IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 4, NO.

1, JANUARY 2007 75

Eigensubspace-Based Filtering With Application in

Narrow-Band Interference Suppression for SAR
F. Zhou, R. Wu, M. Xing, and Z. Bao, Senior Member, IEEE

Abstract—Synthetic aperture radar (SAR) has found wide ap- signal-processing gain in both the range and azimuth dimen-
plications in many areas, e.g., battlefield awareness. However, sion. After matched-filter processing, the strong pulse signal
SAR is vulnerable to various kinds of interference, among which is depressed in the time domain and appears as noise in the
narrow-band interference (NBI) is commonly used. In this letter,
an eigensubspace-based filtering approach is proposed for NBI SAR imagery. However, if the interference power is too high,
suppression in SAR without using passive-sniff data as the refer- the useful signal will be buried in the noise.
ence signal. Moreover, the proposed method can deal with smart Some methods [2]–[6] have been proposed for NBI suppres-
or interrupted NBI. Both simulation and experimental results are sion in SAR. These methods can be classified into two cate-
provided to illustrate the performance of the proposed approach. gories. One is based on temporal-domain interference modeling
Index Terms—Interference suppression, motion-parameter esti- and cancellation. The other one is based on time–frequency
mation, radar imaging, synthetic aperture radar (SAR). domain-interference detection and notch filtering. Temporal-
domain modeling-based approaches are not robust, since their
I. I NTRODUCTION performance relies heavily on the data models used and
the model parameter-estimation accuracy. Frequency-domain

D UE TO its day and night, all-weather, long-range, and

wide-mapping capability, synthetic aperture radar (SAR)
has found wide applications in many areas, especially in the
notch filtering is a nonparametric approach and is easy to
be implemented. However, it will lead to discontinuity of the
frequency-domain data. This will generate artifacts in SAR
field of battlefield awareness. For the purpose of battlefield imagery due to the rise of sidelobes and lower the SNR of
awareness, SAR must have embedded interference-suppression the SAR images. Moreover, this discontinuity will affect the
ability. Doppler parameter estimation needed by the motion compen-
Narrow-band interference (NBI) is commonly used in mod- sation of SAR [7], [8] and, hence, further degrades the SAR
ern warfare. By NBI, we mean the bandwidth of the interfer- image quality. An adaptive coherent method was proposed in
ence signal is much less than that of the transmitted signal [9] and [10] to suppress multiple wide-bandwidth as well as
of SAR (say 1% [1]). The NBI of SAR can be classified as narrow-bandwidth RFI sources in SAR. The approach is based
natural and man-made artificial interference. Low-frequency on a 2-D adaptive filtering of the active SAR data using the
SAR is usually affected by natural radio-frequency interference passive-sniff data as the reference data.
(RFI), including the noncoherent electromagnetic interference In this letter, a method is proposed, which features rough
coming from communication devices, television networks, and NBI detection in the frequency domain and adaptive NBI
other radiation sources whose working frequency falls into the suppression in the temporal domain. It constructs an inter-
frequency band of SAR. This natural NBI tends to reduce ference subspace from the raw data and, then, subtracts the
the signal-to-interference and signal-to-noise ratio (SNR) of projected components of the original data onto the interfer-
the SAR image and leads to great performance degradation ence subspace. It is a nonparametric approach, which is more
of the SAR imagery. In case the power of the NBI is strong, robust than conventional temporal modeling-based approach.
bright lines will appear in the SAR image and the image will Compared with a frequency-domain notch-filtering approach,
be blurred. Man-made NBI usually has large power and is the proposed method keeps the continuity of the frequency-
easy to generate. These strong interference signals would bury domain data, and hence, the drawbacks of the former can be
the useful signals and affect the target detection and imaging avoided, including image artifacts due to high sidelobes and
capability of SAR. For example, strong pulse interference is poor Doppler parameter-estimation performance. Compared
a common NBI type. When its power is very strong, it will with the adaptive-filtering method proposed in [9] and [10], our
produce false targets in conventional radar. However, SAR uses method uses only one-dimensional adaptive eigensubspace [11]
two-dimensional (2-D) matched filtering, which has inherent filtering without additional passive-sniff data. That is to say,
interference-suppression ability, since it has a large coherent the proposed method is more compatible with existing SAR
systems. Moreover, our proposed NBI suppression is a pulse-
by-pulse approach, which can be applied to deal with smart NBI
Manuscript received January 5, 2006; revised August 8, 2006. This work was (or interrupted NBI).
supported by the National Natural Science Foundation of China under Grants
60325102, 60428101, and 60502044. The remainder of this letter is organized as follows. The
F. Zhou, M. Xing, and Z. Bao are with the State Key Laboratory for Radar eigensubspace-based NBI-suppression approach will be pre-
Signal Processing, Xidian University, Shaanxi 710071, China. sented in Section II. Simulation and experimental results will be
R. Wu is with the Tianjin Key Laboratory for Advanced Signal Processing,
Civil Aviation University of China, Tianjin 300300, China. provided in Sections III and IV, respectively. Finally, Section V
Digital Object Identifier 10.1109/LGRS.2006.887033 concludes the whole letter.

