Sei sulla pagina 1di 11

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 109, B03311, doi:10.

1029/2003JB002716, 2004

Joint two-dimensional DC resistivity and seismic travel time inversion


with cross-gradients constraints
Luis A. Gallardo1 and Max A. Meju
Department of Environmental Science, Lancaster University, Lancaster, UK

Received 31 July 2003; revised 29 December 2003; accepted 8 January 2004; published 27 March 2004.
[1] It is now common practice to perform collocated DC resistivity and seismic refraction
surveys that complement each other in the search for more accurate characterization of the
subsurface. Although conventional separate DC resistivity and seismic models can be
diagnostic, we posit that better results can be derived from jointly estimated models. We
make the assumption that both methods must be sensing the same underlying geology and
have developed an innovative resistivity-velocity cross-gradients relationship to evaluate
the structural features common to both methods. The cross-gradients function is
incorporated as a constraint in a nonlinear least squares problem formulation, which is
solved using the Lagrange multiplier method. The resultant iterative two-dimensional
(2-D) joint inversion scheme is successfully applied to synthetic data (serving as
validation tests here) and to field data from collocated DC resistivity and seismic
refraction profiling experiments and also compared to conventional separate inversion
results. The joint inversion results are shown to be superior to those from separate 2-D
inversions of the respective data sets, since our algorithm leads to resistivity and velocity
models with remarkable structural agreement. INDEX TERMS: 0902 Exploration Geophysics:
Computational methods, seismic; 0925 Exploration Geophysics: Magnetic and electrical methods; 0935
Exploration Geophysics: Seismic methods (3025); 3260 Mathematical Geophysics: Inverse theory;
KEYWORDS: joint inversion, DC resistivity, seismic refraction
Citation: Gallardo, L. A., and M. A. Meju (2004), Joint two-dimensional DC resistivity and seismic travel time inversion with cross-
gradients constraints, J. Geophys. Res., 109, B03311, doi:10.1029/2003JB002716.

1. Introduction different physical properties, for which there is no estab-


lished analytical relationship, is much less studied.
[2] The need for improved characterization of the near
[4] In recent times, there have been different approaches
surface has led to an increase in the popularity of collocated
to 2-D joint inversion of disparate data with varying degrees
resistivity and seismic profiling surveys [e.g., Scott et al.,
of success. We may classify them under two different
2000; Meju et al., 2003]. However, the experimental data
philosophical approaches: (1) Inversion methods involving
have so far been interpreted using separate two-dimensional
the use of petrophysical or hydrological characteristics to
(2-D) inversion schemes for each method leading some-
relate two different geophysical properties [e.g., Berge et
times to models that are not in good accord. It is conven-
al., 2000]. For instance, water saturation and porosity have
tional practice to perform 2-D inversion of either DC
been assumed to provide a link between resistivity and
resistivity or seismic refraction data and then use the result
seismic velocity in porous media [see, e.g., Tillmann and
to constrain the inversion of the data for the other method
Stöcker, 2000]. (2) Inversion approaches involving the use
[e.g., Scott et al., 2000] but the resulting model from this
of structural attributes (e.g., boundaries of geological tar-
‘‘sequential inversion’’ [cf. Lines et al., 1988] is commonly
gets) as a common factor between two geophysical models
biased toward the input model. A better model can be
[e.g., Lines et al., 1988; Haber and Oldenburg, 1997; Musil
obtained by simultaneous fitting (i.e., joint inversion) of
et al., 2003; Gallardo-Delgado et al., 2003]. Examples of
the combined data sets from the different methods.
structurally based algorithms for 2-D joint resistivity and
[3] A great deal of work has been done on multidimen-
seismic inversion are those of Zhang and Morgan [1996]
sional inversion of complementary geophysical measure-
and Haber and Oldenburg [1997]. Their algorithms aim to
ments that sense the same physical properties [e.g., Sasaki,
enhance the common boundaries given by the largest
1989]. Multidimensional joint inversion of disparate or
changes in the estimated parameters as measured by a
uncorrelated data from methods based on fundamentally
Laplacian operator, concentrating on the magnitude of the
changes and losing the direction-dependent information.
1 [5] In this paper, we address the problem of how to relate
Also at División de Ciencias de la Tierra, Centro de Investigación
Cientifica y Educación Superior de Ensenada, Mexico. effectively the physical properties sensed by DC resistivity
and seismic refraction methods using a 2-D joint structural
Copyright 2004 by the American Geophysical Union. inversion approach. We provide a generalized framework for
0148-0227/04/2003JB002716$09.00 multidimensional joint inversion of disparate data sets as a

B03311 1 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

follow-up to the algorithm proposed by Gallardo and Meju


[2003] for joint 2-D DC resistivity and seismic P-wave travel
time data inversion. We also undertake experiments that
demonstrate the effectiveness of the inversion procedure.
First, we define our concept of a structural link and how we
incorporate it in a least squares based objective function and
how the problem is solved iteratively. Then the main ele-
ments of the algorithm are validated using data from alter-
native conventional forward modeling codes and then
applied to field data from a site characterized by heteroge-
neous materials.

