C
le
le
S,
at
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004
the comparison of the two ‘high-pass’ images a histogram
matching of the pan-sharpened image to the panchromatic
orthoimage was indispensable.
In Table 6 the correlation coefficients between the
panchromatic band the pan-sharpened bands are presented.
Given that the ideal value is 1, it can be noted that the pan-
sharpened image preserves the characteristics of the initial high-
resolution image it came from.
Correlation coefficient between ‘high-pass’
panchromatic — ‘high-pass’ pan-sharpened image
Pan-sharpened BANDS
/ 2 3 4
Panchromatic BAND 0.99 0.99 0.99 0.99
Table 6. Criteria on the spatial evaluation of the pan-sharpened
image
5. CLASSIFICATION
Having a pan-sharpened image of good spectral and spatial
quality enables us to derive reliable thematic information
through the process of classification. Classification of urban
environment plays a key role in Urban Area Land-use Mapping,
Urban Planning and Management, Establishment and Revision
of GIS Database, Environment and Disaster Monitoring and
Establishment of Telecommunication Network Station (Yu et
al., 2002). With the development of high-resolution satellites it
is now possible to collect and map thematic information from
images of urban areas on large scales. However, high-resolution
data does not mean high classification accuracy. In such cases,
classification is more difficult due to the heterogeneity of urban
Structures.
5.1 The procedure of classification
At first, a supervised classification with the method of
maximum-likelihood classification was applied to the pan-
sharpened image. Although a careful and detailed training was
carried out, problems arose during the evaluation of the
samples. Specifically, in the resulting classified image some
streets were mixed up with buildings and territories of bare soil
were misclassified as burnt forest.
For this reason a different approach was decided. In urban areas
roofs, streets and pavements are built of a similar material, so
their similar reflectance creates problems during the
classification process. The pan-sharpened image was separated
to the urban part and the hilly tree-covered part and the
classification was applied to these two image parts.
Firstly, the set of classes was selected for each part and consists
of streets, roofs, tiles, bare soil, low vegetation, trees and
shadow for the urban part and trees, burnt forest, streets and
rural streets for the hilly part. After the selection of the samples
for each class, they were assessed through their histograms and
the diagrams of their means and found to be representative and
satisfactory.
A fuzzy classification was introduced for both image parts and
a distance file was created. The fuzzy approach was preferred
because it allows more information on the partial class
membership to be made available. The classified images were
189
chosen to have two layers of which Layer 1 contains class
values for the best classification, Layer 2 for the second best.
The pixels of the distance file consist of the values of the
distance of the class means. With the help of the classified
images and the distance files the Fuzzy Convolution operator
creates a single classification layer by calculating the total
weighted inverse distance of all the classes in a window of
pixels, in this case a 3 X 3. Then it assigns the center pixel in the
class with the largest total weighted inverse distance over both
fuzzy classification layers. This reduces the phenomenon of
'salt and pepper' in the classified image which the mixed pixels
produce (Erdas, 1999). The final classified image is presented
in Figure 7.
Classes
Low vegetation
Streets
Bare soil
Trees
Burnt forest
Roofs ’
Tiles
Shadows
Rural streets
MILITE D
Figure 7. The classified image resulted from the application of
the supervised classification
5.2 Accuracy assessment of the classification
The procedure of classification is completed when its accuracy
is estimated. The error of the classification, which expresses the
misclassification of some pixels, determines the degree of
success of the procedure. For this reason a comparison is made
between classified data and ground reference data and the error
matrix, the accuracy report and the Kappa coefficient are
calculated. The results are presented in Table 4.
ACCURACY RESULTS FOR THE URBAN IMAGE PART
ERROR MATRIX
Classes |Low veg.|Streets|Bare soil| Trees | Shadows| Roofs | Tiles
Low veg. 0 | | 0 I 0 0
Streets 7 8 0 0 ] | 0
Bare soil 0 0 8 0 0 | l
Trees 0 0 0 10 0 0 0
Shadows 0 0 0 0 10 0 0
Roofs 0 2 0 0 2 6 0
Tiles 0 0 | 0 1 0 8
Toll test ÿ nl. 10 15 8 9
pixels
Overall Accuracy = 84.43% Kappa coefficient = 78.33%
ACCURACY RESULTS FOR THE HILLY IMAGE PART
ERROR MATRIX
Classes Trees Burnt forest | Streets Rural streets
Trees 2 0 0
Burnt forest 0 8 0 0
Streets 1 0 5 2
Rural streets 0 | 0
Total test pixels 7 11 5 9
Overall Accuracy = 81.25% Kappa coefficient = 75.00%
