See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228397725

SAR image simulation with application to target

recognition

Article in Progress In Electromagnetics Research · January 2011

DOI: 10.2528/PIER11061507

CITATIONS READS

70 2,626

3 authors, including:

Yang-Lang Chang Kun-Shan Chen

National Taipei University of Technology Chinese Academy of Sciences
160 PUBLICATIONS 1,815 CITATIONS 301 PUBLICATIONS 2,887 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Kun-Shan Chen on 25 November 2015.

The user has requested enhancement of the downloaded file.

Progress In Electromagnetics Research, Vol. 119, 35–57, 2011

SAR IMAGE SIMULATION WITH APPLICATION TO

TARGET RECOGNITION

Y.-L. Chang1 , C.-Y. Chiang2 , and K.-S. Chen3, *

1 Department of Electrical Engineering, National Taipei University of
Technology, Taipei, Taiwan
2 Department of Computer Science and Information Engineering,
National Central University, Jhongli, Taiwan
3 Center for Space and Remote Sensing Research, National Central
University, Jhongli, Taiwan

Abstract—This paper presents a novel synthetic aperture radar

(SAR) image simulation approach to target recognition, which consists
of two frameworks, referred to as the satellite SAR images simulation
and the target recognition and identification. The images simulation
makes use of the sensor and target geo-location relative to the Earth,
movement of SAR sensor, SAR system parameters, radiometric and
geometric characteristics of the target, and target radar cross section
(RCS), orbital parameters estimation, SAR echo signal generation and
image focusing to build SAR image database. A hybrid algorithm
that combines the physical optics, physical diffraction theory, and
shooting and bouncing rays was used to compute the RCS of complex
radar targets. Such database is vital for aided target recognition
and identification system Followed by reformulating the projection
kernel in an optimization equation form, the target’s reflectivity field
can be accurately estimated. Accordingly, the target’s features can
be effectively enhanced and extracted, and the dominant scattering
centers are well separated. Experimental results demonstrate that the
simulated database developed in this paper is well suited for target
recognition. Performance is extensively tested and evaluated from real
images by Radarsat-2 and TerraSAR-X. Effectiveness and efficiency of
the proposed method are further confirmed.

Received 15 June 2011, Accepted 5 July 2011, Scheduled 19 July 2011

* Corresponding author: Kun-Shan Chen ([email protected]).
36 Chang, Chiang, and Chen

1. INTRODUCTION

SAR (synthetic aperture radar) image understanding and interpreta-

tion is essential for remote sensing of earth environment and target
detection [1–8]. In the development of aided target recognition and
identification system SAR image database with rich information con-
tent plays important role. It is well recognized that these image base
volumes are enormous and it is impractical to acquire it solely by in-
situ measurements. Hence, SAR simulation is one potential alternative
to alleviating the problem. However, it is highly desirable to develop
a full blown SAR image simulation scheme with high verisimilitude
including the sensor and target geo-location relative to the Earth,
movement of SAR sensor, SAR system parameters, radiometric and
geometric characteristics of the target, and environment clutter The
simulation should at least include the computation of target radar
cross section (RCS), orbital parameters estimation, SAR echo signal
generation, image focusing, and so on. Such simulation scheme is also
well suited for satellite SAR mission planning. As an application ex-
ample of the simulated SAR images, the second part of this paper
is to deal with target recognition. Before doing so, feature enhance-
ment and extraction is presented. It is well known that good feature
enhancement is essential for target identification and recognition of
SAR images [9, 10]. Novel algorithm proposed for spotlight mode SAR
involves the formation of a cost function containing non-quadratic reg-
ularization constraints [11, 12]. In stripmap mode, neither the radar
antenna nor the target being mapped rotate, and the radar flight track
is almost perpendicular to look direction with a small squint angle. By
reformulating the projection kernel and using it in an optimization
equation form, an optimal estimate of the target’s reflectivity field
may be obtained [13]. As a result, the image fuzziness may be re-
duced and image fidelity was preserved. Thus, the target’s features
were adequately enhanced, and dominant scatterers can be well sep-
arated. Finally, target recognition from various commercial aircrafts
was demonstrated through simulated database and a neural classifier.
The neural network is structured to allow its training by Kalman fil-
tering technique [14–18]. Performance evaluation was done from sim-
ulated images various real images by Radarsat-2 and TerraSAR-X in
stripmap mode.

2. SATELLITE SAR IMAGES SIMULATION

Figure 1 illustrates a functional block diagram of a satellite SAR

image simulation and a follow up target recognition. The simulation
Progress In Electromagnetics Research, Vol. 119, 2011 37

Figure 1. SAR Image simulation and target recognition flowcharts.

processing flow is basically adopted from [7, 19, 20]. Generally, the
inputs include satellite and radar parameters setting, target’s CAD
model, and also the clutter model. The computation includes satellite
orbit estimation, imaging geometry, target RCS, SAR echo and raw
signal generation. A critical step of Doppler centroid and rate
estimation for image focusing is also included. The most time
consuming part is in echo and raw signal generation which will be
illustrated in detail below. Then a refined range-Doppler method is
applied to perform the SAR image focusing. The outputs include
sets of images for a desired target for a range of radar looking angles
and orientation angles. Once the image database is built and made
available, feature extraction and enhancement is performed, followed
by a target recognition stage that includes a neural classifier training,
and neural weights storage for later operation stage.
38 Chang, Chiang, and Chen

2.1. SAR Signal Model

A typical side-looking SAR observation geometry is referred to Fig. 2
where the slant range between SAR and target, R is a function of
satellite moving time or slow time η or equivalently depends on squint
angle away from the zero-Doppler plane which corresponds to the slant
range R0 . In SAR processing, it is essential to project R onto R0 .

