ARTICLE IN PRESS

Computers & Geosciences 32 (2006) 1247–1258

www.elsevier.com/locate/cageo

Numerical modeling of ground-penetrating radar

in 2-D using MATLAB$
James Irving, Rosemary Knight
Geophysics Department, Stanford University, Room 360, Mitchell Building, Stanford, CA 94305, USA
Received 29 August 2005; received in revised form 13 November 2005; accepted 14 November 2005

Abstract

We present MATLAB codes for finite-difference time-domain (FDTD) modeling of ground-penetrating radar (GPR) in
two dimensions. Surface-based reflection GPR is modeled using a transverse magnetic (TM-) mode formulation. Crosshole
and vertical radar profiling (VRP) geometries are modeled using a transverse electric (TE-) mode formulation. Matrix
notation is used in the codes wherever possible to optimize them for speed in the MATLAB environment. To absorb waves
at the edges of the modeling grid, we implement perfectly matched layer (PML) absorbing boundaries. Although our codes
are two-dimensional and do not incorporate features such as dispersion in electrical properties, they capture many of the
important elements of GPR surveying and run at a fraction of the computational cost of more elaborate algorithms. In
addition, the codes are well commented, relatively easy to understand, and can be easily modified for the user’s specific
purpose.
r 2005 Elsevier Ltd. All rights reserved.

Keywords: Ground-penetrating radar; GPR; Numerical modeling; Electromagnetics; Finite-difference time domain; FDTD

1. Introduction (ii) borehole surveying, where one or both antennas

are located in boreholes and subsurface properties
Ground-penetrating radar (GPR) is a popular are estimated tomographically. Of interest in our
geophysical method for high-resolution imaging of research is the application of both surface and
the shallow subsurface. The GPR technique can be borehole GPR to hydrogeological problems. Speci-
divided into two main modes of operation: fically, we are interested in using these techniques to
(i) surface-based reflection surveying, where the assist in the development of hydrogeological models
transmitter and receiver antennas are located on the that predict groundwater flow and contaminant
surface of the earth and the subsurface is imaged in transport. A critical step in using GPR for this
terms of changes in its electrical properties, and purpose is determining the link between the hydro-
geological properties that govern these processes,
$
Code available from server at http://www.iamg.org/CGEditor/ and the information contained in a GPR data set.
index.htm Numerical GPR models provide one means of
Corresponding author. Tel.: +1 650 724 9939;
fax: +1 650 725 7344.
exploring the link between subsurface properties
E-mail addresses: [email protected] (J. Irving), and GPR data. We can create a model of a
[email protected] (R. Knight). subsurface region of interest, where we define the

0098-3004/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2005.11.006
 ARTICLE IN PRESS
1248 J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258

subsurface in terms of its lithological or hydro- examples of the use of our codes, one showing
geological properties. We can then transform this modeling of a reflection GPR survey and the other
model into one that represents the subsurface in modeling of a crosshole GPR survey.
terms of its electrical properties. GPR modeling can
then be used to simulate the acquisition of data in
2. Theory
this subsurface region. The synthetic data that are
obtained can be used to advance our understanding
2.1. Governing equations
of the way in which information about the spatial
variability of subsurface properties is captured by,
We begin the theory behind our GPR modeling
and can be extracted from, GPR data.
codes with Maxwell’s curl equations in the fre-
A number of approaches have been presented for
quency domain, which are
the numerical modeling of GPR data. These include
ray-based methods (Goodman, 1994; Cai and r
E ¼ iomH, (1)
McMechan, 1995), frequency-domain methods
(Powers and Olhoeft, 1994; Zeng et al., 1995), r
H ¼ sE þ io
E, (2)
integral methods (Ellefsen, 1999), and pseudospec- pffiffiffiffiffiffiffi
where i ¼ 1, o is angular frequency,
, m, and s
tral methods (Carcione, 1996; Casper and Kung, are the dielectric permittivity, magnetic permeabil-
1996; Lui and Fan, 1999). What has become by far ity, and electrical conductivity parameters, respec-
the most common approach for GPR modeling over tively, and E and H are the electric and magnetic
the past decade, however, is the finite-difference field vectors. To implement PML absorbing bound-
time-domain (FDTD) technique (e.g., Wang and aries in our codes, we consider the general case of a
Tripp, 1996; Bourgeois and Smith, 1996; Bergmann complex stretched coordinate space (e.g., Chew and
et al., 1996; Teixeira et al., 1998; Holliger and Weedon, 1994; Gedney, 1998), where the r operator
Bergman, 2002). Reasons for this include that the takes the following form:
FDTD approach is relatively conceptually simple,
accurate for arbitrarily complex models, and cap- 1 q 1 q 1 q
r ¼ x^ þ y^ þ z^ , (3)
able of accommodating realistic antenna designs sx qx sy qy sz qz
and features such as dispersion in electrical proper-
where
ties (Taflove, 1995). What is lacking, however, are
FDTD modeling codes for GPR, freely available for sk
s k ¼ kk þ ; k ¼ x; y; z (4)
the public use, that are easy to understand and ak þ io
0
modify. are complex coordinate stretching variables that
Here, we present FDTD codes, written in the vary only in the k direction (Kuzuoglu and Mittra,
MATLAB programming language, for basic surface 1996). Here,
0 is the dielectric permittivity of free
reflection and borehole GPR modeling in two space, and sk , kk , and ak are parameters that can be
dimensions. Although 2-D modeling is limited in specified to allow for wave propagation in the
the sense that it cannot fully account for antenna interior of the modeling grid and wave absorption in
behavior and out-of-plane variations in material the PML boundary regions. It should be stressed
properties, our codes capture many of the important that sk , kk , and ak are not true electrical properties.
features of GPR surveying and run at a fraction of Rather they are parameters that, through complex
the computational cost of fully 3-D algorithms. The coordinate stretching, add additional degrees of
codes feature perfectly matched layer (PML) freedom to Maxwell’s equations to allow for PML
absorbing boundaries to avoid reflections from the boundary implementation.
edges of the modeling grid. To optimize the Taking the components of Eqs. (1) and (2) using
programs for speed in MATLAB, matrix notation the identity in Eq. (3), and assuming that there is no
is used wherever possible. To begin, we discuss the variation in the y direction for 2-D modeling, we
theory behind our codes, including the governing arrive at the following two decoupled sets of partial-
analytical equations, their finite-difference approx- differential equations involving the fH x ; H z ; E y g and
imations, numerical stability and dispersion criteria, fE x ; E z ; H y g field components:
and boundary conditions. Next, we briefly discuss
how the FDTD equations are implemented in the 1 qE y
iomH x ¼  , (5a)
MATLAB environment. Finally, we present two sz qz
 ARTICLE IN PRESS
J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258 1249

