Sei sulla pagina 1di 14

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

doi: 10.14355/ijrsa.2015.05.002

www.ijrsa.org

High-Resolution Imaging for Impulse-Based


Forward-Looking Ground Penetrating Radar
Ode Ojowu Jr.*1, Luzhou Xu1, Jian Li,1 John Anderson,2 Lam Nguyen3 , Petre Stoica4
Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL 32611, USA

Department of Electrical and Computer Engineering, Howard University,Washington, DC 20059, USA

Army Research Laboratory, Adelphi, USA

Information Technology Department, Uppsala University, Uppsala, Sweden


ojowuode@ufl.edu;1xuoasis@gmail.com;1li@dsp.ufl.edu;2j-m-anderson@howard.edu;3lam.h.nguyen2.civ@mail;

*1

Peter.Stoica@it.uu.se

Abstract
Forward-Looking Ground Penetrating Radar (FLGPR) has multiple applications, one of which includes its use for detecting
landmines and other buried improvised explosive devices (IEDs). The standard method for generating synthetic aperture radar
(SAR) images for this radar is the backprojection (BP) algorithm, which has poor resolution and high sidelobe problems. In this
paper, we consider using the Sparse Iterative Covariance-based Estimation (SPICE) algorithm and the Spare Learning via
Iterative Minimization (SLIM) algorithm for generating sparse high-resolution images for FLGPR. A pre-processing step, which
involves an orthogonal projection of the received data onto a subspace related to the region of interest is performed, for
decreasing the dimension of the data and for clutter reduction. The SLIM and SPICE algorithms are user-parameter free, and are
capable of providing SAR images with improved resolution. We also use the well-known CLEAN approach for imaging based
on a proposed signal model in the time domain. We show using simulated data that the SPICE and SLIM algorithms provide
higher resolution than CLEAN and the standard BP. Imaging using real data collected via the Synchronous Impulse
Reconstruction (SIRE) radar, a multiple-input multiple-output (MIMO) FLGPR radar developed by the Army Research
Laboratory (ARL), is also presented and used for analysis.
Keywords
Forward-looking Ground Penetrating Radar (FLGPR); SAR Imaging; Landmine Detection; High-Resolution; Orthogonal Projection of Data

Introduction
The global problem of landmines and other buried improvised explosive devices (IEDs) is affecting both military
and civilians alike [1][5], and effective as well as efficient methods for detecting these devices are important in the
world today. Methods for detecting landmines include, but are not limited to, the use of metal detectors, infrared
sensing of the land surface [6], biological sensors such as animals (dogs) [7] and more recently detecting fumes of
the mines using lasers to ionize the air [5]. Radar is an excellent tool for remote sensing applications [8], [9].
Ground penetrating radar (GPR), which transmits an electromagnetic (EM) wave into the ground and examines the
back-scattered returns to determine buried objects, has become a useful tool for effectively detecting landmines
and IEDs [10][12].
When operated in a forward-looking mode, GPR can be applied to the problem of landmine detection as it inspects
the ground surface from a safe stand-off distance as seen in Fig 1. Impulse-based forward-looking ground
penetrating radar (FLGPR) usually transmits a mono-cycle pulse with typical operating frequency range spanning
the UHF and L bands [11], [13]. The low frequency of GPR provides the necessary ground penetrating properties
and the large bandwidth provides the necessary downrange resolution. The cross-range resolution on the other
hand is limited by the antenna beamwidth [14]. Increasing the antenna physical size can improve cross-range
resolution. However, this is limited by physical antenna size constraints. Side-looking synthetic aperture radar
(SAR) techniques improve cross-range resolution by synthesizing a virtual aperture much larger than the physical
aperture. However, in forward-looking mode, the cross-range resolution is limited by the physical radar size [15].

11

www.ijrsa.org

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

A multi-input multi-output (MIMO) radar can be used to enhance this resolution. For example, Fig 1 also shows an
FLGPR for landmine detection built by the Army Research Laboratory (ARL) known as the synchronous impulse
reconstruction (SIRE) radar [13]. This radar consists of 2 transmitter antennas and 16 receiver antennas and exploits
waveform diversity [13], [16], [17] to enhance cross-range resolution by alternatingly transmitting from its two
transmitter antennas.

FIG. 1 SIRE RADAR: IN FORWARD LOOKING MODE [13]

