You are on page 1of 10

APPLIED GEOPHYSICS, Vol.6, No.2 (June 2009), P. 192 - 200 , 10 Figures.

DOI: 10.1007/s11770-009-0021-2

The adaptive chirplet transform and its application in GPR


target detection*
Zeng Zhaofa1,2, Wu Fengshou1, Huang Ling1, Liu Fengshan2, and Sun Jiguang2

Abstract: GPR has become an important geophysical method in UXO and landmine
detection, for it can detect both metal and non-metallic targets. However, it is difficult to
remove the strong clutters from surface-layer reection and soil due to the low signal to noise
ratio of GPR data. In this paper, we use the adaptive chirplet transform to reject these clutters
based on their character and then pick up the signal from the UXO by the transform based on
the Radon-Wigner distribution. The results from the processing show that the clutter can be
rejected effectively and the target response can be measured with high SNR.
Keywords: GPR, target detection, clutter rejection, chirplet transform

Introduction filter is a simple and effective method for removing the


ground surface clutter (Brunzell, 1996). Such clutter
varies with surface roughness and soil conditions and
Target detection, such as UXO and landmine detection, leads to uncertainty in the measurements and processed
has traditionally used metal detectors and EMI methods data. It is therefore necessary to eliminate the surrounding
are serious problems all over the world. However, when effects during GPR data processing for shallow buried
the target objects are made with very low metal content and smaller objects. In this regard, the more important
or are purely non-metallic, it is more challenge for the and widely used approaches for suppression of clutter
traditional detection methods (Singh, 2007). GPR was are the methods based on the autoregressive moving
suggested as an attractive candidate for imaging their average (ARMA) model (van Kempen and Sahli, 2001),
location, shape, and orientation (Fritzsche, 1995). One Kalman filtering (Chowdhury, 1996), and selective
problem is the cluttered signals from the ground surface independent component analysis (SICA) (Fritzsche, 1995;
and underground layers (Bruschini et al., 1998). Clutter van der Merwe. and Gupta, 2000), which all are based on
from the ground surface is dominant in the GPR signals parametric modeling. In this paper, we introduce a sub-
particularly for detecting the shallow buried objects space based adaptive clutter rejection algorithm which
such as landmines. The weak signals from a target such does not need a signal model.
as a non-metal landmine is obscured by these stronger Another problem is how to distinguish the small
signals because the weak target signals and the clutter target signals such as landmine or UXO on a lot of GPR
happen almost in the same frequency bands and at similar data and reduce the manual interpretation workload.
arrival times, so it is very difcult to separate them (van GPR signals have important characteristics in time and
Kempen and Sahli, 2001). The moving spatial average frequency shift because of media dispersion. The electric

Manuscript received by the Editor October 6, 2008; revised manuscript received April 9, 2009.
*This work was supported by U.S. Department of Defense Science Research Fund (Grant No. DAAD 19-03-1-0375) and the
National Natural Science Foundation of China (Grant No. 40774055).
1. Collage of Exploration Science and Technology, Jilin University, Changchun 130026, China.
2. Applied Mathematic Research Center, Delaware State University, Dover, DE 19901, USA.

