You are on page 1of 13

GEOPHYSICS, VOL. 80, NO. 1 (JANUARY-FEBRUARY 2015); P. EN43–EN55, 15 FIGS., 2 TABLES.

10.1190/GEO2013-0476.1

Joint inversion of seismic refraction and resistivity data using


layered models — Applications to groundwater investigation

Niklas Juhojuntti1 and Jochen Kamm2

ABSTRACT lateral smoothing to the layer boundaries and to the materialpara-


meters. Depending on the subsurface conditions, the smoothing
We developed a method for joint inversion of seismic refrac- can be applied either to the depth of the layer boundaries or to the
tion and resistivity data, using sharp-boundary models with few layer thicknesses. The forward responses and Jacobian for refrac-
layers (typically three). We demonstrated the usefulness of the tion seismics were calculated through ray tracing. The resistivity
approach via examples from near-surface case studies involving computations were performed with finite differences and a cell-
shallow groundwater exploration and geotechnical investigations, to-layer transform for the Fréchet derivatives. Our method per-
although it should also be applicable to other types of layered formed well in synthetic tests, and in the case studies, the layer
environments, e.g., sedimentary basins. In our model parameter- boundaries were in good agreement with in situ tests and seismic
ization, the layer boundaries were common for the resistivity and reflection data, although minimum-structure inversion generally
velocity distributions. Within the layers, only lateral variations in has a better data fit due to more freedom to introduce model
the material parameters (resistivity and velocity) were allowed, heterogeneity. We further found that our joint inversion approach
and we assumed no correlation between these. The inversion can provide more accurate thickness estimates for seismic hidden
was performed using a nonlinear least-squares algorithm, using layers.

INTRODUCTION inversion of seismic refraction and resistivity data, seeking to


combine the strengths of the two methods. The seismic method nor-
The interpretation of geophysical data is always connected with mally gives strong constraints on the depth to the major subsurface
some degree of ambiguity. The aim of joint inversion is to combine layer boundaries, whereas the resistivity method can provide high
geophysical methods to reduce nonuniqueness and improve accu- lateral resolution and detect relatively thin layers provided that
racy in subsurface models because the various methods in general these have high electrical conductivity. One particular problem ad-
are sensitive to different physical properties. For shallow targets, dressed in this study is the imaging of thin layers, with application
such as those encountered in engineering or groundwater applica- to groundwater exploration. However, our results should be appli-
tions, many methods have sufficient depth penetration and are prac- cable to various types of geophysical measurements in layered
tical to use (e.g., Kirsch, 2006; McClymont et al., 2011). We have environments.
developed a joint inversion method based on sharp-boundary mod- Inversion algorithms that use mesh-based model parameteriza-
els with a few layers. tion in combination with regularization terms to produce smooth
The seismic refraction method has long been used for near-sur- and well-behaved models are commonly used to interpret geo-
face applications, e.g., to determine the depth to the groundwater physical data, being referred to as minimum-structure inversion
table and the bedrock surface. However, the method fails when (Figure 1). The methodology has been applied to data from, e.g.,
the seismic velocity does not increase with depth, or when layers electromagnetics (de Groot-Hedlin and Constable, 1990), resistivity
are too thin (e.g., Gupta, 1977). We present an approach to joint (Loke and Barker, 1996), and induced polarization (Oldenburg and

Manuscript received by the Editor 20 December 2013; revised manuscript received 27 June 2014; published online 2 January 2015.
1
Formerly Geological Survey of Sweden, Geophysics, Uppsala, Sweden; presently LKAB, Kiruna, Sweden. E-mail: niklas.juhojuntti@lkab.com.
2
Uppsala University, Department of Earth Sciences, Uppsala, Sweden. E-mail: jochen.kamm@geo.uu.se.
© 2015 Society of Exploration Geophysicists. All rights reserved.

EN43
EN44 Juhojuntti and Kamm

Li, 1994). There are also numerous applications of minimum- pure 2D resistivity inversion, but here it is extended to joint
structure models in joint inversion, often relying on cross-gradient inversion, including seismic refraction data. The layer boundaries
constraints to achieve consistent models (Colombo and Stefano, are thus common for the velocity and resistivity parameterizations.
2007; Gallardo, 2007; Gallardo and Meju, 2007; Linde et al., We only allow lateral variations in the physical parameters within
2008; Hu et al., 2009; Moorkamp et al., 2011; Lelievre et al., 2012). a layer, and we assume no correlation between the layers,
In our approach to joint inversion, we instead use sharp-boun- implying that the major changes occur at the common layer boun-
dary, or layered, models with little variation within the layers daries.
and discontinuities at the layer boundaries (Figure 1). This type We argue that there are several advantages of using sharp-
of model is often used for inversion of seismic refraction data boundary models:
(e.g., Zelt and Ellis, 1988), and there are also examples of applica- • In many cases, this is a reasonable first-order representation
tions to resistivity (Auken and Christiansen, 2004; Akca and Baso-
of the subsurface structure.
kur, 2010) and electromagnetic data (de Groot-Hedlin and • The models directly provide interpreted depths to the subsur-
Constable, 2004) inversion. Sharp-boundary models have less often face layer boundaries. These depths are often very important
been used for joint inversion, although there are examples, includ- parameters in shallow groundwater or engineering measure-
ing the application to resistivity and surface wave seismic data by ments. The direct estimates of depths to layer boundaries
Hering et al. (1995) and Misiek et al. (1997), to teleseismic and also makes the models easy to test or calibrate against depths
magnetotelluric data by Moorkamp et al. (2007), to seismic and from in situ tests.
gravity data by Lines et al. (1988), and to gravity and refraction • It is straightforward to include data from geophysical meth-
seismic data by Afnimar et al. (2002). Kis (2002) also performs ods requiring sharp contrasts, e.g., the seismic reflection
joint inversion of resistivity and seismic refraction data. However, method, in the inversion.
their approach uses a 1D approximation and can only be used for • The models normally require few model parameters, which
2D data acquired with a specially designed survey layout, i.e., with potentially can result in more manageable and well-behaved
a large number of profiles parallel to the strike direction of the sub- inverse problems.
surface structures. In this paper, we present a method for 2D joint
inversion of seismic refraction and resistivity data using sharp- From a practical point of view, the common layer boundaries in
boundary models, which to our knowledge has not been tested our joint inversion models are very convenient, because we avoid
previously. the often nontrivial task of harmonizing models from individual in-
We restrict our models to a fixed number of layers (typically terpretations, in particular when the software packages can only
three), with sharp transitions in the model parameters at the layer handle one type of data.
boundaries and smooth lateral variations within the layers. Our ap- The remainder of the paper is structured as follows: We first de-
proach is similar to that used by Auken and Christiansen (2004) for scribe the methodology, and then we validate our method by using a
synthetic example. This is followed by one case study from ground-
water exploration in glacial sediments, including a seismic hidden-
layer scenario. Finally, we apply our method to data from a site
prone to landslides, where thorough geotechnical and geophysical
investigations were carried out.

