004
ims
od)
on
For
ave
ons
tica
1 in
International Archives of the Photogrammetry, Remote Sensing
4.2 Object-based Classification
Object-based segmentations were tried using different scale
parameters given in Table 2. As can be realized that the smaller
scale increases the dimensionality and dividing the object into
the sub-groups, while the larger scale combines the multi-
segments into one (see Figure 5).
Table 2. Segmentation parameters used for Landsat ETM+
Level Scale Color | Shape | Smoothness Compactnes
Parameter s
Level 1 3 0.7 0.3 0.9 0.1
Level 2 10 0.5 0.5 0.5 0.5
Level 3 25 1.0 0 0.5 0.5
ication
and C.
Figure 5. Image segmentation using three different scale
parameters (a= 5, b=10, and c=25).
From the acquired levels, most suitable one, level-3 has been
Selected for the classification of Landsat 7 image. Based on the
Properties of each spectral band, segments have been analysed
With different paprameters in the related classes. As a result, the
Prominent segments are grouped and located in the
‘orreponding classes. Then, classification procedure is
completed by assigning the relevant class colour to segments
and classified image is represented in Figure 7.
1121
and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
Value Name | Color]
_1 [sea |
2 |damlake |
2
=
_5_|open aieas
| EB |coalwaste
1.7 |woodland — |
Figure 6. Results of Object-based classification
After classification phase. eCoginiton software gives the users
accuracy statistics of the acquired classes. Figure 6 shows such
statistics of the classified image by error matrix based on the
samples.
2x4
UserCiytt VSaemple| 1a | dniske | seidemuriyesi dente DoenMePi| COME wosdand | Sum
Een Mar
t^ 1 ü L à i ü ü 1
Sortie 2 y D a n Ü ü 1
[rations ss 0 Ü € 9 ! 4 1 12
— Ü t 0 f ü ü Û 1;
open as 0 0 1 a 4 ü 0 5
Otago e 0 9 9 ! 1 9 5
— u ü 1 ü ji a n 12
lureisseted 2 ü D 0 n ü a ü
Sur 1 1 8 17 6 8 12
Acc
Producer 1 1 ès § 28 0% Dar
Liver 1 f ox ' 0&7 05 Dar
^ nkjeri 1 1 05 f 0727 &615 0n?
het 1 1 Das 1 aan 1 424 m
SLA Pres Cla 1 1 or 1 os ne va
Totals
lveorall Accuracy DUI
KIA 0.768
«| Ja
wan oped
Figure 7. Error matrix and statistical values for Level 3.
4.3 Accuracy of the Classification Results
Classification accuracy in remote sensing is to determine the
agreement between the selected reference materials and the
classified data. For this purpose, 350 pixel in the study have
been selected randomly and their agreement with ground truth
has been analysed. Then, error matrix has been generated and
given in Table 3. This table includes not only the producer’s
and the user’s accuracy values are given but also the kappa
statistics are mentioned.
Looking at the Table 3, settlement, open and green areas
acquired by pixel-based methods have normal user’s accuracy,
but they smaller producer’s accuracy. In general, amongst the
pixel-based . approaches, maximum-likelihood classification
gives the most accurate results. The reason behind is that in this
method, the average vector and the covariance matrices are
estimated with the higher accuracy. Of course, such a condition
depends upon the avaialbility of enough tranining patterns for
each class and this requirement has already been realized.