Figure 2. A typical stripmap SAR observation geometry [19].

A SAR transmits a linear frequency modulated (LFM) signal of

the form: ¡ ¢
st (τ ) = wr (τ, Tp ) cos 2πfc τ + παr τ 2 (1)
where αr is chirp rate, Tp is pulse duration, fc is carrier frequency,
τ is the ADC sampling time or fast time, and wr (·) is a rectangular
function:
µ ¶ ( ¯ ¯
¯ ¯
τ 1, ¯ Tτp ¯ ≤ 0.5
wr (τ, Tp ) = rect = (2)
Tp 0, else
The received signal or raw data is a delayed version of the transmitted
signal (1) [19],
µ ¶ µ ¶
2R 2R
sr (τ ) = A0 st τ − = A0 wr τ − , Tp wa (η)
c c
( µ ¶ µ ¶ )
2R 2R 2
cos 2πf0 τ − + παr τ − +ϕ (3)
c c
where R is distance from antenna to a point target being observed, A0
is slant range backscatter coefficient of the target, ϕ is the phase term,
Progress In Electromagnetics Research, Vol. 119, 2011 39

Figure 3. A typical antenna pattern along the azimuth direction and

received signal.

and wa (η) is the antenna pattern and is a function of slow time. A

commonly used pattern is of the form
(√ )
3θ(η)
wa (η) ∼
= sinc2 (4)
2βaz
where βaz is the azimuth beamwidth, θ(η) is angle measured from
boresight in the slant range plane. Fig. 3 displays such antenna pattern
(upper) and received echo (lower) along the azimuth direction. The
echo strength is varied according to antenna gain and is changed with
flight time η Note that in synthetic aperture concept, the target’s
information is embedded in Doppler frequency and its rate. The
received signal is coherently mixed down to the baseband.
After demodulation, the received signal is given as [19]
µ ¶
2R(η)
s0 (τ, η) = A0 wr τ − , Tp wa (η)
c
( µ ¶ )
4πfc R(η) 2R(η) 2
exp −j + jπαr τ − + ϕ00 (5)
c c
where ϕ00 is lumped sum of phase noise from atmosphere, satellite
altitude error, terrain, etc. and is only of interest for interferometric
SAR. Eq. (5) serves as fundamental signal model for the working
process that follows.

2.2. RCS Computation

To facilitate the radar response from a target, we need the target’s
RCS under radar observation scenario. The radar backscattering
40 Chang, Chiang, and Chen

(a) (b)

Figure 4. Computed RCS of commercial aircrafts for (a) C-band HH

polarization, (b) X-band V V polarization.

characteristics is taken into account. The coherent scattering process

between target and its background is neglected for sake of simplicity.
Also, fully polarimetric response is not considered. The Radar Cross
Section Analysis and Visualization System (RAVIS) [21] is a powerful
algorithm that utilizes the physical optics (PO), physical diffraction
theory (PDT), and shooting and bouncing rays (SBR) to compute
the RCS of complex radar targets [22, 23]. Single scattering and
diffraction from a target are first computed by PO and PDT, followed
by SBR to account for multiple scattering and diffraction. The system
outputs for a given 3D CAD model of the target of interest The
CAD model contains numerous grids or polygons, each associated with
the computed RCS as a function of incident and aspect angles for a
given set of radar parameters. The number of polygons is determined
by target’s geometry complexity and its electromagnetic size. To
realize the imaging scenario, each polygon must be properly oriented
and positioned based on Earth Centered Rotating (ECR) coordinates.
Fig. 4 displays the computed RCSs of these aircraft for the cases of
Radarsat-2 and TerraSAR-X.

2.3. Clutter Model

Although the coherent scattering process between target and its
background is neglected for sake of simplicity, the clutter from
Progress In Electromagnetics Research, Vol. 119, 2011 41

background is incoherently integrated into the radar echo. Extensive

studies on SAR speckle and its reduction have been documented [8, 24–
26]. Many studies [27, 28] suggest that for surface-like clutter such
as vegetation canopy [29, 30] and airport runway, the radar signal
statistics follows fairly well Weibull distribution. For our aircraft
targets of interest, we applied the Weibull distribution for runway
clutter model. Other models can be easily adopted in the simulation
chain.
κ ³ x ´κ−1
p (x) = exp [−(x/λ)κ ] (6)
λ λ
where the shape parameter κ and scale parameter λ may be estimated
from real SAR images over the runway with mean amplitude A [26]
π
κ = √ (7)
std [ln (A)] 6
½ ¾
0.5722
λ = exp [< ln (A) >] + (8)
κ
Fig. 5 displays sample regions of clutter and estimated model
parameters on a TerraSAR-X image. The model parameters are
obtained by averaging several samples.

