Journal of Applied Geophysics 75 (2011) 363–372

Contents lists available at ScienceDirect

Journal of Applied Geophysics

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / j a p p g e o

Unstructured grid based 2-D inversion of VLF data for models including topography
Vikas C. Baranwal a, b,⁎, Antje Franke a, Ralph-Uwe Börner a, Klaus Spitzer a
a
Institute of Geophysics, TU Bergakademie Freiberg, Gustav-Zeuner-Str. 12, D-09596 Freiberg, Germany
b
Dept. of Geology and Geophysics, IIT Kharagpur, West Bengal, Pin-721 302, India

a r t i c l e i n f o a b s t r a c t

Article history: We present a 2-D inversion code incorporating a damped least-squares and a minimum-model approach for
Received 24 June 2009 plane wave electromagnetic (EM) methods using an adaptive unstructured grid finite element forward
Accepted 11 July 2011 operator. Unstructured triangular grids permit efficient discretization of arbitrary 2-D model geometries and,
Available online 23 July 2011
hence, allow for modeling arbitrary topography. The inversion model is parameterized on a coarse parameter
grid which constitutes a subset of the forward modeling grid. The mapping from parameter to forward
Keywords:
Unstructured grids
modeling grid is obtained by adaptive mesh refinement. Sensitivities are determined by solving a modified
Finite elements sensitivity equation system arising from the derivative of the finite element equations with respect to the
Topography model parameters. Firstly, we demonstrate that surface topography may induce significant effects on the EM
VLF response and in the inversion result, and that it cannot be ignored when the scale length of topographic
MT variations is in the order of magnitude of the skin depth. Secondly, the dependency of the inversion on the
Inversion starting model is discussed for VLF and VLF-R data. Thirdly, we demonstrate the inversion of a synthetic data
set obtained from a model with topography. Finally, the inversion approach is applied to field data collected in
a region with undulating topography.
© 2011 Elsevier B.V. All rights reserved.

1. Introduction (Oskooi and Pedersen, 2006; Pedersen and Oskooi, 2004). Sharma and
Kaikkonen (1998) have used a Very Fast Simulated Annealing (VFSA)
The plane wave electromagnetic (EM) methods cover Very Low method for inversion of VLF-R data. Kaikkonen and Sharma (1998) have
Frequency (VLF) EM, Radio Magnetotelluric (RMT) and Magnetotelluric also inverted both VLF and VLF-R data using VFSA. Monteiro Santos et al.
(MT) methods. The VLF method has been successfully used in exploration (2006) have used a 2-D regularized inversion to interpret single frequency
of groundwater, conductive minerals, and ore bodies in various types VLF data. All these inversion approaches have been developed for flat
of host rocks (Arcone, 1978; Paal, 1968; Sharma and Baranwal, 2005), earth models.
in geothermal studies (Baranwal and Sharma, 2006; Zlotnicki et al., Small-scale surface or sea-floor topographic variations have little
2006), in groundwater contamination and waste management studies effect on the responses in plane wave EM methods using low frequencies.
(Monteiro Santos et al., 2006; Poddar and Rathor, 1983) and in However, a large topographic relief can affect the responses even at
archeological problems (Drahor, 2006). frequencies below 10 Hz used in the magnetotelluric method (Chouteau
A quantitative interpretation of VLF or VLF-R data for depth and and Bouchard, 1988; Jiracek et al., 1989; Wannamaker et al., 1986). In
resistivity distribution can be carried out by applying various inversion general, if the mean topographic gradient over a length scale equal to the
schemes. In principle, approaches developed for the inversion of MT skin depth is in the order of unity then there will be significant differences
data can be used to invert VLF and VLF-R data with slight modification in in the responses for flat earth models and for models with topography.
the forward modeling for the computation of VLF responses. Beamish The VLF (5–30 kHz) and RMT (10–240 kHz) methods use rather high
(1994) used a regularized inversion approach well known as OCCAM frequencies therefore a topographic relief with amplitudes in the order of
inversion (deGroot-Hedlin and Constable, 1990) to invert VLF-R data. A 10 m may affect the response depending on the resistivity of the region
further efficient variant of OCCAM inversion called REBOCC inversion and used frequency (Arcone, 1978; Fischer, 1989; Karous, 1979; Liu and
(Siripunvaraporn and Egbert, 2000) is used for inversion of VLF data Becker, 1992). Initially, VLF was thought to work in the range of 3–30 kHz
(Oskooi and Pedersen, 2005) and for the inversion of airborne VLF data (Paal, 1968), in practice, however, it is limited to 15–30 kHz because only
very sporadic transmitters are found in the range of 3–15 kHz.
The finite element (FE) method provides an efficient tool to incorpo-
⁎ Corresponding author at: Geological Survey of Norway (NGU), Postbox 6315, rate arbitrary surface topographic variations in EM simulations. The FE
Sluppen, 7491 Trondheim, Norway. Tel.: + 47 73904447; fax: + 47 73921620.
E-mail addresses: [email protected] (V.C. Baranwal),
method using structured rectangular or triangular grids has been applied
[email protected] (A. Franke), [email protected] to 2-D geo-electromagnetic problems since decades (Coggon, 1971; Fox et
(R.-U. Börner), [email protected] (K. Spitzer). al., 1980; Wannamaker and Stodt, 1987). A more flexible discretization of

