species distribution modelling of western pilbara inshore ... · v distribution modelling for...
TRANSCRIPT
Species Distribution Modelling of
Western Pilbara Inshore Dolphins
Daniella Hanf, BSc Env Mmt
November 2015
This thesis is presented for the degree of
Master of Marine Science (Research)
Murdoch University, Western Australia
ii
Cover image:
Aerial photograph of an Australian humpback dolphin (Sousa sahulensis) group
(S. Sobtzick) with a habitat suitability map produced for this thesis (MaxEnt output).
iii
Declaration I declare that this thesis is my own account of my research and contains as its main
content work that has not previously been submitted for a degree at any tertiary
education institution.
Daniella Marysia Hanf
iv
Abstract
This thesis presents the first insights into inshore dolphin distribution of the western
Pilbara. The region is undergoing rapid coastal development, which has the potential to
threaten Indo-Pacific bottlenose dolphin (Tursiops aduncus) and Australian humpback
dolphin (Sousa sahulensis) populations. Understanding the distribution of these species is
essential for their conservation.
Species distribution models (SDMs) were developed using dolphin sightings data that
were opportunistically collected during dugong aerial surveys. A geographical information
system (GIS) was used to generate the training data, which consisted of the binomial
presence-absence of dolphins, distances from mainland and islands, sea surface
temperature (SST), ocean fronts and bathymetric derivatives. Preliminary models were
developed using generalised additive model (GAM) and component-wise boosting
techniques. Models could not be fit to the data using either technique. It was unclear
whether this was a result of relatively few dolphin sightings across a large study area,
pseudo-absences, weak environmental variables, or a combination of all of these factors.
Maximum Entropy (MaxEnt) software was subsequently used as an alternative modelling
technique to model the presence of dolphins, along with automatically generated
background data, in order to avoid problems associated with unreliable absence data.
Bottlenose and humpback dolphins were sympatric, with overlap in occurrence across the
study area. Bottlenose dolphin presence was associated with the slope at the 20 m contour
and waters around the Muiron Islands. This is likely to be a productive area that could be
important for foraging. Humpback dolphin presence was associated with intertidal areas,
including shallow coastal waters near the mainland and surrounding islands. The presence
of numerous offshore islands would thus explain why humpback dolphins were recorded
more than 50 km from the coastline.
MaxEnt models were limited in their predictive power. Dedicated aerial surveys for
inshore dolphins, using standardised techniques, are required to obtain reliable species
data. In addition to increasing the sample size available for modelling, greater certainty in
group size and composition could allow count, calf and mixed species group data to be
modelled. Adequate species conservation needs to incorporate various ecological
processes that occur at different spatial and temporal scales. Guidance is provided for
undertaking boat-based studies and biopsy sampling, gathering opportunistic sightings
data, and undertaking satellite telemetry research in addition to dedicated aerial surveys.
v
Distribution modelling for inshore dolphins of northern WA is an intrinsically challenging
research project, due to limited habitat data and the elusive nature of the subject species,
particularly humpback dolphins. This research has obtained the first insights into the
distribution of inshore dolphins in northern WA. Through lessons learnt, this research has
paved the way for the development of future models to have a greater predictive ability,
which will be useful for the conservation of threatened inshore dolphin species.
vi
Table of Contents
Chapter 1: Introduction 1
1.1 Description of the study species 4
1.1.1 Population abundance 4
1.1.2 Distribution and habitat 5
1.2 The western Pilbara study area 9
1.2.1 Marine conservation values 9
1.2.2 An absent baseline 10
1.2.3 Industrial setting 11
1.2.4 Threats to inshore dolphins of the western Pilbara 13
1.3 Species distribution models for cetaceans 16
1.4 Research aim and objectives 17
1.5 Thesis structure 17
Chapter 2: Data processing and preliminary modelling 19
2.1 Introduction 19
2.2 Acquisition of dolphin data 21
2.2.1 Study area 21
2.2.2 Sighting conditions 22
2.2.3 Observer protocols 22
2.2.4 Summary of dolphin sightings 25
2.3 Preparation of training data 29
2.3.1 Response variables 29
2.4 Environmental variables 36
2.4.1 Sea surface temperature and ocean fronts 36
2.4.2 Distance from coast 37
2.4.3 Bathymetric derivatives 38
2.4.4 Omitted variables 38
2.5 Data export and exploration 48
2.6 Preliminary (presence-absence) models 51
2.6.1 GAMs 52
2.6.2 Boosted models 55
2.7 Summary and conclusion 57
Chapter 3: First insights into inshore dolphin distribution
of the western Pilbara, using MaxEnt 59
3.1 Introduction 59
3.2 Methods 61
3.2.1 Spatial and temporal extent of the data 61
3.2.2 Spatial data processing for MaxEnt models 62
3.2.3 Running the MaxEnt models 63
3.2.4 Evaluating the MaxEnt models 65
3.3 Results 65
3.3.1 Bottlenose dolphin distribution models 68
3.3.2 Humpback dolphin distribution models 71
vii
3.4 Discussion 74
3.4.1 Model performance 74
3.4.2 Ecological findings 75
3.4.3 Recommendations for further research 78
3.5 Conclusion 86
Chapter 4: Setting the scene for better predictions of inshore
dolphin distribution 88
4.1 Summary and lessons learnt 88
4.2 Setting the scene for better predictions 90
4.2.1 Obtain more informative training data 90
4.2.2 Conduct spatial statistical analysis and
appropriate data processing 91
4.2.3 Conduct exploratory analysis 92
4.2.4 Use appropriate modelling techniques 92
4.2.5 Perform robust model evaluation 93
4.3 Concluding remarks 94
References 95
Appendices
Appendix A: Adjustments for dolphin sighting locations 109
Appendix B: Omitted environmental variables 112
Appendix C: Correlation coefficients between static environmental
variables. 116
Appendix D: Diagnostic plots showing omission rates for MaxEnt
models. 117
viii
List of Figures
Figure 1: Location of Western Australia, its northern regions and the western Pilbara
study area. ................................................................................................................................... 2
Figure 2: Existing developments, foreseeable activities, and current shipping levels
of the western Pilbara study area. ..................................................................................... 12
Figure 3: Pathway from disturbance to population-level effects (New et al., 2013). ..... 15
Figure 4: Study area, transect configuration and survey blocks. ............................................ 21
Figure 5: Strip-width transect as seen from the aircraft by observers. ................................ 23
Figure 6: Count of bottlenose dolphin group sizes per survey................................................. 26
Figure 7: Count of humpback dolphin group sizes per survey. ................................................ 26
Figure 8: Location and group size of bottlenose dolphins per survey. ................................. 27
Figure 9: Location and group size of humpback dolphins per survey. ................................. 28
Figure 10: Spatial processing steps undertaken in order to prepare training data. ...... 31
Figure 11: Process of generating strip width area from the transect line. .......................... 32
Figure 12: Example of effort data calculated for the survey area. ......................................... 35
Figure 13: SST (°C) for each survey period. ..................................................................................... 42
Figure 14: Fronts (SST SD between grid cells) for each survey period. ............................... 43
Figure 15: A-C) SST (°C) for each season; D – F) Fronts (standard deviation in SST
between grid cells) for each season. ................................................................................. 44
Figure 16: Euclidean distance (km) from A) mainland and B) islands. ................................ 45
Figure 17: Bathymetry of the study area: A) depth (m), B) slope (°), C) aspect
(cardinal direction), and D) topographical complexity (standard deviation
in slope between cells). .......................................................................................................... 46
Figure 18: Sample prevalence of for presence-absence of ‘all dolphins’, bottlenose
and humpback dolphins, at each spatio-temporal resolution combination. ... 49
Figure 19: Survey effort, represented as a bias layer for MaxEnt. .......................................... 64
Figure 20: The probability of bottlenose dolphin presence (logistic output) as a
response to the variables with highest contribution for static (A and B),
Spring (C, D and E), and July 2012 (F) MaxEnt models. ........................................... 69
Figure 21: Mean relative habitat suitability for bottlenose dolphins based on A) all
sightings from all surveys and without consideration of SST or front; B, C)
‘seasons’, and D-J) each survey period from 1000 bootstrapped iterations. ... 70
Figure 22: The probability of humpback dolphin presence (logistic output) as a
response to the variables with highest contribution for static (A and B), July
2012 (C and D), and December 2012 (E and F) MaxEnt models. ......................... 72
Figure 23: Mean relative habitat suitability for humpback dolphins based on A) all
sightings from all surveys and without consideration of SST or front; B, C)
‘seasons’, and D-J) each survey period from 1000 bootstrapped iterations. ... 73
Figure 24: Model capabilities vary from describing species distribution to predicting
and testing hypotheses (Redfern et al., 2006). ............................................................. 74
ix
List of Tables
Table 1: Summary of the environmental conditions experienced during each survey. . 22
Table 2: Turbidity scale. ........................................................................................................................... 24
Table 3: Sighting conditions, species sightings and encounter rates and the proportion of
sightings that were of uncertain species for each survey*. ..................................... 25
Table 4: Literature review results for environmental predictors of coastal dolphin
distribution.................................................................................................................................. 40
Table 5: Acquired environmental datasets and derived variables. ........................................ 41
Table 6: Spearman’s rank correlation coefficients between environmental variables. . 48
Table 7: Sample prevalence, expressed as the percentage of presence records, for each
response variable at each spatio-temporal resolution, pre and post 10%
truncation of absence records. ............................................................................................ 50
Table 8: GAM descriptions. ..................................................................................................................... 54
Table 9: Summary of GAMs generated for each response variable at each resolution. Values
are ‘deviance explained’, as a percentage. ...................................................................... 55
Table 10: Summary of the ‘best’ boosted models for each response variable for each
resolution. .................................................................................................................................... 57
Table 11: Summary of MaxEnt model performance. .................................................................... 60
Table 12: Recommended studies to understand Pilbara inshore dolphin distribution and
habitat use. ................................................................................................................................. 83
Table 13: Recommended actions for developing improved SDMs for inshore dolphins in
the western Pilbara, and in northern WA. ...................................................................... 93
x
Acknowledgements I have deep gratitude for everybody who supported me through my Masters by Research
journey. I have gained so much, academically and personally. I am humbled at the support
provided to me by team of supervisors. I appreciate all of their time and energy, and I will
forever be grateful. Thank you Professor Lars Bejder for counsel, strategic advice and for
helping me to strive for high standards in my research. Thank you Dr Halina Kobryn for
skilling me up in GIS, being a fantastic teacher and making me feel well looked after. Thank
you Dr Josh Smith for long and insightful discussions and for teaching me how to use
MaxEnt. Thank you Dr Amanda Hodgson for sharing your data with me, which made this
project possible, and for explaining the technical aspects of marine fauna aerial surveys.
Thank you Robert Rankin for tutoring me in R programming, regression analysis and
component-wise boosting. Thank you also for providing R script for boosting, Box-Cox
transformation and Spearman’s rank correlation coefficients.
Sincere thanks to both Dr Colin MacLeod of GIS in Ecology and Jason Roberts, developer of
Marine Geospatial Ecology Tools (MGET), at Duke University. Their generosity and
detailed advice were essential during the data processing phase. Thank you Colin for
guidance on dealing with survey effort and processing environmental data, in particular.
Thank you Jason for helping me obtain remotely sensed data and teaching me about the
capabilities of MGET, which I will use in future work.
I appreciate the support provided to me in the early part of my candidature from Dr Joe
Fontaine on research techniques. Also, the mentorship of Dr Halina Kobryn and Dr Mike
Van Keulen when undertaking my independent study contracts, ‘Using GIS and remote
sensing to investigate the human use of Eighty Mile Beach’, and ‘Tropical marine ecology
and management’, respectively.
Special thanks to my peers – Simon Allen, Alex Brown, Delphine Chabanne, Jenny Smith,
Kate Sprogis, Nahiid Stevens and Julian Tyne. Truly there is no better crew, than MUCRU!
To all of my family and friends –thank you for your patience and understanding. Thanks
again for all of the little treats and spoils to keep me going. Especially to my parents,
Zbyszek and Rosalind Hanf, and also my sweet sister Joanna, who took delight in keeping
‘the poor uni student’ sustained with crustless crab sandwiches and sparkling wine.
xi
My wonderful Masters experience would not have been the same without Krista Nicholson
– fellow ‘storm rider’, personal Masters coach and good friend. Coffee breaks were kindly
sponsored by Ross Nicholson and Dan Hazell – thank you gentlemen.
I am indebted to my thesis reviewers Professor Phil Hammond and Dr Karin Forney for
their generous feedback. Their insightful and educational comments have significantly
improved the quality of this thesis.
Certainly, I would not have been able to complete this achievement without the support of
my darling Dan, who supported and encouraged me every step of the way. This includes a
gift of two little, loyal and furry study buddies that kept me company while I typed away,
late at night. I will always cherish and appreciate all that has been done for me.
xii
Dedicated to my Grandad,
Czesław Hanf
Who was fascinated in geography and global politics, and would always say
“Darlings, education is your most important thing”.
Chapter 1
1
Chapter 1: Introduction
This study has arisen from the acquisition of opportunistically recorded dolphin sightings
during dugong (Dugong dugon) aerial surveys in north Western Australia (WA). The study
area is situated in the western Pilbara, a remote region approximately 1,400 km north of
the State’s capital city, Perth (Figure 1). Across northern Australia, there is a sense of
urgency to conserve inshore dolphins amidst unprecedented levels of large scale and rapid
development (Allen et al., 2012; Bejder et al., 2012; Bejder et al., 2009; Cagnazzi et al.,
2013; Palmer et al., 2014b; Parra et al., 2006a). This region, north of the tropic of
Capricorn (Figure 1) has until recent years remained remote and largely unmodified
(Halpern et al., 2008). The current Commonwealth Government views the region as
‘economically underutilised’ and is pushing for sizeable growth in tourism and energy
export (Loughane, 2013). This is without strategic planning for sustainable development.
Additionally, adequate environmental impact assessment (EIA) is hindered by a lack of
baseline scientific information and a general understanding of the potential for
population-level effects on inshore dolphin species (Allen et al., 2012; Bejder et al., 2012;
Bejder et al., 2009).
The three inshore dolphin species of northern Australia are the Indo-Pacific bottlenose
dolphin (Tursiops aduncus), Australian humpback dolphin (Sousa sahulensis) and
Australian snubfin dolphin (Orcaella heinsohni) (Allen et al., 2012; Jefferson & Rosenbaum,
2014; Palmer et al., 2014b). Taxonomic uncertainty remains surrounding the Tursiops
genus in Australian waters. However, recent genetic investigation off northern WA
suggests that Indo-Pacific bottlenose dolphins occur within the 20 m bathymetric contour,
whereas common bottlenose dolphins (T. truncatus) most likely occur beyond the 50 m
contour (Allen et al., In review.). The Australian humpback dolphin was recognised as a
new species, distinct from Indo-Pacific humpback dolphins (Sousa chinensis), on 01 August
2014 (Jefferson & Hung, 2004) based on multiple lines of evidence (Frère et al., 2008;
Frère et al., 2011; Mendez et al., 2013; Mendez et al., 2011). Australian humpback and
snubfin dolphins are endemic to the Sahul Shelf, which is situated on the Continental Shelf
of northern Australia and extends to Papua New Guinea (Jefferson & Hung, 2004). Snubfin
dolphins were precluded from this study due to the small sample size collected, perhaps
because the study area represents the southern end of their range. Hereafter, the two
study species are referred to as bottlenose and humpback dolphins.
Chapter 1
2
Figure 1: Location of Western Australia, its northern regions and the western Pilbara study area.
Chapter 1
3
Inshore dolphins of northern WA are protected at the international, national and state
levels. The International Union for the Conservation of Nature (IUCN) Red List of
Threatened Species is a catalogue of global flora and fauna that assigns conservation
listings based on extinction threats and the amount of available data. Under the IUCN,
Bottlenose dolphins are listed as ‘data deficient’, snubfin dolphins as ‘near threatened’ and
the status of humpback dolphin is currently being reassessed (Parra and Cagnazzi, in
prep.).
Under Australia’s national environmental law, the Environment Protection and Biodiversity
Conservation Act 1999 (EPBC Act), all three inshore dolphin species are listed as ‘cetacean’
and ‘migratory’, relating to two different provisions of the Act. As cetaceans, they are
protected within the Australian Whale Sanctuary, where it is illegal for them to be killed,
injured or interfered with. The listing of ‘migratory’ animals under the Act is a provision
relating to the Convention on the Conservation of Migratory Species of Wild Animals
(‘Bonn Convention’), to which Australia is a signatory. As a result of these two listings, all
three species are considered a matter of national environmental significance (MNES). Any
proposed action (e.g. a development, an activity or series of activities) that is likely to have
a ‘significant impact’ on a MNES must be referred to the Federal Minister for the
Environment for approval (DoE, 2013). Nominations for both humpback and snubfin
dolphins to be assessed as threatened under the EPBC Act have been made, but the
assessment cannot occur until data deficiencies are addressed.
Inshore dolphins are protected in WA State waters (defined as lowest astronomical tide
out to 3 nautical miles) under the Wildlife Conservation Act 1950. Under this legislation, it
is illegal to “kill or capture any fauna by any means or disturb or molest any fauna by any
means or to use any method whatsoever to hunt or kill any fauna whether this results in
killing or capturing any fauna or not”.
The Environmental Protection Authority (EPA) of WA has recently formalised specific
environmental objectives that development proponents must be capable of meeting,
should their development proposals be approved. For marine fauna, the objective is “to
maintain the diversity, geographic distribution and viability of fauna at the species and
population levels” (EPA, 2013). However, neither impact assessment nor management on
the basis of populations can occur until quantitative data on these population
demographics and distribution have been collected.
Chapter 1
4
Stock assessment undertaken by the United States’ National Oceanic and Atmospheric
Administration (NOAA) is an example of a marine mammal population-based management
strategy. Under the U.S Marine Mammal Protection Act of 1972, NOAA’s goal is to prevent
the depletion of populations and does so by monitoring human-induced mortality against
a specified limit of mortality (Taylor et al., 2000; Wade, 1998). Assessments are based on
the following inputs (Barlow et al., 1995):
a biological definition, including its geographic range
an abundance estimate,
a potential biological removal estimate (the maximum number of individuals that,
in addition to natural mortality rates, can be removed while allowing that stock to
reach or maintain its optimum sustainable population), and
an understanding of human induced mortality and serious injury levels.
No such information exists for bottlenose or humpback dolphins in northern WA.
1.1 Description of the study species
Indo-Pacific bottlenose and humpback dolphins are both long-lived, reaching over 40
years of age (Parra et al., 2004; Wang & Chu Yang, 2009). They are large animals, reaching
up to 2.8 m in length and 200 kg in weight (Parra et al., 2004; Wang & Chu Yang, 2009).
Sexual maturity for both species is around 11 years of age and females first give birth at
around 12 years of age (Mann et al., 2000; Parra & Ross, 2009a). Gestation lasts for
approximately 12 months. Bottlenose dolphin calves are nursed for three to five years and
it is an average of four years from weaning until the next pregnancy (Mann et al., 2000).
Calving occurs throughout the year for both species but peaks in warmer months,
presumably because adult dolphins have more energy available because their body
temperatures are warmer and prey is more abundant (Mann et al., 2000; Parra & Ross,
2009a).
1.1.1 Population abundance
Population abundance reflects population health and the quality and quantity of available
resources. Abundance estimates may also be used to determine vulnerability to
anthropogenic impact because population level effects are hastened in populations that
are already small (Reed et al., 2003). Parra et al. (2006a) noted that populations of less
than 100 individuals have a high chance of extirpation. Smaller populations are usually
lower in genetic diversity which results in low resilience to morbillivirus and infectious
Chapter 1
5
disease (Sutherland, 1998). Inshore dolphin populations of the western Pilbara are yet to
be identified, let alone have abundance estimates be calculated.
The abundance of bottlenose dolphin populations can vary greatly. A density of 0.14 to
0.23 animals / km2 has been reported from Port Essington, NT (Palmer et al., 2014a),
whereas a density of 4.6 to 6.7 animals / km2 has been recorded at Point Lookout,
Queensland (Chilvers & Corkeron, 2003). In stating this, a direct comparison of population
estimates must be undertaken with caution because different methods will yield different
results, and the reporting of methods is often insufficient to understand the effects of these
differences (Nicholson et al., 2012). Estimates of population abundance do not exist for
any bottlenose dolphins in northern WA but they are likely to be in lower numbers across
the Pilbara due to the oligotrophic and cyclonic conditions that limit resource availability.
Globally, humpback dolphin (Sousa sp.) populations are in decline (Reeves et al., 2008).
Abundance estimates for humpback dolphins have not yet been obtained for WA. In other
parts of Australia, populations tend to be small and discontinuous, and consequently
vulnerable to extirpation (Cagnazzi, 2010; Cagnazzi et al., 2011; Parra et al., 2006a).
Abundance estimates equate to fewer than 0.19 animals / km2 at most study sites
(Cagnazzi, 2010; Cagnazzi et al., 2011; Parra et al., 2006a), but were up to 0.63 animals /
km2 in Port Essington, NT (Palmer et al., 2014a).
1.1.2 Distribution and habitat
Species’ spatial ecology is complex and occurs at a hierarchy of scales. Patterns and
processes are different at each scale, and are further complicated by temporal variation
(Gutzwiller, 2002). The spatial ecology of dolphins is particularly complex because they
are highly mobile and are influenced by the highly connected nature of the marine
environment (Gutzwiller, 2002). The concepts of ‘fundamental’ and ‘realised’ niche are
also pertinent. The fundamental niche is the area where a group of animals can live, based
on having suitable physical habitat characteristics. The realised niche is a subset of the
fundamental niche and incorporates the multidimensional range of conditions that
animals have evolved to actually live within (Morrison et al., 2006). It includes mechanistic
ecological relationships, such as predation and competition. The ecological niche of
western Pilbara inshore dolphins is unknown. The distribution and habitat use of
bottlenose and humpback dolphin is discussed in terms of species geographic range and
habitat requirements, metapopulation dynamics, and fine-scale specialisations and
patterns.
Chapter 1
6
Geographic range and habitat requirements
Bottlenose dolphins are widespread through temperate and tropical waters in the Indo-
Pacific region (Hammond et al., 2012; Wang & Chu Yang, 2009). Until 15 years ago, all
bottlenose dolphins in coastal waters were classified as an inshore form of the common
bottlenose dolphin (T. truncatus) (Möller & Beheregaray, 2001; Wang et al., 1999).
Recently, evidence collected by Allen et al. (In review.) from the western Pilbara and
surrounds demonstrated that bottlenose dolphins inshore of the 20 m isobath are T.
aduncus while those offshore of the 20 m isobath are T. truncatus. Bottlenose dolphins
display high levels of phenotypic plasticity, enabling them to exploit many coastal and
pelagic habitat types (Wang & Chu Yang, 2009). They consume a wide range of benthic and
pelagic fish and cephalopod species (Amir et al., 2005; Heithaus & Dill, 2002; Parra &
Jedensjö, 2013). A third genetically distinct species, T. australis, has recently been
suggested from coastal southern Australia (Charlton-Robb et al., 2011), but this has not
yet been accepted by taxonomists of the Society for Marine Mammalogy.
Humpback dolphins are less widespread than bottlenose dolphins, being restricted to
more coastal tropical / subtropical waters. There are now three recognised species of
humpback dolphins. Distribution of the Australian humpback dolphin is associated with
the Sahul Shelf, which extends around Australia and connects to southern New Guinea
(Merow & Silander, 2014). In addition to the new Australian species, there are those in the
Indo-Pacific (S. chinensis) and those in the Atlantic (S. teuszii), off the west coast of Africa
(Parra & Ross, 2009a). Indo-Pacific humpback dolphins may still consist of two separate
species: S. plumbea from South Africa to the east coast of India, and S. chinensis from the
east coast of India to China (Parra & Ross, 2009a).
Australian humpback dolphins have mainly been studied in Queensland (Cagnazzi, 2010;
Cagnazzi et al., 2011; Parra, 2006; Parra et al., 2006a; Parra & Jedensjö, 2009), so there is
limited knowledge on distribution patterns within WA. Historical stranding data, museum
specimens and opportunistic sightings collected during aerial and boat-based surveys for
other fauna enables the inference that humpback dolphins occur almost continuously from
the WA/NT border to Shark Bay (Allen et al., 2012; Hodgson, 2007; Jenner et al., 2014;
Parra et al., 2004; Parra & Ross, 2009b; Preen et al., 1997; Prince et al., 2001; Ross et al.,
1994; Sleeman et al., 2007). The southern-most sighting of a (likely vagrant) humpback
dolphin, within a group of Indo-Pacific bottlenose dolphins, was recorded on the citizen
science tool ‘Coastal Walkabout’ at Kalbarri (27.7°S) (CW, 2015) (Figure 1).
Chapter 1
7
Humpback dolphins inhabit shallow waters close to coastlines, reef or estuaries
throughout their studied range (Parra, 2006; Parra et al., 2004). In the NT, distribution is
closely associated with presence of a large tidal river (Palmer et al., 2014b). Humpback
dolphins have also been recorded in deeper waters, but in association with islands or reefs
(Corkeron et al., 1997). Allen et al. (2012) suggested that humpback dolphins use a range
of inshore habitats, including both clear and turbid coastal waters across northern WA. At
the NWC, preliminary studies by Brown et al. (2012) recorded humpback dolphins from
269 m to 4.5 km away from shore. Depths ranged from 1.2 to 20 m, with a mean of 7.8 m
(± 0.74 SE). Humpback dolphin reliance on shallow coastal habitats for foraging is
supported through stomach contents analysis, with the consumed fish taxa being from
coastal-estuarine environments (Parra & Jedensjö, 2013).
In some areas, humpback dolphins have been found to take advantage of pulses in
productivity, temporarily moving to areas of increased prey availability (Karczmarski et
al., 1999; Parra et al., 2011). They have fluid groupings within a fission-fusion society
(Allen et al., 2011; Connor, 2000; Parra et al., 2011). Humpback dolphins usually feed in
solitude or in small groups (Parra & Ross, 2009a) but form large aggregations when prey
is abundant (Chilvers et al., 2003).
Other temporal variation in distribution patterns has also been demonstrated. Bottlenose
and humpback dolphins in other areas have both been found to vary in distribution and
niche depending on seasonal reproductive activities (Karczmarski et al., 2000; Mann et al.,
2000; Palmer et al., 2014a). Several studies have found that bottlenose dolphin
distribution changes during warmer months, with an influx of individuals entering certain
areas (Nicholson et al., 2012; Smith et al., 2013; Wilson et al., 1997). Bottlenose dolphin
distribution and realised niche has also been found to be influenced by predator presence
(Heithaus & Dill, 2002) and at a finer scale, tide cycles (Fury & Harrison, 2011).
Metapopulation distribution
A metapopulation can be defined as a cluster of separate populations with some genetic
exchange, and is often associated with ‘patches’ of habitat within a landscape (Gutzwiller,
2002; Hanski & Gaggiotti, 2004; Morrison et al., 2006). In some cases, adult males facilitate
genetic dispersal by travelling between populations to breed with females of those
populations (Gaggiotti & Hanski, 2004). This has been observed for bottlenose dolphins in
Shark Bay and humpback dolphins in South Australia (Karczmarski, 2000; Krützen et al.,
2004; Möller & Beheregaray, 2004). Brown et al. (2014) have recent genetic evidence for a
Chapter 1
8
northern WA humpback dolphin metapopulation. The results are similar to those of
genetic studies from eastern Queensland, which suggest that humpback dolphins exist as a
“metapopulation of small, genetically largely isolated population fragments” (see Cagnazzi,
2010; Cagnazzi et al., 2011). Five small fragmented populations were identified in the
Kimberley and Pilbara regions, with breeding males dispersing up to 350 km (Brown et al.,
2014). They suggest that the North West Cape may be a sink population, receiving low
levels of immigration from the Dampier Archipelago, which is located several hundred
kilometres north-east. Sink populations inhabit less favourable habitat patches than other
populations, resulting in lower reproductive success (Burgman & Lindenmayer, 1998;
Hanski & Gaggiotti, 2004). Source populations are healthier and larger than sink
populations because they inhabit areas of higher habitat quality, so they are able to
genetically ‘subsidise’ sink populations through emigration (Burgman & Lindenmayer,
1998). Maintenance of genetic diversity enables metapopulations and populations to
withstand changes in the environment such as the introduction of disease or pathogens
(Hanski & Gaggiotti, 2004). The low level of genetic dispersal between fragmented
populations in WA is cause for concern because it may be easily disrupted (Brown et al.,
2014). An example of this is the habitat fragmentation along the Great Sandy Straight
coastline that has resulted in highly geographically restricted humpback dolphin
communities (Cagnazzi et al., 2011).
Fine-scale specialisations and patterns
Bottlenose dolphins occupy a broad range of habitats at a species level, but specific,
realised niches are occupied at finer scales (Acevedo-Gutierrez, 2009; Blasi & Boitani,
2012; Sargeant et al., 2007). For example, bottlenose dolphins in Shark Bay exhibit a range
of habitat-specific foraging specialisations, learnt via vertical matrilineal transmission
(Connor, 2001; Heithaus & Dill, 2002; Krützen et al., 2004; Mann et al., 2000; Mann &
Sargeant, 2003; Smolker et al., 1997). This is in part due to habitat heterogeneity
(Sargeant et al., 2007). Observations of humpback dolphins have led to the suggestion that
they too may engage in habitat-specific, specialist behaviours, but this is uncertain as they
are not as well studied as bottlenose dolphins (Whiting, 2011). It is also likely to be related
to density of foraging areas that are available.
