2004 
links 
pairs 
iding 
his is 
it the 
sason, 
object 
ct, all 
ed. Its 
and 
; part. 
re 4). 
art in a 
cy! .of 
S. been 
1e in a 
eatures, 
listance 
ca, their 
ations 
nce 
listance 
tance 
Se 
ince 
stance 
  
arity of 
sponding 
e.g. the 
titute an 
t class in 
values of 
reference 
means of 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
acquisition, date of acquisition, etc., of corresponding 
features 
3.2.2 Deriving similarity measures: The result of an 
integration of corresponding objects is the more significant and 
useful, the clearer and the more reliable the similarity of the 
features can be assessed. If good similarity measures are 
available, then also the applications which are using the results 
of an integration process, namely the conflation, analysis and 
update of corresponding instances, can be optimized. In our 
application, we need similarity measures in order to introduce 
thresholds. These thresholds shall be used to figure out which 
degree of similarity we actually need between instances if we 
want to deduce information about correspondencies between 
schemas. 
A lot of attributes within a MultirepresentationalRelation object 
can also be interpreted as indicators showing the similarity of 
related representations, e.g. the geometric distance, the number 
of adjacent features or the number of corresponding attributes, 
etc. The task is now to figure out how one global similarity 
measure (GSM) can be calculated from all the individual 
similarity measures (ISM). In a first basic approach, we’re using 
a weighted sum: 
GSM =Y ISM; * weight, 
i=0 
As it has been proposed in (Walter and Fritsch 1999), a 
statistical approach in order to exploit combinations of 
measures could be applied as well. 
3.2.3 Difficulties in instance matching: When a matching of 
corresponding instances is performed, we can have simple, non- 
ambiguous cases of cardinality /:/, I:n or n:m. But the process 
can also involve severe difficulties: cases can occur in which 
features of different object classes or with different attributes or 
attribute values are taking part in a /:n or an n:m relation. Thus, 
we have “pure” relations, but we can also have “impure” 
relations (see figure 7). 
  
  
  
  
  
/C ™ 
Class a 
Q----------222-2l-.--2----- 0 
1:n match: 
Class n Class n Class n „pure“ 
N 
Class b Class c 
STEUER JS Sa 0 (Rly EAT I ER EL SI O 
n:m match: 
impure* 
Class o Class p Class p nip 
  
  
  
  
  
Figure 7. Impure and pure relations between instances. 
Impure relations between corresponding representations can 
constrain the usefulness of our approach since they provoke 
ambiguities. For this reason, they have to be dealt with 
appropriately when we infer the correlation between object 
classes or attributes. Pure relations have to have more influence 
than impure. Furthermore, measures to assess the degree of 
impurity have to be found. This is part of our future work. 
155 
4. BUILDING AND ANALYZING RELATIONS 
BETWEEN MULTIPLE REPRESENTATIONS 
In the first phase of this research, a tool has been developed that 
allows building up relations between multiple representations in 
a semiautomatic way. Once the relations are created they can be 
used to automatically derive similarity measures for the schemas 
of the source data sets. This second step is still work in 
progress, only some first results are available. 
The whole software that has been implemented is integrated 
into an open, Java-based software environment, which has been 
developed by the Jump project (JUMP 2004). It consists of 
three modules (in the Jump terminology, they are called plug- 
ins): the Relation Builder module allows to build up relations 
between corresponding instances (see figure 9), the Relation 
Viewer module allows to display these relations and the 
Relation Analyzer module allows to interpret the relations. 
4.1 Building relations 
The first step of our approach consists of generating the 
relations between multiple representations stemming from 
heterogeneous sources. Basically, it would be optimal to realize 
this automatically. However, we are not focusing on the 
automatic creation of relations, but we want to exploit the 
relations in order to deduce information about schema 
correspondencies. Therefore, we have realized a semiautomatic 
approach, where an operator selects corresponding instances in 
the map view. Involving a human operator can cause 
inconsistencies, since two operators might interpret a spatial 
scene differently. Thus, a catalogue of instructions on how to 
deal with certain situations had to be set up in order to achieve 
at least similar and comparable results. For example, a rule has 
to be provided for the matching of street network data that says 
if topologically separated objects can take part in a /:n or an 
n:m relation. In our case, this is possible (see figure 8). 
  
Cardinality: 2:4 
=== Dataseta 
  
Data set b 
i Match 
  
  
  
  
Figure 8. In our case, n:m relations can also be set up between 
separated road segments. 
Up to now, relations have been set up for a test area in the inner 
city of Stuttgart, covering an area of approximately one square 
kilometre. It contains street data of Geographic Data Files 
(GDF) and the Authoritative Topographic Cartographic 
Information System (ATKIS). GDF is mainly used for car 
navigation purposes, whereas ATKIS is a topographic database 
that was set up with the intention to provide spatial data for 
different kinds of applications. Figure 9 shows a clipping of the 
test scenario.