0926-9851/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jappgeo.2011.07.011
364 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372

the geological models including surface topography to solve 2-D geo- Ey is the y-component of the electric field and Hy is the y-component
electromagnetic problems can be obtained applying unstructured of the magnetic field. y denotes the strike direction. ω, μ, i, and σ are
triangular grids. Recently, 2-D FE algorithms for plane wave geo- angular frequency, magnetic permeability, imaginary unit, and
electromagnetic modeling using adaptive unstructured grids have been electrical conductivity, respectively. To solve for the unknown fields,
presented by Franke et al. (2004, 2007) and Key and Weiss (2006). The inhomogeneous Dirichlet boundary conditions are applied that assign
application of adaptive unstructured grids has recently been reported by the field values of a horizontally layered half-space to the outer
Guenther et al. (2006) and Ruecker et al. (2006) for 3-D resistivity forward boundaries. The finite element discretization leads to a system of
modeling and inversion, and also by Börner et al. (2008) for 3-D transient linear equations that can be expressed in matrix–vector form as
EM simulations.
Generally, there are two common approaches for inversion of data ðK + MÞu = f ð3Þ
for models including topography. The first approach considers the effect
of topography as field distortion and effort has been made to correct the where u is either a column vector of the discrete electric field Ey or the
data for topography (Baker and Myers, 1980; Chouteau and Bouchard, magnetic field Hy at each node in E- and H-polarization, respectively, and f
1988; Jiracek et al., 1989) thus allowing any inversion code to be applied represents a source term arising from Dirichlet type boundary conditions.
on the corrected data to obtain a flat earth model. Approaches K and M are referred to as stiffness and mass matrices. The remaining field
developed to correct the observed data for topography involve components Hx, Hz for E-polarization and Ex, Ez for H-polarization can be
approximations which may not be valid under all conditions. In the determined at each grid node by numerical differentiation as follows
second approach, topography is incorporated explicitly during the
1 ∂Ey 1 ∂Ey
inversion (Baba and Chave, 2005; Key et al., 2006; Li et al., 2008). The Hx = and Hz = − ; ð4Þ
development of a forward modeling code which can model topographic iωμ ∂z iωμ ∂x
variations (Franke et al., 2007) gives us opportunity to invert the data
directly for models including topography. There will be no need of either 1 ∂Hy 1 ∂Hy
Ex = − and Ez = : ð5Þ
correcting the data for topographic variations or including the σ ∂z σ ∂x
topography explicitly during inversion. Therefore, the straightforward
damped least-squares and minimum-model inversion techniques are The apparent resistivity ρa and phase ϕ for E- and H-polarizations
applied to develop an inversion code using the forward modeling (in the case of VLF-R and MT methods) can be computed as
algorithm of Franke et al. (2004, 2007) in the present work. Among 0  1
  Im Ey = Hx
various Gauss–Newton methods, the damped least-squares method is a 1  Ey 2 −1 @
ρa = ; ϕ = tan   A; ð6Þ
well established inversion technique which regularizes the solution by ωμ Hx  Re Ey = Hx
putting a bound on the maximum change in the model parameter
update whereas the minimum-model inversion approach minimizes
for E-polarization and
the difference between the model and a reference model and therefore
the final model will be independent of the starting model (Aster et al.,  2 0  1
1  Ex  −1 @
Im Ex = Hy
2005; Farquharson and Oldenburg, 2004; Meju, 1994; Menke, 1989). ρa =   ; ϕ = tan   A; ð7Þ
The paper is organized as follows: First, we demonstrate the ωμ Hy  Re Ex = Hy
topography effects by forward modeling results. Second, we present
results applying our 2-D inversion code on synthetic and field data. The for H-polarization. Inphase and quadrature anomaly calculated from the
region of interest is parameterized by a regular grid type containing magnetic transfer function Hz/Hx are used as data in the VLF method and
rectangular and trapezoidal cells. Logarithmic conductivities are used as are usually represented in percent as
model parameters in the inversion process. The parameter grids are
further partitioned into fine unstructured grids using adaptive refine- Inphase anomaly = 100  ReðHz = Hx Þ and ð8Þ
ment to calculate the forward response with the desired accuracy.
Q uadrature anomaly = 100  ImðHz = Hx Þ:
Therefore, application of different parameterization and discretization
schemes for inversion and forward modeling, respectively, leads to a
two-grid technique. Inversion results obtained by inversion of VLF and
3. Inversion procedure
VLF-R data using various starting models are shown, and the dependency
of inversion results on the starting model is discussed. All the forward
3.1. Damped least-squares and minimum-model approach
modeling and inversion codes are developed using MATLAB®.
We apply a damped least-squares and a minimum-model approach
2. Forward modeling
(Farquharson and Oldenburg, 2004; Loke and Barker, 1996; Sasaki,
1989) for the minimization of the objective function Φ given by
The forward computations for the inversion code are carried out using
a finite element code which incorporates adaptive unstructured triangular Φ = Φd ðmÞ + λΦm ðmÞ: ð9Þ
grids (Franke et al., 2004, 2007). In the case of plane-wave, diffusive, time-
harmonic electromagnetic fields in 2-D conductivity structures Maxwell's Φd and Φm are measures of the data misfit and the model, respectively,
equations yield two decoupled equations of induction reading for E- and and λ is the regularization parameter to balance their effects. We can
H-polarizations, respectively, in a right-handed Cartesian coordinate express them as
system with the positive z-axis pointing upwards as follows
2
Φd ðmÞ = ‖Wd ðΔd−SΔmÞ‖ ; ð10Þ
2 2
∂ Ey ∂ Ey
+ −iωμσ Ey = 0; ð1Þ and
∂x2 ∂z2
2
Φm ðmÞ = ‖M‖ ; ð11Þ
! !
∂ 1 ∂Hy ∂ 1 ∂Hy
− + −iωμHy = 0; ð2Þ where M will be regarded as the model parameter update
∂x σ ∂x ∂z σ ∂z
Δm = m n − m n − 1 for the damped least-squares and as the difference
 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372 365