Many moderate resolution techniques have been considered for GPR and remote sensing in general [18][22]. The
conventional imaging method for this type of FLGPR radar is an approach known as the backprojection (BP)
method [23], [24]. This approach, is also known as the delay-and-sum (DAS) approach, and suffers from high
sidelobe problems and poor resolution. High-resolution imaging is important for separating closely spaced targets
as well as distinguishing targets from clutter. Application of well-known data-adaptive methods as well as
subspace based methods [25] to achieve high resolution imaging for this specific problem is limited, due to the
availability of a single snapshot of data as well as the presence of coherent sources.
In this paper, we focus on sparse high-resolution imaging for impulse-based FLGPR. A signal model in the time
domain is established given that the transmitted impulse is well localized in time. Based on this model, first, the
well-known CLEAN (Matching pursuit) approach [26], [27] is analysed for imaging and compared to the standard
BP. This technique eliminates sidelobes, but it provides no real improvement in imaging resolution over the
standard BP. Sparse imaging can also be achieved via an orthogonal matching pursuit using a dictionary of the
transmitted waveform [28], with similar resolution results as the CLEAN algorithm.
Next, two recently proposed, user-parameter free and data-adaptive methods are considered for high-resolution
imaging: the Sparse Learning via Iterative Minimization (SLIM) [29] and the SParse Iterative Covariance-based
Estimation (SPICE) [30] methods which are capable of providing sparse as well as high-resolution imaging results.
These methods are applied to impulse-based FLGPR data of significantly lower dimension to achieve sparse highresolution imaging.
The data dimension reduction is achieved via a pre-processing technique, which involves an orthogonal projection
of the received data onto the subspace spanned by the dominant singular vectors of a steering matrix
corresponding to the region of interest (ROI). An efficient decomposition of this steering matrix is performed using
the eigenvalue decomposition of a matrix of considerably reduced dimension. For real-valued data, the SPICE
optimization problem can be expressed as a linear program (LP) [31] and solved efficiently to speed up the
computation of SPICE. In the appendix of this paper, a conjugate gradient SPICE (CG-SPICE) algorithm is also
introduced. For FLGPR SAR imaging, the CG-SPICE algorithm provides similar computation improvement as
solving the LP (it can however, be used for applications where the data is not necessarily real-valued).
The improvements in imaging results over the standard BP method are shown via simulated data and also by
using experimental data. For the experimental data, the SIRE radar developed by ARL [13], [32] is used for
analysis.
The remaining sections of this paper are organized as follows. In Section II, a proposed data model is presented for

12

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

www.ijrsa.org

FLGPR based on the ARLs SIRE radar. This MIMO radar is also briefly described therein. In Section III, the
standard BP method for imaging is analyzed as well as the ARLs Recursive Sidelobe Minimization (RSM)
algorithm. The RSM algorithm, which is based on the BP algorithm, can effectively suppress sidelobes [13], [33].
Based on the proposed model, we also show that the CLEAN approach can be used for sparse imaging using
impulse-based FLGPR. In Section IV, we present the sparse adaptive methods for improved resolution as well as
the proposed pre-processing step of orthogonal projection for clutter and data dimension reduction. SectionV
contains the results based on numerically simulated data as well as measured data. Finally the conclusions are
drawn in Section VI.

TABLE 1 NOTATIONS

a vector

a matrix

diag(x)
(.)H

a diagonal matrix with elements of x on the diagonal


conjugate transpose of a matrix or vector

(.)T

transpose of a matrix or vector

(.)(n)

nth iteration for a scalar, vector or matrix

. 2

2 norm

estimate of scalar y

definition
FIG. 2 SIRE FLGPR: DATA ACQUISITION PROCESS

Data Model
For impulse-based FLGPR, we consider the SIRE radar which was designed by ARL and mounted on an SUV for
landmine detection [13]. The radar geometry, as seen in Fig.1, consists of two transmitters and sixteen receivers.
Each transmitter emits an impulse with a frequency range of 0.3-3.0 GHz, which determines the downrange
resolution. The cross-range resolution is limited by the physical 2 m width of the SUV. This system can be
considered as a practical example of a MIMO radar which exploits waveform diversity by transmitting orthogonal
waveforms from the two transmit antennas located at the edges of the receive array [17]. The waveform
orthogonality is achieved by alternatingly transmitting narrow pulses (in ping-pong mode [34]) from each
transmitter antenna. This creates a virtual aperture which is almost double the physical 2 m aperture of the receiver
array, thereby improving the cross-range resolution [35].
The data acquisition process for the SIRE radar involves acquiring data as the system moves slowly towards the
area to be imaged (which consists of L grid points) from a maximum stand-off distance to a minimum stand-off
distance as shown in Fig. 2. As the system moves forward, the right and left transmitters emit a wide-band impulse
alternatingly and repeatedly with the returns from scatterers collected and sampled by the 16 receivers. A second
dimension is added to the physical dimension of the receiver array, forming a 2D synthetic aperture which consists
= 1 measurements (with the kth measurement corresponding to a specific transmit-receive pair) that are
then used for the SAR image formation process [13].
It is important to note that each transmitted signal will

illuminate a region larger than the ROI. Let M denote the total number of grid points outside the ROI illuminated
during the data acquisition process. M will depend not only on the footprint size of the radar, but also on the
distance covered by the radar system as it moves from the maximum to the minimum stand-off distance (Note: M
>> L).
Let () denote the tth sample in time of the kth measurement after sampling by the corresponding receiver. This
measurement can be described by the following equation:
() = + , , + () for = 0 1

(1)

