319 
INVESTIGATING THE PERFORMANCE OF SAR POLARIMETRIC FEATURES IN 
LAND-COVER CLASSIFICATION 
Liang Gao & Yifang Ban 
Division of Geoinformatics, Royal Institute of Technology (KTH) 
Drottning Kristinas vag 30, 10044 Stockholm, Sweden 
mailto: - {liangg, yifang}@infra.kth.se 
Youth Forum 
KEY WORDS: SAR, Polarization, Multi-frequency, Land-Cover, Classification. 
ABSTRACT: 
This paper represents a study on land-cover classification using different polarimetrie SAR features. The experiment is carried out 
using C- and L-band fully polarimetrie EMISAR data acquired on July 5 and 6, 1995 over an agricultural area in Fjàrdhundra, near 
Uppsala, Sweden. The polarimetrie features investigated are coherency matrix, intensity of both C- and L-band SAR, and Cloud 
decomposition product H(l-A) of L-band, and ‘entropy’ texture of L-band HV intensity image. In order to investigate the 
performance of the different features, each feature is classified using a classifier that is best suited for the feature based on previous 
research. HI A! CL Wishart unsupervised classification is used for coherency matrix while neural network is applied to six “mean” 
texture layers of C and L bands fully polarimetrie intensity images. The best classification accuracy was achieved using the intensity 
images combined with H(l-A) and ‘entropy’ texture (overall: 81%; kappa: 0.7). The producer’s accuracy of intensity classification 
result for forest is 100.0% which reveals that the H(l-A) of L-band is a very good indicator for forest. The ’entropy’ texture of L-band 
HV intensity image has the potential to be a good indicator for road with 77.2% user accuracy, while road is not discriminated in 
coherency matrix. The results indicate that the supervised classification of the intensity of both C- and L- bands has a good potential 
for land-cover mapping in this study area. 
1. INTRODUCTION 2. POLARIMETRIC FEATURES 
Synthetic Aperture Radar (SAR) has been proven to be a 
powerful earth observation tool. The emerging Polarimetrie 
SAR (POLSAR) adds another dimension to SAR information 
content, thus makes SAR remote sensing more applicable. 
Polarimetrie SAR has been used in retrieval of soil moisture and 
surface roughness, snow and ice mapping and land-cover 
classification (Martini, 2004; Wakabayashi, 2004; T.Macri, 
2003; J.Shi, 1997). Due to its sensitivity to vegetation, its 
orientations and various land-covers, SAR polarimetry has the 
potential to become a principle mean for crop and land-cover 
classification. 
Many features such as intensities, coherency matrix, correlation 
and phase differences have been used in various classification 
experiments (Dorr, 2003; Hoekman, 2000; Lee & Grunes, 1994; 
Skriver, 2005; Alberga, 2007). As the information in the fully 
polarimetrie data can not be completely represented by one 
single feature, the combination of different polarimetrie features 
according to physical grounds and practical experiences should 
be considered. Most studies have focused on the specific 
methodology and specific polarimetrie feature, few aims at 
systematically comparing the polarimetrie features (Alberga, 
2007). Thus, research is needed to evaluate different 
polarimetrie features in a systematic manner. 
The objective of this research is to evaluate the performance of 
fully polarimetrie multi-frequency SAR features in land-cover 
classification. The investigation is carried out by classification 
of the polarimetrie features and comparing the classification 
results. Coherency matrix, intensity, Cloud decomposition 
product H(l-A)oi L-band, and ‘entropy’ texture of L-band HV 
intensity image will be evaluated and compared. 
The polarimetrie features investigated in this study are reviewed 
in the following sections. 
COVARIANCE MATRIX 
The polarimetrie SAR measures the amplitude and phase of 
backscattered signals in four combinations of the linear receive 
and transmit polarizations: HH, HV, VH and VV (H for 
horizontal and V for vertical polarization, respectively). 
EMISAR data have two available polarimetrie features: 
1) . Scattering matrix data S in slant range projection. 
2) . Covariance matrix data C in pseudo ground range. 
Since the SAR data is stained by speckles, the speckles can be 
filtered at the expense of loss of spatial resolution with 
multi-look processing. In this case, a more appropriate 
representation of S is the covariance matrix in which the 
average properties of a group of resolution cells can be 
expressed in a single matrix (Allan, 2007). It is defined as (van 
Zyl and Ulaby, 1990b): 
<c>= 
< ^hh^hh > 
< S hv Shh > 
< s„s hh > 
< ^hh^hv > < S hh S w > 
<vc> <vc> 
<sjl> <s w sr w > 
(i)