Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 110

To compare and judge the different classification algorithms results, we have to,
as far as possible, exclude the influence of interfering factors. So, while this is a
comparative study, a wider choice in the same training samples (size, number,
location, etc.) in each studied year and for all remotely sensed data, and for all
compared classification algorithms, would be useful. This would not be applicable
when using the masking operation used in the multi stage classification approach.
5.7.1 The Test Area
The four administrative areas of Menbij, Ein Eisa, Al-Journia and Athawra were
selected as testing areas (sub-study-area) for applying various automated super-
vised classification approaches and algorithms. These sites were adopted as they
contained the majority of natural coverage forms and land uses which exist among
the entire ERB area. These areas were also sited within range of the agricultural
stabilization zones in the basin and contained a number of irrigated projects.
Finally, the sites were distributed in only one scene of the LANDSAT-data, which
satisfied the homogeneity in spectral and radiometric characteristics. The result:
this testing area was considered as representative to the whole basin area from the
perspective of natural and climatic characteristics, distribution of natural coverage
and land uses. Therefore, any outcomes resulting from the sub-study-area could be
adopted, generalized and applied to the whole Basin.
5.7.1.1 Unsupervised Classification
Methods of unsupervised classification have the ability to define the different
classes that could be presented in the study area before the going to the field. Then,
the natural objects that are presented in the remotely sensed data can be identified
and linked to the resulting spectral classes of classes of interest (crops, land cover
classes, etc.) (Hoffer 1980 ). For this research, the initial thematic map generated
from this approach helped to identify the features and provide the feel of the study
area, although the images could not be directly used for other analysis without
field-work.
The migrating means (or ISODATA, or nearest mean) algorithm (Ball and Hall
1965 ), is the most commonly used algorithm in unsupervised classification
approaches. It frequently executes a complete classification process; recalculates
statistics; uses lowest spectral distance method (reducing the value of the function
is the average Euclidean distance between each sample point and the matching
cluster mean) repeatedly to classify the pixels; and re-specifies the rules of each
LULC-class or candidate pixel (iterative processes). Naturally, the calculated
minimized value of the average Euclidean distance is equal to creating sphere-
shaped clusters with little difference or dispersity. There is no logical technique for
creating clusters that minimize the value of the average Euclidean distance. So, the