13

www.ijrsa.org

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

where T is number of samples collected for each measurement set; , = 1 (, ) (, )is the propagation
path-loss, with (, ), (, ) being the distances from the transmitter to the ith grid point and from that grid
point to the receiver respectively; is the reflectivity of the ith grid point; ()is the thermal noise associated
with kth measurement; and () denotes the transmitted impulse. The delay is given by
int [ ( (, ) + (, ))] where Fs is the sampling frequency, int[.] denotes the nearest integer operator, and c is
the speed of propagation. The model of the kth measurement in (1) takes into account the contribution of M
scatterers outside the ROI for simplicity (although the scatterers that contribute to the kth measurement maybe
fewer). This allows the model to be rewritten as the following linear equation:

= + = +

(2)

where = [1 (0), , 1 ( 1), , (0), , ( 1)] is the vector of measurements stacked together; = { }=1
is the vector to be estimated containing the reflectivities of the desired imaging grid with corresponding steering
matrix A; = { }
=1 is the vector containing the reflectivities of all the grid points outside the ROI illuminated by
the radar system, with corresponding steering matrix B; and the noise vector is denoted by n. Note also that, C =

[A, B] and = { }+
=1 = [ , ] . The matrix A of dimensions consists of delayed and scaled versions of
the transmitted signal:
A=

1,1 (1,1 ) 1, (1, )

,1 (,1 ) , (, )

(3)

where the vector , = ( , )=0 is the transmitted impulse delayed by , . Fig 3 shows a typical wideband impulse used for transmission. The steering matrix B of dimensions () can be written in a similar
fashion to (3). The value of M (typically ) and the matrix B are difficult to determine due to the changing
footprint of the radar system as a result of its forward motion during the data acquisition process.
The data model in (2) is used for FLGPR SAR imaging

in this paper. In the next section, the BP/DAS and RSM


algorithms as well as the CLEAN algorithm are described for SAR imaging.

FIG. 3 TYPICAL TRANSMITTED IMPULSE

Backprojection/Delay-And-Sum Methods
The standard BP is a well-known and widely used algorithm for FLGPR SAR imaging (also known as the delayand-sum algorithm). This algorithm is limited in downrange resolution by the bandwidth of the transmitted
impulse and in crossrange resolution by the physical (or virtual) aperture of the radar. Another limitation of this
algorithm is that it produces images with high sidelobes. A Recursive Sidelobe Minimization algorithm based on
BP/DAS was proposed in [13], [33] for suppressing sidelobes. The CLEAN approach [26] (which is also based on
the BP/DAS approach) can also be used for eliminating sidelobes [27] as well as estimating weak targets by
iteratively subtracting the contributions of stronger targets from the received data. In this section, we describe
these algorithms as well as analyze the CLEAN approach based on the proposed data model for impulse-based
FLGPR SAR imaging.

14

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

www.ijrsa.org

BP/DAS
For landmine detection, a 2D rectangular grid as shown in Fig. 2 is typically used for image formation [13] as
targets are usually flush buried or lie on the surface of the ground. While a 3D grid can also be used for imaging, in
this paper, we focus on the 2D approach for image formation. However, for computing delays, 3D coordinates are
used because the transmitter and receiver are mounted above the ground on an SUV, hence, they have nonnegligible height coordinates. Also, the surface being imaged may not be perfectly flat.
Based on Fig. 2, the BP algorithm can then be described as follows. Consider the ith grid point in this figure with
location ( , , ) relative to a predefined reference point or origin. For a specific transmit-receive pair, let
(, , , , , ) and (, , , , , ) denote the transmitter and receiver locations respectively in this coordinate system
for = 1 . Therefore, the delay of the transmitted EM pulse from the transmitter corresponding to the kth
measurement to the ith grid point and then back to the corresponding receiver is given by:
, = int acq +

2
2
2
, + , + ,

(4) + , + , + ,

where acq (the data acquisition time delay associated with the radar system), is also taken into account. The BP
estimate of the reflection coefficient at the ith grid point within the ROI is given by:
=

1
1

, ,
=1

(5)

This estimate is simply an average of the aligned measurements with the propagation loss compensated by a
weighting factor 1, . This is referred to as coherent processing of the measurements which improves the signalto-noise ratio (SNR) by a factor of K compared to using a single measurement.

The BP/DAS algorithm is limited in resolution and suffers from poor sidelobes. The RSM algorithm, proposed by
ARL, effectively suppresses sidelobes by generating multiple DAS images using 2D apertures with randomly
missing measurements and selecting the minimum value across all images [13]. In the next subsection, we also
present a CLEAN approach based on the data model in (1) for sparse imaging. However,
these BP/DAS-based algorithms do not improve imaging resolution compared to the standard BP/DAS algorithm.
We then present the SLIM and SPICE algorithms for improved resolution of the imaging process, after a preprocessing step via an orthogonal projection for data dimension and clutter reduction.
Clean
The CLEAN algorithm [26] (also known as matching pursuit) for image formation is briefly described here based
on the data model in (1). As already mentioned, the standard DAS algorithm suffers from high sidelobe problems.
CLEAN can be used to eliminate the sidelobes of strong returns so that weak targets become visible [27].
The CLEAN algorithm is used iteratively to find the grid point of the strongest target and then subtract the
contribution of that target from the data. The next strongest target is then found based on the updated
measurements. A step-by-step description of CLEAN based on the data model in (1) is as follows:
Step 1: Estimate , the strongest scatterer in the BP image ( th location stored).

