4 INVESTIGATIONS
The investigations are carried out on the image processing
system of the Department of Photogrammetry and Carto
graphy of the Technical University of Berlin, where soft-
ware-packages for radiometric and geometric image pro
cessing as well as for multispectral classification are
implemented.
For the investigations a part of the Landsat-scene
210/26 from 9. August 1975 is used. This part covers a
region of Baden-Württemberg, with the Oberrheingraben
and the cities Mannheim and Karlsruhe. The seven most
frequented classes in the concerned area should be dis
tinguished with the conventional maximum-likelihood
classifier.
The following case-studies, which are also presented in
Figures 1-4, are carried out:
Comparison of classification results of original and
contrast enhanced data
- Comparison of classification results of original data,
rectified and resampled with the methods of nearest
neighbour and bilinear interpolation
Comparison of classification results of contrast
enhanced data, rectified and resampled with the
methods of nearest neighbour and bilinear inter
polation
Comparison of classification results of contrast
enhanced data; on one hand, row- and column-doubled
before rectification and resampling (bilinear inter
polation), and on the other hand without doubling
before geometric processing
ence will be expected between the classification results,
because the histogram stretching affects only a linear
extended distribution of the grey values, without any loss
of information.
Original Contrast En-
Class
Data
hanced
1 Water
1.13
1.13
2 Coniferous F.
20.40
20.38
3 Deciduous F.
12.99
12.84
4 Urban areas
8.44
8.69
5 Agriculture
37.45
36.35
6 Grassland
8.77
8.73
7 Vine
10.82
11.88
Table la. Classified pixels in each class in (%)
In Table la it can be seen, that there are only few
differences in the distribution of classes in both images.
Table lb is much more distinct for a comparison of
classification results than Table la, because the exchange
of pixels from one class to another is pointed out. In this
Table the classification result of original data was used as
reference image. The percentages in the lines show the
class changing of pixels in the classification result of the
contrast enhanced data.
Data Contrast Enhanced
Type
For all rectifications the same transformation polynomials
were used. The coefficients were evaluated considering
the coordinates of nine control points. In all cases the
training areas for the classes were defined in a colour
composite of the contrast enhanced data. Subsequently the
edge coordinates of the training areas, which have been
obtained during definition, were calculated for the respec
tive preprocessed data. These new coordinates could be
used to obtain the features (grey-values) of the rectified
and resampled data. This method guarantees comparison of
the results, because all classifications use statistics de
rived from the same training areas. However the sites
differ through their features and/or the number of features
when contrast enhancement, rectification and resampling
is carried out.
Class
1
2
3
4
5
6
7
1
99.49
0.13
_
0.35
_
0.02
0.01
2
0.01
99.70
0.12
0.06
-
-
0.11
3
-
0.29
98.44
-
-
0.09
1.17
4
0.01
-
-
99.39
0.09
-
0.51
5
-
-
-
0.72
96.95
0.50
1.83
6
-
-
0.32
-
0.43
97.25
2.01
7
-
-
-
0.12
-
-
99.88
Sum of pixels : 1048567
Percentage of pixels classified the same : 98.28 %
Table lb. Pixel by pixel comparison of classification results
of original and contrast enhanced data
4.1 Classification Results of Original and Contrast
Enhanced Data
This case-study compares the classification results of orig
inal and contrast enhanced data (Figure 1). For contrast
enhancement of the used part of the scene a linear
histogram stretching was carried out for each channel
separately. The result is more appropriate for visual inter
pretation particulary for defining training areas. No differ-
Figure 1. Comparison of classification results of radiome-
trically preprocessed data
In Table lb the greatest differences occur in the classes
deciduous forest, agriculture and grassland. The mis-
classified pixels are assigned mainly to vine. Visual inter
pretation of the investigated part of the scene shows, that
misclassifications appear in areas, which are highly struc
tured. The creation of a difference image confirms the
visual comparison of both classification results. In homoge-
nious regions like those covered with coniferous forest and
in greater parts of decidous forest and agriculture no
differences can be detected.
The reasons for misclassification can be explained by the
spectral reflectance of the special classes and the contrast
enhancement. It is very difficult to define training areas in
highly structured regions. Such training sites may also
contain grey-values, which are not characteristic to the
considered classes. So the variance of such classes is
relatively high. The contrast enhancement will affect more
in increasing of the variance of highly structured classes,
than that of homogenous classes. Another reason can be
explained by the histogram stretching for every channel
separately. This influences each channel differently, which
results in different covariances for original and contrast
enhanced data. By using a conventional maximum-likeli
hood classifier these facts effect a change of assigning
pixels to such classes.
By means of these facts we can conclude, that differ
ences between classification of original and contrast
enhanced data exist, but they are usually small. The
percentage of pixels, which are assigned to the same
classes in both images, amount to about 98 %.
