6.5 Classification of aerial photos 
a) The P1 orthophoto was classified with another method. In this 
scale (3 m ground resolution) the spectral feature of the classes 
wasn't enough for the exact stand separation, so I had to involve 
the texture, too. The textures of the tree stands at this scale are 
easily interpretable with the ordinary methods. The textures 
mean ages here among the same species. A texture band was 
created by a 5-5 pixel relative richness filter (Equation 
(2))(Turner, 1989). 
R= - 100 (2) 
n 
max 
Relative Richness 
The first classification step was the determination of coniferous 
and deciduous groups on the strength of spectral feature. Then 
three age sub-category were produced in both classes on the 
grounds of the above mentioned texture band. The results of 
classifications are showed in Table 11. 
Deciduous 
A A 
Coniferous 
1-15 91% 1-25 93 % 
16-30 76 % 25-50 72% 
30- 82 % 50- 79 % 
  
Table 11: The accuracy of different age-classes of coniferous and 
deciduous stands. 
Another method which was tried is a kind of Laws filter (Laws, 
1980) with the following convolution (Equation (3)) 
f mE (3) 
where v is a vector (Equation (4)) 
v=(-1-2021) (4) 
This kind of Laws filter gave the best result for forests among 
these filters, but worse than the relative richness. Unfortunately 
the determination of the different species within the two groups 
(coniferous-deciduous) wasn't possible by the texture bands. 
b) The P2 orthophoto was classified with another method. In this 
scale (0.5 m ground resolution) the texture was the main feature 
in addition to the spectral characteristic. At this scale the texture 
is not easily interpretable by common filters, because the 
different amplitude of the texture, which depends on the diameter 
of the tree crown. The tree species have different textures and 
shapes. In the case of coniferous trees I applied the method based 
an the 3D-plot of the brightness numbers (Eckstein, 1996). With 
deciduous trees the method is a little bit more complicated. If the 
forest stand is very dense, few species can be separated by the 
above method. Among the main constructive forest stand tree 
species the sessile oak is well-separable (Figure 12), while the 
beech is purely. 
  
Figure 12: Segmentation of an oak stand. 
The image was smoothed, then it can be segmented using the. 
watershed algorithm with inverse grey values. After the 
segmentation the unit number of the trees (number/ha) can be 
calculated, it's equal to the numbers of the segmented blobs. 
Then a local threshold operation was applied for determining ‘the 
tree' inside the blobs. The diameter of the tree-crown can be 
determined by the following process. 1.) Determination the angle 
of the projecting line for every tree, based on the centre of the 
photo and the blob. 2.) Specify the extension of the trees 
perpendicular to the projecting line. The diameter of the tree 
would bee this distance (Figure 13). I tested this method, and it 
was very hopefully, but wasn't always correct. 
  
Centre 
of the 
Photo 
  
  
  
  
Figure 13: Measuring the diameter. 
A new method was developed for determining the tree species, 
too. I examined the concavity of ‘the trees’ using the number of 
the boundary pixels of the local threshold operation, divided by 
the tree crown diameter. There were three classes separated in 
both class-groups on the basis of the individual trees. In the 
deciduous class-group the segmentable trees with least concavity 
(sessile oak and some other blend species demand on light), the 
unhealthy trees and the non-segmentable trees were separated. In 
the coniferous group a big concavity (spruce, scotch pine), a less 
concavity (fir), and an unhealthy group were isolated. I tried to 
build the forest stands from individual trees, but this method is 
not ready yet. So I applied these procedures only inside the 
Landsat TM thematic classes. 
7. INTEGRATING THE RESULTS TO GIS 
I put an emphasis on integrating the extracted data to GIS. The 
advantages of multi-scale sources can appear in the case of all 
the data are georeferenced to the same system, managed and 
analyzed together. 
The classified Landsat TM image was vectorised and put to the 
Arcview 3.0 as a coverage. The best background for the vectors 
till scale 1:50000 is the Brovey transformed Landsat TM - SPOT 
368 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
1:5( 
Pl 
seve 
shox 
Ti 
grot 
the : 
vect 
far v 
to p 
attri 
heal 
tabk 
extr: 
back 
NFL 
The | 
ortho 
can | 
Thes 
for tl 
long. 
I ack 
DigiT