maximum entropy and species distribution modeling
TRANSCRIPT
![Page 1: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/1.jpg)
Maximum Entropy andMaximum Entropy andMaximum Entropy andMaximum Entropy andMaximum Entropy andSpecies Distribution ModelingSpecies Distribution ModelingSpecies Distribution ModelingSpecies Distribution ModelingSpecies Distribution Modeling
Rob Schapire
Steven PhillipsMiro Dudık
Also including work by or with: Rob Anderson, Jane Elith, Catherine Gra-
ham, Chris Raxworthy, NCEAS Working Group, ...
![Page 2: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/2.jpg)
The Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat Modeling
• goal: model distribution of plant or animal species
![Page 3: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/3.jpg)
The Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat Modeling
• goal: model distribution of plant or animal species
• given: presence records
![Page 4: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/4.jpg)
The Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat Modeling
• goal: model distribution of plant or animal species
• given: presence records
• given: environmental variables
precipitation wet days avg. temp.
· · ·
![Page 5: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/5.jpg)
The Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat ModelingThe Problem: Species Habitat Modeling
• goal: model distribution of plant or animal species
• given: presence records
• given: environmental variables
precipitation wet days avg. temp.
· · ·
• desired output: map of range
![Page 6: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/6.jpg)
Biological ImportanceBiological ImportanceBiological ImportanceBiological ImportanceBiological Importance
• fundamental question: what are survival requirements (niche)of given species?
• core problem for conservation of species
• first step for many applications:
• reserve design• impact of climate change• discovery of new species• clarification of taxonomic boundaries
![Page 7: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/7.jpg)
A Challenge for Machine LearningA Challenge for Machine LearningA Challenge for Machine LearningA Challenge for Machine LearningA Challenge for Machine Learning
• no negative examples
• very limited data• often, only 20-100 presence records• usually, not systematically collected• may be museum records years or decades old
• sample bias• (toward most accessible locations)
![Page 8: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/8.jpg)
Our ApproachOur ApproachOur ApproachOur ApproachOur Approach
• assume presence records come from probability distribution π
• try to estimate π
• apply maximum entropy approach
![Page 9: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/9.jpg)
Maxent ApproachMaxent ApproachMaxent ApproachMaxent ApproachMaxent Approach
• presence records for species X:
altitude July temp. · · ·
record #1 1410m 16◦C · · ·
record #2 1217m 22◦C · · ·
record #3 1351m 17◦C · · ·...
......
![Page 10: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/10.jpg)
Maxent ApproachMaxent ApproachMaxent ApproachMaxent ApproachMaxent Approach
• presence records for species X:
altitude July temp. · · ·
record #1 1410m 16◦C · · ·
record #2 1217m 22◦C · · ·
record #3 1351m 17◦C · · ·...
......
Average 1327m 17.2◦C · · ·
Stand. Dev. 117m 3.2◦C · · ·
![Page 11: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/11.jpg)
Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)
• data allow us to infer many facts, e.g.:• average altitude of species X’s habitat ≈ 1327m• average July temperature of species X’s habitat ≈ 17.2◦C
...• stand. dev. of altitude of species X’s habitat ≈ 117m
...
![Page 12: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/12.jpg)
Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)
• data allow us to infer many facts, e.g.:• average altitude of species X’s habitat ≈ 1327m• average July temperature of species X’s habitat ≈ 17.2◦C
...• stand. dev. of altitude of species X’s habitat ≈ 117m
...• probability species X lives above 1100m ≈ 0.78• probability species X lives above 1200m ≈ 0.62
...
![Page 13: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/13.jpg)
Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)Maxent Approach (cont.)
• data allow us to infer many facts, e.g.:• average altitude of species X’s habitat ≈ 1327m• average July temperature of species X’s habitat ≈ 17.2◦C
...• stand. dev. of altitude of species X’s habitat ≈ 117m
...• probability species X lives above 1100m ≈ 0.78• probability species X lives above 1200m ≈ 0.62
...