between the model m itself and a reference model mref, i.e. mn − mref for where εi and N denote the standard deviation of the ith data and the
the minimum-model approach. n denotes the iteration number. We total number of data, respectively. We stop the iteration process if one
consider the logarithm of the conductivities as model parameters. of the following criteria is met: (1) when the maximum number of
Δd = dobs − d comp describes the discrepancy between the observed data iterations is reached, (2) when the convergence stalls, and (3) when
dobsand the computed data d comp and Wd is a diagonal matrix containing the error approaches the desired target value (e.g. χ 2 ≈ 1).
the standard deviation of the noise in the data. S denotes the sensitivity
matrix. To find the minimum of the objective function Φ, its partial 3.2. Sensitivity calculation
derivatives with respect to the elements of Δm are required to be zero.
Therefore, the resulting normal equations to solve for Δm read The elements of the sensitivity matrix S for the ith observation site
and jth model parameter is calculated using the modified sensitivity
 
T T T T
S Wd Wd S + λI Δm = S Wd Wd Δd; ð12Þ equation method presented by Rodi (1976) which requires (N+ 1)
forward computations for each frequency. We assume two column
vectors ai and bi such that they give the electric and the magnetic field
and components in the case of E-polarization for the ith datum from linear
    combinations of the solution vector u as
T T T T ref
S Wd Wd S + λI Δm = S Wd Wd Δd−λ m−m ; ð13Þ
T T
Ey;i = ai u and Hx;i = bi u; ð16Þ
for the damped least-squares and the minimum-model solution,
respectively. I is the identity matrix. Eqs. (12) and (13) are solved
and vice-versa in the case of H-polarization. In its simplest form the
applying a direct solver at each stage of the iterative inversion process
vector ai is formed by keeping 1 at the position of the ith datum and
and model parameters are updated at each iteration.
zeros at all other nodes. The vector bi is designed in such a way that it
A large value of the regularization parameter λ in early iterations
performs a numerical differentiation over u according to Eqs. (4) and
prevents introduction of undesired structures leading to a smooth
(5). If the observation site is not located exactly at a given grid node
model, however, a small value of λ in the final iteration allows for some
then the desired field values can be obtained by linear interpolation of
structures in the model which are necessary to fit the observed data
adjacent grid nodes expressed by a modification of the entries of ai
well. Starting from a large λ is a good alternative to the line search
and bi.
procedures which produce models with less structure. Therefore, it is
We note from Farquharson and Oldenburg (1996) that sensitivities
reasonable to decrease λ at each iteration starting with a large value
for apparent resistivity and phase in case of VLF-R and MT data can be
(Loke and Barker, 1996). Farquharson and Oldenburg (2004) have
compared various selection methods and have suggested a cooling type
schedule of decreasing λ at each iteration.
We observe in this study that the parameter λ selected by the L-curve
criterion method and, alternatively, by the maximum singular value of
the sensitivity matrix is in the same order of magnitude in the case of
inversion of synthetic data contaminated with Gaussian distributed
random noise of 0.05 standard deviation. However, in the case of
inversion of synthetic data with more random noise, the λ selected by the
L-curve criterion is small in comparison to the maximum singular value
of S which may introduce structures from the very beginning of the
inversion process. Therefore, the inversion departs from a value of λ
equal to the maximum singular value of the sensitivity matrix and will be
reduced by a factor of less than 1, e.g. 0.7 or even 0.1 using the least-
squares or the minimum-model approach, respectively, in each iteration
to reach fast convergence (Loke and Barker, 1996; Oristaglio and
Worthington, 1980).
The χ2 measure and root mean square (RMS) error can be calculated
by

1 N Δd2
2
χ = ∑i = 1 2i ; ð14Þ
N εi

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 N
RMS = ∑i = 1 Δd2i ; ð15Þ
N

