ng
he
iat
ts
nt
as
ch
sat
re
dt
on
ps
ee
ed
nd
ial
be
nd
by
ed
'as
es
Musaoglu, Nebiye
During classification, non-forest areas (settlements, roads etc.) were combined under the heading of settlements class.
Areas covered with water were roughly allocated into 2 groups, being sea and shore. In forest areas, badly deformed
tree stands were combined with mixed woods (copse) class as they had similar characteristics. In this class, there were
also various tree stands such as Carpinus betilus, Qercus spp, Castanea sativa as well as maquis areas.
3.3 Accuracy Assessment
Determining the accuracy of the classification results obtained from the satellite images ensures assessment of quality
and usability of the maps obtained from remote sensing data (Stehmen, 1996).
In the remote sensing field, the accuracy of an image classification refers to the extent to which it agrees with a set of
reference data. Most quantitative methods to assess classification accuracy involve an error matrix built from the two
data sets (classifications are read across the rows and reference data down the columns). The percentage agreement uses
only the main diagonal elements of the error matrix, and, as such, it is a relatively simple and intuitive measure of
agreement. On the other hand, because it does not take into account the proportion of agreement between data sets that
is due to chance alone, it tends to overestimate classification accuracy (Congalton and Mead, 1983; Congalton et al.,
1983; Rosenfield and Fitzpatrick-Lins, 1986; Zhenkui and Redmond, 1995). For this purpose, pixels were selected
during classification or over the classified data followed by examinaton of compatibility of these pixels with reference
data.(Treitz et al. 1992, Skidmore et al., 1996) . Random selection of pixels prevents the user from having the
possibility of obtaining preliminary information about the accuracy to be obtained. In order to determine + 5 %
accuracy of a class, more than 250 pixels to be selected from classified data is needed (Congaltan, 1991).
The Kappa coefficient has come into wide use because it attempts to control for chance agreement by incorporating all
marginal distributions of the error matrix (Cohen, 1960; Congalton, 1991). Kappa coefficient is calculated by using the
total of rows and columns of error matrix and diagonal elements and receives a value between 0 to 1.
Classification accuracy in remote sensing is indeed determining how compatible are the reference data with classified
data (Ma,Z., Redmond,L.R., 1995). For this purpose, different dated 50 pixels classified depending on the size of the
used data groups, and 400 random pixels in area of study were selected, then the compatibility of these pixels with
ground data were examined. In controlling the selected pixels; works of ground facts, tree stand maps, topographic
maps, land use maps, photographs of the area, orthophotos and contacts with people were utilized. For the purpose of
examining the accuracy (A) obtained from classification results, statistical analysis was made and Kappa coefficient (K)
values too were calculated. Values belonging to satellite images dated 1984 and 1997 are shown in Table 1 and Table 2.
Furthermore, misclassified pixels as a result of classification were calculated too. As these pixels were small in number
and located mainly in areas covered with sea, they were ignored.
Class Sea Coastline | Pm | Pn | Copse | Qs Cb Cs Settlement > A (96) K
Sea 94 94
Coastline 15 15
Pm. 6 6
Pn. 5 42 2 49
Copse 40 3 1 44
Qs. 1 13 54 3 6 77
Cb. 3 18 23
Cs. 1 1 28 30
Settlement 62 62
X 94 15 12 | 42 56 62 22 34 6387 400 89 0.87
Table 1 : Accuracy assessment of 1984 dated Landsat TM.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 943