Use of Species Distribution Modeling in the Deep Sea

Abstract

In the last two decades the use of species distribution modeling (SDM) for the study and management of marine species has increased dramatically. The availability of predictor variables on a global scale and the ease of use of SDM techniques have resulted in a proliferation of research on the topic of species distribution in the deep sea. Translation of research projects into
management tools that can be used to make decisions in the face of changing climate and increasing exploitation of deep-sea resources has been less rapid but necessary. The goal of this workshop was to discuss methods and variables for modeling species distributions in deep-sea habitats and produce standards that can be used to judge SDMs that may be useful to meet management and conservation goals. During the workshop, approaches to modeling and environmental data were discussed and guidelines developed including the desire that 1) environmental variables should be chosen for ecological significance a priori; 2) the scale and accuracy of environmental data should be considered in choosing a modeling method; 3) when possible proxy variables such as depth should be avoided if causal variables are available; 4) models with statistically robust and rigorous outputs are preferred, but not always possible; and 5) model validation is important. Although general guidelines for SDMs were developed, in most
cases management issues and objectives should be considered when designing a modeling project. In particular, the trade-off between model complexity and researcher’s ability to communicate input data, modeling method, results and uncertainty is an important consideration for the target audience.

Full text

Use of Species Distribution Modeling in the Deep Sea
E. Kenchington, O. Callery, F. Davidson, A. Grehan, T. Morato, J. Appiott, A.
Davis, P. Dunstan, C. Du Preez, J. Finney, J.M. González-Irusta, K. Howell, A.
Knudby, M. Lacharité, J. Lee, F. J. Murillo, L. Beazley, J.M. Roberts, M.
Roberts, C. Rooper, A. Rowden, E. Rubidge, R. Stanley, D. Stirling, K.R.
Tanaka, J. Vanhatalo, B. Weigel, S. Woolley and C. Yesson
Ocean and Ecosystem Sciences Division
Maritimes Region
Fisheries and Oceans Canada
Bedford Institute of Oceanography
PO Box 1006
Dartmouth, Nova Scotia
Canada B2Y 4A2
2019
Canadian Technical Report of
Fisheries and Aquatic Sciences 3296

Canadian Technical Report of Fisheries and Aquatic Sciences
Technical reports contain scientific and technical information that contributes to existing knowledge but which
is not normally appropriate for primary literature. Technical reports are directed primarily toward a worldwide
audience and have an international distribution. No restriction is placed on subject matter and the series reflects the
broad interests and policies of Fisheries and Oceans Canada, namely, fisheries and aquatic sciences.
Technical reports may be cited as full publications. The correct citation appears above the abstract of each report.
Each report is abstracted in the data base Aquatic Sciences and Fisheries Abstracts.
Technical reports are produced regionally but are numbered nationally. Requests for individual reports will be
filled by the issuing establishment listed on the front cover and title page.
Numbers 1-456 in this series were issued as Technical Reports of the Fisheries Research Board of Canada.
Numbers 457-714 were issued as Department of the Environment, Fisheries and Marine Service, Research and
Development Directorate Technical Reports. Numbers 715-924 were issued as Department of Fisheries and
Environment, Fisheries and Marine Service Technical Reports. The current series name was changed with report
number 925.
Rapport technique canadien des sciences halieutiques et aquatiques
Les rapports techniques contiennent des renseignements scientifiques et techniques qui constituent une
contribution aux connaissances actuelles, mais qui ne sont pas normalement appropriés pour la publication dans un
journal scientifique. Les rapports techniques sont destinés essentiellement à un public international et ils sont
distribués à cet échelon. II n'y a aucune restriction quant au sujet; de fait, la série reflète la vaste gamme des intérêts
et des politiques de Pêches et Océans Canada, c'est-à-dire les sciences halieutiques et aquatiques.
Les rapports techniques peuvent être cités comme des publications à part entière. Le titre exact figure au-dessus
du résumé de chaque rapport. Les rapports techniques sont résumés dans la base de données Résumés des sciences
aquatiques et halieutiques.
Les rapports techniques sont produits à l'échelon régional, mais numérotés à l'échelon national. Les demandes
de rapports seront satisfaites par l'établissement auteur dont le nom figure sur la couverture et la page du titre.
Les numéros 1 à 456 de cette série ont été publiés à titre de Rapports techniques de l'Office des recherches sur
les pêcheries du Canada. Les numéros 457 à 714 sont parus à titre de Rapports techniques de la Direction générale de
la recherche et du développement, Service des pêches et de la mer, ministère de l'Environnement. Les numéros 715
à 924 ont été publiés à titre de Rapports techniques du Service des pêches et de la mer, ministère des Pêches et de
l'Environnement. Le nom actuel de la série a été établi lors de la parution du numéro 925.

ii
Canadian Technical Report of
Fisheries and Aquatic Sciences 3296
2019
USE OF SPECIES DISTRIBUTION MODELING IN THE
DEEP SEA
by
E. Kenchington1, O. Callery2, F. Davidson3, A. Grehan2, T. Morato4, J. Appiott5, A. Davis6, P.
Dunstan7, C. Du Preez8, J. Finney9, J.M. González-Irusta4, K. Howell10, A. Knudby3, M.
Lacharité11, J. Lee5, F.J. Murillo1, L. Beazley1, J.M. Roberts12, M. Roberts6, C. Rooper13, A.
Rowden14, E. Rubidge8, R. Stanley1, D. Stirling15, K.R. Tanaka16, J. Vanhatalo17, B. Weigel17,
S. Woolley7 and C. Yesson18
1 Fisheries and Oceans Canada, Bedford Institute of Oceanography, Dartmouth, NS, Canada
2 National University of Ireland, Galway, Ireland
3 University of Ottawa, Ottawa, Ontario, Canada
4 Institute of Marine Research, Azores, Portugal
5 Secretariat of the Convention on Biological Diversity, Montreal, Quebec, Canada
6 University of Bangor, Bangor, Wales, United Kingdom
7 CSIRO Oceans and Atmosphere, Hobart, Tasmania, Australia
8 Fisheries and Oceans Canada, Institute of Ocean Sciences, Sidney, British Columbia, Canada
9 Fisheries and Oceans Canada, Pacific Biological Station, Nanaimo, British Columbia, Canada
10 University of Plymouth, Plymouth, England, United Kingdom
11 Nova Scotia Community College, Dartmouth, Nova Scotia, Canada
12 University of Edinburgh, Edinburgh, Scotland, United Kingdom
13 Alaska Fisheries Science Center, Seattle, Washington, United States of America
14 National Institute of Water and Atmospheric Research, Wellington, New Zealand
15 Marine Scotland Science, Aberdeen, Scotland, United Kingdom
16 Princeton University, Princeton, New Jersey, United States of America
17 University of Helsinki, Helsinki, Finland
18 Zoological Society of London, London, England, United Kingdom

iii
© Her Majesty the Queen in Right of Canada, 2019.
Cat. N ° Fs97-6/3296E-PDF ISBN 978-0-660-29721-7 ISSN 1488-5379
Correct citation for this publication:
Kenchington, E., Callery, O., Davidson, F., Grehan, A., Morato, T., Appiott, J., Davis, A.,
Dunstan, P., Du Preez, C., Finney, J., González-Irusta, J.M., Howell, K., Knudby, A.,
Lacharité, M., Lee, J., Murillo, F.J., Beazley, L., Roberts, J.M., Roberts, M., Rooper, C.,
Rowden, A., Rubidge, E., Stanley, R., Stirling, D., Tanaka, K.R., Vanhatalo, J., Weigel,
B., Woolley, S. and Yesson, C. 2019. Use of Species Distribution Modeling in the Deep
Sea. Can. Tech. Rep. Fish. Aquat. Sci. 3296: ix + 76 p.

iv
TABLE OF CONTENTS
ABSTRACT ................................................................................................................................... vii
RÉSUMÉ ....................................................................................................................................... viii
1. INTRODUCTION ............................................................................................................... 1
1.1. Setting the Context: Voluntary Specific Workplan on Biodiversity in Cold-water Areas
within the Jurisdictional Scope of the Convention on Biological Diversity ................................... 1
J. Murray Roberts.................................................................................................................... 1
2. THEME 1 SHOWCASE OF APPROACHES TO DEVELOP SPECIES
DISTRIBUTION MODELS IN THE DEEP-SEA ..................................................................... 5
2.1. Model-Based Thinking for Community Ecology .................................................................. 6
2.1.1. Group Discussion ...................................................................................................... 8
2.2. Point Process Framework for Species Distribution Modeling and Joint Species Distribution
Models............................................................................................................................................. 9
Jarno Vanhatalo ...................................................................................................................... 9
2.2.1. Group Discussion .................................................................................................... 17
3. EXTENDED ABSTRACTS .............................................................................................. 18
3.1. What do Environmental Managers and Stakeholders want from a Model? ........................ 19
Ashley A. Rowden ................................................................................................................ 19
3.1.1. Group Discussion .................................................................................................... 19
3.2. Rough Data: Imperfect Data and Directional Rugosity ...................................................... 20
Cherisse Du Preez, Emily Rubidge and Jessica Finney ........................................................ 20
3.3. Determining the best methods for model validation ........................................................... 24
Chris Rooper, Rachel Wilborn and Pam Goddard ................................................................ 24
3.3.1. Group Discussion .................................................................................................... 26
3.4. Distribution Models Applied to Climate Change in the Deep Sea. A Promising but
Challenging Development Field ................................................................................................... 27
José Manuel González-Irusta and Telmo Morato ................................................................. 27

v
3.4.1. Group Discussion .................................................................................................... 30
3.5. Including Predictions of Community Functional Traits in Species Distribution Models using
Hierarchical Modeling of Species Communities (HMSC) ........................................................... 31
Benjamin Weigel .................................................................................................................. 31
3.6. GlobENV: Towards a High-resolution Climatology for the Seafloor................................. 34
Andrew J. Davies and Emyr M.T. Roberts ........................................................................... 34
3.7. Ensembling of Multiple Data Sets and Multiple Models .................................................... 38
Chris Rooper, Duane Stevenson, Ivonne Ortiz, Jerry Hoff .................................................. 38
3.7.1. Group Discussion .................................................................................................... 43
3.8. Resolution of Seabed Features in the Deep-sea: Implications for Habitat Characterization
43
Myriam Lacharité and Craig J. Brown ................................................................................. 43
3.8.1. Group Discussion .................................................................................................... 46
3.9. Best Practices in the Development and Application of Species Distribution Models to
Support Decision Making in Marine Spatial Planning ................................................................. 46
Jessica Finney, Emily Rubidge, Cherisse Du Preez, Jessica Nephin, Candice St. Germain,
Cole Fields, and Edward Gregr ............................................................................................. 46
3.10. Evaluating Effects of Rescaling and Weighting Data on Habitat Suitability Modeling ..... 48
Ying Xue, Lisha Guan, Kisei R. Tanaka, Zengguang Li, Yong Chen, Yiping Ren ............. 48
3.10.1. Group Discussion ................................................................................................ 49
3.11. Developing a Generalized Climate-niche Modeling Framework to Improve Management of
Commercially Important Species in a Climatically Altered Marine Environment: A Case Study
with American lobster and Atlantic scallop in the Gulf of Maine ................................................ 50
K. R. Tanaka, M. P. Torre, V. S. Saba, and Y. Chen ............................................................ 50
3.11.1. Group Discussion ................................................................................................ 53
3.12. Determining Thresholds for Interpretation of Probability Maps ......................................... 53
Chris Rooper ......................................................................................................................... 53
3.12.1. Group Discussion ................................................................................................ 56
3.13. Some Uncertainties of Bathymetry Data ............................................................................. 56

vi
Chris Yesson ......................................................................................................................... 56
3.14. Mapping Uncertainty of SDM Predictions .......................................................................... 59
Fiona Davidson and Anders Knudby .................................................................................... 59
4. THEME 2 - BIOLOGICAL AND ENVIRONMENTAL DATASETS RELEVANT
TO DEEP-SEA SPECIES AND COMMUNITIES .................................................................. 62
5. THEME 3 - TEMPORAL AND SPATIAL SCALES RELEVANT FOR
DEVELOPING SDM IN THE DEEP SEA .............................................................................. 64
6. THEME 4 - MODELING TOOLS IN THE CONTEXT OF DATA-LIMITED
SITUATIONS AND IN THE CONTEXT OF SINGLE SPECIES AND JOINT SPECIES
MODELING ................................................................................................................................ 66
7. DEVELOPMENT OF GUIDELINES .............................................................................. 68
APPENDIX 1. WORKSHOP AGENDA ..................................................................................... 73
APPENDIX 2. LIST OF PARTICIPANTS ................................................................................. 76

vii
ABSTRACT
Kenchington, E., Callery, O., Davidson, F., Grehan, A., Morato, T., Appiott, J., Davis, A., Dunstan,
P., Du Preez, C., Finney, J., González-Irusta, J.M., Howell, K., Knudby, A., Lacharité, M., Lee,
J., Murillo, F.J., Beazley, L., Roberts, J.M., Roberts, M., Rooper, C., Rowden, A., Rubidge, E.,
Stanley, R., Stirling, D., Tanaka, K.R., Vanhatalo, J., Weigel, B., Woolley, S. and Yesson, C. 2019.
Use of Species Distribution Modeling in the Deep Sea. Can. Tech. Rep. Fish. Aquat. Sci. 3296: ix
+ 76 p.
In the last two decades the use of species distribution modeling (SDM) for the study and
management of marine species has increased dramatically. The availability of predictor variables
on a global scale and the ease of use of SDM techniques have resulted in a proliferation of research
on the topic of species distribution in the deep sea. Translation of research projects into
management tools that can be used to make decisions in the face of changing climate and
increasing exploitation of deep-sea resources has been less rapid but necessary. The goal of this
workshop was to discuss methods and variables for modeling species distributions in deep-sea
habitats and produce standards that can be used to judge SDMs that may be useful to meet
management and conservation goals. During the workshop, approaches to modeling and
environmental data were discussed and guidelines developed including the desire that 1)
environmental variables should be chosen for ecological significance a priori; 2) the scale and
accuracy of environmental data should be considered in choosing a modeling method; 3) when
possible proxy variables such as depth should be avoided if causal variables are available; 4)
models with statistically robust and rigorous outputs are preferred, but not always possible; and 5)
model validation is important. Although general guidelines for SDMs were developed, in most
cases management issues and objectives should be considered when designing a modeling project.
In particular, the trade-off between model complexity and researcher’s ability to communicate
input data, modeling method, results and uncertainty is an important consideration for the target
audience.

viii
RÉSUMÉ
Kenchington, E., Callery, O., Davidson, F., Grehan, A., Morato, T., Appiott, J., Davis, A.,
Dunstan, P., Du Preez, C., Finney, J., González-Irusta, J.M., Howell, K., Knudby, A., Lacharité,
M., Lee, J., Murillo, F.J., Beazley, L., Roberts, J.M., Roberts, M., Rooper, C., Rowden, A.,
Rubidge, E., Stanley, R., Stirling, D., Tanaka, K.R., Vanhatalo, J., Weigel, B., Woolley, S. and
Yesson, C. 2019. Recours à la modélisation de la répartition des espèces en haute mer. Can. Tech.
Rep. Fish. Aquat. Sci. 3296: ix + 76 p.
Au cours des deux dernières décennies, le recours à la modélisation de la répartition des espèces
pour étudier et gérer les espèces marines a considérablement augmenté. La disponibilité des
variables prédictives à l’échelle mondiale et la convivialité de ces techniques de modélisation ont
entraîné la multiplication des recherches sur la répartition des espèces en haute mer. La traduction
des projets de recherche en outils de gestion pouvant servir à prendre des décisions dans le contexte
des changements climatiques et de l’exploitation accrue des ressources en haute mer est moins
rapide, quoique nécessaire. Cet atelier visait à discuter des méthodes et variables pour la
modélisation de la répartition des espèces dans les habitats en haute mer, et à établir des normes
pour évaluer les méthodes de modélisation pouvant aider à atteindre les objectifs en matière de
gestion et de conservation. Pendant l’atelier, les approches envers la modélisation et les données
environnementales ont fait l’objet de discussions, et des lignes directrices ont été élaborées.
Celles-ci comprenaient les caractéristiques souhaitées qui suivent : 1) les variables
environnementales devraient être choisies selon leur importance écologique a priori; 2) l’ampleur
et l’exactitude des données environnementales devraient être prises en compte durant la sélection
d’une méthode de modélisation; 3) dans la mesure du possible, les variables substitutives, comme
la profondeur, doivent être évitées si des variables causales sont disponibles; 4) les modèles dont
les résultats sont statistiquement solides et rigoureux sont privilégiés, mais leur utilisation n’est
pas toujours possible; 5) la validation du modèle est importante. Même si des lignes générales sur
la modélisation de la répartition des espèces ont été mises au point, les objectifs et enjeux de
gestion devraient généralement être pris en compte pendant la conception d’un projet de
modélisation. En particulier, le compromis entre la complexité du modèle et la capacité du
chercheur à communiquer les données d’entrée, la méthode de modélisation, les résultats et les
incertitudes sont des facteurs importants pour le public cible.