Chapter 1
9
1.2 The western Pilbara study area
1.2.1 Marine conservation values
The western Pilbara study area is situated on an ancient submerged shoreline, so the
study area is flat and shallow (Wilson, 2013). It is located in an arid region and has a
tropical to subtropical marine environment. Rainfall is generally low, but peaks in both
summer and winter. Cyclonic activity is a major feature of the study area in summer
months, with extreme events occurring approximately every five to ten years. Summer
rainfall causes the expansive Ashburton River Delta to flood, which in turn results in
highly turbid waters close to the coast (Wilson, 2013). Exmouth Gulf, in the southern
portion of the study area, is a wide and semi-enclosed embayment. Strong tidal currents
between the North West Cape and Muiron Islands flush its internal waters and maintain
normal oceanic salinity levels (Orpin et al., 1999).
The major currents that influence the western Pilbara are the Holloway Current and the
Leeuwin Current (Wilson, 2013). The Holloway Current is fed by the warm, oligotrophic
Indonesian Through Flow (ITF), which in turns contributes to the headwaters of the
Leeuwin Current, WA’s prevailing oceanographic feature. The Leeuwin Current is a
southward flowing, narrow, shallow convection current that transports warm water from
Indonesia and is strongest between April and October (Cresswell & Golding, 1980; Godfrey
& Ridgway, 1985). The study area is situated inshore of where the two currents meet. Both
currents influence western Pilbara by warming the sea surface temperature (SST) and
supressing the flow of offshore nutrient-rich upwellings from reaching the coastal area
(Cresswell & Golding, 1980; Godfrey & Ridgway, 1985). This, in addition to relatively low
chlorophyll levels and phytoplankton biomass due to low levels of terrestrial run off
(Ayukai and Miller, 1998), results in oligotrophic waters. A ‘marine heatwave’ was
experienced in early 2011, from the combined influence of the Leeuwin Current and an
extreme La Nina event. SST increased by 5 °C which resulted in large scale fish and benthic
primary producer die off, and the displacement of tropical species such as the whale shark
(Rhincodon typus) (Pearce & Feng, 2013).
Waters of the western Pilbara are relatively nutrient poor, yet support a richness of
marine biodiversity. Water quality is of a very high standard because, until now, the area
has largely remained unmodified (Halpern et al., 2008; Wenziker et al., 2006). Extensive
salt flats, benthic macroalgal communities and mangroves on the eastern side of Exmouth
Gulf invest high levels of organic matter into the embayment that feed low trophic level
Chapter 1
10
grazers, including commercially fished tiger prawns (Penaeus esculentus) (Ayukai & Miller,
1998; Penn & Caputi, 1986). Mangrove communities along the coast also provide critical
nursery habitat for many invertebrate and fish species including endangered sawfish
(Pristis sp.) (Chevron Australia, 2010). Patchy, ephemeral seagrasses provide important
foraging habitat for dugongs, fish and their predators. There is a high species richness of
coastal fish, with many species endemic to WA (Fox & Beckley, 2005). Numerous coral
fringed islands are believed to provide habitat and shelter for rich assemblages of
invertebrates and fish, marine mammals and nesting turtles. The reef and islands protect
the coastline from heavy swells, and the corals act as source reefs for those of Ningaloo
Marine Park (Bancroft & Long, 2008; DEWHA, 2008; Hughes et al., 2003; Kendrick &
Stanley, 2001; Wilson, 2013). Population IV humpback whales (Megaptera novaeangliae)
aggregate in Exmouth Gulf, presumably to rest, during their southward migration from
their breeding grounds in northern Australia to their feeding grounds in Antarctica
(Braithwaite et al., 2015; Chittleborough, 1953; Jenner et al., 2001; Pitman et al., 2014)
1.2.2 An absent baseline
Despite the marine conservation values of the western Pilbara, there is a lack of baseline
environmental data for the region. There is a widely held assumption that Exmouth Gulf
provides important habitat for marine megafauna (Chevron Australia, 2010; DEWHA,
2008; Kendrick & Stanley, 2001). It is concerning that waters north of the gulf have been
continuously overlooked. This was raised by Preen et al. (1997) nearly 20 years ago, when
they recommended that comprehensive fauna studies of the western Pilbara be conducted
prior to the commencement of oil and gas related developments. Regrettably, the area was
excluded from the North West Shelf Joint Environmental Management Study (NWSJEMS),
established by the WA government to inform EIAs and achieve environmentally
sustainable development (Bulman, 2006).
This remote coastline is largely devoid of baseline ecological data for inshore dolphins
(Allen et al., 2012; Bejder et al., 2012). Preliminary boat-based surveys suggested that the
North West Cape may provide important humpback dolphin habitat, and that humpback
and bottlenose dolphins were often found in mixed species groups (Allen et al., 2012;
Brown et al., 2012). Preliminary dolphin distribution maps have been generated from
sightings collected during surveys targeting dugongs and whales, but dolphins were
predominantly identified to the family level, only (Hodgson, 2007; Preen et al., 1997;
Prince et al., 2001; Sleeman et al., 2007).
Chapter 1
11
In 2008, the EPA commenced a risk-based approach to EIA in order to “focus on the
environmental risks and impacts that matter, greater consistency, rigour and transparency
of decision-making” (EPA, 2009), and used Chevron’s Wheatstone Project as its ‘major
project’ test subject. Impact to dugong populations due to the removal of seagrass habitat
from dredging was identified as a key risk. As such, a series of dugong aerial surveys to
obtain population abundance estimates and identify seasonal trends (Hodgson, 2012).
Dolphin populations were not identified as being at high risk, so no targeted studies were
established. This study attempts make use of the inshore dolphin sightings collected
during those dugong aerial surveys.
1.2.3 Industrial setting
Wide-scale habitat alteration is occurring at a rapid pace in the western Pilbara, with high
levels of industrial activity set to become an enduring feature of the landscape (Hatton et
al., 2011). Recent oil and gas discoveries in the Northern Carnarvon Basin (>375,000 km2)
has led the WA Department of State Development (DSD) to designate the ‘Ashburton
North’ development site as a Strategic Industrial Area (ANSIA), (DMP, 2014). Prior to this
decision, Chevron Australia had already identified this site, approximately 12 km west of
Onslow, as suitable for a port and processing facilities for the Wheatstone Project and had
commenced their EIA (Figure 1Figure 2). The ANSIA will be expanded to be an oil and gas
processing ‘hub’ that will support future ventures including BHP Billiton Australia’s
Macedon domestic gas project. Development of marine facilities has resulted in substantial
physical alteration of the coastline. A three-year dredging program has entailed the
removal and dumping of 50 million m3 of sediment to create a 25 km long shipping
channel. Maintenance dredging will continue for the period of operation. Vessel traffic is
set to greatly increase from current levels (AMSA, 2014)(Figure 2). Exploration for new oil
and gas fields will entail seismic and drilling activities. Further subsea construction may
include rock blasting and dumping. Also in the study area is the Gorgon LNG development,
the largest in Australia to date, which has a processing plant situated on Barrow Island
Nature Reserve. Other major developments in the area include the Onslow Salt mine port
and Exmouth Harbour which services tourism, commercial fisheries and oil and gas
industries (DMP, 2014) (Figure 2).
The Pilbara is currently one of the fastest growing regions in Australia (WAPC, 2012).
Karratha, approximately 200 km north of the study area, achieved ‘city status’ on
01 July 2014, with the local human population exceeding 20,000 (Validakis, 2014). Within
the study area, the town of Onslow (Figure 1) has been forecast to increase from
Chapter 1
12
approximately 600 people to at least 3,000, as a direct result of the ANSIA developments.
It is anticipated that Onslow will become a tourist hub for visitors to the Pilbara, with new
facilities and good access to marine and island recreational opportunities (WAPC, 2012).
Figure 2: Existing developments, foreseeable activities, and current shipping levels of the western Pilbara study area. Shipping data are thinned location points of ships
recorded by the satellite ship borne Automatic Information System (AIS) in November 2014.
Chapter 1
13
1.2.4 Threats to inshore dolphins of the western Pilbara
Inshore dolphins are susceptible to a number of threats because their distribution closely
overlaps with that of humans. In the western Pilbara, interactions with commercial
fisheries, tourism activities and recreational boaters are of some concern. Commercial
trawling occurs in the western Pilbara, which could impact upon inshore dolphins through
prey competition, incidental take or modification of natural behaviour (Allen et al., 2014;
Chilvers & Corkeron, 2002; Chilvers et al., 2003). Persistent harassment of dolphins by
wildlife watching operators or recreational boaters could ultimately lead to their
displacement (Bejder et al., 2006b). However, it is thought that these activities do not
currently affect dolphin populations of the Pilbara as they remain at relatively low levels
and management strategies are in place (CALM, 2005; DoF, 2014).
Habitat alteration through coastal development is recognised as the most serious
anthropogenic threat to Australia’s inshore dolphins (e.g., DEWHA 2010). The IUCN ranks
habitat alteration as the largest risk to mammal species survival, world-wide (Morrison et
al., 2006), and this holds true for marine mammals (Harwood, 2001; Reeves et al., 2008).
Humpback (and snubfin) dolphins are particularly susceptible to habitat alteration
because of their close proximity with anthropogenic activity and their relatively small
populations and restricted ranges (Cagnazzi et al., 2013; Karczmarski, 2000; Parra et al.,
2006a; Stensland et al., 2006). Habitat alteration can occur in many ways and includes
changes to size, quality, structure and connectivity (Tuomainen & Candolin, 2011).
In the western Pilbara, loss of foraging habitat, increased anthropogenic noise and
barriers to movement are of ongoing concern due to existing developments and
foreseeable activities (Figure 2). Dredging directly removes benthic primary producer
habitat that supports fish stocks that dolphins prey upon (Acevedo-Gutierrez, 2009; Amir
et al., 2005; Parra & Jedensjö, 2009). This can reduce or redistribute prey availability and
limit dolphins’ energy intake (Jefferson et al., 2009; Pirotta et al., 2013). Increased
turbidity of the water column also increases the susceptibility of dolphins to shark
predation (Heithaus & Dill, 2002).
Sound enables dolphins to communicate with each other and use echolocation for foraging
and predator avoidance (Heithaus & Dill, 2002; Smolker et al., 1993; Van Parijs &
Corkeron, 2001b). Reliance on echolocation would be particularly important in the
western Pilbara, where near-shore waters are very turbid and visibility is poor. Dolphins
employ a range of vocalisations from 0.2 to 22 kHz, and echolocate at frequencies up to
Chapter 1
14
300 kHz (David, 2006; Van Parijs & Corkeron, 2001b). Construction pile driving, seismic
air guns involved in oil and gas exploration, and small boat motors cause disruption within
these bandwidths (David, 2006; Jefferson et al., 2009; Nowacek et al., 2007; Stone &
Tasker, 2006; Van Parijs & Corkeron, 2001a). Consequences vary in type and effect.
Hearing loss, termed threshold shift (TS), can be permanent (PTS) or temporary (TTS)
(Nowacek et al., 2007). Anthropogenic noise can also ‘mask’ other sounds that dolphins
usually hear, such as the signature whistles made by calves (May-Collado & Quiñones-
Lebrón, 2014). The most common response to interruption by anthropogenic sound is for
dolphins move to a quieter area (Nowacek et al., 2007). This is highly problematic if
animals are displaced from critical habitat, and especially if sound intrusion is chronic.
EIA in WA has so far perhaps overestimated the resilience of inshore dolphins to
population level effects which could arise from habitat alteration (Bejder et al., 2012;
Bejder et al., 2009). On a daily basis, animals incorporate the threat of predation into
making decisions about where they feed, rest or undertake reproductive activities (Lima &
Dill, 1990). In a similar fashion, animals may decide that the disturbances such as
underwater noise or increased sedimentation caused by construction activities outweigh
the reasons to stay in an area, and so choose to leave. While avoiding an area for a short
time may be beneficial (e.g. vessel avoidance in order to circumvent strike), prolonged
avoidance from an important habitat is likely to be detrimental. Redistributed animals are
likely to find suboptimal conditions elsewhere. This could be in the form of reduced
foraging opportunities, or increased exposure to competition or predation (Heithaus &
Dill, 2002). As a result, the health and reproductive fitness of individuals may be reduced
(Figure 3; New et al., 2013). In small populations, such as those of humpback dolphins, this
creates a plausible risk that reproductive parameters of the population may be reduced
overall (Banks et al., 2007; Hanski & Gaggiotti, 2004; New et al., 2013; Reed et al., 2003;
Tuomainen & Candolin, 2011).
The pathway from disturbance to population-level effects is difficult to document,
particularly in long-lived species such as dolphins. Quantification of non-lethal
disturbances is difficult because effects are most often subtle, chronic and cumulative over
time. However, evidence does exist. A rare, long-term Before-After Control-Impact (BACI)
baseline dataset demonstrated a decline in bottlenose dolphin abundance due to
cumulative exposure to dolphin watching activities (Bejder et al., 2006b). Similar
disturbance was also found to reduce the time bottlenose dolphins spent resting (Lusseau,
2004) and killer whales spent feeding (Williams et al., 2006b). Pirotta et al. (2013) showed
Chapter 1
15
that dolphins will be displaced by high intensity dredging, even if they were previously
exposed to lower levels.
Figure 3: The multi-step pathway from disturbance to population-level effects (New
et al., 2013).
The alternative, yet equally detrimental, effect to displacement is that a dolphin may
remain in the area of disturbance and sacrifice its vigour. The more reliant that an animal
is on its environment, the more benefit there is to stay (Frid & Dill, 2002). Animals with
specialised ecological interactions or restricted ranges are likely to have stronger habitat
preferences. As such, they will have greater difficulty in leaving so will stay in a disturbed
area for longer. The ability to relocate can be further limited by the dolphin’s health, life
phase or age (Swihart et al., 2003). Increased stress and reduced sustenance will cause a
decline in an individual’s energy budget that could ultimately reduce population viability,
as described above.
Extirpation of a dolphin population could have wide-reaching ecological effects. As top-
order predators, dolphins have an important role in structuring the trophic levels of an
ecosystem (Estes et al., 2011; Heithaus et al., 2008). One possible scenario is trophic
cascade, which occurs when the abundance of one species changes such it causes the
abundance of another species to also change (Heithaus et al., 2008). Such a broad scale
effect is difficult to predict or prove but mounting evidence suggests that trophic-level
effects are occurring more often in both marine and terrestrial ecosystems than previously
recognised (Estes et al., 2011).
1.3 Species distribution models for cetaceans
Identifying the environmental variables that most influence a species niche has become
vital to many conservation programs (Elith & Leathwick, 2009; Redfern et al., 2006). The
use of SDMs to identify causal relationships between species distribution and
Chapter 1
16
environmental features originates from terrestrial flora and fauna studies (Ferrier, 1986;
Hill, 1981; Meents et al., 1983). They became commonly used in marine and freshwater
studies in the 2000s with improved access to remotely sensed data, Geographical
Information Systems (GIS) and an increase in computing power (Elith & Leathwick, 2009;
Guisan & Zimmermann, 2000; Redfern et al., 2006). Making use of a range of data types
including historical and opportunistically obtained species records, SDMs provide a cost-
effective and quantitative tool for understanding ecological patterns over broad spatial
and temporal extents and in remote and unsurveyed areas. As a result, SDMs have become
popular in marine mammal research (Beale & Lennon, 2012; Cañadas et al., 2005; Forney,
2000; Gomez de Segura et al., 2008; Redfern et al., 2006). The purpose of cetacean SDMs
are becoming more varied and sophisticated (Elith & Leathwick, 2009; Forney, 2000;
Guisan & Thuiller, 2005; Guisan & Zimmermann, 2000; Redfern et al., 2006). Cetacean
SDMs have been used to improve estimates of species abundance and densities (Cañadas
& Hammond, 2006, 2008; Forney, 2000; Williams et al., 2006a), improve existing survey
and data analysis methods (Praca et al., 2009), identify critical habitat (Blasi & Boitani,
2012), predict the distribution of species in decline (Cañadas & Hammond, 2008), assess
the inclusivity of critical habitat within existing protected area boundaries (Parra et al.,
2006b), inform time-area closure management measures (Thorne, 2012), delineate
species ecotypes (Torres et al., 2003) and improve monitoring and adaptive management
programs (Dähne et al., 2013).
Complex SDMs generally follow on from the development of earlier, simpler models, using
an iterative process (Redfern, 2006 #84}. This is reflected in the literature, with SDM
intention trending away from explaining a species’ distribution towards predicting it in
another time or place (Elith & Leathwick, 2009). While cetacean SDMs have become a
conventional research technique for both wide-ranging species and small coastal species
in the northern hemisphere and Antarctica (Cañadas et al., 2005; Garaffo et al., 2011;
Williams et al., 2006), they remain a relatively new practice in Australia. This could be
related to the low levels of survey effort that has been undertaken (Jewell et al., 2012).
Chapter 1
17
1.4 Research aim and objectives
The aim of this study is to obtain the first insights into western Pilbara inshore dolphin
distribution, using SDMs built on opportunistically acquired sightings data. The specific
research objectives are as follows.
Identify areas of highest habitat suitability for bottlenose and humpback dolphins.
Determine whether there is difference in habitat preference between bottlenose
and humpback dolphins.
Understand which environmental variables are most important in defining the
distribution of bottlenose and humpback dolphins.
Describe the influence that important environmental variables have in defining the
distribution of bottlenose and humpback dolphins.
These insights will be used to inform survey planning across the expansive and remote
region. Targeting areas known to support higher occurrences of dolphins has several
benefits. From a logistical perspective, shorter survey duration will reduce financial costs
as well as the exposure of observers to safety risks. A more rapid collection of data could
inform management decisions sooner. Stratified surveys can achieve greater precision in
population abundance estimates and allow bespoke detection functions to be developed
for areas of varying sighting conditions (Dawson et al., 2008).
1.5 Thesis structure
Following this introductory chapter, the thesis consists of the following chapters:
Chapter 2: Data processing and preliminary modelling
Chapter 2 describes the methods, decisions and compromises involved in acquiring and
processing dolphin sightings (response) and environmental (predictor) data. This
important stage of the project is responsible for generating the datasets that are used to
train (i.e. develop) and subsequently test the models. A detailed explanation of why the
aerial survey protocols (developed for dugongs) were not suitable for surveying dolphins
is provided. Preliminary model outputs using generalised additive models (GAMs) and
component-wise boosting are also presented, which were used to explore the limitations
of this particular presence-absence dataset on model performance.
Chapter 1
18
Chapter 3: First insights into inshore dolphin distribution of the western Pilbara,
using MaxEnt
Chapter 3 presents the first insights into the distribution of western Pilbara inshore
dolphins, obtained through Maximum Entropy (MaxEnt) modelling. The MaxEnt method is
described in detail. The limiting effect that a) small sample sizes of species sightings, and
b) weak environmental variables, can have on model performance is discussed. The
ecological findings in relation to the research objectives are presented, and
recommendations for survey planning are provided.
Chapter 4: Setting the scene for better predictions of inshore dolphin distribution
Chapter 4 first summarises the thesis, and then outlines the actions that are needed to
develop robust models for predicting distribution of inshore dolphins in northern WA.
Recommended actions are focussed on reducing, and then treating, excess zeroes in the
species sighting dataset.
Chapter 2
19
Chapter 2: Data processing and preliminary modelling
2.1 Introduction
Species distribution models (SDMs) are constructed in order to understand a species’
distribution, based on its relationship with aspects of its environment. A SDM project
involves developing an algorithm that reliably explains (and sometimes also predicts)
species distribution, based on the species and environmental data available (Elith &
Leathwick, 2009). Species sightings are modelled as a response to the explanatory
environmental (e.g. habitat type, sea surface temperature) data. These data on which the
model is based are termed ‘training’ data. Data that is later used to evaluate the reliability
of the model is termed ‘test’ data.
Advancements in technology have enabled scientists to access new sources of
environmental data, which has increased the popularity of SDM use (Elith & Leathwick,
2009). Developments in remote sensing, improved computing power and growing
proficiency amongst scientists and managers in the use of programs that support spatial
data processing (Bailey & Thompson, 2009; Elith & Leathwick, 2009; MacLeod et al.,
2007). Remote sensing is used to acquire information by measuring a feature’s reflected or
emitted electromagnetic or acoustic energy (Jensen, 2007; Purkins & Klemas, 2011). GIS
and programming software such as ‘R’ (R Development Core Team, 2013) that support the
processing of spatial data provide an interface for creating, viewing, managing and
manipulating data (Jensen, 2007; Purkins & Klemas, 2011). The combination of these
technologies allows environmental data to be sourced ex-situ and retrospectively, which
facilitates the modelling of opportunistically acquired species data. They can also offer a
cost-effective, time-efficient and risk-adverse (in terms of human safety) alternative to in-
situ sampling of large and remote study areas (Mumby, Green, Edwards, & Clark, 1999;
Stevens & Collins, 2011; Xie, Sha, & Yu, 2008).
This study has arisen from the acquisition of opportunistic dolphin sightings recorded
during dugong aerial surveys over the western Pilbara region of north WA (Figure 1). The
surveys were designed to understand seasonal trends in abundance of the local dugong
population (Hodgson, 2012). As such, they followed the standard aerial survey strip-width
method for dugongs that were developed by Marsh and Sinclair (1989), and later refined
by Pollock et al. (2006). As dugongs spend most of their time underwater, a specific strip-
width method was designed that also accounts for ‘availability’ bias, with availability being
affected by depth and water turbidity (Marsh & Sinclair, 1989; Pollock et al., 2006). Strip-
Chapter 2
20
width transects are the preferred method for surveying dugongs because dugongs spend
the majority of their time subsurface, in relatively shallow environments. Observers from
aircraft are often able to see dugongs on the sea floor, as the aircraft passes overhead.
Transects are conducted in passing mode which means that the aircraft to remain on the
transect line whilst observations are made. Passing mode easily allows for double
platform observers and avoids introducing the bias that is created by effort being
expended before the craft returns to the transect line (Dawson et al., 2008). The
preference of strip-width transects in passing mode as a survey method for dugongs is in
contrast to the preference of distance sampling which is employed for most other marine
mammals, including inshore dolphins (Hammond, 2010). Dolphins are more mobile,
surface more frequently and tend to spend more time at the surface than dugongs do. For
these reasons, dolphins are more ‘available’ (for detection) at the surface than dugongs
are. Distance sampling involves measuring the distances of animals seen at the surface
from the line transect on which the observer is travelling (Hammond, 2010). It
incorporates the detection function g (0) which is the probability of detecting an object
based on its distance from the transect (or other given line or point) (Buckland et al.,
1993). When conducted in closing mode with the circle-back procedure, distance sampling
yields data with high certainty in species identification, group size and group composition
(Dawson et al., 2008). In addition to the use of strip-width methods in passing mode
instead distance sampling in closing mode, observers recorded dolphin sightings only as a
second priority to dugongs. Thus, it is possible that dolphins were missed or misidentified.
This project does not attempt to use dolphin sightings data to extract an index of dolphin
population abundance. This follows the advice of Anderson (2001) who strongly
discourages the generation of an index from ‘convenience sampling’ as “such numbers are
not trustworthy and cannot lead to valid inferences about the population of interest”.
This chapter outlines the methods of acquiring and processing dolphin sightings and
environmental data, and explores the performance of the data generated in preliminary
models. Spatial data processing can underpin model reliability as it is through this phase
that training and test data are generated (Elith & Leathwick, 2009; Redfern et al., 2006).
Limitations of the opportunistic dolphin data were considered during spatial data
processing. Decisions on how to achieve appropriate resolution, represent survey effort,
treat uncertain sightings and avoid introducing pseudo-absence were also made at this
stage. This chapter describes the problems encountered and explains the compromises
made in generating the training data. It presents the results of preliminary models
Chapter 2
21
(generalised additive models and component-wise boosting) and suggests an alternative
model approach that may better suit the data.
2.2 Acquisition of dolphin data
2.2.1 Study area
The study area is located in the western Pilbara, northern WA (Figure 1). The study area
covered a 10,600 km2, which encompassed Exmouth Gulf and the coastal area (mainland
to the 20 m isobath) as far north-east as the Lowendall Islands (Figure 3). The study was
over a 12 month duration, with data collected over five survey periods: May, July, October
and December 2012, and May 2013. The study area consisted of three blocks: Exmouth,
Onslow and a repeated block, which was situated within the Onslow block. The repeated
block, which was surveyed an additional time during each survey period, allowed
information on inter-seasonal variability in dugong distribution to be obtained (Hodgson,
2012). Transects were of a similar layout to those of previous studies (Hodgson, 2007;
Preen et al., 1997; Prince et al., 2001), in order compare results. Flights were at an altitude
of 500 ft (137 m) above sea level (ASL) (Hodgson, 2012).
Figure 4: Study area, transect configuration and survey blocks.
Chapter 2
22
2.2.2 Sighting conditions
The aerial surveys were conducted in generally good weather conditions for sighting
dugongs and, by default, dolphins (Hodgson, 2012). Wind, cloud cover and Beaufort sea
state (BSS) were low (mode = 5 – 10; 0; 1 – 2, respectively) and air visibility was high
(mode = >10 km) (Table1).
Table 1: Summary of the environmental conditions experienced during each survey.
Environmental condition
Summary Survey 1 2 3 4 5
Wind speed (knots)
Min Max
Mode
2 6 5
4 10 10
7 9 9
5 10 6
3 8 8
Air visibility Min Max
>10 km >10 km
>10 km >10 km
>10 km >10 km
>5 km >10 km
>10 km >10 km
Cloud cover (oktas)
Min Max
Mode
0 3 0
0 7 0
0 1 0
0 8 0
0 2 0
Beaufort sea state
Min Max
Mode
0 2 1
1 3 2
0 4 2
0 3 1
0 4 2
Turbidity Min Max
Mode
1 4 3
1 4 4
2 4 4
1 4 4
1 4 4
Glare Min Max
Mode
0 3 1
0 3 3
0 3 3
0 3 3
0 3 0
1: May 2012; 2: July 2012; 3: Oct 2012; 4: Dec 2012; 5:May 2013
2.2.3 Observer protocols
As described in the introduction, survey procedures followed standardised methods that
have been developed for dugong aerial surveys (Marsh & Sinclair, 1989; Pollock et al.,
2006). Experienced observers were employed to collect the data (Hodgson, 2012). The use
of experienced observers maximises the reliability of sighting records and provides
assurance that observers are able to maintaining focus over long periods of potential
discomfort (Dawson et al., 2008). To maximise accuracy and reduce bias in sightings, two
observers (one at the front and one at the rear) were positioned on each side of the
aircraft. Observers surveyed 200 m strip-width transects on either side of the aircraft
(Figure 5A). To ensure that sightings were made independently, observers were
acoustically isolated with the use of headphones.
Chapter 2
23
Marker rods delineating this strip width were attached to pseudo wing struts on the
outside of the aircraft. Within each strip-width were four 50 m zones marked very high,
high, medium and low, with very high being furthest away from the transect line and low
being the nearest (Figure 5B). The inner marker was placed where the two observers
could begin to see the water through the window, which was not directly underneath the
aircraft (Hodgson, 2012; Marsh & Sinclair, 1989). However, the front seat observers were
able to see a wider space strip because they had bubble windows. This space is referred to
as being ‘inside’ the strip. The space beyond the strip width, further away, is referred to as
being ‘outside’ of the strip width. Dolphin sightings that were recorded as being ‘inside’
and ‘outside’ of the strip were retained within the dataset. A Garmin eTrex 10
Geographical Positioning System (GPS) recorded the location of the every two seconds, to
15 m positional accuracy.
Figure 5: Strip-width transect as seen from the aircraft by observers (A) and zones within the strip-width transect (B). Dolphin sightings were positioned to the centre
of each zone in which they were sighted, using GIS.
A
B
Chapter 2
24
Observers recorded the following parameters for each dolphin sighting, using digital voice
recorders:
genus (or unknown);
genus certainty level (certain, probable or guess);
group size;
number at surface;
number of calves;
zone within the strip-width transect, or whether the sighting was ‘inside’ or
‘outside’ the strip-width transect; and
turbidity (Table 2).
Table 2: Turbidity scale.
Turbidity Water quality Depth range Sea floor visibility
1 Clear Shallow Clearly visible
2 Variable Variable Visible but unclear
3 Clear > 5m Not visible
4 Turbid Variable Not visible
The turbidity scale (Table 2) was specifically developed to overcome bias in detecting
dugongs (Pollock et al., 2006). The scale is not a true measure of turbidity as it represents
a combination of both depth and turbidity. Furthermore, this measure is not relevant for
inshore dolphins as they spend substantially more time at the surface of the water column
than dugongs do.
Post-processing of dolphin sightings data in ArcGIS v 10.1 (ESRI, 2012) involved adjusting
the sighting location from the track line (i.e. the position of the aircraft when the observer
recorded the dolphin sighting) to the centre of the zone in which the animal occurred
(Figure 5B). This adjustment was based on flight direction, observer seating position and
dolphin zone within the strip. Details of this procedure are provided in Appendix A. These
adjustments had to assume that the location of each dolphin was perpendicular to the
aircraft, as the bearing of each sighting was not recorded.
