Modelling seasonal habitat suitability and connectivity
for feral pigs in northern Australia:
towards risk-based management of infectious animal diseases
with wildlife hosts
Jens G. Froese
M.A., Albert-Ludwigs-Universität Freiburg, Germany
A thesis submitted for the degree of Doctor of Philosophy at
The University of Queensland in 2017
School of Agriculture and Food Sciences
Abstract
Infectious animal diseases are a major biosecurity threat in an increasingly connected world.
Wildlife hosts are a well-recognized risk factor for disease introduction, establishment and spread.
Northern Australia is vulnerable to disease incursions from neighbouring countries, and widespread
invasive feral pigs (Sus scrofa) can seriously complicate post-border disease management. The aim
of this thesis was to generate new regional-scale spatial knowledge of feral pig populations in
northern Australia to inform risk-based management of directly transmitted infectious animal
diseases for which feral pigs are a host.
Due to environmental variability and empirical knowledge gaps across this vast region, I
adopted a resource-based modelling approach, based on expert knowledge but rooted in landscape
ecological theory, to answer three research questions at multiple levels of biological organisation.
Specifically, I conceptualized feral pigs in northern Australia as a metapopulation and the landscape
as displaying a patch-corridor-matrix structure. At the level of individual feral pig breeding herds, I
explored the selection of supplementary and complementary resources within home ranges. At the
level of local subpopulations, consisting of several herds with adjacent or overlapping home ranges,
I used a habitat suitability modelling approach to investigate the distribution of potential patches of
breeding habitat emerging from the interactions between resources and home range movements. At
the metapopulation level, I examined potential dispersal pathways between many such patches
using a habitat connectivity modelling approach. As feral pig movements and distributional patterns
vary with conditions, I applied models to two seasonal scenarios (wet and dry) corresponding to
northern Australia’s annual rainfall cycle.
This thesis contributed methodological advances and new ecological insights. I developed a
novel combined methodology, spatial pattern suitability analysis, for capturing feral pigs’ resourceseeking home range movements based on expert-elicited response-to-pattern curves and spatial
moving window analysis. Based on landscape ecological principles, this methodology improves the
application of resource-based Bayesian networks models to mobile animals. I found that habitat
suitability for persistent feral pig breeding in northern Australia is dependent on spatial interactions
between four key habitat requirements: water and food resources as well as protection from heat
and from disturbance. Through scenario analysis and empirical validation I showed that habitat
suitability at the regional scale is most reliably modelled as a function of distance to supplementary
and complementary resource patches. When applied to a wet season and a dry season scenario,
mapped model results indicated that the spatial distribution of feral pig habitat patches varied
markedly. Importantly, empirically validated findings suggest that dry season conditions restrict
overall habitat suitability for feral pig breeding in northern Australia more than previously thought.
i
This thesis also provides the first attempt to describe seasonal habitat connectivity in the entire
northern Australian feral pig population. By linking model assumptions to gender-specific
differences in dispersal ability, I showed that dry season connectivity between habitat patches was
limited for breeding herds, but less constrained for solitary males. Due to greater resource
abundance, wet season connectivity was greater irrespective of dispersal ability. Three broad types
of habitat patches were identified: some were always isolated; some were highly connected to form
large habitat components; and some were mostly isolated but became connected to large
components during the wet season or for wide-ranging males.
By linking the landscape ecological research perspective on feral pigs to a landscape
epidemiological perspective on directly transmitted infectious diseases, risk areas for the
establishment, spread and persistence of disease in wildlife hosts could be identified. I used the
example of classical swine fever to illustrate these links. Following introduction, successful
establishment of such a disease is contingent on locally dense host populations in patches of
breeding habitat; subsequent disease spread requires host dispersal between infected and susceptible
subpopulations; and long-term disease persistence depends on a lasting supply of susceptible
individuals within a regionally connected host metapopulation. Effective post-border disease
management should capitalize on these links. For example, early detection surveillance activities
could be targeted in habitat patches and designed so that each connected habitat component is
sampled. In the event of an incursion, patch connectivity may be used to establish containment
zones, focus population control and support delineation of disease-free compartments. Moreover,
host risk could be combined with disease-specific introduction pathways, transmission rates and
other parameters to generate deeper, dynamic insights into disease-host interactions for better
incursion preparedness.
In conclusion, the research contained in this thesis provides, for the first time, a complete and
coherent, spatially-explicit, seasonally-specific and regional-scale picture of areas most at risk of
disease establishment (via host habitat suitability) and spread (via host habitat connectivity) in feral
pigs in northern Australia. The resource-based modelling approach is transparent and flexible, and
could be applied to other invasive species and wildlife hosts of infectious animal diseases,
especially in data-constrained situations and for wide-ranging species.
ii
Declaration by author
This thesis is composed of my original work, and contains no material previously published or
written by another person except where due reference has been made in the text. I have clearly
stated the contribution by others to jointly-authored works that I have included in my thesis.
I have clearly stated the contribution of others to my thesis as a whole, including statistical
assistance, survey design, data analysis, significant technical procedures, professional editorial
advice, and any other original research work used or reported in my thesis. The content of my thesis
is the result of work I have carried out since the commencement of my research higher degree
candidature and does not include a substantial part of work that has been submitted to qualify for
the award of any other degree or diploma in any university or other tertiary institution. I have
clearly stated which parts of my thesis, if any, have been submitted to qualify for another award.
I acknowledge that an electronic copy of my thesis must be lodged with the University Library and,
subject to the policy and procedures of The University of Queensland, the thesis be made available
for research and study in accordance with the Copyright Act 1968 unless a period of embargo has
been approved by the Dean of the Graduate School.
I acknowledge that copyright of all material contained in my thesis resides with the copyright
holder(s) of that material. Where appropriate I have obtained copyright permission from the
copyright holder to reproduce material in this thesis.
iii
Publications during candidature
Peer-reviewed publications
Froese, JG, Smith, CS, Durr, PA, McAlpine, CA & van Klinken, RD (2017). "Modelling seasonal
habitat suitability for wide-ranging species: invasive wild pigs in northern Australia." PLoS
ONE: e0177018. doi: 10.1371/journal.pone.0177018
Froese, JG, Murray, JV, Perry, JJ & van Klinken, RD (2015). "Spatial modelling to estimate the risk
of feral pigs to pig farm biosecurity in south-eastern Australia." Animal Production Science
55(12): 1456.
Conference proceedings
Froese, JG, Smith, CS, Durr, PA & van Klinken, RD (2016). "Integrating landscape structure into
participatory Bayesian network models of habitat suitability." in S Sauvage, JM Sánchez-Pérez
& AE Rizzoli (eds). Proceedings of the 8th International Congress on Environmental
Modelling and Software. Toulouse: p. 838.
Froese, JG, Smith, CS, McAlpine, CA, Durr, PA & van Klinken, RD (2015). "Modelling habitat
suitability and connectivity of feral pigs for exotic disease surveillance in northern Australia."
in T Weber, MJ McPhee & RS Anderssen (eds). MODSIM2015, 21st International Congress
on Modelling and Simulation. Modelling and Simulation Society of Australia and New
Zealand, Broadbeach: p. 279.
Froese, JG, Smith, CS, McAlpine, CA, Durr, PA & van Klinken, RD (2015). "Moving window
analysis links landscape-scale resource utilization to habitat suitability models of feral pigs in
northern Australia." in T Weber, MJ McPhee & RS Anderssen (eds). MODSIM2015, 21st
International Congress on Modelling and Simulation. Modelling and Simulation Society of
Australia and New Zealand, Broadbeach: pp. 1352-1358.
Froese, J, Smith, C, Durr, PA & van Klinken, RD (2014). "Where can all the pigs be found?
Harnessing expert knowledge for the spatial modelling of feral pig distribution and abundance
in northern Australia." in MN Gentle (ed). Program and abstracts 16th Australasian
Vertebrate Pest Conference. Brisbane: p. 70.
iv
Scientific reports
Murray, J, Froese, J, Perry, J, Navarro Garcia, J & van Klinken, R (2015). Impact modelling for
rabbits and feral pigs in the QMDB. CSIRO Biosecurity Flagship, Brisbane.
Murray, JV, Froese, J, Perry, J & van Klinken, R (2015). Assessing the risk of feral pigs interacting
with domestic pig herds in southeastern Australia. Final report to Australian Pork Limited
(APL), Project 2114/486. CSIRO Biosecurity Flagship, Brisbane.
v
Publications included in this thesis
Froese, JG, Smith, CS, Durr, PA, McAlpine, CA & van Klinken, RD (2017). "Modelling seasonal
habitat suitability for wide-ranging species: invasive wild pigs in northern Australia." PLoS
ONE: e0177018. doi: 10.1371/journal.pone.0177018 – incorporated as Chapter 4
Contributor
Statement of contribution (PPL 4.20.04 Authorship)
Jens Froese (Candidate)
Conceived idea and designed methodology (40%)
Collected data (90%)
Analysed data (100%)
Wrote the paper (80%)
Carl Smith
Conceived idea and designed methodology (20%)
Wrote the paper (5%)
Clive McAlpine
Conceived idea and designed methodology (10%)
Wrote the paper (5%)
Peter Durr
Conceived idea and designed methodology (10%)
Collected data (10%)
Wrote the paper (5%)
Rieks van Klinken
Conceived idea and designed methodology (20%)
Wrote the paper (5%)
vi
Contributions by others to the thesis
A panel of feral pig experts not only helped parameterize the models contained in this thesis, they
also fundamentally shaped the research design and methodology by highlighting the need to take a
landscape ecological view on feral pigs at the outset of the project. Individual experts are
acknowledged by name in the following section. Human ethics approval to conduct expert
elicitation was obtained from the CSIRO Human Ethics Committee and the UQ School of
Agriculture and Food Sciences Ethics Committee (Appendix 1). Justin Perry (CSIRO) contributed R
scripts and helped with access to CSIRO high performance computing facilities needed for some of
the spatial analyses. Justine Murray (CSIRO) contributed expert elicitation methodologies on which
some of the data collection was based and co-facilitated an expert workshop.
Statement of parts of the thesis submitted to qualify for the award of another degree
None
vii
Acknowledgements
Embarking on a PhD is a time of great enthusiasm and anticipation of what lies ahead. Many
experienced advisors – university representatives, supervisors, colleagues and fellow PhD students
– offer much needed guidance about the inevitable challenges in such a long-term project and the
bumpy path that lies ahead in the journey towards completion. I have now been on this path for the
best part of four years. At times, it has been cruisy. At times, it has been exhilarating. At other
times, it has been boring. A few times it has been so dreadfully bumpy that I wanted to turn back.
On the journey I have learned a lot about myself both as an aspiring scientist and as a person. Most
of all, I have learned that there are a lot of people travelling along with you. You may be driving the
car, but others are reading the map, paying for the fuel and providing you with company. I have
enjoyed the support of many wonderful people, while being conscious that I have given back little. I
can only offer my heartfelt gratitude and hope to repay the debts in the personal and professional
life that lies ahead.
Ahead of all, I would like to deeply thank my family: my wonderful wife Kieran for allowing
me to focus on my work when it was most needed while keeping our children alive, advancing her
own career as the best midwife in Brisbane and looking fine at the same time; and my beautiful
daughters Flora, Roxy and Juno for making parenting look easy. I love you all and forever will.
Completion would not have been possible without my outstanding and generous advisory team.
Carl Smith, my principal advisor, has a door that is literally always open! Best of all, whatever the
issue – a technical glitch, methodological question or existential crisis – he always seemed to have
an answer or idea on how to tackle it. Rieks van Klinken was the ‘conceptual engine’ behind my
research, always reminding me of the broader picture when I got bogged down yet again in some
minor detail. He also provided me with invaluable opportunities outside the thesis in his research
team at CSIRO. Peter Durr and Clive McAlpine had a somewhat smaller role in this team, but I was
continuously surprised how much effort both were willing to invest in providing advice, revising
manuscripts, and even attending expert workshops, even though we saw each other so little. I
sincerely thank you all, Carl, Rieks, Peter and Clive, for your support and guidance.
Next, I must offer my deepest gratitude to a panel of feral pig experts, who kindly donated a
significant amount of time to share their knowledge and without whom this research would not have
been possible. Thank you Jim Mitchell (FeralFix), Justin Perry (CSIRO), Col Dollery (Queensland
Parks and Wildlife Service), Andrew Hartwig (Cape York Landcare), Scott Middleton (Gulf
Catchments Biosecurity & Agribusiness Innovation), Tim Kerlin (Northern Australia Quarantine
Strategy), Peter Caley and Cameron Fletcher (both CSIRO), Cassandra Wittwer and Joe Schmidt
(both Northern Australia Quarantine Strategy), Les Harrigan (Rinyirru Land Trust), Glen Sheppard
viii
(Cook Shire Council), Bart Dryden and Jamie Livingston (Terrain Natural Resource Management),
Travis Sydes (Far North Queensland Regional Organisation of Councils), Dave Berman, Darren
Marshall and Vanessa MacDonald (Queensland Murray Darling Committee). I also thank the
Balkanu Cape York Development Corporation, Queensland Parks and Wildlife Service, Northern
Australia Quarantine Strategy and Atlas of Living Australia who shared data used for model
validation, and the Queensland Murray Darling Committee for employing my services on a related
project. It has been great fun and wonderfully rewarding working with you guys.
Further thanks go to my colleagues and fellow researchers: Justin Perry, Russell Richards and
Miguel Angel Villamil Castro helped with R analyses. Justine Murray helped with expert
elicitation. Justin Perry, Justine Murray, Javi Navarro, Sam Nicol and Hawthorne Beyer kindly
reviewed my manuscripts. Tony Pople agreed to chair all my milestone review meetings and Bill
Ellis and Shu Fukai offered their feedback on one milestone occasion each. Last but not least, Grant
Hamilton from the Quantitative Applied Spatial Ecology Group at the Queensland University of
Technology, thank you for trusting in my ability to embark on a full-time position as Research
Associate while working towards completion of my PhD thesis, and for giving me the space that I
needed in the final weeks.
Finally, I would like to thank all funding bodies who supported my research, foremost The
University of Queensland and CSIRO, who topped up an Australian Postgraduate Award with a
generously resourced Integrated Natural Resource Management scholarship. Additional support in
attending two amazing Australian Pig Science Association Conferences in Melbourne was provided
by Australian Pork Limited – thank you for showing such interest in my work.
P.S. And thank Sus scrofa for being such an awesome, interesting study species! I know you
shouldn’t be here in the Australian environment. I also hate to say that my research may contribute,
in a small way, to the effective killing of many of you. Although I haven’t handled you, and only
seen a few, after years of computer modelling I feel that I know you intimately! That is largely
thanks to the many knowledgeable scientists who have studied you before me and the experienced
practitioners who continue to handle you in the wild. For these insights I am deeply grateful.
ix
Keywords
Sus scrofa, landscape ecology, infectious animal disease, wildlife host, habitat suitability, habitat
connectivity, Bayesian network, Circuitscape, expert elicitation, spatial analysis, northern Australia
Australian and New Zealand Standard Research Classifications (ANZSRC)
ANZSRC code: 050103, Invasive Species Ecology, 25%
ANZSRC code: 050104, Landscape Ecology, 50%
ANZSRC code: 070205, Animal Protection (Pests and Pathogens), 25%
Fields of Research (FoR) Classification
FoR code: 0501, Ecological Applications, 75%
FoR code: 0702, Animal Production, 25%
x
Table of Contents
Abstract ................................................................................................................................................ i
Table of Contents .............................................................................................................................. xi
List of Figures .................................................................................................................................. xvi
List of Tables .................................................................................................................................. xvii
List of Abbreviations ....................................................................................................................xviii
Chapter 1 Introduction................................................................................................................... 19
1.1
Research problem ................................................................................................................... 19
1.1.1 Biosecurity threats from infectious animal diseases with wildlife hosts..................... 19
1.1.2 Infectious animal diseases in feral pigs ....................................................................... 20
1.1.3 Risk-based disease management in northern Australia ............................................... 21
1.1.4 A landscape ecological and epidemiological research perspective ............................. 22
1.2
Research aim ........................................................................................................................... 23
1.3
Research questions.................................................................................................................. 23
1.4
Study region ............................................................................................................................ 23
1.5
Research approach .................................................................................................................. 25
1.6
Thesis structure ....................................................................................................................... 27
1.7
Research ethics approval ........................................................................................................ 29
Chapter 2 Literature review .......................................................................................................... 30
Summary ............................................................................................................................................ 30
2.1
Theoretical concepts in landscape ecology............................................................................. 30
2.1.1 Habitat suitability ........................................................................................................ 30
2.1.2 Patch-corridor-matrix model and metapopulation ecology ......................................... 32
2.1.3 Scaling in space and time ............................................................................................ 33
2.1.4 Landscape heterogeneity and home ranges ................................................................. 35
2.1.5 Habitat connectivity .................................................................................................... 36
2.2
Landscape ecology of feral pigs ............................................................................................. 37
2.2.1 Distribution and abundance ......................................................................................... 37
2.2.2 Social organisation and reproduction .......................................................................... 38
2.2.3 Resource selection and home range movements ......................................................... 39
xi
2.2.4 Metapopulation structure and dispersal ....................................................................... 40
2.3
Landscape epidemiology of infectious diseases in feral pigs ................................................. 41
2.4
Habitat suitability models ....................................................................................................... 43
2.4.1 Modelling frameworks ................................................................................................ 43
2.4.2 Bayesian networks ....................................................................................................... 46
2.4.3 Landscape variables .................................................................................................... 47
2.5
Habitat connectivity models ................................................................................................... 49
2.5.1 Modelling frameworks ................................................................................................ 49
2.5.2 Habitat patches ............................................................................................................ 50
2.5.3 Matrix resistance ......................................................................................................... 51
2.6
Expert elicitation..................................................................................................................... 52
Chapter 3 Integrating landscape structure improves habitat models of mobile animals: feral
pigs in Australia ............................................................................................................................... 54
Summary ............................................................................................................................................ 54
3.1
Abstract ................................................................................................................................... 54
3.2
Introduction............................................................................................................................. 55
3.3
Methods .................................................................................................................................. 58
3.3.1 Study region ................................................................................................................ 58
3.3.2 Modelling approach ..................................................................................................... 59
3.3.3 Resource quality models ............................................................................................. 61
3.3.4 Spatial pattern suitability analysis ............................................................................... 61
3.3.5 Habitat suitability model ............................................................................................. 63
3.3.6 Model evaluation and validation ................................................................................. 64
3.4
Results..................................................................................................................................... 65
3.4.1 Spatial pattern suitability analysis ............................................................................... 65
3.4.2 Model evaluation and validation ................................................................................. 66
3.5
Discussion ............................................................................................................................... 70
3.5.1 Ecological significance................................................................................................ 70
3.5.2 Benefits of the modelling approach............................................................................. 72
3.5.3 Limitations................................................................................................................... 73
3.5.4 Conclusion ................................................................................................................... 74
3.6
Appendices ............................................................................................................................. 74
xii
Chapter 4 Modelling seasonal habitat suitability for wide-ranging species: invasive feral pigs
in northern Australia ....................................................................................................................... 75
Summary ............................................................................................................................................ 75
4.1
Abstract ................................................................................................................................... 75
4.2
Introduction............................................................................................................................. 76
4.3
Materials and methods ............................................................................................................ 77
4.3.1 Study region ................................................................................................................ 77
4.3.2 Habitat suitability model ............................................................................................. 78
4.3.3 Model evaluation and validation ................................................................................. 88
4.4
Results..................................................................................................................................... 91
4.4.1 Sensitivity analysis ...................................................................................................... 91
4.4.2 Predictive performance................................................................................................ 91
4.4.3 Seasonal habitat suitability .......................................................................................... 94
4.5
Discussion ............................................................................................................................... 96
4.5.1 Seasonal habitat suitability .......................................................................................... 97
4.5.2 Model evaluation and validation ................................................................................. 99
4.5.3 Management implications ......................................................................................... 100
4.6
Appendices ........................................................................................................................... 101
Chapter 5 Modelling habitat connectivity for biosecurity: the risk of infectious disease spread
in feral pigs in northern Australia ................................................................................................ 102
Summary .......................................................................................................................................... 102
5.1
Abstract ................................................................................................................................. 102
5.2
Introduction........................................................................................................................... 103
5.3
Materials and methods .......................................................................................................... 105
5.3.1 Study region and species ........................................................................................... 105
5.3.2 Habitat connectivity model ....................................................................................... 105
5.3.3 Model evaluation ....................................................................................................... 110
5.4
Results................................................................................................................................... 111
5.4.1 Matrix connectivity ................................................................................................... 111
5.4.2 Patch connectivity ..................................................................................................... 113
5.5
Discussion ............................................................................................................................. 118
xiii
5.5.1 Matrix connectivity ................................................................................................... 118
5.5.2 Patch connectivity ..................................................................................................... 119
5.5.3 Implications for disease management ....................................................................... 121
5.6
Appendices ........................................................................................................................... 122
Chapter 6 General discussion and conclusions .......................................................................... 123
Summary .......................................................................................................................................... 123
6.1
Research question 1 – resource selection by individual breeding herds .............................. 123
6.1.1 Main findings ............................................................................................................ 123
6.1.2 Significance and advances......................................................................................... 124
6.2
Research question 2 – seasonal habitat patches for subpopulations ..................................... 125
6.2.1 Main findings ............................................................................................................ 126
6.2.2 Significance and advances......................................................................................... 127
6.3
Research question 3 – seasonal patch connectivity for metapopulations ............................. 128
6.3.1 Main findings ............................................................................................................ 128
6.3.2 Significance and advances......................................................................................... 129
6.4
Synthesis and applications .................................................................................................... 130
6.4.1 Contributions to the thesis aim .................................................................................. 130
6.4.2 Applications to risk-based disease management ....................................................... 133
6.5
Limitations and future research ............................................................................................ 136
6.5.1 Empirical research ..................................................................................................... 136
6.5.2 Improved habitat suitability and connectivity models .............................................. 138
6.5.3 Spatiotemporal dynamics .......................................................................................... 140
6.5.4 Optimization models ................................................................................................. 141
6.5.5 Other study systems and applications ....................................................................... 142
References ....................................................................................................................................... 143
xiv
Appendices ...................................................................................................................................... 165
Appendix 1 Human ethics approval ................................................................................................ 165
Appendix 3.1 Model variables – parameters .................................................................................. 167
Appendix 3.2 Methodology – spatial pattern suitability analysis ................................................... 170
Appendix 3.3 Methodology – validation ........................................................................................ 180
Appendix 3.4 Validation plots – model scenarios .......................................................................... 190
Appendix 4.1 Model variables – parameters (supplements Fig 4.2) .............................................. 191
Appendix 4.2 Model variables – spatial data (supplements Table 4.2) .......................................... 209
Appendix 4.3 Methodology – spatial pattern suitability analysis ................................................... 214
Appendix 4.4 Methodology – validation ........................................................................................ 220
Appendix 4.5 Validation maps – seasonal habitat suitability ......................................................... 229
Appendix 4.6 Additional analyses – seasonal habitat suitability .................................................... 231
Appendix 5.1 Methodology – omnidirectional current density ...................................................... 233
Appendix 5.2 Methodology – patch connectivity ........................................................................... 239
Appendix 5.3 Analysis – tests of significance ................................................................................ 241
xv
List of Figures
Fig 1.1 Study region in northern Australia.................................................................................... 24
Fig 1.2 Schematic illustration of the multi-level research approach. .......................................... 26
Fig 1.3 Thesis structure.. ................................................................................................................. 29
Fig 2.1 The link between a mobile species’ behavioural levels and corresponding spatial and
temporal scales of analysis. ............................................................................................................. 34
Fig 2.2 Approaches to modelling species distributions, or ‘habitat’, in geographical and
environmental space. ....................................................................................................................... 43
Fig 3.1 Resource-seeking home range movements of mobile animals. ........................................ 56
Fig 3.2 Study region in north-eastern Australia............................................................................ 58
Fig 3.3 Methodology for modelling habitat suitability for feral pig breeding. ........................... 59
Fig 3.4 Computing resource suitability indices using spatial pattern suitability analysis. ....... 63
Fig 3.5 Averaged expert-elicited response-to-pattern curves relating structural patterns of
resource quality to functional suitability indices. ......................................................................... 66
Fig 3.6 Predicted-to-expected (P/E) ratio and corresponding habitat suitability maps for three
alternative models in the Lakefield area. ....................................................................................... 69
Fig 4.1 Study region in northern Australia.................................................................................... 78
Fig 4.2 Feral pig habitat suitability model. .................................................................................... 79
Fig 4.3 Expert-elicited resource suitability in response to distance. ........................................... 88
Fig 4.4 Sensitivity of habitat suitability to four habitat variables and expert opinion. ............. 91
Fig 4.5 Validation plots for individual expert and averaged seasonal habitat suitability
models. ............................................................................................................................................... 93
Fig 4.6 Seasonal habitat suitability for feral pig breeding in northern Australia...................... 94
Fig 4.7 Share of modelled suitable habitat found in different vegetation types and land use
classes for each seasonal scenario. .................................................................................................. 96
Fig 5.1 Study region in northern Australia.................................................................................. 105
Fig 5.2 Methodology for modelling habitat connectivity for feral pigs. .................................... 107
Fig 5.3 Validation plots for seasonal habitat suitability models. ............................................... 109
xvi
Fig 5.4 Seasonal matrix connectivity in northern Australia (for moderate ‘matrix resistance
(ResDisp)’). ...................................................................................................................................... 112
Fig 5.5 Seasonal patch connectivity for feral pigs in northern Australia. ................................ 115
Fig 5.6 Interactions between ‘corridor delineation threshold (CDThresh)’ and ‘inter-patch
dispersal distance (DispDist)’ at each level of ‘matrix resistance (ResDisp)’ for the dry (a-c)
and wet (d-f) season........................................................................................................................ 115
Fig 5.7 Seasonal patch connectivity for two selected habitat patches of feral pigs in northern
Australia.......................................................................................................................................... 118
List of Tables
Table 2.1 Sources of bias in expert elicitation and strategies or techniques for minimizing
them. .................................................................................................................................................. 53
Table 3.1 Validation data sets with ancillary information. .......................................................... 64
Table 3.2 Performance of the eleven habitat suitability models against three validation data
sets. .................................................................................................................................................... 67
Table 4.1 Bayesian network model variables and their states, with definitions......................... 84
Table 4.2 Spatial data proxies linked to model explanatory variables and methods for
reclassifying data attributes into state-specific categories. .......................................................... 86
Table 4.3 Validation data sets with ancillary information. .......................................................... 90
Table 4.4 Validation metrics for individual expert and averaged seasonal habitat suitability
models. ............................................................................................................................................... 92
Table 4.5 Amount of feral pig habitat in each habitat suitability index class per seasonal
scenario. ............................................................................................................................................ 95
Table 5.1 Amount of dispersal habitat in the study region. ....................................................... 113
xvii
List of Abbreviations
Abbreviation
Full term
ALA
Atlas of Living Australia
ALUM
Australian land use and management classification
AUSVETPLAN
Australian Veterinary Emergency Plan
BN
Bayesian network
CBI
Continuous Boyce Index
CPT
Conditional probability table
CSF
Classical Swine Fever
CSIRO
Commonwealth Scientific and Industrial Research Organisation
DOI
Digital object identifier
ESRI
Environmental Systems Research Institute
FGDB
File geodatabase
GIS
Geographic information system
HSI
Habitat suitability index
MVG / MVS
Present major vegetation groups / subgroups
NAQS
Northern Australia Quarantine Strategy
OCD
Omnidirectional current density
P/E ratio
Predicted-to-expected ratio
SDM
Species distribution model
UQ
The University of Queensland
xviii
Chapter 1 Introduction
1.1 Research problem
1.1.1 Biosecurity threats from infectious animal diseases with wildlife hosts
Infectious animal diseases are a major biosecurity 1 threat in an increasingly connected world.
Disease epidemics may seriously impact human, livestock, and wildlife health and incur large costs
on industry and government. Many infectious diseases exist in a host-pathogen continuum
involving wildlife, domestic animals, and humans (Daszak et al. 2000). Over 60% of newly
reported infectious diseases in humans between 1940 and 2004 were zoonotic, and most of these
originated in wildlife (Jones et al. 2008). For example, bats are important reservoir hosts of several
dangerous emerging pathogens such as the rabies, Nipah or Hendra viruses (Calisher et al. 2006).
Migratory birds may contribute to global spread of highly pathogenic avian influenza H5N1, with
devastating consequences to human health and the poultry industry (Olsen et al. 2006; Altizer et al.
2011). Where humans are not implicated in the host-pathogen continuum, emerging animal diseases
may have significant economic consequences, both directly on production and indirectly on market
access and trade. For example, 6.5 million animals were slaughtered in the United Kingdom in
response to an outbreak of foot-and-mouth disease in 2001 and combined costs to industry and
government were estimated at ₤8 billion (National Audit Office 2002).
Australia is currently free from many of these ‘emergency animal diseases’ (Emergency
Animal Disease Response Agreement 2012). Disease-free status has been attributed to both
geographical isolation and a well-functioning, multi-layered biosecurity system (Beale et al. 2008).
This is highly beneficial to Australia’s agricultural industries in terms of livestock welfare,
production advantages and access to competitive domestic and international markets (East et al.
2013; Brookes et al. 2014). However, freedom from disease cannot be guaranteed. Increasing
transboundary movements of humans, animals and animal products carry unavoidable risks of
disease introduction and necessitate effective management of incursions (Beale et al. 2008).
Wildlife hosts may contribute to disease emergence, establishment, spread and persistence and are
recognized as an important focus of risk-based disease management (Daszak et al. 2000; Reisen
2010). In Australia, invasive wildlife poses a particular threat, introducing disease to susceptible
domestic herds of the same species or complicating incursion management in widespread,
uncontrolled and expanding wild populations (Animal Health Australia 2011).
1
Usage of term ‘biosecurity’ varies considerably between disciplines and regions. Here, I adopt an Australian definition
of biosecurity as “the protection of the economy, environment and human health from the negative impacts associated
with entry, establishment or spread of exotic pests (including weeds) and diseases” (Beale et al. 2008, p. 1).
19
1.1.2 Infectious animal diseases in feral pigs
Feral pigs (Sus scrofa) are one of Australia’s worst invaders and among the continent’s most
abundant and widespread terrestrial mammals (West 2008). The species is a susceptible host to
numerous endemic and exotic animal diseases (Animal Health Australia 2011). I use one of these,
classical swine fever (CSF), to illustrate the role of feral pigs as a risk factor and the impacts of an
incursion. CSF is a highly contagious directly transmitted disease, whose potential introduction is of
major concern to the Australian pork industry (Brookes et al. 2014). CSF is classified as an
emergency animal disease in Australia (Emergency Animal Disease Response Agreement 2012),
and as a notifiable disease internationally (World Organisation for Animal Health 2016). While
Australia is currently free from CSF, the disease is occurring throughout the world, including
neighbouring countries in South East Asia (Animal Health Australia 2012; Leslie et al. 2014).
CSF is caused by a virus of the family Flaviviridae and affects only S. scrofa. The disease is
highly contagious in both domestic and feral pigs and can spread by direct contact between live
animals as well as ingestion of contaminated animal products or transfer of contaminated equipment
(Animal Health Australia 2012; World Organisation for Animal Health 2016). Mortality rates and
clinical signs are highly variable: acute forms have an infective period of two weeks and cause high
mortality and morbidity; low virulent strains, however, may cause chronic disease and prolonged
infectiousness with few clinical signs, enabling lasting disease persistence within a susceptible host
population (Kramer-Schadt et al. 2007; Animal Health Australia 2012).
Large outbreaks of CSF have occurred in Europe in the 1990s, resulting, for example, in the
culling of 11 million pigs and economic losses of over US$2 billion in the Netherlands (Artois et al.
2002). An epidemiological analysis of 327 CSF outbreaks in Germany during the 1990s has
identified infected wild boar as the source of 60% of all primary outbreaks (n = 93) (Fritzemeier et
al. 2000). In Australia, the risks associated with feral pigs during a CSF incursion could pertain to
both disease transmission into domestic herds or establishment within wild populations (Animal
Health Australia 2012). Pearson (2012) found that the risk of exposure by domestic herds to directly
transmitted diseases carried by surrounding feral herds is low but not negligible. However, such a
disease may readily establish and spread within widespread feral pig populations themselves
(Animal Health Australia 2012; Cowled et al. 2012). Then, international obligations would require
comprehensive surveillance of feral pigs, additional measures for separating feral from domestic
herds and active control of the disease in feral pigs through containment, vaccination and culling
(Animal Health Australia 2012; World Organisation for Animal Health 2016). Ultimately, these
combined measures are aimed at demonstrating freedom from disease – across the country, or
within epidemiological ‘zones’ or ‘compartments’ (referring to animal populations that are
20
functionally separated by geographical or management boundaries and can be attributed with a
distinct health status) – so that the safety of trade can be ensured (Scott et al. 2006).
1.1.3 Risk-based disease management in northern Australia
Tropical northern Australia is particularly vulnerable to disease introduction from neighbouring
countries due to its vast coastline, remoteness and traditional transboundary movements (Australian
National Audit Office 2012). For CSF, the most likely pathways of introduction to the North are
illegal importation of infected live animals or animal products from commercial shipping, foreign
fishing or yachting vessels. For other infectious animal diseases, important entry risk factors are
wind-borne insects, migratory birds or contaminated fomites (Cookson et al. 2012; East et al.
2013). Northern Australia also contains large and widespread feral pig populations. Its wetland and
floodplain habitats hold among the highest known local feral pig densities in Australia (Choquenot
et al. 1996). There is a particular need for effective post-border disease management (i.e. actions
aimed at actually or potentially introduced pathogens) in feral pigs at all stages of invasion,
including early detection surveillance, preparedness and incursion response (Beale et al. 2008).
Early detection surveillance is currently managed by the Northern Australia Quarantine
Strategy (NAQS) program (Australian National Audit Office 2012). Surveillance activities are
prioritised according to a risk-based framework that considers the likelihood of disease occurrence
across broad risk areas by combining the risks of introduction, establishment and spread (Cookson
et al. 2012; East et al. 2013). For CSF, surveillance predominantly involves aerial surveys and
representative sampling of feral pigs in high risk areas (Cookson pers. comm.). Incursion
preparedness in Australia has been boosted through the Wildlife Exotic Disease Preparedness
Program, which funded research on disease modelling and diagnostic tools, host ecology, disease
epidemiology and novel techniques for population control between 1999 and 2007 (Henderson
2008). Numerous simulation models have investigated the dynamics of potential disease epidemics,
including CSF, specifically in feral pigs and given important insights on their management (Pech &
Hone 1988; Hone et al. 1992; Doran & Laffan 2005; Milne et al. 2008; Cowled et al. 2012; Leslie
et al. 2014). Response to incursions of emergency animal diseases in wildlife hosts is guided by the
Emergency Animal Disease Response Agreement (2012) and the Australian Veterinary Emergency
Plan (AUSVETPLAN) disease strategies (e.g. for CSF; Animal Health Australia 2012) as well as the
Wild Animal Response Strategy (Animal Health Australia 2011). Response involves a range of
measures including: collecting information about susceptible wild animal populations, carrying out
disease surveillance, preventing spread and containing the disease, controlling susceptible
populations to enable disease fadeout via culling or vaccination, and demonstrating freedom from
disease (Animal Health Australia 2011 & 2012).
21
Across all elements of this multi-layered post-border disease management system, the
importance of feral pig distribution, abundance, dispersal as well as seasonal and behavioural
factors for the establishment and spread of infectious diseases has often been highlighted (Cowled
& Garner 2008; Animal Health Australia 2011). Yet, spatially-explicit, regional-scale 2 knowledge
of feral pigs in northern Australia, which duly accounts for seasonal and ecological effects, remains
limited. In this thesis I aimed to fill this knowledge gap.
1.1.4 A landscape ecological and epidemiological research perspective
I adopted a landscape epidemiological perspective on infectious diseases and a landscape
ecological perspective on wildlife hosts. Such an ‘integrated landscape approach’ has been widely
applied in conservation biology but is underutilized in biosecurity research (Glen et al. 2013). In
this framework, I viewed feral pigs in northern Australia as a metapopulation, with local
subpopulations in patches of suitable breeding habitat that are regionally connected by dispersal
corridors 3 (Forman 1995; Hanski 1998; Hastings 2014). A metapopulation perspective on feral pigs
in Australia is supported by landscape-genetic evidence of population structuring, especially in
heterogeneous landscapes (Hampton et al. 2004; Cowled et al. 2008; Lopez et al. 2014).
This metapopulation structure also affects the risk of invasion and persistence by directly
transmitted diseases such as CSF (Anderson et al. 1986; Riley 2007). Following introduction,
successful disease establishment is contingent on locally dense breeding populations with high
contact rates that let an infectious pathogen’s basic reproductive rate R0 exceed 1. Subsequent
disease spread requires feral pig dispersal between infected and susceptible subpopulations. Disease
persistence depends on a lasting supply of susceptible individuals within a regionally connected
feral pig metapopulation (Anderson et al. 1986; Kramer-Schadt et al. 2007). The distinction
between rapid local establishment and slower regional disease spread (Cowled & Garner 2008) is
supported by Zanardi et al.‘s (2003) description of two distinct CSF outbreaks in adjacent wild boar
subpopulations separated by unfavourable habitat in Italy, or Hampton et al.’s (2006) analysis of
infectious waterborne pathogens excreted by feral pigs in Western Australia. Simulation models of
disease dynamics in feral pigs have further suggested that seasonal and gender-specific differences
Here, ‘regional-scale’ refers to a geographical area that is consistent with the total extent of feral pigs’, or another
large mammal species’ natural range (i.e. thousands to millions of km2). Consistent with the usage in the landscape
ecological literature (e.g. Mayor et al. 2009; Yackulic et al. 2016, see also Fig 2.1), this is larger than the landscape
scale (i.e. tens to hundreds of km2, consistent with the extent of a local population’s range), but smaller than the global
scale (i.e. millions to hundreds of millions of km2, exceeding the extent of most species’ ranges).
3
Here, ‘patches’ or ‘habitat patches’ refer to spatially contiguous habitat areas of sufficient quality and size to allow for
feral pig breeding. A habitat patch is inhabited by a subpopulation that persists locally independent of other
subpopulations within the broader metapopulation. For further explanation of these concepts see Sections 2.1 and 2.2.
2
22
in feral pig habitat use, home range movements and dispersal ability may affect disease
establishment and spread (Doran & Laffan 2005; Milne et al. 2008).
1.2 Research aim
The aim of this thesis was to generate spatially-explicit, seasonally-specific and regional-scale
knowledge of feral pig populations in northern Australia to inform risk-based management of
directly transmitted infectious animal diseases with feral pigs as a wildlife host.
1.3 Research questions
To achieve this aim, I address three specific research questions through spatially-explicit,
resource-based models at multiple biological levels.
Research question 1
How can habitat suitability for mobile species such as feral pigs be reliably modelled at the
regional scale, given uncertainty about the processes of habitat selection, in particular the resourceseeking home range movements by individual breeding herds? Answering this question will provide
a suitable methodology for investigating research question 2.
Research question 2
How are patches of suitable feral pig breeding habitat that potentially support locally dense
populations distributed in northern Australia, and is their spatial distribution influenced by seasonal
effects? Answering this question will provide an indicative measure of the seasonal establishment
risk of directly transmitted infectious diseases within feral pigs.
Research question 3
How are patches of suitable feral pig breeding habitat connected by potential dispersal
pathways to form a regionally connected metapopulation, and is their connectivity influenced by
seasonal or gender-specific effects? Answering this question will provide an indicative measure of
the seasonal spread risk of directly transmitted infectious diseases within feral pigs.
1.4 Study region
Our study region extended across tropical Australia north of the Tropic of Capricorn (Fig 1.1).
Similar to a previous study by Cowled et al. (2009), I included all Australian bioregions and
subregions (Department of the Environment 2012) that are fully or partly within agro-climatic
zones I (hot, seasonally wet / dry), J (hot, wet) or H (semi-arid transition to desert climates)
(Hutchinson et al. 2005). The study region encompassed 1.76 million km2 across three jurisdictions:
Queensland, the Northern Territory and Western Australia (Fig 1.1). Northern Australia has a
23
tropical climate, alternating between a wet season (November / December to April / May) and a dry
season (May / June to October / November). Rainfall and primary productivity broadly decline
along a north-south, and to a lesser extent an east-west, gradient (Guerschman et al. 2009; Jones et
al. 2009). Monsoonal savanna woodlands and semi-arid grasslands that are extensively utilized for
cattle grazing are interspersed with riverine channels, coastal wetlands and rainforest fragments
(Fox et al. 2001). Low-lying grasslands are seasonally inundated, transforming into vast floodplains
during the wet season. Intensive uses and population centres are concentrated in fertile coastal
lowlands. Drier inland areas are among the most sparsely populated on earth (Cowled et al. 2009).
Fig 1.1 Study region in northern Australia.
All of the study region appears climatically suitable for feral pigs and has mostly been invaded.
Arid desert regions with insufficient rainfall were not included in this study. Feral pigs are reported
to be widespread in the east and localised in the north and west, where they are expected to expand
their range (Cowled et al. 2009; West 2008). A range of management activities are conducted
throughout northern Australia to mitigate feral pig impacts, including lethal (aerial shooting, poison
baiting, trapping) and non-lethal (exclusion fencing) methods. Yet, efforts are hampered by the
region’s remoteness and inaccessibility. While continuous management has effectively reduced
local densities, there is little evidence of sustained population reduction (Bengsen et al. 2014).
24
1.5 Research approach
The applied aim of this thesis was to enable targeted, risk-based management of infectious
animal diseases with feral pigs as a wildlife host by generating new spatially-explicit knowledge of
the host population. As feral pigs are widespread throughout much of the study region in northern
Australia (West 2008), and an incursion of a disease such as CSF may occur at many different
locations, especially along the coastline (Cookson et al. 2012), effective post-border disease
management requires information at the regional scale. Further, feral pig populations strongly
respond to variability in environmental conditions (Choquenot et al. 1996), which in turn will affect
disease establishment and spread (Milne et al. 2008). Given northern Australia’s annual rainfall
cycle, host information should therefore be seasonally-specific. Unfortunately, continuous empirical
data is rarely available over large areas and for multiple temporal scenarios (Stephens et al. 2015).
This gap in knowledge is also true for feral pigs in northern Australia (Cowled & Garner 2008).
Collecting detailed information through field studies was out of the scope of this thesis.
Instead, I adopted a modelling approach, rooted in landscape ecological theory, to answer three
research questions at multiple biological levels in space and time, whereby ‘level’ refers to the
hierarchically structured organisation or observed response of a species (McGarigal et al. 2016, for
details see Section 2.1 and Section 2.2). Here I focused on three organisational levels (Fig 1.2): At
the level of individual feral pig breeding herds, I explored the selection of resources within home
ranges. At the level of local subpopulations (i.e. many herds), I used a habitat suitability modelling
approach to investigate the distribution of potential patches of breeding habitat emerging from the
interactions between resources and home range movements. At the metapopulation level, I
examined potential dispersal pathways between such patches using a habitat connectivity modelling
approach. Similar combinations of habitat suitability and connectivity models have been used to
study mobile species’ use of entire landscapes (Cianfrani et al. 2013; Dickson et al. 2013; StewartKoster et al. 2015).
25
Fig 1.2 Schematic illustration of the multi-level research approach. At the local subpopulation level, the
distribution of suitable habitat patches (panels B, red polygons) depends on the selection of resources (shown
as coloured patches) by individual breeding herds in home ranges (panels A, red circles). At the regional
metapopulation level, habitat patches (panels C, connector lines) are connected by dispersal. At all levels,
two seasonal scenarios (wet vs. dry) accounted for temporal variability in environmental conditions.
In this thesis, I took a resource-based modelling approach (Hartemink et al. 2015) – building
on the ecological or behavioural factors influencing the spatial population patterns of feral pigs –
and applied models to two seasonal scenarios (wet and dry). Due to empirical uncertainty about the
processes of habitat selection and dispersal at the regional extent, models relied heavily on expert
knowledge and, where possible, subsequent evaluation of assumptions against empirical data.
Importantly, models did not simulate actual population dynamics, spread or abundance in a
spatiotemporal continuum. This level of detail was rejected partly because feral pigs in northern
Australia are subject to frequent and unpredictable control activities that strongly affect actual
population dynamics and distributional patterns. The adopted resource- and scenario-based research
approach was ultimately deemed most appropriate, and cost-effective, for generating regional-scale
spatial knowledge on feral pigs that may usefully inform strategic disease management decisions.
26
1.6 Thesis structure
This thesis is partly comprised of publications in peer-reviewed scientific journals (Chapter 4 is
published and Chapters 3 and 5 are being prepared for submission). These were reproduced here in
their published or prepared manuscript form, with only minor modifications 4. First, in publications,
the study species Sus scrofa was referred to by the internationally recognized term ‘wild pig’; here,
this was substituted by ‘feral pig’, which is more familiar to Australian readers. Second, journal
requirements on formatting and referencing were streamlined to conform to thesis submission
guidelines. Third, all references and appendices were collated at the end of the thesis 5.
In Chapter 1 I introduced the biosecurity threat of infectious animal diseases with wildlife
hosts. I highlighted that limited spatially-explicit knowledge of wildlife hosts at the regional scale is
a major hindrance to effective post-border disease management at all stages of invasion. The aim of
this thesis is to fill this knowledge gap for feral pigs in northern Australia. I identified three research
questions that need to be answered to achieve this aim and provided a brief overview of the study
region. As empirical studies are prohibitive at the regional scale, I address these questions through
landscape ecological models of habitat suitability and connectivity at multiple biological levels.
In Chapter 2 I review the conceptual and methodological underpinnings of this thesis. First, I
review five major theoretical concepts in landscape ecology. Next, I review current knowledge on
the landscape ecology of feral pigs in northern Australia, highlighting that there are still significant
gaps in our understanding across spatial and temporal scales. Then I link this landscape ecological
framework to a landscape epidemiological perspective on infectious diseases in feral pigs. Finally, I
review the literature relating to the adopted research methodologies. This includes spatially-explicit
approaches to modelling habitat suitability and connectivity as well as methods for eliciting robust
knowledge from experts for model development and parameterization.
In Chapter 3 I develop a novel combined methodology, spatial pattern suitability analysis, for
integrating resource-seeking home range movements into habitat models of mobile animals. This
involves measuring structural patterns of resource quality at the home range scale and then relating
these measures to functional values from expert-elicited response-to-pattern curves. I use scenario
analysis and empirical validation in a subsection of the study region in northern Queensland to
evaluate whether this methodology improves model performance and how structural patterns should
Due to the stand-alone nature of journal papers, some repetition in introductory remarks, methodological descriptions
and associated figures, and discussion of findings does occur throughout this thesis. This is especially the case for
Chapters 3 and 4, which used a similar methodology for an exploratory (Chapter 3) and an applied (Chapter 4) analysis.
5
Appendices were numbered according to chapters and order of first mention, i.e. Appendix 1 supplements Chapter 1,
Appendices 3.1 – 3.4 supplement Chapter 3, and Chapter 2 does not have an appendix.
4
27
be measured for feral pig breeding herds. This chapter addresses research question 1, providing a
methodology for reliably modelling habitat suitability for feral pigs at the regional scale under
uncertainty about the processes of habitat selection by individual breeding herds. It is being
prepared for submission to Ecological Modelling.
In Chapter 4 I apply the methodology from Chapter 3 to model and map seasonal habitat
suitability for feral pig breeding and persistence in northern Australia. I provide a detailed
description of the resource-based, spatially-explicit modelling approach using expert-elicited
Bayesian networks and spatial pattern suitability analysis to account for resource-seeking home
range movements. I compare modelled habitat suitability for a wet season and a dry season scenario
and empirically validate model accuracy against four independent distributional data sets per
scenario. This chapter addresses research question 2, providing an indicative measure of the
seasonal establishment risk of directly transmitted infectious diseases within feral pigs in northern
Australia. It has been published in PLoS ONE.
In Chapter 5 I model and map seasonal habitat connectivity for feral pigs in northern Australia.
I apply the results from Chapter 4 to delineate contiguous patches of feral pig breeding habitat,
parameterize matrix resistance to inter-patch dispersal, and model seasonal connectivity in the
landscape matrix using a circuit-theoretic approach. Then, I estimate connectivity between patches
and evaluate its sensitivity to a range of model assumptions that are linked to gender-specific
differences in feral pig dispersal ability. This chapter addresses research question 3, providing an
indicative measure of the seasonal spread risk of directly transmitted infectious diseases within feral
pigs in northern Australia. It is being prepared for submission to Journal of Applied Ecology.
In Chapter 6 I relate the results from Chapters 3 to 5 to the aims of this thesis. For each of the
three research questions, I discuss the main findings of my research as well as their methodological,
ecological and applied significance. Next, I synthesize how the adopted multi-level modelling
approach and its outputs can help to inform risk-based management of infectious animal diseases
for which feral pigs are a wildlife host. I conclude by summarizing limitations and giving
recommendations for future research that may further improve on this thesis’ findings.
28
Publications
Thesis chapter (description)
Results
Chapter 1 Introduce research problem –
infectious animal diseases in feral pigs
Chapter 2 Review study system and
theoretical framework (wildlife hosts and
infectious diseases) and research methods
(habitat and connectivity models, elicitation)
Ecol
Modelling
(in prep.)
Chapter 3 Develop and validate methodology
for integrating home range movements
into feral pig habitat models
Methodology
PLoS ONE
(2017)
Chapter 4 Model and validate
seasonal habitat suitability for feral pigs
in northern Australia
Models
& maps
J Applied
Ecol
(in prep.)
Chapter 5 Model and evaluate
seasonal habitat connectivity for feral pigs
in northern Australia
Models
& maps
Chapter 6 Discuss findings and
applications to risk-based management
of infectious animal diseases in feral pigs
and recommend future research
Fig 1.3 Thesis structure. Broad descriptions of chapter content, types of results and associated publications
(published or being prepared for submission) are also given.
1.7 Research ethics approval
The research published in this thesis was conducted with human ethics approval from the CSIRO
Human Ethics Committee (Approval 075/13, 27 Aug 2013) and the UQ School of Agriculture and
Food Sciences Ethics Committee (9 Oct 2013).
29
Chapter 2 Literature review
Summary
In this chapter I review the conceptual and methodological underpinnings of this thesis. First, I
review five major theoretical concepts in landscape ecology. Next, I review current knowledge on
the landscape ecology of feral pigs in northern Australia, highlighting that there are still significant
gaps in our understanding across spatial and temporal scales. Then I link this landscape ecological
framework to a landscape epidemiological perspective on infectious diseases in feral pigs. Finally, I
review the literature relating to the adopted research methodologies. This includes spatially-explicit
approaches to modelling habitat suitability and connectivity as well as methods for eliciting robust
knowledge from experts for model development and parameterization.
2.1 Theoretical concepts in landscape ecology
Turner & Gardner (2015) broadly state that the discipline of “landscape ecology emphasizes
the interaction between spatial pattern and ecological process – that is, the causes and consequences
of spatial heterogeneity across a range of scales” (p. 2). With regard to an individual organism, this
means that reproduction, dispersal or other behavioural responses are influenced by the way abiotic
and biotic factors are distributed in the landscape and the organism’s responses in turn affect its
own distributional patterns and abundance (Levin 1992). In essence, landscape ecology combines
the spatial focus on observed patterns of biogeographers with the functional focus on underlying
processes of ecologists (Turner & Gardner 2015). The integration of these perspectives – pattern
and process, geography and ecology, structure and function – was also a primary motivation behind
this thesis. Excellent syntheses of the varied theoretical concepts and research approaches in
landscape ecology have been provided by Turner & Gardner (2015), Farina (2006), Forman (1995)
and Wu (2013). It is out of the scope of this thesis to repeat these discussions here. Instead, I will
focus on five broad theoretical concepts, which underpin the adopted research approach.
2.1.1 Habitat suitability
The first concept is the ‘habitat’. To define its use in this thesis, I must review diverse and
often ambiguous research perspectives, which fundamentally also reflect the dichotomy between
pattern and process emphasized above. Understanding the geographic distribution of species has
historically been of interest to ecologists and biogeographers alike. Earliest investigations have
focused on explaining observed distributional patterns from permissive environmental and climatic
conditions (von Humboldt & Bonpland 1807; Grinnell 1917). Hutchinson (1957) formalized
Grinnell’s insights as an “n-dimensional hypervolume, every point in which corresponds to a state
of the environment which would permit [… a species] to exist indefinitely.” (p. 416). He coined this
30
multi-dimensional volume of environmental requirements the ‘fundamental niche’ (since also
known as the ‘Grinnellian niche’; Leibold 1995; Pulliam 2000) and contrasted it with the ‘realized
niche’ (since also known as the ‘Hutchinsonian niche’; Leibold 1995; Pulliam 2000), which was
supposed to be narrower due to the effects of competition. Importantly, Hutchinson’s (1957)
“Concluding remarks” turned the focus from the description of observed distributional patterns to
the underlying, species-specific processes or mechanisms from which these patterns emerge.
Exactly how ecological processes in multi-dimensional ‘environmental (niche) space’ translate into
distributional patterns in two-dimensional ‘geographical space’ remains a complex question, and
depends on both the species and the environment (Hutchinson 1957; Pulliam 2000; Guisan &
Thuiller 2005; Hirzel & Le Lay 2008). Broadly, however, a species fundamental niche can be
linked to its potential distribution (where it may persist in a permissive environment), while the
realized niche gives rise to its actual distribution (where it currently occurs) (Pearson 2007).
The related ‘habitat’ concept has been used predominantly, and often ambiguously, in animal
studies. Firstly, Hall et al. (1997) or Morrison et al. (2006) stress that habitat should be an
organism-centric concept and not be confused with ‘habitat type’ (structurally similar landscape
elements, e.g. a type of vegetation). Secondly, some authors have defined habitat as the physical
area an animal potentially or actually occupies in geographical space (the descriptive ‘distribution’
perspective; Hall et al. 1997; Kearney 2006; Morrison et al. 2006), while others have referred to the
environmental requirements needed for persistence in environmental space (the mechanistic ‘niche’
perspective; Pulliam 2000; Hirzel & Le Lay 2008). Confusion arises when considering that: (a)
some suitable ‘niche habitat’ may actually not be occupied (e.g. due to historical or biogeographical
range limits; short-term extinctions due to stochastic events or, for invasive species, population
control activities; limitations on dispersal and re-colonization; or biotic interactions with other
species); (b) some unsuitable ‘niche habitat’ may consistently be occupied (e.g. due to continuous
immigration from suitable ‘source’ habitat into unsuitable ‘sink’ habitat; or due to regular
movements across unsuitable habitat); (c) permissive environmental conditions as well as actual
occupancy may change over time (e.g. due to climatic conditions and resource availability); and (d)
spatial heterogeneity and behaviour may influence habitat use (e.g. the amount of habitat may be
too small or fragmented to allow persistence or even-short term occupancy; or mobile species may
satisfy different requirements in separate locations, i.e. have a spatially disjunct ‘niche habitat’)
(Pulliam 2000; Guisan & Thuiller 2005; Hirzel & Le Lay 2008; Yackulic & Ginsberg 2016).
Here, I attempted to avoid such confusion by adopting a descriptive definition of habitat as a
geographical place, but limiting its application to feral pigs’ potential distribution and persistence.
In line with the resource-based approach (Hartemink et al. 2015), I defined ‘habitat’ as the
31
geographical area in which a species can potentially persist without immigration based on the
availability of all resources and conditions required for survival and reproduction.
Habitat thus defined can vary in ‘suitability’, which I used here equivalent to the term ‘habitat
quality’ that is preferred by some authors (Hall et al. 1997; Beutel & Beeton 1999; Stephens et al.
2015). I adopted the definition by Hall et al. (1997) who state that habitat quality (here ‘suitability’)
“should be considered a continuous variable, ranging from low to medium to high, based on
[the quality of] resources available [and conditions encountered] for survival, reproduction,
and persistence” (p. 178).
According to these definitions, my investigations of habitat suitability in feral pigs included only
factors that may impact on potential occupancy: (a) abiotic habitat factors such as consumable
resources and climatic conditions; (b) limiting factors such as the presence of competitors, predators
or anthropogenic disturbances; (c) spatiotemporal variability in the availability and quality of these
habitat factors; and (d) behavioural factors, in particular feral pigs’ use of home ranges to satisfy
heterogeneously distributed habitat requirements and avoid disturbances (Section 2.2.3). I excluded
from the analyses additional habitat factors that may only affect actual occupancy or persistence: (e)
dynamic biotic interactions within feral pig populations (e.g. reproductive, metapopulation, sourcesink or density-dependent dynamics; or demographic stochasticity) or with other species (e.g.
trophic, predator-prey or community interactions); and (f) historical anthropogenic factors such as
species introduction, translocation, colonization or population control.
2.1.2 Patch-corridor-matrix model and metapopulation ecology
The second concept is the ‘patch-corridor-matrix’ model (Forman 1995). In this simple spatial
model, the landscape is viewed as a mosaic, with every point being part of either a ‘patch’, a
‘corridor’, or the background ‘matrix’ (Forman 1995, McGarigal et al. 2012). While conceptually
simple, operationalising patches and corridors is by no means trivial (e.g. Girvetz & Greco 2007,
Beier et al. 2008, Galpern et al. 2011). One approach is to regard patches and corridors as
structurally distinct landscape features of a contiguous shape, whereby corridors may be
distinguished from patch polygons by their linear form (e.g. two forest stands connected by a
vegetated stream). Applications in conservation biology and reserve design have often focused on
connecting such ‘natural landscape blocks’ (= contiguous patches, Beier et al. 2011) of broad
preservation value by means of wildlife ‘linkages’ (= linear corridors, Beier et al. 2008).
However, in many landscapes these distinctions may not be obvious. Even where structurally
separate features may be discerned by a human observer, it is widely acknowledged that there is no
single set of patches in a landscape. Rather, “the contrast between a patch[, or a corridor,] and the
32
surrounding matrix is dependent on the perceptual abilities and behavioral responses of a focus
organism” (Girvetz & Greco 2007, p. 1132). From this functional, organism-centric perspective, the
patch-matrix dichotomy is best described in terms of discontinuities in habitat quality or patterns of
use (McGarigal et al. 2012). For mobile animals such as feral pigs, which use home ranges to
access resources at different locations (Börger et al. 2008; Section 2.1.3), meaningful habitat
patches must be delineated functionally. While there are many approaches for doing this (Girvetz &
Greco 2007, Galpern et al. 2011), all aim at identifying patches that are “good enough, big enough,
and close enough together” to function as habitat (Beier et al. 2008, p. 844; Section 2.5.2).
The patch-corridor-matrix model, when functionally defined, can also be linked to the theory
of ‘metapopulation dynamics’ (Hanski 1998). Then, unstable local populations inhabit discrete
patches of suitable breeding habitat that are embedded in a matrix of unsuitable habitat. Certain
areas in the matrix function as dispersal corridors between patches. Metapopulation persistence
depends on local population dynamics and movements in the patch network (Hanski 1998; Hastings
2014). Glen et al. (2013) argue that biosecurity applications and invasive species management
could benefit from using the metapopulation paradigm in reverse, focusing on eradication of local
populations and disruption of inter-patch dispersal. In this thesis I echo their call.
2.1.3 Scaling in space and time
The link to metapopulation ecology points to the importance of scale as a third fundamental
concept to be discussed here (Levin 1992). From a functional perspective, habitat patch networks in
a given landscape vary not only by focal species, but also display a hierarchically nested structure
depending on the biological level and time horizon investigated (Girvetz & Greco 2007). According
to McGarigal et al. (2016), “‘level’ refers to a constructed organizational hierarchy […in] (a) the
environment […] or (b) the focal organism’s behaviour” (p. 1164). For example, a species’
behavioural hierarchy is often described in terms of Johnson’s (1980) four ‘orders’ of habitat
selection (1st order = population range, 2nd order = home range, 3rd order = resource patches such as
feeding sites, 4th order = individual resources such as food items). Each of these levels must be
investigated at different spatial and temporal scales (Mayor et al. 2009; McGarigal et al. 2016).
However, what constitutes a suitable spatiotemporal scale of analysis, including ‘grain’ (=
resolution, smallest unit of observation) and ‘extent’ (= largest entity of analysis), is by no means
clear. Here, Addicott et al. (1987) introduced the general concept of the ‘ecological neighbourhood’
as “the region within which […an] organism is active or has some influence during […an]
appropriate period of time” (p. 341). Addicott et al. (1987) argued that each ecological process has
its own unique ecological neighbourhood and the choice of process will determine the appropriate
timeframe and scale of measurement. Yackulic & Ginsberg (2016) provided an excellent synthesis
33
of how a species’ biological and behavioural levels can be linked with ecological processes and
suitable grains and extents of analysis to explain distributional patterns consistent with the
metapopulation perspective on mobile species adopted here (Fig 2.1): individuals or social groups
(level) select and move between (process) resource patches (grain) within daily home ranges
(extent); local subpopulations (level) form where several home ranges or territories (grain) intersect
to form contiguous habitat patches (extent) that allow for in situ persistence (process);
metapopulations (level) form where several local populations (grain) are connected via dispersal
across geographical ranges (extent) to allow for species persistence (process); ecological processes
and distributional patterns at each level may fluctuate periodically (temporal grain) and stabilize, or
shift, over longer time horizons (extent). A large number of ‘focal patch’ or ‘multi-scale’ studies
have been devoted to finding the “right” ecological neighbourhood at each of these biological levels
(Section 2.4.3; Holland et al. 2004; Jackson & Fahrig 2015; McGarigal et al. 2016).