Figure 5. Samples of clutter and estimated model parameters on a

TerraSAR-X image.
42 Chang, Chiang, and Chen

2.4. Satellite Orbit Determination

In applying Eq. (5), the range between a satellite in space and a target
being imaged on the ground must be precisely determined for matched
filtering. That is to say we need to know the satellite and target
position vectors in order to estimate the Doppler frequency and its rate.
Before doing so we have to define and then determine the reference
systems for time. Also the transformation between geocentric and
geodetic coordinates on the Earth must be computed. Six fundamental
parameters required to determine the orbital position include semi-
major axis a, eccentricity e, inclination angle i, right ascension of
ascending node Ω, argument of perigee ω, true anomaly ν. Three
of the parameters determine the orientation of the orbit or trajectory
plane in space, while three locate the body in the orbital plane [31, 32].
These six parameters are uniquely related to the position and velocity
of the satellite at a given epoch.
To deduce the orbital elements of a satellite, at least six
independent measurements are needed [31]. In this paper, we
adopted simplified perturbations models (SPM) [33–35] to estimate the
orbital state vectors of satellite in the Earth Centered Inertial (ECI)
coordinate system. A C++ version program code by Henry [33] was
used. Once the orbit parameters are determined, we can transform it
to ECR coordinate system, as described below. The transformation
matrix is of the form [31]

UECI
ECR = ΠΘNP (9)
where matrices Π, Θ, N, P represent, respectively, polar motion, Earth
rotation, nutation, and precession matrices:

Θ (t) = Rz (GAST ) (10)

P (T1 , T2 ) = Rz (−z (T, t)) Ry (ϑ (T, t)) Rz (−ξ (T, t)) (11)
where GAST is Greenwich apparent sidereal time. Referring to Fig. 6,
the nutation matrix is

N (T ) = Rx (−ε − ∆ε) Rz (−∆ψ) Rz (ε) (12)

Note that from [31], in computing the derivative of the transformation,
the precession, nutation and polar motion matrix maybe considered as
constant.

dUECI
ECR dΘ
≈Π NP (13)
dt dt
Progress In Electromagnetics Research, Vol. 119, 2011 43

Figure 6. Illustration of nutation.

Then the state vectors in the transformation are

rECR = UECI
ECR rECI (14)
dUECI
vECR = UECI
ECR vECI +
ECR
rECI (15)
dt
¡ ECI ¢T
rECI = UECR rECR (16)
¡ ¢T
¡ ECI ¢T d UECI
ECR
vECI = UECR vECR + rECR (17)
dt
In summary, four key steps are needed:
(i) Find satellite position: calculated from two-line elements data and
imaging time duration;
(ii) Find radar beam pointing vector derived from satellite attitude
(pitch, roll and yaw angle);
(iii) Locate target center position derived from satellite position and
line of sight;
(iv) Locate each target’s polygon position derived from target center
position and target aspect angle.

2.5. Image Focusing

Several focusing algorithms have been proposed including Range-
Doppler (RD), Omega-K, Chirp-Scaling, etc. [19, 36, 37] and their
improved versions Each of them bears pros and cons. We adopted
the refined range-Doppler algorithm with secondary range compression
because of their fast computation while maintaining reasonable good
spatial resolution and small defocusing error. Fig. 7 outlines the
functional block diagram of processing steps in refined RD algorithm.
44 Chang, Chiang, and Chen

START

Orbit Data SAR Raw Data

Baseband Doppler Centroid Frequency Azimuth Matching

Estimation (PB/ACCC) Filter Generation

Doppler Frequency Ambigulity Azimuth Compression

Estimation
(WDA/MLCC/MLBF)

Autofocusing
Absolute Doppler SAR Image (Optional)
Centroid Frequency

Range Compression Good

NO
Enough

Range Migration Cells Correction

YES
(Sinc Insterpolation )

STOP
Doppler Rate Estimation
(Effective Velocity)

PB (Power Balance)
ACCC (Average Cross Correlation Coefficient)
MLCC (Multi-Look Cross-Correlation)
MLBF (Multilook Beat Frequency)
WDA (Wavenumber Domain Algorithm)

Figure 7. SAR image focusing flowchart of a RD-based algorithm.

The power balance method in conjunction with the average cross

correlation coefficient method was used to obtain a rough estimate
of baseband Doppler frequency [38]. To obtain absolute Doppler
centroid frequency, fDC , algorithms such as MLCC (Multi-Look
Cross-Correlation), MLBF (Multilook Beat Frequency), and WDA
(Wavenumber Domain Algorithm) may be utilized to determine the
ambiguity number [19]:
fDC = fDC,base + Mamb P RF (18)
where fDC,base is the baseband part of PRF, Mamb is ambiguity number
and PRF is pulse repetition frequency.
After an effective satellite velocity, Vr , is estimated from orbit
state vectors, the Doppler rate estimation can be computed via
2Vr2 (R0 ) cos (θsq,c )
fR (R0 ) = (19)
R0
Progress In Electromagnetics Research, Vol. 119, 2011 45

where θsq,c is squint angle, and R0 is referred to zero-Doppler plane

(See Fig. 2). Note that in some cases, Eq. (19) is just a rough
estimate. Then autofocus via phase gradient algorithm [19] may be
further applied to refine the Doppler rate given above.