• each tells us something about true distribution
• idea: find distribution satisfying all constraints
• among these, choose distribution closest to uniform(i.e., of highest entropy)
![Page 14: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/14.jpg)
This TalkThis TalkThis TalkThis TalkThis Talk
• theory
• maxent with relaxed constraints• new performance guarantees for maxent
• useful even with very large number of features (orconstraints)
• algorithm and convergence
• experiments and applications
![Page 15: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/15.jpg)
The Abstract FrameworkThe Abstract FrameworkThe Abstract FrameworkThe Abstract FrameworkThe Abstract Framework
• π = (unknown) true distribution
[Della Pietra, Della Pietra & Lafferty]
![Page 16: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/16.jpg)
The Abstract FrameworkThe Abstract FrameworkThe Abstract FrameworkThe Abstract FrameworkThe Abstract Framework
• π = (unknown) true distribution
• given:• samples x1, . . . , xm ∈ X
xi ∼ π• features f1, . . . , fn
fj : X → [0, 1]
[Della Pietra, Della Pietra & Lafferty]
![Page 17: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/17.jpg)
The Abstract FrameworkThe Abstract FrameworkThe Abstract FrameworkThe Abstract FrameworkThe Abstract Framework
• π = (unknown) true distribution
• given:• samples x1, . . . , xm ∈ X
xi ∼ π• features f1, . . . , fn
fj : X → [0, 1]
• goal: find π = estimate of π
[Della Pietra, Della Pietra & Lafferty]
![Page 18: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/18.jpg)
Maxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat Modeling
• xi = presence record
• X = all localities (discretized)
• π = distribution of localities inhabited by species
• ignores sample bias and dependence between samples
![Page 19: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/19.jpg)
Maxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat Modeling
• xi = presence record
• X = all localities (discretized)
• π = distribution of localities inhabited by species
• ignores sample bias and dependence between samples
• features: use raw environmental variables vk
• linear: fj = vk
![Page 20: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/20.jpg)
Maxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat Modeling
• xi = presence record
• X = all localities (discretized)
• π = distribution of localities inhabited by species
• ignores sample bias and dependence between samples
• features: use raw environmental variables vk or derivedfunctions
• linear: fj = vk
• quadratic: fj = v2
k
![Page 21: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/21.jpg)
Maxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat Modeling
• xi = presence record
• X = all localities (discretized)
• π = distribution of localities inhabited by species
• ignores sample bias and dependence between samples
• features: use raw environmental variables vk or derivedfunctions
• linear: fj = vk
• quadratic: fj = v2
k
• product: fj = vkv`
![Page 22: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/22.jpg)
Maxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat Modeling
• xi = presence record
• X = all localities (discretized)
• π = distribution of localities inhabited by species
• ignores sample bias and dependence between samples
• features: use raw environmental variables vk or derivedfunctions
• linear: fj = vk
• quadratic: fj = v2
k
• product: fj = vkv`
• threshold: fj =
{1 if vk ≥ a
0 else
![Page 23: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/23.jpg)
Maxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat ModelingMaxent and Habitat Modeling
• xi = presence record
• X = all localities (discretized)
• π = distribution of localities inhabited by species
• ignores sample bias and dependence between samples
• features: use raw environmental variables vk or derivedfunctions
• linear: fj = vk
• quadratic: fj = v2
k
• product: fj = vkv`
• threshold: fj =
{1 if vk ≥ a
0 else
• # features can become very large (even infinite)
![Page 24: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/24.jpg)
A Bit More NotationA Bit More NotationA Bit More NotationA Bit More NotationA Bit More Notation
• π = empirical distribution• i.e., π(x) = #{i : xi = x}/m
• π[f ] = expectation of f with respect to π(so π[f ] = empirical average of f )
![Page 25: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/25.jpg)
MaxentMaxentMaxentMaxentMaxent
• π typically very poor estimate of π
![Page 26: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/26.jpg)
MaxentMaxentMaxentMaxentMaxent
• π typically very poor estimate of π
• but: π[fj ] likely to be reasonable estimate of π[fj ]
![Page 27: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/27.jpg)
MaxentMaxentMaxentMaxentMaxent
• π typically very poor estimate of π
• but: π[fj ] likely to be reasonable estimate of π[fj ]
• so: choose distribution π such that
π[fj ] = π[fj ]
for all features fj
![Page 28: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/28.jpg)