192
Zeng et al.

properties of near subsurface soil in UXO and landmine where 0 is the attenuation coefcient with no dispersion
detection by GPR are strongly frequency dependent and v is the radar wave velocity in the medium. The
(Turner and Siggins, 1994).The GPR EM pulse reected electric eld is expressed as
from targets such as landmine and UXO travels through
the soil and therefore the EM wave frequency bands will E ( z, t ) E0 e D z e  i (Zt  E z )
Z
 z
be shifted. This character can be analyzed by the chirplet E0 e D0 z e i (Zt  E z ) e 2 vQ* . (4)
transform and can be used for target detection (Qian et al., Z
1995). The chirplet transform is an extension of the well The Q* and v in the  2cQ z
term are both frequency
e v *
known wavelet and Gabor transforms (Mann and Haykin, dependent. From equation (4), we know that the energy
1991; 1992; Alaee and Amindavar, 2008; Lu et al., 2008). of electromagnetic pulse propagation in the medium
Its specic signal representations are based on families of will be attenuated and the frequency will be shifted in
functions related to time and frequency shifting and scale the time-frequency plane (Salvati et al., 1998; Mann and
change, which has been widely used in GPR applications. Haykin, 1995). The frequency shift can be used in the
Compared with the time-shifted and frequency- target detection by the adaptive chirplet transform.
modulated Gaussian function used for the Gabor
expansion (Mann and Haykin, 1992), the chirplet
transform has more freedom and thereby can match the Surface reection rejection by the
signal better. Many natural phenomena, for instance,
signals encountered in radar systems, the impulsive
adaptive subspace algorithm
signal that is dispersed by the ionosphere (Qian et al.,
1995), and seismic signals (Wang and Zhou, 1999) can The GPR signal is a complex signal where the surface
all be modeled as chirp type functions. reflection signal has stronger amplitude than the target
In this paper, we use the adaptive chirplet transform signal and stacks with the target signal in time. To
(Wang et al., 2003) to pick up the UXO signal based on extract the target signal from the combined signal, it
the Radon-Wigner distribution. The test results show is necessary to remove the surface reflection before
the clutter can be rejected effectively and the target applying the adaptive chirplet transform (ACT). We
responses can be enhanced. used an improved signal subspace method to eliminate
the surface reflection (Wang et al., 2003), which does
not need any model and prior knowledge. Suppose
EM pulse propagation in loss medium sc(t) is the received GPR signal in the current scan and
sn1(t), , snN(t) are the GPR detection signals from sc(t)n
neighboring scans.
GPR is a geophysical survey by EM pulse waves which
Define the signal vectors S c as signal space
are widely applied in subsurface target detection and
constructed for the current scan with a target and signal
environmental problem evaluations (Qian et al., 1995). Some
vector Si as the ith neighboring scan from the all samples
equations describing electromagnetic wave propagation in
of the scan:
media are well known (Irving and Knight, 2003).
According to Turner and Siggins (1994), Q represents Sc [ sci (0), sci (1),!, sci ( M  1)]T , (5)
the attenuation quality parameter of radio wave
propagation is closely related to the loss tangent. Si [ sni (0), sni (1),!, sni ( M  1)]T . (6)

Q | 1/ tan G . (1) Then, the covariance matrix R of clutter and external


noise can be estimated as
We define a new parameter Q*, which is obtained
N

K
from the slope of the linear region of the attenuation data J
i Si SiH
in the bandwidth of interest as R i 1
, (7)
N

Ki
J
1 'Z
Q* . (2)
2v 'D i 1
where
Hence the best t line to the data in this region can be ScH Si
described as Ki . (8)
ScH Si
Z
D D0  , (3)
2vQ * The covariance matrix R is the correlation coefcient
193
Application of adaptive chirplet transform in GPR target detection

between the received signal vector at the current scan N N

and the ith neighboring scan where is the position R1 U1 V1 U1H O u u


i 1
i i i
H
O P,
i 1
i i
(12)
scalar and typically takes a value between 1 and 2 and H
is the transpose of the matrix. where i represents the ith largest eigenvalue of R1 and i
The introduction of equation (12) allows us to is its eigenvector and Pi = iiH is the projection operator
normalize the signals with respect to the correlation to the subspace derived by i. From equation (9), we
coefficients. The larger the correlation coefficients know that the stronger the clutter energy distributed in the
between the target and neighboring signals, the larger the subspace vector i, the larger the eigenvalues i are. The
expected weight factors are. The effects from the non- following algorithm provides an efcient way to remove
correlative signals can be removed. strong clutter without any knowledge of the signals:
N
Singular value decomposition (SVD) of the covariance
matrix R can be written as
S proj f (O )PS ,
i 1
i i c
(13)