Fig 2.1 The link between a mobile species’ behavioural levels and corresponding spatial and temporal
scales of analysis. (Source: Mayor et al. 2009, Fig 1, modified after Yackulic & Ginsberg 2016)
When studying habitat suitability for local feral pig subpopulations (research question 2), I
assumed that the most relevant ecological neighbourhood is the home range of individual breeding
herds, i.e. suitability for breeding and persistence is regulated by the same scale-dependent
variations in resource availability that influence habitat selection by individual breeding herds
(Yackulic & Ginsberg 2016). This is because the dominant process at the level of subpopulations is
social aggregation, where individual breeding herds interact with adjacent herds in overlapping
home ranges to form contiguous subpopulations within a habitat patch (Section 2.2.2; Choquenot et
34
al. 1996; Hone 2012; Yackulic & Ginsberg 2016). In territorial animal populations characterized by
repulsion, the ecological neighbourhoods of individuals and local populations may be quite
different (Yackulic & Ginsberg 2016).
2.1.4 Landscape heterogeneity and home ranges
Addicott et al. (1987) also refer to the importance of ‘environmental patterning’ in ecological
neighbourhoods. This can be illustrated with mobile animals’ home range behaviour. Home ranges
can be broadly defined as the “area traversed by an individual in its normal activities of food
gathering, mating, and caring for young” (Burt 1943, p. 352). While many theoretical and
operational questions about home range behaviour remain, the conceptual basis of animals’ space
use is widely agreed upon: mobile animals use home ranges to satisfy a number of heterogeneously
distributed habitat requirements (Börger et al. 2008; Powell & Mitchell 2012). Spatial patterns of,
and relationships between, required resources and other habitat factors – for example the location of
predators in relation to hiding places or the availability of sufficient feeding as well as breeding
sites – will affect space use, and habitat suitability.
The spatial heterogeneity arising from such patterns and relationships has also been referred to
as ‘landscape structure’ (Kupfer 2012; Turner & Gardner 2015). A range of ‘landscape metrics’ for
describing aspects of landscape structure have been developed over the past two decades
(McGarigal & Marks 1995; Uuemaa et al. 2009; McGarigal et al.2012). It is out of the scope of this
thesis to comprehensively review these, yet metrics can be classed into two general categories.
Metrics of landscape ‘composition’ refer to the variety and abundance of elements in a landscape
(e.g. resource or habitat patches), and include measures of amount, density, proportional cover or
diversity. Metrics of landscape ‘configuration’ refer to the spatial position and arrangement of
landscape elements. They can be used to measure both the patterning of individual resources and
the spatial relationships between different types of resources or habitat patches. Typically,
configuration metrics measure properties such as distance / proximity, patch geometry, edges
between patches, fragmentation, aggregation / clumping or isolation (McGarigal & Marks 1995).
Importantly, all metrics of landscape structure, per se, are independent of ecological function, that
is, they carry no implicit meaning to the way a species uses and perceives its environment (Kupfer
2012; Turner & Gardner 2015).
Here, Dunning et al. (1992) described three fundamental ecological processes that link
structural heterogeneity within ecological neighbourhoods (Addicott et al. 1987), or more
specifically mobile animals’ home ranges, to habitat quality (here ‘suitability’). Resource
‘complementation’ occurs when an individual selects resources required for different purposes (e.g.
forage and breeding sites) at different locations within its home range. Therefore, habitat suitability
35
of a focal site depends on the availability of all required resources within the accessible home range
neighbourhood. The individual may also be able to select different types of ‘supplementary’
resources that can be substituted for the same purpose (e.g. different food items). Here, habitat
suitability depends on the quality of the best substitutable resources that are available. Lastly,
habitat suitability is influenced by ‘neighbourhood effects’, that is, the distance of complementary
and supplementary resources from the focal site will affect their value to a mobile animal.
2.1.5 Habitat connectivity
While ‘patches’ have frequently been investigated at all biological levels (e.g. resource, home
range or habitat patch), ‘corridors’ have typically been studied at the metapopulation level in terms
of dispersal within a patch network 6. In the expanding field of connectivity research, which broadly
aims at studying how patches are linked to one another, two broad perspective have emerged
(Tischendorf & Fahrig 2000; Moilanen & Hanski 2001; Kool et al. 2013; Fletcher et al. 2016a):
Metapopulation studies have focused on the population dynamics within and between patches that
are ‘isolated-by-distance’ and investigated how linked patches facilitate metapopulation persistence
(Pulliam & Danielson 1991; Hanski 1998; Hastings 2014). The ‘landscape connectivity’ (here
referred to as ‘habitat connectivity’) perspective has focused on landscape heterogeneity, both
within patches and in the surrounding matrix, and studied how this heterogeneity influences
connectivity between habitat patches that are ‘isolated-by-resistance’ (McRae 2006; Turner &
Gardner 2015). Thus, ‘connectivity’ is conceptualised as an attribute of the landscape, rather than
populations, and has been defined as “the degree to which a landscape facilitates or impedes
movement of organisms among […habitat] patches” (Tischendorf & Fahrig 2000, p. 7).
Nevertheless, habitat connectivity is not independent of the organism, nor the biological level,
investigated (Tischendorf & Fahrig 2000). Kool et al. (2013) distinguished between three analytical
perspectives on connectivity: ‘Structural connectivity’ refers to the physical attributes and
heterogeneity in a landscape. It can be directly measured using landscape metrics – ranging from
simple Euclidean distance to complex measures of landscape composition and configuration
(Kindlmann & Burel 2008) – and carries no implicit ecological meaning. ‘Functional connectivity’,
on the contrary, considers structural heterogeneity as well as species-specific behavioural responses
(Tischendorf & Fahrig 2000; Kindlmann & Burel 2008). A critical concept in studies of functional
connectivity is ‘matrix resistance’. It refers to the permeability of landscape elements outside of
habitat patches to movement or the costs (energy expenditure, mortality risk) associated with
Yet, movements between resource patches within a home range, or movements between home range patches within a
local population, could also be conceptualized as ‘corridors’ (e.g. Mitchell & Powell 2012).
6
36
traversing them. Matrix resistance, and functional connectivity, of a given landscape will vary
according to the requirements, abilities and behavioural responses of an organism (McRae 2006;
Kindlmann & Burel 2008; Zeller et al. 2012). Depending on the biological level investigated, Kool
et al. (2013) and Fletcher et al. (2016a) further distinguish between ‘functional connectivity’ as
defined by Tischendorf & Fahrig (2000), which studies dispersal movements between local
populations over moderate time frames (months to years, Fig 2.1), and ‘genetic connectivity’, which
studies gene flow within or between metapopulations (decades to centuries, Fig 2.1). The ecological
neighbourhood of dispersal or gene flow processes, and corresponding spatiotemporal scales of
analysis, may be quite different (Fletcher et al. 2016a). Finally, Fletcher et al. (2016a) contrast
functional ‘potential connectivity’ (where matrix permeability allows species movement) and
‘realized / actual connectivity’ (where movement currently occurs) (Calabrese & Fagan 2004).
When modelling habitat connectivity for regional feral pig metapopulations (research question
3), I focused on potential functional connectivity. I did not attempt to analyse actual dispersal
movements at a particular point in time (Fletcher et al. 2016a), gene flow over long time horizons
(Fletcher et al. 2016a), or dynamic population-level responses such as source-sink relationships
between patches (Pulliam & Danielson 1991; Dunning et al. 1992) or metapopulation persistence
(Hanski 1998; Hastings 2014). The ecological neighbourhood of connectivity analyses in this thesis
was determined by the ability of individuals to disperse between local subpopulations.
2.2 Landscape ecology of feral pigs
Feral pigs are descendants of domesticated Eurasian wild boar (Sus scrofa). Both wild and
domesticated forms of S. scrofa have been introduced by early settlers to all continents and many
oceanic islands. The species is widely considered as one of the worst invasive species throughout its
introduced range (Barrios-Garcia & Ballari 2012). In Australia, after two centuries of recurrent
introductions, translocations and natural dispersal, feral pigs are now among the most abundant and
widespread terrestrial mammals (West 2008). Consequently, the species’ ecology, impacts and
management options have been widely studied (Choquenot et al. 1996; Hone 2012; Bengsen et al.
2014). Below I review the published knowledge on key ecological processes and spatial patterns
from the individual to the metapopulation level, highlighting that there are still significant gaps in
our understanding of feral pig populations in Australia across spatial and temporal scales.
2.2.1 Distribution and abundance
The distribution and abundance of feral pigs in Australia has been investigated along three
broad lines of enquiry. First, a large number of field studies have empirically investigated local
habitat use and population densities in response to a range of environmental and biotic factors, often
37
dynamically across seasons or years (Hone 1990a; Caley 1993 & 1997; Corbett 1995; Choquenot
1998; Mitchell 2002; Choquenot & Ruscoe 2003; Mitchell et al. 2009). For example, Caley (1993)
and Hone (1990a) showed that feral pig distribution and abundance varies considerably between
seasons and habitat types of varying productivity. Yet, the detailed empirical observations from
such studies remained confined to a few study sites and were rarely integrated with spatial data sets.
Hone (1990b) consolidated five such published field studies to estimate the total size of the feral pig
population in Australia. Yet, with a limited sample size and no spatial integration, this estimate
remained enormously uncertain (95% confidence interval: 3.5 to 23.5 million pigs).
In response to a lack of spatially-explicit information, the second line of enquiry sought to map
the broad distribution and abundance of feral pigs at the regional or continental extent (Mitchell et
al. 1982; Wilson et al. 1992; West 2008; Department of Agriculture and Fisheries 2015). Due to
methodological and resourcing constraints, these efforts typically relied on expert-derived relative
estimates per coarse land parcel and were poorly validated, which limited their usefulness for onground applications (Cowled & Garner 2008). Other ongoing mapping exercises such as the Atlas
of Living Australia (2016) or FeralPigScan (2016) collate and map occurrence records across
Australia from a range of sources, including historical surveys or citizen science. All studies in this
group share two further shortcomings: they were unable to fill data or knowledge gaps and temporal
variability was not considered.
A third type of study – habitat models, which infer species distributions from environmental
predictor variables (Section 2.4.1) – can overcome the limitations of simple mapping exercises. For
feral pigs in northern Australia, one such model was developed by Cowled and colleagues using a
statistical (Generalized Additive Modelling) approach (Cowled & Giannini 2007; Cowled et al.
2009). However, the authors suggested that the model markedly underestimated feral pig
distributions in areas for which no presence / absence data were available. Such extrapolation errors
are a common limitation of statistical models, which rely purely on data with limited consideration
of ecological and behavioural processes (Elith et al. 2010). Further, the model was calibrated from
aggregate occurrence records and did not account for temporal variability.
2.2.2 Social organisation and reproduction
While broad distributional patterns have been mapped and modelled over regional extents,
Cowled & Garner (2008) suggest that, at a finer grain, feral pigs are “rarely distributed
homogenously […], since they have strong social tendencies” (p. 202). There are two distinct social
entities in any feral pig population: related adult sows, sub-adults and juveniles typically live in
social herds called sounders; mature boars usually disperse from their natal area and remain solitary
(Choquenot et al. 1996; Gabor et al. 1999; Mitchell 2008). Female herd sizes can vary considerably
38
between different environments and seasons, and are not stable over time. Choquenot et al. (1996)
reviewed published herd sizes, concluding that they typically contain fewer than 10, but under
exceptional circumstances up to 100 animals. These breeding herds display a highly fluid ‘fissionfusion’ dynamic, with sub-groups splitting off, merging and exchanging individuals across space
and time (Gabor et al. 1999; Spencer et al. 2005). While Gabor et al. (1999) found that sounders
living in adjacent habitat patches did rarely interact, there is little further evidence for territorial
behaviour in either females or males (Choquenot et al. 1996; Hone 2012). In fact, genetic analyses
have shown that individual boars frequently move between sounders for mating and sows accept
multiple mating partners (Gabor et al. 1999; Spencer et al. 2005).
Feral pigs have an extraordinary reproductive capacity, comparable to rabbits rather than other
large mammals in Australia (Choquenot et al. 1996; Mitchell 2008). Under favourable conditions,
breeding may occur throughout the year and sows may produce two weaned litters per year. Each
litter contains on average 5 to 6 piglets and requires 2 to 3 months lactation to weaning (Choquenot
et al. 1996). Results from a number of studies investigating population dynamics of feral pigs in
various environments indicated that populations can increase up to twofold (instantaneous rate of
increase ~0.7) within one year (Choquenot et al. 1996; Hone 1990a; Caley 1993).
2.2.3 Resource selection and home range movements
Site-scale resource selection and habitat use of feral pigs has also been studied widely in the
Australian context. A species’ habitat (resource) requirements are determined by its physiological
attributes. Due to their compact shape and few sweat glands, feral pigs overheat easily (Mitchell
2008). This can have a negative effect on survival (Choquenot & Ruscoe 2003) and fertility (Greer
1983) in both sows and boars. The species’ diet is omnivorous. Yet, pregnant or lactating sows in
particular require a relatively high intake in digestible energy and crude protein (Choquenot et al.
1996). Preferred food items include fresh green vegetation and a range of other protein-rich plant
(fruit, seeds, grains, roots and bulbs) and animal material (eggs, invertebrates) to satisfy nutritional
requirements (Choquenot et al. 1996; Ross 2009). Despite being highly adaptable to a range of local
habitat conditions, resource selection analyses have, unsurprisingly, established that feral pigs
prefer habitats which satisfy their physiological needs for thermoregulation (Dexter 1998), frequent
hydration (Caley 1997) and a protein-rich diet (Caley 1997; Dexter 1998; Wurster et al. 2012).
After Dunning (1992), these resource requirements are both complementary (i.e. are all needed for
different purposes as described above) and supplementary (i.e. can be opportunistically substituted
with resources serving the same purpose, e.g. shady tree vs. muddy wallow for thermoregulation).
Importantly, feral pigs’ habitat use and distributional patterns depend both on the availability
and patterning of resources as well as behavioural factors. Like all large and mobile animals, feral
39
pigs are able to access heterogeneously distributed resources at different locations and times within
the boundaries of their home range (Powell & Mitchell 2012). A number of home range analyses
have found that feral pigs are relatively sedentary for their size and occupy well-defined home
ranges. There were consistent gender-specific and seasonal differences in home range movements
(Caley 1997; Dexter 1999; Mitchell et al. 2009). Dexter (1999) found that female home ranges
decreased significantly under drought conditions. This was interpreted as a “compromise between
the need to forage widely for food and the need to stay close to reliable cover and water” (Dexter
1999, p. 759). Male home ranges were significantly larger and, interestingly, did not vary with
conditions. Such findings suggest that boars are less movement-constrained by daily requirements
for water and protein-rich food (Choquenot et al. 1996). Overall, male and female home range sizes
ranged widely between 1-40 km2 and 1-20 km2 respectively (Choquenot et al. 1996; Hone 2012).
Despite these detailed findings on resource selection and home range movements, few studies
have addressed the role of landscape structure and scaling, i.e. the patterns and spatial relationships
between resources within home range neighbourhoods. Here, a rare insight stems from Choquenot
& Ruscoe (2003), who explicitly investigated resource complementation at different scales and
found that food resources and thermal refuge must co-occur within 5 km to facilitate feral pig
persistence. Caley (1993) investigated food supplementation with cereal crops and found that feral
pig density in riverine woodlands increased almost four-fold when crops were available within a 5
km radius. These site-scale findings are consistent with the observed 20 km2 home range limit for
breeding herds (Hone 2012). Yet, the influence of resource composition and configuration in the
home range on broad distributional patterns of feral pigs has not been studied so far.
2.2.4 Metapopulation structure and dispersal
Several studies have analysed the population structure of feral pigs using landscape genetic
evidence. In a first study covering 4,000 km2 in south-western Queensland, Cowled et al. (2006)
found a genetically contiguous population. When repeating the study over a much larger (500,000
km2) area within the same region, Cowled et al. (2008) did find genetic differentiation into five
subpopulations. A simple Euclidean distance measure explained 50% of the variability in
differentiation. Interestingly, evidence for mixing was most prominent in drier landscapes. This
suggested that feral pigs may utilize rivers and floodplains to disperse widely within unfavourable
landscapes while they are less inclined to disperse where resources are available locally (Cowled et
al. 2008). In a fragmented landscape in south-western Western Australia, Hampton et al. (2004)
revealed genetic fine-grained differentiation at the catchment level. There was high gene flow along
water courses but little population mixing between neighbouring catchments as little as 25 km apart
(Hampton et al. 2004). Likewise, Lopez et al. (2014) identified strong population-structuring at a
40
scale of 25 to 35 km in the coastal lowlands of wet-tropical northern Queensland. Unfortunately,
while population structuring has been extensively documented, dispersal processes responsible for
these emergent patterns often remained unclear. In particular, the influence of landscape
heterogeneity in the matrix on inter-patch connectivity has not been investigated.
In summary, empirical studies indicate that the patch-corridor-matrix and metapopulation
frameworks can adequately describe populations over regional extents. Distributional patterns are
spatially clustered because feral pigs congregate in social breeding herds (also called ‘sounders’ or
‘mobs’). Herds display a dynamic group structure and interact frequently with adjacent herds to
form contiguous local populations within a habitat patch. Contact between subpopulations in
separate habitat patches is generally rare, as the intervening matrix habitat may present a major
obstacle to resource-constrained females. Yet, besides the energetic and survival costs associated
with moving through poor quality habitat, no absolute barriers to feral pig dispersal have been
identified in Australia (Choquenot et al. 1996; Cowled & Garner 2008; Cowled et al. 2008; Hone
2012; Lopez et al. 2014). Hence, subpopulations may be frequently linked via male dispersal
movements, as these individuals are less constrained by daily resource requirements (Choquenot et
al. 1996; Gabor et al. 1999). Importantly, research gaps exist with regard to inter-patch dispersal,
metapopulation dynamics, and connectivity within the habitat patch network.
2.3 Landscape epidemiology of infectious diseases in feral pigs
The field of ‘landscape epidemiology’ (also termed ‘spatial epidemiology’) has recently reemerged from earlier work and received much attention since (Ostfeld et al. 2005). Essentially, it
stipulates the concept of the ‘nidus of pathogen transmission’, where “vector [if applicable], host,
and pathogen populations intersect within a permissive environment” (Reisen 2010, p. 463). That is,
spatial patterns of disease distribution and spread will depend on the spatial patterns of host (and
vector) distribution and the spatial relationships between hosts from the individual to the population
level (Ostfeld et al. 2005; Hartemink et al. 2015). As these patterns and relationships are the
primary focus of landscape ecology (Turner & Gardner 2015; Section 2.1), the disciplines are
closely linked (Reisen 2010). Anderson et al. (1986) distinguished three phases of invasion by
infectious diseases: establishment, persistence and spread. For directly transmitted (as opposed to
indirectly transmitted – e.g. vector-, air-, or soil-borne) diseases with a single host, these phases can
be linked to the general processes in host metapopulations described by Hastings (2014): local
population growth and persistence (Section 2.1.1), connectivity and dispersal (Section 2.1.5), and
regional metapopulation persistence. Following introduction, initial disease establishment is
contingent on locally dense host populations with high contact rates that let an infectious pathogen’s
basic reproductive rate R0 exceed 1. Subsequent disease spread requires host dispersal between
41
infected and susceptible populations. Disease persistence depends on a lasting supply of susceptible
individuals. This could pertain to short-term persistence in a closed subpopulation or long-term
persistence in a regionally connected metapopulation (Anderson et al. 1986; Kramer-Schadt et al.
2007; Riley 2007). In the following I illustrate these links using the example of the directly
transmitted disease classical swine fever (CSF) and its sole host species Sus scrofa.
Feral pigs congregate in social breeding herds that interact frequently with herds in adjacent or
overlapping home ranges to form contiguous, persistent local populations within patches of suitable
breeding habitat (Section 2.2.2). CSF is highly contagious and spreads mainly by direct contact
between live individuals, especially between piglets or from mothers to their young (Kramer-Schadt
et al. 2007; Animal Health Australia 2012). Kramer-Schadt et al. (2007) observed that within-herd
transmission of CSF depends on a persistent supply of young susceptible animals. Therefore,
Cowled & Garner (2008) suggested that local feral pig populations could be regarded as single
epidemiological units, where host density is sufficiently high, and interaction sufficiently frequent,
to let CSF’s basic reproductive rate R0 exceed 1 (Anderson et al. 1986).
Contact between separated feral pig populations is generally rare, as the intervening matrix
habitat may present a major obstacle to dispersal. Yet, contact rates may vary by gender and season,
with resource-constrained female breeders having a lower dispersal ability than often solitary males,
especially under adverse conditions (Section 2.2.4). Kramer-Schadt et al. (2007) observed that CSF
spread between infected and susceptible local populations may be enhanced by increased “contact
during the rutting season, male dispersers and establishment of new social groups” (p. 4). The view
that disease spread in feral pigs may be limited by spatial clustering is also supported by Zanardi et
al.’s (2003) study in Italy. They found that two adjacent populations of wild boar that were
separated by unfavourable habitat experienced two temporally distinct CSF epidemics.
Finally, while contact may be rare, especially during adverse conditions, local feral pig
populations are by no means isolated. Populations may merge, or dispersal movements increase,
during favourable conditions (Section 2.2.3). Distant populations may also be linked by linear
landscape features (Section 2.2.4). In some landscapes, feral pigs may even be distributed
homogeneously (Hone 2012). CSF, on the other hand, occurs in strains of highly variable
infectiousness and impact on the host organism (Section 1.1.2). Low virulent strains may cause
chronic disease and prolonged infectiousness with few clinical signs, enabling long-term disease
persistence within a feral pig metapopulation that seasonally connects local infected and susceptible
subpopulations (Anderson et al. 1986; Kramer-Schadt et al. 2007; Animal Health Australia 2012).
In summary, classical swine fever in feral pigs appears an ideal host-pathogen system to follow
Ostfeld et al.’s (2005) call for “[more studies that consider] the types, sizes and positions of
42
landscape elements (e.g. habitat patches, physical or biotic gradients, and type of matrix
surrounding patches) and their connectivity […as] potentially important drivers of [disease] risk or
incidence […, and that explore] the importance of landscape composition (number and types of
patches) and configuration (spatial relationships among patches) to disease dynamics” (p. 334).
2.4 Habitat suitability models
2.4.1 Modelling frameworks
With increased availability of spatial environmental data, and continued development of new
analytical methods, an enormous body of literature has emerged on the modelling of species
distributions, or ‘habitats’. In Fig 2.2 I attempt to summarize the major approaches based on their
treatment of explanatory variables, formulation of relationships, and modelled responses in either
geographic or environmental space (Section 2.1.1). I emphasize that these categories are arbitrary,
as many hybrid approaches continuously emerge in all lines of research (Franklin 2010; Gallien et
al. 2010), and broadly outline approaches and applications below.
Fig 2.2 Approaches to modelling species distributions, or ‘habitat’, in geographical and environmental
space. The diagram illustrates schematically (by connected points) how different components of the
modelling workflow are treated in either or both geographical and environmental space. Input variables
43
(covariates) can relate to abiotic and biotic spatial entities (e.g. resource patches), or to abiotic and biotic
functional requirements for an organism. The modelling process or algorithm may focus on spatial patterns
of and relationships between, or on the ecological function of and interactions between, abiotic and biotic
variables. The model output (response variable) may describe a species’ functional niche, or its spatial
distributional patterns. The location of each connected point on the abiotic-biotic variable / pattern / process,
fundamental-realized niche, and potential-actual distribution axes is approximate. Placement in the middle of
each axis means that the approach can accommodate either or both extremes.
Statistical habitat models are also known as ‘species distribution models’ (Elith & Leathwick
2009). They correlate a response variable (typically presence, presence / absence or abundance) and
a number of explanatory variables (covariates), both observed in geographical space, via assumed
but non-specified functional links in environmental space to infer the response at non-sampled
locations in geographical space. Neither do they model relationships between variables in
geographical space, nor do they make explicit reference to the functional niche requirements of an
organism. Approaches are methodologically and terminologically diverse (Dormann et al. 2012;
Gallien et al. 2010; Guisan & Thuiller 2005; Venette et al. 2010). Relatively simple ‘bioclimatic
envelope’ models have combined species presence records with abiotic (climatic) variables to
approximate a species potential distribution, usually over broad areas (Guisan & Thuiller 2005;
Venette et al. 2010). More complex methods have integrated species presence or presence / absence
records and abiotic (e.g. vegetation, land use, physical landscape characteristics) or biotic (e.g.
predators, areas with known competitive advantage) covariates, modelled relationships using
machine-learning or regression-based techniques, and been used to estimate potential or actual
distributions over broad or narrow study areas (Guisan & Thuiller 2005; Phillips et al. 2006; Elith
& Leathwick 2009; Jiménez-Valverde et al. 2011). Resource selection functions have quantified
animals’ preferential selection of abiotic resources or habitats by comparing observed patterns of
use with availability and used these insights to estimate actual habitat use, occupancy, or abundance
(Johnson 1980; Manly et al. 2002; Boyce et al. 2016). Although statistical habitat models have also
been referred to as ‘niche models’ (Elith & Leathwick 2009), I concur with Kearney (2006), who
reserves the term ‘niche’ for models that explicitly focus on functional processes.
Some spatially-explicit approaches model only spatial relationships between response and
explanatory variables, with no reference to underlying ecological processes. These models have
been applied to estimate a species’ actual distribution from incomplete presence data, or a mobile
animal’s home range ‘utilization distribution’ from incomplete telemetry data (Powell & Mitchell
2012). Techniques include range maps (Elith & Leathwick 2009), spatial interpolation (Bahn & Gill
2007), and statistical estimators such as convex hulls or α-hulls (Phillips et al. 2008), or kernel
44
smoothing / density estimators (Gormley et al. 2011; Kie et al. 2011). Recent geostatistical methods
such as regression-kriging or geographically weighted regression have emerged to bridge the gap
between modelling spatial or functional relationships (Austin 2007; Hengl et al. 2009).
Statistical techniques can only make robust inferences within the environmental or spatial
gradients contained in the data. Extrapolation to new geographical regions or novel conditions is
fraught with danger (Elith & Leathwick 2009; Paton & Matthiopoulos 2016). To predict species’
potential distribution across regions that have not been fully sampled, a range of approaches that
infer patterns from underlying processes, rather than processes from observed patterns, have been
developed. These have been termed ‘process-based’, ‘resource-based’, ‘deductive’ or ‘mechanistic’
models (Kearney & Porter 2009; Gallien et al. 2010; Venette et al. 2010; Hartemink et al. 2015).
‘Mechanistic niche models’ have focused on the biophysical niche requirements of organisms
and on that basis modelled how abiotic conditions interact to affect fitness (Phillips et al. 2008;
Kearney & Porter 2009; Kearney et al. 2009). As the entire modelling process is based in
environmental space, it has been claimed that such models can truly capture a species’ fundamental
niche (Kearney 2006). Biophysical requirements can subsequently be linked to spatial data proxies
and the modelled niche projected into space as a description of the organism’s potential distribution.
‘Resource-based habitat models’ have also focused on the processes in environmental space
from which habitat patterns emerge. However, different to niche models, resource-based models
have taken a broader view on an organism’s habitat requirements and modelled interactions
between abiotic, behavioural and temporal factors (Hartemink et al. 2015). Resource-based habitat
models have typically not accounted for intra- or inter-specific interactions (Dijak & Rittenhouse
2009). Notable techniques in this framework include habitat suitability index (HSI) models (Brooks
1997; Dijak & Rittenhouse 2009) or Bayesian network models (Section 2.4.2). By linking habitat
requirements to spatial data proxies, models can directly estimate habitat quality or suitability,
which in turn can be related to potential distribution in geographical space. Resource-based models
may not predict species’ actual distributions well (Dijak & Rittenhouse 2009).
Finally, all of the models discussed above essentially provide static snapshots of an organism’s
fundamental / realized niche or potential / actual distribution. Yet, each of these responses, and in
particular the realized niche / actual distribution end of the spectrum, may vary across time. Further,
they may all be influenced by biotic interactions within populations (e.g. density-dependent
population dynamics or demographic stochasticity) or with other species (e.g. trophic, predator-prey
or community interactions). A variety of ‘dynamic simulation models’ have been developed to
account for biotic interactions, often without reference to geographical space (i.e. spatially-implicit
models). Notable examples include ‘population viability analysis’, ‘metapopulation models’,
45
‘matrix models’, or ‘individual-based models’ (Hanski 1998; Larson et al. 2009; Franklin 2010;
Gallien et al. 2010). Spatially-explicit forms of these dynamic models often integrate abiotic and
structural variables (e.g. patch size, or outputs from other habitat models) and may therefore be used
to infer aspects of either the fundamental or realized niche. ‘Mechanistic home range analyses’ are a
widely used spatially-explicit application of the individual-based modelling framework. By
modelling the process of individual movement in environmental space (as a function of behavioural
responses to abiotic and biotic and habitat variables that can be linked to geographical space), such
models can make inferences about home range behaviour (Mitchell & Powell 2004; Börger et al.
2008; Mitchell & Powell 2012; Moorcroft 2012). When projected into geographical space,
simulation models are most useful for inferring dynamic response variables such as actual space
use, abundance, persistence, or population viability.
2.4.2 Bayesian networks
‘Bayesian networks’ are a general, flexible approach and have been applied in diverse contexts
(McCann et al. 2006; Wilhere 2012; Fenton & Neil 2013; Landuyt et al. 2013). Bayesian network
models are graphical influence diagrams, in which explanatory variables (‘parent nodes’) are linked
by causal relationships to response variables (‘child nodes’). Each model variable has at least two
mutually exclusive states, which exhaustively capture its total range of values or conditions. Causal
relationships and interactions between several parent nodes are quantified in ‘conditional
probability tables (CPT)’ behind each child node in the network (Cain 2001; Chen & Pollino 2012;
Fenton & Neil 2013). Because CPTs calculate responses as probability distributions based on
Bayes’ Theorem, they explicitly capture uncertainties about these relationships (Chen & Pollino
2012; Wilhere 2012). CPTs can be parameterized from data (empirical or model-derived) or expert
knowledge (published or elicited).
When applied to habitat modelling, Bayesian networks can be used in either a statistical datadriven, or process-based knowledge-supported framework (Boets et al. 2015). ‘Naïve Bayesian
classifiers’ learn the joint probability of a response variable from its relationships to a range of
explanatory variables in a training data set, similarly to logistic regression or other statistical
machine learning techniques (Lorena et al. 2011; Boets et al. 2015). Bayesian networks have far
more frequently been applied as resource- or process-based habitat models, where they offer unique
strengths (Marcot 2006; McCann et al. 2006; McNay et al. 2006; Douglas & Newton 2014;
Tantipisanuh et al. 2014; Hamilton et al. 2015). Just like most other habitat models (Section 2.4.1),
Bayesian networks model relationships in environmental space. Marcot et al. (2001) has proposed a
spatially-explicit framework, which was subsequently refined by Smith et al. (2007) and explained
in detail by van Klinken et al. (2015). Here, habitat suitability was conditional on a set of habitat
46
variables representing resource requirements (for animals) or invasion processes (for invasive
plants). Each habitat variable was itself influenced by several measurable explanatory variables and
each explanatory variable was linked to one or more remotely sensed or mapped spatial data
proxies.
The advantages of the resource-based, spatially-explicit Bayesian network approach include: a
robust statistical framework for modelling interactions between habitat variables based on species
ecology; flexible data needs, including the ability to integrate unpublished expert knowledge; and
explicit representation and propagation of uncertainty throughout the model, which allows
computation of habitat suitability as a probabilistic index (Uusitalo 2007; Wilhere 2012; Landuyt et
al. 2013). There are also several important limitations. First, model parameterization can be
challenging when the model includes many interactions between parent nodes because a conditional
probability distribution must be specified for each combination of states in all parents linked to a
child node (Cain 2001; Chen & Pollino 2012). If using expert knowledge, simplifying elicitation
techniques must be carefully considered (Cain 2001; Fenton et al. 2007; Section 2.6). Second,
Bayesian networks represent temporal dynamics poorly (McCann et al. 2006). Yet, they lend
themselves to a scenario approach, which has previously been applied to future climates (Murray et
al. 2012), land management (Smith et al. 2012) or population control (Murray et al. 2014), and
could equally be used to represent a series of time horizons (such as the seasonal scenarios used in
this thesis; Chapter 4). Third, while integration with spatial data layers has been frequent, it has
been applied in a non-standardized and often complex form (Johnson et al. 2012a). Yet, this
problem may be alleviated in future by recent technical advances (Landuyt et al. 2015).
2.4.3 Landscape variables
A further important limitation of Bayesian networks, which they share with most other habitat
modelling approaches outlined in Section 2.4.1, is that they are inherently spatially ‘neutral’ or
‘implicit’ (Johnson et al. 2012a). While both explanatory and response variables are routinely
projected into geographical space, relationships are modelled purely in environmental space and are
therefore “blind” to landscape heterogeneity and spatial interactions between variables (Elith &
Leathwick 2009; Yackulic & Ginsberg 2016). For mobile species such as feral pigs, which use
complementary and supplementary resources at different locations within heterogeneous home
ranges, integrating spatial relationships between habitat variables into the modelling process is
critical (Guisan & Thuiller 2005; McGarigal et al. 2016). While several methodologies for doing
this have been developed, these are not routinely integrated into either habitat modelling approach
(Beck & Suring 2009; Moorcroft 2012; McGarigal et al. 2016; Yackulic & Ginsberg 2016). To my
knowledge, resource-based Bayesian network habitat models have rarely addressed this issue
47
(Landuyt et al. 2015). Broadly, three methodologies for integrating ‘landscape variables’ that
measure spatial heterogeneity using various metrics of landscape composition or configuration in
home ranges, or more generally ecological neighbourhoods (Addicott et al. 1987; Holland & Yang
2016), have emerged:
Statistical resource selection functions and empirical habitat studies have adopted a ‘multiscale’ approach, whereby landscape variables were computed in circular areas of different radii
around a sampling site. ‘Scale-optimized’ variables that best explain a given response were then
selected and relationships between them modelled in environmental space (Holland et al. 2004;
Jackson & Fahrig 2015; McGarigal et al. 2016; Miguet et al. 2016). In resource selection functions,
for example, multiple scales effectively constrained the area in which resource availability was
measured and subsequently contrasted with actual use (DeCesare et al. 2012; Martin & Fahrig
2012; Laforge et al. 2016; Paton & Matthiopoulos 2016). Step selection functions have further
allowed this area of availability to shift following an animals’ movements or weighted availability
in the circle according to behavioural criteria, but have usually not used multiple scales of
measurement (Arthur et al. 1996; Rhodes et al. 2005; Moorcroft 2012).
Some recent statistical species distribution models have also adopted the multi-scale approach.
These computed landscape variables in moving windows that are incrementally centred on each
‘focal pixel’ within a study area, thereby essentially converting spatially independent variables into
landscape-scale variables that summarize information contained in each pixel’s neighbourhood
(Guisan & Thuiller 2005; Yackulic & Ginsberg 2016). Generally, scale-optimized landscape
variables were then selected in a univariate model and the relationships between all variables
investigated in a multivariate model (Bellamy et al. 2015; Ducci et al. 2015). One drawback of all
statistical approaches is that the functional relevance of included landscape variables has not always
been explicit and meaning is derived solely from correlations. This may be problematic because
scale-dependent insights may not be transferrable across larger study areas (Paton & Matthiopoulos
2016
Therefore, resource-based ‘landscape HSI models’ have been applied to large landscapes.
These studies used published or expert a priori knowledge to select landscape variables measured at
a specified scale, e.g. based on an animal’s known home range size or habitat use (Mitchell et al.
2002; Dijak et al. 2007; Dijak & Rittenhouse 2009). Similar to the statistical approaches described
above, neighbourhood summary variables were then computed in moving windows for each focal
pixel. Importantly, these structural measures were related to functional suitability indices,
describing the value of a measured spatial pattern to a given species, using expert knowledge or
response functions that were assumed to apply across the entire study area (Dijak & Rittenhouse
48
2009). The main disadvantage of this method is that landscape variables are rarely empirically
scale-optimized and their relationship to habitat suitability or species occurrence rarely validated
(Beck & Suhring 2009; McGarigal et al. 2016).
2.5 Habitat connectivity models
2.5.1 Modelling frameworks
In accordance with the breadth of conceptual perspectives on landscape connectivity (here
‘habitat connectivity’; Section 2.1.5), a variety of modelling frameworks have been developed.
Kool et al. (2013) and Fletcher et al. (2016a) provided excellent syntheses of the most commonly
used approaches. As opposed to the habitat models discussed in Section 2.4.1, most frameworks
explicitly incorporate spatial relationships between habitat patches in geographical space. They
differ, however, in their treatment of structural patterns, and functional dispersal processes, in the
intervening matrix. Approaches also differ in focus, describing connectivity as either an attribute of
the patch or the matrix (Moilanen & Hanski 2001; Calabrese & Fagan 2004; Fletcher et al. 2016a).
Here, I refer to measures of ‘matrix connectivity’ or ‘patch connectivity’ to emphasize this
distinction (Calabrese & Fagan 2004; Kindlmann & Burel 2008; Rayfield et al. 2011).
‘Spatial metapopulation models’ have focused on population dynamics within habitat patches,
depending on their location in relation to other patches (Hanski 1998; Fletcher et al. 2016a).
However, patch connectivity has typically been measured in terms of simple ‘isolation-by-distance’
(e.g. using dispersal kernels and measuring nearest neighbour distance) or patch aggregation (e.g.
using buffer measures) (Moilanen & Nieminen 2002; Calabrese & Fagan 2004; Kindlmann & Burel
2009). Landscape genetic studies have similarly compared genetic distances with Euclidean
distances between discrete subpopulations (McRae & Beier 2007). In either approach, structural
heterogeneity in the matrix and its influence on dispersal was rarely integrated (Kool et al. 2013).
On the other end of the spectrum, ‘spatial pattern indices’ have focused solely on quantifying
aspects of landscape structure in both patches and the matrix. They provided measures of structural
matrix connectivity without explicit reference to an organism or dispersal process (Calabrese &
Fagan 2004; Kindlmann & Burel 2009).
Recently, graph-theoretic approaches have become the dominant paradigm in connectivity
research (Urban et al. 2009; Fletcher et al. 2016a). The basic concept is well-defined by Rayfield et
al. (2011): “In its most basic form, a graph is a set of nodes, some pairs of which are joined by
links. [When applied to habitat patch networks,] a “habitat graph” [is] a collection of nodes (habitat
patches) and links that connect pairs of nodes (representing the potential or frequency of movement
between habitat patches)” (pp. 847 & 849). The framework is extremely flexible and has been used
49
to measure structural or functional connectivity as an attribute of individual patches (‘nodes’),
dispersal corridors (‘links’) or the entire patch network, depending on how nodes and links are
represented (Galpern et al. 2011; Rayfield et al. 2011). For example, links can be directional or
non-directional and informed by Euclidean distance or matrix resistance (Section 2.5.3). Matrix
resistance to dispersal has typically been integrated into patch network graphs via ‘least cost path’
algorithms (Galpern et al. 2011; Etherington & Holland 2013). These algorithms determined the
optimal, most efficient route of dispersal between two nodes (patches) “as a function of the distance
travelled and the costs traversed” (Etherington & Holland 2013, p. 1223).
However, real organisms may rarely use the least cost path or “optimal” route of dispersal.
Further, Moilanen (2011) highlighted that patch network graphs are limited by arbitrary patch
delineation and dispersal thresholds. ‘Circuit-theoretic models’, which are also based on graph and
network theory and have mostly been implemented in the Circuitscape modelling environment
(McRae et al. 2008; McRae et al. 2013), have received increasing attention. Circuitscape treats the
landscape analogous to an electrical circuit, passing current between pairs of nodes (patches)
through an intervening network of resistors (matrix pixels). Circuitscape has been mostly used to
compute the resistance distance of a patch as an improved measure of genetic connectivity (McRae
& Beier 2007). However, the algorithm also computes the current density in each matrix pixel,
which can be interpreted as the probability of a species moving through that pixel via random walk
theory (McRae et al. 2008). As relative flow rates are measured, Circuitscape can identify multiple
alternative connections between nodes rather than a single least cost path. Further, recent technical
developments have allowed computation of movement probabilities (i.e.’ matrix connectivity’)
across large landscapes independent of habitat patches (Koen et al. 2014; Pelletier et al. 2014).
Finally, a disadvantage of all graph-theoretic models is that they are static representations of
reality and typically do not integrate population dynamics in nodes or spatiotemporal variability in
links. More complex ‘dynamic network models’ (Ferrari et al. 2014), ‘spatial metapopulation
models’ (Hanski 1998; Holland et al. 2007; Lurgi et al. 2016), ‘individual-based models’
(Lookingbill et al. 2010; Kanagaraj et al. 2013) or ‘spatially structured diffusion models’ may be
best suited when the goal is to measure actual connectivity and its effects on metapopulation
persistence (Moilanen 2011; Kool et al. 2013; Hastings 2014; Fletcher et al. 2016a).
2.5.2 Habitat patches
Most connectivity modelling frameworks and underlying ecological theories require that
habitat patches are spatially defined as a basic unit of analysis (Girvetz & Greco 2007). Early
approaches have regarded patches as structurally distinct landscape features and attempted to
delineate them based on rules of spatial contiguity. However, structural patches have little basis in
50
ecological theory (Section 2.1.2). Contiguity rules are also sensitive to the spatial grain of analysis,
disregard patch configuration (e.g. shape and edge effects) and composition (e.g. size), and may
overemphasize the disruptive force of narrow intervening areas of poor quality habitat (Beier et al.
2007; Girvetz & Greco 2007). Yet, until Girvetz & Greco’s (2007) important paper, relatively little
progress had been made towards functional patch delineation consistent with ecological theory.
Functional approaches to patch delineation can be applied to multiple biological levels and
time horizons (e.g. resource patches used by an individual for minutes to hours vs. habitat patches
used by local populations for months to years). Patches may also be referenced to specific time
periods to allow for dynamic shifts through time. There is a range of methods available, but all
require the definition of a three quantitative parameters (Girvetz & Greco 2007; Galpern et al. 2011,
Shirk & McRae 2013): (1) a quality threshold above which a landscape element (e.g. a spatial pixel
on a map) is considered as suitable habitat for the organism; (2) a size threshold, which is deemed
necessary for a breeding pair or population to persist without interaction with other patches; and (3)
one or several functions describing the effect of low quality matrix habitat adjacent to or within a
patch. The latter can allow for patch expansion across narrow intra-patch gaps that do not affect the
focal species, or patch removal, for example when edge effects impact negatively on the suitability
of narrow habitat “spurs” (Girvetz & Greco 2007). Parameterization usually relies on prior
empirical knowledge. For example, results from habitat suitability models may be used to define
quality thresholds, while size thresholds are commonly derived from home range estimates or
assumed breeding requirements. Intra-patch gap crossing ability or edge effects may be derived
from empirical knowledge on species mobility and habitat use (Beier et al. 2007; Shirk & McRae
2013). However, Moilanen (2011) warns that multiple thresholding may result in considerable loss
of information.
2.5.3 Matrix resistance
Separating the landscape into suitable patches and an unsuitable matrix can be problematic, for
example when generalist species utilize continuously distributed resources (Hamilton et al. 2006).
Moreover, the shift towards functional perspectives in landscape ecology has led to the recognition
that animal dispersal between patches is influenced by the variable conditions encountered in the
matrix. The purpose of ‘resistance surfaces’ is to quantify structural heterogeneity in the landscape
between patches, or indeed the entire landscape, from the perspective of a specific organism moving
through that landscape (Zeller et al. 2012). Resistance surfaces assign species-specific cost values
to each landscape element (e.g. pixel) and have also been referred to by “combinations of: cost,
friction, permeability, or resistance, and; layer, grid, map, raster, or surface” (Etherington et al.
51
2014, p. 1). They are a crucial, and sometimes the only, input for functional connectivity models
(Beier et al. 2007; Koen et al. 2014).
Resistance surfaces can be parameterized in many ways. Zeller et al. (2012) provide an
excellent review of the different approaches available. The most common types of data used for
model calibration were expert knowledge, followed by information from landscape genetic studies
and single-point observations (Zeller et al. 2012). Direct empirical measures of animals’ long range
(inter-patch) dispersal movements, while conceptually best suited to infer matrix resistance, have
been used infrequently (Zeller et al. 2012; Kool et al. 2013; Fletcher et al. 2016a). The most
common analytical approach was described by Zeller et al. (2012) as the ‘one-stage expert
approach’. Here, biologically informed expert opinion was used to assign resistance values to
landscape elements without further verification against empirical data. When used within broader
resource-based habitat modelling frameworks (Hartemink et al. 2015), landscape resistance has
often been calculated by reversing habitat suitability (Beier et al. 2007; Murray et al. 2014).
However, this relies on the assumption that habitat selection within patches and dispersal between
patches are influenced by the same environmental variables (Beier et al. 2007). For feral pigs, this
assumption may be justified, as dispersal appears to be limited solely by the (gender-specific)
energetic and survival costs associated with moving through poor quality habitat (Section 2.2.4).
2.6 Expert elicitation
Faced with uncertainty, complexity and lack of empirical data, decision-makers have always
relied on the knowledge of experts. The value of expert knowledge to inform landscape ecological
research and modelling has also increasingly been recognised (Perera et al. 2012). The process of
collecting, synthesising and analysing expert knowledge is commonly referred to as ‘expert
elicitation’. Yet, expert knowledge as well as the process of eliciting it are prone to error and bias.
Consequently, a number of recent publications have argued that expert knowledge must not be
perceived as a “cheap, readily available source of knowledge” (McBride 2013, p. 156) and
advocated a structured approach that treats expert elicitation as a form of scientific data collection
(Kuhnert et al. 2010; Johnson et al. 2012b; Martin et al. 2012; McBride et al. 2012a). Like all data,
expert knowledge has limits: it is subjective, incomplete and often unstructured. Elicitation cannot
remove uncertainty or knowledge gaps. Rather, the goal is to minimize cognitive and motivational
biases and provide “an accurate representation of an expert’s true beliefs” (McBride 2013, p. 7). A
structured elicitation approach should follow a three-step protocol: (a) problem definition,
development of an elicitation methodology and expert selection; (b) elicitation pre-training and
actual elicitation; and (c) verification, error analysis, aggregation and validation (McBride 2013).
For expert-based Bayesian networks, Marcot (2006) recommended integrating these steps into an
52
iterative elicitation process that allows for repeated evaluation of model behaviour, validation of
model accuracy and opportunities for expert revision. Within this framework, a range of strategies
and techniques can be used to avoid or minimize bias (Table 2.1).
Table 2.1 Sources of bias in expert elicitation and strategies or techniques for minimizing them as used
in this thesis. (Sources: Kuhnert et al. 2010; McBride et al. 2012a; Fenton & Neil 2013; McBride 2013)
Source of bias
Elicitation strategies and techniques employed in this thesis
Motivation / elicitation fatigue: Experts
⋅ Develop rapport and shared understanding / support for elicitation goals
have other commitments and need to
⋅ Elicit in workshop setting to enhance professional experience
justify time investment to themselves
⋅ Limit time investment and number of answers required
and others.
⋅ Provide rapid feedback and follow-up to maintain interest
Anchoring bias: Expert beliefs are
⋅ Develop a joint system understanding and conceptual models through
influenced by a preconceived estimate
group consensus in a participatory setting
Uncertainty:
⋅ Select experts from diverse backgrounds, preferably well-calibrated ‘expert
⋅ Experts are uncertain of their, or have
practitioners’ that interact with the model on a regular basis
no, knowledge on elements of the
⋅ Stimulate memories and understanding through group interaction
system
⋅ Maintain simplest possible conceptual model structure
⋅ Experts know that there is uncertainty
⋅ Use a probabilistic framework that allows for uncertainty
inherent in elements of the system
⋅ Validate elicitation results against independent data
Overconfidence: Experts overestimate
⋅ Select ‘expert practitioners’
the accuracy of their beliefs
⋅ Use a probabilistic framework
Variability: Experts have different
⋅ Elicit quantitative estimates from individuals and average results
system understanding and uncertainties
⋅ Evaluate model sensitivity to different expert estimates
Dominance: Experts conform to
⋅ Elicit system understanding and quantitative estimates in small groups or
seniority or dominance
individually followed by open review and discussion
Groupthink: Groups want to achieve
⋅ Encourage group participation and consider alternative model scenarios
consensus and dismiss deviations
Linguistic: Experts interpret statements
⋅ Develop clear definitions and shared understanding for all system elements
and questions differently
⋅ Ask simple, unambiguous questions
⋅ Frame probabilities as interval judgements and interactions as weights
⋅ Document elicitation process and results to allow retrospective evaluation
Ambiguity: Experts avoid options where ⋅ Review model structure and behaviour in an iterative process
they have no knowledge or uncertainties ⋅ Delphi approach: elicit individual knowledge, share among the group and
give opportunity for individual revision
53
Chapter 3 Integrating landscape structure improves habitat models of mobile
animals: feral pigs in Australia
Summary
In this chapter I develop a novel combined methodology, spatial pattern suitability analysis, for
integrating resource-seeking home range movements into habitat models of mobile animals 7. This
involves measuring structural patterns of resource quality at the home range scale and then relating
these measures to functional values from expert-elicited response-to-pattern curves. I use scenario
analysis and empirical validation in a subsection of the study region in northern Queensland to
evaluate whether this methodology improves model performance and how structural patterns should
be measured for feral pig breeding herds. This chapter addresses research question 1, providing a
methodology for reliably modelling habitat suitability for feral pigs at the regional scale under
uncertainty about the processes of habitat selection by individual breeding herds. It is being
prepared for submission to Ecological Modelling.
3.1 Abstract
Ecological theory suggests that habitat suitability for mobile animals is influenced by
landscape structure: the spatial patterns of, and relationships between, resource or habitat patches
within heterogeneous home ranges. Yet, many habitat and distribution models of such species either
neglect landscape structure or do not sufficiently evaluate alternative ways of measuring it. We
modelled habitat suitability for feral pigs in northern Australia using a resource-based approach.
Here, we aimed to (a) integrate measures of landscape structure at the home range scale into this
framework, and (b) evaluate which landscape metrics and scales of measurement yield improved
models. We developed spatial pattern suitability analysis, which measured patterns of resource
quality in moving windows and related these structural metrics to functional resource suitability
indices from expert-elicited response-to-pattern curves. We developed eleven alternative habitat
suitability models with and without integrating landscape-scale variables using one of three metrics
(distance-dependent, composition-dependent and combined distance / composition-dependent) and
analysis scales (1, 2 and 3 km radius). All models were empirically validated against independent
distributional data. Habitat suitability models which integrated landscape structure outperformed the
model which did not in 90% of evaluated scenarios. Models that measured resource suitability as a
A detailed description of the methodology for modelling habitat suitability using Bayesian networks, including expert
elicitation procedure, model variables, and spatial data proxies, is given in Chapter 4. This is due to the timing of
journal submissions. The final submission of Chapter 3 to Ecological Modelling will reflect this and refer to Chapter 4
(published in PLoS ONE) for details. Here I focus on the spatial pattern suitability analysis methodology and its
evaluation.
7
54
function of distance performed consistently best against all validation data. Spatial pattern
suitability analysis usefully captured expert knowledge on home range movements. Through model
validation we could determine which landscape variables most improved predictions of habitat
suitability. Our approach is applicable to other mobile species, modelling frameworks and
landscape metrics.
3.2 Introduction
Spatial models of species’ habitat and distributions are now widely applied to inform
management decisions (Millspaugh & Thompson 2009; Venette et al. 2010; Guisan et al. 2013)
from local to global extents (Fletcher et al. 2016b). While methodologically diverse (Guisan &
Thuiller 2005; Beck & Suring 2009; Dormann et al. 2012), all rely on the link between an
organism’s distribution and a combination of habitat variables or environmental requirements
(Grinnell 1917; Hutchinson 1957; Leibold 1995; Pulliam 2000). At the same time, decades of
biogeographical and landscape ecological research have established that species distributions and
habitat use are influenced by landscape structure – the spatial patterns of, and relationships between,
required resources or habitat patches within heterogeneous landscapes (MacArthur & Wilson 1967;
Dunning et al. 1992; Levin 1992; Turner & Gardner 2015) – and that this influence is sensitive to
both the scale and metric used for measuring these patterns (Addicott et al. 1987; Wiens 1989;
McGarigal & Marks 1995; Holland et al. 2004). Despite the theoretical importance of landscape
structure and scale, several recent reviews have highlighted that these issues are not routinely
integrated into contemporary habitat and species distribution models (Beck & Suring 2009;
Moorcroft 2012; McGarigal et al. 2016; Yackulic & Ginsberg 2016). Here, we investigated the
integration of landscape structure into a regional-scale habitat models of mobile animals.
Mobile animals have home ranges, broadly defined as the “area traversed by an individual in
its normal activities of food gathering, mating, and caring for young” (Burt 1943, p. 352). While
many theoretical and operational questions about home range behaviour remain, the conceptual
basis of animals’ space use is widely agreed upon: mobile animals use home ranges to satisfy a
number of heterogeneously distributed habitat requirements (Börger et al. 2008; Powell & Mitchell
2012). Here, we view home ranges as a type of ‘ecological neighbourhood’ (Addicott et al. 1987) in
which resource selection by mobile animals occurs (Fig 3.1). Heterogeneity in this neighbourhood
can be measured using a range of ‘landscape metrics’, which describe aspects of landscape
‘composition’ (the variety and abundance of resources) or ‘configuration (the relative position of
resources) (McGarigal & Marks 1995; Uuemaa et al. 2009; McGarigal et al.2012). Yet, all metrics
of landscape structure, per se, are independent of ecological function, that is, they carry no implicit
meaning to the way a mobile species uses its home range (Kupfer 2012; Turner & Gardner 2015).
55
Here, Dunning et al. (1992) suggested that habitat selection in home ranges is governed by resource
‘complementation’ (several resources are selected for different purposes), ‘supplementation’
(different resources are selected for the same purpose), and ‘neighbourhood effects’ (proximal
resources are selected preferentially to distant resources). The suitability of a site to function as
habitat thus depends on the availability and patterning of complementary and supplementary
resources with the home range, and the species’ behavioural response to these patterns (Fig 3.1).
Fig 3.1 Resource-seeking home range movements of mobile animals. Consider two mobile species (red
and black squiggles) with three resource requirements, all heterogeneously distributed across a landscape
(dark red, light red and grey patches). Landscape portion (A) contains few scattered and portion (B)
abundant contiguous resource patches. Each species uses space according to its mobility and corresponding
home range size (black dashed circles). Suitability of two focal habitat patches (black squares) is influenced
by: the presence of resources in the landscape (greater in portion B than A); the scale of each species’ home
range in which resources are available (red species can access all three resources, black species only two, but
note the difference if home range sizes were reversed (grey dotted circles)); and the behavioural response of
each species to landscape structure (How much of each resource is required? Are distant resources used less
than proximate resources? Are scattered resources used less than contiguous resources?).
Broadly, three methodologies for integrating ‘landscape variables’ that measure spatial
patterning in explanatory variables into habitat models have emerged: Statistical models using
resource selection functions have adopted a ‘multi-scale’ approach, whereby landscape variables
were computed in circular areas of different radii around a sampling site, and ‘scale-optimized’
variables that best explain a response variable were selected (Holland et al. 2004; Jackson & Fahrig
2015; McGarigal et al. 2016; Miguet et al. 2016). Some recent statistical species distribution
models have also adopted the multi-scale approach, but computed landscape variables in moving
windows that are incrementally centred on each ‘focal pixel’ within a study area (Bellamy et al.
2015; Ducci et al. 2015). As scale-dependent insights from empirical multi-scale studies may not be
56
transferrable across large study areas (Paton & Matthiopoulos 2016), resource-based ‘landscape
HSI models’ have used published or expert a priori knowledge to derive suitable landscape metrics
and measurement scales (Dijak et al. 2007; Dijak & Rittenhouse 2009). Landscape variables were
then equally computed in moving windows and, importantly, related to functional suitability indices
that are assumed to hold for entire large landscapes (Dijak et al. 2007; Dijak & Rittenhouse 2009).
Here, we aimed to account for home range movements when modelling habitat suitability for
feral pig breeding and persistence in north-eastern Australia. Feral pigs are descendants of
domesticated Eurasian wild boar. The species has proven a successful invader wherever introduced,
impacting negatively on agriculture, the environment and human and animal health (Barrios-Garcia
& Ballari 2012; Bengsen et al. 2014). Regional-scale information about their distribution and
habitat suitability for persistence is needed to effectively manage these impacts. Feral pigs’ habitat
(resource) selection, space use and distributional patterns have been studied widely. However, the
influence of landscape structure and scaling has rarely been analysed. For example, resource
selection analyses have established that feral pigs prefer habitats which satisfy their physiological
needs for frequent hydration (Caley 1997), thermoregulation (Dexter 1998) and a protein-rich diet
(Giles 1980; Caley 1997; Dexter 1998; Ross 2009; Wurster et al. 2012). Yet, only Choquenot &
Ruscoe (2003) explicitly investigated resource complementation at different scales and found that
food resources and thermal refuge must co-occur within 5 km to facilitate feral pig persistence.
Home range analyses have found movements to be limited by the seasonal availability of these
resource requirements, especially in breeding herds of related sows and their young (Caley 1997;
Dexter 1999; Mitchell et al. 2009). Female home range sizes ranged widely (1-20 km2 – Hone
2012; Choquenot et al. 1996), but possible effects of landscape structure remain unclear.
We adopted a resource-based modelling framework using Bayesian networks. When used to
model habitat (Marcot et al. 2001; Smith et al. 2007; Tantipisanuh et al. 2014; van Klinken et al.
2015), this approach offers flexibility to use expert-elicited knowledge where empirical data
availability is limited, a robust statistical framework for modelling interactions between
complementary resources, and an explicit treatment of uncertainty (Uusitalo 2007; Wilhere 2012).
Our objectives were: (A) To develop a methodology for integrating landscape structure at the home
range scale into this resource-based modelling framework. We limited our analysis to three simple,
yet ecologically meaningful, landscape metrics: distance-weighted resource quality (a measure of
landscape configuration), average resource quality (a measure of landscape composition) and a
combined measure, distance-weighted average resource quality. (B) To test whether such
integration actually improves our model and, if so, which metrics and scales of measurement best
explain observed feral pig presences across the study region. Validation is critical especially for
57
expert-elicited models (van Klinken et al. 2015). We evaluated ten alternative ‘landscape models’
(which used landscape-scale habitat variables using either of the three metrics at either of three
expert-defined measurement scales) and one ‘control model’ (which used site-scale habitat
variables) against independent distributional data, expecting landscape models to perform better.
3.3 Methods
3.3.1 Study region
Our study region in north-eastern Australia (Fig 3.2) is characterised by a hot tropical climate
with a highly seasonal rainfall regime. Ecosystems are dominated by monsoonal savanna
woodlands, extensively utilized for cattle grazing and interspersed with fragments of wet tropical
rainforest, coastal wetlands and open semi-arid grasslands (Fox et al. 2001). Intensive uses are
limited to the fertile coastal lowlands and plateaus of the eastern ranges. According to previous
estimates, feral pigs are common to widespread across the region (West 2008), yet distribution and
abundance varies considerably with climatic and seasonal conditions (Choquenot et al. 1996). We
limited our habitat model to the late dry season (October to November), when resources are most
constrained and scattered across the landscape. Increased heterogeneity under such conditions may
pronounce the influence of landscape structure on habitat suitability.
Fig 3.2 Study region in north-eastern Australia. The study region is shown in grey. Locations of the
independent distributional data sets used for model validation are shown in colour (details in Table 3.1).
58
3.3.2 Modelling approach
In order to evaluate which landscape metric and scale of measurement best explains feral pig
presence in north-eastern Australia, we developed eleven alternative habitat suitability models. Our
modelling approach (Fig 3.3) adapted the conceptual framework described in Marcot et al. (2001),
Smith et al. (2007) and van Klinken et al. (2015): habitat suitability was defined in terms of feral
pigs’ ability to breed and persist over time (Morrison et al. 2006); this depended on a set of habitat
variables representing resource requirements; each habitat variable was itself influenced by several
explanatory variables, which were linked to one or several spatial data proxies.
