Progress In Electromagnetics Research, PIER 81, 447–459, 2008

APPLICATION OF SVD NOISE-REDUCTION

TECHNIQUE TO PCA BASED RADAR TARGET
RECOGNITION

K.-C. Lee, J.-S. Ou, and M.-C. Fang

Department of Systems and Naval Mechatronic Engineering
National Cheng-Kung University
Tainan 701, Taiwan

Abstract—The noise effect is very challenging in radar target

recognition. It usually degrades the accuracy of target recognition
and then makes the recognition unreliable. In this study, we present
a noise-reduction technique to improve the accuracy of radar target
recognition. Our noise-reduction technique is based on the SVD
(singular value decomposition). The PCA (principal components
analysis) based radar recognition algorithm is utilized to verify our
noise-reduction scheme. In our treatment, the received signals are
arranged into a Hankel-form matrix. This Hankel-form matrix is
decomposed into two subspaces, i.e., the noise-related subspace and
clean-signal subspace. The noise reduction is obtained by suppressing
the noise-related subspace and retaining the clean-signal space only.
Simulation results show that the accuracy of target recognition is
greatly improved as the received signals are first processed by the SVD
noise-reduction technique. With the use of proposed noise-reduction
scheme, the radar target recognition can tolerate more noises and then
becomes more reliable. The noise-reduction technique in this study
can also be applied to many other problems in radar engineering.

1. INTRODUCTION

Radar target recognition [1–3] means to identify targets or achieve

the fundamental information of targets by received electromagnetic
signals. In practical measurement, the received signals contain a lot
of random noises in addition to the clean signals. These random
noises may destroy the recognition and then make the recognition
unreliable. In other words, the noise effects are very challenging in
radar target recognition. This then motivates us to develop an efficient
448 Lee, Ou, and Fang

and powerful noise-reduction scheme to improve the accuracy of radar

target recognition.
In this paper, we present a noise-reduction scheme to improve the
accuracy of radar target recognition. Our noise-reduction technique is
based on the SVD (singular value decomposition) [4]. In our treatment,
the received signals are arranged into a Hankel-form matrix. This
Hankel-form matrix is decomposed into two subspaces, i.e., the noise-
related subspace and clean-signal subspace. The noise reduction is
obtained by suppressing the noise-related subspace and retaining the
clean-signal space only.
After the received electromagnetic signals have been processed by
the above noise-reduction technique, the resulting signals are nearly
clean. These nearly clean signals are then utilized for radar target
recognition. In this study, the PCA (principal components analysis) [5]
and angular-diversity based radar target recognition algorithm [6] is
utilized to verify our SVD noise-reduction scheme. Simulation results
show that the accuracy of target recognition is greatly improved as the
received signals are first processed by our noise-reduction scheme. The
SVD based concept has been utilized in speech [7, 8] and imaging [9]
signal processing. However, there is still no research that applies
such a noise-reduction technique to radar target recognition. To our
knowledge, this is the first study that applies SVD noise-reduction
technique to such a problem. With the use of proposed noise-reduction
scheme, the radar target recognition can tolerate more noises and then
becomes more reliable.
In Section 2, the theoretical formulations are given. Numerical
simulation results are given in Section 3. Finally, the conclusion is
given in Section 4.

2. FORMULATIONS

Without loss of generality, we consider a ship on the sea level (X-Y

plane) located at the origin of coordinate for simplicity, as shown in
Figure 1. Note that this study has no limitation in types of targets. In
Figure 1, the front end of ship is in the +x̂ direction and the broadside
of ship is in the ±ŷ direction. The spherical coordinate system is
defined as (R, θ, φ) where R is the distance from observation position
to origin, θ is the elevation angle and φ is the azimuth angle. The
ship is illuminated by a plane wave E i = e+jk0 x ẑ where k0 is the
wavenumber. The bistatic RCS (radar cross section) in the direction
Progress In Electromagnetics Research, PIER 81, 2008 449

Z
observation
position

The broadside of a ship

The front end Y
of a ship
incident wave φ
+jko x
Ei = e z

X X-Y plane = sea level

Figure 1. Schematic diagram of a ship illuminated by an incident

plane wave.

of (θ, φ) is defined as [10]

