Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 105

It was impossible to get training samples for the study area based on accurate
remotely sensed data for 1975 and partially for 1987, as no remote sensing based
research had been carried out in this area. This is one drawback of using the
historical data, where one cannot make any field-work and gather ground truth
data. But, there is the essential advantage in the provision of initial information
about the study area, with which to compare to the present. It was not necessary to
obtain ground truth for the remotely sensed data of LANDSAT-MSS-1975,
because classification can only be done in the broad general classes in the study
area, as they have poor spectral and spatial resolution. Therefore, it was easy to
collect the represented training samples and the accuracy data, from the remotely
sensed data itself using visual interpretation. The ground truth data for LAND-
SAT-TM-1987 were found by ICARDA, but were scarce. Attempts were made to
increase the potential of these truth data by taking advantage of integrating the
remotely sensed data, the historical statistical records and the detailed spatial
schemes of the various irrigation projects (see Sect. 5.10 ).
Twenty GPS points were collected for each class of LULC. These points were
collected along the study area in fields with almost 300 9 300 m dimensions to
ensure the survival of location points in case technology related errors occurred
which would affect the accuracy of the measurements. Photographic images were
taken for several GPS-points to provide extra descriptive information about LU, in
which reference points exist, such as plants' density, length and phenological cases
(when the LULC is agriculture or natural vegetation). As regards to some LULC
such as airports, constructions areas, rivers and lakes, it was easy to find reference
points using the satellite images themselves, topographic maps or Google Earth.
Hence, the majority of reference points represent the more detailed crops types
falling under the more general class of cultivated areas.
Spectral signature generation, analyses and evaluation were processed itera-
tively. As a result, many signature files were produced due to the two classification
approaches (One- and Multi-stage classification), and multi-temporal remotely
sensed data (over many months and years) used in the study. Some results of
spectral separability based on transformed divergence were presented. The pre-
sented training sites here were those used mainly in the training study area (see
Sect. 5.7 ), and for which the optimized classification algorithms MLC, NN, SVM
were chosen. Tables 5.4 , 5.5 , Figs. 5.30 , 5.31 illustrate the increase of spectral
separability in relation to the bands used, and give an illustrated example of how
spectral separability was calculated quantitatively.
The resulting training samples for all classes were checked for normal distribution
of their digital numbers in the remotely sensed data multispectral bands. Where the
training samples' statistical characteristics differed from normal distributions, var-
ious classification algorithms and approaches were experimented with to improve the
relation of the classes and the characteristics of the study area (geographical, location
and its related effects on other sub-characteristics such as climate).
The incapability of actual representation of the studied area regarding the
accuracy ratio of automated classification and the ROIs-separability ratio among
various classes of interest to be classified, can be put down to several reasons,