to using a ‘random’ effect with
only 1 category as a ‘cheat’ (Dormann et al. 2007,
their supplementary material). We ran quasi-Poisson
(and Poisson) GLMMs, but they produced non-posi-
tive definite approximate variance−covariance matri-
ces, making it impossible to check the confidence
intervals of the ‘random’ effect. We therefore did not
pursue this line of analysis.
Additive models
Despite the popularity of generalized additive
models (GAMs) in modeling cetacean habitat use
(e.g. Gómez de Segura et al. 2007, Cañadas & Ham-
mond 2008, Redfern et al. 2008), we did not use them,
for 3 reasons. First, with relatively few samples for
whale species (56 sightings over 4 yr for humpback
whales, 15 for Bryde’s whales, Table 1), techniques
such as GAMs that can handle spatial autocorrela-
tion by seeking nonlinear patterns in the raw data
themselves become inherently less useful. Second,
we wanted to compare model outputs between spe-
cies in order to differentiate between autocorrelation
due to haphazard survey design, and autocorrelation
due to cetacean ecology. As GAMs, by definition, fit
nonlinear curves to the data, we judged that they
were less likely to allow us to make these distinc-
tions. Finally, the SEVM-GLM approach allows a
clear comparison to be made between a model that is
unlikely to successfully account for spatial autocorre-
lation (the GLM) with one that does (the SEVM-
GLM). Differences between these model predictions
allow us to make some inference on the manner in
which spatial autocorrelation influences these pre-
dictions, and so provides another form of insight into
the driver(s) of the autocorrelation. A GAM-based
approach would not allow this.
General niche-environment system factor analysis
Ecological niche factor analysis (ENFA), a form of
niche-environment factor analysis (Calenge & Basille
2008), has recently become a popular tool for devel-
oping habitat models for cetaceans (e.g. Oviedo &
Solís 2008, Praca et al. 2009). We did not use ENFA
on our data for 2 principal reasons. First, ENFA is, by
definition, based on Hutchinsonian niche hyper-
space. We did not attempt to model whale niches, as
we knew that we did not have environmental data
available that would be appropriate for niche model-
ing (see Introduction). Secondly, the mathematical
formulation for ENFA (Hirzel et al. 2002) does not
explicitly account for spatial autocorrelation, espe-
cially that due to haphazard sampling in a ‘design I’
habitat use study (as defined by Thomas & Taylor
1990), such as this one. As the main point of our mod-
eling exercise was to account for this form of spatial
autocorrelation, ENFA was inappropriate.
Details from Oman
Our model results confirm and provide a statisti-
cally robust underpinning for previous work based
on the same survey data (Minton et al. 2011). Our
results clearly demonstrate the importance of the
Dhofar coast, particularly in the region of the Hal-
laniyat Islands and Hasik, for the Arabian Sea popu-
lation of humpback whales over our study period.
Examination of the differences in outputs be -
tween models (i.e. with and without autocorrelation
accounted for), and between species, allow us to
identify the most reliable model outputs and thus the
information that is of greatest value for management.
For humpback whales, there are substantial differ-
ences between the model predictions of the GLM
and the SEVM-GLM (Fig. 4A). The GLM predicts the
greatest relative abundance of humpbacks to be off
Muscat, while the SEVM-GLM predicts most hump-
backs off the Dhofar coast. Examination of model
residuals (Fig. 5A) demonstrates that the SEVM-
GLM prediction is more robust. The GLM prediction
appears driven by the substantial search effort off
Muscat, and the similarity in habitat characteristics
between Muscat and the Dhofar coast (where most
humpback sightings were made). Note that model
bias is introduced into the GLM by focusing survey
effort in an area known to be important for hump-
backs that had similar physiographic characteristics
(i.e. the Dhofar coast). The SEVM-GLM successfully
handles this bias. Both models poorly predicted
humpback occurrence in the shallow coastal shelf
waters of the Gulf of Masirah (Fig. 4A), although the
SEVM-GLM fit is somewhat better.
49
Endang Species Res 15: 39– 52, 2011
The way in which model outputs for Bryde’s whales
differ from humpback whales is of interest. As the
effort data are from the same survey series, one
would expect the confounding effects of haphazard
sampling to be consistent, and as such, model differ-
ences would reflect biological/ecological traits of the
species rather than sampling artefacts. The maxi-
mum cell count for Bryde’s whales is approximately
an order of magnitude less than that for humpbacks.
The pattern of spatial autocorrelation in the data is
also different. The distribution of Bryde’s whales
across the study area is more regular, and they are
more prevalent off Muscat than are humpbacks. This
suggests that the clumped distribution of humpbacks
is real, as is the importance of the Dhofar coast for
humpbacks. This has important local, and regional,
implications for management.