1 qE y we have
iomH z ¼ , (5b)   
sx qx dðtÞ sk t sk
s1
k ðtÞ ¼  exp  þ a k uðtÞ
1 qH z 1 qH x kk
0 k2k
0 kk
sE y þ io
E y ¼  (5c) dðtÞ
sx qx sz qz ¼ þ zk ðtÞ, ð7Þ
kk
and
where dðtÞ is the Dirac delta function and uðtÞ is the
1 qH y Heaviside step function. Using this expression and
sE x þ io
E x ¼ , (6a)
sz qz assuming frequency-independent material proper-
ties, Eqs. (5a)–(5c) and (6a)–(6c) are transformed
1 qH y into the time domain as follows:
sE z þ io
E z ¼  , (6b)
sx qx qH x 1 qE y qE y
m ¼  zz ðtÞ  , (8a)
qt kz qz qz
1 qE x 1 qE z
iomH y ¼  . (6c)
sz qz sx qx qH z 1 qE y qE y
m ¼ þ zx ðtÞ  , (8b)
Eqs. (5a)–(5c) and (6a)–(6c) are the transverse qt kx qx qx
magnetic (TM-) and transverse electric (TE-) mode
qE y 1 qH z 1 qH x
sets of equations in the stretched coordinate space, sE y þ
¼ 
qt kx qx kz qz
respectively. For surface-based reflection GPR
modeling, where the antennas are oriented qH z qH x
þ zx ðtÞ   zz ðtÞ  ð8cÞ
perpendicular to the x2z survey plane, we use the qx qz
TM-mode equations. For crosshole and vertical and
radar profiling (VRP) geometries, where the anten- qE x 1 qH y qH y
nas are contained within the survey plane, the sE x þ
¼ þ zz ðtÞ  , (9a)
qt kz qz qz
TE-mode equations are employed. It is important to
note that when the stretching parameters are set to qE z 1 qH y qH y
unity, Eqs. (5a)–(5c) and (6a)–(6c) become the sE z þ
¼  zx ðtÞ  , (9b)
qt kx qx qx
standard TM- and TE-mode equations in an
unstretched coordinate space. In the interior of the qH y 1 qE x 1 qE z qE x qE z
modeling grid, we therefore set sx ¼ sz ¼ 1. In the m ¼  þ zz ðtÞ   zx ðtÞ  ,
qt kz qz kx qx qz qx
PML boundary regions of the grid, however, sx and (9c)
sz are given complex values, which allows significant
wave absorption to occur. Unlike in the simpler case where  represents convolution. We now describe
of a change in electrical conductivity in these how Eqs. (8a)–(8c) are approximated using finite-
regions (which would also cause absorption), differences in our TM-mode modeling code for
altering the coordinate stretching variables does GPR reflection profiling. The derivations for the
not result in a change in electromagnetic (EM) TE-mode modeling code using Eqs. (9a)–(9c) are
impedance from the interior of the grid, and thus no not included here, as they are performed in a nearly
reflections are created at the edges of the PML identical manner.
domain.
We implement PML absorbing boundaries in our 2.2. Finite-difference approximations
modeling codes using the convolutional PML
(CPML) approach of Roden and Gedney (2000). To numerically model Eqs. (8a)–(8c), we use a
This method makes use of time-domain expressions leap-frog, staggered-grid approach that involves
for 1=sx and 1=sz in the FDTD formulation, and offsetting the electric and magnetic field compo-
avoids the splitting of the electric and magnetic field nents in both space and time such that the finite-
components common in other PML approaches difference approximations of the partial derivatives
(e.g., Berenger, 1994; Fang and Wu, 1996). For this in each equation are centered on the same spatio-
reason, we find the CPML approach to be the most temporal location (Yee, 1966). Fig. 1 shows the
intuitive implementation of PML. Taking the configuration of H x , H z , and E y in space for our
inverse Fourier transform of the inverse of Eq. (4), TM-mode modeling code. All spatial derivatives are
 ARTICLE IN PRESS
1250 J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258

