You are on page 1of 6

NDT&E International 65 (2014) 2227

Contents lists available at ScienceDirect

NDT&E International
journal homepage: www.elsevier.com/locate/ndteint

A well posed inverse problem for automatic pavement parameter


estimation based on GPR data
Diogo B. Oliveira n, Douglas A.G. Vieira, Adriano C. Lisboa, Fillipe Goulart
ENACOM - Handcrafted Technologies, Prof. Jose Vieira de Mendonca, 770, Engenho Nogueira, Belo Horizonte, Minas Gerais, Brazil

art ic l e i nf o

a b s t r a c t

Article history:
Received 2 September 2013
Received in revised form
13 February 2014
Accepted 23 March 2014
Available online 2 April 2014

This work presents a methodology for solving the inverse problem of thickness and permittivity
determination in multilayer problems by using ground penetrating radar (GPR) data. Its main
contribution is to express the inverse problem as well-posed or, similarly, to formulate an error function
with only one global minimum under practical conditions. This result is achieved by establishing the
maximum allowed GPR operation frequency. With the analysis shown in this paper, the complex inverse
problems of multilayer parameters estimation are better understood and the algorithm for minimization
of the error function can be better selected. Therefore, it can reduce results error and decrease the time
for obtaining a solution. The new methodology is tested on two typical situations showing the accuracy
of the proposed method.
& 2014 Elsevier Ltd. All rights reserved.

Keywords:
Ground penetrating radar
Non-destructive testing
Automatic parameter estimation
Multilayer problem
Pavement

1. Introduction
Ground penetrating radar (GPR) is a common tool used for nondestructive testing. It is based on sending an electromagnetic wave
into a dielectric medium and measuring the reections due to its
inhomogeneities. It is used in several areas of knowledge such as
geophysics (e.g. analysis of rocks and soils, search of minerals,
measurement of sedimentation in rivers), archaeology (e.g. search
of buried objects and traces of civilizations), and civil engineering
(e.g. evaluation of conditions in structures such as buildings,
bridges) [1].
The estimation of thicknesses and physical properties of
layered media is a classic problem for GPR evaluation [2] and it
has a direct application in the analysis of roads. The evaluation of
the pavement properties is an important step for certifying that
the construction was done according to the project, and, also, to
measure its degradation. However, few companies supervisory use
GPR in their routine, mainly due to the difculty of analyzing a
large amount of measurements [3]. In this context, automatic tools
are of great importance.
Several papers approach the calculation of thickness in multilayer problems using the GPR by methods that detect the instant
when the reected pulses occur, estimating the thickness based
on the wave traveling time in each medium [3,4]. To detect the
n

Corresponding author. Tel.: 55 31 3401 1044.


E-mail addresses: diogo.oliveira@enacom.com.br (D.B. Oliveira),
douglas.vieira@enacom.com.br (D.A.G. Vieira),
adriano.lisboa@enacom.com.br (A.C. Lisboa), fgsm-c@enacom.com.br (F. Goulart).
http://dx.doi.org/10.1016/j.ndteint.2014.03.008
0963-8695/& 2014 Elsevier Ltd. All rights reserved.

presence of reected pulses, peak detectors [3] and lters are used
that perform a correlation with the type of GPR incident pulse [5].
However, the presence of multiple reections worsens the
performance of these kinds of methods since it is difcult to
associate the reected pulse with its respective interface where
the reection originates. Another disadvantage is that very thin
layers generate overlapped pulses making the task of obtaining the
reected pulses individually very complex.
The determination of electrical properties is usually simplied
by expressions that relate the permittivity with the amplitude of
the reected pulses [4]. In [3], the authors propose an iterative
scheme for detecting the amplitude of the reected pulses using
the least squares method. The accuracy of this method is highly
dependent on the evaluation of the reection instants.
Other approaches are used to solve the inverse problem
considering stratied media. In [6], signal processing techniques
are applied, such as principal component analysis and neural
networks. Moreover, in [79] the authors construct a quadratic
error function between the electric eld phasors obtained from
the GPR measurement and phasors calculated using the mathematical model of the problem considering the thickness and the
permittivity of each medium as variables. However, these methods
do not report the characteristics of the error function and they
apply, without a clear criteria, deterministic optimization methods
based on gradient or stochastic methods considering that the
function is multimodal [10].
This work presents an unprecedented discussion of how the
variables impact the behavior of the error function, and describes
a strategy for solving the inverse problem. Establishing a limit for

