Hybrid classification method 
UNSUPERVISED SUPERVISED 
CLASSIFICATION CLASSIFICATION 
/ Select training 
fieids (47) 
Run 
ISODATA 
      
    
   
Edit/Evaluate 
signatures 
Edit/Evaluate 
signatures 
  
  
     
   
Classify 
image (52) 
Classify 
image (26) 
S 
Identify 
classes (26) 
  
  
  
  
AUXILIARY DATA 
For cover class identification and selection 
Topographic maps 1:50000 
Forest survey maps 1:20000 
Land use maps 1:10000 
Soil maps 1:200000 
Crop inventory books, maps 1:10000 
Field work 1996 and 1997 
  
  
Figure2. Classification methods 
Delineating of training polygons on the image took place 
according to the peaks in numeric histograms drawn on the bases 
of composite image (TM453). Aerial photos, forest survey and 
agricultural land inventory maps were used for placing the trai- 
ning areas of each class in the image. Management plan inven- 
tories of the Alam-Pedja area contained a lot of plant descriptions 
(A) 
  
  
available for use. A greater number of communities (especially 
mire and forest types if compared to the classification scheme), 47 
different cover classes were used in the spectral signature 
composing procedure. After a thorough analysis of all these 
spectral signatures the final decision concerning the separability 
of classes was made. At last 26 signatures were used for classi- 
fication. 
Change detection. For change detection purposes the maps 
(classified to 4 types) of different years were superimposed. For 
better visualization and calculation of changes, all classes on 
classified image of 1988 were multiplied by 10. Class values in 
1995 images were not modofoed.. 
After superimposing two maps of different dates we got change 
map where transition areas between different classes are 
presented spatially and using GIS facilities described quanti- 
tatively. For example, area 12 shows transition between types 1 
and 2 during given (7 years) time span, 43 reveals change from 
type 4 to 3 a.s.o. In this way one can investigate all possible 
transitions between land cover types. All changed areas are 
recommended to check because most of narrow string-like 
transition areas are mainly induced by class edge pixels (mixed 
pixels). 
RESULTS 
Discarding the classification results, made on the basis of 9 cover 
classes we will concentrate the discussion on the 4-type 
classification result only, because this is the source material for 
change detection in this paper. The aggregation of the final 
classification classes - water, natural, seminatural and anthro- 
pogenic depends on the separability of the first level classes but on 
the highest level of generalization this problem is not very 
essential. 
Aggregation of the 26 cover classes into 4 yields two maps (of 
1988 and 1995, Figure 3). These maps were used for 
change detection. First, they were rectified to TM projection and, 
second, the nature reserve area and the buffer zone were cut off. 
Land cover classes 
1. Water 
2. Natural 
3. Seminatural 
4. Anthropogenic 
56 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
(B) 
Aft 
wei 
The 
cha