Chapter 2
25
2.2.4 Summary of dolphin sightings
A total of 646 dolphin sightings were recorded throughout the five survey periods. From
these, approximately 60% were of bottlenose (n=388), 13% were of humpback (n=84),
and 1% was of snubfin dolphins (n=6). The remaining groups were not identified to
species level. No mixed-species groups were recorded. Due to the low sample size, snubfin
dolphins were no longer considered from this point. Encounter rates for both bottlenose
(2.11 – 4.34 sightings / 100 km) and humpback (0.39 – 1.10 sightings / 100 km) dolphins
were low across survey periods (Table 3). These low encounter rates were taken as a
forewarning that the training dataset would likely be characterised by low sample size.
Sighting conditions were consistent between the five surveys (Table 3). More than half
(53.9%) of dolphin sightings from May 2012 and just over one third (31.6%) of dolphin
sightings from July 2012 had to be removed from the genus-specific datasets, due to
uncertainty (Table 3). The highest proportion of uncertain dolphin species recorded
during the first survey was perhaps reflecting the start of a new survey program when
observers were less experienced than they were later on in the program.
Table 3: Sighting conditions, species sightings and encounter rates and the proportion of sightings that were of uncertain species for each survey*.
Survey Conditions (mode) Sightings Number (Encounter rate*)
Uncertain (%)
BSS Turbidity Glare Cloud Bottlenose Humpback Uncertain
1 1 3 3 0 48 (2.11) 11 (0.48) 69 (3.03) 53.9 2 2 4 3 0 64 (2.81) 25 (1.10) 43 (1.89) 31.6 3 2 4 3 0 89 (3.91) 15 (0.66) 16 (0.70) 13.2 4 1 4 3 0 88 (3.86) 9 (0.39) 16 (0.70) 14.2 5 2 4 0 0 99 (4.34) 24 (1.05) 24 (1.05) 16.2
*Effort for each survey was 2,279 km. Survey 1: May 2012; 2: July 2012; 3: Oct 2012; 4: Dec 2012; 5: May 2013. **Sightings / 100 km. BSS 1: Light air, ripples on the sea with no crests forming; BSS 2: Light breeze, small wavelets on the sea with unbreaking crests. Turbidity 3: Clear water, > 5 m deep, seafloor is not visible; Turbidity 4: Turbid water, variable depths, seafloor is not visible. Glare 3: > 50% of view affected. Cloud 0: Completely clear sky.
Chapter 2
26
Bottlenose and humpback dolphin group sizes were generally small, although some larger
groups of bottlenose dolphins were also recorded ( =3, SD=4.2 for bottlenose dolphins;
=2, SD=1.6 for humpback dolphins) (Figure 4; Figure 5). The largest (>25) bottlenose
dolphin groups were recorded in October and December (Figure 4; Figure 6 C, D). The one
humpback dolphin group greater than five (n=7) was recorded in July, near Barrow Island
(Figure 5; 7B).
Bottlenose (Figure 6) and humpback dolphins (Figure 7) were both sighted throughout
the survey region without obvious clumping. In December 2012, humpback dolphins were
only recorded in the north-eastern portion of the study area (Figure 7D).
Figure 6: Count of bottlenose dolphin group sizes per survey.
Figure 7: Count of humpback dolphin group sizes per survey.
0
5
10
15
20
25
30
35
40
45
50
1 2 - 5 6 - 10 11 - 15 16 - 20 21 - 25 >25
Count
Group size
May-12
Jul-12
Oct-12
Dec-12
May-13
-1
4
9
14
1 2 - 5 6 - 10
Count
Group size
May-12Jul-12Oct-12Dec-12May-13
Chapter 2
27
Figure 8: Location and group size of bottlenose dolphins per survey.
Chapter 2
28
Figure 9: Location and group size of humpback dolphins per survey.
Chapter 2
29
2.3 Preparation of training data
2.3.1 Response variables
Sightings data can be processed in different ways in order to create a certain type of
response variable. Response variables may be based on presence-only or presence-
absence information, incorporating survey effort. When survey effort is available,
encounter rates, group sizes and absolute abundance can be modelled (Zuur et al., 2009).
As long as they are reliable, such data may provide stronger statistical modelling power
than measures of just presence and absence. As the dugong aerial surveys employed a
systematic transect design, effort data is available for this modelling project. The choice of
response variable must also consider the purpose of the SDM project and the nature of the
data to be modelled (Elith & Leathwick, 2009; Redfern et al., 2006). The nature of data
collection can affect certainty in species records, sample size, incidence of pseudo-
absences and sampling bias, all which need to be taken into account when deciding how to
process the data (Alvarado‐Serrano & Knowles, 2014). Pseudo-absence in this instance
refers to when a species has incorrectly recorded as absent when really it was just not
detected, or had been misidentified (Gu & Swihart, 2004)1.
One of the aims of this study was to estimate the probability of dolphin occurrence across
the study area, which could have been done using either a binary presence-absence or
encounter rate dataset. A binomial presence-absence response variable is generated
through creating a grid of consistently sized cells across the study area, and then joining
sightings and effort data to each cell. Encounter rate is generated by splitting the survey
track into segments, and then joining sightings and effort data recorded along that
segment to the centre of each segment. It was decided that a binomial dataset be used for
several reasons, but primarily due to the benefits that attaching data to a grid provided.
1. Attaching data to grid cells allowed for a more accurate spatial representation of
strip-width survey effort and the ‘actual location’ of dolphin sightings data, than a
segmented line would allow. Accurately portraying survey effort is essential in
compiling a truthful presence-absence dataset (Gu & Swihart, 2004; Lozier et al.,
2009).
1 The term pseudo-absence can also refer to artificial absence data that has been created by the modeller, as an alternative to true absence data when it is not available (Barbet‐Massin et al., 2012). In order to avoid confusion, the alternative term ‘background data’ has been used within this thesis (see Chapter 3).
Chapter 2
30
2. Grid cells containing ‘uncertain’ dolphin sightings could be removed from the
species-specific datasets, which reduces the problem of pseudo-absence. If
uncertain sightings were left within segments, the encounter rate would be
unreliable. Removal of entire segments that contained uncertain dolphin sightings
would have reduced the sample size, which impedes model performance.
3. Turbidity could not be accounted for, but was known to vary throughout the
survey area. Therefore, detectability of dolphins across the study area was not
homogenous. Segmenting survey tracks would combine turbid and non-turbid
areas, thereby further decreasing reliability in encounter rates. This was not an
issue for small grid cells, but it would have been for large grid cells.
Presence-absence data were generated for ‘all dolphins’, bottlenose and humpback
dolphins. ‘All dolphins’ included certain and probable sightings of bottlenose, humpback
and snubfin dolphins as well as all uncertain sightings. An overview of the processing
steps undertaken to generate the response variable data is presented in Figure 8A.
The spatial area actually surveyed by the strip-width transects was generated following a
number of steps (Figure 9). The planned transect lines feature layer was exported five
times, in order to create a new layer for each survey period. Transects within the repeated
area were exported an extra time. The planned transect lines were manually digitised
using the ‘editor’ toolbar based on the point coordinates of each survey track actually
flown. To represent the areas from which dolphin sightings were recorded, a buffer of
400 m from either side of the transect line was created. To remove the unseen area
underneath the aircraft, a buffer of 125 m from either side of the transect line was created
and then erased from the wider area that was seen by observers.
Chapter 2
31
A
B
Figure 10: Spatial processing steps undertaken in order to prepare training data. A) Preparation of the dolphin presence-absence (P-A) response variables. B) Extraction of environmental predictor variables to each response datum. This process was undertaken for each survey
and combined ‘seasonal’ datasets. It was also undertaken separately for ‘all dolphins’, bottlenose and humpback dolphins. *Cells that contained ‘unknown’ dolphins were removed from species-specific datasets in order to reduce pseudo-absences. SST: Sea surface temperature. MHWM:
Mean high water mark. SD: Standard deviation.
Input data
- Dolphin sightings (point)
- Survey tracks (point)
Rectified data
- True dolphin locations (point)
- True strip-width survey effort
(polygon)
Attached data
- Two grids
500 m x 500m
1,500 x 1,500 m
Modified gridded data
- Summed effort
- Removed unsurveyed cells
- Removed 'unknown' dolphins*
Converted gridded data to
point data
Input data
- Daily SST (gridded)
- Bathymetry (gridded)
- Coastline (MHWM) (line)
Generated new gridded data
- Mean SST (survey / season)
- Mean 'front' (SST SD)
- Distance from islands
- Distance from mainland
- Depth
- Slope
- Aspect
- Topographical complexity
Extracted gridded environmental data to
each P-A point
Chapter 2
32
Figure 11: Process of generating strip width area from the transect line.
A grid of even-sized cells across the survey area was created using the ‘fishnet’ tool. Spatial
resolution refers to the size of these cells and specifies, in spatial terms, how detailed
information is across the study area. Temporal resolution refers to the frequency of data
collection within a given time period. Determining optimum resolution for an SDM is
subjective and without any ‘rules of thumb’ (Guisan et al. 2007; Redfern, Ferguson et al.
2006 Hengl 2006). Choice of resolution for this study was guided by ecological relevance,
cautionary advice (Gregr, 2004; Guisan et al., 2007; Redfern et al., 2006), maximum
positioning error in dolphin sightings (Guisan et al. 2007) and previous studies (Blasi &
Boitani, 2012; Gómez de Segura et al., 2008; Parra et al., 2006b; Torres et al., 2003; Toth et
al., 2011).
Tracing back to the ‘Modifiable Areal Unit Problem (MAUP)’, which Openshaw (1981)
described as “when the same analyses on the same data produces different results because it
has been processed at different scales”, it is possible for spatial and temporal resolution to
be either too coarse or too fine to recognise true distribution or movement patterns. It is
therefore important that the modeller selects a resolution that is neither exceptionally
large or small (Holling, 1992; Redfern et al., 2008; Seo et al., 2009) by considering species
biology, landscape features and data characteristics (Austin, 2002; Redfern et al., 2006).
Spatial resolution should also be at an appropriate size to capture the patchy nature of a
land, or sea, scape (Austin, 2002).
Reducing resolution coarsens data. By decreasing the number of grid cells but increasing
their size, the values of the data attached to the cells are averaged, or ‘smoothed’. This can
result in a loss of information but it can also lead to sightings data being ‘spread’ to an area
where they didn’t occur. For instance, joining a species sighting point to a grid cell inflates
the spatial area that the point information covers. The coarser resolution, the wider the
area that grid cell data is smoothed over. Thus, a larger grid cell would represent a larger
area of dolphin presence, which might not be true. This can result in a ‘forced match’ with
Chapter 2
33
environmental data. This has the potential to lead to a Type 1 error – finding a statistically
significant relationship that does not ‘truly’ exist (Guisan et al. 2007). Gregr (2004)
specifically cautions against this, citing the results of Hamazaki (2002) as an example of a
Type 1 error having occurred due to a too-coarse resolution. The cells were so large that
data were no longer independent, due to issues of spatial autocorrelation and value
averaging over a broad expanse. Increasing resolution by increasing the number of grid
cells but decreasing their size can also create errors in the data. Creating a subset of
smaller cells from existing ones can introduce inaccurate values and unwanted noise.
Thus, cell size must be determined so that the degree of distortion is acceptable and does
not weaken the data integrity.
A priori ecological knowledge must be also be factored into resolution choice (Elith &
Leathwick, 2009). Environmental features can influence species distribution in different
ways, largely dependant on spatial scales As inshore dolphin are generally restricted in
their range and movements, it is likely that their populations are influenced at a fine scale.
Maintaining detail in coastal and bathymetric features was important as humpback
dolphins have a restricted range and are usually distributed close to the coast (Parra et al.,
2006b). In South Africa, humpback dolphins do not venture further than 500 m from shore
(Karczmarski et al., 2000), so it was also important that the cells were small enough to
account for such coastal dependency. This is in contrast to modelling projects for wider
ranging pelagic species. For instance, (Redfern et al., 2008) found that ecological
relationships for four offshore delphinid species were consistent across spatial resolution
ranging from 4 to 120 km. Their environmental variables were sea surface temperature
(SST), ocean fronts, chlorophyll, thermocline depth and strength and bathymetric depth.
(Dawson et al., 2008) emphasise that coastal environments are more complex than the
open ocean and coastal features influence local ecology. Environmental variables selected
for this project are discussed later in this section, under the heading ‘Environmental
variables’.
Understanding maximum positional error allows for judgment-based decisions on
resolution (Guisan et al. 2007). For this study, the potential positional error is how far
ahead or behind a dolphin (or dolphin group) was of the observer’s position in the aircraft.
Grid cells of 500 m by 500 m were large enough to capture the mean (165 m) potential
positional error of dolphin sighting locations. This value includes the potential GPS error
of 15 m. Dolphin locations could not be corrected in relation to them being forward or aft
of the observer’s position in the aircraft because bearings were not recorded. Instead, they
were assumed to be perpendicular to the aircraft).
Chapter 2
34
A fine scale grid across a wide study area has the potential to cause the response data to be
zero-inflated. This reduces sample prevalence, which is the proportion of presence data to
absence data and can be difficult to model (Barbet‐Massin et al., 2012). Training data
were created at alternative resolutions, and test the effect that they had on model fit.
Three spatio-temporal combinations were created. From finest to coarsest resolution,
these were:
500 m x 500 m (0.25 km2) grid cells for each survey period;
1,500 m x 1,500 m (2.25 km2) grid cells for each survey period;
1,500 m x 1,500 m (2.25 km2) grid cells by season.
To increase sample size (and therefore sample prevalence), data were also pooled on a
seasonal basis (autumn, winter and spring/summer). The procedure for this is described
later in this section, under the heading ‘Environmental variables’.
Areas of land were removed from the grid so that absence data from islands would not be
included within the dataset. The ‘intersect’ tool was used to intersect each survey effort
polygon with the grid. This results in feature layer that is of the same shape and size as the
effort polygons but contains the grid cells from the grid. This was done so that survey
effort per grid cell could be calculated. A column ‘area surveyed’ was created for each
polygon in the attribute table, and populated with the area that was surveyed. The new
intersect layers were attached back to the grid using the ‘spatial join’ tool. This was done
individually for each survey period and each season. Once all polygons were attached, a
column ‘total area surveyed’ was created. The sum of effort received by each grid cell was
summed and entered into the ‘total area surveyed’ column. Cells without survey effort
were deleted using the ‘select attributes’ tool. Survey effort polygons were mapped and a
colour ramp was used to symbolise highest effort (red) to lowest effort (blue). These
outputs were visually verified. The highest effort received by any one cell at the end of a
repeated block strip, which overlapped with the start of an Exmouth block strip. Ensuring
that each grid cell contained accurate survey effort information was important so that data
could later be truncated and a weighted regression could be used in the modelling. These
are both treatments that can help reduce statistical problems induced by a zero-inflated
dataset (Elith & Leathwick, 2009; Hirzel et al., 2006).
Chapter 2
35
Figure 12: Example of effort data calculated for the survey area. This represents May 2012 at a 500 m x 500 m grid cell resolution.
Finally, dolphin sightings data were added to the survey effort cells using the ‘spatial join
tool’. A new column ‘Presence’ was added to the attribute table. These cells received ‘1’
when dolphin sightings were joined. Cells not receiving a dolphin sighting retained a ‘0’
value, thus creating a binary code for dolphin presence and absence. Cells containing
sightings of ‘unknown’ dolphin species were removed entirely to avoid having pseudo-
absences in the data. This also followed the advice of Lozier et al. (2009), who
demonstrated that spurious, yet credible-looking, SDMs could be produced if there was
uncertainty in taxonomic records.
The presence-absence gridded data were converted to points within the centre of each
grid cell so that environmental variable values could be later extracted through the
‘extract multiple values to points’ tool at each of these point locations (Figure 8B).
Chapter 2
36
2.4 Environmental variables
It is well documented that model accuracy is improved when good a priori understanding
of the target species’ ecology has been incorporated into its design Austin (2007); (Doniol-
Valcroze et al., 2012; Guisan & Thuiller, 2005; Hernandez et al., 2006; Jiménez‐Valverde
et al., 2008; Robinson et al., 2011). As Gregr (2004) stated, (if) “we are ultimately
interested in better understanding the processes that drive the patterns we see we must
integrate that which we already know”. A literature review was undertaken to examine the
performance of environmental variables used in other SDMs for coastal dolphin species.
However, this review yielded mixed results (Table 4). As such, there were no obvious, or
standard, ‘stand out’ variables to be used, based on these other studies.
As dolphins are apex predators, their distribution is likely to be highly influenced by prey
availability (Benoit-Bird & Au, 2003; Hammond et al., 2013). However, prey data does not
exist for this study area. As prey data are often difficult to obtain, the use of proxy or
surrogate, datasets, are commonplace (Elith & Leathwick, 2009). Selection of
environmental variables was based on ecological relevance, model purpose, data
accessibility, reliability and availability for the study area extent at a sufficient resolution.
This section describes why certain variables were sourced, where they were sourced from,
how they were processed and a summary of what information they contained. Acquisition
and processing is summarised in Table 5 and the resultant layers are presented in Figures
13-17.
2.4.1 Sea surface temperature and ocean fronts
SST and ocean fronts (fronts, from herein) may be used as a proxy for prey distribution.
Oceanographic currents, areas of upwelling and eddies, increase nutrients, attract
zooplankton and transport larval recruitment of finfish. SST is highly dynamic in coastal
environments and has been shown to have a seasonal effect on dolphin distribution,
including a shift from inshore to offshore areas (Brager et al., 2003; Garaffo et al., 2011;
Heithaus & Dill, 2002; Smith, 2012). Standard deviation (SD) in SST is often used to
represent fronts; with the assumption that areas of varying water temperature are areas
where productive oceanographic processes occur (Becker et al., 2010).
Mean SST for each survey period was created from daily, remotely sensed Multiscale
Ultrahigh Resolution (MUR) blended SST (Figure 13; 14; 15). These data were obtained
from the Physical Oceanography Distributed Active Archive Center (PO.DAAC) at the NASA
Jet Propulsion Laboratory, using Marine Geospatial Ecology Tools (MGET) (Roberts et al.,
Chapter 2
37
2010).
To create the third spatio-temporal resolution (1,500 m x 1,500 m grid cells by ‘season’),
SST data for each survey period was visually inspected for consistent patterns between
survey periods. Patterns and temperatures were broadly consistent between May 2012
and 2013 and between October and December 2012 (Figure 13). Using the ‘raster
calculator’, new mean SST and front layers were created to represent each ‘season’. May
2012 and May 2013 were combined to create an autumn layer (Figure 15, C). October
2012 and December 2012 were combined to create a spring / summer layer (Figure 15 B).
The July 2012 was renamed as the winter layer (Figure 15 A).A front raster for each
survey period and season was created by calculating SST SD between each grid cell and
those surrounding it (MacLeod, 2013) (Figure 14; 15 D, E, F).
SST (Figure 15 A, B, C) and fronts (Figure 15 D, E, F) both varied between seasons. The
highest temperatures were recorded in autumn (28.5 OC) and the lowest were in winter
(20 OC). SST was not static through the study area. During winter, temperatures were
coolest in the north-eastern section of the study area. In contrast, temperatures were
coolest in the south-western section of the study area (Exmouth Gulf) during
spring/summer and autumn. The most notable frontal feature was in winter with the
highest value representing SST SD between adjacent cells being 0.24. The front was
positioned at the seaward edge of the study area, in deep water. Fronts during other time
periods were less notable, with values representing SD between adjacent cells being less
than 0.15.
2.4.2 Distance from coast
Distance from coast has the potential to influence coastal dolphin distribution, but a
significant relationship is not always found (Bailey & Thompson, 2009; Torres et al.,
2003). Distance from both the mainland and islands will be modelled, as there appears to
be variation in the distribution between bottlenose and humpback dolphins. It has been
suggested that the waters surrounding islands of the study area are important habitat for
a suite of marine megafauna species, including dolphins (DEWHA, 2008; Kendrick &
Stanley, 2001). However, no formal research has been undertaken to confirm this.
Distances from both islands and the mainland were calculated using the Euclidean
Distance tool, and the mean high water mark within (Geoscience Australia, 2004)
‘GEODATA Coast 100k’ dataset.
Chapter 2
38
2.4.3 Bathymetric derivatives
Bathymetric derivatives are commonly used as surrogates of prey data (Elith & Leathwick,
2009; Guisan & Zimmermann, 2000; Redfern et al., 2006) (e.g. Bailey & Thompson, 2009;
Blasi & Boitani, 2012; Cañadas et al., 2005; Garaffo et al., 2011; Gómez de Segura et al.,
2008; Parra et al., 2006, Table 4). Depth affects temperature, salinity, light attenuation,
pressure and the availability of elements such as calcium which influence finfish presence
(Elith & Leathwick, 2009). Topographical complexity could be modelled in lieu of
substrate data (Bailey & Thompson, 2009). This could indicate locations of limestone
platforms and coral bommies, which are known to exist in the area and are likely to be
important fish habitat at a landscape scale (Brager et al., 2003; MacLeod, 2013). Aspect
and slope can relate to areas of upwelling that often attract fish and cephalopods, principal
prey for coastal dolphins. Aspect offers a more detailed interpretation of interaction
between currents and the seabed (MacLeod, 2013).
Depth, slope, aspect and topographical complexity were derived from Geoscience Australia
(2009) gridded bathymetry data. In lieu of rugosity data, topographical complexity was
modelled from depth using the Focal Statistics tool, which compares the values of each cell
in the raster against those surrounding it, with greater variation indicating higher
complexity (MacLeod, 2013).
The study area is generally flat and shallow, with depths increasing at the outer, seaward,
edge of the study area (Figure 17 A). This is the location of the 20 m bathymetric contour,
and the only notable slope within the study area (Figure 17 A). The large expanse between
Tent Island, Muiron Islands and Onslow (Figure 17 A, B, C, D) is particularly flat and
shallow. The slope, topographical complexity and aspect datasets depicted unexpected
linear features to the south of Barrow Island (Figure 5 B, C, D). Visually checking these
areas on a marine navigational chart revealed that these areas have been unsurveyed,
suggesting that the gridded bathymetry data (Geoscience Australia 2009) have been
interpolated in these areas. It is not known what how the interpolated data may affect
SDM performance, but the data were used as it was the best data available.
2.4.4 Omitted variables
Considerable effort was expended on attempting to source other environmental data that
could influence the distribution of inshore dolphins in the western Pilbara. Unfortunately,
there is a lack of information available. Benthic habitat maps were sought, but there have
been no consistent surveys undertaken for the spatial extent of the study area.
Chapter 2
39
Chlorophyll, salinity and coastal feature (mangroves, saline coastal flats and watercourses)
data were sourced because of their potential to seasonally influence prey abundance or
availability (Redfern et al., 2008; Redfern et al., 2006), but were not able to be utilised. On
investigation, chlorophyll and coastal feature data were deemed to be inaccurate
(Appendix B). Salinity data did not cover the entire study area and was at a scale that is
believed to be ecologically irrelevant in influencing dolphin distribution (Appendix B). The
study area is characterised by high cyclonic activity that can bring high rainfall. Increased
riverine input results in increased turbidity and decreased salinity than usual, which can
in turn alter prey distribution. However, surveys were not undertaken following high
rainfall events.
Turbidity could affect both the occurrence and detectability of inshore dolphins, and is a
major characteristic of the Pilbara region. Near-shore waters of the study area are
brownish-red due to suspended sediments that are rich in iron oxide. Plumes of red
suspended sediments are caused by increased terrestrial run off during rainfall events and
by dredging operations. In contrast, oceanic waters further offshore around the islands are
clear and blue (Viddi et al., 2010). It is not known whether turbidity would affect the
occurrence of either species as both have been recorded in clear and turbid waters within
the study area (Allen et al., 2012). Dolphins could avoid turbid waters due to presence of
predatory shark species (Heithaus, 2001), or be attracted to them when associated with
highly productive mangroves and river mouths (Parra et al., 2006b). This variation is
likely to result in detectability bias but dolphin distribution could also vary between these
areas. While it would be useful to investigate these influences, turbidity data could not be
obtained for this project. The turbidity data collected as part of dugong aerial survey
protocol cannot be used to do so. It is more a measure of ‘visibility’ for detecting dugongs
rather than being a true measure of turbidity, as it incorporates depth. Furthermore, that
data is only available for the locations where dolphins were present. The team leader
recorded changes in turbidity and depth for their side of the aircraft but this dataset was
not informative enough to construct a reliable turbidity spatial layer of information for the
study area. Satellite and aerial imagery were unsuccessfully sought to incorporate
turbidity as a predictive environmental variable. Aerial imagery was not available at a
temporal resolution that matched the survey period. Satellite imagery did not meet the
resolution requirements or budget constraints of this project.
Chapter 2
40
Table 4: Literature review results for environmental predictors of coastal dolphin distribution.
Common bottlenose
dolphin T. truncatus
Humpback dolphin
S. chinensis / S. sahulensis
Snubfin dolphin
O. heinsohni
Hectors dolphin
C. hectori
Chilean dolphin
C. eutropia
Commerson’s dolphin
C. commersonii
Dusky dolphin L. obscurus
Depth Y1; 2 N7 N8 N8 Y3 Y12 N6 N6
Slope Y1 N1;7
Distance from coast Y1; 2 N10; 11 Y8 Y8 Y12 Y6 N6
Distance from river mouth N8 N8 N6 Y6
Proximity to a sea mount Y4
Sea surface temperature N7 Y3 Y6 Y6
Channel width Y12
Coastal complexity Y12
Water clarity Y3
Chlorophyll concentration N7 Y6 Y6
Season Y3; 9
Diurnal v Nocturnal N9
Group size Y10; 11
Individual coloration Y11
Behavioural avoidance Y11
Occurrence of a barnacle sp. Y11 1Bailey & Thompson, 2009; 2Blasi & Boitani, 2012; 3Brager et al., 2003; 4Cañadas et al., 2005; 5Cañadas & Hammond, 2006; 6Garaffo et al., 2011; 7Gómez de Segura et al., 2008; 8Parra et al., 2006; 9Rayment; 10Torres et al., 2003; 11Toth et al., 2012.; 12Viddi et al., 2010
Y : Relationship tested and found. N : Relationship tested but not found.
Chapter 2
41
Table 5: Acquired environmental datasets and derived variables.
Acquired data Derived variable
Product Publisher Method Spatial res. Temporal res. Variable GIS Tool
Australian Bathymetry and Topography Grid
Geoscience Australia (2009)
Swath bathymetry and digitised chart data
250 m N/A (static) Depth Topographic
complexity ‘Focal statistics’ (SD)
Aspect ‘Aspect’ Slope ‘Slope’
GEODATA Coast 100k
Geoscience Australia (2004)
Western Australian Government surveys
100 m N/A (static) Distance to mainland ‘Euclidean distance’ Distance to island ‘Euclidean distance’
GHRSST Level 4 MUR Global Foundation Sea Surface Temperature Analysis* V1.3
NASA Jet Propulsion Laboratory (2013)
Modelled from high resolution satellite mounted remote sensors
1 km Daily Mean SST ‘Raster calculator’ (mean)
1 km Daily Fronts ‘Focal statistics’ (SD) using mean SST layer
*Sourced using Marine Geospatial Ecology Tools (MGET). GHRSST: Group for High Resolution Sea Surface Temperature; MUR: Multiscale Ultrahigh Resolution; SD: Standard deviation; SST: Sea surface temperature
Chapter 2
42
Figure 13: SST (°C) for each survey period.
Chapter 2
43
Figure 14: Fronts (SST SD between grid cells) for each survey period.
Chapter 2
44
Figure 15: A-C) SST (°C) for each season; D – F) Fronts (SST SD between grid cells) for each season.
Chapter 2
45
Figure 16: Euclidean distance (km) from A) mainland and B) islands.
Chapter 2
46
Figure 17: Bathymetry of the study area: A) depth (m), B) slope (°), C) aspect (cardinal direction), and D) topographical complexity (slope SD between grid cells).
Chapter 2
47
2.5 Data export and exploration
The response variables produced for modelling were the presence-absence of ‘all
dolphins’, bottlenose and humpback dolphins. The environmental predictor variables
were SST, fronts, distance from mainland, distance from islands, depth, slope, aspect,
topographical complexity and survey period / season. The ArcGIS ‘Extract multiple values
to points’ tool was used to extract environmental data to each presence-absence point. An
output table for each survey period and season was exported as a ‘comma separated
values’ file for error checking and import into ‘R’ software (R Development Core Team,
2013).
Exploratory data analysis was undertaken to assess whether data would meet the
principle assumptions of regression analysis. These assumptions are that data are
normally distributed and independent, that there is homogeneity of variance and that data
is not zero-inflated (Zuur et al., 2010). If assumptions are violated, there is a non-trivial
chance that analysis will result in a Type 1 (e.g. determining that no ecological relationship
exists when it does) or Type 2 (e.g. determining that an ecological relationship does exist
when it doesn’t) error.
Data were first plotted as box plots to investigate data mean and spread, and to identify
outliers. Next, data were plotted as histograms to understand the distribution of the data.
A Box-Cox power transformation (Box & Cox, 1964) was applied to continuous data to
remove outliers. Data were also mean-centred and scaled to unit variance to achieve
homogeneity. Transformed data were assessed for normality using histograms, q-q plots
and the Kolmogorov-Smirnov test for normality. Spearman’s rank correlation test was
applied to assess for collinearity (i.e. correlations between variables). High collinearity
was identified between slope and topographical complexity (rs = 0.94, Table 6), and
between SST and season (rs = -0.93, Table 6). SST and fronts were not correlated (rs = 0.15).
