This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1

Radar Change Imaging With Undersampled Data

Based on Matrix Completion and Bayesian
Compressive Sensing
Hui Bi, Chenglong Jiang, Bingchen Zhang, Zhengdao Wang, Member, IEEE, and Wen Hong, Member, IEEE

Abstract—Matrix completion (MC) is a technique of recon- In order to mitigate the effect of additive noise and clutter,
structing a low-rank matrix from a subset of matrix elements. This Bayesian CS (BCS) has been used for sparse target reconstruc-
letter proposes an approach for change imaging from undersam- tion [8], [9]. Instead of obtaining a point estimate of an image
pled stepped-frequency-radar data via MC. We demonstrate that [6], [7], a full posterior density function is constructed in the
MC can be used to reconstruct the unknown samples. Based on
the recovered full sample data, we then perform the estimation of BCS approach, which offers sparser imaging than conventional
the change image using a Bayesian compressive sensing (BCS) ap- CS reconstruction methods for cluttered and noisy scenes.
proach. Compared with existing compressive sensing (CS)-based Matrix completion (MC) is a technique for recovering a low-
techniques, which are sensitive to noise and clutter, the proposed rank matrix from a subset of matrix elements [10], [11]. For
method reduces the false-alarm rate and achieves sparser change undersampled data, if the low-rank property of a reshaped data
imaging, which is due to more available data offered by MC matrix is satisfied, the unobserved data can be recovered from
and our explicit consideration of clutter and additive noise in the
the known samples by minimizing an objective function with a
imaging procedure. The effectiveness of the proposed method is
validated with experimental results based on raw radar data. penalty term based on the matrix nuclear norm.
In this letter, we propose a new high-resolution change
Index Terms—Change imaging, compressive sensing (CS), ma- imaging scheme based on MC and the BCS for undersampled
trix completion (MC), stepped-frequency radar.
stepped-frequency-radar data. As depicted in Fig. 1, our scheme
first uses MC to estimate the unobserved frequency samples.
I. I NTRODUCTION Then, the BCS method is utilized to obtain the change image
based on completed data. The main contributions are as follows:
C HANGE imaging tries to recover the change portion of
an imaged region using collected data before and after
a change. It has been widely used in medical, hyperspectral,
1) a novel low-rank formulation of the measurement model that
enables the recovery of unknown frequency samples by MC;
optical, and radar imaging [1]–[4]. 2) the application of the BCS to reconstruct a change image
In conventional radar imaging, a sampling rate higher than based on completed data; 3) the demonstration of the feasibility
that dictated by the Nyquist–Shannon theory is typically em- and superiority of the proposed method using experimental un-
ployed. For high-resolution imaging, the resulting high sam- dersampled stepped-frequency-radar data and the comparison
pling rate poses difficulties for raw data transmission and with competing methods.
storage. Exploiting the fact that change images are naturally The rest of this letter is organized as follows. Section II intro-
sparse in the raw image domain, compressive sensing (CS) [5] duces the basic imaging mechanism of the stepped-frequency
can be used for the reconstruction of the change portion of the radar, the change imaging model, and MC for undersampled
imaged region. For the sparse change scene, CS can achieve change radar data. In Section III, we present the derivation de-
high-resolution change imaging with much less collected data tails of the proposed method. Section IV presents experimental
than those required by the Nyquist–Shannon theory. Recently, results based on undersampled stepped-frequency-radar data,
several works based on CS have been conducted in radar change along with a performance comparison with competing methods.
imaging, e.g., see [6] and [7]. Despite the advantages of CS- Conclusions are drawn in Section V.
based approaches, they are sensitive to noise and clutter, which
leads to an increased false-alarm rate at low signal-to-clutter- II. C HANGE I MAGING M ODEL
and-noise ratio (SCNR) and, hence, low imaging quality. A. Stepped-Frequency Radar Imaging Mechanism
When a stepped-frequency radar scans a region of interest
(the surveillance region) at each scan position, the transmitted
Manuscript received October 21, 2014; revised February 21, 2015; accepted signal at the mth frequency can be written as
March 10, 2015. This work was supported by the National Basic Research  
t
Program of China (973 Program) under Grant 2010CB731905. sm (t) = rect e−j2πfm t (1)
H. Bi, C. Jiang, B. Zhang, and W. Hong are with the National Key Laboratory T
of Microwave Imaging Technology, Institute of Electronics, Chinese Academy
of Sciences, Beijing 100190, China (e-mail: [email protected]). where m = 1, 2, . . . , M , with M being the total number of fre-
Z. Wang is with the Department of Electrical and Computer Engineering, quencies; T is the duration time of the transmitted signal; and
College of Engineering, Iowa State University, Ames, IA 50010 USA. fm = f0 +mΔf , with Δf being the frequency step interval and
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. f0 being the minimum frequency. The rect(t) function is a rec-
Digital Object Identifier 10.1109/LGRS.2015.2412677 tangular pulse function of a duration of 1 s and an amplitude of 1.
1545-598X © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