First, we modelled the quality of each habitat variable (henceforth termed ‘resource quality’),
conditional on its explanatory variables, in separate expert-elicited Bayesian networks. Then, we
used ‘spatial pattern suitability analysis’ to capture feral pigs’ selection of supplementary resources
within home ranges. This involved measuring spatial patterns of resource quality within moving
windows and relating structural metrics to functional ‘resource suitability indices’ (SIr) from expertelicited response curves. Here, we computed distance- (SIDr), composition- (SICr) and combined
distance / composition-dependent (SIDCr) indices at three scales of measurement each (1, 2 and 3 km
moving window radii). Finally, we modelled habitat suitability, conditional on all habitat variables,
in another expert-elicited Bayesian network. Ten ‘landscape models’ integrated landscape-scale
variables (SIr), while one ‘control model’ used site-scale variables (xr) (Fig 3.3).
Fig 3.3 (next page) Methodology for modelling habitat suitability for feral pig breeding. Major steps
with associated methods and implementation tools. Steps 1 and 2 are illustrated using the example of food
resources. Analogous procedures were applied to the other three identified habitat variables (water,
protection from heat stress and protection from disturbance). Bayesian network results are shown as expertelicited probability distributions (bar graphs) and modelled index values ± standard deviation (below bar
graphs), conditional on uniformly distributed prior probabilities. Spatial pattern suitability analysis (here
shortened to PATTSI) was used to compute three alternative landscape-scale resource suitability indices
(SIr): distance- (SIDr), composition- (SIr) or combined distance / composition-dependent (SIDCr). Step 3 is
illustrated using SIDr. Table 3.2 details landscape variables used in habitat suitability models 1-10.
59
Step
Methods
Tools
r {raster} / ArcGIS
– spatial data
pre-processing
Food quality
High
Moderate
Low
High
Moderate
Low
Expert-elicited
Bayesian network
model
Food quality index
1. Model resource
quality indices
19.5
Very good
23.2
Good
22.5
Moderate
13.9
Poor
21.0
Very poor
51.2 ± 29
Scenarios 1 – 10
(with PATTSI)
ArcGIS –
output mapping
2. Spatial pattern
suitability analysis
Expert-elicited
response curves
Focal
(max)
Focal
(mean)
Focal
(mean)
SIDr
SIDCr
SICr
r {raster} –
focal (max) /
ArcGIS Spatial
Analyst –
focal (mean)
Scenarios 1 – 10
(with PATTSI)
Scenario X (without PATTSI)
Food quantity
33.3
33.3
33.3
33.3
33.3
33.3
3. Model habitat
suitability index
Heat protection suitability index
Food suitability index
Very good
20.0
Good
20.0
Moderate
20.0
Poor
20.0
Very poor
20.0
Very good
Good
Moderate
Poor
Very poor
Habitat suitability index
Very high
2.01
High
12.0
Moderate
26.1
Low
21.3
Very low
38.6
33.5 ± 23
Water suitability index
Very good
20.0
Good
20.0
Moderate
20.0
Poor
20.0
Very poor
20.0
Disturbance protection suitability index
Very good
Good
Moderate
Poor
Very poor
20.0
20.0
20.0
20.0
20.0
20.0
20.0
20.0
20.0
20.0
60
Expert-elicited
Bayesian network
model
3.3.3 Resource quality models
Models were calibrated by iteratively eliciting knowledge from a panel of experts with
scientific or field knowledge of feral pigs in north-eastern Australia (Marcot 2006). Elicitation
followed a structured process and used a range of strategies and techniques for minimizing sources
of bias (Table 2.1). Experts identified four resource requirements for feral pig breeding and
persistence: water, food, protection from heat stress and protection from disturbance. We modelled
resource quality as a function of several explanatory variables in probabilistic Bayesian networks
using the Norsys Netica 5.12 and AgenaRisk 6.1 software. Model explanatory variables were linked
to remotely sensed or mapped GIS data layers, which were rasterized and resampled to a common
extent and resolution (100m), then reclassified using the R ‘raster’ package (R Core Team 2015;
Hijmans 2015) and ESRI ArcGIS 10.2 software. Time-specific spatial data were averaged over two
months (October / November) and five years (2010-2014) prior to reclassification to
representatively capture late dry season conditions. We discretised resource quality into five equal
classes and assigned each with a numerical range (0-20 for the poorest class, …, 80-100 for the best
class). Spatially-explicit ‘resource quality indices’ (xr) were computed as expected values by
summing the mid-point value of each class weighted by its probability of occurrence. Accordingly,
xr could range between 10 (mid-point of the poorest class) and 90 (midpoint of the best class). An
example model is shown in Fig 3.3. Tables defining all variables, conditional probabilities and
spatial data proxies are detailed in Appendix 3.1 8.
3.3.4 Spatial pattern suitability analysis
Spatial pattern suitability analysis combined to methodological steps: (a) spatial moving
window analysis to measure spatial patterns of resource quality (b) expert elicitation to relate
structural metrics to functional resource suitability indices (Appendix 3.2).
3.3.4.1 Expert elicitation
We conducted semi-structured interviews with selected individuals from our panel of feral pig
experts (n = 6). First, we asked each expert to define the home range ‘ecological neighbourhood’
(Addicott et al. 1987) of feral pig breeding by specifying a ‘mobility threshold’ beyond which
resources are inaccessible. We assumed that the same threshold applies to all habitat variables.
Experts estimated mobility thresholds at either 1 km (n = 1), 2 km (n = 2) or 3 km (n = 3) which
8
Appendix tables 3.1.1 – 3.1.3 in the present version of this manuscript (prepared for submission to Ecological
modelling) have since been included in the main body and appendix of a manuscript published in PLoS ONE (Chapter
4). They are included as Table 4.1, Table 4.2 and Appendix Table 4.1 in this thesis. The final submission of Chapter 3
will reflect this and refer to Chapter 4 (published in PLoS ONE) for details.
61
corresponded to assumed home ranges of approximately 3, 12 and 28 km2. These behavioural
thresholds determined the three scales of measurement evaluated in this study.
Second, we asked experts to relate structural metrics of resource patterns to their functional
value for feral pig breeding. Similar to other studies (Larson et al. 2009; Dijak & Rittenhouse
2009), we expressed these values as ‘resource suitability indices’ (SIr). Like resource quality,
suitability was discretised into five equal classes (0-20 for the poorest class, … , 80-100 for the best
class). We focused elicitation on two metrics: distance and average quality.
Distance was defined relative to the mobility threshold, and discretised into five equal distance
bands (‘very close’, ‘close’, ‘medium’, ‘far’ and ‘very far’). For each habitat variable, we asked
experts to relate each distance band to a corresponding suitability class under the assumption that
other factors (e.g. amount) do not constrain suitability. Average quality depended on both the
amount and quality of accessible resource patches. Hence, different patterns (e.g. many low quality
resources or few high quality resources) could yield similar average values. Again, we defined five
equal average quality classes (‘very high’, ‘high’, ‘moderate’, ‘low’ and ‘very low’), relative to the
maximum value (100% cover with resources of highest quality xr max). Experts related each to a
corresponding suitability class, assuming no other constraints. Thus, each expert defined two stepwise response-to-pattern curves fDr (distance-dependent measurement) and fCr (compositiondependent measurement) per habitat variable. Finally, we averaged response curves for each habitat
variable across experts for input into moving window analysis (Fig 3.3).
3.3.4.2 Moving window analysis
Moving window analysis has been extensively used to investigate landscape structure (Dijak et
al. 2007; Ducci et al. 2015): it computes the value of a focal pixel by summarizing the values of all
neighbouring pixels contained within a spatial analysis window. This window is incrementally
shifted and centred on each pixel within a study region, enabling analysis of fine-grained patterns
within overlapping neighbourhoods (Larson et al. 2009). The moving window size and shape must
be specified and many landscape metrics are available to compute focal pixel values (McGarigal et
al. 2012). We used circular moving windows with radii corresponding to each of the three elicited
scales of measurement (1, 2 and 3 km). Three SIr were computed in each window (Fig 3.3),
resulting in nine alternative indices overall. Distance-dependent resource suitability (SIDr) was the
highest distance-weighted resource quality index (xr) within the moving window, where weights
were derived from response curves fDr. Composition-dependent suitability (SICr) was computed by
reclassifying average resource quality according to response curves fCr. To compute combined
distance / composition-dependent suitability (SIDCr), we also used fCr to reclassify average resource
quality. However, each individual resource quality index was also weighted according to response
62
curves fDr. Analyses were implemented in R 3.2.2 (R Core Team 2015) and ESRI ArcGIS 10.2
software (Appendix 3.2).
Fig 3.4 schematically illustrates that computed SIr at a focal pixel could vary considerably
depending on how spatial patterns are measured (Froese et al. 2015). Consider spatial resource
patterns A and B, which return the same SIDr (top left-hand plot), but very different SICr and SIDCr
(middle and bottom left-hand plot). Consider pattern C, which returns moderate SIr when measured
in a large window (black boxes in left-hand plots), but SIr = 0 when measured in a smaller window
(red boxes in left-hand plots). Non-linear response curves may alter these effects somewhat.
Fig 3.4 Computing resource suitability indices using spatial pattern suitability analysis. Consider a
resource of highest quality xr max (grey pixels) and various spatial patterns around a focal pixel (black pixel):
A = close and abundant; B = close and scarce; C = far and abundant; D = far and scarce. Three alternative
resource suitability indices (SIr) are computed within large (black dashed line) or small (red dashed line)
moving windows as a function of distance (SIDr, top left-hand plot), average quality (SICr, middle plot) or
distance-weighted average quality (SIDCr, bottom plot). We plot approximate SIr of the focal pixel
corresponding to each pattern A-D and moving window size (red boxes = small, black boxes = large)
assuming linear distance- and composition-dependent responses.
3.3.5 Habitat suitability model
We modelled habitat suitability for feral pig breeding and persistence as a function of four
complementary habitat requirements. Analogous to resource quality, models were implemented as
63
expert-elicited Bayesian networks and spatially-explicit habitat suitability indices (HSI) computed
by summing the mid-point values of five discrete classes weighted by their probabilities of
occurrence (Fig 3.3; Appendix Table 4.1E). We developed eleven alternative habitat suitability
models. For nine ‘landscape models’ we computed all four SIr using the same landscape metric and
scale of measurement. As feral pigs may respond differently to each habitat variable, we added a
‘best knowledge’ landscape model (derived from a combination of expert opinion and literature
review) with resource-specific SIr (distance-dependent SID water and SID heat at 3 km radius, combined
distance / composition-dependent SIDC food and SIDC disturbance at 2 km and 1 km radius respectively).
Lastly, a ‘control’ model used site-scale resource quality indices (xr) without spatial analysis.
3.3.6 Model evaluation and validation
3.3.6.1 Validation data
We validated predicted HSI from all eleven habitat suitability models against three independent
data sets of feral pig presence sourced from collaborators (Table 3.1). We used only presence
records corresponding to breeding herds (identified as female or with a group count greater than
two) and collected during the late dry season. For each data set, we defined a validation background
(shaded in red, blue and black in Fig 3.2) representative of surveyed habitat types from existing
management units (National Park boundaries for Lakefield) or by applying a 15km buffer to
presence records (for Balkanu and NAQS).
Table 3.1 Validation data sets with ancillary information.
Name
Balkanu
Source
Balkanu Cape
No. of Date of
Method and purpose
Background Typical habitat types
records collection
of collection
size (km2)
181
Sep-Nov
Systematic aerial survey 3,954
Eucalyptus woodlands
2013-14
and management
& coastal wetlands
York Development
Corporation
Lakefield
Queensland
(shooting)
350
Parks & Wildlife
Oct-Dec
Systematic aerial
2009-13
management (shooting)
5,788
woodlands, coastal
Service
NAQS
Northern Australia
Quarantine Strategy
Eucalyptus / Melaleuca
wetlands & grasslands
103
Sep-Nov
Opportunistic aerial
11,630
Eucalyptus woodlands,
2007-10
survey and disease
coastal grasslands &
sampling (shooting)
chenopod scrublands
Presence records for mobile species may contain spatial error for many reasons (Hunsacker et
al. 2001; Boyce et al. 2002). As our data sets had been collected by third parties, we faced
64
considerable uncertainty about such error. Similarly, remotely sensed or mapped data layers may
contain positional inaccuracies or misrepresent landscape features due to their mapping resolution
(Hunsacker et al. 2001). Together, such spatial error may yield noisy results that obscure main data
trends. To reduce this noise, we upscaled both predicted HSI and presence records to a 1km
resolution. We subsequently thinned data points to ensure independence, allowing only one
presence record collected on the same day within a given 1 km pixel.
3.3.6.2 Validation metrics
We used the Continuous Boyce Index (CBI) method described by Hirzel et al. (2006) after
Boyce et al. (2002) to evaluate model performance. The CBI was developed specifically for
situations where a numerical prediction such as our HSI is validated against presence-only
observations (Boyce et al. 2002). HSI was partitioned into n overlapping bins b of width w, which
were incrementally shifted upwards along the total range of HSI (HSImax – HSImin) by resolution
factor r (Hirzel et al. 2006). In our analysis, we set w to 10 and r to the default 1/100th of HSImax –
HSImin. The predicted-to-expected (P/E) ratio was then computed as the (predicted) proportion of
presence records in each bin b divided by the (expected) proportion of the validation background
covered by that b and plotted against the average HSI of b. For a good model, this results in a
monotonically increasing curve (Hirzel et al. 2006). The CBI measures the Spearman rank
correlation coefficient of P/E against HSI and varies from 1 (correct model, P/E steadily increases
as HSI increases) to –1 (false model, P/E steadily decreases with increasing HSI), with values close
to zero indicating a random prediction (Hirzel et al. 2006). We computed the P/E ratio and CBI for
each of the three validation data sets independently. We also computed the total proportion of each
validation background expected to be highly or very highly suitable habitat (HSI ≥ 60). Analyses
were implemented in R 3.2.2 (R Core Team 2015; Appendix 3.3).
3.4 Results
3.4.1 Spatial pattern suitability analysis
Expert-elicited response-to-pattern curves fr approximated either of three general shapes (see
Froese et al. 2015): linear decay (SIr declines steadily with increasing distance to, or decreasing
average quality of, a resource), exponential decay (SIr declines at a faster rate) and inverted
exponential decay (SIr declines at a slower rate). Individual experts’ fr are provided in Appendix
Table 3.1.4. After averaging across experts, different responses emerged for the four habitat
variables (Fig 3.5). When describing resource suitability for feral pig breeding as a function of
distance (fDr), these differences were substantial (Fig 3.5A). For instance, at ‘medium’ distance,
water resources were still considered highly suitable (SID water = 80) while disturbance refuges
65
already had poor functional value (SID disturbance = 22). The distance-dependent suitability of food
resources and heat refuges approximated linear decay (both SID food / heat = 58 at ‘medium’ distance).
The relationship between suitability and average quality (fCr) showed similar, but less pronounced,
differences between habitat variables (Fig 3.5B). For example, while 50% average resource quality
always corresponded to high SICr (SIC water = 90, SIC heat = 86, SIC food = 77 and SIC disturbance = 70), the
functional value at 25% and 15% average quality remained considerably higher for water (SIC water =
83 and 50) and heat (SIC heat = 78 and 50) than for food (SIC food = 53 and 30) and disturbance (SIC
disturbance
= 50 and 26).
Fig 3.5 Averaged expert-elicited response-to-pattern curves relating structural patterns of resource
quality to functional suitability indices. Panels A and B compare resource suitability indices for water,
food, protection from heat and protection from disturbance in response to distance (fDr) and average quality
(fCr) respectively. Distance-dependent response-to-pattern curves cross the x axis at a ‘mobility threshold’
beyond which resources are inaccessible.
3.4.2 Model evaluation and validation
Validation using the Continuous Boyce Index (CBI) method showed that integrating landscape
structure improves model performance. Overall, we evaluated 30 landscape model (1-10) /
validation data combinations using various landscape-scale habitat variables and three control
model (X) / data combinations using site-scale habitat variables. Validation metrics are shown in
Table 3.2 and corresponding P/E ratio curves in Appendix Fig 3.4. Results differed between
validation data sets. Yet, when comparing models within each data set, integrating landscape
structure improved CBI values in 27 out of 30 scenarios (Table 3.2): When validated against the
66
Balkanu data, all landscape models 1-10 performed very well (CBI 0.72 to 0.97) while the control
model performed poorly (CBI 0.12). For the NAQS data, landscape models also had higher CBIs
than the control model (CBI -0.88), but performance remained mostly poor (CBI -0.25 to 0.72). For
the Lakefield data, certain landscape variables improved and others depressed model performance
compared to the control model. Yet, the latter’s relatively high CBI (0.66) concealed erratic peaks in
the P/E ratio curve rather than reflecting a steady positive relationship between P/E and HSI (Fig
3.6C).
Table 3.2 Performance of the eleven habitat suitability models against three validation data sets.
Habitat suitability models with
Validation metrics per validation data set
landscape-scale variables
Balkanu (n =239)
Lakefield (n =371)
NAQS (n =103)
ID Measurement
Scale
HSI ≥ 60
CBI
HSI ≥ 60
CBI
HSI ≥ 60
CBI
1
Distance (SIDr)
1 km
34%
0.86
18%
0.85
16%
0.32
2
Distance (SIDr)
2 km
47%
0.96
27%
0.96
27%
0.72
3
Distance (SIDr)
3 km
57%
0.97
35%
0.97
38%
0.30
4
Composition (SICr)
1 km
22%
0.74
8%
0.71
5%
-0.01
5
Composition (SICr)
2 km
21%
0.91
5%
0.71
5%
0.16
6
Composition (SICr)
3 km
20%
0.87
4%
-0.14
4%
0.17
7
Distance / Composition (SIDCr)
1 km
21%
0.82
8%
0.80
5%
-0.25
8
Distance / Composition (SIDCr)
2 km
21%
0.95
6%
0.50
5%
0.05
9
Distance / Composition (SIDCr)
3 km
21%
0.89
5%
0.47
5%
0.31
various
55%
0.88
12%
0.96
16%
-0.20
N/A
6%
0.12
6%
0.66
2%
-0.88
10 a Best Knowledge (various SIr)
X
Site-scale variables (xr)
We show the Continuous Boyce Index (CBI) and proportion of validation background expected to be highly or very
highly suitable habitat (HSI ≥ 60). A CBI = 1 would indicate a perfectly accurate, a CBI ~ 0 a random, and a CBI < 0 a
false model.
a
The following resource-specific landscape variables were used for landscape model 10: SID water and SID heat at 3 km,
SIDC food at 2 km and SIDC disturbance at 1 km.
Results also varied between the ten evaluated landscape models. Only landscape models
computed from distance-dependent SIDr (1-3) consistently performed well (Table 3.2 bold values)
across all scales of measurement (1, 2 and 3 km). Model 2 (using SIDr computed at 2 km moving
window radius) was the only model to yield very good results against all validation data (CBI 0.72
to 0.96). In contrast, landscape models computed from composition-dependent SICr (4-6), combined
67
distance / composition-dependent SIDCr (7-9) or resource-specific SIr (10) had mixed results (Table
3.2). While all these models performed very well against the Balkanu data (CBI 0.74 to 0.95) and
no better than a random model against the NAQS data (CBI -0.25 to 0.31), validation against the
Lakefield data highlighted differences (CBI -0.14 to 0.96). Specifically, performance was generally
below models 1-3 (except model 10) and CBI values decreased at larger scales of measurement.
Even where CBIs were high, P/E ratio curves did not display the monotonic increase characteristic
of good models (Hirzel et al. 2006, e.g. Fig 3.6B: model 5 vs. Fig 3.6A: model 2). Moreover, when
examining each scale of measurement individually (1, 2 or 3 km), landscape models computed from
SIDr always outperformed models computed from SICr or SIDCr. Despite these inconsistent overall
results for models 4-10, some individual model / data combinations were among the best of all
those evaluated (e.g. model 8 / Balkanu data: CBI 0.95 or model 10 / Lakefield data: CBI 0.96, both
with steadily increasing P/E ratio curves, see Appendix Fig 3.4).
Finally, we also analysed the proportion of validation background expected to be highly or
very highly suitable habitat (HSI ≥ 60, Table 3.2). Similar to CBI values, integrating landscape
structure increased the proportion of HSI ≥ 60 in 26 out of 30 cases and results varied between
landscape models and validation data (Table 3.2). All control model (X) / data combinations
expected very low proportions of suitable habitat (HSI ≥ 60 2% to 6%, e.g. Fig 3.6C). Models using
distance-dependent SIDr (1-3) progressively increased this value with rising scales of measurement
(up to 57% for model 3 / Balkanu background, e.g. Fig 3.6A). Models 4-9 (using SICr or SIDCr) also
mostly increased the expected proportion of HSI ≥ 60, but to a lesser degree and somewhat
differently for each validation data set (e.g. Fig 3.6B). Model 10 resulted in moderate to large
increases. Interestingly, three of the landscape models which did not expect a greater proportion of
suitable habitat than the corresponding control model (i.e. models 6, 8 and 9 for the Lakefield
background), also had reduced CBI values.
68
Fig 3.6 Predicted-to-expected (P/E) ratio and corresponding habitat suitability maps for three
alternative models in the Lakefield area. Presence records from Lakefield (see Table 3.1 for details and
Fig 3.2 for location) are also shown. Landscape variables used for modelling are shown in Table 3.2.
69
3.5 Discussion
The results of this study demonstrate that integrating landscape structure improves the
predictive performance of a process-based habitat model of feral pigs in north-eastern Australia. We
are confident in our results, as the predictions from our expert-elicited models were empirically
validated against three truly independent distributional data sets. Improved performance was
observed in 90% of the 30 evaluated landscape model / validation data combinations. However,
results varied between the ten alternative landscape models and three validation data sets. Models
that used distance-dependent resource suitability indices performed consistently best against all
data. In turn, we discuss (a) the ecological significance of observed results, (b) the benefits of our
approach using expert-elicited Bayesian networks and response-to-pattern curves, and (c)
limitations of models, spatial analyses and evaluation methods with suggestions for future research.
3.5.1 Ecological significance
Our study was driven by landscape ecological principles. Because mobile animals use home
ranges to satisfy their habitat requirements (Burt 1943; Powell & Mitchell 2012), and these
requirements are heterogeneously distributed across landscapes (Turner & Gardner 2015), spatiallyexplicit habitat models must integrate landscape structure at the home range scale. We aimed to
integrate these principles into a resource-based modelling framework using expert-elicited Bayesian
networks that is suitable for regional-scale applications, in our case feral pigs in northern Australia.
Feral pigs have distinct habitat requirements affecting their ability to breed and persist – water,
food, heat refuge and disturbance cover (Choquenot et al. 1996) – which were also identified by our
panel of experts. Moreover, they are both mobile, i.e. can access required resources at different
locations in the landscape, and sedentary, i.e. restrict their resource-seeking movements to defined
home ranges (Caley 1997; Dexter 1999; Mitchell et al. 2009). The results from this study indicate
that spatial pattern suitability analysis can usefully capture feral pigs’ resource selection within
home ranges and allow for improved predictions of habitat suitability. We suggest that our control
model underestimated habitat suitability by failing to account for the use of supplementary and
complementary resources (Dunning et al. 1992), thereby limiting high HSI to locations that
provided all habitat requirements. We expect similar findings for other mobile animals.
We also sought to explore which landscape metrics and scales of measurement are best suited
to reliably modelling habitat suitability for feral pig breeding over a large study region. Here, we
evaluated ten alternative landscape models integrating different metrics and scales. Validation
results showed that not all models were equally able to explain observed feral pig presences across
the study region.
70
Firstly, results varied considerably among evaluated landscape models. Models computed from
distance-dependent SIDr accurately described observed feral pig presences at all measurement
scales. Yet, only one model (intermediate scale, r = 2 km) performed very well against all data. In
contrast, models integrating composition-dependent SICr or combined distance / compositiondependent SIDCr displayed mixed, scale-dependent results. Both of the latter metrics emphasized the
value of large contiguous resource patches and discounted that of small scattered patches (Froese et
al. 2015). Predicted suitable habitat was therefore confined to resource-rich core areas and may
have failed to capture two types of actual habitat, and feral pig presences, in the validation data: (a)
the wider neighbourhood of core areas was not modelled as suitable especially at a narrow scale of
measurement; and (b) small resource patches used by feral pigs to move between core areas were
“averaged out” and modelled as unsuitable especially at a broad measurement scale. Individual
well-performing model / data combinations may have reflected a balance between broadening core
areas and capturing marginal resource patches under certain local conditions. In contrast, distancedependent measures simply extended the “accessibility” of resources to larger neighbourhoods,
increasingly so at larger measurement scales (Froese et al. 2015), thereby predicting more suitable
habitat and explaining additional feral pig presences. We suggest that this simple landscape metric
may be less prone to model error from expert uncertainty about resource utilization and movement
capacity. It could be suited to situations where empirical data is scarce or where some
‘overprediction’ of suitable habitat is preferred to the false omission of actual habitat (e.g. invasive
species management; Jiménez-Valverde et al. 2011).
Secondly, there were considerable differences between validation data sets. Most notably,
model performance was generally high against Balkanu or Lakefield and poor against NAQS data.
This may have been due to (a) error in the NAQS data, (b) error in the spatial data linked to model
explanatory variables in this area, or (c) failure of Bayesian networks models to adequately describe
local feral pig habitat (van Klinken et al. 2015). Interestingly, NAQS-validated landscape models
still improved considerably on the control model, with model 2 achieving a very high CBI. This
suggests that poor performance is unlikely due to a false model, but problems in predicting feral pig
presences from the location of resource patches alone. The NAQS area is characterised by drier
grassland and woodland habitats than the woodland and wetland-dominated Lakefield and Balkanu
areas. Under such conditions, feral pigs may more often traverse, and be detected in, marginal
habitat between scattered resource patches. Time-averaged spatial data layers may also have failed
to capture resources which were actually available at the time of data collection. The NAQS data
with its smaller sample size (n = 103) and opportunistic approach to data collection may have been
more susceptible to such bias than the comprehensive Lakefield and Balkanu data (n = 350 and
181). Landscape models 1-3, which used distance-dependent landscape variables, may have been
71
better able to reflect inter-patch movements and weaken bias. Nevertheless, other sources of error
such as local differences in habitat requirements or behaviour cannot be dismissed. Van Klinken et
al. (2015) describe a range of methods to systematically test for error and uncertainty in spatiallyexplicit Bayesian network models. We suggest that empirical studies of feral pigs’ movement
ecology and habitat utilization in resource-poor environments as well as remote or field validation
of spatial data layers may contribute most to improving our models.
3.5.2 Benefits of the modelling approach
Methodologically, the application of resource-based habitat models using Bayesian network or
other frameworks to mobile animals was improved by accounting for resource-seeking home range
movements. Spatial pattern suitability analysis essentially converted site-scale into landscape-scale
variables similar to previous statistical or expert-based approaches (Dijak et al. 2007; Martin &
Fahrig 2012; Ducci et al. 2015; McGarigal et al. 2016). However, our methodology offered four
main advantages: First, resource suitability indices were computed from response-to-pattern curves.
These were explicitly referenced to feral pig behaviour (i.e. the ability of breeding herds to access
and survive on available resources), which enhances transferability across large study regions
(Larson et al. 2009). Second, eliciting this information from experts was cost-effective and provided
transparent hypotheses (Appendix Table 3.1.4) that can be easily updated if new knowledge
becomes available. The structured elicitation method for relating measured spatial patterns to
corresponding functional response (suitability) values was easily implemented and understood by
expert practitioners. It may prove a useful addition to the landscape ecology toolbox, especially in
situations where multi-scale empirical studies are not feasible due to time, resourcing or data
constraints. Scales of measurement were also elicited a priori from experts based on their field
knowledge. While validation did not indicate selection of an ‘optimal’ scale for any of the
landscape variables, expert assumptions were broadly consistent with the range of empirical home
range estimates from Australia (3-28 km2 in this study compared to 1-20 km2 in Choquenot et al.
1996). Third, landscape variables were derived from numerical site-scale indices rather than
categorical spatial proxies (e.g. vegetation type, land use). These indices directly represented the
quality of resources as perceived by feral pigs (Moorcroft 2012). They were also modelled in
Bayesian networks from a suite of spatially-explicit explanatory variables. Hence, different sitescale conditions could yield similar quality values (e.g. food with high biomass and energy content
vs. high quality and protein content), which usefully reflected Dunning et al.’s (1992) concept of
resource supplementation. Finally, we have systematically tested how models based on different
landscape metrics and scales of measurement affect model performance. While this is common in
statistical multi-scale studies, resource-based habitat models that may be preferred for large study
72
regions or when data is limited, have rarely rigorously evaluated assumptions behind explanatory
variables (Beck & Suring 2009).
3.5.3 Limitations
3.5.3.1 Modelling approach
Firstly, we focused on three simple landscape metrics that were ecologically meaningful and
easy to relate to experts’ field knowledge. Yet, resource suitability for feral pig breeding may
depend on other metrics such as fragmentation (Jackson & Fahrig 2016), edge proximity (Dijak &
Rittenhouse 2009) or heterogeneity (McClure et al. 2015). Secondly, we asked each expert to
specify one scale of measurement applicable to all habitat variables and landscape metrics. Yet,
theory (Addicott et al. 1987; Wiens 1989) and evidence (Jackson & Fahrig 2014; Miguet et al.
2016) suggest that species may in fact respond to different habitat variables and aspects of
landscape structure at different scales. Both could be explored through further expert elicitation or
multi-scale empirical studies (Jackson & Fahrig 2015), or by evaluating other combinations of
resource-specific SIr than in our ‘best knowledge’ model. Thirdly, although widely used, circular
moving windows may not reflect the shape of feral pigs’ actual home ranges (Dexter 1999).
Knowledge from movement ecological studies (Morelle et al. 2014) in combination with technical
advances that allow for dynamic moving windows (Berry 2013) may further improve results.
Finally, we generalized local knowledge to the regional extent by averaging. While it has been
suggested that average expert estimates are more accurate than either individual or consensus
estimates (McBride et al. 2012b), model uncertainty from divergent knowledge could be further
analysed (Johnson et al. 2012b; Murray et al. 2012). In particular, future research could aim to
adapt spatial pattern suitability analysis for probabilistic habitat variables rather than our
deterministic resource quality indices xr.
3.5.3.2 Model evaluation and validation
We used the Continuous Boyce Index to evaluate model performance (Boyce et al. 2002;
Hirzel et al. 2006). While this method is well suited to presence-only validation data, it provides no
information on model specificity, i.e. its ability to correctly predict absences and minimize false
positives (Hirzel et al. 2006; Jiménez-Valverde et al. 2011). Secondly, CBI values alone may be
misleading in some instances. For example, P/E ratio curves that are steadily increasing along a flat
slope may result in a high CBI but not necessarily reflect a good model. Systematically collected
presence / absence data, especially from resource-poor environments, could facilitate confusion
matrix-based validation of sensitivity vs. specificity and is suggested as a future research priority.
73
3.5.4 Conclusion
Our results highlight that habitat models of mobile animals must integrate landscape structure.
We developed a novel combined methodology, spatial pattern suitability analysis, for capturing
feral pigs’ resource-seeking home range movements based on expert-elicited response-to-pattern
curves and spatial moving window analysis. Based on landscape ecological principles, this
methodology improved the application of resource-based Bayesian networks models to mobile
animals. Through comprehensive scenario analysis and validation, we were able to determine which
of various landscape-scale variables allowed for improved predictions of habitat suitability. Our
approach could be applied to other mobile species, modelling frameworks and landscape metrics,
for example to compute ecologically meaningful landscape predictors for statistical analyses or to
integrate expert knowledge on species mobility and space use into resource-based habitat models.
Built on a priori knowledge, our methodology may be particularly useful to support regional-scale
management decisions where empirical data is limited.
3.6 Appendices
Appendix 3.1 Expert-elicited response-to-pattern curves.
Appendix 3.2 Methodology for spatial pattern suitability analysis (annotated R code).
Appendix 3.3 Methodology for model validation (annotated R code).
Appendix 3.4 Validation plots for eleven habitat suitability models.
74
Chapter 4 Modelling seasonal habitat suitability for wide-ranging species:
invasive feral pigs in northern Australia
Summary
In this chapter I apply the methodology from Chapter 3 to model and map seasonal habitat
suitability for feral pig breeding and persistence in northern Australia. I provide a detailed
description of the resource-based, spatially-explicit modelling approach using expert-elicited
Bayesian networks and spatial pattern suitability analysis to account for resource-seeking home
range movements. I compare modelled habitat suitability for a wet season and a dry season scenario
and validate model accuracy against four independent distributional data sets per scenario. This
chapter addresses research question 2, providing an indicative measure of the seasonal
establishment risk of directly transmitted infectious diseases within feral pigs in northern Australia.
It has been published in PLoS ONE.
4.1
Abstract
Invasive wildlife often causes serious damage to the economy and agriculture as well as
environmental, human and animal health. Habitat models can fill knowledge gaps about species
distributions and assist planning to mitigate impacts. Yet, model accuracy and utility may be
compromised by small study areas and limited integration of species ecology or temporal
variability. Here we modelled seasonal habitat suitability for feral pigs, a widespread and harmful
invader, in northern Australia. We developed a resource-based, spatially-explicit and regional-scale
approach using Bayesian networks and spatial pattern suitability analysis. We integrated important
ecological factors such as variability in environmental conditions, breeding requirements and home
range movements. The habitat model was parameterized during a structured, iterative expert
elicitation process and applied to a wet season and a dry season scenario. Model performance and
uncertainty were evaluated against independent distributional data sets. Validation results showed
that an expert-averaged model accurately predicted empirical feral pig presences in northern
Australia for both seasonal scenarios. Model uncertainty was largely associated with different
expert assumptions about feral pigs’ resource-seeking home range movements. Habitat suitability
varied considerably between seasons, retracting to resource-abundant rainforest, wetland and
agricultural refuge areas during the dry season and expanding widely into surrounding grassland
floodplains, savanna woodlands and coastal shrubs during the wet season. Overall, our model
suggested that suitable feral pig habitat is less widely available in northern Australia than previously
thought. Mapped results may be used to quantify impacts, assess risks, justify management
75
investments and target control activities. Our methods are applicable to other wide-ranging species,
especially in data-poor situations.
4.2
Introduction
Where mammalian wildlife becomes invasive, it is often detrimental to the economy and
agriculture as well as environmental, human and animal health (Beale et al. 2008, Hone 2007). To
effectively mitigate impacts, spatially-explicit knowledge on invaders’ distribution and habitat use
is needed (Ostfeld et al. 2005; Venette et al. 2010). This can be particularly challenging for wideranging species, as continuous empirical information is rarely available over broad geographic
regions (Stephens et al. 2015). With rapid developments in spatial environmental data availability
and new analytical methods, habitat models that infer species distributions from environmental
predictor variables have proliferated to fill knowledge gaps (Guisan & Thuiller 2005; Elith &
Leathwick 2009). Research is methodologically and terminologically diverse – depending on the
research perspective, “habitat models” are also known as “species distribution models”, “ecological
niche models”, “habitat suitability models”, “resource selection functions” and variations thereof
(Guisan & Thuiller 2005; Morrison et al. 2006; Beck & Suring 2009; Elith & Leathwick 2009;
Dormann et al. 2012; Venette et al. 2010). However, important ecological considerations such as
temporal variability or behavioural factors are often missed, especially in statistical, correlative
models. This can affect model accuracy and utility for decision-making (Franklin 2010). Here, we
developed a resource-based, spatially-explicit approach to modelling seasonal habitat suitability for
a widespread and harmful mobile invader, the feral pig (Sus scrofa), in northern Australia.
Feral pigs, originally native to Eurasia, are one of the most widespread terrestrial mammals
(Barrios-Garcia & Ballari 2012). Both wild and domesticated forms were introduced by early
settlers to all continents and many oceanic islands (Barrios-Garcia & Ballari 2012). In its introduced
range, S. scrofa is also known as feral pig, feral swine, wild hog or razorback and has often been
associated with severe negative impacts (Choquenot et al. 1996; Barrios-Garcia & Ballari 2012;
Bengsen et al. 2014). In Australia, invasive feral pigs are a major threat to unique ecosystems and
agricultural industries (Bengsen et al. 2014; Department of the Environment 2015). They are most
widespread in the tropical north, yet spatial knowledge is either empirical, detailed, and local scale
(Hone 1990a; Caley 1993 and 1997; Mitchell 2002; Mitchell et al. 2009; Wurster et al. 2012;
Elledge et al. 2013), or expert-based, coarse, broad scale, and poorly validated (Wilson et al. 1992;
West 2008; Department of Agriculture and Fisheries 2015). Improved regional-scale knowledge of
feral pig distribution could be used to delineate management units and limit re-invasion of
conservation sites following local eradication (Bengsen et al. 2014; Dexter & McLeod 2015). It
could also help assess the magnitude of environmental and economic impacts or the risk of
76
establishment of infectious animal diseases, especially when abundance estimates are derived
(Kramer-Schadt et al. 2007; Cowled & Garner 2008; Krull et al. 2016; Weber et al. in press).
Statistical habitat models for feral pigs have been developed for northern Australia (Cowled &
Giannini 2007; Cowled et al. 2009) as well as parts of Europe and North America with similar
knowledge gaps (Bosch et al. 2014; McClure et al. 2015; Morelle & Lejeune 2015; Alexander et al.
2016). However, some general limitations of statistical models are also apparent in these studies.
First, correlative models calibrated from species presence or presence / absence records can only
reliably predict species distribution within, and not outside, the environmental gradients used for
model calibration (Elith et al. 2010). Second, except for Morelle & Lejeune’s (2015) study in
Belgium, all models were calibrated from aggregate species records and did not consider temporal
variability or ecological factors. In northern Australia such models may yield misleading results
when reflecting previous research: Caley (1993) and Hone (1990a) showed that feral pig
distribution and abundance varies considerably between the wet and dry season; Caley (1997) and
Mitchell et al. (2009) found that habitat use and home range movements differ distinctly between
breeding herds (consisting of related sows and their young) and solitary boars; and Choquenot &
Ruscoe’s (2003) work suggested that feral pig persistence depends on complementary access to key
resources within the boundaries of such home ranges.
Here we adapted a resource-based modelling framework using Bayesian networks that allowed
us to address these issues. This general approach has previously been applied to habitat models
(Marcot et al. 2001; Smith et al. 2007; Tantipisanuh et al. 2014; van Klinken et al. 2015) and offers
several advantages over correlative methods: a robust statistical framework for modelling
interactions between habitat variables based on species ecology rather than distributional data;
flexible data requirements with the ability to integrate unpublished expert knowledge; and explicit
treatment of the uncertainty in parameter estimates (Uusitalo 2007; Wilhere 2012; Landuyt et al.
2013). Our objectives were: (1) to model seasonal habitat suitability for feral pig breeding and
persistence in northern Australia at the regional scale whilst integrating behavioural factors as well
as temporal variability; and (2) to rigorously evaluate accuracy and uncertainty in our expertelicited models by validating spatial predictions of habitat suitability against truly independent
distributional data sets.
4.3
Materials and methods
4.3.1 Study region
Our study region covered 1.76 million km2 north of the Tropic of Capricorn spanning three
Australian states (Fig 4.1). The climate is tropical with seasonal rainfall, alternating between a wet
77
and a dry season. Rainfall and primary productivity broadly decline on a north-south, and to a lesser
extent, on an east-west, gradient (Guerschman et al. 2009; Jones et al. 2009). Monsoonal savanna
woodlands and semi-arid grasslands are interspersed with riverine channels, seasonally inundated
floodplains, coastal wetlands and rainforest fragments (Fox et al. 2001). Intensive human uses are
concentrated in fertile coastal lowlands. The semi-arid inland is sparsely populated.
At a coarse scale, all of the study region appears climatically suitable for feral pigs and has
mostly been invaded (West 2008). Arid regions with insufficient rainfall were not included in our
study. Feral pigs are reported to be widespread in the east and localised in the north and west (West
2008). Highest local densities have been recorded in resource-abundant wetlands and floodplains,
yet these populations fluctuate considerably with climatic conditions (Hone 1990a; Caley 1993;
Choquenot et al. 1996). A wide range of management activities are conducted throughout northern
Australia to mitigate feral pig impacts, including lethal and non-lethal methods. Yet, effective
management is hampered by the region’s remoteness and there is little evidence of sustained
population reduction (Bengsen et al. 2014).
Fig 4.1 Study region in northern Australia. The study region is shown in grey. Locations of the
independent distributional data sets used for model validation are shown in colour. Refer to Table 4.3 for
details about these data sets.
4.3.2 Habitat suitability model
Our modelling approach consisted of three main steps (Fig 4.2). First, we modelled ‘resource
quality indices’ for a suite of habitat variables, referenced specifically to the resource requirements
78
of feral pig breeding herds, in separate Bayesian networks. Second, we used ‘spatial pattern
suitability analysis’ to capture feral pigs’ ability to access complementary resources at different
locations within their home range. Finally, we modelled a ‘habitat suitability index’ by combining
all ‘resource suitability indices’ in another Bayesian network. The model was calibrated using a
structured, iterative elicitation process (Marcot 2006) with a panel of experts. Experts were
practitioners with field knowledge of feral pigs from various localities and professional
backgrounds (Perera et al. 2012). We combined techniques for eliciting system understanding
through group consensus and for eliciting quantitative estimates from individuals with opportunities
for Delphi-style revision (Kuhnert et al. 2010; Martin et al. 2012; Murray et al. 2012; van Klinken
et al. 2015). Expert elicitation was approved by the CSIRO Human Ethics Committee (Project
075/13) and written consent obtained from all participants.
Fig 4.2 (next page) Feral pig habitat suitability model. Resource quality indices for each habitat variable
were modelled in Bayesian networks. Spatial pattern suitability analysis was used to compute resource
suitability indices as a weighted function of distance to resource patches (fDr). Habitat suitability was
modelled in another Bayesian network. An average habitat suitability index was computed and mapped (Fig
4.6) from six individual expert models. Bar graphs show expert-elicited conditional probabilities and values
below graphs show modelled index values ± standard deviation. Probabilities and indices change once
evidence about the states of each explanatory variable at a given study area pixel is inserted (i.e. prior
probabilities are no longer uniformly distributed).
79
Expert1
Expert2
Expert3
Expert4
Expert
25.0
25.0
25.0
25.0
Terrain ruggedness
33.3
Low
33.3
Moderate
High
33.3
50 ± 27
Expert
Food quality
High
33.3
Moderate
33.3
Low
33.3
50 ± 27
Freshwater presence
Yes
50.0
No
50.0
50 ± 25
Expert1
Expert2
Expert3
Expert4
Expert5
Food quality index
Very good
19.5
Good
23.2
Moderate
22.5
Poor
13.9
Very poor
21.0
51.2 ± 29
Water quality index
Very good
33.4
Good
8.21
Moderate
6.29
Poor
4.96
Very poor
47.2
45.1 ± 37
20.0
20.0
20.0
20.0
20.0
Food quantity
High
33.3
Moderate
33.3
Low
33.3
50 ± 27
Spatial pattern suitability analysis
(using each expert’s response function fDr)
Food suitability index
20.0
Very good
20.0
Good
20.0
Moderate
20.0
Poor
20.0
Very poor
50 ± 29
Water suitability index
20.0
20.0
20.0
20.0
20.0
50 ± 29
Very good
Good
Moderate
Poor
Very poor
Habitat suitability index
2.01
Very high
12.0
High
26.1
Moderate
21.3
Low
38.6
Very low
33.5 ± 23
Heat protection suitability index
20.0
20.0
20.0
20.0
20.0
50 ± 29
Very good
Good
Moderate
Poor
Very poor
Expert1
Expert2
Expert3
Expert4
Expert5
Expert6
Expert
16.7
16.7
16.7
16.7
16.7
16.7
Disturbance protection suitability index
20.0
Very good
20.0
Good
20.0
Moderate
20.0
Poor
20.0
Very poor
50 ± 29
Spatial pattern suitability analysis
(using each expert’s response function fDr)
Heat stress
Low
Moderate
High
Expert1
Expert2
Expert3
Expert4
Expert5
33.3
33.3
33.3
50 ± 27
Expert
20.0
20.0
20.0
20.0
20.0
Heat protection quality index
Very good
55.8
Good
14.4
Moderate
14.5
Poor
5.73
Very poor
9.49
Disturbance protection quality index
Very good
Good
Moderate
Poor
Very poor
70.3 ± 27
18.4
20.0
20.7
16.2
24.7
Dense vegetation cover
Good
Moderate
Poor
50 ± 27
48.2 ± 29
Shady vegetation cover
33.3
33.3
33.3
50 ± 27
Expert
Disturbance stress
Good
Moderate
Poor
Low
Moderate
High
Expert1
Expert2
Expert4
Expert5
Expert6
0
6.60
93.4
18.7 ± 13
Frequency of control
Low
High
Intensity of control
0
100
Low
High
25
Expert
80
0
100
25
Expert1
Expert2
Expert4
Expert5
Expert6
20.0
20.0
20.0
20.0
20.0
33.3
33.3
33.3
20.0
20.0
20.0
20.0
20.0
Predator presence
Low
High
50.0
50.0
50 ± 25
4.3.2.1 Bayesian network models
We adapted the Bayesian network modelling framework proposed by Marcot et al. (2001),
refined by Smith et al. (2007) and explained in detail by van Klinken et al. (2015). Here, habitat
suitability was conditional on a set of habitat variables representing resource requirements. Each
habitat variable was itself influenced by several measurable key explanatory variables, and each
explanatory variable was linked to one or more remotely sensed or mapped spatial data proxies. Our
model was implemented in the Norsys Netica v.5.12 software.
Expert elicitation
During an initial expert workshop, a preliminary model was developed (Marcot 2006). A panel
of experts (n = 18) constructed a conceptual model, defined each model variable (habitat suitability,
habitat variables and explanatory variables including spatial data proxies) and assigned it mutually
exclusive states. We quantified causal relationships in the network by eliciting conditional
probability tables (CPTs) behind each response variable (child node). We used the CPT calculator
software (Cain 2001), which reduces the number of elicited response probabilities to key scenarios,
i.e. combination of states in explanatory variables (parent nodes), and interpolates all other
combinations. Each step was performed in break-out groups or individually, followed by panel
discussion and consensus formation (except for the CPT calculator) (Murray et al. 2012; van
Klinken et al. 2015).
Following preliminary application, sensitivity analysis and validation of the preliminary model,
we conducted semi-structured interviews with a self-selected subsample from our panel of experts
(n = 6). Model structure, spatial data proxies and evaluation results were reviewed against each
expert’s knowledge and simplified CPTs were parameterized. We asked experts to revise prior
CPTs from the preliminary model rather than parameterizing new ones (Kuhnert et al. 2010; Martin
et al. 2012). As interviews were less time-constrained than the workshop, experts could utilize
either or both of two elicitation methods that were more flexible and robust to error than the CPT
calculator (Cain 2001). Method A was implemented in the AgenaRisk v.6.1 software and made the
simplifying assumption that any response follows a truncated normal distribution (TNormal)
centred on the weighted mean of its explanatory variables (Fenton et al. 2007). In order to use this
method, we converted all model variables into “ranked nodes”, whose states were assigned with
equal intervals on a numerical scale from 0 to 100 (Fenton et al. 2007). Experts only defined: (a) the
weight of each explanatory variable, (b) overall uncertainty in making this judgement (determining
the variance of TNormal), and (c) whether the weighted mean function should be replaced by either
a weighted minimum (to describe limiting factor relationships), or maximum (to describe
substitution relationships) function (Fenton et al. 2007). Method B restricted elicitation to key
81
scenarios as in Cain (2001). However, instead of directly assigning probabilities to each state of the
response variable, we used interval judgements (Speirs-Bridge et al. 2010; McBride et al. 2012a),
asking experts for their best estimate and the outer bounds of a 95% confidence interval. To
maintain consistency with method A, we allowed only TNormal distributions centred on the best
estimate. Post-elicitation, we discretised interval judgements into probabilities for each response
state using a binning algorithm.
Habitat suitability and resource quality indices
The final Bayesian network model (Marcot 2006) is shown in Fig 4.2. Expert-elicited CPTs are
provided in Appendix Table 4.1. Definitions for all model variables and their states are given in
Table 4.1. Spatial data proxies that determined the state of each explanatory variable at each pixel in
the study region are described in Table 4.2. Experts identified four key resource requirements for
sustained feral pig breeding: food, water, protection from heat and protection from disturbance (Fig
4.2). Each of these habitat variables had five states with corresponding equal numerical intervals (020 for the poorest state, …, 80-100 for the best state). For each habitat variable, we computed
expert-averaged ‘resource quality indices’ (xr) as model expected values from an equal-weighted
average CPT (Martin et al. 2012; McBride et al. 2012b; Fig 4.2) by summing the mid-point value
of each state interval weighted by its probability. Accordingly, xr could range between 10 (midpoint of the poorest state) and 90 (mid-point of the best state) and varied spatially according to the
combination of states of explanatory variables at a given pixel. Following spatial pattern suitability
analysis of pixel-scale xWater, xFood, xHeat and xDisturbance (see below), we computed a ‘habitat
suitability index’ (HSI) from the derived landscape-scale habitat variables (Martin & Fahrig 2012;
Jackson & Fahrig 2015). The method was analogous to the resource quality indices. However, we
used each individual expert’s CPT to compute model expected HSI. This allowed us to evaluate
average results as well as model uncertainty from diverging expert knowledge.
4.3.2.2 Spatial pattern suitability analysis
In order to capture feral pigs’ ability to access their four resource requirements at different
locations within heterogeneous home ranges, we converted pixel-scale resource quality indices (xr)
into landscape-scale ‘resource suitability indices’ (SIr). Apart from various empirical estimates of
female home range sizes (1-20 km2 in Choquenot et al. 1996; Hone 2012) and a previous finding
that pasture and riverine woodlands must co-occur within 5 km (Choquenot & Ruscoe 2003), we
had limited a priori knowledge about feral pigs’ resource-seeking home range movements in
northern Australia.
82
We therefore elicited distance-dependent response-to-pattern curves (fDr) for each habitat
variable from individual experts (n = 6) during interviews (Fig 4.3 and Appendix 4.3). This
involved first specifying a feral pig ‘mobility threshold’ (or home range boundary) beyond which
resources are inaccessible to breeding herds. Experts defined these at 1 km (n = 1), 2 km (n = 2) or
3 km (n = 3), corresponding to circular home ranges of approximately 3, 12 and 28 km2. Second,
experts described the functional value of a given resource for feral pig breeding in response to
distance (Fig 4.3). Third, we applied spatial moving window analysis to compute SIr at a focal pixel
as the highest distance-weighted xr of all resources contained within an analysis window. This
window was shifted and centred on each pixel within the study region. We used circular moving
windows (Jackson & Fahrig 2015) with radii defined by elicited mobility thresholds and distance
weights defined by response curves fDr. We termed this combined methodology ‘spatial pattern
suitability analysis’ (Appendix 4.3). 9
4.3.2.3 Seasonal scenarios
We applied each individual expert’s final model (Marcot 2006) to two seasonal scenarios by
linking model explanatory variables to seasonally-specific spatial data proxies (Table 4.2). We were
most interested in the late wet season (March to April), when resources required by feral pigs are
abundant and widely distributed, and the late dry season (October to November), when resources
are scarce and scattered across the region. Suitable remotely sensed or mapped spatial proxies were:
(a) discussed with experts and sourced from various agencies (Table 4.2); (b) rasterized and
resampled to a common extent and a fine resolution (1 ha) for capturing spatial heterogeneity
relevant to feral pigs (van Klinken et al. 2015); and (c) averaged for the two months corresponding
to each seasonal scenario over five years (2010-2014) to reflect average conditions. Finally, spatial
data attributes were reclassified to match the states of explanatory variables (Table 4.2 and
Appendix 4.2). Some model variables were linked to static spatial proxies without seasonal
variability (e.g. terrain ruggedness). For three variables determining Disturbance stress (Fig 4.2 and
Table 4.1), no spatial proxies were available for the study region. We applied a global (spatially
uniform) scenario [42], with “high” Intensity of control and Frequency of control and no selected
state for Predator presence (Fig 4.2). While this assumption likely overestimated disturbance in our
models, it approximated conditions under which most of the local validation data were collected.
A detailed description of the methodology for spatial pattern suitability analysis including theoretical framework,
expert elicitation and moving window analysis is given in Chapter 3.
9
83
Table 4.1 Bayesian network model variables and their states, with definitions.
Node & definition
State
Very high
⋅ Habitat suitability index
(HSI)
High
Whether feral pig mobs are
Moderate
able to breed and persist
⋅ Water suitability index
(SIWater)
Whether there is sufficient
potable water available to
meet drinking requirements
of feral pig mobs
(interval) State definition
(80-100) Feral pig mobs always able to breed, strong population growth
(60-80) ... usually able to breed, population growth
(40-60) ... occasionally able to breed, population maintenance
Low
(20-40) ... usually unable to breed, population decline
Very low
(0-20)
… always unable to breed, strong population decline
Very good (80-100) Drinking water fully sufficient to meet requirements
Good
(60-80) ... usually sufficient to meet requirements
Moderate
(40-60) ... occasionally sufficient to meet requirements
Poor
(20-40) ... usually limiting breeding and persistence
Very poor
(0-20)
… strongly limiting breeding and persistence
⋅ Water quality index (xWater) Very good (80-100) Potable water present and fully accessible
Presence of accessible
Good
(60-80) ... present and usually accessible
sources of potable water
Moderate
(40-60) ... present and accessible with some difficulty
Poor
(20-40) ... present and usually inaccessible
Very poor
(0-20)
… absent or fully inaccessible
⋅ Freshwater presence
Presence of potable water
Yes
⋅ Terrain ruggedness
Accessibility of water
source due to terrain
ruggedness
Low
(67-100) Level and / or low-lying terrain
Moderate
(33-67) Moderately rugged terrain
High
(0-33)
⋅ Food suitability index
(SIFood)
Whether there are sufficient
food resources available to
meet nutritional and energy
requirements of breeding
sows
Very good (80-100) Food resources fully sufficient to meet requirements
⋅ Food quality index (xFood)
Presence and nutritional
value of food resources
Very good (80-100) Food resources present and of very high nutritional value
⋅ Food quality
Quality (protein content)
and accessibility of food
resources
No
(50-100) Potable water present
(0-50)
Water absent or non-potable (salty or brackish)
Highly rugged terrain
Good
(60-80) ... usually sufficient to meet requirements
Moderate
(40-60) ... occasionally sufficient to meet requirements
Poor
(20-40) ... usually limiting breeding and persistence
Very poor
(0-20)
… strongly limiting breeding and persistence
Good
(60-80) ... present and of high nutritional value
Moderate
(40-60) ... present and of some nutritional value
Poor
(20-40) ... present and of low nutritional value
Very poor
(0-20)
… absent or of very low nutritional value
High
(67-100) High quality food resources and readily accessible
Moderate
(33-67) Moderate quality food resources and / or restricted accessibility
Low
(0-33)
High
⋅ Food quantity
Quantity (energy content) of
Moderate
food resources
Low
Low quality food resources or inaccessible
(67-100) High quantity (energy-rich) of food resources
(33-67) Moderate quantity (some energy) of food resources
(0-33)
Low quantity (little or no energy) of food resources
84
⋅ Heat protection suitability Very good (80-100) Heat refuge fully sufficient to meet requirements
index (SIHeat)
Good
(60-80) ... usually sufficient to meet requirements
Whether there is sufficient
(40-60) ... occasionally sufficient to meet requirements
heat refuge available to meet Moderate
protection requirements of Poor
(20-40) ... usually limiting breeding and persistence
feral pig mobs
Very poor
(0-20) … strongly limiting breeding and persistence
⋅ Heat protection quality
index (xHeat)
Presence and quality of
refuge from heat stress
conditions
Very good (80-100) Insignificant heat stress or refuge offers full protection
Good
(60-80) Refuge offers good protection from heat stress conditions
Moderate
(40-60) Refuge offers some protection ...
Poor
(20-40) Refuge offers little protection ...
Very poor
(0-20)
Refuge offers no or very little protection ...
Low
⋅ Heat stress
Heat stress conditions from Moderate
daytime temperatures
High
(67-100) Insignificant heat stress
Good
⋅ Shady vegetation cover
Cool microclimate provided Moderate
by shady vegetation canopy
Poor
(67-100) Deep shading provided by a dense vegetation canopy
(33-67) Some heat stress from prolonged moderate daytime temperatures
(0-33)
Significant heat stress from prolonged high daytime temperatures
(33-67) Dappled shading provided by an open vegetation canopy
(0-33)
No or little shading provided by a very sparse vegetation canopy
⋅ Disturbance protection
suitability index (SIDisturbance)
Whether there is sufficient
disturbance refuge available
to meet protection
requirements of feral pig
mobs
Very good (80-100) Disturbance refuge fully sufficient to meet requirements
⋅ Disturbance protection
quality index (xDisturbance)
Presence and quality of
refuge from disturbance
stress
Very good (80-100) Insignificant disturbance stress or refuge offers full protection
Good
(60-80) ... usually sufficient to meet requirements
Moderate
(40-60) ... occasionally sufficient to meet requirements
Poor
(20-40) ... usually limiting breeding and persistence
Very poor
(0-20)
… strongly limiting breeding and persistence
Good
(60-80) Refuge offers good protection from disturbance stress
Moderate
(40-60) Refuge offers some protection ...
Poor
(20-40) Refuge offers little protection ...
Very poor
(0-20)
Refuge offers no or very little protection ...
⋅ Dense vegetation cover
Cover provided by
understory vegetation
Good
(67-100) Good cover provided by dense understory vegetation
Moderate
(33-67) Moderate cover provided by open or medium-height vegetation
Poor
(0-33)
⋅ Disturbance stress
Disturbance stress from
human or predator
interference
Low
(67-100) Insignificant disturbance stress
Moderate
(33-67) Some disturbance stress
High
(0-33)
⋅ Intensity of control
Low
High
⋅ Frequency of control
Low
High
⋅ Predator presence
Low
No or little cover provided by sparse or tall vegetation
Significant disturbance stress
(50-100) No or low to moderate impact control activities
(0-50)
High impact control activities
(50-100) No or infrequent to occasional control activities
(0-50)
Sustained control effort (throughout and across years)
(50-100) No or insignificant numbers of predators are present
High
(0-50) Significant numbers of predators are present
The first state listed is considered most favourable to feral pig breeding. Numerical intervals corresponding to each state
are also listed in brackets.
85
86
WRel1 > 20
[Good]
n/a
WRel1 ≤ 20
[Poor]
Calculated average WRel1 for each scenario, classified
CSIRO, Raupach et al.
visually into good / poor accessibility of below-ground
2009
protein sources
Monthly relative soil
moisture upper layer
(WRel1), historical run
26j / operational v26
2015, ESRI
float (0.05°)
Classification see
Appendix Table 4.2.2
Federal Dept. of
Estimated food quality in modified systems (land use
Agriculture, ABARES class) under good / poor accessibility of below-ground
2011
protein, elicited from experts & Murray et al. 2015
TRI >24
[High]
Catchment scale Land Use 2015, FGDB
of Australia (CLUM)
grid (50m)
TRI 12-24
[Moderate]
MVS other
[No]
⋅ dry: none +
ephemeral
⋅ wet: none
[No]
⋅ dry/wet:
none
[No]
Classification see
Appendix Table 4.2.1
TRI < 12
[Low]
n/a
n/a
⋅ dry: perennial
⋅ wet: perennial
+ ephemeral
[Yes]
MVS 2 & 6
[Yes]
n/a
⋅ dry/wet:
perennial
[Yes]
Worst
⋅ dry: none +
ephemeral
⋅ wet: none
[No]
Estimated food quality in natural systems under good /
poor accessibility of below-ground protein, elicited
from experts & Murray et al. 2015
Calculated terrain ruggedness index (TRI) using
gdaldem algorithm (based on Wilson et al. 2007),
classified visually as thresholds scale dependent
Water bodies inadequately mapped in tropical, subtropical or warm temperate rainforest, classified as
water present (non-limiting)
Classified HydroAreas as ‘ephemeral/none’ and
WaterBody as ‘perennial/ephemeral/none’ from
AHGFType, Perennial & SrcFType attributes
Classified as ‘perennial’ except abandoned/dry or saline
n/a
Middle
Federal Dept. of the
Environment, DEWR
2007
2010, FGDB Geoscience Australia,
grid (3 arcsec) Gallant et al. 2011
Best
Classified areas as ‘perennial/ephemeral/none’ from
⋅ dry: perennial
Wetclass & WtrRegime attributes, excluded non-potable ⋅ wet: perennial
water from Salinmod attribute, classified points as
+ ephemeral
‘perennial’ except if dormant
[Yes]
Rationale, method
Reclassification into state-specific categories [name of state]
Present Major Vegetation
2012, FGDB
Subgroups (MVS), NVIS
grid (100m)
V.4.1
3 sec SRTM derived
Terrain
Digital Elevation Model,
ruggedness
V.1.0
Food
quality
[seasonal]
Bureau of
Meteorology, BOM
2012
2012, FGDB
feature class
(1:250,000)
Federal Dept. of the
Environment, DEWR
2007
Bureau of
Meteorology, BOM
2012
Dept. of Environment
& Heritage Protection
QLD, EPA 2005
2012, FGDB
feature class
(1:250,000)
2012, ESRI
shapefile
(1:100,000)
Date, file type Source, reference
Spatial layer
Present Major Vegetation
2012, FGDB
Subgroups (MVS), NVIS
grid (100m)
V.4.1
Cartography V.2.1 –
HydroArea / WaterBody
Geofabric Surface
Cartography V.2.1 –
Freshwater HydroPoint / Dam
presence
[seasonal] Geofabric Surface
Queensland wetland data
V.3 - wetland areas /
points (springs)
Node
[seasonal?] Name
Table 4.2 Spatial data proxies linked to model explanatory variables and methods for reclassifying data attributes into state-specific categories.