3. APPLICATION TO TARGET RECOGNITION

The proposed working flows and algorithms were validated by

evaluating the image quality including geometric and radiometric
accuracy using simple point targets first, followed by simulating four
types of commercial aircrafts: A321 B757-200, B747-400, and MD80
The satellite SAR systems for simulation include ERS-1, ERS-2,
Radarsat-2, TerraSAR-X, and ALOS PALSAR, but others can be
easily realized. Table 1 lists the key image quality indices for a
simulated image of TerraSAR-X and PLASAR satellite SAR systems.
Quality indices include 3 dB azimuth beamwidth (m), peak-to-side
lobe ratio (dB), integrated side lobe ratio (dB), scene center accuracy
in different coordinate systems. As can be seen from the table, the
simulated images are all well within the nominal specifications for
different satellite systems.

Table 1. Comparison of simulated image quality for a point target.

ALOS PALSAR TerraSAR-X

Index
Simulation Nominal Simulation Nominal
3 dB Azimuth beamwidth [m] 4.26 5.10 3.49 4.51
Slant range − 24.85 −22.85
PSLR [dB] −20.5 −13
Azimuth − 25.42 −31.16
Slant range − 30.62 −31.62
ISLR [dB] −15.0 −17
Azimuth − 25.62 −43.98
121.49927 E, 121.49930 E, 119.395090 E, 119.39506oE,
o o o
Lon, Lat [degree]
24.764622 oN 24.764477oN 24.894470oN 24.894620 oN
−3027814.2, −3027808.4, −2817396.77, −2817397.45,
ECR(x,y,z) [m] 4941079.3, 4941075.0, 5608995.60, 5608996.11,
Scente cener 2655407.1 2655421.7 2830631.97 2830630.29
Lon/Lat difference -5 -4
−3x10 , 1.45x10 −3x10 , 1.5x10-4
-6
[degree]
ECR difference
(ground range, 0.98, −16.26 0.06, 16.97
azimuth) [m]
46 Chang, Chiang, and Chen

3.1. Feature Enhancement

The essential information for target classification, detection, recog-
nition, and identification is through feature selection for training
phase. For our interest, depending on the spatial resolution offered
by Radarsat-2 and TerraSAR-X in stripmap mode, we aim at target
recognition. Before feature extraction, feature enhancement is per-
formed first. To enhance the SAR data using the non-quadratic regu-
larization method [11, 12], one has to modify the projection operator
kernel accordingly. The received signal so in Eq. (5) may be expressed
as [13]:
so = Tf + sn (20)
where sn represents noise, and T is the projection operation kernel with
dimension M N × M N which plays a key role in contrast enhancement
if the signal in Eq. (5) is of dimension M ×N . It has been shown that [1]
non-quadratic regularization is practically effective in minimizing the
clutter while emphasizing the target features via:
n o
f̂ = arg min ksr − Tf k22 + γ 2 kf kpp (21)

where kkp denotes `p -norm (p ≤ 1), and γ 2 is a scalar parameter and

n o
kso − Tf k22 + γ 2 kf kpp is recognized as the cost or objective function.
To easily facilitate the numerical implementation, both so and f
may be formed as long vectors, with T a matrix. Then, from Eq. (5)
and Eq. (20), we may write the projection operation kernel for the
stripmap mode as [13]
  ³ ´  ³ ´ 2 

 (x−x0 )2 (x−x0 )2 

 2 r+ 2r 
2 r+ 2r
 
T = exp −iω0 −α t−  (22)

 c c 


where t, r, x − x0 are matrices to be described below by firstly defining

the following notations to facilitate the formation of these matrices:
N : the number of discrete sampling points along the slant range
direction,
M : the number of discrete sampling points along the azimuth
direction,
Dt: the sampling interval along the slant range,
`r : the size of the footprint along the slant range,
`x : the size of the footprint along the azimuth direction,
1l : the column vector of dimension l × 1 and all of elements equal
to one, in which l = M or l = M N .
Progress In Electromagnetics Research, Vol. 119, 2011 47

With these notions, we can obtain explicit t, r, x − x0 forms, after

some mathematical derivations [13].
t = [1M ⊗ M1 ]M N ×M N (23)
   
0 0
 1   ∆t 
v1 =  .
 ..
 ∆t = 
  .. 
 (24a)
.
N −1 (N − 1)∆t N ×1
 
0 ··· 0
 ∆t ∆t 
M1 = V1 · 1TM N =
 ... .. 
 (24b)
.
(N − 1)∆t · · · (N − 1)∆t N ×M N
r = [M2 ⊗ V2T ]M N ×M N (25)
 −` 
2 + r0
r

 −`r + r0 + ∆`  `r
 
V2 =  . 2  , ∆` = (26a)
 ..  N
−`r
2 + r0 + (N − 1)∆` N ×1
M2 = 1TM ⊗ 1M N = [ 1M N · · · 1M N ]M N ×M N (26b)
x − x0 = [M3 ⊗ M4 ]M N ×M N (27)
   −` 
x
0 2
`x  
`x `x
 1  `x  M − 2 
V3 =  .
 ..

M − =  .  (28a)
2  .. 
M −1 (M − 1) M `x
− `2x M ×1
· ¸ ¹ º
V3 [δ − 1 : (M − 1)]δ×1 M −1
V4 = 0 ; δ= +1 (28b)
(M −δ)×1 M ×1 2
· ¸
Mirror(V3 [0 : δ − 1])δ×1
V5 = 0 (28c)
(M −δ)×1 M ×1
£ ¤
M3 = Toep(V5T , V4T ) M ×M (28d)
£ T
¤
M4 = 1N ×1 · 1N ×1 N ×N (28e)
In the above equations, each of the bold-faced letters denotes
a matrix, and ⊗ represents the Kronecker product. The operation
V[m : n] takes element m to element n from vector V, and Toep( · )
converts the input into a Toeplitz matrix [39, 40]. Note that in (28c),
48 Chang, Chiang, and Chen

by Mirror [a : b, c : d]p×q we mean taking elements a to b along the rows

and elements c to d along the columns, so that the resulting matrix is
of size p × q.

