Modeling Range Distributions of Terrestrial Vertebrates from Species Occurrences and Landscape Variables

1
Modeling Range Distributions of Terrestrial
Vertebrates from Species Occurrences and
Landscape Variables
Geoffrey M. Henebry Brian C. Putz Milda R.
Vaitkus Amanda K. Holland and James W. Merchant
2
The Challenge of Habitat Modeling Range
Forecasting

• Recent national efforts to regionalize species
  models by mosaicking range distributions of
  adjacent states have revealed significant
  differences in predicted species distributions
  both within and across state borders1.
• A primary reason for this lack of concordance is
  variation in modeling methodologies.
• To generate seamless regional and national range
  distribution, unified and generalizable
  approaches to modeling are required.

1Brannon, R. 2000. An exploratory look at
combining vertebrate models from several states
An overview of vertebrate modeling in the western
states. GAP Analysis Program Bulletin 921-24.
3
Recursive Partitioning Algorithms Grow
Statistical Trees from Multivariate Data

• QUEST (Quick, Unbiased and Efficient Statistical
  Trees)2,3
• Similar to CART (Classification and Regression
  Trees) algorithm
• QUEST has several advantages for habitat
  modeling
• (1) much faster than CART
• (2) unbiased variable selection
• (3) handles missing values robustly
• (4) handles categorical predictor variables with
  many categories and
• (5) automated cross-validation.

2Shih, Y.-S. 1999. Families of splitting criteria
for classification trees. Statistics and
Computing 9309-315. 3Lim, T.-S., Loh, W.-Y., and
Shih, Y.-S. 2000. A comparison of prediction
accuracy, complexity, and training time of
thirty-three old and new classification
algorithms. Machine Learning Journal 40203-228.
4
Animal Modeling Using Statistical Trees
Species Occurrence Data
Soils Data
Terrain Data
Temperature Data
Rescale Data to 40 km2 Hexagonal Coverage
Variable Coverage
Precipitation Data
Land Cover Classification
Wildlife-Habitat Relationship Model
QUEST Tree Pruning
Model Inversion
Expert Review
Range Distribution Map
5
Nebraskas Approaches to Habitat Modeling
Occurrence Data Data Type Modeling Method
Birds BBS CBC 1970-1999 Presence/ Absence QUEST on P/A
Reptiles Amphibians NE Museum vouchers 1970-1999 Presence only QUEST on aggregates
Mammals NE Museum vouchers 1970-1999 Presence only Literature gestalt
6
(No Transcript)
7
Amphibian and Reptile Voucher Specimens in
Nebraska State Museum
Collected from 1970-1999 and georeferenced in
2000 46 species with gt10 specimens totaling
12,497 occurrences
8
Breeding Bird Survey Routes and Christmas Bird
Count Circles in Nebraska
9
Types of Occurrence Patterns

• Distribution
• Statewide
• Delimited
• Latitudinal
• Longitudinal
• Elevational
• Patchy
• Riparian
• Peripheral
• Regular
• Erratic

• Density
• Common
• Sporadic
• Rare
• Absent

X
10
Habitat Modeling Using Regional Contrasts
Statewide Delimited Peripheral
Common V. Hard V. Easy Easy
Sporadic V. Hard Easy-Moderate Easy-Hard
Rare V. Hard Easy-Moderate Easy-Hard
Hard cases make for ugly or complicated models
11
Occurrence data for Mourning Dove (Zenaida
macroura )
n 70,143
Common statewide occurrence
12
Occurrence data for Ring Necked Pheasant
(Phasianus colchicus)
n 32,628
Common statewide occurrence
13
Occurrence data for Baltimore Oriole (Icterus
galbula)
n 5,649
Sporadic statewide occurrence
14
Occurrence data for Woodhouses Toad (Bufo
woodhousii )
n579
Common statewide occurrence
15
Occurrence data for Gopher Snake (Pituophis
catenifer)
n109
Sporadic statewide occurrence
16
Occurrence data for Milk Snake (Lampropeltis
triangulum)
n49
Rare statewide occurrence
17
Occurrence data model for Willet
(Catoptrophorus semipalmatus)
n 156
Rare delimited occurrence
18
Occurrence data model for Short Horned
Lizard(Phrynosoma douglasii )
n21
escapee!
Rare delimited/peripheral occurrence
19
Occurrence data model for Sharp Tailed
Grouse(Tympanuchus phasianellus)
n557
Sporadic delimited occurrence
20
Occurrence data model for Tree Swallow
(Tachycineta bicolor)
n120
Rare peripheral occurrence
21
Occurrence data model for Northern Cricket Frog
(Acris crepitans)
n396
Sporadic delimited occurrence
22
Occurrence data model for Willow Flycatcher
(Empidonax traillii)
n72
Rare peripheral occurrence
23
Occurrence data model for Many-lined Skink
(Eumeces multivirgatus)
n55
Rare delimited occurrence
24
Occurrence data model for Prairie Skink
(Eumeces septentrionalis)
n67
Sporadic delimited occurrence
25
Occurrence data model for Great Plains Skink
(Eumeces obsoletus)
n28
Rare peripheral occurrence
26
Concluding Thoughts

• Transparency and repeatability is better than
  modeling via literature gestalt.
• Inverting the habitat model to forecast range
  distribution predicts only the occurrence of
  modeled habitat, not species presence/absence or
  abundance. Implications for model accuracy
  assessment (potential vs. realized vs. realizable
  habitat), especially the interpretation of
  commission error.
• Habitat modeling from statistical trees lend a
  greater degree of objectivity to data analysis
  but there is still substantial subjectivity/judgme
  nt in the pruning (generalization) phase of the
  modeling process.
• Ability to incorporate museum voucher specimens
  and curated surveys into the habitat modeling
  strengthens the base for biodiversity planning.