LANDUSE CLASSES DISCRIMINATION WITH SATELLITE IMAGES 
8ASE0 ON SPECTRAL KNOWLEDGE 
Vladimir Cervenka , Karel Charvbt 
Geodetic and Cartographic Enterprise, Prague 
Earth Remote Sensing Centre 
Introduction 
At present, muitispectrai 
in many remote sensing 
attention has also been 
automatic interpretation 
principal approaches to 
unsupervised one. 
image data are frequently exploited 
applications. Therefore, a great 
paid to the development of their 
(classification). There are two 
the classification: supervised and 
A method of supervised classification of 
data, based on spectral knowledge, is 
contribution. Training data are collected f 
using the supervised classification. The 
samples has to be representative, but ra 
creation of sufficiently representative tr 
a serious problem. Satellite images cover 
nevertheless it is difficult to find suitabl 
which cover the whole feature space. 
muitispectrai image 
described in this 
or every class when 
choice of training 
ndom. However, the 
aiming sets may be 
some hundreds km 2 
e training samples, 
It is necessary to find such classification rules, which 
generalize the properties of training samples (being localized 
in a certain part of image only) for the whole area of 
interest. The approach based on the creation of data base of 
spectral knowledge seems to be an appropriate solution of the 
problem mentioned. Such classification system characterizes the 
target classes in terms of numerical rules, which reflect 
characterictic relations between spectral bands. The 
band-to-band relations describe the shape of the spectral 
reflectance curves [ 1 ]. 
The spectral knowledge based classification - overview 
The spectral knowledge based approach prefers the description 
of target classes on the basis of certain relations between 
individual spectral features [ 1 ]. This approach can be used 
to avoid the scene-specific limitations - the data base of 
spectral knowledge can be used (to a certain extent only) for 
classification of further scenes. ine target classes are 
described using the inequalities (so called spectral 
relations), which can be written in the general form 
SI > T , 
where SI is a spectral index (a numerical expression in terms 
of spectral features) and T is a threshold. Every spectral 
relation is evaluated in regard to the ability to separate 
various subsets of target classes. It is necessary to find 
out, which target classes fulfil a certain spectral relation. 
This knowledge can be used for target classes discrimination 
using a binary tree classifier. Ihe whole classification 
procedure could be divided into following steps: 
- collection of suitable training samples for every class 
325