3.2. Feature Vectors and Extraction

The feature vector contains two types: fractal geometry and scattering
characteristics. In fractal domain, the image is converted into
fractal image [41]. It has been explored that SAR signals may be
treated as a spatial chaotic system because of the chaotic scattering
phenomena [42, 43]. Applications of fractal geometry to the analysis
of SAR are studied in [44, 45]. There are many techniques proposed to
estimate the fractal dimension of an image. Among them, the wavelet
approach proves both accurate and efficient. It stems from the fact
that the fractal dimension of an N -dimensional random process can be
characterized in terms of fractional Brownian motion (fBm) [41]. The
power spectral density of fBm is written as
P (f ) ∝ f −(2H+D) (29)
where 0 < H < 1 is the persistence of fBm, and D is the topological
dimension (= 2 in image). The fractal dimension of this random
process is given by D = 3 − H. As image texture, a SAR fractal
image is extracted from SAR imagery data based on local fractal
dimension. Therefore, wavelet transform can be applied to estimate
the local fractal dimension of a SAR image. From (29), the power
spectrum of an image is therefore given by
³p ´−2H−2
P (u, v) = υ u2 + v 2 (30)
where υ is a constant. Based on the multiresolution analysis, the
discrete detailed signal of an image I at resolution level j can be written
as [46, 47]:
­ ¡ ¢®
Dj I = I (x, y) , 2−j Ψj x − 2−j n, y − 2−j m
¡ ¢¡ ¢
= I (x, y) ⊗ 2−j Ψj (−x, −y) 2−j n, 2−j m (31)
¡ ¢
where ⊗ denotes a convolution operator, Ψj (x, y) = 22j Ψ 2j x, 2j y
and Ψ (x, y) is a two-dimensional wavelet function. The discrete
detailed signal, thus, can be obtained by filtering the signal with
2−j Ψj (−x, −y) and sampling the output at a rate 2−j . The power
spectrum of the filtered image is given by [47]
¯ ¯2
¯ ¯
Pj (u, v) = 2−2j P (u, v) ¯Ψ̃j (u, v)¯ (32)
Progress In Electromagnetics Research, Vol. 119, 2011 49

¡ ¢
where Ψ̃j (u, v) = Ψ̃ 2−j u, 2−j v and Ψ̃ (u, v) is the Fourier transform
of Ψ (u, v). After sampling, the power spectrum of the discrete detailed
signal becomes
XX ¡ ¢
Pjd (u, v) = 2j Pj u + 2−j 2kπ, v + 2−j 2lπ (33)
k l

Let σj2 be the energy of the discrete detailed signal

ZZ
2 2−j
σj = Pjd (u, v) dudv (34)
(2π)2
By inserting (32) and (33) into (34) and changing variables in this
integral, (34) may be expressed as σj2 = 2−2H−2 σj−1
2 . Therefore, the

fractal dimension of a SAR image can be obtained by computing the

ratio of the energy of the detailed images:
1 σj2
D= log2 2 + 2 (35)
2 σj−1
A fractal image indeed represents information regarding spatial
variation; hence, its dimension estimation can be realized by sliding a
preset size window over the entire image. The selection of the window
size is subject to reliability and homogeneity considerations with the
center pixel of the window replaced by the local estimate of the fractal
dimension. Once fractal image is generated, features of angle, target
area, long axis, short axis, axis ratio are extracted from target of
interest. As for scattering center (SC), features of major direction
[X], major direction [Y ], minor direction [X], minor direction [Y ],
major variance, and minor variance are selected, in addition to radar
look angle and aspect angle. Fig. 8 displays such feature vector from
simulated MD80 and B757-200 aircrafts from Radarsat-2. Finally, the
recognition is done by a dynamic learning neural classifier [14–18].
This classifier is structured with a polynomial basis function model.
A digital Kalman filtering scheme [48] is applied to train the network.
The necessary time to complete training basically is not sensitive to
the network size and is fast. Also, the network allows recursive training
when new and update training data sets is available without revoking
training from scratch. The classifier is structured with 13 input nodes
to feed the target features, 350 hidden nodes in each of 4 hidden layers,
and 4 output nodes representing 4 aircraft targets (A321, B757-200,
B747-400, and MD80).
50 Chang, Chiang, and Chen

(a) (b)

Figure 8. Selected features of simulated Radarsat-2 images of (a)

A321 and (b) B757-200 for incidence angle of 35◦ and aspect angle
from −180◦ ∼ +179 o (1◦ step).

