43
.ssification results,
;cts only a linear
s, without any loss
Contrast En
hanced Data
1.13
20.38
12.84
8.69
36.35
8.73
11.88
n (%)
lere are only few
es in both images.
a comparison of
ause the exchange
pointed out. In this
.1 data was used as
:he lines show the
ation result of the
need
5 6 7
-
0.02
0.01
-
-
0.11
-
0.09
1.17
0.09
-
0.51
>6.95
0.50
1.83
0.43
97.25
2.01
-
-
99.88
: 1048567
le : 98.28 %
assification results
:cur in the classes
issland. The mis-
vine. Visual inter-
: scene shows, that
h are highly struc-
nage confirms the
'esults. In homoge-
niferous forest and
nd agriculture no
De explained by the
;s and the contrast
ne training areas in
lg sites may also
iracteristic to the
>f such classes is
nt will affect more
structured classes,
her reason can be
for every channel
differently, which
ginal and contrast
il maximum-likeli-
lange of assigning
elude, that differ-
inal and contrast
isually small. The
;ned to the same
?8 96.
4.2 Classification Results of Rectified and Resampled
Original Data
The proceeding of data processing for a comparison of
original and rectified and resampled data is shown in
Figure 2. It is very difficult to compare classification
results of original and geometrically preprocessed data,
because rectifying the images causes another scale and
another form. Therefore comparison is only possible by
studying the classified images and the percentage distribu
tion of classes in the whole image.
Figure 2. Comparison of classification results of geo
metrically preprocessed data
In Table 2a it can be seen, that classification results of the
original and nearest neighbourhood resampled data differ
only slightly. A visual examination confirms this result.
The classification result of resampled data with bilinear
interpolation shows greater differences compared to that
of the original data. The apparent differences are caused
by geometric transformations, which create more or less
new pixels between the rectified pixels.
Original
Original Data
Data
Nearest
Bilinear
Class
Neighbour
Interpolation
1 Water
1.13
1.15
0.87
2 Coniferous F.
20.40
20.46
20.47
3 Decidous F.
12.99
12.99
12.27
4 Urban Areas
8.44
8.51
8.66
5 Agriculture
37.45
37.53
38.94
6 Grassland
8.77
9.01
8.21
7 Vine
10.82
10.35
10.58
Table 2a. Classified pixels in each class in (%)
Because the classification of original and nearest neighbour
resampled data are very similar, the following investi
gations are carried out only with the rectified data. For
this purpose we can refer to Table 2b which shows pixel to
pixel comparison, using nearest neighbourhood resampled
data as reference image. The Table shows great differ
ences between the classification results. Only about 86% of
the whole image part are classified the same. The altering
can be explained by the method of bilinear interpolation,
which uses for the generation of the new pixels the
surrounding pixels, while the nearest neighbour algorithm
only takes already existing grey-values. Especially a class
changing of isolated pixels and other small areas, e. g.
Data
Original,
Bilinear
Interpolated
Type
Class
1
2
3
4
5
6
7
D
O
_Q
1
68.72
5.84
_
22.47
0.27
2.65
0.04
SZ
—T DO
2
0.11
92.72
3.86
1.12
0.09
0.04
2.07
S3
3
-
8.41
81.63
0.03
0.40
4.35
5.19
4
0.67
1.46
0.01
83.15
10.61
0.01
4.08
ô jö
5
-
0.01
0.02
1.74
92.90
2.28
3.05
6
-
0.02
5.59
0.02
18.22
71.54
4.61
z
7
-
2.01
3.55
4.17
14.14
2.99
73.15
Sum
of pixels
:
1915289
Percentage of pixels classified the same : 86.32 %
Table 2b. Pixel by pixel comparison of classification results
of resampled data with nearest neighbour and bilinear
interpolation.
lines, borders, etc. can be recognized. Mainly the changing
from riverbanks and lakesides to urban areas is evident.
Moreover small areas situated in a larger region have a loss
of space, while the greater surfaces increase.
In the visual comparison of both classification results it
can be seen that the result of nearest neighbour resampled
data is more differentiated, but classification of bilinear
interpolated data looks more homogeneous. The subtraction
of both images from each other shows very distinctly the
effects of bilinear interpolation to edges and lines.
However only large areas with homogeneous spectral
reflectances can be classified correctly (coniferious forest,
some areas of agriculture), while highly structured regions
will tend to more misclassification. An additional inspec
tion of the feature space shows that changing of pixels
tends mostly to adjoining classes.
4.3 Classification Results of Rectified and Resampled
Contrast Enhanced Data
In this case-study for rectification and resampling the
contrast enhanced data are used (Figure 3). It is expected,
that classification results of this data show only small
differences in comparison to the results of chapter 4.2,
because the classification of the original and the contrast
enhanced data also differ slightly.
Figure 3. Comparison of classification results of radiome-
trically and geometrically preprocessed data
