ork: 
ting 
1ing 
Vol. 
pre- 
ipu- 
ork: 
Per- 
res, 
tion 
ung, 
ls in 
ted, 
ollte 
Ob- 
utt- 
con- 
con- 
ern. 
"y 
»en- 
n. 
In- 
bust 
ls.), 
THEMATIC MAP COMPILATION USING NONPARAMETRIC 
CLASSIFICATION METHODS. 
Vladimir Cervenka 
Institut of Surveying and Mapping, Prague 
Czechoslovakia 
Commission No.: III/3 
ABSTRACT: 
A method combining unsupervised clustering and supervised nonparametric classification of multispectral 
image data will be described. The creation of sufficiently representative training sets for supervised 
classification may be a serious problem - it is difficult to find training samples, which cover the whole 
feature space. Therefore results of unsupervised classification are used for completion of terrestrial 
investigation. Then the training data are verified 
using generalized entropy measure and mutual 
information. Finally the principles of nonparametric Bayesian decision based on Parzen windows are applied. 
Nonparametric methods have been shown to yield excellent results in applications other than remote sensing 
for the present. These methods are suitable especially when there is a poor knowledge about real 
probability densities or about their functional form. Unfortunately, they require storage and computation 
proportional to the number of samples in the training set. 
KEY WORDS: Algorithm, Artificial Intelligence, Classification, Feature Extraction, Image Interpretation, 
Thematic, Training 
1. INTRODUCTION 
Gathering of information on the land use belongs to 
the main goals of remote sensing methods. This task 
is of special importance in regions with 
complicated structural zoning, e.g. in urban 
aglomerations and their surrounding. At present, 
Thematic Mapper (TM) data are frequently exploited 
for these purposes. A great attention has also been 
paid to the development of their automatic 
interpretation (classification). There are two 
principal approaches to the classification: 
supervised and unsupervised one. 
Any computer classification that will lead to 
a ground-cover thematic map is based on the ground 
truth data gathered from selected area. The choice 
of training samples has to be representative, but 
random. However, the creation of sufficiently 
representative training sets may be a serious 
problem. Satellite images cover some hundreds km? 
nevertheless it is difficult to find suitable 
training samples, which cover the whole feature 
space. Therefore results of unsupervised 
classification are used for completion of 
terrestrial investigation when significantly 
different spectral classes are determined. The 
unsupervised classification enables to reduce the 
extent of subsequent supervised classification to 
a selected subset of spectral classes. 
The notion of unsupervised classification will be 
presented in Section 2. The interpretation of 
clustering results in terms of mutual information 
will be proposed in Section 2.1. The practical 
aspects of nonparametric classification methods and 
various approaches are discussed in Section 3. 
2.  UNSUPERVISED CLASSIFICATION 
The clustering method ISODATA has been used to 
analyze satellite data (Charvat, 1987a). Using this 
method approximately 50 % sample of pixels in the 
scene is clustered. In the k-means ISODATA method 
the pixels are placed in k groups (clusters) 
according to the similarity of digital features. 
The cluster centres are established during the 
iterative clustering. Then all pixels are mapped 
onto the original spatial domains using the nearest 
neighbour classifier. To avoid the excessive CPU 
873 
time requirements, a threedimensional histogram is 
used when all samples in the feature space with the 
same feature values are represented with a specific 
histogram cell. The clustering process is realized 
in the reduced feature space only. A feature 
reduction technique is necessary for this reason - 
usually three new synthetic features (images) are 
computed. 
2.1 Feature reduction 
There are two basic reasons for incorporating the 
feature reduction procedure into the classification 
process. The ISODATA method uses threedimensional 
histogram, so the maximal number of features is 
three. A color composite production is the second 
reason for transformation of all disposable 
spectral bands into the three ones. The color 
composite created on the basis of the three 
uncorrelated features preserves great deal of 
spectral information from all original spectral 
bands. The method used improves the contrast of the 
color composite significantly. The color composites 
seems to be a useful tool for collection and 
verification of training samples as well as for the 
visual verification of classification results. 
The use of "Tasseled Cap” transformation (Crist, 
1984a) or the principal component method for this 
purpose has been described. A method based on 
neural networks can be utilized successfully 
(Charvat, 1990) when the back propagation algorithm 
(Hinton, 1987) is used. The neural net proposed 
consists of three layers. Input and output layer 
has the same number of nodes (neurons) equal to the 
number of spectral bands, the middle layer has 
three nodes in our case. Each node in the middle 
layer is connected with all nodes in preceding and 
succeeded layer. The neural net can be described by 
a unidirectional graph where nodes (neurons) bear 
some value. A certain weight is assigned to every 
connection. In the course of adaptation the feature 
values of selected samples are assigned to the 
nodes in the input layer and the values xi in the 
middle and output layer are computed according to 
the expression 
Xi = 114 Xj), (1) 
je