Working with what we know – presence-only / ecological ... · WHAT? What are they? • Predict...

Working with what we know –presence-only / ecological niche models

in marine mammal science

Kristin Kaschner, Colin MacLeod, Laura Mandleberg, Ross Comptom

Sea Around Us Project, Fisheries Centre, University of British Columbia, Canada &FTZ Büsum, Kiel Christian-Albrechts-University, Germany

School of Biological Sciences (Zoology), University of Aberdeen, UK

Outline

• What are presence-only models?

• Why do we use them?

• Which ones are there?

• Do they work?

• What can we do with them?

OUTLINE

What are they?WHAT?

• Predict ecological niches

• Use only presence data

Why do we use them?WHY?

• Data paucity

• Absence data issues

• Niche modeling vs. distribution

Data paucity

Hawaii

WHY?

Data paucityWHY?

OBIS-SEAMAP (http://seamap.env.duke.edu/species)

- compilation & storage of marine mammal occurrence data

- out of 115 species, geo-referenced / effort corrected data

- available/accessible for ~ 50%

- representative coverage = ~ 2 %

Data paucityWHY?

Model evaluationPresence-absence confusion matrix

Predicted presence

Predicted absence

Recorded presence

Recorded absence

a (true presence)

c (false absence)

b (false presence)

d (true absence)

WHY? Absence Data Issues


Predicted presence

Predicted absence

Recorded presence

Recorded absence

a (true presence)

c (false absence)

b (false presence)

d (true absence)


Omission error / Model overfitting

Commission error / Model overprediction


Predicted presence

Predicted absence

Recorded presence

Recorded absence

a (true presence)

c (false absence)

b (false presence)

d (true or perceived absence????)


Site with environmental

value X

Species present?

Site visited? Species detected?

Presence-only data

M. Nakamura, CONABIO, 2005GBIF Ecological Niche Modelling Workshop, KU

Absence Data IssuesWHY?


value X

Species present?


Presence-absence data




value X

Species present?


True absence data


Absence Data Issues

False absence data

WHY?


0

0.5

1

1.5

2

2.5

0 20 40 60 80Depth [m]

Mea

n de

nsity

[# a

nim

als

/ km

2]

Harbour porpoise density

No animals in deeper waters!

Scheidat, Gilles et al, (unpublished data)


0

0.5

1

1.5

2

2.5

0 20 40 60 80Depth [m]

Mea

n de

nsity

[# a

nim

als

/ km

2]

0

2

4

6

8

10

12

0 20 40 60 80Depth [m]

Mea

n Ef

fort

[km

2]

Harbour porpoise density

Aerial survey effort

No animals in deeper waters?

True absences???


0

10

20

30

40

50

60

0 5 10 15 20 25

Effort [km2]

# si

ghtin

gs


Spring 2002-2005



Spring

0

10

20

30

40

50

60

70

80

0 10 20 30Effort category [km2]

Mea

n #

sigh

tings

Summer

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings

Fall

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings



Spring

R2 = 0.4597

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings

Summer

R2 = 0.6622

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings

Fall

R2 = 0.6069

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings



Spring

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings

Summer

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings

Fall

0

10

20

30

40

50

60

70

80


Mea

n #

sigh

tings

True absences???


.20

Variable A

Species #1

Presence

.8Bias

.16Obs. rate

Example: 100 trials

16Observed

.80

Variable B

.1

.08

8

0

Variable C

.1

0

0

× ××

= ==

.32

Variable A

Species #2

.5

.16

16

.16

Variable B

.5

.08

8

.52

Variable C

0

0

0

×× ×

== =



Model Complexity

Pre

dict

ion

Err

or

Low High

Low Variance High Variance

Training sample

Test sample

Hastie et al. (2001)

WHY? Ecological Niche vs Distribution

Model Complexity

Pre

dict

ion

Err

or

Low High

Low Variance High Variance

Training sample

Test sample

Hastie et al. (2001)


Ecological Niche Model

Distribution Model


J. Soberon, CONABIO, 2005GBIF Ecological Niche Modelling Workshop, KU

Geographic space(2 dimensions)

Ecological space(n dimensions / hypervolume)

Which ones are out there?

Envelope models• BioClim• DOMAIN• Fuzzy bioclimatic envelope model• RES*

Machine-learning• Garp**• Maxent**

• ENFA (Biomapper)*

WHICH?

* Models have been applied to and tested for marine mammals ** Preliminary applications to marinemammals

Which ones are out there?Web-based applications• AquaMaps*• KGS-Mapper**• WhyWhere?

* Models have been applied to and tested for marine mammals ** Preliminary applications to marinemammals

WHICH?

Ecological Niche ModelsWHICH?

