e the 
rtion. 
'ound 
g the 
ained 
were 
lasses 
rmed 
nixed 
| was 
y MD 
Lypes, 
ighest 
'sults, 
> high 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
species” was less than the first species, a mixed type was 
obtained (Gorji, Bahri, 2000). 
On the basis of the data, resulted from the fieldwork, a vector 
ground truth map possessing aforesaid forest types was prepared 
using a GIS approach. 
In order to assess the accuracy of the maps resulted from forest 
classification, the vector GT was then rasterized. In this Process, 
cell size of the raster GT was decided to be 5 meters, to keep 
the details. 
  
  
  
ÜBBBOSQU 
   
B 
  
  
  
  
  
  
462000 
  
  
Figure 1. The ground truth map for 9 forest types 
4.2 Digital image processing 
It is necessary to be aware of geomatric and radiometric 
situation of the image, before performing digital processing 
(Darvishsefat, 1994). Therefore, different geomatric and 
radiometric distortions such as striping, banding, sweep error, 
duplicate pixels and also atmospheric error of existing clouds 
were inspected, while any noticeable distortion was found. 
Orthorectification was then implemented utilizing 14 GCPs, 
ETM4 ; 
Y = 0.034937 + 0.999854 X 
mi . a 
2031 
1614 
coeff of det (2) = 0.99 
H 
HH | T4 sdof X (Ex) 318005901 
te Fi sdofY (Sy) =31-4027444 
1145 se. of estimate = 00575366 
se. of beta = 0.0000224 
922 
t stat forrorbeta = 30831 2101563 
$39 | t stat forbeta©>1 = -10.6668125 
vs 1 27 n = 3175 
ms 1 Ar of 3173 
30 T T T T T T T T TU EM 
  
  
30 254 438 702 926 1150 1374 1598 1822 2046 2270 
containing bare soil 
ETM3 (X axis) and ETM4 (Y axis) 
4.3 Image classification 
Based on the divergence between class signatures which is 
calculated from the class sample means and the class covariance 
matrices, the best band set were selected (Richards, 1999: 
Ann., 2001), using training areas. A supervised classification 
was then implemented utilizing maximum likelihoo (ML), 
Minimum Distanct to Mean (MD), Parallelepiped (PPD) and 
Spectral Angle Mapper (SAM) classifieres. In order to eliminate 
single pixels deviating from the neighborhood a mode filter 
(7x7 pixels) was done on the resulted maps. Ultimately, 
accuracy assessment was carried out through a pixel by pixel 
comparison of the classified outputs by the ground truth map, 
considering overall accuracy, kappa coefficient, user and 
producer accuracies. 
00 to digital elevation model and ephemeris data using toutin model 
; (Ann., 2001). In this regard while performing resampling, pixel 5. Results 
ES are size was distinguished to be 5 meters. In order to seperated I- The RMS error resulted from orthorectification was 
perus forest types more efficiently, various synthetic bands were 6.9m (0.23 pixed) along with the X axis and 6 m (0.2 
ndary created applying band ratioing (Terrill, 1994; Sandison, 1999), pixel) along with Y axis. Desired coincidence 
s and principal component analysis, tasseled cap transformation and between the roads and rivers layers of digital 
bands fusion (Darvishsefat, 2002). topographic maps and the rectified image indicated 
In addition, different vegetation indices such as PVI, SAVI, high precision of the orthorectification. 
MSAVII, MSAVI2, TSAVII, TSAVI2 and WDVI were 2- The best result of classifying 9 forest types was 
bands produced using soil line parameters, to reduce soil effect obtained by MD classifier with overall accuracy and 
digital (Terrill, 1994; Sandison, 1999; Jelenak, 2001). kappa coefficient equal to 18.45% and 11.05% 
Iso to The soil line relation was as below: respectively. 
3- Because of undesirable separability between pure and 
dominant types, understood from Bhattacharrya 
Y — 0.034937 + 0.999654 X : (1) distance and trasformed divergence critera and 
confusion matrices, these types merged and the 
classification was iterated with 5 types. The best 
sround Where X= ETM3 overall accuracy and kappa coefficient inferred from 
909 ha Y 5 ETMA MD classifier were 47.13% and 22.83% respectively. 
in the r=0,99 In this case, separability criteria, user and producer 
Large accuracies for amygdalus scoparia were better than 
me of those of the others. 
ntly 9 4- Finally by merging all of the types except for 
|, Acer Amygdalus scoparia, because of its better results, the 
ia), 4 classification was performed with 2 types consisting 
lanum, Am. Scoparia and the others as a mixed type. The best 
e were overall accuracy and kappa coefficient obtained by 
es was MD classifier were 92.16% and 67.58% respectivley. 
etween 
area of 
second Figure 2. Scatter diagram of digital numbers of pixels 
411