curso lichos dia1
TRANSCRIPT
BEGIN WITH A DETOUR … NSB
Estimate Ecological Niche
No Silver Bullets in ENM
• Single algorithms may perform ‘best’ on average• The best algorithm in any given situation, however,
may be other than the ‘best’• NSB thinking suggests that we should not use a
single approach• Use a suite of approaches (e.g., as implemented in
OM, BIOMOD, BIOENSEMBLES, etc.), challenge to predict, choose best for that situation
• Maxent is good, but it is not the only algorithm …
COMPLEXITY OF MODELS AND OVERFITTING
Independent variables
Mod
eled
resp
onse
Independent variables
Mod
eled
resp
onse
Independent variables
Mod
eled
resp
onse
Independent variables
Mod
eled
resp
onse
Which Model is Best?
Which Model is Best?
Least calibration error
Most calibrationerror
Moderate calibrationerror
Which Model is Best?
Least calibration error
Most calibrationerror
Moderate calibrationerror
NEED INDEPENDENT DATA!
Five Goals of Niche Modeling
1. ESTIMATE THE FUNDAMENTAL NICHE2. ESTIMATE THE FUNDAMENTAL NICHE3. ESTIMATE THE FUNDAMENTAL NICHE4. ESTIMATE THE FUNDAMENTAL NICHE5. ESTIMATE THE FUNDAMENTAL NICHE
How Would the Fundamental Niche Look?
• In any one dimension, expected to be unimodal
• In multiple dimensions, expected to be convex• So, simple models are probably better• Need to take sampling and incomplete
representation into account carefully
http://d1vn86fw4xmcz1.cloudfront.net/content/royptb/367/1596/1665/F1.large.jpg
The Area of Distribution
G Physiological requirements
(Abiotic)
A
Favorable bioticenvironment
(Biotic)B
Accessible to dispersal(Movements)
M
Fundamental niche
Existing fundamental
niche
Realized ecological
niche
Existing niches are irregular, concave, and complex
Fundamental niches should be simple and convex
Maxent Model Responses
Current Methods• Popular methods fit highly complex objects to
estimate niches … but which niche?• Complex objects are more likely to correspond
to the existing niche, rather than the fundamental
Simplify:
• Complex response forms created by many current algorithms do not fit well with our understanding of the fundamental niche
• Simple, convex response forms may be much more appropriate as approximations to the fundamental niche
• This thinking will most likely require algorithms that can incorporate incomplete data, as well as uncertainty in inference, but fit simple response forms
Which Model is Best?
UNDERFITTING OVERFITTING
AIC: A Means of Choosing
Complex Models• Fit better with realized
niche and SDM• Risk overfitting• Frequently will be limited in
predictive ability
Simple Models• Fit better with ecological
niche theory• Risk underfitting• May not address the full
complexity of distributional ecology
THRESHOLDING MODELS
Absence DataAbiotic niche
Biotic interactionsAccessibility
“Presence”
E parameter
ENM Thresholding
• Least Training Presence Thresholding seeks the highest threshold level that includes all calibration data, or T100
• The T100 approach fails when there is error in the occurrence data set
• E is a summary of the expected proportion of data in the calibration dataset that will include significant errors
• And “Adjusted LTPT” approach would seek T100-E as a thresholding approach ideal for ENM
POST-PROCESSING MODELS
You Fit Your ENM… And Now?
• Niches to distributions• Consideration of dispersal (non-equilibrium
situations) • Add consideration of factors not able to be
incorporated directly in the model– E.g., land-use change, human presence
You Fit Your ENM… And Now?
• Niches to distributions• Consideration of dispersal (non-equilibrium
situations) • Add consideration of factors not able to be
incorporated directly in the model– E.g., land-use change, human presence
Potential DistributionAbiotic niche
Biotic interactions
Potential and Actual DistributionsAbiotic niche
Biotic interactionsAccessibility
Niches to Distributions
• In geographic space, distributions are the results of the BAM intersection:
GO = A ∩ B ∩ M• In “canonical” ENM, the model output is something
like A, or A∩B if the Eltonian Noise Hypothesis holds
• What does this say about the semantics?– Ecological niche modeling– Species distribution modeling
Niches to Distributions
GO = A ∩ B ∩ M• Need to restrict model outputs via some
hypothesis of M … post hoc?
Test Arena: The Lawrence Species
M and Model Training
M and Model Validation
Model Evaluation
M and Model Comparison
Model Comparison
The Area of Distribution
G Physiological requirements
(Abiotic)
A
Favorable bioticenvironment
(Biotic)B
Accessible to dispersal(Movements)
M
Sampling Effort
Interesting Result
• If model calibration is constrained to M, and if the Eltonian Noise Hypothesis holds, then …
GO = A• Simplifies the modeling process enormously,
and eliminates the need for post-modeling assumptions about M
You Fit Your ENM… And Now?
• Niches to distributions• Consideration of dispersal (non-equilibrium
situations) • Add consideration of factors not able to be
incorporated directly in the model– E.g., land-use change, human presence
What About Non-equilibrium Situations?
• Non-equilibrium:– Species does not inhabit the entire spatial
footprint of its habitable area• E.g., a projection of future-climate potential
distribution of a species• E.g., an invading species that has only established
populations in a small part of its potential range• E.g., a species being evaluated in terms of invasive
potential on other continents
• Need to bring in consideration of dispersal
https://youtu.be/dKdNP42B0Aw
You Fit Your ENM… And Now?
• Niches to distributions• Consideration of dispersal (non-equilibrium
situations) • Add consideration of factors not able to be
incorporated directly in the model– E.g., land-use change, human presence
Often Cannot Include Land Cover…
• Occurrence data are too old, or are heterogeneous temporally
• Occurrence data are not sufficiently finely resolved spatially to permit link to land cover
• Land use change is ongoing, such that there is not a single, stable state
• Solution: incorporate after model calibration
Adding Fine-scale Effects
• When ENMs cannot be calibrated at the fine resolutions of land use
• Can make the assumption that species have particular land-use associations, and that they will be found only under certain land use types
• The coarse, climate-based ENM output can be “clipped” by the finer-resolution land use outputs