Spatial autocorrelation, when data consist of similar values as a result of their geographic
proximity, is another concern for SDMs. This was not anticipated to be a problem for this
project as the survey design ensured that transects were evenly spaced across the study
area and were perpendicular to the bathymetry of the area. To confirm this was not an
issue for the project, sightings of all, bottlenose and humpback dolphins were plotted
against each environmental variable. No patterns in data distribution were apparent.
Chapter 2
48
Table 6: Spearman’s rank correlation coefficients between environmental variables.
sst fro edml edis dep sl asp toco
sst
fro -0.15
edml 0.17 0.21
edis 0.01 0.07 0.2
dep -0.06 -0.18 -0.39 -0.57
sl 0.03 0.07 0.08 0 -0.34
asp 0.01 0.04 0.08 0.08 -0.32 0.55
toco 0.03 0.08 0.09 0.02 -0.39 0.94 0.56
survey 0.48 -0.23 0.00 0.00 0.00 0.00 0.00 0.00
season -0.93 0.23 0.00 0.00 0.00 0.00 0.00 0.00
fro: front; sst: sea surface temperature; edis: distance from islands; edml: distance from mainland; dep: depth; toco: topographical complexity; asp: aspect; sl: slope.
Low sample prevalence, as a result of zero-inflation, is problematic for modelling a
binomial presence-absence dataset (Barbet‐Massin et al., 2012). This was identified as a
potential problem for this project in the early stages of spatial data processing. All models
were run at three alternative spatio-temporal resolution combinations. This was to
determine whether model fit to data could be improved, by increasing sample size. For
each response variable at each resolution, the number of both presence and absence
records were plotted as a bar chart (Figure 18). Use of a coarser resolution was successful
in increasing sample prevalence. Increasing the cell size from 500 x 500m and 1,500 x
1,500 m resulted in a nine-fold increase in grid cell size, which reduced the amount of
absences in the dataset. Pooling the data from survey period to a seasonal basis also
decreased the proportion of absences in the dataset. However, there was also a loss of
some presence records as resolution was increased (Figure 18), as an unwanted artefact
of interpolation. This is because reprocessing data at alternative resolutions entailed
merging cells across space and time. Data were truncated by removing 10% of absence
records (Table 7). Those removed represented grid cells that had received very little
survey effort.
Chapter 2
49
Figure 18: Sample prevalence of for presence-absence of ‘all dolphins’, bottlenose
and humpback dolphins, at each spatio-temporal resolution combination.
Chapter 2
50
Table 7: Sample prevalence, expressed as the percentage of presence records, for each response variable at each spatio-temporal resolution, pre and post 10%
truncation of absence records.
A B C
Pre Post Pre Post Pre Post
‘All dolphins’ 1.04 1.15 1.7 1.89 7.94 8.82 Bottlenose dolphins 0.62 0.69 1.02 1.14 4.63 5.15 Humpback dolphins 0.14 0.15 0.22 0.24 0.98 1.09
A: 500 m x 500 m, by survey; B: 500 m x 500 m, by season; C: 1,500 m x 1,500 m, by season; Pre: pre 10% absence data truncation; Post: post 10% absence data truncation (these were the data that were modelled).
An additional problem identified during spatial data processing was the potential for
pseudo-absence data to exist within the dataset as a result of availability bias (primarily
due to turbidity), perception bias and disparity in observer effort received by each cell.
Variation in turbidity could be a major source of detection bias both spatially (i.e. across
the study area) and temporally (i.e. between survey periods and especially between those
occurring in different seasons). Areas of higher turbidity could mean that dolphins are less
available to be seen by the observer. Unfortunately, this bias could not be accounted for
because turbidity values were only available for the locations where dolphins were
present. As previously explained, the team leader recorded changes in turbidity and depth
for their side of the aircraft but this dataset was not informative enough to construct a
reliable turbidity spatial layer of information for the study area. Perception bias was
considered to be less of an issue because surveys were undertaken in relatively good
weather conditions. Glare and sea state, like turbidity, can also reduce dolphin
detectability but the surveys were conducted at times when these conditions were
suitable (Table 3). To reduce the chance of dolphins being ‘missed’ by observers, only
observers experienced in aerial surveys were used. The availability of dolphins to be seen
is likely to have also been affected by surfacing times, associated with behaviour but this
could not be accounted for. For instance, dolphins could spend greater time underwater in
areas where they are foraging.
Absence records with low survey effort were down-weighted within models in an attempt
to account for variation in the amount of survey effort received by each cell. Values to
represent survey effort received by each cell with an absence record were calculated.
During an aerial survey, it is always possible for observers, due to perception bias, to miss
a dolphin. Therefore, there was never 100% certainty that a dolphin was ‘truly’ absent
from a grid cell. Presence records were based on certainty, so they were assigned a
Chapter 2
51
weighting of one. Thus, the remaining absence data were weighted with values less than
one, but greater than zero, based on the amount of effort received. To calculate this for
each absence record, the amount of survey effort (in m2) received by each cell was divided
by the rounded amount of maximum survey effort (850,000 m2) received by any one cell.
This is a cumulative value that represents the spatial area of effort coverage over all
survey periods. The cells that received the highest effort were situated within the end of a
repeated block strip, which overlapped with the start of an Exmouth block strip, which
means that the area was seen by observers an additional time. This method was
undertaken ad-hoc with the purpose of applying a weighting scale that was relative and
based on survey effort alone.
2.6 Preliminary (presence-absence) models
Preliminary presence-absence models were built using Generalized Additive Models
(GAMs) and component-wise boosting (‘boosting’) methods. GAMs and boosting were run
in parallel, to test their performance in achieving good model fit to the training data. Both
are regression based, facilitating the relationships between response and predictor
variables to be investigated. GAMs are a commonly used SDM technique in cetacean
research (Cañadas & Hammond, 2006, 2008; Forney, 2000; Redfern et al., 2008), whereas
boosting has so far generally been limited to the use of regression trees (BRT) (Elith &
Leathwick, 2009). Detailed BRT methods are provided by Elith et al. (2008), and an
example of its use in a cetacean SDM is by Panigada et al. (2008). A GAM is a semi-
parameterised extension of the Generalised Linear Model (GLM), and is able to model
weighted data. Following the rationale of Guisan et al. (2002), GAMs were used instead of
GLMs because they have a smoothing function, which allows predictors to have a more
flexible fit and more closely resemble ecological relationships. Rather than a straight line
representing the sum effect that environmental covariates have on the response variable,
a GAM will fit a line that is a sum of the covariates’ smooth functions. This negates the
need to explore various polynomial term options and alternative transformations that is
central to the GLM process (Guisan et al., 2002).
Boosting was applied in an attempt to obtain reliable model results from small samples of
response data but with many predictor variables (Buehlmann & Hothorn, 2007; Schmid et
al., 2010). It entails ‘ensemble learning’, a type of ‘machine learning’ method, in which a set
of hypotheses is automatically constructed and computed from a dataset of response and
predictor variables (Zhou, 2009). The hypotheses are the many combinations of predictor
data available from the training dataset. These variable combinations are known as ‘base-
Chapter 2
52
learners’. The modeller specifies the degrees of freedom and the base-model algorithm
(e.g. ‘bols’ which is a linear fit). Boosting ‘boosts’ the power of weak base-learners (those
which have a relatively lower influence in the model algorithm) so that they are as
informative as the stronger ones (Zhou, 2009). It does so through regularisation,
simultaneously shrinking coefficient estimates and selecting relevant predictors (Schmid
et al., 2010). The disadvantages of boosting, and perhaps why it is as yet infrequently used,
are that it is computationally intensive and that expertise are required to prepare code
and extract results. It is expected that the use of boosting in SDMs will grow over time
(Elith & Leathwick, 2009).
2.6.1 GAMs
GAMs were implemented in the Mixed GAM Computation Vehicle (MGCV) R Package
(Wood, 2004). The model formulae were fit with a binomial probability distribution and a
logit link function, which is appropriate for binary presence-absence data. Truncated data
and weights were assigned to absence data based on survey effort, as previously
described. The default setting of ‘NULL’ was specified for subset, offset, knots, sp, min.sp,
H, paraPen and G arguments. Neither offsets or knots (degrees of freedom) were specified
because MGCV automates an integrated smoothness estimator (Wood, 2004). MGCV
allows the use of shrinkage splines so that smooth functions reduce to linear ones if there
is not support for smoothing. The integrated smoothness estimator was automated by
specifying ‘GCV’ as the method within the R script. Other arguments were left as the
default setting (e.g. smoothing parameters including sp and min.sp) in order to avoid
overfitting.
A total of 102 GAMs were run. At each of the three spatio-temporal resolutions, 11 models
were run for both ‘all dolphins’ and bottlenose dolphins, and 12 models were run for
humpback dolphins. The approach taken was to first identify whether any of the
environmental variables had a linear relationship with the response data, then test for
interactions between continuous and categorical environmental variables, and then refine
the models for each response variable as required.
Models M0 – M2 were run for each response variable. The first model (M0) included all
covariates, and a smoothing function for each continuous variable. As collinearity between
slope and topographical complexity had been identified during data exploration (Table 6),
each of these variables were removed in turn to ascertain which variable contributed
more information to the model. Model M1 included slope and excluded topographical
Chapter 2
53
complexity while M2 included topographical complexity and excluded slope. As
topographical complexity was found to perform marginally better, slope was not
incorporated into subsequent models (Tables 8; 9). Aspect was also dropped as it did not
appear to add much information to the model.
The effective degrees of freedom (EDF) provided by model summaries were used to assess
which covariates would be better described as linear functions. As described by Forney
(2000), one degree of freedom (DF) represents a linear response. Consequently,
topographical complexity was modelled as a linear function for all response variables. The
front variable was applied as a linear function in models for ‘all dolphins’ (M3, M13).
Depth, front, and distance from islands were applied as a linear function in humpback
dolphin models (M5, M15) (Table 8).
Interactions between each continuous environmental variable and survey period (a
categorical variable) were tested (M7 – M12) for each response variable (Table 8). GAMs
were limited to only one interaction because of the amount of data available. As there
appeared to be a strong interaction between SST and survey period, this interaction was
incorporated into later models (M13 – M15) (Table 8).
Model fit to the data were assessed using ‘deviance explained’. Deviance explained was
expressed as a percentage, with higher values indicating that the model did a better job of
fitting to the data (Table 9). This value is a generalisation of the coefficient of
determination (R2), and a surrogate for R2 when it is not available. Models poorly fit the
data, with each explaining less than 6% of deviance, and most explaining less than 1% of
deviance (Table 9). It is not clear whether reducing the proportion of absence records
improved model fit to data. GAMs for ‘all dolphins’ and bottlenose dolphins performed
marginally better at the coarsest resolution (M0 = 4.05; 5.97 at resolution C) (Table 9).
However, models for humpback dolphins performed best at the finest resolution (M0 =
4.23% DE at resolution A) (Table 9). As the low deviance explained values indicated that
all models had a poor fit to the data, GAMs were not progressed further to the model
selection phase, which would have involved extracting and comparing GCV scores.
Chapter 2
54
Table 8: GAM descriptions. Models M0-M2 and M7-M12 were run for all response variables. M13 was run for ‘all dolphins’, M14 for bottlenose dolphins and M15 for humpback dolphins.
GAM Description Notes
M0 y ~ aspF + surveyF + s(sst_bc) + s(edml_bc) + s(dep_bc) + s(fro_bc) + s(edis_bc) + s(toco_bc) +s(sl_bc)
Collinearity between toco and sl.
M1 y ~ aspF + surveyF + s(sst_bc) + s(edml_bc) + s(dep_bc) + s(fro_bc) + s(edis_bc) + s(sl_bc) Smoothers for all cont. variables; with sl, no toco.
M2 y ~ aspF + surveyF + s(sst_bc) + s(edml_bc) + s(dep_bc) + s(fro_bc) + s(edis_bc) + s(toco_bc) Smoothers for all cont. variables; with toco, no sl.
M3 y ~ aspF + surveyF + s(sst_bc) + s(edml_bc) + s(dep _bc) + fro_bc + s(edis_bc) + toco_bc Fro, edml, toco and dep as linear variables (run for ‘all dolphins’).
M4 y ~ aspF + surveyF + s(sst_bc) + edml_bc + dep_bc + fro_bc + s(edis_bc) + s(toco_bc)
M5 y ~ aspF + surveyF + s(sst_bc) + s(edml_bc) + s(dep_bc) + s(fro_bc) + s(edis_bc) + toco_bc Fro and sl are linear (run for bottlenose dolphins).
M6 y ~ aspF + surveyF + s(sst_bc) + s(edml_bc) + dep_bc + fro_bc + edis_bc + toco_bc Fro, edis, toco and dep are linear (run for humpback dolphins).
M7 y ~ s(sst_bc, by = surveyF) Interaction exists between sst and survey.
M8 y ~ s(fro_bc, by = surveyF)
M9 y ~ s(edis_bc, by = surveyF)
M10 y ~ s(edml_bc, by = surveyF)
M11 y ~ s(dep_bc, by = surveyF)
M12 y ~ s(toco_bc, by = surveyF)
M13 y ~ s(edml_bc) + s(dep_bc) + fro_bc + s(edis_bc) + toco_bc + s(sst_bc, by = surveyF) Fro, edml, toco and dep are linear, sst by survey (run for ‘all dolphins’.
M14 y ~ s(edml_bc) + s(dep_bc) + s(fro_bc) + s(edis_bc) + toco_bc + s(sst_bc, by = surveyF) Fro and sl are linear, sst by survey (run for bottlenose dolphins).
M15 y ~ s(edml_bc) + dep_bc + fro_bc + edis_bc + toco_bc +s(sst_bc, by = surveyF) Fro, edis, toco and dep are linear, sst by survey (run for humpback dolphins).
y: response (i.e. ‘all dolphins’, bottlenose dolphins, or humpback dolphins); s: smoother applied; by: interaction applied. Environmental variables: edml: distance from mainland; edis: distance from islands; dep: depth; sl: slope; asp: aspect; toco: topographical complexity; sst: sea surface temperature; fro: front; _bc: Box-Cox transformed; F: categorical variable specified as a factor.
Chapter 2
55
Table 9: Summary of GAMs generated for each response variable at each resolution. Values are ‘deviance explained’, as a percentage.
‘All dolphins’ Bottlenose dolphins Humpback dolphins
GAM A B C A B C A B C
M0 2.19 2.68 4.05 3.50 4.35 5.97 4.23 2.40 2.48
M1 1.90 2.66 3.90 3.42 4.25 5.81 3.97 2.28 2.25
M2 2.19 2.68 4.04 3.45 4.35 5.91 3.98 2.28 2.27
M3 1.92 2.67 4.01 - - - - - -
M4 - - - - - - - - -
M5 - - - 3.45 4.35 5.91 - - -
M6 - - - - - - 3.98 2.28 2.26
M7 0.92 1.27 1.69 1.69 2.45 3.31 2.44 0.89 1.15
M8 0.70 0.67 1.49 1.44 1.31 2.84 0.63 0.12 0.10
M9 0.19 0.27 0.05 0.29 0.14 0.23 1.63 0.48 0.89
M10 1.21 1.61 1.93 1.83 1.91 2.42 1.58 1.83 2.65
M11 0.57 0.66 0.98 1.02 0.99 1.29 0.57 0.45 1.01
M12 0.42 0.40 0.93 0.75 0.94 2.26 0.77 0.82 1.36
M13 2.18 2.57 3.85 - - - - - -
M14 - - - 3.52 3.87 5.74 - - -
M15 - - - - - - 3.95 1.90 2.05
A: 500 m x 500 m, by survey; B: 500 m x 500 m, by season; C: 1,500 m x 1,500 m, by season. M0 – M17: each GAM is described in Table 8.
2.6.2 Boosted models
Boosting with binomial error function and logit link function was implemented in the
Mboost R Package (Hothorn et al., 2013). Boosting utilised bols (linear), btree (decision
tree) and bbs (p-spline) (Schmid & Hothorn, 2008), base-learners so that the model of best
fit to the data could be found from a range of algorithms. The bspatial base-learner was
also used, which considers ‘space’ as a variable. All environmental variables were included
in the models so that the output could inform the later decisions on which variables to
include in the final models.
A total of 108 boosted models were constructed and run in sequence. Each response
variable at each resolution was modelled with both a bols and a btree baselearner, and
each with three degrees of freedom (df) options (6, 7 and 8). This culminated in 54 models
being run. A further six models (also for each response variable at each resolution) were
manually constructed with bbs base-learners. This was done separately so that bbs
baselearner df could be specified following Schmid and Hothorn (2008)’s general rule of
thumb. Essentially, for continuous environmental variables, df=4 for a single predictor,
Chapter 2
56
and df=7 for two predictors. For categorical environmental variables, degrees of freedom
were based on the number of categories within the variable. For instance, when
considering aspect, which consist of four categories, ‘north’, ‘east’, ‘south’ and ‘west’,
degrees of freedom were multiplied by four. This is a conservative approach in reducing
the ‘wiggliness’ of the line.
Cross-validation was used to identify the optimum stopping rate (m-stop) for each
response variable (Mayr et al., 2011). The m-stop was specified to prevent boosting
iterations from running until convergence, which is likely to result in overfit models (Mayr
et al., 2011; Schmid et al., 2010). Experimental models were run at various appropriate
shrinkage rates, to identify the point of convergence. There is a compromise between
shrinkage rates and computational efficiency. Shrinkage rates range between 0.01 and 0.1.
A lower rate provides higher optimisation of risk functions but can take a long time to
achieve. Final models were run at shrinkage rates of 0.05.
Cross-validated negative log-likelihood (risk), Weighted Brier and Pseudo R2 scores were
extracted and used to assess model fit to the data. Cross validation coefficients (0 to 1)
represent the risk value at the optimum m-stop, with a lower value indicating a lower
empirical risk of an overfit model. Weighted Brier scores (0 to 1) are used to assess
prediction reliability, with a lower score indicating higher reliability (Schmid et al., 2010).
Nagelkerke pseudo R2 scores were generated as an alternative to R2 scores. R2 scores
cannot be obtained from boosted models because they are a logistic, rather than an
ordinary least squares (OLS), method of regression. Pseudo R2 scores have a similar
meaning to R2 scores in that a model with a score closest to one explains greatest variance,
thus providing best fit to data. A negative pseudo R2 score indicates that the model
provides a worse explanation of the data than a random explanation would.
All models had a negative pseudo R2 score, indicating that they provided a worse fit to the
data than would have been fit at random (Table 10).
Chapter 2
57
Table 10: Summary of the ‘best’ boosted models for each response variable for each resolution.
Response Res DF Base-learner
m-stop w Brier CV CV SD P R2 (%)
‘All dolphins’ A 6 bbs 361 0.074 0.219 0.005 -0.90 B 6 bbs 283 0.074 0.218 0.003 -0.80 C 7 bbs 60 0.200 0.562 0.006 -3.00
Bottlenose A 6 bbs 570 0.049 0.151 0.007 -1.00 B 8 bbs 924 0.049 0.149 0.005 -1.10 C 8 bbs 196 0.154 0.444 0.008 -5.00
Humpback A 8 btree 94 0.008 0.044 0.003 -0.20 B 6 bols 339 0.008 0.043 0.005 -0.30 C 8 bbs 20 0.056 0.169 0.013 -0.40
Res: Resolution; A: 500 m x 500 m, by survey; B: 500 m x 500 m, by season; C: 1,500 m x 1,500 m, by season; DF: Degrees of freedom (6, 7, 8); m stop: optimum m stop; w Brier: weighted Brier score; CV: Cross-validation risk score; CV SD: Cross-validation risk score standard deviation; P R2
: Pseudo R2.
2.7 Summary and conclusion
The performance of a SDM is greatly influenced by the training data on which it is based
(Elith & Leathwick, 2009; Redfern et al., 2006). This chapter has described how dolphin
sightings and environmental data were acquired and processed, using GIS. It presents the
results of preliminary models that were run in order to understand the performance of the
training data that were generated at this stage of the modelling project.
There were inherent limitations in the use of the dolphin sightings data, due to the surveys
having been designed for dugongs. Many sightings were removed from the dataset due to
observer uncertainty, which reduced the sample size available for modelling. Uncertainty
also has meant that count data cannot be used in the models. As observers were focussed
on searching for dugongs, it is possible that dolphins were missed, which is a source of
pseudo-absence.
The systematic nature of the aerial survey design allowed effort data to be generated. The
presence-absence of ‘all dolphins’, bottlenose dolphins and humpback dolphins across the
study area was created, by populating a grid. Grid cells with uncertain sightings were
removed in order to reduce the problem of pseudo-absence data. Zero-inflation and low
encounter rates was recognised at this stage, so data were processed at three alternative
spatio-temporal resolutions in order to reduce zero-inflation of the data.
Chapter 2
58
Environmental variables were generated based on data availability, coverage and
reliability. In the absence of prey (i.e. fish distribution) data, proxy variables were sourced.
The lack of reliable habitat data for the extent of the study area was identified to be a
potential limitation for this project. SST, fronts, bathymetric features and distances from
mainland and islands were created at each of the three spatio-temporal resolutions.
GAM and boosting techniques were used to create preliminary models. This was done so
that the performance of the training data at different resolutions could be explored. To
increase sample prevalence, data were truncated and weighting was applied. Model fit to
the data were poor in all instances. It is not clear whether the poor fit was a result of
a) small sample of dolphin sightings across a large study area,
b) pseudo-absence data (caused by dolphins having been missed by observers),
c) weak environmental variables,
d) effects of weather conditions, particularly turbidity, on dolphin detectability, or
e) a combination of all or some of the factors above.
The number of certain or probable species sightings across the wide study area were
relatively low. This problem is common for data collected over a wide study area and often
causes difficulties in model fitting (Arab et al., 2012; Barbet‐Massin et al., 2012;
McPherson et al., 2004). This was likely exacerbated by the training data being created at a
fine scale across. However, the fine scale seemed to be appropriate for coastal dolphin
ecology, especially for humpback dolphins, as it allowed fine scale variation in distances
from coast.
The lack of environmental data can only be resolved through further data collection. This
could be done by undertaking in-situ habitat and fish distribution surveys, or through the
purchase of expensive satellite products (e.g. hyperspectral imagery) from which habitat
data could be interpreted.
Chapter 3 revisits the modelling process through the use Maximum Entropy (MaxEnt).
This is an alternative model approach that uses dolphin sightings and background data as
the response variable, rather than dolphin presence-absence data.
Chapter 3
59
Chapter 3: First insights into inshore dolphin distribution of the western Pilbara,
using MaxEnt
3.1 Introduction
Choosing the most appropriate model technique is an essential component of the SDM
process, and should consider model purpose and data characteristics (Araujo & Guisan,
2006; Elith & Leathwick, 2009; Guisan & Zimmermann, 2000; Redfern et al., 2006). The
previous chapter presented preliminary results of GAMs and boosted models that used
presence-absence data. Good fit to the training data could not be achieved, for a range of
possible reasons including unreliable absence data (see Chapter 2). This chapter revisits
the modelling process through the use of MaxEnt, an alternative technique that utilises
presence and ‘background’ data. The models in this chapter include additional dolphin
sightings data (October 2013 and May 2014) that became available after the preliminary
GAMs and boosted models were run.
MaxEnt is a machine learning technique that uses presence-only data to model the
probability of species presence, conditioned by environmental constraints (Elith et al.,
2011; Phillips et al., 2004). It is based on the maximum-entropy principle which is “the
best approximation of an unknown probability distribution is to ensure that the
approximation satisfies any constraints on the unknown distribution that we are aware of,
and that subject to those constraints, the distribution should have maximum entropy”
(Jaynes, 1957 in Phillips et al., 2006). In a SDM context, MaxEnt’s goal is to find the least
complicated description of a species’ distribution. It works by generating ‘features’, which
are combinations of environmental variables with alternative transformations and model
terms (Elith et al., 2006; Phillips et al., 2004). The modeller can either set a feature type
(i.e. linear, quadratic, product, threshold or discrete) or use the default setting which
allows the software to find the most appropriate feature type, or combination of feature
types, for the data. These settings are fine-tuned so that they work well for a variety of
datasets (Phillips & Dudík, 2008). General empirically derived rules have been
programmed into the software so that the ‘auto features’ setting identifies a feature type
that is appropriate for the number of presence records. It will not use a feature type that is
too complex. As with other forms of machine learning, including boosting (see Chapter 2),
regularisation is used to prevent models from being overfit by adding a complexity penalty
to the loss function. The automatic regularisation setting in MaxEnt versions, has been
fine-tuned to suit the number of feature types and samples locations (Phillips et al., 2006).
MaxEnt generates background data in place of absence data, sampling environmental
Chapter 3
60
variables from points throughout the landscape. Through a process of regularisation and
optimisation, a series of model iterations are run which achieves higher gain. Similarly to
the GAM ‘deviance’ goodness-of-fit measure, higher gain represents better model fit to the
data (Phillips et al., 2004). During this process, MaxEnt generates a probability
distribution and populates this information into cells of a geographic grid, based on the
sampled background data (Elith et al., 2006; Phillips et al., 2004). Models are evaluated
through the comparison of features at sample and background points.
MaxEnt performs well even when sample sizes are small (Elith et al., 2006; Phillips et al.,
2004). Elith et al. (2006) used presence data for 226 species from six regions across the
world to compare 16 model techniques and found that MaxEnt consistently outperformed
more traditional methods including GAMs. They evaluated model performance using
Receiver Operating Characteristic (ROC), area under the curve (AUC), area under the
correlation (COR) and Kappa statistics and further examined geospatial relationships
using a series of geospatial metrics. MaxEnt has been used for a variety of studies where
an understanding of habitat preference is of high importance to marine mammal species
conservation. For instance, the software has been used to predict spinner dolphin resting
habitat around Hawaii Island (Thorne, 2012); identify humpback whale breeding areas at
the Great Barrier Reef (Smith et al., 2012); and determine areas of niche partitioning
between krill predators in the Western Antarctic Peninsula (Friedlaender et al., 2011).
MaxEnt models have also informed spatial risk assessments, including for harbour
porpoises (Phocoena phocoena) in coastal Danish waters subject to development (Edrén et
al., 2010), and dugongs at risk to bycatch in Sabah (Briscoe et al., 2014).
It is often suggested that presence-absence models are more reliable than those that only
use presence data. However, this is only true if the absence data is reliable. In this study,
care was taken during spatial data processing to avoid introducing pseudo-absence data
(see Chapter 2), but some uncertainty remains due to dolphins being highly mobile and
often difficult to detect. The background data generated by MaxEnt from random points
throughout the study area are not treated as absence data typically would be, which
avoids the problem of introducing pseudo-absences. A key difference between MaxEnt and
presence-absence techniques is that MaxEnt does not attempt to predict a relative index of
environmental suitability based on species occurrence. MaxEnt instead models a
probability distribution across the study area based on favourable environmental
conditions (Phillips et al., 2006; Wiley et al., 2003). Accordingly, the maximum achievable
AUC for a MaxEnt model will always be less than one, since the occurrence of a species can
Chapter 3
61
never be known without accurate absence data (Elith & Leathwick, 2009; Phillips et al.,
2006; Wiley et al., 2003). A major potential pitfall of using a model based only on presence
data is that the presence data could have been collected in an uneven fashion; with effort
biased to areas that are easily accessible (Phillips et al., 2006). Such bias is not an issue for
this project as the systematic transect design ensured even sampling coverage across the
study area, with the repeated sampling effort in the repeated block to be accounted for.
Phillips et al. (2006) state that too few presence data points across a study area, and a lack
of descriptive environmental variables could also reduce the reliability of models built
upon only presence data, but these problems are also pertinent for presence-absence
models.
This chapter presents the first insights into western Pilbara inshore dolphin distribution,
as obtained through the development of MaxEnt models for bottlenose and humpback
dolphins. The research objectives were to:
Identify areas of highest habitat suitability for bottlenose and humpback dolphins.
Determine whether there is difference in habitat preference between bottlenose
and humpback dolphins.
Understand which environmental variables are most important in defining the
distribution of bottlenose and humpback dolphins.
Describe the influence that important environmental variables have in defining the
distribution of bottlenose and humpback dolphins.
The ecological findings in relation to the research objectives are presented, and
recommendations for future survey planning are provided.
3.2 Methods
3.2.1 Spatial and temporal extent of the data
The 10,600 km2 study area is situated in the western Pilbara approximately 1,200 km
north of Perth, WA (see Chapter 2). The area encapsulates the wide protected embayment
of Exmouth Gulf and coastal waters to the Lowendall Islands (Figure 1), to the 20 m
isobath. The central third portion of the study area (i.e. ‘repeated block’) was repeated
during each survey period (Figure 4). Five dugong aerial surveys were conducted over a
thirteen-month period, approximately three months apart (May, July, October and
December 2012 and May 2013), to help account for seasonal variation. Two follow-up
Chapter 3
62
surveys were conducted in October 2013 and May 2014. Data were pooled into the
following datasets so that the influence of temporally dynamic variables (i.e. SST and
front) could be explored:
Static (all data, no consideration of seasonality),
Autumn (May 2012, 2013 and 2014), and
Spring (October 2012 and 2013).
A hierarchy of models at different temporal periods (i.e. all surveys combined, surveys
combined by season and individual surveys) was used to determine whether temporal
patterns presented by output could be reliable. All model outputs were visually inspected
to determine whether patterns were consistent across or between the varying temporal
seasons.
3.2.2 Spatial data processing for MaxEnt models
A detailed description of the spatial data processing using GIS was provided in Chapter 2.