The Arabian Sea population of humpback whales
is the smallest population of humpback whales
known to exist, the only population known not to
undertake an extensive seasonal migration, and one
of the most endangered baleen whale populations
(Minton et al. 2008). The Dhofar coast, in particular in
the region of the Hallaniyat Islands and Hasik, was
identified previously (Minton et al. in press) as likely
to be an important habitat for this population. Our
modeling work quantifies the significance of this
area for these whales.
Caveats
We considered it inappropriate to attempt to make
inference on parts of the Oman coast not covered by
the surveys. Although it is theoretically possible for
us to project model predictions into other areas, we
consider this inadvisable, as our basic design was not
to make inference about the distribution of hump-
back whales along the entire Oman coast. Given the
constraints under which field work was undertaken,
both logistic and financial, a synoptic survey of the
entire coast was impossible. We focused our survey
effort on areas which available information sug-
gested were likely the most important areas for
humpback whales.
An extension of these caveats is that although we
make recommendations on the relative importance
for conservation of the Dhofar coast (see below), we
cannot state with certainty that other areas will not
prove equally important. Nevertheless, the informa-
tion available is sufficient to note the importance of
starting the process of mitigating inadvertent anthro-
pogenic mortality on Arabian Sea humpback whales,
and that the science available suggests that the best
place to start is off the Dhofar coast. Recent (March
2011) fieldwork off the Dhofar coast, focusing near
the village of Hasik (Fig. 1) was planned based in
part on the results of our model. This field season
resulted in regular, multiple sightings of humpback
whales, and observations of feeding and breeding
behavior, confirming the area’s continued relative
importance (A. Willson pers. obs.). Unfortunately, the
threat of pirate activity offshore prevented field work
around the Hallaniyat Islands.
Implications for humpback whale conservation
The coastal zone of Oman is experiencing rapid
transformation as the country moves beyond a wholly
petroleum-dependent economy. Oman’s population
growth rate is among the highest in the world (3.14%
per annum), and there is a continuing demographic
shift towards coastal areas (Oman Ministry of
National Economy 2009). Fishing effort off the coast
of Oman and in other parts of the Arabian Sea is
increasing dramatically (Oman Ministry of Agricul-
ture and Fisheries 2002, FAO 2007, Oman Ministry of
National Economy 2009), and drifting and set gillnets
as well as traps are already widely used (Stengel & Al
Harthy 2002).
One of the most important findings of this study is
that the clustering of humpback whales along part of
the Dhofar coast and the Hallaniyat Islands is not a
sampling artefact, but a result of the whales’ ranging
behavior. This suggests that a spatially-explicit man-
agement program should be implemented along this
section of the Dhofar coast, as a preliminary step to
larger-scale marine conservation planning in Oman.
There are instances where declaring a marine pro-
tected area for cetacean conservation has not led to
cessation of threatening processes, particularly gill-
netting (e.g. Notarbartolo di Sciara et al. 2008). There
are also examples where it has been successful: the
implementation of netting restrictions to protect
dugongs, and the general process of rezoning in the
Great Barrier Reef Marine Park (Dobbs et al. 2008,
Grech et al. 2008) provides an example of how to
achieve spatially-explicit restrictions on netting. We
suggest that this process, suitably modified for
Omani cultural norms and local capacity for manage-
ment, start as soon as possible.
Finally, we suggest that other researchers working
with spatially autocorrelated data on cetacean (and
other marine wildlife) distribution give serious con-
sideration to using spatial eigenvector models.
50
Corkeron et al.: Spatial models for cetacean conservation planning
There will be instances where sampling regimes are
just too haphazard, and there are no options to dis-
tinguish between the likely causes of spatial auto-
correlation, where these models may prove ineffec-
tive. But in those instances, it is possible that any
other spatial modeling approach will be equally
ineffective. Researchers need to ensure that they do
not make inference beyond their data with this tool,
as is the case for all spatial or habitat models. Tools
to run spatial eigenvector models on all operating
systems are available for free download as part of
the R statistical language (R Development Core
Team 2010), and example code to run eigenvector
filtering is provided as an appendix to the paper by
Dormann et al. (2007).
Acknowledgements. Logistic support and research permits
for surveys were provided by Oman's Ministry of Environ-
ment and Climate Affairs, the Oman Natural History
Museum, and Oman's Ministry of Agriculture and Fisheries.
Fieldwork was supported by: The Ford Environmental
Grants, The UK Foreign and Commonwealth Office, Shell
Marketing Oman, Petroleum Development Oman, Veritas
Geophysical, The Peter Scott Trust for Education and
Research in Conservation, and the Marina Bandar al Row-
dah. The manuscript was improved by comments from
Daniel Palacios and two anonymous reviewers.