Fig. 1. Flow diagram of the proposed change imaging method.

We assume that the surveillance region is rectangular, with the change imaging model can be expressed as
Nu pixels in azimuth and Nv pixels in range, and we char-
acterize a point by its 2-D index (u, v), where 1 ≤ u ≤ Nu , ΔyM N ×1 = ΦM N ×L ΔxL×1 + ΔwM N ×1 (7)
and 1 ≤ v ≤ Nv . Let L = Nu Nv , and let σ(u, v) be the com-
plex reflectivity at point (u, v). Let X denote an Nu × Nv where Δy = y(2) − y(1) is the change data, Δx = x(2) − x(1)
matrix whose (u, v)th entry is σ(u, v). Let x = vec(X), where is the change image of the surveillance region, and Δw is the
vectorization operation vec(·) stacks the columns of X one Gaussian noise, which is assumed to be circularly symmetric
after another. Let a represent the floor of a nonnegative real complex Gaussian distributed with independent and identically
number a. For 1 ≤ l ≤ L, define distributed entries of zero mean and variance η 2 .

vl := (l − 1)/Nu  + 1 ul := l − (vl − 1)Nu . (2)

C. MC
The lth entry of x is then σ(ul , vl ). In a practical imaging procedure, which is motivated by the
Considering the imaging geometry between the radar and the need to reduce the amount of data to be transmitted and stored,
surveillance region, the collected data of the mth frequency at undersampling may be employed. The reduction of samples, if
the nth scan position can be expressed as not treated, can lead to the appearance of false-alarm targets,
L   increase the reconstruction error, and thereby lower the quality
t − τn (ul , vl ) of image. In order to address this problem, we propose to use
ym,n (t) = σ(ul , vl )rect
T MC in this letter for the recovery of missing frequency samples
l=1
· e−j2πfm τn (ul ,vl ) (3) based on known data.
At the nth scan position, the mth element of the change data
where n = 1, 2, . . . , N , with N being the total number of scan Δyn = [Δy1,n , Δy2,n . . . , ΔyM,n ]T , in the absence of noise,
positions; and Rn,l and τn (ul , vl ) = 2Rn,l /c are the distance can be expressed as
and the signal delay time between the nth radar scan position L  
t − τn (ul , vl )
and point (ul , vl ), respectively. Δym,n (t) = Δσ(ul , vl )rect
Let yn = [y1,n , y2,n . . . , yM,n ]T represent the total collected T
l=1
data at the nth scan position. We can express (3) as · e−j2πfm τn (ui ,vi ) (8)
yn = Φn x, x ∈ CL×1 (4) where Δσ(ul , vl ) is the complex reflectivity of the lth point of
the change image.
where Φn ∈ CM ×L is the measurement matrix according to In order to represent the phase relationship accurately, we
the imaging geometry at the nth position, with Φn (m, l) = ignore the complex reflectivity that is the constant term. Let us
e−j2πfm τn (ul ,vl ) . consider a typical change point at first, e.g., (u1 , v1 ). We can
Similarly, considering the imaging formation of all scan rewrite (8) as
positions, we can express the total imaging process as  
t − τn (u1 , v1 ) −j2πfm τn (u1 ,v1 )
yM N ×1 = ΦM N ×L xL×1 (5) Δym,n (t) = rect e . (9)
T
