Internation
LICE HAE 1
-
SUPERVISED AND UNSUPERVISED NEURAL MODELS FOR MULTISPECTRAL
tolerating t
IMAGE CLASSIFICATION from the |
several dif
are capabl
Chih-Cheng Hung *°, Tommy L. Coleman ^, and Orrin Long * signals. %
computatic
in parallel
ANN is ch
learning ru
are connec
each neurc
[4,5]. AN
because o
elements.
* School of Computing and Software Engineering, Southern Polytechnic State University
Marietta, GA 30060 USA — chung@spsu.edu
? Center for Hydrology, Soil Climatology and Remote Sensing (HSCaRS), Alabama A&M University,
Normal, AL 35762 USA — tcoleman@aamu.edu
* Z/I Imaging Corporation, Madison, AL 35757 USA — olong@ziimaging.com
KEY WORDS: Land Cover, Classification, Identification, Algorithms, Artificial Intelligence, Imagery, Pixel, Multispectral
One type
property,
different c
learning ne
(KSFM) :
(FSCL) ne
some reco,
algorithms
the natural
ABSTRACT:
Competitive learning neural networks have been successfully used as unsupervised training methods. It provides a way to discover the
salient general features that can be used to classify a set of patterns in neural networks. Competitive learning models have shown
superior training results compared with the K-means or ISODATA algorithms. In this study, the model is extended for image
classification. A new layer is added to this one-layer competitive learning to form a two-layer complete classification system. In
addition, a modified competitive learning (CL) using simulated annealing is proposed. The preliminary results show that this model
for image classification is encouraging. As the backpropogation multilayer perceptron (MLP) neural networks have been used for
image analysis, a comparative study is provided for images on these two different models. The models were tested on Landsat TM
data.
1. INTRODUCTION
Multispectral image classification is one of the important
techniques in the quantitative interpretation of remotely sensed
images. Remotely-sensed images usually involve a pixel
(picture element) having its characteristics recorded over a
number of spectral channels [1]. A pixel can be defined as a
point in n-dimensional feature (spectral) space. The thematic
information can then be extracted in multispectral image
classification. Hence, the output from a multispectral
classification system is a thematic map in which each pixel in
the original imagery has been classified into one of several
spectral classes. Multispectral image classification may be
subjected to either supervised or unsupervised analysis, or to a
hybrid of the two approaches [2,3].
A hybrid multispectral image classification system for
quantitative analysis consists of unsupervised training for
defining feature classes and supervised classification for
assigning an unknown pixel to one of several feature classes. In
the training stage, the objective is to define the optimal number
of feature classes and the representative prototype of each
feature class. Feature classes are groups of pixels that are
uniform with respect to brightness in several spectral, textural,
and temporal bands. Unsupervised training is a critical step in
the hybrid image classification system. Once feature classes are
defined, each pixel in the image is then evaluated and assigned
to the class in which it has the highest likelihood of being a
member using a classification decision rule in the classification
stage.
Artificial neural networks have been employed for many years
in many different application areas such as speech recognition
and pattern recognition [4,5]. In general, these models are
composed of many nonlinear computational elements (neural-
nodes) operating in parallel and arranged in patterns
104
reminiscent of biological neural nets. Similar to pattern
recognition, there exist two types of modes for neural networks
— unsupervised and supervised. The unsupervised type of these
networks, which possesses the self-organizing property, is
called competitive learning networks [5]. A competitive
learning provides a way to discover the salient, general features
which can be used to classify a set of patterns [5].
Because of the variations of object characteristics, atmosphere
condition, and noise, remotely sensed images may be regarded
as samples of random processes. Thus, each pixel in the image
can be regarded as a random variable. It is extremely difficult
to achieve a high classification accuracy for most per-pixel
classification algorithms (classifiers). Photo interpreters have
had pre-eminence in the use of context-dependent information
for remote sensing mapping.
Neural networks have been recognized as an important tool for
constructing membership functions, operations on membership
functions, fuzzy inference rules, and other context-dependent
entities in fuzzy set theory. On the other hand, attempts have
been made to develop alternative neural networks, more attuned
to the various procedures of approximate reasoning. These
alternative neural networks are usually referred to as fuzzy
neural networks. In this study, the competitive learning neural
networks and Backpropagation neural networks will be
explored for the multispectral classification.
2. ARTIFICIAL NEURAL NETWORKS FOR
MULTISPECTRAL IMAGE CLASSIFIERS
Artificial neural networks (ANNs), a brain-style computation
model, have been used for many years in different application
areas such as vector quantization, speech recognition and
pattern recognition [4,5].
In general, ANN is capable of
of the Koh
2-Dimensi
is connecte
patterns d
output no
particular |
both the
frequency-
process of
sequentiall
the center
between in
The follov
competitiv
training [8
which for 1
2-D (N x ?
is chosen s
or equal to
Step 1: In
to small ra
training da
Step 2: Pr
Step 3: Cc
node using
X; are simi
Step 4: Sc
(Le. the wi
