; not known 
y an iterative 
set to some 
eight value is 
1 local DEM. 
ower then the 
rdinates of a 
nodel to the 
the following 
ant the terrain 
ethods using 
ighly increase 
olution image 
cy of exterior 
ubstitution of 
points with a 
f images of 
to a medium 
e accuracy of 
e accuracy of 
ax can be 0.01 
0 10m on the 
For exterior 
y can be 1-1.5 
on KVR-1000 
JNOS images 
ior orientation 
racy limit of 
ial from 1 to 3 
Nedra”, 1984. 
ibpixel Cross- 
lem // Optical 
September- 2 
  
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002 
FUZZY SUPERVISED CLASSIFICATION OF REMOTE SENSING IMAGE IN THE 
PRESENCE OF UNTRAINED CLASSES USING NEURAL NETWORK 
M. Kalita?, R. Devi? 
? Research Scholar, Civil Engineering Department, IIT Delhi, India — (mkalita99 9 yahoo.com) 
? Professor, Civil Engineering Department, IIT Delhi, India — (devirema42 Q yahoo.com) 
KEY WORDS: Spectral unmixing; Fuzzy membership; Trained and untrained classes; back-propagation 
ABSTRACT: 
Mixed pixels have well been recognized as one of the major problems affecting the accuracy of classification of remotely sensed 
digital data. Use of neuro-fuzzy technique to handle this problem has emerged recently. However, most studies have focussed in 
deriving fuzzy membership values only with respect to a set of predefined land cover categories used for training the neural network. 
There exists a major limitation in this approach since a mixed pixel may have some other classes, which are not defined for the 
analysis. The research being reported herein attempts to demonstrate a technique in a neural network framework to incorporate the 
unknown classes in a supervised scheme of fuzzy classification. A simulation experiment has been carried out to demonstrate the 
potential of the technique using back-propagation algorithm. Sets of synthetic data used for training the network and testing the 
potential of the technique are generated using the concept of linear mixture model for three land cover classes with respect to the 
statistics of these classes extracted from a IRS 1C LISS III sub-scene. Further, in order to appreciate the behavior of the model in the 
classification of a real image, outputs for six real classes, including three undefined classes over and above the three defined classes 
are derived for analysis. The results obtained indicate significant potential for use in case of real image classification. 
1.0 INTRODUCTION 
Classification of mixed pixels is one of the well-known 
problems in deriving thematic maps with better accuracy from 
remotely sensed data. A mixed pixel embodies more than one 
land cover categories within the spatial coverage of the pixel. 
The conventional crisp concept of ‘one-pixel-one-class’ in the 
classification leads to erroneous representation of the cover 
classes in the classified output; therefore, hampers the effective 
use of this multifaceted data source (Wang, 1990). To handle 
this problem, attention of researchers is focussed towards the 
use of a soft or fuzzy classification strategy that allows pixels to 
have multiple and partial class membership (Wang, 1990; 
Foody, 2000). With this approach, the strength of membership, 
of each pixel to each class derived from the classification is 
used as a surrogate for the fractional coverage of the associated 
classes in the area represented by the pixel (Foody, 1996; 
Bastin, 1997; Foody, 2000). A range of approaches, for 
example, softened maximum likelihood, fuzzy c-means 
algorithm etc. may be used to derive a fuzzy classification. 
Use of artificial neural network for fuzzy classification (neuro- 
fuzzy classification) of remote sensing image has emerged as a 
subject of recent interest and a number of studies are already 
reported (Foody, 1996; Moody et al., 1996; Warner and Shank, 
1997; Kalita and Devi, 2001a). However in these studies, it has 
been observed that the classifiers are capable of producing the 
strength of membership only to the classes defined for training 
the neural network. It is important to note that a mixed pixel 
may have fractional presence of some land cover features that 
are not defined as the classes of interest. Thus, since, neural 
network is not trained to handle such a situation, it is compelled 
to ignore the presence of untrained classes; thus result in 
erroneous thematic maps. This inadequacy of the classification 
techniques demands alternative technique to efficiently handle 
the problem. 
In view of above discussion, the present research attempts to 
devise a technique basically to estimate the degree of 
membership of any pixel to the trained classes when some 
unknown cover features are also present within the pixel 
dimensions. Neural network with the widely accepted back- 
propagation learning algorithm has been used. Training of the 
neural network and subsequent evaluation of the model are 
accomplished using synthetic data that correspond to three 
defined classes present in a real IRS 1C LISS III data scene. 
These synthetic data sets are generated employing the concept 
of linear mixture model (Shimabukuro and Smith, 1991; Settle 
and Drake, 1993). Additionally to present the brief glimpse of 
the behavior of the model in case of real image, tests on real 
data sets including both, the defined and undefined categories 
are also carried out. 
2.0 BACKGROUND 
2.1 Linear Mixture Model 
Linear mixture modeling (LMM) assumes that each field within 
a ground pixel contributes to the signal received at the satellite 
sensor an amount characteristic of the cover type in that field 
and proportional to the area of the field (Quarmby et al., 1992). 
Thus the signal in the ith band, x; (say) is given by 
Where f; is the proportion of the pixel’s area covered by the jth 
ground cover type, n is the number of bands and c is the 
number of cover types. In vector-matrix notation, 
X= Me ibaa. ol. (2)