dinate systan 
sat subsence, 
transforma- 
S by means 
he training 
e terminal 
evels excee= 
c picture 
e statistics 
ard devia- 
es of the 
jsibility to 
cas extracted 
0 delete some 
"Ce S 
the training 
sification 
represented 
rent Super- 
fication, 
sification. 
nd thresholds 
file. The 
resenting 
fied colour 
otating drum 
ariability of 
p an opera- 
tain the re- 
in this prac- 
objective. The 
ntioned diffi- 
data proces- 
n is done by 
ystem has 
first cluster- 
ctral eharac- 
group pixels 
posed by H. 
o consists in 
n. Firstly the 
which must 
es. Then, the 
s" is per- 
of the regions 
   
  
2. 
nMng-up classifiod images 
^ S2 
+ 
The result from à supervised classification and some unsuper- 
vised classifications are pixelwise classified image 
S using multi- 
Spectral information only. This can result in à classified image 
having single or isolated pixels with a class assignment that diff- 
ers from those of tho surrounding. In many eases, Such detailed in- 
formation is not required, instead a generalised classified image 
is more useful. Such images may be obtained by comparing the class 
assignment of one pixel to its neighbours. The EPDCS system for the 
processing of a digital thematic map will eliminate all regions 
within an area less than a predefined minimum. Regions which ropre- 
sent different features can have different minima. 
8. Accuracy evaluation 
In order to compare the landsat classified image with ground 
truth, we developed this programme which transfer data between image 
and map by ground control points, then we can calcul 
- ing paramoters!:i 
AR. Difference map. 
ate the follow- 
The map will show the pixels, which have been correctly classi- 
fied, by means of an unique symbol. It will also show the pixels, 
which have been incorrectly classified, by means of different sym- 
bols representing the change of class. 
B. Confusion matrix. 
The table shows the amount of changed pixels of different class- 
es. 
C. Over all classification accuracy. 
D. Mapping accuracy of every class. 
5m. Over all mapping accuracy. 
ACKNOWLEDGE 
I am very grateful to professor Leit Wastenson, Dr. Walter Arnberg, 
Mr. Góran Alm and Miss Laine Boresjo at thé Department of Physical Geography, 
University of Stockholm, professor Juri Talts at LMV for their very useful 
ideas and experiences. 
37