Following the exploration of preliminary MaxEnt models, two new environmental
variables were created for use in subsequent models. These variables were a) distance
from the slope at the 20 m isobath, and b) distance from intertidal areas, using the
Euclidean Distance tool. The slope at the 20 m isobath was hand digitised from
bathymetry data and the intertidal areas identified from Geoscience Australia’s (2004)
‘GEODATA Coast 100k’ dataset.
A raster layer representing survey effort (i.e. ‘bias’ in MaxEnt terminology) was created
following MaxEnt specifications (Phillips, 2005; Phillips et al., 2009). Cells within the
repeated block were appointed a value of ‘2’ and all other cells were appointed a value
of ‘1’ (Figure 19). This is representative of the repeated survey block having received
twice the amount of survey effort than the rest of the study area. MaxEnt incorporates this
information about sample bias in order to “vastly improve model performance and
reliability”, as it ensures that the assumption of unbiased samples is upheld (Phillips et al.,
2009). So that data were readable by MaxEnt, each raster was reclassified to be the same
resolution, extracted to be the same extent and exported in American Standard Code for
Information Interchange (ASCII) format.
Chapter 3
63
Figure 19: Survey effort, represented as a bias layer for MaxEnt.
3.2.3 Running the MaxEnt models
Twenty MaxEnt models were run. For both bottlenose and humpback dolphins, a model
was generated for each individual survey period, for all survey periods together (i.e. the
‘static’ model), for Autumn (May 2012, 2013 and 2014), and for Spring (October 2012 and
2013).
Initially, MaxEnt models were run with the same environmental variables used in the
preliminary MaxEnt models (see Chapter 2). From those, it appeared that distance from
the mainland could have had an effect on dolphin distribution (particularly bottlenose
dolphins), as a proxy for distance to the slope at the 20 m isobath. Also, the influence of
distance from the mainland and distance from islands seemed to have a similar effect on
the probability of dolphin occurrence. Having these as separate variables caused the
output map of suitable habitat areas to be ambiguous. This initiated the creation of the two
Chapter 3
64
new environmental variables, ‘distance to the slope at the 20 m contour’, and ‘distance
from intertidal areas’.
The final models were based on SST, fronts, depth, slope, distance to the slope at the 20 m
contour and distance from intertidal areas. Slope was used instead of topographical
complexity because it was deemed to be more reliable, since it was the raw variable from
which topographical complexity was derived. Aspect was excluded because initial results
were confused when using aspect and slope together, even though collinearity was not
identified (Table 11). The slope at the 20 m contour is perhaps the most notable
bathymetric feature in the relatively flat study area so could be of ecological significance.
Distance to intertidal areas data amalgamated distance from mainland and islands, and
also included a large intertidal area that is present between Barrow Island and the
mainland. Spearman’s rank correlation coefficients between pre-existing and new static
variables were generated to confirm that the new variables reduced the potential for
collinearity (Table C1, Appendix C).
Default feature and regularisation settings were selected, as they are known to produce
reliable results. The default settings have been finely tuned so that they perform well for a
range of datasets and ensure that a model will not be overly complex for the size of the
dataset that is available (Phillips & Dudík, 2008). The prevalence setting was set at the
default value of 0.5 (Phillips et al., 2006). In this context, prevalence refers to the species’
natural occurrence and is the proportion of sites occupied by the species across the
landscape (Elith et al., 2011). The prevalence value of 0.5 represents a 50% chance that
the species may occur in any given grid cell. A maximum of 10,000 background data points
were randomly generated. The bias layer was incorporated into the model which specified
areas of greater survey effort (Phillips et al., 2006; Phillips et al., 2009). The importance of
environmental variables in informing the models was measured with a permutation test.
MaxEnt measures the contribution of variables by the resulting decrease in training AUC
for each permutation, and then normalises the resulting values to give percentages that
are easily interpretable. Response curves were created for each variable in order to
understand the probability of suitable conditions for dolphins to be present. Raster data
indicating mean relative habitat suitability across the study area were generated in logistic
output. Probability of presence estimates ranging from 0 to 1 were allocated to each
500 m x 500 m grid cell. A probability value of 0.5 indicates that there is a 50% chance of
species occurrence at that location (Elith et al., 2011). Rasters were imported into ArcMap
v 10.1 (ESRI, 2012) and maps showing habitat suitability and sighting locations were
Chapter 3
65
created. A colour ramp of red to blue was used to indicate the predicted probability of
suitable conditions for dolphin presence, with red representing high probability and light
blue representing low probability. MaxEnt generates maps that are a generalised
prediction and broader than the species’ realised niche, which is the area that it truly
occupied and is related to the species’ role in the ecosystem (Phillips et al., 2006; Phillips
et al., 2004).
3.2.4 Evaluating the MaxEnt models
Each model was run over a maximum of 500 iterations (Table 12) for 1,000 bootstrapped
replicates. Bootstrapping is a method for evaluating the variance of an estimator by
randomly resampling training data at presence locations, with replacement. Each replicate
used 75% of dolphin sightings at random to train models and the remaining 25% of
sightings to test models. Each final model consisted of a mean output of the 1,000
replicates. Diagnostic plots for omission (rate of underprediction, or potential for false
negatives) and commission (rate of overprediction, or potential for false positives) were
generated. A good model should have a low omission rate, which is indicated by the mean
omission on training data falling below the predicted omission (Phillips et al., 2006). Test
data that falls well below the predicted omission for individual models may suggest that
training and test data are spatially auto-correlated (Friedlaender et al., 2011; Phillips,
2005). The second diagnostic plot provided the ROC curve and AUC metric for both
training and test data, which were used to assess model performance in predicting areas
of high habitat suitability, compared to a random selection of points. The training data line
(in red) and AUC show indicate how well the model fits the training data. The test data line
(in blue) and AUC indicate how reliable the model would be in its predictive power.
3.3 Results
Twenty MaxEnt models were built and run on 1,000 bootstrapped replicates. The number
of training and test samples varied between models, for both species, and training samples
were fewer for humpback dolphins than for bottlenose dolphins (Table 11).
All models had omission rates that were both close to and less than the predicted omission
rate (Table D1, Appendix D), which indicates low chances of omission error and spatial
autocorrelation. All models achieved fit to the training data, but only four models (two for
each species) had acceptable test performance, following Hosmer and Lemeshow (1989)
(Table 11). For bottlenose dolphins, spring and July 2012 models had excellent fit to
Chapter 3
66
training data (AUC = 0.81; 0.82), and acceptable test performance (AUC = 0.71; 0.72). The
static model had poor fit to training data (AUC = 0.68), and the remaining models had
acceptable fit to training data (AUC = 0.72 to 0.78). For humpback dolphins, all May (2012,
2013 and 2014), October 2013, July 2012 and December 2012 models had excellent fit to
training data (AUC = 0.81 to 0.87). Of these, July 2012 and December 2012 achieved
acceptable test performance (AUC = 0.86 for each). However, the test data set was small
for both periods, particularly December 2012 (n = 2) (Table 11). The remaining models
had an acceptable fit to data (AUC = 0.73 to 0.80) but poor test performance (AUC = 0.58 to
0.70).
Chapter 3
67
Table 11: Summary of MaxEnt model performance. Discrimination of model performance is according to Hosmer and Lemeshow (1989).
Time period
Training samples
Iterations Training AUC
Model performance
Test samples
Test AUC
Test performance
Variable importance (%)
dep sl ed20m edin sst fro
Bottlenose dolphins Static 363 484 0.68 Poor 120 0.63 Poor 1.7 10.8 48.4 29.7 N/A N/A
Autumn 138 500 0.78 Acceptable 45 0.66 Poor 13.9 9.6 20.2 18.2 20.0 18.0
Spring 130 500 0.81 Excellent 43 0.71 Acceptable 9.5 10.4 20.0 21.2 14.1 24.9
May-12 30 367 0.78 Acceptable 10 0.64 Poor 6.8 6.3 9.7 29.8 38.6 8.8
May-13 68 398 0.72 Acceptable 22 0.63 Poor 19.3 9.6 14.7 14.4 14.6 27.5
May-14 57 480 0.80 Acceptable 14 0.68 Poor 7.9 5.0 11.7 34.7 31.0 9.7
Oct-12 66 454 0.77 Acceptable 22 0.69 Poor 10.9 17.0 23.2 17.3 6.6 25.1
Oct-13 64 440 0.75 Acceptable 21 0.66 Poor 9.0 5.6 29.9 19.2 22.8 13.5
Jul-12 47 440 0.82 Excellent 15 0.72 Acceptable 30.2 7.9 18.4 13.8 13.6 16.2
Dec-12 60 447 0.74 Acceptable 20 0.63 Poor 18.8 17.1 31.3 14.5 6.0 12.3
Humpback dolphins Static 88 413 0.75 Acceptable 29 0.63 Poor 14.3 25.1 33.2 27.5 N/A N/A
Autumn 36 385 0.79 Acceptable 11 0.65 Poor 14.9 7.8 25.7 17.7 23.7 10.3
Spring 30 390 0.80 Acceptable 9 0.66 Poor 13.6 8.1 36.1 12.9 19.4 9.8
May-12 7 72 0.81 Excellent 2 0.70 Poor 11.3 4.1 48.3 17.2 7.7 10.8
May-13 16 362 0.84 Excellent 5 0.65 Poor 11.1 12.0 13.4 21.3 24.6 17.5
May-14 17 377 0.87 Excellent 4 0.62 Poor 13.4 28.4 23.7 5.2 8.8 20.4
Oct-12 11 84 0.73 Acceptable 3 0.58 Poor 7.1 7.8 5.9 17.7 33.6 25.9
Oct-13 19 337 0.83 Excellent 6 0.67 Poor 6.9 12.5 36.0 17.0 20.8 6.9
Jul-12 18 354 0.86 Excellent 6 0.72 Acceptable 18.1 26.3 11.4 8.0 27.0 9.1
Dec-12 7 81 0.86 Excellent 2 0.77 Acceptable 6.6 4.0 3.4 16.4 65.3 4.2 AUC: area under the curve; dep: depth; edin: Euclidean distance from intertidal areas; ed20m: Euclidean distance from the slope at the 20 m isobath; fro: front; sl: slope; sst: sea surface temperature. Discrimination of model performance: <0.5 = none; 0.5 to 0.7 = poor; 0.7 to 0.8 = acceptable; 0.8 to 0.9 = excellent; >0.9 = outstanding.
Chapter 3
68
3.3.1 Bottlenose dolphin distribution models
The percentage contribution of environmental variables towards explaining habitat
suitability was inconsistent across models (Table 11). For bottlenose dolphins, when
dynamic variables (i.e. SST and front) were omitted from the model, the variable with
highest contribution was distance from the slope at the 20 m isobath (48.4 %
contribution), followed by distance from intertidal areas (29.7 % contribution). The
response curves for these variables suggested that habitat suitability would increase both
immediately at the slope and at greater distances away from the slope (Figure 20A). The
response curve for distance from intertidal areas was uninformative, with the mean value
of 0.6 or less (Figure 20B). Of the models with acceptable test data performance, distance
from the slope at the 20 m isobath, SST and front had nearly equal contribution (20.0 %,
21.2 % and 24.9 % contribution, respectively) in Spring.
The response curves showed that probability of bottlenose dolphin presence is likely to
increase close to the slope at the 20 m isobath (Figure 20C), in cooler SST (Figure 20D)
and in areas where there is variation in water temperature (i.e. fronts) (Figure 20E).
However, the mean value of the response curves was 0.65 or less. Depth had the highest
contribution in July 2012 (30.2%) (Table 11), but the response curve was uninformative,
with a mean value of 0.4 or less.
At a broad scale, relative habitat suitability maps for bottlenose dolphins generally
showed that environmental conditions were more favourable further away from the coast
(Figures 21A to 21J). The static map suggests that areas of highest habitat suitability are
within Exmouth Gulf, near the slope at the 20 m isobath and north-east of Barrow Island
(Figure 21A), but this model obtained low AUC score for the test data.
Areas of highest habitat suitability varied between seasons, with the area north-west and
west of the North West Cape being of high habitat suitability in Spring, but not Autumn
(Figures 21B and C). Variation in habitat suitability was depicted between the months of
Autumn (i.e. May 2012, 2013 and 2014) (Figures 21D; 21E; 21F) while the months of
Spring (i.e. October 2012 and 2013) were more consistent (Figures 21G and H).
Chapter 3
69
Static
A) ed20m B) edin
Spring
C) ed20m D) sst E) fro
July 2012
F) dep
ed20m: Euclidean distance from the slope at the 20 m isobaths; edin: Euclidean distance from intertidal areas; sst: sea surface temperature; dep: depth; fro: front.
Figure 20: The probability of bottlenose dolphin presence (logistic output) as a
response to the variables with highest contribution for static (A and B), Spring (C, D and E), and July 2012 (F) MaxEnt models. The red line is the mean response of all
bootstrapped replicates, while the blue banding represents ±1SD.
( )
Chapter 3
70
Figure 21: Mean relative habitat suitability for bottlenose dolphins based on A) all sightings from all surveys and without consideration of SST or front; B, C) ‘seasons’,
and D-J) each survey period from 1000 bootstrapped iterations.
Chapter 3
71
3.3.2 Humpback dolphin distribution models
The percentage contribution of environmental variables towards explaining habitat
suitability was largely inconsistent across models for humpback dolphins, as was the case
for bottlenose dolphins (Table 11). Also similar to the bottlenose dolphins, the variable
with highest contribution to the static model was distance from the slope at the 20 m
isobath (33.2 % contribution), followed by distance from intertidal areas (27.5 %
contribution). The response curves for these variables suggested that habitat suitability
was highest further away from the slope (Figure 21A), and nearer to intertidal areas
(Figure 21B), which is in contrast to the results for bottlenose dolphins. However, values
were weak for the distance to intertidal area response curve (<0.6). Of the models with
acceptable test data performance, SST and slope had the greatest contribution in July 2012
(27.0 %; 26.3 % contribution) while SST had by far the greatest contribution that any
other variable in December 2012 (65.3 %) (Table 11). In both July and December 2012,
the SST response curve for showed that probability of humpback dolphin presence
increased as temperatures increased (Figures 22 C and E). The July 2012 response was
weaker than the December 2012 response, the mean value being 0.6 or less (Figure 22C).
The response curve for slope was weak with a small increase for probability for humpback
dolphin presence in areas with steepest slope, but the value was little more than 0.6
(Figure 22D).
The static habitat suitability map suggested that highest habitat suitability for humpback
dolphins was in waters close to intertidal areas including around islands or certain areas
on the mainland. (Figure 23A). Particular locations of note include the North West Cape,
Muiron Islands, north-eastern portion of Exmouth Gulf, Barrow Island and the intertidal
area between Barrow Island and the mainland. These patterns are consistent between the
seasonal models for Autumn and Spring (Figure 23B and C). However, these patterns were
inconsistent within seasons, with May 2012 (Figure 23D) depicting the outer edge of the
study area to be of highest habitat suitability while May 2013 and May 2014 looking more
similar to each other and the Autumn model (Figure 23 E, F and B). The maps for October
2012 and 2013 are different from each other (Figure 23 G and H). The Spring map (Figure
23 C), which incorporates data from both of those study periods, predicts habitat
suitability to be highest at the areas shown by both October models. Southern Exmouth
Gulf was depicted to be of low relative habitat importance for all models, apart from July
2012 where it was of high importance.
Chapter 3
72
Static
A) ed20m B) edin
July 2012
C) sst D) sl
December 2012
E) sst
ed20m: Euclidean distance from the slope at the 20 m isobath; edin: Euclidean distance from intertidal areas; sst: sea surface temperature; sl: slope.
Figure 22: The probability of humpback dolphin presence (logistic output) as a
response to the variables with highest contribution for static (A and B), July 2012 (C and D), and December 2012 (E and F) MaxEnt models. The red line is the mean
response of all bootstrapped replicates, while the blue banding represents ±1 SD.
Chapter 3
73
Figure 23: Mean relative habitat suitability for humpback dolphins based on A) all sightings from all surveys and without consideration of SST or front; B, C) ‘seasons’,
and D-J) each survey period from 1000 bootstrapped iterations.
Chapter 3
74
3.4 Discussion
3.4.1 Model performance
MaxEnt was largely successful in building models that fit the training data, but they were
not found to be reliable through the evaluation process. Their predictive power was
generally poor, as indicated by the low AUC test scores. The reliability of humpback
dolphin models that did achieve good fit to test data (July 2012 and December 2012), were
questionable because the test datasets were very small, which could suggest that they
were overfit to the data.
Models with good fit to training data but poor fit to test data can be used to describe
dolphin distribution patterns at the time that they were observed. Poor predictive power
was also reflected in the inconsistencies between areas depicted as high habitat suitability
across models. The patterns could be purely ecological, or they could reflect random
sampling variation or uneven detectability across the study area. They cannot, however,
be relied upon to predict patterns of distribution, or areas of habitat importance, outside
of that period.
These are the first distribution models that have been developed for inshore dolphins of
northern WA. As such limited a. priori information was available. As described by Redfern
et al. (2006), the amount of a. priori available at the beginning of a modelling project will
often dictate the capability of a model in being able to describe, predict or test hypotheses
(Figure 6). However, a model’s predictive power is also dependant on the quality of
species and environmental data on which it is built (Elith & Leathwick, 2009).
Figure 24: Model capabilities vary from describing species distribution to predicting and testing hypotheses about a species distribution is largely dependant on the
amount of a. priori knowledge that is available (Redfern et al., 2006).
Chapter 3
75
There are several reasons as to why the majority of MaxEnt models were unable to
achieve acceptable fit to the test data. Models incorporated information to account for
greater survey effort so that bias was not a problem. MaxEnt creates background data and
treats that in a way that is different to how presence-absence techniques would treat
absence data, in order to avoid problems of pseudo-absence data. The models might have
been influenced by small sample size of dolphin sightings relative to the study area, a lack
of informative environmental predictors, or the inability to account for factors that affect
dolphin detectability including turbidity. A combination of these factors may have been
affected model reliability.
Uncertainty in dolphin sightings, due to the aerial surveys being designed for dugongs and
not dolphins, led to data being removed from the dataset, which reduced the sample size
available for modelling. Other reasons for a small sample size are that
a) dolphins are highly mobile and, by chance, they may be outside of the observers’
view from the transect line;
b) dolphins may be near the transect line but not available to be recorded by
observers because they are deep underwater or in turbid waters;
c) inshore dolphins, particularly humpback dolphins, are in relatively low population
sizes which means they are not commonly encountered across the region; or
d) a combination of some or all of the above.
MaxEnt can fit models be successful in fitting models to small datasets (Phillips et al.,
2006), but only to a certain point. According to Wisz et al. (2008), decreasing sample sizes
reduce model accuracy and increase variability within models (Wisz et al., 2008). This is
because there is less information available about where species are found. This issue was
encountered during this project, with inconsistencies in the importance of environmental
variables between models built on seasonal and individual survey data.
The large study area exacerbates the problem of a small sample size as regional scale
models also have the potential for inaccuracies inherent to their dimension. Jiménez-
Valverde et al. (2008) warn against using regional scale model output as a direct estimate
of species distribution, particularly because important fine scale information can be
overlooked. Limited inferences can be made from a sparse dataset as sample sizes can be
insufficient to describe the full range of conditions that shape habitat suitability (Wisz et
al., 2008). Despite this, MaxEnt can still obtain high AUC values if background data are
selected from a broad area (Edrén et al., 2010)
Chapter 3
76
Non-detection of animals that are actually present (but missed) during a survey is often a
concern for SDMs (Gu & Swihart, 2004a). As MaxEnt models are based on the pattern of
presence records available, non-detection of an animal at a given time can lead to false-
absences being presented in the habitat suitability maps. These models could have been
affected by reduced detectability of dolphins in near shore waters. Near-shore waters of
the study area are brownish-red due to high levels of turbidity from sediment and organic
matter input from the Ashburton River Delta. In contrast, oceanic waters further offshore
around the islands are clear, which could result in a detectability bias (Viddi et al., 2010).
When interpreting the predictive maps it was therefore important to remember that
dolphins may have been ‘missed’ in the near shore turbid waters, and that no corrections
for availability had been applied. The mobile nature of dolphins also means that they
might not have been in the area being surveyed at the time the survey was underway. Such
temporary absence is an intrinsic and often encountered difficulty when modelling the
distribution of vagile species (Elith et al., 2011; Martin et al., 2005). This problem is likely
to be exacerbated for humpback dolphins, which occur in lower numbers and would
therefore have less chance of being encountered.
The models would be improved with environmental data that is more informative. The
response curves were often weak, with low values, or had complex patterns that are likely
to be spurious. There were inconsistencies in the importance of environmental variables
between models built on seasonal and individual survey data. It is unclear as to why the
static model for bottlenose dolphins had only a poor fit to training data. It could have been
that the dynamic variables (i.e. SST and front) were needed to explain the distribution of
bottlenose dolphins. However, when considering the seasonal models (i.e. autumn and
spring), the dynamic variables were of equal or lesser importance than the other variables.
Typically, depth and slope contributed little to the models. It is likely that, being
bathymetric derivatives within a relatively flat study area, they are weak predictors of
dolphin distribution in the region.
3.4.2 Ecological findings
Species’ spatial ecology is complex and occurs at a hierarchy of scales (Gutzwiller, 2002).
Broad scale models, such as these presented for inshore dolphins of the western Pilbara,
are useful in explaining broad patterns in distribution but cannot inform on local
processes. Further, the capability for the model to explain fundamental or realised niche
must be understood. Fundamental niche is the area where a species could hypothetically
be present, based on the environmental conditions that exist there, whereas realised niche
Chapter 3
77
is where the species actually occurs (Jiménez‐Valverde et al., 2008). MaxEnt output tends
to be somewhere between the potential and realised niche (Jiménez‐Valverde et al.,
2008). It is important to bear these concepts in mind when developing and interpreting
models, or else model output may be misconstrued. Thus, models developed for inshore
dolphins of the western Pilbara provide a broad scale understanding of where these
dolphins may occur.
The map outputs showed both similarities and differences between habitat suitability for
bottlenose and humpback dolphins. There was some overlap in suitable habitat between
species, suggesting that they are sympatric. This was expected to be the case around the
North West Cape, where a high proportion (~25%) of mixed bottlenose-humpback
dolphin groups were encountered by Brown et al. (2012). Sympatric top-order predator
species are able to co-exist within the same area because they have adapted to utilise
resources in a different way, thereby reducing interspecific competition (Parra, 2006).
Ecological niche segregation is enabled by differences in diet composition and timeframes
in which resources are being targeted (Kiszka et al., 2011). At a broad scale, bottlenose
dolphin distribution appeared to be linked to the slope at the 20m contour while
humpback dolphin distribution was more often near intertidal areas. Similar spatial
separation between the two species exists elsewhere (Palmer et al., 2014b; Stensland et
al., 2006). Both species are ‘opportunistic-generalist feeders’, consuming a wide range of
benthic and pelagic fish species, but bottlenose dolphins target cephalopods whereas
humpback dolphins usually do not (Amir et al., 2005; Heithaus & Dill, 2002; Parra &
Jedensjö, 2013). Localised surveys and distribution models would reveal finer scale
patterns in resource partitioning between the two species (see section 3.4.3).
Relatively high habitat suitability for bottlenose dolphins was depicted at the slope at the
20 m isobath and in waters around the Muiron Islands. Slopes are often important
foraging areas because prey are usually in higher abundance (Hastie et al., 2004). The
slope here is a prominent feature, dropping down to 100 m and extending for
approximately 3,000 km from the North West Shelf to the Arafura Sea (Wilson, 2013). It
could act as a navigational cue for large migrating finfish that dolphins predate upon.
Cephalopods, which are also target prey of bottlenose dolphins, are in high abundance
along this slope and at the Muiron Islands (Amir et al., 2005; Heithaus & Dill, 2002;
Jackson et al., 2008). Nutrient rich waters from offshore are channelled into Exmouth Gulf
between the North West Cape and the Muiron islands, which may facilitate the
accumulation of dolphin prey (Jackson et al., 2008). The potentially high occurrence of
bottlenose dolphins at the 20 m isobath location could also be a convergence between the
Chapter 3
78
Indo-Pacific and common bottlenose dolphin species (T. truncatus) that are found further
offshore (Allen et al., In review.).
Areas of relatively high humpback dolphin habitat suitability were depicted around
islands (particularly Barrow Island), the north-eastern portion of Exmouth Gulf and the
large intertidal area between Barrow Island and the mainland. This area is dominated by
coral reef and macroalgae, and is situated within the Barrow Island Marine Management
Area (DEC, 2007; Wilson, 2013). The association with islands would explain why
humpback dolphins were recorded more than 50 km from the coastline during the aerial
surveys. This is in line with Corkeron et al. (1997) who notes that while humpback dolphin
are usually found in shallow near shore waters close to coast, they have also been
recorded in deeper waters, but in association with islands or reefs.
Southern Exmouth Gulf was of relatively high habitat importance for humpback dolphins
only in July 2012, and of low habitat importance at all other times. This inconsistency may
reflect random sampling variation or be a result of seasonal turbidity. Southern Exmouth
Gulf receives terrestrial run off from small creeks and a wide intertidal area. It supports
mangrove communities and is highly productive area – habitat features that are important
for humpback dolphins elsewhere in Australia. Data collected from dedicated dolphin
aerial surveys would be more reliable and produce more consistent model output (see
section 3.4.3), which is needed if the conservation value of Exmouth Gulf for inshore
dolphins is to be understood. Localised boat-based surveys would be useful in (see
section 3.4.3).
3.4.3 Recommendations for further research
A key purpose of this modelling project was to inform future survey planning across the
expansive and remote region. The ability to target areas known to support higher
occurrences of dolphins would have had several benefits of informed survey planning (see
Chapter 1). Unfortunately, the low predictive power of the MaxEnt models has made it
difficult to pinpoint areas of high dolphin occurrence with much certainty. What this does
highlight is the need for dedicated dolphin aerial surveys to be conducted over the study
area. As described in Chapter 2, the ability to detect dolphins and record dolphins to
species level with certainty was hindered by the fact that dolphins were recorded as a
second priority to dugongs. Also, the lack of informative environmental variables
highlights the general lack of baseline information available for the northern WA coastline,
which is a gap that needs to be filled with some urgency if useful distribution models are
to be developed.
Chapter 3
79
Recommendations for further research are provided in Table 12. Adequate species
conservation needs to incorporate various ecological processes that occur at different
spatial and temporal scales. It is essential that future dolphin studies in the region are
designed and implemented correctly. They should follow best practice techniques in order
to maximise applicability of the data and meet assumptions of analysis. It is essential that
clear objectives are established and that the studies are planned accordingly (Pollock et al.,
2002). Select requisite design features and references for further guidance and
benchmarks studies are presented in Table 12. Also, the rigour of these surveys should be
open to peer review (Allen et al., 2012; Bejder et al., 2012).
Dedicated inshore dolphin aerial surveys
The number one priority would be to undertake aerial surveys across the region that
specifically target inshore dolphins. This would obtain regional abundance and density for
inshore dolphin species and enable new SDMs to be developed with a greater sample size
of species data. Targeted dolphin aerial surveys should be designed so that estimates of g
(0) (perception and availability bias) can be incorporated into analysis for absolute
estimates of abundance. Long-term, targeted dolphin aerial surveys could identify intra-
seasonal and interannual variations in distribution. Tropical lows and cyclones are a major
feature of the region, occurring between December and March. Tropical lows redistribute
prey at local scales, whereas cyclones can cause regional-scale and long lasting habitat
alteration (McKinnon et al., 2003). A major cyclone in the study area has previously caused
regional displacement of dugongs, through the removal of seagrass habitat (Gales et al.,
2004). Also, post cyclonic phytoplankton blooms that bolster prey abundance can occur
(McKinnon et al., 2003).
Targeted dolphin aerial surveys are needed for the western Pilbara and the wider
northern WA coastline. The purpose of this is to obtain relative abundance and densities,
validate the model results and obtain data for unsurveyed areas. Relative abundance and
density estimates over broad spatial and temporal scales will support monitoring
(Hammond et al., 2013). Monitoring trends over time is as important aspect of mitigating
potential impacts of coastal development (Jefferson et al., 2009). Obtaining a broader scale
dataset will also help in understanding the metapopulation structure. Surveys over the
western Pilbara can be stratified, based on the results from this survey. A power analysis
should be undertaken, with the results used to inform survey design and ensure that a
sufficient sample size is obtained. Using circle-back procedures will increase observer
certainty of species and group size, which will increase the sample size available for
Chapter 3
80
modelling, and reduce the number of zeroes in the dataset (Dawson et al., 2008).
Correcting for availability will allow abundance estimates to be obtained and improve the
reliability of SDMs. Slooten et al. (2004) demonstrates a method for devising dolphin
species-specific based on the time spent underwater. A standardised grading system for
recording turbidity, without incorporating depth, is also required. Analysis of trends can
be strengthened by incorporating SDM techniques (Cañadas & Hammond, 2006; Forney,
2000; Hammond et al., 2013). Conditions (e.g. Glare, sea state, turbidity) that could affect
the detectability of dolphins should be recorded regularly as the effects of these variables
can be included as covariates in subsequent analyses (Dawson et al., 2008).
Boat-based studies and biopsy sampling
Comprehensive boat-based studies are recommended to obtain basic data on population
parameters that are needed for the conservation of inshore dolphin species. Boat-based
surveys obtain fine scale and in-situ data that aerial surveys cannot. Because boat-based
surveys are conducted at slower speeds (~ 6 knots), than aerial surveys, more time is
available for dolphins to surface and be seen by an observer (Dawson et al., 2008).
