ACQUISITION OF RULES FOR THE TRANSITION BETWEEN MULTIPLE REPRESENTATIONS IN A GIS 
Monika Sester 
Institute of Photogrammetry 
Stuttgart University 
P.O.B. 106037 
70049 Stuttgart / Germany 
monika.sester Q ifp.uni-stuttgart.de 
Commission IV Working Group | 
KEY WORDS: multiple representation, multi scale representation, model acquisition, machine learning 
ABSTRACT: 
Multiple or multi-scale representation is an issue of growing interest and importance in GIS. It deals with the representation 
of different spatial entities that describe the same physical objects in one common information system. The need for such 
a representation - instead of a description of objects on the most detailed level of resolution - results from various reasons. 
The main reason being the fact that spatial phenomena usually only occur on a certain scale - which is not necessarily the 
most detailed one. Changes in scale lead not only to changes in geometry, but also in topology and semantic. Multiple 
representations also result from different interpretations of the given reality according to scale or thematic emphasis, and also 
due to the date of capture. Since geographical phenomena have multiscale aspects, they should also be represented as such 
- and not only at one level. This then allows for an inspection of spatial data on various levels of detail - logically zooming in 
and out. Multiple representation affects data modelling and data capture, integration, storage, analysis and presentation, i.e. 
all parts of a GIS. Whereas multiple representation first was considered to be a mere cartographic problem, it is getting more 
and more obvious that it is an important issue in GIS as well. 
The paper first introduces into the problem of multiple representation and tries to clarify the terms used. The main emphasis is 
put on possible realizations of such a representation. This presumes to have a means to generate different levels of detail and 
provide links between these representations. The paper finally presents a concept for the transition between different scales 
based on an object-oriented representation. In order to go from one scale to the next, certain rules are required. These rules 
are partly given a priori, partly they are acquired automatically from given data sets with techniques from Machine Learning. 
The concept is a extension of a program developed for the derivation of object models for map and image interpretation. 
1 INTRODUCTION AND OVERVIEW cartography ever since: the national mapping agencies store 
multiple scale versions of data. It is only recently that there 
is a consensus in GIS research community that apart from 
graphics-oriented generalization there is a need for model 
generalization in a database. Thus also in spatial databases 
generalization operations have to be applied in order to result 
in a higher level view of the same phenomena. In this way the 
understanding and applicability of the data is improved. 
Geographic phenomena are highly scale dependent. This 
fact is obvious in our everyday life, consider e.g. our in- 
trinsic rules of stepping back to get an overview of a given 
scene, and getting closer in order to distinguish details. Each 
phenomenon has its corresponding level, where it is best 
understood: e.g. a sentence cannot be understood on the 
level of letters. Even individual sentences need the higher 1.1 
order structuring of sections, captions and a table of con- 
tents. Such a hierarchical multi-scale representation is used 
to guide the paths of perception - from coarse to fine. The The problem of multiple representation is straightforward and 
same holds for information represented in a data base. Usu- Well known in the domain of cartographic generalization. Mul- 
ally the information is captured for a certain purpose - which — tiscale representation has however many other aspects, just 
often determines the data model. Thus e.g. in ordertoinves- ^ to name a few (see also e.g. [Weibel 1995]): 
tigate Waldsterben, individual trees have to be modelled and 
captured, for landuse classification on a general level how- 
ever there is no need to identify a single tree, but the forest 
area as a whole is described. 
Implications and Related Topics 
> Multiscale representation allows for a controlled data 
reduction concerning spatial, semantic and/or time di- 
mension. In this way data abstraction leads to a re- 
The perception of our surrounding varies with scale. Both duction of spatial and semantic resolution and to data 
type and appearance of objects differ when getting closer or bases at multiple levels of accuracy and resolution. 
going away, resp. A given phenomenon thus is not fixed, This in turn has the effect of a reduction of storage 
but scale dependent. Dealing with this fact is an issue in space and also of a speed-up of calculations. 
768 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996 
  
In ( 
dis 
Fig 
sua 
DLI 
Sys 
File 
traf 
mo 
car 
con