Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 111

data will be continuously classified until either a maximum number of iterations
have been executed or a maximum percentage of unchanged pixels have been
achieved between two iterations (Jensen 2005 ). The process starts with an iden-
tified number of random cluster means or the means of existing signatures, and
then it processes iteratively, so that those means move to the means of the clusters
in the data. ''The ISODATA-classifier filters cluster by splitting (if the cluster
standard deviation exceeds a predefined value and the number of pixels is twice
the threshold for the minimum number of members) and merging (if either the
number of members (pixel) in a cluster is less than a certain threshold or if the
centers of two clusters are closer than a certain threshold)'' (Jensen 2007 ). There
are various forms of this technique, but in all of them at least two factors have to
be defined by the analyst: clusters number; and the iterations maximum number
(this ensures the method will stop if convergence is not achieved).
It has some drawbacks. A few of the generated clusters are not important in
regard to reality as they represent a mix of unlike LULC-features or ''on the
ground'' classes. It is also not unusual that some spectral classes build one func-
tional class, and it has to be remerged. And, there is a causal bond between the
functionality of this algorithm and the ability of the user to identify the number of
present spectral classes (Hoffer 1980 ). Many of the data characteristics that a photo
interpreter would use to identify an individual LULC-feature (such as: shape, size,
texture, shadow, etc.) are not used in classification of the data that operated based
on the computer digitally (Hoffer 1980 ).
The ISODATA-algorithm has proved useful as an indicator and guide as it
provides an idea of the relative stability of each category (McCoy 2005 ). The
individual data are processed using the unsupervised ISODATA-algorithm to
generate a large number of class assortments. These so-called clusters are then
supposed to represent classes in the image and are utilized to compute statistics of
the class signatures. It is helpful to define relatively homogeneous features to be
used as training sites in the potential supervised classification approach (Scho-
wengerdt 2007 ), where pixels that always arise jointly in the same cluster are
strong and are a very homogeneous category.
It was found that the hybrid-procedure integrating ISODATA-clustering with
the supervised classification algorithms such as MLC seemed to be the most
satisfactory and effective procedure to follow as it simplified the work and pro-
duced better results. This was the case mainly in land areas with wild habitat where
the fields were small, or where the LULC-categories and spectral classes were
complex (Hoffer 1980 ). The classification approach is illustrated in Fig. 5.33 .
The parameters for the performance of ISODATA-algorithm were given as
follows: Number of classes = 25; Maximum iterations = 20; Convergence
threshold = 0.98. A thematic raster layer and a signature file (identifiable) were
created from the ISODATA-clustering. As, it was found that water bodies, bare
areas, artificial surfaces and fallow ground could be clearly identified using the
ISODATA-clustering. It gave general information about the spectral mixture
between the various LULC-features. Mixtures were between built-up areas and
dark color-tones bare areas; dark color-tones bare areas and fallow on dark soils;