h
nþ1 nþ1=2
Ey Hz Hx E y ji;j ¼ C a ji;j ½E y jni;j  þ C bx ji;j H z jiþ3=2;j
0 i
nþ1=2 nþ1=2 nþ1=2
þ27H z jiþ1=2;j  27H z ji1=2;j þ H z ji3=2;j
h
1 nþ1=2 nþ1=2
 C bz ji;j H x ji;jþ3=2 þ 27H x ji;jþ1=2
i
nþ1=2 nþ1=2
27H x ji;j1=2 þ H x ji;j3=2
2 h i
nþ1=2 nþ1=2
þ C c ji;j CE yx ji;j  CE yz ji;j , ð10cÞ
Index j

3 where the subscripts indicate spatial position and

nþ1=2
the superscripts indicate time (i.e., H x ji;jþ1=2 repre-
sents the H x field component at position ðx; zÞ ¼
4
ðiDx; ðj þ 1=2ÞDzÞ and time t ¼ ðn þ 1=2ÞDt, where
Dx and Dz are the horizontal and vertical field
5 discretization intervals, and Dt is the time step).
FDTD modeling using these equations is accom-
plished by alternately updating the electric and
6 magnetic fields. This moves forward in time since
0 1 2 3 4 5 6 the two fields are temporally staggered by Dt=2. As
Index i can be seen from the equations, all field updates are
Fig. 1. Spatial arrangement of H x , H z , and E y field components fully explicit.
for TM-mode modeling. Electric and magnetic field components The update coefficients C a , C bx , C bz , C c , Dbx , Dbz ,
are also staggered in time by Dt=2. and Dc in Eqs. (10a)–(10c) are given in terms
of the electrical properties and grid parameters as
approximated using fourth-order-accurate finite- follows:
difference expressions. The time derivatives are   
sDt sDt 1
approximated with second-order-accurate expres- Ca ¼ 1  1þ , (11a)
sions. This Oð2; 4Þ scheme is identical to that 2
2

described in Bergmann et al. (1996) for 1-D GPR  
Dt sDt 1
modeling. The conduction current density term in C bk ¼ 1þ ð24kk DkÞ1 , (11b)

2

Eq. (8c) is modeled using a semi-implicit approx-  
imation (e.g., Taflove, 1995, p. 64) which has been Dt sDt 1
Cc ¼ 1þ , (11c)
shown to have superior numerical properties over
2

one-sided expressions for this term (Bergmann
Dt
et al., 1996). Dbk ¼ ð24kk DkÞ1 , (11d)
After substituting the appropriate finite-differ- m
ence expressions into Eqs. (8a)–(8c) and solving for Dt
Dc ¼ . (11e)
the updated electric and magnetic field components, m
we arrive at the following FDTD update equations: Although not explicitly indicated, these coefficients
h are all functions of position since
, m, s, kx , and kz
nþ1=2 n1=2
H x ji;jþ1=2 ¼ H x ji;jþ1=2  Dbz ji;jþ1=2 E y jni;jþ2 are, in general, spatially varying. The convolution
i terms in Eqs. (8a)–(8c) are modeled using the
þ27E y jni;jþ1  27E y jni;j þ E y jni;j1 recursive convolution technique (Luebbers and
h i
 Dc ji;jþ1=2 CH xz jni;jþ1=2 , ð10aÞ Hunsberger, 1992), and are accounted for in Eqs.
(10a)–(10c) through the CH xz , CH zx , CE yx , and CE yz
terms. These are defined as
h h i
nþ1=2 n1=2
H z jiþ1=2;j ¼ H z jiþ1=2;j þ Dbx jiþ1=2;j E y jniþ2;j CH xz jni;jþ1=2 ¼ Bz ji;jþ1=2 CH xz jn1
i;jþ1=2
i h
þ27E y jniþ1;j  27E y jni;j þ E y jni1;j þ Az ji;jþ1=2 E y jni;jþ2 þ 27E y jni;jþ1
h i i
þ Dc jiþ1=2;j CH zx jniþ1=2;j , ð10bÞ 27E y jni;j þ E y jni;j1 , ð12aÞ
 ARTICLE IN PRESS
J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258 1251