where y = [y1T , y2T , . . . , yN ] and Φ = [ΦT1 , ΦT2 , . . . , ΦTN ]T
T T
In (9), we can see that the entries of the change vector Δyn
represent the total collected data and the measurement matrix, satisfy Δyp+q,n = Δyp,n Δyq,n . If we reshape Δyn into a matrix
respectively. A1,n whose size is p×q, where p and q are two integers such
that p×q = M is satisfied, the reshaped matrix can be expressed
B. Change Imaging Model as the product of a column vector and a row vector as follows:
Our goal is to reconstruct the change image of the surveil- ⎡ ⎤
Δy1,n Δyp+1,n ··· Δy(q−1)p+1,n
lance region by using the undersampled frequency data of the ⎢Δy2,n
⎢ Δyp+2,n ··· Δy(q−1)p+2,n ⎥
⎥
stepped-frequency radar before and after a change. We assume A1,n =⎢ . . .. ⎥ (10)
⎣ .. .. .. ⎦
that the imaging geometry and the radar-emitted frequencies are . .
unchanged in the two data collection processes. The imaging Δyp,n Δy2p,n ··· Δyqp,n
formulation before and after a change can be expressed as ⎡ ⎤
Δy1,n
⎢Δy2,n ⎥
y(1) = Φx(1) + w(1) y(2) = Φx(2) + w(2) (6) ⎢ ⎥
=⎢ . ⎥ 1, Δyp,n , . . . , Δy(q−1)p,n . (11)
(1) (2) ⎣ .. ⎦
where w and w are the additive zero-mean Gaussian noise.
According to the stepped-frequency-radar imaging model (5), Δyp,n
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BI et al.: RADAR CHANGE IMAGING WITH UNDERSAMPLED DATA BASED ON MC AND BCS 3

The rank of the product of the two matrices is upper bounded

by the minimum of the ranks of the two matrices. Since matrix
A1,n is a product of two vectors (rank-1 matrices), the rank of
matrix A1,n is 1 as long as A1,n is nonzero. We know that each
change point contributes to the reshaped change image, i.e., a
term similar to (10). The reshaped matrix due to all change
points is the sum of several such rank-1 matrices. Therefore, Fig. 2. Experimental stepped-frequency radar.
the rank of the reshaped change image matrix An satisfies
 the full posterior because the Laplacian prior is not a conjugate
 prior1 for a Gaussian likelihood function; hence, the associated
rank(An ) = rank Al,n (12) Bayesian inference may not be performed in the closed form.
 l Instead of using a Laplacian prior, a hierarchical prior is
≤ rank(Al,n ) = Δx0 (13) used in [8]. The full posterior density function of Δx can be
l expressed as a multivariate Gaussian distribution as follows:
where Al,n is the reshaped matrix due to the lth change point. p(Δx|α, α0 ) = N (Δx|μ, Σ) (17)
Compared with the number of nonzero entries in the whole
surveillance region, the change portion is much smaller, i.e., where α = (α1 , α2 , . . . , αl , . . . αL ), αl is the inverse variance
Δx0  L. Therefore, the reshaped matrix An is a low- of the Gaussian density function, and α0 is the inverse of the
rank matrix with partially observed elements. Let Ω denote noise. The conditional mean of Δx given Δy can be written as
the indices of the observed elements in An ; we can recover
μ = α0 ΣΦT Δy (18)
the unknown entries of An via MC by solving the following
optimization problem: where
 
