normalized/weighted to the spectral depth of the image 
channels as separate channels instead of recalculating the 
image channels themselves. A maximum-likelihood 
classification was conducted on the 5 data channels 
consisting of the three SPOT-channels and two for the 
slope and aspect values. 
3.1 Object-based post classification 
An object-based approach was implemented to achieve a 
better result than the pixel-based interpretation gave. 
Object-based interpretation has several advantages 
compared to a pixel-based interpretation, as summarized 
by [Lócherbach 1992]. Object boundaries have also been 
used to reduce the effect from mixed pixels along the 
boundaries. How to establish the objects? Stands could be 
delineated with the information contents in the image by 
the use of the program SKOGIS [Hagner 1990 and 1991, 
and Kolstad 1993]. An other possibility was to use the 
stands in the 10 year old forest inventory. As the plan was 
to update the previous forest inventory an estimation of 
the reliability of the old boundaries was done. A 
comparison was made between the 1981 and 1991 
re SPOT-XS 
Statistical 
classification 
of pixels 
  
Nes 
ym 
Object boundaries 
Object 
Level 
Analysis/Edit 
Pixel : 
level = 
E 
   
Classification 
of objects 
inventories. Boundaries demarcating physical differences 
are normally unchanged except where forestry activities 
have changed them. The opposite was boundaries for 
administrative or planning reasons which could differ 
widely, but it seems to be of less importance from à 
classification point of view, as it was merely the same 
forests on both sides of these kinds of boundaries. 
The use of field boundaries for pre-object and post-object 
classification had been tested [Jansen, Van Amsterdam 
1991], where the post-object method gave the best result, 
As shown in Figure 1 a two-step object-based post 
classification was used. In the first phase (left-hand side) 
the pixel-based classification used a maximum likelihood 
statistical algorithm. 
Also other inventory class information from the previous 
registration was registered (object attributes) and rules 
were implemented to have a knowledge-based control 
function. But they had not been fully used at this stage as 
not all of the inventory information were marked at same 
level on the whole map. 
Training-fields 
  
   
Neural Network 
training and 
classification 
       
   
    
      
    
- - ^.^ Knowledge-based 
rules 
into 
CE nah. a = "gm 5 | pt SY 
I 
To 
— 
as es es. (p v 5 
Lnd m — $1599 
m^ 
a 122 CP 
rex "S — Mu 
was mad: ud 
  
    
classes or 
superclasses 
  
  
Figure 1. Flow chart for post—object classification 
544 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996