D.B. Oliveira et al. / NDT&E International 65 (2014) 2227

each thickness variable and a maximum operating frequency of


the GPR or increasing the number of frequencies in the GPR output
the error function is presented with only one global minimum, so
that the inverse problem becomes well-posed. This facilitates the
process of obtaining a solution and improves the results accuracy.

2. Problem denition
Fig. 1 shows a typical pavement modeled as a set of dielectric
layers. Each layer is characterized by its thickness and its electric
permittivity, which may be determined by the GPR evaluation.
For dry paved roads, the structure materials have low conductivity [11]. Hence, the electromagnetic wave propagates easily
through the structure, which leads to a better context for property
estimation. This work is interested in this context, so that conductive
and polarization losses are not taken into account.

23

where Ei and Ei are the elds in the interface i propagating in


the directions b
z and  b
z , respectively, according to Fig. 2.
At each interface, between two layers, an elementary reection
coefcient can be dened as

i  i 1
;
i i 1

i M; M 1; ; 1

p
where i 0 =i and 0 is the air magnetic permeability.
The eld at the interface i is related to that of interface i1
through the matrix below:
" #
"
#"
#
jk d
Ei
i e  jki di Ei 1
1 e i i

; i M; M  1; ; 1
3
Ei 1
Ei
i i ejki di e  jki di
where at each interface the coefcient i obeys the following
p
relationship: i i 1. The variable ki i is the propagation
constant and 2 f is the angular frequency.
Simultaneously solving (1), (2) and (3) results in the following
recursive relationship:

i Ri 1 e  2jki di
1 i Ri 1 e  2jki di

3. Mathematical model

Ri

A general pavement can be modeled as M1 dielectric layers


where the electromagnetic wave propagates. In Fig. 2 the electromagnetic model of this structure is shown. The thickness of each
layer is represented by d1 ; d2 ; ; dM and i is the characteristic
impedance and ki is the propagation constant of layer i.
To obtain the analytic solution of the electromagnetic elds
to this problem, a well-known methodology [12] is used. The
method computes the electric eld phasor reected by multilayer
structure.
Basically, the solution is a recursive expression that computes
the coefcient, Ri

where i M; M  1; ; 2; 1 and RM 1 M 1 .
The analysis of the pavement problem via GPR does not use the
information about the eld distribution throughout the structure.
The only eld used in the analysis is the phasor reected on the
surface of the pavement which is the phasor R1 .
For obtaining this phasor we should rst determine RM 1 then
RM , and so on recursively until reaching the rst interface. Therefore, the phasor R1 is a complex and nonlinear function dependent
on thickness variables, d1 ; d2 ; ; dM , the permittivity variables,
1 ; 2 ; ; M 1 , and the frequency. This feature of the R1 function
introduces much difculty for solving the inverse problem of the
pavement evaluation. To not overburden the nomenclatures in
paper the phasor R1 be only called R after that section.

