You are on page 1of 2

SVD-based Clutter Removal Technique for GPR

Maria Garcia-Fernandez, Yuri Alvarez-Lopez, Yolanda Rodriguez-Vaqueiro, Borja Gonzalez-Valdes,


Ana Arboleya-Arboleya, Fernando Las-Heras Antonio Pino-Garcia
Area of Signal Theory, University of Oviedo AtlanTIC Research Center, University of Vigo
Campus Viesques, Modulo 8, 33203 Gijon, Spain Esc. Ing. Telecomunicaciones, 36310 Vigo, Spain
mariagarcia@tsc.uniovi.es yrvaqueiro@com.uvigo.es

Abstract—The main goal of this contribution is to present constitutive parameters are known or estimated). Assuming
a clutter reduction technique for Ground Penetrating Radar a multi-monostatic scenario where the scattered field Es is
(GPR) imaging. The proposed technique is based on a subspace measured at Nf frequencies on No observation points, the
projection method applied to the focused Synthetic Aperture
Radar (SAR) image. It consists of decomposing the image in reflectivity at point r0 is calculated as follows:
its different eigenimages and classifying them as belonging to Nf No
clutter, target or noise subspaces. Several scenarios have been
X X
ρ(r0 ) = Es (fn , rm ) exp(+j2(φ1 + φ2 )) (1)
simulated in order to test this technique and its performance has
n=1 m=1
been measured analyzing the variation in the Signal-to-Clutter
Ratio (SCR). In addition, it has been compared with background where φ1 is the phase shift of the propagation between the
subtraction and other subspace projection methods. antenna position rm and the refraction point at the soil surface
ri , and φ2 is the phase shift between ri and r0 [5].
I. I NTRODUCTION
B. SVD
Ground Penetrating Radar (GPR) is a well-known technique
for non-invasive detection of metallic and dielectric buried A rectangular matrix X, whose dimensions are M ×N (with
targets [1]. This is possible due to the different constitutive M ≥ N ), can be decomposed using SVD as X = U ΣV H ,
parameters of targets and soil. Nevertheless, one of the main where U and V are M × M and N × N unitary matrices
issues of imaging buried objects is the strong clutter produced that contain the left and right singular vectors, Σ is a diagonal
by the specular reflection from the ground surface. Several matrix that contains the singular values of X in descending
clutter rejection methods, such as time-gating, average mean order (σ1 ≥ σ2 ≥ ...), and H refers to the Hermitian transpose.
subtraction and subspace projection methods, among others, According to this decomposition, X can be expressed as
have been proposed to overcome this issue in subsurface [2] the sum of the N eigenimages, where the ith eigenimage is
and through-wall imaging [3]. Mi = ui σi viH . Therefore, the goal of the subspace projection
Subspace projection methods, such as Singular Value De- method is to find the subsets of eigenimages that compose the
composition (SVD), estimate the clutter subspace and then clutter, target and noise subspaces.
project the B-scan data (i.e. the 2D raw measures) orthogonally C. Subspace Projection Method
onto this subspace. It has been assumed that the first eigen-
The first step of the method is to estimate the distance
image represents the clutter and the second one the target [4].
between the observation domain and the ground surface hsoil .
However, both clutter and targets can span a multidimensional
Once the average B-scan in the space-domain rav is computed,
subspace, depending on the surface roughness, soil composi-
hsoil is estimated as the location of its global maximum. The
tion, size, location and number of targets, etc.
radar returns at a distance smaller than the threshold th =
In this contribution, a subspace projection method is applied
hsoil +λmax /2 (where λmax is the maximum wavelength) are
to the focused Underground - Synthetic Aperture Radar (SAR)
considered to correspond to the air and the air-soil interface.
image, instead of to the B-scan data (which helps to reduce
Next, the B-scan data is focused using the Underground-
the dimensionality of surface and target subspaces). The
SAR imaging method and the SVD of the obtained image
proposed method classifies each eigenimage as belonging to
is computed. The singular values that correspond to the 99%
clutter, target or noise subspaces. Two dimensional Method of
of the total power (where the power represented by the ith
Moments (2D-MoM) simulations are used to test the proposed
singular value is σi /tr(Σ)) are considered to belong to clutter
method.
and target subspaces, whereas the other singular values are
II. M ETHODOLOGY assumed to belong to the noise subspace and are discarded.
For each remaining eigenimage, the location of its maximum
A. Underground-SAR imaging hmax is found. If the position of this maximum lies within
The measured scattered field has been focused using an the air-soil interface area (hmax ≤ th), this eigenimage is
Underground-SAR technique that takes into account the dif- classified as belonging to the clutter subspace. Otherwise, it
ferent wave velocities in the air and in the soil (provided the is assumed that it belongs to the target subspace.