Fig. 2. Three models used to generate synthetic responses for demonstrating the effect
Fig. 1. Parameterizing the model in rectangles and trapezoids (thick lines). The of topography: (a) with topography and conductive regions (b) only with topography
unstructured initial forward modeling grid is indicated by thin lines. and (c) only with conductive regions within a flat earth.
366 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372

calculated by derivation of electric and magnetic fields with respect to Using Eq. (16) and the solution vector u from Eq. (3), the expression
model parameters (ln σj) as of SijMT will result in
( )T !T
i   1 ∂ai 1 ∂bi ai bi −1 ∂ðK + MÞ
∂ ln ρa MT
− − ðK + MÞ
Sij ρa Sij = u− u:
ln MT
= = 2Re Sij aTi u ∂ ln σj bTi u ∂ ln σj aTi u bTi u ∂ ln σj
∂ ln σj
ð18Þ
and

∂ϕi   VLF-R surveys are generally performed in E-polarization, which is

Sϕij =
MT
= Im Sij ð17Þ the case when the profile is perpendicular to the strike direction of the
∂ ln σj
2-D conductivity structure. The first term ({…}) in Eq. (18) will vanish
where in the case of E-polarization because vectors ai and bi are independent
! of the conductivity (refer Eq. (4)).
MT 1 ∂Ei 1 ∂Hi An analogous strategy is used to calculate sensitivities for the
Sij = − :
Ei ∂ ln σj Hi ∂ ln σj inphase and quadrature anomaly in the case of VLF that corresponds

Fig. 3. Synthetic VLF-R and VLF responses for three models shown in Fig. 2.
 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372 367

to E-polarization case again. In this case, the vectors ai and bi will be

designed in such a way that they perform numerical differentiation
over u to get the magnetic field components Hz and Hx according to
Eqs. (4) and (5). Therefore SijVLF (for VLF data) is calculated by
!T
VLF 1 aT u −1 ∂ðK + MÞ
Sij = ai − iT bi ðK + MÞ u: ð19Þ
T
bi u bi u ∂ ln σj

The sensitivities for inphase and quadrature anomaly can be

calculated using real and imaginary parts of SijVLF as
   
RE VLF IM VLF
Sij = 100  Re Sij ; Sij = 100  Im Sij : ð20Þ

The reader is referred to Baranwal (2007) and Farquharson and

Oldenburg (1996) for details of these derivations.

4. Parameterization of a model including surface topography

In the following, we discuss parameterizing a model whose surface is

associated with a varying topography. We may perform the parame- Fig. 5. Damped least-squares inversion results of VLF-R data generated from a model
terization by partitioning either in rectangles and trapezoids that form a including topography and conductive regions (Fig. 3a): (a) using a flat earth model,
rather regular type of grid, or in unstructured triangular cells that closely (b) using a model including topography.
correspond to the forward modeling grid. However, we only show the
rectangles and trapezoids case in Fig. 1.
The grid forms a set of parameters which represent the spatial In a second step, we take data of a homogeneous earth model with
distribution of electrical conductivity. The rectangular or trapezoidal surface topography and perform a flat earth inversion to point out
grid comprises non-conforming, locally refined parameter cells which topography induced artifacts. Finally, we show inversion results using
may further enhance the flexibility with respect to resolution. It is flat earth and topographic model for synthetic data corresponding to
partitioned into an unstructured triangular grid for forward modeling, model having topography.
and in a subsequent step, may be adaptively refined on the basis of an a-
posteriori error estimation obtained from the current solution on a given 5.1. Decomposing the response from surface topography and subsurface
FE grid. The parameter grid is indicated by thick lines and the first conductivity structures
refinement stage of the unstructured triangular forward modeling grid
is indicated by thin lines (Fig. 1). Note, that the latter has been further The synthetic model displayed in Fig. 2a consists of two anomalous
refined using adaptive refinement to reach the desired numerical regions having resistivities of 100 Ω·m and 20 Ω·m, respectively,
accuracy of the solution. within a 1000 Ω·m half-space with a smooth, but pronounced
topography. The observation sites are located at 50 m intervals from
5. The topography effect − 575 m to 575 m and marked by arrows. Fig. 2b presents a
homogeneous half-space with topography removing the perturbing
We now investigate the influence of the topography on the VLF-R bodies and Fig. 2c presents the same anomalous regions as depicted in
and VLF response, and demonstrate the effect of topography in the Fig. 2a but now embedded within a flat earth.
inverse process. For this purpose, we disassemble a model in a first step For all these models, synthetic data are generated for three
to separately examine its response originating from the conductive frequencies: 5, 16 and 25 kHz. At these frequencies, the skin depths
subsurface structures and from topographic undulations. Since these for a 1000 Ω·m half-space are ≈225, 125 and 100 m, respectively. In
features are inductively coupled, both superposed responses are Fig. 3a, the total synthetic response of the model combining topography
certainly not giving the total response. However, this study attempts with the conducting bodies (Fig. 2a) is displayed in terms of apparent
to instructively display the order of magnitude of the associated effects. resistivity and phase according to Eq. (6) and inphase and quadrature

Fig. 4. Flat earth damped least-squares inversion results of VLF-R data generated from a model including topography. For reasons of comparison, the original topography is plotted at
the top. Resistive anomalies appear below the hills whereas the anomalies associated with the valleys are conductive.
368 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372

Fig. 8. Inverted model (dipping bodies) obtained by damped least-squares inversion of

Fig. 6. Inverted model (dipping bodies) obtained by damped least-squares inversion of
VLF-R data starting with a 2000 Ω·m half-space.
VLF data starting with a 2000 Ω·m half-space.

