MULTISPECTRAL CLASSIFICATION OF LANDSAT TM IMAGES BY NEURAL NETWORKS
Ivica Martic
Remote Sensing Laboratory, Zagreb, Croatia,
Slobodan Ribarié
Faculty of Electrical Engineering and Computing, Zagreb, Croatia
Vladimir Kusan
Faculty of Forestry, Zagreb, Croatia
Commission VIL Working Group 4
.KEY WORDS: Neural network, Back-propagation, Classification, Multispectral images, Remote sensing, Landsat TM.
ABSTRACT
This work presents the results of experiments with the use of artificial neural networks for land cover classification. The goal of this
research was to examine the operation of back-propagation neural network and its application in the classification of remotely sensed
data. A back-propagation network with one hidden layer is used to classify a Landsat TM image. The results are compared to Gaussian
maximum likelihood classifier. We show that neural network is better in performing than the maximum likelihood classifier.
1. INTRODUCTION
The use of artificial neural networks in remote sensing is
relatively new, dating back eight years. Traditionally, supervised
classification of multispectral remote sensing images is
performed by statistical classifiers. Parametric statistical
classifiers, such as maximum likelihood, require that the data
have a Gaussian distribution. It is theoretically optimal if the
assumptions about the probability functions are correct (Bischof
et al. 1992). Neural networks are distribution-free and there is no
need for prior knowledge of the statistical distribution of the
classes in the data.
Few studies compared the back-propagation neural network
classifier to statistical classifiers. (Benediktsson et al, 1990), and
(Civco, 1991) found that the statistical classifiers are able to
perform better, while (Bischof et al. 1992), (Heerman et al.
1992), (Kanellopoulas et al. 1993) and (Liu et al. 1993) found
that the classification obtained by neural networks is similar or
better than one obtained by Gaussian maximum likelihood
classifier. All these researches showed the usefulness of
application of artificial neural networks in the analysis of data
collected by remote sensors.
In this research our primary goal has been to investigate the
operation of back-propagation neural network and compare it
with Gaussian maximum likelihood classifier. We have
examined the various architectures of back-propagation neural
networks ( by changing the number of nodes in a hidden layer) in
order to find out the best one suitable for given Landsat TM
scene.
In this paper, we first discuss the fundementals of two classifiers:
statistical (parametric) Gaussian maximum likelihood and neural
network (non-parametric) back-propagation. The following
sections present the data used in this work and their results.
Finally, conclusions are given in the last section.
2. METHODOLOGY
2.1 Maximum Likelihood Classifier
Bayes decision theory is a fundamental statistical approach to the
problem of pattern classification. This approach is based on the
376
assumption that the decision problem is posed in probabilistic
terms, and that all of the relevant probability values are known
(Duda, Hart, 1973). The basic decision function for each pattern
(class) o; is
g,(x) = In p(x/w;) * In p(w;) (1)
where p(x/ o ;) are the conditional densities and p(o ;) are
a priori probabilities. When it is known or it is reasonable to
assume that the conditional densities are Gaussian normal
1
p(x) = Gy pom de Au mj) UA (zr - m)] 2)
where each density is completely specified by its mean vector
m, and covariance matrix 2; which are defined as
m, 7 Ej {x} (3)
Zi = E; ec - m; Xa " n) | e
where E; {e} denotes the expectation operator over the patterns
of class c»; , the Equation (1) becomes:
8;(x)=1n fw;)-= in 25 - inf; |-3[6- mj) = (x m)
(5)
Since the term (N/2)1n 22 does not depend on i it can be
eliminated from the expression. If the a priori probabilities are
assumed to be equal the first term can be ignored as well as the
factor y 2.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
"- ($9 M" 1