1545-598X/$25.00 © 2007 IEEE

76 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 4, NO. 1, JANUARY 2007

II. E IGENSUBSPACE -B ASED A DAPTIVE

NBI-S UPPRESSION A PPROACH
To effectively suppress NBI, we need to first detect and
recognize the interference signal. The conventional notch-
filtering method requires that we first detect the interference
position in the frequency domain of the original SAR data.
Conventional detection and recognition methods include the
time–frequency and range–spectrum methods [2]. Our method
requires only rough inspection of the original data in the
frequency domain to declare if NBI exists or not. Take range
fast Fourier transform (FFT) over the original data. In the range
frequency domain, if the magnitude has an abrupt change at a
particular or some frequency bins and the peak value is much
higher than others, then we can declare that the data may have
NBI interference. We do not need the exact position of the
interference for the followed suppression. The check is pulse-
by-pulse based so that we can deal with the smart or interrupted
NBI. Below, we present our NBI-suppression approach.
The SAR data containing NBI can be expressed as [4]

X=S+F+N (1)

where S denotes interested signal, F denotes NBI, and N

denotes receiver noise. The basic signal processing of SAR
is range compression and azimuth coherent integration. For Fig. 1. Block diagram of the proposed NBI-suppression method.
a linear frequency-modulated signal transmitted by SAR, the
· · · > λL and assume their corresponding eigenvectors are
product of the time width and bandwidth is very large, which
denoted by u1 , u2 , . . . , uL , respectively. Then, we can find
requires the NBI to have a very large power so that it can
interfere with the SAR system, i.e., the NBI has very high r(r < L) dominant eigenvalues, and the subspace constructed
peak in the narrow bandwidth. Before 2-D matched filtering, by the corresponding eigenvectors is called the interference
compared with the NBI, the interested signal has white noise subspace, which can be expressed as
like slow-varying power spectrum.
F = span{u1 , u2 , . . . ur }, 1 ≤ r < L. (5)
SAR transmits and receives a series of pulses during the
coherent processing interval. Assume that there are M range Since in SAR, the power of the NBI is much higher than
samples (fast time samples) at each pulse position and there are that of target signal and noise, the above interference subspace
N pulses (slow time or azimuth samples) during the coherent corresponds to the interference data. The signal and noise lie in
processing interval. For simplicity and without loss of general- the subspace spanned by the rest of other eigenvectors.
ity, below, we ignore the index related to the pulse number. Now, project xk onto the interference subspace F, we can get
Assume that xm denotes the mth range sample (m = the interference data, which has the form
1, 2, . . . , M ) and x =
x1 , x2 , . . . , xM T denotes the received
data vector at some pulse, where (•)T denotes transpose. r


We can divide data vector x into K subvectors with dimen- fk = uT
i xk ui . (6)
sion L with K = M − L + 1. Define the kth subvector (k = i=1

1, 2, . . . , K) as
Then, subtract the contribution of the interference from the
T
xk = [xk , xk+1 , . . . , xk+L−1 ] . (2) original data vector xk ; we can obtain the data with NBI
removed, which has the form
Then, we can construct an L × K data X r


x̂k = xk − fk = xk − uT
i xk ui . (7)
X =
x1 , x2 , . . . , xk , k = 1, 2, . . . , K. (3)
i=1

Form (3), we can obtain the estimate of the interference corre- Similar to (3), construct a new data matrix corresponding to the
lation matrix, which has the form data with NBI removed
K
1
X̂ =
x̂1 , x̂2 , . . . , x̂k .
R̂ = xk xT
k. (4) (8)
M
k=1
Then, rearranging the data matrix X̂(L × K), we can get a
Then, we perform eigenvalue decomposition with R̂. new data vector x (M × 1), which corresponds to the data after
Arrange the eigenvalues in the descending order λ1 > λ2 > NBI suppression.
ZHOU et al.: EIGENSUBSPACE-BASED FILTERING 77

Fig. 2. Block diagram of the SAR imaging algorithm with NBI suppression.