Dolphins can be photo-identified and oceanographic and other environmental data can be
collected. Presence and absence data collected during boat-based surveys are highly
reliable so can be used to verify results of SDMs that use aerial survey data.
By incorporating biopsy sampling into boat-based survey programs, information on
habitat use and genetic connectivity can be obtained cost-effectively. Biopsy sampling can
provide data on genetics, prey preferences, foraging ecology, contaminant loads, and
physiological processes (Noren & Mocklin, 2012). Stable isotope analysis (SIA) has become
a widely used technique to understand foraging ecology and habitat use and is useful for
examining diet and trophic level among and within individuals of species (Newsome et al.,
2010). SIA would be useful in examining how resources are partitioned between
bottlenose and humpback dolphins. Such an approach was used by Kiszka et al. (2011) to
understand niche partitioning between sympatric dolphins. SIA has also been used to
distinguish discrete population units (Barros et al., 2010), monitor effects of
anthropogenic disturbance impacts coastal food webs (Rossman et al., 2013). Preliminary
genetic work has been undertaken in the area, but this needs to be expanded (Allen et al.,
2012; Allen et al., in prep; Brown et al., 2014). Levels of connectivity and dispersal rates
within the metapopulation need to be understood.
Localised research should target inner and southern Exmouth Gulf, waters between the
North West Cape and Muiron Islands, the slope at the 20 m isobath, the nearshore waters
Chapter 3
81
of Barrow Island and the intertidal area between Barrow Island and the mainland. Studies
at the slope at the 20 m isobath should obtain genetic data particularly, to determine
whether the area is a convergence of the coastal and offshore bottlenose dolphin species
(i.e. T. aduncus and T. truncatus). Boat-based reconnaissance and biopsy sampling should
also be undertaken at the Ashburton River delta because the area remains unsurveyed. As
the only river in the area, it could be important habitat for humpback and / or snubfin
dolphins as these species are often associated with estuarine environments (Palmer et al.,
2014b; Parra et al., 2006b).
Boat-based focal follows that record dolphin behaviour are needed to understand habitat
use. Data can be incorporated into SDMs in order to identify areas of critical habitat.
Critical habitats are locations where essential behaviours for survival, such as breeding
and foraging, are carried out (Harwood, 2001). Identification of such areas was prioritised
during the Commonwealth funded multi-stakeholder ‘Tropical Inshore Dolphin Workshop’
that was held in 2010, as summarised by Bejder et al. (2010).
Opportunistically collected presence data
Opportunistically collected presence data can supplement data obtained through targeted
studies with species occurrence data in areas and times outside of dedicated study areas.
Utilising data collected that is already being collected by dedicated MFOs has been found
to be useful in obtaining rough measures of species occurrence in unsurveyed areas
(Evans & Hammond, 2004; Kiszka et al., 2007). Dedicated MFOs are usually experienced
observers with considerable training and perform no other tasks so there is reasonable
level of certainty in their sightings and record keeping. Sightings data collected by
commercial tour operators, citizen scientists, and rangers can deliver a form of ‘coastal
surveillance’, and provide a cost-effective solution for monitoring trends across broad
spatial and temporal extents.
Satellite telemetry
Combining satellite telemetry data with data obtained from biopsy sampling is a powerful
way to understand drivers of distribution (Newsome et al., 2010). An understanding of
movement patterns in relation to both genetic dispersal and resource use is essential in
conserving coastal dolphins at population, metapopulation and species levels. This is
especially important for humpback dolphins as their populations are small and prone to
fragmentation (Cagnazzi, 2010; Cagnazzi et al., 2011; Cagnazzi et al., 2013; Parra et al.,
2006a).
Chapter 3
82
It is possible that broad scale movements are an important aspect of dolphin foraging
ecology in the western Pilbara. The area is relatively oligotrophic, without barriers to
movement and characterised by seasonal cyclonic activity and pulses in productivity
(Furnas, 2007; McKinnon et al., 2003; Wilson, 2013). Long-term telemetry studies would
assist in understanding the response to prey distribution caused by episodic climatic
events. Such information would help to equip decision makers considering the potential
effects of climate change and the expected associated increase in cyclones and
redistribution of prey (DoEH & AGO, 2006; MacLeod, 2009).
Satellite tagging can also obtain data to understand diurnal variation in dolphin
distribution. For instance, Elwen et al. (2006), found that the coastally restricted
Heaviside's dolphin (Cephalorhynchus heavisidii), has an onshore-offshore diurnal pattern
of distribution. Such information is important in informing EIA because many
development activities continue at night without MFOs keeping attendance.
Chapter 3
83
Table 12: Recommended studies to understand Pilbara inshore dolphin distribution and habitat use.
Priority Recommended study Purpose Location Requisite design features Examples and references
1 Regional aerial surveys targeting dolphins
Obtain regional abundance and density for all three coastal dolphin species. Construct new SDMs with better dolphin sightings data.
Pilbara inshore waters.
Informed by a priori power analysis.
Systematic line transects.
Transects undertaken in closing mode (i.e. with ‘circle-back’ procedure).
Dual platform observers.
Ensure that estimates of g (0) (perception and availability bias) can be obtained. - Regularly obtain data on
weather conditions likely to affect detectability (e.g. turbidity).
- Obtain data on time dolphin species spend submerged.
- Correct for availability (time submerged and turbidity).
Conduct interannual surveys.
Incorporate modelling in trend analysis.
Forney (1995).
Pollock et al. (2002).
Evans and Hammond (2004).
Slooten et al. (2004).
Hammond et al. (2013).
(Dawson et al., 2008)
2 Biopsy sampling for SIA and molecular analyses.
Characterise diet (the trophic level of prey consumed). Understand genetic connectivity within the (suggested)
Slope at the 20 m bathymetric contour, Pilbara islands including Barrow,
Consider seasonal cycles in food intake and energy demands.
Adopt standardised methods for collecting and preparing tissues.
Wursig and Jefferson (1990).
Krützen et al. (2002).
Barros et al. (2010).
Newsome et al. (2010).
Allen et al. (2012).
Chapter 3
84
Priority Recommended study Purpose Location Requisite design features Examples and references
metapopulation, especially between islands. Understand resource partitioning / niche segregation between bottlenose and humpback dolphins. Determine spatial overlap or partitioning of Tursiops sp.
Montebello, Serrurier, north and south Muiron.
Noren and Mocklin (2012).
Rossman et al. (2013).
Brown et al. (2014).
Rossman et al. (2014).
3 Localised boat-based studies for - photo-identification, - in-situ environmental
sampling, - biopsy, and - focal follows.
Obtain fine scale and localised population abundance (including immigration and emigration rates), habitat data for finer scale SDMs, opportunistic biopsy, identify critical habitat. Obtain data to verify results of SDMs based on aerial surveys.
Areas of North West Cape (underway), Muiron Islands, southern Exmouth Gulf, Barrow Island, Ashburton River Delta.
Abundance:
Systematic line transects
Mark-recapture techniques
Follow Pollock’s robust design
Follow best practice for photo ID Focal follow:
Devise and adhere to a protocol
Use standardised terms for describing behaviour
Record the number of animals observed
Define ‘group’ and record size
Abundance:
Pollock (1982)
Kendall et al. (1997)
Nicholson et al. (2012)
Smith et al. (2013)
Urian et al. (2014) Focal follow:
Mann (1999)
Lusseau and Higham (2004)
Hastie et al. (2004)
Bejder et al. (2006a)
4 Opportunistic presence data (i.e. as collected by commercial tour operators, MFOs, citizen
Supplement data obtained through targeted studies with species occurrence data
Platforms of opportunity; Community groups, tour
Utilise records of dedicated (i.e. professional) MFOs
Reliable sightings (evidence provided by photographs and, if
Evans and Hammond (2004)
Kiszka et al. (2007)
Chapter 3
85
Priority Recommended study Purpose Location Requisite design features Examples and references
scientists, rangers) in areas and times outside of dedicated study areas
operators and rangers based at Exmouth and Onslow.
possible, GPS coordinates to be provided)
5 Satellite telemetry Understand broad scale movements
North West Cape, Barrow Island, Onslow.
Careful consideration of tag type.
Duty cycle needs to meet study objectives.
Target males of breeding age (for movement related to gene dispersal).
Target times following high rainfall (for movement related to episodic feeding events).
Mate et al. (1995).
Elwen et al. (2006).
Balmer et al. (2011).
SIA: stable isotope analysis; WA: Western Australia; SDM: species distribution model; MFO: marine fauna observer.
Chapter 3
86
3.5 Conclusion
MaxEnt was used to obtain the first insights into western Pilbara coastal dolphin
distribution. The key findings were as follows.
Bottlenose and humpback dolphins are sympatric, with overlap in occurrence
across the study area.
Bottlenose dolphin presence is associated with the slope at the 20 m contour and
waters around the Muiron Islands. This is likely to be a productive area and could
therefore be an important dolphin foraging area.
Humpback dolphin presence is associated with intertidal areas, including shallow
coastal waters near the mainland and surrounding islands. With numerous
offshore islands being a key characteristic of the western Pilbara, this would
explain why humpback dolphins were recorded more than 50 km from the
coastline.
The objectives to understand which environmental variables are most important in
defining the distribution of bottlenose and humpback dolphins, and describe the influence
that those variables exert, were partially met. The variability across models and
unconvincing response curves (either spurious or weak) suggested that there is missing
information. It is reasonable to assume from this that inshore dolphin distribution is
influenced by other environmental variables that were not incorporated in the models
(e.g. habitat).
Models were fit to training data but had limited predictive power, with few achieving
acceptable fit to test data. Poor predictive power was reflected in inconsistencies of high
habitat suitability across models. As was the case for the GAMs and boosted models, it is
not clear what the limiting factors that led to poor fit to data were. MaxEnt models were
built in order to avoid problems caused by of unreliable pseudo-absence data, as the
software generates its own background data and then performs slightly different analyses
that acknowledge that true absences across the study area are unknown. The most likely
explanations for largely unreliable models are low sample size and the lack informative
environmental predictors. As a key characteristic of much of the survey area, periods of
high turbidity are also likely to effect dolphin detectability.
A prioritised list of study recommendations has been provided. Ultimately, the low
Chapter 3
87
predictive power of current models highlight the need for dedicated dolphin aerial surveys
to be conducted over the study area. Too few dolphin sightings across a wide study area
were largely a result of the aerial survey targeting dugongs as the primary species, and not
dolphins. Also, the lack of informative environmental variables highlights the general lack
of baseline information available for the northern WA coastline. A habitat dataset that
covered at least the western Pilbara coastline would be of much use for many marine and
coastal conservation research projects. In order to produce reliable and descriptive
distribution models for inshore dolphins of northern WA, better species and
environmental data is required.
Chapter 4
88
Chapter 4: Setting the scene for better predictions of inshore dolphin distribution
4.1 Summary and lessons learnt
This study met its aim, which was to obtain insights into the distribution of inshore
dolphins off western Pilbara. This was accomplished by developing models based on
opportunistically acquired sightings of dolphins acquired during aerial surveys targeting
dugongs. The information obtained through this study is important as it provides the first
insights of dolphin distribution that can inform the conservation of Indo-Pacific bottlenose
and Australian humpback dolphins, in a region of increasing human population and
coastal industrial activity (see Chapter 1).
Acquisition and processing of dolphin sightings and environmental datasets involved a
number of technical decisions, compromises and methods (Chapter 2). There were
intrinsic problems in the dolphin dataset because the aerial surveys were aimed to
determine trends in dugong population abundance. Standardised methods to record
dolphins usually involve distance sampling, which enables a detection function to be
estimated, and a circle-back procedure that enables observers to record species and group
size and composition with certainty (Buckland et al., 1993; Dawson et al., 2008;
Hammond, 2010). Dugong aerial surveys do not incorporate these procedures. As a result,
dolphin availability could not be accounted for, the species of many sightings was not
certain and group size and composition was questionable. Only records with ‘certain’ or
‘probable’ levels of species certainty were retained for modelling which reduced the
sample size. GIS was used to obtain and process the environmental variable data:
distances from coast and islands; SST; front; and bathymetric derivatives. Choice of
environmental variables was limited to data that was available, reliable and affordable. GIS
was also used to generate a binomial presence-absence dataset to represent dolphin
sightings across the study area, and then extract environmental data to each presence or
absence point. Preliminary models were developed using GAMs and component-wise
boosting in order to explore the performance of the presence-absence datasets. Model fits
to the data were poor in all instances. It was not clear whether poor fit was a result of:
a) small sample of dolphin sightings across a large study area,
b) pseudo-absence data (caused by dolphins having been missed by observers),
c) weak environmental variables,
d) effects of weather conditions, particularly turbidity, on dolphin detectability, or
Chapter 4
89
e) a combination of all or some of the factors above.
MaxEnt was used to as an alternative technique to GAMs and boosting, to reduce problems
caused by pseudo-absence data. MaxEnt was used to model dolphin presence and
background data, with algorithms that are different to those for absence data (Chapter 3).
The output takes into account that background data is not true absence data, which avoids
introducing pseudo-absences. MaxEnt was largely successful in building models that fit the
training data but, when evaluated using test data, they were found to unreliable in
predicting dolphin distribution. There was no clear evidence that any particular
environmental variable was consistently important in defining the distribution of either
dolphin species. Variability across models made it difficult to understand which
environmental variables were the most important in defining the distribution of
bottlenose and humpback dolphins. The output maps were used to describe dolphin
distribution patterns, noting that those patterns could also reflect random sampling
variation or uneven detectability across the study area.
Whilst there were constraints in the predictive power of the models, the following insights
into inshore dolphin distribution of the western Pilbara were obtained.
Bottlenose and humpback dolphins are sympatric, with overlap in occurrence across
the study area.
Bottlenose dolphin presence is associated with the slope at the 20 m contour and
waters around the Muiron Islands. This is likely to be a productive area and could
therefore be an important dolphin foraging area.
Humpback dolphin presence is associated with intertidal areas, including shallow
coastal waters near the mainland and surrounding islands. With numerous offshore
islands being a key characteristic of the western Pilbara, this would explain why
humpback dolphins were recorded more than 50 km from the coastline.
Recommendations, with a list of important design features, were presented for dedicated
dolphin aerial surveys, boat based studies with biopsy sampling, collecting of
opportunistic data and satellite telemetry studies (Chapter 3; Table 12). A range of studies
would enable an understanding of various ecological processes that occur at different
spatial and temporal scales, which is needed for adequate species conservation.
Chapter 4
90
4.2 Setting the scene for better predictions
The development of useful SDMs for conservation is often a result of an iterative process,
where findings and limitations of earlier projects can inform subsequent projects. This
leads to continual improvement and models that have greater predictive power (Elith &
Leathwick, 2009; Redfern et al., 2006) (Chapter 3). Eventually, models can be used to test
specific hypotheses, such as how climate change could shift dolphin distribution. The
results of this research project can guide the development of future predictive SDMs for
inshore dolphins of the western Pilbara, and northern WA.
Major constraints for modelling were met during the early phases of the project, which
affected the subsequent phases. The opportunistic nature of dolphin sighting data
collection resulted in many uncertainties in the dataset – both in presences and absences.
The lack of environmental data availability for the study area was also prohibitive.
Distances from coasts and islands seemed as though they would be most useful, but only if
the information could be kept at a fine scale. Creating a fine scale grid across a wide study
area exacerbated zero-inflation. Recommended actions for developing improved inshore
dolphin SDMs across the western Pilbara are presented here, and summarised in Table 13.
These are key actions that should be undertaken in conjunction with usual SDM
procedures.
4.2.1 Obtain more informative training data
Ultimately, a model’s output can only be as good as its input. A colloquial acronym often
used in modelling is ‘GIGO’, which means ‘garbage in is garbage out’. Dedicated aerial
surveys for inshore dolphins, using standardised techniques, are required to obtain
reliable species data (see Chapters 2 and 3). In addition to increasing the sample size
available for modelling, greater certainty in group size and composition could allow count,
calf and mixed species group data to be modelled. Recording environmental data (e.g.
turbidity, weather conditions, and unexpected phenomena such as algal blooms)
frequently and in a way that is appropriate for inshore dolphins could later assist in
obtaining a detection function and understanding if there are natural fluctuations (i.e.
outliers) in the data (Dawson et al., 2008; Zuur et al., 2010).
More informative environmental variables are needed for models to reliably predict
dolphin distribution. Habitat and turbidity data would be particularly useful. The
environmental variables used did contain some information but are unlikely to be key
Chapter 4
91
drivers of inshore dolphin distribution. Oceanographic variables of SST and fronts are
more likely to influence pelagic species with wider distributions. Inshore dolphins are
likely to be driven by currents and fluctuations occurring at more localised scales.
Bathymetric variables in this area were likely weak because they lack variation in data
values, because the study area is predominantly flat. While turbidity could affect
detectability, it could also affect dolphin occurrence. To understand these different effects,
satellite-derived turbidity data could be modelled in conjunction with detectability
functions that account for dolphin availability for observers to record them.
4.2.2 Conduct spatial statistical analysis and appropriate data processing
Response and environmental data must be processed at an appropriate spatial scale for
the modelling project. Subsequent models should consider a coarser resolution than what
was used for this project – if variables that are more informative are obtained. A coarser
resolution is likely to further reduce the potential issue of zero-inflation. Care must be
taken to find the optimal spatial resolution, which is one that captures the scale of
variation in the study area. This can be done objectively using statistical tools such as a
variogram or spatial correlogram. It must be remembered that a predictive model over a
regional area can only provide broad scale predictions. It cannot be used to reveal
distribution patterns at a fine scale. For instance, incorrect predictions for the absence of
the rare butterfly Maculinea nausithous at a fine scale were attributed to missing data on
fine-scale population structure or metapopulation dynamics (Jiménez-Valverde et al.,
2008). Thus, a clear model purpose relative to the study area extent is imperative and
must be upheld as a guiding principal throughout the modelling project.
With better dolphin sightings data and an understanding of optimum resolution, response
data can be processed as encounter rate of total animals or groups. Such data are more
informative than binary presence-absence data, which would assist in building a more
informative model. Dividing the transect line into segments and attaching the number of
individuals or groups will generate encounter rates. Environmental data should then be
extracted as mean values to the centre of each transect line.2
2 To later map predictions, a grid across the study area can be populated with results obtained from models based on training data extracted from transect line.
Chapter 4
92
4.2.3 Conduct exploratory analysis
In addition to the normal exploratory analysis required to understand model
characteristics and determine the best modelling technique, special attention should be
paid to zero-inflation and overdispersion. The use of standardised aerial survey methods
for inshore dolphins will reduce uncertainty in the data, which will mean that the sample
size available for modelling will not be reduced. However, a high number of zeroes (i.e.
absences) in the data may remain because:
a) dolphins are highly mobile and, by chance, they may be outside of the observers’
view from the transect line;
b) dolphins may be near the transect line but not available to be recorded by
observers because they are deep underwater or in turbid waters;
c) inshore dolphins, particularly humpback dolphins, are in relatively low population
sizes which means they are not commonly encountered across the region; or
d) a combination of some or all of the above.
4.2.4 Use appropriate modelling techniques
Once data has been explored so that data characteristics are well understood, the model
technique that best fits those characteristics can be chosen. Finding the right technique
will avoid violation of model assumptions and enable a robust model to be developed.
Datasets for species that are rare or hard to detect typically contain naturally high
proportions of absence records (Arab et al., 2012; Martin et al., 2005). Models that use
encounter rate are likely to have a more informative output than MaxEnt. This is because
reliable absence data can increase the explanatory power and useful information available
for modelling (Arab et al., 2012). If new training data is obtained as recommended and
thorough exploratory analysis does not find that the data is zero-inflated or is
overdispersed, the use of GAMs should be revisited, by specifying a Poisson distribution. If
data is found to be over dispersed, GAMs may still be revisited but a Negative-Binomial
distribution should be specified (Zuur et al., 2010). If data is still zero-inflated, a Zero
Inflated Poisson (ZIP) model is recommended (Martin et al., 2005). Alternatively, a
Bayesian approach incorporating detectability could also be employed. New techniques
are continually being developed to handle complicated datasets. For instance, Kassahun et
al. (2014) present a multi-level marginalised model approach for count data that is zero-
inflated, over-dispersed and correlated. Literature should be regularly reviewed in order
to stay abreast of such developments.
Chapter 4
93
4.2.5 Perform robust model evaluation
Following normal procedures (e.g. using AIC) to select the best-fitting, parsimonious
model and then carrying out diagnostic tests to ensure model assumptions have been met,
it is essential to undertake robust evaluation. Resampling techniques (including
bootstrapping which was used for this project) are good for assessing the explanatory
power of the model. To ascertain the true predictive power of a model, however, further
evaluation with an independent dataset is required. This final step is especially important
when predictions will be used to inform conservation (Jiménez‐Valverde et al., 2008).
Table 13: Recommended actions for developing improved SDMs for inshore dolphins in the western Pilbara and in northern WA. NB: These are key actions that
should be undertaken in conjunction with usual SDM procedures.
Action Purpose
Obtain more informative training data 1. Undertake targeted aerial surveys for inshore dolphins (see Chapters 2 and 3)
Use closed transects with circle-back procedures
Regularly collect data on weather conditions at set intervals on both sides of the plane.
Team leader to take descriptive field notes about unusual environmental phenomena (e.g. algal blooms).
Increase sample size, as a result of greater certainty in the species data recorded. - Increasing sample size would reduce
the problem of zero-inflation.
Reduce the potential for pseudo-absences in the dataset.
Obtain accurate group size and composition data.
Account for availability (i.e. detection and perception bias, see Chapter 2).
Field notes can later help understand whether outliers in the data represent real fluctuations in environmental variables, as opposed to being errors.
Obtain more informative training data 2. Obtain additional environmental data
Obtain reliable and complete habitat data.
Obtain fit-for-purpose turbidity data.
Generate better informed (and therefore more informative) models.
3. Conduct spatial statistical analysis and appropriate data processing
Use spatial statistic tools to understand spatial patterns of the study area (e.g. variograms, spatial correlograms).
Process the response data as encounter rate and group size (enabled by accurate species data).
Determine the optimum resolution for spatial data processing that will adequately capture the scale of environmental variation in the region.
Increase statistical power of models (as encounter rate and group sizes are more informative than presence-absence response variables).
4. Conduct exploratory analysis of training data
Check whether zero-inflation and/or
Deal with overdispersion appropriately (if identified), which can be caused by a zero-inflated dataset
Chapter 4
94
Action Purpose
overdispersion are still a problem.
If outliers in the environmental data exist, determine whether they are ‘real’ or error induced.
of animals that exist in small numbers.
Only transform data if outliers are thought to be errors. Outliers in data that represent true ecological phenomena provide important information to models.
5. Use appropriate modelling techniques, based on outcome of exploratory analysis Consider*:
GAM with Poisson distribution.
GAM specifying a negative binomial (NB) distribution.
Zero Inflated Poisson (ZIP) model.
Bayesian approach incorporating detectability.
* NB: As new techniques are being continually developed to handle complicated datasets (e.g. Kassahun et al., 2014), it is important that a literature review be undertaken.
GAMs with Poisson distribution are appropriate if all other model assumptions are met (e.g. data is not zero-inflated).
GAMs with NB distribution can be appropriate for overdispersed data.
NB distributions are good for modelling environmental data with outliers, which enables data transformation to be avoided.
ZIPs are appropriate for zero-inflated count data, and can deal with excessive zeroes better than GAM with NB distribution.
The Bayesian approach can account for detectability so may be appropriate for datasets with both real and pseudo-absences.
5. Perform robust model evaluation, preferably with an independent dataset.
Achieve explanation and prediction of inshore dolphin distribution that can reliably inform EIA and management plans.
4.3 Concluding remarks
This research presents the first distribution models for inshore dolphins of northern WA.
In doing so, it has paved the way for obtaining a better understanding of coastal dolphin
ecology in a region undergoing rapid development and human population growth.
Distribution modelling for inshore dolphins of northern WA is an intrinsically challenging
research project, due to limited habitat data and the elusive nature of humpback dolphins.
Models can be a powerful tool for conservation, but they can be dangerous if interpreted
incorrectly or misapplied. Incorporating lessons learnt from this project into future
modelling projects will help achieve robust predictions of inshore dolphin distribution.
References
95
References
Acevedo-Gutierrez, A. (2009). Habitat use. In W. F. Perrin, W. Bernd & J. G. M. Thewissen (Eds.), Encyclopaedia of Marine Mammals (Second Edition) (pp. 524-528). London: Academic Press.
Allen, S. J., Bryant, K., Loneragan, N. R., Kopps, A. M., Brown, A., Gerber, L. R., & Krützen, M. (In review.). Genetic isolation between coastal and fishery-impacted offshore bottlenose dolphin (Tursiops spp.) populations of north-western Australia. Mol. Ecol.
Allen, S. J., Cagnazzi, D. D., Hodgson, A. J., Loneragan, N. R., & Bejder, L. (2012). Tropical inshore dolphins of north-western Australia: Unknown populations in a rapidly changing region. Pacific Conservation Biology, 18(1), 56-63.
Allen, S. J., Cagnazzi, D., Krützen, M., Bejder, L., & Parra, G. J. (2011). Alliance formation by male Indo-Pacific humpback dolphins of north-western Australia. Paper presented at the The 48th Annual Conference of the Australian Marine Sciences Association, Fremantle, WA.
Allen, S. J., Tyne, J. A., Kobryn, H. T., Bejder, L., Pollock, K. H., & Loneragan, N. R. (2014). Patterns of dolphin bycatch in a north-western Australian trawl fishery. PLoS ONE, 9(4), e93178.
Allen, S.J., Loneragan, N.R., Kopps, A. M., Kobryn, H.T., & Krützen, M. (in prep). What genes tell us about trawler-associated dolphins in the Pilbara Demersal Scalefish Fishery off north-western Australia.
Alvarado‐Serrano, D. F., & Knowles, L L. (2014). Ecological niche models in
phylogeographic studies: applications, advances and precautions. Molecular ecology resources, 14(2), 233-248.
Amir, O. A., Berggren, P., Ndaro, S. G. M., & Jiddawi, N. S. (2005). Feeding ecology of the Indo-Pacific bottlenose dolphin (Tursiops aduncus) incidentally caught in the gillnet fisheries off Zanzibar, Tanzania. Estuarine, Coastal and Shelf Science, 63(3), 429-437.
AMSA. (2014). Craft tracking system November 2014. Anderson, D. R. (2001). The need to get the basics right in wildlife field studies. Wildlife
Society Bulletin, 1294-1297. Arab, A., Holan, S. H., Wikle, C. K., & Wildhaber, M. L. (2012). Semiparametric bivariate zero
‐inflated Poisson models with application to studies of abundance for multiple species. Environmetrics, 23(2), 183-196.
Araujo, M. B., & Guisan, A. (2006). Five (or so) challenges for species distribution modelling. Journal of Biogeography, 33(10), 1677-1688.
Austin, M. P. (2002). Spatial prediction of species distribution: an interface between ecological theory and statistical modelling. Ecological Modelling, 157(2-3), 101-118. doi: 10.1016/s0304-3800(02)00205-3
Austin, M. P. (2007). Species distribution models and ecological theory: a critical assessment and some possible new approaches. Ecological Modelling, 200(1), 1-19.
Ayukai, T., & Miller, D. (1998). Phytoplankton biomass, production and grazing mortality in Exmouth Gulf, a shallow embayment on the arid, tropical coast of Western Australia. Journal of Experimental Marine Biology and Ecology, 225(2), 239-251.
Bailey, H., & Thompson, P. M. (2009). Using marine mammal habitat modelling to identify priority conservation zones within a marine protected area. Marine Ecology-Progress Series, 378, 279-287. doi: 10.3354/meps07887
Balmer, B. C., Wells, R. S., Schwacke, L. H., Rowles, T. K., Hunter, C., Zolman, E. S., Townsend, F. I., Danielson, B., Westgate, A. J., & McLellan, W. A. (2011). Evaluation of a single-pin, satellite-linked transmitter deployed on bottlenose dolphins (Tursiops truncatus) along the coast of Georgia, USA. Aquatic Mammals, 37(2), 187-192.
References
96
Bancroft, K. P, & Long, S. (2008). Modelling suggests connectivity between the Ningaloo Reef and coral reefs of the Pilbara. Ningaloo Research Program Progress Report. (pp. 61-64). Perth: Department of Environment and Conservation.
Banks, S. C., Piggott, M. P., Stow, A. J., & Taylor, A. C. (2007). Sex and sociality in a disconnected world: a review of the impacts of habitat fragmentation on animal social interactions. Canadian Journal of Zoology, 85(10), 1065-1079.
Barbet‐Massin, M., Jiguet, F., Albert, Cecile H., & Thuiller, W. (2012). Selecting pseudo‐
absences for species distribution models: how, where and how many? Methods in Ecology and Evolution, 3(2), 327-338.
Barlow, J., Brownell Jr, L., DeMaster, D. P., Forney, K. A., Lowry, M. S., Osmek, S., Ragen, T. J., Reeves, R. R., & Small, R. J. (1995). US Pacific marine mammal stock assessments. NOAA Technical Memorandum., NMFS-SWFSC-219, 162.
Barros, N. B., Ostrom, P. H., Stricker, C. A., & Wells, R. S. (2010). Stable isotopes differentiate bottlenose dolphins off west‐central Florida. Marine Mammal
Science, 26(2), 324-336. Becker, E. A., Forney, K. A., Ferguson, M. C., Foley, D. G., Smith, R. C., Barlow, J., & Redfern, J.
