Roeland de Kok 
  
6.2 Object based classification of the SPOT images 
Additionally to the three SPOT channels used in the pixel-based classification, a 
fourth channel with an NDVI was used for classification. As in object-oriented 
classification also form, texture and other parameters were used for classification, the 
fact that the NDVI channel is derived from two SPOT channels poses no problems. 
The Classification can be divided into three steps: 
In the first step, a separation of water and non-water areas was achieved. For this 
purpose, a segmentation of solely the NDVI channel with a relatively high scale 
parameter of 25, so that large image objects resulted and form and compactness 
parameter of 0.2 each was made. The objects were classified using the spectral 
values of the NDVI channel and area parameter. The classified objects of each class 
were fused in a knowledge-based classification and the resulting objects were used as 
' the highest object level (Figure 4 A). 
In the next step, a second segmentation was made of the red and near infrared 
channel with a scale parameter of 15 and Form and compactness parameter of 0.3 
each, producing objects suited for the separation of forest and non-forest. All objects 
with super objects of the non-water class were classified into forest and non-forest, 
using the nearest neighbor classifier in combination with training areas. The objects 
with super-objects belonging to the class water were also classified as water. The 
classified objects of each class were also fused building the second level of the class 
hierarchy (Figure 4 B). 
The third and last step was to separate N. Pumilio and N. Antarctica. Another 
segmentation was made, using all three SPOT channels in combination with a low 
scale parameter of 5. The form criteria were set to 0.2, the compactness to 0.4. This 
time five classes were separated: Water, non-forest, non-forest II, N. Pumilio and N. 
Antarctica. Water and non-forest were classified according to their relation to their 
super objects. Non-forest II was used to correct the last classification of forest non- 
forest, so objects with super objects of forest, and specific spectral values were 
assigned to this class. The remaining two forest classes were divided using the 
relations to super objects and the nearest neighbor classifier. In addition, the class 
Nothofagus Antarctica had a border attribute, which denied objects to be classified as 
N. Antarctica if they were surrounded completely by objects of the class N. Pumilio 
(Figure 4 C). 
  
Ii 
vs. 
ve 
       
       
Figure 4: 
Samples of the different 
classification stages 
6.3. Change detection with aerial photos and satellite images 
Image processing 
The first concept to detect changes in the vegetation cover was to 
compare a vegetation map digitized manually from the aerial photos 
with the classification results of the satellite images. But by doing so, 
only statements about the changes that have occurred in just the small 
part, which was covered by the Map, could have been made. The 
software eCognition gave a method for change detection that was able 
to quickly produce accurate and reproducible information for nearly 
any spot on Tierra del Fuego. 
SPOT Data Aerial Photos 
First, the data material for classification had to be produced. The aerial 
photos were neither georectified nor information about the flying height 
or the lens focus was given. So the easiest way was to correct the 
scanned photos using the satellite images as master data. The Satellite 
images were then resampled over the SPOT data resulting in a pixel 
size of 2.5 meters. The results achieved were far from perfect but 
sufficient for the task. 
  
Four Layer Image 
Figure 5: Samples of the different pre-processing stages 
  
218 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.