ix

1
1. INTRODUCTION
The use of species distribution modeling (SDM; also known as habitat suitability modeling or HSM)
in marine ecology has grown enormously in recent years with the accumulation of appropriate
response and predictor data sets, and software that is freely available and often global in scope.
Exploitation of the deep sea is increasing and diversifying, while knowledge of the actual distribution
of deep-sea biodiversity is still poor due to the limited spatial footprint of sampling to date. There is
now a great need for robust species distribution modeling that can inform decision-making and
anticipate the influence of global change. In particular, to support the Voluntary Specific Workplan
on Biodiversity in Cold-water Areas within the Jurisdictional Scope of the Convention as adopted at
CBD COP13. However, the variables that may be useful predictors for species and communities
living in shallow waters and on the continental shelves may not be appropriate for modeling
distributions below 100 m. For example, in the northwest Atlantic temperature and salinity are
relatively constant below this depth but are key determinants of distributions at shallower depths.
The goal of this workshop was to discuss appropriate variables and methods for modeling of species
and communities of deep-sea habitats, including seamounts, and to produce a set of standards for
publication that will lead to more comparable work in this field.
This document summarizes the discussions of a workshop that brought together global experts
(including from the EU Horizon 2020 research projects ATLAS and SponGES) on species
distribution modeling and deep-sea biology to discuss this topic. The main objectives of the workshop
were to: 1) showcase existing attempts to develop species distribution modeling in the deep sea; 2)
to discuss biological and environmental datasets relevant to the deep sea; 3) discuss temporal and
spatial scales relevant for developing species distribution modeling in the deep sea; 4) discuss
appropriate modeling tools in the context of data-limited situations, and in the context of single
species and joint species modeling; and 5) establish more effective collaborations within the deep-
sea modeling community.
Participants were asked to prepare short working papers addressing agenda items and to make ten
minute presentations to stimulate discussion. Those presentations are presented in the Extended
Abstract section below.
Co-chairs: Ellen Kenchington (DFO, Canada), Anthony Grehan (NUI, Galway, Ireland), Telmo
Morato (IMAR, Azores, Portugal)
1.1. Setting the Context: Voluntary Specific Workplan on Biodiversity in Cold-water
Areas within the Jurisdictional Scope of the Convention on Biological Diversity
J. Murray Roberts
ATLAS project coordinator, University of Edinburgh, School of GeoSciences,
Grant Institute, James Hutton Road, Edinburgh, UK

2
In 2012 the 11th Conference of the Parties (COP) to the Convention on Biological Diversity (CBD)
requested that the CBD Executive Secretary prepare a draft specific workplan on biodiversity and
ocean acidification in cold-water areas of the ocean. Developed through a collaborative process
involving other governments and relevant organisations, this workplan would build upon elements
of a previous workplan on physical degradation and destruction of coral reefs, including cold-
water corals (CBD, 2004). The ‘cold-water areas’ workplan should link closely with relevant work
under the CBD, including the description of areas meeting the scientific criteria for Ecologically
or Biologically Significant marine Areas (EBSAs) and the UN Food & Agriculture Organisation’s
(FAO) identification of Vulnerable Marine Ecosystems (VMEs).
This process began when the CBD Executive Secretary issued a notification in May 2015
requesting scientific and technical information and suggestions from Parties, other Governments
and relevant organisations. Information was received from Argentina, Australia, Brazil, Canada,
Colombia, France, Mexico, New Zealand, the United Kingdom of Great Britain and Northern
Ireland (UK), the European Union, the International Atomic Energy Agency, the OSPAR
Commission and the UN Division on Ocean Affairs and the Law of the Sea. Alongside updates to
the 2014 CBD Technical Series report on ocean acidification (CBD, 2014), these contributions
were used to prepare a background document for peer review. Following reviews from Canada,
the UK and FAO this background document was considered at the 20th meeting of the CBD’s
Subsidiary Body on Scientific, Technical and Technological Advice (SBSTTA) in April 2016
(CBD, 2016a).
The background document and SBSTTA meeting in April informed discussion at the 13th COP in
December 2016 where Decision XIII/11, a ‘voluntary specific workplan on biodiversity in cold-
water areas within the jurisdictional scope of the Convention’, was formally adopted (CBD,
2016b). COP Decision XIII/11 helps set the wider context for the expert workshop on deep-sea
SDM organised through the European Union’s ATLAS and SponGES projects in May 2018 at the
CBD Secretariat in Montreal, Canada. Compared to other biomes, the need for reliable species
distribution modeling in the deep sea is particularly pressing. This need is because the vast extent,
remoteness, technical challenge and expense of deep-sea research greatly limit the amount of
species and habitat distribution data available. Reliable deep-sea species distribution models offer
a potentially valuable resource to managers struggling to develop policies to manage rapidly
increasing human activities in the deep ocean, including on-going fisheries and hydrocarbon
extraction through to possible future deep-sea mining. However, while the potential of species
distribution modeling is great, there are many issues to be worked through to optimise underlying
species occurrence/absence information, environmental data and the modeling approaches applied
– issues addressed at the Montreal workshop and described in this report.
The background document underpinning COP Decision XIII/11 considered the biodiversity of
deep, cold areas of the global ocean excluding polar regions. Implications of human activities and
global change were summarised, and activities to monitor key parameters and relevant policy

3
instruments were reviewed in that document. It also included analysis of knowledge gaps in both
the scientific evidence-base and the policy instruments available. The background document
highlighted several key points, many of which are now subject to increased research and
management action in several parts of the world:
(a) Cold-water areas sustain ecologically important habitats, such as cold-water corals and
sponge fields, which play important functional biological and ecological roles, including
supporting rich communities of fish as well as suspension-feeding organisms such as
bryozoans and hydroids;
(b) Ocean acidification, increases in ocean temperature and deoxygenation can have
significant impacts on biodiversity and ecosystems in cold-water areas, including
decreased ocean mixing, changes in nutrient cycling and oxygen supply, community shifts,
and impacts on habitat structure and range, and organism physiology;
(c) Ocean acidification, in particular, presents a significant threat to ecosystems in cold-water
areas, including through weakening and dissolution of cold-water coral skeletons, and
impacts on diverse taxa such as sponges, squid species, pteropods, krill and fish. Because
the aragonite saturation horizon is projected to become much shallower by 2100, more
cold-water ecosystems and habitats are expected to be exposed to the impacts of ocean
acidification in the coming years;
(d) There are existing and potential pressures on biodiversity in cold-water areas from
anthropogenic sources, including destructive fishing practices, deep-sea marine mining,
hydrocarbon exploitation, shipping and bioprospecting, as well as impacts related to the
accumulation of plastic microfibers and other pollutants;
(e) Although knowledge on biodiversity and acidification in cold-water areas is growing and
global monitoring of ocean acidification is increasing, there is a need for further research
in this area, including on the interactions among species within food webs, impacts of
ocean acidification on different life stages of cold-water organisms, impacts of multiple
stressors on biodiversity and ecosystems, the goods and services they provide and
variability in the response by organisms to various pressures;
(f) There is a need to address pressures through targeted policy and management responses at
the global, regional and national levels, supported by identification of specific areas that
are of high biological and ecological importance.
Decision XIII/11 encourages Parties to the CBD, other Governments and competent
intergovernmental organizations to:
(a) Avoid, minimize and mitigate the impacts of global and local stressors, and especially the
combined and cumulative effects of multiple stressors;

4
(b) Maintain and enhance the resilience of ecosystems in cold-water areas in order to
contribute to the achievement of Aichi Biodiversity Targets 10, 11 and 15, and thereby
enable the continued provisioning of goods and services;
(c) Identify and protect refugia sites and areas capable of acting as refugia sites, and adopt, as
appropriate, other area-based conservation measures, in order to enhance the adaptive
capacity of cold-water ecosystems;
(d) Enhance understanding of ecosystems in cold-water areas, including by improving the
ability to predict the occurrence of species and habitats and to understand their
vulnerability to different types of stressors as well as to the combined and cumulative
effects of multiple stressors;
(e) Enhance international and regional cooperation in support of national implementation,
building on existing international and regional initiatives and creating synergies with
various relevant areas of work within the Convention.
The activities proposed in Decision XIII/11 included enhanced policy integration, strengthened
management, development and application of Marine Protected Areas, improved monitoring,
improved research coordination and capacity building and initiatives that would secure and sustain
funding.
The explicit inclusion of predictive mapping approaches was expanded in Annex 3 ‘Monitoring
and research needs’ under Subsection 5 to develop or expand upon predictive model research to
determine how projected climate change will impact cold-water biodiversity over different time
scales:
5.1 Improve ocean carbonate models to understand the temporal and three-dimensional spatial
changes in carbonate saturation state and its main drivers, including changing atmospheric
CO2 conditions and ocean currents;
5.2 Document existing gaps in knowledge on global, regional and national scales that limit the
predictive power of models;
5.3 Couple ocean carbonate chemistry mapping and oceanographic models to biophysical and
ecological information to predict the temporal and spatial variability of acidification impacts
in order to help identify areas under the greatest threat as well as possible refugia;
5.4 Optimise habitat modeling to predict key habitats and biodiversity occurrence from seawater
carbonate chemistry, oceanographic and water mass modeling and larval dispersal.
The CBD’s voluntary specific workplan on biodiversity in cold-water areas is a significant
demonstration of how important it has become to not only understand the implications of global
change on marine biodiversity and ecosystem function, but to enhance our abilities to monitor and
make predictions of how marine ecosystems could change. Given the great complexity and

5
interconnected nature of deep and open ocean ecosystems the challenge is great but, as the
discussions in this workshop demonstrate, the underlying data and techniques required for robust
deep-sea predictive species distribution modeling have developed rapidly in the last decade. The
onus is now on the marine scientific and policy communities to take up these challenges and
deliver integrated assessments that allow us to develop a truly predictive ability to forecast not
only where key species occur but how their distributions will change in the future.
Acknowledgements
This paper is a contribution to the European Union’s Horizon 2020 research and innovation
programme under grant agreement No 678760 (ATLAS). It reflects solely the author’s view, and
the European Union and CBD cannot be held responsible for any use that may be made of the
information contained herein.
References
CBD. 2004. Decision adopted by the Conference of the Parties to the Convention on Biological
Diversity. VII/5 Marine and coastal biological diversity.
https://www.cbd.int/doc/decisions/cop-07/cop-07-dec-05-en.pdf
CBD. 2014. An updated synthesis of the impacts of ocean acidification on marine biodiversity
(Eds: S. Hennige, J.M.Roberts and P. Williamson). Montreal, Technical Series No. 75, 99
pages. https://www.cbd.int/doc/publications/cbd-ts-75-en.pdf
CBD. 2016a. Background document on biodiversity and acidification in cold-water areas. CBD-
SBSTTA 20. https://www.cbd.int/doc/meetings/sbstta/sbstta-20/information/sbstta-20-inf-
25-en.pdf
CBD. 2016b. Decision adopted by the Conference of the Parties to the Convention on Biological
Diversity. Voluntary specific workplan on biodiversity in cold-water areas within the
jurisdictional scope of the Convention. CBD-COP 13.
https://www.cbd.int/doc/decisions/cop-13/cop-13-dec-11-en.pdf
2. THEME 1 SHOWCASE OF APPROACHES TO DEVELOP SPECIES
DISTRIBUTION MODELS IN THE DEEP SEA
The workshop opened with two invited speakers who gave keynote presentations relevant to the
workshop theme. The discussion following those presentations is summarized herein.

6
2.1. Model-Based Thinking for Community Ecology
Piers Dunstan
Commonwealth Scientific and Industrial Research Organisation (CSIRO), Oceans
and Atmosphere, Hobart, Tasmania, Australia
One of the key challenges of ecology is to understand the distribution and abundance of species
through space and time. Ecologists have been searching for the most parsimonious explanations
of these patterns for decades. One of the earliest analytical approaches was developed by Bray and
Curtis (1957), and application of similar methods has been extensive in ecology since then.
However, recent developments have demonstrated that these distance-based methods may have
serious flaws (Warton et al., 2011; Woolley et al., 2017) and alternatives are needed.
Fundamentally, when we are observing ecological systems we are observing individuals from
different species – this is the data that can be used in analysis. From these data, we infer
assemblages, communities, bioregions and other higher-level properties. Warton et al. (2015a, b)
suggests that an appropriate approach to this problem is through model-based ecology. The key
concept is that a statistical model can be developed that describes the relationship between the
observations (individuals of species) and the higher-level properties of interest (e.g., assemblages,
bioregions). These models use latent variables to capture the higher-level properties and use well-
described covariates (e.g., environmental) to estimate the functional responses of the latent
variables.
Two examples of this approach are Species Archetype Models (SAM, Dunstan et al., 2011;
Dunstan et al., 2013) and Regions of Common Profiles (RCP, Foster et al., 2014a; Foster et al.,
2017; Hill et al., 2017). Species Archetype Models are multivariate mixture regression models
where species are grouped together based on a common response to environmental covariates.
Species that have a statistically identical response to environmental gradients are grouped together
and a regression model is estimated for each species group. The species groups can then be used
to predict the distribution of the group into space using environmental covariates. The model
retains information on the species in each group and propagates uncertainty from data to
prediction. The models can be estimated with presence only data (Inhomogeneous Poisson Point
Models), presence/absence data (binomial models), abundance data (negative binomial models)
and biomass data (Tweedie models). Examples of SAMs have been on distribution of species in
Western Australia (Woolley et al., 2013) and the response of species to trawl fisheries (Foster et
al., 2014b).
RCP models are similar to SAM models, except that the grouping is on sites rather than species.
The RCP models group sites based on biological content and see how these groups vary with the
environment – these models suppose that assemblages exist. The assumption is that an assemblage
can be characterised by its mean expectation of all species (its profile) and that different
assemblages have different profiles. This assumption means that different assemblages will have

7
different expected biological content. Each site contains one assemblage, which we don't observe,
but can be described as a mixture of sites (characterised by biological profiles). We label this
model as `region of common profiles', but it is also known as a `mixture of experts' model. In a
similar way to SAM models, RCP models can handle a range of different data types (presence
only data (Inhomogeneous Poisson Point Models), presence/absence data (binomial models),
abundance data (negative binomial models) and biomass data (Tweedie models) and propagate
uncertainty from the observations to the predictions. Examples of RCPs are the distribution of
demersal fish on Kerguelen Plateau (Hill et al., 2017) and predictions where sampling artefacts
are present (Foster et al., 2017).
SAM and RCP models often give slightly different predictions as a result of differences in the
assumptions. Assemblages often share species with different expected abundances. This property
means that there will be multiple species groups from SAMs at any site. However, sites are often
the unit of spatial management and distinguishing between sites is an important tool for planning
and management decisions. We expect that SAMs and RCPs will give similar results when there
are very strong environmental gradients acting across the area of interest, which will cause groups
of species to be tightly correlated with each other under specific environmental conditions. As
gradients weaken, more species groups should be present at any site.
Future applications of this work will start to link the outputs from both SAMs and RCP with other
models types, such as qualitative and quantitative ecosystem models.
References
Bray, J.R., and Curtis, J.T. 1957. An Ordination of the Upland Forest Communities of Southern
Wisconsin. Ecol. Monogr. 27(4): 326-349.
Dunstan, P.K., Foster, S.D., and Darnell, R. 2011. Model Based Grouping of Species across
Environmental Gradients. Ecol. Modell. 222: 955-963.
Dunstan, P.K., Foster, S.D., Hui, F.K.C., and Warton, D.I. 2013. Finite Mixture of Regression
Modeling for high-dimensional count and biomass data in Ecology. J. Agric. Biol. Ecol.
Stat. 18(3): 357-375.
Foster, S.D., Givens, G.H., Dornan, G.J., Dunstan, P.K., and Darnell, R. 2014a. Modeling
biological regions from multi-species and environmental data. Environmetrics DOI:
10.1002/env.2245
Foster, S.D., Dunstan, P.K., Althaus, F., and Williams, A. 2014b. The cumulative effect of trawl
fishing on a multispecies fish assemblage in south-eastern Australia. J. Appl. Ecol.
52:129-139.
Foster, S.D., Hill, N.A., and Lyons, M. 2017. Ecological grouping of survey sites when sampling
artefacts are present. Appl. Stat. 66: 1031-1047.