V. (2010). Comparing California Current cetacean-habitat models developed using in situ and remotely sensed sea surface temperature data. Marine Ecology-Progress Series, 413, 163-183. doi: 10.3354/meps08696
Bejder, L., Hodgson, A. J., Loneragan, N. R., & Allen, S. J. (2012). Coastal dolphins in north-western Australia: The need for re-evaluation of species listings and short-comings in the Environmental Impact Assessment process. Pacific Conservation Biology, 18, 22-25.
Bejder, L., Samuels, A., Whitehead, H., Finn, H., & Allen, S. (2009). Impact assessment research: use and misuse of habituation, sensitisation and tolerance in describing wildlife responses to anthropogenic stimuli. Marine Ecol. Progress Series, 395, 177–185.
Bejder, L., Samuels, A., Whitehead, H., & Gales, N. (2006a). Interpreting short-term behavioural responses to disturbance within a longitudinal perspective. Animal Behaviour, 72(5), 1149-1158.
Bejder, L., Samuels, A., Whitehead, H., Gales, N., Mann, J., Connor, J., Heithaus, M., Watson-Capps, J., Flaherty, C., & Krützen, M. (2006b). Decline in relative abundance of bottlenose dolphins exposed to long-term disturbance. Conservation Biology, 20(6), 1791-1798.
Benoit-Bird, K. J., & Au, W. W. L. (2003). Prey dynamics affect foraging by a pelagic predator (Stenella longirostris) over a range of spatial and temporal scales. Behavioral Ecology and Sociobiology, 53(6), 364-373. doi: 10.1007/s00265-003-0585-4
Blasi, M. F., & Boitani, L. (2012). Modelling fine-scale distribution of the bottlenose dolphin Tursiops truncatus using physiographic features on Filicudi (southern Thyrrenian Sea, Italy). Endangered Species Research, 17(3), 269-288.
Box, G. E. P., & Cox, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical Society. Series B (Methodological), 211-252.
Brager, S., Harraway, J. A., & Manly, B. F. J. (2003). Habitat selection in a coastal dolphin species (Cephalorhynchus hectori). Marine Biology, 143(2), 233-244. doi: 10.1007/s00227-003-1068-x
Braithwaite, J. E., Meeuwig, J. J., & Hipsey, M. R. (2015). Optimal migration energetics of humpback whales and the implications of disturbance. Conservation Physiology, 3(1). doi: 10.1093/conphys/cov001
Briscoe, D. K., Hiatt, S., Lewison, R., & Hines, E. (2014). Modeling habitat and bycatch risk for dugongs in Sabah, Malaysia. Endangered Species Research, 24, 237-247.
References
97
Brown, A. M., Bejder, L., Cagnazzi, D., Parra, G. J., & Allen, S. J. (2012). The North West Cape, Western Australia: A potential hotspot for Indo-Pacific humpback dolphins 'Sousa chinensis'? Pacific Conservation Biology, 18(4), 240.
Brown, A. M., Kopps, A. M., Allen, S. J., Bejder, L., Littleford-Colquhoun, B., Parra, G. J, Cagnazzi, D., Thiele, D., Palmer, C., & Frère, C. H. (2014). Population differentiation and hybridisation of Australian snubfin (Orcaella heinsohni) and Indo-Pacific humpback (Sousa chinensis) dolphins in north-western Australia. PLoS ONE.
Buckland, S. T., Anderson, D .R., Burnham, K. P., & Laake, J. L. (1993). Distance sampling: estimating abundance of biological populations: Chapman & Hall.
Buehlmann, P., & Hothorn, T. (2007). Boosting algorithms: regularization, prediction and model fitting (with discussion). Statistical Science, 22(4), 477 - 505.
Bulman, C. (2006). Trophic webs and modelling of Australia's North West Shelf. North West Shelf Joint Environmental Management Study, Technical Report No 9: CSIRO Marine and Atmospheric Research.
Burgman, M. A., & Lindenmayer, D. B. (1998). Conservation biology for the australian environment. . New South Wales, Australia: Surrey Beatty and Sons.
Cagnazzi, D. (2010a). Conservation Status of Australian snubfin dolphin, Orcaella heinsohni, and Indo-Pacific humpback dolphin, Sousa chinensis, in the Capricorn Coast, Central Queensland, Australia. PhD thesis., Southern Cross University, Lismore, Australia.
Cagnazzi, D., Harrison, P. L., Ross, G. J. B., & Lynch, P. (2011). Abundance and site fidelity of Indo-Pacific Humpback dolphins in the Great Sandy Strait, Queensland, Australia. Marine Mammal Science, 27(2), 255-281.
Cagnazzi, D., Parra, G. J., Westley, S., & Harrison, P. L. (2013). At the heart of the industrial boom: Australian snubfin dolphins in the Capricorn Coast, Queensland, need urgent conservation action. PLoS ONE, 8(2).
CALM. (2005). Management Plan for the Ningaloo Marine Park and Muiron Islands Marine Management Area 2005 - 2015. Perth: Western Australian Government Department of Conservation and Land Management.
Cañadas, A., & Hammond, P. S. (2006). Model-based abundance estimates for bottlenose dolphins off southern Spain: implications for conservation and management. Journal of Cetacean Research and Management, 8(1), 13.
Cañadas, A., & Hammond, P. S. (2008). Abundance and habitat preferences of the short-beaked common dolphin Delphinus delphis in the southwestern Mediterranean: implications for conservation. Endangered Species Research, 4(3), 309.
Charlton-Robb, K., Gershwin, L., Thompson, R., Austin, J., Owen, K., & McKechnie, S. (2011). A new dolphin species, the Burrunan dolphin Tursiops australis sp. nov., endemic to southern Australian coastal waters. PLoS ONE, 6(9), e24047.
Chevron Australia. (2010). Wheatstone Project Environmental Impact Statement (EIS) / Environmental Review and Management Programme (ERMP). Perth.
Chilvers, B. L., & Corkeron, P. J. (2002). Association patterns of bottlenose dolphins (Tursiops aduncus) off Point Lookout, Queensland, Australia. Canadian Journal of Zoology-Revue Canadienne De Zoologie, 80(6), 973-979. doi: 10.1139/z02-075
Chilvers, B. L., & Corkeron, P. J. (2003). Abundance of Indo‐Pacific bottlenose dolphins, Tursiops aduncus, off point lookout, Queensland, Australia. Marine Mammal Science, 19(1), 85-095.
Chilvers, B. L., Corkeron, P. J., & Puotinen, M. L. (2003). Influence of trawling on the behaviour and spatial distribution of Indo-Pacific bottlenose dolphins (Tursiops aduncus) in Moreton Bay, Australia. Canadian Journal of Zoology-Revue Canadienne De Zoologie, 81(12), 1947-1955. doi: 10.1139/z03-195
Chittleborough, R. G. (1953). Aerial observations on the humpback whale, Megaptera nodosa (Bonnaterre), with notes on other species. Mar Freshwater Res, 4, 219–226.
Connor, R. C. (2000). Group living in whales and dolphins. Chicago: The University of Chicago Press.
References
98
Connor, R. C. (2001). Individual foraging specializations in marine mammals: Culture and ecology. Behavioral and Brain Sciences, 24(2), 329-330.
Corkeron, P. J., Morissette, N. M., Porter, L, & Marsh, H. (1997). Distribution and status of hump-backed dolphins, Sousa chinensis, in Australian waters. Hong Kong: Hong Kong University Press.
Cresswell, G. R., & Golding, T. J. (1980). Observations of a south-flowing current in the southeastern Indian Ocean. Deep Sea Research Part A. Oceanographic Research Papers, 27(6), 449-466.
CW. (2015). Coastal Walkabout: Smart data collection for the engaged coastal citizen. Retrieved 5th February, 2015, from http://www.coastalwalkabout.org
Dähne, M., Gilles, A., Lucke, K., Peschko, V., Adler, S., Krügel, K., Sundermeyer, J., & Siebert, U. (2013). Effects of pile-driving on harbour porpoises (Phocoena phocoena) at the first offshore wind farm in Germany. Environmental Research Letters, 8(2), 025002.
David, J. A. (2006). Likely sensitivity of bottlenose dolphins to pile‐driving noise. Water
and Environment Journal, 20(1), 48-54. Dawson, S., Wade, P., Slooten, E., & Barlow, J. (2008). Design and field methods for sighting
surveys of cetaceans in coastal and riverine habitats. Mammal Review, 38(1), 19-49.
DEC. (2007). Management Plan for the Montebello/Barrow Islands Marine Conservation Reserves 2007 - 2017. Perth: Western Australian Government Department of Environment and Conservation.
DEWHA. (2008). The north-west marine bioregional plan: bioregional profile. Canberra: Australian Government Department of the Environment Water Heritage and the Arts.
DMP. (2014). Major Resource Projects. DoE. (2013). Matters of National Environmental Significance: Significant impact guidelines
1.1. Environment Protection and Biodiversity Conservation Act 1999. Australian Government, Department of the Environment, Canberra.
DoEH, & AGO. (2006). Impacts of climate change on Australian marine life: Part B. Technical Report. In A. J. Hobday, T. A. Okey, E. S. Poloczanska, T. J. Kunz & A. J. Richardson (Eds.), Report to the Australian Greenhouse Office. Canberra: Australian Government Department of Environment and Heritage Australian Greenhouse Office.
DoF. (2014). Status reports of the fisheries and aquatic resources of Western Australia 2013/14. Perth: Western Australian Government Department of Fisheries.
Doniol-Valcroze, T., Lesage, V., Giard, J., & Michaud, R. (2012). Challenges in marine mammal habitat modelling: evidence of multiple foraging habitats from the identification of feeding events in blue whales. Endangered Species Research, 17(3), 255-268.
Edrén, S., Wisz, M. S., Teilmann, J., Dietz, R., & Söderkvist, J. (2010). Modelling spatial patterns in harbour porpoise satellite telemetry data using maximum entropy. Ecography, 33(4), 698-708.
Elith, J., Graham, C. H., Anderson, R. P., Dudik, M., Ferrier, S., Guisan, A., Hijmans, R. J., Huettmann, F., Leathwick, J. R., Lehmann, A., Li, J., Lohmann, L. G., Loiselle, B. A., Manion, G., Moritz, C., Nakamura, M., Nakazawa, Y., Overton, J. M., Peterson, A. T., Phillips, S. J., Richardson, K., Scachetti-Pereira, R., Schapire, R. E., Soberon, J., Williams, S., Wisz, M. S., & Zimmermann, N. E. (2006). Novel methods improve prediction of species' distributions from occurrence data. Ecography, 29(2), 129-151. doi: 10.1111/j.2006.0906-7590.04596.x
Elith, J., & Leathwick, J. R. (2009). Species distribution models: ecological explanation and prediction across space and time. Annual Review of Ecology, Evolution, and Systematics, 40, 677-697.
References
99
Elith, J., Leathwick, J. R., & Hastie, T. (2008). A working guide to boosted regression trees. Journal of Animal Ecology, 77(4), 802-813.
Elith, J., Phillips, S. J., Hastie, T., Dudík, M., Chee, Y. E., & Yates, C. J. (2011). A statistical explanation of MaxEnt for ecologists. Diversity and Distributions, 17(1), 43-57.
Elwen, S., Meÿer, M. A., Best, P.B., Kotze, P. G. H., Thornton, M., & Swanson, S. (2006). Range and movements of female Heaviside's dolphins (Cephalorhynchus heavisidii), as determined by satellite-linked telemetry. Journal of Mammalogy, 87(5), 866-877.
EPA. (2009). Review of the environmental impact assessment process in Western Australia. Perth: Environmental Protection Authority of Western Australia.
EPA. (2013). Environmental assessment guideline 8: environmental factors and objectives. Perth: Environmental Protection Authority of Western Australia.
ESRI. (2012). ArcGIS Desktop: Release 10.1. Redlands, CA.: Environmental Systems Research Institute.
Estes, J. A., Terborgh, J., Brashares, J. S, Power, M. E, Berger, J., Bond, W. J., Carpenter, S. R., Essington, T. E., Holt, R. D., & Jackson, J. B. C. (2011). Trophic downgrading of planet Earth. Science, 333(6040), 301-306.
Evans, P. G. H., & Hammond, P. S. (2004). Monitoring cetaceans in European waters. Mammal Review, 34(1-2), 131-156. doi: 10.1046/j.0305-1838.2003.00027.x
Forney, K. A. (1995). A decline in the abundance of harbor porpooise, Phocoena phocoena, in nearshore waters off California, 1986-93. La Jolla: National Marine Fisheries Service, NOAA.
Forney, K. A. (2000). Environmental models of cetacean abundance: Reducing uncertainty in population trends. Conservation Biology, 14(5), 1271-1286. doi: 10.1046/j.1523-1739.2000.99412.x
Fox, N. J., & Beckley, L. E. (2005). Priority areas for conservation of Western Australian coastal fishes: a comparison of hotspot, biogeographical and complementarity approaches. Biological Conservation, 125(4), 399-410.
Frère, C. H., Hale, P. T., Porter, L., Cockcroft, V. G., & Dalebout, M. L. (2008). Phylogenetic analysis of mtDNA sequences suggests revision of humpback dolphin (Sousa spp.) taxonomy is needed. Marine and Freshwater Research, 59(3), 259-268.
Frère, C. H., Seddon, J., Palmer, C., Porter, L., & Parra, G. J. (2011). Multiple lines of evidence for an Australasian geographic boundary in the Indo-Pacific humpback dolphin (Sousa chinensis): population or species divergence? Conservation Genetics, 12(6), 1633-1638.
Frid, A., & Dill, L. M. (2002). Human-caused disturbance stimuli as a form of predation risk. Conservation Ecology, 6(1), 11.
Friedlaender, A. S., Johnston, D. W., Fraser, W. R., Burns, J., & Costa, D. P. (2011). Ecological niche modeling of sympatric krill predators around Marguerite Bay, Western Antarctic Peninsula. Deep Sea Research Part II: Topical Studies in Oceanography, 58(13), 1729-1740.
Furnas, M. (2007). Intra-seasonal and inter-annual variations in phytoplankton biomass, primary production and bacterial production at North West Cape, Western Australia: Links to the 1997–1998 El Niño event. Continental Shelf Research, 27(7), 958-980.
Fury, C. A., & Harrison, P. L. (2011). Seasonal variation and tidal influences on estuarine use by bottlenose dolphins (Tursiops aduncus). Estuarine Coastal and Shelf Science, 93(4), 389-395. doi: 10.1016/j.ecss.2011.05.003
Gales, N., McCauley, R. D., Lanyon, J., & Holley, D. (2004). Change in abundance of dugongs in Shark Bay, Ningaloo and Exmouth Gulf, Western Australia: evidence for large-scale migration. Wildlife Research, 31(3), 283-290.
Garaffo, G. V., Dans, S. L., Pedraza, S. N., Degrati, M., Schiavini, A., Gonzalez, R., & Crespo, E. A. (2011). Modeling habitat use for dusky dolphin and Commerson's dolphin in Patagonia. Marine Ecology Progress Series, 421, 217-227.
References
100
Geoscience Australia. (2004). GEODATA Coast 100k. Geoscience Australia. (2009). Australian bathymetry and topography grid. Godfrey, JS, & Ridgway, KR. (1985). The large-scale environment of the poleward-flowing
Leeuwin Current, Western Australia: Longshore steric height gradients, wind stresses and geostrophic flow. Journal of Physical Oceanography, 15(5), 481-495.
Gómez de Segura, A., Hammond, P. S., & Raga, J. A. (2008). Influence of environmental factors on small cetacean distribution in the Spanish Mediterranean. Journal of the Marine Biological Association of the UK, 88(06), 1185-1192.
Gregr, E. J. (2004). Modeling species-habitat relationships in the marine environment: A comment on Hamazaki (2002). Marine Mammal Science, 20(2), 353-355. doi: 10.1111/j.1748-7692.2004.tb01165.x
Gu, W., & Swihart, R. K. (2004). Absent or undetected? Effects of non-detection of species occurrence on wildlife–habitat models. Biological Conservation, 116(2), 195-203.
Guisan, A., Edwards Jr, T. C., & Hastie, T. (2002). Generalized linear and generalized additive models in studies of species distributions: setting the scene. Ecological Modelling, 157(2), 89-100.
Guisan, A., Graham, C. H., Elith, J., & Huettmann, F. (2007). Sensitivity of predictive species distribution models to change in grain size. Diversity and Distributions, 13(3), 332-340.
Guisan, A., & Thuiller, W. (2005). Predicting species distribution: offering more than simple habitat models. Ecology letters, 8(9), 993-1009.
Guisan, A., & Zimmermann, N. E. (2000). Predictive habitat distribution models in ecology. Ecological Modelling, 135(2–3), 147-186. doi: http://dx.doi.org/10.1016/S0304-3800(00)00354-9
Gutzwiller, K. (2002). Applying landscape ecology in biological conservation. New York: Springer.
Halpern, B. S., Walbridge, S., Selkoe, K.A., Kappel, C.V., Micheli, F., D'Agrosa, C., Bruno, J. F., Casey, K. S., Ebert, C., Fox, H. E., Fujita, R., Heinemann, D., Lenihan, H. S., Madin, E. M. P., Perry, M. T., Selig, E. R., Spalding, M., Steneck, R., & Watson, R. (2008). A global map of human impact on marine ecosystems. Science, 319(5865), 948-952. doi: 10.1126/science.1149345
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(4), 920-939. doi: 10.1111/j.1748-7692.2002.tb01082.x
Hammond, P. S. (2010). Estimating the abundance of marine mammals. In I. L. Boyd, W. D. Bowen & S. J. Iverson (Eds.), Marine mammal ecology and conservation: a handbook of techniques. Oxford: Oxford University Press.
Hammond, P. S., Macleod, K., Berggren, P., Borchers, D. L., Burt, L., Cañadas, A., Desportes, G., Donovan, G. P., Gilles, A., Gillespie, D., Gordon, J., Hiby, L., Kuklik, I., Leaper, R., Lehnert, K., Leopold, M., Lovell, P., Øien, N., Paxton, C. G. M., Ridoux, V., Rogan, E., Samarra, F., Scheidat, M., Sequeira, M., Siebert, U., Skov, H., Swift, R., Tasker, M. L., Teilmann, J., Van Canneyt, O., & Vázquez, José A. (2013). Cetacean abundance and distribution in European Atlantic shelf waters to inform conservation and management. Biological Conservation, 164(0), 107-122. doi: http://dx.doi.org/10.1016/j.biocon.2013.04.010
Hammond, P.S., Bearzi, G., Bjørge, A., Forney, K.A., Karkzmarski, L., Kasuya, T., Perrin, W.F., Scott, M.D., Wang, J.Y., Wells, R.S., & Wilson, B. . (2012). Tursiops aduncus. In: IUCN 2014. IUCN Red List of Threatened Species. Version 2014.1.
Hanski, I., & Gaggiotti, O. E. (2004). Ecology, genetics and evolution of metapopulations. San Diego: Elsevier Academic Press.
Harwood, J. (2001). Marine mammals and their environment in the twenty-first century. Journal of Mammalogy, 82(3), 630-640.
References
101
Hastie, G. D., Wilson, B., Wilson, L. J., Parsons, K. M., & Thompson, P. M. (2004). Functional mechanisms underlying cetacean distribution patterns: hotspots for bottlenose dolphins are linked to foraging. Marine Biology, 144(2), 397-403.
Hatton, T., Cork, S., P., Harper., Joy, R., Kanowski, P., Mackay, R., McKenzie, N., Ward, T., & Wienecke, B. (2011). The Australian State of the Environment (SoE) 2011: Report for the Department of Sustainability, Environment, Water, Population and Communities (DSEWPaC). Canberra: Australian Government.
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. doi: 10.1017/s0952836901000061
Heithaus, M. R., & Dill, L. M. (2002). Food availability and tiger shark predation risk influence bottlenose dolphin habitat use. Ecology, 83(2), 480-491.
Heithaus, M. R., Frid, A., Wirsing, A. J., & Worm, B. (2008). Predicting ecological consequences of marine top predator declines. Trends in ecology & evolution, 23(4), 202-210.
Hernandez, P. A., Graham, C. H., Master, L. L., & Albert, D. L. (2006). The effect of sample size and species characteristics on performance of different species distribution modeling methods. Ecography, 29(5), 773-785.
Hirzel, A. H., Le Lay, G., Helfer, V., Randin, C., & Guisan, A. (2006). Evaluating the ability of habitat suitability models to predict species presences. Ecological Modelling, 199(2), 142-152.
Hodgson, A. (2007). The distribution, abundance and conservation of dugongs and other marine megafauna in Shark Bay Marine Park, Ningaloo Reef Marine Park and Exmouth Gulf. Report to the Western Australia Department of Environment and Conservation September.
Hodgson, A. (2012). Wheatstone Project dugong research plan. Perth: Murdoch University. Holling, C. S. (1992). Cross-scale morphology, geometry, and dynamics of ecosystems.
Ecological Monographs, 62(4), 447-502. doi: 10.2307/2937313 Hosmer, D. W., & Lemeshow, S. . (1989). Applied logistic regression. . New York: John Wiley
& Sons. Hothorn, T., Buehlmann, P., Kneib, T., Schmid, M., & Hofner, B. (2013). mboost: Model-
based boosting (Version R package version 2.2-3). Hughes, T. P., Baird, A. H., Bellwood, D. R., Card, M., Connolly, S. R., Folke, C., Grosberg, R.,
Hoegh-Guldberg, O., Jackson, J., & Kleypas, J. (2003). Climate change, human impacts, and the resilience of coral reefs. Science, 30(5635), 929-933. .
Jackson, G. D., Meekan, M. G., Wotherspoon, S., & Jackson, C. H. . (2008). Distributions of young cephalopods in the tropical waters of Western Australia over two consecutive summers. ICES Journal of Marine Science: Journal du Conseil, 65, 140 - 147.
Jefferson, T. A., Hung, S. K., & Wursig, B. (2009). Protecting small cetaceans from coastal development: Impact assessment and mitigation experience in Hong Kong. Marine Policy, 33(2), 305-311.
Jefferson, T. A., & Rosenbaum, H. C. (2014). Taxonomic revision of the humpback dolphins (Sousa spp.), and description of a new species from Australia. Marine Mammal Science, 'EarlyView'. doi: 10.1111.mms.12152
Jefferson, Thomas A., & Hung, Samuel K. (2004). A review of the status of the Indo-Pacific humpback dolphin (Sousa chinensis) in Chinese waters. Aquatic Mammals, 30(1), 149-158. doi: 10.1578/am.30.1.2004.149
Jenner, C. K. S., Jenner, M. N., & Pirzl, R. (2014). Cetaceans. pp 401-454 in Comrie-Greig, J. and Abdo, L.J. (eds). Ecological studies of the Bonaparte Archipelago and Browse Basin. INPEX Operations Australia Pty Ltd., Perth, Western Australia.
References
102
Jenner, K. C. S., Jenner, M. N., & McCabe, K. A. (2001). Geographical and temporal movements of humpback whales in Western Australian waters. APPEA Journal, 38, 692–707.
Jewell, R., Thomas, L., Harris, C. M., Kaschner, K., Wiff, R., Hammond, P. S., & Quick, N. J. (2012). Global analysis of cetacean line-transect surveys: detecting trends in cetacean density. Marine Ecology-Progress Series, 453, 227-240. doi: 10.3354/meps09636
Jiménez-Valverde, A., Gómez, J. F., Lobo, J. M., Baselga, A., & Hortal, J. (2008). Challenging species distribution models: the case of Maculinea nausithous in the Iberian Peninsula. Annales Zoologici Fennici, 45(3), 200-210.
Jiménez‐Valverde, A., Lobo, J. M., & Hortal, J. (2008). Not as good as they seem: the importance of concepts in species distribution modelling. Diversity and Distributions, 14(6), 885-890.
Karczmarski, L. (2000). Conservation and management of humpback dolphins: the South African perspective. Oryx, 34(3), 207-216.
Karczmarski, L., Cockcroft, V. G., & McLachlan, A. (1999). Group size and seasonal pattern of occurrence of humpback dolphins Sousa chinensis in Algoa Bay, South Africa. South African Journal of Marine Science, 21(1), 89-97.
Karczmarski, L., Cockcroft, V. G., & McLachlan, A. (2000). Habitat use and preferences of Indo‐Pacific humpback dolphins Sousa chinensis in Algoa Bay, South Africa. Marine Mammal Science, 16(1), 65-79.
Kassahun, W., Neyens, T., Faes, C., Molenberghs, G., & Verbeke, G. (2014). A zero-inflated overdispersed hierarchical Poisson model. Statistical Modelling, 14(5), 439-456.
Kendall, W. L., Nichols, J. D., & Hines, J. E. (1997). Estimating temporary emigration using capture-recapture data with Pollock's robust design. Ecology, 78(2), 563-578.
Kendrick, P., & Stanley, F. (2001). Pilbara 4 (PIL4 - Roebourne synopsis) A Biodiversity Audit of Western Australia's 53 Biogeographic Subregions in 2002. Perth: Department of Conservation and Land Management.
Kiszka, J., Macleod, K., Van Canneyt, O., Walker, D., & Ridoux, V. (2007). Distribution, encounter rates, and habitat characteristics of toothed cetaceans in the Bay of Biscay and adjacent waters from platform-of-opportunity data. ICES Journal of Marine Science: Journal du Conseil, 64(5), 1033-1043.
Kiszka, J., Simon-Bouhet, B., Martinez, L., Pusineri, C., Richard, P., & Ridoux, V. (2011). Ecological niche segregation within a community of sympatric dolphins around a tropical island. Marine Ecology Progress Series, 433(1).
Krützen, M., Barré, L. M., Möller, L. M., Heithaus, M. R., Simms, C., & Sherwin, W. B. (2002). A biopsy system for small cetaceans: Darting success and wound healing in Tursiops spp. Marine Mammal Science, 18(4), 863-878.
Krützen, M., Sherwin, W. B., Berggren, P., & Gales, N. (2004). Population structure in an inshore cetacean revealed by microsatellite and mtDNA analysis: Bottlenose dolphins (Tursiops sp.) in Shark Bay, Western Australia.
Lima, S. L., & Dill, L. M. (1990). Behavioral decisions made under the risk of predation: a review and prospectus. Canadian Journal of Zoology, 68(4), 619-640.
Loughane, B. (2013). The Coalition's 2030 vision for developing northern Australia. Canberra: Liberal Party.
Lozier, J. D., Aniello, P., & Hickerson, M. J. (2009). Predicting the distribution of Sasquatch in western North America: anything goes with ecological niche modelling. Journal of Biogeography, 36(9), 1623-1627.
Lusseau, D. (2004). The hidden cost of tourism: detecting long-term effects of tourism using behavioral information. Ecology and Society, 9(1), 2.
References
103
Lusseau, D., & Higham, J. E. S. (2004). Managing the impacts of dolphin-based tourism through the definition of critical habitats: the case of bottlenose dolphins (Tursiops spp) in Doubtful Sound, New Zealand. Tourism Management, 25(6), 657-667.
MacLeod, C. D. (2009). Global climate change, range changes and potential implications for the conservation of marine cetaceans: a review and synthesis. Endangered Species Research, 7(2), 125-136.
MacLeod, C. D. (2013). An introduction to using GIS in marine mammal research. Glasgow: Pictish Beast Publications.
MacLeod, C. D., Weir, C. R., Pierpoint, C., & Harland, E. J. (2007). The habitat preferences of marine mammals west of Scotland (UK). Journal of the Marine Biological Association of the United Kingdom, 87(1), 157-164. doi: 10.1017/s0025315407055270
Mann, J. (1999). Behavioral sampling methods for cetaceans: a review and critique. Marine Mammal Science, 15(1), 102-122. doi: 10.1111/j.1748-7692.1999.tb00784.x
Mann, J., Connor, R. C., Barre, L. M., & Heithaus, M. R. (2000). Female reproductive success in bottlenose dolphins (Tursiops sp.): life history, habitat, provisioning, and group size effects. Behavioural Ecology, 11, 210-219.
Mann, J., & Sargeant, B. L. (2003). Like mother, like calf: The ontogeny of foraging traditions in wild Indian Ocean bottlenose dolphins (Tursiops sp.). In D. Fragaszy & S. Perry (Eds.), The biology of traditions: models and evidence (pp. 236-266). Cambridge: Cambridge University Press.
Marsh, H., & Sinclair, D. Franklin. (1989). An experimental evaluation of dugong and sea turtle aerial survey techniques. Wildlife Research, 16(6), 639-650.
Martin, T. G., Wintle, B. A., Rhodes, J. R., Kuhnert, P. M., Field, S. A., Low‐Choy, S. J., Tyre, A. J., & Possingham, Hu. P. (2005). Zero tolerance ecology: improving ecological inference by modelling the source of zero observations. Ecology letters, 8(11), 1235-1246.
Mate, B. R., Rossbach, K. A., Nieukirk, S. L., Wells, R. S., Blair Irvine, A., Scott, M. D., & Read, A. J. (1995). Satellite‐monitored movements and dive behavior of a bottlenose
dolphin (Tursiops truncatus) in Tampa Bay, Florida. Marine Mammal Science, 11(4), 452-463.
May-Collado, L. J., & Quiñones-Lebrón, S. G. (2014). Dolphin changes in whistle structure with watercraft activity depends on their behavioral state. The Journal of the Acoustical Society of America, 135(4), EL193-EL198.
Mayr, A., Hofner, B., & Schmid, M. (2011). The importance of knowing when to stop. A sequential stopping rule for component-wise gradient boosting. Methods of Information in Medicine, 51(2), 178-186.