3.3. Performance Evaluation

3.3.1. Simulated SAR Images
Both simulated Radarsat-2 and TerraSAR-X images for 4 targets, as
described above, are used to evaluate the classification and recognition
performance. For this purpose, Radarsat-2 images with incidence angle
of 45◦ and TerraSAR-X with 30◦ of incidence were tested. Both are
with spatial resolution of 3 m in stripmap mode. The training data
contains all 360 aspect angles (1◦ a step), among them 180 samples were
randomly taken to test. Fast convergence of neural network learning
is observed. The confusion matrix for Radarsat-2 and TerraSAR-X in
classifying 4 targets is given in Table 2 and Table 3, respectively. The
overall accuracy and kappa coefficient are very satisfactory. Higher
confusion between B757-200 and MD80 was observed. It has to be
noted that with Radarsat-2 and TerraSAR-X systems, comparable
results were obtained as long as classification rate is concerned.
Progress In Electromagnetics Research, Vol. 119, 2011 51

Table 2. Confusion matrix of classifying 4 targets from simulated

Radarsat-2 images.
Producer
A321 B747-400 B757-200 MD80
accuracy
A321 165 2 6 7 0.917
target B747-400 0 175 3 2 0.972
B757-200 4 1 164 11 0.911
MD80 4 2 11 163 0.906
User accuracy 0.954 0.972 0.891 0.891
overall accuracy: 0.926, kappa coefficient: 0.902

Table 3. Confusion matrix of classifying 4 targets from simulated

TerraSAR-X images.
Producer
A321 B747-400 B757-200 MD80
accuracy
A321 166 0 5 9 0.922
target B747-400 0 179 0 1 0.994
B757-200 2 0 168 10 0.933
MD80 1 0 8 171 0.950
User accuracy 0.954 0.982 1.000 0.928 0.895
overall accuracy: 0.950, kappa coefficient: 0.933

3.3.2. Real SAR Images

Data acquisition on May 8, 2008 from TerraSAR-X over an airfield
was processed into 4 looks images with spatial resolution of 3 m in
stripmap mode. Feature enhancement by an operator kernel in (22)
was performed. Meanwhile, ground truth collections were conducted
to identify the targets. Fig. 9 displays such acquired images where the
visually identified targets are marked and their associated image chips
were later fed into neural classifier that is trained by the simulated
image database. Among all the targets, 3 types of them are contained
in simulated database: MD80, B757-200, and A321. From Table 4,
these targets are well recognized, where the numeric value represents
membership. Winner takes all approach was taken to determine the
classified target and whether they are recognizable. More sophisticated
schemes such as type-II fuzzy may be adopted in the future. It is
realized that for certain poses, there exists confusion between MD80
and B757-200, as already demonstrated in simulation test above.
As another example, a blind test was performed. Here blind means
52 Chang, Chiang, and Chen

Table 4. Membership matrix for target recognition on TerraSAR-X

image of Fig. 9.
PP
PPoutput A321 B747-400 B757-200 MD80 recognizable
input PP
P
MD80 1 16.61% 7.26% 0.00% 76.12% yes
MD80 2 0.00% 9.63% 18.50% 71.85% yes
MD80 3 0.00% 3.63% 8.78% 87.58% yes
MD80 4 0.00% 3.64% 23.08% 73.27% yes
MD80 5 17.88% 12.28% 0.00% 69.82% yes
MD80 6 5.64% 0.00% 46.72% 47.63% yes
B757-200 1 12.05% 11.84% 76.09% 0.00% yes
B757-200 2 0.00% 19.41% 78.55% 2.03% yes
A321 63.28% 15.09% 0.00% 21.61% yes

Figure 9. A TerraSAR-X image over an airfield acquired on May 15,

2008. The identified targets are marked from ground truth.

the targets to be recognized is not known to the tester, nor is the image
acquisition time. The ground truth was collected by the third party
and was only provided after recognition operation was completed. This
is very close to the real situation for recognition operation. A total
of 12 targets (T1-T12) were chosen for test, as indicated in Fig. 10.
Unlike in Fig. 9, this image was acquired in descending mode but was
unknown at the time of test. Table 5 gives the test results, where
the recognized targets and truth are listed. It is readily indicated
that all the MD80 targets were successfully recognized, while 4 types
of targets were wrongly recognized. T12 target (ground truth is a
Progress In Electromagnetics Research, Vol. 119, 2011 53

Figure 10. A Radarsat-2 image over an airfield. The identified targets

are marked.

Table 5. Membership matrix for target recognition on Raradsat-2

image of Fig. 10.

HH
HH A321 B747-400 B757-200 MD80 Recognized Truth
H
T1 19.25% 0.00% 22.88% 57.86% MD80 MD80
T2 0.01% 0.00% 0.00% 99.98% MD80 MD80
T3 14.99% 0.00% 11.59% 73.40% MD80 MD80
T4 21.31% 10.81% 15.27% 52.60% MD80 MD80
T5 0.00% 3.09% 95.10% 1.79% B757-200 E190
T6 96.82% 1.70% 0.00% 1.46% A321 B737-800
T7 0.13% 0.00% 0.17% 99.69% MD80 MD80
T8 0.10% 0.04% 0.00% 99.84% MD80 MD80
T9 63.28% 15.09% 0.00% 21.61% A321 FOKKER-100
T10 6.48% 2.26% 0.00% 91.24% MD80 MD80
T11 43.61% 15.59% 40.59% 0.00% A321 DASH-8

C130) was completely not recognizable by the system. These are

mainly attributed the lack of database. Consequently enhancement
and update of the target database is clearly essential.