2. Definition of the Resistivity-Velocity


Cross-Gradients Function
[6] The philosophy in our approach to joint 2-D inversion
is that even in the absence of an analytical relationship
between the physical properties exploited by different
geophysical methods, we expect a degree of structural
similarity in the images that they furnish. In general, the Figure 1. Basic 2-D grid used to represent our models of
properties of the subsurface vary with position and these the subsurface. The three-cell scheme used to define the
variations can occur in any direction. Thus, at any position, discrete version of the cross-gradients at any cell position in
the changes can be characterized in terms of two attributes, the model is also depicted.
the intensity or magnitude and the specific direction. A
commonality of distribution of these changes determines [9] In the two-dimensional case of interest, the x- and
whether or not electrical resistivity and seismic velocity z-components of ~t vanish. Thus we are interested in its
images are perceived as being structurally similar. These y-component, henceforth referred to as t:
attributes can be represented mathematically by the vector      
field of the gradients of the electrical and seismic properties. @mr ð x; zÞ @ms ð x; zÞ @mr ð x; zÞ @ms ð x; zÞ
t ð x; zÞ ¼  :
The key issue here is the following: Can these attributes be @z @x @x @z
used to develop a generalized mathematical scheme for ð3Þ
quantifying the structural similarity between resistivity
and seismic models of the heterogeneous subsurface?
[10] We estimate the derivatives in equation (3) using
[7] The structural differences between seismic velocity
forward differences (see Figure 1) yielding the formula
and electrical resistivity images may be measured using an
[Gallardo and Meju, 2003]
angular function such as

  4
tffi ðmrc ðmsb  msr Þ þ mrr ðmsc  msb Þ þ mrb ðmsr  smsc ÞÞ;
1 rmr ð x; y; zÞ  rms ð x; y; zÞ DxDz
qð x; y; zÞ ¼ cos ; ð1Þ
jrmr ð x; y; zÞjjrms ð x; y; zÞj ð4Þ

or related functionals such as 1/(1  cos q), where where the second subscript c, b, or r on mr (logarithm of
rmr(x, y, z) and rms(x, y, z) are the resistivity and seismic resistivity) or ms (slowness) denotes center, bottom or right
property gradients, respectively. However, these nonlinear cell in our heterogeneous 2-D grid as depicted in Figure 1.
functions have discontinuities arising from their angular Dx and Dz are the horizontal and vertical dimensions of the
nature and singularities in areas where rmr(x, y, z) or cells (Figure 1) and serve to normalize for grid differences
rms(x, y, z) vanish, as in homogeneous materials or local in equation (4).
maxima or minima of the physical properties. Instead, the [11] In the rest of this paper, we will develop and evaluate
preferred approach in this paper is to use the cross-gradients an algorithm that is underpinned by the cross-gradients
function [see Gallardo and Meju, 2003], which is given by concept so as to gauge the effectiveness of t as a structural
link between DC resistivity and seismic refraction models in
multidimensional joint inversion.
~
t ð x; y; zÞ ¼ rmr ð x; y; zÞ  rms ð x; y; zÞ: ð2Þ
3. Regularized Least Squares Inversion With
This nonlinear second-order function has no problems of
discontinuity or singularity other than those particular to the
Cross-Gradients Constraint
adopted model parameterizations. [12] The conventional regularized inverse problem for-
[8] Using this cross-gradients function, we deem the mulations for separate two-dimensional seismic or DC
resistivity and seismic models to be structurally identical resistivity inversion [e.g., Loke and Barker, 1995; Zelt
if ~ t(x, y, z) = ~
t(x, y, z) vanishes everywhere, i.e., ~ 0, as it and Barton, 1998; Pérez-Flores et al., 2001] involve least
implies full collinearity of simultaneous changes on the squares minimization of data misfit and smoothness con-
resistivity and seismic parameters. straints. The smoothness measures help to overcome the