METHODOLOGY
Model parameterization
Figure 1. Schematic comparison of (a) sharp-boundary and (b) min-
imum-structure models. We use models with sharp transitions in the model parameters at
the layer boundaries and smooth lateral variations within the layers,
similar to the approach used by Auken and Christiansen (2004) for
resistivity inversion. The key feature is that the boundary nodes for
the layers are common for the resistivity and velocity distributions,
and we assume no connection between velocity and resistivity
within the layers. We place the boundary nodes at the same lateral
positions for all layer boundaries (Figure 2). The resistivity and
velocity nodes can, however, be distributed in any manner and inde-
pendently of each other. Velocity v, thickness h, and the logarithm
of the resistivity lgðρÞ are linearly interpolated between the nodes
within the layers, where lg denotes the base 10 logarithm. Using
logarithmic resistivity is common in resistivity inversion to avoid
negative resistivity and to capture the large variability of the param-
eter more adequately. Note that our method does not require acquis-
ition arrays of similar lengths for the resistivity and the seismic
measurements because this is often difficult to achieve in practice.
Figure 2. Layout of the sharp-boundary model parameterization. The combined model vector is defined as
Layer-based seismic/DC joint inversion EN45

2 3
v ΦðmÞ ¼ Φd ðmÞ þ Φm ðmÞ; (4)
m ¼ 4 lg ðρÞ 5: (1)
lg ðhÞ where the first term is a measure of the data misfit and the second
term is a measure of how much the model deviates from the regu-
We have chosen to use logarithms of layer thicknesses because larization criteria. The data misfit term is defined as
this is often used for resistivity modeling and precludes negative
values (e.g., Vozoff and Jupp, 1975). Due to the logarithmic repre- Φd ¼ ðdobs − dpre ÞT Cd −1 ðdobs − dpre Þ; (5)
sentation, the layers can be very thin but not closed.
where dobs and dpre are the observed and predicted data sets. The
latter are obtained from a forward calculation for model m; i.e.,
Forward modeling and calculation
of Fréchet derivatives dpre ¼ gðmÞ: (6)
Two preexisting software packages have been interfaced to ob-
tain forward responses as well as Fréchet derivatives. For the refrac- The data covariance matrix Cd defines the measurement error
tion seismics, we use software from the Rayinvr package (Zelt and model, where Gaussian error statistics are implied by equation 5.
Ellis, 1988). The seismic waves are assumed to travel as head waves By assuming uncorrelated errors, Cd becomes a diagonal matrix
along the internal layer boundaries. Diving waves could sometimes whose inverse is easily obtained. Then, Cd is a normalization
be more appropriate; however, this requires vertical gradients within of the individual data misfit terms on their respective error,
the layers. rendering the measurements dimensionless and scaled to be com-
The resistivity forward responses and derivatives are obtained us- parable.
ing the 2.5D finite-difference modeling introduced by Kalscheuer It is not straightforward to estimate an error model for the data.
et al. (2010). The modeling assumes a flat terrain; hence, the calcu- The first component is the errors in the measured data, which are in
lated response will be inaccurate in the presence of severe topog- themselves not easy to quantify. Some contributing factors can be
raphy. Long-wavelength variations are less problematic. quantified to a certain degree, e.g., the traveltime picking uncer-
The finite-difference code requires a 2D resistivity model consist- tainty, uncertainty in electrode positioning, or the contact resistance.
ing of rectangular cells within which the respective resistivity has to The latter can in principle be remedied through reciprocal measure-
be constant, and it computes the Fréchet derivatives with respect to ments (Slater et al., 2000; Koestel et al., 2008), so that more reliable
cell resistivity. Therefore, we implemented a procedure for trans- statistical data errors could be obtained. The second component is
forming cell-based sensitivities to node-based sensitivities (see Ap- the bias introduced during the inversion due to the model paramet-
rization and the regularization, which is impossible to address ob-
pendix A). In addition, to obtain the sensitivities toward the model
jectively. Therefore, in practice, the error model is estimated
parameters m as specified in equation 1, the inverse operation of the
through several trial inversions with different regularization param-
logarithmic transformation has to be linearized. This procedure can
eters, determining how well the data can be fit with reasonable
be found in various places in the literature (e.g., in Auken and Chris-
models.
tiansen, 2004).
The regularization term in the objective function (equation 4) is
The seismic and resistivity data are joined by concatenation of
given by
traveltimes t and logarithms of apparent resistivities lg ρa :
  Φm ðmÞ ¼ lT Cl −1 l; (7)
t
d¼ : (2)
lg ρa where

Correspondingly, the joint Fréchet derivate matrix has the form lðmÞ ¼ 0 (8)
 
Gt;v 0 Gt;lg h are constraint equations to be fulfilled. We apply lateral smoothness
G¼ ; (3)
0 Glg ρa ;lg ρ Glg ρa ;lg h constraints on model parameters within the layers in a similar fash-
ion as in Auken and Christiansen (2004), explained in Appendix B.
where Gt;v and Gt;lg h represent Fréchet derivatives for the seismic For the velocity and the logarithm of resistivity, this is implemented
traveltimes with respect to velocity and layer thickness, whereas by standard first-difference operators along the horizontal direction
Glg ρa ;lg ρ and Glg ρa ;lg h represent Fréchet derivatives for the appar- for each layer. The smoothness constraints for the layer boundaries
ent resistivity with respect to resistivity and layer thickness (com- are more complicated to implement. In some cases, it is reasonable
pare equation 1). to assume little variation in the layer thicknesses, whereas in other
cases, it is more reasonable to assume that the layer boundaries are
Inversion relatively flat (e.g., the groundwater table in permeable sediments),
and we allow both options to be used. When constraining the ab-
The inverse problem is formulated as a standard nonlinear opti- solute value of the thickness (rather than the logarithm), we must
mization problem to be solved in a least-squares sense (Tarantola apply the inverse log transform, which makes the regularization
and Valette, 1982). Auken and Christiansen (2004) approach layer- nonlinear.
based resistivity inversion in a similar way, the difference of the The diagonal weighting matrix Cl −1 normalizes and balances the
stated scheme being mainly the joint inversion aspect. We seek smoothness constraints contained in l. It may be modified to trade
to minimize the objective function Φ, defined as off smoothness of different model parameters. The minimization of
EN46 Juhojuntti and Kamm