h i
CH zx jniþ1=2;j ¼ Bx jiþ1=2;j CH zx jiþ1=2;j
n1 (e.g., Moghaddam et al., 1991; Xu and McMechan,
h 1997). Nevertheless, as stated previously, our 2-D
þ Ax jiþ1=2;j E y jniþ2;j þ 27E y jniþ1;j codes capture many of the important features of
i reflection GPR surveying, and can provide much
27E y jni;j þ E y jni1;j , ð12bÞ insight into the interaction of EM waves with a
h i complex subsurface. They should be used with
nþ1=2 n1=2
CE yx ji;j ¼ Bx ji;j CE yx ji;j caution, however, when antenna radiation patterns
h play a critical role in how the data are analyzed.
nþ1=2 nþ1=2
þ Ax ji;j H z jiþ3=2;j þ 27H z jiþ1=2;j A key example is crosshole GPR attenuation
i tomography.
nþ1=2 nþ1=2
27H z ji1=2;j þ H z ji3=2;j , ð12cÞ
h i 2.3. Numerical stability and dispersion criteria
nþ1=2 n1=2
CE yz ji;j ¼ Bz ji;j CE yz ji;j
h
nþ1=2 nþ1=2
þ Az ji;j H x ji;jþ3=2 þ 27H x ji;jþ1=2 An important step in FDTD modeling is choos-
i ing appropriate time and spatial discretization
nþ1=2 nþ1=2 intervals for a simulation. Ideally, we would like
27H x ji;j1=2 þ H x ji;j3=2 , ð12dÞ
to have Dx, Dz, and Dt as large as possible to make
where the simulation run most quickly. However, if Dt is
sk too large, the FDTD scheme presented above will
Ak ¼ ðBk  1Þ, (13a)
sk kk þ ak k2k become numerically unstable. In addition, if Dx or
   Dz are too large, the electric and magnetic fields will
Dt sk
Bk ¼ exp  þ ak (13b) be inadequately sampled in space and numerical

0 kk
dispersion will be the result. For the O(2,4) scheme
are PML update coefficients that, again, vary with presented here, the maximum time step that can be
location in the modeling grid. As can be seen from used, in order for the scheme to remain numerically
Eqs. (12a)–(12d), the values of the convolution stable, is (Georgakopoulos et al., 2002)
terms at the current time step are computed from sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
those at the previous time step. Therefore CH xz , 6 mmin
min
Dtmax ¼  , (14)
CH zx , CE yx , and CE yz must be stored, in addition to 7 1=Dx2 þ 1=Dz2
the H x , H z , and E y field components, during the
FDTD simulation. Throughout the above equa- where mmin and
min are the minimum magnetic
tions, we have kept our notation consistent with permeability and dielectric permittivity values pre-
that of Roden and Gedney (2000) so that their sent in the modeling grid. To control numerical
paper can be consulted for further details on CPML dispersion, the O(2,4) scheme must allow for five
implementation. field samples per minimum wavelength (Bergmann
To introduce an electric field source into the grid et al., 1996; Georgakopoulos et al., 2002). The
during FDTD modeling using Eqs. (10)–(13), we MATLAB programs finddt.m and finddx.m
add a source pulse function to the update for the E y determine the maximum possible Dt, Dx, and Dz
field component at the desired spatial location. This given a model’s electrical properties as input, and
amounts to adding the source function to the should be run prior to starting any simulation.
y-component of the current density term in Max-
well’s equations. To model receivers in our code, we 2.4. PML absorbing boundaries
simply record the E y field component as a function
of time at the receiver locations. It must be stressed Compared with other absorbing boundary types,
that, because our modeling codes are 2-D, all such PML boundaries possess a number of significant
sources and receivers are actually line elements, advantages. First, the PML approach offers super-
extending to positive and negative infinity in the ior attenuation of reflections from the edges of the
dimension perpendicular to the survey plane. As a modeling grid, and requires only a small number of
result, the radiation patterns and geometrical cells to be very effective (Gedney, 1998). Second, to
spreading for realistic, dipole-type GPR antennas implement PML, only the coordinate stretching
cannot be properly modeled with our codes; this variables need to be changed in the boundary
would require either a fully or pseudo 3-D approach regions, and not the FDTD update equations
 ARTICLE IN PRESS
1252 J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258

