Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 98

recommended that the satellite imagery for the study area was brought with the
same referred date to use in estimating crop area of summer crops. The best
spectral range for the separation of winter crops (wheat, barley and sugar beet) was
also found to be the near infrared domain. The best period to distinguish these
crops was found to be during the month of May.
''There are more short-term temporal variations in the spectral responses of crops
and other ground surface features, such as differences in spectral behavior at different
times of the day or night. Differences in the angle of the sun cause variations in
atmospheric damping. Sometimes, vegetation that is not under moisture stress early
in the morning will show severe symptoms of this later in the day'' (Hoffer 1980 ).
''Researchers have also found the problem of temporal definition of a particular cover
type of interest, for example, the use of remotely sensed data to classify corn. At what
stage of growth do you define a particular agricultural field as being corn?; do you call
field (X) a field of corn after it has been planted or after emergence, or when the corn-
stems are 15 cm high?; or is it not until the corn covers 25 % of the ground surface?;
or indeed 50 %?'' (Hoffer 1980 ).
5.5.4 Choice of the most Appropriate Bands Composite
of the Satellite Images
The optimal selection of spectral bands for classification was broadly discussed in
a variety of literature (Jensen 2007 ). There are two general kinds of techniques: (1)
graphic analysis (e.g., bar graph spectral plots, two-dimensional feature space plot,
and ellipse plots); and (2) statistical methods (e.g., average divergence, trans-
formed divergence, Bhattacharyya distance, Jeffreys-Matusita distance). They
were both applied to find an optimal subset of spectral bands (Jensen 2007 ).
Generally, it may appear that three spectral bands may be more suitable than
two, as more information is offered. Also, data that have a broader radiometry field
may provide improved results, since some of the problems related to parametric
models are avoided, whose support significantly falls outside of the data domain.
Yet, by using three spectral bands instead of two with broader data domain instead
of the standard one, classification and estimation may in fact be much slower.
''Classification accuracy does not increase linearly, or even increase at all, as
the number of spectral bands used is increased'' (Hoffer 1980 ). However, this is
not true for the spectral separability of crops or other Earth surface features, which
increases steadily as the spectral bands increased.
5.6 Training Samples: Selection, Analysis and Evaluation
The main factor in selecting training sites for supervised classification is that all
the variability within classes is representative. Only a few sites will be required in
some homogeneous classes, and more sites in classes with high variability.