87
Fig 4.3 Expert-elicited resource suitability in response to distance. Resource suitability indices (SIr) were
computed from elicited response-to-pattern curves (fDr) for each of the four habitat variables in the model:
water (A), food (B), heat protection (C) and disturbance protection (D). Curves cross the x-axis at different
points because experts defined different home range boundaries.
4.3.3 Model evaluation and validation
4.3.3.1 Sensitivity analysis
Behaviour of each individual expert model as well as an expert-averaged consensus model was
evaluated using the “Sensitivity to findings” algorithm in the Norsys Netica 5.12 software. We
focused on the variance reduction metric recommended for numerical variables (Marcot 2012). It
assessed the relative influence of our four habitat variables on predicted habitat suitability by
88
calculating how much the variance of HSI was reduced by entering a particular finding (i.e. SIr
value) for one of the habitat variables. Greater variance reduction means that HSI was more
sensitive to a change in the state of the habitat variable (Marcot 2012; Smith et al. 2012).
4.3.3.2 Predictive performance
Predictive model performance was validated against four data sets of feral pig presence per
seasonal scenario (Table 4.3). Most data were independently collected by external agencies in
conjunction with aerial management activities. As aerial survey data were only available for the
eastern state of Queensland (Fig 4.1), we also obtained presence records (search term “Sus scrofa”)
from the national Atlas of Living Australia. This database contained only one dated record in the
Western Australian portion of our study area. Hence, we confined the downloaded ‘ALA’ data set
to records from the Northern Territory, which appeared adequately sampled (748 unfiltered
records). Where possible, we matched data to model assumptions, using only presence records that
corresponded to breeding herds (identified as female or with a group count greater than two) and
were dated in the late wet and late dry season respectively. To reduce unwanted noise from spatial
error in validation data or spatial proxies used for modelling (Hunsacker et al. 2001; Boyce et al.
2002), we upscaled both predicted HSI and feral pig presence records to a 1 km resolution. We
subsequently thinned data to ensure independence among data points, allowing only one data point
collected on the same day within a given 1 km pixel.
For each data set, we also defined a validation background, which served to contrast presences
with areas from which feral pips could be considered ‘absent’ (Fig 4.1). In doing so, we aimed to
strike a compromise between (a) evaluating performance across sufficiently large areas to justify
inferences about the models’ discriminatory power and (b) restricting evaluation to similar
environmental gradients as contained in the presence data so that validation metrics are not
artificially inflated (VanDerWal et al. 2009). Where possible, we defined backgrounds from
existing management units in which surveys were conducted (National Park boundaries for
‘Lakefield’ and ‘Oyala Thumotang’). Otherwise, we arbitrarily applied a 15 km buffer to data
points (for ‘Balkanu’, ‘NAQS’ and ‘ALA’), which corresponded to five times the highest expertelicited mobility threshold (home range boundary) in this study.
89
Table 4.3 Validation data sets with ancillary information.
Name
Source
No. of
Date of collection Method and purpose Background Typical habitat types
records
Balkanu
Balkanu Cape
of collection
⋅ dry: 181 ⋅ Sep-Nov 2013-14 Systematic aerial
York Development ⋅ wet: 67
survey and
⋅ May 2015
Corporation
size (km2)
⋅ dry: 3,954 Eucalyptus woodlands
⋅ wet: 3,089 & coastal wetlands
management
(shooting)
Lakefield
Queensland Parks ⋅ dry: 350 ⋅ Oct-Dec 2009-13 Systematic aerial
⋅ dry: 5,788 Eucalyptus / Melaleuca
& Wildlife Service ⋅ wet: 124 ⋅ Feb-May 2010-13 management
⋅ wet: 5,788 woodlands, coastal
(shooting)
Oyala
Queensland Parks wet: 263 Apr-May 2010-13 Systematic aerial
Thumotang & Wildlife Service
ALA
wet: 3,819
Eucalyptus woodlands,
management
riparian Melaleuca
(shooting)
forests & rainforests
Atlas of Living
⋅ dry: 111 ⋅ Sep-Dec 98-2012 Surveys and
⋅ dry: 36,024 Eucalyptus woodlands,
Australia
⋅ wet: 144 ⋅ Feb-May 91-2012 sightings (purpose
⋅ wet: 41,511 floodplains, Melaleuca
unknown)
NAQS
wetlands & grasslands
forests & mangroves
Northern Australia dry: 103 Sep-Nov 2007-10 Opportunistic aerial dry: 11,630 Eucalyptus woodlands,
Quarantine
survey and disease
floodplain grasslands
Strategy
sampling (shooting)
& chenopod scrublands
We used the Continuous Boyce Index (CBI) to evaluate model performance (Appendix 4.4).
This method was developed specifically for evaluation against presence-only observations (Boyce
et al. 2002; Hirzel et al. 2006). A predicted-to-expected (P/E) ratio was computed as the (predicted)
proportion of presence records coinciding with a specified range of HSI values (bin) divided by the
(expected) proportion of the validation background covered by that bin. The CBI measures the
Spearman rank correlation coefficient of P/E against HSI and varies from 1 (correct model, P/E
steadily increases as HSI increases) to –1 (false model, P/E steadily decreases with increasing HSI).
Values close to zero indicate a random prediction (Hirzel et al. 2006). We computed the P/E ratio
and CBI for each expert model / validation data combination. We also computed the proportion of
each validation background expected to be highly or very highly suitable (HSI ≥ 60). This resulted
in 28 sets of validation metrics (P/E ratio + CBI + HSI ≥ 60) per seasonal scenario (6 expert models
x 4 validation data + 1 expert-averaged model x 4 validation data; Appendix 4.4). This allowed us
to evaluate uncertainty from diverging expert knowledge and test the expectation that an expertaveraged model performs best (McBride et al. 2012b).
90
4.4
Results
4.4.1 Sensitivity analysis
Sensitivity analysis revealed that habitat suitability in the Bayesian network model was most
strongly influenced by water and food resources (24% and 26% variance reduction respectively)
and to a lesser degree by protection from heat (6%) and protection from disturbance (3%) (Fig 4.4).
Habitat suitability was least sensitive to expert opinion (< 1% variance reduction). Hence, experts
were mostly in agreement about the relative importance of the four habitat variables and quantified
model relationships similarly. However, one expert weighted the heat protection requirement as
highly as water and food resources (Expert 1: 15% variance reduction; Fig 4.4).
Fig 4.4 Sensitivity of habitat suitability to four habitat variables and expert opinion. Sensitivity to
findings was calculated as % variance reduction for each individual expert model and an averaged model
(red bar and percentages).
4.4.2 Predictive performance
In general, expert-elicited habitat suitability models performed well against the validation data
(Table 4.4 and Fig 4.5). For some seasonal scenario / validation data combinations, all expert
models performed well (e.g. wet or dry season model / Lakefield data). For others, there were
considerable differences (e.g. wet or dry season model / ALA data) (Table 4.4). As the HSI
Bayesian network model was not sensitive to expert opinion, model uncertainty stemmed largely
91
from disagreement about feral pigs’ resource-seeking home range movements, i.e. expert-elicited
response-to-pattern curves (Fig 4.3). Expert models that assumed high mobility thresholds (3 km for
experts 2, 4 and 6; Fig 4.3) predicted the highest proportions of suitable habitat in all validation
backgrounds. Average HSI ≥ 60 for these three expert models ranged between 71-78% in the wet
season and 36-42% in the dry season. Model accuracy was also generally highest for these experts,
with average CBI ranging between 0.58-0.85 in the wet season and 0.69-0.86 in the dry season.
Both metrics were lowest for the expert model that assumed the least mobility (1 km for expert 1;
Fig 4.3). Here, average HSI ≥ 60 was 47% (wet) and 15% (dry) and average CBI was 0.33 (wet) and
0.44 (dry).
The expert-averaged model performed similar to, or better than, the best individual expert
models for most validation data, except for the dry season model / NAQS data and dry season
model / ALA data combinations. Its average CBI across all validation data sets was 0.85 in the wet
season and 0.67 in the dry season. The predicted proportion of suitable habitat was modest
compared to individual expert models (except for expert 1), with an average HSI ≥ 60 of 60% (wet)
and 26% (dry). While averaging did not increase model accuracy as expected, it produced
consistently accurate results for all validation data (Table 4.4, Fig 4.5 and Appendix Fig 4.5).
Table 4.4 Validation metrics for individual expert and averaged seasonal habitat suitability models.
Validation metrics per model scenario / validation data combination
Habitat
suitability
model
Wet season scenario
Balkanu
Lakefield
Dry season scenario
Oyala Thum
ALA
Balkanu
Lakefield
NAQS
ALA
CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60
Expert 1
-0.22 59%
0.69 60%
0.66 41%
0.17 26%
0.59 28%
0.66 15%
0.27 12%
0.25
Expert 2
0.94 83%
0.94 91%
0.61 77%
0.9
58%
0.94 64%
0.9
38%
0.94 44%
0.66 20%
Expert 3
0.73 80%
0.97 85%
0.64 68%
0.7
54%
0.94 58%
0.88 30%
0.75 44%
0.75 15%
Expert 4
0.83 82%
0.94 83%
0.7
0.86 49%
0.98 56%
0.97 33%
0.72 42%
0.76 12%
Expert 5
0.38 74%
0.63 56% -0.24 69%
42%
0.89 50%
0.96 34%
0.74 30%
0.56
9%
Expert 6
0.44 83%
0.94 91%
0.43 77%
0.52 60%
0.94 63%
0.96 38%
0.35 44%
0.5
22%
Averaged 0.87 71%
0.92 70%
0.88 59%
0.71 39%
0.85 44%
0.97 26%
0.59 25%
0.27
7%
68%
0
3%
Validation was performed against four data sets per seasonal scenario (Table 4.3). We show the Continuous Boyce
Index (CBI) and proportion of validation background expected to be highly or very highly suitable habitat (HSI ≥ 60,
here shortened to HSI60). A CBI = 1 would indicate a perfectly accurate, a CBI ~ 0 a random, and a CBI < 0 a false
model.
92
Fig 4.5 Validation plots for individual expert and averaged seasonal habitat suitability models.
Validation was performed against four validation data sets per seasonal scenario (Table 4.3). The predictedto-expected (P/E) ratio (y axis) measures the proportion of feral pig presences relative to the proportion of
background pixels on a continuous scale of predicted HSI (x axis).
93
4.4.3 Seasonal habitat suitability
We present and discuss seasonal results only for the expert-averaged consensus model, which
produced consistently accurate results across the study region. Predicted habitat suitability varied
considerably between seasonal scenarios (Table 4.5 and Fig 4.6). Overall the model predicted
suitable habitat (HSI ≥ 40) in 36% of the study region (~640,000 km2) during the wet season and <
10% (~170,000 km2) during the dry season. Of this, about one quarter was highly or very highly
suitable habitat (HSI ≥ 60, 8% during the wet season and < 3% during the dry season, Table 4.5).
Fig 4.6 Seasonal habitat suitability for feral pig breeding in northern Australia. Habitat suitability
indices were averaged across all expert models and mapped for a wet (March / April – A) and dry (October /
November – B) season scenario.
94
Table 4.5 Amount of feral pig habitat in each habitat suitability index class per seasonal scenario.
Seasonal scenario
Total area covered by each habitat suitability index class (km2 and %)
Very high
High
Moderate
Low
Very low
Wet season
16,726 (1%)
130,112 (~7%)
489,371 (~28%)
239,948 (~14%)
880,902 (~50%)
Dry season
11,147 (< 1%)
39,512 (~2%)
116,273 (~7%)
71,532 (~4%)
1,518,597 (~86%)
Habitat suitability also varied spatially between administrative units, vegetation types and land
use classes (Fig 4.6, Fig 4.7 and Appendix 4.6).
When analysed by states (shown in Fig 4.1), highly and very highly suitable feral pig habitat
was located mostly in Queensland, especially during the dry season. It was largely absent from
Western Australia in either scenario. The Northern Territory’s share increased more than three-fold
during the wet season. Moderately suitable habitat (40 ≤ HSI < 60) was somewhat more evenly
distributed across the study region, especially during the wetter months (covering about 40% of the
study region in Queensland, 20% in the Northern Territory and 10% in Western Australia,
Appendix Fig 4.6.1). Within each state, suitable habitat was concentrated in coastal environments
during the dry season (except for some large inland riverine and wetland systems), and expanded
widely during the wet season (Fig 4.6).
By broad vegetation types (Department of the Environment and Water Resources 2007; Fig
4.7), highly suitable habitat (HSI ≥ 60) was disproportionately found in rainforests, wetlands,
mangroves and modified (agricultural) vegetation, especially during the dry season (23%, 7%, 4%
and 17% share of total suitable habitat respectively). Yet, these vegetation types covered only 6.2%
of the study region combined. A large share of suitable habitat – 42% (dry) and 64% (wet) – was
also contained in the region’s dominant vegetation type, savanna woodlands, which was broadly
proportional to its overall cover (68% of the study region). Very highly suitable habitat was even
more concentrated in rainforests (68% (dry) and 54% (wet) share of HSI ≥ 80) and less frequently
found in savanna woodlands (22% (dry) and 29% (wet) share of HSI ≥ 80; Fig 4.7). During the dry
season, the vast majority of grasslands (98%), shrublands (93%) and woodlands (91%) were
modelled as unsuitable for feral pig breeding (HSI ≤ 40; Appendix Fig 4.6.2). During the wet
season, habitat concentration was weaker and suitable habitat was somewhat more evenly
distributed among vegetation types (Fig 4.7).
Habitat suitability was also unevenly distributed between land use classes (Australian Bureau
of Agricultural and Resource Economics and Sciences 2011; Fig 4.7). During the dry season, high
value land used for water resources, irrigated or dryland production together contained 17% of
95
suitable habitat on less than 2% of the study region. Interestingly, the 26% of the study region set
apart for nature conservation also contained a disproportionate amount of suitable habitat (52%
(dry) and 43% (wet) share of HSI ≥ 60). The dominant land use type “grazing natural vegetation”
contained a greater share of suitable habitat during the wet season (42%) than the dry season (24%)
on 71% of the study region. As with vegetation types, habitat concentration was further increased
for very highly suitable habitat (HSI ≥ 80) and weaker in the wet than in the dry season (Fig 4.7).
Fig 4.7 Share of modelled suitable habitat found in different vegetation types and land use classes for
each seasonal scenario. For each vegetation (panel A) or land use (panel B) category, we show its percent
share of highly (HSI ≥ 60) and very highly (HSI ≥ 80) suitable habitat during the dry (top bars) or wet
(bottom bars) season, compared to its percent share of the total study region (central bars). Spatial analyses
in panel (A) were based on Present Major Vegetation Groups (MVG V.4.1) and analyses in panel (B) on
Australian Land Use and Management classes (ALUM V.7).
4.5
Discussion
Effective management of invasive wildlife requires spatially-explicit knowledge of their
distribution and habitat use, especially for wide-ranging species (Venette et al. 2010). Yet,
continuous empirical information is rarely available over large areas (Stephens et al. 2015). Habitat
models can fill this knowledge gap, but their utility may be compromised by small study areas and
limited integration of species ecology or temporal variability (Franklin 2010). Here, we modelled,
for the first time, seasonally-specific habitat suitability for feral pigs, a widespread and harmful
96
invader, in northern Australia. Rigorous evaluation and validation showed that our resource-based,
expert-elicited model, which integrated important ecological factors such as home range movements
and breeding requirements, accurately predicted feral pig presence across the study region. Results
suggest that suitable feral pig habitat is more constrained in northern Australia than previously
thought, especially during the dry season. Mapped results may be used by land managers to
quantify impacts, assess risks, justify investments and target control activities. Our transparent
methodology could be applied to other wide-ranging species, especially in data-poor situations.
4.5.1 Seasonal habitat suitability
Our habitat model confirmed previous site-scale findings that the distribution and habitat use of
feral pigs in northern Australia is highly seasonal (Mitchell et al. 1982; Hone 1990a; Caley 1993;
Choquenot et al. 1996) for the entire region. Modelled habitat suitability for feral pig breeding and
persistence was mainly driven by seasonal availability of food and water resources, both of which
ultimately vary with northern Australia’s annual rainfall cycle. Our model indicated a four-fold
increase in suitable habitat during the wet season. Inter-annual climatic variability, which has been
shown to greatly affect feral pig populations in drier parts of Australia (Mitchell et al. 1982;
Choquenot 1998; Dexter 1998), was not investigated in this study. Our scenario approach could be
usefully extended to model cyclical, as well as seasonal, variability in feral pig distribution.
Seasonal fluctuations in habitat suitability were expressed at different spatial scales. At the
regional scale, habitat suitability varied along an east-west gradient. Suitable feral pig habitat was
concentrated in the eastern study region throughout the year. During the dry season, suitable habitat
contracted more in the west than in the east (eleven-fold in Western Australia, five-fold in the
Northern Territory and three-fold in Queensland). These patterns correspond well with prevailing
rainfall gradients and harsher dry season conditions in the west (Guerschman et al. 2009; Jones et
al. 2009). At the landscape scale, contiguous patches of suitable habitat were located predominantly
along the coastline during the dry season and expanded widely across the study region during the
wet season. If all suitable wet season habitat was to be used by feral pigs, this points to long
distance seasonal migration. However, such migratory behaviour in Australian feral pigs has been
rejected by previous research (Caley 1997; Mitchell et al. 2009; Hone 2012). Rather, empirical
findings suggest that feral pigs may expand and contract their home range in response to changing
conditions (Dexter 1999; Mitchell et al. 2009), or shift it entirely if adverse conditions persist
(Caley 1997; Choquenot et al. 1996). Hence, not all suitable wet season habitat may be within reach
of feral pig breeding herds dispersing from dry season refuges (Caley 1997). At the local scale, dry
season habitat was concentrated in productive rainforest, wetland and mangrove ecosystems as well
as high value agricultural lands, where resources remain abundant. While these dry season refuges
97
continued to provide suitable habitat during the wet season, habitat was more evenly distributed
among vegetation types and land use classes. This suggests that feral pigs forage widely into
grassland floodplains, savanna woodlands and coastal shrublands when conditions permit. Our
regional model showed that rainforests are a key year-round habitat for feral pig breeding (Fig 4.7).
This was partly due to model assumptions. For example, as actual freshwater presence was
inadequately mapped in densely vegetated rainforests, we assumed that water was uniformly
available (Table 4.2). Site-scale studies have already provided a more detailed understanding of the
fine-grained variations in actual habitat use within this broadly suitable habitat type (Mitchell 2002;
Mitchel et al. 2009; Wurster et al. 2012; Elledge et al. 2013).
While all parts of northern Australia contained at least some suitable feral pig habitat, our
model suggests that feral pigs are less widely distributed in the region than previously thought. For
example, moderately suitable habitat in Western Australia totalled between 3000 km2 (dry) and
32,000 km2 (wet) while Cowled et al.’s (2009) model predicted 89,000 km2 of suitable habitat in
the same area. Similarly, West (2008) reported that feral pigs are widespread throughout
Queensland while our model found that only 16% (dry) to 54% (wet) of this area contained all
required resources for persistent feral pig breeding. We note that (a) such area estimates are highly
dependent on habitat thresholds and therefore are unreliable and difficult to compare between
studies, and (b) our figures are possibly overestimates – they refer to a threshold of HSI ≥ 40, yet
validation plots (Fig 4.5) suggested that a higher, more restrictive threshold (e.g. HSI ≥ 60) may
better discriminate between suitable and unsuitable habitat in most environments. We suggest our
lower estimates are defensible when considering a number of methodological improvements in our
study. First, we specifically modelled resource requirements of feral pig breeding herds, which are
more limiting than those of solitary boars (Caley 1997; Choquenot et al. 1996; Mitchell et al. 2009).
Previous statistical models (Cowled et al. 2009) and mapping studies (Mitchell et al. 1982; West
2008) did not make this distinction, although long-term occupancy critically relies on breeding.
Second, previous work did not distinguish between seasonal scenarios but included any location
where feral pigs have occurred or may occur. This may have approximated wet season habitat and
likely led to overestimations by failing to consider the dry season as a limiting factor for feral pigs
in Australia’s north (Hone 1990a; Caley 1993). Finally, we used a finer spatial resolution (1 ha)
than previous studies (25 km2 in Cowled et al. 2009 or ~250 km2 in West 2008), resulting in more
detailed predictions and less upscaling error (e.g. one “suitable” pixel equalled 25 km2 in Cowled et
al. 2009, even if only a fraction of this area actually contained suitable habitat).
98
4.5.2 Model evaluation and validation
Model results were robust to both expert opinion and a range of independently collected
validation data sets of feral pig presence. All six experts contributing to the final consensus model
provided similar parameter estimates for the Bayesian network model of habitat suitability. That is,
they all agreed that water and food resources are more important to persistent feral pig breeding in
northern Australia than protection from heat and protection from disturbance. However, validation
revealed inconsistent performance between expert models in some instances. We attributed this
model uncertainty mainly to different expert assumptions about feral pigs’ resource-seeking home
range movements, described in elicited response-to-pattern curves (Fig 4.3) 10. Our simple approach
to reducing uncertainty, equal-weighted expert averaging (Martin et al. 2012; McBride et al.
2012b), yielded similar, or more accurate, results than the best expert models for most validation
data.
While the averaged habitat suitability model always performed well, there were differences
between seasonal scenarios and validation data. This points to some limitations of our study. First,
most validation data were collected in environments with an above-average proportion of suitable
feral pig habitat. Second, model parameters were elicited from experts with field knowledge mostly
from resource-abundant environments in the eastern study region and applied to all of northern
Australia. Model accuracy in resource-poor inland environments and in those portions of the study
region for which no validation data were available and that were outside the expertise of our panel
of experts (especially Western Australia) needs further investigation (Murray et al. 2009). For
example, feral pigs may also sustain themselves in ‘unsuitable’ habitat during the dry season from
resources not included in our model (e.g. carrion and other animal matter). Third, reporting bias in
the data points used for validation (Phillips et al. 2009) and our definition of validation
backgrounds may have affected validation results. For example, most data sets were biased towards
highly suitable sites as management-focused survey efforts mostly concentrated on sites known to
be impacted by feral pigs. Less suitable sites, which nevertheless support feral pig breeding, may
therefore have few or no data points recorded, resulting in somewhat inflated CBI values. Further,
different buffering choices for defining the ALA, NAQS or Balkanu validation backgrounds may
have yielded poorer (likely for backgrounds that are more narrowly defined around presence
records than our 15 km buffer) or enhanced (likely for larger backgrounds that encompass high
proportions of unsuitable habitat especially in the spatially disjunct ALA data set; Appendix Fig
4.5) performance results. Fourth, the Continuous Boyce Index validation method is well suited to
In Chapter 3 I showed that such assumptions affect the amount and distribution of suitable habitat as well as model
performance.
10
99
presence-only validation data but cannot evaluate model specificity, i.e. its ability to correctly
predict absences and minimize false positives (Boyce et al. 2002; Hirzel et al. 2006). Finally, we
applied the same expert-elicited mobility thresholds and distance-dependent response-to-pattern
curves to the entire study region and both seasonal scenarios. Yet, feral pigs may adjust their
resource-seeking home range movements to environmental conditions (Caley 1997; Dexter 1999;
Mitchell et al. 2009) and in fact respond to aspects of landscape structure other than distance such
as resource composition or heterogeneity (Turner & Gardner 2015).
A listing of study limitations is most useful for guiding future research. We refer to van
Klinken et al. (2015) who suggest methods to evaluate possible errors in expert-elicited models. In
particular, our study may be usefully improved by: (1) systematically collecting presence / absence
data, also from resource-poor environments, to eliminate reporting bias in validation data and
enable evaluation of both model sensitivity and specificity; (2) field-validating the accuracy of
spatial data proxies and whether they match the states of model explanatory variables; (3) revising
model parameters with experts from the Northern Territory and Western Australia; and (4)
conducting a multi-scale study (Jackson & Fahrig 2015; McGarigal et al. 2016) to better understand
feral pigs’ response to spatial patterns of resources with varying quality in different types of
environments.
4.5.3 Management implications
Our regional-scale, seasonally-specific and rigorously validated results can serve to better
manage the impacts of feral pigs in northern Australia. For example, we estimated habitat suitability
per broad vegetation and land use types. When combined with information on impacts or costs in
each category (Barrios-Garcia & Ballari 2012; Mitchell 2010; Krull et al. 2016), this may help to
more accurately quantify environmental or economic impacts across any area of interest. Habitat
suitability could also be analysed in other management units to justify investments in population
control or, if verified by stratified field surveys, to serve as a monitoring baseline in adaptive
management programs. Further, because habitat suitability was explicitly referenced to feral pig
breeding herds, it is a useful indicator of establishment risk for infectious animal diseases, which
often depend on a persistent supply of young susceptible animals (Kramer-Schadt et al. 2007;
Cowled & Garner 2008). Habitat connectivity for feral pigs is also a critical parameter in
understanding disease spread, but has not been explicitly studied in northern Australia. Future
models may integrate validation results from this study to derive habitat quality thresholds for patch
delineation (Hirzel et al. 2006) and reverse habitat suitability to define resistance to movement, i.e.
the cost of traversing habitat of different quality (Zeller et al. 2012). Finally, our resource-based
model captures the relative importance of four habitat requirements as well as spatial interactions
100
such as landscape complementation or supplementation (Dunning et al. 1992). This knowledge may
be used to manipulate resource access at strategic locations (e.g. by exclusion fencing or local
eradication) and model the effect on habitat suitability and connectivity within the broader
landscape.
4.6
Appendices
Appendix 4.1 Bayesian network conditional probability tables (supplements Fig 4.2).
Appendix 4.2 Classification of categorical spatial data attributes to match the states of model
explanatory variables (supplements Table 4.2).
Appendix 4.3 Methodology for spatial pattern suitability analysis (annotated R code).
Appendix 4.4 Methodology for model validation (annotated R code).
Appendix 4.5 Maps of seasonal habitat suitability per validation area.
Appendix 4.6 Seasonal habitat suitability per state and per broad vegetation type.
101
Chapter 5 Modelling habitat connectivity for biosecurity: the risk of infectious
disease spread in feral pigs in northern Australia
Summary
In this chapter I model and map seasonal habitat connectivity for feral pigs in northern
Australia. I apply the results from Chapter 4 to delineate contiguous patches of feral pig breeding
habitat, parameterize matrix resistance to inter-patch dispersal, and model seasonal connectivity in
the landscape matrix using a circuit-theoretic approach. Then, I estimate connectivity between
patches and evaluate its sensitivity to a range of model assumptions that are linked to genderspecific differences in feral pig dispersal ability. This chapter addresses research question 3,
providing an indicative measure of the seasonal spread risk of directly transmitted infectious
diseases within feral pigs in northern Australia. It is being prepared for submission to Journal of
Applied Ecology.
5.1
Abstract
Directly transmitted diseases in wildlife hosts can spread between infected and susceptible
populations via host dispersal. Risk-based disease management requires spatial knowledge of the
functional links between populations, often over areas too large for empirical studies. Yet, habitat
connectivity models have not been widely applied to address this problem. We modelled seasonal
habitat connectivity for feral pigs, an abundant invader that is a susceptible host to numerous
infectious diseases, in northern Australia. We used a circuit-theoretic approach to model potential
dispersal paths in the landscape matrix and measured connectivity between patches of breeding
habitat as an indicator of disease spread risk. To account for gender-specific differences in dispersal
ability, we evaluated a range of model assumptions on the level of matrix resistance and on two
parameters for delineating dispersal paths. Habitat connectivity was significantly affected by
changes in these assumptions. Specifically, results indicated that dispersal paths of feral pigs, and
connectivity between habitat patches, differ considerably with climatic conditions and between
breeding herds and solitary boars. We identified three broad types of patches: some were always
isolated; some were always connected to large habitat components; and some were mostly isolated
but became connected to large habitat components only during the wet season or for wide-ranging
boars. Our results can directly inform risk-based management of infectious animal diseases at all
stages of invasion, including early detection surveillance, preparedness and incursion response. The
resource-based approach could readily be applied to other wildlife hosts, especially in regionalscale and data-constrained studies.
102
5.2
Introduction
Infectious animal diseases are a major biosecurity threat in an increasingly connected world
(Jones et al. 2008). New incursions may seriously impact human, livestock and wildlife health and
incur large costs on industry and government (Daszak et al. 2000). Wildlife hosts are a wellrecognized risk factor for disease introduction, establishment and spread (Daszak et al. 2000; Jones
et al. 2008; Reisen 2010). For example, bats are important reservoir hosts of several emerging
pathogens, including the rabies, Nipah or Hendra viruses (Calisher et al. 2006); migratory birds
may contribute to global spread of highly pathogenic avian influenza H5N1, with devastating
consequences to human health and the poultry industry (Olsen et al. 2006); wild boar infected with
classical swine fever caused repeated outbreaks in domestic herds in Germany during the 1990s,
necessitating costly interventions (Fritzemeier et al. 2000; Kramer-Schadt et al. 2007). Invasive
wildlife species can be particularly problematic, introducing disease to susceptible hosts (Daszak et
al. 2000; Macpherson et al. 2016) or complicating disease management in widespread and
expanding populations. However, realistic spatially-explicit descriptions of wildlife hosts’
distribution and habitat use are rarely integrated into risk-based analyses of disease establishment
and spread (Ostfeld et al. 2005; Ward et al. 2011; Rees et al. 2013; Macpherson et al. 2016). Here,
we apply a resource-based approach (Hartemink et al. 2015) to model habitat connectivity for feral
pig (Sus scrofa) populations in northern Australia. The species is an abundant invader that is
susceptible to numerous emergency diseases such as foot-and-mouth disease, classical swine fever
and Japanese encephalitis (Animal Health Australia 2011).
We built on the theoretical ‘patch-corridor-matrix’ (Forman 1995) and metapopulation (Hanski
1998) frameworks, where local wildlife populations are viewed as persisting in patches of suitable
breeding habitat that are embedded in a matrix of unsuitable habitat. Some areas in the matrix
function as dispersal corridors between patches. Metapopulation persistence depends on local
population dynamics and inter-patch movements (Hanski 1998; Hastings 2014). A metapopulation
perspective on feral pigs in Australia is supported by landscape-genetic evidence of population
structuring, especially in heterogeneous landscapes (Hampton et al. 2004; Cowled et al. 2008;
Lopez et al. 2014). Patches and corridors also affect the risk of invasion and persistence by directly
transmitted diseases in a wildlife host metapopulation (Anderson et al. 1986; Riley 2007). Initial
establishment is contingent on locally dense breeding populations with high contact rates that let an
infectious pathogen’s basic reproductive rate R0 exceed 1. Subsequent disease spread requires host
dispersal between infected and susceptible subpopulations. Disease persistence depends on a lasting
supply of susceptible individuals within a regionally connected host metapopulation (Anderson et
al. 1986; Kramer-Schadt et al. 2007). Risk-based management of infectious diseases with wildlife
103
hosts would benefit from spatially-explicit knowledge of habitat connectivity at the regional scale
(Ostfeld et al. 2005; Riley 2007; Cowled & Garner 2008; East et al. 2013).
Connectivity research, which broadly aims at identifying links between patches, has
proliferated in recent decades (Tischendorf & Fahrig 2000; Galpern et al. 2011; Zeller et al. 2012;
Kool et al. 2013). Applications have mostly examined metapopulation dynamics (Moilanen &
Hanski 2001; Urban et al. 2009) or conservation benefits such as reserve networks (Beier et al.
2011; Doerr et al. 2011; Rayfield et al. 2011). More recently, research has also focused on the
management of invasive species (Glen et al. 2013; Stewart-Koster et al. 2015) or infectious diseases
(Riley 2007; Rioux Paquette et al. 2014; Macpherson et al. 2016). Modelling methods vary broadly
according to research aims and definitions of connectivity. Tischendorf & Fahrig (2000) defined
connectivity as “the degree to which the landscape facilitates or impedes movement” (p. 7), while
Kool et al. (2013) distinguished between structural, functional and genetic connectivity. Functional
approaches that consider structural heterogeneity in the landscape as well as species-, gender-,
scale- and time-specific factors have been increasingly recommended (Hamilton & Mather 2009;
Doerr et al. 2011; Rayfield et al. 2016). When modelling habitat connectivity for disease
management, identifying multiple dispersal paths between infected and susceptible patches is of
greater interest than optimal (least-cost) dispersal paths or genetic connectivity over long
timeframes (McRae et al. 2008; Kool et al. 2013). A variety of connectivity metrics have been
developed (Moilanen & Nieminen 2002; Kindlmann & Burel 2008; Rayfield et al. 2011), broadly
describing connectivity either as an attribute of the matrix, or the patch (Moilanen & Hanski 2001;
Calabrese & Fagan 2004). For the management of infectious diseases both interpretations are highly
relevant. Here, we emphasize this conceptual distinction by referring to ‘matrix connectivity’
(Where in the matrix is disease spread via host dispersal likely to occur?) and ‘patch connectivity’
(To which other patches is an infected patch connected?). We use the term ‘habitat connectivity’ as
an overarching concept that integrates both perspectives in the patch-corridor-matrix framework.
Here, we modelled the risk of spread of directly transmitted diseases within feral pigs in
northern Australia. We focused on understanding habitat connectivity for the wildlife host, as such
knowledge gaps have impeded epidemiological models and disease management in Australia and
elsewhere (Ostfeld et al. 2005; Cowled & Garner 2008; Rees et al. 2013; Macpherson et al. 2016).
Previously, we modelled habitat suitability for feral pig breeding in northern Australia (Chapter 4).
Our first objective was to apply the resource-based, empirically validated results of this earlier
study to model ‘matrix connectivity’ across the entire region. Our second objective was to measure
‘patch connectivity’ (Rayfield et al. 2011) as an indicator of disease spread risk. Previous
simulation models of disease spread in feral pigs have suggested that seasonal (distributional
104
patterns and dispersal behaviour vary according to conditions) and gender-specific (dispersal by
mostly solitary boars appears much less resource-constrained than that of breeding herds) factors
may affect disease transmission (Doran & Laffan 2005; Milne et al. 2008). Hence, our third
objective was to apply the model to two seasonal scenarios and to evaluate the effect of changes in
model parameters that may be reflective of gender-specific differences in dispersal ability.
5.3
Materials and methods
5.3.1 Study region and species
Our study region (1.76 million km2) in northern Australia spanned three states: Queensland,
Western Australia, and the Northern Territory (Fig. 5.1). The region has a tropical climate with a
seasonal rainfall cycle, alternating between a wet (November / December to April / May) and a dry
(May / June to October / November) season. Dominant monsoonal savanna woodlands and semiarid grasslands are interspersed with riverine channels, coastal wetlands, fragments of wet tropical
rainforest and seasonally inundated floodplains (Fox et al. 2001). Feral pigs (Sus scrofa) have
invaded most of the region. They are now reported to be widespread in the east and localised in the
north and west, achieving highest local densities in resource-abundant wetlands and floodplains
(West 2008). Savanna woodlands are not a preferred habitat (Choquenot et al. 1996).
Fig 5.1 Study region in northern Australia (shaded in grey).
5.3.2 Habitat connectivity model
5.3.2.1 Overview
We followed four broad methodological steps to model patch connectivity (Fig 5.2). Each step
was applied to two seasonal scenarios: the late wet season (March to April), when resources
required by feral pigs are abundant and widely distributed, and the late dry season (October to
November), when resources are scarce and scattered across the landscape. First, we parameterized
matrix resistance to dispersal from a previous resource-based habitat suitability model (Chapter 4).
105
Second, we delineated habitat patches large enough to sustain local breeding populations based on
patch quality and size thresholds also derived from the habitat suitability model. Third, we modelled
matrix connectivity using the methodology developed by Pelletier et al. (2014) for regional-scale
applications of the circuit-theoretic approach (McRae et al. 2008). Fourth, we estimated patch
connectivity, by computing, for each habitat patch, the total area the patch is connected to via
dispersal corridors. For each seasonal scenario, we computed 27 alternative patch connectivity
values based on three-way changes in ‘matrix resistance to dispersal’ (ResDisp), ‘inter-patch
dispersal distance’ (DispDist) and ‘corridor delineation threshold’ (CDThresh). Modelling
procedures and parameter estimates are explained in detail below.
5.3.2.2 Matrix resistance
Matrix resistance refers to the permeability of landscape elements to movement or the costs
(energy expenditure, mortality risk) associated with traversing them (Kindlmann & Burel 2008;
Zeller et al. 2012). Resistance surfaces, which assign species-specific cost values to each matrix
pixel, can be parameterized in several ways (Zeller et al. 2012). Here, we adopted a resource-based
approach (Hartemink et al. 2015) using a previous seasonal habitat suitability model (Chapter 4).
The model integrated expert-elicited Bayesian networks and spatial pattern suitability analysis to
characterize habitat quality for feral pigs as a function of complementary access to four key
resource requirements (water, food, heat refuge and disturbance refuge) within heterogeneous home
ranges. Mapped ‘habitat suitability indices (HSI)’ were empirically validated (Chapter 4).
Here, we computed ‘matrix resistance to dispersal (ResDisp)’ at a fine spatial resolution of 1 ha
by reversing the modelled HSI (Chapter 4; see Beier et al. 2007 for a similar approach). The
resource-based model specifically characterized habitat quality for feral pig breeding. To account
for gender-specific differences between dispersal-constrained breeding herds and wide-ranging
boars (Choquenot et al. 1996; Gabor et al. 1999; Milne et al. 2008), we parameterized three
alternative resistance levels. We linearly converted HSI to range between either 1-5 (‘low’
ResDisp), 1-21 (‘moderate’ ResDisp) or 1-100 (‘high’ ResDisp), while reversing values so that
higher HSI obtained lower resistance values using the formula:
ResDisp = ResDispmax – (((HSI – HSImin) / (HSImax – HSImin)) * (ResDispmax – ResDispmin))
Narrower ranges imply greater ease of movement through low quality matrix habitat, which may
approximate dispersal by boars. Wider ranges place higher penalties on such movements, reflecting
more constrained movements by breeding herds. An intermediate scenario was applied to reflect
uncertainty and behavioural variability.
106
Fig 5.2 Methodology for modelling habitat connectivity for feral pigs. For each of four broad
methodological steps, we show methods used, workflow and outputs generated. Methods for the habitat
suitability model (step 1a) are detailed in Chapter 4.
107
5.3.2.3 Habitat patches
First, contiguous habitat patches of sufficient quality and size to allow breeding populations to
persist without needing to interact with other patches were delineated. We used Gnarly Landscape
Utilities: Core Mapper software (Shirk & McRae 2013) with parameters specific to feral pigs.
First, we derived seasonal ‘habitat quality thresholds’ from HSI model validation using the
Continuous Boyce Index method (Chapter 4; Hirzel et al. 2006). We plotted the predicted-toexpected (P/E) ratio, which compares, for each HSI, the proportion of observed feral pig presences
to the proportion of all pixels in a validation background (Fig 5.3). HSI values may be pooled when
the P/E ratio shows a flat or negative slope (Hirzel et al. 2006). We specified a threshold value for
‘high quality breeding habitat’ at HSI ≥ 60 and HSI ≥ 65 for the dry and wet season respectively
(Fig 5.3). Core Mapper applied thresholds to individual pixels and neighbourhood averages in order
to exclude small suitable areas embedded in low quality habitat (Beier et al. 2007; Shirk & McRae
2013). We calculated neighbourhood averages within a 200 m radius and using a slightly reduced
threshold (HSI ≥ 59 (dry) and HSI ≥ 64 (wet)) so that linear features such as rivers, which are an
important feral pig habitat especially in semi-arid areas (Cowled et al. 2008), were not removed.
108
Fig 5.3 Validation plots for seasonal habitat suitability models. Predicted-to-expected (P/E) ratios are
plotted for each of four independent validation data sets (thin labelled lines) and as a smoothed average of all
data sets (thick line). Vertical dotted lines indicate HSI thresholds used for patch delineation.
Next, Core Mapper ‘patch expansion’ was used to merge suitable areas separated by narrow
gaps of low quality habitat (Girvetz & Greco 2007; Shirk & McRae 2013). We assumed that feral
pigs can cross these gaps because of their home range mobility. We derived a movement radius of
2.33 km by averaging expert estimates on home range boundaries (Chapter 4). In Core Mapper,
these movements were cost-weighted according to matrix resistance (Shirk & McRae 2013).
Finally, merged patches below a specified ‘minimum patch size’ were considered insufficient to
support independent breeding populations and removed (Beier et al. 2007; Shirk & McRae 2013).
We defined this at 17.1 km2 by averaging the home range sizes suggested by experts in Chapter 4,
assuming that herds can persist within one resource-rich home range neighbourhood (Choquenot et
al. 1996). Patch delineation in Core Mapper was affected by matrix resistance, resulting in networks
of patches differing in shape, distribution and overall number for each resistance level.
5.3.2.4 Matrix connectivity
We used a circuit-theoretic approach, implemented in the Circuitscape model (McRae & Beier
2007; McRae et al. 2008; McRae et al. 2013) to model ‘matrix connectivity’ across the region.
Circuitscape treats the landscape analogous to an electrical circuit, passing current between pairs of
nodes (patches) through an intervening network of resistors (matrix pixels). Circuitscape computes
the current density for each pixel, which can be interpreted as the probability of a species moving
through that pixel via random walk theory (McRae et al. 2008). We used a recent extension of
Circuitscape described by Pelletier et al. (2014), which models movement probabilities across
entire landscapes rather than between nodes. It does not require delineated habitat patches
(Moilanen 2011; Koen et al. 2014) and overcomes computational limitations for regional scale
applications (see Leonard et al. 2016 for a high-performance computing solution).
We implemented the methodology for each level of matrix resistance using the R 3.2.2 and
Circuitscape 4.0 software (McRae et al. 2013; R Core Team 2015; Appendix 5.1). It involved (a)
dividing the 175 million study region pixels into 127 square tiles (150 x 150 km); (b) adding a 150
km buffer around each tile, filled with values drawn randomly from the resistance surface to
minimize border effects (Koen et al. 2010); (c) running the Circuitscape model in pairwise mode,
passing directional current between two parallel nodes placed as thin (one pixel-wide) strips along
the outer (west-east or north-south) borders of each buffered tile; (d) removing buffers and
reassembling tiles into directional mosaics spanning the entire study region and (e) computing
109
‘omnidirectional current density (OCD)’ by multiplication (Pelletier et al. 2014). Finally, we log10
transformed OCD (OCDlog10) for better visualisation and identification of consistent ‘corridor
delineation thresholds (CDThresh)’ for all three levels of matrix resistance. The CDThresh
parameter was used to distinguish between dispersal paths and those parts of the matrix where no
dispersal occurs. We selected three arbitrary threshold levels reflecting different levels of habitat
quality required for dispersal by breeding herds (‘strict’ CDThresh where dispersal is possible only
in high quality habitat with an OCDlog10 ≥ -7.25) and solitary boars (‘loose’ CDThresh where
dispersal can occur in most habitats with an OCDlog10 ≥ -7.75). An intermediate scenario
(‘moderate’ CDThresh with -7.25 > OCDlog10 ≥ -7.75) was again applied to reflect uncertainty and
behavioural variability.
5.3.2.5 Patch connectivity
After pairing delineated patches and modelled OCD based on the level of matrix resistance
(high, moderate or low), we computed patch connectivity for each delineated habitat patch and for
the entire patch network (Rayfield et al. 2011). At the patch level we measured the total habitat area
(including breeding habitat within patches and dispersal habitat within the matrix) that each patch is
connected to (Moilanen & Nieminen 2002; Rayfield et al. 2011). First, we allocated delineated
dispersal habitat (see above) to the nearest habitat patch. We defined three alternative ‘inter-patch
dispersal distances (DispDist)’ at 5, 10 and 15 km, again reflecting differences for breeding herds
and solitary boars (Choquenot et al. 1996; Cowled & Garner 2008). These were implemented as
accumulative cost distances within uniformly weighted corridors (each pixel having a cost of ‘100
m’) to allow for non-linear corridor dispersal (Moilanen & Nieminen 2002; Galpern et al. 2011).
Next, we merged habitat patches that were connected via cost-weighted dispersal into ‘habitat
components’ (Rayfield et al. 2011), and assigned each individual habitat patch a value reflecting the
total size of its habitat component. Finally, we computed network-level patch connectivity by
aggregating patch-level values across the study region (Galpern et al. 2011). Analyses were
implemented in the ESRI ArcGIS 10.2 Spatial Analyst and ModelBuilder software (Appendix 5.2).
5.3.3 Model evaluation
To account for a lack of empirical data about model parameters, and intra-specific differences
between breeding herds and solitary boars, we evaluated patch connectivity under a range of model
assumptions. For each season, we computed 27 alternative patch connectivity scenarios in response
to three-way changes in three key parameters affecting various modelling steps (Fig 5.2; Galpern et
al. 2011, Moilanen 2011): ‘matrix resistance to dispersal’ (ResDisp, levels: high, moderate, low)
influenced patch delineation and matrix connectivity; ‘corridor delineation threshold’ (CDThresh,
110
levels: strict, moderate, loose) and ‘inter-patch dispersal distance’ (DispDist, levels: 5 km, 10 km,
15 km) influenced patch connectivity.
We evaluated network-level effects analogous to a factorial experiment. We focused on
differences in median, rather than mean, size of habitat components because we anticipated nonGaussian distributions dominated by many small components. First, we used the Kruskal-Wallis
rank sum test to probe for significant effects of changes in ResDisp irrespective of DispDist and
CDThresh. Where significant differences in the distribution of component sizes were found, we
performed two-way comparisons between ResDisp levels using the Mann-Whitney U test. Next we
applied the same non-parametric tests of significance to the parameters CDThresh and DispDist at
each ResDisp level. Separate evaluation was necessary because the shape, location and number of
patches to which component sizes were assigned differed between levels. We also investigated
possible interactions between CDThresh and DispDist. To visualise changes in patch connectivity
due to interaction effects, we linearly transformed absolute component sizes into relative values
(range 0-1: minimum to maximum area each patch is connected to). All analyses were implemented
in the R 3.2.2 software (R Core Team 2015).
5.4
Results
5.4.1 Matrix connectivity
There were large seasonal differences in matrix connectivity for feral pig dispersal. According
to variations in resource availability, large tracts of additional dispersal habitat became available in
inland areas during the wet season (Fig 5.4). The level of matrix resistance (ResDisp) assumed
during modelling also affected the amount of dispersal habitat (Table 5.1). As matrix resistance
influenced the ease of movement through low quality habitat rather than the spatial distribution of
resources, lower resistance levels mostly widened dispersal paths without adding dispersal habitat
in different regions. A notable exception occurred when matrix resistance was ‘low’ and the
corridor delineation threshold was ‘loose’. With this parameter combination, all matrix habitat was
considered suitable for feral pig dispersal in both seasons (Table 5.1). Spatial data files showing
seasonal matrix connectivity for each ResDisp level will be openly available upon publication of
this Chapter (to be submitted to Data Dryad).
111
Fig 5.4 Seasonal matrix connectivity in northern Australia (for moderate ‘matrix resistance
(ResDisp)’). Dispersal habitat is shown for three ‘corridor delineation thresholds (CDThresh)’. A
methodological artefact increased ‘omnidirectional current density (OCDlog10)’ at some tile corners. This
occurred mostly in low quality habitat (white areas) and had minimal impact on patch connectivity.
112
Table 5.1 Amount of dispersal habitat in the study region.
Wet season scenario
ResDisp
CDThresh (% of study region)
strict
mod
loose no disp.
high (1-100)
4.7
19.4
44.8
55.2
mod (1-21)
6.3
27.5
57.6
42.4
low (1-5)
12
41
100
0
Dry season scenario
ResDisp
CDThresh (% of study region)
strict
mod
loose no disp.
high (1-100)
1.8
4.3
12.5
87.5
mod (1-21)
2.2
5.9
20.1
79.9
low (1-5)
3.1
11.2
100
0
For each level of ‘matrix resistance (ResDisp)’, we compare the proportion of dispersal habitat for each ‘corridor
delineation threshold (CDThresh)’, compared to the proportion of non-dispersal habitat (shortened to ‘no disp.’).
5.4.2 Patch connectivity
5.4.2.1 Network-level connectivity
At the network level, patch connectivity also varied strongly between seasons (Fig 5.5 and
Appendix Table 5.3.1). The median and mean size of connected habitat components ranged
between 341-411 km2 and 7,820-45,174 km2 respectively during the dry season. During the wet
season, habitat component sizes increased considerably to 2,697-6,932 km2 (median) and 32,46386,139 km2 (mean). Sizes were distributed unevenly at all levels of matrix resistance, with many
small (isolated), some intermediate and few large (well-connected) habitat components (Fig 5.5).
Interestingly, while mean size increased steadily with decreasing resistance, median size remained
near constant (Appendix Table 5.3.1). This reflects a large increase in size of few well-connected
habitat components, and the addition of many small and isolated patches exceeding the specified
‘minimum patch size’ (doubling the overall patch count). When matrix resistance was ‘high’ or
‘moderate’, resource-poor environments (e.g. inland areas and all of Western Australia; Fig 5.1)
contained few or no patches. When testing for statistical significance, we found that the distribution
of component sizes was significantly different for ‘low’ resistance compared to both other levels in
both seasons (p < .001). There was no significant difference between ‘moderate’ and ‘high’ matrix
resistance in either the dry (p = .704) or the wet (p = .098) season (Appendix Table 5.3.1).
113
At each level of matrix resistance, we plotted and analysed median component size in response
to two-way changes in the ‘corridor delineation threshold (CDThresh)’ and ‘inter-patch dispersal
distance (DispDist)’ parameters (Fig 5.6 and Appendix Table 5.3.2). Using the Mann-Whitney U
test, we found significant effects for both parameters, as well as interaction effects between them,
but not for each factorial combination. For example, during the dry season, DispDist significantly
affected patch connectivity only when CDThresh was loose or when CDThresh was moderate and
ResDisp was low or moderate. Otherwise DispDist was not significant. Conversely, CDThresh had
significant effects at each factorial combination of ResDisp and DispDist. Yet, a major increase in
median component size was only observed when CDThresh changed to loose (Fig 5.6a-c and
Appendix Table 5.3.2a). During the wet season, DispDist and CDThresh interacted significantly at
all factorial combinations except when CDThresh was strict and ResDisp was high or moderate (no
significant effects of DispDist) or when CDThresh was strict and ResDisp was low (significant
decrease in patch connectivity only for DispDist = 5 km) (Fig 5.6d-f and Appendix Table 5.3.2b).
5.4.2.2 Patch-level connectivity
Individual habitat patches were affected differently by changes in the three key parameters
ResDisp, CDThresh and DispDist. We show results for two target patches (Fig 5.7). Patch A was
relatively large and surrounded by numerous other patches in a resource-abundant region in the
north-eastern study region. Patch B was much smaller and isolated, but in the vicinity of seasonally
resource-abundant coastal floodplains in the northern study region (Fig 5.7).
The total area of breeding and dispersal habitat connected to patch A increased steadily during
the wet season (median range 77,444 to 207,276 km2) and somewhat more abruptly during the dry
season (median range 2,322 to 94,269 km2) as model parameters became more permissive to
dispersal (Fig 5.7 boxplots and map panels d-f). This was partly due to increasing patch size via
intra-patch gap crossing and partly due to greater connectivity via inter-patch dispersal. In the dry
season, patch B was below delineation thresholds when ResDisp was high. It remained small and
isolated except for a sharp increase in component size (from values < 400 km2 to 40,852 km2) when
all model parameters were at their most permissive levels (ResDisp low, CDThresh loose and
DispDist 10 or 15 km). During the wet season, component size also dramatically increased (from
values < 3,000 km2 to values > 30,000 km2) under moderate levels of ResDisp and CDThresh when
DispDist was at its highest level 15 km (Fig 5.7 boxplots and map panels a-c).
Other, more isolated, habitat patches in inland areas had limited patch connectivity irrespective
of model assumptions. Spatial data files showing seasonal patch connectivity for all patches and
each factorial combination (two seasonal scenarios and three ResDisp / CDThresh / DispDist levels)
will be openly available (to be submitted to Data Dryad upon publication).
114
Fig 5.5 Seasonal patch connectivity for feral pigs in northern Australia. Median, interquartile range,
outlier and mean size of connected habitat components are shown for each ‘matrix resistance’ level.
Fig 5.6 Interactions between ‘corridor delineation threshold (CDThresh)’ and ‘inter-patch dispersal
distance (DispDist)’ at each level of ‘matrix resistance (ResDisp)’ for the dry (a-c) and wet (d-f) season.
We plot relative changes to patch connectivity, where 0 and 1 corresponded to the minimum and maximum
component size assigned to each patch. While this allowed comparing interactions across components of
different absolute size (~20-330,000 km2), it obscured the magnitude of effects.
115
116
117
Fig 5.7 Seasonal patch connectivity for two selected habitat patches of feral pigs in northern Australia.
For each ‘matrix resistance (ResDisp)’ level, we report median, interquartile range, outlier and mean size of
habitat components connected to the target patch (note different y-axis scale). Map panels a-f show the target
patch (blue) in the context of surrounding patches (colour-coded by the median size of their habitat
component) and dispersal habitat (shown for one ‘corridor delineation threshold (CDThresh)’ per panel and
colour-coded in shades of grey for different ‘inter-patch dispersal distances (DispDist)’).
5.5
Discussion
Directly transmitted animal diseases in wildlife can spread between infected and susceptible
populations via host dispersal (Anderson et al. 1986). Risk-based disease management requires
better spatial knowledge of the links (dispersal paths or ‘corridors’) between populations (Riley
2007; Cowled & Garner 2008; Rioux Paquette et al. 2014). Habitat connectivity models are well
suited to deliver such knowledge, but have not been widely applied to investigate risk of disease
spread (Kool et al. 2013). Here, we modelled habitat connectivity for feral pigs, an abundant
invasive species that is susceptible host to numerous infectious diseases, in northern Australia. Our
resource-based (Hartemink et al. 2015) modelling approach could readily be applied to other
wildlife hosts with a metapopulation structure, especially when empirical data is limited and risk is
investigated over large extents. Results highlighted seasonal variability in ‘patch connectivity’, an
indicative measure of disease spread risk. Model evaluation further suggested that wide-ranging
boars are a greater risk factor for disease spread than resource-constrained breeding herds. First, we
discuss results on ‘matrix’ and ‘patch’ connectivity, focusing on ecological interpretation and
methodological evaluation. Second, we discuss implications for risk-based disease management.
5.5.1 Matrix connectivity
We modelled matrix connectivity for two seasonal scenarios using the circuit-theoretic
approach (McRae & Beier 2007; McRae et al. 2008). This framework was suited to our study aims
as it (a) modelled movement probabilities along all possible dispersal paths (which is indicative of
disease spread risk) and (b) did not rely on strict distinctions between nodes (patches), edges
(corridors) and the matrix (which is likely not how generalist omnivores such as feral pigs perceive
and use the landscape; Moilanen 2011). The approach may be less useful for other systems, e.g.
when modelling movement along established migration routes (McClure et al. 2016).
Circuit-theoretic models can be computationally prohibitive when applied over large areas
(McRae et al. 2013). We used a recent extension (Pelletier et al. 2014) that models matrix
connectivity (as ‘omnidirectional current density’) on buffered tiles. These can join to form
arbitrarily large study areas (here 175 million pixels). Our automated implementation (Appendix
118
5.1) may be useful to future regional-scale connectivity models. We detected increased current
density at some tile corners. This artefact of tile buffering (Koen et al. 2010; Pelletier et al. 2014)
occurred mostly in low quality habitat, when resistance in the tile greatly exceeded resistance in its
randomly filled buffer. While this had minimal impact on our analyses of patch connectivity
(affected tiles contained no or few patches), it may disturb other applications. In heterogeneous
study areas, we suggest filling buffers with resistance values drawn randomly from each tile rather
than the entire study area (Koen et al. 2010) to ensure matching value ranges and eliminate border
effects.
Matrix connectivity was modelled independently of habitat patches (Koen et al. 2014), based
solely on resistance to dispersal. However, we did not have empirical data from movement or
landscape genetic studies (Kool et al. 2013) to parameterize matrix resistance. Instead, we used
converted results from a previously developed habitat suitability model (Chapter 4). While this is
not unusual (Beier et al. 2007; Zeller et al. 2012), and justified for feral pigs in Australia where
inter-patch dispersal does not appear to be limited by anything but habitat quality (Choquenot et al.
1996; Cowled & Garner 2008), it is arguably subjective. Moreover, modelled ‘omnidirectional
current density’ is a relative measure which has not yet been directly linked to ecological factors
that distinguish between dispersal corridors and non-dispersal matrix (Pelletier et al. 2014).
Our scenario approach evaluated the effects of a range of parameter estimates. We explicitly
linked these scenarios to different dispersal abilities between two distinct social entities in feral pig
populations (Choquenot et al. 1996; Gabor et al. 1999): ‘high’ resistance and a ‘strict’ corridor
delineation threshold may approximate dispersal by resource-constrained breeding herds; ‘low’
resistance and a ‘loose’ corridor delineation threshold may reflect dispersal by wide-ranging boars;
intermediate scenarios reflected empirical uncertainty and behavioural variability. Results showed
that season, matrix resistance and corridor delineation threshold all had a major effect on matrix
connectivity. This suggests that dispersal patterns in northern Australia may differ considerably
with climatic conditions and between feral pig breeding herds and solitary boars. Importantly,
matrix connectivity should be validated against movement data, or to test the assumed link to
disease spread, against epidemiological observations. However, at the time of this study such data
was unavailable or incomplete. Our results highlight that future data collection must consider
seasonal and intra-specific differences and should be stratified by broad environmental gradients to
capture variability in feral pig dispersal between resource-poor and resource-abundant landscapes.
5.5.2 Patch connectivity
We used modelled matrix connectivity to compute and evaluate connectivity between habitat
patches. Patch delineation in Core Mapper required definition of multiple thresholds for ‘habitat
119
quality’ and ‘minimum patch size’ (Shirk & McRae 2013). For feral pigs, we aimed to maintain
narrow patches (e.g. rivers in semi-arid environments) and allow for considerable gap crossing
abilities. Nevertheless, large inland tracts and all of Western Australia contained no or very few
delineated patches. This is inconsistent with previous research (West 2008; Cowled et al. 2009).
Our quality thresholds were derived from a model that was validated mostly in resource-abundant
environments in the eastern and northern parts of the study region (Chapter 4). Different quality or
size thresholds may in fact apply to habitat patches in inland environments or Western Australia.
Alternatively, previous research may have overestimated distribution in drier landscapes that, at
best, support small and isolated populations. Empirical research is needed to better understand
habitat use and movements (both intra-patch gap crossing movements and inter-patch dispersal) of
feral pigs in resource-poor regions of northern Australia.
For connecting habitat patches, we introduced a third parameter, inter-patch dispersal distance.
All three evaluated model parameters significantly affected patch aggregation into habitat
components. Interestingly, during the dry season, patch connectivity increased most as each
parameter changed to its most permissive level (‘low’ matrix resistance, ‘loose’ corridor delineation
threshold, or ‘15 km’ inter-patch dispersal distance). During the wet season, however, patches
became steadily more connected as parameters placed fewer constraints on dispersal. This suggests
that, during the dry season, patch connectivity for dispersal-constrained breeding herds of feral pigs
is low, even when allowing for uncertainty about their dispersal ability. However, unconstrained
dispersal by solitary boars in all habitat types and over long distances may dramatically increase
connectivity between habitat patches for these individuals. During the wet season, when resources
are more abundant and widely distributed, patch connectivity may still be greatest for boars, but less
different to that of breeding herds.