(i.e., the same update equations are used everywhere kkmax is the maximum value. Similarly, for sx and sz ,
in the grid). For this reason, PML is well suited to we have
parallel implementations. Finally, the CPML ap- 8
>
< 0 in grid interior;
proach that we have chosen for our codes has the 
advantage of being media independent (Roden and sk ¼ d m (16)
: d skmax inside PML region;
>
Gedney, 2000). That is, the approach is implemen-
ted in the same manner no matter what the
where skmax is the maximum value. Here, kk and sk
properties of the materials being modeled. For
vary only along the k direction since the coordinate
example, our FDTD formulation could be modified
stretching variables are one-dimensional functions.
to allow for dispersion in
, m, and s (say, using the
For example, sx will be zero throughout the interior
technique of Bergmann et al. (1998)), and the C
of the modeling grid and non-zero only in the left
terms for invoking CPML in Eqs. (10a)–(10c) would
and right PML regions. The kz parameter, on the
remain unchanged.
other hand, will be one throughout the interior of
As mentioned previously, in the interior of the
the grid and greater than one only in the top and
modeling grid, the coordinate stretching variables,
bottom PML regions. In our codes, we use m ¼ 4
sx and sz , are set to unity such that Eqs. (10)–(13)
and set kxmax ¼ kzmax ¼ 5 by default, which we have
become the standard, TM-mode, FDTD update
found is adequate for the effective absorption of
equations in an unstretched coordinate space. This
evanescent waves. The maximum value for sk is
requires that kx ¼ kz ¼ 1 and sx ¼ sz ¼ 0 in these
determined using the following criterion (Gedney,
regions (see Eq. (4)). In the PML boundary regions,
1998):
however, these parameters are given different
values. Setting sx and sz greater than zero (which mþ1
skmax ¼ pffiffiffiffi , (17)
makes the stretching variables complex) allows 150p
r Dk
propagating waves to be absorbed. Making kx and
kz greater than one, on the other hand, allows the where
r is the relative dielectric permittivity
PML regions to absorb evanescent waves. Setting ax (i.e., the permittivity normalized to its value in free
and az greater than zero may also improve the space) in the interior of the grid nearest to the PML
absorption of evanescent modes (Kuzuoglu and boundary. In other words, in a heterogeneous
Mittra, 1996; Roden and Gedney, 2000), although medium, skmax may vary depending on the permit-
by default these parameters are set to zero in our tivity values bounding the interior of the modeling
codes. domain. The natural boundaries in the grid interior
In theory, because there is no change in electro- are extended into the PML region by padding with
magnetic impedance associated with a change in sx PML cells before running a simulation (Chen and
and sz , values for kx , kz , sx , and sz should be set as Chew, 1997).
high as possible in the PML regions to achieve the
most complete absorption of propagating and 3. MATLAB implementation
evanescent waves. In practice, however, in the
discrete FDTD space, numerical reflections occur The 2-D, TM-mode, finite-difference formulation
when electrical properties change too much between presented above has been implemented in the
nodes. Consequently, the PML parameters must be MATLAB environment in the code TM_mo-
set to gradually increase from their values in the del2d.m, for reflection GPR modeling. The
interior of the grid to some maximum value at the MATLAB code for TE-mode modeling, suitable
grid edges. For kx and kz , we have for crosshole and VRP surveys, is TE_model2d.m.
8 To perform the FDTD simulations in MATLAB,
<1  
> in grid interior; we use matrices to store all field components and
kk ¼ d m electrical properties. The property matrices are
: 1 þ d ðkkmax  1Þ inside PML region;
>
double the size of the field matrices because
, m,
and s are required at every electric and magnetic
(15)
field component location (see Fig. 1). As inputs,
where d is the distance into the PML region from TM_model2d.m and TE_model2d.m require the
the interior/PML boundary, d is the thickness of the electrical property matrices, all source and receiver
PML region, m is known as the PML exponent, and locations, a time vector and corresponding source
 ARTICLE IN PRESS
J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258 1253

pulse sampled with the appropriate Dt, and the maximum time step that ensures numerical stability
number of PML boundary cells to use. All finite- given the model’s electrical properties, Dx, and Dz.
differences in the FDTD update equations are If necessary, the electrical property matrices are
computed using matrix notation instead of loops interpolated to half the spatial discretization inter-
to optimize the codes for speed in MATLAB. In val, since they are double the size of the field
addition, the FDTD and PML update coefficients matrices, using the code gridinterp.m. As a final
are calculated and stored in matrices the same size step before running a simulation, the property
as the electrical property matrices before starting a matrices are then padded around the edges with
simulation, so that the number of calculations the proper number of PML absorbing boundary
required during each iteration is reduced. Finally, cells using the program padgrid.m. The values of
the C terms in Eqs. (10a)–(10c) are computed and s,
, and m in the PML regions are set to be simply
added to the electric and magnetic fields only in the an extension of those in the interior of the grid.
PML regions, as in the interior of the modeling grid, In the code TM_model2d.m, an outer loop runs
these terms become zero. over the number of sources and an inner loop over
For the source pulse in our codes, we use the the number of FDTD iterations required for each
normalized first derivative of a Blackman–Harris source. In each FDTD iteration, the H x , H z , and E y
window function (Harris, 1978), which is described field matrices are first updated, in that order, using
in Chen and Chew (1997) for geophysical FDTD Eqs. (10)–(13). The appropriate time sample of the
modeling. This pulse is created using the program source pulse function is then fed into the E y field
blackharrispulse.m. Fig. 2 shows Blackman– component at the current source location, and the
Harris pulses having dominant frequencies of 50, E y wavefield is recorded at all of the receiver
100, and 200 MHz. When fed into the E y field locations. The output from our codes is a data cube
component at the desired source location in our containing a series of common-source gathers
TM-mode code, the resulting pulse that travels (i.e., we record multi-offset data). Extraction of
through the grid and is recorded at the receiver common-offset data, which is typically acquired
locations roughly resembles a Ricker wavelet. during reflection GPR profiling, from the multi-
Fig. 3 is a flowchart showing the sequence of steps offset data cube is a trivial matter in MATLAB.
involved in simulating a reflection GPR survey with
our TM-mode modeling code. First, the maximum 4. Examples
spatial discretization intervals that can be used, in
order to control numerical dispersion, are deter- We now show examples of reflection and cross-
mined using finddx.m given the electrical proper- hole GPR modeling using the codes TM_mo-
ties in the model and the source pulse as inputs. del2d.m and TE_model2d.m, respectively. The
Next, the code finddt.m is used to determine the programs are called with the example run scripts
TM_run_example.m and TE_run_example.m.
1 Fig. 4 shows the electrical property model used for
50 MHz the TM-mode, reflection GPR example. The subsur-
100 MHz face consists of two layers separated by a dipping
0.5 200 MHz boundary. The upper layer, representative of vadose
zone sediment, has
r ¼ 9 and s ¼ 1 mS=m. The
Amplitude