Φ can be performed by setting the gradient of the function equal to Uncertainty estimates
zero, resulting in an iterative Gauss-Newton scheme. In every iter-
ation, an updated solution mkþ1 is obtained from the previous Following Auken and Christiansen (2004) and Tarantola and Val-
ette (1982), we calculate linearized a posteriori uncertainty esti-
one by
mates of the model parameters as
pffiffiffiffiffiffiffiffiffiffiffiffi
mkþ1 ¼ mk þ Δmk ; (9) σðmk Þ ¼ Cest;kk ; (12)

where Cest is given by


where Δmk is an update term that is computed by solving the linear
system
Cest ¼ ðGT Cd −1 G þ LT Cl −1 LÞ−1 : (13)

ðGT Cd −1 G þ LT Cl −1 L þ λCm −1 ÞΔmk For a linear problem that converges to a solution in a single step,
(10) and has known a priori distributions and data errors, this estimate
¼ ½GT Cd −1 ðdobs − gðmk ÞÞ − LT Cl −1 lðmk Þ: becomes exact. Although none of these assumptions are strictly ful-
filled, the estimates can be useful to examine differences between
The matrices G and L contain the Fréchet derivatives of g and l at single and joint inversion results, and to assess how the addition or
removal of data affect the models. Note that we have not included
m ¼ mk ; i.e.,
the damping term from equation 10 because it is a numerical sta-
bilizer that damps the model update in every iteration, but in prin-
∂gi ðmk Þ ∂li ðmk Þ ciple does not bias the solution toward any particular model.
Gij ¼ Lij ¼ : (11)
∂mj ∂mj
SYNTHETIC EXAMPLE
−1
Note that a stabilization term Cm has been added, to introduce We first test the performance of our algorithm on a synthetic ex-
weighted Levenberg-Marquardt damping. We define Cm −1 as a ample. We have chosen a synthetic model with three layers, resem-
diagonal matrix, with entries scaled to account for the different units bling a typical scenario in a groundwater investigation (Figure 3).
of the model parameters. The weighting is chosen to be identical to On the left side of the model, the middle layer is very thin, and
the weighting Cl −1 of the corresponding smoothness constraint therefore not visible as a separate velocity branch in the seismic
equations l. This implies that for the damping of the logarithmic traveltime curves. Such seismic hidden layers are frequently en-
thickness parameter, the weights must be modified by the Jacobian countered, and we are particularly interested in investigating
of the logarithmic transform in every iteration. The regularization whether the layer could be better constrained by our joint inversion
parameter λ that scales the damping term has to be chosen manually approach rather than by individual inversions. We have chosen a
to trade off convergence and stability of the inversion. The param- fairly low resistivity in the middle layer, and introduced mild var-
eter can be adjusted between the inversion iterations; however, in iations in the uppermost layer. The boundary between the upper-
most cases it is sufficient to use a constant value. In some formu- most two layers is relatively flat, resembling a groundwater table
lations of the inverse problem, the damping term is used to constrain in permeable sediments.
the solution to some reference model; however, we assume no such We have added Gaussian noise with standard deviations of
a priori information. The iteration is stopped when the data are fit 5% for the apparent resistivities and 0.5 ms for the traveltimes.
within the prescribed confidence limits, or when a minimum is Table C-1 in Appendix C summarizes the acquisition parameters
reached. for this synthetic example and for the following case studies.

Figure 3. Setup for the synthetic example. The upper row shows the (a) resistivity and (c) velocity distributions for the true model. Panels (b
and d) show the corresponding distributions for the starting model.
Layer-based seismic/DC joint inversion EN47

The individual inversions and the joint inversions were carried out Using equation 13, we have also calculated uncertainty estimates
using the same weighting parameters for the regularization. For this (Figure 5). With respect to the layer boundaries, the estimates
particular model, we chose the regularization weights to primarily indicate that the joint inversion result has the lowest uncertainty.
constrain the vertical position (or elevation) of the layer boundaries. The resistivity in the deepest layer is much better constrained
See Table C-2 in Appendix C for additional information about the by the joint inversion, and the seismic velocity is better constrained
inversion parameters. for the middle layer and the deepest layer. Because all inversions
The model from the individual resistivity inversion rather accu- were carried out using the same regularization weights, the
rately determines the thickness of the middle layer to the left in the uncertainty estimates from the different inversions are com-
model, where this layer is very thin (Figure 4), although the vertical parable.
position of the layer is somewhat wrong at the end of the profile. The calculated and predicted responses are shown in Figure 6. As
The individual seismic inversion gives a too-large thickness, be- mentioned earlier, the middle layer is not visible as a separate veloc-
cause the layer is not visible in the seismic traveltimes. The seismic ity branch in the left part of the profile. Table C-2 summarizes the
inversion gives a more accurate thickness for the middle layer in the data misfit values for the different inversion models.
right part of the profile, where it is much better constrained by the
traveltimes. CASE STUDIES
The joint inversion reconstructs the synthetic model more accu-
rately with respect to the layer boundaries, in particular regarding Field example I: Groundwater exploration and the
the thin middle layer in the left part of the model, which is now seismic hidden-layer problem
accurately imaged both regarding to the thickness and the vertical This field example is from a site in western Sweden that was in-
position. It should be pointed out that the resistivity contrasts are vestigated for groundwater exploration purposes. At the surface, the
rather large for this synthetic example. For other synthetic examples overburden is dominated by fine-grained sediments (silt and clay),
with less-pronounced resistivity contrasts, it is more difficult to im- which often have poor hydraulic conductivity. The investigations
age thin layers. were conducted to determine the overburden thickness and to detect

Figure 4. Inversion results for the synthetic example. The upper row show (a) resistivity and (c) velocity models from individual inversions. In
panels (b and d), the corresponding joint inversion models are shown.

Figure 5. Uncertainty estimates, calculated from equation 13. We display 2σ, corresponding to 95% confidence interval. The uncertainty on
the boundaries is shown with thin lines above and below the layer boundaries. The panels correspond to the models in Figure 6.
EN48 Juhojuntti and Kamm

possible overburden layers with higher hydraulic conductivity at The initial seismic measurements showed that the overburden at
depths. The bedrock in the area is interpreted to be granitic, the site is fairly thin; however, the measurements did not conclu-
although there are no outcrops at the site. sively answer the question of whether an aquifer was present or
not in the sediments. Therefore, further investi-
gations were carried out, including a resistivity
a) 100 profile, gravity measurements, and drilling. For
practical reasons, the seismic and resistivity ar-
80 rays have different lengths (Table C-1 in Appen-
dix C), and there is also an 80-m offset in the start
Traveltime (ms)

60 points of the arrays.


