IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, SEP 2016 1
Clustering-based Geometrical Structure Retrieval of
Man-made Target in SAR Images
Jiani Wu, Yongguang Chen, Dahai Dai, Member, IEEE, Siwei Chen, Member, IEEE,
and Xuesong Wang, Member, IEEE
Abstract—In Synthetic Aperture Radar (SAR) images, scatter-
ing centers from the same geometric structure of the man-made
target usually have the same scattering type and similar coor-
dinates. Inspired by this observation, a novel Clustering based
Geometrical Structure Retrieval (C-GSR) method is proposed
to estimate the geometrical structure of targets by clustering
scattering centers according to their types and coordinates. The
C-GSR method considers each peak in a SAR image as a single
scattering center and extracts both frequency and polarization
features for scattering center classification. Then, scattering
centers are efficiently clustered using the Density-Distance (DD)
based clustering algorithm. Finally, the geometrical structure
corresponding to each canonical scatterer can be retrieved by
computing the coordinates of scattering centers associated to the
corresponding cluster. Experiment results have demonstrated the
feasibility and accuracy of the proposed C-GSR method.
Index Terms—Synthetic Aperture Radar (SAR) images, ge-
ometrical structure retrieval, scattering center classification,
clustering algorithm.
I. INTRODUCTION
I
T has been recognized that the profile of targets can be
commonly represented by a combination of well-known
canonical scatterers such as plate, sphere, cylinder, dihedral
and so on, which makes great convenience for target inter-
pretation as well as target recognition. Thus, retrieving the
geometrical structure, such as the type, location and size of
the canonical scatterer of targets, is a key problem in target
interpretation. At sufficiently high frequencies, the scattering
of a target can be approximated by a sum of responses of
several individual components, termed as scattering centers
of the target. The typical Attributed Scattering Center (ASC)
model [1], which considers the relationship between the am-
plitude and frequency as well as aspect of scattering centers,
characterizes each scattering center by a parameters set of
location, amplitude, shape, orientation, etc. Therefore, ASC
model provides a physical description of the target and enables
the ASC model based Geometrical Structure Retrieval (A-
GSR) by attribute parameter estimation. Extensive researches
have been conducted to investigate ASC model based param-
eter estimation and geometrical structure retrieval [2]-[7].
However, there are two serious problems in practical imple-
mentations of the A-GSR. First, it is difficult to distinguish be-
tween distributed scattering centers and local scattering centers
J. Wu, D. Dai, S. Chen, and X. Wang are with State Key Laboratory of
Complex Electromagnetic Environment Effects on Electronics and Informa-
tion System, National University of Defense Technology, Changsha 410073,
China (email: wujiani06@nudt.edu.cn).
Y. Chen is with Beijing Institute of Tracking and Telecommunications
Technology, Beijing 100094, China.
in SAR images. In ASC model, local scattering center, such
as vertical dihedral (i.e., the dihedral with a fold-line oriented
vertical to the imaging plane), vertical cylinder, trihedral, top
hat, and sphere, has a slowly varying amplitude response to
changes in azimuth. Distributed canonical scatterer, such as
horizontal dihedral (i.e., the dihedral with a fold-line oriented
parallel to the imaging plane), horizontal cylinder and plate,
has a rapidly varying sinclike amplitude response in azimuth
[1]. The local scattering center appears as a single peak while
distributed scattering center often appears as a set of in-line
peaks in SAR image. Multiple adjacent local scattering centers
can produce a SAR image with multiple adjacent peaks, which
is similar to that produced by a single distributed scattering
center [6]. Therefore, it is difficult to distinguish between
these two cases. Furthermore, it becomes even challenging
to select the proper model order, which denotes the number
of scattering centers. Second, with the increasing dimensions
of the parameters, the computation complexity for attribute
parameter estimation increases exponentially, while the esti-
mation accuracy decreases due to the coupling among multiple
parameters [8]. These two problems make it highly challenging
for A-GSR to retrieve the geometrical structure of man-made
target efficiently and accurately.
In order to overcome aforementioned problems involved
in A-GSR, this letter proposes a novel Clustering based
Geometrical Structure Retrieval (C-GSR) method for man-
made target in SAR images. First, C-GSR considers each peak
in the SAR image as a single scattering center and classifies
each scattering center based on the extracted frequency and
polarization features. Then, a Density-Distance (DD) based
clustering algorithm [9] is used considering both the class
types and the coordinates of the scattering centers. Specifically,
the adjacent scattering centers with the same class type are
considered to be produced by the same canonical scatterer
and thus they are clustered together. Finally, based on the
coordinates of scattering centers within the same cluster, the
geometrical structure of the target is retrieved. The main
contributions of this paper are two-fold:
1) Each peak in a SAR image is considered as a single
scattering center and the frequency and polarization features
are extracted for scattering center classification based on the
peak intensity. It avoids the judgment of the local and distribut-
ed scattering centers and reduces computation complexity of
attribute parameter estimation.
2) The scattering centers are mapped into the classification-
coordinate space and clustered by using the DD based clus-
tering algorithm. The geometrical structure can be retrieved