lower layer, representative of material in the

0 saturated zone, has
r ¼ 25 and s ¼ 5 mS=m. With-
in the upper layer there are three anomalous blocks
of different sizes having
r ¼ 16 and s ¼ 1 mS=m.
-0.5 For all materials, m was set equal to its free space
value, m0 . An air–earth interface is included in the
model at z ¼ 0; this was accomplished by simply
-1 adding a thin upper layer with
r ¼ 1 and s ¼ 0 to
0 5 10 15 20 25 the grid. Sources and receivers were located along
Time (ns) the air–earth interface every 0.2 m for the reflection
Fig. 2. Examples of Blackman–Harris pulses fed into grid as survey. The Blackman–Harris source pulse had a
sources in our FDTD modeling codes. dominant frequency of 100 MHz. For this pulse and
 ARTICLE IN PRESS
1254 J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258

determine maximum ∆x,∆z (finddx.m)

determine maximum ∆t (finddt.m)

interpolate ε, µ, σ grids (gridinterp.m)

pad grids for PML (padgrid.m)

begin FDTD simulation for new source

update Hx, Hz, and Ey (in that order)

add source pulse to Ey at desired location

TM_model2d.m
store Ey field component at receiver locations

NO
iterations complete?

YES
NO
done all sources?

YES
DONE

Fig. 3. Flowchart of FDTD modeling procedure.

εr= 1, σ = 0 mS/m the source located at x ¼ 10 m. Because of the PML

0
absorbing boundaries implemented in our code, no
2 εr= 9, σ = 1 mS/m
reflections can be seen coming from the edges of the
Depth (m)

4 εr= 16, σ = 1 mS/m modeling domain in any of the panels. At t ¼ 30 ns,

we capture the wavefield as it is spreading outwards
6 from the source before it has encountered any
8 εr= 25, σ = 5 mS/m heterogeneities within the earth. Head waves link
the energy traveling more rapidly through the air
10
0 5 10 15 20 with that traveling through the ground. At
Position (m) t ¼ 50 ns, the wavefield has clearly encountered the
large anomalous block in the middle of the upper
Fig. 4. Electrical property model used for TM-mode reflection layer, and has been partly reflected back towards the
GPR example.
surface. The direct wave traveling through the air
from the source location has also left the boundaries
the electrical properties in the model, finddx.m of the modeling domain at this point. At t ¼ 70 ns,
yielded a maximum possible spatial discretization of the energy reflected from the block in the upper
0.0423 m. We used Dx ¼ Dz ¼ 0:04 m. The max- layer has reached the air–earth interface. At
imum possible time step, determined using t ¼ 90 ns, we see the wavefield being reflected from
finddt.m, was 0.0801 ns. We used Dt ¼ 0:08 ns. the dipping boundary between the top and bottom
Fig. 5 shows snapshots of the E y field component layers. Fig. 5 clearly shows that, even for the
at various times during the FDTD simulation for relatively simple earth model shown in Fig. 4, the
 ARTICLE IN PRESS
J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258 1255

Fig. 5. Snapshots showing amplitude of E y wavefield at different times during TM-mode FDTD modeling. Source is located at x ¼ 10 m,
z ¼ 0 m.

propagating wavefield becomes quite complicated

very quickly.
Fig. 6 shows one of the common-source gathers
(consisting of the recorded E y field component in
time at all of the receiver locations) for the source
located at x ¼ 10 m in Fig. 4. This is one slice of the
output multi-offset data cube. The linear events in
the image are the direct air and ground arrivals. The
events with hyperbolic moveout are reflections from
the top and bottom of the block anomaly in the
middle of the upper layer, and the dipping boundary
between the upper and lower layers. Fig. 7, on the
other hand, shows the common-offset reflection
GPR data that were extracted from the data cube. Fig. 6. Common-source gather for source located at x ¼ 10 m,
z ¼ 0 m in Fig. 4.
The source–receiver offset for this image is 1 m. In
the image, diffractions are clearly seen originating
from each corner of the block anomalies. Also, the and ground waves are merged together and present
dipping boundary between the upper and lower at the top of Fig. 7.
layers is strongly visible, although it is not a In our second example, we demonstrate FDTD
perfectly straight interface in the GPR section modeling of a crosshole GPR survey using TE_mo-
because of time shifts caused by the velocity del2d.m. Fig. 8 shows the subsurface EM-wave
anomalies in the upper layer. Finally, the direct air velocity of the survey plane, which was obtained
 ARTICLE IN PRESS
1256 J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258

