banner literature about projects tools meetings search links BBS GAP home GAP home
home

<-Previous | First | Next ->

 

Predictive Modeling pf Rare Plant Species

A Report on GAP Research Project #1434-HQ-97-RU-01542

 

WALTER FERTIG

Department of Botany, University of Wyoming, Laramie

Resource managers and conservation biologists are faced with a
critical shortage of information on the potential distribution of rare
plant species over large areas. Predictive modeling can be a cost-
effective means of identifying and prioritizing likely areas of rare
plant habitat for more efficient and productive ground surveys.

Traditionally, rare plant ranges have been inferred from dot (RMH
bution maps (Figure 1), but these maps may represent only a sites
tion of the species’ actual range or reflect sampling bias. An present
native is to model the potential range of a species by identifying
correlations between the plant’s known distribution and relevant
environmental variables using geostatistical methods (Franklin
1995). These empirical models can be derived from herbarium or
Natural Heritage Program location records and state- or regional-
scale coverages of substrate, topography, and climate in a Geo-
graphic Information System (GIS).


Figure 1. Dot distribution map of the known range of Physaria eburniflora.
To illustrate this technique, I have developed an empirical model of
the potential distribution of a rare Wyoming endemic, Physaria
eburniflora Rollins. This low-growing perennial mustard is re-
stricted to barren calcareous ridges and granite outcrops in the North
Platte and Sweetwater River drainages in central Wyoming (Figure
1). Of 29 known populations, 28 were selected for modeling be-
cause these sites could be placed within 0.1 - 2.5 km of their actual
location. As absence data have not been routinely collected for this
species, I used the Rocky Mountain Herbarium’s (RMH) database
of over 9,000 sampling points in Wyoming to identify areas where
P. eburniflora has not been collected and is thus presumed absent.
researchers have systematically established these collection
across the state for the purpose of collecting all plant taxa
in the immediate area.) Of these putative absence sites,
1,270 were randomly selected for modeling. The combined pres-
ence and absence data set was randomly split into model-building
and model-validation pools.

Environmental attributes for each presence and absence point were
derived from digital coverages in ArcView version 3.1. For climate
data, I used PRISM mean monthly precipitation data in 4 km raster
format (Daly et al. 1994) and unpublished PRISM mean monthly
temperature data in 2 km raster format. Topographic data, includ-
ing elevation, slope, and aspect, were derived from 30 m Digital
Elevation Model (DEM) coverages of the state. An index of land-
scape position for each 30 m pixel was calculated using the proto-
col of Fels and Matson (1996) and then reclassified into four ter-
rain position categories based on overall slope and shape (concave,
slope, flat, and convex) (Ken Driese, pers. comm.). Lastly, vector
coverages of bedrock geology and GAP land cover were used.

Using classification tree analysis in S-plus version 1.1, I developed
a model of P. eburniflora distribution using seven environmental
variables as predictors (mean April and July precipitation, mean
January and July temperature, bedrock geology, aspect, and eleva-
tion). Classification trees use a recursive partitioning algorithm to
identify the values of continuous and categorical environmental
variables that best explain the differences in predicted presence or
absence of a species (Breiman et al. 1984). From the model output,
I created a potential range map in ArcView by intersecting the envi-
ronmental values that best predicted the presence of P. eburniflora
(Figure 2). The model correctly classified 13 of the known pres-
ence points (93%) and 608 (95.7%) of the known absence points in
the model-building data set and seven (50%) of the known pres-
ence points and 614 (96.7%) of the known absence points in the
validation data set.

 

Figure 2. Predicted distribution of Physaria eburniflora in Wyoming
based on correlational modeling.

According to the predicted distribution map, potential habitat for P.
eburniflora may occur along the flanks of the Bighorn, Wind River,
Owl Creek, Medicine Bow, and Uinta mountains in northern, west-
ern, and southeastern Wyoming, and on isolated buttes and valleys
elsewhere in the state (Figure 2). Surveys to date have shown that
other, closely related Physaria species occur in these areas, per-
haps reflecting their superior competitive ability or localized ex-
tinction or incomplete dispersal of P. eburniflora .

The predictive ability of correlational models may be hampered by
errors inherent in the input data sets. Imprecise location points,
errors in converting map data to digital format, and horizontal and
vertical errors in DEMs may all reduce prediction accuracy (Franklin
1995). Potentially useful environmental factors such as local soil

pH, soil texture, or extremes in precipitation or temperature are
unavailable in statewide coverages or are masked when
macroclimate data are averaged over diurnal cycles and monthly
periods. Equally useful spatial data sets for the distribution of pol-
linators, seed dispersal vectors, predators, and soil symbionts are
also unavailable. Spatial autocorrelation can inflate the explana-
tory power of models when location points for a species are natu-
rally clustered, although this problem may be lessened if this spa-
tial patterning is related entirely to spatial patterning in the explana-
tory environmental variables (Franklin 1998). Lastly, an inadequate
number of sample points may be available for some extremely rare
plants to meet the minimum data input requirements for a statisti-
cally useful model. Despite these caveats, GIS-based correlational
models can be a powerful tool for developing testable and ecologi-
cally meaningful distribution maps of rare species and for identify-
ing areas of potential habitat for field surveys.

 

Literature Cited

Breiman, L., J. Friedman, R. Olshen, and C. Stone. 1984. Classi-
fication and regression trees. Chapman and Hall, New York.

Daly, C., R.P. Neilson, and D.L. Phillips. 1994. A statistical-topo-
graphic model for mapping climatological precipitation over
mountainous terrain. Journal of Applied Meteorology 33:140-
158.

Fels, J.E., and K.C. Matson. 1996. A cognitively-based approach
for hydrogeomorphic land classification using digital terrain
models. Third International Conference/Workshop on Integrat-
ing GIS and Environmental Modeling, Santa Fe, New Mexico.

Franklin, J. 1995. Predictive vegetation mapping: Geographic
modelling of biospatial patterns in relation to environmental gra-
dients. Progress in Physical Geography 19:474-499.

Franklin, J. 1998. Predicting the distribution of shrub species in
southern California from climate and terrain-derived variables.
Journal of Vegetation Science 9:733-748.

<-Previous | First | Next ->Previous | First | Next ->