International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
color. In the forest signature, colors are equally presented, i.e.
none of the scattering mechanisms which are detected by the
Pauli method, dominates. In the vegetation signature, the blue
colour dominates. This means that the surface scattering type is
the most presented. The runway signature is presented by dark
tones. This means that none of the scattering type dominates.
4. CLASSIFICATION
4.1 Classification based on original polarimetric data
Based on the absolute value of the Sy, and S,, and (Sy, + Syn)/2
of the polarimetric components, we applied the supervised
Maximum Likelihood classifier in order to classify the study
area into four classes: urban, forest, vegetation, and runways.
Classification was applied a) on raw data and b) on data without
speckle, generated by the application of the Lee filter. In order
to investigate the most appropriate window size, we first
applied the Lee filter 9 times, each time increasing the size of
the widow by a step equal to 2. The windows used were 3x3,
5x5, ... 17x17. Then classifications were performed, and their
accuracy was calculated based on a set of test areas. The total
accuracy of the classification applied on the raw data is 83.60%,
and the results obtained after the Lee filter application are given
in figure 8.
91
90
88
87
86
85
% Toatal accuracy
3 5 7 9 11 13 15 17
The window size
Figure 9. % Total accuracy of the polarimetric data based
classification as a function of the size of the Lee filter
We observe that the highest accuracy, 91%, is obtained for a
window size 15x15. In this classification, the accuracy obtained
for the class of the: urban area is 45,84%, forest 94.72%,
vegetation 95.43%, and runway 99.32%. The urban class is
most confused with the forest. Although the polarization
signatures of the two categories are quite different (figure 3 and
4), due to the fact that Bayes classifier is based on the
magnitudes only, the urban and forest classes are confused. For
the other classes, the magnitude based classification produces
high accuracies.
4.2 Classification based on Pauli decomposition analysis
Bayes classifier was also applied on the absolute values of the
three Pauli decomposition components, by using the same
training set as in the previous classification. Classification was
applied a) on the initial Pauli components, and b) on the Pauli
components after the application of the Lee filter. The most
appropriate window size was investigated by the method
described in the previous section. Accuracy was tested by the
same test set as in the previous classification. The total
accuracy of the classification applied on the initial Pauli
components is 80.00%, and the results obtained after the Lee
filter application are given in figure 9.
% Total acaracy
œ c
o o
3 5 7 9 11 13 15 17 19 21 23 25
Me window size
Figure 9. % Total accuracy of the Pauli based classification as
a function of the size of the Lee filter
We observe that the highest accuracy, 91.4196, is obtained for a
window size 23x23. In this classification, the accuracy obtained
for the class of the: urban area is 81,0896, forest 78.06%,
vegetation 97.3596, and runway 99.0195. The urban and forest
classes are confused. Although the accuracy for the urban class
is significantly increased in comparison to the classification
based on original polarimetric data, the accuracy of the forest
class is reduced. The scattering mechanisms that Pauli
decomposition analyzes, and especially the even bounce and
45? titled even bounce mechanisms, reinforce the discrimination
of the urban class. On the other hand, the weakness of the Pauli
decomposition in analyzing the volume scattering mechanism
(based on dipoles) affects the accuracy for the forest class,
which is analyzed on the basis of dihedrals. For the other
classes, the Pauli based classification produces high accuracies.
4.3 Classification based on original and Pauli
decomposition analysis
To reduce confusion between urban and forest areas, Bayes
classifier was applied on the absolute values of the a)
polarimetric data, and b) data generated by Pauli decomposition
analysis. In this way, odd and even bounce scattering
mechanisms, as well as, original full polarimetric information
participate in a magnitude based classification. Classification
was applied a) on the initial data, and b) on data after the
application of the Lee filter. The most appropriate window size
was investigated by the method described in section 4.l.
Accuracy was tested by the same test set as in the previous
classification. The total accuracy of the classification applied on
the initial data is 82.14%, and the results obtained after the Lee
filter application are given in figure 10.
% Total accuracy
82
3 5 7 9 11 13 15 17 19°: 21 23
The window size
Figure 10. % Total accuracy of the classification as a function
of the size of the Lee filter
We observe that the highest accuracy, 91.53%, is obtained for a
window size 23x23. In this classification the accuracy obtained
for the class of the: urban area is 81,65%, forest 78.3 1%,
vegetation 97.20%, and runway 99.25%. The accuracy of the
urban class is slightly increased although urban and forest class
284
Internatioi
HE ads
are still cc
18.1996 of
account fo
The poten
different |
processing
signatures
interprete
Polarimetr
this purps
recognized
mechanisn
samples ol
given a po
behaviour
difficult tc
centres thé
scattering
make it d
characteris!
The non «
indicated E
surface ar
determine.
recognized
ability to it
which pre:
scattering «
after their a
Regarding
filter whicl
based metl
determinist
window ha.
original dat
images ger
documents
filler windc
be achieved
To evaluat
polarimetric
applied on :
the Pauli d
data. In all
accuracies (
The very hi
this. Howe
mediocre
pronouncing
data to disci
by Pauli d
classificatio
accuracy eq
C, respective
the forest cl
well discrin
urban class |
by the Paul
Likelihood