Fig. 7. Common-offset gather for TM-mode reflection GPR

modeling example. Fig. 8. Subsurface velocity model for TE-mode crosshole GPR
example. Source and receiver locations are marked by an ‘
’.

from the distribution of relative dielectric permit- graphy code. Despite this, low-amplitude reflections
tivity using the following low-loss approximation: are clearly seen in Fig. 9, which arise from the
c velocity heterogeneities.
v ¼ pffiffiffiffi , (18) Fig. 10 shows one of the common-source gathers

r
that was obtained in this example for the E z source
where c is the velocity of EM waves in free space located at z ¼ 5 m in Fig. 8. The first-arrival times in
(0.3 m/ns). The permittivity model was constructed each common-source gather (i.e., the arrival times
using the GSLIB geostatistical software package for each source–receiver configuration in the survey)
(Deutsch and Journel, 1992). Relative permittivity were picked automatically from the data based on
values were set to vary between 20 and 32, a range the first point in each trace where the amplitude
which is typical of materials in the saturated zone. A exceeded 1% of the trace maximum. These noise-
constant conductivity of 5 mS/m was assumed, and free travel-time data were then inverted using a
m was again set equal to its value in free space. straight-ray, least-squares, tomography algorithm
Sources were located every 0.25 m from 0.5 to with a small amount of second derivative smooth-
10.5 m depth down a borehole located at x ¼ 0:5 m. ness regularization added for stability. Fig. 11
Receivers were located at the same depths in a shows the resulting velocity tomogram. Compared
borehole located at x ¼ 5:5 m. Again, a Black- with Fig. 8, we see that the inversion of the noise-
man–Harris pulse with a dominant frequency of free travel-time data, as expected, produces a very
100 MHz was used as a source function. For this good image of the subsurface velocity field. There is
example, finddx.m yielded a maximum spatial a small amount of smearing in the image along some
discretization interval of 0.0374 m. We used of the high-angle raypath directions. Also, the
Dx ¼ Dz ¼ 0:025 m. The maximum possible time resolution of the inversion result is slightly worse
step determined using finddt.m was 0.0226 ns. We than that of the true velocity model, due to the
used Dt ¼ 0:02 ns. The air–earth interface was not resolution limits of ray-based tomography, the
modeled in this example to facilitate the automatic aperture-limited nature of the survey, and errors
picking of travel times through the earth in the resulting from our straight ray approximation.
resulting data.
Fig. 9 shows shapshots of the E z field component 5. Conclusions
at various times during the crosshole modeling
when the E z source was located at z ¼ 2 m. The The codes presented here provide a relatively
propagating wavefront in this case is much more easy-to-understand, flexible package for 2-D GPR
circular than that shown in Fig. 5 because the modeling in MATLAB. Although the use of
magnitudes of the velocity heterogeneities in this MATLAB means that our codes are slightly slower
example are quite small. Small variations in velocity than similar compiled C or Fortran routines, the
were purposely used in this example so that the MATLAB environment offers significant benefits
resulting travel time data could be inverted reason- over these other programming languages such as
ably accurately using a simple, straight-ray tomo- increased code readability and easier plotting and
 ARTICLE IN PRESS
J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258 1257

Fig. 9. Snapshots showing amplitude of E z wavefield at different times during TE-mode FDTD modeling. Source is located at x ¼ 0:5 m,
z ¼ 2 m.

Fig. 10. Common-source gather for source located at x ¼ 0:5 m, Fig. 11. Straight-ray velocity tomogram obtained from synthetic
z ¼ 5 m in Fig. 8. crosshole data created using our TE-mode code.

the run scripts that we have provided to be modified

manipulation of data. Our codes are well suited to so that only the subsurface region contributing to
be modified for more complex modeling if desired, the traces being recorded is considered. This could
an example being to alter them to allow for be done without altering the main code, TM_mo-
dispersion in electrical properties. The possibility del2d.m, and could significantly decrease the
also exists, for very long reflection GPR surveys, for modeling time for very long survey lines.
 ARTICLE IN PRESS
1258 J. Irving, R. Knight / Computers & Geosciences 32 (2006) 1247–1258

