tion carried by the original cells should then be 
transferred to the new cell. 
- class driven generalization: in this strategy regions 
are identified, consisting of mutually adjacent 
objects belonging to the same class. These objects 
will then be aggregated to form larger spatial units 
with uniform thematic characteristics. The general- 
ization is then driven by the thematic information 
of the spatial data. 
- functional generalization: spatial objects at a low 
aggregation level are aggregated to form new 
objects at a higher level. The objects are functional 
units with respect to some process defined at their 
aggregation level, the processes at the different 
aggregation levels are related. 
- structural generalization: the main aim of the 
process is to simplify the description of a spatial 
system, such as drainage networks, while keeping 
the overall structure intact. This to the fact that 
after generalization the total functioning of the 
same system can be understood at a less detailed 
ievel. 
Each strategy has its own range of applications. Data base 
users should be well aware of why they are generalizing 
spatial data, so that they can chose which strategy is to 
be used. The first strategy is in most cases used when 
the geometric resolution of aspatial description is reduced 
without a clear semantic motivation. The latter three 
strategies, however, are semantically defined and 
motivated. They will be explained in some more detail 
now. 
3.1. Class Driven Object Generalization 
CLASS GENERALIZATION STEP 1 OBJECT AGGREGATION . STEP 1 
2,7 [5] nat. grassiand — [7] nat. grassiand 
i 1, 5. 6. 8, 12 c 23 
[7] deciduous forest T 
3,9. 1r EZ al forest 
2 um 2 
4,10 = [J contferous foret —7 2,4 iu 38 
arable land 
16 2 le lan TU an iure i er 7 
5,8, 124] pasture land , 10, 11 —— 52 
  
  
  
  
  
  
  
  
fig. 4: Class driven object aggregation. 
Suppose that a database contains the situation A of figure 
4, this is a detailed description of a terrain situation with 
agricultural fields, forrest areas and natural grasslands. 
This description might be too detailed for a structural 
analysis which should give information about the areas 
covered by the different major types of land use and their 
spatial distribution. A less detailed spatial description can 
then be obtained, if the original objects area aggregated 
to form larger spatial regions per major land use class. 
Figures 4 and 5 show that this less detailed description 
can be obtained in two steps: 
- first the objects are assigned to more general 
classes representing the major land use types this 
results in situation B of figure 4, 
- then mutually adjacent objects are combined per 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
class to form regions, this results in situation C of 
figure 4. 
These final regions can be considered as aggregated objects. 
The functions D/(O, 7) express then that objects should be 
aggregated per (super)class, i.e. if aggregated objects should 
be formed for agriculture then 
if O € Agriculture D(O,Agriculture) 
else D(O, Agriculture) 
7 
0. 
il 
il 
The output of the aggregation process are regions in the 
sense of section 2.3. Each region is an aggregate of objects 
that belong to one land use class, so if R,is an agricultural 
region then: 
- for all objects O, € R, is D(O; Agriculture) = 1 
- if O; € R, and ADJACENT[O,, O;] ^ 1 and D(O, 
Agriculture) = 1 
then O, € R, 
A consequence of this rule is that after the aggregation 
process there can be no two adjacent regions that are of 
the same type, i.e. that represent the same land use class. 
  
Thematic and Geometric Resolution 
The example represented in the figures 4 and 5 shows a 
situation where the thematic aspects of the newly 
aggregated objects can still be handled within the original 
class hierarchy. It might be that the same classes can be 
used as for the original objects, but the example shows 
a situation where it is quite clear that with each database 
generalization step the class hierarchy is adjusted; per step 
the occurring lowest level of classes is removed, only the 
more general classes remain, see also figures 5 and 12. 
That means that the thematic resolution is adjusted to the 
geometric resolution of the terrain description. 
There might be situations whereitis not necessary to jump 
to more general classes with each aggregation step. In 
those cases the new objects can be assigned to the original 
classes with consequence that they have the same 
attributes as the objects from which they have been 
composed. This is in fact the case if we consider the step 
from B to C in figure 4 in isolation. There the objects 
1,5,6,8 and 12 all belong to the class "agriculture". 
Therefore they have the same attribute structure. They 
are distinct because they had different attribute values. 
Within this class they are aggregated to form the composite 
object 23, i.e. 
0, « AGGRIO,, O;, Oy, 05, 0,5). 
This new object still belongs to the same class "agriculture" 
and has attributes in common with original objects. The 
attribute values of the original objects will then be 
transferred to the new object as in figure 6. 
That will always have the effect that the spatial variability 
of the attribute values will be reduced, because after each 
aggregation step the attribute values that were assigned 
per object will then be merged into one value for a larger 
object. That means that the relationship between spatial 
and thematic resolution is not only expressed through the 
link between class level and aggregation level, it also 
548 
    
fig. 
expres 
When 1 
the agg 
or by 
compo 
AIO, ] 
Examp 
and po 
or popt 
should 
fig. 
3.2. 
It is ce 
be dor 
many 
differe 
classe 
where