We have performed individual and joint inver-
40 sions of the data sets, using models with two and
three layers (Figure 7) and calculated corre-
20 sponding uncertainty estimates (Figure 8). In this
case, we use a starting model with sloping inter-
0 nal boundaries, approximately following the sur-
0 50 100 150 200 250
Distance (m) face. We have also adjusted the lateral constraints
Predicted
to impose constant layer thicknesses rather than
b) Observed c) horizontal layer boundaries, by applying the lat-
0 0
eral smoothing to the absolute value of the layer
50 50 thickness. Additional inversion parameters are
Pseudodepth (m)

Pseudodepth (m)

given in Table C-2 in Appendix C.


100 100 The second layer is not visible as a separate
velocity in the seismic traveltimes (Figure 9).
150 150 As is apparent from the misfit values (Ta-
ble C-2), the seismic traveltimes can also be ex-
200 200
plained almost equally well using models with
either two or three layers (Figure 7), and thus
0 100 200 300 0 100 200 300
Distance (m) Distance (m) it would not be possible to conclude that an aqui-
fer was present solely based on the seismic data.
Apparent resistivity ( Ω m) However, the resistivity data suggest a three-lay-
10 32 100 316 1000 3162 ered structure, with a conductive middle layer,
interpreted as an aquifer. The drilling also con-
Figure 6. Calculated and predicted responses for the synthetic example. Panel (a) shows firmed that a thin aquifer is present at the site.
seismic traveltimes, calculated (blue curve) and predicted from joint inversion (red Furthermore, the drilling shows a transition in
curve). The lower row shows apparent resistivities, (b) calculated and (c) predicted from overburden composition at roughly the same
joint inversion (right).
depth as the groundwater table, from the mainly

Figure 7. Layer-based inversion results for field example I. The uppermost row shows the two-layer models from (a) individual resistivity
inversion and (d) seismic inversion, whereas the middle (b and e) shows corresponding three-layer models. In the lowermost row (c and f), the
corresponding three-layer joint inversion model is shown. The vertical line shows depths to the groundwater table and the bedrock surface,
determined via drilling.
Layer-based seismic/DC joint inversion EN49

fine-grained sediments close to the surface to sand and gravel at We interpret the model layers as, starting from the surface: dry
depths. Our three-layered model from joint inversion is in good overburden with high resistivity and low velocity, water-saturated
agreement with the information from the drilling (Figure 7). The overburden with low velocity and midrange velocity, and bedrock
three-layer seismic inversion also agrees almost equally well with with high resistivity and high velocity. The resistivity models sug-
the drilling; however, in this case the middle layer has a virtually gest that there are some lateral variations near the surface, possibly
constant thickness, indicating that there are only weak constraints due to some sand or gravel layers. The variations appear to be par-
on the thickness variations from the data. ticularly strong around the borehole position, where the joint inver-

Figure 8. Uncertainty estimates for field example I. See Figure 5 for further explanation. The panels correspond to the models in Figure 7. The
vertical line shows depths to the groundwater table and the bedrock surface, determined via drilling.

a) 100 Figure 9. Observed data and predicted responses


for field example I. Panel (a) shows seismic trav-
80 eltimes, observed (blue curve) and predicted from
joint inversion (red curve). The lower row shows
Traveltime (ms)

apparent resistivities, (b) observed and (c) pre-


60
dicted from joint inversion.
40

20

0
0 50 100 150 200 250 300 350 400 450 500
Distance (m)
b) Observed c) Predicted
0 0
Pseudodepth (m)

Pseudodepth (m)

100 100

200 200

300 300

0 100 200 300 400 500 0 100 200 300 400 500
Distance (m) Distance (m)

Apparent resistivity (Ω m)
100 316 1000 3162
EN50 Juhojuntti and Kamm

sion models also show variations in the middle layer. However, the for the three-layer individual seismic inversion. Interestingly, the
explanation for the variations in the middle layer could be that the velocity in this layer becomes better constrained after the addition
lateral smoothing constraints on the resistivity in the uppermost of the resistivity data in the joint inversion. The model geometries
layer prevents the near-surface variations to be properly accounted are similar, so this change cannot be explained solely by a difference
for in this layer. in the depth to the layers.
The uncertainty analysis shows, as expected, that the resistivity For comparison, we also carried out a minimum-structure inver-
and velocity models are better determined toward the center of the sion of the resistivity data using the same forward modeling routine
profile, where the depth penetration is best (Figure 8). In general, as for the layer-based inversion (without the cell-to-layer conver-
there is also a trade-off between the depth of a layer and the uncer- sions), and following a inversion approach similar to that of
tainty of the physical parameters (resistivity and the velocity). It is Oldenburg and Li (1994) and Loke and Barker (1996). The mini-
clear that the velocity is very poorly constrained in the middle layer mum-structure model shows the same trends as the three-layered
resistivity models (Figure 10); however, the model suggests too-
large depths to the layer boundaries. Interestingly, all the layered
resistivity models have almost as good a data fit as the minimum-
Depth (m)

−10
0 structure model.
10
20
0 50 100 150 200 250 300 350 400 450 500
Field example II: Geotechnical investigation
Distance (m)

Here, we examine data from an area in western Sweden, close to


Resistivity (Ω m)
100 316 1000 3162 the Göta River, which has been investigated using a variety of meth-
ods, including the cone penetration test (CPT) method and numer-
Figure 10. Resistivity model for field example I, based on mini- ous geophysical techniques, e.g., radio magnetotelluric, slingram,
mum-structure inversion. The layer boundaries from the joint inver- resistivity, and seismic measurements (Kamm et al., 2013; Mal-
sion have been overlaid for comparison. ehmir et al., 2013a, 2013b; Shan et al., 2014). We use data from

Figure 11. Layer-based inversion results for field example II. The upper row shows the models from (a) individual resistivity inversion and
(c) seismic inversion, and panels (b and d) show the corresponding joint model.