Acknowledgements Gedney, S., 1998. The perfectly matched layer absorbing

medium. In: Taflove, A. (Ed.), Advances in Computational
This research was supported by funding to R. Electrodynamics: The Finite-Difference Time-Domain Meth-
od. Artech House, Norwood, MA, pp. 263–343.
Knight from the National Science Foundation, Georgakopoulos, S.V., Birtcher, C.R., Balanis, C.A., Renaut,
Grant Number EAR-0229896-002. J. Irving was R.A., 2002. Higher-order finite-difference schemes for elec-
also supported during this work through a Depart- tromagnetic radiation, scattering, and penetration, Part 1:
mental Chair’s Fellowship at Stanford University. theory. IEEE Antennas and Propagation Magazine 44,
134–142.
Goodman, D., 1994. Ground-penetrating radar simulation in
Appendix A. Supplementary data engineering and archeology. Geophysics 59, 224–232.
Harris, F.J., 1978. On the use of windows for harmonic analysis
Supplementary data associated with this article with the discrete Fourier transform. Proceedings of the IEEE
can be found in the online version, at 10.1016/ 66, 51–83.
Holliger, K., Bergman, T., 2002. Numerical modeling of borehole
j.cageo.2005.11.006.
georadar data. Geophysics 67, 1249–1257.
Kuzuoglu, M., Mittra, R., 1996. Frequency dependence of the
constitutive parameters of causal perfectly matched aniso-
tropic absorbers. IEEE Microwave and Guided Wave Letters
References 6, 447–449.
Luebbers, R.J., Hunsberger, F., 1992. FDTD for Nth-order
Berenger, J.P., 1994. A perfectly matched layer for the absorption dispersive media. IEEE Transactions on Antennas and
of electromagnetic waves. Journal of Computational Physics Propagation 40, 1297–1301.
114, 185–200. Lui, Q.H., Fan, G., 1999. Simulations of GPR in dispersive
Bergmann, T., Robertsson, J.O.A., Holliger, K., 1996. Numerical media using a frequency-dependent PSTD algorithm. IEEE
properties of staggered finite-difference solutions of Max- Transactions on Geoscience and Remote Sensing 37,
well’s equations for ground-penetrating radar modeling. 2317–2324.
Geophysical Research Letters 23, 45–48. Moghaddam, M., Yannakakis, E.Y., Chew, W.C., 1991. Model-
Bergmann, T., Robertsson, J.O.A., Holliger, K., 1998. Finite- ing of subsurface interface radar. Journal of Electromagnetic
difference modeling of electromagnetic wave propagation in Waves and Applications 5, 17–39.
dispersive and attenuating media. Geophysics 63, 856–867. Powers, M.H., Olhoeft, G.R., 1994. Modeling dispersive ground-
Bourgeois, J.M., Smith, G.S., 1996. A fully three-dimensional penetrating radar data. Proceedings of the 5th International
simulation of a ground-penetrating radar: FDTD theory Conference on Ground-Penetrating Radar, Waterloo, Ontar-
compared with experiment. IEEE Transactions on Geoscience io, pp. 173–183.
and Remote Sensing 34, 36–44. Roden, J., Gedney, S., 2000. Convolution PML (CPML): an
Cai, J., McMechan, G.A., 1995. Ray-based synthesis of bistatic efficient FDTD implementation of the CFS-PML for
ground-penetrating radar profiles. Geophysics 60, 87–96. arbitrary media. IEEE Microwave and Optical Technology
Carcione, J.M., 1996. Ground-penetrating radar: wave theory Letters 27, 334–339.
and numerical simulation in lossy anisotropic media. Geo- Taflove, A., 1995. Computational Electrodynamics: The Finite-
physics 61, 1664–1677. Difference Time-Domain Method. Artech House, Norwood,
Casper, D.A., Kung, K.S., 1996. Simulation of ground-penetrat- MA, 599pp.
ing radar waves in a 2-D soil model. Geophysics 61, Teixeira, F.L., Chew, W.C., Straka, M., Oristaglio, M.L., Wang,
1034–1049. T., 1998. Finite-difference time-domain simulation of ground-
Chen, Y.H., Chew, W.C., 1997. Application of perfectly matched penetrating radar on dispersive, inhomogeneous, and con-
layers to the transient modeling of subsurface EM problems. ductive soils. IEEE Transactions on Geoscience and Remote
Geophysics 62, 1730–1736. Sensing 36, 1928–1936.
Chew, W.C., Weedon, W.H., 1994. A 3-D perfectly matched Wang, T., Tripp, A.C., 1996. FDTD simulation of EM wave
medium from modified Maxwell’s equations with stretched propagation in 3-D media. Geophysics 61, 110–120.
coordinates. IEEE Microwave and Optical Technology Xu, T., McMechan, G.A., 1997. GPR attenuation and its
Letters 7, 599–604. numerical simulation in 2.5 dimensions. Geophysics 62,
Deutsch, C.V., Journel, A.G., 1992. GSLIB: Geostatistical 403–414.
Software Library and User’s Guide. Oxford University Press, Yee, K., 1966. Numerical solution of initial boundary value
New York, NY, 340pp. problems involving Maxwell’s equations in isotropic media.
Ellefsen, K.J., 1999. Effects of layered sediments on the guided IEEE Transactions on Antennas and Propagation AP-14,
wave in crosswell radar data. Geophysics 64, 1698–1707. 302–307.
Fang, J., Wu, Z., 1996. Generalized perfectly matched layer for Zeng, X., McMechan, G.A., Cai, J., Chen, H.W., 1995.
the absorption of propagating and evanescent waves in Comparison of ray and Fourier methods for modeling
lossless and lossy media. IEEE Transactions on Microwave monostatic ground-penetrating radar profiles. Geophysics
Theory and Techniques 44, 2216–2222. 60, 1727–1734.