Step 2: Subtract the contribution of strongest scatterer from the measurements = { ()}1
=0

= , , = 1

Iteration: Repeat Steps 1 to 3 with the regenerated image until the following condition is met:

Step 3: Use the updated measurements from Step 2 to regenerate the BP image.

2 < with:
o

= threshold (typically chosen as 1).


2
2 = 1 (Noise variance).
=1

where =1 are the estimates of a small subset of grid points in the BP image with no target present.

15

www.ijrsa.org

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

The CLEAN approach is limited in resolution, which is similar to the standard BP/DAS algorithm (see the
numerical examples in Section V). We therefore, investigate superresolution methods for FLGPR SAR imaging in
the next section.
Super-Resolution Methods:
In this section, a new approach for high-resolution imaging is presented for impulse-based FLGPR. The 2D
aperture of measurements in Fig. 2, used for imaging the specified grid, will result in a data vector in (1) 1
whose dimension is large (on the order of 106), making practical applications of adaptive techniques infeasible.
Also, the availability of only a single data vector makes direct application of well known high-resolution dataadaptive approaches, such as CAPON and APES as well as subspace-based methods [25], impossible Another
challenge is that the data vector contains clutter reflections from regions outside the ROI, which need to be
suppressed prior to imaging. These clutter reflections are given by = { }
=1 with a corresponding unknown and
fairly large steering matrix B as described in Section II.
We propose a data filtering and data dimension reduction approach via time gating and orthogonal projection to
reduce interference from scatterers outside the ROI. This approach involves a singular value decomposition of the
steering matrix corresponding to the ROI and a projection of the data using the dominant singular vectors of the
said matrix, which reduces the data dimension to practical levels. A computationally efficient method for this
projection via an eigenvalue decomposition is introduced to obtain an updated data model.

Using the updated model, two recently proposed algorithms (SPICE and SLIM) are used to produce highresolution imaging results, even with a single data vector. The process is described in the next subsections.

FIG. 4 TIME GATING

Time gating and orthogonal projection


1
For time gating, consider the kth measurement { ()}=0
. Based on (4), the delays of the grid points within the

ROI, = ,1 , ,2 , , , , to the corresponding kth transmit-receive pair for this measurement can be computed.
The minimum and maximum delays between this transmit-receive pair and the grid points in the ROI are denoted

by = max
( ) and = min
( )respectively. The kth measurement can then be updated to { ()}=

by discarding data outside these delays.

Without loss of generality, this procedure is shown in

Fig. 4 for a colocated transmit-receive pair centered below (with a pre-specified stand-off range) the imaging grid
of the ROI. In this way, interference from the regions below the minimum delay line and above the maximum
delay
line
are
discarded.
The
data
vector
in
(1)
can
then
be
updated
to

= [1 (1 ), , 1 (1 ), , ( ), , ( )] and the updated data model is:


= + +

with and ( < ).

(6)

Note from Fig. 4, that interference from scatterers outside the ROI still contributes to the measurements. In this
paper, an orthogonal projection of the data onto relevant singular vectors of the steering matrix is performed to
reduce the effects of the scatterers outside this the ROI as well as to reduce the data dimension significantly (by a

16

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

www.ijrsa.org

factor of 103 in a practical setting), allowing for practical applications of high-resolution imaging techniques. This
interference reduction via orthogonal projection is performed in a computationally efficient manner using an
eigen-decomposition and is described as follows.
Let = be the singular value decomposition of B (we omit the subscript g from (6) for notational
simplicity). The matrices and are the left and right singular vectors of B respectively; contains the
singular values. To eliminate the contributions of clutter outside the ROI from the measurements, the data vector y
can be projected onto the null space of BT given by the set of vectors orthogonal to . However, this approach is of
limited interest for the following reasons. First, the steering matrix B and hence is unknown. Also, even if B was
known, projecting onto its null space will also eliminate some target contributions that lie within the ROI. This is
due to the fact that, for a given delay within the imaging area (see Fig 4), the received signal will have contributions
from scatterers outside the ROI. Hence, the subspaces spanned by A and B are not truly orthogonal to each other.
An alternative for clutter reduction is to project the data along the subspace spanned by A, hence reducing
interference from clutter outside the ROI. The process can be described as follows. Consider the following singular
value decomposition of the steering matrix with Q > L in practice,

[] =
= =

(7)

] are the left singular vectors of A, the columns of V are their right counterparts, and
where the columns of [ ,
the singular values of A are on the diagonal of the diagonal matrix .

Due to the fine grid used for the ROI, some of the singular values of A in are quite small. By discarding the
small singular values of A, we approximate A as = . Then using for orthogonal projection yields an
updated data model:
or equivalently:

= +

(8)

+
=

(9)

= and is the noise and interference residue. Via a series of simulations, we found that the
where ,
image formation performance is quite robust to the number of columns in . We choose the dimension by
where (a factor of 103
analyzing the metric . Note that 1 and
reduction in practical applications).