R UVU H , (9) where f (i) is a weighting function that takes a smaller


where U is a unitary matrix and V is a diagonal matrix. value for a larger value of i . In this paper, f (i) is chosen as
Columns of U are eigenvectors of R and the elements in 1
f (Oi ) , (14)
the diagonal of V are the corresponding eigenvalues. As Oi
the target responses are usually in a small range and the
Then equation (14) is substituted into the equation (13)
clutter is the dominant component in the received GPR
to get
signals, the eigenvectors u1, u2, , uM corresponding to N N
1
the M largest eigenvalues can be reasonably associated S proj f (Oi )PS PS R11Sc ( R  V 2 I ) 1 Sc.
i 1 Oi
i c i c
to the clutter. The signals to be extracted can be obtained i 1
(15)
by removing the effects from the rst M eigenvectors in
the GPR signals. The expression is: The noise variance 2 controls the rejection effects
against the surface reection components.
M
H Figure 1a shows a raw GPR profile. In the GPR
S proj I  uni u ni Sc . (10)
survey, we used the GSSI SIR-3000 system and selected
i 1
an antenna with the center frequency of 270 MHz.
Equation (5) is rank deficient because the number of Figure 1b shows the profile after surface reflection
neighboring scans is usually smaller than the dimension of rejection using the adaptive signal subspace algorithm.
the variance matrix R to be estimated. With consideration From Figure 1b, we can clearly see that the reflection
of the existence of noise, the full-rank covariance matrix from the surface is rejected and also other reflections
R1 for clutter and noise can be written as from the soil are also reduced, while the responses from
the targets are clearer and have higher SNR. Figure 2
R1 R  V 2I , (11)
shows a comparison of one of the raw traces and the
where 2 is the noise variance, which can be roughly trace after clutter removal acquired over one target. The
estimated, and I is an identity matrix. Performing single target responses at 10 ns are almost the same and the
value decomposition of R1 yields surface reections from 2 to 8 ns are greatly reduced.

0 0

10 10
Time (ns)

Time (ns)

20 20

30 30

40 40
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Distance (feet) Distance (feet)
(a) (b)
Fig. 1 A comparison between raw data (a) and surface reection rejected data by the adaptive method (b)

194
Zeng et al.
1
a
b
0.8

Normal amplitude
0.6

0.4

0.2

0
0 5 10 15 20 25 30
Time (ns)
Fig. 2 A comparison between two signal traces extracted from the raw data
in Figure 1 (a) and the surface reection rejected data in Figure 1 (b).

Adaptive chirplet transform and include functions with different times, frequency
bandwidths, and central locations. For example, the
target signal extraction following modulated Gaussian function is usually used:
Jk J E
exp  k t  tk  j Ik  k t  tk ,
2
g k (t ) 4
Adaptive chirplet transform S 2 2
The adaptive chirplet transform was introduced in (18)
Yin et al. (2002), and Wang et al. (2003). In a GPR where k,and k are parameters that control the envelope
profile, the signals reflected from targets are no longer and phase of the chirplet and k and t k denote the
linear thanks to complex background conditions. When frequency and time centers, respectively.
the time-varying frequency component is a higher The next step is to construct the chirplet frame. The
order polynomial of time, the signals can be expressed Radon-Wigner transform can be used to estimate these
as a combination of several linear chirps (frequency- chirp rates. For a given signal s(t), the chirp rate 1 is
modulated signals) at different time intervals. The obtained by searching the largest peak in the RWT plane
procedure of the chirplet decomposition of a signal is after taking the RWT of the signal s(t). Modulating
rst to estimate the chirp rates 1, 2, , N of s(t) over the frame {hk(t)} in equation (17), we then obtain the
different segments and then to construct the respective chirplet frame {hk(t)ui(t)} with
chirps following the relation (Irving and Knight, 2003): 1
u1 (t ) exp j D1t 2 . (19)
1 2
ui (t ) exp j D i t 2 . (16)
2 Next, the signal which optimally matches the signal
For a given frame ^hk (t ), k Z ` and N0 chirp rate, a s(t) in the modified frame {hk(t)u1(t)} is estimated and
new chirplet frame ^hk (t )ui (t ), k , i Z ` is obtained. Based denoted as hk (t)u1(t), where
1

on this chirplet frame ^hk (t )ui (t )`, s(t) is decomposed as s(t ), u(t )hk1 (t )
N0 hk (t ) arg min s(t )  u1 (t )hk1 (t ) . (20)
s (t ) C
i 1 k
i ,k ] i hk (t )ui (t ), (17) k
hk (t )
The signal s1(t) is dened as
where Ci ,k s(t ), h ' k (t )ui (t ) is the frame decomposition
s (t ), u(t )hk1 (t )
and ^h ' k (t ), k Z ` is the dual frame of ^hk (t ), k Z ` , , s1 (t ) s (t )  u1 (t )hk1 (t ). (21)
represents the inner product, and is an arbitrary weight hk (t )
which satises By repeating the same procedure of s(t) to s 1(t),
N0
the chirp rate 2 corresponding to the second largest
]
i 1
i 1, 0 d ] i d 1. component of s(t) is obtained. Letting