8
Hill, N.A., Foster, S.D., Duhamel, G., Welsford, D., Koubbi, P. and Johnson, C.R. 2017. Model-
based mapping of assemblages for ecology and conservation management: A case study
of demersal fish on the Kerguelen Plateau. Divers. Distrib. 23: 1216-1230.
Warton, D.I., Wright, S.T., Wang, Y. 2011. Distance-based multivariate analyses confound
location and dispersion effects. Methods Ecol. Evol. 3: 89-101. doi: 10.1111/j.2041-
210X.2011.00127.x
Warton, D.I., Foster, S.D., De’ath, G., Stoklosa, J., Dunstan, P.K. 2015a. Model-based thinking
for community ecology. Plant Ecol. 216:669–682.
Warton, D.I., Blanchet, F.G., O’Hara, R.B., Ovaskainen, O., Taskinen, S., Walker, S.C., Hui,
F.K.C. 2015b. So Many Variables: Joint Modeling in Community Ecology. Trends Ecol.
Evol. 30: 766-799. http://dx.doi.org/10.1016/j.tree.2015.09.007
Woolley, S.N.C., McCallum, A.W., Wilson, R., O’Hara, T.D., Dunstan, P.K. 2013. Fathom out:
biogeographical subdivision across the Western Australian continental margin – a
multispecies modeling approach. Divers. Distrib. 19: 1506-1517. DOI:
10.1111/ddi.12119
Woolley, S.N.C., Foster, S.D., O’Hara, T.D., Wintle, B.A., Dunstan, P.K. 2017. Characterising
uncertainty in generalised dissimilarity models. Methods Ecol. Evol. 8: 985-995. doi:
10.1111/2041-210X.12710
2.1.1. Group Discussion
Following Dr. Dunstan’s presentation, there was a discussion comparing and contrasting SAMs
and RCP as approaches to the statistical analysis and prediction of multispecies data. It was agreed
that each have potential benefits, and the more suitable approach depends on the intended use of
the outputs and questions to which answers are being sought. For example, for the purposes of
marine spatial planning where we might want to group sites which share common sets of species
with common probabilities of occurrence, the RCP approach may be more suitable, while for the
purposes of understanding how species and groups of species respond to their environment and
various pressures, SAMs may provide more useful predictions. One can expect these two methods
(grouping by sites and grouping by species) occasionally will give similar results, for example,
when there are strong environmental gradients or strong associations between species. It was
agreed that there were obvious benefits to multispecies modeling approaches; not least of which
is the reduction in the number of models required to describe an ecosystem comprising large
numbers of species. Also, many workshop participants noted that they had often observed multi-
species modeling approaches to have better predictive power to describe the distributions of
individual species than single species models targeted at that species alone. For example, there are
occasions where rare species, for which occurrence data is extremely limited, may be fit to an
archetype thereby improving both the overall archetype model as well as its ability to make

9
predictions for that rare species. It was conceded, however, that where data for a species was so
scant as to impede its being fit to an archetype, it may be better to exclude those data entirely.
SAMs offer improvements over previous multi-species modeling as the processes of species
assembly and model prediction are performed simultaneously. This approach is in contrast to
previously implemented modeling approaches where either 1) species are assembled first and
predictions made based on grouped species, or 2) where model predictions are made first and
multiple predictions assembled thereafter. The primary advantage of this simultaneous
prediction/assembly approach is that users can obtain an estimate of uncertainty – an element that
is too often ignored by other modeling approaches despite its importance to the use of model
predictions in subsequent decision-making processes.
To facilitate implementation of both SAMs and RCP modeling, two R packages - “SpeciesMix”
and “RCPmod” – have been made available on the CRAN repository – these are currently being
combined together into a single interface, “EcoMix”, to make them more user-friendly and to
provide more flexibility. It is envisaged that future work in the area of multispecies modeling will
focus on using SAMs combined with mechanistic models with the goal of building whole
ecosystem-level models which have both descriptive/predictive and explanatory utility.
2.2. Point Process Framework for Species Distribution Modeling and Joint Species
Distribution Models
Jarno Vanhatalo
Department of Mathematics and Statistics and Organismal and Evolutionary
Biology Research Programme, University of Helsinki, Finland
Point process framework for species distribution modeling
Herein, I briefly review the point process modeling approach for species distribution modeling
and illustrate it with few case examples. The benefits from this approach for deep-sea species
distribution modeling could be that they i) cover both occurrence and abundance modeling, ii)
allow integration of heterogeneous data from different survey setups (including presence only),
and iii) can be straightforwardly extended to current joint species distribution models (JSDM). I
will consider log-Gaussian Cox processes (LGCP; Banerjee et al., 2015) whose connection with a
common species distribution modeling method, MaxEnt, was shown by Renner and Warton
(2013). I will assume spatial LGCP, but the same framework can be extended to spatio-temporal
cases.
Let’s denote by  as spatial locations (e.g., longitude-latitude coordinates – and possibly depth),
and  to be the vector of environmental covariates associated with that location (e.g., depth,

10
temperature etc.). A key component of point process models is the intensity function 
which, loosely speaking, is the probability that one individual is present at location . The number
of animals (abundance) in any fixed area/volume of water, , is Poisson distributed with expected
value 
 which is the “total intensity” over . For example, for benthic animals the
intensity function can be directly interpreted as the expected number of animals per unit area (m2),
whereas for pelagic animals this function is interpreted as the expected number of animals per
volume of water (m3). Let’s assume that the sampling effort and, hence, observation probability,
varies in space, and denote this by . In this case the expected number of observed individual
animals in an area is 
 . Now, the sampling effort can be treated as known or
as an uncertain parameter of the model. In many cases the observation probability is not known
but can be assumed to be constant, in which case  corresponds to relative intensity.
In practice LGCPs are typically discretized so that observations are considered to be made over
finite areas (e.g., discrete lattice grid), over which the intensity is treated to be constant (Renner
and Warton, 2013; Banerjee et al., 2015). Hence, given a set of species observations  over
a set of  finite areas indexed with locations   , the likelihood for an intensity and
effort function is
  


where  is the (center) location of ’th sampling area, and  is the size of that area (or volume of
water). See Figure 2.2.1 for an illustration. To finalize the species distribution model we build a
model for the intensity function which is typically constructed using a log linear model with spatial
random effects so that
     
where the intercept  corresponds to the average log intensity,  describes the effect of
environmental covariates and  is a spatial random effect that explains spatially correlated
patterns not explained by environmental covariates. The spatial random effect is modelled with
zero mean Gaussian process (Banerjee et al., 2015), and it has been shown to improve the
explanatory and predictive performance of species distribution models in many studies (Latimer
et al., 2006; Vanhatalo et al., 2012; Clark et al., 2014; Kallasvuo et al., 2017). The linear
environmental predictors can also be replaced by non-linear functions (e.g., Kallasvuo et al.,
2017). If we assume independent random noise per sampling location/area, the Poisson likelihood
extends naturally to a Negative-Binomial model that allows for over-dispersion (Vanhatalo et al.,
2017).

11
Hence, from an analysis point of view, simple point process models are generalized linear models.
However, the added value is that their interpretation is ecologically meaningful, and allows
integration of alternative data sets and prediction of relative biomasses.
Figure 2.2.1. A schematic illustration of a point process model implemented over a lattice grid.
The survey data has resulted in a table of count observations at surveyed cells (  ). After
solving the posterior distribution of model parameters, we can predict the species intensity in cells
not surveyed.
Bayesian predictive inference and example applications
Relative effect on intensity
Given training data, we can calculate the posterior distribution (Banerjee et al., 2015) of the model
parameters, examine the effect of environmental covariates and use the model for making
predictions in locations not covered by data. Due to the log-link function, the additive model
components have particularly simple interpretation as log relative change in intensity (Vanhatalo
et al., 2017). For example, the relative effect to intensity by the spatial random effect is:
  
 
    
which tells how much larger (>0) or smaller (<0) the density of a species is at location  compared
to a density estimate based only on (environmental) covariates. Similarly, the relative effect of the
’th covariate is  regardless of the other covariates or spatial locations.
Vanhatalo et al. (2017) used LGCP to analyze spatial-temporal fluctuations of Crown of Thorns
Starfish (COTS) in the Great Barrier Reef, Australia. The log intensity included spatial and spatio-
temporal random effects. The discrete sampling areas  varied by observations and corresponded
to the surveyed reef area. One of the key results was the visualization of the spatial and spatial-
Long
Lat
1
2
3
12 3
XX
Id Long Lat Effort Count
observation
1 1 1 1 0
2 1 2 2 0
3 2 2 2 2
4 3 2 1 3
5 3 3 1 0
z=0
z=0
z=0
Survey
transect
XX
Individual observations
observations along transect
X
Count data from survey transect
Likelihood
Prediction for unsurveyed grid cell

12
temporal relative effects, which revealed high/low intensity COTS hot/cold spots and COTS
outbreak dynamics (Figure 2.2.2).
Bayesian predictive inference and biomass estimates
We can use point process models for (relative) biomass predictions. Kallasvuo et al. (2017)
analyzed larval areas of four commercially important coastal fish species. Each of their
observations corresponded to number of fish larvae in a volume of water, , sampled by sampling
nets and, hence, the intensity function  corresponded to relative larvae density (number
per m3 multiplied by catch probability). Based on the relative density predictions, Kallasvuo et al.
(2017) divided the study region into three classes: important regions (the highest density areas
producing 80% of total number of larvae), suitable regions (regions with >50% probability to
observe larvae) and non-suitable regions (regions with <50% probability to observe larvae). In
practice this analysis was done by predicting  in approximately 15 million 50 m  50 m grid cells
and calculating the total (relative) number of larvae in the study region,   where
the sum goes through all the grid cells and  is the volume of water (surface layer) in the grid cell.
The highest intensity grid cells, whose total relative number of larvae was  , comprise
the important larval production region. One of the key findings of their study was that the
important regions can be much smaller in size than regions suitable for larval production.
Moreover, the difference between important and suitable areas varied dramatically between
species, which shows that using only presence/absence models to predict species distributions
might give a biased view of the important areas (see Figure 2.2.3).
Accounting for unequal sampling effort
For a last example, we consider integration of heterogeneous data and modeling the survey effort.
Mäkinen and Vanhatalo (2018) conducted marine mammal distribution modeling in the Arctic
using heterogeneous data collected from earlier publications and free data sources. The data were
comprised of species sightings made during survey cruises, but the exact survey protocol during
the surveys was unknown, for which reason they also explicitly modeled the survey effort, . In
their model, each survey had its own, unknown effort, which was considered constant throughout
the study region and was given a log-Gaussian prior distribution. Other examples of analyses with
non-constant survey effort are presented by Renner and Warton (2013) and Warton et al. (2013),
who consider modeling presence only data. In their applications, the effort is a function of
environmental covariates and spatial location,  Yuan et al. (2017) consider LGCP models
for transect surveys where the observation probability decreases with distance from survey vessel.

13
Figure 2.2.2. Illustration of spatial (left) and spatial-temporal (right) random effects reported as
relative change/effect in intensity in the Great Barrier Reef, Australia. The spatial-temporal
random effect is projected on the reef system’s center line. In plot c) the outbreak corresponds to
a 2-fold increase in Crown of Thorns starfish abundance relative to the local mean abundance.
Reproduced with permission from (Vanhatalo et al., 2017).
Figure 2.2.3. Illustration of abundance predictions and classification of spatial regions based on
their relative importance to fish production. Reproduced with permission from (Kallasvuo et al.,
2017).
Total number of larvae
in
surface water, 109
mean (95% credible
interval)
Percent of
water
area
suitable
for larvae
Percent of
water area
producing
80% of
larvae
perch
1.56 (0.89, 2.55)
13.66
3.03
pikeperch
0.54 (0.12, 1.56)
3.68
1.37
Baltic
herring
8.72 (5.65, 12.86)
99.79
52.89
smelt
5.91 (2.88, 10.81)
22.50
4.44

14
Joint species distribution models
Joint species distribution models (JSDM) have gained increasing interest in recent years (Warton
et al., 2015). The key assumption behind most of the JSDMs is that the model should include
interspecific dependence between species-specific environmental effects and spatial random
effects. In our context, extension to JSDMs corresponds to denoting the species-specific log
intensity with:
  
where  denotes a species and gives the species-specific parameters a hierarchical prior/model that
introduces the dependence. For example, in the hierarchical model of species communities
(HMSC, Ovaskainen et al., 2017), the species-specific linear weights are given a Gaussian prior
, where  is the mean vector and  the covariance matrix. The interspecific
dependence is then included by, for example, modeling the mean vector as a function of species
traits; that is  where  is the ’th trait of the species and  are the trait weights. The
HMSC allows also for a generic multivariate Gaussian prior for , as well as inclusion of
phylogenetic information into the correlation between . In species archetype modeling the
responses of the species to the environment are clustered into a few archetype models
corresponding to shared  between group of species (Dunstan et al., 2013; Hui et al., 2013).
Similarly, the spatial random effects are extended to include interspecific correlations. Vanhatalo
et al. (manuscript) extended point process based JSDMs to semiparametric models, where the
effects of environmental covariates are allowed to follow a semiparametric Gaussian process
model.
Key advantages of JSDMs compared to single species models can be summarized as follows. The
inclusion of interspecific dependence allows information flow between species, which improves
the estimates for covariate effects especially for species with only scarce data (e.g., Ovaskainen et
al., 2017; Hui et al., 2013; Clark et al., 2014). From the predictive point of view, this approach has
added benefit that the models predictive accuracy also improves in both interpolation (Ovaskainen
et al., 2017; Hui et al., 2013) and extrapolation (Vanhatalo et al., manuscript). JSDMs have also
been demonstrated to more accurately estimate the joint distribution of multiple species, whereas
single species models typically predict larger distribution areas than JSDMs (see Figure 2.2.4).
Moreover, the distribution of species is not governed only by their environment and stochastic
processes and species interactions play a significant role in the realized distribution (Ovaskainen
et al., 2017). The interspecific correlations in the spatial random effects account also for the
species-to-species interactions. By examining the interspecific correlation matrices, we can gain
understanding of the similarities in the response of species to the environment, and of species-to-
species interactions (see Figure 2.2.5 for illustration).

15
Figure 2.2.4. Abundance predictions with single and joint species distribution models for the
three-spined stickleback in the Gulf of Bothnia, northern Baltic Sea. The JSDM typically predicts
smaller distribution ranges. Reproduced with permission from (Vanhatalo et al., manuscript).
Figure 2.2.5. Interspecific correlations between the fixed effects, , (first 7 plots) and spatial
random effect. Reproduced with permission from (Vanhatalo et al., manuscript).

16
Discussion
Point process models are classical tools in spatial statistics as well as in some areas of applied
ecology. In recent years, they have gained increasing attention in the general species distribution
modeling literature as well. In this working paper, I discussed their connection to traditional
generalized linear model-based species distribution models, as well as to current developments in
joint species distribution models. I also highlighted some of their useful properties from a practical
modeling viewpoint. Namely, the point process framework provides results that are directly
interpretable as (relative) species densities.
The case studies reviewed in this work included classical count data observed either visually or
collected with fishing nets. However, an interesting area for future development would be to
extend the methods to other types of data as well. In the deep-sea context this could include, for
example, combining trawling and visual data with acoustic survey data. Juntunen et al. (2012)
proposed a Bayesian method to integrate trawling and acoustic survey data to estimate species
composition and abundance (biomass) of multiple fish in the Baltic Sea. Their approach does not
follow the point process framework as such, but is very closely related to it, and could in principle
be extended to it as well.
References
Banerjee, S.Carlin, B. P., and Gelfand, A. E. 2015. Hierarchical Modeling and Analysis for Spatial
Data. Second Edition. CRC Press/Chapman & Hall. Monographs on Statistics and Applied
Probability 135, Boca Raton, Florida, 562 pp.
Clark, J.S., Gelfand, A.E., Woodall, C.W. and Zhu, K. 2014. More than the sum of the parts: forest
climate response from joint distribution models. Ecol. Appl. 24: 990-999.
Dunstan, P. K., Foster, S. D., Warton, D. I. and Hui, F. K.C. 2013. Finite mixture of regression
modeling for high-dimensional count and biomass data in ecology. J. Agric. Biol. Environ.
Stat. 18: 357-375.
Hui, G.K.C., Warton, D.I., Foster, S.D. and Dunstan, P.K. 2013. To mix or not to mix: comparing
the predictive performance of mixture models vs. separate species distribution models. Ecol.
94: 1913-1919.
Juntunen, T., Vanhatalo, J., Peltonen, H., and Mantyniemi, S. 2011. Bayesian spatial multispecies
modeling to assess pelagic fish stocks from acoustic- and trawl-survey data. ICES J. Mar. Sci.
69: 95–104. https://doi.org/10.1093/icesjms/fsr183
Kallasvuo, M., Vanhatalo, J., and Veneranta, L. 2017. Modeling the spatial distribution of larval
fish abundance provides essential information for management. Can. J. Fish. Aquat. Sci. 74:
636-649.

17
Latimer, A., Wu, Shanshan, Gelfand, A.E. and Silander, J. 2006. Building statistical models to
analyze species distributions. Ecol. Appl. 16: 33-50.
Mäkinen, J. A.-E., and Vanhatalo, J.P. 2018. Hierarchical Bayesian model reveals the distributiona
shifts of Arctic marine mammals. Divers. Distrib. 2018: 1-14;
https://doi.org/10.1111/ddi.12776
Ovaskainen, O., Tikhonov, G., Norberg, A., Blanchet, F.G., Duan, L., Dunson, D., Roslin, T. and
Abrego, N. 2017. How to make more out of community data? A conceptual framework and
its implementation as models and software. Ecol. Lett. 20: 561-576.
Renner, I.W., and Warton, D.I. 2013. Equivalence of MAXENT and Poisson point process models
for species distribution modeling in ecology. Biometrics 69: 274-281.
Vanhatalo, J., Veneranta, L., and Hudd, R. 2012. Species distribution modeling with Gaussian
processes: a case study with the youngest stages of sea spawning whitefish (Coregonus
lavaretus L. s.l.) larvae. Ecol. Model. 228: 49-58.
Vanhatalo, J., Hosack, G., and Sweatman, H. 2017. Spatio-temporal progression of outbreaks of
the crown-ofthorns starfish on the Great Barrier Reef, 1985–2014. J. Appl. Ecol.54: 188-
197.
Vanhatalo J., Hartmann, M. and Veneranta, L. (manuscript). Joint species distribution modeling
with additive multivariate Gaussian process priors and heteregenous data. arXiv:1809.02432v1
Warton, D.I., Blanchet, F.G., O-Hara, R.B., Ovaskainen, O., Taskinen, S., Walker, S.C., and Hui,
F.K.C. 2015. So many variables: joint modeling in community ecology. Trends Ecol. Evol.
30: 766-779.
Warton, D. I., Renner, I. W., and Ramp, D. 2013. Model-based control of observer bias for the
analysis of presence-only data in ecology. PLoS ONE, 8(11).
https://doi.org/10.1371/journal.pone.0079168
Yuan, Y., Bachl, F. E., Lindgren, F., Brochers, D. L., Illian, J. B., Buckland, S. T., Ruie, H., and
Gerrodette, T. 2017. Point process models for spatio-temporal distance sampling data from a
large-scale survey of blue whales. Ann. Appl. Stat. 11(4), 2270–2297.
https://doi.org/10.1214/16-AOAS1011
2.2.1. Group Discussion
The discussion following Dr. Vanhatalo’s presentation examined issues surrounding specifics of
species distribution modeling and joint species distribution modeling (JSDM), as well as exploring
the user-friendliness of various models, and how this attribute affects model usefulness. Dr.
Vanhatalo posited that the point process approach described in his presentation was in essence no

18
more complicated than a standard GLM approach in R, and felt that any user who could use the
latter would have no difficulty in implementing the former. An instructional/guidance document
complete with sample data and worked examples will be made publicly available to improve the
ease with which interested parties might adopt the point process approach described in Dr.
Vanhatalo’s talk.
Point process models can be very useful where a user wants to make use of data with high degrees
of uncertainty to improve modeling performance. For example, data from acoustic monitoring
equipment can be helpful in detecting how many distinct species of fish are in a given area
(however, further data would be required to try to identify those species etc.). Observational data
(e.g., data from acoustic monitoring) has quantitative aspects, which can aid in analysis, and the
amount of information with these data reduces the residual variance because more aspects of these
data are explained. The discussion also focused on how the MaxEnt (maximum entropy) model
approach compares with the spatial point process. The point was raised that many scientists have
short timeframes to produce results and MaxEnt is a very user-friendly option of model. Other,
more complicated models are being developed but are not as easy to use. The discussion indicated
that, if MaxEnt is used properly, it can return excellent results and can produce a similarly smart
analysis. However, there are benefits to investing time in learning more about model-based
approaches because in the long term they will likely give more information alongside the model
output. Following this discussion of MaxEnt’s user-friendly nature, the point was raised that no
matter which type of model is being used, it is still being written in R as a type of Generalized
Linear Model (GLM) at the base level. The challenge arrives once the user goes beyond R
packages and focuses on more complex data types and spatial correlations. It is important to
consider the analysis before fieldwork is completed because often datasets do not correspond well
with the analysis that will be done. One method of helping users become more familiar with these
methods is producing test datasets alongside publications, which can be used as trial projects for
users and gives them the tools to do such analyses properly. It was agreed that the provision of
trial datasets is an important aspect of increasing the uptake of new complex methods.
3. EXTENDED ABSTRACTS
Participants were invited to provide extended abstracts for this publication on a topic of their
choice. Some authors volunteered to present their subject more formally in the Participants Forum
on Day 1 (Appendix 1), and the topic was discussed further by the group. For those abstracts a
summary of the group discussion follows the extended abstract. The formatting follows the
original style of the submission, and figures are numbered consecutively within each abstract with
the numbering restarting with the next.