2 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

problems of nonuniqueness and instability that bedevil The travel time computation is carried out using our
geophysical inversion methods and can be combined with implementation of the progressive finite differences method
the cross-gradients concept to great effect in problem of Vidale [1990], incorporating some of the improvements
formulation. We therefore define our objective function as developed by C. Zelt [Zelt and Barton, 1998]. The Jacobian
[cf. Gallardo and Meju, 2003]: matrix As, composed by the fraction of distance of every ray
   path in each model cell, is computed efficiently by ray
min F mr ; ms ¼ ½dr  f r ðmr Þ T C1
rr ½dr  f r ðmr Þ
tracing through the velocity field generated during the
forward modeling process.
þ ½ds  f s ðms Þ T C1
ss ½ds  f s ðms Þ [15] Similarly, the linearization of the cross-gradients
þ a2r mTr DT Dmr þ a2s mTs DT Dms constraint in equation (5) is accomplished using a first-
    order Taylor expansion (neglecting higher orders), namely,
mr  mRr T 1 mr  mRr
þ CRR 0 1
ms  mRs ms  mRs mr  m0r
subject to t ðmr ; ms Þ ¼ 0: ð5Þ t ðmr ; ms Þ ffi t ðm0r ; m0s Þ þ B@ A: ð8Þ
ms  m0s
Here, dr (logarithm of apparent resistivity) and ds (seismic
travel times) are the observed data, fr(mr) are the computed In this case, we require a reference resistivity model m0r,
apparent resistivities, fs(ms) are the computed travel times, a reference slowness model m0s and the derivatives of t
Crr is the covariance matrix of the resistivity data (which are with respect to the model parameters given by B. The
assumed to be uncorrelated) and Css is the covariance relevant expressions for computing B are obtained from
matrix of the observed travel times (also assumed the 2-D discrete version of t given in equation (4). These
uncorrelated). t(mr, ms) contains the cross-gradients for all are
the cells making up the model. D is the smoothness matrix,
ar and as are weighting factors that define the level of @t 4 @t 4
smoothness required in the models, and mRr and mRs are the ffi ðmsb  msr Þ; ffi ðmrr  mrb Þ;
@mrc DxDz @msc DxDz
a priori model parameters with covariance CRR. The @t 4 @t 4
superscripts T and 1 denote matrix transpose and matrix ffi ðmsc  msb Þ; ffi ðmrb  mrc Þ; ð9Þ
@mrr DxDz @msr DxDz
inverse, respectively. Note that reliable prior knowledge of
@t 4 @t 4
petrophysical relationships linking resistivity and seismic ffi ðmsr  msc Þ; and ffi ðmrc  mrr Þ:
parameters [e.g., Berge et al., 2000; Kozlovskaya, 2001; @mrb DxDz @msb DxDz
Meju et al., 2003] can be incorporated into the objective
function via the a priori model parameters (mRr and mRs) [16] Using equations (6), (7), and (8), the linearized
and off-diagonal elements of their covariance matrix CRR, weighted equivalent of equation (5) is stated as
but this issue will not be addressed in this paper.
[13] Our objective function (5) is nonlinear since the DC 
resistivity and seismic forward problems as well as the 1
min FL ðmr ; ms Þ ¼ 2 ½dr  Ar mr T C1
rr ½dr  Ar mr
cross-gradients constraint are nonlinear. Our solution to b
equation (5) is thus achieved by linearization. For the DC 1
þ 2 ½ds  f s ðm0s ÞAs ðms  m0s Þ T
resistivity forward problem, we adopt the approach of b
Pérez-Flores et al. [2001] founded on the nonlinear integral C1
ss ½ds  f s ðm0s ÞAs ðms  m0s Þ
equations for electromagnetic inverse problems of Gómez-
Treviño [1987]. In this approach, the DC resistivity forward þ a2r mTr DT Dmr þ a2s mTs DT Dms
   
computation simplifies to a linear problem of the form mr  mRr T 1 mr  mRr
þ CRR
f r ðmr Þ ffi Ar mr ; ð6Þ ms  mRs ms  mRs
 
mr  m0r
where Ar defines the Jacobian matrix, evaluated using the subject to t ðm0r ; m0s Þ þ B ¼ 0; ð10Þ
ms  m0s
formulas given by Pérez-Flores et al. [2001]. Note that Ar is
independent of any particular mr thus it is computed only
once and is recurrently used in later calculations. We have where b is an auxiliary damping factor.
adopted this approximate model for the 2-D DC resistivity [17] The solution of equation (10) is determined, using
forward problem because of its computational speed; Lagrange multipliers [see e.g., Menke, 1984; Tarantola,
however, a more rigorous 2-D forward model can be used 1987], by solving the system of equations:
in the cross-gradients inversion approach. ( " #)
[14] The relationship between seismic travel time and @ X
n X2n
slowness in the functional fs(ms) is nonlinear since the FL þ 2 lj bj;k ðmk  m0k Þ þ t ðm0 Þj ¼0
@mi j¼1 k¼1 ð11Þ
trajectory of the seismic ray path in the subsurface depends
upon the slowness [see Hole, 1992], but it can be linearized for i ¼ 1; 2n
by considering a small perturbation of the slowness about a
reference slowness model, m0s. The resulting linearized
expression is X
2n  
bp;j mj  m0j þ t ðm0 Þp ¼ 0 for p ¼ 1; n; ð12Þ
f s ðms Þ ffi f s ðm0s Þ þ As ðms  m0s Þ: ð7Þ j¼1

