PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2016
719
Distance Measure Based Change Detectors
for Polarimetric SAR Imagery
Yonghong Zhang, Hong’an Wu, Huiqin Wang, and Shanshan Jin
Abstract
Change detection based on multi-temporal
SAR
images is a fun-
damental process in many practical applications. Popular
SAR
change detectors include ratio and logarithmic-ratio (log-ratio)
operators, and those based on a statistical similarity between
temporal images. The ratio and log-ratio operators are not ideal
for polarimetric
SAR
(
POLSAR
) images, as only the intensity or
amplitude information is used. Change detectors based on
similarity comparison of probability distribution functions are
difficult to implement and not reliable because of the uncertain-
ties in estimating distribution parameters. Our research aims
to find a reliable and computationally simple change detector
from among three typical polarimetric distance measures. The
change detection potential and abilities of these distance mea-
sures are analyzed from a mathematical point of view, and then
compared through a test dataset composed of two
RADARSAT-2
fine-quad polarized images. The symmetric revised Wishart
(
SRW
) distance, originally developed for image segmentation, is
found to be an effective change detector. Based on the test data,
the change map derived from the
SRW
distance achieves 93.24
percent change rate and 5.67 percent false alarm rate. Further-
more, the eigendecompostion of the
SRW
distance is given for the
first time, which uncovers the linkage of the
SRW
distance with
the scattering mechanisms and the corresponding amplitudes
embedded in two polarimetric covariance matrices, forming a
theoretical explanation for the superiority of the
SRW
distance as
a change detector. Our research indicates the general applicabil-
ity of the
SRW
distance for
POLSAR
change detection.
Introduction
Change detection is a process that analyzes a pair of remote
sensing images acquired for the same geographical area at
different times in order to identify changes that may have
occurred between acquisition dates (Bovolo and Bruzzone,
2005). Synthetic Aperture Radar (
SAR
) images can be obtained
almost independent of atmospheric and sunlight conditions.
SAR
has been used frequently in recent decades for change de-
tection. This capability is sometimes crucial in time sensitive
applications. For example, bad weather sometimes coincides
with some emergency events such as floods, landslides, and
earthquakes. In such circumstances, it might not be possible
to obtain timely optical data and change detection from
SAR
images is the only way for timely damage assessment.
As summarized by Bovolo and Bruzzone (2005) and Bazi
et
al
. (2005), unsupervised change detection from multitemporal
SAR
images usually consists of three steps: (1) preprocessing; (2)
image comparison; and (3) image segmentation. The preprocess-
ing step includes co-registration, and speckle noise filtering. The
image comparison step generates a so-called “difference image”
using a specific change detector, and the last image segmentation
step segments the difference image into two classes: changed pix-
els and unchanged pixels using either a manual trial-and-error
procedure or various automatic techniques. This paper focuses
on the second step of unsupervised change detection; the change
detectors for multitemporal polarimetric
SAR
(
POLSAR
) images.
Several types of
SAR
change detectors have been developed
since 1990s when the
ERS-1
satellite started to continuously
capture
SAR
images. The most common
SAR
change detector
is the ratio operator (Rignot and Van Zyl, 1993; Villasenor
et al
., 1993) or equivalently the logarithmic-ratio (Log-ratio)
operator (Bazi
et al.
, 2005; Dekker, 1998). They are deemed
preferable for
SAR
intensity data because of the multiplicative
nature of speckle noise and because these approaches are
robust when calibration errors exist (Dierking and Skriver,
2002). But they are not ideal for fully polarized
SAR
data as
only the intensity information is exploited. The second type
of change detectors are based on a similarity comparison
between the probability density function (
PDF
) of the tempo-
ral images. For example, Inglada and Mercier (2007) present
the Kullback-Leibler (
KL
) divergence as a similarity measure
and change detector to derive changes in multitemporal
SAR
data by assuming that intensity follows a Gaussian
PDF
and/
or more generally a Pearson distribution. Erten
et al
. (2012)
extended this work and proposed another similarity measure,
the Mutual Information (
MI
) scalar, for temporal multichannel
SAR
images, e.g.
POLSAR
images. Mario
et al
. (2013) developed
a polarimetric change detector by testing the equi-scattering
mechanism (
ESM
) hypothesis based on geometrical perturba-
tion filters. Recently, Liu
et al
. (2014) proposed a measure to
describe the similarity of sliding windows on two multilook
POLSAR
images, assuming that both the two
POLSAR
images
follow a spherically invariant random vector (
SIRV
) distribu-
tion model. This kind of change detectors are able to effec-
tively discriminate changes between temporal
POLSAR
images.
However, difficulties exist in accurately estimating the
PDF
s,
especially in two aspects: (1) how to choose an appropriate
statistical distribution model which can fit the data well;
and (2) how to determine a homogeneous neighborhood over
which the parameters of the distribution are estimated. The
third type of change detector is specifically tailored for
POL-
SAR
data and based on statistically testing the equality of two
polarimetric covariance matrices. A good example is the test
statistic developed by Conradsen
et al
. (2003) by assuming
POLSAR
data follows the Wishart distribution. A distance mea-
sure for
POLSAR
data, the Bartlett distance, was then derived
from this test statistic and used for change detection.
In this paper, we analyze in detail the potential of three
typical polarimetric distance measures in change detection:
Yonghong Zhang, Hong’an Wu, Huiqin Wang, and Shanshan Jin
are with the Chinese Academy of Surveying and Mapping, No.28
Lianhuachi West Road, Beijing, China (
).
Ms. Huiqin Wang is with Shanghai M&D Technical Measure-
ment Co., Ltd, No. 3058 Pusan Road, Shanghai, China.
Ms. Shanshan Jin is with Geomatics Center of Zhejiang, No.83
Baochu North Road, Hangzhou, Zhejiang Province, China.
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 9, September 2016, pp. 719–727.
0099-1112/16/719–727
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.9.719