Ri

Ei
Ei

Tx Rx

Air

4. Inverse problem

4.1. Denitions

d1

d2

d3

d4

The signal measured by a GPR is represented by a vector of n


phasors:
R0 R0 f 1 ; ; R0 f n ;

Fig. 1. Typical composition of the exible pavement.

where the vector f f 1 ; ; f n  contains frequencies that are


composed of the signal emitted by the GPR and R0 f j is the
measurement of the electric eld phasor reected by the multilayer at frequency f f j , j 1; ; n.
The signal evaluated by a mathematical model is represented
by a vector of n phasors:
Rd; Rf 1 ; ; Rf n ;

where d d1 ; ; dM  and 1 ; ; M 1  are vectors that contain


the thickness and permittivity of the M1 composing layers,
respectively, and Rf j is the electric eld phasor reected by
the multilayer at frequency f f j , j 1; ; n. The methodology
to calculate, in frequency domain, the electric eld phasor
reected by multilayer dielectric was shown in Section 3. The
incident electric eld on the multilayer structure is a plane
electromagnetic wave.
Then, we can dene the quadratic error function
erf Rd;  R0 22 ;

Fig. 2. General pavement modeled as a M 1 dielectric layers.

which is the difference between the predicted vector computed


using the mathematical model, Rd; , and the vector R0 obtained
from the measurement.

24

D.B. Oliveira et al. / NDT&E International 65 (2014) 2227

The minimization of erf d; is, in general, a nonlinear and


multimodal problem with 2M 1 variables (the thickness of
the deepest layer is considered innite). Just minimizing the
error function (7) is not enough to guarantee the correct solution
because the inverse problem may have many global minima, i.e., it
may be ill-posed. However, in the following section we show a
methodology to present the inverse problem as well-posed. This is
achieved by considering limits on the variables of the problem
and/or by selecting the GPR working frequencies.
The whole discussion presented in the following sections is also
very important for the selection of the optimization algorithms
used in the minimization process. The presence or the absence of
multiple local or global minimum inuences that selection.
4.2. Analysis of the error function with respect to the layer
thickness variables
The electric eld phasor reected by the multilayer, computed
by mathematical model and expressed through Rd; , depends on
thickness and permittivity of each layer, as well as the frequency
emitted by the GPR. Considering this inuence, the erf also
depends on the thickness and permittivity variables (as seen in
(7)). In this section, the error function behavior with respect to
thickness variables is presented.
Consider a exible pavement with ve layers described in
Section 5 (see Table 1) as a practical example to demonstrate
how the phasor R changes its magnitude as the variable d2 is
changed. The phasor R is represented by Rd2 to highlight this
dependency. Basically, what we do is to evaluate the expression (4)
considering the values of variables, d1 ; d3 ; d4 ; 1 ; ; 5 with the
same values as those presented in Table 1 and changing the values
of the variable d2 . Fixing the values of some variable is used only to
help the visualization of the nal result.
For this situation, Fig. 3 shows the magnitude of the phasor Rd2 ,
represented by jRd2 j, considering a GPR frequency of fj 1.0 GHz.
The jRd2 j function is a periodic function with a period of 2;f j =2,
where 2;f j is the wavelength at layer 2 , considering a frequency
Table 1
Flexible pavement parameters.
Layer

Thickness (m)

Relative permittivity

1st
2nd
3rd
4th
5th

0.04
0.15
0.075
0.15
1

4.0
5.0
4.0
9.0
11.0

equal to f j computed as follows:


1

2;f j pp
f j 0 0 2

where 0 is the air electric permittivity.


