Musaoglu, Nebiye
3.2 Classification
In multispectral data analyses, spectral excess is formed due to the correlation made up of the similarities among
spectral channels, therefore, differentiability of topographic features is decreased. This correlation arises out of the
similarities in reflections of objects, proximity among the spectral channels and the topographic effects, indicating that
the spectral bands are proximate visually and digitally (Lillesand and Kiefer, 1987). Use of low correlation channels in
classification of satellite images enhances the differentiability of topographic features of the earth and positively affects
the classification accuracy.
For the purpose of determining the suitable channel combination before classification, firstly spectral curves in different
vegetation cover types were drawn on two different dated satellite images, then variance and covariance analysis was
made to calculate the correlation among the channels. When the correlation coefficients obtained as a result of such
analyses were examined, it was found out that the lowest correlation were obtained in 4* and 5" channels. In Landsat
TM data, low correlation of 1* and 214 channels with 4* and 5* channels and of 3“ channel with 4% channel, were
considered and assessed jointly with the blank areas and areas covered with water in the area of study . As a result, it
was decided that 1, 2", 3, 4™ and 5™ channels were to be used (Musaoglu, 1999).
Classification of satellite images was done in 2 stages. In Stage 1, for the purpose of having preliminary information
about the region, ISODATA uncontrolled classification algorithm was applied to all the satellite data and 30 groups
were obtained for each data. Data group obtained as a result of uncontrolled classification were compared with tree
stand maps, land use plans and other data, followed by elimination of some groups and combination of some others to
use as sample area in controlled classification.
In Stage 2, Maximum Likelihood controlled classification algorithm was applied to all the satellite data. In controlled
classification, 20 sample zones were determined on each satellite image. In determining the sample zones and
controlling the classification results, tree stand maps, ground data, orthophotos, regional photographs and personal
contacts were utilized.
In selecting the sample area in test sites, the accepted basis consisted of 1/5000 scale tree stand map, digitized to be
transferred into computer media, of Education and Research Forest of Faculty of Forestry of Istanbul University and
field measurements done on the spots . It was found out that 10 groups whose accuracy had been determined by
analysing as a result of uncontrolled classification and 10 sample areas selected from the tree stand maps have provided
adequate differentiation in the test site . Then, together with 20 classes, Maximum Likelihood controlled algorithm was
applied to the satellite images dated 1984 and 1997. While the color attributions were being made to the images
obtained as a result of classification, groups with similar characteristics were given the same colors and number of
classes was decreased (Fig. 4).
© Pinus nigra
Qin maritima
© Copse
@ Quercus spp.
1) Carpinus betulus
© Castanea sativa
C) Settlements
Qe Sea
€» Coastline
(a) (b)
Figure 4 : Classified images a) 1984 b) 1997
942 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
Duri
tree
also
3.3
Dete
and
In th
refer
data
only
agre
is di
198:
durit
data.
poss
accu
marg
total
Clas
data
used
grou
map:
exan
valu
Furt]
and