  Σ = (α0 ΦT Φ + A)−1 (19)
min Ân ∗ s.t. PΩ (Ân − An ) < δ (14)
F
with A = diag(α1 , α2 , . . . , αL ). By marginalizing over Δx,
where Ân is the estimated matrix with completed entries,  · ∗ the marginal likelihood for α and α0 can be expressed as
is the nuclear norm that is the sum of singular values, and L(α, α0 ) = − L log π + log |C| + ΔyT C−1 Δy (20)
δ is a constant that is decided based on the noise level. For
example, if the noise variance is σ 2 , we should 2 −1 T
 choose δ to be with C = σ I + ΦA Φ . Using a maximum likelihood ap-
a small positive number (e.g., 2–4) times |Ω|σ 2 , where |Ω| proximation, α and α0 can be estimated as [8]
denotes the cardinality of set Ω. In (14), PΩ : Cp×q → Cp×q is γl
αlnew = , l = 1, 2, . . . , L (21)
a projection operator defined by the following: μl 
 L − l γl
  (Ân −An )n1 ,n2 , (n1 , n2 ) ∈ Ω α0new = (22)
PΩ (Ân −An ) = (15) Δy − Φμ22
n1 ,n2 0, otherwise
where γl = 1 − αl Σll . From (18), (19), (21), and (22), we can
where n1 = 1, 2, . . . , p, and n2 = 1, 2, . . . , q. After the afore- see that μ and Σ are functions of α and α0 , respectively,
mentioned recovery, Ân should be reshaped back into a vector. whereas αnew and α0new are functions of μ and Σ, respectively.
We remark that, in order for the proposed method to work, Therefore, we can use an iterative algorithm to reconstruct the
matrix An should be rank deficient. Since its rank can be as change image.
large as Δx0 , cf., (13), we should choose min(p, q) > Δx0 . In summary, our proposed method first takes the difference
of the undersampled data before and after the change. Then,
III. BCS FOR C HANGE I MAGING the unobserved change data are estimated by the MC method
through solving the problem in (14). The BCS iterations based
In order to reduce the impact of noise and clutter, we next on (18), (19), (21), and (22) are then performed to reconstruct
apply the BCS to the complete data to obtain the final change the change image.
image. The method applies the existing results in literature [8],
[9] and provides an estimate of the full posterior density for
IV. E XPERIMENT
change image Δx instead of merely a point estimate.
Based on the assumption of the noise term in (7), we can In this section, experimental results based on raw radar data
write the likelihood function of the collected data as are reported to illustrate the validity and effectiveness of the
   −M N −(1/η2 )Δy−ΦΔx2 proposed change imaging method. For comparison, we also
p Δy|Δx, η 2 = πη 2 e . (16) show the reconstructed results by back projection (BP) [13] and
the CS [14] method.
With the prior knowledge that Δx is sparse, the change imag- Our experiments exploit the stepped-frequency radar with
ing problem can be converted to a linear regression problem. one transmit–receive pair that is located on a linear orbit with
With the BCS, our aim is to construct a full posterior density a length of 2 m, as illustrated in Fig. 2. We use ten metallic
function for change image Δx and noise variance η 2 .
In a Bayesian formulation, the fact that Δx is sparse is 1 In the Bayesian probability theory, if the resulting posterior distributions
formulated by placing a sparsity-promoting prior on Δx. The
p(θ|x) are in the same family as prior probability distributions p(θ), we say
Laplace density function is a broadly used sparseness prior [12]. that a class of p(θ) is a conjugate prior for the class of likelihood functions
However, the Laplacian prior may not offer us an estimate of p(x|θ).
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

Fig. 3. Surveillance scene of the experiment. (a) Scene before the change.
(b) Scene after the change.
TABLE I
E XPERIMENTAL PARAMETERS

Fig. 4. Change imaging results with 100% measured frequencies interrupted

by the 5-dB noise and clutter by using the (a) CS and (b) BCS methods.