This relation of periodicity is observed for any thickness variable.
If the analysis is in the thickness variable di the period of the
function jRdi j is i;f j =2, i.e., the period is half of the wavelength of
the electromagnetic wave with frequency f j in layer i.
For obtaining the phasor from the measurement, R0 , we use the
mathematical model described in Section 3. Considering that the
mathematical model computes the electric eld with accuracy this
procedure is valid. The procedure is computing the phasor
reected by the multilayer with all the thickness and permittivity
n
variables taken from Table 1 including d2 , i.e., R0 Rd ; n in
n
n
which d ; are the parameters of the assessed pavement.
Fig. 3 shows the magnitude of the R0 phasor. The constant value
is obtained and it is shown in gure for all d2 values. This is done
in order to show where the jRd2 j and jR0 j functions intersect
each other.
With the phasors Rd2 and R0 obtained the erf function which
can be computed through (7). Fig. 3 also shows the error function.
Note that the behavior of erf is governed by the behavior of Rd2
and the same periodicity is observed. In half times that the jRd2 j
and jR0 j functions cross each other the erf has null value because
the phasors R and R0 are equal. In the others situations where the
functions cross each other, the magnitude of these phasors is equal
and the phase is different, then, the expression (7) is not null.
To solve the inverse problem, we have to minimize the error
function. Since the erf function has many global minima, the inverse
problem is ill-posed. However, if, instead of using only one frequency
to excite the multilayer structure, we use two frequencies, the error
function takes the form as shown in Fig. 4. In this gure we have two
signals emitted by the GPR with frequencies equal to fj 1.0 GHz and
fk 1.5 GHz and their respective error functions, erf1 and erf2. Each
erf has different periods and their sum yields a function with only
one global minimum.
Another way to make the problem well-posed is to consider
only a vicinity of the optimum point. If we consider a range of
i;f k =2 around the optimum point (d2 0.15 m), as seen in Fig. 4, it
is easy to see that to the left of this point the error function is
monotonically decreasing and to the right the function is monotonically increasing, where the fk is the highest frequency emitted
by the GPR equipment.
Then, let us assume that each layer i has thickness di and a
range of di (feasible deviation for each thickness variable). The
erf function would have one global minimum if half of the
0.25

0.7

erf (fj=1.0GHz)

|R(d )|
2

0.6

|R |
1/2

erf = erf1 + erf2

erf(d )

2,f

0.5

erf2 (fk=1.5GHz)

0.2

0.15

2,f

erf

0.4

1/2

1/2 2,f

0.1

0.3
0.2

0.05

0.1
0

0
0.05
0

0.1

0.15

0.2

0.3

0.4

thickness, d2 [m]
Fig. 3. Analysis the of error function with respect to layer thickness.

0.1

0.15

0.2

0.25

thickness, d2 [m]
Fig. 4. Behavior of the quadratic error function considering two frequencies
emitted by the GPR.

D.B. Oliveira et al. / NDT&E International 65 (2014) 2227

wavelength in each medium is equal or greater than di , considering the maximum frequency that GPR emits. This means that
the previous analysis done for layer 2 must be done for each layer.
Under this condition and considering that the range of the
thickness variables is a prior determined, the maximum frequency
irradiated by the GPR must be
1
2

di r i;f max

25

the presence of one global minimum of the erf despite the


presence of the local minimum. In the vicinity of the optimum,
the erf function tends to a convex function.
So, the erf function has one global minimum in relation to
permittivity variable. However, if the materials used in the pavement are well characterized and a small range for feasible permittivity variables is used the error function has a unique minimum.

1 1
1 1
f max r p p
2 0 0 i di

10

which considers the approximation of a plane electromagnetic


wave in each medium. The maximum frequency is calculated to
each layer and the lowest one is used to set up the GPR, ensuring
one global minimum for the error function.
In short, if more than one frequency is used in the GPR to excite
the pavement the error function has a unique global minimum and
has also many local minima. If the limit of the thickness variables
is known and the relation (10) is used to limit the maximum
frequency the error function has one global minimum using only
one frequency.
4.3. Analysis of the error function with respect to the layer
permittivity variables
The analysis of the error function with respect to the relative
permittivity is a more complicated task. Several characteristics
of the wave propagation in each medium of the multilayer are
affected by a variation in permittivity, such as wavelength, reection
and transmission coefcients of each interface and the propagation
velocity.
All these physical features make it difcult to understand the
behavior of erf as a function of permittivity only. It is very hard to
derive a similar deterministic analysis here because the variation
in the thickness of each layer introduces only a variation in the
phase of the electromagnetic wave. In the case of permittivity
variable, many characteristics of the electromagnetic wave are
affected. Nevertheless, some interesting notes can be deducted.
Again, the erf analysis considers a single variable, erf 2 , as the
same way done in the previous section for the thickness variable.
The pavement described in Section 5 and Table 1 is used.
Fig. 5 exemplies the dependency between the permittivity
variable and the error function. The erf function is zero in the
optimum (2 5.0) and it increases as the permittivity increases
or decreases moving away from the optimum. This behavior
is presented for high and low frequencies, even when the erf
function is more oscillatory at high frequency. This fact suggests