McKinnon, A. D., Meekan, M. G., Carleton, J. H., Furnas, M. J., Duggan, S., & Skirving, W. (2003). Rapid changes in shelf waters and pelagic communities on the southern Northwest Shelf, Australia, following a tropical cyclone. Continental Shelf Research, 23(1), 93-111.
McPherson, J., Jetz, W., & Rogers, D. J. (2004). The effects of species’ range sizes on the accuracy of distribution models: ecological phenomenon or statistical artefact? Journal of Applied Ecology, 41(5), 811-823.
Mendez, M., Jefferson, T. A., Kolokotronis, S.-O., Krützen, M., Parra, G. J., Collins, T., Minton, G., Baldwin, R., Berggren, P., Särnblad, A., Amir, O. A., Peddemors, V. M., Karczmarski, L., Guissamulo, A., Smith, B.D., Sutaria, D., Amato, G., & Rosenbaum, H. C. (2013). Integrating multiple lines of evidence to better understand the evolutionary divergence of humpback dolphins along their entire distribution range: a new dolphin species in Australian waters? Mol Ecol. doi: 10.1111/mec.12535
References
104
Mendez, M., Subramaniam, A., Collins, T., Minton, G., Baldwin, R., Berggren, P., Särnblad, A., Amir, O. A., Peddemors, V. M., & Karczmarski, L. (2011). Molecular ecology meets remote sensing: environmental drivers to population structure of humpback dolphins in the Western Indian Ocean. Heredity, 107(4), 349-361.
Merow, C., & Silander, J. A. (2014). A comparison of Maxlike and Maxent for modelling species distributions. Methods in Ecology and Evolution, 5(3), 215-225.
Möller, L. M., & Beheregaray, L. B. (2001). Coastal bottlenose dolphins from southeastern Australia are Tursiops aduncus according to sequences of the mitochondrial DNA control region. Marine Mammal Science, 17(2), 249-263.
Möller, L. M., & Beheregaray, L. B. (2004). Genetic evidence for sex-biased dispersal in resident bottlenose dolphins (Tursiops aduncus). Mol Ecol, 13(6), 1607-1612.
Morrison, M. L., Marcot, B., & Mannan, W. (2006). Wildlife-habitat relationships: concepts and applications. Washington, D.C: Island Press.
Movable Type Ltd. (Accessed 2014). Calculate distance, bearing and more between Latitude/Longitude points. Movable type scripts. 2014
New, L. F., Harwood, J., Thomas, L., Donovan, C., Clark, J. S., Hastie, G. D., Thompson, P. M., Cheney, B., Scott‐Hayward, L., & Lusseau, D. (2013). Modelling the biological significance of behavioural change in coastal bottlenose dolphins in response to disturbance. Functional Ecology, 27(2), 314-322.
Newsome, S. D., Clementz, M. T., & Koch, P. L. (2010). Using stable isotope biogeochemistry to study marine mammal ecology. Marine Mammal Science, 26(3), 509-572.
Nicholson, K., Bejder, L., Allen, S. J., Krützen, M., & Pollock, K. H. (2012). Abundance, survival and temporary emigration of bottlenose dolphins (Tursiops sp.) off Useless Loop in the western gulf of Shark Bay, Western Australia. Marine and Freshwater Research, 63(11), 1059-1068.
Noren, D. P., & Mocklin, J. A. (2012). Review of cetacean biopsy techniques: Factors contributing to successful sample collection and physiological and behavioral impacts. Marine Mammal Science, 28(1), 154-199.
Nowacek, D. P., Thorne, L. H., Johnston, D. W., & Tyack, P. L. (2007). Responses of cetaceans to anthropogenic noise. Mammal Review, 37(2), 81-115. doi: 10.1111/j.1365-2907.2007.00104.x
Orpin, A. R, Haig, D. W, & Woolfe, K. J. (1999). Sedimentary and foraminiferal facies in Exmouth Gulf, in arid tropical northwestern Australia. Australian Journal of Earth Sciences, 46(4), 607-621.
Palmer, C., Brooks, L., Parra, G. J., Rogers, T., Glasgow, D., & Woinarski, J. C. Z. (2014a). Estimates of abundance and apparent survival of coastal dolphins in Port Essington harbour, Northern Territory, Australia. Wildlife Research, 41(1), 35-45.
Palmer, C., Parra, G. J., Rogers, T., & Woinarski, J. (2014b). Collation and review of sightings and distribution of three coastal dolphin species in waters of the Northern Territory, Australia. Pacific Conservation Biology, 20(1), 116-125.
Panigada, S., Zanardelli, M., MacKenzie, M., Donovan, C., Mélin, F., & Hammond, P. S. (2008). Modelling habitat preferences for fin whales and striped dolphins in the Pelagos Sanctuary (Western Mediterranean Sea) with physiographic and remote sensing variables. Remote Sensing of Environment, 112(8), 3400-3412.
Parra, G. J. (2006). Resource partitioning in sympatric delphinids: Space use and habitat preferences of Australian snubfin and Indo-Pacific humpback dolphins. Journal of Animal Ecology, 75(4), 862-874.
Parra, G. J., Corkeron, P. J., & Arnold, P. (2011). Grouping and fission-fusion dynamics in Australian snubfin and Indo-Pacific humpback dolphins. Animal Behaviour, 82(6), 1423-1433.
Parra, G. J., Corkeron, P. J., & Marsh, H. (2004). The Indo-Pacific humpback dolphin, Sousa chinensis (Osbeck, 1765), in Australian waters: a summary of current knowledge. Aquatic Mammals, 30(1), 197-206.
References
105
Parra, G. J., Corkeron, P. J., & Marsh, H. (2006a). Population sizes, site fidelity and residence patterns of Australian snubfin and Indo-Pacific humpback dolphins: Implications for conservation. Biological Conservation, 129(2), 167-180.
Parra, G. J., & Jedensjö, M. (2009). Feeding habits of Australian Snubfin (Orcaella heinsohni) and Indo-Pacific humpback dolphins (Sousa chinensis) Project Report to the Great Barrier Reef Marine Park Authority, Townsville and Reef & Rainforest Research Centre Limited, Cairns.
Parra, G. J., & Jedensjö, M. (2013). Stomach contents of Australian snubfin (Orcaella heinsohni) and Indo‐Pacific humpback dolphins (Sousa chinensis). Marine Mammal Science.
Parra, G. J., & Ross, G. J. B. (2009a). Humpback dolphins: S. chinensis and S. teuszii. In W. F. Perrin, W. Bernd & J. G. M. Thewissen (Eds.), Encyclopaedia of Marine Mammals (Second Edition ed., pp. 576-582). London: Academic Press.
Parra, G. J., Schick, R., & Corkeron, P. J. (2006b). Spatial distribution and environmental correlates of Australian snubfin and Indo‐Pacific humpback dolphins. Ecography,
29(3), 396-406. Parra, G. J, & Ross, G. J. B. (2009b). Humpback Dolphins, S. chinensis and S. teuszii. In W.
Perrin, B. Würsig & J. Thewissen (Eds.), Encyclopedia of Marine Mammals, 2nd Edition (pp. 1100-1103). San Diego, California: Academic Press.
Pearce, A. F., & Feng, M. (2013). The rise and fall of the “marine heat wave” off Western Australia during the summer of 2010/2011. Journal of Marine Systems, 111–112(0), 139-156. doi: http://dx.doi.org/10.1016/j.jmarsys.2012.10.009
Penn, J. W., & Caputi, N. (1986). Spawning stock-recruitment relationships and environmental influences on the tiger prawn (Penaeus esculentus) fishery in Exmouth Gulf, Western Australia. Marine and Freshwater Research, 37(4), 491-505.
Phillips, S. J. (2005). A brief tutorial on Maxent AT&T Research. Phillips, S. J., Anderson, R. P., & Schapire, R. E. (2006). Maximum entropy modeling of
species geographic distributions. Ecological Modelling, 190(3), 231-259. Phillips, S. J., & Dudík, M. (2008). Modeling of species distributions with Maxent: new
extensions and a comprehensive evaluation. Ecography, 31(2), 161-175. Phillips, S. J., Dudík, M., Elith, J., Graham, C. H., Lehmann, A., Leathwick, J. R., & Ferrier, S.
(2009). Sample selection bias and presence-only distribution models: implications for background and pseudo-absence data. Ecological Applications, 19(1), 181-197.
Phillips, S. J., Dudík, M., & Schapire, R. E. (2004). A maximum entropy approach to species distribution modeling. Paper presented at the Proceedings of the twenty-first international conference on Machine learning.
Pirotta, E., Laesser, B. E., Hardaker, A., Riddoch, N., Marcoux, M., & Lusseau, D. (2013). Dredging displaces bottlenose dolphins from an urbanised foraging patch. Marine Pollution Bulletin, 74(1), 396-402.
Pitman, R. L., Totterdell, J. A., Fearnbach, H., Ballance, L. T., Durban, J. W., & Kemps, H. (2014). Whale killers: Prevalence and ecological implications of killer whale predation on humpback whale calves off Western Australia. Marine Mammal Science, 31(2), 629 - 657.
Pollock, K. H. (1982). A capture-recapture design robust to unequal probability of capture. The Journal of Wildlife Management, 752-757.
Pollock, K. H., Nichols, J. D., Simons, T. R., & Farnsworth, G. L. (2002). Large scale wildlife monitoring studies: statistical methods for design and analysis. Environmetrics, 13(2), 105-119. doi: 10.1002/env.514
Pollock, K. H, Marsh, H. D, Lawler, I. R, & Alldredge, M. W. (2006). Estimating animal abundance in heterogeneous environments: an application to aerial surveys for dugongs. Journal of Wildlife Management, 70(1), 255-262.
References
106
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 Research Part I: Oceanographic Research Papers, 56(4), 648-657. doi: 10.1016/j.dsr.2008.11.001
Preen, A. R., Marsh, H., Lawler, I. R., Prince, R. I. T., & Shepherd, R. (1997a). Distribution and abundance of dugongs, turtles, dolphins and other Megafauna in Shark Bay, Ningaloo Reef and Exmouth Gulf, Western Australia. Wildlife Research, 24(2), 185-208.
Prince, R. I. T., Lawler, I. R., & Marsh, H. D. (2001). The distribution and abundance of dugongs and other megavertebrates in Western Australian coastal waters extending seaward to the 20 metre isobath between North West Cape and the DeGrey River Mouth, Western Australia, April 2000. Report for Environment Australia.
R Development Core Team. (2013). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Retrieved from URL http://www.R-project.org/
Redfern, J. V., Barlow, J., Ballance, L. T., Gerrodette, T., & Becker, E. A. (2008). Absence of scale dependence in dolphin-habitat models for the eastern tropical Pacific Ocean. Marine Ecology Progress Series, 363, 1-14. doi: 10.3354/meps07495
Redfern, J. V., Ferguson, M. C., Becker, E. A., Hyrenback, K. D., Good, C., Barlow, J., Kaschner, K., Baumgartner, M. F., Forney, K. A., Balance, L. T., Fauchald, P., Halpin, P., Hamazaki, T., Pershing, A. J., Qian, S. S., Read, A., Reilly, S. B., Torres, L., & Werner, F. (2006). Techniques for cetacean-habitat modeling. Marine Ecology Progress Series, 310(271-295).
Reed, D. H., O'Grady, J. J., Brook, B. W., Ballou, J. D., & Frankham, R. (2003). Estimates of minimum viable population sizes for vertebrates and factors influencing those estimates. Biological Conservation, 113, 23-34.
Reeves, R. R., Dalebout, M. L., Jefferson, T. A., Karczmarski, L., Laidre, K., O’Corry-Crowe, G., Rojas-Bracho, L., Secchi, E. R., Slooten, E., Smith, B. D., Wang, J. Y., & Zhou, K. (2008). Sousa chinensis. Retrieved 09 August 2013, from IUCN
Robinson, L. M., Elith, J., Hobday, A. J., Pearson, R. G., Kendall, B. E., Possingham, H. P., & Richardson, A. J. (2011). Pushing the limits in marine species distribution modelling: lessons from the land present challenges and opportunities. Global Ecology and Biogeography, 20(6), 789-802. doi: 10.1111/j.1466-8238.2010.00636.x
Ross, G. J. B, Heinsohn, G. E., & Cockcroft, V. G. (1994). Humpback dolphins Sousa chinensis (Osbeck, 1765), Sousa plumbea (G. Cuvier, 1829) and Sousa teuszii (Kukenthal, 1892). In S. Ridgway & R. Harrison (Eds.), Handbook of marine mammals (Vol. 5, pp. 23-42): Academic Press, London.
Rossman, S., Barros, N. B., Ostrom, P. H., Stricker, C. A., Hohn, A. A., Gandhi, H., & Wells, R. S. (2013). Retrospective analysis of bottlenose dolphin foraging: A legacy of anthropogenic ecosystem disturbance. Marine Mammal Science, 29(4), 705-718.
Rossman, S., Berens McCabe, E., Barros, N. B., Gandhi, H., Ostrom, P. H., Stricker, C. A., & Wells, R. S. (2014). Foraging habits in a generalist predator: Sex and age influence habitat selection and resource use among bottlenose dolphins (Tursiops truncatus). Marine Mammal Science.
Sargeant, B. L., Wirsing, A. J., Heithaus, M. R., & Mann, J. (2007). Can environmental heterogeneity explain individual foraging variation in wild bottlenose dolphins (Tursiops sp.)? Behavioral Ecology and Sociobiology, 61(5), 679-688. doi: 10.1007/s00265-006-0296-8
Schmid, M., & Hothorn, T. (2008). Boosting additive models using component-wise P-splines. Computational Statistics & Data Analysis, 53(2), 298-311.
References
107
Schmid, M., Potapov, S., Pfahlberg, A., & Hothorn, T. (2010). Estimation and regularization techniques for regression models with multidimensional prediction functions. Statistics and Computing, 20(2), 139-150.
Seo, C., Thorne, J. H., Hannah, L., & Thuiller, W. (2009). Scale effects in species distribution models: implications for conservation planning under climate change. Biology Letters, 5(1), 39-43.
Sleeman, J. C., Meekan, M. G., Wilson, S. G., Jenner, C. K. S., Jenner, M. N., Boggs, G. S., Steinberg, C. C., & Bradshaw, C. J. A. (2007). Biophysical correlates of relative abundances of marine megafauna at Ningaloo Reef, Western Australia. Marine and Freshwater Research, 58(7), 608-623. doi: 10.1071/mf06213
Slooten, E., Dawson, S. M., & Rayment, W. J. (2004). Aerial surveys for coastal dolphins: abundance of Hector's dolphins off the South Island west coast, New Zealand. Marine Mammal Science, 20(3), 477-490.
Smith, H. (2012). Population dynamics and habitat use of bottlenose dolphins (Tursiops aduncus), Bunbury, Western Australia. (PhD), Murdoch University, Murdoch University.
Smith, H. C., Pollock, K. H., Waples, K., Bradley, S., & Bejder, L. (2013). Use of the robust design to estimate seasonal abundance and demographic parameters of a coastal bottlenose dolphin (Tursiops aduncus) population. PLoS ONE, 8(10), e76574.
Smith, J. N., Grantham, H. S., Gales, N., Double, M. C., Noad, M. J., & Paton, D. (2012). Identification of humpback whale breeding and calving habitat in the Great Barrier Reef. Marine Ecology: Progress Series, 447, 259-272.
Smolker, R. A., Mann, J., & Smuts, B. B. (1993). Use of signature whistles during separations and reunions by wild bottlenose dolphin mothers and infants. Behavioral Ecology and Sociobiology, 33(6), 393-402.
Smolker, R. A., Richards, A. F., Connor, R., Mann, J., & Berggren, P. (1997). Sponge carrying by dolphins (Delphinidae, Tursiops sp.): a foraging specialization involving tool use? Ethology, 103(6), 454-465.
Stensland, E., Carlén, I., Särnblad, A., Bignert, A., & Berggren, P. (2006). Population size, distribution, and behavior of Indo-Pacific bottlenose (Tursiops aduncus) and humpback (Sousa chinsensis) dolphins off the south coast of Zanzibar. Marine Mammal Science, 22(3), 667-682.
Stone, C. J., & Tasker, M. L. (2006). The effects of seismic airguns on cetaceans in UK waters. Journal of Cetacean Research and Management, 8(3), 255.
Sutherland, W. J. (1998). The importance of behavioural studies in conservation biology. Animal Behaviour, 56(4), 801-809.
Swihart, R. K., Gehring, T. M., Kolozsvary, M. B., & Nupp, T. E. (2003). Responses of ‘resistant’ vertebrates to habitat loss and fragmentation: the importance of niche breadth and range boundaries. Diversity and Distributions, 9(1), 1-18.
Taylor, B. L., Wade, P. R., De Master, D. P., & Barlow, J. (2000). Incorporating uncertainty into management models for marine mammals. Conservation Biology, 14(5), 1243-1252.
Thorne, L. H., Johnston, D. W., Urban, D. L., Tyne, J., Bejder, L., Baird, R. W., Yin, S., Rickards, S.H., Deakos, M., Mobley Jr. J. R., Pack, A. A., and Hill, M. C. (2012). Predictive modeling of spinner dolphin (Stenella longirostris) resting habitat in the main Hawaiian islands. PLoS ONE, 7(8). doi:10.1371/journal.pone.0043167
Torres, L. G., Rosel, P. E., D'Agrosa, C., & Read, A. J. (2003). Improving management of overlapping bottlenose dolphin ecotypes through spatial analysis and genetics. Marine Mammal Science, 19(3), 502-514. doi: 10.1111/j.1748-7692.2003.tb01317.x
Toth, J. L., Hohn, A. A., Able, K. W., & Gorgone, A. M. (2011). Patterns of seasonal occurrence, distribution, and site fidelity of coastal bottlenose dolphins (Tursiops
References
108
truncatus) in southern New Jersey, USA. Marine Mammal Science, 27(1), 94-110. doi: 10.1111/j.1748-7692.2010.00396.x
Tuomainen, U., & Candolin, U. (2011). Behavioural responses to human‐induced environmental change. Biological Reviews, 86(3), 640-657.
Urian, K. W., Gorgone, A. M., Read, A., Balmer, B. C., Wells, R. S., Berggren, P., Durban, J. W., Eguchi, T., Rayment, W. J., & Hammond, P. S. (2014). Recommendations for photo‐
identification methods used in capture‐recapture models with cetaceans. Marine Mammal Science.
Validakis, V. (2014). Karratha receives new status, now known as a city, Australian Mining. Van Parijs, S. M., & Corkeron, P. J. (2001a). Boat traffic affects the acoustic behaviour of
Pacific humpback dolphins, Sousa chinensis. Journal of the Marine Biological Association of the United Kingdom, 81(3), 533-538.
Van Parijs, S. M., & Corkeron, P. J. (2001b). Vocalizations and behaviour of Pacific humpback dolphins, Sousa chinensis. Ethology, 107(8), 701-716.
Viddi, F. A, Hucke-Gaete, R, Torres-Florez, J P, & Ribeiro, S. (2010). Spatial and seasonal variability in cetacean distribution in the fjords of northern Patagonia, Chile. ICES Journal of Marine Science: Journal du Conseil, 67(5), 959-970.
Wade, P. R. (1998). Calculating limits to the allowable human‐caused mortality of cetaceans and pinnipeds. Marine Mammal Science, 14(1), 1-37.
Wang, J. Y., Chou, L‐S., & White, B. N. (1999). Mitochondrial DNA analysis of sympatric morphotypes of bottlenose dolphins (genus: Tursiops) in Chinese waters. Mol Ecol, 8(10), 1603-1612.
Wang, J. Y., & Chu Yang, S. (2009). Indo-Pacific bottlenose dolphin Tursiops aduncus. In P. W. F., W. B. & T. J. G. M. (Eds.), Encyclopedia of Marine Mammals. Second ed. (pp. 602-608). San Diego: Academic Press.
WAPC. (2012). Western Australia Tomorrow, Population Report No 7, 2006 to 2026, Forecast Summary. Perth: West Australian Planning Commission (Department of Planning).
Wenziker, K., McAlpine, K., Apte, S., & Masine, R. (2006). Background quality for coastal marine waters of the North West Shelf, Western Australia North West Shelf Joint Environmental Management Study, Technical Report No 18: CSIRO Marine and Atmospheric Research.
Whiting, S. D. (2011). Shallow water foraging using a shoreline boundary by the Indo-Pacific Humpback Dolphin Sousa chinensis in northern Australia. Northern Territory Naturalist, 23(1), 39-44.
Wiley, E. O., McNyset, K. M., Peterson, A. T., Robins, C. R., & Stewart, A. M. (2003). Niche modeling and geographic range predictions in the marine environment using a machine-learning algorithm. Oceanography, 16(3), 120-127.
Williams, R., Hedley, S. L., & Hammond, P. S. (2006a). Modeling distribution and abundance of Antarctic baleen whales using ships of opportunity. Ecology and Society.
Williams, R., Lusseau, D., & Hammond, P. S. (2006b). Estimating relative energetic costs of human disturbance to killer whales (Orcinus orca). Biological Conservation, 133(3), 301-311.
Wilson, B. (2013). The biogeography of the Australian north west shelf : environmental change and life's response (1 ed.). Sydney: Elsevier.
Wilson, B., Thompson, P. M., & Hammond, P. S. (1997). Habitat use by bottlenose dolphins: seasonal distribution and stratified movement patterns in the Moray Firth, Scotland. Journal of Applied Ecology, 34(6), 1365-1374. doi: 10.2307/2405254
Wisz, M. S., Hijmans, R. J., Li, J., Peterson, A. T., Graham, C. H., & Guisan, A. (2008). Effects of sample size on the performance of species distribution models. Diversity and Distributions, 14(5), 763-773.
References
109
Wood, S. N. (2004). Stable and efficient multiple smoothing parameter estimation for generalized additive models. Journal of the American Statistical Association, 99, 673-686.
Wursig, B., & Jefferson, T. A. (1990). Methods of photo-identification for small cetaceans. Reports of the International Whaling Commission. Special(12), 42-43.
Zhou, Z.-H. (2009). Ensemble learning. In S. Z. Li (Ed.), Encyclopedia of Biometrics (pp. 270 - 273). Berlin: Springer.
Zuur, A. F., Ieno, E. N., & Elphick, C. S. (2010). A protocol for data exploration to avoid common statistical problems. Methods in Ecology and Evolution, 1(1), 3-14.
Zuur, A. F., Ieno, E. N., Walker, N. J. , Saveliev, A. A. , & Smith, G. M. (2009). Mixed effects models and extensions in ecology with R. New York: Springer.
Appendix A
110
Appendix A: Adjustments for dolphin sighting locations
The location of dolphin sightings were adjusted so that they were at a more accurate
distance from the aircraft, rather than on the transect line. The 200 m strip width
(Table A1) consisted of four 50 m zones. Sightings were repositioned according to the
zone that they were recorded in. Sightings ‘inside’ and ‘outside’ the strip width transect
were retained. The inside area is the area under the aircraft that could still be seen. The
unseen area underneath the aircraft was also accounted for.
The following steps were undertaken in order to calculate the offset value for each dolphin
sighting (Table A1):
1. Divided the unseen gap (300m) underneath the aircraft by the number of
aircraft sides (2).
2. For sightings within the 200 m strip width,
a. the distance of the further side of the zone was added,
b. the sighting was placed into the middle of the zone by subtracting half
the zone width (25 m).
3. Half the zone width (25 m) was removed for sightings ‘inside’ the transect.
4. Half the zone width (25 m) was added to the distance of the outer side of the
strip width transect (200 m +25 m).
Table A1: Offset value based on zone and removed unseen area.
Zone Calculation Offset value (m)
Inside (300/2) – 25 125 Low (300/2) + (50 - 25) 175 Medium (300/2) + (100 - 25) 225 High (300/2) + (150 - 25) 275 Very high (300/2) + (200 - 25) 325 Outside (300/2) + (200 + 25) 375
These adjusted positions assume that the dugongs were recorded at the time that they
were abeam (perpendicular) to the aircraft, as the bearing of each sighting was not
recorded. This assumption is known to not have always been true so is therefore a source
of error that has not been accounted for.
Once offset values had been calculated, the directional bearing of dolphin sightings from
the aircraft was determined. Bearings were assigned based on direction of travel and
observer seating position (Table A2).
Appendix A
111
Table A2: Assigned bearing based on direction of travel and observer position.
Exmouth Block Onslow Block
East West ‘North’ ‘South’
Port 360° 180° 240° 60° Starboard 180° 360° 60° 240°
Offset values and bearings were added as columns of data into the Microsoft Excel dolphin
sighting data spread sheets. They were used in the following formulas to calculate latitude
and longitude (Movable Type Ltd), with 6,371 referring to the mean radius of the earth (in
kilometres). These new geographical coordinates were saved for each sighting.
New latitude
ASIN (SIN(Radians of Latitude) * COS(offset value/ 6371) + COS(radians of latitude) *
SIN(offset value/ 6371) * COS(radians bearing))
New longitude
Radians of longitude + ATAN2(COS(offset value/ 6371) - SIN(radians of latitude) * SIN(New
radians of latitude), SIN(radians bearing) * SIN(offset value / 6371) * COS(radians of
latitude)
Finally, sightings were plotted in GIS using new coordinates and verified using the
measure tool, to confirm they were at correct distances from the transect line.
Appendix B
112
Appendix B: Omitted environmental variables
Chlorophyll concentrations
Chlorophyll data (Figure A1) obtained from the Commonwealth Scientific and Industrial
Research Organisation (CSIRO) was rejected due to its unreliability. It has not been
validated and is likely to include suspended sediment and detritus such as decaying
mangrove and salt flat flora material. The algorithm on which the data is based was
developed for the open ocean and has an error of 35%, which would be much greater for
the coastal waters of the study area.
Mean: mg/m3 for each grid cell; stdv: standard deviation between cells
Figure B1: December 2012 chlorophyll data for the study area.
Appendix B
113
Salinity
Salinity was available for only a portion of the study area and only at a low resolution
(approximately 11 km x 11 km grid cells) (Figure A2). It was hypothesised that salinity
could be a predictor of coastal dolphin distribution, with higher levels experienced in the
semi-enclosed embayment of Exmouth Gulf. While exploration of the data did show that
such a gradient existed, it was only of a small variation with values remaining between
34.8 – 35.8 parts per million (ppm) (Table A1). This was not deemed to be at a level of
ecological relevance for dolphins, as its still within the range for ‘normal’ euhaline marine
waters (30 – 36 ppm).
Figure B2: December 2012 salinity data for the study area.
Table B1: Salinity values for the study area at each survey period.
Survey Pattern observed Range (entire study area)
Value (ppm) Description
May 12 Inner gulf most saline 34.8 – 35.8 Euhaline July 12 Inner gulf most saline 34.5 – 35.6 Euhaline Oct 12 Inner gulf most saline 34.8 – 35.6 Euhaline Dec 12 Inner gulf most saline 34.9 – 35.7 Euhaline May 13 Inner gulf most saline 34.9 – 35.8 Euhaline
Appendix B
114
Coastal features
The coastal features dataset (Geoscience Australia, 2004) consists of mangroves, saline
coastal flats and watercourses that all have potential ecological relevance to coastal
dolphins in this study area. The series of snapshots that follow illustrate the incomplete
nature of dataset, the reason why it was not utilised in this project. The digitised
mangroves, saline coastal flats and watercourse layers were exported as Keyhole Markup
Language (KML) files and viewed in Google Earth, against high-resolution, full colour
satellite imagery (Google Earth, 2013). As depicted in these snapshots of the Ashburton
River mouth, verification revealed that there are unmapped tracts of mangroves and
watercourses along the coast (Table B2).
Figure B3: Coastal features data for the study area.
Appendix B
115
Table B2: Demonstrated incompleteness of coastal features data, using snapshots of the Ashburton River Delta area from Google Earth imagery.
Snapshot Explanation
A) Coastal features overlaid over Google Earth imagery
B) Mangroves and watercourses By removing the saline coastal flat data layer it becomes obvious that sections of mangrove, and waterways, have not been digitized.
C) Watercourses Removing the mangrove data layer enables a comparison of areas that were digitized and those that were not. There is not explanation for the discrepancy other than that they were missed or that the dataset is no longer current.
Appendix C
116
APPENDIX C: Correlation coefficients between static environmental variables.
Table C1: Spearman’s rank correlation coefficients between static environmental variables, those shaded in light grey were used in the initial MaxEnt models but not
the refined MaxEnt models.
dep sl asp toco edml edis edin* edsl*
dep sl -0.32
asp -0.29 0.55 toco -0.37 0.93 0.55
edml -0.34 0.05 0.05 0.05 edis -0.57 -0.02 0.08 0.00 0.18
edin* -0.57 -0.08 0.06 -0.07 0.40 0.75 edsl* 0.38 -0.15 -0.02 -0.17 -0.55 -0.14 -0.21
dep: depth; sl: slope; asp: aspect; toco: topographical complexity; edml: distance from mainland; edis: distance from islands; edin: distance from intertidal areas; edsl: distance from the slope at the 20 m isobath; *variables used in refined models only.
Appendix D
117
APPENDIX D: Diagnostic plots showing omission rates for MaxEnt models.
Table D1: Omission and predicted area for the final models that are a mean of 1,000 replicated models.
Model Bottlenose dolphins Humpback dolphins
Static
Autumn'
Spring'
May-12
May-13
Appendix D
118
Model Bottlenose dolphins Humpback dolphins
May-14
Oct-12
Oct-13
Jul-12
Dec-12