2004
lon as
| for a
tained
06%,
forest
| class
cation
forest
Pauli
e and
nation
: Pauli
anism
class,
other
acles.
Pauli
Bayes
he a)
"sition
tering
nation
cation
er the
W Size
n 4.1.
evious
ied on
ne Lee
inction
d for a
tained
8.31%,
of the
st class
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
are still confused. Urban pixels classified as forest account for
18.19% of the urban sample. Forest pixels classified as urban
account for 21.69% of the forest sample.
S. CONCLUSIONS
The potential of the polarimetric data to discriminate among
different land uses was investigated in this study. Simple
processing methods, such as extraction of polarimetric
signatures and Pauli decomposition analysis were used to
interprete the scattering mechanisms of each land use.
Polarimetric signatures were proved to be a powerful tool for
this purpose. Several scattering mechanisms have been
recognized on each signature. The large number of scattering
mechanisms produced by non-deterministic targets, such as
samples of land use, was their main disadvantage. Although
given a polarimetric signature we can recognize the scattering
behaviour of the specific land use, the opposite is sometimes
difficult to achieve. Due to the large number of scattering
centres that a land use sample includes, several secondary
scattering mechanisms are simultaneously produced. This can
make it difficult for the interpreter to distinguish the most
characteristic mechanism for the sample under interpretation.
The non deterministic nature of land use targets was also
indicated by the Pauli decomposition analysis. Dihedrals and
surface are the scatterers that Pauli decomposition can
determine. The limited number of scattering mechanisms
recognized by the Pauli analysis is its main disadvantage. Its
ability to indicate dihedrals is its main advantage. Urban areas
which present a complexity by including several different
scattering centres in the same target can easily be interpreted
after their analysis in dihedrals.
Regarding the definition of the size of the window of the Lee
filter which is used for speckle suppression, a classification
based method was developed and applied. Due to the non
deterministic approach of the scatterers, the optimum size of the
window has been found to be quite large, 17x17 pixels for the
original data and 23x23 pixels for the data sets which include
images generated by the Pauli decomposition method. This
documents the non deterministic nature of land uses, as a large
filler window is necessary for high classification accuracies to
be achieved.
To evaluate classifications based on the magnitude of the
polarimetric data, the maximum Likelihood classifier was
applied on a) the full polarimetric data, b) the data produced by
the Pauli decomposition method, and c) both previous cases
data. In all cases speckle suppression ^ preceded. The total
accuracies obtained in all cases were satisfactory (91-91.53%).
The very high resolution of the E-SAR data also contributed to
this. However, the accuracy obtained for the urban class was
mediocre (45.8496) in the first classification results,
pronouncing the weakness of the magnitude of the polarimetric
data to discriminate urban areas from forest. The data generated
by Pauli decomposition contributed to the improvement of
classification results regarding the urban class, producing an
accuracy equal to 81.0896 and 81.65% for classifications b and
6, respectively. On the other hand, they reduced the accuracy of
the forest class by approximately 15%, (78.31%). Dihedrals are
well discriminated by the Pauli analysis and characterize the
urban class more than other classes. When the images generated
by the Pauli decomposition are introduced in the Maximium
Likelihood classification contribute significantly to the
285
statistical definition of this category but they produce
misclassifications to the other categories which also contain
dihedrals, such as the forest.
The key subject for further work will be the introduction of the
phase of the full polarimetric data in the classification
algorithm.
REFERENCES
Cloude, S.R., Pottier, E., Boerner W., 1997. Unsupervised
Image Classification using the Entropy/Alpha/Anisotropy
Method in Radar Polarimetry. In AIRSAR Earth Science and
applications Workshop, March 1997.
Cloude, S.R., Pottier, E., 1997, An entropy based classification
scheme for land applications of polarimetric SAR. JEEE
Transactions on Geoscience and Remote Sensing, Vol 35, n°1,
pp. 68-78.
Hellmann, M., 1999, Classification of Full Polarimetric SAR-
Data for Cartographic Application. Thesis, Technical
University of Dresden, August 1999.
Hellmann, M., Kratzschmar, E., 1998. Interpretation of SAR-
Data using Polarimetric Techniques. In: The retrieval of Bio-
and Geo-physical parameters from SAR Data for land
application Workshop, ESTEC, 21-23 October 1998.
Lee, K.Y., Liew, S.C., Kwoh, L.K., and Nakayama, M., 2001.
Land Cover Classification Using NASA/JPL Polarimetric
Synthetic Aperture Radar (POLSAR) Data. In the 22" Asian
Conference on Remote Sensing, Singapore, 5-9 November
2001.
Lee, J.S., Grunes, M., Kwok, R., 1994. Classification of multi-
look polarimetric SAR imagery based on complex Wishart
distribution. /nternational Journal of Remote Sensing, Vol. 15,
no 11, pp. 2299-2311.
Scheuchl, B., Caves, R., Cumming, I., and Staples, G., 2001.
Automated Ice Classification Using Spaceborne Polarimetric
SAR Data. In: /GARSS Proceedings, Australia.
Smith, A. J., van den Broek, A. C., and Dekker R. J., 1998.
Landuse Classification Using PHARUS Polarimetric Radar. In
Proceedings of the 2nd International Workshop on Retrieval of
Bio- & Geo-physical Parameters from SAR Data for Land
Applications, 21-23 October 1998, Netherlands.
Titin-Schnaider, C., 1999, Radar polarimetry for vegetation
observation. In: CEOS SAR Workshop, Netherlands, 26-29
October 1999.
ACKNOWLEDGEMENTS
Authors are grateful to Dr J. Moreira (DLR) who kindly
provided the E-SARpolarimetric data.