Dugnol et al. (2008) have introduced and conducted 1


u2 (t ) exp j D 2t 2 , (22)
an efficient frame decomposition, ^hk (t ), k Z ` should 2

195
Application of adaptive chirplet transform in GPR target detection

we get domain.
WVD has a very high resolution for a single chirp
hk2 (t ) signal but its major disadvantage is the presence
s1 (t ), u2 (t )hk1 (t )u2 (t )hk (t ) of artificial cross-terms caused by the quadratic
arg min s1 (t )  , multiplication nature. Radon-Wigner transform (RWT),
k
hk (t ) which takes advantage of the above oscillating properties
(23) by integrating the WVD along lines with different
and combinations of chirp rate and frequency shift. A large
s1 (t ), u2 (t )hk12 (t ) part of the WVD cross-terms is canceled by each other
s2 (t ) s2 (t )  u2 (t )hk2 (t ) . (24) through the integration and the residual part of the cross-
hk (t ) terms can be further reduced in the Radon-Wigner plane
Repeating this procedure, all signal components by noting the fact that the RWT auto- and cross-terms
can be obtained and the signal s(t) can be expressed have different character. Therefore, the cross-terms with
minimum distortion to the auto-terms are obtained. A
as s2 (t ) si (t ) . Based on this decomposition, the
i
WVD with substantially suppressed cross-terms can be
instantaneous frequencies of all signal components can obtained by transforming the masked RWT back to the
be obtained and then reconstructed based on the target Wigner plane. It is proved that the WVD auto-terms after
responses. the inverse Radon transform of the masked coefcients
are the same as those in the original WVD.
For multi-component signals with approximately
Chirp signal extraction equal magnitudes, RWT filtering in the Radon-Wigner
plane is effective. However, when the magnitudes of the
The Wigner-Ville distribution (WVD) of a signal s(t) signal components are signicantly different, the method
is dened as: may not be effective because of the cross-terms between
stronger signal components. The weaker signals may
f
W W be shaded in the presence of the stronger cross-terms
WVDs (t , Z ) s t  2 s * t  2 exp  jZW dW ,
f
(20)
and can hardly be detected. In GPR systems, the signal
echoes from small targets are often much weaker than
where variables t and represent the time and frequency, the clutter, even after clutter cancellation. In this case,
respectively. The WVD can be thought of as a joint time- the method we introduced above can be used to detect
frequency energy density function which describes the the stronger signal components and then remove them
signals energy distribution in the joint time-frequency from the original recorded GPR data.

60 (feet)
130 180 220 330 400 440 490 Unit:Inch

Diameter 8.5 Diameter 8.5 Diameter 6


Diameter 8.5 Hight 3.2 Hight 6.5
Hight 3.2 Hight 3.2
Dept 13 Dept 11
25 (feet)

B Dept 12 Dept 11 B
`

120 Diameter 6 Diameter 8.5


Diameter 8.5 Hight 6.5
Hight 3.2 Hight 3.2
Dept 7 Dept 11
Dept 15
A Diameter 6.5 A
`

Legth 29
40 Diameter 3.5
Dept 15.5
Fig. 3 The map of the experimental site and the buried targets.

Application field experiment. There are ten targets buried in the


near-surface including three UXO (same type with
different orientations, targets 1 and 2 are placed
Target detection horizontally and target 3 is slanted to surface level)
In the example, we use the GPR data from a real and seven landmine analogs (two metal and ve non-

196
Zeng et al.

metal with different depths). Figure 3 shows the with about 9% water content. The conductivity of the
target distributions and the target attributes are listed soil is shown in Figure 5, where we can see that the
in Table 1. Figure 4 shows pictures of the UXO and conductivity and the dielectrical properties change
landmines. The soil at the test site is clay-sand soil with frequency.

