u 2004
Is not
for the
ergetic
tail by
a first
es will
exts Or
output)
or the
merely
iay be
| object
t itself.
er facts
rope or
ons of
ries of
, and in
g. This
eet the
wey of
nent of
; of real
seen in
'B
vner Z
imes of
"data
extual
ed from
perty of
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
owner A and a different property of owner A (formerly
belonging to owner Z). Starting from this point, it follows a
straight line for a distance of 301.00 m; at this point it turns left
and follows a small creek along its windings for a length of
68.70 m, while it is adjacent to the property A (formerly
belonging to Z). From this point it continues along the small
creek and accompanies its windings along a distance of 158.60
m, while it is adjacent to the property of B (formerly belonging
to Z). At this point, it turns left and follows a straight line for
351.00 m, while it is adjacent to the property of C. At this point,
it turns left and follows a straight line for 190.20 m and turns
slightly left and follows an already existing fence for 86.00 m
up to the starting point and closes the existing boundaries while
on this last part it is adjacent to the property of A."
As can be seen in the example, the texts describe the boundary
parts of the parcels with precise lengths and names of owners of
adjacent properties. Sometimes, the texts can include additional
information about buildings, coffee trees, and other objects of
value on the land.
4.3 Scientific challenges
The scenario of the cadastre was selected as an apparently
"easy" application where the input (the describing text) and the
output (the associated map) are a one-dimensional problem
(line-following) and known in advance. The main goal of this
project is to analyse in how far a transformation from text to
map is possible. Related to this aim is the question which levels
of abstraction are used in the text and in the map. Are they
comparable? Is it possible to transfer the information of the text
and the map to a common symbolic level which allows a direct
comparison? How can topological and context information be
represented?
More possibilities and questions arise after such a common
symbolic level has been found: is each representation of a land
parcel complete and free of redundancy? Are both
representations consistent? Are representations of neighbouring
land parcels consistent in their common border arca? Is it
possible to generate a new text starting from different border
points and following the border in the contrary direction than
originally described?
A closer inspection of the presented example reveals that the -
at first glance - “easy” task of the Brazilian cadastre already
comprises a number of difficulties and challenges due to rather
compact and fuzzy text passages. Originally, each description
referred to a visual inspection of the property, thus passages in
the text such as “follows a small creek along its windings for a
length of 68.70 m” or “turns slightly left and follows an already
existing fence for 86.00 m” should be clear in the field.
However, a transformation to a map based solely on the
information of the text will have to cope with such incomplete
and vague spatial information.
4.4 Transformation
In order to represent both text and map on a common symbolic
level, a suitable form of knowledge representation has to be
chosen. In this project, semantic networks were selected since
they allow an explicit structuring of concepts and their relation
within nodes and connecting links. Examples of successful
applications of semantic networks in the area of speech and
image analysis can be found in Mast et al., 1994; Paulus et al.
2000; Miiller et al. 2003 and Kumar et al. 2004. In this work,
the applied system shell for semantic networks is ERNEST
(Niemann et al., 1990; Quint and Bähr, 1994). Figure 3 shows
an excerpt of a semantic network for the Brazilian cadastre.
219
property |
part
| boundary |
part
| b part
MET E
lx "spec spec A spec
| line | | creek | | fence |
Figure 3. Excerpt of a semantic network for the Brazilian
cadastre. Individual concepts (the nodes of the network, e.g.
property or b. part) are connected via part- and specification-
links.
This example already presents concepts of the common
symbolic level of maps and texts. Both forms of knowledge
representation are concerned with a certain "property". Besides
some possible additional information about objects on the
property, the main part of each text and the essential features of
a map are related to the “boundary” of the property. Within the
text, as well as in the map, the boundary is divided into
segments (boundary parts, “b_parts”) that are described in more
detail or annotated with additional information. One of the
details that are mentioned textually (also as text in the map
within the legend to each chosen pattern) is the type
(specialisation) of the boundary part. Besides these typical links
(part, specialisation) that are used by most approaches for
semantic networks, ERNEST allows e.g. to explicitly establish
relationships between different levels of abstraction via a
"concrete" link. A different level of abstraction occurs e.g. for
the map representation between the concepts in Fig. 3'and their
concrete realisation at a basic geometrical level (Fig. 4).
>>
b part |
spec peut
| line | | creek | | fence |
NC S. con / con
= Dx /
A ; “AB
| straight line | | irregular shape |
Figure 4. Different levels of abstraction (realised in ERNEST by
“concrete” links) within the representation of maps of the
Brazilian cadastre: boundary parts and their geometric
realisation.
This geometrical level is one example where the difference
between texts and maps becomes obvious. The only geometrical
information that is always mentioned in the texts, is the length
of a boundary part. There are other texts in the given data base
of cadastre texts where the precise direction of a line is also
given. However, such examples seem to be exceptions. Usually,
if some information about the direction exists in the text, it is
some relative and often vague description, e.g. “slightly left”
(compare section 4.3). Analysing the texts, fences seem to be
regarded as rather stable objects in this Brazilian region because