MaxentMaxentMaxentMaxentMaxent
• π typically very poor estimate of π
• but: π[fj ] likely to be reasonable estimate of π[fj ]
• so: choose distribution π such that
π[fj ] = π[fj ]
for all features fj
• among these, choose one closest to uniform,i.e., of maximum entropy [Jaynes]
![Page 29: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/29.jpg)
MaxentMaxentMaxentMaxentMaxent
• π typically very poor estimate of π
• but: π[fj ] likely to be reasonable estimate of π[fj ]
• so: choose distribution π such that
π[fj ] = π[fj ]
for all features fj
• among these, choose one closest to uniform,i.e., of maximum entropy [Jaynes]
• problem: can badly overfit, especially with a large number offeatures
![Page 30: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/30.jpg)
A More Relaxed VersionA More Relaxed VersionA More Relaxed VersionA More Relaxed VersionA More Relaxed Version
• generally, only expect π[fj ] ≈ π[fj ]
![Page 31: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/31.jpg)
A More Relaxed VersionA More Relaxed VersionA More Relaxed VersionA More Relaxed VersionA More Relaxed Version
• generally, only expect π[fj ] ≈ π[fj ]
• usually, can estimate upper bound on
|π[fj ] − π[fj ]|
![Page 32: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/32.jpg)
A More Relaxed VersionA More Relaxed VersionA More Relaxed VersionA More Relaxed VersionA More Relaxed Version
• generally, only expect π[fj ] ≈ π[fj ]
• usually, can estimate upper bound on
|π[fj ] − π[fj ]|
• so: compute π to maximize H(π) (= entropy)subject to
∀j : |π[fj ] − π[fj ]| ≤ βj
where βj = known upper bound
[Kazama & Tsujii]
![Page 33: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/33.jpg)
DualityDualityDualityDualityDuality
• can show solution must be Gibbs distribution:
π(x) = qλ(x) ∝ exp
∑
j
λj fj(x)
![Page 34: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/34.jpg)
DualityDualityDualityDualityDuality
• can show solution must be Gibbs distribution:
π(x) = qλ(x) ∝ exp
∑
j
λj fj(x)
• in unrelaxed case, solution is Gibbs distribution thatmaximizes likelihood, i.e., minimizes:
−1
m
∑
i
ln qλ(xi )
︸ ︷︷ ︸
negative log likelihood
![Page 35: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/35.jpg)
DualityDualityDualityDualityDuality
• can show solution must be Gibbs distribution:
π(x) = qλ(x) ∝ exp
∑
j
λj fj(x)
• in unrelaxed case, solution is Gibbs distribution thatmaximizes likelihood
• in relaxed case, solution is Gibbs distribution that minimizes:
−1
m
∑
i
ln qλ(xi )
︸ ︷︷ ︸
negative log likelihood
+∑
j
βj |λj |
︸ ︷︷ ︸
“regularization”
![Page 36: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/36.jpg)
Equivalent MotivationsEquivalent MotivationsEquivalent MotivationsEquivalent MotivationsEquivalent Motivations
• maxent with relaxed constraints
• log loss with regularization
• MAP estimate with Laplace prior on weights λ
![Page 37: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/37.jpg)
How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?
• want to bound distance between π and π(measure with relative entropy)
RE(π ‖ π) ≤
![Page 38: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/38.jpg)
How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?
• want to bound distance between π and π(measure with relative entropy)
• can never beat “best” Gibbs distribution π∗
RE(π ‖ π) ≤ RE(π ‖ π∗) +
![Page 39: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/39.jpg)
How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?How Good Is Maxent Estimate?
• want to bound distance between π and π(measure with relative entropy)
• can never beat “best” Gibbs distribution π∗
• additional term• → 0 as m → ∞• depend on
• number or complexity of features• “smoothness” of π∗
RE(π ‖ π) ≤ RE(π ‖ π∗) + additional term
![Page 40: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/40.jpg)
Bounds for Finite Feature ClassesBounds for Finite Feature ClassesBounds for Finite Feature ClassesBounds for Finite Feature ClassesBounds for Finite Feature Classes
• with high probability, for all λ∗
RE(π ‖ π) ≤ RE(π ‖ π∗) + O
(
‖λ∗‖
1
√
lnn
m
)
(for choice of βj based only on n and m)
• π∗ = qλ∗ = “best” Gibbs distribution• ‖λ∗‖
1measures “smoothness” of π∗
• very moderate in number of features
![Page 41: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/41.jpg)
Bounds for Infinite Binary Feature ClassesBounds for Infinite Binary Feature ClassesBounds for Infinite Binary Feature ClassesBounds for Infinite Binary Feature ClassesBounds for Infinite Binary Feature Classes
• assume binary features with VC-dimension d
• then with high probability, for all λ∗:
RE(π ‖ π) ≤ RE(π ‖ π∗) + O
(
‖λ∗‖1
√
d
m
)
(for choice of βj based only on d and m)
• e.g., infinitely many threshold features, but very lowVC-dimension
![Page 42: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/42.jpg)
Main TheoremMain TheoremMain TheoremMain TheoremMain Theorem
• both bounds follow from main theorem:
• assume ∀j : |π[fj ] − π[fj ]| ≤ βj
• thenRE(π ‖ π) ≤ RE(π ‖ π∗) + 2