The projection described in (9) can be obtained in a computationally more efficient way via an eigenvalue
decomposition in lieu of an SVD. Note that in (7) , therefore the computational cost of the SVD
decomposition is (2 ).
Consider the following eigenvalue decomposition

(10)

where and = 2 . The computation cost of the matrix product and the eigenvalue decomposition in
(10) is also (2 ). However, numerical simulations showed an improvement in computation by a constant factor
of approximately 40 using the eigenvalue decomposition, which is significant when dealing with large matrices.
Additionally, for standard personal computers (PCs), the matrix A maybe too large to store directly in memory,
however can be computed and stored. Based on the decomposition in (10), the left singular vectors of A can be
computed as = 1 , where the diagonal matrix 1 is given by:
1
1 =

(11)

with the diagonal of the sub-matrix 1 consisting of the inverse of the dominant singular values of A. The
corresponding dominant left singular vectors of A are given by = 1 , with corresponding to the
dominant eigenvectors of .
17

www.ijrsa.org

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

TABLE II SLIM ALGORITHM

Initialization:
SLIM (n+1)th Iteration:

Obtain initial estimates (0) with the DAS algorithm and (0) based on (15).
Repeat the following steps until convergence.

() =
diag(() )
+ () =

()
+ ()
Step 1: Compute:
(+1) =


1
Step 2: Compute:

()
1
(+1) 2

Step 3: Update: (+1) =

The updated data vector in (9) is given by:

= = 1

(12)

= = 1 = 1 =

(13)

whereas the updated steering matrix is given by:

This decomposition can be performed offline as long as prior knowledge on the imaging grid is available, which is
typically the case.
SLim
The SLIM method is a maximum-aposteriori (MAP) approach for sparse signal recovery. It considers a hierarchial
Bayesian model with a sparsity promoting prior for sparse parameter estimation. SLIM estimates the sparse vector
and the variance of the noise and interference residue , by minimizing the following cost function [29]:
1
2 + 2(| | 1)
(, ) = log +
2

(14)

Minimizing this cost function yields the following estimates:

=

+ =
1


=
2

(15a)
(15b)

where P = diag(p) and p = || (elementwise). These estimates are computed in an iterative manner as described in
Tab. II.
Spice
The SPICE method [30], [36] re-writes the model in (9) as follows:
+ =
=

(16)

, and = [ , ] . This method minimizes the following covariance fitting cost function:
where =
12 ( )

(17)

where = is the postulated covariance matrix of the data and the diagonal matrix P is now given by:

(18)

with = diag( ) and = diag( ) being the diagonal matrices containing the power estimates of and the noise
and interference residue , respectively.
+

The estimates of = =1 and = [ , ] = =1 are given by (see [36] for details):


=

p = ,

= 1, , +

(19a)
(19b)

where = and is the jth column of D. These estimates are computed iteratively till convergence
[36]. The steps of the SPICE algorithm are described in Tab. III. To improve on the computational efficiency of the
SPICE algorithm for FLGPR SAR imaging, we note that the data is real-valued and hence the covariance fitting

18

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

www.ijrsa.org

problem in (17) can be re-formulated as the following linear program (LP) [31] [37] [38]:
x = +

=1 s.t. =

(20)

This LP formulation leads to a more efficient solver than the approach presented in Tab. III. Based on numerical
simulations, using (20) for FLGPR SAR imaging, we can achieve an improvement in computation by a factor of
approximately two. In the appendix, we introduce a CG-SPICE algorithm for FLGPR SAR imaging with similar
improvements in computations as the LP; however, this CG-SPICE is not limited to applications with real-valued
data, unlike (20).
Note that the estimates in SLIM and SPICE have similar forms with the difference lying in the estimation of the
variance of the noise and interference residue. SPICE estimates the reflection coefficients and the noise and
interference residue simultaneously, with the variances of different elements being not necessarily equal, unlike
SLIM. However, simulation results show a similar performance for the two algorithms with SPICE being a bit less
susceptible to the noise and interference residue. Both algorithms appear to be robust and effective for generating
sparse results.
In the next section, these robust user-parameter free algorithms are used for FLGPR SAR imaging. The analysis is
performed on both simulated and measured SIRE FLGPR data.
TABLE III SPICE ALGORITHM

Initialization:
SLIM (n+1)th Iteration:

Obtain initial estimates (0) = with (0) = (0)


Repeat the following steps until convergence.
() = diag
()
() =
Step 1: Compute:
1

Step 2: Compute: (+1) = ()


