. Istanbul 2004
ted from three
nonforest, A=
real situation of
sification. The
el classification
ig forest, AA is
rban areas. The
vel class. The
uping of certain
ito thematically
grouping brings
t class is formed
om young forest
:lass from urban
adapted to real
om to region to
nentation allows
segmentation for
je regions was
the first image-
and the channel
size formed the
es in the higher-
second image-
1e for lower scale
he segmentation
ied from higher
er value for more
eity played more
ified into classes.
as done by the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
nearest neighbor classifier. The classification signature space
used mean segment values of orthophotograph, channels
calculated from median filter and Gauss filter with three
different standard deviation values (equal to 2, 3, and 4). The
last part of signature space was created by channels calculated
as texture measures. Texture measures — mean, dissimilarity and
standard deviation using three Haralick functions were
calculated for three window sizes — 5x5 pixels, 11x11 pixels
and 21x21 pixels. Classifications in both levels used the same
signature space.
3. RESULTS
The higher-level classification had to be corrected in several
cases (several segments). It is a relatively quick part of
processing being performed during visual control and being
done manually. Fig. 1 shows the original image data and Fig. 2
shows result of the higher-level segmentation. The second-level
classification result is on Fig. 3.
The accuracy of classification result was controlled in random
sample areas. The accuracy was calculated for producer’s
accuracy PA(class 1) defined by
PA(class zi ) +
N
k=1
and for user's accuracy UA(class i)
UA(class _i)= —+
2,4.
k=
The producer's accuracy estimates the probability that a pixel,
which is of class i of the reference classification, is correctly
classified. The total number of pixels of class. i in the reference
classification is obtained as the sum of column 7. The user's
accuracy estimates the probability that a pixel classified as
class, i is actually of class. i. It compares the correctly classified
number of pixels of class i with the total number of pixels
classified as class i . The total number of pixels classified as
class iis in the row i.
Classified Reference classes
classes house | tree road field [forest up|deciduous| coniferous | Z of pixels
to 7y
house 8354 0 680 0 0 0 0 9034
tree 358 | 1609 199 288 0 0 0 2454
road 1831 0 4681 0 0 0 0 6512
field 792 275 264 {297253 0 0 0 298584
forestupto7y| 0 0 0 0 49813 0 0 49813
deciduous 29 855 0 0 32792 | 24173 580 58429
coniferous 0 415 0 0 9 964 81086 82474
Z of pixels |11364| 3154 5824 | 297541 | 82614 | 25137 81666 507300
Accuracies for individual classes
Producer's PA| 0.74 | 0.51 0.80 1.00 0.60 0.96 0.99
Users UA | 0.92 | 0.66 0.72 1.00 1.00 0.41 0.98
Table 1. Results of classification accuracy. The best results are in yellow and the worst in gray
The overall accuracy was 0.9 with kappa coefficient equal to 0.87.