(a) (b) (c)


Fig. 4 Picture of UXO and the non-metal and metal landmine analogs used in the test site.

Table 1 Target description in the GPR detection test


(The unit of length, height, depth and diameter in inches)
Targets 1 2 3 4 5 6 7 8 9 10
Cylinder with conic
Shape Round case Cylinder
head
Length or height 29 3.2 6.5
Diameter 6.5 8.5 6
Depth in soil 13.5 15 13 10 13 9 11 9 7 9
Material Metal with plastic Metal Plastic
Content in target Air Air Air
Orientation Horizontal Slant Vertical Vertical

3. We can see that responses from targets 1 to 3 are


Sample1
Sample2 different and they are clearer after removing the surface
0.004 reflection. Figure 7 shows the result after adaptive
chirplet transform on the GPR data. We can clearly see
Conductivity (S/m)

0.003
the target response.
In target location identification, the accumulated
energy is usually used to acquire the target much
0.002 faster from a large amount of data. It is efficient in
practical application of target detection. Figure 8 is the
0.001 comparison between Kirchhoff migration and chirplet
transform. We can see that the data after adaptive chirplet
0 transform shows lower background and larger target
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
responses and the signal noise ratio is improved. Figure
Frequency(GHz)
9 shows the accumulated energy of targets in map view.
Fig. 5 The soil conductivity change with frequency of two The accumulated energy data are normalized to between
samples at the test site.
0 and 1000. We can see from Figure 9 that the targets
are much clearer even if there are some reections from
Data processing unknown targets.
We apply the algorithms discussed above to process In order to show the signal to noise ratio of the target
the GPR data acquired at the test site. Figure 6 is the responses, we use the normalized signal to noise ratio
target response WVD of profile AA shown in Figure (NSNR). The NSNR is dened as

197
Application of adaptive chirplet transform in GPR target detection
T p/2 where s(n) is the time sequence of signals. T is the target

n T p/2
s 2 (n) location and p is the target response width. Figure 10 shows
NSNR N , (30) the NSNR results of ten targets. We can see that the NSNR
s
n 1
2
( n) of the target responses after the adaptive chirplet transform
has been greatly improved over the raw and migrated data.

500

400
Frequency (MHz)

300

200

100

0
0 1.6 3.2 4.8 6.4 8.0 9.6 11.2 12.8 14.4 16
Time (ns)
Fig. 6 The target response WVD of prole AA.
0

10

15
0 10 20 30 40 50 60
Distance (feet)
Fig. 7 The images of targets 1-3 from the prole AA.

Conclusions resolution than Kirchhoff migration results.


1.00
Noramlized accumulation energy

We apply the adaptive subspace and adaptive chirplet 0.75


transforms in GPR data processing. They are efcient in
removing clutter and target detection. 0.50
In the clutter rejection, the adaptive subspace method
requires no prior knowledge and no model. It is easily
0.25
used in practice. The surface reflection removal is
efcient and the target response is well preserved.
0.00
The target signals can effectively be extracted from 0 10 20 30 40 50 60
a large amount of GPR data by the adaptive chirplet Distance (Feet)
transform and therefore it will be benecial for automatic Fig. 8 The accumulated energy of GPR data after
interpretation. The chirplet transform processing results Kirchhoff migration (solid blue line) and adaptive chirplet
show that the detection energy has higher SNR and transform (dashed red line).

198
Zeng et al.

20 800

15 600
Distance (feet)
10 400

5 200

0 0
0 5 10 15 20 25 30 35 40 45 50 55
Distance (feet)
(a)
20 240
200
15
Distance (feet)

160
10 120
80
5
40
0 0
0 5 10 15 20 25 30 35 40 45 50 55
Distance (feet)
(b)
Fig. 9 Map views of accumulated GPR energy after Kirchhoff migration (a) and chirplet
transform (b).

