1. Classification techniques used
Unsupervised classification and classification with human assistance
are used in this study. The LANDSAT multispectral data are first analysed by
an unsupervised classification. Then several test zones are delineated in the
image and the multispectral data in these zones are used as prior information
for the classification with human assistance, the goal of which is to improve
the classification result.
a) Unsupervised classification
The method used (DOTU and INSTALLE) (1) combines the histogram me thod
and the ISODATA method and may be described by the following three steps.
1) Each pixel is represented by a 4-dimensional vector the component
of which is the measure in each spectral band. Since the measured values in
these spectral bands are largely correlated, a Karhunen-Loeve transformation
is first used in order to reduce the dimension from 4 to 2 (principal component
analysis and the first two components the most important being retained).
2) The 2-dimensional histogram of the transformed image is computed.
This histogram visualizes the location of pixels in this new and reduced space
(2 dimensions) and informs the number of pixels which have same spectral values
in this space. A class appears as a grouping close to the maxima of the histo-
gram. The unsupervised classification consists in realizing an optimal grouping
in this histogram, each group corresponding to a class.
3) An unsupervised classification (clustering) method based on the
modification of the ISODATA method is used where each "pattern" is a point in
the histogram, and the value of which represents a weighting factor. The following
four points describe the detail of the algerithm : i) find the maxima of the
histogram which are used as initial centers, ii) move the centers by use of the
ISODATA method, modified by the weighting factors ; the distance used is the
Euclidian distance, iii) then compute the intra-class distance. If it is large
for a specific class, this class will be split in two; if it is small, this
(1) H. DOTU and M. INSTALLE : A fast clustering procedure based on ISODATA
algorithm with application to remote-sensing, Tech.Rep.,Feb. 78 (also
appeared in a condensed form in the proceedings of the 4th Int. Joint
Conf. on Pattern Rec., Kyoto, Nov. 7-10, 78).
clas
clas
they
ciou
orde
of c
thes
cent
new
the
the
this
as w
assi
a li
is v
tanc
dist
dist