978-1-5386-3284-0/17/$31.00 ©2017 IEEE 2369 AP-S 2017


TABLE I
In order to measure the performance of the method, the C HANGE IN SCR (G) FOR DIFFERENT CLUTTER REDUCTION TECHNIQUES .
change in Signal-to-Clutter Ratio (SCR) after applying the Scenario Target GBS [dB] GP [dB] GI [dB] GB [dB]
clutter reduction technique is computed (2). The value of SCR
Metallic 10.5 6.8 5.9 5.5
is given by (3) where At is the target region (i.e. the pixels (a)
Dielectric 25.1 22.1 11.7 10.8
where the target is placed), Ac is the rest of the image, and Nt (b)
Metallic 10.1 6.4 5.3 4.9
and Nc are the number of pixels in At and Ac , respectively. Dielectric 14.2 13.3 9.0 8.6
Metallic 11.2 7.9 4.7 3.2
(c)
Dielectric 25.7 16.3 5.4 4.2
 
SCRwith clutter reduction
G [dB] = 10 log (2) (d)
Metallic 18.9 14.5 13.8 13.0
SCRinitial Dielectric 31.8 28.9 17.3 15.9
1 2
P
N (x,z)∈At |ρ(x, z)|
SCR = 1t P 2
(3)
Nc (x,z)∈Ac |ρ(x, z)|

III. R ESULTS
Several scenarios have been simulated to analyze the be-
haviour of the proposed method. Unless otherwise stated,
the following parameters are considered: frequency ranges
from 3.5 to 5.5 GHz (sampled every 20 MHz), the observation (a) Without clutter reduction (b) Background Subtraction
domain is a straight line 50 cm above the surface sampled
every 2.5 cm (∼ 0.5λmin ), the soil constitutive parameters are
r,soil = 3 and σsoil = 0.05 S/m and the target is a 19 × 2 cm2
rectangular object buried 10 cm below the surface. Standard
deviation of the surface roughness is 1 cm and measurements
are contaminated with white gaussian noise resulting in a
Signal-to-Noise Ratio (SNR) of 30 dB. Both metallic and
(c) Proposed method (d) Removing first eigenimage
dielectric targets (with r,target = 4.5) have been considered.
The proposed method has been compared with Background Fig. 1. Images of scenario (c) with a buried dielectric target.
Subtraction (BS) and with the removal of the first eigenimage
in the SAR image and in the B-scan data (as proposed in
literature). BS is the ideal clutter removal technique but it IV. C ONCLUSIONS
requires accurately know the field scattered by the soil without The power is more widespread among the different eigen-
targets (not available in real applications). Table I shows the images in the B-scan decomposition than in the SAR image,
change in the SCR for all these techniques (GBS , GI and GB , which is in agreement with the fact that GI > GB in all
respectively), as well as for the proposed technique (GP ). scenarios. When the target is metallic, the performance of
In the initial scenario, named (a), when the metallic target the proposed method is usually similar to just removing the
is present, the first eigenimage belongs to clutter subspace first eigenimage. However, when the target is dielectric, the
whereas the second one belongs to target subspace. Due to proposed method is notably better (GP >> GI ).
the high reflectivity of the metallic target, the elimination
ACKNOWLEDGEMENT
of just the first eigenimage is almost as effective as the
proposed method. However, when the target is dielectric, the This work has been partially supported by Spain government
first two eigenimages belong to the clutter subspace. Hence, under projects TEC2014-54005-P and TEC2014-55290-JIN,
the removal of the first eigenimage is not enough to properly and grant FPU15/06341; by Asturias government under project
detect the target. In this case, the proposed method is much GRUPIN14-114; and by Galicia government under projects
more effective than the removal of the first eigenimage. CN2012/279, CN20127260 and Plan I2C (2011-2015).
In the second scenario, denoted as (b), the SNR is decreased R EFERENCES
to 10 dB to test the influence of noise. In this case, the
[1] D. J. Daniels, Ground Penetrating Radar, 2nd ed. London, UK: IET, 2004.
proposed method is also closer to the background subtraction [2] R. Solimene, A. Cuccaro, A. Dell’Aversano, I. Catapano, and F. Sol-
technique than the other subspace projection methods. dovieri, “Ground Clutter Removal in GPR Surveys”, IEEE JSTARS, vol.
In scenario (c), the surface roughness standard deviation is 7, no. 3, pp. 792–798, March 2014.
[3] P. K. Verma, A. N. Gaikwad, D. Singh, and M. J. Nigam, “Analysis of
increased to 3 cm. This increase yields a higher dimensional clutter reduction techniques for through wall imaging in UWB range”,
clutter subspace. When the target is dielectric, the improve- PIER, vol. 17, pp. 29–48, January 2009.
ment of the proposed method is particularly noticeable. Fig. 1 [4] F. Abujarad, G. Nadim, and A. Omar, “Clutter reduction and detection
of landmine objects in ground penetrating radar data using singular value
shows the original image as well as the one obtained with the decomposition (SVD)”, in Proc. of the 3rd IWAGPR, May 2005.
proposed method, background subtraction and with removing [5] J. A. Martinez-Lorenzo, C. M. Rappaport, and F. Quivira, “Physical
the first eigenimage in SAR domain. Limitations on Detecting Tunnels Using Underground-Focusing Spotlight
Synthetic Aperture Radar”, IEEE Trans. Geosci. Remote Sens., vol. 49,
Finally, in scenario (d), the target shape and size is changed no. 1, pp. 65–70, January 2011.
to a circular object with radius 2 cm.
2370

You might also like