Figure 12. Uncertainty estimates for field example II. See Figure 5 for further explanation. The panels correspond to the models in Figure 11.
Layer-based seismic/DC joint inversion EN51

a)
line 2 in Malehmir et al. (2013a) and compare 140
our results to a coincident reflection seismic pro- 120
file and CPT data. The motivation for the mea-

Traveltime (ms)
surements was to investigate a certain type of 100
clay sediments, commonly referred to as quick- 80
clay, which can become unstable and cause land- 60
slides. However, we do not study the quick-clay
aspects here. 40
The inversion was regularized using constant 20
thickness constraints, because the sloping topog- 0
raphy makes horizontal layers an unrealistic 0 50 100 150 200 250 300 350 400 450
Distance (m)
assumption. Likewise, we chose a starting model
with constant layer thicknesses. b) Observed c) Predicted
As for the previous case study, we present the 0 0
layer-based inversion models (Figure 11) to-
gether with the associated uncertainty plots 100 100
Pseudodepth (m)

Pseudodepth (m)
(Figure 12), the observed data and predicted re-
sponses (Figure 13), and a minimum-structure 200 200
model (Figure 14).
The joint inversion model shows a similar type 300 300
of seismic and velocity distribution as in the pre-
vious examples (Figure 11). The uppermost layer 400 400
likely represents dry overburden, whereas the
middle layer represents water-bearing sediments,
0 100 200 300 400 0 100 200 300 400
although in this case we do not have any indepen- Distance (m) Distance (m)
dent information regarding depth to the water-
saturated layer. At the northern part of the model, Apparent resistivity ( Ω m)
100 316 1000 3162
the middle layer becomes much more conduc-
tive, suggesting an overall compositional change
to more fine-grained sediments. Figure 13. Observed data and predicted responses for field example II. Panel (a) shows
seismic traveltimes, observed (blue curve) and predicted from joint inversion (red
A comparison with the reflection seismic data curve). The lower row shows apparent resistivities, (b) observed and (c) predicted from
indicates that the deepest layer in our model rep- joint inversion.
resents bedrock along the southern part of the
Depth (m)

profile (Figure 15). Further to the north, the interpretation is likely −20
more complicated. According to the interpretation of the CPT and 0
the reflection seismic data from Malehmir et al. (2013b), a sedimen- 20
0 50 100 150 200 250 300 350 400 450
tary unit is present at depth along the northern part of the profile. We Distance (m)
suggest that the deepest layer in our model corresponds to this unit
Resistivity (Ω m)
within the distance interval 250–360 m (Figures 11 and 15). This 100 316 1000 3162
unit likely has a relatively high seismic velocity and a considerably
higher resistivity than the overlying sediments. In the northernmost Figure 14. Resistivity model for field example II, based on mini-
part of the profile, it is more difficult to see a correlation between the mum-structure inversion. The layer boundaries from the joint inver-
deepest layer in our model and the CPT and the reflection seismic sion have been overlaid for comparison.
section. The individual seismic inversion shows a
rather large depth to the deepest layer in the SE NW
25
northern part of the profile, which appears poorly
Approx. elevation (m)

10
constrained and suggests a lack of depth penetra- –5
tion in the seismic refraction data. The CPT fur- –20
–35
thest to the north did not reach any hard material;
–50
however, the seismic reflection data do not sup- –65
port a dramatic increase in the depth to the deep- –80
est layer in the north. The individual resistivity –95
–110
inversion does not show any resistive deep layer –125
in the north, possibly due to a lack of depth pen- 0 50 100 150 200 250 300 350 400 450
etration, and the model geometry is quite differ-
ent in comparison to the other models. However,
Figure 15. Velocity model from joint inversion for field example II, with seismic re-
the resistivity in the deepest layer appears well flection section in the background (modified from Malehmir et al., 2013b). The vertical
constrained based on the uncertainty analysis, bars show depths from geotechnical drill holes (CPT data). With the exception of the
probably due to the thin overlying layers. northernmost drill hole, these all reached hard material.
EN52 Juhojuntti and Kamm