3 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

where lj are the required Lagrange multipliers, bj,k are the the data as b is varied while a substage minimization
coefficients of B and n refers to the total number of cells in ensures structural similarity by seeking the solution satisfy-
the rectangular model grid. In these expressions mi (i = 1, n) ing the cross-gradients constraint for a constant b.
are the elements of mr, mi (i = n + 1, 2n) are the elements of [20] To track the evolution of the misfit at every iterative
ms and the zero subscript refers to the initial model and step the rms values of the normalized residuals of the fitted
thereafter to the previous iterate. data are computed as
[18] Using matrix notation, let us define the variables
2 3 2 3 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
mr m0r ½dr  f r ðmr Þ T C1 ddr ½dr  f r ðmr Þ
rmsr ¼ ; ð15Þ
m¼4 5; m0 ¼ 4 5 nr
ms m0s

and + = {li} where i = 1, n. The solutions to equations (11) sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi


and (12), after some algebra, are ½ds  f s ðms Þ T C1 dds ½ds  f s ðms Þ
rmss ¼ ; ð16Þ
ns
   1
T 1

L ¼ BN1
1 B BN1 n2  Bm0 þ t 0 ; ð13Þ
where nr and ns are the number of DC resistivity and seismic
travel time data, respectively. The convergence of the main
m ¼ N1 1 T
1 n2  N1 B L: ð14Þ iterative process is based on reducing equations (15) and
(16) to their desired values of unity, while that of the
Here, substage minimization is given by the relative differences
between the parameters of the models at two consecutive
2 3
1 T 1 iterations
A C A þ a2r DT D þ C1 0
6 b2 r rr r RRr 7
N1 ¼ 6
4
7
5 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 T 1 uX
0 A C
2 s ss
A s þ a 2 T
s D D þ C1
RRs u n ðmi  m0i Þ2
b u
t m20i þ e
2 3 convr ð%Þ ¼ 100 i¼1 ; ð17Þ
1 T 1 n
Ar Crr fdr g þ C1
RRr mRr
6 b2 7
and n2 ¼ 6
4
7
5
1 T 1 1
A C f d s  f s ðm 0s Þ þ A m
s 0s g þ C m
RRs Rs vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
b 2 s ss u 2n
u X ðmi  m0i Þ2
u
u
are the matrices of the normal equations for the regularized ti¼nþ1 m20i þ e
convs ð%Þ ¼ 100 ; ð18Þ
problem, and t0 is t(m0r, m0s) in equation (8). The first term n
on the right-hand side of equation (14) corresponds to the
regularized (but structurally unlinked) least squares solution where the parameters involved are described in equations
while the second term (N11BT+) is the linking contribution (11) and (12) and e is a small positive number that serves to
from the cross-gradients constraint. Note that N1 should be prevent division by zero.
positive definite to use the above equations and this is
assured by the use of Twomey-Tikhonov-type [Tikhonov
and Arsenin, 1977; Twomey, 1977] derivative regularization 4. Separate Versus Joint 2-D Inversion of
measures and the covariances CRRr and CRRs of the a priori Synthetic Data
models mRr and mRs, respectively. [21] We tested our implemented seismic forward model-
[19] In the iterative solution process, the search for the ing algorithm and the adopted DC resistivity code using the
optimal solution is started with an initial model m0 (not simple 2-D model depicted in Figure 2. The model consists
necessarily the same as mR and preferably a smooth model), of two rectangular boxes embedded in a half-space and
which is then updated at every iterative step. Important having the following properties:
considerations are the stability and convergence character-
istics of the algorithm and the role played by the regular-
Half -space r ¼ 100 ohm m; Vp ¼ 1000 m=s;
ization factors (ar and as) and auxiliary damping factor b.
The main purpose of the regularization is to stabilize the
Box 1 r ¼ 10 ohm m; Vp ¼ 2000 m=s;
process and avoid local minima. Note that in equation (14)
the values of the linearized cross-gradients are effectively
Box 2 r ¼ 1000 ohm m; Vp ¼ 2000 m=s:
equal to zero. However, the original (i.e., nonlinear) con-
straint in equation (5) is dependent on the convergence of
the model; it will tend toward zero as the model converges [22] The values of the model parameters were chosen so
to a stable solution. We carried out comparative tests in that one box had host-target contrasts of the same sign for
which b, ar, and as were initially assigned large values and both resistivity and velocity and the other box had opposite
then decreased in successive iterations and found that contrasts. The model was discretized into rectangular cells
satisfactory convergence was obtained when predetermined of variable sizes in the area of interest, as shown in Figure 2,
values of ar and as are held fixed while b was allowed to and the bordering zones were padded with cells of increas-
decrease (until its threshold value of unity is reached). We ing dimensions up to a lateral and depth extent of 500 m, a
use a two-stage minimization process; the main iteration fits necessary feature for the DC resistivity modeling.

4 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

Figure 3. Dipole-dipole resistivity response for the test


model computed using the code of Dey and Morrison
[1979] constituting our synthetic data for 2-D inversion.

Figure 2. Theoretical test model consisting of two


rectangular boxes embedded in a uniform host medium.
(a) Resistivity model. (b) Seismic velocity model. The plotted
cells correspond to those used to define our test model.