19
3.1. What do Environmental Managers and Stakeholders want from a Model?
Ashley A. Rowden
National Institute of Water & Atmospheric Research, Wellington, New Zealand
Human activities are increasingly impacting the deep-sea environment, and there is a need to
manage the effects of these impacts on deep-sea species. To design effective spatial management
measures to conserve and protect deep-sea species it is necessary to know their distribution
patterns. Environmental managers and stakeholders understand that it is necessary to use
numerical models to determine distribution patterns because of the paucity of deep-sea species
records. However, the acceptance and use of species distribution models in spatial management
planning can depend on several factors that model-building scientists don’t always perceive or
address. Recent experience producing species distribution models for Vulnerable Marine
Ecosystems (VMEs), as part of a process to design management measures for deep-sea bottom
trawling in the South Pacific, has highlighted what environmental managers and stakeholders want
from a model. These need to include the following: A model should to be relevant; models for
species that act as proxies or indicators for VMEs are not as acceptable as models that attempt,
even coarsely, to model the ecosystem feature itself (e.g., coral reef). A model should be made at
the appropriate spatial scale; because it is possible, models are sometimes made at a resolution
that is smaller or bigger than the use to which they will be put, but this impacts upon their
usefulness and acceptance (e.g., making a model at 1 km2 resolution when spatial closures will be
applied at a scale of >1000 km2). The modeling approach should be conservative and
parsimonious; multiple model or ensemble approaches are preferred over reliance on a single
model approach (even if considered by scientists to be the latest and best), and models using fewer
predictor variables are more easily understood and accepted. A model should to be believable;
field validation of a model is highly valued (even for a small portion of the modelled area), and
the more broadly understood metrics used for internal model validation are preferred over those
that are more model-approach specific (e.g., correlation metrics versus AUC). A model’s
limitations must be expressed spatially; quantifying and mapping the spatial uncertainty of a model
helps acceptance (acknowledgment that model not equally good everywhere), and is a particularly
useful output for designing spatial management measures. In the main, these needs are similar or
the same as the concerns of a model-building and evaluating scientist, but there are subtle
differences in viewpoint that impact acceptance of models by managers and stakeholders. This
presentation will expand upon these differences with the aim of assisting in the making of models
and their outputs which are more readily accepted and used by environmental managers and
stakeholders for spatial management planning.
3.1.1. Group Discussion
The discussion following this talk focused on concerns arising from designating vulnerable marine
ecosystems (VMEs) and which thresholds should be used to quantify these boundaries (e.g.,

20
where a VME indicator species is at a “significant concentration” that represents a VME). A
comment was raised questioning how best to move from a species approach to a VME approach
and whether the threshold of abundance of a VME indicator species should be used.
A second comment questioned the possibility of jointly modeling species and traits, and whether
this can be used to designate VMEs. This approach would be useful because the FAO Guidelines
for the identification of VME indicators provides a list of traits that such species should possess.
Finally, it was suggested that model predictions for species habitat should (when published) be
compared to areas where stakeholders and managers want to ban fishing. These maps can often
show different areas, and for this reason, communication between model creators, managers, and
stakeholders has to be increased to ensure that management decisions are properly informed by
the best available information.
3.2. Rough Data: Imperfect Data and Directional Rugosity
Cherisse Du Preez, Emily Rubidge and Jessica Finney
Pacific Region, Fisheries and Oceans Canada (DFO)
This working paper covers two topics applicable to species distribution models (SDMs) in the
deep sea: imperfect data and seafloor roughness (rugosity). The following is based on the
experiences of Fisheries and Oceans Canada scientists who modeled vulnerable marine
ecosystems and ecologically or biologically significant areas in the deep Northeast Pacific.
Imperfect data and transparency about it
Imperfect detection
The probability of detecting a species at a site when it is present is calculated through repeated
sampling of the same site, and combined with statistical models to estimate species’ occupancy.
Although factoring in the probability of detection is standard in terrestrial ecology, it has proven
impractical in the deep sea. Why? Cost, time, accessibility… If you haven’t already, may I suggest
reading “How long should we ignore imperfect detection of species in the marine environment
when modeling their distribution?” by Jacquomo Monk (2014). In response to ideas like those
discussed by Monk, a team of DFO researchers are planning a repeat survey study using remotely
operated vehicles and visual benthic surveys.
Uncertainty in extrapolations vs. interpolations
It is common practice to model beyond the range of data sampled for one or more environmental
variables, be it carefully and within reason. A shaded overlay on an uncertainty map is often used
to identify such extrapolations (Figure 1). An extrapolated area analyses is a standard SDM output,

21
but there isn’t an equivalent for interpolation, to identify data gaps within a range sampled. During
the 2012 Cobb Seamount expedition, the team sampled between 34 and 1154 m depth, but
technical issues prevented them from surveying between 211 and 472 m depth. The ~250 m gap
occurred over problematic depths, at the transition region from flat plateau to steep flanks, and the
known narrow depth ranges of unique cold-water coral gardens and reefs and dense lost-fishing
gear (discussed in Du Preez et al., 2016). To clearly identify the unreliable interpolation within
the range of data sampled, the team manually added a hatching overlay to their uncertainty map
(Figure 1).
Figure 1. Outputs from the Random Forest model on Cobb Seamount illustrating the extrapolated
areas (grey shading) and the unreliable interpolated area (i.e., 211 to 472 m depth gap not
surveyed; hatching) (Du Preez et al., 2016).
Combination models
If you are fortunate enough to have density input data for modeling, consider the value of applying
a presence probability threshold to your density predictions (e.g., only display predicted density
if output presence probability is > 0.5). While the drawback is that densities will not be displayed

22
over the full study area, the advantage is that they will only be displayed where there is a higher
probability that the species is present. A similar result can be achieved by applying an uncertainty
threshold (a common species distribution modeling output).
Directional rugosity
Seafloor roughness (or complexity) is a common environmental variable used in species
distribution models. There are many metrics for measuring seafloor roughness. One useful metric
is the arc-chord ratio (ACR) (Du Preez, 2015), which is part of the ESRI ArcGIS Benthic Terrain
Modeller toolbox (Walbridge et al., 2018). Regardless of the specific roughness metric used, scale
and directionality are two important factors to consider. The former is a fairly common
consideration, while the latter is often overlooked, especially when the 3-D bathymetric
geoprocessing uses a moving window or circular polygon. In environments with dominant flow
directionality (uni- or bi-directional), using a linear metric of rugosity upstream may provide a
strong indicator for species distributions and other biological characteristics (e.g., biodiversity)
downstream (Figure 2). Physical mechanisms may include changes in pressure, dissolved gases,
shear stress, particulate load, entrained organisms and larvae, and water stratification. Changes in
the local bottom flow regime (e.g., changes to flow velocity, energy, and directionality) affect
epibenthic communities through larval delivery and recruitment, delivery of oxygen and nutrients,
feeding opportunities, removal of waste, the passive collection or dispersal of organisms,
suspension and deposition of sediment (i.e., available substratum and turbidity), scouring and
erosion of sediment and of organisms, and levels of biotic and abiotic disturbance (likely
influencing sessile fauna more than mobile fauna). Incorporating directionality into roughness
metrics may improve the predictive power of species distribution models over conventional
rugosity metrics. Directional linear ACR rugosity will be released within the ESRI ArcGIS
Benthic Terrain Modeller soon.

23
Figure 2. In a study comparing different scales, dimensions, and directionality, large-scale linear
rugosity upstream was found to be the best predictor of local benthic biodiversity (Du Preez,
2014).
References
Du Preez, C. 2014. Resolving relationships between deep-sea benthic diversity and multi-scale
topographic heterogeneity (Doctoral dissertation, Biology Dept., University of Victoria,
British Columbia, Canada). https://dspace.library.uvic.ca/handle/1828/5828?show=full
Du Preez, C. 2015. A new arc–chord ratio (ACR) rugosity index for quantifying three-dimensional
landscape structural complexity. Landscape Ecol. 30: 181-192.
Du Preez, C., Curtis, J.M. and Clarke, M.E. 2016. The structure and distribution of benthic
communities on a shallow seamount (Cobb Seamount, Northeast Pacific Ocean). PloS
One 11(10), p.e0165513.
Monk, J. 2014. How long should we ignore imperfect detection of species in the marine
environment when modeling their distribution? Fish Fish.15: 352-358.

24
Walbridge, S., Slocum, N., Pobuda, M. and Wright, D.J. 2018. Unified geomorphological analysis
workflows with Benthic Terrain Modeler. Geosci. 8: 94.
3.3. Determining the best methods for model validation
Chris Rooper*, Rachel Wilborn and Pam Goddard
Alaska Fisheries Science Center, National Marine Fisheries Service, NOAA,
Seattle, Washington, USA
*Current affiliation: Pacific Biological Station, Fisheries and Oceans Canada,
Nanaimo, British Columbia, Canada
Model validation for species distribution modeling usually takes one of three forms; resubstitution
or resampling, testing with new or held back data, or testing with independent data collected from
a different area or an independent survey. The general purpose of model validation exercises is
most often to determine the robustness of the model and the confidence that can be placed in a
model’s predictions.
Within-sample model validation (resubstitution or resampling) is the most commonly used method
for model evaluation, typically as a leave-one-out analysis or a k-fold cross-validation, where the
effects of different partitions of the modeled data on the results are examined. Testing with held
back data from the same modeled area (either from a new time or a random selection of data) is
also fairly common. Transferring the model to a new area or a newly collected data set designed
to test model predictions are fairly uncommon methods, possibly because of the associated high
cost of additional sampling. The literature (and practical experience) indicates that predicting an
entirely new data set from a new area is probably the most difficult test of any species distribution
model.
Models are usually evaluated using the same criteria as the initial models (e.g., using AUC, TSS
and R2). An example of a model validation is shown in Figure 1. These results are from SDMs of
structure-forming invertebrates in the Aleutian Islands of Alaska. Predictions of species
distributions were initially developed using Generalized Additive Models and bottom trawl survey
data. The models were tested on a year of data that was held back (2012). The results of this testing
were good for both presence-absence models and abundance models (with the exception of
Stylasteridae).

25
Figure 1. AUC values and R2 values for coral and sponge models developed using bottom trawl
survey data in the Aleutian Islands. Test data was a year (2012) that was held back for testing
purposes and camera data was an independent camera survey completed in 2014. Details of these
results can be found in Rooper et al. (2014) and Rooper et al. (2018). Sponges are combined
classes (Hexactinellida and Demospongiae) and corals include the order Antipatharia, suborders
Holaxonia (families Plexauridae, Acanthogorgiidae), Calcaxonia (families Primnoidae and
Isididae), Scleraxonia (family Paragorgiidae), family Paramuriceidae, and hydrocorals from the
family Stylasteridae.
In 2012 and 2014 an independent validation survey was conducted at 216 randomly selected sites
in the Aleutian Islands. These data used a different gear (underwater camera) and surveyed a
somewhat different area, as camera transects were able to be conducted in rough, rocky areas
where the bottom trawl could not be deployed. The results were still adequate for presence absence
and abundance models of coral. However, the ability of the bottom trawl survey models to predict
sponge presence or absence and abundance in the camera survey was not good. This result
demonstrates that challenging a species distribution model with new and different data can

26
sometimes reduce the confidence in its performance relative to the performance on resampled or
held back data.
Some potential questions for discussion are:
 What are the best methods for testing species distribution models (especially when
collecting new data is not feasible)
 Are there better ways to evaluate model performance in validation exercises?
 How good does model performance in validation exercises need to be in order to accept a
model?
References
Rooper, C.N., Wilborn, R.E., Goddard, P., Williams, K., Towler, R., and Hoff, G.R. 2018.
Validation of deep-sea coral and sponge distribution models in the Aleutian Islands,
Alaska. ICES J. Mar. Sci. 75:199-209.
Rooper, C.N., Zimmermann, M., Prescott, M., and Hermann, A. 2014. Predictive models of coral
and sponge distribution, abundance and diversity in bottom trawl surveys of the Aleutian
Islands, Alaska. Mar. Ecol. Prog. Ser. 503:157-176.
3.3.1. Group Discussion
Dr. Rooper’s presentation explored three main validation techniques: resubstitution or resampling,
testing with new or held back data, or testing with independent data collected from a different area
or independent survey. The discussion following this presentation focused mainly on bias inherent
with different types of data collection and model validation. To begin, it was noted that one must
use the same method of data collection to both test and train the data (for example camera data,
trawl data, etc.). Next, trawl data, ROV (Remotely Operated Vehicle) data, drop cameras and other
methods of data collection all come with biases in terms of the area they describe. Validating
models with ROV data is problematic due to the data often being comprised of a 500 m by 500 m
cell, while the ROV moves only through the middle of this area, therefore missing large amounts
of seafloor. Validating models with drop camera data was mentioned as resulting in less bias for
habitat mapping than ROVs due to their blind sampling, however, drop cameras stop working well
on steep slopes and potentially in other areas such as very soft sediments where plumes can be
created. This limitation may result in a similar bias if such areas are included in the study area,
because it limits possible deployment areas.

27
3.4. Distribution Models Applied to Climate Change in the Deep Sea. A Promising but
Challenging Development Field
José Manuel González-Irusta and Telmo Morato
Departamento de Oceanografia e Pescas, Universidade dos Açores
PT-9901-862 HORTA, Portugal
Species distribution models are being currently used to inform marine spatial planning worldwide,
modeling the distribution of relevant species (González-Irusta et al., 2015; Greathead et al., 2015;
Parra et al., 2016), the distribution of benthic assemblages as a proxy to biological habitats
(Serrano et al., 2017) or the abundance of structural species such as Lophelia pertusa as a proxy
to the habitat itself (Howell et al., 2011; Fernandes et al., unpublished). These models have also
been successfully used to delineate essential fish habitats of commercial fish species, modeling
the abundance of juveniles (nursery areas, Aires et al., 2014; Asjes et al., 2016), the abundance of
adults at their spawning stage (spawning grounds, González-Irusta and Wright, 2016a, 2016b,
2017) and fish egg distribution (e.g., Loots et al., 2011; Lelièvre et al., 2014). More recently, these
models have been combined with other type of approaches to answer different questions than
species presence or habitat location. For instance, distributions have been combined with
Biological Traits Analysis and effort maps (from Vessel Monitoring Systems) to determine and
map the impact of trawling disturbance on sensitive species (González-Irusta et al., 2018). They
have also been combined with ecosystem simulation models to produce maps of ecosystem
production across entire ecosystems such as the Gulf of Mexico (Grüss et al., 2018), and their use
combined with particle tracking analysis in connectivity studies has been already proposed (e.g.,
Gallego et al., 2016). Finally, these models have been combined with future climatic scenarios to
predict distribution changes for different marine species (e.g., Tittensor et al., 2010; Sequeira et
al., 2014; Gallego et al., 2017). The combination of species distribution models with other
techniques (such as biological traits analysis or community analysis), and its use to predict climate
change impacts on marine biodiversity, are probably two of the most promising fields for the
implementation of the ecosystem approach into the management of the marine ecosystems. In the
framework of the ATLAS project, we are currently working on the use of these models to predict
changes in future distributions of cold-water coral species and fish under different climate change
scenarios at a North Atlantic basin scale. Although some of the challenges we are facing are
common to most of the model exercises (lack of biological data, sampling bias, poor resolution in
some environmental layers, etc.) others are specific to the application of these models to predict
future changes in distribution. One of the challenges is the environmental variable selection and
its impact on the models’ sensitivity to future changes. One example of this issue, especially
important for the deep sea, is the inclusion or not of depth into the models. In fact, depth is not an
environmental variable stricto sensu but usually is the variable with the most accurate information,
and a very good proxy for several other relevant variables (such as temperature near bottom,
salinity, light intensity, pressure, etc). To include depth in the models will improve them in most

28
cases, but can also reduce the sensitivity of these models to future changes in environmental
conditions. A better understanding of the relationship between the modelled species and depth is
a key factor to make the right decision. Another important challenge is the lack of appropriate
tools to predict the accuracy of these models to extrapolate their present prediction into the future.
Current evaluation methods such as AUC and TSS are controversial even for present predictions
(Lobo et al., 2008; Jiménez-Valverde, 2014). Fourcade et al. (2018) has recently demonstrated the
inability of current evaluation metrics to assess the biological significance of the relationship
between the response variable and the explanatory variables in distribution models, if they are not
based on a solid knowledge of the species and its ecological niche. This issue is especially relevant
when these models are used to extrapolate present predictions into the future based on these
relationships. Species distribution models can predict accurately the distribution of habitat and
species even based on spurious relationships, but they will only be able to predict future changes
if the model is really based on the ecological niche of the species. A sound knowledge of the
environmental requirements of the species, the support of experimental work on this topic, and the
inclusion of the whole range of environmental variability for the target species are key for
producing a correct distribution model that is able to predict future changes in species distribution
under different climate change scenarios.
References
Aires, C., González-Irusta, J.M., and Watret., R. 2014. Updating Fisheries Sensitivity Maps in
British Waters. Scottish Marine and Freshwater Science Vol 5 No 10. 88 pp. DOI:
10.7489/1555-1.
Asjes, A., González-Irusta, J. M., and Wright, P. J. 2016. Age-related and seasonal changes in
haddock Melanogrammus aeglefinus distribution: Implications for spatial management.
Mar. Ecol. Progr. Ser. 553: 203-217.
Fernandes, P. G., McIntyre, F.D., González-Irusta, J.M., and Neat, F.C. (2018). Estimating the
abundance of deep cold-water coral: the Northeast Atlantic’s great barrier reef. Unpublished
Fourcade, Y., Besnard, A.G., and Secondi, J. 2018. Paintings predict the distribution of species,
or the challenge of selecting environmental predictors and evaluation statistics. Global Ecol.
Biogeogr 27: 245-256.
Gallego, A., Gibb, F. M., Tullet, D., and Wright, P. J. 2016. Bio-physical connectivity patterns of
benthic marine species used in the designation of Scottish nature conservation marine
protected areas. ICES J. Mar. Sci. 74: 1797–1811.
Gallego, R., Dennis, T. E., Basher, Z., Lavery, S., and Sewell, M. A. 2017. On the need to consider
multiphasic sensitivity of marine organisms to climate change: a case study of the Antarctic
acorn barnacle. J. Biogeography 44: 2165–2175.