TABLE I
SYSTEM PARAMETERS USED IN THE SIMULATION

Fig. 3. Comparison of SAR images of simulated data obtained via different

methods. (a) Without NBI suppression. (b) NBI suppression with notch filter-
Repeat the above process on a pulse-by-pulse basis until N ing. (c) NBI suppression with the proposed method.
pulses have been processed. The block diagram of the proposed
NBI-suppression method is shown in Fig. 1.
The choice of L is very important. If L is too small, then range-Doppler (R-D) algorithm [12] is shown in Fig. 3(a), from
the frequency resolution of NBI is too low, which leads to the which it can be noted that due to the high power of the NBI,
decease of the peak sidelobe ratio of the impulse response of the the target is completely buried in the interference and cannot be
range compressor. If L is too large, the estimation performance detected. The SAR image obtained using notch filtering (zero-
of the interference correlation matrix R̂ becomes poor and the filling those frequency bins whose magnitude is higher than a
interference subspace has more perturbation, which leads to the threshold, say five to eight times of the averaged magnitude of
performance degradation of the NBI suppression. Moreover, those interference-free regions) and R-D imaging algorithm is
too large an L will increase the computational load. There is shown in Fig. 3(b), from which it can be noted that artifacts
a tradeoff in the choice of L. Usually, we choose K ≥ 2L. have appeared in the SAR imagery. Because we only take range
As pointed out before, NBI suppression is pulse-by-pulse FFT over the data vector received at each pulse, the frequency
based. After all pulses have been processed, we can obtain the resolution is not very high, hence, the notch width is very wide,
data with NBI removed. Then, we can finish the SAR imaging which gives rise to high sidelobes. Moreover, the wide notch
task with these data. Fortunately, our proposed NBI-removal width will degrade the followed Doppler parameter-estimation
algorithm has very good compatibility with the existing SAR performance, which is vital for the automated focus of SAR
imaging algorithms. A combination of our proposed NBI sup- imagery. The SAR image obtained using our proposed method
pression with the SAR imaging is shown in Fig. 2. is shown in Fig. 3(c), from which it can be noted that the SAR
image is well focused to one point. In our method, we have
used L = 64.
III. S IMULATION R ESULTS
Assume that the SAR is working at broadside mode. We
IV. E XPERIMENTAL R ESULTS
simulate a point target. The system parameters used in our
simulation are listed in Table I. The signal-processing gain In this section, we present experimental results to demon-
of the SAR is the product of the time width and bandwidth. strate the performance of the proposed method. The data con-
In our simulation, the range and azimuth gain is 25 and taining NBI was collected by an airborne stripmap SAR with a
31 dB, respectively. We add two NBI sources working at 6 and resolution of 1 × 1 m. The scene to be imaged is in a rural area.
1.6 GHz, respectively, whose amplitude is 52 dB higher than First, we consider the NBI-detection issue. Fig. 4(a) shows
that of the point target. the amplitude response of a pulse data in the range-frequency
The SAR images obtained via different algorithms are com- domain, from which it can be noted that the magnitude in the
pared in Fig. 3. The SAR image obtained via the conventional middle frequency part is much higher than other regions, and in
78 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 4, NO. 1, JANUARY 2007

Fig. 4. Analysis of measured data with NBI. (a) Amplitude responses of a

pulse data with NBI in the range-frequency domain. (b) Variation of amplitude
at the NBI frequency bin respect to the pulse number.