∑
j
βj |λ∗
j |
• preceding results are simple corollaries using standard uniformconvergence results
• in practice, theorem tells us how to set βj parameters:use tightest bound available on |π[fj ] − π[fj ]|
![Page 43: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/43.jpg)
Finding an AlgorithmFinding an AlgorithmFinding an AlgorithmFinding an AlgorithmFinding an Algorithm
• want to minimize
L(λ) = −1
m
∑
i
ln qλ(xi) +∑
j
βj |λj |
• no analytical solution
• instead, iteratively compute λ1,λ2, . . . so that L(λt)converges to minimum
• most algorithms for maxent update all weights λj
simultaneously
• less practical when very large number of features
![Page 44: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/44.jpg)
Sequential-update AlgorithmSequential-update AlgorithmSequential-update AlgorithmSequential-update AlgorithmSequential-update Algorithm
• instead update just one weight at a time
• leads to sparser solution
• sometimes can search for best weight to update very efficiently
• analogous to boosting• weak learner acts as oracle for choosing function (weak
classifier) from large space
• can prove convergence to minimum of L
![Page 45: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/45.jpg)
Experiments and ApplicationsExperiments and ApplicationsExperiments and ApplicationsExperiments and ApplicationsExperiments and Applications
• broad comparison of algorithms• improvements by handling sample bias
• case study
• discovering new species
• clarification of taxonomic boundaries
![Page 46: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/46.jpg)
NCEAS ExperimentsNCEAS ExperimentsNCEAS ExperimentsNCEAS ExperimentsNCEAS Experiments[Elith, Graham, et al.]
• species distribution modeling “bake-off” comparing 16methods
• 226 plant and animal species from 6 world regions
• mostly 10’s to 100’s of presence records per species• min = 2, max = 5822, average = 241.1, median = 58.5
• design:• training data:
• incidental, non-systematic, presence-only• mainly from museums, herbaria, etc.
• test data:
• presence and absence data• collected in systematic surveys
![Page 47: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/47.jpg)
ResultsResultsResultsResultsResults
mean AUC (all species)
0.656
0.699
0.699
0.716
0.722
0.725
0.60 0.62 0.64 0.66 0.68 0.70 0.72 0.74 0.76
bioclim
garp
gen'd additivemodels
gen'd dissimilaritymodels
MAXENT
boosted regressiontrees
![Page 48: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/48.jpg)
ResultsResultsResultsResultsResults
mean AUC (all species)
0.656
0.699
0.699
0.716
0.722
0.725
0.60 0.62 0.64 0.66 0.68 0.70 0.72 0.74 0.76
bioclim
garp
gen'd additivemodels
gen'd dissimilaritymodels
MAXENT
boosted regressiontrees
• newer statistical/machine learning methods (includingmaxent) performed better than more established methods
• reasonable to use presence-only incidental data
![Page 49: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/49.jpg)
Maxent versus Boosted Regression TreesMaxent versus Boosted Regression TreesMaxent versus Boosted Regression TreesMaxent versus Boosted Regression TreesMaxent versus Boosted Regression Trees
• very similar, both mathematically and algorithmically, asmethods for combining simpler features
• differences:
• maxent is generative; boosting is discriminative• as implemented, boosting uses complex features;
maxent uses simple features
• open: which is more important?
![Page 50: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/50.jpg)
The Problem with CanadaThe Problem with CanadaThe Problem with CanadaThe Problem with CanadaThe Problem with Canada• results for Canada are by far the weakest:
mean AUC (all regions vs. Canada)
0.632
0.560
0.582
0.601
0.656
0.699
0.699
0.722
0.725
0.549
0.551
0.716
0.54 0.58 0.62 0.66 0.70 0.74
bioclim
garp
gen'd additivemodels
gen'd dissimilaritymodels
MAXENT
boosted regressiontrees
allCanada
• apparent problem: very bad sample bias• sampling much heavier in (warm) south than (cold) north
![Page 51: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/51.jpg)
Sample BiasSample BiasSample BiasSample BiasSample Bias
• can modify maxent to handle sample bias
• use sampling distribution (assume known) as “default”distribution (instead of uniform)
• then factor bias out from final model
• problem: where to get sampling distribution
![Page 52: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/52.jpg)
Sample BiasSample BiasSample BiasSample BiasSample Bias
• can modify maxent to handle sample bias
• use sampling distribution (assume known) as “default”distribution (instead of uniform)
• then factor bias out from final model
• problem: where to get sampling distribution
• typically, modeling many species at once
• so, for sampling distribution, use all locations where anyspecies observed
![Page 53: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/53.jpg)
Results of Sample DebiasingResults of Sample DebiasingResults of Sample DebiasingResults of Sample DebiasingResults of Sample Debiasing