= 1, , +
Step 3: Update: (+1) = (+1)
= 1, , +

= 1, , +

FIG. 5 ORTHOGONAL PROJECTION ANALYSIS: FLGPR SAR IMAGING


(A) EIGENVALUES OF , (B) SPICE ( = 100), (C) SPICE ( = 200), AND (D) SPICE ( = 300)
TABLE IV SUBSPACE APPROXIMATION

No. of singular vectors -

300

0.182

100
200

0.512
0.257

19

www.ijrsa.org

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

FIG. 6 FLGPR SAR IMAGING - DETECTION OF WEAK TARGET (A) BP, (B) RSM, (C) CLEAN, (D) SLIM, AND (E) SPICE

Numerical And Experimental Results


In this section, we perform sparse high-resolution imaging for FLGPR using orthogonal projection as a preprocessing step for data dimension and clutter reduction. The SPICE and SLIM algorithms are considered for highresolution imaging.
We also use the CLEAN approach for imaging based on the proposed data model and show the ability of this
algorithm to yield sparse results. As we will see, however, the CLEAN approach does not improve the resolution
over the standard BP method.
For practical FLGPR SAR imaging analysis, we use the SIRE radar designed by ARL for imaging based on the
setup in Figs. 1 and 2. For simulations, we consider an imaging area which has a range swath of 4 m and a crossrange swath of 5 m. A minimum stand-off distance of 8 m is used for simulations, and a maximum stand-off of 14
m. For each simulation, targets are placed at various [x, y, z] locations (in meters) marked in each image by the X
symbols. The imaging area consists of L = 10000 grid points.
An analysis of the robustness of the imaging procedure with respect to (the number of dominant singular values
of A) is performed first. Targets at locations [0, 0, 0], [-0.3, 1, 0], and [1.5, 1.5, 0] are simulated. Fig. 5 shows the
eigenvalues of the matrix as well as the results obtained with the SPICE algorithm used for imaging with
different values of . Tab. IV shows that thresholds as large as 0.5 based on the criterion
(normalized scale of 0 to 1) yield satisfactory results with all the targets being detected as can be seen in Fig. 5.

The next analysis involves detecting a weak target buried by the sidelobes of much stronger targets. Three targets
are again simulated at locations [0, 0, 0], [-0.1, 0.8, 0], and [1.5, 1.5, 0], with the target at [-0.1, 0.8, 0] ten times
weaker than the other two targets. From Fig. 6, we can see that in the BP and RSM images, the weak target is
buried in the sidelobes of the stronger target near it. On the other hand, the CLEAN approach can detect the weak
target. The SPICE and SLIM algorithms are also applied post orthogonal projection (with threshold
= 0.2) and the weak target is detected as well.
20

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

www.ijrsa.org

FIG. 7 FLGPR SAR IMAGING - RESOLUTION IMPROVEMENT (A) BP/DAS, (B) RSM, (C) CLEAN, (D) SLIM, AND (E) SPICE

FIG. 8 REAL DATA - SIRE FLGPR SAR IMAGING: (A) BP, (B) RSM,
AND (C) SPICE

FIG. 9 RECEIVER OPERATING CHARACTERISTICS (ROC)


COMPARISON FOR THE PROPOSED METHOD

The CLEAN approach, although effective in providing sparser results than BP, is limited in resolution and has little
improvement over the standard BP approach. SLIM and SPICE, on the other hand, provide noticeable
improvement in resolution over the BP-based algorithms (BP, RSM and CLEAN). Fig. 7 shows three targets at
locations [0, 0, 0], [0, 0.75, 0], and [1.5, 1.5, 0]. Two of these targets are closely spaced and even so are clearly
resolved by SLIM and SPICE.
Next, the proposed approach is tested using experimentally measured data. Results based on the SIRE data
provided by the ARL can be seen in Fig. 8. In this figure, subimages 2 m in range are continuously formed, based
on overlapping 2D apertures, to generate the entire image [13]. As mentioned earlier, the SPICE and SLIM
approaches yield similar results for the simulated data, with the SPICE approach being less susceptible to noise
and interference residue. The SPICE approach is therefore used for the experimental data. We can see from Fig. 8,
that SPICE provides sparse results with all the targets detected (based on prior knowledge); some targets of interest
are encircled in red oval circles in the figure.
Finally, a quantitative numerical analysis is performed to show the effectiveness of the high-resolution SPICE
approach for FLGPR SAR imaging. Here nine targets with randomly selected strengths are simulated and the

