inbul 2004 
ation is to 
each land 
on-screen 
vith strong 
composite 
composite 
nages was 
| pixels so 
ng class as 
d be made 
re training 
wever, this 
on and the 
| land class 
CS. 
next step is 
tion. These 
‚ide, 2004). 
ich of each 
ite a signal 
's contain à 
lasses they 
PF file that 
IG file. It is 
displaying 
mes of the 
ization was 
and the full 
rultispectral 
1ide,2003). 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
2.3 Classification of the Image 
The classification of the image is the third and the final step. 
This can be done with any of the hard or soft classifiers 
described below. 
The Parellelpiped procedure (PIPED) is used for special 
pedalogic reasons only. Generally this procedure is not used for 
landuse mapping . When training sites are known to be strong, 
the MAXLIKE procedure is used (Richards 1995). However, if 
there are concerns about the quality of the training sites, the 
MINDIST procedure with standardized distances should be 
used (Richards 1995). The MINDIST module with the 
standardized distances option is a very strong classifier and one 
that is less susceptible training site problems than MAXLIKE. 
The FISHER Classifier can perform exceptionally well when 
there are not substantial areas of unknown classes and when the 
raining sites are strongly representative of their informational 
classes (IDRISI Klimanjaro Guide 2004). 
2.4 Genereal Properties of Classifiers 
In this study, supervised classification classifiers have been 
used to classify the image of the study area for land cover 
classification. The — parellelpiped, maximum likelihood, 
minimum distance and fisher (lineer discrimination) classifiers 
are used for this purposes. 
The parellelpiped classifier is a very simple supervised 
classifier that is, in principle, trained by inspecting histograms 
of the individual spectral components of the available training 
data (Richards, 1995). 
Whilst the parellepiped method is, in principle, a particularly 
simple classifier to train and use, it has several drawbacks. One 
is that there can be considerable gaps between the 
parellelpipeds, and the pixels in those regions will not be 
classfied. By comparision the minumum distance and 
maximum likelihood classifiers will label all pixels in an image, 
unless thresholding methods are used. Another limitation is that 
prior probabilities of class membership are not taken into 
account of;nor are they for minimum distance classification. 
Finally, for the correlated data there can be overlap of the 
parellelpipeds since their sides are parallel to the spectral axes 
(Richards 1995), 
The Minimum distance classifier is based on training site data. 
This classifier characterizes each class by its mean position on 
&àch band (IDRISI Klimanjaro Guide 2004). 
  
Minimum distance classifier is highly recommended in all 
1995) The 
classification is performed by placing a pixel in the class of the 
image classification applications (Richards 
nearest mean. The minimum distance algorithm is also more 
attractive since it is a faster technique than the maximum 
likelihood classification. 
The maximum likelihood classification is the most common 
supervised classification method used with remote sensing 
image data (Richards 1995). This classifier is based on 
Bayesian probability theory (IDRISI Klimanjaro Guide 2004). 
The Fisher classifier conducts a linear discriminant analysis of 
the training site data to form a set of linear functions that 
express the degree of support for each class. It is more difficult 
to describe graphically (IDRISI Klimanjaro Guide 2004). 
3. IMAGE CLASSIFICATION AND RESULTS 
Landsat 7 ETM+ images of the Ayvalik, were classified to 
obtain the landuse map of the area using above mentioned four 
classifiers. Of these hard supervised classifiers used in this 
study, the maximum likelihood and Fisher are clearly the most 
powerful as they make more reliable classification. But these 
realiabilities can change according to purpose of the study. In 
order to make an image classification for landuse mapping, 
selection of the most proper image is the first step. For this, 
Landsat 7 ETM + images processed with IDRISI Klimanjaro 
GIS and image processing package. Firstly, all visible and 
infared bands were corrected atmospherically and 
geometrically. These images can be used for different 
interpretations such as geomorphological, geologcial, landuse 
and land cover mapping. We have seen that using only normal 
composite and false color composite images to interprate may 
be missleading in view of discrimination of objects on the 
image. To eliminate this discrepancy, visible and infrared bands 
have been processed by principal component analysis. After 
then, the composite images were made by different PCA bands. 
20 PCA composite images were formed to chose the most 
approppriate images to classify them for landuse mapping. The 
composite image composed with PC2, PC4 and PC5 was used 
to map landuse features. In this image, land properties such as 
agricultural sites, vegetation cover, settlement areas, bare lands, 
wetlands and others was more clear than the same compositon 
of false color composite image. After digitizing of the training 
sites, the signature file from defined training sites was 
1093