[23] We computed the resistivity response of the model,


using the forward codes of Pérez-Flores et al. [2001] and
that of Dey and Morrison [1979], for a dipole-dipole array
recording from position 100 up to +100 m with an
electrode spacing (a) of 10 m and interval separation (na)
of 10 m up to 180 m. The maximum differences between
the computed responses were no more than 4.8% in the
logarithm of apparent resistivity, which we consider appro-
priate for the level of noise expected in typical field data.
Henceforth, we use the data depicted in Figure 3 generated
using the Dey and Morrison [1979] code as the synthetic
test data for inversion.
[24] In the case of the forward computations of seismic
P-wave travel times, the responses of the test model were
obtained for a deployment consisting of reversed profiles
running from 100 m up to 100 m with geophones located
every 5 m and seismic sources every 50 m along the profile.
The synthetic data from our code are compared with those
obtained using C. Zelt’s code in Figure 4a. The two data
sets are in accord, any minor differences arising mainly
from truncating and rounding off the numerical data values.
We plotted the ray paths in Figure 4b, which resulted from Figure 4. (a) Computed travel time responses for the test
the direct, critically refracted and diffracted waves. The ray model using the seismic code implemented in our inversion
tracing in the seismic model reflects, in the conventional scheme (crosses) and using the code of C. Zelt [Zelt and
procedure, the sensitivity of the synthetic data to the Barton, 1998] (solid lines). (b) Ray tracing showing the
different cells in the model. Notice that the rays are coverage associated with the test data.

5 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

concentrated in the upper surface of the boxes. Thus the


synthesized data will have the necessary sensitivity to
define the upper part of the boxes but not their depth extent.
[25] We inverted the synthetic data without adding ran-
dom noise, but uniform standard deviations of 1% and
0.1 ms were used for the data covariance matrices for the
apparent resistivity and travel times, respectively. The data
were first inverted simultaneously but without the cross-
gradients constraint, thus mimicking the conventional ap-
proach of separately inverting the resistivity and seismic
data sets. The inversion process was then repeated incorpo-
rating the cross-gradients constraint ( joint inversion proper)
with the same regularization parameters for comparison.

4.1. Simultaneous Inversion of Synthetic Data Without


Cross-Gradients Constraint
[26] For the separate inversion, the process was initiated
using a half-space model with a resistivity of 100 ohm m and
a seismic velocity of 1000 m/s. The multiplication factor b
was selected as 31.62, as larger values ended in flat models
and it was set to decrease in 6 steps down to 1 (the threshold
value required to fit the data at the level defined by their
standard deviations). The final seismic and resistivity mod-
els after six iterations did not change by more than 2% from
those of the penultimate iterate, based on equations (17) and
(18), and fit the data satisfactorily according to equations (15)
and (16). The misfits of the final models are rmsr = 0.403 for
the resistivity data and rmss = 1.329 for the seismic data. The
fit of the responses of these optimal models to the test data is
shown in Figures 5a and 5b.
[27] The optimal resistivity and velocity models obtained
are shown in Figures 6a and 6b. Both techniques recovered
parts of the original model as allowed by their respective
data coverage, the regularization (smoothness) measures,
and model errors (differences between responses from the
implemented forward modeling codes and those used to
generate the test data). Note that the resultant resistivity
model recovers the gross structure of the test model better
than the seismic model as the DC resistivity data contain Figure 5. (a) Computed response (contours) for the
measurements from large enough electrode spacing to sense resistivity model obtained from separate data inversion.
deeper structures enabling the definition of the bottom of The point values are the corresponding relative differences in
the boxes. The detailed sampling of the top of the boxes by percent (compare Figure 3). (b) Comparison of travel time
the seismic rays resulted in a relatively flat structure that test data (crosses) and computed travel times (solid lines)
resembles better the upper parts of the boxes and their from the separate inversion model. Selected residual values
lateral extents. However, the basal parts and vertical walls (in ms) are annotated above their corresponding positions.
of the boxes are poorly resolved in the seismic model and
are mainly controlled by the combined effects of smooth-
ness measures and the a priori model. multiplication factor b at every iteration of the joint inver-
sion process is shown in Figure 7.
4.2. Joint Inversion of Synthetic Data [29] The final resistivity model obtained by joint inver-
[28] To evaluate the performance of our joint cross- sion is shown in Figure 6c, and the corresponding seismic
gradients inversion algorithm, we inverted the same data velocity model is shown in Figure 6d. These jointly recon-
sets using the same initial models (m0 = mR = half-space) structed resistivity and velocity models show improved
and regularization factors (ar, as, b). The models recovered features of interest. These are as follows: (1) The structural
after 6 iterations satisfied the target misfit with rmsr = 0.535 aspect of the boxes (defined by the shapes of the contours)
and rmss = 1.989 and the prescribed convergence criterion is very similar in both models, as required by our joint
(convr < 2% and convs < 2%). The misfit values are inversion algorithm. (2) The bottom of each box now
somewhat larger than those of the separate inversion exper- becomes defined in the seismic model, this being an indirect
iment because of the cross-gradients constraint. However, contribution propagated from the resistivity model by the
the data fit of the optimal models is not significantly cross-gradients constraint. (3) The top and lateral extent of
different from that of Figures 5a and 5b and is therefore the boxes are better delineated; the resistivity model has
not shown here. The evolution of the rms misfit and the improved on the round shape in the separate inversion

