International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
constituted. And then, this signature file were used for four 
classifier. 
CORINE method was used (CORINE, 1995) to make landuse 
map. In application of this method, 7 landuse classes were 
determined. Water bodies (sea and lake), wetlands, forest, 
urban areas, agricultural areas, bare lands were selected firstly. 
As the olive trees covers the study area spreadingly, a seventh 
class was added to the above mentioned six classes. 
The Parellelpiped, Minimum distances, Maximum likelihood 
and Linear discriminant classifiers were applied respectively. 
Results obtained from the classified images were compared and 
each of these images were controlled by field verification. 
The parellelpiped classification results (Figure 6) was simple 
and have not reflected the real features on the land. For 
example, the urban sites on the map could not be identified, 
bare lands and olive trees could not be distinguished from each 
other. At some locations the vegetation cover has been seen as 
black color. Because of these anomalies, this classifier was not 
found proper enough for landuse mapping purpose. 
The map derived using the minimum distances classifier 
(Figure 7) seemed more reliable than the map produced by 
parellelpiped method. In this map, settlement sites were 
selectable, borders of vegetation cover, agricultural areas and 
olive trees were more clear than the parellelpiped classifier 
map. The Maximum likelihood classification result was much 
better than the previous two maps. In the maximum likelihood 
map, barelands-olive trees discrimination could be seen clearly, 
boundries of agricultural areas and forest were more apparent 
than both minimum distances and parellelpiped maps (Figure 
8). In this map, some wetlands areas were indicated by orange 
color scale. This was a cause of sedimantation on that locations. 
The map that has been produced by application of the linear 
discriminant classifier (Figure 9) was more suitable than the 
minimum distances map and was less proper than maximum 
likelihood map. In this map, urban sites were more clear than 
both maximum likelihood, minimum distances and 
parellelpiped maps. Olive trees and barelands borders were also 
identifiable. However, forest and olive trees boundries were not 
clear according to ground truth studies. 
Inter) 
—— 
  
Y= 206711+ 0660863X r= 0632118 
Regression Parameters: 
X axis: maxlike 
Y axis: fisher 
Coeff. of Det. 47.90 1 
Std. Dev. of X 2.087050 
Std. Dev. of Y 1.992803 
¢.E. of Estimate 1.43034 
Std. Error of Beta 0.000645 
t Stat for r or Beta 1024.343038 
-525.664731 
1141160 
1141158 
t Stat for Beta <> I 
Sample Size ín) 
"^a ww w un " ww. 
Apparent df 
  
  
æ ET RE 
low frequency high frequency 
  
Figure 2. Regression of Maximum Likelihood and Linear 
  
Regression Parameters: 
Figur 
discriminant classifier. tar 
distar 
Classification results have compared with one another and pu 
regression. analysis were made. The best corelation was M 
obtained between the Maximum likelihood map and the Ir 
minimum distance map (= 0.79). Other regression results are a 
Psp 
r= 0.69, r= 0.52 and r= 0.76 for maximum likelihood-linear L 
discriminant classifier, minimum distances-linear discriminant s 
classifiers and maximum  likelihood-parellepiped classifiers n 
respectively. An interesting point related with these results is v 
that the corelation between maximum likelihood and Ls 
parellelpiped maps results have high corelation coefficient than 5 = 
maximum likelihood and linear discriminant classifiers 
corelation. This might be thingking usefullnes of parellelpiped Figure 
map. But field studies have showed that parallelpiped map classii 
results do not reflect real properties on the land surface. It is 
thought that it may be due to the classification algorithm 
differences. For m 
Ye 12518504 0760582 Dr 0759755 manag 
       
ia X axis: maxlike satelli 
Y axis: mindist 
57.72 5 mappi 
2.087056 
2.11123? proper 
1.372786 
Coeff. of Det. = 
Std. Dev. of X = 
Std. Dev. of Y = 
S.E. of Estimate = 
Std. Error of Beta f ; X monit« 
t Stat for r or Beta = 1248.228723 
t Stat for Beta <> 1 = -375, 836643 
Saxple Size (n) = 1141160 techni 
Apparent df = 1141158 
satelli 
eo 
digital 
^» 
  
proper 
= {437 PE RNR Very ii 
low frequency high frequency 
and di; 
; ; ; jid nd com 
Figure 3. Regression of Maximum likelihood and Minimu po 
. s these 
distances classifiers. I 
maxim 
1094