 
E s (θ, φ)2
2
σ(θ, φ) = lim 4πR  2 . (1)
R→∞ E i 

where E s (θ, φ) is the scattered electric field. The bistatic RCS of a

ship at a fixed elevation angle θ and different N azimuth angles of φ
are collected to constitute an RCS vector u = [ u1 u2 · · · uN ]T ,
where “T ” denotes the transpose.
In practical measurement, the collected RCS u contains a lot
of noises. In this study, the measured u is processed by the SVD
technique [4] to reduce the noise. In general, the noisy u consists
of the clean signal v = [ v1 v2 · · · vN ]T and the addition noise
n = [ n1 n2 · · · nN ]T , i.e.,

u = v + n. (2)

In (2), v is deterministic and n is random. The goal is to extract the

clean v from the noisy u. To achieve this goal, the subspace concept is
450 Lee, Ou, and Fang

utilized for noise reduction and a Hankel-form matrix is defined as

 
u1 u2 ··· uJ
 u2 u3 ··· uJ+1 
Hu =  ... .. ..  . (3)
. . 
uI uI+1 · · · uI+J−1 I×J

The dimension of H u is I × J, where I + J = N + 1 and I ≥ J.

From (2)–(3), the noisy H u can be represented as the summation of
two Hankel-form matrices as
H u = H v + H n, (4)

where H v and H n represents the contributions from clean signals and

random noises, respectively. The structures of H v and H n are similar
to (3) except that ui (i = 1, 2, . . . , I + J − 1) is replaced by vi and ni ,
respectively. The clean signal v can be successfully extracted as H v is
obtained. By using the SVD technique, we can decompose the matrix
H u into

T
Hu = L Σ R . (5)

In (5), the columns of L (dimension I × I) and R (dimension

J × J) are called left and right singular vectors, respectively. All of
these column vectors are orthogonal. The matrix Σ is diagonal with
dimension I×J and is defined as Σ = diag{s1 , s2 , · · · , sM in(I,J) }, where
si (i = 1, 2, . . . , M in(I, J)) is the singular value and s1 ≥ s2 . . . ≥
sM in(I,J) ≥ 0. Physically, the largest singular value contributes almost
only clean signal information, whereas the smallest singular value
contributes almost only noise information. To extract the clean signal,
the largest K singular values are viewed as the clean components.
The remaining M in(I, J) − K singular values are viewed as the noise
components and are discarded. Therefore, the noise reduction can be
obtained by adjusting the singular values as
 
s1 0 · · · 0 0 ···
 .. .. 
 0 s2 . . 0 · · ·
. 
. .. .. 
Σ’ =  . . . 0 0 · · ·  . (6)

 0 · · · 0 sK 0 · · ·
0 ··· 0 0 0 · · ·
 
..
0 ··· ··· ··· ··· . I×J
Progress In Electromagnetics Research, PIER 81, 2008 451

A new matrix H v
is then constructed as

T
H v
= L Σ R (7)

to approximate the clean contribution. Note that H v
in (7) is no longer

Hankel-form. In our treatment, the anti-diagonal components of H v

are averaged to generate the modified H v
and the resulting matrix
will be Hankel-form. The estimation of the clean signal v is denoted
as v  and will be reconstructed from the modified H v
. In other words,
v  is the approximation for the clean signal v. Conventional radar
recognition utilized the noisy signal u for target identification, whereas
we utilized the nearly clean signal v  for target identification.
In this study, the PCA and angular-diversity based radar target
recognition algorithm [6] is utilized to verify our SVD noise-reduction
scheme. The recognition procedures are given in [6] and the results are
given in the following section.

3. NUMERICAL SIMULATION RESULTS

In this section, numerical examples are given to verify our noise-

reduction scheme in radar target recognition. Note that our noise-
reduction scheme has no limitation in types of targets. To easily obtain
the RCS data by simulation, simplified ship models are utilized instead
of real ships. Assume there are three types of known ships (P = 3)
including type I (to model the container vessel), type II (to model the
naval ship) and type III (to model the fishing boat). The geometrical
models for the three types of known ships are shown in Figure 2. The
ship length l is chosen to be k0 l = 9.4 for the ship of type I, k0 l = 6.3
for the ship of type II, and k0 l = 3.1 for the ship of type III. All ships
are on rough sea surface (X-Y plane). The characteristic for surface
roughness of sea water is assumed to be

4 15π 15π 8
z(x, y) = l · sin x sin y + l. (8)
75 4 4 75

The sea water has dielectric constant 81 and conductivity 4 S/m.