29
González-Irusta, J. M., González-Porto, M., Sarralde, R., Arrese, B., Almón, B., and Martín-Sosa,
P. 2015. Comparing species distribution models: A case study of four deep sea urchin
species. Hydrobiologia 745: 43–57.
González-Irusta, J. M., and Wright, P. J. 2016a. Spawning grounds of Atlantic cod (Gadus
morhua) in the North Sea. ICES J. Mar. Sci. 73: 304–315.
González-Irusta, J. M., and Wright, P. J. 2016b. Spawning grounds of haddock (Melanogrammus
aeglefinus) in the North Sea and West of Scotland. Fish. Res., 183: 180–191.
González-Irusta, J. M., and Wright, P. J. 2017. Spawning grounds of whiting (Merlangius
merlangus). Fish Res. 195: 141–151.
González-Irusta, J. M., De la Torriente, A., Punzón, A., Blanco, M., and Serrano, A. 2018.
Determining and mapping species sensitivity to trawling impacts: the BEnthos Sensitivity
Index to Trawling Operations (BESITO). ICES J. Mar. Sci.
https://doi.org/10.1093/icesjms/fsy030
Greathead, C., González-Irusta, J. M., Clarke, J., Boulcott, P., Blackadder, L., Weetman, A., and
Wright, P. J. 2015. Environmental requirements for three sea pen species: relevance to
distribution and conservation. ICES J. Mar. Sci. 72: 576–586.
Grüss, A., Drexler, M. D., Ainsworth, C. H., Babcock, E. A., Tarnecki, J. H., and Love, M. S.
2018. Producing Distribution Maps for a Spatially-Explicit Ecosystem Model Using Large
Monitoring and Environmental Databases and a Combination of Interpolation and
Extrapolation. Front. Mar. Sci. 5: 1–20.
Howell, K. L., Holt, R., Endrino, I. P., and Stewart, H. 2011. When the species is also a habitat:
Comparing the predictively modelled distributions of Lophelia pertusa and the reef habitat
it forms. Biol. Cons. 144: 2656–2665.
Jiménez-Valverde, A. 2014. Threshold-dependence as a desirable attribute for discrimination
assessment: implications for the evaluation of species distribution models. Biodivers.
Conserv. 23: 369–385.
Lelièvre, S., Vaz, S., Martin, C. S., and Loots, C. 2014. Delineating recurrent fish spawning
habitats in the North Sea. J. Sea Res. 91: 1–14.
Lobo, J. M., Jiménez-Valverde, A., and Real, R. 2008. AUC: a misleading measure of the
performance of predictive distribution models. Glob. Ecol. Biogeogr. 17: 145–151.
Loots, C., Vaz, S., Planque, B., and Koubbi, P. 2011. Understanding what controls the spawning
distribution of North Sea whiting (Merlangius merlangus) using a multi-model approach.
Fish. Oceanogr. 20: 18-31.

30
Parra, H. E., Pham, C. K., Menezes, G. M., Rosa, A., Tempera, F., and Morato, T. 2016. Predictive
modeling of deep-sea fish distribution in the Azores. Deep Sea Res. Part 2. Top. Stud.
Oceanogr. 145: 49-60.
Sequeira, A. M. M., Mellin, C., Fordham, D. A., Meekan, M. G., and Bradshaw, C. J. A. 2014.
Predicting current and future global distributions of whale sharks. Global Change Biol. 20:
778–789.
Serrano, A., González-Irusta, J. M., Punzón, A., García-Alegre, A., Lourido, A., Ríos, P., Blanco,
M., et al. 2017. Deep-sea benthic habitats modeling and mapping in a NE Atlantic seamount
(Galicia Bank). Deep Sea Res. Part 1. Oceanogr. Res. Pap. 126: 115–127.
http://dx.doi.org/10.1016/j.dsr.2017.06.003.
Tittensor, D. P., Baco, A. R., Hall-Spencer, J. M., Orr, J. C., and Rogers, A. D. 2010. Seamounts
as refugia from ocean acidification for cold-water stony corals. Mar. Ecol. 31: 212–225.
3.4.1. Group Discussion
To begin the discussion, the concept of evolutionary adaptability was introduced. It was noted that
most correlative models assume to capture the process of evolutionary adaptability, unlike
mechanistic models. A question was raised as to whether there is a way to incorporate the
mechanistic component into statistical models and not completely rely on correlative models.
Mechanistic distribution models look at the physiological attributes of an animal and can
determine factors needed for survival (i.e., water, food, etc.), and potentially suitable areas can
then be determined based on variable thresholds (a fundamental niche). However, there are often
insufficient data on the physiological attributes of deep-sea species. It was also noted that it is
important to consider for what use future predictions are being developed; many scientists
extrapolate to 2050, 2100 and dates such as this, but managers and stakeholders are likely to be
working on a shorter time frame, generally decadal, so what would a 2100 prediction be used for?
If one wanted to look specifically at smaller sections of one larger area, it may be beneficial to
look at areas where one knows conditions will change (i.e., salinity, temperature), so perhaps only
these regions could be used to apply species distribution models. Next, it was mentioned that as a
field of study, we need a greater experimental understanding of how animals respond to changes
in specific variables; many animals can still survive certain levels of change and more
experimental tests would be helpful in order for models to be more applicable. Finally, it was
mentioned that many studies that model the future distribution of VME indicator taxa do not take
into account reproduction and conditions necessary for reproduction, and rather focus on mature
animals. This focus has led to a large gap in knowledge and possibly unrealistic predictions.

31
3.5. Including Predictions of Community Functional Traits in Species Distribution
Models using Hierarchical Modeling of Species Communities (HMSC)
Benjamin Weigel
Research Centre for Ecological Change, Organismal and Evolutionary Biology
Research Programme, University of Helsinki, Helsinki 00014, Finland
Background
The ever-increasing impact of human-induced pressures on marine ecosystems causes
reorganization of communities by reshuffling species compositions, altering distribution ranges
and affecting species interactions, through changing environmental parameters and physical
disturbances. Geographic distribution, as well as species interactions within a community,
ultimately depend on the individual traits of a species. Functional traits may reflect information
on the species’ life history strategies, as well as morphological, physiological and behavioural
characteristics. Such information provides insight into, for example, preferred habitat types,
feeding modes or size distribution, among various other aspects that can be linked to ecosystem
functions and services (Weigel et al., 2016). Hence, considering species traits in distribution
modeling is a promising tool to better understand and predict community developments under
environmental changes, while also including the functional implications of community changes,
and therefore may be a useful framework for management and conservation.
Aim and Approach
The aim of this working paper is to highlight the possible application of including functional trait
information within Hierarchical Modeling of Species Communities (HMSC; Ovaskainen et al.,
2017), an approach that belongs to the class of joint species distribution models (JSDM; Warton
et al., 2015). To illustrate some of the possibilities HMSC allows for, I used zoobenthos
community datasets extracted from the EMODnet data base (www.emodnet-biology.eu) for the
Gulf of Finland (Finnish Environment Institute SYKE; 2018), including a total of 93 species at
541 locations that have been sampled between 2009 and 2014. I used time and depth as
explanatory variables. The trait data was obtained from the MERP Trait Explorer
(http://www.marine-ecosystems.org.uk/Trait_Explorer, Bruggeman et al., 2009; Brey et al., 2010;
Webb and Hosegood, 2013; Webb et al., 2017) offering inferred trait information for all marine
species, while indicating the uncertainty of their values based on available data. I included traits
such as maximum size and maximum mass of the species (or family) as well as feeding types, to
illustrate this approach.
Preliminary Results and Conclusions
In Figure 1, I highlight some relevant outputs that can be generated in HMSC exemplified on
coastal zoobenthos data from the Baltic Sea, but applicable and relevant to all community types

32
and habitats. The results focus on the inclusion of functional traits in community analysis and the
species interaction network, to illustrate species associations. Species-to-species interactions, and
hence their associations ultimately depend on the set of functional traits they express. Not being
fully included yet, one future task will be to better incorporate traits into the species interaction
network structure (Figure. 1b). However, there is already now a huge potential to predict
geographic ranges of community-weighted traits in relation to their environment that could
Figure 1. a) Results of variance partitioning. Variation in species occurrence is partitioned into
responses to fixed and random covariates. Fixed effects include depth and year, with the sampling
unit set as the random effect. The bar-plot shows the species-specific results whereas the legend
indicates averages over all species. Traits explain 16 % of the fixed effect. b) Estimates of species
associations measured by residual correlation. Species-to-species association matrices
highlight species pairs, with positive associations indicated in red and negative ones in blue, here
with statistical support of at least 75% posterior probability (the remaining cases are shown in
white). Species are ordered to emphasize the network structure. c) Mean predicted trait values
of zoobenthos assemblages for each sampling site, here exemplified for the traits: size, as
maximum length and two feeding types, deposit feeders and predators. Blue indicates low and red
indicates high values. d) Predicted community weighted mean of traits as response to included
covariates, here exemplified with the predicted development of passive suspension feeders over
time, showing a general decrease, and the maximum length against depth, which indicates that
species tend to be bigger in deeper waters. Both predictions show high confidence intervals and
can here only point out broad trends for a general illustration of the approach.

33
support conservation and management programmes. Especially being able to predict the impact
of changing environmental factors on functional aspects of communities will be a key element to
better understand the effect of climate-induced changes on ecosystem functioning and services.
HMSC provides a framework for the inclusion of functional traits in community analysis as a
hierarchical Bayesian joint species distribution model (Ovaskainen et al., 2017) and is
implemented in R and a Matlab package.
References
Brey, T., Müller-Wiegmann, C., Zittier, Z.M.C., and Hagen, W. 2010. Body composition in
aquatic organisms — A global data bank of relationships between mass, elemental
composition and energy content. J. Sea Res. 64: 334-340.
Bruggeman, J., Heringa, J., and Brandt, B.W. 2009. PhyloPars: estimation of missing parameter
values using phylogeny. Nucleic Acids Res. 37 (Web Server issue): W179-84.
doi: 10.1093/nar/gkp370.
European Marine Observation Data Network (EMODnet) Biology project (www.emodnet-
biology.eu), funded by the European Commission’s Directorate - General for Maritime
Affairs and Fisheries (DG MARE). Available online at www.emodnet-biology.eu
Finnish Environment Institute SYKE; 2018; Finnish Baltic Sea benthic monitoring, POHJE
database
Ovaskainen, O., Tikhonov, G., Norberg, A., Guillaume Blanchet, F., Duan, L., Dunson, D., Roslin,
T. and Abrego, N. 2017. How to make more out of community data? A conceptual
framework and its implementation as models and software. Ecol. Lett. 20: 561–576.
Warton, D.I., Blanchet, F.G., O’Hara, R.B., Ovaskainen, O., Taskinen, S., Walker, S.C. and Hui,
F.K.C. 2015. So Many Variables: Joint Modeling in Community Ecology. Trends Ecol.
Evol. 30: 766–779.
Webb, T. et al. Body sizes of UK marine species, draft dataset for MERP v2017_05_09.
Supplemented with data provided by Beth Mindel on 29/6/2017.
Webb, T., and Hosegood, J. 2013. Integrating biological traits of European marine benthic taxa
into the World Register of Marine Species. EMODnet Project Final Report.
http://www.marinespecies.org/traits/docs/reports_pilots/Benthos.pdf
Weigel, B., Blenckner, T. and Bonsdorff, E. 2016. Maintained functional diversity in benthic
communities in spite of diverging functional identities. Oikos 125: 1421–1433.

34
3.6. GlobENV: Towards a High-resolution Climatology for the Seafloor
Andrew J. Davies1,2 and Emyr M.T. Roberts1,3
1School of Ocean Sciences, Bangor University, Menai Bridge, Anglesey, Wales.
From August 2018: 2Department of Biological Sciences, University of Rhode
Island, Kingston, USA, davies@uri.edu.
3Department of Biological Sciences & KG Jebsen Centre for Deep-Sea Research,
University of Bergen (UiB), Norway
Rationale
The ocean is the largest habitat on earth, covering approximately 70% of the planet. Our
knowledge of patterns within surface waters is fairly extensive, principally driven by the
development of earth observing satellites. Whilst marine scientists have clearly benefitted from
such technologies, effectiveness is limited to the upper parts of the water column (e.g., ocean
colour: Behrenfield and Falkowski, 1997) or coarse gravity estimates of the sea-surface that
correlate with ocean depth (Smith and Sandwell, 1997). Accurate data regarding conditions at the
seafloor remain scarce and are generally concentrated around developed countries (Ramirez-
Llodra et al., 2010). Even with the now widespread adoption of technologies such as multibeam
echosounders, remotely operated vehicles and autonomous underwater vehicles (Danovaro et al.,
2014), only approximately 5% of the seafloor has been mapped, and a far smaller area has been
investigated in great detail (Ramirez-Llodra et al., 2010). Recently, there has been renewed interest
in ocean exploration, driven by the need to have a better understanding of geological features,
underwater resources and species distributions. However, studies in many parts of our ocean
remain constrained by the availability of high quality and validated data on seafloor conditions.
Large-scale ocean mapping requires significant infrastructure and investment, and as a result, our
understanding of the ocean floor significantly lags behind terrestrial environments.
Towards a new deep-sea climatology
Several marine climatologies are currently available. For example, Bio-ORACLE initially
provided a data package that focusses on surface waters (Tyberghein et al., 2012) which was
recently extended to include some benthic and future climate data (Assis et al., 2017). The
MARSPEC dataset was based upon a higher resolution bathymetric dataset (i.e., SRTM30 Becker
et al., 2009), and provides several benthic terrain variables and temperature/salinity for the sea
surface (Sbrocco and Barber, 2013). GlobENV aims to extend these previous climatologies by
providing an up-scaling approach that can be applied to any bathymetric dataset available of any
resolution by using the best available environmental data in the ocean. It extends previously
upscaled datasets (Davies and Guinotte, 2011; Guinotte and Davies, 2014), by providing a more

35
accurate and computationally efficient methodology and by implementing a robust validation
framework that can be applied to each variable.
Methodology
Trilinear interpolation on a three-dimensional regular grid was used to estimate conditions on the
seafloor from various environmental data sources. This approach interpolates the value of a point
at depth z with coordinates x and y from eight surrounding points obtained from regularly gridded
environmental data (Figure 1). GlobENV is designed to take gridded environmental data of
varying resolution, and bathymetric data of an equal or finer resolution, producing an output that
is an estimate of the variable at that particular depth and position. Finally, after computation of a
variable (for an example see Figure 2), a validation process is conducted that compares
performance against various environmental data. Several key parameters are calculated, 1) overall
root-mean-square error, 2) spatial error calculations, 3) error by depth bin and 4) correlation
metrics.
Conclusion
To date, GlobENV variables have been created for 10 different bathymetric datasets (5 global and
5 regional), and 7 environmental variables extracted from the World Ocean Database. Several
more are planned, and both code and data products will be released under an open source license
in the near future. Overall, the validation of the various variables indicates that finer resolution
bathymetry produces a consistently lower RMSE than coarse bathymetric products, indicating that
multibeam data would have particular value when used with GlobENV. However, the approach
does suffer from several limitations as observed in Davies and Guinotte (2011). Depth and spatial
error validation shows that the approach does not resolve well in shallower coastal seas, but
performance improves rapidly at depths below 50 m irrespective of the bathymetric layer used. As
such, this climatology may have particular value for the study of deep-sea species at varying
resolutions depending on the data available.

