ing of
ngs "
p. 53-
design
e and
126.
f SAR
ematic
1985,
or the
ational
nes, a
lution,
nthetic
nsing,
CLASSIFICATION AND SEGMENTATION FOR RADAR IMAGERY
USING GAUSSIAN MAKOV RANDOM MODEL*
Yunhan Dong!, Bruce Forster! and Anthony Milne?
School of Geomatic Engineering, The University of New South Wales, Sydney 2052, Australia
Telephone: +61 2 9385 4209, +61 2 9385 4172, Facsimile: +61 2 9313 7493
Email: y.dong@unsw.edu.au, b.forster@unsw.edu.au
20ffice of Postgraduate Studies, The University of New South Wales, Sydney 2052, Australia
Telephone: +61 2 9385 2731, Facsimile: +61 2 9385 3733, Email: t.milne@unsw.edu.au
ISPRS, Budapest 1998
Commission VII, Working Group 6
KEY WORDS: Radar Images, Segmentation, Classification, Markov Random Field Model.
ABSTRACT:
Classification of synthetic aperture radar (SAR) images based on the information provided by individual
pixels cannot generally give satisfactory results because the images are highly speckled due to coherent
processing. In order to obtain reliable and accurate results, the classification must be based on the
statistics of clusters rather than individual pixels. This paper approaches SAR image classification via
two steps: 1) to segment an image into uniform areas (clusters); and 2) to classify the segmented clusters.
Since the classification is based on the segmented imagery, segments (a group of uniform pixels) provide
reliable statistics such as mean, standard deviation and texture characteristics. It can, therefore, be
expected that the accuracy of the classification is greatly improved.
1. INTRODUCTION
Since spaceborne SAR systems became opera-
tional, there have been demands for classifying
and monitoring geophysical parameters both lo-
cally and globally using remtely sensed SAR im-
age data. Excellent work has been done on clas-
sifications using SAR images for both natural
and built targets (Van Zyl 1989, Freeman and
Durden 1992, Pierce et al 1994, Dobson et al
1995, 1996, Touzi et al 1992, Nezry et al 1996,
Dong et al 1996). However, the classification of
SAR images based on the information provided
by individual pixels cannot generally give sat-
isfactory results because the images are highly
speckled due to coherent processing. In order to
increase the accuracy of classification, Pierce et
al (1994) used information averaged from multi-
pixels (2 x 2 pixels, for example), whereas Dob-
son et al (1996) performed segmentation before
classification. Segmentation segments an image
into disjoined regions corresponding to objects,
or parts of objects that differ from their sur-
roundings, and thus enables further classification
to be performed based on the information pro-
*This work was supported by the Australian Research
Council.
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
vided by clusters rather than individual pixels.
Various methods for image segmentation have
been developed based on textures, statistical val-
ues, such as the mean and the standard deviation
and covariances (Lee and Jurkevich 1989, Rig-
. not and Chellappa 1992, Serpico and Roli 1995,
Baraldi and Parmiggiani 1996, Derin et al 1990,
and Geman and Geman 1984). The Gaussian
Markov random field model (GMRF) has been
studied extensively and shown to be an accu-
rate compact representation for a range of im-
ages (Panjwani and Healey 1995). Segmentation
using the GMRF model segments objects based
on the region information of up to the second
order statistics and the spatial relationships. In
other words, the GMRF model considers two re-
gions to be different if one or more than one of
the following conditions is true:
e The first order statistics (the means for a
single channel image or the mean vectors
for a multi-channel image) are different;
* The second order statistics (the variances
for a single channel image or the covariance
matrices for a multi-channel image) are dif-
ferent;
517
