Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 100

study for: (1) identification and labeling the broadly general classes (e.g., water
surfaces) and some sub-classes (e.g., trees, since they change slowly over the
time); (2) help in drawing the out-borders of the irrigated projects; and (3)
assistance in assessing the classification accuracy (especially for general classes).
Ground-reference data were compiled from ICARDA for the remotely sensed data
obtained in the year 1987, from GORS for the remotely sensed data coverage for
2005, and from the two excursions carried out in the years 2007 and 2009. Parts of
these ground truth data were used in the generation of training samples and others
were used for accuracy assessment at the end of the classification.
Several measures of class separability have been suggested as way to isolate
optimal or near-optimal subsets of features for use with classification algorithms.
Swain (1978) found three approaches: divergence; Jeffries-Matusita distance; and
transformed divergence. The general concept is that the used approach can make a
quality measure of the discrimination ratio of a group of spectral features, when
achieved over all classes. By comparing between all the achievable combinations
of subsets of the spectral features (e.g., which three out of nine available spectral
bands), the one that presents the highest quality metric can be used. Only the
reduced subset of spectral bands is then used in the overall image classification
process. A potential problem is that if one combination of spectral bands creates
classes with a large divergence values for some classes and small values for other
classes, and a second creates a small divergence values for all classes, which
represents a better overall pair-wise selection of features. This suggests that
increasing the pair-wise divergence has a decreasing return (Schott 2007 ). Swain
(1978) invented the Jeffries-Matusita distance to overcome this problem, but it had
the disadvantage of time-consuming computing. A more commonly used heuristic
approach is the transformed divergence that has the mathematical statement:
Div ij ¼ 21 e ð Div ij = 8 Þ
''This has the characteristic of exponential saturation of the divergence measure
and scales the transformed divergence over the range 0-2 00 (Schott 2007 ). Mausel
et al. ( 1990 ), in assessing separability measures, used the scaling factor of 2000
rather than 2 that gave larger additions for differences between small divergence
values (Schott 2007 ).
For example, when classifying crops, it is important to train not only the crop
classes of interest but also the other classes of no interest such as urban, water, etc.
if they occur in the region. Alike, when we focus on a few existing crops (e.g.,
wheat, barley, etc.), we also have to classify all other crops (e.g., lentil, cumin,
etc.) and list them under ''other crops'', for example. Failure in the training phase
generally results in cases of the untrained classes being commissioned. This means
that the analyst must spend considerable time and effort in training the classes of
no interest.
Training samples selection also depends on many factors which affect classi-
fication results and their accuracy. They are, according to Foody et al. ( 2006 ): (1)