36
Figure 1. GlobENV workflow (a) and a schematic of the trilinear interpolation technique (b).
World Ocean Atlas
IPCC Climate Model Data
Vertically resolved hydrodynamic
model output
Environmental Data
Inputs Preparation Analysis
GlobENV
Validation
Global Data Products
Regional Modelled Bathymetries
Multibeam Data from Local Areas
Bathymetric Data
Extracted from 3D
vertically resolved grids
interpolated into
continuous rasters.
XY coordinates extracted.
Cut into workable
chunks to allow
parallel processing of
individual rasters and
to bypass in memory
errors.
Applies a trilinear prediction
of conditions based on the
depth of a bathymetric cell
in relation to its surrounding
environmental data.
Produces a standardised
validation approach to
determine the accuracy
of the environmental layer.
x step(-1), ystep(-1), z step(-1)
z step(- n)
z step(+n)
x step(-n)x st ep(+n)
x step(-1), ystep(-1), z step(+1)
y step(+n)
y step(-n)
x step(+1), ystep(-1), z step(+1)
x step(+1), ystep(+1), z step(+1)
x step(+1), ystep(+1), z step(-1)
x step(-1), ystep(+1), z step(-1)
x step(-1), ystep(+1), z step(+1)
Bathymetry
Centriod of selected bathymetry cell (red)
Selected centroids of environmental data raster
Unselected centroids of environmental data raster
Trilinear calculation points
a)
b)

37
Figure 2. Example output from GlobENV at approximately 100 m XY grid resolution for the
variable temperature (°C) based INFOMAR bathymetry and World Ocean Atlas 2013 data.
References
Assis, J., Tyberghein, L., Bosch, S., Verbruggen, H., Serrao, E.A. and De Clerck, O. 2017. Bio-
ORACLE v2.0: Extending marine data layers for bioclimatic modeling. Global Ecol.
Biogeogr. 27:277–284.
Becker, J. J., Sandwell, D. T., Smith, W. H. F., Braud, J., Binder, B., Depner, J., Fabre, D., Factor,
J., Ingalls, S., Kim, S-H., Ladner, R., Marks, K., Nelson, S., Pharaoh, A., Trimmer, R., Von
Rosenberg, J., Wallace, G. and Weatherall, P. 2009. Global Bathymetry and Elevation Data
at 30 Arc Seconds Resolution: SRTM30_PLUS. Mar. Geod. 32: 355–371.
Behrenfield, M.J., and Falkowski, P.G. 1997. Photosynthetic reates derived from satellite-based
chlorophyll concentration. Limnol. Oceanogr. 42: 1–20.
Danovaro, R., Snelgrove, P.V. and Tyler, P. 2014. Challenging the paradigms of deep-sea ecology.
Trends Ecol. Evol. 29: 465–475.
Davies AJ, and Guinotte, J.M. 2011. Global habitat suitability for framework-forming cold-water
corals. PloS One 6:e18483.
Guinotte, J.M., and Davies, A.J. 2014. Predicted Deep-Sea Coral Habitat Suitability for the U.S.
West Coast. PLoS ONE 9(4): e93918. https://doi.org/10.1371/journal.pone.0093918.
Ramirez-Llodra, E., Brandt, A., Danovaro, R., De Mol, B., Escobar, E., German, C. R., Levin, L.
A., Martinez Arbizu, P., Menot, L., Buhl-Mortensen, P., Narayanaswamy, B. E., Smith, C.
R., Tittensor, D. P., Tyler, P. A., Vanreusel, A., and Vecchione, M. 2010. Deep, diverse and

38
definitely different: unique attributes of the world's largest ecosystem. Biogeosciences 7:
2851–2899.
Sbrocco, E. J., and Barber, P. H. 2013. MARSPEC: Ocean climate layers for marine spatial
ecology. Ecology 94: 979.
Smith, W.H.F., and Sandwell, D.T. 1997. Global seafloor topography from satellite altimetery and
ship depth soundings. Science 277: 1957–1962.
Tyberghein, L., Verbruggen, H., Pauly, K., Troupin, C., Mineur, F., and De Clerck, O. 2012. Bio-
ORACLE: A global environmental dataset for marine species distribution modeling. Ecol.
Biogeogr. 21:272–281.
3.7. Ensembling of Multiple Data Sets and Multiple Models
Chris Rooper1*, Duane Stevenson1, Ivonne Ortiz2, Jerry Hoff1
1Alaska Fisheries Science Center, National Marine Fisheries Service, NOAA,
Seattle, Washington, USA;
2Joint Institute for the Study of the Atmosphere and Oceans, University of
Washington, Seattle, Washington, USA
*Current affiliation: Pacific Biological Station, Fisheries and Oceans Canada,
Nanaimo, British Columbia, Canada
One of the advantages of using species distribution models is their ability to take advantage of
diverse data sources using a variety of different methods. The data used for modeling the
distribution of species in the deep-sea can range from statistically robust and well-designed
surveys (e.g., Laman et al., 2018) to compilations of data from museum specimens, opportunistic
collections or multiple sources (e.g., Davies and Guinotte, 2011). Modeling methods can
incorporate different types of records, such as presence-only data (e.g., maximum entropy models)
or presence/absence and abundance records using either statistical (e.g., generalized linear models
and generalized additive models) or computer-learning (e.g., random forest and boosted regression
tree models) methods. Species distribution models can also account for spatial autocorrelation
explicitly in the method (e.g., kernel density models and vector auto-regressive spatial-temporal
models), implicitly (e.g., maximum entropy models) or not at all. The choice of which data to use
in the modeling and which modeling method to pursue can sometimes result in different
representations of the species distribution (Figure 1).
Often ensembling data into a species distribution model can be problematic. The most common
example of this in the deep-sea environment is combining data collected with different gear types

39
into a single modeling framework. This is routinely done with maximum entropy models, but can
also be achieved with statistical and computer learning models where a term for gear type is
included in the model formulation. An example of this can be found in vector-auto-regressive-
spatial-temporal (VAST) models where a separate “catchability coefficient” can be set for each
gear type (Thorson et al., 2015).
Another example of using ensembled data in species distribution models is for models predicting
future changes in species distribution due to climate change effects. In this case there may be
multiple climate scenarios each providing a time series of future temperatures. These data could
be fed into a single model to produce multiple potential representations of species distributions.
An example of this is shown for multiple climate scenarios acting on Pacific cod distributions in
the eastern Bering Sea (Figure 2) or Araujo and New (2007).
Ensemble models have seen wide use in climatology, where multiple scenarios are combined to
come up with the best predictions attainable. Ensembling is also becoming more prevalent in the
fisheries stock assessment realm, where multiple types of assessment models with different
assumptions can be combined into a trend in population status (e.g., Anderson et al., 2017 and
Rosenberg et al., 2018 papers on “super-ensembles”). In the species distribution modeling
literature there are a number of examples of using multiple models in an ensemble (Robert et al.,
2016; Rowden et al., 2017; Rooper et al., 2017). The methods used to combine models into an
ensemble have varied, but most are weighted by some type of measure of variability in the
goodness-of-fit (e.g., AUC, SD, or R2). In some cases, some sort of scaling of predictions may be
necessary to combine model types into an ensemble. A particular subset of ensembling, that is
interesting and relatively underutilized, is spatially explicit model ensembles. In some cases
models have been constructed at different spatial scales, for example a PICES working group has
been conducting broad-scale modeling of corals and sponges for the entire North Pacific ocean.
This is an effort that is fairly data poor for some regions and uses presence only data. However,
within the North Pacific Ocean there are a number of EEZ’s and seamounts for which better
models have been produced. For example, Miyamoto et al. (2017) constructed models for the
Emperor Seamount chain. It would be beneficial to ensemble these smaller scale and larger scale
models to come up with better overall predictions. Importantly, this type of ensemble should
somehow represent spatially explicit confidence estimates for the predictions.
Key questions for ensembling data and models are:
 What are the advantages and disadvantages of combining data sets into a single modeling
method versus modeling each data set independently and combining predictions?
 How to best weight or scale models when ensembling (especially across different data
types)
 How to represent error in ensemble models that reflects individual model prediction
variability

40
An example of ensembling methods with different data and models can be found with the
distribution of skate nursery areas in the eastern Bering Sea. Several species of skates in the eastern
Bering Sea deposit eggs in highly concentrated areas that are persistent through time (Hoff, 2010).
The skate eggs incubate and develop for up to several years before a small but fully developed
juvenile emerges. The persistence and concentration of eggs in specific nursery locations has been
well documented, but the distribution of the nursery areas is relatively unknown. The data set for
these nurseries is limited to 26 locations that have been discovered since 2008 during camera
surveys and bottom trawl surveys. Recorded absences from the same camera and trawl surveys
can be used, however, the catchability of skate eggs is presumed to be low, so these are probably
not all “true” absences.
Figure 1. A) Map of the best model of probability of suitable habitat for skate nursery areas based
on environmental variables, presence observations (n = 26), and absences for the eastern Bering
Sea outer shelf and slope. The contour line indicates the threshold probability (0.60) at which
presence of suitable habitat was determined. B) Kriged surface of the number of skate egg cases
in the commercial fisheries catch across multiple gears, seasons and years. The data was collected
by observers onboard the fishing vessels.
A second data set is available on the distribution of skate egg cases recorded by observers on
commercial fishing vessels. This data set includes abundance (counts) and locations of catches of
skate eggs. There are no absence records for these data and because of confidentiality issues and
the relatively coarse position data available for the presence records, they are resolved on a scale

41
of ~5-10 km. These data are very different to the bottom trawl survey catches and camera survey
data.
Figure 2. Predictions from a species distribution model for Pacific cod in the eastern Bering Sea
using 6 climate forecasts. A) shows the distribution of Pacific cod in 2016, B) shows the predicted
distribution of Pacific cod in 2099, C) shows the reduction (average and standard deviation) in
area occupied by Pacific cod from 1980 to 2099 and D) shows the latitudinal and longitudinal
shifts of the center of distribution for each of the climate scenarios.
In response to management concerns species distribution models were developed that predicted
potential nursery locations based on the 26 known areas and the fisheries data. Combining the two

42
data sets was unreasonable given their fundamentally divergent properties, thus two modeling
approaches were used. A maximum entropy model was used for the camera and bottom trawl
surveys (with recorded absences as pseudo-absences). This model resulted in a fairly narrow depth
band along the upper continental slope that was predicted to be suitable habitat (Figure 1). For the
commercial fisheries data a kriging method was used that incorporated the abundance data.
These two outputs (Figure 1) reflect similar patterns in some cases, but are fundamentally
different. How can these outputs best be combined in an ensemble to inform management on where
skate nurseries are likely to be found?
References
Anderson, S.C., Cooper, A.B., Jensen, O.Pl, Minto, C., Thorson, J.T., Walsh, J.C., Afflerbach, J.,
Dickey-Collas, M., Kleisner, K.M., Longo, C., Osio, G.C., Ovando, D., Mosqueira, I,
Rosenberg, A.A. and Selig, E.R. 2017. Improving estimates of population status and trend
with superensemble models. Fish Fish. 18: 732-741. https://doi.org/10.1111/faf.12200
Araújo, M.B., and New, M. 2007. Ensemble forecasting of species distributions. Trends Ecol.
Evol. 22: 2–47. doi: 10.1016/j.tree.2006.09.010
Davies, A.J., and Guinotte, J.M. 2011. Global habitat suitability for framework-forming cold-
water corals. PLOS One 6:1-15.
Hoff, G.R. 2010. Identification of skate nursery habitat in the eastern Bering Sea. Mar. Ecol. Prog.
Ser. 403: 243-254.
Laman, E.A., Rooper, C.N., Turner, K., Rooney, S., Cooper, D.W., and Zimmermann, M. 2018.
Using species distribution models to define essential fish habitat in Alaska. Can. J. Fish.
Aquat. Sci. Early online. https://doi.org/10.1139/cjfas-2017-0181
Miyamoto, M., Kiyota, M., Murase, H., Nakamura, T., and Hayashibara, T. 2017. Effects of
bathymetric grid-cell sizes on habitat suitability analysis of cold-water gorgonian corals on
seamounts. Mar. Geod. 40: 205-223. DOI: 10.1080/01490419.2017.1315543
Robert, K., Jones, D.O.B., Roberts, M.J., and Huvenne, V.A.J. 2016. Improving predictive
mapping of deep-water habitats: considering multiple model outputs and ensemble
techniques. Deep Sea Res. I 113: 80–89 doi: 10.1016/j.dsr.2016.04.008
Rooper, C.N., Zimmermann, M., and Prescott, M. 2017. Comparisons of methods for modeling
coral and sponge distribution in the Gulf of Alaska. Deep Sea Res. II 126:148-161.
Rosenberg, A.A., Kleisner, K.M., Afflerbach, J., Anderson, S.C., Dickey-Collas, M., Cooper,
A.B., Fogarty, M.J., Fulton, E.A., Gutierrez, N.L., Hyde, K.J.W., Jardim, E., Jensen, O.P.,
Kristiansen, T., Longo, C., MInte-Vera, C.V., Minto, C., Mosquiera, I., Osio, G.C., Ovando,

43
D., Selig, E.R., Thorson, J.T., Walsh, J.C. and Ye, Y. 2018. Applying a new ensemble
approach to estimating stock status of marine fisheries around the world. Cons. Letters 11:
1-9, e12363.
Rowden, A.A., Anderson, O.F., Georgian, S.E., Bowden, D.A., Clark, M.R., Pallentin, A., and
Miller, A. 2017. High-resolution habitat suitability models for the conservation and
management of vulnerable marine ecosystems on the Louisville seamount chain, South
Pacific Ocean. Front. Mar. Sci. 4: 1-19.
Thorson, J.T., Shelton, A.O., Ward, E.J., and Skaug, H.J. 2015. Geostatistical delta-generalized
linear mixed models improve precision for estimated abundance indices for West Coast
groundfishes. ICES J. Mar. Sci. J. Cons. 72(5): 1297–1310.
3.7.1. Group Discussion
Dr. Rooper’s presentation focused on the difficulties associated with ensembling data into species
distribution modeling work and using this method to predict future changes in species
distributions. Two comments were made concerning climate change predictions. First, it was
suggested that it is important to understand the difference between climate predictions and climate
projections. Projections are averaged statistical properties over a period of time, while projections
deemed most likely become predictions or forecasts. Second, it was noted that one should use the
same climate model when training data as when making predictions. It was suggested that it would
be better to produce six models for six climate prediction data sets, and then ensemble the model
output, rather than ensembling the data and producing a single output from that. Also, to obtain
finer-scale climate predictions it was suggested that coarse climate change predictions could be
upscaled by applying the data in “large pixels” to “finer pixels” which are available in present
“nowcasts”.
3.8. Resolution of Seabed Features in the Deep-Sea: Implications for Habitat
Characterization
Myriam Lacharité and Craig J. Brown
Applied Oceans Research Group, Nova Scotia Community College, Dartmouth,
NS, Canada
Seabed features – i.e., shape of the seafloor and substratum type - play a central role in habitat
characterization for deep-sea benthos (e.g., Tong et al., 2012; Zeppili et al., 2016; and many
others). Marine geomorphometry quantifies patterns of the shape of seafloor with derived
bathymetric (i.e., ‘terrain’) variables (e.g., slope, roughness, curvature), which are also often used
as proxy of substratum type, in particular in complex benthic ecosystems to predict the presence
of hard substratum.

44
Using high-resolution seafloor data increases quality and accuracy of species distribution models
(SDMs) in the deep-sea benthos (Rengstorf et al., 2012). Coarser environmental data can omit
fine- (i.e., sub-grid) scale variability relevant to habitat characterization, introducing error into
predictive models by mis-representing the distribution of a species’ suitable habitat (Vierod et al.,
2014). Because of data scarcity in the deep sea, SDMs are increasingly used at scales relevant to
management (e.g., regional/territorial seas) to predict the occurrence of specific taxa of interest
(e.g., cold-water corals) in unsurveyed regions. However, Vierod et al. (2014) suggest that the lack
of accurate fine-scale bathymetric data limits the use of broad-scale models in the deep sea, and
clear discrepancies between predicted and observed presence can arise during model validation
(Anderson et al., 2016). While bathymetric gridded data are available at the global scale, their
resolution remains coarse (30 arc-seconds: ~1 km at the equator; Weatherall et al., 2015), and the
availability of higher-resolution bathymetric data from local surveys (sub-meter to 10s of meters)
is sporadic (Mayer et al., 2018).
Here, we first suggest that in addition to fine-scale bathymetric data in the deep sea, care must be
given to fine-scale substratum type, the characterization of which may not necessarily be captured
using derived bathymetric variables. For example, recent studies have demonstrated the role of
glacial dropstones (sporadic large boulders) at temperate and high latitudes in fostering diverse
deep-sea benthos (Lacharité and Metaxas, 2017; Ziegler et al., 2017), the presence of which is
captured with optical imagery or very-high bathymetric resolution (sub-meter). A detailed
depiction of a wider extent of the surficial geology is detected with the reflectivity of the seafloor
(i.e., acoustic backscatter), and advances in its standardization make these datasets increasingly
more comparable (Brown et al., 2011; 2012).
Second, we suggest open and shared high-resolution bathymetric data available to multiple end-
users would benefit habitat characterization for deep-sea benthos. The benefits of this approach
for deep-sea SDMs goes beyond increased model accuracy by also supporting interdisciplinary
research, in particular the development of high-resolution ocean hydrodynamic and
biogeochemical models. Derived outputs from these models can enhance local habitat
characterization and support spatially-explicit distribution models incorporating connectivity.
Finally, we support the view that broad-scale habitat characterization in the deep sea would benefit
from the availability of multiresolution (spatially-varying) and multi-layered (bathymetry and
backscatter) seabed data. Recent efforts aim to increase the bathymetric resolution at the global
scale (Mayer et al., 2018). However, we suggest that high-resolution bathymetric datasets should
be available wherever they have been collected, including the raw data and backscatter return for
substratum characterization. Future research in deep-sea SDMs could explore how developing
multiresolution modeling approaches would limit dependence of broad-scale high-resolution
bathymetry.

