|
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
4.2 Automatic Derivation of Implicit Data
As mentioned above there are rules implicit in spatial data,
however there are two different ways of approaching the goal of
extraction of implicit knowledge. These two kinds of extraction
models are on the one hand to define the rules a priori
(association rules) and to apply them to the data, on the other
hand to let the computer find the rules by itself by exploring the
data. Both ways lead to more knowledge, but in the first case it
is knowledge, which we were especially searching for, like the
concepts of chapter 3.1. The second case brings up unknown
knowledge or inherent information, which may be useful to
learn more about the data set, but can be not useful as well.
Both methods are usually known as data mining (Witten and
Frank, 2000) and will be described and examined. They are
discerned into supervised and unsupervised classification.
4.2.1 Supervised Classification: implies knowledge
discovery on the basis of predetermined models respectively
spatial association rules. Supervised classification starts from a
set of classified examples for a concept to be learnt. From this
set classification schemes for the concepts are derived, e.g.
using machine learning approaches (Michalski et al., 1998), or
also Maximum Likelihood classification (Lillesand and Kiefer,
1994). In principle every kind of knowledge representation can
be used to form a classification scheme, especially rule-based
systems or semantic networks. We will depict the process by
the help of decision trees. Every branch symbolises the
existence of a distinctive classification feature. Depending on
the result of the inquiry the adequate branch will be followed
further. In the end the model leads to a classification into
different categories of one issue. However the scheme includes
some essential problems. The sequence of the validation of a
distinctive classification feature is one determining factor. The
use of such a step by step algorithm without the possibility to
go backwards holds the endangerment of abandoning important
elements or a proper solution at an early stage. The
determination of thresholds respectively stop criterions can lead
to problems. Therefore the need of high quantitative and
qualitative data is necessary to be able to calibrate the model.
The concepts of *the centre of a city" can be implemented by
using such supervised methods. For example, we could
determine a point as a city centre, if it fulfils following
conditions:
- major streets will meet in the centre
- the buildings in the centre are larger in comparison to
areas outside
- non-existence of industrial areas
= etc. etc.
The weak point of such specifications can easily be recognised:
- the descriptions are given in natural language, which
is not directly usable by a computer
- the specifications are vague
- not all conditions might be needed in all cases
- some conditions can be more important, some less
important
- there is no guarantee, that the model composed by
humans is accurate, proper and especially complete
- possibly there are much more criteria, which we have
ignored and did not take into account. On the
contrary, we could have included distinctive features,
which do not correspond to the reality, and have only
been valid for a small test data set.
Basically we expect to retrieve a special information as a result
of predefined inputs. However, the classification model will
338
fail, if the perceptions will not agree with reality. The above
mentioned difficulty of combining the criteria and their values
is already hidden in the scheme. In the case of inadequate
combination and insufficient provision of characteristics
misinformation will be generated. On the other hand the quality
of deliberately formed models depends highly on the human
creativity and ability to reason. Spatial phenomena and
relationships have to be recognised by humans a priori to
implement them into a supervised classification algorithm.
This implies, that the setup of such models has to be done very
carefully, possibly using large test data sets in order to gain the
information from and to perform tests for verification of the
derived rules. Furthermore, a specific inference scheme has to
be designed to apply the rules to the data, that takes the
probability or the importance of a condition to a rule into
account.
4.2.2 Unsupervised Classification: The method aims at
leaving the process of knowledge discovery to the computer
itself. That means the computer has to discover rules,
separations into categories, similarities in data sets without any
predefined restrictions. Koperski & Han, 1995, describe an
approach, where spatial associations between objects have been
analysed automatically leading to the derivation of a rule stating
that “all large cities lie close to a river”. Since such rules are
induced from a finite set of examples, they cannot be verified,
but only falsified. Thus, there has to be a validation of the
utility of the detected information. It may happen, that rules
will be found, which are obvious and do not give further
knowledge. It is another process of learning to distinguish
useful and non-useful rules.
One form of Data Mining is clustering in order to find
regularities or similarities in data sets. We used it for the
following investigation:
A way to analyze geometric objects is to determine their
characteristics and to try to find regularities among them. Such
regularities then, in turn, can be considered as representatives
for a certain class of objects or a class of objects in a certain
context or environment. For linear objects or even networks of
linear objects the nodes are such a characteristic, including the
node degree, i.e. the number of outgoing lines from the node.
Furthermore, also the angles of the outgoing lines can be
important. Different types of nodes can be distinguished and
classified, as shown in figure 4:
os TY Ko Y X
ELL TEE ARW FRK | CRS KAY PSI
Figure 4. Different node types
We made some investigations analyzing the node types of
linear networks.
Three examples will point out the process:
I. While investigating the concept of the city centre with
supervised models, we introduced the criteria of crossroads in
the centre. A crossroad is a node with at least four outgoing
lines, which were expected primarily in the city centre, as there
many roads come together. The tests turned out in an
unexpected result of this investigation. The condition to find
crossroads in the city centre depends on the size of the town.
There seems to be a rule regarding the relation between the
structure of the centre, the spatial arrangement of streets and the
size of the city.
In figure 5 typical structures in the city centre are shown,
depending on the dimension of the town. In small towns often a
big street leads through and mainly TEE-junctions can be
Internat
found, |
dominat
the city
observe
itself o
located.
This co
can be f
Figu
2. June
existen:
to look
settlem
approxi
Ámong
represe
junction
there a
betwee
As sho
highwa
their c
validate
areas ir
