paul eastwood centre for environment, fisheries and aquaculture science (cefas) lowestoft, uk...
TRANSCRIPT
Paul Eastwood
Centre for Environment, Fisheries and Aquaculture Science (Cefas)Lowestoft, UK
INCOFISH WP3 Brazil workshop
Species distribution modelling
• All started in early 1980s with US Fish and Wildlife Service
• Framework for predicting habitat suitability based on known preferences and tolerances
• Habitat Suitability Index (HSI) modelling
• HSI models formulated from word, graphical or mathematical expressions that described the relationship between a species’ life-history stage and its environment
HSI modelling
• Early HSI models were non-spatially structured
• GIS & digital spatial data were not widely available
• Models developed primarily for terrestrial species
0
2
4
6
8
10
0 0.2 0.4 0.6 0.8 1
Habitat factor
Spe
cies
den
sity
0.0
0.2
0.4
0.6
0.8
1.0
Sui
tabi
lity
inde
x
0
2
4
6
8
10
12
0 0.2 0.4 0.6 0.8 1 1.2
Habitat factor
Spe
cies
den
sity
0.0
0.2
0.4
0.6
0.8
1.0
Sui
tabi
lity
inde
xSI1 SI2 Geometric mean HSI
= (SI1 + SI2)0.5
HSI & GIS modelling
1.0
0.5
028 29 30 31 32
1.0
0.5
0
7 8 9 10 11
1.0
0.5
010 20 30 40 50
1.0
0.5
0
A B
Temperature
Substrate type
Salinity
Depth
Modelled fish-habitat relationships (SI’s)
Temperature SI map
Depth SI map
Salinity SI map
Substrate SI map
HSI =
1/4
Low suitability
High suitability
Unsuitable Medium
Habitat suitability index map
Digital environmental maps recoded with the SI’s
Many ways to skin the cat…
From Guisan and Thuiller (2005)
Why so many methods?
• Distributional data come in different forms
– Relative abundance
– Presence-absence
– Presence only
• Try and improve predictions
• Resolve some of the (false) assumptions made by HSI models, e.g. all habitat variables selected independently
• And also because we’re scientists and are always looking for better and more efficient solutions
Habitat factor
Response e.g. catch
density Average but non-limiting effect
Quantile regression for SDM
Limiting effect
Common sole in the eastern Channel Absent High
Catch density
Spawning Spawning
Nursery Nursery
Adult summerfeeding
Adult summerfeeding
Models based on upper RQs Models based on central RQsNon-limitingLimiting
Model selection
• Model construction is not an exact science
• Environmental factors can be few or many
• Models fitted using linear and non-linear functions, parametric and non-parametric
From Oksanen and Minchin (2002)
Modelling procedure
• Typical procedure for constructing a species distribution model
Define input variables• species data• environmental data
Model construction• Selection of variables• Significance tests• Assessment of fit (AIC)
Model validation • Internal• External
Final model anddistribution map
Success Fail?????
Model validation
• Measures of predictive performance are generally all based on a confusion matrix:
Model validation
• Performance measures based on confusion matrix
From Fielding and Bell (1997)
Model validation
• Some measures influenced by species prevalence
• Not an issue for INCOFISH as only have presence data
From Fielding and Bell (1997)
Model validation
• Issues for Aquamaps…
• Maps generated at global scale using all data
• Therefore, validation measures would be internal not based on external data
• Would either have to
– accept this
– generate bootstrap samples
– withhold some data for model testing