45
References
Anderson, O.F., Guinotte, J.M., Rowden, A.A., Clark, M.R., Mormede, S., Davies, A.J., and
Bowden, D.A. 2016. Field validation of habitat suitability models for vulnerable marine
ecosystems in the South Pacific Ocean: Implications for the use of broad-scale models in
fisheries management. Ocean Coast. Manag. 120: 110–126.
Brown, C.J., Sameoto, J.A., and Smith, S.J. 2012. Multiple methods, maps, and management
applications: Purpose made seafloor maps in support of ocean management. J. Sea Res. 72:1-
13.
Brown, C.J., Smith, S.J., Lawton, P., and Anderson, J.T. 2011. Benthic habitat mapping: A review
of progress towards improved understanding of the spatial ecology of the seafloor using
acoustic techniques. Estuar. Coast. Shelf Sci. 92: 502–520.
Lacharité, M., and Metaxas, A. 2017. Hard substrate in the deep ocean: How sediment features
influence epibenthic megafauna on the eastern Canadian margin. Deep. Res. Part I
Oceanogr. Res. Pap. 126: 50–61.
Mayer, L., Jakobsson, M., Allen, G., Dorschel, B., Falconer, R., Ferrini, V., Lamarche, G., Snaith,
H., and Weatherall, P. 2018. The Nippon Foundation—GEBCO Seabed 2030 Project: The
Quest to See the World’s Oceans Completely Mapped by 2030. Geosciences 8: 63.
Rengstorf, A.M., Grehan, A., Yesson, C., and Brown, C. 2012. Towards high-resolution habitat
suitability modeling of vulnerable marine ecosystems in the deep-sea: Resolving terrain
attribute dependencies. Mar. Geod. 35: 343–361.
Tong, R., Purser, A., Unnithan, V., and Guinan, J. 2012. Multivariate statistical analysis of
distribution of deep-water gorgonian corals in relation to seabed topography on the
Norwegian margin. PLoS One 7: e43534.
Vierod, A.D.T., Guinotte, J.M., and Davies, A.J. 2014. Predicting the distribution of vulnerable
marine ecosystems in the deep sea using presence-background models. Deep. Res. Part II
Top. Stud. Oceanogr. 99: 6–18.
Weatherall, P., Marks, K.M., Jakobsson, M., Schmitt, T., Tani, S., Arndt, J.E., Rovere, M.,
Chayes, D., Ferrini, V., and Wigley, R. 2015. A new digital bathymetric model of the
world’s oceans. Earth Space Sci. 2: 331-345.
Zeppilli, D., Pusceddu, A., Trincardi, F., and Danovaro, R. 2016. Seafloor heterogeneity
influences the biodiversity-ecosystem functioning relationships in the deep sea. Sci. Rep. 6:
1–12.

46
Ziegler, A.F., Smith, C.R., Edwards, K.F., and Vernet, M. 2017. Glacial dropstones: Islands
enhancing seafloor species richness of benthic megafauna in West Antarctic Peninsula
fjords. Mar. Ecol. Prog. Ser. 583: 1–14.
3.8.1. Group Discussion
Discussion following the presentation focused on the lack of high-quality broad-scale seabed data
and possible methods of adjusting to this lack of data while still providing well-informed model
outputs. Variables vary at different spatial scales and different variables are more or less useful as
predictors depending 1) on the species/habitat whose distribution is being modeled, and 2) the
questions that are being asked of the model (i.e., the model’s intended use). Fine-scale data can be
highly relevant to a given species’ distribution, yet data at such a resolution may not (and likely
won’t) be available. One solution to this problem could be the use of proxy data, for example using
satellite-based glacier tracking as a proxy for dropstones. A second method mentioned creating a
substrate predictive map for small areas such as the coast of British Columbia; using substrate data
from small areas to predict other areas is a good option when fine-scale data is unavailable.
Finally, it is important to distinguish between the scale at which the habitat varies compared to the
scale at which the biota is sampled. Multiscale models were suggested as a possible answer for
this problem. An additional example was given: it would be expected that rock would vary at a
fine scale compared to other covariates that vary at large scales. If available data is limited, such
as specific rock location, certain habitats can be eliminated for certain species. Models are
conditional on covariates, and if these are absent it will show in the error.
3.9. Best Practices in the Development and Application of Species Distribution Models
to Support Decision Making in Marine Spatial Planning
Jessica Finney1, Emily Rubidge1, Cherisse Du Preez1, Jessica Nephin1,
Candice St. Germain1, Cole Fields1, and Edward Gregr2
1 Pacific Region, Fisheries and Oceans Canada (DFO)
2 SciTech Consulting, Vancouver, BC
The Marine Spatial Ecology and Analysis (MSEA) Section is a part of Fisheries and Oceans
Canada (DFO) in the Pacific Region. The MSEA Section is involved in several areas of research
that use SDMs to support ocean management, including marine protected area network design, oil
spill response planning, forecasting the impacts of climate change, and identifying sensitive
benthic areas. The Species Distribution Modeling Working Group (SDM WG) was formed to
provide coordination and advice to help facilitate the use and development of species distribution
models within MSEA and more generally in DFO in the Pacific Region.

47
A key output of the SDM WG will be a best practices workbook. The intent of this workbook
is to provide users with a well-referenced guide to the development of species distribution models.
The workbook will include discussions on appropriate data usage, pre-processing and data
preparation, modeling approaches and development, model validation, and interpretation of
results. The workbook will outline best practices to standardize the species distribution modeling
approach within the MSEA section (to ensure
consistent quality and rigor in the models
developed in the section). It will also include
important information about how to interpret
model outputs and how to convey underlying
uncertainty to managers. The workbook will
facilitate and simplify the model building and
interpretation process, support the use of best
practices, and provide consistent outputs,
measures of uncertainty, and evaluation
metrics for managers.
The workbook will present three types of
commonly used models of increasing
complexity and data needs to highlight the importance of the four basic steps of species
distribution models development (data preparation, model development, model validation, and
interpretation) and to identify when a simple approach may be more appropriate than a complex
approach. The three models that will be included in the workbook to demonstrate these points are:
1) Habitat Suitability Index (HSI) models (simple model); 2) Generalized Linear Models
(GLMs) (medium complexity); and 3) Boosted Regression Tree (BRT) models (most complex).
This tiered approach increases the utility of the workbook by making it applicable to a wide range
of data availability scenarios. It also lets users decide how complex they need or want their models
to be. For example, if a simple model provides good results, it may not be necessary to spend the
time and effort required to build a more complex model.
An additional component of the workbook will be automated code for both the GLM and BRT
models. The species distribution modeling procedure for both models is run in two stages. The
first stage processes species (presence/absence or abundance) and environmental data, produces
diagnostic plots, and maps environmental data. The second stage builds and evaluates the GLM
and BRT models using k-fold cross validation, produces validation plots, estimates the marginal
effects and relative influence of predictor variables in the model and uses models to predict the
presence or abundance of species in new locations.
The workbook will initially be tested out using a set of 12 benthic marine species representing a
variety of life-history strategies, habitats and data availability scenarios. Results of this analysis,
as well as the completed workbook, are expected in early 2019.
Model interpretation
Uncertainty Application Limitations
Model validation
Numerical Spatial
Model development
HSI GLM BRT
Data preparation
Environmental Species

48
3.10. Evaluating Effects of Rescaling and Weighting Data on Habitat Suitability
Modeling
Ying Xue1,2, Lisha Guan2, Kisei R. Tanaka2, Zengguang Li1,2, Yong Chen2,
Yiping Ren1
1 Fisheries College, Ocean University of China, Qingdao 266003, China
2 School of Marine Sciences, University of Maine, Orono, ME 04469, USA
Background
Habitat suitability index (HSI) models are important tools in identifying suitable habitats of living
marine resources (LMR). Abundance indices (AI) derived from fishery-independent surveys are
often used to calibrate HSI models. However, AIs tend to have a highly right-skewed distribution
as a result of large spatial heterogeneity of LMR distributions, which can further affect subsequent
HSI model performance. Furthermore, traditional HSI models are often based on unrealistic
assumptions that environmental variables have equal impacts on the AIs. Using American lobster
(Homarus americanus) in the inshore Gulf of Maine as a case species, this study evaluates the
performance of different approaches in developing HSI models.
Objectives
The objectives of this study are to compare the performance of HSI models derived using
untransformed abundance indices (AIs) versus log-transformed AIs and evaluate the effectiveness
of a boosted regression tree (BRT)-based weighting approach.
Approaches
The existence of large AI values might result in underestimation of HSI values for most sampling
stations. One approach to avoid this problem is to use rescaled AIs (e.g., log-transformed AIs) to
reduce the impact of large AI values on SIs. In this study, boosted regression tree (BRT) models
were used to determine the weights of environmental variables in HSI modeling.
Key findings
Both cross-validation and predicted habitat suitability maps suggested that the weighted HSI
model based on log-scaled AI data tended to yield a more reliable prediction of optimal habitats
for American lobster (Figure 1). The unweighted HSI model based on the original AI data,
however, tended to underestimate optimal habitats and overestimate suboptimal habitats. We
recommend using log-transformed AIs and determining the weights of different environmental
variables based on the BRT method in HSI modeling, especially when AI data are highly skewed.

49
The approach demonstrated in this study can improve the quality of HSI modeling, leading to
better definitions of suitable habitats (Xue et al., 2017).
Figure 1. Spatial distribution of observed abundance indices (AIs) for American lobster in fall of 2013
overlaid on two predicted habitat suitability index (HSI) maps derived from HSI-AI and weighted HSI-
lnAI models in Maine–New Hampshire inshore bottom trawl survey areas.
Reference
Xue, Y., Guan, L., Tanaka, K., Li, Z., Chen, Y., and Ren, Y. 2017. Evaluating effects of rescaling
and weighting data on habitat suitability modeling. Fish. Res., 188: 84–94.
https://doi.org/10.1016/j.fishres.2016.12.001
3.10.1. Group Discussion
The discussion following Dr. Tanaka’s presentation focused on the use of a simple rank-
based (i.e., non-statistical) Habitat Suitability Index (HSI) model to predict species

50
distribution. It was pointed out that this type of qualitative method can be useful for
introducing non-statisticians to the field of species modeling – this is particularly important
in helping non-modeller scientists who engage with stakeholders and need to understand
species distribution models to a sufficient degree to simply explain the methodologies and
their outputs to stakeholders. One problem with this method, however, is that when
calibrating HSI with highly skewed zero-inflated data, species abundance and environmental
variables may not be associated appropriately in terms of computing the habitat suitability
of marine species. Furthermore, if environmental variables are weighted equally, the model
output can be problematic as it ignores the difference in ecological importance of each
environmental variable. The presentation showed that rescaling skewed observations (e.g., log-
transformation) and applying machine-learning techniques (e.g., boosted regression tree) can
address these known shortcomings in the HSI modeling. These approaches are important
because weighted environmental variable contributions are generally an output which is
sought after from species distribution modeling work, rather than used as an input.
3.11. Developing a Generalized Climate-niche Modeling Framework to Improve
Management of Commercially Important Species in a Climatically Altered Marine
Environment: A Case Study with American lobster and Atlantic scallop in the Gulf
of Maine
K. R. Tanaka*1,2, M. P. Torre*1, V. S. Saba3, and Y. Chen1
*Both authors contributed equally
1School of Marine Sciences, University of Maine, Orono, ME 04469, USA
2Atmospheric and Oceanic Sciences Program, Princeton University, 300 Forrestal
Road, Sayre Hall, Princeton, New Jersey 08544, USA
3National Oceanic and Atmospheric Administration, National Marine Fisheries
Service, Northeast Fisheries Science Center, c/o Geophysical Fluid Dynamics
Laboratory, 201 Forrestal Road, Princeton University Forrestal Campus,
Princeton, New Jersey 08540, USA
Background
Climate change has been identified as a key driver of ecology and population dynamics of
commercially important fish and shellfish stocks within marine ecosystems. Accounting for
changing biogeography of fish populations is essential to the effective assessment and
management of fisheries; however, the majority of fishery management frameworks in the United
States have yet to incorporate information about how environmental conditions influence

51
distribution and abundance of commercial stocks (Saba et al., 2015; Skern-Mauritzen et al.,
2015). That observed biogeographic ranges shift as consequences of ongoing climate change has
raised a lot of concerns and uncertainty over the effectiveness of the current survey programs and
subsequently in associated stock assessment models. The development of a forecasting ability to
evaluate climate change impacts on the ecology of commercially valuable marine stocks has been
advocated as an important step in mitigating uncertainty in climate change adaptation strategies.
Objectives
Using two commercially important stocks (American lobster and Atlantic scallop) in the northeast
USA continental shelf as case studies, the objective of this research is to develop a generalized
climate-niche modeling framework that can be a part of a coordinated regional effort to assess
climate change effects on commercial fish and their fisheries. The proposed climate-niche
modeling framework will improve our management of marine fish and shellfish stocks by
providing (1) enhanced hind/now/forecasting capacity of spatio-temporal changes in their
biogeography, and (2) a space/time varying estimate of their availability to the existing fishery
surveys and management zones. Such a modeling framework will provide excellent tools for
climate change-related research in fisheries, thus greatly enhancing foundations for future climate-
change-related research proposals in fisheries.
Approach
The proposed climate-niche modeling framework will consist of using a delta-generalized linear
mixed model (delta-GLMM: Thorson et al., 2015) and ensemble niche model (biomod2: Thuiller
et al., 2016) coupled with a dispersion simulation (MigClim: Engler et al., 2012) for
hindcasting/forecasting spatio-temporal changes in distribution of American lobster and Atlantic
scallop on the northeast USA continental shelf. In this project we would like to address the
following questions: (1) How did abundance and distribution of American lobster and Atlantic
scallop change over time? (2) What are the management implications of changing biogeography
of the lobster and scallop stocks to the existing fishery survey designs and management zones?
Preliminary results
The ensemble niche model was used to predict the change in spatial distribution of American
lobster and Atlantic scallop based on an 80-year time series of future bottom temperatures and
salinity changes based on a transient climate response (2xCO2) simulation where atmospheric CO2
is increased by 1% per year (Figures 1, 2). The model results will be used to (1) evaluate the past,
present and future performance of the existing fishery-independent survey design in capturing
annual variability of lobster and scallop stock abundance, and (2) discuss management
implications of changing biogeography of the lobster and scallop stocks to the existing fishery
management zones.

52
Year 1-10
Year 71-80
High
probability of
presence
Low
probability of
presence
Figure 1. Predicting the change in American lobster probability of occupancy in the U.S.A.
northeast continental shelf region. Predictions represent the mathematical average of ensemble
models of probability of presence from years 1984–2015. Model predictions from the first and last
10 years are compared.
Year 1-10
Year 71-80
High
probability of
presence
Low
probability of
presence
Figure 2. Predicting the change in sea scallop probability of occupancy in the U.S.A. northeast
continental shelf region. Predictions represent the mathematical average of ensemble models of
probability of presence from the years 1984–2015. Model predictions from the first and last 10
years are compared.
References
Engler, R., Hordijk, W., and Guisan, A. 2012. The MIGCLIM R package – seamless
intergration of dispersal constraints into projections of species distribution models.
Ecography 35: 872–878.

53
Saba, V.S., Griffies, S.M., Anderson, W.G., Winton, M., Alexander, M.A., Delworth, T.L., Hare,
J.A., Harrison, M.J., Rosati, A., and Vecchi, G.A. 2015. Enhanced warming of the
Northwest Atlantic Ocean under climate change. J. Geophys. Res. Ocean., 120: 118–132.
https://doi.org/10.1002/2015JC011346.Received
Skern-Mauritzen, M., Ottersen, G., Handegard, N.O., Huse, G., Dingsør, G.E., Stenseth, N.C., and
Kjesbu, O.S. 2015. Ecosystem processes are rarely included in tactical fisheries
management. Fish Fish. 17: 165–175. https://doi.org/10.1111/faf.12111
Thorson, J.T., Shelton, A.O., Ward, E.J., and Skaug, H.J. 2015. Geostatistical delta-generalized
linear mixed models improve precision for estimated abundance indices for West Coast
groundfishes. ICES J. Mar. Sci. 72: 1297-1310.
Thuiller, W., Georges, D., Engler, R., and Breiner, F. 2016. biomod2: Ensemble Platform for
Species Distribution Modeling. R package version 3.3-7. https://CRAN.R-
project.org/package=biomod2, (2), 1–104. Retrieved from https://cran.r-
project.org/web/packages/biomod2/index.html
3.11.1. Group Discussion
Following Dr. Tanaka’s presentation, a suggestion was made that time-frames for predictions
should be quantified scientifically rather than arbitrarily. For example, using timelines relevant to
lobster fisheries stock assessments would be beneficial rather than picking 80-100 years to look
at, simply because that timeframe is commonly used. Variables will likely change drastically on a
smaller timeframe than 100 years, for example SST will likely change drastically on the 10-15
year timeframe. A second comment focused on scale: when a regional scale is used, more attention
must be paid to individual variable responses. When a global scale is used, larger patterns must be
looked at such as global fish decline. Finally, a third comment focused on how to quantify change.
The Gulf of Maine was provided as an example, and how the baseline should be set was
questioned, for example which date to start the analysis at when projecting into the next five or
ten years.
3.12. Determining Thresholds for Interpretation of Probability Maps
Chris Rooper
Alaska Fisheries Science Center, National Marine Fisheries Service, NOAA,
Seattle, Washington, USA
*Current affiliation: Pacific Biological Station, Fisheries and Oceans Canada,
Nanaimo, British Columbia, Canada

54
One of the primary uses of species distribution models is to predict where habitat for a deep-sea
organism occurs and where it does not. This is a binary solution requiring a “yes” or “no” for each
location of interest. Most species distribution models use data to predict a probability of presence
or absence. The movement from probability of presence to presence or absence requires
implementation of a threshold probability. Secondarily, many of the metrics used to evaluate
performance of models are based on predictions of presence or absence (e.g., Cohen’s Kappa,
TSS).
The choice of a threshold for use in a management strategy can have impacts on the results that
are communicated to decision makers. For example, Figure 1 shows an analysis of thresholds for
four models related to eastern Bering Sea trawl fisheries. In this case, the question was asked
“should we be protecting Pribilof Canyon from bottom trawling to protect deep-sea corals”. An
important decision point was how much of the total coral habitat was contained within the canyon.
Three data sources were separately modeled (using two disparate methods) and an ensemble model
was constructed as well. Using a number of different thresholding methods found in the literature,
the proportion of coral habitat in Pribilof Canyon was predicted to range from ~17% to 90%. This
example shows some of the major questions and concerns that arise when using thresholds with
species distribution models. An incomplete list of these concerns that could be discussed are:
 Should we be doing this and are there alternatives?
 Which method should be chosen or how should the determination of method be made?
 Are there general guidelines that can be followed?
 How to best represent uncertainty in conclusions resulting from the choice of a specific
threshold?
A growing body of literature has sought to evaluate some of these questions (e.g., Manel et al.,
2001; Liu et al., 2005; Bean et al., 2011; Lawson et al., 2014). The consensus of most of these
evaluations is that the best methods for thresholding are objective; meaning they set the threshold
using statistics derived from the data (such as the prevalence approach) or the model predictions
(such as sensitivity = specificity approach). However, it appears that the distribution of the
modeled data and the modeling method are important considerations when choosing a
thresholding method.