21

www.ijrsa.org

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

receiver operating characteristics (ROC) is shown in Fig. 9. This curve, which shows the probability of detection
versus the number of false alarms, is generated by using a threshold detector. The image under analysis is
segmented into target-size regions, and the maximum pixel value in each region is retained. The threshold is
incremented in steps and for each threshold the number of pixels exceeding the threshold are recorded. The alarms
that fall outside the regions with targets present (based on prior knowledge) are considered false alarms. As shown
in Fig. 9, the SPICE method renders perfect detection at a much smaller false alarm rate than the BP algorithm
does.
Conclusions
In this paper, we have considered new approaches to imaging for FLGPR. The proposed pre-processing step
involves a data model in the time domain which takes into account the contributions of clutter that lie outside the
imaging area. An orthogonal projection of the measured data onto a subspace spanned by the steering matrix
corresponding to the ROI is then used for clutter reduction as well as significant data size reduction, making it
feasible to use high-resolution methods for FLGPR SAR imaging. The steering matrix decomposition is performed
efficiently and requires only prior knowledge of the desired imaging area (hence it can be performed offline). Two
data-adaptive approaches, SPICE and SLIM, are used for FLGPR SAR imaging. They are user-parameter free
algorithms that have the ability to provide sparse high resolution images using a single data vector, unlike manyot
her high-resolution methods. The results using simulated data show that SLIM and SPICE provide improvement in
resolution compared to the BP-based algorithms viz. BP, RSM and CLEAN. The SPICE optimization problem can
be reformulated as a linear program and solved efficiently, in the case of real-valued data. A new conjugate
gradient based SPICE algorithm is introduced in the appendix which also provides improved computation
compared to the standard SPICE algorithm, and can be used even in applications with complex-valued data.
Apprendix: CG Spice
1 in Step 2 of the iterative SPICE algorithm
The CG-SPICE uses the conjugate gradient method to compute =
in Tab. III.

CG Initialization: (0) = ; (0) = (0) =


CG (mth) Iteration:

o
o
o
o

( ) = ( ) ( ) ( ) ( )
( +1) = ( ) + ( ) ( )
() ( )
( +1) = ( ) ( )

( +1) = ( ) + ( +1) ( +1) ( ) ( )

This algorithm can be used for both real and complex-valued data. (For complex valued data, the operator ()
should be replaced by () .

ACKNOWLEDGMENT

This material is based upon work supported by, or in part by, the U.S. Army Research Laboratory and the U.S.
Army Research Office under contract/grant No. W911NF-11-2-0039. The views and conclusions contained herein
are those of the authors and should not be interpreted as necessarily representing the official policies or
endorsements, either expressed or implied, of the U.S. Government. The U.S. Government is authorized to
reproduce and distribute reprints for Governmental purposes notwithstanding any copyright there on.
REFERENCES

[1]

S. M. Bird and C. B. Fairweather, Military fatality rates (by cause) in Afghanistan and Iraq: a measure of hostilities,
International Journal of epidemiology, vol. 36, no. 4, pp. 841846, 2007.

[2]

R. Gunaratna, Suicide terrorism: a global threat, Janes Intelligence Review, vol. 12, no. 4, pp. 5255, 2000.

[3]

Hidden killers: The global landmine crisis, in 1998 report to the US Congress, US Department of State, Bureau of

22

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

www.ijrsa.org

Political-Military Affairs, US GPO, 1998.


[4]

A. Soroush, F. Falahati, M. Zargar, M. Soroush, S. Khateri, and A. Khaji, Amputations due to landmine and unexploded
ordnances in post-war iran, Arch Iran Med, vol. 11, no. 6, pp. 595597, 2008.

[5]

R. B. Miles, A. Dogariu, and J. B. Michael, Bringing bombs to light, IEEE Spectrum, vol. 49, no. 2, pp. 3843, 2012.

[6]

B. N. Nelsen, A forward looking infrared sensor for landmine detection that incorporates a novel method of identifying
regions of interest and a fuzzy inference system, in Fuzzy Information Processing Society, 1999. NAFIPS. 18th
International Conference of the North American, pp. 408 412, 1999.

[7]

J. A. MacDonald and J. Lockwood, Alternatives for landmine detection. No. 1608, Santa Monica, CA: RAND corporation,
2003.

[8]

M. I. Skolnik, Introduction to radar systems. New York, NY: McGrawhill, 2001.

[9]

Metwaly, Mohamed and Ismail, Ahmed and Matsushima, Jun, Evaluating some factors that affect feasility of using
ground penetrating radar for landmine detection, Applied Geophysics, vol. 4, no. 3, pp. 221230 .

