data, the map including the forest boundaries of 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
las on points, and by using a first-order polynomial 
eet, are Nour Forest Park from 1993 is updated to the transformation led to an RMSE at an acceptable 
using a situation of the forest in the year of satellite data rate (0.64 pixel). Overlay of vector data of 
During acquisition (1988) by interviewing local forest streets on the rectified image indicates that the 
els are workers. The so updated map is digitized and geocoding performed well within the required 
nearest then used as ground truth. This ground truth is limits. 
prepared in two stages as forest and non-forest 
layer, and then converted to raster format. Classification: The results of the classification 
Results of different classifications are in different stages and using sets of suitable 
compared on a pixel by pixel basis with respect bands are compared with the ground truth of 
formed to this ground truth and the errors matrices are the entire area. It indicates that the ML classifier 
is ‘and obtained including the overall, producer and leads to the best results in almost all cases. The 
y Koch user accuracy. . highest accuracy results from the classification 
with TM band 3 and 4 resulting in a final 
accuracy of 95.14%. 
5. RESULTS Figure 2 shows the result of this classification 
vs. ground truth. To find classification errors, 
Quality Control: The qualitative control of the classified image is visually compared to the 
striping and banding errors showed that ground truth and the satellite image It showed 
banding exists only in TM band 4 and 1s that the existence of very small forest spots 
ation neglectable (e.g. *1DN) (see Figure 1). existing sporadically in the area (but not 
Further on, in TM band 7, some undesired included as forest stands in the ground truth 
ses, the pixels are observed. While this error may be map) as well as clear cutting and fresh 
e data is removed in considerable extent using filters, it aforestation areas are included as forest in the 
»d with was not performed. Filtering has a an effect on map (but considered as non-forest because of 
lepiped all DN's of the image and since TM band 7 is their similar reflection in forest stands) decrease 
1ce data not used in the analysis and due to its high the accuracy of the classification. Further on, 
) colour correlation with TM band 5 it was omitted from the mixed pixels in boundaries between 
reas for this study. forested and non-forested areas increase the 
sen. The classification errors. 
water, Geometric Correction: The geometric correction 
Il places using 10 well distributed ground control 
villagers 9 T7 
ater for 
Special 8.5 + 
ck of a Ed 
or these al 
if these EMI PM ERR ER 
st class. E 
in two - e 151 
. well as = £ St HIR 
ET 
chosen x I 
rability u 
,adopted : 
61 
5.5 + 
5 here mare rm mm EE EEE 
LTRABBDERBRRLADESORSSSNRARA 
y of the - T Line no. 
satellite Fis. 1: Striping and banding in a homogenous area within the Caspian sea 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 225