Geographic spaceEcological / Environmental space

J. Soberon

BioClim / Climate Envelope Range

• Based on presence cells

• Very simple & intuitive

• No interactions betweenvariables

• Unweighted variables

• Binary predictions

• No extrapolations

• Tends to overpredict

Lindenmayer et al. 1991 J. Biogeog. 18: 371-383.Arcscript: http://arcscripts.esri.com/details.asp?dbid=13745

WHICH?

J. Soberon

RES

Kaschner et al. 2006, MEPS / www.seaaroundus.orgSkov & Svenning, 2004, Ecography, 27:366-380

Relative Environmental Suitability Model(Fuzzy Bioclimatic Envelope Model)

Assigned habitat usage categories: Depth, SST, Ice edge

0 -200 -1000 -2000 -3000 -4000 -5000 -6000 -7000 -8000

Depth [m]

Prob

abilit

y

Pmax

Mainly continental slope

-2 0 5 10 15 20 25 30

Subpolar – warm temp.

SST [° C]

Pmax

*-1 0 1 500 1000 2000 8000

No association with ice edge

Distance from ice edge [km]

Pmax

*

• Based on relative occurrences

• Very simple, transparent & intuitive

• Expert knowledge based (no point data required)



• Continuous output

• Tends to overpredict

Sowerby‘s beaked whaleSowerby‘s beaked whale

WHICH?

DOMAIN


• Cluster algorithm in environmental space



• Non-binary predictions

• Tends to overfit

Carpenter et al. 1993 Biodiv. Conservation 2: 667-680.Freeware: http://www.cifor.cgiar.org/docs/_ref/research_tools/domain/

WHICH?

J. Soberon

GARP Genetic Algorithm of Rule-set Prediction


• Machine learning / automated model optimization (not transparent)

• Generates pseudo-absencedata

• Optimization using training / test data sets

• Interactions between weightedvariables


• Tends to overpredict?

Stockwell & Noble 1992, Math. & Comp. in Simulation 33:385-390Freeware: http://www.lifemapper.org/desktopgarp/

WHICH?

J. Soberon

GARP Genetic Algorithm of Rule-set Prediction

Stockwell & Noble 1992, Math. & Comp. in Simulation 33:385-390Freeware: http://www.lifemapper.org/desktopgarp/

Minke whale (IWC whaling data)



• Generates pseudo-absencedata




• Tends to overpredict?

WHICH?

MaxentMaximum entropy






• Tends to overfit?

Phillips et al. 2006, Ecological ModellingFreeware: http://www.cs.princeton.edu/~schapire/maxent/

Minke whale (IWC whaling data)

WHICH?

ENFAEcological Niche Factor Analysis


• Compares species occurrence to all available habitat

• Multivariate (Interactionsbetween weighted variables)

• Non-binary predictions (HIS)

Hirzel et al. 2002, Ecology 83: 2027-2036.Biomapper freeware: http://www2.unil.ch/biomapper/

WHICH?

Species occurrence cells Available habitat



• Compares species occurrence to all available habitat

• Multivariate (Interactionsbetween weighted variables)

• Non-binary predictions (HIS)

Hirzel et al. 2002, Ecology 83: 2027-2036.Biomapper freeware: http://www2.unil.ch/biomapper/

WHICH?

Marginality

Specialization


Northern bottlenose whale

(NW Atlantic & Total N Atlantic)

Compton, 2004, MRes Thesis, U of PlymouthMacLeod, 2005, PhD Thesis, U of Aberdeen

WHICH?

MarginalCore

UnsuitableHabitat Suitability

AquaMaps (Coming soon....)

Kaschner et al, in prep,www.fishbase.de

WHICH?

Do they work?Model Evaluations

• Test statistics

• Cross-validation

• External testing

• Comparison with other models

DO THEY WORK?

Predicted presence

Predicted absence

Recorded presence

Recorded absence

a (true presence)

c (false absence)

b (false presence)

d (true absence)

Test statisticsDO THEY WORK?

Sensitivity: Proportion of observed presences correctly predicteda/(a + c)

R. Pearson, AMNH, 2005GBIF Ecological Niche Modelling Workshop, KU

Predicted presence

Predicted absence

Recorded presence

Recorded absence

a (true presence)

c (false absence)

b (false presence)

d (true absence)


Sensitivity: Proportion of observed presences correctly predicteda/(a + c)


Specificity: Proportion of observed (or assumed) absences correctly predicted d/(b + d)

Test statistics: Receiver Operator Curve

DO THEY WORK?

AUC = 0.5 = random

Area Under Curve (AUC)

Predicted presence

Predicted absence

Recorded presence

Recorded absence

a (true presence)

c (false absence)

b (false presence)

d (true absence)



Cohen’s Kappa:

)]/)))(())(((([)]/)))(())(((()[(

ndcdbbacanndcdbbacadak

+++++−+++++−+=

Araújo et al. 2005 Gl. Ch. Biol.