[10] D. J. Daniels, Ground penetrating radar. London, UK: The Institution of Engineering and Technology, 2007.
[11] S. Vitebskiy, L. Carin, M. A. Ressler, and F. H. Le, Ultra-wideband, short-pulse ground-penetrating radar: simulation and
measurement, IEEE Transactions on Geoscience and Remote Sensing, vol. 35, no. 3, pp. 762772, 1997.
[12] L. Peters, J. Daniels, and J. Young, Ground penetrating radar as a subsurface environmental sensing tool, Proceedings of
the IEEE, vol. 82, no. 12, pp. 18021822, 1994.
[13] L. Nguyen, Signal and Image Processing Algorithms for the U.S. Army Research Laboratory Ultra-wideband (UWB)
Synchronous Impulse Reconstruction (SIRE) Radar, ARL Technical Report, ARL-TR- 4784, April 2009.
[14] E. Byron, Radar Principles, Technology, Applications. Englewood Cliffs, New Jersey: PTR Prentice hall, 1993.
[15] C. V. Jakowatz, D. E. Wahl, P. H. Eichel, D. C. Ghiglia, and P. A. Thompson, Spotlight-mode synthetic aperture radar: a
signal processing approach, vol. 101. Norwell, MA: Kluwer Academic Publishers, 1996.
[16] J. Li and P. Stoica, MIMO Radar with Colocated Antennas, IEEE Signal Processing Magazine, vol. 24, pp. 106114, Sept.
2007.
[17] O. Ojowu, Y. Wu, J. Li, and L. Nguyen, SIRE: a MIMO radar for landmine/IED detection, in Proc. SPIE, Radar Sensor
Technology XVII, vol. 8714, 2013.
[18] Kotarba, Andrzej Z, Evaluation of ISCCP cloud amount with MODIS observations, Atmospheric Research, vol. 153, pp.
310317, 2015.
[19] Garca-Mora, Tziztiki J and Mas, Jean-Francois and Hinkley, Everett A, Land cover mapping applications with MODIS: a
literature review, International Journal of Digital Earth, vol. 5, no. 1, pp. 6387, 2012.
[20] Rahimi, Saeid and Gholami Sefidkouhi, Mohammad Ali and Raeini-Sarjaz, Mahmoud and Valipour, Mohammad,
Estimation of actual evapotranspiration by using MODIS images (a case study: Tajan catchment), Archives of Agronomy
and Soil Science, pp. 115, 2014.
[21] Gader, Paul D and Nelson, Bruce N and Frigui, Hichem and Vaillette, Gary and Keller, James M, Fuzzy logic detection of
landmines with ground penetrating radar, Signal Processing, vol. 80, no. 6, pp. 10691084, 2000.
[22] Carri`ere, Simon D and Chalikakis, Konstantinos and Senechal, Guy and Danquigny, Charles and Emblanch, Christophe,
Combining electrical resistivity tomography and ground penetrating radar to study geological structuring of karst
unsaturated zone, Journal of Applied Geophysics, vol. 94, pp. 3141, 2013.
[23] H. L. Van Trees, Detection, Estimation, and Modulation Theory, Optimum Array Processing. New York, NY: John Wiley &
Sons, 2002.
[24] L. A. Gorham and L. J. Moore, SAR image formation toolbox for MATLAB, in SPIE Defense, Security, and Sensing, vol.
7699, 2010.
[25] P. Stoica and R. L. Moses, Spectral Analysis of Signals. Upper Saddle River, NJ: Prentice-Hall, 2005.

23

www.ijrsa.org

International Journal of Remote Sensing Applications (IJRSA) Volume 5, 2015

[26] J. Hogbom, Aperture Synthesis with a Non-Regular Distribution of Interferometer Baselines, Astronomy and
Astrophysics Supplement, vol. 15, pp. 417426, 1974.
[27] E. Karpat, CLEAN technique to classify and detect objects in subsurface imaging, International Journal of Antennas and
Propagation, 2012.
[28] L. Nguyen and T. Tran, Robust recovery of synthetic aperture radar data from uniformly under-sampled measurements,
in IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pp. 35543557, July 2011.
[29] X. Tan, W. Roberts, J. Li, and P. Stoica, Sparse Learning via Iterative Minimization with Application to MIMO Radar
Imaging, IEEE Transactions on Signal Processing, vol. 59, pp. 10881101, March 2011.
[30] P. Stoica, P. Babu, and J. Li, New Method of Sparse Parameter Estimation in Separable Models and Its Use for Spectral
Analysis of Irregularly Sampled Data, IEEE Transactions on Signal Processing,vol. 59, pp. 3547, Jan. 2011.
[31] P. Babu and P. Stoica, A combined linear programming-maximum likelihood approach to radial velocity data analysis for
extrasolar planet detection, in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.
43524355, Prague, 2011.
[32] M. Ressler, L. Nguyen, F. Koenig, D. Wong, and G. Smith, The Army Research Laboratory (ARL) Synchronous Impulse
Reconstruction (SIRE) Forward-Looking Radar., in Proc. SPIE, Unmanned Systems Technology IX, vol. 6561, 2007.
[33] L. Nguyen and R. Innocenti, Suppression of sidelobes and noise in airborne SAR imagery using the Recursive Sidelobe
Minimization technique, in IEEE Radar Conference, pp. 522525, May 2010.
[34] J. Ender and J. Klare, System architectures and algorithms for radar imaging by MIMO-SAR, in IEEE Radar Conference,
pp. 16, May 2009.
[35] J. Li and P. Stoica, eds., MIMO Radar Signal Processing. Hoboken, NJ: John Wiley & Sons, 2009.
[36] P. Stoica, P. Babu, and J. Li, SPICE: A Sparse Covariance-Based Estimation Method for Array Processing, IEEE
Transactions on Signal Processing, vol. 59, no. 2, pp. 629638, 2011.
[37] P. Stoica and P. Babu, Algebraic Derivation of Elfving Theorem on Optimal Experiment Design and Some Connections
With Sparse Estimation, IEEE Signal Processing Letters, vol. 17, no. 8, pp. 743745, 2010.
[38] P. Stoica and P. Babu, SPICE and LIKES: Two hyperparameter-free methods for sparse-parameter estimation., Signal
Processing vol. 92, pp. 15801590, Oct 2012.

24

Potrebbero piacerti anche