LAND COVER CLASSIFICATION USING E-SAR POLARIMETRIC DATA 
V. Karathanassi”, M. Dabboor 
Laboratory of Remote Sensing, School of Rural and Surveying Engineering, National Technical University of Athens, 
Heroon Polytechniou 9, Zographos, 15780, Greece 
(karathan*,md2003)@(survey).ntua.gr 
Commission VII, WG VII/3 
KEY WORDS: Land cover, Polarization, Detection, SAR, Accuracy, Theory, Test 
ABSTRACT: 
Different decomposition approaches have been proposed in order to analyse and interpret SAR polarimetric images. These are based 
cither on the complex voltage reflection matrix, like Pauli, or on power reflection matrix, like the covariance or coherency matrix. 
They produce polarimetric parameters which are appropriate to retrieve information on the scattering process of the target. If the 
target is distributed, polarimetric parameters are affected by speckle. 
The objectives of this work are to search and point out the parameters most appropriate for the interpretation of different land uses in 
a ESAR image; to evaluate Maximum Likelihood (ML) classification results produced by two different polarimetric input sets: the 
full polarimetric, and the Pauli images; to investigate the most appropriate size of the Lee filter window for polarimetric speckle 
reduction. 
Based on the full polarimetric L-band, polarization signatures were extracted and analyzed for four land cover classes: urban, forest, 
vegetation and smooth surfaces. The scattering mechanism of these land cover classes was also analysed based on the images 
generated by Pauli decomposition analysis. The Maximum Likelihood classification was performed on the “magnitude content" of 
the a) original polarimetric data, b) images produced by the Pauli analysis, and c) both previous cases data. The accuracy of each 
class confirmed the contribution of polarimetric data and Pauli parameters in the interpretation of the scattering mechanism. 
To reduce speckle effects and improve classification results, the Lee filter was applied on the above images several times, each time 
increasing the size of the moving window. The ML classification was performed on the despeckled images. Classification accuracy 
pointed out the most appropriate size of the filter window for speckle reduction. 
1. INTRODUCTION matrix and measures an appropriate distance, d, according to 
maximum likelihood classification (Lee, Grunes, Kwok, 1994), 
Full polarimetric data can define the scattering behaviour of has been investigated for supervised land use/cover 
land use/cover, thus giving better land use/cover classification classification (Lee, Liew, Kwoh, Nakayama, 2001). An 
results than single-channel SAR (Smith, Broek, Dekker, 1998). unsupervised classification method based on the Wishart 
Several parameters have been proposed to assist the classifier was also developed (Scheuchl, Caves, Cumming, 
interpretation and the classification of polarimetric SAR data. Staples, 2001), as well as, a weighed Wishard classifier 
These parameters are deduced from the decomposition of either according to which each polarimetric component is weighed 
the complex voltage reflection matrix or the power reflection relative to the absolute amplitude of the measurements (Smith, 
matrices. The first. decomposition category deals with the Broek, Dekker, 1998). 
sphere/deplane/helix decomposition and the Pauli 
decomposition. The latter category can be divided into the Focusing on the evaluation of the polarimetric information 
Muller/Kennaugh matrix and covariance or coherency matrix deduced by the simplest processing methods, the objectives of 
(Hellmann, 1999). Among them the basic Cloude-Pottier this work are to: 
parameters, entropy and alpha, (Cloude, Pottier, 1997) deduced 1. Interpret land use/cover scattering behaviour by 
from the Cloude decomposition theorem applied on the analysing the a) polarization signatures regarding to 
coherency matrix are the most investigated (Hellmann, the ellipticity angle, orientation and intensity, and b) 
Kratzschmar, 1998; Titin-Schnaider, 1999; Scheuchl, Caves, signatures extracted by the Pauli decomposition 
Cumming, Staples, 2001) for land use/cover interpretation. The method 
number of these parameters was farther increased by the 2. Classify land use/cover of the test area based on the 
addition of two polarizing parameters, the propagation and magnitude content of a) the original full polarimetric 
helicity phase angles and three depolarising parameters, the data, b) the data produced by the Pauli decomposition 
anisotropy A and two depolarising eigenvector angles (Cloude, method, and c) both previous cases data. 
Potier, Boerner, 2002). Parameters deduced by the span 3. Define the most appropriate size of the Lee filter 
normalisation of the Mueller matrix have also been investigated window, applied for speckle reduction. 
in order to retrieve scattering electromagnetic mechanisms 
(Titin-Schnaider, 1999) and interpret polarimetric data. The full polarimetric airborne data sets were acquired with 
DLR's Experimental SAR (E-SAR). The test region is the arca 
Based on the coherency matrix, the complex Wishart classifier, of Oberpfaffenhofen, Germany. For the study needs, the L band 
which uses the complex Wishart distribution of the coherency was used which has a resolution of 3 meters. 
280 
  
Internati 
lel vir, 
2. PC 
2.1 Wa 
For a ple 
the elect 
propagati 
called po 
system. 1 
E withir 
an ellipse 
elliplicall 
the desci 
reference 
with full 
linear po 
introduce 
plane trar 
horizonta 
field. The 
function ¢ 
The vecto 
componer 
the propag 
These cor 
amplitude 
where | 
also be wr 
E,(F,1)= Re| 
=d4, COS(T +6 
where 7 
phase diffi 
results that 
E (Ft) > 
d = 
=cos(r +0, 
Substitutin,