Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 141

planted crop. After all the representative training samples for each agriculture-
category (permanent, winter, and summer) were collected, the supervised classi-
fication was conducted under the vegetation-mask generated from NDVI-values.
This resulted in the collection of several training samples within the irrigation
agricultural land projects, which represented the majority of variations in the
LULC-categories, especially the agriculture class. Some of training-samples were
used in the automated classification process, whereas others were used to evaluate
the accuracy of these classifications.
At this point, a query arose over what requirement was needed to link statistical
records with spatial records concerning the irrigation projects. This was needed for
the training-samples representing the LUs which would be used later in automated
classification, despite the availability of other referential GPS-points. There were
three reasons for this. Firstly, the GPS-points did not totally cover all the study
sites. Points of ICARDA-1987 were all located within Aleppo governorate, but no
points existed in the Arraqqa and Deir Azzour. The GORS-points for 2005 were all
located within Deir Azzour. The only points distributed over the three govern-
orates were those taken in 2007. Secondly, because of the relatively large exten-
sion of the study area which lies within various natural regions (climatic: rains,
temperature, humidity and soil), there was too much variation in the cultivations
and plants along the basin. As, points of ICARDA-1987 did not represent poplar
tree farms which existed only within the irrigation agriculture projects area in the
Arraqqa (Al-Asad institution project). Thirdly, the variation in the method of
collecting the GPS-points and the training-samples proved of issue, as well as the
potential error in measurements of the GPS, relating to technical reasons. It is
hypothesized that if one of the points measured a wheat field which neighbored a
field of barley, then if the GPS-device was inaccurate enough or the satellite image
that received the point was not correctly referenced, then the point may be shown
as lying within the closely bordered barley field. So, mistakes could occur in the
classification process.
After the collection of enough training samples for the existing various LULC-
features from the state farm, the whole 21,000 ha project was generalized via the
supervised classification (Fig. 5.57 ). Here, the multi stage classification was fol-
lowed with various created masks. Initially, both the unsupervised and supervised
classification were used to classify the five general classes (see Sect. 5.8 ). Then,
the approaches were combined into two more general classes, i.e., uncultivated and
cultivated areas. The subset ''cultivated areas'' represented the actual planted fields
and the fallow and/or drilled lands, and displayed the irrigated area in project-
scale. A mask was then created to represent the spatial distribution of this culti-
vated areas class, and to eliminate the uncultivated areas and their negative
spectral influence on the other features. It also reduced the computer processing-
time of the data (though the user-data-interaction time was increased due to the
greater number of processing-steps). After applying the masking-process and the
supervised classification, the three classes were obtained (trees, herbaceous and
fallow). The tree class was extracted from the next classification steps. Two masks
were then built; the first for the herbaceous and the second for herbaceous and