4. CONCLUSION
In this paper, we present a full blown satellite SAR image simulation
approach to target recognition. The simulation chain includes
orbit state vectors estimation, imaging scenario setting, target
54 Chang, Chiang, and Chen

RCS computation, clutter model, all specified by a SAR system

specification. As an application example, target recognition has been
successfully demonstrated using the simulated images database as
training sets. Extended tests on both simulated image and real
images were conducted to validate the proposed algorithm. To this
end, it is suggested that a more powerful target recognition scheme
should be explored for high-resolution SAR images. Extension to fully
polarimetric SAR image simulation seems highly desirable as more
such data are being available for much better target discrimination
capability. In this aspect, further improvement on the computational
efficiency by taking advantage of graphic processing unit (GPU) power
may be preferred and is subject to further analysis.

REFERENCES

1. Rihaczek, A. W. and S. J. Hershkowitz, Theory and Practice of

Radar Target Identification, Artech House, 2000.
2. Huang, C. W. and K. C. Lee, “Frequency-diversity RCS
based target recognition with ICA projection,” Journal of
Electromagnetic Waves and Applications, Vol. 24, No. 17–18,
2547–2559, 2010.
3. Guo, K. Y., Q. Li, and X. Q. Sheng, “A precise recognition method
of missile warhead and decoy in multi-target scene,” Journal of
Electromagnetic Waves and Applications, Vol. 24, No. 5–6, 641–
652, 2010.
4. Tian, B., D. Y. Zhu, and Z. D. Zhu, “A novel moving target
detection approach for dual-channel SAR system,” Progress In
Electromagnetics Research, Vol. 115, 191–206, 2011.
5. Wang, X. F., J. F. Chen, Z. G. Shi, and K. S. Chen, “Fuzzy-
control-based particle filter for maneuvering target tracking,”
Progress In Electromagnetics Research, Vol. 118, 1–15, 2011.
6. Lee, J. S. and E. Pottier, Polarimetric Radar Imaging: From
Basics to Applications, CRC Press, 2009.
7. Margarit, G., J. J. Mallorqui, J. M. Rius, and J. Sanz-Marcos, “On
the usage of GRECOSAR, an orbital polarimetric SAR simulator
of complex targets, to vessel classification studies,” IEEE Trans.
Geoscience and Remote Sens., Vol. 44, 3517–3526, 2006.
8. Lee, J. S., “Digital image enhancement and noise filtering by use of
local statistics,” IEEE Trans. Pattern Anal. Mach. Intell., Vol. 2,
165–168, 1980.
9. Chang, Y. L., L. S. Liang, C. C. Han, J. P. Fang, W. Y. Liang, and
K. S. Chen, “Multisource data fusion for landslide classification
Progress In Electromagnetics Research, Vol. 119, 2011 55

using generalized positive boolean functions,” IEEE Trans.

Geosci. Remote Sensing, Vol. 45, 1697–1708, 2007.
10. Wang, J. and L. Sun, “Study on ship target detection and
recognition in SAR imagery,” CISE’09 Proceedings of the 2009
First IEEE International Conference on Information Science and
Engineering, 1456–1459, 2009.
11. Çetin, M. and W. C. Karl, “Feature-enhanced synthetic aperture
radar image formation based on nonquadratic regularization,”
IEEE Trans. on Image Processing, Vol. 10, 623–631, 2001.
12. Çetin, M., W. C. Karl, and A. S. Willsky, “Feature-preserving
regularization method for complex-valued inverse problems with
application to coherent imaging,” Optical Engineering, Vol. 45,
017003-1–11, 2006.
13. Chiang, C. Y., K. S. Chen, C. T. Wang, and N. S. Chou,
“Feature enhancement of stripmap-mode SAR images based on
an optimization scheme,” IEEE Geoscience and Remote Sensing
Letters, Vol. 6, 870–874, 2009.
14. Tzeng, Y., K. S Chen, W. L. Kao, and A. K. Fung, “A dynamic
learning neural network for remote sensing applications,” IEEE
Trans. Geosci. Remote Sensing, Vol. 32, 1096–1102, 1994.
15. Chen, K. S., Y. C. Tzeng, C. F. Chen, and W. L. Kao, “Land-cover
classification of multispectral imagery using a dynamic learning
neural network,” Photogrammet. Eng. Remote Sensing, Vol. 61,
403–408, 1995.
16. Chen, K. S., Y. C. Tzeng, and W. L. Kao, “Retrieval of surface
parameters using dynamic learning neural network,” International
Geoscience and Remote Sensing Symposium 1993, IGARSS ’93.
Better Understanding of Earth Environment, Vol. 2, 505–507,
1993.
17. Chen, K. S., W. P. Huang, D. W. Tsay, and F. Amar,
“Classification of multifrequency polarimetric SAR image using a
dynamic learning neural network,” IEEE Trans. Geosci. Remote
Sensing, Vol. 34, 814–820, 1996.
18. Tzeng, Y. C. and K. S. Chen, “A fuzzy neural network to
SAR image classification,” IEEE Trans. Geosci. Remote Sensing,
Vol. 36, 301–307, 1998.
19. Cumming, I. and F. Wong, Digital Signal Processing of Synthetic
Aperture Radar Data: Algorithms and Implementation, Artech
House, 2004.
20. Curlander, J. C. and R. N. McDonough, Synthetic Aperture
Radar: Systems and Signal Processing, Wiley-Interscience, 1991.
56 Chang, Chiang, and Chen