These network-level generalizations omit differences between patches, depending on their
location in the patch-network. We distinguished three broad types of patches: First, all modelled
scenarios contained many small isolated habitat patches. These were often located in inland areas
not connected to other patches irrespective of assumptions. Second, many small patches interacted
with surrounding patches to form larger habitat patches (via intra-patch gap crossing) or connected
habitat components (via inter-patch dispersal). These were mostly located in resource-abundant
areas and became increasingly connected to additional breeding habitat as parameters became more
permissive to dispersal (e.g. patch A in Fig 5.7). Third, some habitat patches were isolated under
most dispersal scenarios but became connected to large habitat components during the wet season
or for wide-ranging boars. These were usually located in marginal habitat adjacent to resourceabundant regions and could act as “stepping stones” between components (e.g. patch B in Fig 5.7).
120
Our interpretations were based on measured patch connectivity in response to three-way
changes in key parameters. However, it is uncertain whether evaluated levels reflect the full
parameter space of feral pig dispersal. More empirical research is needed to improve model
parameterization (Kool et al. 2013), especially for the wet season when any difference in dispersal
ability had a significant effect on patch connectivity. Further, recasting delineated patches and
modelled matrix connectivity as a network graph with strictly defined nodes and links (Pelletier et
al. 2014) may allow further analyses: link thresholding could determine critical dispersal thresholds
at which network connectivity changes most; patch prioritisation could help to rank patches based
on their contribution to connectivity (Galpern et al. 2011).
5.5.3 Implications for disease management
Our results can directly inform risk-based management of infectious animal diseases at all
stages of invasion (Cowled & Garner 2008; Milne et al. 2008).
First, effective early detection surveillance should focus on patches with high risk of disease
occurrence (East et al. 2013). Spatial data files showing seasonal patch connectivity across northern
Australia for a range of assumptions about feral pigs’ dispersal ability will be openly available upon
publication of this chapter (to be submitted to Data Dryad). From this data, connected habitat
components can be identified. When combined with information on pathways for disease
introduction and, ideally, with estimates of potential feral pig abundance within each component,
surveillance programs could be designed so that each habitat component is adequately sampled for
disease occurrence. Our data could also be further analysed using a graph-theoretic approach to
prioritize individual patches that contribute most to connectivity within and between habitat
components for surveillance (Galpern et al. 2011). Finally, surveillance activities could be
optimized across time and space for such patch networks according to mathematical efficiency rules
(Chades et al. 2011).
Second, we modelled patch connectivity, which can be interpreted as an indicator of disease
spread risk. Deeper, dynamic insights into disease-host interactions and the spread of epidemics for
better incursion preparedness could be gained through epidemiological simulation models (Ostfeld
et al. 2005; Riley 2007; Milne et al. 2008). Our data on patch and matrix connectivity could be
integrated to realistically constrain pathways of spread or the extent of epidemiologically connected
zones (Cowled & Garner 2008; Rees et al. 2013, Macpherson et al. 2016). Epidemiological models
could also elucidate the relationship between connectivity and disease transmission or persistence in
wildlife hosts. While we implicitly assumed that connectivity increases risk, this is not necessarily
the case (Plowright et al. 2011; Huang et al. 2015).
121
Third, our results could guide management responses to disease incursions. For example, patch
connectivity data could be used to identify the type of infected patch and derive suitable
interventions. In isolated patches, disease spread could be easily contained and risk is minimal.
Patches that are mostly isolated but connected to large components under dispersal-permissive
assumptions could act as “stepping stones”. Management should aim to isolate the patch and
constrain boar and wet season dispersal, which may spread disease to susceptible nearby
components. Large patches or habitat components consisting of several highly connected patches
should be treated as contiguous management units with limited opportunities for containment. The
focus should be on efficient disease management within the infected component. Our matrix
connectivity data could help to identify transmission pathways between infected and susceptible
patches of any type, and design interventions that reduce connectivity. This could include
population control or vaccination in dispersal corridors, fencing as a dispersal barrier or to prevent
access to key resources. Importantly, effective control of connectivity corridors around an infected
habitat component may help justify declaration of disease-free compartments (Scott et al. 2006).
Functional connectivity models can yield useful insights on the spread risk of infectious
diseases with wildlife hosts and should be in the toolbox of epidemiologists, ecologists and
practitioners interested in combating this growing biosecurity threat.
5.6
Appendices
Appendix 5.1 Methodology for modelling omnidirectional current density (annotated R code).
Appendix 5.2 Methodology for computing patch connectivity. (ModelBuilder process)
Appendix 5.3 Two-way tests of significance between levels of ResDisp, CDThresh and DispDist
122
Chapter 6 General discussion and conclusions
Summary
In this chapter I relate the results from Chapters 3 to 5 to the aims of this thesis. For each of the
three research questions, I discuss the main findings of my research as well as their methodological,
ecological and applied significance 11. Next, I synthesize how the adopted multi-level modelling
approach and its outputs can help to inform risk-based management of infectious animal diseases
for which feral pigs are a wildlife host. I conclude by summarizing limitations and giving
recommendations for future research that may further improve on this thesis’ findings.
6.1
Research question 1 – resource selection by individual breeding herds
How can habitat suitability for mobile species such as feral pigs be reliably modelled at the
regional scale, given uncertainty about the processes of habitat selection, in particular the
resource-seeking home range movements by individual breeding herds?
The purpose of the first research question was to lay the foundation for modelling habitat
suitability for feral pigs at the regional scale under empirical uncertainty. Like other mobile
animals, feral pigs are able to access required resources at different locations within heterogeneous
home ranges (Powell & Mitchell 2012). The suitability of a focal site to function as habitat thus
depends on landscape structure – that is, the availability and patterning of complementary and
supplementary resources (Section 2.1.3; Addicott et al. 1987; Dunning et al. 1992).
6.1.1 Main findings
In Chapter 3 I developed an integrated methodology that can be used to model habitat
suitability for mobile animals at regional extents, while also capturing resource-seeking home range
movements by individuals. I adapted a resource-based framework (Hartemink et al. 2015) using
Bayesian networks that has increasingly been applied to habitat models (Section 2.4.2). However,
Bayesian networks are inherently spatially ‘neutral’ or ‘implicit’. Few studies have accounted for
spatial interactions between variables (Section 2.4.3). Here, I developed a novel combined
methodology, spatial pattern suitability analysis, for capturing feral pigs’ selection of
supplementary resources within home ranges. This involved: (a) measuring structural patterns of
‘resource quality’ within home range-sized moving windows around each study region pixel, and
(b) using expert-elicited ‘response-to-pattern’ curves to relate these structural measures to
11
Due to the stand-alone nature of journal submissions, most of the key findings and discussions of their significance
have similarly been stated in Chapters 3 to 5 (submitted or being prepared for submission). Here, I attempt to reiterate
their relevance with regard to answering the research questions of this thesis.
123
functional ‘resource suitability’ indices that reflect the availability of resources in the landscape.
After applying this method to each resource required for sustained feral pig breeding (food, water,
protection from heat and protection from disturbance), I modelled habitat suitability as a function of
all complementary resource suitability indices.
For the purpose of this thesis, I investigated whether this integrated methodology improved
model accuracy in predicting habitat suitability for feral pig breeding and, if so, which aspects of
landscape structure and measurement scales are best suited to describing resource-seeking home
range movements by individual feral pig breeding herds in northern Australia. I explored this
question through comprehensive scenario analysis and model validation against three independent
distributional data sets from a subsection of the study region in northern Queensland. ‘Landscape
models’ that integrated landscape-scale ‘resource suitability indices’ based on either distance-,
composition- or combined distance / composition-dependent measures explained observed feral pig
presences better than a ‘control model’ that used site-scale ‘resource quality indices’ in 90% of
evaluated scenarios. However, there were differences between landscape models and validation data
sets. Overall, models that measured resource suitability as a function of distance (i.e. assigned
greater value to adjacent than distant resources) performed consistently best against all validation
data. Validation statistics were similar at all measurement scales (moving window radius = 1, 2 or 3
km), but most consistent at the intermediate scale (r = 2 km). Models that measured resource
suitability as a function of landscape composition or a combination of distance and composition (i.e.
the averaged quality or distance-weighted averaged quality of all resources in the home range)
displayed mixed results, ranging from very high to worse than random accuracy.
6.1.2 Significance and advances
Methodologically, the application of Bayesian network habitat models to mobile animals was
improved by accounting for resource-seeking home range movements. The Bayesian network
approach was preferred over more widely used correlative or mechanistic methods because: (a) it
could be applied at the regional extent, under empirical uncertainty and independent of
distributional data by utilizing unpublished expert knowledge; (b) it provided a robust statistical
framework for modelling habitat suitability as a function of complementary resources; and (c)
uncertainty in expert-elicited parameter estimates was explicitly accounted for (Uusitalo 2007;
Wilhere 2012; Landuyt et al. 2013). Spatial pattern suitability analysis essentially converted sitescale into landscape-scale variables that could be used to model habitat suitability consistent with
landscape ecological principles. Compared to previous statistical or expert-based approaches for
integrating ‘landscape variables’ into habitat models (Section 2.4.3), the methodology offered three
main advantages: First, landscape-scale resource suitability indices were computed from response124
to-pattern curves, which provided ecological meaning to structural metrics and enhanced
transferability across large study areas. Second, response-to-pattern curves were elicited from
experts using a structured method that was easily implemented and understood by expert
practitioners and may be useful to future studies that are unable to utilize or generate empirical data.
Third, landscape variables were derived from modelled site-scale indices of resource quality that
reflected the species-specific value of supplementary resource items (Dunning et al. 1992).
Through comprehensive scenario analysis and validation, the methodologies’ application to
feral pigs also yielded ecologically significant results. The finding that habitat suitability for feral
pig breeding may be reliably described by measuring the distance to a suite of supplementary and
complementary resource patches confirms previous empirical results (Caley 1993; Choquenot &
Ruscoe 2003). However, it has also been reported that feral pig densities respond to the amount of
available food (Caley 1993). In this study, observed feral pig presences in some validation data sets
were also accurately explained by models based on composition-dependent landscape variables.
Yet, results were inconsistent and scale-dependent. Appropriate measurement scales for describing
the mobility limits of feral pig breeding herds in their regular resource-seeking activities were
elicited from experts at either 1 km (n = 1), 2 km (n = 2) or 3 km (n = 3). While evaluation did not
indicate selection of an ‘optimal’ scale, these a priori assumptions (Miguet et al. 2016) are broadly
consistent with the range of empirical home range estimates from Australia (Section 2.2.3). Overall,
findings may not allow conclusive inferences on the mechanisms of resource selection by feral pig
breeding herds. Yet, they do suggest that simple measures of distance are less vulnerable to sitespecific differences in behaviour or data and can be reliably used for integrating resource-seeking
home range movements into regional-scale habitat suitability models.
6.2
Research question 2 – seasonal habitat patches for subpopulations
How are patches of suitable feral pig breeding habitat that potentially support locally dense
populations distributed in northern Australia, and is their spatial distribution influenced by
seasonal effects?
The purpose of the second research question was to fill an existing knowledge gap about the
spatial distribution of feral pigs in northern Australia at spatial and temporal scales (resolution and
extent) that are useful for on-ground applications (Section 2.2.1), in particular the risk-based
management of infectious animal diseases. There are likely seasonal and gender-specific differences
in distribution and habitat use (Caley 1997; Dexter 1999; Mitchell et al. 2009). The research
question focused on those habitat patches that support persistently breeding local subpopulations.
125
6.2.1 Main findings
In Chapter 4 I applied the integrated methodology from Chapter 3 to model and map, for the
first time, seasonal habitat suitability for feral pig breeding and persistence in northern Australia. I
evaluated model sensitivity to habitat variables as well as expert opinion and validated the
seasonally-specific predictions from six individual expert models as well as an expert-averaged
‘consensus’ model against four independent distributional data sets per seasonal scenario. Finally, I
compared the mapped wet and dry season predictions from the consensus model, which produced
consistently accurate results across the study region, at different spatial scales.
Habitat suitability for persistent feral pig breeding was modelled as a function of four
complementary resource requirements – food, water, protection from heat, and protection from
disturbance – each measured as a distance-dependent landscape variable. The consensus model was
least sensitive to expert opinion. All experts parameterized the model so that water and food
resources were more important than heat or disturbance refuges. Overall, model predictions were
robust to validation data. However, there were some differences between expert models, seasonal
scenarios and validation data sets. Model uncertainty was largely attributed to different expert
assumptions about feral pigs’ resource-seeking home range movements as described in elicited
response-to-pattern curves. By averaging across experts, I was able to reduce uncertainty and
produce consistently accurate results for all validation data. As validation data sets were biased
towards the eastern portion of the study region and environments with high proportions of suitable
feral pig habitat, I caution that model accuracy in resource-poor inland environments and in the
western portion of the study region needs further evaluation.
Predicted habitat suitability from the expert-averaged model fluctuated considerably between
seasons. Overall, models revealed a four-fold increase in suitable breeding habitat during the wet
season (36% of the study region), when resources are abundant and widely distributed, compared to
the dry season (9.5% of the study region), when resources are scarce and scattered. Of this, about
one quarter was considered highly or very highly suitable habitat. At the regional scale, patches of
suitable breeding habitat were concentrated in the eastern study region, where they also contracted
less dramatically during the dry season than in the western study region. At the landscape scale,
habitat patches were located predominantly along the coastline and interior riverine and wetland
systems during the dry season and expanded widely across the study region during the wet season. I
suggest that not all suitable wet season habitat may be realized by feral pig subpopulations due to
dispersal constraints. At the local scale, patches of suitable habitat were highly concentrated in
resource-abundant rainforest, wetland and agricultural refuge areas during the dry season. During
the wet season, habitat patches were more evenly distributed among vegetation types and land use
126
classes. This suggests that feral pigs may forage widely into grassland floodplains, savanna
woodlands and coastal shrublands when conditions permit.
6.2.2 Significance and advances
Methodologically, I demonstrated that the integrated modelling approach using Bayesian
networks and spatial pattern suitability analysis can be applied to accurately model habitat
suitability for feral pigs at the regional scale. The resource-based approach (Hartemink et al. 2015)
could readily be adapted to the study of other wide-ranging mobile species, especially in data-poor
situations. Previous habitat models using Bayesian networks have usually been applied to overcome
empirical uncertainty and lack of data. Consequently, models have frequently utilized expert
knowledge, but rarely validated results against independent data (Marcot et al. 2001; Smith et al.
2007; Murray et al. 2012; Tantipisanuh et al. 2014; van Klinken et al. 2015). The present study
developed and implemented a range of innovative techniques for the structured and repeatable
elicitation of expert knowledge (Section 2.6), including: (a) combining the CPT calculator (Cain
2001) and interval judgement (Speirs-Bridge et al. 2010; McBride et al. 2012a) approaches to elicit
conditional probabilities; and (b) using elicited response-to-pattern curves to derive ecologically
meaningful landscape variables. Nevertheless, some ecologists continue to question the value of
expert models. To my knowledge, this study is the first to apply the Continuous Boyce Index
method (Boyce et al. 2002, Hirzel et al. 2006) to empirically validate spatial predictions from a
Bayesian network model against truly independent species observations. Further, I evaluated the
performance of a range of models based on the opinions of individual experts as well as an average
of all expert opinions. Validation added considerably to the expert-based modelling approach,
allowing an evaluation of model uncertainty and providing confidence in the findings.
This study was ecologically significant as it provided the first seasonally-specific estimate of
feral pig breeding habitat in northern Australia. Although seasonal variability in distribution,
density and habitat use has been widely documented in site- or landscape-scale studies (Mitchell et
al. 1982; Hone 1990a; Caley 1993; Choquenot et al. 1996), validated model results confirmed these
findings at the regional scale. Moreover, spatial analyses of mapped results highlighted specific
vegetation types and land use classes that contribute most to seasonal shifts in suitable habitat
patches. Importantly, the model suggested that suitable feral pig habitat is less widely available in
northern Australia than previously thought (e.g. Mitchell et al. 1982; West 2008; Cowled et al.
2009). I suggest that previous overestimates could be confidently corrected because this study (a)
accounted for ecological constraints arising from the dry season and the resource requirements for
breeding herds, and (b) improved model resolution to avoid upscaling error.
127
Finally, the study’s findings have applied significance. The model was specifically designed to
be useful for feral pig management in northern Australia, featuring a regional analysis extent, a
sufficiently fine resolution to capture local variability in conditions, seasonally-specific predictions,
and comprehensive accuracy assessment. Mapped results may thus be confidently used to quantify
impacts, assess risks, justify management investments and target control activities. Further, because
habitat suitability was explicitly referenced to local breeding populations, it is a useful indicator of
the seasonal establishment risk for infectious animal diseases within feral pigs, which often depend
on a persistent supply of young susceptible animals. Results could help to target disease
management activities at high-risk areas (Section 6.4.2).
6.3
Research question 3 – seasonal patch connectivity for metapopulations
How are patches of suitable feral pig breeding habitat connected by potential dispersal
pathways to form a regionally connected metapopulation, and is their connectivity influenced
by seasonal or gender-specific effects?
The purpose of the third research question was to fill a second knowledge gap in our
understanding of feral pig populations in northern Australia, namely the dispersal between local
feral pig subpopulations within larger metapopulations. While population structuring in feral pigs
has been extensively documented, the emergence of population patterns from dispersal processes
and landscape heterogeneity in between patches of suitable breeding habitat has received little
attention (Section 2.2.4). In particular, there are likely seasonal and gender-specific differences in
inter-patch dispersal (Choquenot et al.1996; Gabor et al. 1999; Cowled et al. 2008; Hone 2012).
6.3.1 Main findings
In Chapter 5 I modelled and mapped, for the first time, seasonal habitat connectivity for feral
pigs in northern Australia using a functional, resource-based modelling approach (Hartemink et al.
2015). I applied the empirically validated habitat suitability indices from Chapter 4 to delineate
contiguous patches of feral pig breeding habitat (by identifying suitability thresholds; Hirzel et al.
2006), and to parameterize matrix resistance to inter-patch dispersal, i.e. the cost of traversing
habitat of different quality (by reversing suitability indices; Zeller et al. 2012). Then, I modelled
seasonal ‘matrix connectivity’ using a circuit-theoretic approach, which identified relative
movement probabilities along all possible dispersal paths in the entire landscape (McRae & Beier
2007; McRae et al. 2008). Finally, I computed seasonal ‘patch connectivity’ as the total size of the
habitat component (including breeding and dispersal habitat) that each patch is connected to
(Rayfield et al. 2011). To account for gender-specific differences in dispersal ability as well as
empirical uncertainty in parameter estimates, a range of model assumptions relating to ‘matrix
128
resistance’, ‘dispersal distance’ and a ‘corridor delineation threshold’ (which arbitrarily described
the boundary between dispersal paths and non-dispersal matrix) were explored.
Both matrix and patch connectivity were significantly affected by season as well as model
assumptions. Large tracts of additional dispersal habitat became available in inland areas during the
wet season, when resources were widely distributed across the study region. The assumed level of
matrix resistance mostly affected the width of dispersal paths, but not their spatial distribution
across the region. Patch connectivity was analysed for each delineated habitat patch and in the patch
network by aggregating patch-level values across the study region. At the network level, the median
size of connected habitat components increased from 341-411 km2 in the dry season to 2,697-6,932
km2 in the wet season. Within each seasonal scenario, there were interesting differences in the
effect of model assumptions. During the dry season, patch connectivity increased significantly only
under the most permissive dispersal parameters. During the wet season, however, any change in
model assumptions affected patch connectivity. Overall, there were many small (isolated), some
intermediate and few large (well-connected) habitat components. Although habitat patches were
delineated with the aim of maintaining small, narrow patches (e.g. inland rivers), large tracts of the
interior study region and all of Western Australia contained no or very few delineated habitat
patches in both seasonal scenarios. At the patch level, I identified three broad types of habitat
patches that were affected similarly by model assumptions, depending on their location in the patchnetwork: some were always isolated; some were always connected to very large habitat
components; and some were mostly isolated but became connected to large habitat components
only during the wet season or under the most permissive parameter estimates.
6.3.2 Significance and advances
Methodologically, matrix connectivity was modelled using a recently developed extension of
the circuit-theoretic approach that computes “omnidirectional current density” on buffered tiles
(McRae et al. 2008; Pelletier et al. 2014). This methodology overcomes computational limitations
when applied over large areas (McRae et al. 2013). The automated implementation in this study
(provided as Appendix 5.1) may be useful to future regional-scale connectivity models.
Ecologically significant inferences could be made by linking model assumptions to the genderspecific differences in dispersal ability between two distinct social entities in feral pig populations
(Choquenot et al.1996; Gabor et al. 1999). I assumed that ‘high’ matrix resistance, a ‘strict’
corridor delineation threshold and a ‘low’ maximum dispersal distance may approximate dispersal
by resource-constrained breeding herds; ‘low’ matrix resistance, a ‘loose’ corridor delineation
threshold and a ‘high’ maximum dispersal distance may reflect dispersal by wide-ranging boars;
intermediate scenarios reflected empirical uncertainty and behavioural variability. Given this, the
129
study’s findings suggest that many habitat patches are disconnected for dispersal-constrained
breeding herds during the dry season, even when allowing for uncertainty. However, unconstrained
dispersal by solitary boars may dramatically increase connectivity between patches for these
individuals. During the wet season, connectivity between patches may still be greatest for boars, but
less different to that of breeding herds. Lastly, the paucity of habitat patches in the interior and
western study region is inconsistent with previous research (West 2008; Cowled et al. 2009). This
may reflect errors in patch delineation in this thesis or previous overestimations regarding the
distribution of feral pigs in resource-poor environments that, at best, support small and isolated
populations.
This study is also of applied significance. To my knowledge, it is one of the first to investigate
functional connectivity for a widely distributed invasive species in Australia. Yet, biosecurity and
invasive species management could benefit greatly from such analyses and aim to reduce habitat
connectivity for harmful species (Glen et al. 2013). In this thesis, seasonal patch connectivity for
breeding herds and solitary boars could be interpreted as an indicator of disease spread risk between
infected and susceptible subpopulations (Kramer-Schadt et al. 2007; Cowled & Garner 2008).
Results can inform risk-based management of infectious animal diseases at all stages of invasion,
including early detection surveillance, preparedness and incursion response (Section 6.4.2).
6.4
Synthesis and applications
6.4.1 Contributions to the thesis aim
In the following, I synthesize how the research approach, conceptual framework and presented
findings contributed to achieving the thesis aim: to generate spatially-explicit, seasonally-specific
and regional-scale knowledge of feral pig populations in northern Australia to inform risk-based
management of directly transmitted infectious animal diseases with feral pigs as a wildlife host.
6.4.1.1 Synthesis of the research approach
In summary, this thesis generated new spatial knowledge of feral pig populations in northern
Australia at multiple levels of biological organisation and associated scales (resolutions and extents)
of analysis. The multi-level research approach was specifically designed to be useful for post-border
management of directly transmitted infectious animal diseases with feral pigs as a wildlife host.
(A) Analyses were conducted at the regional extent, covering all of tropical Australia north of
the Tropic of Capricorn. This study region was selected because it is vulnerable to disease
introduction from neighbouring countries and contains widespread feral pig populations.
Consequently, there is a need for effective post-border disease management in feral pigs, yet
130
spatially-explicit knowledge with sufficient detail to inform risk-based management has so far been
limited (Choquenot et al. 1996; Australian National Audit Office 2012).
(B) I focused on improving three types of detail. First, analyses were conducted at a
sufficiently fine spatial resolution (1 ha pixel size) to capture local variability and enable on-ground
applications. Second, analyses introduced a temporal resolution, which has often been missing from
regional-scale studies (West 2008; Cowled et al. 2009). For northern Australia, a scenario approach
encompassing the dominant climatic wet / dry season cycle was deemed adequate, while interannual climatic variability was not considered. Third, analyses contained an ecological resolution
based on feral pigs’ social organisation. Gender-specific differences in feral pig habitat use, home
range movements and dispersal ability were explicitly considered at all levels of analysis
(Choquenot et al. 1996; Gabor et al. 1999; Mitchell 2008).
(C) Due to empirical knowledge gaps at the regional extent, I adopted a resource-based
approach to modelling habitat suitability and connectivity based largely on expert knowledge.
Although such an approach is arguably better suited to making regional inferences than
extrapolations from statistical models (Hartemink et al. 2015), uncertainty is inherent to expert
models. Thus, I attempted to rigorously evaluate model accuracy and assumptions where possible.
In conclusion, efforts were made to balance generality against detail, and maintain
transparency for decision-makers who may ultimately use results to inform management. Specific
contributions to the field of habitat modelling include: extending previous modelling approaches
using Bayesian networks and circuit theory in scope and application; and developing a novel
methodology for integrating home range movements into resource-based habitat models, which may
be usefully applied to other mobile animals in data-poor situations.
6.4.1.2 Synthesis of the conceptual framework
The multi-level perspective on feral pig populations was explicitly linked to a multi-level
perspective on infectious diseases. I conceptualized feral pigs in northern Australia as a
metapopulation and the landscape as displaying a patch-corridor-matrix structure (Forman 1995;
Hanski 1998). Individual breeding herds select supplementary and complementary resources within
home ranges (Dunning et al. 1992). Several herds with adjacent or overlapping home ranges
interact frequently to form contiguous, persistent subpopulations within patches of suitable breeding
habitat. Certain areas in the matrix function as dispersal corridors between patches. However,
regional connectivity between habitat patches varies by gender, with resource-constrained female
breeders having a lower dispersal ability than often solitary males (Choquenot et al. 1996; Gabor et
al. 1999). Movements and distributional patterns at each level may fluctuate seasonally. This
131
structure in the feral pig host population also affects the risk of invasion and persistence by directly
transmitted diseases such as classical swine fever: following introduction, successful disease
establishment is contingent on a locally dense host subpopulation; subsequent disease spread
requires host dispersal between infected and susceptible subpopulations; long-term disease
persistence depends on a persistent supply of susceptible individuals within a regionally connected
host metapopulation (Anderson et al. 1986; Kramer-Schadt et al. 2007; Riley 2007).
Three research questions were formulated to capitalize on these conceptual links by
investigating disease risk as an attribute of the wildlife host. The purpose of research question 1 was
to lay the foundation for modelling subpopulations in habitat patches by understanding how
individual breeding herds select resources within home ranges and what site-level conditions
determine habitat suitability. Research question 2 aimed to generate knowledge about the seasonal
distribution of persistent subpopulations within patches of suitable breeding habitat. Landscapelevel habitat suitability for breeding served as an indicator of disease establishment risk. The
purpose of research question 3 was to generate knowledge about the seasonal connectivity between
local breeding populations. Regional-level patch connectivity served as an indicator of disease
spread risk. Similar combinations of habitat suitability and connectivity models have been applied
to gain a more complete understanding of mobile species’ use of entire landscapes (Cianfrani et al.
2013; Dickson et al. 2013; Stewart-Koster et al. 2015). Resource-based habitat approaches such as
the one adopted in this thesis have recently been recommended for application to the holistic study
of pathogen-host interactions (Hartemink et al. 2015).
6.4.1.3 Synthesis of findings
This thesis showed that habitat suitability for persistent feral pig breeding is dependent on
spatial interactions between four key habitat requirements: water and food resources as well as
protection from heat and from disturbance. Individual breeding herds can access supplementary and
complementary resources within home ranges. Here I showed that habitat suitability of a focal site
is most reliably modelled as a function of distance to each of these complementary requirements.
The availability and quality of resources in northern Australia varies seasonally according to
climatic conditions. This was also reflected in model results. The spatial distribution of feral pig
habitat patches, which support locally dense populations and pose the highest risk of disease
establishment, varied markedly between seasons. Disease establishment risk may be locally highest
during the dry season, when breeding herds must congregate in resource-abundant refuge sites.
Habitat suitability analyses further showed that during the wet season, breeding subpopulations are
able to expand widely. These shifts in patch size and distribution may lead to less locally intense
(due to lower local densities), but more widely distributed (due to increased contact within merged
132
patches) risks of disease establishment. Importantly, findings suggest that dry season conditions
restrict habitat suitability for feral pig breeding much more than previously thought, with many
areas in the region’s interior and western portions containing no or only a few small habitat patches
under that scenario. Due to dispersal constraints, breeding herds may be unable to realize much of
the potentially suitable wet season habitat. Large tracts of the study region may thus present a
minimal risk to disease establishment.
Patch size and distribution also affect the connectivity between local subpopulations, and hence
the risk of disease spread across the region. Connectivity analyses showed that habitat patches were
smaller and less connected during the dry season, especially for breeding herds. Under this scenario,
less dispersal-constrained solitary boars are likely the main carrier of disease between infected and
susceptible local subpopulations. During the wet season, however, many isolated dry season patches
became connected to form larger habitat components. As patch connectivity was less constrained by
dispersal ability, both breeding herds and boars may contribute to disease spread, which greatly
increases overall risk. Nevertheless, many habitat patches remained isolated regardless of season or
gender-specific dispersal. Should a disease be introduced and establish in these local
subpopulations, there is a low risk of subsequent disease spread within the larger metapopulation.
In conclusion, the research contained in this thesis provides, for the first time, a complete and
coherent, spatially-explicit, seasonally-specific and regional-scale picture of areas most at risk of
disease establishment (via host habitat suitability) and spread (via host habitat connectivity) in feral
pigs in northern Australia. The resource-based modelling approach is transparent and flexible, and
could be applied to other invasive species and wildlife hosts of infectious animal diseases,
especially in data-constrained situations and for wide-ranging species.
6.4.2 Applications to risk-based disease management
In the following, I synthesize how these findings could be applied to inform the risk-based
management of directly transmitted infectious animal diseases with feral pigs as a wildlife host. I
use the specific example of classical swine fever to illustrate recommendations. Spatial data files
showing seasonal habitat suitability and patch connectivity to support these applications will be
openly available to decision-makers upon publication of the manuscripts contained in this thesis.
6.4.2.1 Early detection surveillance
Early detection surveillance for classical swine fever in feral pigs in northern Australia is
administered by the Northern Australia Quarantine Strategy (Australian National Audit Office
2012). The program’s risk-based surveillance framework targets broad areas that are deemed to
have a high likelihood of disease occurrence based on estimated risks of introduction, establishment
133
and spread (Cookson et al. 2012; East et al. 2013). Within these areas, surveys are conducted and
feral pigs representatively sampled for classical swine fever prevalence (Cookson pers. comm.).
Our findings could be incorporated into this framework as improved spatially-explicit parameters
describing the risks of disease establishment and spread in feral pigs. Firstly, efficient surveillance
could be more narrowly targeted at seasonally delineated habitat patches instead of broad risk areas,
as these patches have the highest likelihood of disease establishment and detection of an infected
individual. Secondly, likelihood of disease persistence as a function of establishment and spread is
highest in patches that are connected to larger habitat components throughout the year. Such
patches could be prioritised for sampling. Thirdly, when combined with information on the risks of
disease introduction (e.g. ship arrivals from infected zones overseas) and, ideally, with estimates of
feral pig abundance within each habitat component, surveillance programs could be designed so that
each connected metapopulation is adequately sampled for disease occurrence.
6.4.2.2 Incursion preparedness
Incursion preparedness aims at increasing our understanding of the course and impacts of a
disease incursion prior to its occurrence. Given that classical swine fever is currently not present in
Australia, the research contained in this thesis contributes to incursion preparedness by highlighting
those areas most at risk of disease establishment (via host availability) and spread (via host
dispersal) across northern Australia. Here, disease risk was investigated purely as an attribute of the
wildlife host. This view is justified for directly transmitted diseases such as classical swine fever
and allows generalizations to other diseases with feral pigs as a wildlife host. However, it omits
other important factors of disease risk. Epidemiological simulation models could combine host risk
with disease-specific introduction pathways, transmission rates and other factors to generate deeper,
dynamic insights into disease-host interactions and the spread of epidemics for better incursion
preparedness (Section 6.5.3.2; Ostfeld et al. 2005; Riley 2007; Milne et al. 2008). Epidemiological
models could also elucidate the relationship between connectivity and disease transmission or
persistence in wildlife hosts. While it was implicitly assumed in this thesis that connectivity
increases risk, this is not necessarily the case. In some circumstances increased connectivity could
actually reduce the risk of epidemics by encouraging host diversity and immunity or ‘migratory
escape’ (Plowright et al. 2011; Huang et al. 2015)
6.4.2.3 Incursion response
Response to an incursion of classical swine fever in Northern Australia is regulated by the
Emergency Animal Disease Response Agreement (2012), the AUSVETPLAN: Disease strategy
classical swine fever (Animal Health Australia 2012) and the Wild Animal Response Strategy
134
(Animal Health Australia 2011). These guidelines prescribe a range of measures (Animal Health
Australia 2011 & 2012) that can be informed by the findings in this thesis:
(A) Collecting information about feral pig host populations: This thesis provides spatiallyexplicit information about feral pig populations in northern Australia at multiple biological levels
and scales of analysis. Data generation has been designed to be useful for disease management
(regional extent, fine spatial resolution, seasonal scenarios, ecological basis and accuracy
assessment) and data will be openly available upon publication of manuscripts.
(B) Carrying out disease surveillance: (see Section 6.4.2.1)
(C) Preventing spread and containing the disease: Patch connectivity could be used to identify
the type of infected patch and derive suitable interventions. In patches that are always isolated,
disease spread could be easily contained and risk is minimal. Patches that are mostly isolated but
connected to large habitat components under assumptions that place few constraints on dispersal
could act as “stepping stones”. Management should aim to isolate the patch and constrain boar and
wet season dispersal, which may enable disease spread into susceptible habitat components. Large
patches or habitat components consisting of several highly connected patches should be treated as
contiguous management units with limited opportunities for containment.
(D) Controlling susceptible populations to enable disease fadeout via culling or vaccination: If
infected individuals are found within large or well-connected patches, the focus should be on
efficient disease management within the habitat component. Further, matrix connectivity could help
to identify all possible pathways of spread between infected and susceptible patches of any type and
design interventions that reduce connectivity. This could include population control or vaccination
in dispersal corridors, and fencing as a direct dispersal barrier or to prevent access to key resources
required for dispersal.
(E) Demonstrating freedom from disease: Contiguous patches of suitable feral pig breeding
habitat should be regarded as single epidemiological units (Cowled & Garner 2008). Patch
connectivity could help to identify subpopulations with potentially different disease status and
design sampling strategies that demonstrate this. Further, effective control of connectivity corridors
around an infected habitat component may help justify declaration of disease-free compartments
(Scott et al. 2006).
135
6.5
Limitations and future research
6.5.1 Empirical research
Perhaps most importantly, the findings presented in this thesis were derived from resourcebased models that were applied to a large study region (Hartemink et al. 2015). Due to empirical
knowledge gaps about feral pigs’ resource use within habitat patches and dispersal between patches,
models were heavily reliant on expert-elicited knowledge. This was ultimately deemed most
appropriate for the aims of this thesis. While I attempted to rigorously evaluate model results
(Chapters 3 and 4) and assumptions (Chapters 3 and 5), uncertainty is inherent to any expert model.
Further empirical research is needed to test findings and reduce uncertainty.
6.5.1.1 Resource selection by individual breeding herds
While the methodology developed in Chapter 3, spatial pattern suitability analysis, usefully
improved the application of expert-based Bayesian networks to mobile animals, knowledge about
resource selection by individuals in home ranges is more robustly derived from empirical ‘multiscale’ studies (Jackson & Fahrig 2015; McGarigal et al. 2016). For example, here I evaluated feral
pigs’ response to three simple landscape variables (distance-weighted resource quality, averaged
resource quality or distance-weighted averaged resource quality) that were relatively easy to relate
to experts’ field knowledge. Yet, habitat suitability for feral pig breeding may depend on other
aspects of landscape structure such as fragmentation (Jackson & Fahrig 2016), edge proximity
(Dijak & Rittenhouse 2009) or heterogeneity (McClure et al. 2015). Secondly, an ‘optimal’ scale of
measurement for these landscape variables could not be identified. This may have been due to an
insufficient range of elicited scales (experts defined mobility thresholds of 1km, 2km or 3km) or
due to the fact that most models measured selection of all resources (water, food, heat refuge and
disturbance refuge) at the same scale (Jackson & Fahrig 2015; McGarigal et al. 2016). In fact,
individuals may select different resources at different scales (Wiens 1989; Miguet et al. 2016).
Finally, I was only able to determine the ‘most reliable’ landscape variable for describing resourceseeking home range movements by feral pig breeding herds via validation against multiple data
sets. Without empirical validation data, the accuracy of expert estimates remains unclear. A
dedicated empirical multi-scale study could compare a response variable (for example the presence
of tagged female pigs) to a suite of landscape variables measuring different aspects of landscape
composition and configuration at scales differing in orders of magnitude. Empirical data could also
be collected from different types of environments (e.g. resource-poor vs. resource-abundant) and
across seasons and years, so that a differentiated, ‘scale-optimized’ understanding of feral pigs’
habitat use can be gained (McGarigal et al. 2016).
136
6.5.1.2 Seasonal habitat patches for subpopulations
Model-predicted seasonal habitat suitability for feral pig breeding was validated against
observed feral pig presences. Although efforts were made to match the data to model assumptions,
the relationship between model output (potential habitat suitability) and response variable (actual
herd presence) is, at best, indirect (Section 2.1.1). Furthermore, validation data were collected by
third parties, often for management purposes rather than science and in a non-systematic fashion.
Hence, there was considerable uncertainty about potential error (e.g. GPS recording or timing error,
limited ability to detect in closed vegetation, movement and scattering upon aerial detection) and
bias (e.g. ‘sample selection’ or ‘reporting’ bias (Phillips et al. 2009) because highly suitable,
densely populated sites were overrepresented in targeted surveys) in the data. These issues almost
certainly impacted validation results. Finally, all data were collected in environments with an
above-average proportion of suitable feral pig habitat. Model accuracy in resource-poor inland
environments needs further investigation. Systematically collected presence / absence data from a
range of environments could actively seek to minimize sampling bias and increase confidence in
model accuracy. Even more beneficially, data that are more directly related to habitat suitability
(e.g. home range movement trajectories of individual breeding herds acquired from telemetry
studies) would greatly enhance the quality of validation.
A related issue is the accuracy of spatial proxies linked to model explanatory variables. Here,
potential problems may pertain to mapping, resampling (all proxies required a common model
resolution), averaging (proxies for seasonally-specific model variables reflected average conditions
over a five-year time period and may have failed to capture resources which were actually available
at the time of a given detection in the validation data) or reclassification (proxies may have
inadequately reflected the defined states of model variables, Appendix 4.2) error (Hunsacker et al.
2001; Boyce et al. 2002; van Klinken et al. 2015). None of these issues could be quantified in this
study due to lacking information. I suggest that field-validating the mapping accuracy of resampled
spatial proxies and whether they match the states of model explanatory would be highly beneficial.
6.5.1.3 Seasonal patch connectivity for metapopulations:
No empirical data was available to inform the habitat connectivity model (Chapter 5). I used
the expert-based, but empirically validated habitat suitability model (Chapter 4) to parameterize
matrix resistance and delineate habitat patches. While somewhat arbitrary procedures are not
unusual in connectivity models due to data constraints (Beier et al. 2007; Girvetz & Greco 2007;
Moilanen 2011; Zeller et al. 2012), they are arguably subjective and non-desirable. Further, I
evaluated the effect of a range of model scenarios to address uncertainty in parameterization, but (a)
whether these assumptions reflect the full parameter space of feral pig dispersal remained unknown
137
and (b) model results were not validated. Kool et al. (2013) suggested a variety of techniques for
collecting empirical connectivity data that can be used for model parameterization or validation. For
the purpose of estimating disease spread, telemetry studies focusing on regular or seasonal interpatch dispersal movements would be most useful. The results in this thesis highlight that movement
studies must consider seasonal and intra-specific variability and be stratified by resource-poor and abundant environments. Finally, the assumed link between patch connectivity and disease spread
must be specifically tested against spatially-explicit epidemiological data collected during future
incursions.
6.5.2 Improved habitat suitability and connectivity models
Empirical research could yield improved data for both model parameterization and evaluation.
This could ultimately also allow adoption of more sophisticated modelling and validation methods.
6.5.2.1 Statistical habitat suitability models
The modelling approach in this thesis was reliant on expert-elicited knowledge. I made every
effort to maintain a structured approach to elicitation, carefully documented methods used (in
Chapters 3 and 4) and results obtained to ensure reproducibility (Bayesian network CPTs in in
Appendix 4.1 and response-to-pattern curves in Appendix 3.1) and evaluated model uncertainty (in
Chapters 3, 4 and 5). Nevertheless, due to the inherent uncertainty and biases in expert-based
models, many ecologists remain sceptical of their value and prefer repeatable statistical analyses
(Johnson et al. 2012b). Many techniques are available to model species’ habitat and distributions
statistically (Section 2.4.1). These ‘species distribution models’, and their own pitfalls, have been
discussed at length elsewhere (Guisan & Thuiller 2005; Austin 2007; Elith & Leathwick 2009; Elith
et al. 2010 McGarigal et al. 2016; Yackulic & Ginsberg 2016). Given a suitable data set on the
response variable (e.g. feral pig presence / absence), and inclusion of ecologically meaningful
‘scale-optimized’ landscape variables (Section 2.4.3), I suggest that statistical multi-scale models
could also be applied to feral pigs in northern Australia and yield powerful inferences. However,
these may pertain to actual distributions (as evident in the data) rather than potential (resourcebased) habitat quality, and must be interpreted as such (Section 2.1.1).
I used the Continuous Boyce Index to evaluate model performance (Boyce et al. 2002; Hirzel et
al. 2006). This method was well suited to the presence-only validation data that I was able to source
for this study. However, it provides no information on model specificity, that is, its ability to
correctly predict absences and minimize false positives (Hirzel et al. 2006; Jiménez-Valverde et al.
2011). Presence-only validation also necessitates definition of a ‘background’. In the CBI method,
these serve to evaluate model deviation from a random prediction (e.g. if 20% of the background
138
contains highly suitable habitat, a useful model should find a higher proportion of actual presences
in predicted highly suitable habitat). However, background definition may influence validation
results. Generally, “too small an area can produce spurious models while […] too large of an area
can lead to artificially inflated test statistics” (VanDerWal et al. 2009, p. 592). A systematically
collected presence / absence data set could help to overcome these limitations and pitfalls of
presence-only validation techniques and facilitate a confusion matrix-based validation of model
sensitivity as well as specificity and is suggested as a future research priority.
6.5.2.2 Improved Bayesian network / spatial analysis integration
A further area of improvement relates to the integration of spatial pattern suitability analysis
within the probabilistic Bayesian network modelling framework proposed in Chapter 3. Firstly, a
major theoretical advantage of the Bayesian network approach to habitat modelling lies in the
explicit representation and propagation of uncertainty throughout a model (Wilhere 2012). Here,
resource quality at a given site was modelled in Bayesian networks and therefore expressed as a
conditional probability distribution. However, spatial pattern suitability analysis required
deterministic site-scale input values (resource quality indices, computed as model expected value)
and converted them into equally deterministic landscape-scale output values (resource suitability
indices). Model uncertainty was then re-introduced when habitat suitability was modelled in another
Bayesian network as a function of all landscape-scale indices. Future research should work towards
an uninterrupted workflow and propagation of uncertainty throughout an integrated model (Wilhere
2012).
6.5.2.3 Improved habitat connectivity models
This study used the circuit-theoretic approach to connectivity modelling, which is well suited
to identifying multiple, rather than optimal, dispersal paths throughout the entire landscape (McRae
& Beier 2007; McRae et al. 2008; Pelletier et al. 2014). Yet, the model output, ‘omnidirectional
current density’, is not easily interpreted based on an underlying ecological mechanism (Pelletier et
al. 2014). Other methods for modelling connectivity (for example graph-theoretic network graphs;
Section 2.5.1) may provide a more robust platform for further analyses. Most importantly, model
validation is critical (Kool et al. 2013; Fletcher et al. 2016a), but relies on empirical knowledge that
was unavailable or incomplete at the time of this study. Procedures for patch delineation in this
study must be better supported by data about home range movements (e.g. intra-patch gap crossing
abilities), especially in resource-poor environments that contained no modelled patches. Matrix
connectivity must be validated against observed inter-patch dispersal paths. However, a likely
mismatch between predictions (potential connectivity) and empirical data (actual dispersal) applies
to connectivity models as it does to habitat suitability models (Section 2.1.5).
139
6.5.3 Spatiotemporal dynamics
6.5.3.1 Host dynamics
The models developed in this thesis generated new spatially-explicit ‘snapshot’ knowledge
about seasonal habitat suitability for feral pig breeding and habitat connectivity between patches of
breeding habitat. In line with the resource-based habitat definition in Section 2.1.1, modelled
indices can be used to infer potential distributions and dispersal and associated risks of disease
establishment and spread. Yet, these static indices, although seasonally-specific, may not always
correspond well to actual occupancy, and less likely to local abundance or population viability
(Pulliam 2000; Larson et al. 2009; Stephens et al. 2015). For example, the models in Chapter 4
showed a four-fold increase in suitable habitat during the wet season. Yet, how much of this can
actually be realized by feral pig breeding herds dispersing from dry season refuges remains unclear.
Similarly, the connectivity models in Chapter 5 showed seasonal and gender-specific differences.
Yet, actual dispersal between habitat patches may depend on the occupancy of each patch at a given
time, the directionality between source and sink habitats, and the density-dependent population
pressure in source habitats (Pulliam 2000; Calabrese & Fagan 2004; Fletcher et al. 2016a).
Ecological simulation models that integrate density-dependent, intra-specific or metapopulation
dynamics are needed to provide deeper insights on local abundance, population growth, dispersal,
and metapopulation persistence (Hastings 2014; O’Reilly-Nugent et al. 2016). While many
techniques are available for different purposes (Section 2.4.1), dynamic models are necessarily
complex and require a large amount of empirical data for parameterization (Gallien et al. 2010,
Franklin 2010, Kool et al. 2013). Our results on habitat suitability for breeding and connectivity
could be used to constrain and inform models (Gallien et al. 2010; O’Reilly-Nugent et al. 2016), for
example by defining spatially-explicit patches and functional distances between them in
metapopulation models (Guisan & Thuiller 2005). Graph-theoretic dynamic network models, which
integrate local population growth, dispersal of individuals between patches, and allow for temporal
stochasticity in network connections, may be a useful tool to integrate our seasonal scenarios into
one dynamic framework (Ferrari et al. 2014). Lastly, recent models incorporating the effects of
management interventions on feral pig population densities in New Zealand (Krull et al. 2016) or
invasive rabbit metapopulation persistence in Australia (Lurgi et al. 2016) could also be used to
investigate how management can reduce the risk of disease establishment and risk in wildlife hosts.
6.5.3.2 Disease dynamics
Similarly, inferences about the establishment and spread risk of directly transmitted infectious
diseases were made from habitat measures describing potential for breeding, and connectivity
measures describing potential pathways of dispersal between habitat patches, respectively.
140
However, these seasonally static measures may not equate to the actual establishment, spread and
persistence of a given disease at a given point in time. For example, actual disease establishment
may depend on local ‘threshold’ densities and contact rates rather than habitat suitability, and actual
disease spread on the directionality, timing, frequency and number of dispersal movements as well
as many disease-specific parameters (e.g. basic reproductive rate, infectiousness, infectious period,
host mortality, or effects on host behaviour; Doran & Laffan 2005; Cowled & Garner 2008; Milne
et al. 2008). To effectively and rapidly respond to an incursion, spatially-explicit dynamic
epidemiological simulation models must be formulated based on the on-ground conditions at the
time, or ideally, for multiple scenarios of on-ground conditions in advance (Cowled & Garner
2008). Yet, such models are even more complex than ecological simulation models as they require
specification of dynamic parameters related to both host persistence and spread, and disease
transmission between infected and susceptible hosts (McCallum et al. 2001). The results from this
thesis could be used to improve the spatially-explicit ‘host’ component of such investigations
(Doran & Laffan 2005). For example, assumptions of random mixing could be more realistically
constrained by spatially clustered habitat patches and dispersal corridors (Cowled & Garner 2008).
Our results could also be used to generate improved estimates of host densities (albeit with the
considerable uncertainties discussed above), which in turn could ‘seed’ spatially-explicit cellular
automata or individual-based models (Doran & Laffan 2005; Ward et al. 2011; Rees et al. 2013).
6.5.4 Optimization models
This thesis generated improved regional-scale data about feral pig populations and suggested a
number of options how these could be used to inform risk-based disease management in northern
Australia (Section 6.4.2). However, these recommendations remained suggestive and unevaluated.
A large field of research has emerged on structured decision-making and deriving mathematically
‘optimal’ management recommendations, often in the context of systematic conservation planning
(Januchowski-Hartley et al. 2011; Glen et al. 2013). For invasive species, Baxter & Possingham
(2011) traded off search effort and improved spatial knowledge about species occurrence to derive
optimal management investments into either widespread uninformed or focused well-informed
surveillance. By investigating patterns in connectivity networks, Chades et al. (2011) showed “how
to prioritize management and survey effort across time and space for networks of susceptible–
infected–susceptible subpopulations” (p. 8323). The results in this thesis could be used to spatially
inform such management-focused optimization models. For example, individual patches that
contribute most to connectivity within and between habitat components could be prioritized using a
graph-theoretic approach (Galpern et al. 2011). Surveillance activities could then be optimized
across time and space within such patch networks according to mathematical efficiency rules
141
(Chades et al. 2011). Tools for structured decision-making could also combine the disease risk
indices developed in this thesis (establishment and spread) with other important criteria such as
likely impact of disease occurrence or budgetary constraints (Glen et al. 2013).
6.5.5 Other study systems and applications
Finally, despite the limitations and potential improvements discussed above, the multi-level
conceptual framework and expert-based modelling approaches presented in this thesis provide a
useful framework for investigating potential distributions (via resource-based habitat suitability)
and dispersal (via resource-based habitat connectivity) of mobile animals that could be readily
applied to other wide-ranging species, especially in situations where empirical data is limited.
Moreover, spatially-explicit, seasonally-specific and ecologically justified model predictions were
explicitly linked to an applied management problem, the risk-based management of directly
transmitted infectious animal diseases with feral pigs as a wildlife host. The integrated study of
organisms at multiple behavioural and organisational levels and scales in space and time is a current
frontier in ecology (McGarigal et al. 2016; Yackulic & Ginsberg 2016). Similar combinations of
habitat suitability and connectivity models to the one presented in this thesis have increasingly been
applied to aid conservation decisions (Cianfrani et al. 2013; Dickson et al. 2013) and, more rarely,
invasive species management (Stewart-Koster et al. 2015). Yet, Glen et al. (2013) suggested that
‘integrated landscape approaches’ based on metapopulation theory, connectivity models and spatial
optimization are currently underutilized to inform applied biosecurity problems.
The approaches developed here could be adapted, for example, to other widespread vertebrate
pest animals in Australia (e.g. deer, horses, wild dogs, foxes, cats), for which regional-scale
landscape ecological knowledge may be equally limited. To my knowledge, few connectivity
models have been applied to such species in Australia, although many beneficial insights for
invasion management may be gained. For example, results could be used to establish connected
management units or target local population control activities in patches while minimizing
reinvasion along dispersal corridors. The resource-based approaches could also be applied to other
host-pathogen systems. As disease risk was conceptualized as an attribute of the host, the
framework may be applied to directly transmitted diseases in a range of species. Yet, it could also
be extended to study more complex interactions. For example, by modelling habitat patches for
wildlife hosts (and vectors) and combining this information with spatial patterns of domestic animal
or human use, hotspots of disease emergence may be identified and efficient sampling campaigns
designed (Hartemink et al. 2015). It is the hope of the author that the research presented in this
thesis may inspire other scientists and practitioners to continue the journey towards a spatiallyinformed, risk-based management of infectious animal diseases.
142
References
Addicott, JF, Aho, JM, Antolin, MF, Padilla, DK, Richardson, JS et al. (1987). "Ecological
neighborhoods: scaling environmental patterns." Oikos 49(3): 340-346.
Alexander, NS, Massei, G & Wint, W (2016). "The European distribution of Sus scrofa. Model
outputs from the project described within the poster – where are all the boars? An attempt to
gain a continental perspective." Open Health Data 4: e1.
Allaire, JJ (2016). Package ‘rmarkdown’: dynamic documents for R. URL
http://rmarkdown.rstudio.com.
Altizer, S, Bartel, R & Han, BA (2011). "Animal migration and infectious disease risk." Science
331(6015): 296.
Anderson, RM, May, RM, Joysey, K, Mollison, D, Conway, GR et al. (1986). "The invasion,
persistence and spread of infectious diseases within animal and plant communities."
Philosophical Transactions of the Royal Society B: Biological Sciences 314(1167): 533-570.
Animal Health Australia (2011). Wild animal response strategy (Version 3.3). Australian
Veterinary Emergency Plan (AUSVETPLAN), Edition 3. Primary Industries Ministerial
Council, Canberra.
Animal Health Australia (2012). Disease strategy: classical swine fever (Version 4.0). Australian
Veterinary Emergency Plan (AUSVETPLAN), Edition 4. National Biosecurity Committee,
Canberra.
Arthur, SM, Manly, BFJ, McDonald, LL & Garner, GW (1996). "Assessing habitat selection when
availability changes." Ecology 77(1): 215-227.
Artois, M, Depner, K, Guberti, V, Hars, J, Rossi, S et al. (2002). "Classical swine fever (hog
cholera) in wild boar in Europe." Revue scientifique et technique (International Office of
Epizootics) 21(1): 287-304.
Atlas of Living Australia, ALA (2017). URL http://www.ala.org.au/.
Austin, M (2007). "Species distribution models and ecological theory: a critical assessment and
some possible new approaches." Ecological Modelling 200(1-2): 1-19.
Australian Bureau of Agricultural and Resource Economics and Sciences, ABARES (2011).
Guidelines for land use mapping in Australia: principles, procedures and definitions, fourth
edition. Commonwealth of Australia, Canberra.
143
Australian National Audit Office (2012). Administration of the Northern Australia Quarantine
Strategy: ANAO audit report no. 46 2011–12. Commonwealth of Australia (ANAO), Canberra.
Bahn, V & McGill, BJ (2007). "Can niche-based distribution models outperform spatial
interpolation?" Global Ecology and Biogeography 16(6): 733-742.
Barrios-Garcia, MN & Ballari, SA (2012). "Impact of wild boar (Sus scrofa) in its introduced and
native range: a review." Biological Invasions 14(11): 2283-2300.
Bates, D & Maechler, M (2015). Package 'Matrix': sparse and dense matrix classes and methods.
URL http://Matrix.R-forge.R-project.org/.
Baxter, PWJ & Possingham, HP (2011). "Optimizing search strategies for invasive pests: learn
before you leap." Journal of Applied Ecology 48(1): 86-95.
Beale, R, Fairbrother, J, Inglis, A & Trebeck, D (2008). One biosecurity: a working partnership.
Commonwealth of Australia, Canberra.
Beck, JL & Suring, LH (2009). "Chapter 10: Wildlife habitat-relationships models: description and
evaluation of existing frameworks." in JJ Millspaugh & FR Thompson III (eds). Models for
planning wildlife conservation in large landscapes. Academic Press, San Diego: pp. 251-285.
Beier, P, Majka, D & Jenness, J (2007). Conceptual steps for designing wildlife corridors.
Retrieved June 2014 from http://corridordesign.org/designing_corridors.
Beier, P, Majka, DR & Spencer, WD (2008). "Forks in the road: choices in procedures for
designing wildland linkages." Conservation Biology 22(4): 836-851.
Beier, P, Spencer, W, Baldwin, RF & McRae, BH (2011). "Toward best practices for developing
regional connectivity maps." Conservation Biology 25(5): 879-892.
Bellamy, C & Altringham, J (2015). "Predicting species distributions using record centre data:
multi-scale modelling of habitat suitability for bat roosts." PLoS ONE 10(6): e0128440.
Bengsen, AJ, Gentle, MN, Mitchell, JL, Pearson, HE & Saunders, GR (2014). "Impacts and
management of wild pigs Sus scrofa in Australia." Mammal Review 44(2): 135-147.
Berry, JK (2013). Beyond mapping III: a compilation of Beyond Mapping columns from 1996 to
2013. Basis Press, Fort Collins.
Beutel, TS, Beeton, RJS & Baxter, GS (1999). "Building better wildlife-habitat models." Ecography
22(2): 219-219.
144
Boets, P, Landuyt, D, Everaert, G, Broekx, S & Goethals, PLM (2015). "Evaluation and comparison
of data-driven and knowledge-supported Bayesian belief networks to assess the habitat
suitability for alien macroinvertebrates." Environmental Modelling & Software 74: 92-103.
Börger, L, Dalziel, BD & Fryxell, JM (2008). "Are there general mechanisms of animal home range
behaviour? A review and prospects for future research." Ecology Letters 11(6): 637-650.
Bosch, J, Mardones, F, Pérez, A & Muñoz, MJ (2014). "A maximum entropy model for predicting
wild boar distribution in Spain." Spanish Journal of Agricultural Research 12(4): 984-999.
Boyce, MS, Vernier, PR, Nielsen, SE & Schmiegelow, FKA (2002). "Evaluating resource selection
functions." Ecological Modelling 157(2-3): 281-300.
Boyce, MS, Johnson, CJ, Merrill, EH, Nielsen, SE, Solberg, EJ et al. (2016). "REVIEW: can habitat
selection predict abundance?" Journal of Animal Ecology 85(1): 11-20.
Broennimann, O (2015). Package 'ecospat': spatial ecology miscellaneous methods. URL
http://cran.r-project.org/web/packages/ecospat/.
Brookes, VJ, Hernández-Jover, M, Cowled, B, Holyoake, PK & Ward, MP (2014). "Building a
picture: prioritisation of exotic diseases for the pig industry in Australia using multi-criteria
decision analysis." Preventive Veterinary Medicine 113(1): 103-117.
Brooks, RP (1997). "Improving habitat suitability index models." Wildlife Society Bulletin 25(1):
163-167.
Bureau of Meteorology, BOM (2012). Australian hydrological geospatial fabric (geofabric) data
product specification: surface cartography Version 2.1. Bureau of Meteorology, Canberra.
Burt, WH (1943). "Territoriality and home range concepts as applied to mammals." Journal of
Mammalogy 24(3): 346-352.
Cain, J (2001). Planning improvement in natural resource management. Centre for Ecology and
Hydrology, Wallingford, Oxon.
Calabrese, JM & Fagan, WF (2004). "A comparison-shopper's guide to connectivity metrics."
Frontiers in Ecology and the Environment 2(10): 529-536.
Caley, P (1993). "Population dynamics of feral pigs (Sus Scrofa) in a tropical riverine habitat
complex." Wildlife Research 20(5): 625-636.
Caley, P (1997). "Movements, activity patterns and habitat use of feral pigs (Sus scrofa) in a
tropical habitat." Wildlife Research 24(1): 77-87.
145
Calisher, CH, Childs, JE, Field, HE, Holmes, KV & Schountz, T (2006). "Bats: important reservoir
hosts of emerging viruses." Clinical Microbiology Reviews 19(3): 531-545.
Chades, I, Martin, TG, Nicol, S, Burgman, MA, Possingham, HP et al. (2011). "General rules for
managing and surveying networks of pests, diseases, and endangered species." Proceedings of
the National Academy of Sciences 108(20): 8323–8328.
Chen, SH & Pollino, CA (2012). "Good practice in Bayesian network modelling." Environmental
Modelling & Software 37: 134-145.
Choquenot, D, McIlroy, J & Korn, T (1996). Managing vertebrate pests: feral pigs. Commonwealth
of Australia, Canberra.
Choquenot, D (1998). "Testing the relative influence of intrinsic and extrinsic variation in food
availability on feral pig populations in Australia's rangelands." Journal of Animal Ecology
67(6): 887-907.
Choquenot, D & Ruscoe, WA (2003). "Landscape complementation and food limitation of large
herbivores: habitat-related constraints on the foraging efficiency of wild pigs." Journal of
Animal Ecology 72(1): 14-26.
Cianfrani, C, Maiorano, L, Loy, A, Kranz, A, Lehmann, A et al. (2013). "There and back again?
Combining habitat suitability modelling and connectivity analyses to assess a potential return
of the otter to Switzerland." Animal Conservation 16(5): 584-594.
Cookson, B, Sergeant, ESG & Martin, PA (2012). "Risk-based prioritisation of surveillance for
exotic animal diseases in northern Australia." Proceedings of the 13th International
Symposium on Veterinary Epidemiology and Economics. Maastricht, Netherlands: p. 379.
Corbett, L (1995). "Does dingo predation or buffalo competition regulate feral pig populations in
the Australian wet-dry tropics? An experimental study." Wildlife Research 22(1): 65-74.
Cowled, BD, Lapidge, SJ, Hampton, JO & Spencer, PBS (2006). "Measuring the demographic and
genetic effects of pest control in a highly persecuted feral pig population." The Journal of
Wildlife Management 70(6): 1690-1697.
Cowled, BD & Giannini, F (2007). Spatial predictive modelling for feral pigs in northern Australia
(2.054R). Department of Agriculture, Fisheries and Forestry, Canberra.