Environmental envelope

Evaluation

Calibration

Projection

Same region

New region

New resolution

New time


Evaluation

Calibration

Projection

Same region

New region

New resolution

New time

Same region

New region

New resolution

New time

DO THEY WORK? Cross-validation

0

2

4

6

8

10

12

14

16

0-33 34-66 67-100

HS bin range

Mea

n ar

e-ad

just

ed fr

eque

ncy

Compton, 2004, MRes Thesis, U of Plymouth

Northern bottlenose whale (NW Atlantic)

• Same data set

• Jack-knife procedure


Evaluation

Calibration

Projection

New region

New resolution

New time

100%Environmental envelope

Evaluation

Calibration

Projection

New region

New resolution

New time

New region

New resolution

New time

100%

Araújo et al. 2005 Gl. Ch. Biol.

DO THEY WORK? External Testing

Kaschner et al, 2006, MEPS

Southern bottlenosewhale

RES category

Avg

SP

UE

per

cel

l Common nameComparison with random data sets

rho p Simulated p-valueNorthern fur seal 0.54 < 0.0001 0Harbour porpoise 0.59 < 0.0001 0Sperm whale 0.66 < 0.0001 0Antarctic minke whale 0.71 < 0.0001 0Killer whale 0.56 < 0.0001 0.54Fin whale 0.53 < 0.0001 0Blue whale 0.48 < 0.0001 0.268Humpback whale 0.20 <0.05 0.006Hourglass dolphin 0.68 < 0.0001 0Southern bottlenose whale 0.83 < 0.0001 0Dwarf minke whale -0.77 < 0.0001 0

Spearman's non-parametric rank correlation analysis of RES vs mean SPUE

• Different data set

• Permutation

DO THEY WORK? Model Comparison

Mandelbaum, 2005, MSc Thesis, U of Aberdeen

PCA ENFA

GARP GLM

Presence

Absence

Harbour porpoise



0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

PCA ENFA

GARP GLM

AUC = 0.746 AUC = 0.745

AUC = 0.773 AUC = 0.828

1-specificity

sens

itivi

ty

Significantly different



43210

Combined model

Do they work?Limitations

• Presence cells vs. encounter rates

• Effort / Sampling biases

DO THEY WORK?

DO THEY WORK? Presence cells vs Encounter Rate

Minke whale

Maxent prediction (IWC whaling data)

Kaschner et al, in prep

SOWER sightingsWhaling data

DO THEY WORK? Presence cells vs Encounter Rate

Minke whale

Maxent prediction (IWC whaling data)

Kaschner et al, in prep

SOWER SPUE

DO THEY WORK? Sampling Biases


Minke whale Blue whale Humpback whale

DO THEY WORK?

Minke whale Blue whale Humpback whale


Sampling Biases

What can we do with them? Potential Applications

• Biodiversity Mapping

• Management / Research Prioritiziation

– Risk mitigation

– Marine mammal-fisheries interactions

– Marine Protected Areas

– Climate change

WHAT FOR?

What can we do with them? Potential Applications

• Biodiversity Mapping


– Risk mitigation



– Climate change

WHAT FOR?

Kaschner & WormWednesday, 15:15


– Risk mitigation

– Marine mammal fisheries interactions


– Climate change

Applications

RES > 0.8

Beaked whales

WHAT FOR?


– Risk mitigation



– Climate change

Applications

RES > 0.8

Beaked whales

M. densirostris

All Mesoplodon Spp.

MacLeod, 2005, PhD Thesis, U of Aberdeen

WHAT FOR?

Applications


– Risk mitigation



– Climate change

APPLICATIONS

FisheriesFood consumption

Resource overlap

Kaschner et al, accepted, CJFAS

Applications


– Risk mitigation



– Climate change

& Seamounts

Species richness

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0 5 10 15 20 25 30 35 40

Species richness category

Mea

n se

amou

nt d

ensi

ty

Kaschner, 2006, Ecology of Seamounts

WHAT FOR?

-0.50-0.40-0.30-0.20-0.100.000.100.200.300.400.50

Applications


– Risk mitigation



– Climate change

Small toothed whales

Predicted relative changes in distribution size by 2020

WHAT FOR?

• require less data / maximize available data

• less affected by false absences

• more general, useful to investigate large scale patterns & ecological interactions

• time & cost efficient starting points

• can supplement small scale studies and help to focus research and management efforts

Presence-only / Ecological niche models

ConclusionsCONCLUSIONS

• Ed Gregr, MMRU, UBC, Vancouver

• Anita Gilles & Roger Mundry, FTZ Buesum

• Instructors & Participants of the GBIF Ecological Niche Modelling Workshop, 5-9 Dec 2005, Lawrence, Kansas

Acknowledgements

Working with what we know – presence-only / ecological ... · WHAT? What are they? • Predict...

Documents

Transcript of Working with what we know – presence-only / ecological ... · WHAT? What are they? • Predict...