inversion process. For this purpose, synthetic VLF-R data shown in

anomaly according to Eq. (8). In Fig. 3b, the perturbing conductive Fig. 3a is considered. The data is generated for a model with topography
bodies are removed so that a homogeneous model incorporating and two anomalous regions of 100 and 20 Ωm embedded in a 1000 Ωm
topographic variations remains. The response clearly shows the half-space. The data is first inverted using a flat earth model assumption
influence of the topography. In Fig. 3c, the topography undulations are and second using a model including topography. In both cases, the
replaced by an average flat earth level so that the remaining lateral starting model is a 2000 Ωm resistive half-space. The inversion result
variation in the response is only due to the perturbing bodies. Note, that obtained after 9 iterations with RMS error 0.025 for the flat earth model
the order of magnitude of both effects is comparable. and the inversion result obtained after 11 iterations with RMS error
0.021 for the model with topography are shown in Fig. 5a and b,
5.2. Flat earth inversion of VLF-R data obtained from models with topography respectively.
It is clear from Fig. 5a that the depth, especially the bottom of the
In this section, we investigate how data from a model with topography conductive bodies is not well resolved and the position of the region
influence the results of our damped least-squares inversion algorithm if with the lower resistivity is shifted to the left of its actual position due
the topography is not taken into account. For this purpose, we consider to the coupling of the response between the conductive region and
the synthetic VLF-R data set shown in Fig. 3b which is generated for a topographic valley. Actual positions of the anomalous regions are
homogeneous 1000 Ω·m model with topography. We invert this data set marked by dashed lines in Fig. 5a and b. When topographic variations
using a flat earth assumption. The starting model is a 2000 Ω·m half- are considered in the model then the inversion resolves the position
space. The inversion result obtained after 11 iterations with RMS error and the depth of both of the conductive bodies (Fig. 5b). These
0.017 is shown in Fig. 4. There are clear artifacts associated with the examples clearly demonstrate that the topography effect may become
topography undulations. Conductive structures having resistivities of significant when the scale length of topographic variations (≈100 m)
about 500 Ω·m appear below the central valley and the transitions from is in the order of magnitude of the skin depth (100–200 m). It is
the hills to the planes on the left- and right-hand side whereas resistive therefore necessary to take into account any arbitrary topography for
anomalies around 2500 Ω·m show up beneath the hills. Knowing the true simulation and inversion. Approximate data correction schemes to
resistivity, the maximum deviation of the inverted resistivities is more remove the topography effect then become needless.
than a factor of 2 in both directions of the resistivity scale.

5.3. Inversion of topography affected VLF-R data with anomalous regions: 6. Inversion of noisy synthetic data for models including topography
using a flat earth model and a model with topography
We have obtained reasonable inversion results of the synthetic
The damped least-squares inversion results of synthetic data from a data shown in Fig. 3a after applying Gaussian distributed noise with a
model including topographic variation and anomalous conductivity standard deviation of 0.05 to 0.1. Because the VLF data resolves
regions are presented here. It is shown that the inverted models may vertical dyke type structures very well, the performance of recovering
differ significantly if topography is not accounted for during the dipping conductive bodies from VLF data will be examined. To

Fig. 7. Fitting between noise added synthetic data and modeled data for VLF data damped least-squares inversion starting at 2000 Ω·m half-space. Symbols and solid line represent
synthetic and computed data, respectively.
 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372 369

Fig. 9. Fitting between noise added synthetic data and modeled data for VLF-R data damped least-squares inversion starting at 2000 Ω·m half-space. Symbols and solid line represent
synthetic and computed data, respectively.

