( 2004
—
age, at
age of
)enalty
ty role
te and
late of
y and
on is a
ng less
TS
subset
irborne
) “of
ver an
as 220
om 0.4
image
map is
eparate
ids. the
sed on
ch had
ym was
cent of
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
original image content information. These features are used for
classification. In order to compute classification parameters and
endemember selection the training sample with proper schema
were introduced; also test sample for computing overall
accuracy and classification assessment were picked. At the end,
the image was classified by ML (figure 5) and LSU algorithms.
The results of classification were sequence, rule images and
fraction images which used for post classification.
Figure 5. Maximum likelihood classification
We implemented two algorithms of cellular learning automata
and probability label relaxation. These two algorithms were
used for post classification of the result of MLC and LSU
classification which following results (table 1) for test samples
were obtained.
Classification ML LSU
Algorithm
Post Without Without
processing post PLR| CA post PLR | CA
algorithm process process
Overs 68.01 |72.2| 84,3 | 74.40 | 783 | 87.5
accuracy
Table 1. The result of post processing algorithm
Noticing to the result, it could be realised that the accuracy of
images was increased, and the cellular learning automata
adapted itself better to the environment as compared of
probability relaxation algorithm. Another interesting result is
that the result of CLA for two classification algorithms is close
together. Therefore it could be said that the CLA could
overcome the result of poor classification such as MLC in
Hyperspectral images. And also CLA could be used for transit
fraction images computed by LSU to image classified and it is
useful for accuracy assessment of sub pixel classifier. Another
advantage of LA is that the CLA compensates the poor result of
classification algorithm and it isn’t so sensitive to initial
probability (state) but PLR is too sensitive to the initial
probability. In addition the result of CLA algorithm is almost
independent from initial probabilities and with respect to two
parameters of entropy and omission error the CLA algorithm
tries to optimize these parameters and to reach a global
optimization. However in PLR it is possible to algorithm
satisfied in local optimization. The two parameters of a and b in
equation 11 affect to the results of post processing and it
depends on our bias to local or global optimization. The
prosperous of algorithm depends on schema which designs
environment so active in which way. response actually to the
action of automata and compute penalty and reward in a real
Way. One of the disadvantage which experiments showed was
995
that the CLA takes more time as compared with PLR and we
should control it by the number of iterations.
7. CONCLOSION
The algorithm developed is so flexible that can change label
pixels to reach an agreement between neighbour pixels and
decrease chaos in environment of image classified. Therefore
cellular automata have good potential for dealing with problems
which need to find the best choice until transiting from chaos
environment to order environment.
REFERENCES
References from Books:
Mather, paul M 1999. Computer processing of remotely-sensed
images, john wiely& sons,pp.202-203
Richards, A. John, 1993. Remote sensing digital image analysis,
Third edition, Springer- Verlag Berlin Heidelberg, Printed in
Germany, pp .196-201
References from websites:
Franciscus Johannes Maria van der Wel, 2000. Assessment and
visualisation of uncertainty in remote sensing land cover
classifications Faculteit Ruimtelijke Wetenschappen
Universiteit Utrecht Netherlands
www library.uu.nl/digiarchief/dip/diss/1903229/ref.pdf
(accessed 20 sep. 2003)
Fei Qian ,Yue Zhao Hironori Hirata, 2001. Learning Cellular
Automata for Function Optimization Problems T.IEE Japan,
Vol. 121-C, No.1.
http://www.iee.or.jp/trans/pdf/2001/0101C 261.pdf
(accessed 20 sep. 2003)
Oommenl,B. John and T. Dale Roberts,2003 continuous
learning automata solutions to the capacity assignment problem,
School of Computer Science Carleton University Ottawa;
http://www.scs.carleton.ca/-oommen/papers/CascaJn.PDF
(accessed 20 sep. 2003)
zur Erlangung des Grades, 1999. A New Information Fusion
Method for Land-Use Classification Using High Resolution
Satellite Imagery An Application in Landau, Germany
Dissertation
archimed.uni-mainz.de/pub/2000/0004/diss.pdf
(accessed 20 sep. 2003)