The minimum-structure resistivity model (Figure 14) is geomet- this sense, and even though these models are often associated with
rically quite similar to the joint inversion model. The deep high-re- low data misfits, the ambiguous interpretation will introduce uncer-
sistivity unit again disappears at the northern end. The resistivity tainty. For our sharp-boundary models, we can immediately com-
data fit is much better for the minimum-structure inversion than pare our results to, e.g., depths from drilling. An obvious extension
for the joint inversion (see Table C-2), most likely due to difficulties to our method would be to also use the drilled interfaces to constrain
in imaging the shallow resistivity structure with the layered model. our models. The fairly good agreement between our layer bounda-
It could perhaps have been possible to improve the resistivity data ries and the drillings suggests that we are correct in assuming that
fit by using a model with four layers; however, the layer boundaries these boundaries coincide with the major changes in the physical
would likely have been poorly constrained. parameters.
The layer-based approach could perhaps also facilitate the char-
DISCUSSION acterization of different subsurface units in terms of their physical
parameters. We can easily study the lateral resistivity and seismic
We have demonstrated that sharp-boundary inversion can be used velocity distribution for a specific layer, e.g., the one we interpret to
to obtain readily interpretable geophysical models, which in terms be the aquifer. This would be more complicated when using a mini-
of the layer boundaries agree well with other independent data. The mum-structure inversion. We assume that the major changes in the
necessary a priori assumption is that the subsurface is approxi- physical parameters occur simultaneously at the layer boundaries;
mately layered, such as in the near-surface environments in our however, we make no assumptions about relationships between
study areas, and potentially in sedimentary bedrock environments. these parameters within the layers. Another extension to our inver-
In an environment with smooth subsurface variations, it would sion method could be to assume some relationship between the
likely be more appropriate to use a minimum-structure inversion. physical parameters, perhaps by requiring correlation between
Following Occam’s principle, the minimum-structure inversion the horizontal derivatives.
could also be more justified when there is little or no support Our models allow us to include other types of geophysical signals
for a layered structure. In this case, the sharp-boundary models that require sharp layer boundaries, e.g., seismic reflections. In field
could potentially be misleading, implying an unrealistically high example II, the lowermost layer boundary would likely have been
accuracy of the interface positions. Minimum-structure models better constrained in the northern part of the model (Figure 11) if we
could then be more reliable in capturing the large-scale structures had included seismic reflection traveltimes in the inversion. We
and trends. have used our inversion code for a combined interpretation of seis-
Comparing the data misfit of our layer-based algorithm to that of mic refraction and reflection traveltimes for a data set not shown in
minimum-structure inversion, we generally observe that the latter this study, but thus far we have not applied it to a joint inversion
approach gives smaller misfit values, as can be expected. In the example.
minimum-structure inversion, there is often more freedom to intro-
duce heterogeneity to explain the observed data, whether this CONCLUSIONS
heterogeneity is real or not. For both approaches, the choice of
proper regularization weights is important. In our approach, we We have developed a joint inversion method based on sharp-
have many different weights, which can be adjusted according to boundary models with a small, fixed number of layers. The method
the a priori assumptions. Our experience from applying the algo- has been applied to seismic refraction and resistivity data from near-
rithm to a variety of data sets suggests that stable inversion results surface investigations. In synthetic and field examples, we show
often can be achieved fairly easily, although some trial runs are al- that our inversion method provides models that are inherently
ways required to find proper regularization weights. We often find and conveniently interpreted in terms of depths to major layer boun-
that the layer boundaries are the parameters that are most difficult to daries, such as the groundwater table or the bedrock surface. The
constrain and are most likely to become unstable in the inversion models agree well with independent information, such as in situ
process unless smoothed properly. In general, we try to minimize tests and seismic reflection data. Minimum-structure inversion will
the difference in the smoothing parameters between the layers; how- usually give a better data fit than the sharp-boundary inversion, due
ever, it is often necessary to choose these values individually. to more freedom to introduce heterogeneity, in particular in the shal-
Because our inverse problem is nonlinear, there is a risk of only low subsurface. We find that the geometries of the layer interfaces in
finding a local minimum to the objective function, making the our models are mostly determined by the seismic data. However,
choice of a proper starting model quite important. A valuable ad- there are situations where the resistivity data provide additional con-
dition to our algorithm could be some routines for determining the straints on the layer thicknesses, e.g., for thin layers, thus suggesting
starting models, perhaps via approximate 1D inversions. It would that our approach could help solving the seismic hidden-layer
also be interesting to try some other way of solving the inverse prob- problem.
lem, e.g., the direct-search method based on Voronoi cells proposed
by Sambridge, 1999a, 1999b), which also includes straightforward ACKNOWLEDGMENTS
uncertainty estimates. Another option would be some type of ge-
netic algorithm similar to that used by Moorkamp et al. (2007). N. Meqbel at GeoForschungsZentrum in Potsdam kindly allowed
Perhaps the most significant advantage of the layer-based ap- us to use the finite-difference code for resistivity forward modeling.
proach is the very convenient inherent interpretation of the depths We thank M. Bastani and L. Persson at the Geological Survey of
to subsurface layer boundaries. In many practical applications, it is Sweden for support during various stages of this study. The project
necessary to produce interpretations including depths to major in- Geoscientists Without Borders, sponsored by SEG, provided us
terfaces, e.g., the bedrock surface or the groundwater table. Mini- with data for field example II. We also thank the reviewers for their
mum-structure models are much less straightforward to interpret in constructive comments that improved the paper.
Layer-based seismic/DC joint inversion EN53

APPENDIX A the thickness node j lies at the same horizontal position as the depth
node n and in a layer above zðnode nÞ . It has the value
TRANSFORMATION OF CELL-BASED TO
NODE-BASED SENSITIVITIES FOR ∂aðcell kÞ ∂aðcell kÞ
¼ ; (A-5)
RESISTIVITY FORWARD MODELING ∂hðnode jÞ ∂zðnode nÞ
Here, we demonstrate a scheme to transform the sensitivity of
cells, or grid nodes, from the finite-difference code to sensitivities and it is obtained through elementary geometry. The derivation is
lengthy due to the many ways that a boundary can cut through a
of the layer nodes of the layer-based parameterization.
cell, but straightforward. Therefore, we omit to state it.
We assume the cell resistivity to equal the layer resistivity at the
location of the cell midpoint. Only in the case of a layer boundary
cutting a cell are the logarithmic values above and below the boun- APPENDIX B
dary averaged according to the area size of the two segments. The LATERAL SMOOTHNESS CONSTRAINTS
derivatives with respect to the resistivity of cells k (cell midpoints at
yc ðkÞ, zc ðkÞ) are translated to a derivative with respect to the (log-
arithmic) resistivity at a particular node j (located at yn ðjÞ) within We impose lateral smoothness by minimizing the difference in
layer l via the model parameters between neighboring nodes. A first difference
matrix R serves as a suitable discrete approximation (Constable
∂ρa X ∂ρa ∂ρðcell kÞ ∂lgðρðcell kÞ Þ et al., 1987). It is applied to the velocity, the logarithm of the re-
¼
∂ lgðρðnode jÞ Þ cells k in layer l ∂ρðcell kÞ ∂ lgðρðcell kÞ Þ ∂lgðρðnode jÞ Þ sistivity and the layer thickness. The constraint equations are

X ∂ρa ∂lgðρðcell kÞ Þ lv ðvÞ ¼ Rv ðv0 − vÞ ¼ 0; (B-1)


¼ ρðcell kÞ ;
cells k in layer l
∂ρðcell kÞ ∂lgðρðnode jÞ Þ
llg ρ ðlg ρÞ ¼ Rlg ρ ðlg ρ0 − lg ρÞ ¼ 0; (B-2)
(A-1)
where lh ðhÞ ¼ Rh ðh0 − hÞ ¼ 0: (B-3)
∂lgðρðcell kÞ Þ
Here, the user-defined parameters v0 , lg ρ0 , and h0 are usually
∂lgðρðnode jÞ Þ chosen to be constant for maximum smoothness, but they can be
8
> yn ðkÞ − yc ðj − 1Þ used to impose lateral discontinuities, as well. Particularly in the
>
> yn ðj − 1Þ < yc ðkÞ < yn ðjÞ;
> case of layer thickness, it is often useful to impose horizontally flat
< yn ðjÞ − yn ðj − 1Þ
>
layers underneath a variable topography, e.g., for a free groundwater
¼ yn ðj þ 1Þ − yc ðkÞ table. Then, h0 is chosen to reflect the topography accordingly.
>
> yn ðjÞ < yc ðkÞ < yn ðj þ 1Þ;
>
> yn ðj þ 1Þ − yn ðjÞ Even though the model is parameterized using logarithmic thick-
>
:
0 otherwise; ness, we often choose to constrain the absolute thickness value be-
cause constraining the logarithm of the thickness would impose a
(A-2) relative measure and thus allow more variation for the boundaries of
thick layers than of thin layers. This choice makes it impossible to
and to a derivative with respect to (logarithmic) layer thickness at a
apply equation B-3 directly to the model parameter vector. Instead,
particular node j at yðjÞ for layer boundary l via
the logarithmic transform must be inverted leading to a nonlinear
∂ρa equation in the logarithmic thickness
∂lgðhðnodejÞ Þ
X ∂ρa ∂ρðcellkÞ ∂lgðρðcellkÞ Þ ∂aðcellkÞ ∂hðnodejÞ llg h ðlg hÞ ¼ Rh ðh0 − 10lg h Þ ¼ 0: (B-4)
¼
∂ρ
cellskcut byboundaryl ðcellkÞ
∂lgðρ ðcellkÞ Þ ∂aðcellkÞ ∂hðnodejÞ ∂lgðhðnodejÞ Þ
The composite equations for the model parameter vector m
X ∂ρa ∂lgðρðcellkÞ Þ ∂aðcellkÞ read
¼hðnodejÞ ρðcellkÞ ;
cellskcutbyboundaryl
∂ρ ðcellkÞ ∂aðcellkÞ ∂hðnodejÞ 2 3
lv ðvÞ
(A-3) 6 7
lðmÞ4 llg ρ ð lg ρÞ 5: (B-5)
where aðcell kÞ is the area ratio of the segments of the cell k, which llg h ð lg hÞ
has been intersected by the layer boundary. The values ∂ρa ∕∂ρðcell kÞ
are delivered by the finite-difference code. With z pointing positive The diagonal weighting matrix
downward, 2 3
  C ∂y v
∂ lgðρðcell kÞ Þ ρbelow ðyc ðkÞÞ Cl ¼ 4 C∂y lg ρ 5 (B-6)
¼ lg ; (A-4)
∂aðcell kÞ ρabove ðyc ðkÞÞ C∂y h