5. Results
The automatic method of parameters estimation in the pavement problem is tested in two cases in this section. The values of
R0 are computed through the analytic expression of a mathematical model in order to simulate the real word measurement.
5.1. Flexible pavement
The pavement analyzed here is described in [3] and Table 1
shows its properties on each layer.
Table 2 shows the range of thickness and permittivity variables,
which determines the feasible domain where the problem of minimizing (7) is solved. The maximum frequency is readily computed for
each layer, using the expression (3), ensuring that the erf function has
only one global minimum. The lowest frequency in this table is used
as the highest frequency emitted by the GPR.
With these guidelines, it follows an analysis of the error function
considering one and two variables. For the thickness variables, the
erf function has one global minimum as shown in Fig. 6 and, for
permittivity variables, the erf is shown in Fig. 7. Fig. 8 shows the
level sets of erf considering 1 and d1 as variables. These gures
show that the error function has only a single global minimum.
Table 2
Variable limits and the computation of the maximum frequency of the GPR.
Layer

d (m)

fmax (GHz)

1st
2nd
3rd
4th
5th

0.030.05
0.11250.1875
0.0550.095
0.11250.1875

3.05.0
2.57.5
3.05.0
7.011.0
8.514.5

3.75
0.89
1.87
0.66

Table 3
Calculation of the maximum frequency.
Layer

d (m)

f max (GHz)

1st
2nd

0.0750.125

2.04.0
6.010.0

1.75

erf1 (f = 200 MHz)

0.4

erf 2 (fk = 1.5 GHz)

0.35

0.04

erf(d )
1

erf(d )
2

0.03

0.25

erf(d3)
erf(d )

0.2

erf

erf

0.3

0.15

0.02

0.1
0.01

0.05
0

10

12

14

16

18

relative permittivity,2
Fig. 5. Behavior of the error function with respect to relative permittivity for low
and high frequencies.

0.05

0.1

0.15

0.2

0.25

thickness [m]
Fig. 6. Function behavior considering each thickness variable separately.

26

D.B. Oliveira et al. / NDT&E International 65 (2014) 2227

x 103

0.8

erf(2)

erf(3)

0.6

erf( )

10

erf

erf

0.4

5
0.2
0

10

relative permittivity

relative permittivity
0.025

erf(4)

0.15

erf( )
5

0.1

erf

erf

0.02
0.015
0.01

0.05

0.005
4

9 10

12

14

10 11 12

14

16

relative permittivity

relative permittivity

Fig. 7. Function behavior considering each permittivity variable separately.

Again, the ellipsoid algorithm nds the optimum within the


machine epsilon error for thickness and permittivity. In this case,
the GPR output signal has only one frequency as presented in
Table 4.

relative permittivity 1

4.5

6. Conclusion

3.5

3
0.03

0.035

0.04

0.045

0.05

thickness d1 [m]
Fig. 8. Function behavior considering two variables: level sets of erf d1 ; 1 .

Table 4
Frequencies emitted by the GPR.
Case

Frequencies (GHz)

Flexible pavement
Simple pavement

0.60.550.50.450.40.350.3
1.75

The ellipsoid method [13] has been used to solve the minimization problem of erf. The algorithm nds the global optimum
within the machine epsilon error for thickness and permittivity.
The frequencies contained in the output signal of the GPR are
presented in Table 4. These frequencies respect to the maximum
frequency of 0.66 GHz which is given in Table 2.