In the simulation of RCS, the commercial software Ansoft HFSS is
exploited. This software was proved to be accurate by many researchers
in this field. As the arrangement in Figure 1, the bistatic RCS from
each type of known ship at a fixed elevation angle θ and 181 (i.e.,
N = 181) azimuth angles of φ = 0◦ , 1◦ , . . . , 180◦ are calculated by
Ansoft HFSS software. To model the practical measurement including
452 Lee, Ou, and Fang

(a) (b)

(c)

Figure 2. Geometrical models for three known classes of ships: (a)

type I, (b) type II and (c) type III.

random noises, we add to each RCS a quantity of independent random

noise having a Gaussian distribution with zero mean. The standard
derivation of noise is 4×10−1 normalized with respect to the root-mean
square value of the RCS. This noisy RCS vector ui is first processed
by the SVD noise-reduction scheme and the result is v  i (nearly clean).
In the SVD, the dimension of the Hankel-form matrix is I = 91 and
J = 91. The number of reserved singular values in (6) is chosen as
K = 15, i.e., the remaining 76 singular values are discarded. This
implies that the largest 15 singular values are viewed as contribution
of clean signals and the remaining 76 singular values are viewed as
contribution of noises. The choice for the number of reserved singular
values depends on experiences. By sampling the elevation angles at
θ = 61◦ , 63◦ , . . . , 89◦ , we have M1 = M2 = M3 = 15 and M = 45.
The number of principal components in the PCA recognition is chosen
Progress In Electromagnetics Research, PIER 81, 2008 453

to be M  = 2.
Initially, the projected features of training RCS data in Step-1
are given. Figure 3 shows the projected features on the 2-dimensional
PCA space for the training RCS data of the three known ships. In
Figure 3(a), the noisy RCS data are not processed by any noise-
reduction scheme. From Figure 3(a), it shows that projected features
of type I and type II distribute close together. As the noisy RCS data
are first processed by our SVD noise-reduction scheme, the projected
features of all the three classes are well separated and are shown in
Figure 3(b). This example convinces us that the SVD based noise-
reduction scheme can really improve the target recognition. From
Figure 3, it shows there may be confusion in identifying the targets
of type I and type II. Therefore, we give further testing for these two
types of targets.
In the first testing, we assumed that the unknown target is just the
target of type I. The testing is implemented 15 times at θ = 62◦ , 64◦ ,
. . . , 90◦ , respectively. It should be noted that these testing elevation
angles are different from those for collecting the training data in Step-
1. Figure 4 shows the distance to feature centers (i.e., class error)
for the three known classes of ship RCS under 15 testing elevation
angles at θ = 62◦ , 64◦ , . . . , 90◦ , respectively. According to [6], the
magnitude of distance (class error) is in inverse proportion to the degree
of similarity. The smallest distance (class error) means that the target
ship has the highest degree of similarity with the corresponding type
of ship. Therefore, the lowest plot line represents the prediction for the
unknown target (should be type I in this example), because this plot
line has the shortest distance (to the unknown target) among the three
known targets. In Figure 4(a), the noisy RCS data are not processed
by any noise-reduction scheme. From Figure 4(a), it shows that the
lowest two plots (for type I and type II) distribute close together. One
may be confused in identifying the two targets. As the noisy RCS data
are first processed by our SVD noise-reduction scheme, the distance
(i.e., class error) is shown in Figure 4(b). From Figure 4(b), it shows
that the lowest two plots (for type I and type II) are well separated.
In the second testing, we assumed that the tested target is just the
target of type II. The remaining procedures are the same as those of the
previous example. Figure 5 shows the distance to feature centers (i.e.,
class error) for the three known classes of ship RCS under 15 testing
elevation angles respectively. The lowest plot line is expected to be type
II in this example. In Figure 5(a), the noisy RCS data are not processed
by any noise-reduction scheme. From Figure 5(a), it shows that the
lowest two plots (for type I and type II) distribute close together. One
may be confused in identifying the two targets. As the noisy RCS data
454 Lee, Ou, and Fang