i.e., the relative vertical resistivity contrast across boundary l at the implements the a priori estimate or expectation on the lateral varia-
lateral position of cell k. The last factor is only different from zero if tion of the regularized quantities from one node to the next. The
EN54 Juhojuntti and Kamm

diagonal entries can be understood as the inverses of the variances REFERENCES


σ 2 and have to be supplied in the according units. Keeping the node
spacing in consideration, this is useful to guide the process of find- Afnimar, A., K. Koketsu, and K. Nakagawa, 2002, Joint inversion of refrac-
tion and gravity data for the three-dimensional topography of a sediment-
ing suitable parameters by physical considerations. The values σ ∂y v basement interface: Geophysical Journal International, 151, 243–254, doi:
for the velocities are supplied in m∕s, the thickness variations σ ∂y h 10.1046/j.1365-246X.2002.01772.x.
are in meters, and the resistivity variation is given by σ ∂y lg ρ ¼ Akca, I., and A. T. Basokur, 2010, Extraction of structure-based geoelectric
models by hybrid genetic algorithms: Geophysics, 75, no. 1, F15–F22,
 lg ½1 þ ðX∕100%Þ, where X is the percentage of change ex- doi: 10.1190/1.3273851.
pected for the resistivity in Ωm. Auken, E., and A. V. Christiansen, 2004, Layered and laterally constrained
2D inversion of resistivity data: Geophysics, 69, 752–761, doi: 10.1190/1
.1759461.
Colombo, D., and M. D. Stefano, 2007, Geophysical modeling via simulta-
APPENDIX C neous joint inversion of seismic, gravity and electromagnetic data: Appli-
cation to prestack depth imaging: The Leading Edge, 26, 326–331, doi: 10
TECHNICAL PARAMETERS .1190/1.2715057.
Constable, S. C., R. L. Parker, and C. G. Constable, 1987, Occam’s inver-
sion: A practical algorithm for generating smooth models from electro-
magnetic sounding data: Geophysics, 52, 289–300, doi: 10.1190/1
Table C-1. Acquisition parameters for the geophysical surveys .1442303.
(including the synthetic example). All resistivity measurements de Groot-Hedlin, C., and S. Constable, 1990, Occam’s inversion to generate
were carried out using the Wenner configuration. smooth, two-dimensional models from magnetotelluric data: Geophysics,
55, 1613–1624, doi: 10.1190/1.1442813.
de Groot-Hedlin, C., and S. Constable, 2004, Inversion of magnetotelluric
Seismic Resistivity data for 2D structure with sharp resistivity contrasts: Geophysics, 69, 78–
86, doi: 10.1190/1.1649377.
Gallardo, L. A., 2007, Multiple cross-gradient joint inversion for geospectral
Example Geophone Array Electrode Array imaging: Geophysical Research Letters, 34, L19301, doi: 10.1029/
distance (m) length (m) distance (m) length (m) 2007GL030409.
Gallardo, L. G., and M. Meju, 2007, Joint two-dimensional cross-gradient
Synthetic 5 235 2 240 imaging of magnetotelluric and seismic traveltime data for structural and
Field 10 460 5 400 lithological classification: Geophysical Journal International, 169, 1261–
example I 1272, doi: 10.1111/j.1365-246X.2007.03366.x.
Field 4 476 5 490 Gupta, S., 1977, Hidden layer problem in seismic refraction work: Geophysi-
cal Prospecting, 25, 385–387, doi: 10.1111/j.1365-2478.1977.tb01176.x.
example II Hering, A., R. Misiek, A. Gyulai, T. Ormos, M. Dobroka, and L. Dresen,
1995, A joint inversion algorithm to process geoelectric and surface wave
seismic data. Part I: Basic ideas: Geophysical Pro-
specting, 43, 135–156, doi: 10.1111/j.1365-2478
.1995.tb00128.x.
Table C-2. Data misfit and inversion parameters. See Appendix B for Hu, W., A. Abubakar, and T. M. Habashy, 2009, Joint
explanation. Numbers separated by forward slashes refer to electromagnetic and seismic inversion using struc-
parameters for individual layers or layer boundaries. tural constraints: Geophysics, 74, no. 6, R99–
R109, doi: 10.1190/1.3246586.
Kalscheuer, T., M. de los Angeles Garcia Juanatey, N.
Meqbel, and L. B. Pedersen, 2010, Non-linear model
Misfit Regularization error and resolution properties from two-dimen-
sional single and joint inversions of direct current re-
Resistivity sistivity and radiomagnetotelluric data: Geophysical
(% of Journal International, 182, 1174–1188, doi: 10.1111/
apparent Seismic σ ∂y lg ρ σ ∂y v (m/s σ ∂y h (% σ ∂y h j.1365-246X.2010.04686.x.
resistivity) (ms) λ (% per m) per m) per m) (cm/m) Kamm, J., M. Becken, and L. B. Pedersen, 2013, In-
version of slingram electromagnetic induction data
Synthetic using a Born approximation: Geophysics, 78,
no. 4, E201–E212, doi: 10.1190/geo2012-0484.1.
Seismic — 0.66 8 — 8/4/3 0.8/0.8 — Kirsch, R., 2006, Groundwater geophysics: Springer-
Verlag.
Resistivity 5.1 — 8 0.8/0.4/0.3 — 0.8/0.8 — Kis, M., 2002, Generalised series expansion (GSE)
Joint 5.1 0.58 8 0.8/0.4/0.3 8/4/3 0.8/0.8 — used in DC geoelectric-seismic joint inversion: Jour-
nal of Applied Geophysics, 50, 401–416, doi: 10
Field example I .1016/S0926-9851(02)00167-2.
Resistivity, 2- 12.6 — 20 35/35 — — 15 Koestel, J., A. Kemna, M. Javaux, A. Binley, and H.
Vereecken, 2008, Quantitative imaging of solute trans-
layer. port in an unsaturated and undisturbed soil monolith
Seismic, 2-layer. — 1.41 50 — 100/100 — 15 with 3-D ERT and TDR: Water Resources Research,
Resistivity, 3- 10.7 10 35/35/35 — 15/15 44, W12411, doi: 10.1029/2007WR006755.
layer. Lelievre, P., C. Farquharson, and C. Hurich, 2012, Joint
inversion of seismic traveltimes and gravity data on
Resistivity, 3- — 1.51 50 — 100/100/100 — 20/20 unstructured grids with application to mineral explo-
layer. ration: Geophysics, 77, no. 1, K1–K15, doi: 10
Joint 11.4 1.24 20 35/35/35 100/100/100 — 30/30 .1190/geo2011-0154.1.
Resistivity MS 10.2 — 1 1 — — — Linde, N., A. Tryggvason, J. E. Peterson, and S. S.
Hubbard, 2008, Joint inversion of crosshole radar
Field example II and seismic traveltimes acquired at the South Oyster
Bacterial Transport Site: Geophysics, 73, no. 4,
Resistivity 22.1 — 100 5/15/15 — — 6/6 G29–G37, doi: 10.1190/1.2937467.
Seismic — 1.85 500 — 15/5/5 — 4/4 Lines, L., A. Schultz, and S. Treitel, 1988, Cooperative
inversion of geophysical data: Geophysics, 53, 8–20,
Joint 19.8 2.76 500 16/50/50 15/5/5 — 4/4 doi: 10.1190/1.1442403.
Loke, M., and R. Barker, 1996, Rapid least-squares in-
Resistivity MS 12.2 — 1 1 — — — version of apparent resistivity pseudosections using a
Layer-based seismic/DC joint inversion EN55

