SELF-ORGANIZING NEURAL NETWORKS IN FEATURE EXTRACTION 
Mr. Markus Törmä 
Institute of Photogrammetry and Remote Sensing 
Helsinki University of Technology 
Espoo, Finland 
markus@mato.hut.fi 
Commission II, Working group 3 
KEY WORDS: Feature Extraction, Neural Networks, Classification 
ABSTRACT 
Due to large datavolumes when remote sensing or other kind of images are used, there is need for methods to 
decrease the volume of data. Methods for decreasing the feature dimension, in other words number of channels, are 
called feature selection and feature extraction. In the feature selection, important channels are selected using some 
search technique and these channels are used for current problem. In the feature extraction, original channels are 
transformed to lower dimensional channels and these are used for problem. Widely used feature extraction method 
is Karhunen-Löwe transformation. In this study Karhunen-Löwe transformation is compared to transformation made 
by Kohonen self-organizing feature map. Tests made using artificially generated datasets show that the differences 
between compared methods are small. 
1. INTRODUCTION 
Usually remote sensing instruments carry out 
measurements using several areas of the 
electromagnetic spectrum. As a result, image provided 
by a remote sensing instrument consist of several 
spectral channels. The number of the channels can be 
seven like in LANDSAT TM-image, but it can go as high 
as several hundred when spectrometers (e.g. AVIRIS, 
224 channels) are used. Important step in data 
processing before e.g. land use classification is to find 
relevant channels for the current problem, so that 
feature dimension would decrease. 
We can choose relevant channels using knowledge about 
spectral properties of the targets represented in the 
image. For example, if we want to separate land areas 
from water areas we can use LANDSAT TM channel 4, 
because the reflectance of water is nearly zero on the 
near-infrared part of the spectrum. But usually in the 
more complicated problems we do not have this kind of 
a priori information, or it is quite time consuming to 
utilize a priori information to channel selection. In this 
case, we can perform mathematical feature selection. 
The structure of this paper is as follows: chapter 2 
represents different approaches for the feature selection 
and chapter 3 one of these methods, Karhunen-Löwe 
transformation, called also principal component analysis, 
is reviewed. In chapter 4 self-organizing neural network 
called Kohonen self-organizing feature map (SOM) is 
presented and its use in the feature selection is 
discussed. Chapter 5 presents experiments made for 
comparing Karhunen-Löwe transformation and SOM 
and chapter 6 discusses about results. Finally, chapter 
7 represents conclusions. 
374 
2. FEATURE SELECTION 
The methods for feature selection are divided into two 
groups: feature selection in feature space and feature 
selection in transformed space. Feature selection in 
feature space is made by choosing those features, which 
contain useful information and deleting those features 
which contain redundant or unnecessary information. In 
other words, we have all features in featureset Y and we 
seek the best subset of Y called X. The best subset of Y 
is chosen by maximizing some criterion function. In the 
ideal case this best subset maximizes the probability of 
correct classification compared to other possible 
combinations. Usually feature selection in the feature 
space is simply called feature selection. Feature selection 
in the transformed space is made by transforming the 
original measurement vector y to lower dimensional 
feature vector x. In this case the decrease of redundant 
and unnecessary information depends on used 
transformation. Transformation can be any kind of 
vector function of y, but usually linear transformations 
are used. Linear transformation can be written as 
x = Ay, (1) 
where À is transformation matrix. The problem is how 
to determine a good matrix À, so that useful information 
is not destroyed. Feature selection in transformed space 
is also called feature extraction. 
The best subset of all features in the feature selection is 
chosen using criterion function and search algorithm. 
Criterion function J to be maximized can based on 
probability of error, interclass distance, probabilistic 
distance, probabilistic dependence or entropy. The idea 
in all these criterion functions is to measure the 
separability of classes. The best subset could be found by 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996 
FR VS A N Fi bd NN 
© ©“ OO FHM ct oí — ce 23 
pd + 
Y^ 
40 Q We Me 
VE