![](bg2.png)
2 SYSTEM ARCHITECTURE
This paper presents an enhanced version of
LSGENSYS (Nair, 2007). LSGENSYS is an
integrated system that allows the user to analyse
images, extract feature descriptors such as area,
length, location etc of patterns and then use these
descriptors to form linguistic summaries of these
patterns. This system also provides an interactive
environment, wherein, the user may suggest some
possible linguistic summaries for the image patterns.
The system would evaluate the fitness of these user-
summaries. Alternately, the system could, without
user intervention, generate some possible summaries
and evaluate their fitness and suitability with respect
to the image patterns.
The system architecture is shown in Figure 1.
The input image is analysed and some feature
descriptors extracted. These descriptors are stored
thereafter in a relational table in the database. The
blackboard holds the current state in the process of
developing summaries. The user has the choice of
suggesting concepts such as descriptions of area,
length, location of patterns etc. The knowledge base
uses geographic facts to define feature descriptors
using fuzzy sets. It interacts with a built-in library of
linguistic labels, which also interacts with the
summariser as it supplies the necessary labels to it.
The summariser receives input from these
components and performs a comparison between
actual feature descriptors of the image patterns
stored in the database, the concepts suggested by the
user, and the feature definitions stored in the
knowledge base. After this comparison, the
summariser uses the linguistic labels supplied by the
library to formulate some possible summaries for
each pattern/object in the database. These summaries
are stored in the blackboard. From among these
summaries, the most suitable one describing each
pattern is selected by interaction with the engine
(genetic algorithm). As the GA evolves through
several generations, it generates better summaries
(indicated by higher fitness, as defined in Section 4)
which are then stored and indicated on the
blackboard. Thus, the system includes some
elements of supervised classification and
summarisation. The following set of rules is
developed to perform pattern classification and
identification in multi-band satellite images.
1. If a pattern/object is to be classified as an
island, it should have a water envelope
surrounding it such that it has a uniform band
ratio at at least eight points on this envelope
(corresponding to directions E, W, N, S, NE,
NW, SE, SW). Also grey level values on the
envelope could be lower than the grey level
values on the object.
2. If an object does not have an envelope in all
directions as described in rule (1) above, then
it is classified as land.
3. If an object is to be classified as water body
(expanse of water, river), it is necessary that
it should have a uniform band ratio.
4. Fire is classified as a separate pattern. It is
identified by applying colour density slicing
to the image and by viewing the histogram of
the affected area. The histogram would show
a majority of pixels at lower intensity for the
burnt scar area near the fire.
5. The enhanced system presented in this paper
includes the capability to recognise and
summarise urban settlements in an image.
This new rule for classification is described
as follows. Urban settlements are classified
based on their shape and their grey level
intensity. In this paper, an attempt is made to
identify only rectangular and circular-shaped
urban settlements. The shape ratio (defined as
ratio of the area of the urban settlement to the
area of the bounding rectangle or bounding
circle, as the case may be, provided both
shapes equal in perimeter or circumference)
is a major parameter in this classification
process. Shape ratio closest to a value of 1 is
ideal. The grey level intensity (closeness to
white colour, where the triplet (R,G,B) =
(255,255,255), for urban settlement) is
another parameter used in this process.
(Younes et al., 2004) provided the basis for
the development of this criterion. A
percentage of accuracy is calculated (Section
4) and associated with this classification. No
attempt is made to describe the size of the
urban settlement.
The summaries generated by the system for patterns
such as land, island, water body, and river are also
more descriptive as shown in Section 4.
3 APPROACH
The attributes of patterns that are used to develop
their linguistic summaries are area, length, location
(X, Y pixel co-ordinates of centroid of pattern in
image), Additional Information or Pattern Id, grey
level intensity, and Shape Id. These attributes are
calculated /extracted automatically by the GUI tool.
ICEIS 2008 - International Conference on Enterprise Information Systems
6