- Mapping Mountain vegetation - 499
Applied Vegetation Science 11: 499-508, 2008
doi: 10.3170/2008-7-18560, published online 4 June 2008
© IAVS; Opulus Press Uppsala.
Mapping mountain vegetation using species distribution modeling,
image-based texture analysis, and object-based classifcation
Dobrowski, Solomon Z.
1*
; Safford, Hugh D.
2
; Cheng, Yen Ben
2
& Ustin, Susan L.
3
1
Department of Forest Management, College of Forestry and Conservation, University of Montana, Missoula, MT 59812, USA;
2
Center for Spatial Technologies and Remote Sensing, One Shields Ave, The Barn, University of California, Davis,
CA 95616, USA;
3
USDA-Forest Service Pacifc Southwest Region, 1323 Club Drive, Vallejo, CA 94592, USA;
*
Corresponding author; E-mail solomon.dobrowski@cfc.umt.edu
Abstract
Objective:The objective of this study was to map vegetation
composition across a 24 000 ha watershed.
Location: The study was conducted on the western slope of
the Sierra Nevada mountain range of California, USA.
Methods: Automated image segmentation was used to deline-
ate image objects representing vegetation patches of similar
physiognomy and structure. Image objects were classifed us-
ing a decision tree and data sources extracted from individual
species distribution models, Landsat spectral data, and life
form cover estimates derived from aerial image-based texture
variables.
Results: A total of 12 plant communities were mapped with an
overall accuracy of 75% and a κ-value of 0.69. Species distribu-
tion model inputs improved map accuracy by approximately
15% over maps derived solely from image data. Automated
mapping of existing vegetation distributions, based solely on
predictive distribution model results, proved to be more ac-
curate than mapping based on Landsat data, and equivalent in
accuracy to mapping based on all image data sources.
Conclusions: Results highlight the importance of terrain,
edaphic, and bioclimatic variables when mapping vegetation
communities in complex terrain. Mapping errors stemmed
from the lack of spectral discernability between vegetation
classes, and the inability to account for the confounding effects
of land use history and disturbance within a static distribution
modeling framework.
Keywords: Decision tree; GAM; Image segmentation; Sierra
Nevada; Topographic convergence index; Vegetation map-
ping.
Nomenclature: Hickman (ed.) 1993.
Abbreviations: DOQQ = Digital ortho-photo quarter quads;
DT = Decision tree; GAM = General additive modeling;
SDM = Species Distribution Modeling; TCI = Topographic
convergence index.
Introduction
Image classifcation and predictive distribution
modeling (referred to as species distribution modeling,
gradient modeling, niche modeling, among others) are
common approaches to mapping vegetation. Both ap-
proaches have limitations, for instance, mapping based
on the classifcation of spectral data is often hindered
by the lack of spectral discernability between vegetation
types (Dirnböck et al. 2003; Treitz et al. 1992), whereas
predictive distribution modeling characterizes potential
rather than actual vegetation distributions (Austin 2002;
Guisan & Zimmerman 2000). When used in combina-
tion, both image analysis and predictive modeling can be
complementary. Terrain variables have long been used in
image classifcation to improve map accuracies (J. Fran-
klin 1995). Similarly, image variables have been used in
predictive distribution modeling (Dirnböck et al. 2003;
Lees & Ritman 1991) and are more recently referred to as
functional gradients (Muller 1998). A challenge to auto-
mated mapping of vegetation, is devising a methodology
that leverages the strengths of both predictive modeling
and image-based approaches, as well as divides fnite
resources between tasks associated with each.
Approaches to incorporating environmental variables
into existing vegetation mapping efforts include but are
not exclusive to: (1) the direct use of environment and
image variables in a nominal classifer, (2) constrained
ordination techniques using inventory data, image and
environmental variables (e.g. Ohmann & Gregory
2002; Dirnböck et al. 2003), and (3) Probabilistic or
consensus theoretic approaches to combining image
analysis and gradient modeling results (e.g. Strahler
1980; Dobrowski et al. 2006). The use of environmental
variables in these and others studies has been shown
to improve map accuracies given that these types of
data are related to biophysical factors that affect the
distribution of species (J. Franklin 1995; Richards et
al. 1982; Strahler 1980).
Mapping vegetation at high spatial resolution presents