IONS
n widely
eospatial
n the use
message
omes not
'st results
level as a
danalyse)
Graphik
u.a. beim
on zwei
igements)
Beispiele,
nisse des
nnen, auf
AL
integrates
logue and
and other
vorld”.
is already
instance,
avigation
output is
ining to a
Piloting a
ool. Map-
dangerous
e satellite
is]y been
s. Classes
' even by
auch less
ng of that
n pointed
f a class
M of the
ompletely
eometrical
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
“Real World”
| |
Observation, Projection
“Real World” Level
Transformation physical
laws
Iconic/Verbal
Level Image
Transformation
| d
rules and procedures
cartographic verbal
compositions
i i
Map Text
|
uonoensqe
Modelling
(topological and logical structures, grammars, graphs, ...)
: |
' Symbolic Level
. Formal Representation
+
abstraction me
Figure |. Knowledge transformation in different levels of abstraction.
Nota bene: The name is not at all a self-explaining, an error-free
parameter: What is the meaning of a sealed surface, of a
forest.....? In class descriptions, there is a contradiction of
rigorously modelled physics and mathematics at the one side
and of neglected linguistics and semantics at the other.
Finally, most primary information is obviously given by
language. This holds true for geospatial descriptions, too. In
case of further computer processing or machine vision, spoken
or written language has first to be coded and then integrated in
the analysis process. Language is an essential element in
geospatial analysis. Systems for a more automatic processing
and analysis of messages, still at very low level, are under
development for a number of environments, e.g. in military
applications or emergency/disaster management (Hecking,
2003; Allen et al., 2000). It is the latter field which is addressed
in this paper.
The main question is how language can be fused with spatial
data, whereas the problem of automatic reading and
understanding is left aside.
3. TERMINOLOGY
When performing fusion, or even comparison of language and
spatial data, terminological traditions of many disciplines are
shared: from linguistics and cartography, cognition sciences and
image analysis, from Al and even philosophy (Bähr, 2001; Bähr
and Lenk, 2002). Therefore, terms have to be thoroughly
defined.
3.1 The three levels: “reality”, iconic/verbal, and symbolic
Both language and graphs represent knowledge, but coded by
different tools. In order to define knowledge, Makato Nagao's
definition (Nagao, 1990) is taken: “Knowledge = Cognition +
Logic”. “Logic” implies representation in formal structures, e.g.
Minsky's “frames” (Minsky, 1975): There is no knowledge if
217
not assigned to organised patterns.
More terms will be discussed by Fig.l which explains the
sequence of reasoning starting from the “real world” level. It
should be clear that human perception of “reality” is an indirect
process, limited to mere observation of projections. There is no
principal difference whether they are projected “on-line” into
the human's eye or indirectly transmitted via maps, pictorial, or
even by verbal descriptions.
In any case, man's process of perception cannot be separated
from his (individual) brain-based inference machine. Hence
geospatial analysis starts with secondary information laid down
in images, maps, graphs, texts etc.. This second level, accessible
to human analysis, shall be called “iconic” or “verbal”, relating
to imagery/graphs or texts, respectively (Bähr and Lenk, 2002).
Nevertheless, direct comparison or even fusion of knowledge at
this low iconic level is not possible by using computers, though
the human easily reads and compares maps, texts, and imageries
based on his a-priory knowledge.
In order to organise knowledge in a computer-accessible form, a
second transformation is required for its formal representation
on a third, symbolic level. This step goes far beyond mere
digitisation. Explicit modelling is preferred, where knowledge
acquisition and knowledge use are strictly separated, like for
semantic or Bayesian networks or for blackboard-controlled
production systems. The reason for explicit modelling is the
more rigorous model, which allows, among others, to determine
directly the quality of image analysis. This is less stringent for
implicit models, where the computer machine is trained “on the
job” by an operator, like multispectral classification or neural
networks. For explicit models, quality measure is given by the
particular model itself, for implicit models it is kept in the
operator's brain.
In Fig. 1 the word “abstraction” indicates scaling in two
directions: vertically from the first level (*real world") to the
third level (formal representation) as well as horizontally from