investigate the dependency of the damped least-squares inversion of (χ 2 ≈ 1) is met after 16 iterations for inversion of VLF-R data. Fitting
VLF and VLF-R data on the choice of the starting models we compute between synthetic and computed data is shown in Fig. 9. The VLF-R
synthetic data generated for a model containing two bodies of 100 and data recovers the true model as well as the background resistivity
20 Ω·m dipping in different directions at three frequencies 5, 16 and significantly better than VLF data and it is less dependent on the
25 kHz. Gaussian distributed noise with a standard deviation of 0.05 starting model even in the case of dipping bodies. The VLF data
has been applied to apparent resistivity, inphase and quadrature contains information of magnetic fields only and responds well for
anomaly and an equivalent of approximately 1.45° to the phase prior lateral contrasts in resistivity. Nevertheless, the inversion of VLF data
to inversion. We invert VLF and VLF-R data separately starting with is also able to resolve anomalous regions and its dipping nature when
homogeneous 500, 2000 and 5000 Ω·m half-space models. However, started with a half-space model close to the true host resistivity.
for clarity we exclusively show the resulting models obtained by The models depicted in Figs. 6 and 8 exhibit strong undesired
inversion of VLF and VLF-R data starting with the homogeneous oscillations of the background resistivity especially in the case of VLF
model of 2000 Ω·m in Figs. 6 and 8 respectively. The original dipping data inversion (Fig. 6) applying the damped least-squares approach
anomalous regions are indicated by dashed lines. The value of χ 2 (Eq. (12)). The minimum-model solution described in Eq. (13) yields
converges to ≈3 in the damped least-squares inversion of VLF data a smoother background, however, does not recover the true model
starting with various homogeneous models before it stagnates. We resistivities sufficiently as shown in Figs. 10 and 12 for the noise added
observe that more resistive starting models recover the true model synthetic VLF and VLF-R data, respectively. We have used a 2000 Ω·m
better than less resistive starting models due to the larger skin depth half-space as starting and reference model and both inversions are
of the EM fields. However, the model closest to the true model is stopped after the second iteration when χ 2and Φ start to increase. The
recovered by inversion of noisy VLF data using a starting model of fitting between the noisy synthetic VLF and VLF-R data and the
2000 Ω·m (Fig. 6) because this starting model is closest to the host computed data is shown in Figs. 11 and 13, respectively.
resistivity (1000 Ω·m) of the true model. The fitting between noisy In this study, we have only shown results for the parameterization
synthetic data and computed data is shown in Fig. 7. We would like to scheme containing rectangular and trapezoidal cells. However, the great
point out that the inversion of VLF data can provide useful information advantage of the parameterization into unstructured grids may come
about conductivity distributions if an adequate value of the starting into play if we succeed in adapting our meshes to the conductivity
model resistivity is available e.g. from DC resistivity or other surveys structures that evolve during the course of the inversion process.
(Monteiro Santos et al., 2006). From experiments with different parameter grids we have learned
In contrast to noisy VLF data damped least-squares inversion, the that the inversion result is affected by the size and the location of the
true model has been recovered sufficiently well by damped least- grid cells. We choose the vertical and horizontal grid size according to
squares inversion of noisy VLF-R data using starting models of 500, the skin depth of the EM fields and the spatial data sampling,
2000 and 5000 Ω·m resistivities. Here, we show the inverted model respectively. The vertical extent of the parameterized model is
obtained by the starting model of 2000 Ω·m in Fig. 8. The χ 2-criterion determined to be 1 to 2 times the maximum skin depth. In vertical

Fig. 10. Inverted model (dipping bodies) obtained by minimum-model inversion of VLF data starting with the reference model of a 2000 Ω·m half-space.
370 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372

Fig. 11. Fitting between noise added synthetic data and modeled data for minimum-model inversion of VLF data starting with the reference model of a 2000 Ω·m half-space. Symbols
and solid line represent synthetic and computed data, respectively.

Fig. 12. Inverted model (dipping bodies) obtained by minimum-model inversion of VLF-R data starting with the reference model of a 2000 Ω·m half-space.

direction, the cell size is increased with depth by a factor of 1.2 to 1.5 Bräunsdorf, Germany. From geological mapping a pyrite mineraliza-
such that 10 to 15 cells are obtained in the subsurface. In horizontal tion of hydrothermal origin is known there. Monteiro Santos et al.
direction, the grid is arranged so that the observation points do not (2006) have reported that the inversion of even single frequency VLF
coincide with grid nodes. More specifically, they are placed at the data can yield useful models if it is carried out using a background
mid-points of two adjacent nodes. resistivity in the starting model that is close to the true host resistivity.
We have chosen a homogeneous half-space of 500 Ω·m as starting
7. Inversion of field data from a region with topography model which appears to be reasonable from early DC resistivity and
EM surveys carried out in this area. Inversion of field VLF data is
Finally, we present the inversion of field VLF data for a frequency of executed for models with and without topography using rectangular
21.8 kHz collected from a region with undulating topography in and trapezoidal parameter cells. The models obtained after damped

Fig. 13. Fitting between noise added synthetic data and modeled data for minimum-model inversion of VLF-R data starting with the reference model of a 2000 Ω·m half-space.
Symbols and solid line represent synthetic and computed data, respectively.
 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372 371

Fig. 14. Inverted model of Bräunsdorf data for a model without topography. 5% error Fig. 17. Inverted model of Bräunsdorf data for a model including topography in the
floor is assumed in the observed data and damped least-squares inversion is started region. 5% error floor is assumed in the observed data and the minimum-model
with 500 Ω·m half-space model. inversion is started with the reference model of 500 Ω·m half-space.

The damped least-squares inversion result for the model including

topography displayed in Fig. 15 is validated by the minimum-model
solution depicted in Fig. 17. Both models show the same anomalous
body between x = 50 m and x = 100 m which is even more pro-
nounced in terms of decreased resistivity in the case of the minimum-
model approach yielding a smoother background as expected from
the previous studies.