120

80

type I
40 type II
2nd eigenspace of PCA

type III

-40

-80

-120
-100 -50 0 50 100 150 200
1st eigenspace of PCA

(a)
40

20

type I
2nd eigenspace of PCA

type II
type III

-20

-40
-100 -50 0 50 100 150
1st eigenspace of PCA

(b)

Figure 3. The projected features on the 2-dimensional PCA space for

the training RCS data of the three known ships: (a) without any noise-
reduction processing, and (b) with SVD noise-reduction processing.
Progress In Electromagnetics Research, PIER 81, 2008 455

200

distance to class center (class error)

150
testing target : type I
without any noise-reduction processing
distance to type I
distance to type II
distance to type III
100

50

0
60 70 80 90
elevation angle θ (degree)

(a)

200
distance to class center (class error)

150
testing target : type I
with SVD noise-reduction processing
distance to type I
distance to type II
distance to type III
100

50

0
60 70 80 90
elevation angle θ (degree)

(b)

Figure 4. The distance (i.e., class error) to feature centers for the
three known classes of ship RCS at different elevation angles θ by using
ship of type I as the testing target: (a) without any noise-reduction
processing, and (b) with SVD noise-reduction processing.
456 Lee, Ou, and Fang

200

distance to class center (class error)

150
testing target : type II
without any noise-reduction processing
distance to type I
distance to type II
distance to type III
100

50

0
60 70 80 90
elevation angle θ (degree)

(a)

200
distance to class center (class error)

150
testing target : type II
with SVD noise-reduction processing
distance to type I
distance to type II
distance to type III
100

50

0
60 70 80 90
elevation angle θ (degree)

(b)

Figure 5. The distance (i.e., class error) to feature centers for the
three known classes of ship RCS at different elevation angles θ by using
ship of type II as the testing target: (a) without any noise-reduction
processing, and (b) with SVD noise-reduction processing.
Progress In Electromagnetics Research, PIER 81, 2008 457

are first processed by our SVD noise-reduction scheme, the distance

(i.e., class error) is shown in Figure 5(b). From Figure 5(b), it shows
that the lowest two plots (for type I and type II) are well separated.
The results of the above two examples convince us that the SVD noise-
reduction scheme can improve the target recognition.
The above results of RCS data are calculated on a personal
computer with Pentium-3.0 CPU. The signal-processing programs are
coded using the Matlab-7 software.

4. CONCLUSIONS

The random noises of measured electromagnetic signals are challenging

problems in radar target recognition. In this paper, a noise-reduction
scheme based on SVD is proposed to treat the PCA radar target
recognition. In noise-reduction procedures, the subspace concepts
are utilized, i.e., suppressing the noise-related subspace and reserving
only the clean-signal subspace. Simulation results show that the
recognition accuracy is greatly improved as the RCS data are fist
processed by the SVD noise-reduction scheme. With the use of our
noise-reduction scheme, the target recognition can tolerate random
noises and then its results become more reliable. The models of
targets in this study are somewhat simple. However, this is not
important because our main purpose is to illustrate that the SVD-
based noise-reduction scheme can reduce the noise effects in target
recognition. From physical points of view, only the clean signal can
give information of the target. Therefore, noise-reduction processing
on collected data becomes necessary. This study can be applied to
many other applications in radar target recognition [11–21].

ACKNOWLEDGMENT

The work in this paper was supported by the National Science Council,
Taiwan, under Grant NSC NSC 96-2628-E-006-250-MY3, and by the
Landmark Program for NCKU’s Top-University Project under grant
96-B041. The author would like to express his sincere gratitude to
them.

REFERENCES

1. Hajduch G., J. M. Le Caillec, and R. Garello, “Airborne

high-resolution ISAR imaging of ship targets at sea,” IEEE
Transactions on Aerospace and Electronic Systems, Vol. 40, No. 1,
378–384, 2004.
458 Lee, Ou, and Fang