This paper presents a method for the automatic analysis of


multilayer structures using GPR data. It has been shown that, from
the physical knowledge of the problem, it is possible to deduce
some behaviors about the error function. Information like quantities of global minimum, the inuence of the GPR output
frequency and the range of the variables are studied. This information implies in some simplications in the solution of the inverse
problem and in the selection of the optimization algorithm for
solving the minimization problem.
The performance of the method on two typical situations has
been shown. In both cases the thickness and permittivity parameters were found within the machine epsilon error.
One important implication of this work is to show that GPR
equipment used in the pavement evaluation which can use an
antenna with narrow band. The error function can have a single
global minimum using only one single GPR frequency. This allows
the construction of a simpler GPR equipment at a lower cost.

Acknowledgments
This work was supported by National Council for Scientic and
Technological Development (CNPq) and Foundation for Research
Support in Minas Gerais (FAPEMIG), Brazil.
References

5.2. Simple pavement


In this case, the pavement is composed of two layers: dry
asphalt (1 3.0) and dry clay (2 8.0) [6]. The thickness of the
asphalt layer is d1 0.10 m. Table 3 shows the variable limits and
the highest frequency emitted by that GPR such as erf contains
only one global minimum.

[1] Daniels D. Ground penetrating radar, IEE radar, sonar, navigation, and avionics
series, 2nd ed. The Institution of Engineering and Technology; 2004.
[2] Gordon M, Broughton K, Hardy M. The assessment of the value of GPR imaging
of exible pavements. NDT&E Int 1998;31(6):42938.
[3] Lahouar S, Al-Qadi IL. Automatic detection of multiple pavement layers from
gpr data. NDT&E Int 2008;41:6981.
[4] Loizos A, Plati C. Accuracy of pavement thicknesses estimation using different
ground penetrating radar analysis approaches. NDT&E Int 2007;40(2):14757.

D.B. Oliveira et al. / NDT&E International 65 (2014) 2227

[5] Zhou H, Li S, Zhu J. Automatic layer-interface detection of pavement based on


matched lter. In: 2010 Sixth international conference on wireless communications networking and mobile computing (WiCOM), 2010. p. 13.
[6] Pantoja MF, Rodriguez JB, Bretones AR, de Jong CM, Garcia SG, Martin RG, et al.
Application of neural network and principal component analysis to GPR data.
In: 2011 Sixth international workshop on advanced ground penetrating radar
(IWAGPR), 2011. p. 14.
[7] Spagnolini U. Permittivity measurements of multilayered media with monostatic pulse radar. IEEE Trans Geosci Remote Sens 1997;35(2):45463.
[8] Loulizi A, Al-Qadi I, Lahouar S. Optimization of ground-penetrating radar data to
predict layer thicknesses in exible pavements. J Transp Eng 2003;129(1):939.
[9] Andre F, Tran A, Mourmeaux N, Lambot S. Integrated modeling of near-eld
ground-penetrating radar and electromagnetic induction data for reconstructing

[10]

[11]
[12]
[13]

27

multilayered media. In: 2012 14th International conference on ground penetrating radar (GPR), 2012. p. 40712.
Travassos X, Vieira D, Ida N, Vollaire C, Nicolas A. Inverse algorithms for
the GPR assessment of concrete structures. IEEE Trans Magn 2008;44(6):
9947.
Fauchard C, Drobert X, Cariou J. GPR performances for thickness calibration
on road test sites. NDT&E Int 2003;36(2):6775.
Balanis CA. Advanced electromagnetic engineering. Wiley, New York, USA;
1989.
Vieira DAG, Lisboa A, Saldanha R. An enhanced ellipsoid method for electromagnetic devices optimization and design. IEEE Trans Magn 2010;46(8):
284351.

You might also like