INTEGRATION AND USAGE OF INDICES, FEATURE COMPONENTS AND TOPOGRAPHY 
IN VEGETATION CLASSIFICATION FOR REGIONAL BIODIVERSITY ASSESSMENT 
. 1 +2 . 3 ne CH d 
Aysegül Domaç , Ugur Zeydanli”, Ertan Yesilnacar”, M. Lütfi Süzen* 
! Conservation GIS Laboratory, Department of Biology, Middle East Technical University, Ankara, TURKEY 
aysegul@kafkaskoruma.net 
? Conservation GIS Laboratory, Department of Biology, Middle East Technical University, Ankara, TURKEY 
* Department of Geomatics, University of Melbourne, 3010, Melbourne, Victoria, AUSTRALIA 
ertan@sunrise.sli.unimelb.edu.au 
* RS —GIS Lab. Department of Geology, Middle East Technical University, Ankara, TURKEY suzen@metu.edu.tr 
  
KEY WORDS: Remote Sensing, Vegetation, Classification, Accuracy, Ecology, Spectral 
ABSTRACT: 
The classification of vegetation has been an important research subject in botany, ecology, geography, and other disciplines 
to map the differences in vegetation types. Classifying vegetation by remote sensing is valuable because it can determine 
vegetation distribution and occurrence for very large areas in a short time. Advances in technology have led to developments 
in methods of vegetation classification, leading to the creation of new and more sophisticated components and powerful 
techniques. Classifying original bands and/or image components may cause unsatisfactory results in spectrally chaotic fields. 
In such cases, the demand for accurate land-use, land-cover, vegetation, and forestry information may require more 
explanatory components those components should represent specific information for target land-covers and not contain 
redundant knowledge. 
In this study, spectral bands of Landsat Thematic Mapper and topographic data were used as an input. Different image 
components and indices were produced and then used in the Maximum Likelihood Classification method. In order to find out 
proper inputs for our case, newly produced components and indices were statistically compared and the bands that include 
the information about vegetation are selected. Overall accuracy parameter that is obtained from the Error Matrix helped to 
evaluate the results of the classification. Results obtained in this study suggest that using these spectrally improved bands and 
indices; the accuracy of the classification could be increased up to 10-15 percent. 
1. Introduction 
With the aid of classical vegetation indices or raw input 
The need to map wide areas with limited resources forced bands hardly any classification can fulfill this need; hence 
the improvement of vegetation classification methods by some improvements should have to be made either by post 
using satellite images. Conservation agencies use these classification sorting or by adding new components 
images to extract variety of vegetation types in order to derived from the original input bands to the classification 
assess the biodiversity of a region. Among the possible process. The extraction of spectral information related to 
commercial satellite systems, Landsat images have got this type of target from Landsat TM imagery has been 
some serious advantages over other systems such that: 30 achieved through the use of image processing techniques 
m ground resolution yields as a convenient resolution for such as band rationing and principal component analysis 
regional vegetation studies with a minimum mapping unit (Sabine1999). The major fact behind this new component 
of 100 ha., the spectral coverage fits well to the vegetation adding is to create a spectral subset of the data itself and to 
spectra, and the wide swath width yields in less number of create more explanatory variables which can be used to 
images to process which maintains the coherence of the exploit the variance of the vegetation types that are desired 
imagery. Furthermore Landsat system is a mature system to be mapped from imagery. Due to the high similarity 
dating back to early 1970's, hence plenty of researchers among individual bands of a multispectral image, 
have exploited many mapping methods. However still the statistical data compression tools like principal component 
classification results are way off the desired levels, only analysis (PCA) are often applied in image analysis and 
major homogenized groups of forest can be discriminated, image classification to reduce the amount of redundant 
yet the conservation measures require a more detailed information (Ricotta et a/., 1999). The objective of the 
legend. Subdivision of this multi-spectral continuum into study was to improve the accuracy of vegetation 
meaningful vegetation classes is a major challenge that classification by using future components which were 
requires careful consideration (Brook and Kenkel, 2002). constituted by using raw bands and various vegetation 
For instance, visual analysis of different bands/colour indices. 
composites from a multispectral dataset with constant 
pixel resolution still reflects the same spatial structure, 3. Studv Area 
even if the contrast between different scene elements (i.e. : 
forest patches versus  non-forest patches) might Study area is located in the Southern part of Turkey in the 
considerably vary for the different band combinations Mediterranean region and covers approximately 235km’ 
(Bryan, 1988). (Figure 1). Elevation values of the region vary between 
204 
  
  
Internatic 
Ner ee 
300 
chara 
forest 
for th 
4 070 ( 
4 065 
4 000 Ut 
  
3. Me 
Using 
a wid 
statist 
spatia 
somet 
the or 
Lands 
classi 
classi 
and tl 
accou 
classe 
The « 
availa 
reconi 
trainir 
differ 
Tauru 
Were « 
To ch 
with 
000 s 
the tr 
overal 
satisfz 
   
Ground Tr 
  
  
Hane 
  
  
  
At thi 
increa 
indice