bul 2004
USING LEARNING CELLULAR AUTOMATA FOR POST CLASSIFICATION
SATELLITE IMAGERY
B. Mojaradi * , C.Lucas ^, M. Varshosaz *
? Faculty of Geodesy and Geomatics Eng., KN Toosi University of Technology, Vali Asr Street, Mirdamad Cross,
Tehran, Iran,
Mojaradi@alborz.kntu.ac.ir , varshosazm(@kntu.ac.ir
° Dept. of Electrical and Computer science Engineering, University of Tehran ,Amirabad Cross, Tehran, Iran,
Lucas@imp.ir
KEY WORDS: Cellular Automata, Expert System, Entropy, Hyper Spectral, Information Extraction, Post Classification,
Reliability, Uncertainty
ABSTRACT:
When classifying an image, there might be several pixels having near among probability, spectral angle or mahalanobis distance
which are normally regarded as unclassified or misclassified. These pixels so called chaos pixels exist because of radiometric
overlap between classes, accuracy of parameters estimated, etc. which lead to some uncertainty in assigning a label to the pixels. To
resolve such uncertainty, some post classification algorithms like Majority, Transition matrix and Probabilistic Label Relaxation
(PLR) are traditionally used. Unfortunately, these techniques are inflexible so a desired accuracy can not be achieved. Therefore,
techniques are needed capable of improving themselves automatically.
Learning Automata have been used to model biological learning systems in computer science to find an optimal action offered by an
environment. In this research, we have used pixels as the cellular automata and the thematic map as the environment to design a self-
improving post classification technique. Each pixel interacts with the thematic map in a series of repetitive feedback cycles. In each
cycle, the pixel chooses a class (an action), which triggers a response from the thematic map (the environment); the response can
either be a reward or a penalty. The current and past actions performed by the pixel and its neighbours define what the next action
should be. In fact, by learning, the automata (pixels) change the class probability and choose the optimal class adapting itself to the
environment. For learning, tow criteria for local and global optimization, the entropy of each pixel and Producer's Accuracy of
classes have been used.
Tests were carried out using a subset of AVIRIS imagery. The results showed an improvement in the accuracy of test samples. In
addition, the approach was compared with PLR, the results of which suggested high stability of the algorithm and justified its
advantages over the current post classification techniques.
1. INTRUDUCTION need some background for using. Their accuracy depends on
the knowledge; therefore, techniques are needed capable of
There are many techniques for hyper spectral image analysis in improving themselves automatically and compensate the lack of
order to extract information. Classification is one of these complete knowledge. In this paper, at first we express different
which are used frequently in remote sensing. Maximum techniques of post processing, then we introduce components of
Likelihood (ML), Spectral Angle Mapper, Linear Spectral learning automata and their structures. We followed by
Unmixing (LSU), fuzzy and binary encoding are conventional discussing about cellular learning automata and the way of
algorithms for multi spectral and hyperspectral image learning cellular automata. As cellular learning automata is goal
classification. These algorithms have their own accuracy which oriented and try to change its action with respect many
should be investigated. In order to produce thematic map it is parameters such as its experiments, action of its neighbours and
necessary to performed post processing algorithms on the result the response of environment, it could be used for different
of classification. There are many parameters that tend to make purpose. In this research cellular learning automata is used for
uncertainty in remote sensed data. These parameters arise from post processing of result of classification which performed by
sensor system, complexity of the area that is covered by image, maximum likelihood and linear spectral unmixing algorithms.
geometric and atmospheric distortions (Franciscus Johannes, At the end the result of algorithm is compared with probability
2000).Furthermore training data, size of sample data for label relaxation.
estimating of statistic such as mean and standard deviation,
statistical model for computing statistic parameters, radiometric
overlap and also classification algorithms effect on label 2. CONVENTIONAL POST CLASSIFICATION
classified pixels. ALGORITHM
These parameters cause to decrease accuracy of classification 2.1 Majority Filter
which should be improved in post processing stage. There are : sui
many conventional techniques such as, majority filter, Tomas’s ~~ Majority filter is a logical filter which relabel centre pixel, if it
filter, transition matrix, Probability Label Relaxation (PLR) is not a member of majority class; in other word the label of
Model which are used to improve accuracy of classification majority class is given to center pixel.This algorithm perform in
results. Most of these algorithms are limited and inflexible or the following expression .
991