21. Chen, C. C., Y. Cheng, and M. Ouhyoung, “Radar cross section

analysis and visualization system,” Proceeding of Computer
Graphics Workshop 1995, 12–16, Taipei, Taiwan, 1995.
22. Lee, H., R. C. Chou, and S. W. Lee, “Shooting and bouncing
rays: Calculating the RCS of an arbitrarily shaped cavity,” IEEE
Trans. Antennas Propag., Vol. 37, 194–205, 1989.
23. Chen, S. H. and S. K. Jeng, “An SBR/image approach for
indoor radio propagation in a corridor,” Trans. IEICE Electron.,
Vol. E78-C, No. 8, 1058–1062, 1995.
24. Lee, J. S., “Refined filtering of image noise using local statistics,”
Comput., Vis., Graph., Image Process., Vol. 15, 380–389, 1981.
25. Lee, J. S., M. R. Grunes, and R. Kwok, “Classification of
multi-look polarimetric SAR imagery based on complex Wishart
distribution,” Int. J. Remote Sensing, Vol. 15, 2299–2311, 1994.
26. Lee, J. S., P. Dewaele, P. Wambacq, A. Oosterlinck, and
I. Jurkevich, “Speckle filtering of synthetic aperture radar images:
A review,” Remote Sens. Rev., Vol. 8, 313–340, 1994.
27. Schleher, D. C., “Radar detection in Weibull clutter,” IEEE
Trans. Aerosu. Electron. Svst., Vol. 12, 736–743, 1976.
28. Boothe, R. R., “The Weibull distribution applied to the ground
clutter backscattering coefficient,” Automatic Detection and
Radar-data Processing, D. C. Schleher (ed.), 435–450, Artech
House, 1980.
29. Fung, A. K., Microwave Scattering and Emission Models and
Their Applications, Artech House, 1994.
30. Chen, K. S. and A. K. Fung, “Frequency dependence of signal
statistics from vegetation components,” IEE Processings —
Radar, Sonar and Navigation, Vol. 142, No. 6, 301–305, 1996.
31. Montenbruck, O. and E. Gill, Satellite Orbits: Models, Methods,
and Applications, Springer-Verlag, 2000.
32. Vallado, D. A. and W. D. McClain, Fundamentals of Astrody-
namics and Applications, 3rd edition, Microcosm Press, Springer,
2007.
33. Henry, M. F., “NORAD SGP4/SDP4 implementations,” Avail-
able: http://www. zeptomoby.com/satellites/.
34. Vallado, D., P. Crawford, R. Hujsak, and T. S. Kelso, “Revisiting
spacetrack report# 3,” AIAA, Vol. 6753, 1–88, 2006.
35. Miura, N. Z., “Comparison and design of simplified general
perturbation models,” California Polytechnic State University,
San Luis Obispo2009, Earth Orientation Centre, Available:
http://hpiers.obspm.fr/eop-pc/.
Progress In Electromagnetics
View publication stats
Research, Vol. 119, 2011 57

36. Cafforio, C., C. Prati, and F. Rocca, “SAR data focusing using
seismic migration techniques,” IEEE Transactions on Aerospace
and Electronic Systems, Vol. 27, 194–207, 1991.
37. Moreira, A., J. Mittermayer, and R. Scheiber, “Extended chirp
scaling algorithm for air- and spaceborne SAR data processing
in stripmap and ScanSAR imaging modes,” IEEE Trans. Geosci.
Remote Sensing, Vol. 34, 1123–1136, 1996.
38. Li, F.-K., D. N. Held, J. C. Curlander, and C. Wu, “Doppler
parameter estimation for spaceborne synthetic-aperture radars,”
IEEE Trans. Geosci. Remote Sensing, Vol. 23, 47–56, 1985.
39. Strang, G., Introduction to Applied Mathematics, Cambridge
Press, Wellesley, 1986.
40. Bini, D., Toeplitz Matrices, Algorithms and Applications, ECRIM
News Online Edition, No. 22, Jul. 1995.
41. Mandelbrolt, A. B. and J. W. van Ness, “Fractional Brownian
motion, fractional noises and applications,” SIAM Review, Vol. 10,
422–437, 1968.
42. Leung, H. and T. Lo, “A spatial temporal dynamical model for
multipath scattering from the sea,” IEEE Trans. Geosci. Remote
Sensing, Vol. 33, 441–448, 1995.
43. Leung, H., N. Dubash, and N. Xie, “Detection of small objects in
clutter using a GA-RBF neural network,” IEEE Transactions on
Aerospace and Electronic Systems, Vol. 38, 98–118, 2002.
44. Tzeng, Y. C. and K. S. Chen, “Change detection in synthetic
aperture radar images using a spatially chaotic model,” Opt. Eng.,
Vol. 46, 086202, 2007.
45. Chou, N. S., Y. C. Tzeng, K. S. Chen, C. T. Wang, and K. C. Fan,
“On the application of a spatial chaotic model for detecting
landcover changes in synthetic aperture radar images,” J. Appl.
Rem. Sens., Vol. 3, 033512, 1–16, 2009.
46. Mallat, S. G., “A theory for multiresolution signal decomposition:
The wavelet representation,” IEEE Trans. Pattern Anal. Mach.
Intell., Vol. 11, 674–693, 1989.
47. Mallat, S., A Wavelet Tour of Signal Processing, 2nd edition,
Academic Press, 1999.
48. Brown, R. and P. Hwang, Introduction to Random Signal Analysis
and Kalman Filtering, Wiley, New York, 1983.