Cowled, BD, Aldenhoven, J, Odeh, IOA, Garrett, T, Moran, C et al. (2008). "Feral pig population
structuring in the rangelands of eastern Australia: applications for designing adaptive
management units." Conservation Genetics 9(1): 211-224.
146
Cowled, B & Garner, G (2008). "A review of geospatial and ecological factors affecting disease
spread in wild pigs: considerations for models of foot-and-mouth disease spread." Preventive
Veterinary Medicine 87(3-4): 197-212.
Cowled, BD, Giannini, F, Beckett, SD, Woolnough, A, Barry, S et al. (2009). "Feral pigs:
predicting future distributions." Wildlife Research 36(3): 242-251.
Cowled, B, Garner, M, Negus, K & Ward, M (2012). "Controlling disease outbreaks in wildlife
using limited culling: modelling classical swine fever incursions in wild pigs in Australia."
Veterinary Research 43(3): 1-16.
Daszak, P, Cunningham, AA & Hyatt, AD (2000). "Emerging infectious diseases of wildlife:
threats to biodiversity and human health." Science 287(5452): 443.
DeCesare, NJ, Hebblewhite, M, Schmiegelow, F, Hervieux, D, McDermid, GJ et al. (2012).
"Transcending scale dependence in identifying habitat with resource selection functions."
Ecological Applications 22(4): 1068-1083.
Department of Agriculture and Fisheries, DAF (2015). Annual pest distribution survey 2013-14.
The State of Queensland.
Department of the Environment and Water Resources, DEWR (2007). Australia’s native
vegetation: a summary of Australia’s major vegetation groups, 2007. Commonwealth of
Australia, Canberra.
Department of the Environment (2012). Interim Biogeographic Regionalisation for Australia
(Subregions) v.7 (IBRA). URL https://www.environment.gov.au/land/nrs/science/ibra#ibra.
Department of the Environment (2015). Threat abatement plan for predation, habitat degradation,
competition and disease transmission by feral pigs (Sus scrofa). Commonwealth of Australia,
Canberra.
Department of Science, IT, Innovation and the Arts, DSITIA (2012). Land cover change in
Queensland 2009–10: a Statewide Landcover and Trees Study (SLATS) report. The State of
Queensland, Brisbane.
Dexter, N (1998). "The influence of pasture distribution and temperature on habitat selection by
feral pigs in a semi-arid environment." Wildlife Research 25(5): 547-559.
Dexter, N (1999). "The influence of pasture distribution, temperature and sex on home-range size of
feral pigs in a semi-arid environment." Wildlife Research 26(6): 755-762.
147
Dexter, N & McLeod, SR (2015). "Modeling ecological traps for the control of feral pigs." Ecology
and Evolution 5(10): 2036-2047.
Dickson, BG, Roemer, GW, McRae, BH & Rundall, JM (2013). "Models of regional habitat quality
and connectivity for pumas (Puma concolor) in the southwestern United States." PLoS ONE
8(12): e81898.
Dijak, WD, Rittenhouse, CD, Larson, MA, Thompson, FR & Millspaugh, JJ (2007). "Landscape
habitat suitability index software." The Journal of Wildlife Management 71(2): 668-670.
Dijak, WD & Rittenhouse, CD (2009). "Chapter 14: Development and application of habitat
suitability models to large landscapes." in JJ Millspaugh & FR Thompson III (eds). Models for
planning wildlife conservation in large landscapes. Academic Press, San Diego: pp. 367-389.
Doerr, VAJ, Barrett, T & Doerr, ED (2011). "Connectivity, dispersal behaviour and conservation
under climate change: a response to Hodgson et al." Journal of Applied Ecology 48(1): 143147.
Doran, RJ & Laffan, SW (2005). "Simulating the spatial dynamics of foot and mouth disease
outbreaks in feral pigs and livestock in Queensland, Australia, using a susceptible-infectedrecovered cellular automata model." Preventive Veterinary Medicine 70(1-2): 133-152.
Dormann, CF, Schymanski, SJ, Cabral, J, Chuine, I, Graham, CH et al. (2012). "Correlation and
process in species distribution models: bridging a dichotomy." Journal of Biogeography
39(12): 2119-2131.
Douglas, SJ & Newton, AC (2014). "Evaluation of Bayesian networks for modelling habitat
suitability and management of a protected area." Journal for Nature Conservation 22(3): 235246.
Dowle, M, Srinivasan, A, Short, T & Lianoglou, S (2015). Package 'data.table': extension of
data.frame. URL https://github.com/Rdatatable/data.table/wiki/.
Ducci, L, Agnelli, P, Di Febbraro, M, Frate, L, Russo, D et al. (2015). "Different bat guilds perceive
their habitat in different ways: a multiscale landscape approach for variable selection in species
distribution modelling." Landscape Ecology 30(10): 2147-2159.
Dunning, JB, Danielson, BJ & Pulliam, HR (1992). "Ecological processes that affect populations in
complex landscapes." Oikos 65(1): 169-175.
East, IJ, Wicks, RM, Martin, PAJ, Sergeant, ESG, Randall, LA et al. (2013). "Use of a multi-criteria
analysis framework to inform the design of risk based general surveillance systems for animal
disease in Australia." Preventive Veterinary Medicine 112(3-4): 230-247.
148
Elith, J & Leathwick, JR (2009). "Species distribution models: ecological explanation and
prediction across space and time." Annual Review of Ecology, Evolution and Systematics 40:
677-697.
Elith, J, Kearney, M & Phillips, SJ (2010). "The art of modelling range-shifting species." Methods
in Ecology and Evolution 1(4): 330-342.
Elledge, AE, McAlpine, CA, Murray, PJ & Gordon, IJ (2013). "Modelling habitat preferences of
feral pigs for rooting in lowland rainforest." Biological Invasions 15(7): 1523-1535.
Emergency Animal Disease Response Agreement (2012). Government and livestock industry cost
sharing deed in respect of emergency animal disease responses (Variation No. 12/01 25/05/2012), Canberra.
Environmental Protection Agency, EPA (2005). Wetland mapping and classification methodology.
Overall framework: a method to provide baseline mapping and classification for wetlands in
Queensland, version 1.2. The State of Queensland, Brisbane.
Etherington, TR & Holland, EP (2013). "Least-cost path length versus accumulated-cost as
connectivity measures." Landscape Ecology 28(7): 1223-1229.
Etherington, TR, Perry, GLW, Cowan, PE & Clout, MN (2014). "Quantifying the direct transfer
costs of common brushtail possum dispersal using least-cost modelling: a combined costsurface and accumulated-cost dispersal kernel approach." PLoS ONE 9(2): e88293.
Farina, A (2006). Principles and methods in landscape ecology. Chapman & Hall, London.
Fenton, NE, Neil, M & Caballero, JG (2007). "Using ranked nodes to model qualitative judgments
in Bayesian networks." IEEE Transactions on Knowledge and Data Engineering 19(10): 14201432.
Fenton, NE & Neil, M (2013). Risk assessment and decision analysis with Bayesian networks. CRC
Press, Boca Raton.
FeralPigScan (2016). URL http://www.feralscan.org.au/feralpigscan/default.aspx.
Ferrari, JR, Preisser, EL & Fitzpatrick, MC (2014). "Modeling the spread of invasive species using
dynamic network models." Biological Invasions 16(4): 949-960.
Fletcher, RJ, Burrell, NS, Reichert, BE, Vasudev, D & Austin, JD (2016a). "Divergent perspectives
on landscape connectivity reveal consistent effects from genes to communities." Current
Landscape Ecology Reports 1(2): 67-79.
149
Fletcher, RJ, McCleery, RA, Greene, DU & Tye, CA (2016b). "Integrated models that unite local
and regional data reveal larger-scale environmental relationships and improve predictions of
species distributions." Landscape Ecology 31(6): 1369-1382.
Forman, RT (1995). "Some general principles of landscape and regional ecology." Landscape
Ecology 10(3): 133-142.
Fox, ID, Neldner, VJ, Wilson, GW & Bannink, PJ (2001). The vegetation of the Australian tropical
savannas. Queensland Herbarium, Environmental Protection Agency, Brisbane.
Franklin, J (2010). "Moving beyond static species distribution models in support of conservation
biogeography." Diversity and Distributions 16(3): 321-330.
Fritzemeier, J, Teuffert, J, Greiser-Wilke, I, Staubach, C, Schluter, H et al. (2000). "Epidemiology
of classical swine fever in Germany in the 1990s." Veterinary Microbiology 77(1-2): 29-41.
Froese, JG, Smith, CS, McAlpine, CA, Durr, PA & van Klinken, RD (2015). "Moving window
analysis links landscape-scale resource utilization to habitat suitability models of feral pigs in
northern Australia." in T Weber, MJ McPhee & RS Anderssen (eds). MODSIM2015, 21st
International Congress on Modelling and Simulation. Modelling and Simulation Society of
Australia and New Zealand, Broadbeach: pp. 1352-1358.
Gabor, TM, Hellgren, EC, Van Den Bussche, RA & Silvy, NJ (1999). "Demography, sociospatial
behaviour and genetics of feral pigs (Sus scrofa) in a semi-arid environment." Journal of
Zoology 247(3): 311-322.
Gallant, JC, Dowling, TI, Read, AM, Wilson, N, Tickle, P et al. (2011). 1 second SRTM derived
products user guide, Version 1.0.4. Geoscience Australia, Canberra.
Gallien, L, Münkemüller, T, Albert, CH, Boulangeat, I & Thuiller, W (2010). "Predicting potential
distributions of invasive species: where to go from here?" Diversity and Distributions 16(3):
331-342.
Galpern, P, Manseau, M & Fall, A (2011). "Patch-based graphs of landscape connectivity: a guide
to construction, analysis and application for conservation." Biological Conservation 144(1):
44-55.
Giles, JR (1980). The ecology of the feral pig in western New South Wales. A thesis submitted for
the degree of Doctor of Philosophy, University of Sydney.
Gill, T, Heidenreich, S & Guerschman, JP (2014). MODIS monthly fractional cover: product
creation and distribution. Joint Remote Sensing Research Program Publication Series.
150
Girvetz, E & Greco, S (2007). "How to define a patch: a spatial model for hierarchically delineating
organism-specific habitat patches." Landscape Ecology 22(8): 1131-1142.
Glen, A, Pech, R & Byrom, A (2013). "Connectivity and invasive species management: towards an
integrated landscape approach." Biological Invasions 15(10): 2127-2138.
Gormley, AM, Forsyth, DM, Griffioen, P, Lindeman, M, Ramsey, DSL et al. (2011). "Using
presence-only and presence–absence data to estimate the current and potential distributions of
established invasive species." Journal of Applied Ecology 48(1): 25-34.
Greer, EB (1983). "Heat stress and reproduction in pigs: its role in seasonal infertility." RAAN
Conference Proceedings. pp. 216-230.
Grinnell, J (1917). "Field tests of theories concerning distributional control." The American
Naturalist 51(602): 115-128.
Guerschman, JP, Hill, MJ, Renzullo, LJ, Barrett, DJ, Marks, AS et al. (2009). "Estimating fractional
cover of photosynthetic vegetation, non-photosynthetic vegetation and bare soil in the
Australian tropical savanna region upscaling the EO-1 Hyperion and MODIS sensors." Remote
Sensing of Environment 113(5): 928-945.
Guisan, A & Thuiller, W (2005). "Predicting species distribution: offering more than simple habitat
models." Ecology Letters 8(9): 993-1009.
Guisan, A, Tingley, R, Baumgartner, JB, Naujokaitis-Lewis, I, Sutcliffe, PR et al. (2013).
"Predicting species distributions for conservation decisions." Ecology Letters 16(12): 14241435.
Hall, LS, Krausman, PR & Morrison, ML (1997). "The habitat concept and a plea for standard
terminology." Wildlife Society Bulletin 25(1): 173-182.
Hamilton, GS, Mather, PB & Wilson, JC (2006). "Habitat heterogeneity influences connectivity in a
spatially structured pest population." Journal of Applied Ecology 43(2): 219-226.
Hamilton, GS & Mather, PB (2009). "Differential impacts of habitat heterogeneity on male and
female connectivity in a spatially structured pest system." Austral Ecology 34(1): 41-49.
Hamilton, SH, Pollino, CA & Jakeman, AJ (2015). "Habitat suitability modelling of rare species
using Bayesian networks: model evaluation under limited data." Ecological Modelling 299: 6478.
151
Hampton, JO, Spencer, PBS, Alpers, DL, Twigg, LE, Woolnough, AP et al. (2004). "Molecular
techniques, wildlife management and the importance of genetic population structure and
dispersal: a case study with feral pigs." Journal of Applied Ecology 41(4): 735-743.
Hampton, JO, Spencer, PBS, Elliot, AD & Thompson, RCA (2006). "Prevalence of zoonotic
pathogens from feral pigs in major public drinking water catchments in Western Australia."
EcoHealth 3(2): 103-108.
Hanski, I (1998). "Metapopulation dynamics." Nature 396(6706): 41-49.
Hartemink, N, Vanwambeke, SO, Purse, BV, Gilbert, M & Van Dyck, H (2015). "Towards a
resource-based habitat approach for spatial modelling of vector-borne disease risks." Biological
Reviews 90(4): 1151-1162.
Hastings, A (2014). "Persistence and management of spatially distributed populations." Population
Ecology 56(1): 21-26.
Henderson, WR (2008). Research on wildlife disease preparedness in Australia. Invasive Animals
Cooperative Research Centre, Canberra.
Hengl, T, Sierdsema, H, Radović, A & Dilo, A (2009). "Spatial prediction of species’ distributions
from occurrence-only records: combining point pattern analysis, ENFA and regressionkriging." Ecological Modelling 220(24): 3499-3511.
Hijmans, RJ (2015). Package 'raster': geographic data analysis and modeling. URL http://cran.rproject.org/web/packages/raster/.
Hirzel, AH, 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.
Hirzel, AH & Le Lay, G (2008). "Habitat suitability modelling and niche theory." Journal of
Applied Ecology 45(5): 1372-1381.
Holland, EP, Aegerter, JN & Smith, GC (2007). "Spatial sensitivity of a generic population model,
using wild boar (Sus scrofa) as a test case." Ecological Modelling 205(1-2): 146-158.
Holland, JD, Bert, DG & Fahrig, L (2004). "Determining the spatial scale of species' response to
habitat." BioScience 54(3): 227-233.
Holland, JD & Yang, S (2016). "Multi-scale studies and the ecological neighborhood." Current
Landscape Ecology Reports 1(4): 135-145.
Hone, J (1990a). "Note on seasonal changes in population density of feral pigs in three tropical
habitats." Australian Wildlife Research 17(2): 131-134.
152
Hone, J (1990b). "How many feral pigs in Australia?" Australian Wildlife Research 17(6): 571-572.
Hone, J, Pech, R & Yip, P (1992). "Estimation of the dynamics and rate of transmission of classical
swine fever (hog cholera) in wild pigs." Epidemiology and Infection 108(2): 377-386.
Hone, J (2007). Wildlife damage control. CSIRO Publishing, Collingwood.
Hone, J (2012). Applied population and community ecology: the case of feral pigs in Australia.
John Wiley & Sons.
Huang, ZYX, van Langevelde, F, Prins, HHT & de Boer, WF (2015). "Dilution versus facilitation:
Impact of connectivity on disease risk in metapopulations." Journal of Theoretical Biology
376: 66-73.
Hunsaker, CT, Goodchild, MF, Friedl, MA & Case, TJ (2001). Spatial uncertainty in ecology:
implications for remote sensing and GIS applications. Springer, New York.
Hutchinson, GE (1957). "Concluding remarks." Cold Spring Harbor Symposia on Quantitative
Biology 22: 415-427.
Hutchinson, MF, McIntyre, S, Hobbs, RJ, Stein, JL, Garnett, S et al. (2005). "Integrating a global
agro-climatic classification with bioregional boundaries in Australia." Global Ecology and
Biogeography 14(3): 197-212.
Jackson, HB & Fahrig, L (2015). "Are ecologists conducting research at the optimal scale?" Global
Ecology and Biogeography 24(1): 52-63.
Jackson, ND & Fahrig, L (2014). "Landscape context affects genetic diversity at a much larger
spatial extent than population abundance." Ecology 95(4): 871-881.
Jackson, ND & Fahrig, L (2016). "Habitat amount, not habitat configuration, best predicts
population genetic structure in fragmented landscapes." Landscape Ecology 31(5): 951-968.
Januchowski-Hartley, SR, Visconti, P & Pressey, RL (2011). "A systematic approach for
prioritizing multiple management actions for invasive species." Biological Invasions 13(5):
1241-1253.
Jiménez-Valverde, A, Peterson, A, Soberón, J, Overton, J, Aragón, P et al. (2011). "Use of niche
models in invasive species risk assessments." Biological Invasions 13(12): 2785-2797.
Johansen, K, Gill, T, Trevithick, R, Armston, J, Scarth, P et al. (2012). "Validation of Landsat based
time-series of persistent green-vegetation fraction for Australia." Proceedings of the XXII
Congress of the International Photogrammetry and Remote Sensing Society 2012. Melbourne.
153
Johnson, DH (1980). "The comparison of usage and availability measurements for evaluating
resource preference." Ecology 61(1): 65-71.
Johnson, S, Low-Choy, S & Mengersen, K (2012a). "Integrating Bayesian networks and geographic
information systems: Good practice examples." Integrated Environmental Assessment and
Management 8(3): 473-479.
Johnson, CJ, Hurley, M, Rapaport, E & Pullinger, M (2012b). "Using expert knowledge effectively:
lessons from species distribution models for wildlife conservation and management." in AH
Perera, CA Drew & CJ Johnson (eds). Expert knowledge and its application in landscape
ecology. Springer, New York: pp. 153-171.
Jones, DA, Wang, W & Fawcett, R (2009). "High-quality spatial climate data-sets for Australia."
Australian Meteorological and Oceanographic Journal 58(4): 233-248.
Jones, KE, Patel, NG, Levy, MA, Storeygard, A, Balk, D et al. (2008). "Global trends in emerging
infectious diseases." Nature 451(7181): 990-993.
Kanagaraj, R, Wiegand, T, Kramer-Schadt, S & Goyal, SP (2013). "Using individual-based
movement models to assess inter-patch connectivity for large carnivores in fragmented
landscapes." Biological Conservation 167: 298-309.
Kearney, M (2006). "Habitat, environment and niche: what are we modelling?" Oikos 115(1): 186191.
Kearney, M & Porter, W (2009). "Mechanistic niche modelling: combining physiological and
spatial data to predict species' ranges." Ecology Letters 12(4): 334-350.
Kearney, M, Porter, WP, Williams, C, Ritchie, S & Hoffmann, AA (2009). "Integrating biophysical
models and evolutionary theory to predict climatic impacts on species' ranges: the dengue
mosquito Aedes aegypti in Australia." Functional Ecology 23(3): 528-538.
Kie, JG, Matthiopoulos, J, Fieberg, J, Powell, RA, Cagnacci, F et al. (2010). "The home-range
concept: are traditional estimators still relevant with modern telemetry technology?"
Philosophical Transactions of the Royal Society of London B: Biological Sciences 365(1550):
2221-2231.
Kindlmann, P & Burel, F (2008). "Connectivity measures: a review." Landscape Ecology 23(8):
879-890.
Koen, EL, Garroway, CJ, Wilson, PJ & Bowman, J (2010). "The effect of map boundary on
estimates of landscape resistance to animal movement." PLoS ONE 5(7): e11785.
154
Koen, EL, Bowman, J, Sadowski, C & Walpole, AA (2014). "Landscape connectivity for wildlife:
development and validation of multispecies linkage maps." Methods in Ecology and Evolution
5(7): 626-633.
Kool, JT, Moilanen, A & Treml, EA (2013). "Population connectivity: recent advances and new
perspectives." Landscape Ecology 28(2): 165-185.
Kramer-Schadt, S, Fernandez, N & Thulke, HH (2007). "Potential ecological and epidemiological
factors affecting the persistence of classical swine fever in wild boar Sus scrofa populations."
Mammal Review 37(1): 1-20.
Krull, CR, Stanley, MC, Burns, BR, Choquenot, D & Etherington, TR (2016). "Reducing wildlife
damage with cost-effective management programmes." PLoS ONE 11(1): e0146765.
Kuhnert, PM, Martin, TG & Griffiths, SP (2010). "A guide to eliciting and using expert knowledge
in Bayesian ecological models." Ecology Letters 13(7): 900-914.
Kupfer, JA (2012). "Landscape ecology and biogeography: rethinking landscape metrics in a postFRAGSTATS landscape." Progress in Physical Geography 36(3): 400-420.
Laforge, MP, Brook, RK, Beest, FM, Bayne, EM & McLoughlin, PD (2016). "Grain-dependent
functional responses in habitat selection." Landscape Ecology 31(4): 855-863.
Landuyt, D, Broekx, S, D'Hondt, R, Engelen, G, Aertsens, J et al. (2013). "A review of Bayesian
belief networks in ecosystem service modelling." Environmental Modelling & Software 46: 111.
Landuyt, D, Van der Biest, K, Broekx, S, Staes, J, Meire, P et al. (2015). "A GIS plug-in for
Bayesian belief networks: towards a transparent software framework to assess and visualise
uncertainties in ecosystem service mapping." Environmental Modelling & Software 71: 30-38.
Larson, MA, Millspaugh, JJ & Thompson III, FR (2009). "Chapter 9: A review of methods for
quantifying wildlife habitat in large landscapes." in JJ Millspaugh & FR Thompson III (eds).
Models for planning wildlife conservation in large landscapes. Academic Press, San Diego:
pp. 225-250.
Leibold, MA (1995). "The niche concept revisited: mechanistic models and community context."
Ecology 76(5): 1371-1382.
Leonard, PB, Duffy, EB, Baldwin, RF, McRae, BH, Shah, VB et al. (2017). "gflow: software for
modelling circuit theory-based connectivity at any scale." Methods in Ecology and Evolution
8(4): 519-526.
155
Leslie, E, Cowled, B, Graeme Garner, M, Toribio, J-ALML & Ward, MP (2014). "Effective
surveillance strategies following a potential classical swine fever incursion in a remote wild pig
population in north-western Australia." Transboundary and Emerging Diseases 61(5): 432442.
Levin, SA (1992). "The problem of pattern and scale in ecology: the Robert H. MacArthur Award
Lecture." Ecology 73(6): 1943-1967.
Lookingbill, TR, Gardner, RH, Ferrari, JR & Keller, CE (2010). "Combining a dispersal model with
network theory to assess habitat connectivity." Ecological Applications 20(2): 427-441.
Lopez, J, Hurwood, D, Dryden, B & Fuller, S (2014). "Feral pig populations are structured at fine
spatial scales in tropical Queensland, Australia." PLoS ONE 9(3): e91657.
Lorena, AC, Jacintho, LFO, Siqueira, MF, Giovanni, RD, Lohmann, LG et al. (2011). "Comparing
machine learning classifiers in potential distribution modelling." Expert Systems with
Applications 38(5): 5268-5275.
Lurgi, M, Wells, K, Kennedy, M, Campbell, S & Fordham, DA (2016). "A landscape approach to
invasive species management." PLoS ONE 11(7): e0160417.
MacArthur, RH & Wilson, EO (1967). The theory of island biogeography. Princeton University
Press.
Macpherson, MF, Davidson, RS, Duncan, DB, Lurz, PW, Jarrott, A et al. (2016). "Incorporating
habitat distribution in wildlife disease models: conservation implications for the threat of
squirrelpox on the Isle of Arran." Animal Conservation 19(1): 3-14.
Manly, B, McDonald, L, Thomas, D, McDonald, TL & Erickson, WP (2002). Resource selection by
animals: statistical design and analysis for field studies. Springer Netherlands.
Marcot, BG, Holthausen, RS, Raphael, MG, Rowland, MM & Wisdom, MJ (2001). "Using
Bayesian belief networks to evaluate fish and wildlife population viability under land
management alternatives from an environmental impact statement." Forest Ecology and
Management 153(1-3): 29-42.
Marcot, BG (2006). "Characterizing species at risk I: modeling rare species under the Northwest
Forest Plan." Ecology and Society 11(2): 10.
Marcot, BG (2012). "Metrics for evaluating performance and uncertainty of Bayesian network
models." Ecological Modelling 230: 50-62.
156
Margosian, ML, Garrett, KA, Hutchinson, JMS & With, KA (2009). "Connectivity of the American
agricultural landscape: assessing the national risk of crop pest and disease spread." BioScience
59(2): 141-151.
Martin, AE & Fahrig, L (2012). "Measuring and selecting scales of effect for landscape predictors
in species–habitat models." Ecological Applications 22(8): 2277-2292.
Martin, TG, Burgman, MA, Fidler, F, Kuhnert, PM, Low-Choy, S et al. (2012). "Eliciting expert
knowledge in conservation science." Conservation Biology 26(1): 29-38.
Mayor, SJ, Schneider, DC, Schaefer, JA & Mahoney, SP (2009). "Habitat selection at multiple
scales." Écoscience 16(2): 238-247.
McBride, MF, Garnett, ST, Szabo, JK, Burbidge, AH, Butchart, SHM et al. (2012a). "Structured
elicitation of expert judgments for threatened species assessment: a case study on a continental
scale using email." Methods in Ecology and Evolution 3(5): 906-920.
McBride, MF, Fidler, F & Burgman, MA (2012b). "Evaluating the accuracy and calibration of
expert predictions under uncertainty: predicting the outcomes of ecological research." Diversity
and Distributions 18(8): 782-794.
McBride, M (2013). Delivering expert knowledge for conservation: tools for enhancing the quality
of expert judgement. A thesis submitted for the degree of Doctor of Philosophy, The University
of Melbourne.
McCallum, H, Barlow, N & Hone, J (2001). "How should pathogen transmission be modelled?"
Trends in Ecology & Evolution 16(6): 295-300.
McCann, RK, Marcot, BG & Ellis, R (2006). "Bayesian belief networks: applications in ecology
and natural resource management." Canadian Journal of Forest Research 36(12): 3053-3062.
McClure, ML, Burdett, CL, Farnsworth, ML, Lutman, MW, Theobald, DM et al. (2015). "Modeling
and mapping the probability of occurrence of invasive wild pigs across the contiguous United
States." PLoS ONE 10(8): e0133771.
McClure, ML, Hansen, AJ & Inman, RM (2016). "Connecting models to movements: testing
connectivity model predictions against empirical migration and dispersal data." Landscape
Ecology 31(7): 1419-1432.
McGarigal, K & Marks, BJ (1995). FRAGSTATS: spatial pattern analysis program for quantifying
landscape structure. US Department of Agriculture, Forest Service, Pacific Northwest
Research Station, Oregon.
157
McGarigal, K, Cushman, SE & Ene, E (2012). FRAGSTATS v4: spatial pattern analysis program
for categorical and continuous maps. University of Massachusetts, Amherst.
McGarigal, K, Wan, HY, Zeller, KA, Timm, BC & Cushman, SA (2016). "Multi-scale habitat
selection modeling: a review and outlook." Landscape Ecology 31(6): 1161-1175.
McNay, RS, Marcot, BG, Brumovsky, V & Ellis, R (2006). "A Bayesian approach to evaluating
habitat for woodland caribou in north-central British Columbia." Canadian Journal of Forest
Research 36(12): 3117-3133.
McRae, BH (2006). "Isolation by resistance." Evolution 60(8): 1551-1561.
McRae, BH & Beier, P (2007). "Circuit theory predicts gene flow in plant and animal populations."
Proceedings of the National Academy of Sciences 104(50): 19885-19890.
McRae, BH, Dickson, BG, Keitt, TH & Shah, VB (2008). "Using circuit theory to model
connectivity in ecology, evolution, and conservation." Ecology 89(10): 2712-2724.
McRae, BH, Shah, VB & Mohapatra, TK (2013). Circuitscape 4 user guide. The Nature
Conservancy.
Miguet, P, Jackson, HB, Jackson, ND, Martin, AE & Fahrig, L (2016). "What determines the spatial
extent of landscape effects on species?" Landscape Ecology 31(6): 1177-1194.
Millspaugh, JJ & Thompson III, FR (eds) (2009). Models for planning wildlife conservation in
large landscapes. Academic Press, San Diego.
Milne, G, Fermanis, C & Johnston, P (2008). "A mobility model for classical swine fever in feral
pig populations." Veterinary Research 39(6): 53.
Mitchell, J, Allen, L & Merrell, P (1982). Vertebrate pests of Queensland. Stock Routes and Rural
Lands Protection Board, Queensland Department of Lands, Brisbane.
Mitchell, J (2002). Ecology and management of feral pigs (Sus scrofa) in rainforests A thesis
submitted for the degree of Doctor of Philosophy, James Cook University.
Mitchell, J (2008). Feral pig control: a practical guide to pig control in Queensland. The State of
Queensland (DPI&F), Charters Towers.
Mitchell, J, Dorney, W, Mayer, R & McIlroy, J (2009). "Migration of feral pigs (Sus scrofa) in
rainforests of north Queensland: fact or fiction?" Wildlife Research 36(2): 110-116.
Mitchell, J (2010). Experimental research to quantify the environmental impact of feral pigs within
tropical freshwater ecosystems. Final Report to the Department of the Environment, Water,
Heritage and the Arts, Canberra.
158
Mitchell, MS, Zimmerman, JW & Powell, RA (2002). "Test of a habitat suitability index for black
bears in the southern Appalachians." Wildlife Society Bulletin 30(3): 794-808.
Mitchell, MS & Powell, RA (2004). "A mechanistic home range model for optimal use of spatially
distributed resources." Ecological Modelling 177(1-2): 209-232.
Mitchell, MS & Powell, RA (2012). "Foraging optimally for home ranges." Journal of Mammalogy
93(4): 917-928.
Moilanen, A & Hanski, I (2001). "On the use of connectivity measures in spatial ecology." Oikos
95(1): 147-151.
Moilanen, A & Nieminen, M (2002). "Simple connectivity measures in spatial ecology." Ecology
83(4): 1131-1145.
Moilanen, A (2011). "On the limitations of graph-theoretic connectivity in spatial ecology and
conservation." Journal of Applied Ecology 48(6): 1543-1547.
Moorcroft, PR (2012). "Mechanistic approaches to understanding and predicting mammalian space
use: recent advances, future directions." Journal of Mammalogy 93(4): 903-916.
Morelle, K, Lehaire, F & Lejeune, P (2014). "Is wild boar heading towards movement ecology? A
review of trends and gaps." Wildlife Biology 20(4): 196-205.
Morelle, K & Lejeune, P (2015). "Seasonal variations of wild boar Sus scrofa distribution in
agricultural landscapes: a species distribution modelling approach." European Journal of
Wildlife Research 61(1): 45-56.
Morrison, ML, Marcot, BG & Mannan, RW (2006). Wildlife-habitat relationships: concepts and
applications. Island Press, Washington.
Murray, JV, Goldizen, AW, O’Leary, RA, McAlpine, CA, Possingham, HP et al. (2009). "How
useful is expert opinion for predicting the distribution of a species within and beyond the
region of expertise? A case study using brush-tailed rock-wallabies Petrogale penicillata."
Journal of Applied Ecology 46(4): 842-851.
Murray, JV, Stokes, KE & van Klinken, RD (2012). "Predicting the potential distribution of a
riparian invasive plant: the effects of changing climate, flood regimes and land-use patterns."
Global Change Biology 18(5): 1738–1753.
Murray, JV, Berman, DM & van Klinken, RD (2014). "Predictive modelling to aid the regionalscale management of a vertebrate pest." Biological Invasions 16(11): 2403-2425.
159
Murray, J, Froese, J, Perry, J, Navarro Garcia, J & van Klinken, R (2015). Impact modelling for
rabbits and feral pigs in the QMDB. CSIRO Biosecurity, Brisbane.
National Audit Office (2002). The 2001 outbreak of foot and mouth disease. Stationary Office,
London.
National Committee on Soil and Terrain, NCST (2009). Australian soil and land survey field
handbook. CSIRO Publishing, Collingwood.
O’Reilly-Nugent, A, Palit, R, Lopez-Aldana, A, Medina-Romero, M, Wandrag, E et al. (2016).
"Landscape effects on the spread of invasive species." Current Landscape Ecology Reports
1(3): 107-114.
Olsen, B, Munster, VJ, Wallensten, A, Waldenström, J, Osterhaus, ADME et al. (2006). "Global
patterns of Influenza A virus in wild birds." Science 312(5772): 384.
Ostfeld, RS, Glass, GE & Keesing, F (2005). "Spatial epidemiology: an emerging (or re-emerging)
discipline." Trends in Ecology & Evolution 20(6): 328-336.
Paton, RS & Matthiopoulos, J (2016). "Defining the scale of habitat availability for models of
habitat selection." Ecology 97(5): 1113-1122.
Pearson, HE (2012). Understanding and mitigating the risk of pathogen transmission from wild
animals to domestic pigs in Australia. A thesis submitted for the degree of Doctor of
Philosophy, The University of Sydney.
Pearson, RG (2007). Species’ distribution modeling for conservation educators and practitioners:
synthesis. American Museum of Natural History.
Pech, RP & Hone, J (1988). "A model of the dynamics and control of an outbreak of foot and
mouth disease in feral pigs in Australia." Journal of Applied Ecology 25(1): 63-77.
Pelletier, D, Clark, M, Anderson, MG, Rayfield, B, Wulder, MA et al. (2014). "Applying circuit
theory for corridor expansion and management at regional scales: tiling, pinch points, and
omnidirectional connectivity." PLoS ONE 9(1): e84135.
Perera, AH, Drew, CA & Johnson, CJ (eds) (2012). Expert knowledge and its application in
landscape ecology. Springer, New York.
Phillips, BL, Chipperfield, JD & Kearney, MR (2008). "The toad ahead: challenges of modelling
the range and spread of an invasive species." Wildlife Research 35(3): 222-234.
Phillips, SJ, Anderson, RP & Schapire, RE (2006). "Maximum entropy modeling of species
geographic distributions." Ecological Modelling 190(3–4): 231-259.
160
Phillips, SJ, Dudík, M, Elith, J, Graham, CH, Lehmann, A et al. (2009). "Sample selection bias and
presence-only distribution models: implications for background and pseudo-absence data."
Ecological Applications 19(1): 181-197.
Plowright, RK, Foley, P, Field, HE, Dobson, AP, Foley, JE et al. (2011). "Urban habituation,
ecological connectivity and epidemic dampening: the emergence of Hendra virus from flying
foxes." Proceedings of the Royal Society B: Biological Sciences 278(1725): 3703.
Powell, RA & Mitchell, MS (2012). "What is a home range?" Journal of Mammalogy 93(4): 948958.
Pulliam, HR & Danielson, BJ (1991). "Sources, sinks, and habitat selection: a landscape perspective
on population dynamics." The American Naturalist 137(Supplement): S50-S66.
Pulliam, HR (2000). "On the relationship between niche and distribution." Ecology Letters 3(4):
349-361.
Raupach, M, Briggs, P, Haverd, V, King, E, Paget, M et al. (2009). Australian Water Availability
Project (AWAP), CSIRO Marine and Atmospheric Research component: final report for phase
3. Technical Report No. 013. Centre for Australian Weather and Climate Research (Bureau of
Meteorology and CSIRO), Melbourne.
Rayfield, B, Fortin, M-J & Fall, A (2011). "Connectivity for conservation: a framework to classify
network measures." Ecology 92(4): 847-858.
Rayfield, B, Pelletier, D, Dumitru, M, Cardille, JA & Gonzalez, A (2016). "Multipurpose habitat
networks for short-range and long-range connectivity: a new method combining graph and
circuit connectivity." Methods in Ecology and Evolution 7(2): 222-231.
RCoreTeam (2015). R: a language and environment for statistical computing. R Foundation for
Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.
Rees, EE, Pond, BA, Tinline, RR & Bélanger, D (2013). "Modelling the effect of landscape
heterogeneity on the efficacy of vaccination for wildlife infectious disease control." Journal of
Applied Ecology 50(4): 881-891.
Reisen, WK (2010). "Landscape epidemiology of vector-borne diseases." Annual Review of
Entomology 55(1): 461-483.
Rhodes, JR, McAlpine, CA, Lunney, D & Possingham, HP (2005). "A spatially explicit habitat
selection model incorporating home range behaviour." Ecology 86(5): 1199-1205.
161
Riley, S (2007). "Large-scale spatial-transmission models of infectious disease." Science
316(5829): 1298.
Rioux Paquette, S, Talbot, B, Garant, D, Mainguy, J & Pelletier, F (2014). "Modelling the dispersal
of the two main hosts of the raccoon rabies variant in heterogeneous environments with
landscape genetics." Evolutionary Applications 7(7): 734-749.
Ross, B (2009). Diet selectivity and feeding ecology of feral pigs (Sus scrofa) in Lakefield National
Park, Cape York Peninsula. A thesis submitted for the degree of BSc with Honours, James
Cook University.
Scott, A, Zepeda, C, Garber, L, Smith, J, Swayne, D et al. (2006). "The concept of
compartmentalisation." Revue scientifique et technique (International Office of Epizootics)
25(3): 873-907.
Shirk, AJ & McRae, BH (2013). Gnarly Landscape Utilities: Core Mapper user guide. The Nature
Conservancy, Fort Collins.
Smith, CS, Howes, AL, Price, B & McAlpine, CA (2007). "Using a Bayesian belief network to
predict suitable habitat of an endangered mammal: the Julia Creek dunnart (Sminthopsis
douglasi)." Biological Conservation 139(3-4): 333-347.
Smith, CS, van Klinken, RD, Seabrook, L & McAlpine, C (2012). "Estimating the influence of land
management change on weed invasion potential using expert knowledge." Diversity and
Distributions 18(8): 818-831.
Speirs-Bridge, A, Fidler, F, McBride, M, Flander, L, Cumming, G et al. (2010). "Reducing
overconfidence in the interval judgments of experts." Risk Analysis 30(3): 512-523.
Spencer, PBS, Lapidge, SJ, Hampton, JO & Pluske, JR (2005). "The sociogenetic structure of a
controlled feral pig population." Wildlife Research 32(4): 297-304.
Stankus, T (2012). "Razorbacks: feral pigs as agricultural pests, disruptors of ecosystems, reservoirs
of contagion, and favored game for sport and subsistence hunters: a review of the literature,
2005-2011." Journal of Agricultural & Food Information 13(3): 283-301.
Stephens, PA, Pettorelli, N, Barlow, J, Whittingham, MJ & Cadotte, MW (2015). "Management by
proxy? The use of indices in applied ecology." Journal of Applied Ecology 52(1): 1-6.
Stewart-Koster, B, Olden, JD & Johnson, PTJ (2015). "Integrating landscape connectivity and
habitat suitability to guide offensive and defensive invasive species management." Journal of
Applied Ecology 52(2): 366-378.
162
Tantipisanuh, N, Gale, GA & Pollino, C (2014). "Bayesian networks for habitat suitability
modeling: a potential tool for conservation planning with scarce resources." Ecological
Applications 24(7): 1705-1718.
Tischendorf, L & Fahrig, L (2000). "On the usage and measurement of landscape connectivity."
Oikos 90(1): 7-19.
Turner, MG & Garder, RH (2015). Landscape ecology in theory and practice: pattern and process.
Springer, New York.
Urban, DL, Minor, ES, Treml, EA & Schick, RS (2009). "Graph models of habitat mosaics."
Ecology Letters 12(3): 260-273.
Uuemaa, E, Antrop, M, Roosaare, J, Marja, R & Mander, Ü (2009). "Landscape metrics and
indices: an overview of their use in landscape research." Living Reviews in Landscape
Research 3(1): 1-28.
Uusitalo, L (2007). "Advantages and challenges of Bayesian networks in environmental modelling."
Ecological Modelling 203(3-4): 312-318.
van Klinken, RD, Murray, JV & Smith, C (2015). "Process-based pest risk mapping using Bayesian
networks." in RC Venette (ed). Pest risk modelling and mapping for invasive alien species.
CABI International: pp. 171-188.
VanDerWal, J, Shoo, LP, Graham, C & Williams, SE (2009). "Selecting pseudo-absence data for
presence-only distribution modeling: how far should you stray from what you know?"
Ecological Modelling 220(4): 589-594.
Venette, RC, Kriticos, DJ, Magarey, RD, Koch, FH, Baker, RHA et al. (2010). "Pest risk maps for
invasive alien species: a roadmap for improvement." BioScience 60(5): 349-362.
von Humboldt, A & Bonpland, A (1807). Ideen zu einer Geographie der Pflanzen : nebst einem
Naturgemälde der Tropenländer : auf Beobachtungen und Messungen gegründet, welche vom
10ten Grade nördlicher bis zum 10ten Grade südlicher Breite, in den Jahren 1799, 1800, 1801,
1802 und 1803 angestellt worden sind. Bey F.G. Cotta/Bey F. Schoell, Tübingen/Paris.
Ward, MP, Laffan, SW & Highfield, LD (2011). "Disease spread models in wild and feral animal
populations: application of artificial life models." Revue scientifique et technique (International
Office of Epizootics) 30(2): 437-446.
Weber, MM, Stevens, RD, Diniz-Filho, JAF & Grelle, CEV (in press). "Is there a correlation
between abundance and environmental suitability derived from ecological niche modelling? A
meta-analysis." Ecography.
163
West, P (2008). Assessing invasive animals in Australia 2008. Invasive Animals Cooperative
Research Centre, Canberra.
Wiens, JA (1989). "Spatial scaling in ecology." Functional Ecology 3(4): 385-397.
Wilhere, GF (2012). "Using Bayesian networks to incorporate uncertainty in habitat suitability
index models." The Journal of Wildlife Management 76(6): 1298-1309.
Wilson, G, Dexter, N, O'Brien, P & Bomford, M (1992). Pest animals in Australia: a survey of
introduced wild mammals. Kangaroo Press, Kenthurst.
Wilson, MFJ, O’Connell, B, Brown, C, Guinan, JC & Grehan, AJ (2007). "Multiscale terrain
analysis of multibeam bathymetry data for habitat mapping on the continental slope." Marine
Geodesy 30(1-2): 3-35.
World Organisation for Animal Health, OIE (2016). Terrestrial animal health code, 25th edition.
URL http://www.oie.int/international-standard-setting/terrestrial-code/access-online/.
Wu, J (2013). "Landscape ecology." in R Leemans (ed). Ecological systems. Springer New York:
pp. 179-200.
Wurster, CM, Robertson, J, Westcott, DA, Dryden, B, Zazzo, A et al. (2012). "Utilization of
sugarcane habitat by feral pig (Sus scrofa) in northern tropical Queensland: evidence from the
stable isotope composition of hair." PLoS ONE 7(9): e43538.
Xie, Y (2016). Package 'knitr': a general-purpose package for dynamic report generation in R.
URL http://yihui.name/knitr/.
Yackulic, CB & Ginsberg, JR (2016). "The scaling of geographic ranges: implications for species
distribution models." Landscape Ecology 31(6): 1195-1208.
Zanardi, G, Macchi, C, Sacchi, C & Rutili, D (2003). "Classical swine fever in wild boar in the
Lombardy region of Italy from 1997 to 2002." Veterinary Record 152(15): 461-465.
Zeller, KA, McGarigal, K & Whiteley, AR (2012). "Estimating landscape resistance to movement:
a review." Landscape Ecology 27(6): 777-797.
164
Appendices
Appendix 1 Human ethics approval
Letter 1: The University of Queensland, School of Agriculture and Food Sciences Ethics Committee
9 October 2013
Dear Jens
Re: Ethical Research Application
On behalf of the SAFS Ethics Committee this letter is to formally advise that your application for ethical
research involved with your PhD research project titled ‘Regional-scale spatial modeling of feral pig
distribution and abundance in northern Australia’ has been approved.
We wish you every success with your research.
Yours sincerely
Kaelene Matts | Academic Administration Officer
School of Agriculture & Food Sciences | Faculty of Science
The University of Queensland | Gatton Queensland 4343 | Australia
T: +61 7 5460 1471 F: +61 7 5460 1324
E: safs_admin@uq.edu.au W: http://www.uq.edu.au/agriculture/
165
Letter 2: CSIRO Social Science Human Research Ethics Committee
27 August 2013
Subject: Ethics Clearance 075/13 Assessing Regional-scale spatial modelling of feral pig distribution and
abundance in northern Australia
Dear Jens
We have now undertaken an initial ethics review and risk assessment of your project proposal entitled
“Assessing Regional-scale spatial modelling of feral pig distribution and abundance in northern Australia"
(075/13).
Please be advised that based on the information you have provided your project has been assessed against the
requirements of the National Statement as posing a low risk to participants. As such ethical clearance has
been granted for you to undertake the research and for the project to be managed within CSIRO's existing
project management processes. Ethical clearance has been granted for the period 01/09/13 – 31/01/16 subject
to the following condition:
A copy of the participant information / consent materials that will be used for the project should be
provided for Executive review and clearance prior to use.
In granting this clearance we would like to remind you of the need to adhere to the requirements of the
National Statement at all times during the life of the project. Should any adverse events occur to participants
during or resulting from the research or any ethically relevant variations be needed regarding the project's
implementation or completion you are required to notify us immediately for further advice or amended
clearance. You will also be required to complete a brief report about the project upon its completion. A
template for this report will be forwarded to you in due course.
We wish you all the best with the research.
Sincerely
Cathy Pitkin
Manager, Social Responsibility and Ethics
CSIRO
EcoSciences Precinct, 41 Boggo Rd, Dutton Park, QLD 4102
GPO Box 2583, Brisbane QLD 4001
Ph: +61-7-3833 5693 M: +61-409-441-055 Fax: +61-7-3833 5504 Web: http://www.ces.csiro.au/
166
Appendix 3.1 Model variables – parameters
(Appendix Table 3.1.1 included as Table 4.1 in this thesis.)
(Appendix Table 3.1.2 included as Table 4.2 in this thesis.)
(Appendix Table 3.1.3 included as Appendix Table 4.1 in this thesis.)
Appendix Table 3.1.4 Expert-elicited response-to-pattern curves. For each habitat variable – water (A),
food (B), protection from heat stress (C) and protection from disturbance (D) – tables show resource
suitability indices for distance-dependent measurement (SIDr) and composition-dependent measurement
(SICr) as elicited from each individual expert.
167
(A)
Water suitability index (SID water) per distance band
Expert
Very close
Close
Medium
Far
Very far
Expert 1
100
50
30
30
10
Expert 2
100
100
100
50
10
Expert 3
100
100
100
70
50
Expert 4
100
70
50
30
10
Expert 5
100
100
100
70
50
Expert 6
100
100
100
70
50
Expert
(B)
Water suitability index (SIC water) per average quality class (to % of max)
Very high
High
Moderate
(-100)
(-80)
(-60)
(-40)
(-30)
(-25)
(-20)
(-15)
(-10)
(-5)
Expert 1
90
90
90
90
90
70
50
30
10
10
Expert 2
90
90
90
90
90
90
70
70
50
10
Expert 3
90
90
90
90
90
90
70
70
50
10
Expert 4
90
90
90
90
90
90
70
50
30
10
Expert 5
90
90
90
90
90
70
50
30
30
10
Expert 6
90
90
90
90
90
90
70
50
30
10
Expert
Low
Very low
Food suitability index (SID food) per distance band
Very close
Close
Medium
Far
Very far
Expert 1
100
70
50
30
10
Expert 2
100
70
50
30
10
Expert 3
100
70
50
30
10
Expert 4
100
70
50
30
10
Expert 5
100
100
100
70
50
Expert 6
100
70
50
30
10
Expert
Food suitability index (SIC food) per average quality class (to % of max)
Very high
High
Moderate
Low
(-100)
(-80)
(-60)
(-40)
(-30)
(-25)
(-20)
(-15)
(-10)
(-5)
Expert 1
90
90
90
70
50
50
30
30
30
10
Expert 2
90
90
90
70
50
50
30
30
10
10
Expert 3
90
90
70
50
50
50
30
30
30
10
Expert 4
90
90
90
70
70
70
50
50
30
10
Expert 5
90
90
70
70
70
70
50
30
10
10
Expert 6
90
70
50
30
30
30
10
10
10
10
168
Very low
(C)
(D)
Heat protection suitability index (SID heat) per distance band
Expert
Very close
Close
Medium
Far
Very far
Expert 1
100
30
10
10
10
Expert 2
100
50
10
10
10
Expert 3
100
100
100
70
30
Expert 4
100
70
70
70
30
Expert 5
100
100
100
100
70
Expert 6
Not elicited
Expert
Heat protection suitability index (SIC heat) per average quality class (to % of max)
Very high
High
Moderate
Low
(-100)
(-80)
(-60)
(-40)
(-30)
(-25)
(-20)
(-15)
(-10)
(-5)
Expert 1
90
90
70
50
50
50
30
30
10
10
Expert 2
90
90
90
70
70
70
50
30
30
10
Expert 3
90
90
90
90
90
90
70
70
50
10
Expert 4
90
90
90
90
90
90
70
50
30
10
Expert 5
90
90
90
90
90
90
70
70
50
10
Very low
Expert 6
Not elicited
Expert
Disturbance protection suitability index (SID disturbance) per distance band
Very close
Close
Medium
Far
Very far
Expert 1
100
30
10
10
10
Expert 2
100
70
50
30
10
Expert 3
Not elicited
Expert 4
100
30
10
10
10
Expert 5
100
50
10
10
10
Expert 6
100
50
30
10
10
Expert
Disturbance protection suitability index (SIC disturbance) per average quality class (to % of max)
Very high
High
Moderate
Low
(-100)
(-80)
(-60)
(-40)
(-30)
(-25)
(-20)
(-15)
(-10)
(-5)
Expert 1
90
70
50
30
30
30
10
10
10
10
Expert 2
90
90
90
70
70
70
50
30
30
10
Expert 3
Very low
Not elicited
Expert 4
90
90
70
50
50
50
30
30
10
10
Expert 5
90
70
50
30
30
30
10
10
10
10
Expert 6
90
90
90
70
70
70
50
50
30
10
169
Appendix 3.2 Methodology – spatial pattern suitability analysis
This page is intentionally blank.
170
R code - spatial pattern suitability analysis
(PATTSI)
Jens G. Froese
10 August 2016
This document provides a detailed, reproducible description of the spatial pattern suitability analysis methodology. It is
Supporting Information to the manuscript:
Froese JG, Smith CS, McAlpine CA, Durr PA, van Klinken RD. Integrating landscape
structure improves habitat models of mobile animals: feral pigs in Australia.
It is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. A
.zip file containing data inputs to reproduce analyses can be requested from the authors. Some parts of the code
used to print this document have been suppressed to enhance readability. A generalized version of the code is
available at URL https://github.com/jgfroese/PATTSI (https://github.com/jgfroese/PATTSI).
Load required R packages
R packages raster ([4]) and Matrix ([5]) and their dependencies are required for spatial pattern suitability
analysis. Session information incl. package versions are listed at the bottom of this document.
require(raster) # for all analyses of raster objects incl. moving window analysis
require(Matrix) # for function `nnzero`
1. Expert elicitation
Overview
For each habitat variable (i.e. its modelled resource quality indices xr ), we elicited two response-to-pattern curves
from experts, relating structural metrics of resource patterns to functional suitability indices SIr . We focused on a
distance-dependent and a composition-dependent response curve. The information below supplements the
description of elicitation procedures in Chapter 3. Elicitation results from each indivudal expert are provided in
Appendix Table 3.1.4. Expert-averaged response-to-pattern curves are shown in Fig 3.5.
Step 1: Distance-dependent response-to-pattern curve
Distance-dependent curves followed a step-wise pattern, because we discretised both:
distance into five equal distance bands (“very close”, “close”, “medium”, “far” and “very far”), relative to each
expert’s defined mobility threshold (i.e. 1km, 2km or 3km).
resource suitability indices ( SIr ) into five equal classes (“very good (80-100)”, “good (60-80)”, “moderate
(40-60)”, “poor (20-40)” and “very poor (0-20)”)
We asked experts to relate each distance band to a corresponding suitability class under the assumption that other
variables do not constrain suitability.
To derive distance weights for computation of 2. Distance-dependent resource suitability indices, Step 4, we used
the mid-points of elicited suitability index classes divided by 100 (e.g. class “moderate (40-60)” = SIr 50 = weight
0.5). For class “very good (80-100)” we did not use the mid-point SIr = 90 but assigned SIr = 100 (= weight 1.0)
to avoid unintended distance penalties (i.e. an adjacent resource of quality xr = 60 should compute as distanceweighted suitability SIr = 60 (if weight is 1.0) and not SIr = 54 (if weight is 0.9)).
Figure 1 Step-wise distance-dependent response curve for habitat variable “food” elicited from expert 4 (A) and
distance weights derived for computation of food suitability indices (B)
Step 2: Composition-dependent response-to-pattern-curve
Composition-dependent curves also followed a step-wise pattern, because we discretised both:
averagen quality into five equal classes (“very high (80-100%)”, “high (60-80%)”, “moderate (40-60%)”, “low
(20-40%)” and “very low (0-20%)”), relative relative to 100% cover with highest quality resources.
resource suitability indices ( SIr ) into five equal classes (“very good (80-100)”, “good (60-80)”, “moderate
(40-60)”, “poor (20-40)” and “very poor (0-20)”)
We asked experts to relate each average quality class to a corresponding suitability class under the assumption that
other variables do not constrain suitability.
Prior to computation of 3. Composition-dependent resource suitability indices, Step 2, we slightly adjusted
response curves to avoid unintended results. For example, some experts related “very low” (= 0-20%) average quality
to “moderate” suitability to highlight feral pigs’ ability to effectively utilize small quantities of a high quality resource. As
such a function would also describe absent resources as “moderately” suitable, we adjusted it to 0-5% average quality
= “very poor” suitability, 5-10% average quality = “poor” suitability and 10-20% average quality = “moderate” suitability.
Figure 2 Step-wise composition-dependent response curve for habitat variable “food” elicited from expert 4 (A) and
adjusted curve used for computation of food suitability indices (B)]
2. Distance-dependent resource suitability indices
Goal
Focal pixel resource suitability index SIr depends on the distance of a (numerical) habitat variable.
Method:
Generate a circular moving window where each position is weighted by its distance from the focal pixel
(radius/weights derived from 1. Expert elicitation).
compute the focal pixel SIr as the highest weighted value ( xr ) of a habitat variable within this moving
window.
Step 1
Define a function that returns a circular matrix of given radius and resolution and assigns value 1 if matrix position <=
radius and value NA if matrix position > radius (Source: [6])
make_circ_filter <- function(radius, res){
circ_filter <- matrix(NA, nrow=1+(2*radius/res), ncol=1+(2*radius/res))
dimnames(circ_filter)[[1]] <- seq(-radius, radius, by=res)
dimnames(circ_filter)[[2]] <- seq(-radius, radius, by=res)
sweeper <- function(mat){
for(row in 1:nrow(mat)){
for(col in 1:ncol(mat)){
dist <- sqrt((as.numeric(dimnames(mat)[[1]])[row])^2 +
(as.numeric(dimnames(mat)[[1]])[col])^2)
if(dist<=radius) {mat[row, col]<-1}
}
}
return(mat)
}
out <- sweeper(circ_filter)
return(out)
}
Step 2
Apply function to generate five matrices with different radii (= distance bands), relative to each expert’s defined
mobility threshold (i.e. 1km for Expert1, 2km for Experts 3/5 and 3km for Experts 2/4/6).
res <- 1 # resolution (= pixel size, e.g. 100m)
mr <- 10 # matrix radius (= mobility threshold, must be multiple of res, e.g 1km = 10 x 100
m)
m.vf <- make_circ_filter(mr, res) # distance band 'very far' (= mobility threshold)
m.f <- make_circ_filter((mr/5)*4, res) # distance band 'far'
m.m <- make_circ_filter((mr/5)*3, res) # distance band 'medium'
m.c <- make_circ_filter((mr/5)*2, res) # distance band 'close'
m.vc <- make_circ_filter((mr/5), res) # distance band 'very close'
Replace value==1 with unique temp value in ascending order from largest to smallest matrix
m.vf[m.vf == 1] <- 1
m.f[m.f == 1] <- 2
m.m[m.m == 1] <- 3
m.c[m.c == 1] <- 4
m.vc[m.vc == 1] <- 5
Step 3
Combine the five matrices into one (two at a time starting with the smallest):
a.1 <- array(NA, dim(m.c), dimnames(m.c)) # create temp array of size = larger matrix
a.1[rownames(m.vc), colnames(m.vc)] <- m.vc # ... with values = smaller matrix
m.c <- pmax(m.c, a.1, na.rm = TRUE) # combine values: larger matrix + temp array
a.2 <- array(NA, dim(m.m), dimnames(m.m)) # repeat with: output + next-larger matrix
a.2[rownames(m.c), colnames(m.c)] <- m.c
m.m <- pmax(m.m, a.2, na.rm = TRUE)
a.3 <- array(NA, dim(m.f), dimnames(m.f))
a.3[rownames(m.m), colnames(m.m)] <- m.m
m.f <- pmax(m.f, a.3, na.rm = TRUE)
a.4 <- array(NA, dim(m.vf), dimnames(m.vf))
a.4[rownames(m.f), colnames(m.f)] <- m.f
m.band <- pmax(m.vf, a.4, na.rm = TRUE)
m.band
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
-10
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
8
9
10
-10
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
1
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
-9
NA
NA
NA
NA
NA
NA
1
1
1
1
1
1
1
1
1
NA
NA
NA
NA
NA
NA
-8
NA
NA
NA
NA
1
1
1
1
1
1
2
1
1
1
1
1
1
NA
NA
NA
NA
-7
NA
NA
NA
1
1
1
1
2
2
2
2
2
2
2
1
1
1
1
NA
NA
NA
-6
NA
NA
1
1
1
2
2
2
2
2
3
2
2
2
2
2
1
1
1
NA
NA
-5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
NA NA NA NA NA 1 NA NA NA NA NA NA NA
NA 1 1 1 1 1 1 1 1 1 NA NA NA
1 1 1 1 1 2 1 1 1 1 1 1 NA
1 1 2 2 2 2 2 2 2 1 1 1 1
2 2 2 2 2 3 2 2 2 2 2 1 1
2 2 3 3 3 3 3 3 3 2 2 2 1
2 3 3 3 3 4 3 3 3 3 2 2 1
3 3 3 4 4 4 4 4 3 3 3 2 2
3 3 4 4 4 5 4 4 4 3 3 2 2
3 3 4 4 5 5 5 4 4 3 3 2 2
3 4 4 5 5 5 5 5 4 4 3 3 2
3 3 4 4 5 5 5 4 4 3 3 2 2
3 3 4 4 4 5 4 4 4 3 3 2 2
3 3 3 4 4 4 4 4 3 3 3 2 2
2 3 3 3 3 4 3 3 3 3 2 2 1
2 2 3 3 3 3 3 3 3 2 2 2 1
2 2 2 2 2 3 2 2 2 2 2 1 1
1 1 2 2 2 2 2 2 2 1 1 1 1
1 1 1 1 1 2 1 1 1 1 1 1 NA
NA 1 1 1 1 1 1 1 1 1 NA NA NA
NA NA NA NA NA 1 NA NA NA NA NA NA NA
8
NA
NA
NA
NA
1
1
1
1
1
1
2
1
1
1
1
1
1
NA
NA
NA
NA
9
NA
NA
NA
NA
NA
NA
1
1
1
1
1
1
1
1
1
NA
NA
NA
NA
NA
NA
10
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
1
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
Step 4
Replace temp values with averaged expert-elicited weight for each distance band ( ESMdata.zip file containing
f_DFood.csv and other weights derived from 1. Expert elicitation, Step 1 can be requested from the authors)
csv.DFood = read.csv("ESMdata/MWA/f/f_DFood.csv")
m.band[m.band == 1] <- mean(csv.DFood$very.far)
m.band[m.band == 2] <- mean(csv.DFood$far)
m.band[m.band == 3] <- mean(csv.DFood$medium)
m.band[m.band == 4] <- mean(csv.DFood$close)
m.band[m.band == 5] <- mean(csv.DFood$very.close)
## very.close
##
1.00
close
0.75
medium
0.58
far
0.37
very.far
0.17
Step 5
Perform moving window analysis using function focal {raster} with parameters ( ESMdata.zip file containing
raster layers with resource quality indices can be requested from the authors):
r = raster("ESMdata/MWA/GIS/Food-quality.tif") # raster layer with numerical resource quali
ty index
w = m.band # moving window is banded weights matrix
fun = max # focal pixel takes highest weighted resource quality index within moving window
WARNING! The following process may take several hours depending on the size of r and w
r.D <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padValue = NA) # na.rm = TRUE ignores No
Data
r.SI.D <- mask(r.D, r) # extract by r to remove padded edges
writeRaster(r.SI.D, filename = paste("ESMdata/MWA/out/Food_SI_D1k.tif", sep="")) # save oup
ut raster
3. Composition-dependent resource suitability indices
Overview
For each focal pixel, we computed the average resource quality index within a moving window and reclassified this
average value to a numerical suitability index from a reclassification matrix (here: circular window, five equal average
quality classes, radius/reclassifcation matrix elicited from experts in 1. Expert elicitation, Step 2)
Goal
Focal pixel suitability index SIr depends on the amount of a (numerical) habitat variable
Method:
For each focal pixel, compute the average value ( xr ) of a habitat variable within a moving window
reclassify the average value to a numerical focal pixel SIr from a reclassification matrix (here: circular window,
five equal average value classes, radius/reclassifcation matrix elicited from experts in 1. Expert elicitation,
Step 2)
Step 1
Perform moving window analysis using the ESRI ArcGIS Focal Statistics tool with parameters:
in_raster = “ESMdata/MWA/GIS/Food-quality.tif”: raster layer with numerical resource quality index
( ESMdata.zip file containing raster layers with resource quality indices can be requested from the authors)
neighborhood = NbrCircle, radius = mr: circular moving window with radius = mobility threshold (see above)
statistics_type = mean: focal pixel takes average resource quality index within moving window
ignore_nodata = TRUE: ignores NoData
## Currently not functioning: implementation using function "focal {raster}" with parameter
s:
r = raster("ESMdata/MWA/GIS/Food-quality.tif") # raster layer with numerical resource quali
ty index
w = m.vf # moving window is circular matrix with radius = moblity threshold
fun = mean # focal pixel takes average resource quality index within moving window
r.C <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padValue = NA) # na.rm = TRUE ignores No
Data
## returns empty result due to `NA` values at matrix positions > radius
## changing `NA` to `0` values distorts computations of mean value
r.C = raster("ESMdata/MWA/GIS/Food_C1k.tif") # load ArcGIS output raster with average resou
rce quality indices
r.C <- mask(r.C, r) # extract by r to remove padded edges
Step 2
Determine minimum and maximum averaged resource quality index values
fm.min <- minValue(r) # all resource quality indices within the moving window take minimum
value
fm.max <- maxValue(r) # all resource quality indices within moving window take maximum valu
e
Step 3
Replace averaged resource quality indices with averaged expert-elicited SIr for each average quality class
( ESMdata.zip file containing f_CFood.csv and other SIr derived from 1. Expert elicitation, Step 2 can be
requested from the authors)
csv.CFood = read.csv("ESMdata/MWA/f/f_CFood.csv")
rcl.C <- c(minValue(r.C), fm.min + (fm.max - fm.min) * 0.05, mean(csv.CFood$X5),
fm.min + (fm.max-fm.min) * 0.05, fm.min + (fm.max-fm.min) * 0.10, mean(csv.CFood$X10),
fm.min + (fm.max-fm.min) * 0.10, fm.min + (fm.max-fm.min) * 0.15, mean(csv.CFood$X15),
fm.min + (fm.max-fm.min) * 0.15, fm.min + (fm.max-fm.min) * 0.20, mean(csv.CFood$X20),
fm.min + (fm.max-fm.min) * 0.20, fm.min + (fm.max-fm.min) * 0.25, mean(csv.CFood$X25),
fm.min + (fm.max-fm.min) * 0.25, fm.min + (fm.max-fm.min) * 0.30, mean(csv.CFood$X30),
fm.min + (fm.max-fm.min) * 0.30, fm.min + (fm.max-fm.min) * 0.40, mean(csv.CFood$X40),
fm.min + (fm.max-fm.min) * 0.40, fm.min + (fm.max-fm.min) * 0.60, mean(csv.CFood$X60),
fm.min + (fm.max-fm.min) * 0.60, fm.min + (fm.max-fm.min) * 0.80, mean(csv.CFood$X80),
fm.min + (fm.max-fm.min) * 0.80, fm.min + (fm.max-fm.min) * 1.00, mean(csv.CFood$X100))
m.rcl.C <- matrix(rcl.C, ncol=3, byrow=TRUE) # reclassification matrix (3 columns: "from" /
"to" average quality, "SI" )
## X100
X80
X60
X40
X30
X25
X20
X15
X10
X5
## 90.00 86.67 76.67 60.00 53.33 53.33 33.33 30.00 20.00 10.00
r.SI.C <- reclassify(r.C, m.rcl.C) # apply function "reclassify {raster}"
writeRaster(r.SI.C, filename = paste("ESMdata/MWA/out/Food_SI_C1k.tif", sep="")) # save oup
ut raster
4. Combined composition/distance-dependent
resource suitability indices
Goal
Focal pixel suitability index SIr depends on both distance and amount of a (numerical) habitat variable
Method:
Generate a circular moving window where each position is weighted by its distance from the focal pixel
(radius/weights derived from 1. Expert elicitation).