10 Raw Data
88(7), 1817 1826.
8 Migrated Data
Chirplet Transform Data Fritzsche, M., 1995, Detection of buried landmines using
6
NSNR

ground penetrating radar: Proceedings of the SPIE, 2496,


4 100 109.
2 Irving, J. D., and Knight, R. J., 2003, Removal of
0 wavelet dispersion from ground-penetrating radar data:
1 2 3 4 5 6 7 8 9 10 Geophysics, 68(3), 960 970.
Target Number
Lu, Y., Oruklu, E., Saniie, J., 2008, Fast chirplet transform
Fig. 10 The NSNR of the ten targets in different processing
with FPGA-based implementation: IEEE Signal
methods on GPR data.
Processing Letters, 15, 577 580.
Mann, S., and Haykin, S., 1991, The adaptive chirplet: An
References adaptive generalized wavelet-like transform: Proceedings
of the SPIE, 1565, 402 413.
Mann, S. and Haykin, S., 1992, Adaptive chirplet
Alaee, M., and Amindavar, H., 2008, Chirplet-based target transform: an adaptive generalization of the wavelet
recognition using RADAR technology: 5th IEEE Sensor transform: Optical Engineering, 31(6), 1243 1256.
Array and Multichannel Signal Processing Workshop, Mann, S. and Haykin, S., 1995, The chirplet transform:
SAM. 21 23 July, 451 454. Physical considerations: IEEE Trans. Signal Processing,
Bruschini, C., Gros, B., Guerne, F., Piece, P.Y., and 43, 2745 2461.
Carmona, O., 1998, Ground penetrating radar and Qian, S., Dunham, M. E., and Freeman, M. J., 1995, Trans-
imaging metal detector for antipersonnel mine detection: ionospheric signal recognition by joint time-frequency
J. Appl. Geophys., 40, 59 71. representation: Radio Sci., 30(6), 1817 1829.
Chowdhury, F., 1996, Kalman filter with hypothesis Salvati, J. L., Chen, C. C., and Johnson, J. T., 1998,
testing: A tool for estimating uncertain parameters: Theoretical study of a surface clutter reduction
Circuits Systems Signal Processing., 15(3), 291 311. algorithm: Proc. of 1998 IEEE International Geoscience
Dugnol, B., Fernndez, C., Galiano, G., and Velasco, J., and Remote Sensing, 3, 1460 1462.
2008, On a chirplet transform-based method applied to Singh, S., 2007, Sensors-An effective approach for the
separating and counting wolf howls: Signal Processing, detection of explosives: Journal of Hazardous Materials,
199
Application of adaptive chirplet transform in GPR target detection

144(1 2), 15 28. Signal Processing, 50(6), 1298 1306.


Turner, G., and Siggins, A. F., 1994, Constant Q attenuation
of subsurface radar pulses: Geophysics, 59(8), 1192
1200.
van Kempen, L., and Sahli, H., 2001, Signal processing
techniques for clutter parameters estimation and clutter
removal in GPR data for landmine detection: Proceedings Zhaofa Zeng received a B.S. (1988), an M.S. (1991)
of the 11th IEEE Signal Processing Workshop on in Applied Geophysics from
Statistical Signal Processing, 158 161. Changchun University of Earth
van der Merwe, A., and Gupta, I. J., 2000, A novel signal Sciences, and a Ph.D. (1997) in
processing technique for clutter reduction in GPR
Geophysics from Changchun
measurement of small, shallow land mines: IEEE
University of Science and
Transactions on Geoscience and Remote Sensing, 38(6),
Te c h n o l o g y, C h i n a . F r o m
2627 2637.
Wang, G., Xia, X., Root, B.T., Chen, V. C., Zhang, Y., and
October 2001 through October,
Amin, M., 2003, Maneuvering target detection in over- 2002, he was a visiting scholar
the-horizon radar using adaptive clutter rejection and at Tohoku University, Japan.
adaptive chirplet transform: IEEE Proc. Radar Sonar From 2006 to present, he is a
Navig., 150(4), 292 298. visiting professor at Delaware State University, USA.
Wang, J., and Zhou, J., 1999, Chirplet-based signal He is currently also a Professor at Jilin University. His
approximation for earthquake ground-motion model: research work is focusing on GPR, electromagnetic,
IEEE Signal Processing Mag., 16, 94 99. and other geophysical methods in engineering and
Yin, Q., Qian, S., and Feng, A., 2002, A fast renement for environment applications. He is a member of Chinese
adaptive Gaussian chirplet decomposition: IEEE Trans. Geophysical Society.

200
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

You might also like