55
Figure 1. The proportion the total area of coral habitat in eastern Bering Sea that occurs in Pribilof
Canyon predicted by four models as a function of the threshold probability for presence. Lines
indicate the results for Generalized Additive Models of coral presence based on bottom trawl
survey data (bottom trawl), camera survey data (camera survey) and both (Unified model) from
Rooper et al. (2016) and a maximum entropy model based on camera data (Miller et al., 2015).
The black arrows indicate the threshold where the predicted error rates for presence and absence
are equal. The red numbers indicate two suggested thresholds based on high probabilities.
References
Bean, W.T., Stafford, R., and Brashares, J.S. 2011. The effects of small sample size and sample
bias on threshold selection and accuracy assessment of species distribution models. Ecogr.
35: 250–258
Lawson, C.R., Hodgson, J.A., Wilson, R.J., and Richards, S.A. 2014. Prevalence, thresholds and
the performance of presence-absence models. Methods Ecol. Evol. 5: 54-64.
Liu, C.R., Berry, P.M., Dawson, T.P., and Pearson, R.G. 2005. Selecting thresholds of occurrence
in the prediction of species distributions. Ecogr. 28: 385–393.

56
Manel, S., Williams, H.C., and Ormerod, S.J. 2001. Evaluating presence-absence models in
ecology: the need to account for prevalence. J. Appl. Ecol. 38: 921-931.
Miller, R.J., Juska, C., and Hocevar, J. 2015. Submarine canyons as coral and sponge habitat on
the eastern Bering Sea slope. Glob. Ecol. Conserv. 4: 85−94.
Rooper, C.N., Sigler, M., Goddard, .P, Malecha, P.W., Towler, R., Williams, K., Wilborn, R., and
Zimmermann, M. 2016. Validation and improvement of species distribution models for
structure forming invertebrates in the eastern Bering Sea with an independent survey. Mar.
Ecol. Prog. Ser. 551: 117-130.
3.12.1. Group Discussion
The discussion following the presentation focused on thresholds, as well as model validation and
uncertainty measures. To begin with, the concept of deriving a threshold without human selection
was brought up, and whether it is better to use a mathematical based approach rather than a human-
derived approach for thresholds. As well it was noted that the research question supporting the
work will impact how conservative a threshold selection must be, whether it is more dangerous to
have false positives or false negatives is an important consideration. When a threshold is selected,
information is immediately lost, and this is important for managers and stakeholders to be aware
of. An alternative approach was offered which suggested picking for example 80% of density
(where 80% of observations will be occurring) rather than picking a threshold. Following this
discussion, the concept of model validation was briefly covered. In order to calibrate a model, it
must be validated, which can be expensive and unlikely for deep-sea situations. Independent data
can be used for validation: often with a statistical model, available data is used to teach the model
and then in order to validate it, cross-validation can be used rather than retrieving additional field
data for validation. Finally, the concept of measuring uncertainty was covered. One question was
raised concerning using data from one area to predict to another, and if these areas should be
environmentally similar in order to do this. As well it was mentioned that a good idea is assigning
a level of certainty to outputs based on which areas of the final prediction had input data compared
to which areas were completely extrapolated towards, because uncertainty in the data will be
translated to uncertainty in the model. In regards to the research question once more, which is
better, 50% probability with 90% certainty, or 80% probability with lower certainty. This will
depend upon the manager/stakeholder and what the model is being created for.
3.13. Some Uncertainties of Bathymetry Data
Chris Yesson
Institute of Zoology, Zoological Society of London, Regent's Park, London NW1
4RY United Kingdom

57
Bathymetric data are important constituents of many habitat models, both directly in the form of
seabed depth measurements and indirectly as constituents of other environmental layers. For
example, oceanographic models have bathymetric grids as inputs, which help shape predictions of
currents and temperature. Furthermore, we can directly infer seabed characteristics from
bathymetry data by examining local variation in depth to create layers such as slope, aspect and
terrain roughness, which are useful predictors of benthic organisms (e.g., Dolan et al., 2008). We
can also examine the shape of the seabed to predict topographic features such as seamounts (e.g.,
Yesson et al., 2011), which are regarded as important habitats for many species (e.g., Bouchet et
al., 2014). These topographic layers are often the highest resolution datasets upon which to base
predictions of habitat suitability, and can be very important constituents of the models. However,
it is important to consider uncertainties in the underlying bathymetry and consider the implications
of this uncertainty on any resulting model.
Global bathymetry grids such as GEBCO (Weatherall et al., 2015) and SRTM (Becker et al., 2009)
are widely used in large-scale habitat modeling studies. However, the grids themselves are models
based on a combination of soundings (i.e., high resolution acoustic surveys) and satellite altimetry
(lower resolution data from satellite sensors). Satellite- altimetry provides global coverage and is
the foundation of bathymetry models, but these sensors cannot determine small features (i.e.,
seamounts under 1.5km, Wessel et al., 2010). Acoustic surveys generate data best suited for
determining seabed depth and these are used to constrain models used to create bathymetry grids
(Becker et al., 2009). Soundings are limited to a small proportion of the ocean and the majority of
bathymetry grid data is derived from the underlying model, rather than acoustically surveyed. For
example only 18% of GEBCO grid cells are directly supported by acoustic surveys (Weatherall et
al., 2015). With so little sounding data available, there is a premium on making full use of the data
available, and historical soundings (based on weighted lines) have been extracted from nautical
charts to expand these data (Becker et al., 2009).
An example of the shortcomings of these data are presented as an anecdote. A survey was
conducted in February 2016 in the British Indian Ocean Territory to visit seamounts around the
Chagos Archipelago. Two shallow seamounts (summit <500m) predicted by Yesson et al., (2011),
derived from a global bathymetric grid, were visited, but no seabed feature was detected above
1,500m. An examination of Admiralty charts for the area showed that the summits of these features
had soundings reporting “no bottom detected at depth X” where the bathymetry grid (and
subsequent seamount predictions) showed seamount summits at depth X. We believe these “no
seabed” soundings have been misinterpreted as seabed depths in bathymetry grids such as
GEBCO. In this case, spatially isolated, erroneously interpreted sounding data can lead to the false
prediction of seabed features such as seamounts. We estimate that 12.5% of seamounts predicted
to be in the British Indian Ocean Territory by Yesson et al., (2011) are “phantoms” based on
falsely interpreted soundings.

58
Bathymetry grids are continually improving, whether that be from new multibeam acquisition,
such as that collected during the search for Malaysian Airlines flight MH370 (Smith and Marks,
2014), or improved satellite gravity data (Sandwell et al., 2014). However, these bathymetry grids
still rely on sparse sounding data for many regions. It is important that we consider the
uncertainties in the bathymetry when using these to make predictive habitat maps.
References
Becker, J. J., Sandwell, D. T., Smith, W. H. F., Braud, J., Binder, B., Depner, J., Fabre, D., Factor,
J., Ingalls, S., Kim, S-H., Ladner, R., Marks, K., Nelson, S., Pharaoh, A., Trimmer, R., Von
Rosenberg, J., Wallace, G. and Weatherall, P. 2009. Global Bathymetry and Elevation Data
at 30 Arc Seconds Resolution: SRTM30_PLUS. Mar. Geod. 32: 355–371.
Bouchet, P.J., Meeuwig, J.J., Salgado Kent, C.P., Letessier, T.B., and Jenner, C.K. 2014.
Topographic Determinants of Mobile Vertebrate Predator Hotspots: Current Knowledge and
Future Directions. Biol. Rev. 90: 699–728. https://doi.org/10.1111/brv.12130.
Dolan, M.F., Grehan, A.J., Guinan, J.C., and Brown, C. 2008. Modeling the local distribution of
cold-water corals in relation to bathymetric variables: Adding spatial context to deep-sea
video data. Deep Sea Res. Part I: Oceanogr. Res. Pap. 55: 1564-1579.
Sandwell, D.T., Muller, R.D., Smith, W.H.F., Garcia, E., and Francis, R. 2014. New Global
Marine Gravity Model from CryoSat-2 and Jason-1 Reveals Buried Tectonic Structure.
Science 346: 65–67. https://doi.org/10.1126/science.1258213.
Smith, W.H.F., and Marks, K.M. 2014. Seafloor in the Malaysia Airlines Flight MH370 Search
Area. Eos Trans. AGU 95: 173–74. https://doi.org/10.1002/2014eo210001.
Weatherall, P., Marks, K.M., Jakobsson, M., Schmitt, T., Tani, S., Arndt, J.E., Rovere, M.,
Chayes, D., Ferrini, V., and Wigley, R. 2015. A New Digital Bathymetric Model of the
Worlds Oceans. Earth Space Sci. 2: 331–45. https://doi.org/10.1002/2015ea000107.
Wessel, P., Sandwell, D., and Kim, S.S. 2010. The Global Seamount Census. Oceanogr. 23: 24–
33. https://doi.org/10.5670/oceanog.2010.60.
Yesson, C., Clark, M.R., Taylor, M.L., and Rogers, A.D. 2011. The Global Distribution of
Seamounts Based on 30 Arc Seconds Bathymetry Data. Deep Sea Res. Part I: Oceanogr.
Res. Papers 58: 442–53. https://doi.org/10.1016/j.dsr.2011.02.004.

59
3.14. Mapping Uncertainty of SDM Predictions
Fiona Davidson and Anders Knudby
Department of Geography, Environment and Geomatics, University of Ottawa
Introduction
To help decision-makers use species distribution modeling for conservation planning, the accuracy
of predictions made by the model must be estimated. While overall model accuracy is commonly
evaluated using metrics such as the AUC, the uncertainty of predictions is better estimated
spatially, especially when models are used to extrapolate predictions to a new area. The relative
confidence in predictions across separate regions of the modeling extent can be assessed with a
bootstrap method (Anderson et al., 2016; Rowden et al., 2017). This working paper presents the
initial results of such a bootstrap method used to map uncertainty for a spatially extrapolated
model.
Data
The data used for this project included hexactinellid sponge presence and absence data from the
Pacific Ocean along the North American coast (Figure 1).
Methods
We trained a MaxEnt model on data from British Columbia using 19 environmental data layers,
and extrapolated predictions for Alaska. We then trained 200 bootstrapped models, and calculated
the standard deviation (SD) and the coefficient of variation (CV) of the predictions on a cell-by-
cell basis for Alaska. We also applied the sum of sensitivity and specificity threshold to each
bootstrapped model, and quantified the proportion of predictions falling in the least frequent class
multiplied by two, and called this measure categorical uncertainty (CU; e.g., with 40 ‘presence’
and 160 ‘absence’ predictions, CU = 40/200 * 2 = 0.4). We then compared model predictions with
the presence/absence data from Alaska, to show the relationship between estimated uncertainty
(SD/CV/CU) and the actual accuracy of predictions.

60
Figure 1. Sponge Data. Glass sponge presences in British Columbia, Canada (Red) and Alaska,
USA (Blue), and absences (Green).
Results
Figure 2 shows the predictions made for the Alaska region, extrapolated using the MaxEnt model
trained with data from British Columbia, as well as the three uncertainty metrics. Figure 3 shows
the overall accuracy of the predictions, binned into 10 equally spaced ranges of uncertainty
estimates.
Discussion
For the SD and CU metrics, data bins with higher uncertainty (on the right side of each plot) have
lower accuracies, and vice versa, suggesting that these two metrics can effectively be used to
quantify the amount of confidence that should be placed in the predictions. For the CV metric, on
the other hand, data bins with high uncertainty have high accuracy of predictions, suggesting that
this metric should not be used to guide confidence in predictions.

61
Figure 2. Top left: Mapped predictions of sponge presence/absence in Alaska. Top right: Standard
deviation (SD) of bootstrapped predictions. Bottom left: Coefficient of variation (CV) of
bootstrapped predictions. Bottom right: Categorical uncertainty (CU) based on bootstrapped
predictions.
Figure 3. SD, CV and CU values, binned, and the corresponding accuracy of predictions.

62
References
Anderson, O.F., Guinotte, J.M., Rowden, A.A., Tracey, D.M., Mackay, K.A., and Clark, M.R.
2016. Habitat suitability models for predicting the occurrence of vulnerable marine
ecosystems in the seas around New Zealand. Deep-Sea Res. Part I: Oceanogr. Res. Pap. 115:
265–292.
Rowden, A.A., Anderson, O.F., Georgian, S.E., Bowden, D.A., Clark, M.R., Pallentin, A., and
Miller, A. 2017. High-Resolution Habitat Suitability Models for the Conservation and
Management of Vulnerable Marine Ecosystems on the Louisville Seamount Chain, South
Pacific Ocean. Front. Mar. Sci. 4: 335.
4. THEME 2 - BIOLOGICAL AND ENVIRONMENTAL DATASETS RELEVANT
TO DEEP-SEA SPECIES AND COMMUNITIES
Chair: Dr. Telmo Morato
Issues related to data gaps and data quality hampering the development of improved species
distribution models in the deep sea were discussed. This discussion focused largely on data
quality, accuracy, extent, resolution, ecological relevance, and temporal scales.
Some of the points made are summarised below:
1. Overall data availability is improving. Archiving the available environmental data and
derived data so that it is open access would be important moving forward.
2. Scientists should be encouraged to publish new environmental data in ‘data journals’
rather than as supplementary data linked to the original research paper to facilitate more
ready access.
3. There is a need for a shared data repository, for example linked to an expert forum similar
to the group assembled for this workshop. Data layers from many institutes could be
continuously updated, however quality control and oversight would be required for such
a project. Also, a critique of new papers coming out would also be helpful. A web portal
could be created with metadata of relevant datasets. Users could rank their usefulness
using a simple 1 to 5 star system. An example of an existing repository is BioOracle
(http://www.bio-oracle.org). Action: Dr. Andrew Davies agreed to set up a simple
website where relevant datasets could be listed in order to facilitate species distribution
modeller’s efforts.
4. Required data quality relates directly to the scale at which the model will be used.

63
Biological data needs to be georeferenced with much greater care and attention; often,
very high-quality bathymetry is available for an area and the georeferencing of biological
data becomes the weak point. Actual position reference data is required for samples,
rather than data aggregated at a 1 km cell size. Often, global databases contain limited
metadata describing what and how data were collected. Often only the geographic
coordinates describing the location of the organism are given with no information about
gear type/length of tow used to collect the sample available - this increases uncertainty.
OBIS (Ocean Biogeographic Information System) facilitates entry of comprehensive
metadata associated with biological data if available.
5. In terms of ecologically relevant data, for species such as corals and sponges, process
studies are very important in identifying variables which are important to individual
species. It was noted that adding hydrodynamic data into species distribution models will
be an important next step (e.g., Mohn et al., 2014).
A short discussion on the topic of new data sources proposed the inclusion of detection
probability. To do this, repeat sampling surveys would be needed, but this will only be possible
for some species. Occupancy counts are often lower than actual occupancy due to the probability
of missing some counts; this should be quantified in some way, e.g., using still images taken
from the same sampling sites. For cryptic deep-water sponge species - repeat images could yield
a value for detection probability. How this differs across the study area is also potentially
significant.
The importance of biological traits was then discussed and how they might be incorporated into
modeling approaches. A major issue identified was the availability of sufficient data. It was
mentioned that information pertaining to other closely related species could be used to help fill
gaps in knowledge for certain species (see e.g., the Marine Ecosystem Research Program
(MERP); the Trait Explorer).
The Hierarchical Modeling of Species Communities (HMSC) package opens the door for
including traits and modeling the response to covariates in the model.
1. Joint Species Distribution Models (JSDMs) are still rather new; it has been demonstrated
that when trait information is included, the result is similar to responses of species to
their environment. In some cases, it has been demonstrated as a positive influence on the
model, but other cases show including species-to-species interactions also helps; this
does not need to come from trait inform