Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 101

number of training sites for each category; (2) method of sampling (random or
systematic sampling); (3) source of the used data for labeling training sites (ground
data, air photographs, etc.); and (4) timing of data collection.
Several authors have proven that good separability values between the LULC-
features to be classified will improve classification accuracy, because there is no
narrow relation between the average transformed divergence for a feature set and
the accuracy reached during classification (Chen et al. 2004 ). The reason is
because the separability measures are usually calculated only from the training
sites. So, these measures cannot predict the exact classification accuracy for
classified LULC-features in the whole image, if the training sites are not fully
representative for all spectral ground surface feature variations in the remotely
sensed image, including areas of potential edge effects. In general, a specified
value of an obtained separability measure can estimate a certain range of possible
classification accuracies for the examined training sites (Landgrebe 2003 ).
The training sites were chosen in a way to give the broadest possible range that
can represent all, or almost all, existing LULC-categories (especially crops) spa-
tially and spectrally. Crop fields with various planted and fallow areas (on light
soil, on dark soil, etc.) were visited. The size of the training areas was chosen to be
at least 50 9 50 m, since some studies have concluded that this is a suitable size
for training sites in semi-arid areas (Olsson 1985 ). Larger training sites were
selected, when it was possible, to reduce the effect of possible technical geo-
metrical noise in satellite data and GPS-data. Homogeneous agricultural fields
smaller than about 100 9 100 m were excluded, while they were too small in
contrast to LANDSAT-pixels of 30 9 30 m. The training site plots were taken in
the centre of the homogeneous area. The GPS-measurements were taken twice in
the middle of the field to obtain a mean value and reduce possible noise related to
the GPS-type (Fig. 5.29 ).
The size of samples also has a great importance, together with distribution, for
providing representative training sites. Justice et al. ( 1981 ) recommended that the
using of a model that takes advantage of using the characteristics of the spatial
image to define the size of a training site. The suggested model can approximate
the size of any sample quadrant as a function of the pixel size and the predicted
geometric accuracy of the images.
L = P(1 ? 2 G), Or: A = P(1 ? 2G) 2 ; where: (L: length of any side, A: area
to be sampled, P: pixel size, and G: geometric accuracy of the image). So, using
TM images with 1-pixel geometric accuracy, the size of the training site will be
0.81 hectare, the equivalent to a 3 9 3 pixel kernel area.
Generally, two procedures were used: (1) ground truth data based approach:
here, the crops to be classified were defined in addition to some of their attributes
(e.g., statistical records, agricultural calendar, etc.). Of key interest were the
strategic crops, such as the winter crops of wheat, barley, and sugar beet, and the
summer crops of cotton and corn. Random GPS-measurements were then taken at
the study area and other historical agricultural information was obtained from local
farmers in Aleppo in the Upper-Euphrates and in Deir Azzour in the Lower
Euphrates Basin. The training sites were analyzed statistically using the two