LITERATURE CITED
Baddeley A, Turner R (2005) Spatstat: an R package for ana-
lyzing spatial point patterns. J Stat Softw 12: 1−42
Basille M, Calenge C, Marboutin É, Andersen R, Galliard
JM (2008) Assessing habitat selection using multivariate
statistics: some refinements of the ecological-niche factor
analysis. Ecol Model 211: 233−240
Bivand R, Altman M, Anselin L, Assunção R and others
(2010) spdep: Spatial dependence: weighting schemes,
statistics and models. R package version 0.5-16. Avail -
able at http: //CRAN.R-project.org/package=spdep
Bivand RS, Pebesma EJ, Gómez-Rubio V (2008) Applied
spatial data analysis with R. Springer, New York, NY
Braulik GT, Ranjbar S, Owfi F, Aminrad T, Dakhteh SMH,
Kamrani E, Mohsenizadeh F (2010) Marine mammal
records from Iran. J Cetacean Res Manag 11: 49−64
Burkill PH (1999) ARABESQUE: an overview. Deep-Sea Res
II 46: 529−547
Calenge C, Basille M (2008) A general framework for the
statistical exploration of the ecological niche. J Theor
Biol 252: 674−685
Cañadas A, Hammond PS (2008) Abundance and habitat
preferences of the short-beaked common dolphin Del-
phinus delphis in the southwestern Mediterranean:
implications for conservation. Endang Species Res 4:
309−331
Crowder LB, Osherenko G, Young OR, Airame S and others
(2006) Resolving mismatches in U.S. ocean governance.
Science 313: 617−618
Dobbs K, Fernandes L, Slegers S, Jago B and others (2008)
Incorporating dugong habitats into the marine protected
area design for the Great Barrier Reef Marine Park,
Queensland, Australia. Ocean Coast Manag 51: 368−375
Dormann CF (2009) Response to Comment on ‘Methods to
account for spatial autocorrelation in the analysis of spe-
cies distributional data: a review’. Ecography 32: 379−381
Dormann CF, McPherson JM, Araújo MB, Bivand R and
others (2007) Methods to account for spatial autocorrela-
tion in the analysis of species distributional data: a
review. Ecography 30: 609−628
FAO (Food and Agriculture Organisation of the United
Nations) (2007) The state of world fisheries and aquacul-
ture 2006. FAO, Rome
Fernandes L, Day J, Lewis A, Slegers S and others (2005)
Establishing representative no-take areas in the Great
Barrier Reef: large scale implementation of theory on
marine protected areas. Conserv Biol 19: 1733−1744
Gómez de Segura A, Hammond PS, Cañadas A, Raga JA
(2007) Comparing cetacean abundance estimates de -
rived from spatial models and design-based line transect
methods. Mar Ecol Prog Ser 329: 289−299
Grech A, Marsh H (2008) Rapid assessment of risks to a
mobile marine mammal in an ecosystem-scale marine
protected area. Conserv Biol 22: 711−720
Grech A, Marsh H, Coles R (2008) A spatial assessment of
the risk to a mobile marine mammal from bycatch.
Aquatic Conserv 18: 1127−1139
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). Mar Mamm Sci 18: 920−939
Hirzel AH, Hausser J, Chessel D, Perrin N (2002) Ecological-
niche factor analysis: how to compute habitat-suitability
maps without absence data? Ecology 83: 2027−2036
Hutchinson GE (1957) Concluding remarks. Cold Spring
Harbor Symp Quant Biol 22: 415−427
Johnston DW, Westgate AJ, Read AJ (2005) Effects of fine-
scale oceanographic features on the distribution and
movements of harbour porpoises Phocoena phocoena in
the Bay of Fundy. Mar Ecol Prog Ser 295: 279−293
Lewin-Koh NJ, Bivand R, Pebesma EJ, Archer A and others
(2010) maptools: Tools for reading and handling spatial
objects. R package version 0.7-34. Available at http://
CRAN.R-project.org/package=maptools
Macleod K, Fairbairns R, Gill A, Fairbairns B, Gordon J,
Blair-Myers C, Parsons ECM (2004) Seasonal distribution
of minke whales Balaenoptera acutorostrata in relation
to physiography and prey off the Isle of Mull, Scotland.