2. Tello, M., C. Lopez-Martinez, and J. J. Mallorqui, “A novel

algorithm for ship detection in SAR imagery based on the wavelet
transform,” IEEE Geoscience and Remote Sensing Letters, Vol. 2,
No. 2, 201–205, 2005.
3. Farhat, N. H., “Microwave diversity imaging and automated
target identification based on models of neural networks,” IEEE
Proceedings, Vol. 77, No. 5, 670–681, 1989.
4. Moon, T. K. and W. C. Stirling, Mathematical Methods and
Algorithms for Signal Processing, Prentice Hall, 2000.
5. Duda, R. O., P. E. Hart, and D. G. Stork, Pattern Classification,
2nd edition, John Wiley & Sons Inc., 2001.
6. Lee, K. C., J. S. Ou, and C. H. Huang, “Angular-diversity
radar recognition of ships by transformation based approaches —
including noise effects,” Progress In Electromagnetics Research,
PIER 72, 145–158, 2007.
7. Jensen, S. H., P. C. Hansen, S. D. Hansen, and J. A. Sorensen,
“Reduction of broad-band noise in speech by truncated QSVD,”
IEEE Transactions on Speech and Audio Processing, Vol. 3, 439–
448, 1995.
8. Hermus, K., I. Dologlou, P. Wambacq, and D. V. Compernolle,
“Fully adaptive svd-based noise removal for robust speech
recognition,” European Conference on Speech Communication and
Technology, 1951–1954, 1999.
9. Konstantinides, K., B. Natarajan, and G. S. Yovanof, “Noise es-
timation and filtering using blockbased singular value decomposi-
tion,” IEEE Transactions on Image Processing, Vol. 6, 479–483,
1997.
10. Ruck, G. T., D. E. Barrick, W. D. Stuart, and C. K. Krichbaum,
Radar Cross Section Handbook, Vol. 1, Plenum, New York, 1970.
11. Cui, B., J. Zhang, and X. W. Sun, “Single layer microstrip antenna
arrays applied in millimeter-wave radar front-end,” Journal of
Electromagnetic Waves and Applications, Vol. 22, No. 1, 3–15,
2008.
12. Xue W. and X. W. Sun, “Target detection of vehicle volume
detecting radar based on Wigner-Hough transform,” Journal of
Electromagnetic Waves and Applications, Vol. 21, No. 11, 1513–
1523, 2007.
13. Wang, C. J., B. Y. Wen, Z. G. Ma, W. D. Yan, and X. J. Huang,
“Measurement of river surface currents with UHF FMCW radar
systems,” Journal of Electromagnetic Waves and Applications,
Vol. 21, No. 3, 375–386, 2007.
Progress In Electromagnetics Research, PIER 81, 2008 459

14. Jung, J. H., H. T. Kim, and K. T. Kim, “Comparisons

of four feature extraction approaches based on Fisher’s linear
discriminant criterion in radar target recognition,” Journal of
Electromagnetic Waves and Applications, Vol. 21, No. 2, 251–265,
2007.
15. Wang, S. G., X. P. Guan, X. Y. Ma, D. W. Wang, and Y. Su,
“Calculating the poles of complex radar targets,” Journal of
Electromagnetic Waves and Applications, Vol. 20, No. 14, 2065–
2076.
16. Alivizatos, E. G., M. N. Petsios, and N. K. Uzunoglu, “Towards
a range-doppler UHF multistatic radar for the detection of non-
cooperative targets with low RCS,” Journal of Electromagnetic
Waves and Applications, Vol. 19, No. 15, 2015–2031, 2005.
17. Capineri, L., D. Daniels, P. Falorni, O. Lopera, and C. Windsor,
“Estimation of relative permittivity of shallow soils by using the
ground penetrating radar response from different buried targets,”
Progress In Electromagnetics Research Letters, Vol. 2, 63–71, 2008.
18. Abdelaziz, A. A., “Improving the performance of an antenna array
by using radar absorbing cover,” Progress In Electromagnetics
Research Letters, Vol. 1, 129–138, 2008.
19. Razevig, V. V., S. I. Ivashov, A. P. Sheyko, I. A. Vasilyev,
and A. V. Zhuravlev, “An example of holographic radar
using at restoration works of historical building,” Progress In
Electromagnetics Research Letters, Vol. 1, 173–179, 2008.
20. Hebeish, A. A., M. A. Elgamel, R. A. Abdelhady, and
A. A. Abdelaziz, “Factors affecting the performance of the
radar absorbant textile materials of different types and structus,”
Progress In Electromagnetics Research B, Vol. 3, 219–226, 2008.
21. Chan, Y. K. and V. C. Koo, “An introduction to synthetic
aperture radar (SAR),” Progress In Electromagnetics Research B,
Vol. 2, 27–60, 2008.