6 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

Figure 6. Recovered models for the hypothetical two-box example. Shown are (a) resistivity and
(b) seismic velocity models from separate inversion. Also shown are the (c) resistivity model and
(d) seismic velocity model from joint inversion.

(compare Figure 6a) while the slanted top margin of the


seismic model (compare Figure 6b) has become horizontal,
yielding the desired shape of the boxed structure with only
minor vertical smearing. This smearing may have resulted
from the differences of the adopted DC resistivity approach
propagated into the estimated model. (4) The values of the
seismic velocity in the boxes are closer to the true values
than those of the model from separate inversion (compare
Figure 6b).
[30] We can compare the computed values of the cross-
gradients function for the two optimal models from the
separate 2-D inversion (Figure 8a) with those for the joint
inversion models (Figure 8b). Figure 8a reveals zones of
structural incongruence in the separate inversion models
(see Figures 6a and 6b). It is pertinent to mention that the
cross-gradients function tends to decay to zero in areas far
from those jointly constrained by both sets of data. This
implies that one insensitive method will allow its counter-
part to continue defining the proper model on its own until
neither data set constrains the model, whereafter it gradually
tends toward the a priori (usually flat) model. The values of
the cross-gradients function for the joint 2-D models
(Figure 8b) are more than 1 order of magnitude closer to
zero than those computed using the results of separate
inversion and more randomly distributed. This suggests that Figure 7. RMS values of the normalized residuals at every
the cross-gradients constraint leads to joint inversion models iteration for the test experiment with joint inversion.

7 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

velocity. We selected a half-space resistivity of 200 ohm m,


this being the average value of the observed apparent
resistivity data. The initial seismic velocity was set to
5000 m/s, this being typical of the velocity of granodiorites.
Thus the slowness model ms is an order of magnitude
different from mr for the starting models.
[33] The subsurface was discretized using irregular cells
with a finer mesh in the areas sampled by the field measure-
ments and padded with thicker cells outside the areas of
interest. To make the most of the resolution capability of the
data and enhance the expected geological structure (espe-
cially at shallow depth), we selected an unnormalized dis-
crete Laplacian operator (D), which is independent of the cell
spacing. This allows the models to produce horizontally
elongated structures near the surface. The factors ar = 1
and as = 10, selected after some trials, define the relative
weight of this operator. The difference in these values is
comparable to the 1 order of magnitude difference in the
initial resistivity and velocity model parameters, which we
find appropriate to keep the level of smoothness in both
models similar. In this example we found it appropriate to
vary the multiplication factor b from 100 to 1 in 8 predeter-
mined steps. To obtain the maximum structural concordance
between the models, each substage minimization was re-
quired to continue for 10 iterations unless a global conver-
gence (convr and convs) of 2% is achieved.
[34] Figure 9 shows the convergence characteristics of the
algorithm for 8 iterations for these field data sets. Notice

Figure 8. Mapping of the cross-gradients function.


(a) Results for the models obtained by separate inversion
(Figures 6a and 6b). (b) Results for the models obtained by
the cross-gradients joint inversion (Figures 6c and 6d). See
color version of this figure in the HTML.

with enhanced common structural attributes relative to the


conventional separate 2-D inversion schemes.

5. Joint Inversion of Field Data


[31] We have inverted field data sets from collocated DC
resistivity and seismic refraction experiments along a 200 m
long profile at a test site in Quorn in England. The
exploration target is the upper 40 m of the subsurface.
The geology consists of a highly fractured and eroded
granodiorite bedrock that is successively overlain by Mer-
cian Mudstone and heterogeneous glacial drift deposits
[Meju et al., 2003]. The resistivity data consist of six
vertical electric soundings (VES) using the Schlumberger
configuration with maximum current electrode spread
length (AB/2) of 90 m. The seismic data consist of a set
of forward and reversed refraction profiles over a 165 m
transect with a 2 m geophone spacing and 4 shot points. As Figure 9. RMS values of the normalized residuals and
observational errors were not available for the Quorn data multiplication factor b per iteration for the joint inversion
sets, uniform standard deviations of 5% and 1 ms were used experiment for Quorn data. Note the gradual decrease of the
for the data covariance matrices for the apparent resistivity misfit at each iteration as b is varied. The inset shows the
and travel times, respectively. convergence measures (equations (17) and (18)) and the trend
[32] We started our joint inversion of the Quorn data of the root-mean-square values of the cross-gradients
using an initial model with constant resistivity and seismic function (i.e., substage minimization) for the last iteration.