Mar Ecol Prog Ser 277: 263−274
Margules CR, Pressey RL (2000) Systematic conservation
planning. Nature 405: 243−253
Mikhalev YA (2000) Whaling in the Arabian Sea by the
whaling fleets Slava and Sovetskaya Ukraina. In:
Yablokov AV, Zemsky VA, Tormosov DD (eds) Soviet
whaling data (1949−1979). Centre for Russian Environ-
mental Policy, Moscow, p 141−181
Minton G, Collins TJQ, Pomilla C, Findlay KP, Rosenbaum
HC, Baldwin R, Brownell RL Jr (2008) Megaptera novae -
angliae, Arabian Sea subpopulation. IUCN Red List of
Threatened Species. Available at www.iucnredlist.org/
details/132835
Minton G, Cerchio S, Collins T, Ersts P and others (2010) A
note on the comparison of humpback whale tail fluke
catalogues from the Sultanate of Oman with Madagascar
and the East African Mainland. J Cetacean Res Manag
11: 65−68
51
➤
➤
➤➤
➤
➤➤
➤
➤
➤
➤
➤
➤➤
➤
➤
➤
➤
➤
➤➤
Endang Species Res 15: 39– 52, 2011
52
Minton G, Collins TJQ, Findlay KP, Baldwin R (2011)
Cetacean distribution in the coastal waters of the Sul-
tanate of Oman. J Cetacean Res Manag 11: 301−313
Minton G, Collins T, Findlay K, Ersts P, Rosenbaum H,
Berggren P, Baldwin R (in press) Seasonal distribution,
abundance, habitat use and population identity of hump-
back whales in Oman. J Cetacean Res Manag (Spec
Issue Southern Hemisphere Humpback Whales)
Notarbartolo di Sciara G, Agardy T, Hyrenback D, Scovazzi
T, Van Klaveren P (2008) The Pelagos Sanctuary for
Mediterranean marine mammals. Aquatic Conserv 18:
367−391
Oman Ministry of Agriculture and Fisheries (2002) Fisheries
statistical year book 2001. Ministry of Agriculture and
Fisheries, Muscat
Oman Ministry of National Economy (2009) 2009 statistical
yearbook. Ministry of National Economy, Muscat
Oviedo L, Solís M (2008) Underwater topography deter-
mines critical breeding habitat for humpback whales
near Osa Peninsula, Costa Rica: implications for marine
protected areas. Rev Biol Trop 56: 591−602
Parra GJ, Schick R, Corkeron PJ (2006a) Spatial distribution
and environmental correlates of Australian snubfin and
Indo-Pacific humpback dolphins. Ecography 29: 396−406
Parra GJ, Corkeron PJ, Marsh H (2006b) Population sizes,
site fidelity and residence patterns of Australian snubfin
and Indo-Pacific humpback dolphins: implications for
conservation. Biol Conserv 129: 167−180
Praca E, Gannier A, Das K, Laran S (2009) Modelling the
habitat suitability of cetaceans: example of the sperm
whale in the northwestern Mediterranean Sea. Deep-
Sea Res I 56: 648−657
R Development Core Team (2010) R: a language and en -
vironment for statistical computing. R Foundation
for Statistical Computing, Vienna. Available at www.
R-project.org
Redfern JV, Barlow J, Ballance LT, Gerrodette T, Becker EA
(2008) Absence of scale dependence in dolphin−habitat
models for the eastern tropical Pacific Ocean. Mar Ecol
Prog Ser 363: 1−14
Rosenbaum HC, Pomilla C, Mendez M, Leslie MS and oth-
ers (2009) Population structure of humpback whales from
their breeding grounds in the South Atlantic and Indian
Oceans. PLoS ONE 4: e7318
Stafford KM, Citta JJ, Moore SE, Daher MA, George JE
(2009) Environmental correlates of blue and fin whale
call detections in the North Pacific Ocean from 1997 to
2002. Mar Ecol Prog Ser 395: 37−53
Stengel H, Al Harthy A (2002) The traditional fishery of the
Sultanate of Oman (fishing gear and methods). Ministry
of Agriculture and Fisheries, Directorate General of Fish-
eries Resources, Marine Science and Fisheries Center,
Muscat
Taylor BL, Wade PR, DeMaster DP, Barlow J (2000) Incorpo-
rating uncertainty into management models for marine
mammals. Conserv Biol 14: 1243−1252
Thomas D, Taylor E (1990) Study designs and tests for com-
paring resource use and availability. J Wildl Manag 54:
322−330
Thompson PM, Wilson B, Grellier K, Hammond PS (2000)
Combining power analysis and population viability
analysis to compare traditional and precautionary
approaches to conservation of coastal cetaceans. Con-
serv Biol 14: 1253−1263
Venables WN, Ripley BD (2002) Modern applied statistics
with S, 4th edn. Springer, New York, NY
Wade PR (1998) Calculating limits to the allowable human-
caused mortality of pinnipeds and cetaceans. Mar
Mamm Sci 14: 1−37
Wray P, Martin KR (1983) Historical whaling records from
the western Indian Ocean. Rep Int Whal Comm 5 (Spec
Issue): 213−241
Editorial responsibility: Daniel Palacios,
Pacific Grove, California, USA
Submitted: January 16, 2011; Accepted: July 8, 2011
Proofs received from author(s): October 6, 2011
➤
➤
➤
➤
➤
➤➤
➤
➤
➤
➤
➤