For each focal pixel, compute the average value ( xr ) of a habitat variable within this weighted moving window
reclassify the average value to a numerical focal pixel SIr from a reclassification matrix (here: circular window,
five equal average value classes, radius/reclassifcation matrix elicited from experts in 1. Expert elicitation,
Step 2)
Step 1
Save banded weights matrix (as TXT, add header with number of rows/columns after export (e.g. 21 for matrix with 1
km radius)
m.exp <- m.band
m.exp[is.na(m.exp)] <- 0 # change NAs to 0s
write.table(m.exp, file = "ESMdata/MWA/out/m_Food_D1k.txt", row.names = F, col.names = F)
Step 2
Perform moving window analysis using the ESRI ArcGIS Focal Statistics tool with parameters:
in_raster = “ESMdata/MWA/GIS/Food-quality.tif”: raster layer with numerical resource quality index
( ESMdata.zip file containing raster layers with resource quality indices can be requested from the authors)
neighborhood = NbrWeight: select exported banded weights matrix .txt
statistics_type = mean: focal pixel takes weighted average resource quality index within moving window
ignore_nodata = TRUE: ignores NoData
## Currently not functioning: implementation using function "focal {raster}" with parameter
s:
r = raster("ESMdata/MWA/GIS/Food-quality.tif") # raster layer with numerical resource quali
ty index
w = m.band # moving window is banded weights matrix
fun = mean # focal pixel takes average resource quality index within moving window
r.DC <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padValue = NA) # na.rm = TRUE ignores N
oData
## returns empty result due to `NA` values at matrix positions > radius
## changing `NA` to `0` values distorts computations of mean value
r.DC = raster("ESMdata/MWA/GIS/Food_DC1k.tif") # load ArcGIS output raster with distance-we
ighted average resource quality indices
r.DC <- mask(r.DC, r) # extract by r to remove padded edges
Step 3
Determine minimum and maximum distance-weighted averaged resource quality index values as: (number of pixels in
each distance band * (minValue(r)/maxValue(r) * weight for distance band)) / total number of pixels in moving window
m.t <- table(m.vf) # number of pixels in each distance band
fwm.min <- (m.t[5] * minValue(r) * mean(csv.DFood$very.close) + m.t[4] * minValue(r) * mean
(csv.DFood$close) + m.t[3] * minValue(r) * mean(csv.DFood$medium) + m.t[2] * minValue(r) *
mean(csv.DFood$far) + m.t[1] * minValue(r) * mean(csv.DFood$very.far)) / nnzero(m.band, na.
counted = FALSE)
fwm.max <- (m.t[5] * maxValue(r) * mean(csv.DFood$very.close) + m.t[4] * maxValue(r) * mean
(csv.DFood$close) + m.t[3] * maxValue(r) * mean(csv.DFood$medium) + m.t[2] * maxValue(r) *
mean(csv.DFood$far) + m.t[1] * maxValue(r) * mean(csv.DFood$very.far)) / nnzero(m.band, na.
counted = FALSE)
as.vector(round(fwm.min, digits = 1)); as.vector(round(fwm.max, digits = 1)) # print as rou
nded vector
Here, we computed distance-weighted average minima/maxima as the average across all moving window sizes
(radius = 1/2/3 km). The following values were used for analysis:
Water: minimum = 6.1 (xrmin = 11); maximum = 49.5 (xrmax = 89)
Food: minimum = 4.5 (xrmin = 11); maximum = 35.9 (xrmax = 88)
Heat protection: minimum = 7.4 (xrmin = 15); maximum = 43.6 (xrmax = 89)
Disturbance protection: minimum = 3.6 (xrmin = 17); maximum = 18.9 (xrmax = 89)
Step 4
Replace distance-weighted averaged resource quality indices with averaged expert-elicited SIr for each average
quality class ( ESMdata.zip file containing f_CFood.csv and other SIr derived from 1. Expert elicitation, Step 2
can be requested from the authors)
csv.CFood = read.csv("ESMdata/MWA/f/f_CFood.csv")
rcl.DC <- c(minValue(r.DC), fwm.min + (fwm.max - fwm.min) * 0.05, mean(csv.CFood$X5),
fwm.min + (fwm.max - fwm.min) * 0.05, fwm.min + (fwm.max - fwm.min) * 0.10, mean(csv.CFoo
d$X10),
fwm.min + (fwm.max - fwm.min) * 0.10, fwm.min + (fwm.max - fwm.min) * 0.15, mean(csv.CFoo
d$X15),
fwm.min + (fwm.max - fwm.min) * 0.15, fwm.min + (fwm.max - fwm.min) * 0.20, mean(csv.CFoo
d$X20),
fwm.min + (fwm.max - fwm.min) * 0.20, fwm.min + (fwm.max - fwm.min) * 0.25, mean(csv.CFoo
d$X25),
fwm.min + (fwm.max - fwm.min) * 0.25, fwm.min + (fwm.max - fwm.min) * 0.30, mean(csv.CFoo
d$X30),
fwm.min + (fwm.max - fwm.min) * 0.30, fwm.min + (fwm.max - fwm.min) * 0.40, mean(csv.CFoo
d$X40),
fwm.min + (fwm.max - fwm.min) * 0.40, fwm.min + (fwm.max - fwm.min) * 0.60, mean(csv.CFoo
d$X60),
fwm.min + (fwm.max - fwm.min) * 0.60, fwm.min + (fwm.max - fwm.min) * 0.80, mean(csv.CFoo
d$X80),
fwm.min + (fwm.max - fwm.min) * 0.80, fwm.min + (fwm.max - fwm.min) * 1.00, mean(csv.CFoo
d$X100))
m.rcl.DC <- matrix(rcl.DC, ncol=3, byrow=TRUE) # reclassification matrix (3 columns: "from"
/ "to" distance-weighted average quality, "SI" )
r.SI.DC <- reclassify(r.DC, m.rcl.DC)
writeRaster(r.SI.DC, filename = paste("ESMdata/MWA/out/Food_SI_DC1k.tif", sep="")) # save o
uput raster
5. Example plots
References
[1] Allaire, J.J. et al. 2016. Package ‘rmarkdown’: dynamic documents for R. URL http://rmarkdown.rstudio.com/
(http://rmarkdown.rstudio.com/).
[2] Xie, Y. 2016. Package ‘knitr’: a general-purpose package for dynamic report generation in R. URL http://yihui.name
/knitr/ (http://yihui.name/knitr/).
[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for Statistical
Computing, Vienna, Austria. URL http://www.R-project.org/ (http://www.R-project.org/).
[4] Hijmans, R.J. 2015. Package ‘raster’: geographic data analysis and modeling. URL http://cran.r-project.org
/web/packages/raster/ (http://cran.r-project.org/web/packages/raster/).
[5] Bates, D. and Maechler, M. 2015. Package ‘Matrix’: sparse and dense matrix classes and methods. URL
http://Matrix.R-forge.R-project.org/ (http://Matrix.R-forge.R-project.org/).
[6] Scroggie, M. 2012. Applying a circular moving window filter to raster data in R. URL
https://scrogster.wordpress.com/2012/10/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/
(https://scrogster.wordpress.com/2012/10/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/).
Session information
## Session info --------------------------------------------------------------
##
##
##
##
##
##
##
##
setting
version
system
ui
language
collate
tz
date
value
R version 3.1.3 (2015-03-09)
x86_64, mingw32
RTerm
(EN)
English_Australia.1252
Australia/Brisbane
2017-04-17
## Packages ------------------------------------------------------------------
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
package
devtools
digest
evaluate
formatR
htmltools
knitr
lattice
magrittr
Matrix
memoise
raster
Rcpp
rgdal
rmarkdown
sp
stringi
stringr
yaml
* version
1.10.0
0.6.8
0.8
1.2.1
0.3
1.12.3
0.20-30
1.5
* 1.2-3
1.0.0
* 2.4-20
0.12.1
1.1-1
0.9.2
* 1.2-1
1.0-1
1.0.0
2.1.13
date
2016-01-23
2014-12-31
2015-09-18
2015-09-18
2015-12-29
2016-01-22
2015-02-22
2014-11-22
2015-11-28
2016-01-29
2015-09-08
2015-09-10
2015-11-02
2016-01-01
2015-10-18
2015-10-22
2015-04-30
2014-06-12
source
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
Appendix 3.3 Methodology – validation
This page is intentionally blank.
180
R code - model evaluation and validation
Jens G. Froese
10 August 2016
This document provides a detailed, reproducible description of the methodology used to evaluate and validate
performance of habitat suitability models using the Continuous Boyce Index. It is Supporting Information to the
manuscript:
Froese JG, Smith CS, McAlpine CA, Durr PA, van Klinken RD. Integrating landscape
structure improves habitat models of mobile animals: feral pigs in Australia.
The document is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible
reports. A .zip file containing data inputs to reproduce analyses can be requested from the authors. Some parts of
the code used to print this document have been surpressed to enhance readability. A generalized version of the
source code is available at URL https://github.com/jgfroese/HSI-CBI-validation (https://github.com/jgfroese/HSI-CBIvalidation).
Load required R packages
R packages ecospat ([4]), data.table ([5]) and zoo ([6]) and their dependencies are required for HSI-CBIvalidation. Session information incl. package versions are listed at the bottom of this document.
require(data.table) # for function `setnames`
require(ecospat) # for function `ecospat.boyce`
1. Data preparation
Step 1
This R script requires two .TXT files for each model / validation data combination, which have to be prepared in
package {raster} or alternative GIS software as described below. These files can be requested from the authors.
1. Expected HSI across validation background:
define three validation backgrounds ( Balkanu = BLKW , Lakefield = LKNP , NAQS - see Table 3.1)
mask raster layers of the eleven alternative habitat suitability models ( 1 - 10 with PATTSI, X without
PATTSI - see manuscript Table 3.2) by each validation background
export raster attribute tables to .TXT with 3 columns: [ID], [HSI], [pixel count]
2. Predicted HSI at species presence records:
convert presence records of three validation data sets into raster layer
combine masked raster layers of the eleven alternative models with each set of presence records
export raster attribute tables to .TXT with 5 columns: [ID], [Value], [pixel count], [HSI], [number of
presence records per pixel]
Step 2
Compute model-predicted HSI at feral pig presence records
First, read .TXT files for each model (m1, m2, …) / validation presences (BLKW, LKNP, NAQS) combination as data
frame,
pred.m1.BLKW = read.csv("ESMdata/Validation/BLKW/Predicted/SiPred_D1_BLKW1000.txt")
...
pred.m10.BLKW = read.csv("ESMdata/Validation/BLKW/Predicted/SiPred_BK_BLKW1000.txt")
pred.mX.BLKW = read.csv("ESMdata/Validation/BLKW/Predicted/SiPred_X_BLKW1000.txt")
...
pred.m10.LKNP = read.csv("ESMdata/Validation/LKNP/Predicted/SiPred_BK_LKNP1000.txt")
pred.mX.LKNP = read.csv("ESMdata/Validation/LKNP/Predicted/SiPred_X_LKNP1000.txt")
...
pred.m10.NAQS = read.csv("ESMdata/Validation/NAQS/Predicted/SiPred_BK_NAQS1000.txt")
pred.mX.NAQS = read.csv("ESMdata/Validation/NAQS/Predicted/SiPred_X_NAQS1000.txt")
and combine all data frames in a list for faster analysis.
pred.list <- list (pred.m1.BLKW, pred.m2.BLKW, pred.m3.BLKW, pred.m4.BLKW, pred.m5.BLKW, pr
ed.m6.BLKW, pred.m7.BLKW, pred.m8.BLKW, pred.m9.BLKW, pred.m10.BLKW, pred.mX.BLKW, pred.m1.
LKNP, pred.m2.LKNP, pred.m3.LKNP, pred.m4.LKNP, pred.m5.LKNP, pred.m6.LKNP, pred.m7.LKNP, p
red.m8.LKNP, pred.m9.LKNP, pred.m10.LKNP, pred.mX.LKNP, pred.m1.NAQS, pred.m2.NAQS, pred.m3
.NAQS, pred.m4.NAQS, pred.m5.NAQS, pred.m6.NAQS, pred.m7.NAQS, pred.m8.NAQS, pred.m9.NAQS,
pred.m10.NAQS, pred.mX.NAQS)
n.list <- 33 # the number of data frames in your list
Then, homogenise the five column names for all data frames,
for (i in seq_along(pred.list)) {
setnames(pred.list[[i]], c("ID", "Value", "Pixelcount", "HSI", "Presences"))
}
and calculate the total number of presence records per HSI value (one pixel may contain multiple records)
pred.sum.list <- vector("list", n.list)
for (i in seq_along(pred.list)) {
pred.sum.list[[i]] <- aggregate(cbind(Pixelcount*Presences)~HSI, data = pred.list[[i]], s
um)
}
Finally, add descriptive column names to the new list of data frames
for (i in seq_along(pred.sum.list)) {
setnames(pred.sum.list[[i]], c("HSI", "Presences"))
}
and convert it into a list of vectors (= HSI at feral pig presence records)
pred.v.list = vector("list", n.list)
for (i in seq_along(pred.sum.list)) {
pred.v = vector()
for (j in 1:length(pred.sum.list[[i]]$Presences)) {
for (k in 1:pred.sum.list[[i]][j, 2]) {
pred.v <-append(pred.v, pred.sum.list[[i]][j, 1])
}
}
pred.v.list[[i]] <- append(pred.v.list[[i]], pred.v)
}
e.g. HSI of model 1 at Balkanu presence records (first in list)
##
[1] 11 11
## [24] 48 48
## [47] 58 58
## [70] 66 67
## [93] 71 71
## [116] 71 71
## [139] 71 71
## [162] 75 75
11
48
58
67
71
71
71
75
11
48
58
67
71
71
71
75
11
48
58
67
71
71
71
75
11
48
60
67
71
71
71
75
11
49
61
67
71
71
71
75
11
49
61
67
71
71
71
75
11
49
61
67
71
71
71
75
35
49
61
71
71
71
71
75
40
49
61
71
71
71
71
75
43
49
61
71
71
71
72
75
43
49
61
71
71
71
72
75
43
49
61
71
71
71
75
75
44
49
66
71
71
71
75
75
44
49
66
71
71
71
75
75
44
49
66
71
71
71
75
75
46
49
66
71
71
71
75
75
46
52
66
71
71
71
75
77
46
53
66
71
71
71
75
77
46
57
66
71
71
71
75
46
58
66
71
71
71
75
48
58
66
71
71
71
75
Step 3
Compute model-predicted HSI across validation backgrounds
First, read .TXT files for each model (m1, m2, …) / validation background (BLKW, LKNP, NAQS) combination as data
frame
exp.m1.BLKW = read.csv("ESMdata/Validation/BLKW/Expected/SiExp_D1_BLKW1000.txt")
...
exp.m10.BLKW = read.csv("ESMdata/Validation/BLKW/Expected/SiExp_BK_BLKW1000.txt")
exp.mX.BLKW = read.csv("ESMdata/Validation/BLKW/Expected/SiExp_X_BLKW1000.txt")
...
exp.m10.LKNP = read.csv("ESMdata/Validation/LKNP/Expected/SiExp_BK_LKNP1000.txt")
exp.mX.LKNP = read.csv("ESMdata/Validation/LKNP/Expected/SiExp_X_LKNP1000.txt")
...
exp.m10.NAQS = read.csv("ESMdata/Validation/NAQS/Expected/SiExp_BK_NAQS1000.txt")
exp.mX.NAQS = read.csv("ESMdata/Validation/NAQS/Expected/SiExp_X_NAQS1000.txt")
and combine all data frames in a list for faster analysis.
exp.list <- list (exp.m1.BLKW, exp.m2.BLKW, exp.m3.BLKW, exp.m4.BLKW, exp.m5.BLKW, exp.m6.B
LKW, exp.m7.BLKW, exp.m8.BLKW, exp.m9.BLKW, exp.m10.BLKW, exp.mX.BLKW, exp.m1.LKNP, exp.m2.
LKNP, exp.m3.LKNP, exp.m4.LKNP, exp.m5.LKNP, exp.m6.LKNP, exp.m7.LKNP, exp.m8.LKNP, exp.m9.
LKNP, exp.m10.LKNP, exp.mX.LKNP, exp.m1.NAQS, exp.m2.NAQS, exp.m3.NAQS, exp.m4.NAQS, exp.m5
.NAQS, exp.m6.NAQS, exp.m7.NAQS, exp.m8.NAQS, exp.m9.NAQS, exp.m10.NAQS, exp.mX.NAQS)
Then, homogenise the three column names for all data frames
for (i in seq_along(exp.list)) {
setnames(exp.list[[i]], c("ID", "HSI", "Pixelcount"))
}
and convert it into a list of vectors (= HSI across validation backgrounds [potentially very large])
exp.v.list = vector("list", n.list)
for (i in seq_along(exp.list)) {
exp.v = vector()
for (j in 1:length(exp.list[[i]]$Pixelcount)) {
for (k in 1:exp.list[[i]][j, 3]) {
exp.v <-append(exp.v, exp.list[[i]][j, 2])
}
}
exp.v.list[[i]] <- append(exp.v.list[[i]], exp.v)
}
2. Data analysis
Step 4
Apply function boyce {ecospat} to each model (m1, m2, …) / validation data (BLKW, LKNP, NAQS) combination
with parameters:
exp.v.list[[i]] # Expected HSI across background ([[1]] = m1.BLKW, ... , [[33]] = mX.NAQS)
pred.v.list[[i]] # Predicted HSI at presences ([[1]] = m1.BLKW, ... , [[33]] = mX.NAQS)
nclass = 0 # defaults to moving window (continuous, classification-independent) computation
with arguments
window.w = 10 # moving window width (i.e. 10 adjacent HSI values are considered in each com
putation)
res = 100 # resolution factor (i.e. 100 computations across the total range of HSI)
PEplot = F # no PEplot is generated (customised plots - see Fig 3.6 and Appendix 3.4)
boyce.m1.BLKW <- ecospat.boyce(exp.v.list[[1]], pred.v.list[[1]], nclass, window.w, res, PE
plot)
...
boyce.m10.BLKW <- ecospat.boyce(exp.v.list[[10]], pred.v.list[[10]], nclass, window.w, res,
PEplot)
boyce.mX.BLKW <- ecospat.boyce(exp.v.list[[11]], pred.v.list[[11]], nclass, window.w, res,
PEplot)
...
boyce.m10.LKNP <- ecospat.boyce(exp.v.list[[21]], pred.v.list[[21]], nclass, window.w, res,
PEplot)
boyce.mX.LKNP <- ecospat.boyce(exp.v.list[[22]], pred.v.list[[22]], nclass, window.w, res,
PEplot)
...
boyce.m10.NAQS <- ecospat.boyce(exp.v.list[[32]], pred.v.list[[32]], nclass, window.w, res,
PEplot)
boyce.mX.NAQS <- ecospat.boyce(exp.v.list[[33]], pred.v.list[[33]], nclass, window.w, res,
PEplot)
Step 5
Investigate results of CBI analysis
Combine all results in list for faster analysis,
boyce.list <- list (boyce.m1.BLKW, boyce.m2.BLKW, boyce.m3.BLKW, boyce.m4.BLKW, boyce.m5.BL
KW, boyce.m6.BLKW, boyce.m7.BLKW, boyce.m8.BLKW, boyce.m9.BLKW, boyce.m10.BLKW, boyce.mX.BL
KW, boyce.m1.LKNP, boyce.m2.LKNP, boyce.m3.LKNP, boyce.m4.LKNP, boyce.m5.LKNP, boyce.m6.LKN
P, boyce.m7.LKNP, boyce.m8.LKNP, boyce.m9.LKNP, boyce.m10.LKNP, boyce.mX.LKNP, boyce.m1.NAQ
S, boyce.m2.NAQS, boyce.m3.NAQS, boyce.m4.NAQS, boyce.m5.NAQS, boyce.m6.NAQS, boyce.m7.NAQS
, boyce.m8.NAQS, boyce.m9.NAQS, boyce.m10.NAQS, boyce.mX.NAQS)
and print CBI ($Spearman.cor) for all model / validation data combinations (see Table 3.2)
CBI.list = vector("list", n.list)
for (i in seq_along(boyce.list)) {
CBI.list[[i]] <- append(CBI.list[[i]], round(boyce.list[[i]]$Spearman.cor, digits = 2))
}
CBI.list
e.g. CBI of model 1 at Balkanu presence records (first in list)
## [1] 0.86
Step 6
Compute proportion of validation background expected to be highly or very highly suitable habitat (HSI >= 60) for all
model / validation data combinations (see Table 3.2)
t.HSI <- 59.99 # HSI threshold
HSI.60 <- vector("list", n.list)
for (i in seq_along(HSI.60)) {
HSI.60[[i]] <- aggregate(Pixelcount~HSI > t.HSI, data = exp.list[[i]], sum) / sum(exp.lis
t[[i]]$Pixelcount)
}
HSI.60.list <- vector("list", n.list)
for (i in seq_along(HSI.60.list)) {
HSI.60.list[[i]] <- append(HSI.60.list[[i]], (round(HSI.60[[i]][2, "Pixelcount"] * 100, d
igits = 0)))
}
HSI.60.list
e.g. HSI >= 60 of model 1 at Balkanu presence records (first in list)
## [1] 34
Plots comparing the P/E ratio of the three validation data sets for each of the eleven alternative habitat suitability
models are provided as Appendix 3.4.
Additional analysis
Fit linear regression model between CBI and HSI >= 60
dat <- read.csv("ESMdata/Validation/lmHSI-CBI.csv")
str(dat)
## 'data.frame':
33 obs. of 2 variables:
## $ HSI: int 2 4 4 5 5 5 5 5 5 5 ...
## $ CBI: num -0.88 -0.14 0.17 0.71 0.47 -0.01 0.16 -0.25 0.05 0.31 ...
plot(CBI ~ HSI, data = dat) # plot CBI against HSI >= 60
lmodel <- lm(CBI ~ log10(HSI), data = dat) # log transform HSI >= 60
plot(CBI ~ log10(HSI), data = dat); abline(lmodel) # plot again with regression line
summary(lmodel) # summary statistics
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
Call:
lm(formula = CBI ~ log10(HSI), data = dat)
Residuals:
Min
1Q
-0.81480 -0.14624
Median
0.05406
3Q
0.19734
Max
0.51376
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.3829
0.1718 -2.229
0.0332 *
log10(HSI)
0.8286
0.1487
5.573 4.15e-06 ***
--Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.332 on 31 degrees of freedom
Multiple R-squared: 0.5005, Adjusted R-squared: 0.4844
F-statistic: 31.06 on 1 and 31 DF, p-value: 4.15e-06
par(mfrow = c(2,2))
plot(lmodel) # plot model diagnostics
References
[1] Allaire, J.J. et al. 2016. Package ‘rmarkdown’: dynamic documents for R. URL http://rmarkdown.rstudio.com/
(http://rmarkdown.rstudio.com/).
[2] Xie, Y. 2016. Package ‘knitr’: a general-purpose package for dynamic report generation in R. URL http://yihui.name
/knitr/ (http://yihui.name/knitr/).
[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for Statistical
Computing, Vienna, Austria. URL http://www.R-project.org/ (http://www.R-project.org/).
[4] Broennimann, O. 2015. Package ‘ecospat’: spatial ecology miscellaneous methods. URL http://cran.r-project.org
/web/packages/ecospat/ (http://cran.r-project.org/web/packages/ecospat/).
[5] Dowle, M. et al. 2015. Package ‘data.table’: extension of data.frame. URL https://github.com/Rdatatable/data.table
/wiki/ (https://github.com/Rdatatable/data.table/wiki/).
[6] Zeileis, A. et al. 2015. Package ‘zoo’: S3 infrastructure for regular and irregular time series. URL http://zoo.RForge.R-project.org/ (http://zoo.R-Forge.R-project.org/).
Session information
## Session info --------------------------------------------------------------
##
##
##
##
##
##
##
##
setting
version
system
ui
language
collate
tz
date
value
R version 3.1.3 (2015-03-09)
x86_64, mingw32
RTerm
(EN)
English_Australia.1252
Australia/Brisbane
2017-04-17
## Packages ------------------------------------------------------------------
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
package
abind
acepack
ade4
adehabitatHR
adehabitatLT
adehabitatMA
ape
biomod2
boot
chron
CircStats
class
cluster
codetools
colorspace
data.table
deldir
devtools
digest
dismo
ecodist
ecospat
evaluate
foreach
foreign
formatR
Formula
gam
gbm
ggplot2
goftest
gridExtra
gtable
hexbin
Hmisc
htmltools
iterators
knitr
lattice
latticeExtra
magrittr
maptools
MASS
Matrix
MatrixModels
mda
memoise
mgcv
multcomp
munsell
mvtnorm
nlme
nnet
plyr
polspline
polyclip
pROC
proto
quantreg
randomForest
raster
* version
1.4-3
1.3-3.3
* 1.7-2
0.4.14
0.3.20
0.3.10
* 3.3
3.1-64
1.3-15
2.3-47
0.2-4
7.3-12
2.0.1
0.2-10
1.2-6
* 1.9.6
0.1-9
1.10.0
0.6.8
1.0-12
1.2.9
* 1.1
0.8
* 1.4.3
0.8-63
1.2.1
1.2-1
* 1.12
* 2.1.1
2.1.0
1.0-3
2.0.0
0.1.2
1.27.1
3.17-0
0.3
1.0.8
1.12.3
* 0.20-30
0.6-26
1.5
0.8-37
7.3-39
1.2-3
0.4-1
0.4-7
1.0.0
1.8-4
1.4-1
0.4.2
1.0-3
3.1-120
7.3-9
1.8.3
1.1.12
1.3-2
1.8
0.3-10
5.19
4.6-12
2.4-20
date
2015-03-13
2013-05-03
2015-04-14
2015-07-22
2015-07-22
2015-07-22
2015-05-29
2014-12-10
2015-02-24
2015-06-24
2012-10-29
2015-02-11
2015-01-31
2015-01-17
2015-03-11
2015-09-19
2015-03-09
2016-01-23
2014-12-31
2015-03-15
2013-12-03
2015-03-06
2015-09-18
2015-10-13
2015-02-20
2015-09-18
2015-04-07
2015-05-13
2015-03-11
2016-03-01
2015-07-03
2015-07-14
2012-12-05
2015-08-19
2015-09-21
2015-12-29
2015-10-13
2016-01-22
2015-02-22
2013-08-15
2014-11-22
2015-09-29
2015-02-24
2015-11-28
2015-08-22
2015-05-25
2016-01-29
2014-11-27
2015-07-23
2013-07-11
2015-07-22
2015-02-20
2015-02-11
2015-06-12
2015-07-14
2015-05-27
2015-05-05
2012-12-22
2015-08-31
2015-10-07
2015-09-08
source
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
rasterVis
0.37
RColorBrewer
1.1-2
Rcpp
0.12.1
reshape
0.8.5
rmarkdown
0.9.2
rms
4.4-0
rpart
4.1-9
sandwich
2.3-4
scales
0.3.0
sp
* 1.2-1
SparseM
1.7
spatstat
1.41-1
stringi
1.0-1
stringr
1.0.0
survival
* 2.38-1
tensor
1.5
TH.data
1.0-6
yaml
2.1.13
zoo
1.7-12
2015-09-06
2014-12-07
2015-09-10
2014-04-23
2016-01-01
2015-09-28
2015-02-24
2015-09-24
2015-08-25
2015-10-18
2015-08-15
2015-02-27
2015-10-22
2015-04-30
2015-02-24
2012-05-05
2015-01-05
2014-06-12
2015-03-16
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
Appendix 3.4 Validation plots – model scenarios
Appendix Fig 3.4 Validation plots for eleven habitat suitability models (panels A-K, refer to Table 3.2
for corresponding landscape metrics and scales of measurement). The predicted-to-expected (P/E) ratio is
plotted for each validation data set (black lines) and as a smoothed average of all data sets (red line).
190
Appendix 4.1 Model variables – parameters (supplements Fig 4.2)
Appendix Table 4.1 Bayesian network conditional probability tables. Used to model resource quality
indices for water (A), food (B), protection from heat stress (C) and protection from disturbance (D1 and D2)
as well as an index of habitat suitability for feral pig breeding and persistence (E*) as shown in Fig 4.2.
(A)
Water quality index (% probability)
Expert
Freshwater
presence
Terrain
ruggedness
Expert1
Yes
Low
97
Expert1
Yes
Moderate
50
Expert1
Yes
High
3
Expert1
No
Low
Expert1
No
Expert1
No
Expert2
Moderate
Poor
Very poor
3
0
0
0
40
10
0
0
23
48
23
3
0
0
0
16
84
Moderate
0
0
0
16
84
High
0
0
0
16
84
Yes
Low
97
3
0
0
0
Expert2
Yes
Moderate
97
3
0
0
0
Expert2
Yes
High
33
33
23
9
2
Expert2
No
Low
0
0
0
3
97
Expert2
No
Moderate
0
0
0
3
97
Expert2
No
High
0
0
0
3
97
Expert3
Yes
Low
97
3
0
0
0
Expert3
Yes
Moderate
97
3
0
0
0
Expert3
Yes
High
33
33
23
9
2
Expert3
No
Low
0
0
0
3
97
Expert3
No
Moderate
0
0
0
3
97
Expert3
No
High
0
0
0
3
97
Expert4
Yes
Low
97
3
0
0
0
Expert4
Yes
Moderate
97
3
0
0
0
Expert4
Yes
High
3
47
47
3
0
Expert4
No
Low
0
0
0
3
97
Expert4
No
Moderate
0
0
0
3
97
Expert4
No
High
0
0
0
3
97
Very good
Good
* Due to size constraints in this Appendix, the CPT for the habitat suitability index is shown for only one expert (Expert
1) for illustrative purposes. Full CPTs can be requested from the author in .xlsx format.
191
(B)
Expert
Food
quality
Food
quantity
Expert1
High
Expert1
Food quality index (% probability)
Very good
Good
Moderate
Poor
Very poor
High
68
29
3
0
0
High
Moderate
43
46
11
0
0
Expert1
High
Low
29
50
20
1
0
Expert1
Moderate
High
33
52
15
0
0
Expert1
Moderate
Moderate
2
31
51
15
1
Expert1
Moderate
Low
0
15
47
34
4
Expert1
Low
High
13
50
34
3
0
Expert1
Low
Moderate
0
8
44
42
6
Expert1
Low
Low
0
0
0
3
97
Expert2
High
High
97
3
0
0
0
Expert2
High
Moderate
36
50
14
0
0
Expert2
High
Low
17
51
30
2
0
Expert2
Moderate
High
15
54
29
2
0
Expert2
Moderate
Moderate
1
26
54
18
1
Expert2
Moderate
Low
0
9
45
40
6
Expert2
Low
High
0
0
0
26
74
Expert2
Low
Moderate
0
0
0
15
85
Expert2
Low
Low
0
0
0
3
97
Expert3
High
High
97
3
0
0
0
Expert3
High
Moderate
36
50
14
0
0
Expert3
High
Low
17
51
30
2
0
Expert3
Moderate
High
15
54
29
2
0
Expert3
Moderate
Moderate
1
26
54
18
1
Expert3
Moderate
Low
0
9
45
40
6
Expert3
Low
High
0
3
23
48
26
Expert3
Low
Moderate
0
0
12
26
62
Expert3
Low
Low
0
0
0
3
97
Expert4
High
High
97
3
0
0
0
Expert4
High
Moderate
33
37
22
7
1
Expert4
High
Low
23
35
28
12
2
Expert4
Moderate
High
23
35
28
12
2
Expert4
Moderate
Moderate
10
26
33
23
8
Expert4
Moderate
Low
5
18
32
30
15
Expert4
Low
High
9
24
34
24
9
Expert4
Low
Moderate
2
12
28
35
23
Expert4
Low
Low
0
0
0
3
97
Expert5
High
High
97
3
0
0
0
Expert5
High
Moderate
50
47
3
0
0
Expert5
High
Low
3
47
47
3
0
Expert5
Moderate
High
3
47
47
3
0
Expert5
Moderate
Moderate
1
27
53
18
1
Expert5
Moderate
Low
0
12
47
36
5
Expert5
Low
High
0
0
3
47
50
Expert5
Low
Moderate
0
0
2
25
73
Expert5
Low
Low
0
0
0
3
97
192
(C)
Expert
Heat stress
Shady vegetation
cover
Expert1
Low
Expert1
Heat protection quality index (% probability)
Very good
Good
Moderate
Poor
Very poor
Good
97
3
0
0
0
Low
Moderate
97
3
0
0
0
Expert1
Low
Poor
97
3
0
0
0
Expert1
Moderate
Good
65
33
2
0
0
Expert1
Moderate
Moderate
26
48
23
3
0
Expert1
Moderate
Poor
3
23
48
23
3
Expert1
High
Good
26
47
24
3
0
Expert1
High
Moderate
20
35
40
4
1
Expert1
High
Poor
0
0
0
3
97
Expert2
Low
Good
97
3
0
0
0
Expert2
Low
Moderate
97
3
0
0
0
Expert2
Low
Poor
97
3
0
0
0
Expert2
Moderate
Good
97
3
0
0
0
Expert2
Moderate
Moderate
3
47
47
3
0
Expert2
Moderate
Poor
0
3
47
47
3
Expert2
High
Good
97
3
0
0
0
Expert2
High
Moderate
0
16
68
16
0
Expert2
High
Poor
0
0
0
3
97
Expert3
Low
Good
97
3
0
0
0
Expert3
Low
Moderate
97
3
0
0
0
Expert3
Low
Poor
97
3
0
0
0
Expert3
Moderate
Good
97
3
0
0
0
Expert3
Moderate
Moderate
62
26
11
1
0
Expert3
Moderate
Poor
26
48
23
3
0
Expert3
High
Good
97
3
0
0
0
Expert3
High
Moderate
26
48
23
3
0
Expert3
High
Poor
0
0
0
3
97
Expert4
Low
Good
97
3
0
0
0
Expert4
Low
Moderate
97
3
0
0
0
Expert4
Low
Poor
97
3
0
0
0
Expert4
Moderate
Good
97
3
0
0
0
Expert4
Moderate
Moderate
3
47
47
3
0
Expert4
Moderate
Poor
0
3
47
47
3
Expert4
High
Good
97
3
0
0
0
Expert4
High
Moderate
3
23
48
23
3
Expert4
High
Poor
0
3
23
48
26
Expert5
Low
Good
97
3
0
0
0
Expert5
Low
Moderate
97
3
0
0
0
Expert5
Low
Poor
97
3
0
0
0
Expert5
Moderate
Good
97
3
0
0
0
Expert5
Moderate
Moderate
16
68
16
0
0
Expert5
Moderate
Poor
0
16
68
16
0
Expert5
High
Good
97
3
0
0
0
Expert5
High
Moderate
3
47
47
3
0
Expert5
High
Poor
0
0
0
3
97
193
Disturbance Dense vegetation
stress
cover
Disturbance protection quality index (% probability)
Very good
Good
Moderate
Poor
Very poor
Expert1
Low
Good
97
3
0
0
0
Expert1
Low
Moderate
97
3
0
0
0
Expert1
Low
Poor
97
3
0
0
0
Expert1
Moderate
Good
74
26
0
0
0
Expert1
Moderate
Moderate
26
71
3
0
0
Expert1
Moderate
Poor
0
16
68
16
0
Expert1
High
Good
74
26
0
0
0
Expert1
High
Moderate
0
1
53
45
1
Expert1
High
Poor
0
0
0
3
97
Expert2
Low
Good
97
3
0
0
0
Expert2
Low
Moderate
97
3
0
0
0
Expert2
Low
Poor
97
3
0
0
0
Expert2
Moderate
Good
97
3
0
0
0
Expert2
Moderate
Moderate
3
47
47
3
0
Expert2
Moderate
Poor
0
16
68
16
0
Expert2
High
Good
16
68
16
0
0
Expert2
High
Moderate
0
3
47
47
3
Expert2
High
Poor
0
0
0
3
97
Expert4
Low
Good
97
3
0
0
0
Expert4
Low
Moderate
97
3
0
0
0
Expert4
Low
Poor
97
3
0
0
0
Expert4
Moderate
Good
97
3
0
0
0
Expert4
Moderate
Moderate
50
47
3
0
0
Expert4
Moderate
Poor
0
3
47
47
3
Expert4
High
Good
74
26
0
0
0
Expert4
High
Moderate
0
50
50
0
0
Expert4
High
Poor
0
0
3
95
2
Expert5
Low
Good
97
3
0
0
0
Expert5
Low
Moderate
97
3
0
0
0
Expert5
Low
Poor
97
3
0
0
0
Expert5
Moderate
Good
97
3
0
0
0
Expert5
Moderate
Moderate
26
48
23
3
0
Expert5
Moderate
Poor
3
23
48
23
3
Expert5
High
Good
3
47
47
3
0
Expert5
High
Moderate
0
3
47
47
3
Expert5
High
Poor
0
0
0
3
97
Expert6
Low
Good
97
3
0
0
0
Expert6
Low
Moderate
97
3
0
0
0
Expert6
Low
Poor
97
3
0
0
0
Expert6
Moderate
Good
97
3
0
0
0
Expert6
Moderate
Moderate
86
14
0
0
0
Expert6
Moderate
Poor
74
26
0
0
0
Expert6
High
Good
74
26
0
0
0
Expert6
High
Moderate
3
47
47
3
0
Expert6
High
Poor
0
0
0
3
97
(D1) Expert
194
Intensity of
control
Frequency
of control
Predator
presence
Expert1
Low
Low
Expert1
Low
Expert1
Expert1
(D2) Expert
Disturbance stress (% probability)
Low
Moderate
High
Low
79
21
0
Low
High
56
44
0
Low
High
Low
3
86
11
Low
High
High
0
71
29
Expert1
High
Low
Low
33
67
0
Expert1
High
Low
High
13
85
2
Expert1
High
High
Low
0
24
76
Expert1
High
High
High
0
24
76
Expert2
Low
Low
Low
100
0
0
Expert2
Low
Low
High
50
48
2
Expert2
Low
High
Low
0
50
50
Expert2
Low
High
High
0
49
51
Expert2
High
Low
Low
50
50
0
Expert2
High
Low
High
25
73
2
Expert2
High
High
Low
0
0
100
Expert2
High
High
High
0
0
100
Expert4
Low
Low
Low
97
3
0
Expert4
Low
Low
High
97
3
0
Expert4
Low
High
Low
50
48
2
Expert4
Low
High
High
50
48
2
Expert4
High
Low
Low
2
48
50
Expert4
High
Low
High
2
48
50
Expert4
High
High
Low
0
3
97
Expert4
High
High
High
0
3
97
Expert5
Low
Low
Low
97
3
0
Expert5
Low
Low
High
97
3
0
Expert5
Low
High
Low
0
50
50
Expert5
Low
High
High
0
50
50
Expert5
High
Low
Low
50
50
0
Expert5
High
Low
High
50
50
0
Expert5
High
High
Low
0
3
97
Expert5
High
High
High
0
3
97
Expert6
Low
Low
Low
97
3
0
Expert6
Low
Low
High
50
48
2
Expert6
Low
High
Low
2
48
50
Expert6
Low
High
High
1
48
51
Expert6
High
Low
Low
97
3
0
Expert6
High
Low
High
50
48
2
Expert6
High
High
Low
0
3
97
Expert6
High
High
High
0
3
97
195
196
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
197
12
7
5
3
0
7
5
3
1
0
3
2
1
0
0
0
0
0
0
0
0
0
0
0
0
2
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
62
58
51
42
3
58
51
42
33
3
46
37
28
21
3
18
12
8
5
3
0
0
0
0
0
33
25
18
12
3
25
18
12
8
3
18
12
8
5
3
9
5
3
2
3
0
0
0
0
0
25
34
42
51
23
34
42
51
58
23
48
56
61
64
23
64
62
58
51
23
0
0
0
0
0
58
62
64
62
23
62
64
62
58
23
64
62
58
51
23
60
54
46
37
23
0
0
0
0
0
1
1
2
4
49
1
2
4
8
49
3
5
10
15
49
18
25
33
42
49
26
26
26
26
26
7
12
18
25
49
12
18
25
33
49
18
25
33
42
49
30
39
48
56
49
26
26
26
26
26
0
0
0
0
25
0
0
0
0
25
0
0
0
0
25
0
1
1
2
25
74
74
74
74
74
0
0
0
1
25
0
0
1
1
25
0
1
1
2
25
1
2
3
5
25
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
198
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
44
34
25
15
0
28
20
14
7
0
7
5
3
1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
52
58
63
65
3
62
65
64
59
3
58
51
42
33
3
25
18
12
8
3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
4
8
12
20
23
10
15
22
33
23
34
42
51
58
23
62
64
62
58
23
0
0
0
0
0
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
0
0
0
0
49
0
0
0
1
49
1
2
4
8
49
12
18
25
33
49
26
26
26
26
26
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
0
0
0
0
25
0
0
0
0
25
0
0
0
0
25
0
0
1
1
25
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
199
28
20
14
7
0
18
12
7
3
0
5
3
1
1
0
0
0
0
0
0
0
0
0
0
0
7
5
3
1
0
5
3
1
1
0
2
1
0
0
0
0
0
0
0
0
0
0
0
0
0
62
65
64
59
3
65
64
58
47
3
51
42
33
24
3
18
12
8
5
3
0
0
0
0
0
58
51
42
33
3
51
42
33
24
3
37
28
21
14
3
12
8
5
3
3
0
0
0
0
0
10
15
22
33
23
17
24
34
47
23
42
51
58
63
23
64
62
58
51
23
0
0
0
0
0
34
42
51
58
23
42
51
58
63
23
56
61
64
64
23
62
58
51
42
23
0
0
0
0
0
0
0
0
1
49
0
0
1
3
49
2
4
8
12
49
18
25
33
42
49
26
26
26
26
26
1
2
4
8
49
2
4
8
12
49
5
10
15
22
49
25
33
42
51
49
26
26
26
26
26
0
0
0
0
25
0
0
0
0
25
0
0
0
0
25
0
1
1
2
25
74
74
74
74
74
0
0
0
0
25
0
0
0
0
25
0
0
0
0
25
1
1
2
4
25
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
200
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
25
18
12
8
3
18
12
8
5
3
12
8
5
3
3
5
3
2
1
3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
62
64
62
58
23
64
62
58
51
23
62
58
51
42
23
54
46
37
28
23
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
12
18
25
33
49
18
25
33
42
49
25
33
42
51
49
39
48
56
62
49
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
0
0
1
1
25
0
1
1
2
25
1
1
2
4
25
2
3
5
9
25
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
201
12
7
5
3
0
7
5
3
1
0
3
2
1
0
0
0
0
0
0
0
0
0
0
0
0
8
5
3
1
0
5
3
1
1
0
2
1
0
0
0
0
0
0
0
0
0
0
0
0
0
62
58
51
42
3
58
51
42
33
3
46
37
28
21
3
18
12
8
5
3
0
0
0
0
0
58
51
42
33
3
51
42
33
24
3
37
28
21
14
3
12
8
5
3
3
0
0
0
0
0
25
34
42
51
23
34
42
51
58
23
48
56
61
64
23
64
62
58
51
23
0
0
0
0
0
34
42
51
58
23
42
51
58
63
23
56
61
64
64
23
62
58
51
42
23
0
0
0
0
0
1
1
2
4
49
1
2
4
8
49
3
5
10
15
49
18
25
33
42
49
26
26
26
26
26
0
2
4
8
49
2
4
8
12
49
5
10
15
22
49
25
33
42
51
49
26
26
26
26
26
0
0
0
0
25
0
0
0
0
25
0
0
0
0
25
0
1
1
2
25
74
74
74
74
74
0
0
0
0
25
0
0
0
0
25
0
0
0
0
25
1
1
2
4
25
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
202
3
2
1
0
0
2
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
46
37
28
21
3
37
28
21
14
3
25
18
12
7
3
8
5
3
1
3
0
0
0
0
0
18
12
8
5
3
12
8
5
3
3
8
5
3
1
3
3
2
1
0
0
0
0
0
0
0
48
56
61
64
23
56
61
64
64
23
63
65
64
58
23
58
51
42
33
23
0
0
0
0
0
64
62
58
51
23
62
58
51
42
23
58
51
42
33
23
46
37
28
20
14
0
0
0
0
0
3
5
10
15
49
5
10
15
22
49
11
17
24
34
49
33
42
51
58
49
26
26
26
26
26
18
25
33
42
49
25
33
42
51
49
33
42
51
58
49
48
56
62
65
64
26
26
26
26
26
0
0
0
0
25
0
0
0
0
25
0
0
0
1
25
1
2
4
8
25
74
74
74
74
74
0
1
1
2
25
1
1
2
4
25
1
2
4
8
25
3
5
9
15
22
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
203
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
33
25
18
12
3
25
18
12
8
3
18
12
8
5
3
9
5
3
2
3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
58
62
64
62
23
62
64
62
58
23
64
62
58
51
23
60
54
46
37
23
0
0
0
0
0
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
7
12
18
25
49
12
18
25
33
49
18
25
33
42
49
30
39
48
56
49
26
26
26
26
26
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
0
0
0
1
25
0
0
1
1
25
0
1
1
2
25
1
2
3
5
25
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
204
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
25
18
12
8
3
18
12
8
5
3
12
8
5
3
3
5
3
2
1
3
0
0
0
0
0
18
12
8
5
3
12
8
5
3
3
8
5
3
1
3
3
2
1
0
0
0
0
0
0
0
62
64
62
58
23
64
62
58
51
23
62
58
51
42
23
54
46
37
28
23
0
0
0
0
0
64
62
58
51
23
62
58
51
42
23
58
51
42
33
23
46
37
28
20
14
0
0
0
0
0
12
18
25
33
49
18
25
33
42
49
25
33
42
51
49
39
48
56
62
49
26
26
26
26
26
18
25
33
42
49
25
33
42
51
49
33
42
51
58
49
48
56
62
65
64
26
26
26
26
26
0
0
1
1
25
0
1
1
2
25
1
1
2
4
25
2
3
5
9
25
74
74
74
74
74
0
1
1
2
25
1
1
2
4
25
1
2
4
8
25
3
5
9
15
22
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
205
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
9
5
3
2
3
5
3
2
1
3
3
2
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
60
54
46
37
23
54
46
37
28
23
46
37
28
20
14
34
25
18
12
7
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
30
39
48
56
49
39
48
56
62
49
48
56
62
65
64
58
63
65
64
59
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
23
1
2
3
5
25
2
3
5
9
25
3
5
9
15
22
7
11
17
24
34
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
77
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Good
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
206
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
207
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
23
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
77
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Expert1
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Very good
Very good
Very good
Very good
Good
Good
Good
Good
Good
Moderate
Moderate
Moderate
Moderate
Moderate
Poor
Poor
Poor
Poor
Poor
Very poor
Very poor
Very poor
Very poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
Very good
Good
Moderate
Poor
Very poor
208
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
26
23
26
26
26
21
15
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
74
77
74
74
74
79
85
Appendix 4.2 Model variables – spatial data (supplements Table 4.2)
Appendix Table 4.2.1 Classification of Present Major Vegetation Subgroups (MVS V.4.1 *) into food
quality under good and poor accessibility of below-ground protein sources.
MVS
MVS Name
No.
Type
Food quality (access)
(good)
(poor)
2
Tropical or sub-tropical rainforest
Rainforest
High
High
4
Eucalyptus open forests, shrubby understorey
Woodland
Low
Low
5
Eucalyptus open forests, grassy understorey
Woodland
Moderate
Low
6
Warm Temperate Rainforest
Rainforest
Moderate
Moderate
7
Tropical Eucalyptus forest & woodlands, tall annual grassy understorey
Woodland
Low
Low
8
Eucalyptus woodlands, shrubby understorey
Woodland
Low
Low
9
Eucalyptus woodlands, tussock grass understorey
Woodland
Low
Low
10
Eucalyptus woodlands, hummock grass understorey
Woodland
Low
Low
11
Tropical mixed spp forests & woodlands
Woodland
Moderate
Low
12
Callitris forests & woodlands
Woodland
Low
Low
13
Brigalow forests & woodlands
Woodland
Low
Low
14
Other Acacia forests & woodlands
Woodland
Low
Low
15
Melaleuca open forests & woodlands
Woodland
Moderate
Low
16
Other forests & woodlands
Woodland
Low
Low
18
Eucalyptus low open woodlands with hummock grass
Woodland
Low
Low
19
Eucalyptus low open woodlands with tussock grass
Woodland
Moderate
Low
20
Mulga woodlands +/- tussock grass +/- forbs
Shrubland
Low
Low
21
Other Acacia tall open shrublands & [tall] shrublands
Shrubland
Low
Low
23
Acacia open woodlands & shrublands with hummock grass
Shrubland
Low
Low
24
Acacia open woodlands & shrublands +/- tussock grass
Shrubland
Low
Low
25
Acacia open woodlands & sparse shrublands, shrubby understorey
Shrubland
Low
Low
26
Casuarina & Allocasuarina forests & woodlands
Woodland
Low
Low
27
Mallee with hummock grass
Shrubland
Low
Low
28
Low closed forest or tall closed shrublands
Shrubland
Low
Low
30
Heath
Shrubland
Low
Low
31
Saltbush & Bluebush shrublands
Shrubland
Low
Low
32
Other shrublands
Shrubland
Low
Low
33
Hummock grasslands
Grassland
Low
Low
34
Mitchell grass tussock grasslands
Grassland
Moderate
Low
35
Blue grass & tall bunch grass tussock grasslands
Grassland
High
Low
37
Other tussock grasslands
Grassland
Moderate
Low
38
Wet tussock grassland with herbs, sedges or rushes, herblands or ferns
Wetland
High
High
39
Mixed chenopod, samphire +/- forbs
Shrubland
Low
Low
40
Mangroves
Mangroves
High
High
41
Saline or brackish sedgelands or grasslands
Wetland
High
High
42
Naturally bare, sand, rock, claypan, mudflat
None
Low
Low
209
43
Salt lakes & lagoons
Wetland
Moderate
Moderate
44
Freshwater, dams, lakes, lagoons or aquatic plants
Wetland
Moderate
Moderate
45
Mulga open woodlands & sparse shrublands +/- tussock grass
Shrubland
Low
Low
46
Sea, estuaries (includes seagrass)
None
Low
Low
47
Eucalyptus open woodlands with shrubby understorey
Woodland
Low
Low
48
Eucalyptus open woodlands, grassy understorey
Woodland
Moderate
Low
49
Melaleuca shrublands & open shrublands
Shrubland
Low
Low
51
Mulga woodlands & shrublands with hummock grass
Shrubland
Low
Low
52
Mulga open woodlands & sparse shrublands with hummock grass
Shrubland
Low
Low
53
Eucalyptus low open woodlands, shrubby understorey
Woodland
Low
Low
56
Eucalyptus open woodlands, chenopod or samphire understorey
Woodland
Low
Low
57
Lignum shrublands & wetlands
Wetland
Moderate
Moderate
59
Eucalyptus woodlands, ferns, herbs, sedges, rushes, wet tussock grasses
Woodland
High
Moderate
60
Eucalyptus open forests, ferns, herbs, sedges, rushes, wet tussock grasses Woodland
High
Moderate
62
Dry rainforest or vine thickets
Rainforest
Moderate
Low
63
Sedgelands, rushs or reeds
Wetland
High
High
64
Other grasslands
Grassland
Moderate
Low
70
Callitris open woodlands
Woodland
Low
Low
71
Casuarina & Allocasuarina open woodlands, tussock grass understorey
Woodland
Low
Low
74
Casuarina & Allocasuarina open woodlands, shrubby understorey
Woodland
Low
Low
75
Melaleuca open woodlands
Woodland
Moderate
Low
79
Other open Woodlands
Woodland
Low
Low
80
Other sparse shrublands & sparse heathlands
Shrubland
Low
Low
90
Regrowth or modified forests & woodlands
Unknown
Low
Low
92
Regrowth or modified graminoids
Unknown
Low
Low
96
Unclassified Forest
Woodland
Low
Low
97
Unclassified native vegetation
Unknown
Low
Low
98
Cleared, non-native vegetation, buildings
None
Low
Low
99
Unknown/No data
Unknown
Low
Low
* Department of the Environment and Water Resources (2007). Australia’s native vegetation: a summary of Australia’s
major vegetation groups, 2007. Commonwealth of Australia, Canberra.
210
Appendix Table 4.2.2 Classification of Australian Land Use and Management (ALUM V.7 *) primary
(and secondary / tertiary where applicable) classes into food quality under good and poor accessibility
of below-ground protein sources.
Food quality (access)
Primary (secondary, tertiary) ALUM class
(good)
(poor)
1 Conservation and natural environments
Low
Low
2 Production from relatively natural environments
Low
Low
3 Production from dryland agriculture and plantations
3.1 Plantation forestry, 3.6 Land in transition
Low
Low
3 Production from dryland agriculture and plantations
3.2 Grazing modified pastures
Moderate
Low
3 Production from dryland agriculture and plantations
3.3 Cropping
High
Moderate
3 Production from dryland agriculture and plantations
3.4 Perennial horticulture, 3.5 Seasonal horticulture
High
High
4 Production from irrigated agriculture and plantations
4.1 Irrigated plantation forestry, 4.5.5 Irrigated turf farming & 4.6 Irrigated land
in transition
Low
Low
Moderate
Moderate
4 Production from irrigated agriculture and plantations
4.3 Irrigated cropping, 4.4 Irrigated perennial horticulture, 4.5 Irrigated seasonal
horticulture (except 4.5.5)
High
High
5 Intensive uses
Low
Low
6 Water (except 6.3.0 & 6.3.1)
Low
Low
Moderate
Moderate
4 Production from irrigated agriculture and plantations
4.2 Grazing irrigated modified pastures
6 Water
6.3.0 River, 6.3.1 River - conservation
* Australian Bureau of Agricultural and Resource Economics and Sciences (2011). Guidelines for land use mapping in
Australia: principles, procedures and definitions, fourth edition. Commonwealth of Australia, Canberra.
211
Appendix Table 4.2.3 Classification of Present Major Vegetation Groups (MVG V.4.1 *) into dense
vegetation cover.
MVG No.
MVG Name
Type
Cover
1
Rainforests and Vine Thickets
Rainforest
Good
2
Eucalypt Tall Open Forests
Woodland
Moderate
3
Eucalypt Open Forests
Woodland
Moderate
4
Eucalypt Low Open Forests
Woodland
Moderate
5
Eucalypt Woodlands
Woodland
Moderate
6
Acacia Forests and Woodlands
Woodland
Moderate
7
Callitris Forests and Woodlands
Woodland
Moderate
8
Casuarina Forests and Woodlands
Woodland
Moderate
9
Melaleuca Forests and Woodlands
Woodland
Moderate
10
Other Forests and Woodlands
Woodland
Moderate
11
Eucalypt Open Woodlands
Woodland
Poor
12
Tropical Eucalypt Woodlands/Grasslands
Woodland
Moderate
13
Acacia Open Woodlands
Shrubland
Poor
14
Mallee Woodlands and Shrublands
Shrubland
Moderate
15
Low Closed Forests and Tall Closed Shrublands
Shrubland
Good
16
Acacia Shrublands
Shrubland
Moderate
17
Other Shrublands
Shrubland
Moderate
18
Heathlands
Shrubland
Good
19
Tussock Grasslands
Grassland
Poor
20
Hummock Grasslands
Grassland
Poor
21
Other Grasslands, Herblands, Sedgelands and Rushlands
Wetland
Poor
22
Chenopod Shrublands, Samphire Shrublands and Forblands
Shrubland
Poor
23
Mangroves
Mangroves
Good
24
Inland aquatic - freshwater, salt lakes, lagoons
Wetland
Poor
25
Cleared, non-native vegetation, buildings
None/modified
Poor
26
Unclassified native vegetation
None/modified
Poor
27
Naturally bare - sand, rock, claypan, mudflat
None/modified
Poor
28
Sea and estuaries
None/modified
Poor
29
Regrowth, modified native vegetation
None/modified
Poor
30
Unclassified Forest
Woodland
Moderate
31
Other Open Woodlands
Woodland
Poor
32
Mallee Open Woodlands and Sparse Mallee Shrublands
Shrubland
Poor
99
Unknown/no data
None/modified
Poor
* Department of the Environment and Water Resources (2007). Australia’s native vegetation: a summary of Australia’s
major vegetation groups, 2007. Commonwealth of Australia, Canberra.
212
Appendix Table 4.2.4 Classification of Australian Land Use and Management (ALUM V.7 *) primary
(and secondary / tertiary where applicable) classes into dense vegetation cover.