balls with the same size and reflectivity, which are uniformly
distributed along the azimuth within the surveillance region.
The distance of adjacent two balls is 15 cm. Eight balls are Fig. 5. Experimental results. (a) BP change imaging result with 100% mea-
sured frequencies. (b) BP change imaging result with 25% random measured
chosen as the change portion of the scene. Images showing frequencies. (c) BP change imaging result based on MC with 25% random mea-
the surveillance scene before and after the change are given in sured frequencies. (d) CS change imaging result with 25% random measured
Fig. 3. Experimental parameters are listed in Table I. frequencies. (e) Proposed method change imaging result with 25% random
In the experiment, we first collect all frequency data and then measured frequencies. (f) Change imaging results of CS and the proposed
method with 25% random measured frequencies in an azimuth slice.
choose the frequency samples according to an undersampled
ratio by randomly and identically discarding the collected fre-
quency bins before and after the change. This is equivalent to of the BCS compared with that of the CS at a higher SCNR,
only collecting random undersampled frequency data, and it which is the case in our experiment due to the good reflectivity
enables us to perform controlled experiments by varying the of the metallic balls. To perform meaningful comparisons, we
undersampled ratio. artificially introduced some noise and clutter with a 5-dB SCNR
to the raw data, and then, we used the CS and BCS methods to
reconstruct the change scene with 100% measured frequencies
A. Noise and Clutter Suppression of BCS disturbed by the 5-dB noise and clutter. As shown in Fig. 4,
Due to the fact that the proposed method is a combination we can see that both the BCS and CS methods can reconstruct
of MC and the BCS, it is critical to isolate their contributions to the change scene, whereas the BCS can suppress the noise and
the improvement of the reconstructed change image quality. We clutter effectively.
first illustrate the effectiveness of the BCS for the suppression
of noise and clutter. The contributions of MC will be introduced
B. Change Imaging
in the next section. Define the SCNR as
Pdata To illustrate the validity of the proposed method comprehen-
SCNR = 10 × log (23) sively, as shown in Fig. 5, we compare the reconstructed results
Pnc of the different methods with 100% frequency data and 25% un-
where Pdata and Pnc are the power of the received frequency dersampled data. Fig. 5(a) depicts the change imaging result by
data and the power of noise and clutter, respectively. It is the BP method with 100% frequency data. Eight metallic balls
difficult to distinguish the effectiveness of the noise suppression are focused in the center clearly. The change images formed by
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BI et al.: RADAR CHANGE IMAGING WITH UNDERSAMPLED DATA BASED ON MC AND BCS 5

Fig. 6(a) and (b) shows the undersampled ratio, i.e., SCNR,
phase diagrams formed by CS and the proposed method, and
Fig. 6(c) represents the 50% successful recovery probability
curve of the two phase diagrams. The results in Fig. 6 show
that the successful recovery region of the proposed method is
larger than that of CS, i.e., the proposed method outperforms
CS in change imaging when the false-alarm rate and the missed
detection rate are used for gauging the imaging performance.