8 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

Figure 10. Evolution of the joint inversion process. Shown are the resultant resistivity and velocity
models for each iteration. Note the gradual development of common structural features in both sets of
models during the process. See color version of this figure in the HTML.

that the rmsr and rmss values per iteration were gradually local minima but also controlled the development of im-
reduced as the b factor was varied. Figure 10 shows the portant common features in both models. As a result, the
resistivity and seismic models obtained for each iteration. final resistivity and seismic models (at iteration 8) show a
Note that this gradual process of joint solution reconstruc- remarkable structural resemblance, which can be gleaned
tion not only reduced the possibility of being trapped in from the shape of the contours in Figure 10.

9 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

log of resistivity and 0.5 to 0.2 s/km in slowness). In


comparison, Figure 12b shows the cross-gradients values
of our jointly inverted models. Note the similar values of t
in both the top and the middle portions of the model where
there is good data coverage. However, there are some
differences between the resistivity and velocity images in
the bottom part of our joint models. These correspond to
areas with the smallest gradients of the slowness and log
resistivity parameters and where the data coverage is low. In
addition to the gained structural conformity of the cross-
gradients inversion models, they can be shown to highlight
important resistivity-velocity trends [see Gallardo and
Meju, 2003], and can complement petrophysical measure-
ments for improved subsurface characterization.

6. Conclusions
[37] We have demonstrated in this paper that the cross-
gradients function offers a quantitative means to evaluate
structural similarities between two smooth images. The
cross-gradients function can be applied to geophysical
images from heterogeneous geological environments, and
can be used to provide a link between two seemingly
disparate geophysical models. We have incorporated it as
a constraint into a simple 2-D joint inversion procedure for
Figure 11. The fit between the Quorn field data and the DC resistivity and seismic travel time data from collocated
responses of the seismic and resistivity models resulting surveys.
from joint inversion. (a) DC apparent resistivity data. The [38] The results from inversion experiments using syn-
positions of the DC soundings on the survey line are thetic and field data revealed several interesting features of
indicated in the top right-hand corners of the plots. our joint inversion procedure. The algorithm yields resis-
(b) Seismic travel time data. In both cases, the cross tivity and seismic models that are consistent with the
symbols represent field data, while the solid lines represent experimental data and have improved structural similarity.
the computed response of the final models of Figure 10. Importantly, this conformity is reached without forcing or

[35] Figures 11a and 11b show the fit obtained for the
resistivity and seismic data in the final iteration of this joint
inversion exercise. The normalized values of the residuals
for the final iteration were rmsr = 1.004 for the apparent
resistivity data and rmss = 0.864 for the travel time data.
This suggests that the data were fitted to the expected levels
even though there were no actual field data errors available.
Note that there are some larger misfits in particular seg-
ments of the travel time data that can only be fitted by more
complex models or are manifestations of 3D effects on the
data. In the final iteration, some parts of the models still
show slight differences from that of the previous iterates
(defining a convergence factor of 3%), but they do not
substantially change the fit of the data and will only
marginally affect the cross-gradients function.
[36] The same field data sets were previously inverted
separately by Meju et al. [2003] and the resultant models
were interpreted as having similar structural features. How-
ever, we show that there are subtle differences in structure
between the separate models of Meju et al. [2003] as
evidenced by the values of the computed cross-gradients
function of their models (Figure 12a). The largest values of t
in the top 2 meters are due to the high gradients shown by
both sets of model parameters (from 4 to 1 in the log of Figure 12. Comparison of cross-gradients values com-
resistivity and from 3 to 0.1 s/km in slowness). On the other puted for the (a) separate inversion and (b) joint inversion of
hand, the smallest t values at the bottom of the resistivity the Quorn field data. See color version of this figure in the
and seismic models relate to less abrupt gradients (2 to 3 in HTML.

