a Av 
So 
ne 
ne 
by 
ed 
of 
ng 
er 
st 
of 
on 
al 
00 
AS 
Y: 
ne 
TM 
er 
ire 
al 
st 
and 
ah, 
vas 
the 
at 
ive 
the 
the 
rk 
le, 
/er 
the 
on 
SAT 
Ise 
ASS 
der 
nd- 
Nas 
and 
ion 
3)] 
  
z 
  
  
  
  
STUDY AREA 
BOLU PROVINCE 
  
YIGILCA 
lgr\d 
Hasan am MENGE 
S cale : 1.000. 000 
ISTANBUL © 777777 GEREDE 
S. oO 
DUZCE PSS SINAS LI BON 
7 
Olkol] rezer voir 
A ANKARA 
bat lake 
Le OMUDURNU 
SOYNUR KIBRISCIK 
SEBEN 
STUDY AREA 
  
BLACK SEA 
    
   
   
    
   
  
  
Ankara 
TURKEY 
Adana 
  
  
MEDITERRANEAN 
  
  
  
Figure 1. Map of Bolu province showing the location of the study area 
provided the most useful 
for the study area. 
dispose the 
temporal discrimination 
By doing so it was aimed to 
vegetation cover in overall image in 
order to enhance the ability of visual 
interpretation. 
Selection of multiple band combination was based 
on the decision to test TM data which are 
representative of the three major spectral 
regions, namely Near Infrared, Band 4; Red, Band 
3; Green, Band 2 (Karteris, 1990). 
Supervised and unsupervised classification 
techniques were performed for spectral pattern 
recognition of LANDSAT TM multispectral data by 
using ERDAS software package. 
classification is based on the 
use of various training data sets. 
The size and number of these data for each 
category are dependent mainly upon the spectral 
variability within that category throughout the 
study area and should be unique in terms of 
effectively defining the category. In this study 
28 training sets were selected by displaying a 
composite of bands 4, 3 and 2. All these sets were 
located well within the boundar ies of the 
corresponding categories. The training data were 
processed statistically and spectral signature 
files containing the means, the standard 
deviations and the variance-covariance matrices 
were generated for each category and used as input 
to the classifier for the classification of the 
whole study area (1.047.552 pixel in total). 
Classification stage was performed by using two 
different approaches; Gaussian maximum likelihood 
classifier and minimum-distance to means 
classifier. Output stage was presented in the form 
of tables of area statistics. 
Supervised 
selection and 
291 
An alternative approach 
clustering was also 
area. 
called unsupervised 
carried out for the study 
An accuracy assessment was 
some degree of confidence to the classification 
results. Overall accuracy of each individual 
category were calculated for each set of analyzed 
data. The procedure was accomplished by overlaying 
and registering on the projected classification 
images the land-cover map and then collecting the 
required data. Data collection was done by 
stratifying the area into the classification 
categories and conducting a random sampling of 
points within each of the stratified  land-cover 
categories. 
per formed to provide 
Evaluation of spectral separability was provided 
by a confusion matrix for the purpose of large 
area accuracy analysis of study area data set that 
are different from, and considerably more 
extensive than, the training area (Lillesand and 
Kiefer, 1979). From this information, 
classification error of omission and commission 
was studied. Accuracy estimates based only upon 
diagonal elements of these tables may produce 
inflated accuracies. Therefore, the Kappa 
statistic was used as a coefficient of agreement 
since it corrects for chance agreements and 
accounts for errors for both 
commission (Hudson, 1987). 
omission and 
4. RESULTS AND DISCUSSION 
4.1. Land-Cover Pattern 
According to the information subtracted from land- 
cover map of the study area, highlighting the