mean AUC (Canada only)
0.632
0.551
0.549
0.560
0.582
0.601
0.723
0.50 0.55 0.60 0.65 0.70 0.75
bioclim
garp
gen'd additive models
gen'd dissimilarity models
MAXENT
boosted regression trees
MAXENT (w/ debiasing)
• huge improvements possible using debiasing
![Page 54: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/54.jpg)
Results of Sample DebiasingResults of Sample DebiasingResults of Sample DebiasingResults of Sample DebiasingResults of Sample Debiasing
mean AUC (all species)
0.656
0.699
0.699
0.716
0.722
0.725
0.752
0.60 0.62 0.64 0.66 0.68 0.70 0.72 0.74 0.76
bioclim
garp
gen'd additive models
gen'd dissimilarity models
MAXENT
boosted regression trees
MAXENT (w/ debiasing)
• huge improvements possible using debiasing
![Page 55: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/55.jpg)
Interpreting Maxent ModelsInterpreting Maxent ModelsInterpreting Maxent ModelsInterpreting Maxent ModelsInterpreting Maxent Models
• recall qλ(x) ∝ exp(∑
j
λj fj(x))
• for each environmental variable, can plot total contribution toexponent
value of environmental variable
addi
tive
cont
ribut
ion
to e
xpon
ent
threshold, β=1.0 threshold, β=0.01linear+quadratic, β=0.1
aspect elevation slope
precipitation avg. temperature no. wet days
![Page 56: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/56.jpg)
Case Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutus[with Phillips & Anderson]
![Page 57: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/57.jpg)
Case Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutus[with Phillips & Anderson]
![Page 58: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/58.jpg)
Case Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutusCase Study: Microryzomys minutus[with Phillips & Anderson]
![Page 59: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/59.jpg)
Case Study (cont.)Case Study (cont.)Case Study (cont.)Case Study (cont.)Case Study (cont.)
• accurately captures realized range
• did not predict other wet montane forest areas where could
live, but doesn’t• examined predictions of maxent on six of these (chosen
by biologist)• found all had characteristics well outside typical range for
actual presence records
• e.g., four sites had July precipitation ≥ 5 standarddeviations above mean
![Page 60: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/60.jpg)
Finding New SpeciesFinding New SpeciesFinding New SpeciesFinding New SpeciesFinding New Species[Raxworthy et al.]
• build models of severalgekkos and chameleons ofMadagascar(≈ 10-20 presence recordseach)
• identify isolated regionswhere predicted but notfound
![Page 61: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/61.jpg)
Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)
• combine all identifiedregions into single map
![Page 62: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/62.jpg)
Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)
• combine all identifiedregions into single map
• many regions already wellknown areas of localendemism
![Page 63: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/63.jpg)
Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)Finding New Species (cont.)
• combine all identifiedregions into single map
• many regions already wellknown areas of localendemism
• survey regions not previouslystudied
?
?
??
?
?? ?
?
![Page 64: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/64.jpg)
New SpeciesNew SpeciesNew SpeciesNew SpeciesNew Species
• result: discovery of many new species (possibly 15-30)
Thanks to Chris Raxworthy for all maps and photos!
![Page 65: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/65.jpg)
Clarifying Taxonomic BoundariesClarifying Taxonomic BoundariesClarifying Taxonomic BoundariesClarifying Taxonomic BoundariesClarifying Taxonomic Boundaries[Raxworthy et al.]
• “cryptic species”: classified as single species, but suspectedmixture of ≥ 2 species
• sometimes, maxent model is wildly wrong if trained on allrecords
• but model is much more reasonable if trained on eachsub-population separately
• gives strong evidence actually dealing with multiple species
• can then follow up with morphological or genetic study
![Page 66: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/66.jpg)
Phelsuma madagascariensis subspecies
Phelsuma mad. grandis
Phelsuma mad. kochi
Phelsuma mad. madagascariensis
One species Three species
?
?
This slide courtesy of Chris Raxworthy.
![Page 67: Maximum Entropy and Species Distribution Modeling](https://reader030.vdocuments.site/reader030/viewer/2022012705/61a69ef9996f7c003f1e6e28/html5/thumbnails/67.jpg)
SummarySummarySummarySummarySummary
• maxent provides clean and effective fit to habitat modelingproblem
• works with positive-only examples
• seems to perform well with limited number of examples
• theoretical guarantees indicate can be used even with a verylarge number of features
• other nice properties:• easy to interpret by human expert• can be extended to handle sample bias
• many biological applications