this method can be used for RS image interpretation. Figure 6.a
Gives a part of a SPOT XSS image of Ameland, one the islands in
the North of the Netherlands. Figure 6.b. shows a first image
classification result with segments that have been identified under
different spectral classes which could be related to land cover
classes. The variety of spectral classes in this area is often due to
local variations which were not relevant for land cover
classification, this resulted in the identification of too many and
too small area land cover segments (objects). Therefore similarity
measures have been formulated between these spectral classes;
adjacent segments could now be aggregated into larger units by a
stepwise relaxation procedure based on these similarity measures.
The results are shown in Figure 6.c and d, the last result proved to
give a relevant output for land use mapping in this area.
5. FUNCTIONAL OBJECT AGGREGATION
51 Object aggregation and generalization
It is certainly not always so that object aggregation can be
achieved within the framework of one class hierarchy. In many
cases objects will be aggregated to form new functional units.
This is illustrated in Figure 7 where houses and factories are
assigned to a more general class of buildings, but then these
buildings and the lots they are on form real estate units and a
contiguous set of these units forms a street block. Similarly
sidewalks and roadways form streets.
These aggregation steps follow a bottom-up procedure in the
sense that starting from the elementary objects composite objects
of increasing complexity are constructed in an upward direction,
in Figure 7 from left to right (Smaalen, 2003). These rules for the
construction of aggregates are based on topological (adjacency)
relations between objects, as in the previous section; in Figure 7
the elementary. objects that form an aggregated object are linked
by a connected adjacency graph. But the semantic rules are no
longer specified within the context of a class generalization
hierarchy or based on thematic similarity rules.
Object Classes
for Collections)
Cblind alls?
Selector
* instances & Topology
(adjacency Grapir) t M
4 + A
* Georeference Ÿ x + mem
= | m}
m Howe
Offre biidhg
Lot
[S3 sibwak
Tox
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
This method has been applied to the topographic data of Figure 8.
This figure shows how houses gardens and streets are aggregated
in two steps to urban land use units. Similarly farm plots are
aggregated to from generalised units under the class of
agricultural land use. The source data for these figures have been
taken from TOP10 Vector Dataset from (1:10 000 topographic
map) the Netherlands Topographic Service.
Aggregation level Aggregation level 2 Aggregation level 3 Aggregation level 4
(Sup«l)
Figure 9: Classification hierarchies related to
aggregation levels (Molenaar, 1998).
5.2 Aggregation levels and classification systems
Each aggregation level within such a hierarchy will have its own
(sub) context within a thematic field, expressed through a
classification system with related attribute structures. This means
that each aggregation level requires its own thematic definitions '
implying that each aggregation level will have its own
classification hierarchy. :
This should be structured so that the generated attribute structures
provide the information which is relevant for the objects at each
aggregation level, see Figure 9. In the example of
Figure 7. the levels refer to urban land use. The
different (sub)contexts are related by the fact that
sets of objects at one level can be aggregated to
form the objects at the next higher level. There
are often also relationships between various
(stest) classes of the different class hierarchies related to
the aggregation levels.
There are bottom-up relationships between the
objects at different levels in the sense that the
state information of the lowest level objects, as
contained in the attribute data, can be transferred
upwards to give state information about the
objects at higher levels. There are also tóp-down
relationships in the sense that the behaviour of
lower level objects will be constrained by the
information contained in the higher level objects.
e Such vertical relationships have also been
mans defined in the context of gencral systems theory
Dl sweat (Bertalanfy, 1968). In (Klijn, 1995) they have
been observed in the context of landscape
ecology.
Figure 7: Functional object aggregations in an urban land use context
(Smaalen, 2003)
Interne
53
Fun
rem
of (
Spe
lase
sou
pro
in g
COV
wat
be i
pav
labe
seg!
iden
and
figu
clus
clus
unit
with
obta
base
ager
a hi;
go f
then
The
seve
spati
min«
appr
l.
The
aspe
exan
of ot
Such
natur
role «
speci
sema
unde
impli
objec
The :
form
modc
under
to av
founc
appe:
objec