Primary (secondary, tertiary) ALUM class
Cover
1 Conservation and natural environments
Poor
2 Production from relatively natural environments
2.1 Grazing natural vegetation
Poor
2 Production from relatively natural environments
2.2 Production forestry
Moderate
3 Production from dryland agriculture and plantations
3.1 Plantation forestry, 3.3.5 Sugar, 3.4 Perennial horticulture (except 3.4.0)
Moderate
3 Production from dryland agriculture and plantations
3.2 Grazing modified pastures, 3.3 Cropping (except 3.3.5), 3.4.0 Perennial horticulture,
3.5 Seasonal horticulture, 3.6 Land in transition
Poor
4 Production from irrigated agriculture and plantations
4.1 Irrigated plantation forestry, 4.3.5 Irrigated sugar, 4.4 Irrigated perennial horticulture
(except 4.4.0 & 4.4.7), 4.5.1 Irrigated seasonal fruits
Moderate
4 Production from irrigated agriculture and plantations
4.2 Grazing irrigated modified pastures, 4.3 Irrigated cropping (except 4.3.5), 4.4.0
Irrigated perennial horticulture, 4.4.7 Irrigated perennial vegetables & herbs, 4.5 Irrigated
seasonal horticulture (except 4.5.1), 4.6 Irrigated land in transition
Poor
5 Intensive uses
Poor
6 Water
Poor
* Australian Bureau of Agricultural and Resource Economics and Sciences (2011). Guidelines for land use mapping in
Australia: principles, procedures and definitions, fourth edition. Commonwealth of Australia, Canberra.
213
Appendix 4.3 Methodology – spatial pattern suitability analysis
This page is intentionally blank.
214
S3 Appendix. Rcode spatial pattern suitability
analysis (PDF)
Jens G. Froese
23 February 2017
This document provides a detailed, reproducible description of the spatial pattern suitability analysis methodology. It is
Supporting Information (S3) to the manuscript:
Froese JG, Smith CS, Durr PA, McAlpine CA, van Klinken RD. Modelling seasonal habitat
suitability for wide-ranging species: invasive wild pigs in northern Australia. Submitted to
PLoS ONE.
It is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. A
.zip file containing data inputs to reproduce analyses can be downloaded from Dryad (http://dx.doi.org/10.5061
/dryad.v103v (http://dx.doi.org/10.5061/dryad.v103v)). Some parts of the code used to print this document have been
surpressed to enhance readability. A generalized version of the code is available at URL https://github.com/jgfroese
/PATTSI (https://github.com/jgfroese/PATTSI).
Load required R packages
R package raster ([4]) and its dependencies are required for spatial pattern suitability analysis. Session information
incl. package versions are listed at the bottom of this document.
require(raster) # for all analyses of raster objects incl. moving window analysis
1. Expert elicitation
For each habitat variable (i.e. its modelled resource quality indices xr ), we elicited a distance-dependent responseto-pattern curve ( fDr ) from each individual expert (see manuscript Fig 3). These curves followed a step-wise pattern,
because we discretised both:
distance into five equal distance bands (“very close”, “close”, “medium”, “far” and “very far”), relative to each
expert’s defined mobility threshold (i.e. 1km, 2km or 3km).
resource suitability indices ( SIr ) into five equal classes (“very good (80-100)”, “good (60-80)”, “moderate
(40-60)”, “poor (20-40)” and “very poor (0-20)”), see manuscript S2.1 Table.
We asked experts to relate each distance band to a corresponding suitability class under the assumption that other
variables do not constrain suitability. To fill two elicitation gaps (expert 3 did not define fDDisturbance and expert 6
did not define fDHeat ) we
applied all other experts’ fDr to the missing expert’s defined mobility threshold (Expert3 = 2km, Expert6 = 3km)
computed the average fDr and used it for the missing expert’s model
To derive distance weights for computation of 2. Resource suitability indices, Step 4, we used the mid-points of
elicited suitability index classes divided by 100 (e.g. class “moderate (40-60)” = SIr 50 = weight 0.5). For class “very
good (80-100)” we did not use the mid-point SIr = 90 but assigned SIr = 100 (= weight 1.0) to avoid unintended
distance penalties (i.e. an adjacent resource of quality xr = 60 should compute as distance-weighted suitability
SIr = 60 (if weight is 1.0) and not SIr = 54 (if weight is 0.9)).
Figure S3.1 Step-wise distance-dependent response-to-pattern curve for habitat variable “food” elicited from expert 4
(A) and distance weights derived for computation of food suitability indices (B).
2. Resource suitability indices
Goal
Focal pixel resource suitability index ( SIr ) depends on the distance of a (numerical) habitat variable.
Method:
Generate a circular moving window where each position is weighted by its distance from the focal pixel
(radius/weights derived from 1. Expert elicitation).
compute the focal pixel SIr as the highest weighted value ( xr ) of a habitat variable within this moving window.
Step 1
Define a function that returns a circular matrix of given radius and resolution and assigns value 1 if matrix position <=
radius and value NA if matrix position > radius (Source: [5]).
make_circ_filter <- function(radius, res){
circ_filter <- matrix(NA, nrow=1+(2*radius/res), ncol=1+(2*radius/res))
dimnames(circ_filter)[[1]] <- seq(-radius, radius, by=res)
dimnames(circ_filter)[[2]] <- seq(-radius, radius, by=res)
sweeper <- function(mat){
for(row in 1:nrow(mat)){
for(col in 1:ncol(mat)){
dist <- sqrt((as.numeric(dimnames(mat)[[1]])[row])^2 +
(as.numeric(dimnames(mat)[[1]])[col])^2)
if(dist<=radius) {mat[row, col]<-1}
}
}
return(mat)
}
out <- sweeper(circ_filter)
return(out)
}
Step 2
Apply function to generate five matrices with different radii (= distance bands), relative to each expert’s defined mobility
threshold (i.e. 1km for Expert1, 2km for Experts 3/5 and 3km for Experts 2/4/6).
res <- 1 # resolution (= pixel size, e.g. 100m)
mr.1 <- 10 # matrix radius (= mobility threshold, must be multiple of res, e.g 1km = 10 x 10
0m)
m.1 <- make_circ_filter(mr.1, res)
m.2 <- make_circ_filter((mr.1/5)*4, res)
m.3 <- make_circ_filter((mr.1/5)*3, res)
m.4 <- make_circ_filter((mr.1/5)*2, res)
m.5 <- make_circ_filter((mr.1/5), res)
Replace value==1 with unique temp value in ascending order from largest to smallest matrix.
m.1[m.1
m.2[m.2
m.3[m.3
m.4[m.4
m.5[m.5
==
==
==
==
==
1]
1]
1]
1]
1]
<<<<<-
1
2
3
4
5
Step 3
Combine the five matrices into one (two at a time starting with the smallest):
a.5 <- array(NA, dim(m.4), dimnames(m.4)) # create temp array of size = larger matrix
a.5[rownames(m.5), colnames(m.5)] <- m.5 # ... with values = smaller matrix
m.4 <- pmax(m.4, a.5, na.rm = TRUE) # combine values: larger matrix + temp array
a.4 <- array(NA, dim(m.3), dimnames(m.3)) # repeat with: output + next-larger matrix
a.4[rownames(m.4), colnames(m.4)] <- m.4
m.3 <- pmax(m.3, a.4, na.rm = TRUE)
a.3 <- array(NA, dim(m.2), dimnames(m.2))
a.3[rownames(m.3), colnames(m.3)] <- m.3
m.2 <- pmax(m.2, a.3, na.rm = TRUE)
a.2 <- array(NA, dim(m.1), dimnames(m.1))
a.2[rownames(m.2), colnames(m.2)] <- m.2
m.band.1 <- pmax(m.1, a.2, na.rm = TRUE)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
-10
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
8
9
10
-10
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
1
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
-9
NA
NA
NA
NA
NA
NA
1
1
1
1
1
1
1
1
1
NA
NA
NA
NA
NA
NA
-8
NA
NA
NA
NA
1
1
1
1
1
1
2
1
1
1
1
1
1
NA
NA
NA
NA
-7
NA
NA
NA
1
1
1
1
2
2
2
2
2
2
2
1
1
1
1
NA
NA
NA
-6
NA
NA
1
1
1
2
2
2
2
2
3
2
2
2
2
2
1
1
1
NA
NA
-5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
NA NA NA NA NA 1 NA NA NA NA NA NA NA
NA 1 1 1 1 1 1 1 1 1 NA NA NA
1 1 1 1 1 2 1 1 1 1 1 1 NA
1 1 2 2 2 2 2 2 2 1 1 1 1
2 2 2 2 2 3 2 2 2 2 2 1 1
2 2 3 3 3 3 3 3 3 2 2 2 1
2 3 3 3 3 4 3 3 3 3 2 2 1
3 3 3 4 4 4 4 4 3 3 3 2 2
3 3 4 4 4 5 4 4 4 3 3 2 2
3 3 4 4 5 5 5 4 4 3 3 2 2
3 4 4 5 5 5 5 5 4 4 3 3 2
3 3 4 4 5 5 5 4 4 3 3 2 2
3 3 4 4 4 5 4 4 4 3 3 2 2
3 3 3 4 4 4 4 4 3 3 3 2 2
2 3 3 3 3 4 3 3 3 3 2 2 1
2 2 3 3 3 3 3 3 3 2 2 2 1
2 2 2 2 2 3 2 2 2 2 2 1 1
1 1 2 2 2 2 2 2 2 1 1 1 1
1 1 1 1 1 2 1 1 1 1 1 1 NA
NA 1 1 1 1 1 1 1 1 1 NA NA NA
NA NA NA NA NA 1 NA NA NA NA NA NA NA
Repeat steps 2 and 3 for mobility thresholds 2km and 3km.
8
NA
NA
NA
NA
1
1
1
1
1
1
2
1
1
1
1
1
1
NA
NA
NA
NA
9
NA
NA
NA
NA
NA
NA
1
1
1
1
1
1
1
1
1
NA
NA
NA
NA
NA
NA
10
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
1
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
mr.2 <- 20 # moving window radius 2km (= 20 x 100m)
mr.3 <- 30 # moving window radius 3km (= 30 x 100m)
...
m.band.2
m.band.3
Step 4
Replace temp values with expert-elicited weight for each distance band ( SIdata.zip file containing f_DFood.csv
and other weights derived from 1. Expert elicitation can be downloaded from Dryad (http://dx.doi.org/10.5061
/dryad.v103v (http://dx.doi.org/10.5061/dryad.v103v))).
csv.Food = read.csv("SIdata/S3Appendix/fD/f_DFood.csv")
csv.Food.E1 <- subset(csv.Food, Expert == 'Expert1') # for Expert1 use m.band.1
m.band.1[m.band.1 == 1] <- csv.Food.E1$X1
m.band.1[m.band.1 == 2] <- csv.Food.E1$X0.8
m.band.1[m.band.1 == 3] <- csv.Food.E1$X0.6
m.band.1[m.band.1 == 4] <- csv.Food.E1$X0.4
m.band.1[m.band.1 == 5] <- csv.Food.E1$X0.2
m.band.Food.E1 <- m.band.1
csv.Food.E2 <- subset(csv.Food, Expert == 'Expert2') # for Expert2 use m.band.3
m.band.3[m.band.3 == 1] <- csv.Food.E2$X3
m.band.3[m.band.3 == 2] <- csv.Food.E2$X2.4
m.band.3[m.band.3 == 3] <- csv.Food.E2$X1.8
m.band.3[m.band.3 == 4] <- csv.Food.E2$X1.2
m.band.3[m.band.3 == 5] <- csv.Food.E2$X0.6
m.band.Food.E2 <- m.band.3
csv.Food.E3 <- subset(csv.Food, Expert == 'Expert3') # for Expert3 use m.band.2
m.band.2[m.band.2 == 1] <- csv.Food.E3$X2
m.band.2[m.band.2 == 2] <- csv.Food.E3$X1.6
m.band.2[m.band.2 == 3] <- csv.Food.E3$X1.2
m.band.2[m.band.2 == 4] <- csv.Food.E3$X0.8
m.band.2[m.band.2 == 5] <- csv.Food.E3$X0.4
m.band.Food.E3 <- m.band.2
Step 5
Perform moving window analysis using function focal {raster} with parameters ( SIdata.zip file containing raster
layers with resource quality indices xr can be downloaded from Dryad (http://dx.doi.org/10.5061/dryad.v103v
(http://dx.doi.org/10.5061/dryad.v103v))).
r =
rce
w =
fun
raster("SIdata/S3Appendix/GIS/Food-quality-dry.tif") # raster layer with numerical resou
quality index, e.g. Food quality in dry season scenario
m.band.Food.E1 # moving window is banded weights matrix, e.g. Expert 1 f_Dfood
= max # focal pixel takes highest weighted resource quality index within moving window
WARNING! The following process may take several hours depending on the size of r and w.
r.f <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padValue = NA) # na.rm = TRUE ignores NoD
ata
r.m <- mask(r.f, r) # extract by r to remove padded edges
writeRaster(r.m, filename = paste("SIdata/S3Appendix/out/Food-SI-dry_E1.tif", sep="")) # sav
e ouput raster
Repeat steps 4 and 5 for all four habitat variables and six experts in two seasonal scenarios:
Water wet/dry = 12 runs
Food wet/dry = 12 runs
Heat wet/dry = 12 runs
Disturbance global scenario = 6 runs
References
[1] Allaire, J.J. et al. 2016. Package ‘rmarkdown’: dynamic documents for R. URL http://rmarkdown.rstudio.com/
(http://rmarkdown.rstudio.com/).
[2] Xie, Y. 2016. Package ‘knitr’: a general-purpose package for dynamic report generation in R. URL http://yihui.name
/knitr/ (http://yihui.name/knitr/).
[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for Statistical Computing,
Vienna, Austria. URL http://www.R-project.org/ (http://www.R-project.org/).
[4] Hijmans, R.J. 2015. Package ‘raster’: geographic data analysis and modeling. URL http://cran.r-project.org
/web/packages/raster/ (http://cran.r-project.org/web/packages/raster/).
[5] Scroggie, M. 2012. Applying a circular moving window filter to raster data in R. URL https://scrogster.wordpress.com
/2012/10/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/ (https://scrogster.wordpress.com/2012/10
/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/).
Session information
## Session info --------------------------------------------------------------
##
##
##
##
##
##
##
##
setting
version
system
ui
language
collate
tz
date
value
R version 3.1.3 (2015-03-09)
x86_64, mingw32
RTerm
(EN)
English_Australia.1252
Australia/Brisbane
2017-02-23
## Packages ------------------------------------------------------------------
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
package
* version
devtools
1.10.0
digest
0.6.8
evaluate
0.8
formatR
1.2.1
htmltools
0.3
knitr
1.12.3
lattice
0.20-30
magrittr
1.5
memoise
1.0.0
raster
* 2.4-20
Rcpp
0.12.1
rgdal
1.1-1
rmarkdown
0.9.2
sp
* 1.2-1
stringi
1.0-1
stringr
1.0.0
yaml
2.1.13
date
2016-01-23
2014-12-31
2015-09-18
2015-09-18
2015-12-29
2016-01-22
2015-02-22
2014-11-22
2016-01-29
2015-09-08
2015-09-10
2015-11-02
2016-01-01
2015-10-18
2015-10-22
2015-04-30
2014-06-12
source
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
Appendix 4.4 Methodology – validation
This page is intentionally blank.
220
S4 Appendix. Rcode validation (PDF)
Jens G. Froese
17 April 2017
This document provides a detailed, reproducible description of the methodology used to evaluate and validate performance of habitat
suitability models using the Continuous Boyce Index (CBI). It is Supporting Information (S4 Appendix) to the manuscript:
Froese JG, Smith CS, Durr PA, McAlpine CA, van Klinken RD. Modelling seasonal habitat suitability
for wide-ranging species: invasive wild pigs in northern Australia. Submitted to PLoS ONE.
The document is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. A
.zip file containing data inputs to reproduce analyses can be downloaded from Dryad
( http://dx.doi.org/10.5061/dryad.v103v ). Some parts of the code used to print this document have been suppressed to
enhance readability. A generalized version of the code is available at URL https://github.com/jgfroese/HSI-CBI-validation .
Load required R packages
R packages ecospat ([4]), data.table ([5]) and zoo ([6]) and their dependencies are required for HSI-CBI-validation. Session
information incl. package versions are listed at the bottom of this document.
require(data.table) # for function `setnames`
require(ecospat) # for function `ecospat.boyce`
require(zoo) # for plotting function `na.fill`
1. Data preparation
Step 1
This R script requires two .TXT files for each model / validation data combination, which have to be prepared in package {raster}
or alternative GIS software as described below. These files can be downloaded from Dryad
( http://dx.doi.org/10.5061/dryad.v103v ).
1. Expected HSI across validation background:
define validation backgrounds ( Balkanu = BLKW , Lakefield = LKNP , Oyala Thumotang = OTNP , NAQS , ALA see manuscript Table 2)
mask raster layers of individual expert models and an average model by each validation background
export raster attribute tables to .TXT with 3 columns: [ID], [HSI], [pixel count]
2. Predicted HSI at wild pig presence records:
convert presence records of each validation data set into raster layer
combine masked raster layer of each expert/average model with each set of presence records
export raster attribute tables to .TXT with 5 columns: [ID], [Value], [pixel count], [HSI], [number of presence records per
pixel]
Step 2
Compute model-predicted HSI at wild pig presence records.
First, read .TXT files for each model (Expert1, Expert2, …, averaged) / validation presences (BLKW, LKNP, OTNP, NAQS, ALA)
combination as data frame,
pred.E1.wetBLKW = read.csv("SIdata/S4Appendix/wet/Predicted/SiPred_wet_E1_BLKW1000.txt")
...
pred.E6.wetBLKW = read.csv("SIdata/S4Appendix/wet/Predicted/SiPred_wet_E6_BLKW1000.txt")
pred.av.wetBLKW = read.csv("SIdata/S4Appendix/wet/Predicted/SiPred_wet_Eall_BLKW1000.txt")
pred.E1.dryBLKW = read.csv("SIdata/S4Appendix/dry/Predicted/SiPred_dry_E1_BLKW1000.txt")
...
pred.E6.dryBLKW = read.csv("SIdata/S4Appendix/dry/Predicted/SiPred_dry_E6_BLKW1000.txt")
pred.av.dryBLKW = read.csv("SIdata/S4Appendix/dry/Predicted/SiPred_dry_Eall_BLKW1000.txt")
...
pred.av.wetLKNP = read.csv("SIdata/S4Appendix/wet/Predicted/SiPred_wet_Eall_LKNP1000.txt")
...
pred.av.dryLKNP = read.csv("SIdata/S4Appendix/dry/Predicted/SiPred_dry_Eall_LKNP1000.txt")
...
pred.av.wetOTNP = read.csv("SIdata/S4Appendix/wet/Predicted/SiPred_wet_Eall_OTNP1000.txt")
...
pred.av.dryNAQS = read.csv("SIdata/S4Appendix/dry/Predicted/SiPred_dry_Eall_NAQS1000.txt")
...
pred.av.wetALA = read.csv("SIdata/S4Appendix/wet/Predicted/SiPred_wet_Eall_ALA1000.txt")
...
pred.av.dryALA = read.csv("SIdata/S4Appendix/dry/Predicted/SiPred_dry_Eall_ALA1000.txt")
and combine all data frames in a list for faster analysis.
pred.list <- list (pred.E1.wetBLKW, pred.E2.wetBLKW, pred.E3.wetBLKW, pred.E4.wetBLKW, pred.E5.wetBLKW,
pred.E6.wetBLKW, pred.av.wetBLKW, pred.E1.dryBLKW, pred.E2.dryBLKW, pred.E3.dryBLKW, pred.E4.dryBLKW, p
red.E5.dryBLKW, pred.E6.dryBLKW, pred.av.dryBLKW, pred.E1.wetLKNP, pred.E2.wetLKNP, pred.E3.wetLKNP, pr
ed.E4.wetLKNP, pred.E5.wetLKNP, pred.E6.wetLKNP, pred.av.wetLKNP, pred.E1.dryLKNP, pred.E2.dryLKNP, pre
d.E3.dryLKNP, pred.E4.dryLKNP, pred.E5.dryLKNP, pred.E6.dryLKNP, pred.av.dryLKNP, pred.E1.wetOTNP, pre
d.E2.wetOTNP, pred.E3.wetOTNP, pred.E4.wetOTNP,pred.E5.wetOTNP, pred.E6.wetOTNP, pred.av.wetOTNP, pred.
E1.dryNAQS, pred.E2.dryNAQS, pred.E3.dryNAQS, pred.E4.dryNAQS, pred.E5.dryNAQS, pred.E6.dryNAQS, pred.a
v.dryNAQS, pred.E1.wetALA, pred.E2.wetALA, pred.E3.wetALA, pred.E4.wetALA, pred.E5.wetALA, pred.E6.wetA
LA, pred.av.wetALA, pred.E1.dryALA, pred.E2.dryALA, pred.E3.dryALA, pred.E4.dryALA, pred.E5.dryALA, pre
d.E6.dryALA, pred.av.dryALA)
n.list <- 56 # the number of data frames in your list
Then, homogenise the five column names for all data frames,
for (i in seq_along(pred.list)) {
setnames(pred.list[[i]], c("ID", "Value", "Pixelcount", "HSI", "Presences"))
}
and calculate the total number of presence records per HSI value (one pixel may contain multiple records).
pred.sum.list <- vector("list", n.list)
for (i in seq_along(pred.list)) {
pred.sum.list[[i]] <- aggregate(cbind(Pixelcount*Presences)~HSI, data = pred.list[[i]], sum)
}
Finally, add descriptive column names to the new list of data frames,
for (i in seq_along(pred.sum.list)) {
setnames(pred.sum.list[[i]], c("HSI", "Presences"))
}
and convert it into a list of vectors (= HSI at feral pig presence records),
pred.v.list = vector("list", n.list)
for (i in seq_along(pred.sum.list)) {
pred.v = vector()
for (j in 1:length(pred.sum.list[[i]]$Presences)) {
for (k in 1:pred.sum.list[[i]][j, 2]) {
pred.v <-append(pred.v, pred.sum.list[[i]][j, 1])
}
}
pred.v.list[[i]] <- append(pred.v.list[[i]], pred.v)
}
e.g. HSI of model 1 at Balkanu presence records (first in list).
Step 3
Compute model-predicted HSI across validation backgrounds.
First, read .TXT files for each model (Expert1, Expert2, …, averaged) / validation background (BLKW, LKNP, OTNP, NAQS, ALA)
combination as data frame,
exp.E1.wetBLKW = read.csv("SIdata/S4Appendix/wet/Expected/SiExp_wet_E1_BLKW1000.txt")
...
exp.E6.wetBLKW = read.csv("SIdata/S4Appendix/wet/Expected/SiExp_wet_E6_BLKW1000.txt")
exp.av.wetBLKW = read.csv("SIdata/S4Appendix/wet/Expected/SiExp_wet_Eall_BLKW1000.txt")
exp.E1.dryBLKW = read.csv("SIdata/S4Appendix/dry/Expected/SiExp_dry_E1_BLKW1000.txt")
...
exp.E6.dryBLKW = read.csv("SIdata/S4Appendix/dry/Expected/SiExp_dry_E6_BLKW1000.txt")
exp.av.dryBLKW = read.csv("SIdata/S4Appendix/dry/Expected/SiExp_dry_Eall_BLKW1000.txt")
...
exp.av.wetLKNP = read.csv("SIdata/S4Appendix/wet/Expected/SiExp_wet_Eall_LKNP1000.txt")
...
exp.av.dryLKNP = read.csv("SIdata/S4Appendix/dry/Expected/SiExp_dry_Eall_LKNP1000.txt")
...
exp.av.wetOTNP = read.csv("SIdata/S4Appendix/wet/Expected/SiExp_wet_Eall_OTNP1000.txt")
...
exp.av.dryNAQS = read.csv("SIdata/S4Appendix/dry/Expected/SiExp_dry_Eall_NAQS1000.txt")
...
exp.av.wetALA = read.csv("SIdata/S4Appendix/wet/Expected/SiPred_wet_Eall_ALA1000.txt")
...
exp.av.dryALA = read.csv("SIdata/S4Appendix/dry/Expected/SiPred_dry_Eall_ALA1000.txt")
and combine all data frames in a list for faster analysis.
exp.list <- list (exp.E1.wetBLKW, exp.E2.wetBLKW, exp.E3.wetBLKW, exp.E4.wetBLKW, exp.E5.wetBLKW, exp.E
6.wetBLKW, exp.av.wetBLKW, exp.E1.dryBLKW, exp.E2.dryBLKW, exp.E3.dryBLKW, exp.E4.dryBLKW, exp.E5.dryBL
KW, exp.E6.dryBLKW, exp.av.dryBLKW, exp.E1.wetLKNP, exp.E2.wetLKNP, exp.E3.wetLKNP, exp.E4.wetLKNP, exp
.E5.wetLKNP, exp.E6.wetLKNP, exp.av.wetLKNP, exp.E1.dryLKNP, exp.E2.dryLKNP, exp.E3.dryLKNP, exp.E4.dry
LKNP, exp.E5.dryLKNP, exp.E6.dryLKNP, exp.av.dryLKNP, exp.E1.wetOTNP, exp.E2.wetOTNP, exp.E3.wetOTNP, e
xp.E4.wetOTNP, exp.E5.wetOTNP, exp.E6.wetOTNP, exp.av.wetOTNP, exp.E1.dryNAQS, exp.E2.dryNAQS, exp.E3.d
ryNAQS, exp.E4.dryNAQS, exp.E5.dryNAQS, exp.E6.dryNAQS, exp.av.dryNAQS, exp.E1.wetALA, exp.E2.wetALA, e
xp.E3.wetALA, exp.E4.wetALA, exp.E5.wetALA, exp.E6.wetALA, exp.av.wetALA, exp.E1.dryALA, exp.E2.dryALA,
exp.E3.dryALA, exp.E4.dryALA, exp.E5.dryALA, exp.E6.dryALA, exp.av.dryALA)
Then, homogenise the three column names for all data frames,
for (i in seq_along(exp.list)) {
setnames(exp.list[[i]], c("ID", "HSI", "Pixelcount"))
}
and convert it into a list of vectors (= HSI across validation backgrounds [potentially very large - see manuscript Table 1]).
exp.v.list = vector("list", n.list)
for (i in seq_along(exp.list)) {
exp.v = vector()
for (j in 1:length(exp.list[[i]]$Pixelcount)) {
for (k in 1:exp.list[[i]][j, 3]) {
exp.v <-append(exp.v, exp.list[[i]][j, 2])
}
}
exp.v.list[[i]] <- append(exp.v.list[[i]], exp.v)
}
2. Data analysis
Step 4
Apply function boyce {ecospat} to each model (Expert1, Expert2, …, averaged) / validation data (BLKW, LKNP, OTNP, NAQS,
ALA) combination with parameters:
exp.v.list[[i]] # Expected HSI across background ([[1]] = E1.wetBLKW, ... , [[56]] = av.dryALA)
pred.v.list[[i]] # Predicted HSI at presences ([[1]] = E1.wetBLKW, ... , [[56]] = av.dryALA)
nclass = 0 # defaults to
nts:
window.w = 10 # moving
res = 100 # resolution
PEplot = F # no PEplot
moving window (continuous, classification-independent) computation with argume
window width (i.e. 10 adjacent HSI values are considered in each computation)
factor (i.e. 100 computations across the total range of HSI)
is generated (customised plots - see manuscript Figure 6)
boyce.E1.wetBLKW <- ecospat.boyce(exp.v.list[[1]], pred.v.list[[1]], nclass, window.w, res, PEplot)
...
boyce.E6.wetBLKW <- ecospat.boyce(exp.v.list[[6]], pred.v.list[[6]], nclass, window.w, res, PEplot)
boyce.av.wetBLKW <- ecospat.boyce(exp.v.list[[7]], pred.v.list[[7]], nclass, window.w, res, PEplot)
boyce.E1.dryBLKW <- ecospat.boyce(exp.v.list[[8]], pred.v.list[[8]], nclass, window.w, res, PEplot)
...
boyce.E6.dryBLKW <- ecospat.boyce(exp.v.list[[13]], pred.v.list[[13]], nclass, window.w, res, PEplot)
boyce.av.dryBLKW <- ecospat.boyce(exp.v.list[[14]], pred.v.list[[14]], nclass, window.w, res, PEplot)
...
boyce.av.wetLKNP <- ecospat.boyce(exp.v.list[[21]], pred.v.list[[21]], nclass, window.w, res, PEplot)
...
boyce.av.dryLKNP <- ecospat.boyce(exp.v.list[[28]], pred.v.list[[28]], nclass, window.w, res, PEplot)
...
boyce.av.wetOTNP <- ecospat.boyce(exp.v.list[[35]], pred.v.list[[35]], nclass, window.w, res, PEplot)
...
boyce.av.dryNAQS <- ecospat.boyce(exp.v.list[[42]], pred.v.list[[42]], nclass, window.w, res, PEplot)
...
boyce.av.wetALA <- ecospat.boyce(exp.v.list[[49]], pred.v.list[[49]], nclass, window.w, res, PEplot)
...
boyce.av.dryALA <- ecospat.boyce(exp.v.list[[56]], pred.v.list[[56]], nclass, window.w, res, PEplot)
Step 5
Investigate results of CBI analysis.
Combine all results in list for faster analysis,
boyce.list <- list (boyce.E1.wetBLKW, boyce.E2.wetBLKW, boyce.E3.wetBLKW, boyce.E4.wetBLKW, boyce.E5.we
tBLKW, boyce.E6.wetBLKW, boyce.av.wetBLKW, boyce.E1.dryBLKW, boyce.E2.dryBLKW, boyce.E3.dryBLKW, boyce.
E4.dryBLKW, boyce.E5.dryBLKW, boyce.E6.dryBLKW, boyce.av.dryBLKW, boyce.E1.wetLKNP, boyce.E2.wetLKNP, b
oyce.E3.wetLKNP, boyce.E4.wetLKNP, boyce.E5.wetLKNP, boyce.E6.wetLKNP, boyce.av.wetLKNP, boyce.E1.dryLK
NP, boyce.E2.dryLKNP, boyce.E3.dryLKNP, boyce.E4.dryLKNP, boyce.E5.dryLKNP, boyce.E6.dryLKNP, boyce.av.
dryLKNP, boyce.E1.wetOTNP, boyce.E2.wetOTNP, boyce.E3.wetOTNP, boyce.E4.wetOTNP, boyce.E5.wetOTNP, boyc
e.E6.wetOTNP, boyce.av.wetOTNP, boyce.E1.dryNAQS, boyce.E2.dryNAQS, boyce.E3.dryNAQS, boyce.E4.dryNAQS,
boyce.E5.dryNAQS, boyce.E6.dryNAQS, boyce.av.dryNAQS, boyce.E1.wetALA, boyce.E2.wetALA, boyce.E3.wetALA
, boyce.E4.wetALA, boyce.E5.wetALA, boyce.E6.wetALA, boyce.av.wetALA, boyce.E1.dryALA, boyce.E2.dryALA,
boyce.E3.dryALA, boyce.E4.dryALA, boyce.E5.dryALA, boyce.E6.dryALA, boyce.av.dryALA)
and print CBI ( $Spearman.cor ) for all model / validation data combinations (see manuscript Table 3),
CBI.list = vector("list", n.list)
for (i in seq_along(boyce.list)) {
CBI.list[[i]] <- append(CBI.list[[i]], round(boyce.list[[i]]$Spearman.cor, digits = 2))
}
CBI.list
e.g. CBI of expert model 1 validated against wet season Balkanu presence records (first in list).
Step 6
Compute proportion of validation background expected to be highly or very highly suitable habitat (HSI >= 60) for all model /
validation data combinations (see manuscript Table 3),
t.HSI <- 59.99 # HSI threshold
HSI.60 <- vector("list", n.list)
for (i in seq_along(HSI.60)) {
HSI.60[[i]] <- aggregate(Pixelcount~HSI > t.HSI, data = exp.list[[i]], sum) / sum(exp.list[[i]]$Pixel
count)
}
HSI.60.list <- vector("list", n.list)
for (i in seq_along(HSI.60.list)) {
HSI.60.list[[i]] <- append(HSI.60.list[[i]], (round(HSI.60[[i]][2, "Pixelcount"] * 100, digits = 0)))
}
HSI.60.list
e.g. HSI 60 of expert model 1 in wet season Balkanu validation background (first in list).
Step 7
Compare P/E ratio between individual expert models and an averaged model for each validation data set.
Create a nested list (first list expert/average models per validation data set, then list validation data sets),
boyce.wetBLKW <- list (boyce.E1.wetBLKW, boyce.E2.wetBLKW, boyce.E3.wetBLKW, boyce.E4.wetBLKW, boyce.E5
.wetBLKW, boyce.E6.wetBLKW, boyce.av.wetBLKW)
boyce.dryBLKW <- list (boyce.E1.dryBLKW, boyce.E2.dryBLKW, boyce.E3.dryBLKW, boyce.E4.dryBLKW, boyce.E5
.dryBLKW, boyce.E6.dryBLKW, boyce.av.dryBLKW)
boyce.wetLKNP <- list (boyce.E1.wetLKNP, boyce.E2.wetLKNP, boyce.E3.wetLKNP, boyce.E4.wetLKNP, boyce.E5
.wetLKNP, boyce.E6.wetLKNP, boyce.av.wetLKNP)
boyce.dryLKNP <- list (boyce.E1.dryLKNP, boyce.E2.dryLKNP, boyce.E3.dryLKNP, boyce.E4.dryLKNP, boyce.E5
.dryLKNP, boyce.E6.dryLKNP, boyce.av.dryLKNP)
boyce.wetOTNP <- list (boyce.E1.wetOTNP, boyce.E2.wetOTNP, boyce.E3.wetOTNP, boyce.E4.wetOTNP, boyce.E5
.wetOTNP, boyce.E6.wetOTNP, boyce.av.wetOTNP)
boyce.dryNAQS <- list (boyce.E1.dryNAQS, boyce.E2.dryNAQS, boyce.E3.dryNAQS, boyce.E4.dryNAQS, boyce.E5
.dryNAQS, boyce.E6.dryNAQS, boyce.av.dryNAQS)
boyce.wetALA <- list (boyce.E1.wetALA, boyce.E2.wetALA, boyce.E3.wetALA, boyce.E4.wetALA, boyce.E5.wetA
LA, boyce.E6.wetALA, boyce.av.wetALA)
boyce.dryALA <- list (boyce.E1.dryALA, boyce.E2.dryALA, boyce.E3.dryALA, boyce.E4.dryALA, boyce.E5.dryA
LA, boyce.E6.dryALA, boyce.av.dryALA)
plot.v.list <- list ("Balkanu (wet season)" = boyce.wetBLKW, "Balkanu (dry season)" = boyce.dryBLKW, "L
akefield (wet season)" = boyce.wetLKNP, "Lakefield (dry season)" = boyce.dryLKNP, "Oyala Thumotang (wet
season)" = boyce.wetOTNP, "NAQS (dry season)" = boyce.dryNAQS, "ALA NT (wet season)" = boyce.wetALA, "
ALA NT (dry season)" = boyce.dryALA)
and plot from nested list (see manuscript Fig 5).
graphics.off()
par(mfrow = c(4, 4), mar = c(2, 2, 3, 0), oma = c(4, 3, 0, 0))
for (i in seq_along(plot.v.list)) {
plot(plot.v.list[[i]][[1]]$HS, plot.v.list[[i]][[1]]$F.ratio, type = "n", xlab = '', ylab = '',
xlim = c(15, 80), ylim = c(0, 3), main = paste(names(plot.v.list[i])))
lines(plot.v.list[[i]][[1]]$HS, na.fill(plot.v.list[[i]][[1]]$F.ratio, 0), col = "black", lty = 1)
lines(plot.v.list[[i]][[2]]$HS, na.fill(plot.v.list[[i]][[2]]$F.ratio, 0), col = "black", lty = 2)
lines(plot.v.list[[i]][[3]]$HS, na.fill(plot.v.list[[i]][[3]]$F.ratio, 0), col = "black", lty = 3)
lines(plot.v.list[[i]][[4]]$HS, na.fill(plot.v.list[[i]][[4]]$F.ratio, 0), col = "black", lty = 4)
lines(plot.v.list[[i]][[5]]$HS, na.fill(plot.v.list[[i]][[5]]$F.ratio, 0), col = "black", lty = 5)
lines(plot.v.list[[i]][[6]]$HS, na.fill(plot.v.list[[i]][[6]]$F.ratio, 0), col = "black", lty = 6)
lines(plot.v.list[[i]][[7]]$HS, na.fill(plot.v.list[[i]][[7]]$F.ratio, 0), col = "red", lty = 1, lwd
= 2)
mtext(LETTERS[i], side = 3, line = -2, adj = 0.05)
plot.new()
legend("left", c((paste("Expert 1, CBI =", round(plot.v.list[[i]][[1]]$Spearman.cor, digits = 2))),
(paste("Expert 2, CBI =", round(plot.v.list[[i]][[2]]$Spearman.cor, digits = 2))),
(paste("Expert 3, CBI =", round(plot.v.list[[i]][[3]]$Spearman.cor, digits = 2))),
(paste("Expert 4, CBI =", round(plot.v.list[[i]][[4]]$Spearman.cor, digits = 2))),
(paste("Expert 5, CBI =", round(plot.v.list[[i]][[5]]$Spearman.cor, digits = 2))),
(paste("Expert 6, CBI =", round(plot.v.list[[i]][[6]]$Spearman.cor, digits = 2))),
(paste("Averaged, CBI =", round(plot.v.list[[i]][[7]]$Spearman.cor, digits = 2))))
,
bty = "n", lty = c(1,2,3,4,5,6,1), col = c("black", "black", "black", "black", "black", "black
", "red"), lwd = c(1, 1, 1, 1, 1, 1, 2), y.intersp = 1.3, title = expression(bold(Legend)), title.adj =
0.05)
}
mtext(expression(italic(Habitat~suitability~index~(HSI))), side = 1, outer = TRUE, cex = 1.2, line = 2.
2)
mtext(expression(italic(Predicted-to-expected~(P/E)~ratio)), side = 2, outer = TRUE, cex = 1.2, line =
0.8)
References
[1] Allaire, J.J. et al. 2016. Package ‘rmarkdown’: dynamic documents for R. URL http://rmarkdown.rstudio.com/ .
[2] Xie, Y. 2016. Package ‘knitr’: a general-purpose package for dynamic report generation in R. URL http://yihui.name/knitr/ .
[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna,
Austria. URL http://www.R-project.org/ .
[4] Broennimann, O. 2015. Package ‘ecospat’: spatial ecology miscellaneous methods. URL
http://cran.r-project.org/web/packages/ecospat/ .
[5] Dowle, M. et al. 2015. Package ‘data.table’: extension of data.frame. URL
https://github.com/Rdatatable/data.table/wiki/ .
[6] Zeileis, A. et al. 2015. Package ‘zoo’: S3 infrastructure for regular and irregular time series. URL
http://zoo.R-Forge.R-project.org/ .
Session information
## Session info --------------------------------------------------------------
##
##
##
##
##
##
##
##
setting
version
system
ui
language
collate
tz
date
value
R version 3.1.3 (2015-03-09)
x86_64, mingw32
RTerm
(EN)
English_Australia.1252
Australia/Brisbane
2017-04-18
## Packages ------------------------------------------------------------------
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
package
abind
acepack
ade4
adehabitatHR
adehabitatLT
adehabitatMA
ape
biomod2
boot
chron
CircStats
class
cluster
codetools
colorspace
data.table
deldir
devtools
digest
dismo
ecodist
ecospat
evaluate
foreach
foreign
formatR
Formula
gam
gbm
ggplot2
goftest
gridExtra
gtable
hexbin
Hmisc
htmltools
iterators
knitr
lattice
latticeExtra
magrittr
maptools
MASS
Matrix
MatrixModels
mda
memoise
mgcv
multcomp
munsell
mvtnorm
nlme
nnet
plyr
polspline
polyclip
pROC
proto
quantreg
randomForest
raster
rasterVis
RColorBrewer
Rcpp
reshape
rmarkdown
rms
rpart
* version
1.4-3
1.3-3.3
* 1.7-2
0.4.14
0.3.20
0.3.10
* 3.3
3.1-64
1.3-15
2.3-47
0.2-4
7.3-12
2.0.1
0.2-10
1.2-6
* 1.9.6
0.1-9
1.10.0
0.6.8
1.0-12
1.2.9
* 1.1
0.8
* 1.4.3
0.8-63
1.2.1
1.2-1
* 1.12
* 2.1.1
2.1.0
1.0-3
2.0.0
0.1.2
1.27.1
3.17-0
0.3
1.0.8
1.12.3
* 0.20-30
0.6-26
1.5
0.8-37
7.3-39
1.2-3
0.4-1
0.4-7
1.0.0
1.8-4
1.4-1
0.4.2
1.0-3
3.1-120
7.3-9
1.8.3
1.1.12
1.3-2
1.8
0.3-10
5.19
4.6-12
2.4-20
0.37
1.1-2
0.12.1
0.8.5
0.9.2
4.4-0
4.1-9
date
2015-03-13
2013-05-03
2015-04-14
2015-07-22
2015-07-22
2015-07-22
2015-05-29
2014-12-10
2015-02-24
2015-06-24
2012-10-29
2015-02-11
2015-01-31
2015-01-17
2015-03-11
2015-09-19
2015-03-09
2016-01-23
2014-12-31
2015-03-15
2013-12-03
2015-03-06
2015-09-18
2015-10-13
2015-02-20
2015-09-18
2015-04-07
2015-05-13
2015-03-11
2016-03-01
2015-07-03
2015-07-14
2012-12-05
2015-08-19
2015-09-21
2015-12-29
2015-10-13
2016-01-22
2015-02-22
2013-08-15
2014-11-22
2015-09-29
2015-02-24
2015-11-28
2015-08-22
2015-05-25
2016-01-29
2014-11-27
2015-07-23
2013-07-11
2015-07-22
2015-02-20
2015-02-11
2015-06-12
2015-07-14
2015-05-27
2015-05-05
2012-12-22
2015-08-31
2015-10-07
2015-09-08
2015-09-06
2014-12-07
2015-09-10
2014-04-23
2016-01-01
2015-09-28
2015-02-24
source
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
##
##
##
##
##
##
##
##
##
##
##
##
sandwich
scales
sp
SparseM
spatstat
stringi
stringr
survival
tensor
TH.data
yaml
zoo
2.3-4
0.3.0
* 1.2-1
1.7
1.41-1
1.0-1
1.0.0
* 2.38-1
1.5
1.0-6
2.1.13
* 1.7-12
2015-09-24
2015-08-25
2015-10-18
2015-08-15
2015-02-27
2015-10-22
2015-04-30
2015-02-24
2012-05-05
2015-01-05
2014-06-12
2015-03-16
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
CRAN
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
(R
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
Appendix 4.5 Validation maps – seasonal habitat suitability
229
Appendix Fig 4.5 Seasonal habitat suitability for feral pig breeding in the four validation
backgrounds. Presence records used for validation are shown for: Balkanu wet (A) and dry season (B),
Lakefield wet (C) and dry (D) season, Oyala Thumotang wet season (E), NAQS dry season (F) and ALA wet
(G) and dry (H) season. Descriptions of data sets are in Table 4.3. Validation backgrounds were defined from
existing management units or by buffering data points.
230
231
Fig 4.1.
Appendix Fig 4.6.1 Share of seasonal habitat per state. Percentages were calculated for each suitability class separately. Location of states is shown in
Appendix 4.6 Additional analyses – seasonal habitat suitability
232
network model (Table 4.1). Percentages were calculated separately for each vegetation type (from Present Major Vegetation Groups (MVG V.4.1).
Appendix Fig 4.6.2 Distribution of modelled habitat suitability for each broad vegetation type. Habitat suitability classes were taken from the Bayesian
Appendix 5.1 Methodology – omnidirectional current density
This page is intentionally blank.
233
Rcode - modelling omnidirectional current
density
Jens G. Froese
18 April 2017
This document provides a detailed, reproducible description of the omnidirectional current density methodology. It is
Supporting Information to the manuscript:
Froese JG, Smith CS, Durr PA, McAlpine CA & van Klinken RD (in prep). “Modelling habitat
connectivity for biosecurity: the risk of infectious disease spread in wild pigs in northern
Australia.”
It is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. Some
parts of the code used to print this document have been suppressed to enhance readability.
The methodoology was adapted from:
Pelletier D, Clark M, Anderson, MG, Rayfield, B, Wulder, MA et al. (2014). “Applying circuit
theory for corridor expansion and management at regional scales: tiling, pinch points, and
omnidirectional connectivity.” PLoS ONE 9(1): e84135.
It extends the circuit-theoretic Circuitscape (CS) model to regional-scale applications (millions of study area pixels).
While access to a high-performance computing cluster is not required, running Circuitscape is computationally demanding,
a high powered computer with as much RAM as possible is recommended (we had 40GB).
Step 1: Preparation
Prepare your workspace and load required packages
rm(list=ls()) # clear workspace
dev.off() # clear graphics
require(raster) # load packages
require(rgdal)
Set raster options (incl. tmpdir for easily deleting temporary files that otherwise clog up your
C:/ drive)
rasterOptions (tmpdir = "[your-temp-dir]", progress = "text", timer = TRUE)
Load study area wide resistance and random raster
Prepare input raster data in GIS software of your choice (we used .tif format, must have same spatial reference, extent,
resolution)
matrix resistance for entire study area res
matrix resistance tiles (divide study area into suitably sized tiles, collated in one directory, see
Pelletier et al. 2014 for guidance)
random resistance values for entire study area ran (must have same value range as resistance raster, WARNING!
for heterogeneous study areas, random values should reflect the range of resistance values in each individual tile to
avoid border effects)
res = raster("[your-resistance-raster.tif]")
res; #plot(res)
ran = raster("[your-random-raster.tif]")
ran; #plot(ran)
Create buffered resistance and current source tiles
Set wd to your tiles dir (if you choose base wd, then for loop print(i) produces error) and load resistance .tif tiles
into list
setwd("[your-tiles-dir]")
l.rtl = dir(getwd(), pattern = ".tif")
Loop through each tile in list and write out buffered resistance and current source (two directions WE and NS) .tif files
to tifdir
tifdir = "[your-tif-dir]"
for (i in l.rtl) {
print(i)
rtl = raster(i) # read tile
rtl.temp <- extend(rtl, c(1500,1500), value=NA) # extend by 100% buffer (our tiles were 1500
*1500 pixels, need to adjust value), buffer is NA
ran.buf <- crop(ran, rtl.temp) # crop random raster by buffered tile
rtl.buf <- cover(rtl.temp, ran.buf) # replace NA in buffer with random values
we.buf <- rtl.buf
we.buf[] <- NA; we.buf[,1] <- 1; we.buf[,4500] <- 2 # current source tile (size = buffered t
ile, values = NA except thin strip west/east)
ns.buf <- rtl.buf
ns.buf[] <- NA; ns.buf[1,] <- 1; ns.buf[4500,] <- 2 # current source tile (size = buffered t
ile, values = NA except thin strip north/south)
writeRaster(rtl.buf, datatype='INT4S', filename = paste(tifdir, "Tiles_buf/", i, ".asc", sep
= "")) # write buffered resistance tile as .asc
writeRaster(we.buf, datatype='INT4S', filename = paste(tifdir, "SourceWE_buf/SourceWE_", i,
".asc", sep = "")) # write current source WE tile as .asc
writeRaster(ns.buf, datatype='INT4S', filename = paste(tifdir, "SourceNS_buf/SourceNS_", i,
".asc", sep = "")) # write current source NS tile as .asc
}
Clean up your temp files (to avoid memory issues)
rtemp = "[your-temp-dir]"
setwd(rtemp)
file.remove(dir(rtemp))
gc()
Step 2: Directional Circuitscape (CS) run
Code was adapted from
Marrotte 2015: http://robbymarrotte.weebly.com/blog/running-circuitscape-in-r-windows-os
(http://robbymarrotte.weebly.com/blog/running-circuitscape-in-r-windows-os)
Prepare CS inputs from Step 1 in lists
Set wd to your tiles dir
setwd("[your-tif-dir]")
Create two lists for resistance and current source (here WE direction, repeat for NS) .asc files
l.tile = dir("[your-tif-dir]/Tiles_buf/", pattern = ".asc")
l.source = dir("[your-tif-dir]/SourceWE_buf/", pattern = ".asc")
Prepare CS .ini files for each tile
For each listed resistance tile, write lines with CS options, input .asc and output .out … (WE or NS)
l.tile.ini = list() # same as: vector("list")
for (i in l.tile) {
tile.ini <- c("[circuitscape options]",
"data_type = raster",
"scenario = pairwise",
"write_cur_maps = True",
"write_cum_cur_map_only = True",
paste("habitat_file = ", getwd(), "/Tiles_buf/", i, sep = ""),
paste("output_file = ", getwd(), "/CS_outWE/", i, ".out", sep = ""))
l.tile.ini[[i]] <- append(l.tile.ini[[i]], tile.ini)
}
For each listed current source tile (WE or NS), write line with input .asc
l.source.ini = list()
for (i in l.source) {
source.ini <- c(paste("point_file = ", getwd(), "/SourceWE_buf/", i, sep = ""))
l.source.ini[[i]] <- append(l.source.ini[[i]], source.ini)
}
For each list element, combine lines and write lines out to .ini file (to CS_runWE or CS_runNS dir), then combine all
written out .ini files in a run list
l.run <- list()
for (i in 1:length(l.tile.ini)) {
l.tile.ini[[i]] <- c(l.tile.ini[[i]], l.source.ini[[i]])
writeLines(l.tile.ini[[i]], paste(getwd(), "/CS_runWE/Tile", i, ".ini", sep = ""))
l.run[i] <- paste(getwd(), "/CS_runWE/Tile", i, ".ini", sep = "")
}
Run the Circuitscape model
For instructions on software installation see
McRae BH, Shah VB, Mohapatra TK (2013). Circuitscape 4 user guide: The Nature
Conservancy. http://www.circuitscape.org (http://www.circuitscape.org)
Make a CS run CMD and run command for each .ini file in the run list (this runs the model)
CS_exe <- 'C:/"Program Files"/Circuitscape/cs_run.exe' # Don't forget quotation marks in "Prog
ram Files"
sapply(l.run, function(x) system(paste(CS_exe, x)))
Step 3: Directional current density mosaic
Prepare workspace and load packages
rm(list=ls()) # clear workspace
tifdir = "[your-tif-dir]/MosaicWE/"
setwd("[your-tif-dir]/CS_outWE/")
require(raster)
require(rgdal)
projection <- "+proj=aea +lat_1=-18 +lat_2=-36 +lat_0=0 +lon_0=132 +x_0=0 +y_0=0 +ellps=GRS80
+units=m +no_defs" # define a projection system (should match your input rasters, we used Aust
ralian Albers prjection)
Load cumulative current .asc tiles (WE or NS), combine in a list and crop buffer by rows and columns
l.in = dir(getwd(), pattern = "curmap.asc")
l.cur = list()
for (i in l.in) {
cur = raster(i)
cur.crop <- crop(cur, extent(cur, 1501, 3000, 1501, 3000)) # this extracts the orginal tile
area (here 1500*1500 pixels), from the larger buffer tile (here 4500*4500 pixels)
crs(cur.crop) <- projection
writeRaster(cur.crop, filename = paste(tifdir, i, ".tif", sep = ""))
l.cur[[i]] <- cur.crop
}
Mosaic (use merge because tiles do not overlap) cropped cumulative current tiles into one raster and write out as .tif
x <- list()
for (i in 1:length(l.cur)){
x[i] <- l.cur[i]
}
x$filename <- '[your-[WE/NS]mosaic-raster.tif]'
x$overwrite <- TRUE
cur.mos <- do.call(merge, x)
writeRaster(cur.mos, filename = paste(tifdir, "[your-[WE/NS]mosaic-raster.tif]", sep = ""))
Step 4: Omnidirectional current density mosaic
Load resistance and directional cumulative current mosaic rasters (WE and NS)
res = raster("[your-resistance-raster.tif]")
res; #plot(res)
setwd("[your-tif-dir]")
tifdir = "[your-tif-dir]/Current/"
mos.WE = raster(paste(getwd(), "/MosaicWE/[your-NSmosaic-raster.tif]", sep = ""))
mos.NS = raster(paste(getwd(), "/MosaicNS/[your-WEmosaic-raster.tif]", sep = ""))
Mask mosaic rasters (WE and NS) by resistance raster
res <- crop(res, mos.WE)
mos.WE <- mask(mos.WE, res)
mos.WE; #plot(mos.WE)
writeRaster(mos.WE, filename = paste(tifdir, "[your-masked-WEmosaic-raster.tif]", sep = ""))
mos.NS <- mask(mos.NS, res)
mos.NS; #plot(mos.NS)
writeRaster(mos.NS, filename = paste(tifdir, "[your-masked-NSmosaic-raster.tif]", sep = ""))
Create omnidirectional current density mosaic by multiplication and log10 transformation
mos <- overlay(mos.NS, mos.WE, fun = function(x, y) {return(log10(x * y))} )
writeRaster(mos, filename = paste(tifdir, "[your-OCDlog10-raster.tif", sep = ""))
Optional: explore mosaic histogram
hist(mos,
col="springgreen4",
main="Histogram of omnidirectional current density",
ylab="Number of Pixels",
xlab="current density")
Clean up your temp files (to avoid memory issues)
rtemp = "[your-temp-dir]"
setwd(rtemp)
file.remove(dir(rtemp))
gc()
Session information
## Session info --------------------------------------------------------------
##
##
##
##
##
##
##
##
setting
version
system
ui
language
collate
tz
date
value
R version 3.1.3 (2015-03-09)
x86_64, mingw32
RTerm
(EN)
English_Australia.1252
Australia/Brisbane
2017-04-18
## Packages ------------------------------------------------------------------
##
##
##
##
##
##
##
##
##
##
##
##
##
package
* version
devtools
1.10.0
digest
0.6.8
evaluate
0.8
formatR
1.2.1
htmltools
0.3
knitr
1.12.3
magrittr
1.5
memoise
1.0.0
rmarkdown
0.9.2
stringi
1.0-1
stringr
1.0.0
yaml
2.1.13
date
2016-01-23
2014-12-31
2015-09-18
2015-09-18
2015-12-29
2016-01-22
2014-11-22
2016-01-29
2016-01-01
2015-10-22
2015-04-30
2014-06-12
source
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
CRAN (R
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
3.1.3)
Appendix 5.2 Methodology – patch connectivity
This page is intentionally blank.
239
Appendix 5.3 Analysis – tests of significance
Appendix Table 5.3.1 Two-way tests of significance between levels of ResDisp for the dry (a) and wet (b)
season.
(a)
Dry season scenario
Matrix resistance (ResDisp)
high
moderate
low
p value
393
341
-
0.704
-
341
411
< .001
393
-
411
< .001
Mean (km )
7,820
15,282
45,174
(b)
Wet season scenario
Median (km )
2
2
Matrix resistance (ResDisp)
Median (km )
2
Mean (km )
2
high
moderate
low
p value
5,640
2,697
-
0.098
-
2,697
6,932
< .001
5,640
-
6,932
< .001
32,463
44,961
86,139
Median and mean patch connectivity within the network of feral pig habitat patches in northern Australia is shown for
each level (size of connected habitat components in km2). Pairwise p values have been calculated using the MannWhitney U-test of significance. *p < .05, **p < .01, ***p < .001.
241
242
(299)
mod
(4,065)
loose
CD
Thresh
(356)
mod
(1,901)
loose
CD
Thresh
(279)
mod
(1,472)
loose
CD
Thresh
(a)
season.
***
***
***
(38)
strict
(299)
mod
5k
***
***
***
(43)
strict
(356)
mod
5k
***
***
***
(53)
strict
(279)
mod
5k
mod
(874)
loose
(29,681)
CD
Thresh
mod
(543)
loose
(47,172)
CD
Thresh
(369)
mod
loose
(5,843)
CD
Thresh
***
mod
(874)
10k
DispDist
***
mod
(543)
10k
DispDist
***
mod
(369)
10k
DispDist
***
***
(38)
strict
***
***
(43)
strict
***
***
(53)
strict
mod
(1,969)
loose
(275,592)
CD
Thresh
mod
loose
(112,943)
(735)
***
mod
(580)
***
***
(53)
strict
(53)
10k
(53)
15k
Disp
Dist
***
***
(43)
strict
(43)
10k
(43)
15k
Disp
Dist
***
mod
(1,969)
15k
***
***
(38)
strict
(38)
10k
(38)
15k
Disp
Dist
5k
(43)
5k
p = 0.998
(38)
10k
strict
(38)
p = 0.988
(43)
10k
strict
Dry season scenario / ResDisp low (1-5)
***
mod
(735)
15k
5k
(53)
p = 0.953
(53)
10k
strict
Dry season scenario / ResDisp moderate (1-21)
CD
Thresh
(580)
mod
loose
(33,394)
CD
Thresh
15k
Dry season scenario / ResDisp high (1-100)
10k
(874)
15k
(1,969)
10k
(543)
15k
(735)
Disp
Dist
(369)
10k
15k
(580)
Disp
Dist
5k
(279)
***
10k
(874)
moderate
CDThresh
*
10k
(543)
moderate
CDThresh
***
***
5k
(299)
*
***
5k
(356)
p = 0.142
10k
(369)
moderate
CDThresh
10k
(29,681)
15k
(275,592)
Disp
Dist
10k
(47,172)
15k
(112,943)
Disp
Dist
10k
(5,843)
15k
(33,394)
Disp
Dist
***
10k
(29,681)
loose
***
10k
(47,172)
loose
***
10k
(5,843)
loose
***
***
5k
(4,065)
***
***
5k
(1,901)
***
***
5k
(1,472)
Appendix Table 5.3.2 Two-way tests of significance between levels of CDThresh and DispDist at each level of ResDisp for the dry (table a) and wet (table b)
***
***
***
(79)
strict
***
***
(64)
strict
(4,811)
mod
5k
***
mod
(4,238)
5k
***
***
***
(97)
strict
(5,640)
mod
5k
mod
(36,017)
loose
(76,942)
CD
Thresh
mod
(13,588)
loose
(18,814)
CD
Thresh
mod
(10,737)
loose
(15,563)
CD
Thresh
***
mod
(36,017)
10k
DispDist
***
mod
(13,588)
10k
DispDist
***
mod
(10,737)
10k
DispDist
***
***
(95)
strict
***
***
(71)
strict
***
***
(100)
strict
loose
(208,590)
mod
(40,027)
loose
(334,644)
CD
Thresh
(139,246)
mod
***
***
(101)
strict
10k
(100)
15k
(101)
Disp
Dist
15k
***
***
(71)
10k
(72)
15k
Disp
Dist
***
mod
(40,027)
15k
***
***
(100)
strict
(95)
10k
15k
(100)
Disp
Dist
5k
(64)
p = 0.186
(95)
10k
strict
*
**
5k
(79)
p = 0.353
(71)
10k
strict
Wet season scenario / ResDisp low (1-5)
***
(72)
strict
5k
(97)
p = 0.443
10k
(100)
strict
Wet season scenario / ResDisp moderate (1-21)
***
mod
(12,545)
CD
mod
Thresh (139,246)
mod
(12,545)
loose
(18,455)
CD
Thresh
15k
Wet season scenario / ResDisp high (1-100)
10k
(36,017)
15k
(40,027)
10k
(13,588)
15k
(139,246)
10k
(10,737)
15k
(12,545)
***
10k
(36,017)
moderate
CDThresh
***
10k
(13,588)
moderate
CDThresh
***
10k
(10,737)
moderate
CDThresh
***
***
5k
(4,811)
***
***
5k
(4,238)
***
***
5k
(5,640)
10k
(76,942)
15k
(334,644)
Disp
Dist
10k
(18,814)
15k
(208,590)
Disp
Dist
10k
(15,563)
15k
(18,455)
Disp
Dist
***
10k
(76,942)
loose
***
10k
(18,814)
loose
***
10k
(15,563)
loose
***
***
5k
(19,191)
***
***
5k
(6,266)
***
***
5k
(8,813)
Due to observed interaction effects (Fig. 6), we tested changes to either variable for each level of the other two variables. Median patch connectivity for each factorial
combination is reported in brackets (size of connected habitat components in km2). Pairwise p values have been calculated using the Mann-Whitney U-test of significance.
These are only reported if a three-way Kruska-Wallis rank sum test indicated a significant effect within a factorial subset. Otherwise, the p value of this test is reported.
*p < .05, **p < .01, ***p < .001.
(4,811)
mod
(19,191)
loose
CD
Thresh
(4,238)
mod
(6,266)
loose
CD
Thresh
(5,640)
mod
(8,813)
loose
CD
Thresh
(b)
243
Download PDF
Similar pages