8. Conclusions

We have developed a 2-D inversion code for inverting plane wave

EM data for models including topography. By means of forward
Fig. 15. Inverted model of Bräunsdorf data for a model including topography in the modeling studies, we have demonstrated that the topography effect
region. 5% error floor is assumed in the observed data and damped least-squares may become significant. For this purpose, we have decomposed the total
inversion is started with 500 Ω·m half-space model.
response into individual contributions from the topography and from
the subsurface conductivity structures. A flat earth damped least-
least-squares inversion not accounting for topography and including squares inversion of data generated from a homogeneous model
topography are shown in Figs. 14 and 15, respectively. Inversion for including topography exhibits characteristic artifacts and, thus, corrob-
the flat earth model yields an expanded split conductor at a larger orates the necessity to incorporate the topography into the inversion
depth than the single conductor obtained in the model accounting for process.
topography located between x = 50 m and x = 100 m. The flat earth After demonstrating the effect of topography, we have studied the
model also shows another conductive body at x = −50 m which effect of varying starting models on the damped least-squares inversion
seems to be due to topography alone. The conductive body on the left- of VLF and VLF-R data. Inversion of VLF-R data is less dependent on the
hand model margin is due to a Quadrature anomaly only and more starting model in comparison to VLF data. In the case of VLF inversion,
pronounced in the inverted flat earth model. We attribute it to the resistivity of the homogeneous starting model should be close to the
artificial electromagnetic noise. Fitting between field data and true host resistivity and provide a corresponding maximum skin depth
modeled data for the inverted model incorporating topography is that is sufficiently large to yield the target structure. By employing a
shown in Fig. 16. An almost similar fitting is obtained for the inverted minimum-model inversion scheme we have eliminated undesired
flat earth model as well. oscillations of the background resistivity in the resulting models.

Fig. 16. Fitting between observed data and modeled data obtained after damped least-squares inversion of Bräunsdorf data starting with 500 Ω·m half-space. Symbols and solid lines
represent observed and computed data, respectively.
372 V.C. Baranwal et al. / Journal of Applied Geophysics 75 (2011) 363–372

