Proceedings of International Conference on Microwave - 08
Classification of Polarimetric Synthetic Aperture Radar
Images from SIR-C and ALOS PALSAR
Varsha Turkar and Y.S. Rao
Centre of Studies in Resources Engineering, lIT, Bombay, Powai, Mumbai-400 076, India
Abstract - SIR-C quad-pol MLC data and ALOS
PALSAR quad-pol and dual pol SLC data over Indian
sites have been processed using PolSARpro software for
classification of various land features. The land features
include ocean, clear water, settlements, agriculture fields,
arid lands, grown and young forest, hilly terrain,
mangrove forest, etc. The test sites used are SIR-C L-band
and C-band Kolkata city and its surroundings, ALOS
PALSAR data over West Bengal, Haryana, Rajasthan,
Uttar Pradesh, and Mumbai. For Kolkata city we observed
that classification results for L- and C-bands are slightly
different. For Mumbai dual pol data classification
accuracy is poor due to overlap of backscattering values
for bare ground and mangrove vegetation with ocean
backscattering values.
Index Terms - Radar polarimetry, synthetic aperture
radar, speckle, target decomposition, terrain classification.
I. INTRODUCTION
Classification of polarimetric SAR images has
become a very important topic after the availability
Polarimetric SAR images through ENVISAT ASAR,
ALOS PALSAR and Radarsat-2. Classification is the
task of assigning a set of given data elements to a given
set of labels or classes such that the cost of assigning the
data element to a class is minimum. Remote-sensing
research focusing on image classification is important
classification results are the basis for many
envIronmental and socioeconomic applications.
Classifying remotely sensed data into a thematic map is
very challenging because of many factors such as the
complexity of the landscape in a study area, selected
remotely sensed data, image-processing and
classification approaches may affect the success of a
classification. The major steps of image classification
may include determination of a suitable classification
system, selection of training samples, image
preprocessing and feature extraction, selection of
suitable classification approaches, post-classification
processing and accuracy assessment.
The two main techniques for image classification
are supervised and unsupervised classification
Unsupervised classification technique,
claSSIfies the image automatically by finding the clusters
based on certain criterion. On the other hand in
supervised classification technique the location and the
identity of some cover type, for example urban, forest,
and water are known before. The data is collected by a
field work, maps, and personal experience. The analyst
tries to locate these areas on the remotely sensed data.
These areas are known as "training sites". An analyst
can guide a classifier with the help of these training sites
978-1-4244-2690-4444/08/$25.00©2008 IEEE
438
to learn the relationship between the data and the
classes. manual technique of selecting training sets
could be dIfficult when ground truth is not available.
this paper, both supervised and unsupervised
technIques are used for classification. The steps
followed for the classifications are:
1. Calculation of Coherency or Covariance matrix.
2. Speckle removal with different Speckle filters.
3. Target Decomposition (using H, A and a)
4. Supervised and Unsupervised Image Classification
Techniques.
The data acquired is processed by PolSARpro
ver3.0 software for classification of various land
features. The software is freely available on the internet
developed by ESA.
II. COHERENCY AND COVARIANCE MATRICES
The reflectivity of the area being observed at a
given radar wavelength can be represented by
"scattering matrix". Each of the four complex elements
of this matrix is the amplitude and phase of the
backscattered radiation as measured at one of four
orthogonal transmit/receive polarizations: HH, HV, VH,
andVV.
There are different target polarimetric descriptors:
Sinclair, Vectors, Kennaugh, Muller, Coherency
and Covanance. We have used Coherency matrix for
and Covariance matrix for SIR-C data. By using
dIfferent basis (Lexiogrphic or Pauli) we can get two
types of target vectors. The scattering vector or
covariance vector is a vectorized version of the
scattering matrix. It is easy to construct a power domain
representation of the scattering properties, which is done
by forming the product of this vector itself. The result is
a covar!ance matrix which fully describes the scattering
propertIes of the target. The coherency matrix is similar
to the covariance matrix. To obtain the coherency matrix
the matrix is vectorized in different way using
the Pault spIn elements. The eigenvalues of covariance
and the coherency matrix are real and are the same. The
sum of the diagonal elements (trace) of both matrices is
also the same and represents the total power of the
The .first diagonal element of coherency
matnx gIves the Information about single bounce
scattering, the second diagonal element gives the
information about double bounce scattering and the third
diagonal element gives the information about volume
scattering. In order to get Pauli RGB image which is
very much used, these diagonal elements are given blue,
red and green colour respectively. Both covariance and