V. C ONCLUSION
In this letter, we have proposed a new method for change
imaging from stepped-frequency-radar undersampled data. Ex-
ploiting the low-rank property of the stepped-frequency-radar
data reshaped matrix, it recovers the undersampled frequency
data by the MC method and then uses the completed data to
reconstruct the change image by means of the BCS method.
Compared with the conventional BP and CS methods, this new
approach can suppress the sidelobe and reduce the false-alarm
rate significantly, and it can achieve high-resolution change
imaging with undersampled frequency data. The experimental
results verified the validity of the proposed method.
Fig. 6. Two-dimensional undersampled ratio, i.e., SCNR, phase diagrams.
The blue region means the 100% successful recovery, the red region means R EFERENCES
the 0% successful recovery, and the transition area means the success from 1
to 0. (a) Phase diagram of the CS method. (b) Phase diagram of the proposed [1] J. Thirion and G. Calmon, “Deformation analysis to detect and quan-
method. (c) 50% successful recovery probability curve of the two methods. tify active lesions in three-dimensional medical image sequences,” IEEE
Trans. Med. Imag., vol. 18, no. 5, pp. 429–441, May 1999.
[2] M. Molinier, J. Laaksonen, and T. Hame, “Detecting man-made structures
and changes in satellite imagery with a content based information retrieval
four methods with 25% frequencies are shown in Fig. 5(b)–(e), system built on self-organizing maps,” IEEE Trans. Geosci. Remote Sens.,
respectively. We can see that the BP and CS methods cannot vol. 45, no. 4, pp. 861–874, Apr. 2007.
recover the change scene owing to the appearance of several [3] W. Liang, D. Hoja, M. Schmitt, and U. Stilla, “Comparative study of
change detection for reconstruction monitoring based on very high res-
false-alarm targets in the same azimuth cell of change balls, olution optical data,” in Proc. JURSE, 2011, pp. 73–76.
whereas the BP method can recover the image with less false- [4] H. Wan, “Novel change detection in SAR imagery using local connectiv-
alarm targets after MC for the undersampled data, as shown ity,” IEEE Trans. Geosci. Remote Sens. Lett., vol. 10, no. 1, pp. 174–178,
in Fig. 5(c). Compared with the aforementioned algorithms, Jan. 2013.
[5] D. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52,
the proposed method reconstructs the image with 25% random no. 4, pp. 1289–1306, Apr. 2006.
frequency samples successfully [see Fig. 5(e)]. To illustrate the [6] C. Kreucher and M. Brennan “A compressive sensing approach to multi-
validity of the proposed method intuitively, one azimuth slice static radar change imaging,” IEEE Trans. Geosci. Remote Sens., vol. 52,
corresponding to the center of a change ball is extracted. In no. 2, pp. 1107–1112, Feb. 2014.
[7] S. Zhu, G. Liao, W. Wang, C. Zeng, and D. Yang, “Wideswath syn-
Fig. 5(f), we can see that the recovered result of the proposed thetic aperture radar ground moving targets indication with low data rate
method is quite close to the full-sampled reconstructed image, based on compressed sensing,” IET Radar, Sonar Navig., vol. 7, no. 9,
without any false-alarm target. pp. 1027–1034, Dec. 2013.
[8] S. Ji, Y. Xue, and L. Carin, “Bayesian compressive sensing,” IEEE Trans.
Signal Process., vol. 56, no. 6, pp. 2346–2356, Jun. 2008.
[9] J. Xu, Y. Pi, and Z. Cao, “Bayesian compressive sensing in synthetic
C. Performance Versus Undersampled Ratio and SCNR aperture radar imaging,” IET Radar, Sonar Navig., vol. 6, no. 1, pp. 2–8,
Jan. 2012.
To compare the proposed method with CS in terms of the re- [10] E. Candès and B. Recht, “Exact matrix completion via convex op-
constructed quality, we conducted the experiment with the fixed timization,” Found. Comput. Math, vol. 98, no. 6, pp. 925–936,
Jun. 2010.
sparsity of the change scene, but we varied the undersampled [11] E. Candès and Y. Plan, “Matrix completion with noise,” Proc. IEEE,
ratio of the frequency and the SCNR, and obtained the phase vol. 98, no. 6, pp. 717–772, Jun. 2009.
diagram of two methods [5]. Suppose Pfa and Pmd represent [12] J. Bernardo and A. Smith, Bayesian Theory. New York, NY, USA:
the false-alarm rate and the missed detection rate, respectively, Wiley, 1994.
[13] L. M. H. Ulander, H. Hellsten, and G. Stenstrom, “Synthetic-aperture
and ε is a threshold constant. In a fixed combination of the radar processing using fast factorized back-projection,” IEEE Trans.
undersampled ratio and the SCNR, if condition Pfa + Pmd < ε Aerosp. Electron. Syst., vol. 39, no. 3, pp. 760–776, Jul. 2003.
is satisfied, we view the change image as reconstructed suc- [14] C. Jiang, B. Zhang, W. Hong, and Y. Wu, “Efficient Lq regularization
cessfully. For each set of the aforementioned two parameters, algorithm with range-azimuth decoupled for SAR imaging,” Electron.
Lett., vol. 50, no. 3, pp. 204–205, Jan. 2014.
the successful recovery rate is taken over 50 replications. We [15] D. Donoho and V. Stodden, “Breakdown point of model selection when
then plot the successful recovery rates for all sets of parameters the number of variables exceeds the number of observations,” in Proc.
to generate the phase diagram. IJCNN, 2006, pp. 1916–1921.