other regions, the spectrum is very flat (like noise). Fig. 4(b)
shows the magnitude of the highest frequency bin shown in
Fig. 4(b) as a function of the pulse index number, from which
abrupt changes can be observed. From Fig. 4 we can declare
that there exists a very strong smart NBI (or interrupted NBI).
The SAR images obtained using the experimental data via
different algorithms are compared in Fig. 5. Fig. 5(a) shows
the image obtained via the R-D algorithm without NBI sup- Fig. 5. Comparison of SAR images of measured data obtained via different
pression, from which it can be noted that there are multiple methods. (a) Without NBI suppression. (b) NBI suppression with notch filter-
ing. (c) NBI suppression with the proposed method.
interference strips distributed along the range dimension. Part
of the targets was masked by these interference strips. Fig. 5(b)
shows the images obtained via notch filtering plus R-D imaging from which it can be concluded that the SAR-image quality is
algorithm, from which it can be noted that although the inter- much better than the other two.
ference was effectively suppressed, due to the same reason as The eigenvalue distribution of the covariance matrix with and
explained in the previous section, artifacts appeared in the SAR without NBI is compared in Fig. 6. Fig. 6(a) shows the results
imagery and the targets were not well focused, just as predicted. corresponding to the data collected at some azimuth position
Fig. 5(c) shows the image obtained using the proposed method, that NBI does not exist, from which it can be noted that the
ZHOU et al.: EIGENSUBSPACE-BASED FILTERING 79

to the dominant eigenvalue, we can construct the interference

subspace for NBI suppression.

V. C ONCLUSION
In this letter, an eigensubspace-based adaptive filtering is
proposed for NBI suppression in SAR. Compared with existing
algorithms, the proposed method has very good compatibility
with existing SAR imaging algorithms. Moreover, it can be
used to deal with smart or interrupted NBI. Both simulation
and experimental results have proven the effectiveness of the
proposed method.

R EFERENCES
[1] J. D. Taylor, Introduction to Ultra-Wideband Radar Systems. Boca
Raton, FL: CRC, 1994.
[2] S. Buckreuss and R. Horn, “E-SAR P-band SAR subsystem
design and RF-interference suppression,” in Proc. IGARSS, 1998,
vol. 1, pp. 466–468.
[3] A. Potsis, A. Reigber, and K. P. Papathanassiou, “A phase preserving
method for RF interference suppression in P-band synthetic aperture radar
interferometic data,” in Proc. IGARSS, 1999, vol. 5, pp. 2655–2657.
[4] T. Miller, L. Potter, and J. McCorkle, “RFI suppression for ultra wideband
radar,” IEEE Trans. Aerosp. Electron. Syst., vol. 33, no. 4, pp. 1142–1156,
Oct. 1997.
[5] R. T. Lord and M. R. Inggs, “Approaches to RF interference
suppression for VHF/UHF synthetic aperture radar,” in Proc. COMSIG,
1998, pp. 95–100.
[6] ——, “Efficient RFI suppression in SAR using LMS adaptive filter in-
tegrated with range/Doppler algorithm,” Electron. Lett., vol. 35, no. 8,
pp. 629–630, Apr. 1999.
[7] X. Mengdao and B. Zheng, “Motion parameter estimation based SAR
imaging,” Acta Electron. Sinica, vol. 29, no. 12A, pp. 1824–1828,
Dec. 2001.
[8] S. D. Madsen, “Estimating the Doppler centroid of SAR data,” IEEE
Trans. Aerosp. Electron. Syst., vol. 25, no. 2, pp. 134–140, Mar. 1989.
[9] L. Nguyen, M. Ressler, D. Wong, and M. Soumekh, “Enhancement of
back projection SAR imagery using digital spotlighting preprocessing,”
Fig. 6. Eigenvalue distribution of the covariance matrix (a) without and
in Proc. IEEE Radar Conf., Apr. 2004, pp. 53–58.
(b) with interference.
[10] L. Nguyen, T. Tuan, D. Wong, and M. Soumekh, “Adaptive coherent sup-
pression of multiple wide-bandwidth RFI sources in SAR,” Proc. SPIE,
eigenvalue spread is very small. Fig. 6(b) shows the results vol. 5427, pp. 1–16, 2004.
corresponding to the data collected at some azimuth position [11] R. Klemm, Space-Time Adaptive Processing Principles and Applications.
that NBI exists, from which it can be noted that the eigenvalue London, U.K.: Short Run, 1998.
[12] R. Bamler, “A comparison of range-Doppler and wavenumber domain
spread is very large and there is a dominant eigenvalue that SAR focusing algorithm,” IEEE Trans. Geosci. Remote Sens., vol. 30,
corresponds to the NBI. From the engenvector corresponding no. 4, pp. 706–713, Jul. 1992.