GREE 
Axis N GVI IPVI ND TVI 
NESS 
I 0.7340 | -0.9673 | -0.2756 | 0.8746 | 0.8622 
2 0.1738 | 0.1960 | -0.9519 | 0.1359 | 0.2095 
3 0.5125 | 0.1605 | 0.1340 | 0.4491 | 0.4354 
4 |-0.4101 | 0.0071 | 0.0015 | 0.1178 | 0.0580 
5. |-0.0131 | 0.0000 | 0.0016 | -0.0321 | 0.1404 
Table 2. Correlation Between Input Rasters and Principal 
Components 
The relationship between vegetation cover and the indices 
appears to change over the area according to the certain 
conditions such as soil cover type. To minimize the effect 
of soil on vegetation reflectance, second set of indices 
were used. These indices were Soil Adjusted Vegetation 
Index (SAVI), Modified Soil Adjusted Vegetation Index 1 
(MSAVII), and Modified Soil Adjusted Vegetation Index 
2 (MSAVI2) (Table 1). For this study the first PC acquired 
from the 3 soil indices contains the spectral information 
adequate for the classification to normalize the effects that 
emerge due to the different soil types of the areas with low 
canopy of vegetation. 
(Table 3) 
  
  
  
  
  
Axis| MSAVI | MSAVD | SAVI ses 
0 
L| 0,9999 | 0.0789 | 0.0046 | 84.4805 
2 | 0.0145 | 0.9969 |-0.0629| 15.5186 
3 | -0.0000 | 0.0000 | 0.9980 | 0.0009 
  
  
  
  
  
  
  
Table 3. Correlation Between Soil Adjusted Vegetation 
Indices and Principal Components 
Besides these feature extraction oriented indices, PCA 
were performed on raw bands in order to find if vegetation 
related information could be collected in few explanatory 
bands. In this transformation, examination of principal 
components eigenvector loadings determine which PC 
possesses information related directly to the spectral 
signatures of vegetation. Eigenvector loadings for PC2 in 
Table 4 indicate that PC2 describes the difference between 
the visible channels (TM1, 2, and 3) and the infrared (IR) 
channels (TM5 and 7) and also this component is 
commonly thought to be related to vegetation. Eigenvector 
loadings for PC3 (in Table 3) indicate that PC3 is 
dominated by vegetation. In this component both the 
loading values of TM3 and TM4 is negative but the 
difference between these two band were high because in 
TM3 chlorophyll is absorbed, on the contrary chlorophyll 
is highly reflected in the near infrared band. Therefore 
PC2 and PC3 were selected as feature components. 
206 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
  
  
  
  
  
  
Eigen 
Axis | TMI TM2 TM3 TM4 TM7 TMS | values 
(%) 
: : 4 85.731 
] 0.9160 | 0.9667 | 0.9781 0.6093 | 0.9615 | 0.9329 0 
2 -0.3436 | -0.2114 | -0.1570 | 0.6718 | 0.1201 | 0.3008 | 9.2461 
3 ]|-0.1460 | -0.1245 | -0.0459 | -0.4201 | 0.2347 | 0.1756 | 3.9920 
4 -0.1408 | 0.0019 | 0.1247 | -0.0047 | -0.0164 | -0.0336 | 0.6456 
5 -0.0100 | 0.0107 | -0.0155 | 0.0286 | 0.0756 | -0.0846 | 0.2556 
6 0.0394 | -0.0720 | 0.0284 | 0.0104 | 0.0070 | -0.0099 | 0.1296 
  
  
  
  
  
  
  
  
  
Table 4. Correlation Between Input Rasters and Principal 
Components 
In addition to principal component bands, Decorrelation 
Stretched (DS) bands were used in this study. Even though 
these bands still show the properties of the original bands, 
the color separation of these bands are enhanced with 
significant band to band correlation. Decreasing the 
correlation of spectral data corresponds to exaggerating 
the color saturation without changing the distribution of 
hues (or relative color composition) (Gillespe et al., 1987). 
At the end of these analyses it is assumed that; selecting 
PCI and PC2 of vegetation indices, PCI of soil indices, 
PC2 and PC3 of raw bands and DC3 and DC4 as feature 
components will remove the redundant data among 
multivariate datasets, such as multispectral remote sensing 
images and increase the accuracy of the classification. 
  
Classification with Raw Bands 
  
  
  
  
  
  
  
  
  
  
  
  
  
      
  
  
  
  
EN Black Pine 
EEE] Calabrian Pine 
(7) Taurus Cedar 
2] Taurus Fir | 
CZ) Sparse Vegetation | 
LL Bare Area i 
L i 
ES Fannland 
  
  
  
  
Figure 2. Classification results of both raw bands and 
feature components. 
By using these feature components overall accuracy was 
increased to 76.92 %. This rise shows that the new formed 
bands were very successful in the discrimination of 
vegetation classes with very similar spectral reflectance 
values. 
Internatic 
IR CHE 
Ground 7 
   
The. ı 
Com 
extra 
by cr 
corre] 
meth: 
with 
consi: 
| classi 
| of cla 
This 
imple 
4. Ac 
We v 
neces 
Fores 
5, Re 
Brool 
Bryar 
Drake 
Fahsi 
Gilles