g stages 
Roeland de Kok 
Classification 
The actual change detection is a quite straightforward process. A project with four layers, one being the aerial 
photos and the other three the satellite image channels is created and a first segmentation, which considers only 
the photo channel, is made. The parameters were a scale parameter of 70 and color and compactness parameter 
of 0.3. By using the nearest neighbor classifier containing spectral as well as texture and brightness features, the 
image objects were classified to forest and non-forest. The availability of the object variability makes this 
extraction of a forest mask much different from density slicing in a normal panchromatic imagery. The statistical 
parameters of the forest area encapsulate the textural information of the forest. This principle difference cannot 
be overstressed. 
A second segmentation that uses the three SPOT channels provides the second level. The scale parameter was set 
to 15 and the Color and compactness parameters were set to 0.3 each. In this level all objects with super objects - 
of forest and the spectral information of non-forest in the SPOT channels are classified as deforest classes and 
subsequently merged. 
In the resulting level, a final classification was done, separating forest, non-forest and deforested areas. This time 
an additional size criterion was included in the classification, eliminating all deforest objects smaller than 800 
pixels. So, the focus could be set to the objects of interest, the substantial changes in the forest vegetation. 
The result is a map, which shows all major forest losses over the past sixty years. By varying the size parameter 
in the last classification, the scale of deforested areas that are of interest can be varied. 
  
  
Forest Non-forest 
Deforested Areas 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
1960 1960 1960 |+| 1996 
  
  
  
  
  
  
  
Four layer image Separation of forest and Classification of 
comprised of the three 
SPOT channels and one 
aerial photo. 
  
  
  
non-forest by means of 
classification of objects 
derived from the aerial 
photo. 
  
  
deforested areas by the 
use of objects derived 
from the SPOT channels 
in combination with the 
preceding classification 
  
  
  
  
Figure 6: Aerial photo classification stages 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
219