10 of 11
B03311 GALLARDO AND MEJU: JOINT 2-D RESISTIVITY AND SEISMIC INVERSION B03311

assuming the form of the relationship between electrical Hole, J. A. (1992), Non-linear high-resolution three-dimensional travel time
tomography, J. Geophys. Res., 97, 6553 – 6562.
resistivity and seismic velocity. The algorithm is also Kozlovskaya, E. (2001), Theory and application of joint interpretation of
flexible in the sense that it can admit boundaries that are multimethod geophysical data, Ph.D. dissertation, Univ. of Oulu, Oulu,
only constrained by one method and not the other. The Finland.
inversion scheme is robust since it showed adequate con- Lines, L. R., A. K. Schultz, and S. Treitel (1988), Cooperative inversion of
geophysical data, Geophysics, 53, 8 – 20.
vergence characteristics. We conclude that our algorithm is Loke, M. H., and R. D. Barker (1995), Least-squares deconvolution of
an effective procedure for jointly interpreting resistivity and apparent resistivity pseudosections, Geophysics, 60, 1682 – 1690.
seismic data sets from complex two-dimensional environ- Meju, M. A., L. A. Gallardo, and A. K. Mohamed (2003), Evidence for
correlation of electrical resistivity and seismic velocity in heterogeneous
ments. A drawback of our present inversion scheme is that it near-surface materials, Geophys. Res. Lett., 30(7), 1373, doi:10.1029/
uses an approximate 2-D resistivity forward model for 2002GL016048.
computational efficiency. For an alternative and more rig- Menke, W. (1984), Geophysical Data Analysis: Discrete Inverse Theory,
orous approach we recommend using the approximate Academic, San Diego, Calif.
Musil, M., H. R. Maurer, and A. G. Green (2003), Discrete tomography and
method for initial convergence and then switching to a joint inversion for loosely connected or unconnected physical properties:
proper resistivity finite difference solution for the final Application to crosshole seismic and georadar data sets, Geophys. J. Int.,
iterations. 153, 389 – 402.
Pérez-Flores, M. A., S. Méndez-Delgado, and E. Gómez-Treviño (2001),
[39] Although we have demonstrated the cross-gradients Imaging low-frequency and DC electromagnetic fields using a simple
method for the 2-D case and using only resistivity and linear approximation, Geophysics, 66, 1067 – 1081.
seismic data, the method should also be extendable to any Sasaki, Y. (1989), Two-dimensional joint inversion of magnetotelluric and
dipole-dipole resistivity data, Geophysics, 54, 254 – 262.
other geophysical methods and in three-dimensions. Scott, J. B. T., R. D. Barker, and S. Peacock (2000), Combined seismic
refraction and electrical imaging, paper presented at 6th Meeting of En-
[40] Acknowledgments. We acknowledge SUPERA-ANUIES in vironmental and Engineering Geophysics, Environ. and Eng. Geophys.
Mexico for a scholarship award to Luis A. Gallardo. We also thank Colin Soc., Bochum, Germany.
Zelt and Marco A. Pérez-Flores for making their codes available. We are Tarantola, A. (1987), Inverse Problem Theory, Elsevier Sci., New York.
grateful to K. Whaler, S. K. Park, and J. Berryman for their incisive Tikhonov, A. N., and V. Y. Arsenin (1977), Solutions of Ill-Posed Problems,
comments, which improved the clarity of this paper. John Wiley, Hoboken, N. J.
Tillmann, A., and T. Stöcker (2000), A new approach for the joint inversion
References of seismic and geoelectric data, paper presented at 63rd EAGE Conference
and Technical Exhibition, Eur. Assoc. of Geosci. and Eng., Amsterdam.
Berge, P. A., J. G. Berryman, H. Bertete-Aguirre, P. Bonner, J. J. Roberts, Twomey, S. (1977), An Introduction to the Mathematics of Inversion in
and D. Wildenschild (2000), Joint inversion of geophysical data for site Remote Sensing and Indirect Measurement, Elsevier Sci., New York.
characterization and restoration monitoring, LLNL Rep. UCRL-ID- Vidale, J. E. (1990), Finite-difference calculation of traveltimes in three-
128343, Proj. 55411, Lawrence Livermore Natl. Lab., Livermore, Calif. dimensions, Geophysics, 55, 521 – 526.
Dey, A., and H. F. Morrison (1979), Resistivity modelling for arbitrarily Zelt, C. A., and P. J. Barton (1998), Three-dimensional seismic refraction
shaped two-dimensional structures, Geophys. Prospect., 27, 106 – 136. tomography: A comparison of two methods applied to data from the
Gallardo, L. A., and M. A. Meju (2003), Characterization of heterogeneous Faeroe Basin, J. Geophys. Res., 103, 7187 – 7210.
near-surface materials by joint 2D inversion of DC resistivity and seismic Zhang, J., and F. D. Morgan (1996), Joint seismic and electrical tomography,
data, Geophys. Res. Lett., 30(13), 1658, doi:10.1029/2003GL017370. paper presented at EEGS Symposium on Applications of Geophysics to
Gallardo-Delgado, L. A., M. A. Pérez-Flores, and E. Gómez-Treviño Engineering and Environmental Problems, Environ. and Eng. Geophys.
(2003), A versatile algorithm for joint 3-D inversion of gravity and mag- Soc., Keystone, Colo.
netic data, Geophysics, 68, 949 – 959.
Gómez-Treviño, E. (1987), Nonlinear integral equations for electromag- 
netic inverse problems, Geophysics, 52, 1297 – 1302. L. A. Gallardo and M. A. Meju, Department of Environmental Science,
Haber, E., and D. Oldenburg (1997), Joint inversion: A structural approach, Lancaster University, Lancaster LA1 4YQ, UK. (lgallard@cicese.mx;
Inverse Problems, 13, 63 – 77. m.meju@lancaster.ac.uk)

11 of 11

Potrebbero piacerti anche