quasi-Newton method: Geophysical Prospecting, 44, 131–152, doi: 10 Oldenburg, D., and Y. Li, 1994, Inversion of induced polarization data: Geo-
.1111/j.1365-2478.1996.tb00142.x. physics, 59, 1327–1341, doi: 10.1190/1.1443692.
Malehmir, A., M. Bastani, C. M. Krawczyk, M. Gurk, N. Ismail, U. Polom, Sambridge, M., 1999a, Geophysical inversion with a neighbourhood algo-
and L. Persson, 2013a, Geophysical assessment and geotechnical inves- rithm — I. Appraising the ensemble: Geophysical Journal International,
tigation of quick-clay landslides — A Swedish case study: Near Surface 138, 727–746, doi: 10.1046/j.1365-246x.1999.00900.x.
Geophysics, 11, 341–350, doi: 10.3997/1873-0604.2013010. Sambridge, M., 1999b, Geophysical inversion with a neighbourhood algo-
Malehmir, A., M. Saleem, and M. Bastani, 2013b, High-resolution reflection rithm – I. Searching a parameter space: Geophysical Journal International,
seismic investigations of quick-clay and associated formations at a land- 138, 479–494, doi: 10.1046/j.1365-246X.1999.00876.x.
slide scar in southwest Sweden: Journal of Applied Geophysics, 92, 84– Shan, C., M. Bastani, A. Malehmir, L. Persson, and M. Engdahl, 2014, In-
102, doi: 10.1016/j.jappgeo.2013.02.013. tegrated 2D modeling and interpretation of geophysical and geotechnical
McClymont, A. F., J. W. Roy, M. Hayashi, L. R. Bentley, H. Maurer, and G. data to image quick-clays at a landslide in southwest Sweden: Geophys-
Langston, 2011, Investigating groundwater flow paths within proglacial ics, 79, no. 4, EN61–EN75, doi: 10.1190/geo2013-0201.1.
moraine using multiple geophysical methods: Journal of Hydrology, 399, Slater, L., A. M. Binley, W. Daily, and R. Johnson, 2000, Cross-hole elec-
57–69, doi: 10.1016/j.jhydrol.2010.12.036. trical imaging of a controlled saline tracer injection: Journal of Applied
Misiek, R., A. Liebig, A. Gyulai, T. Ormos, M. Dobroka, and L. Dresen, Geophysics, 44, 85–102, doi: 10.1016/S0926-9851(00)00002-1.
1997, A joint inversion algorithm to process geoelectric and surface wave Tarantola, A., and B. Valette, 1982, Generalized nonlinear inverse problems
seismic data. Part II: Applications: Geophysical Prospecting, 45, 65–85, solved using the least squares criterion: Reviews of Geophysics, 20, 219–
doi: 10.1046/j.1365-2478.1997.3190241.x. 232, doi: 10.1029/RG020i002p00219.
Moorkamp, M., B. Heincke, M. Jegen, A. Roberts, and R. Hobbs, 2011, A Vozoff, K., and D. L. B. Jupp, 1975, Joint inversion of geophysical data:
framework for 3-D joint inversion of MT, gravity and seismic refraction Geophysical Journal International, 42, 977–991, doi: 10.1111/j.1365-
data: Geophysical Journal International, 184, 477–493, doi: 10.1111/j 246X.1975.tb06462.x.
.1365-246X.2010.04856.x. Zelt, C. A., and R. M. Ellis, 1988, Practical and efficient ray tracing in two-
Moorkamp, M., A. Jones, and D. Eaton, 2007, Joint inversion of teleseismic dimensional media for rapid traveltime and amplitude forward modelling:
receiver functions and magnetotelluric data using a genetic algorithm: Are Canadian Journal of Exploration Geophysics, 24, 16–31.
seismic velocities and electrical conductivities compatible?: Geophysical
Research Letters, 34, L16311, doi: 10.1029/2007GL030519.

You might also like