ware in the 
he National 
olygons of 
lade buffer- 
reasing the 
the stand's 
ining areas. 
ing time of 
1e different 
shaped at 
hairy oak, 
ne, Spruce. 
> first band 
ause of the 
remove the 
I” strips are 
and the VI. 
d from the 
opographic 
d the DTM 
| vegetation 
s. 
  
relation 
cient 
SRT EE) 
  
  
on with the 
, when the 
able. 
Then the radiometric normalisation of each bands with the above 
parameters was made by the following equation (Equation (1)) 
P 
m mM 1 
m-cosa +b 0) 
I started the classification with the above 10 classes, at first. The 
classification method was based on maximum likelihood 
algorithm using empirical probability density (Czimber, 1995). 
Then I joined and deleted some classes because of the overlaps. 
Thus seven classes left, which were the followings: robinia, oak, 
hombeam, beech, larch, scotch pine - spruce, austrian pine. The 
error matrix of the hole forested test area classified by the above 
method is shown in Table 6. The error matrices of the common 
minimum distance and maximum likelihood classification are 
showed in the Table 7. and 8. for the comparison. 
  
Rob. |Beech |Sc. p. -JA. p. |Hb. |Oak |Larch 
Sp. 
  
Ref. 0.0%] 0.2%] 0.0%] 0.0%] 0.0%] 0.0% | 0.0% 
  
Rob. [92.7% 1.9%] 0.3%] 0.0%] 0.4% | 14% | 0.0% 
  
Beech | 1.8%] 73.0% 0.2%] 0.0%12.8% [31.3% | 0.5% 
  
Sc. p. -| 0.0%] 0.0%] 55.6%) 13.0%] 0.0% | 0.0% [12.5% 
Sp. 
  
A. p. 0.0%|_0.0%|22.0% | 79.7%| 0.0% | 0.2% | 3.0% 
  
Hb. 0.0%| 12.6%| 0.0%| 0.0% |85.7%|10.0% | 0.0% 
  
Oak 5.5%| 11.1%| 0.2%| 0.0%| 1.1% |55.4%| 1.5% 
  
  
  
  
  
  
  
  
  
  
Larch | _0.0%| 1.3%|218%| 7.4%| 0.0% | 18% |82.5% 
  
Table 6: The error matrix of ‘maximum likelihood empirical’ 
classifier. 
  
Rob. (Beech |Sc.p.-|A.p. |Hb. |Oak [Larch 
Sp. 
  
Ref. 0.0%] 0.0%| 0.0%] 0.0%] 0.0%| 0.0%] 0.0% 
  
Rob. [82.7%] 9.0%] 0.0%] 0.0%] 3.8%] 33.0%| 0.0% 
  
Beech | 5.5%(54.4%)| 0.1%] 0.0%] 9.1%] 24.1%| 0.0% 
  
Sc. p. -| 0.0%| 0.0%|60.6%| 24.3%| 0.0%| 0.0%] 20.5% 
  
A. p. 0.0%] 0.0%] 21.9%|67.8%)| 0.0%| 0.0%| 0.0% 
  
Hb. 0.0%] 13.2%| 0.0%] 0.0%|86.8%| 7.1%| 0.0% 
  
Oak 11.8%] 22.6%] 0.1%] 0.0%] 0.4%|34.5%| 1.5% 
  
  
  
  
  
  
  
  
  
  
Larch | 0.0%| 0.8%] 17.2%] 7.9%| 0.0%| 1.3%|78.0% 
  
Table 7: The error matrix of ‘minimum distance’ classifier. 
  
Rob. |Beech |Sc. p. -JA. p. |Hb. |Oak [Larch 
Sp. 
  
Ref. 0.0%|_0.2%| 15%|_0.0%| 0.0%| 0.1%| 0.0% 
  
Rob. _189.7%| 1.5%| 0.0%| 0.0%| 0.4%| 1.1%| 0.0% 
  
Beech | 1.5%|63.6%| 0.1%| 0.0%| 13.2%| 48.6%| 0.0% 
  
Sc. p. -| 0.0%| 0.0%|49.9%| 9.0%| 0.0%| 0.0%| 18.5% 
Sp. 
  
À. p. 0.0%| 0.0%] 29.1%|83.6%)| 0.0%| 0.0%| 2.0% 
  
Hb. 0.0%] 13.0%] 0.0%] 0.096|86.09/6| 11.6%| 0.0% 
  
Oak 8.8%] 20.9%] 0.3%| 0.0%| 0.4%|36.5%| 2.0% 
  
  
  
  
  
  
  
  
  
  
Larch | 0.0%] 0.8%] 19.1%] 7.3%| 0.0%] 2.2%|77.5% 
  
Table 8: The error matrix of ‘maximum likelihood’ classifier. 
The main advantage of this method is to separate classes, which 
do not have a normal distribution. The distribution of a lot of 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
vegetation classes are two-peak featured, and the conventional 
maximum likelihood classifier is unusable for (Figure 9). 
frequency / 
probability 
> 
  
  
  
T T T T T T T T T de 
58 60 62 64 66 68 70 72 74 76 Band IV. 
Figure 9: The distributions of two forest classes. 
6.3 Classification of SPOT Pan 
The correlation between the SPOT P and the DTM was studied 
at first, and was found insignificant, so topographic 
normalisation (I mean radiometric) wasn't applied. Then a lot of 
classification procedures were tried with the image. A good and 
very simple method is a reclassification of the original image 
with two classes. The wooded areas are up to 63 brightness 
numbers, and the brighter (64-255) is non forested. The accuracy 
of this classification method is 94.3 % for this test area, and 
about 3 % of the error are the water surfaces. Finally three forest 
classes were separated with this reclassification of the forested 
area. It was astonishing, that the coniferous and deciduous stands 
were not separable on the grounds of the SPOT P, but the 
brightness numbers of the image correlate with the age and the 
crown closure (Table 10), but almost impossible to determine the 
characteristics of the classes. 
  
Correlation coefficient (r) 
Age 0.68 
Crown closure 0.56 
  
  
  
  
  
  
Table 10: The correlation between the stands' age, crown closure 
and SPOT P brightness numbers. 
6.4 Classification of Brovey-transformed Landsat TM and 
SPOT Pan 
The Brovey-transformed Landsat and SPOT image was not used 
for forest stand classification purposes, because there were some 
problems from the different date. There were big clear felling in 
the four-year period (‘91-‘95) because of the great woodborer 
disasters. The changes are easily demonstrable, but hard to 
interpret in the forest stand classification. I tried to apply an edge 
detection for the boundaries of stands. An edge-preserving- 
smoothing was the first step. Then an edge-detector filter and a 
segmentation method was applied (Narendra, Goldberg, 1980). 
The best result was showed by the band VI. (!!!), but only the 
major roads were extractable. 
367