International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
satisfies this requirement. The sets of objects representing a
real-world object are modeled as aggregated objects, which are
associated via “Matching” to the aggregated object representing
the same real world object in the database to be linked. To
improve the performance of spatial queries the aggregated
geometry for each aggregated object is stored. The topological
relations between the single objects are modeled in the class
“Relation”.
| Objects A NE Relation | Objects B |
1.* qut
1 1
| Aggregated Object A | | Aggregated Object B |
0.1 0.1
7 Matching 7
Figure 3: Link structure
3. MATCHING PROCESS
To find the links between corresponding objects representing
the same real world object we propose a stepwise process as
follows. First the input sets for the geometric algorithms should
be as small as possible without losing quality of results.
Therefore the first step is to divide the object sets into sets of
comparable object types, that is to accomplish a semantic
classification on the object sets in the component databases.
The details of this step are described in paragraph 3.1.
The next step is to find the geometrically possible matchings,
that is the pairs of object sets which are geometrically likely to
represent the same real world object, within the comparable
object types. An algorithm for this purpose is detailed in
paragraph 3.2.
The mentioned algorithm computes more than just the “correct”
links, so that subsets of “confirmed” (which means correct) and
“discarded” matchings need to be selected from the result set.
This should be widely automated, as suggested in paragraph
33.
After the automatic selection procedures there will remain some
matchings, which could not automatically be confirmed or
discarded. For such cases of doubt an interface is needed which
provides an operator with tools to manually handle this set. The
requirements for this interface are presented in paragraph 3.3.3.
3.1 Semantic classification
To reduce the necessary amount of computations filtering
should be done which defines the input for the following
matching algorithm. Such a filter should separate all object
classes which can never represent the same real-world object,
but must not exclude any possible n:m matching. Consider for
example the two database schemas in figure 4 a). In database A
traffic routes are modelled in the classes highway, street and
alley. The differentiation between streets and alleys is made by
the importance of the roads for transit traffic. In database B
traffic routes are modeled in the classes street and alley,
whereas the differentiation is made by means of paving, that is
streets have tramac, alleys not. An algorithm for determination
174
Database A
Highway
Street
Street
Alley
Alley
Figure 4: Semantically corresponding
classes
of possible matchings should obviously compare both streets
and highways of database A with the streets of database B as
well as the alleys of database A with the alleys of database B.
Furthermore the comparison must be drawn between the streets
of database A and the alleys of database B. In figure 4 b) all
direct comparisons are shown as associations between the
classes.
Database A:
highway| street |alley
Database B: à
street alley
Figure 5
Beyond this, if a situation like in figure 5 occurs on the object
level, one cannot set aside the indirect associations (shown with
dashed lines in figure 4 b)) between highways in database A
and alleys in database B. Therefore the filter should in the first
step only separate such classes, that are not even indirectly
associated with one another, e. g. Such aggregated streets from
(maybe aggregated) railroad lines. We call the result of this step
coarse class matching, the considered object sets in the
databases coarse compare sets.
In the next step, it has to be examined, wether there is an
attribute in both coarse compare sets dividing these sets into
disjoint comparable sets, in our example say an attribute which
says if the traffic route is inside an urban area or out of town.
The classes within the coarse compare sets can be divided into
smaller, disjoint sets. And the corresponding coarse compare
sets and the coarse class matching can be divided by direct
derivation from these without the risk to lose an essential input
for the matching algorithm. We call such characterizing
attributes — "partitioning attributes". When all partitioning
attributes are applied, the resulting sets are called (fine) class
matching and compare sets respectively. The subsets of classes
forming the compare sets are called object types.
The description of the semantic classification is stored in the
integration database according to the schema in figure 6.
Internc
Obje«
[o
3.2
An als
objects
assum]
have s
transfo
a purel
should
respect
more s
The al;
every «
totally
aggreg:
partner:
