44
By examining Table 3a it is apparent, that only a little
change in the distribution of each class in respect to the
classification result of the original data, has taken place.
This is especially evident for data, which is resampled with
the nearest neighbour algorithm. Moreover the classifica
tion results of the bilinear interpolated data seems to have
only small deviations relative to the original data, but
visual interpretation shows significant differences.
Original
Class
Data
1 Water
1.13
2 Coniferous F.
20.40
3 Decidious F.
12.99
4 Urban Areas
8.44
5 Agriculture
37.45
6 Grassland
8.77
7 Vine
10.82
Contrast Enhanced Data
Nearest Bilinear
Neighbour Interpolation
1.15
1.04
20.43
20.78
13.06
12.30
8.61
9.05
37.36
37.61
8.99
8.45
10.40
10.76
Table 3a. Classified pixels in each class in (%)
In order to investigate the exchange of pixels assignment
to classes of the differentially resampled data, Table 3b,
using data resampled with the nearest neighbour method as
reference image, is referred. Generally this Table shows
the same tendencies as Table 2b, but with 89% pixels
classified the same, the difference in classification results
between the resampling algorithms is smaller than in
Table 2b. Mainly the agreement of water pixels increases
remarkably.
Data Contrast Enhanced, Bilinear Interpolated
Type
„ Class 12 3 4 5 6 7
Figure 4. Comparison of classification results of doubled
and not doubled data, radiometrically and geometrically
preprocessed
u
c
o
jQ
1
83.34
6.74
-
9.88
-
rtf
jC
XL
60
2
0.12
94.69
3.34
0.53
-
C
LLÎ
’5
3
-
8.62
84.06
0.02
0.05
•M
z
4
0.63
0.94
0.01
88.88
5.26
rtf
00
d)
5
-
-
-
1.95
93.56
+-*
C
L.
6
-
0.02
4.58
0.01
13.51
o
O
z
7
-
1.48
2.19
4.29
9.45
Sum of pixels :
Percentage of pixels classified the same :
_
0.03
Contrast
Rectified Data
_
1.31
Enhanced
None-
Doubled
3.37
3.87
Class
Data
Doubled
0.03
4.26
1.81
2.69
1 Water
4.22
4.59
4.03
77.81
4.07
2 Coniferous F.
9.12
9.06
9.02
3.22 79.39
3 Deciduous F.
8.89
7.90
8.25
4 Urban areas
28.66
28.05
28.54
1915289
5 Agriculture
34.82
34.16
34.90
89.13 %
6 Grassland
14.29
16.01
15.04
Table 3b. Pixel by pixel comparison of classification results
of resampled contrast enhanced data with nearest neigh
bour and bilinear interpolation.
4.4 Classification Results of Contrast Enhanced Data,
Doubled and None-Doubled before Rectification
Classification of rectified data with resampling of nearest
neighbour delivers results with only less accuracy loss. But
if classification of resampled data with bilinear inter
polation is carried out the results will be of unsatisfying
accuracy. In order to obtain better classification accuracy,
the experiment of doubling the input data before rectifica
tion is carried out (Figure 4).
Considering the available disk memory only a little part
of the so far investigated region is choosen. Consequently
new training areas for now six classes are defined. Subse
quent classification of contrast enhanced original data and
rectified doubled and none-doubled data is executed. Table
4 shows the percentual distribution of classes in the
results.
It can be pointed out that the classification outcome of
rectified data with doubled input is much more similiar to
the outcome of geometrically none-processed data. In
studying the results visually, better agreement of these
data with the original data can be recognized, especially in
lines and edges. An additional evaluated difference image
shows mainly distinctions in riverbanks, lakesides and edges
Table 4. Classified pixels in each class in (%)
of greater areas. Also isolated pixels, which disappear in
normally rectified data, occur in the result of doubled
data. Consequently this shows that doubling of data before
rectifying produces an obvious improvement of classifica
tion accuracy.
5 CONCLUSIONS
It can be concluded that the various kinds of data prepro
cessing before classification influence the results. The
main results of the above investigations are briefly dis
cussed. They underline the different influences of the
various preprocessing methods.
Classification of original and contrast enhanced data
shows only small differences, if linear histogram
stretching to each channel is carried out separately.
Classification results of nearest neighbourhood re
sampled data show only little change in respect to
the results of original data. This is valid for re
sampling of original and contrast enhanced data.
Classification results of bilinear interpolated data
show great differences in respect to the result of
original data. Most of the changing occurs in isolated
pixels, lines and edges. Therefore classes with a lot
of I
ing.
tior
sho'
- Cla:
lati'
doul
The above
influence:
geometric
of neare;
which is *
great gee
map grid)
If bilim
fication r
those of t
Bilineai
if a doub
improverr
ing of cla
gâtions.
In the \
the influe
the purpo
of image
REFEREi'
Forster, E
Landsa
Photo.
Kahler, M
ly Pro
Mappin
1986 (ir
Realnutzi
1:5000C
Neckar
Frankfi