However, in this case the inversion result is affected by the reference Extended abstract, 17th Workshop on Electromagnetic Induction in the Earth,
Hyderabad, India, pp. 1–6 www-documenthttp://www.emindia2004.org. S.2-O.01.
model parameters instead of those of the starting model. Franke, A., Börner, R.-U., Spitzer, K., 2007. Adaptive unstructured grid finite element
Our implementation of VLF and VLF-R inversion schemes is able to simulation of two-dimensional magnetotelluric fields for arbitrary surface and
resolve anomalous conductivity regions in the presence of topography. seafloor topography. Geophysical Journal International 171, 71–86.
Guenther, T., Ruecker, C., Spitzer, K., 2006. Three-dimensional modelling and inversion
However, VLF (contrary to VLF-R) is not able to recover the background of dc resistivity data incorporating topography—II. Inversion. Geophysical Journal
resistivity very well. The example of a field VLF data inversion has shown International 166, 506–517.
efficacy of our inversion code. Concluding, the inversion of models Jiracek, G.R., Reddig, R.P., Kojima, R.K., 1989. Application of the Rayleigh-FFT technique
to magnetotelluric modeling and correction. Physics of the Earth and Planetary
including surface or subsurface topography, i.e., seabed topography, Interiors 53, 365–375.
voids, mining galleries, tunnels, caves etc. opens up new ways for field Kaikkonen, P., Sharma, S.P., 1998. 2-D nonlinear joint inversion of VLF and VLF-R data
surveys and specific applications and enhances the interpretation using simulated annealing. Journal of Applied Geophysics 39, 155–176.
Karous, M.R., 1979. Effects of relief in EM methods with very distant source.
techniques available at present.
Geoexploration 17, 33–42.
Key, K., Weiss, C., 2006. Adaptive finite-element modeling using unstructured grids: the
Acknowledgements 2D magnetotelluric example. Geophysics 71, g291–g299.
Key, K., Constable, S., Weiss, C.J., 2006. Mapping 3D salt using the 2D marine
agnetotelluric method: case study from Gemini Prospect, Gulf of Mexico.
We acknowledge the German Academic Exchange Service Programme Geophysics 71, B17–B27.
(DAAD) for funding VCB to carry out research at TU Bergakademie Li, S., Booker, J.R., Aprea, C., 2008. Inversion of magnetotelluric data in the presence of
strong bathymetry/topography. Geophysical Prospecting 56, 259–268.
Freiberg, Germany. Our gratitude goes to the G.W. Hohmann Memorial Liu, G., Becker, A., 1992. Evaluation of terrain in AEM surveys using the effects boundary
Trust for supporting AF by a graduate scholarship. We are also grateful to element method. Geophysics 57, 272–278.
the German Research Foundation (DFG) for supporting our research. Loke, M.H., Barker, R.D., 1996. Rapid least-squares inversion of apparent resistiity
pseudosections by a quasi-Newton method. Geophysical Prospecting 44, 131–152.
Finally, we acknowledge Dr. Colin Farquharson and two anonymous Meju, M.A., 1994. Geophysical data analysis: understanding inverse problem theory
reviewers for their critical and helpful comments which improved the and practice, Vol. 6. Society of Exploration Geophysicists, Tulsa, Oklahoma.
manuscript to the present form. Menke, W., 1989. Geophysical Data Analysis: Discrete Inverse Theory. Academic Press
Inc., San Diego, California.
Monteiro Santos, F.A., Mateus, A., Figueiras, J., Gonçalves, M.A., 2006. Mapping
References groundwater contamination around a landfill facility using the VLF-EM method — a
case study. Journal of Applied Geophysics 60, 115–125.
Arcone, S.A., 1978. Investigation of a VLF airborne resistivity survey conducted in Oristaglio, M.L., Worthington, M.H., 1980. Inversion of surface and borehole electromag-
northern Maine. Geophysics 43, 1399–1417. netic data for two dimensional electrical conductivity models. Geophysical Prospecting
Aster, R.C., Borchers, B., Thurber, C., 2005. Parameter Estimation and Inverse Problems. 28, 633–657.
Elsevier Academic Press. Oskooi, B., Pedersen, L.B., 2005. Comparison between VLF and RMT methods: a
Baba, K., Chave, A.D., 2005. Correction of seafloor magnetotelluric data for topographic combined tool for mapping conductivity changes in the sedimentary cover. Journal
effects during inversion. Journal of Geophysical Research 110, B12105. of Applied Geophysics 57, 227–241.
Baker, H.A., Myers, J.O., 1980. A topographic correction for VLF-EM profiles based on Oskooi, B., Pedersen, L.B., 2006. Resolution of airborne VLF data. Journal of Applied
model studies. Geoexploration 18, 135–144. Geophysics 58, 158–175.
Baranwal V.C., 2007. Integrated interpretation of VLF data with other geophysical data Paal, G., 1968. Very low frequency measurements in northern Sweden. Geoexploration
and study of two-dimensional VLF modeling and inversion. Ph.D. Thesis, Dept. Of 6, 141–149.
Geology and Geophysics, IIT Kharagpur, India. Pedersen, L.B., Oskooi, B., 2004. Airborne VLF measurements and variations of ground
Baranwal, V.C., Sharma, S.P., 2006. Integrated geophysical studies in the East-Indian conductivity: a tutorial. Surveys in Geophysics 25, 151–181.
geothermal province. Pure and Applied Geophysics 163, 209–227. Poddar, M., Rathor, B.S., 1983. VLF survey of the weathered layer in southern India.
Beamish, D., 1994. Two-dimensional, regularised inversion of VLF data. Journal of Geophysical Prospecting 31, 524–537.
Applied Geophysics 32, 357–374. Rodi, W.L., 1976. A technique for improving the accuracy of finite element solutions for
Börner, R.-U., Ernst, O.G., Spitzer, K., 2008. Fast 3D simulation of transient magnetotelluric data. Geophysical Journal of the Royal Astronomical Society 44,
electromagnetic fields by model reduction in the frequency domain using Krylov 483–506.
subspace projection. Geophysical Journal International 173, 766–780. Ruecker, C., Guenther, T., Spitzer, K., 2006. Three-dimensional modeling and inversion
Chouteau, M., Bouchard, K., 1988. Two-dimensional terrain correction magnetotelluric of DC resistivity data incorporating topography — I. Modelling. Geophysical Journal
surveys. Geophysics 53, 854–862. International 166, 495–505.
Coggon, J.H., 1971. Electromagnetic and electrical modeling by the finite element Sasaki, 1989. Two-dimensional joint inversion of magnetotelluric and dipole–dipole
method. Geophysics 36, 132–155. resistivity data. Geophysics 54, 254–262.
deGroot-Hedlin, C., Constable, S., 1990. Occam's inversion to generate smooth, two- Sharma, S.P., Baranwal, V.C., 2005. Delineation of groundwater bearing fracture zones in
dimensional models from magnetotelluric data. Geophysics 55, 1613–1624. a hard rock area integrating very low frequency electromagnetic and resistivity
Drahor, M.G., 2006. Integrated geophysical studies in the upper part of Sardis data. Journal of Applied Geophysics 57, 155–166.
archaeological site, Turkey. Journal of Applied Geophysics 59, 205–223. Sharma, S.P., Kaikkonen, P., 1998. Two-dimensional nonlinear inversion of VLF-R data
Farquharson, C.G., Oldenburg, D.W., 1996. Approximate sensitivities for the electro- using simulated annealing. Geophysical Journal International 133, 649–668.
magnetic inverse problem. Geophysical Journal International 126, 235–252. Siripunvaraporn, W., Egbert, G., 2000. An efficient data-subspace inversion method for
Farquharson, C.G., Oldenburg, D.W., 2004. A comparison of automatic techniques for 2-D magnetotelluric data. Geophysics 65, 791–803.
estimating the regularization parameter in non-linear inverse problems. Geophysical Wannamaker, P.E., Stodt, J.A., 1987. A stable finite element solution for two-dimensional
Journal International 156, 411–425. magnetotelluric modeling. Geophysical Journal of the Royal Astronomical Society 88,
Fischer, G.A., 1989. A strong topographic valley effect in AMT and VLF-R measurements. 277–296.
Geophysical Journal International 96, 469–475. Wannamaker, P.E., Stodt, J.A., Rijo, L., 1986. Two-dimensional topographic responses in
Fox, R.C., Hohmann, G.W., Killpack, T.J., Rijo, L., 1980. Topographic effects in resistivity magnetotellurics modeled using finite elements. Geophysics 51, 2131–2144.
and induced-polarization surveys. Geophysics 45, 75–93. Zlotnicki, J., Vargemezis, G., Mille, A., Bruere, F., Hammouya, G., 2006. State of the
Franke, A., Börner, R.-U., Spitzer, K., 2004. 2D finite element modelling of plane-wave hydrothermal activity of Soufriere of Guadeloupe volcano inferred by VLF surveys.
diffusive time-harmonic electromagnetic fields using adaptive unstructured grids. Journal of Applied Geophysics 58, 265–279.