Journal of Applied Geophysics 79 (2012) 2737

Contents lists available at SciVerse ScienceDirect

Journal of Applied Geophysics

journal homepage: www.elsevier.com/locate/jappgeo

3D stochastic joint inversion of gravity and magnetic data

Pejman Shamsipour , Denis Marcotte, Michel Chouteau
Dpartement des Gnies Civil, Gologique et des Mines, cole Polytechnique de Montral, C.P. 6079, Succursale Centre-Ville, Montreal, Qubec, Canada H3C 3A7

a r t i c l e

i n f o

Article history:
Received 20 July 2011
Accepted 19 December 2011
Available online 3 January 2012
Keywords:
3D
Joint inversion
Multivariate
Cokriging

a b s t r a c t
A novel stochastic joint inversion method based on cokriging is applied to estimate density and magnetic
susceptibility distributions from gravity and total magnetic eld data. The method fully integrates the physical relations between densitygravity, on one hand, and magnetic susceptibilitytotal magnetic eld, on the
other hand. As a consequence, when the data are considered noise-free, the responses from the inverted density and susceptibility data exactly reproduce the observed data. The required density and magnetic susceptibility auto- and cross covariance are assumed to follow a linear model of coregionalization (LCM). The
parameters of the LCM are estimated from vv plot tting of the gravity and total magnetic experimental
covariances. The model is tested on two synthetic cases and one real data set, the Perseverance mine (Quebec,
Canada). Joint inversions are compared to separate inversions. The joint inversions better recover the known
models in the synthetic cases. With the real data set, better denition and location of the mineralized lenses
are achieved by joint inversion.
2012 Elsevier B.V. All rights reserved.

1. Introduction
Interpretation of geophysical data needs to bring together different types of information to make the proposed model geologically
realistic. Multiple data sets can reduce ambiguity and non uniqueness
present in separate geophysical data inversions. Potential eld surveys (gravity and magnetic) are suitable candidates for joint inversion as they are among the most economical methods in geophysics.
Moreover, gravity and magnetic elds are closely related quantities,
yet complementary. Most often, the magnetic minerals are dense so
they cause also gravimetric anomalies. However, many dense rocks
are not magnetic. Also, magnetic eld display higher frequency variations than gravity data, so they are often more effective than gravity
at resolving shallow or complex structures. The method presented
exploits the complementary nature of the two potential elds to provide joint inversions that are better than separate inversion of each
eld. For this purpose, a 3D stochastic inversion approach based on
cokriging is used.
3D joint inversion of magnetic and gravity data was rst described
in the work by Zeyen and Pous (1993). They applied a priori information to reduce ambiguity of potential eld inversion and interpolation. Gallardo-Delgado et al. (2003) proposed the method based on
optimization to minimize the joint data mist. The stochastic formulation for joint inversion is presented by Bosch et al. (2006). The
method is based on lithology discrimination and classication. This
method allows for joint inversion of gravity, magnetic and other a
Corresponding author.
E-mail addresses: pejman.shamsipour@polymtl.ca (P. Shamsipour),
Denis.Marcotte@polymtl.ca (D. Marcotte), Chouteau@geo.polymtl.ca (M. Chouteau).
0926-9851/$ see front matter 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.jappgeo.2011.12.012

priori information to provide a model for major layers and properties

inside layers. Gravity and magnetic data were inverted also jointly by
Pilkington (2006) where the model consists of an interface separating
two layers with different, but constant, densities and magnetization.
In Fregoso and Gallardo (2009), the authors propose using cross gradients for 3D inversion of gravity and magnetic data. Most of these
methods directly, or indirectly, adapt the Tarantola and Valette (1982)
techniques to the generalized minimization problem in a Bayesian
framework. Their approach builds on the earlier works of Franklin
(1970) who was the rst in geophysics to introduce the concept of
stochastic inversion. Geostatistical methods in geophysical inversion
were applied by Asli et al. (2000), Gloaguen et al. (2005, 2007),
Giroux et al. (2007) and Hansen et al. (2006). Bosch et al. (2006) also
used Monte Carlo techniques in gravity inversion for generating a posterior probability density function describing acceptable models.
Chasseriau and Chouteau (2003) advocate 3D inversion of gravity
data using an a priori model of covariance. Shamsipour et al. (2010,
2011) proposed geostatistical techniques of cokriging and conditional
simulation for the separate three-dimensional inversion of gravity and
magnetic data respectively, including geological constraints.
We propose a new method, based on cokriging, for the joint inversion of gravity and magnetic data. The method is a useful extension of
Shamsipour et al. (2010, 2011). After the description of the method, it
is tested on two synthetic models: a model consisting of two prisms
buried in a homogeneous background and a model with stochastic
distribution of parameters. The results illustrate the capability of the
method to improve the inverted model compared to the separate
inverted models with either gravity or magnetic data. In addition,
the survey data collected over the Perseverance and Equinox deposits
in the Matagami mining camp (Quebec, Canada) are studied. The

28

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

recovered 3D joint-models better identify the known deposits within

the survey area than do the separate inverted models. They also provide useful information to analyze and re-interpret the geology of the
area under study.
2. Methodology
2.1. Forward modeling
The purpose of forward modeling is to compute the magnetic eld
T and the gravimetric response g at the surface due to a susceptibility
distribution in the sub-surface and a density distribution respec
tively. The magnetization vector m can be obtained as a vector sum:

Cgg Gg C Gg C0 ;
CgT Gg C GT ;

where C0 in Eqs. (6) and (7) are diagonal matrices containing the socalled nugget effect in geostatistics. They represent the data noise
variances which can be constant for each observation or vary for different observation subsets (e.g. if different data acquisitions are
used). Note that in Eq. (8) C0 = 0 as noise in the gravity and total
magnetic eld observations can be considered uncorrelated. We also
have:
CT GT C ;

Cg Gg C ;

10

where H is Earth's magnetic eld and m r is the remanent component.

When remanent magnetization is absent, m is in the direction

of the

Earth's eld and can be obtained simply as: m H . We assume

there is no remanent magnetization in this paper, an assumption
deemed valid for the Archean rocks in Quebec.
The most common method used to evaluate the potential responses
is to break down the 3D domain into geometrically simple bodies having constant susceptibilities or densities. In our case, and for the sake
of simplicity, the domain studied is divided into a nite number of rectangular prisms of uniform parameters. It should be noted that this classic method can consume memory space for a ne domain discretization.
Note however that the method can easily be generalized to accommodate for prisms of different sizes or for other shapes than prisms.
Closed form solutions for magnetic forward modeling were rst
presented by Bhattacharyya (1964), and later simplied in Rao and
Babu (1991) into a form more suitable for fast computer implementation. For closed form of gravity modeling also several equivalent analytical forms were proposed (see Li and Chouteau (1998) for a critical
review of different formulas). Using Eq. (1) in Shamsipour et al.
(2010) and Eq. (3) in Shamsipour et al. (2011), the response at the
observation point (x0, y0, z0) of all the prisms included within the subsurface model is the sum of the contribution of each prism:

C T GT C ;

11

m H m r

T0

m
X

T i x; y; z

g i x; y; z

i1

g0

m
X
i1

Considering that there are n total magnetic eld or gravity observations and m rectangular prisms, the preceding relationship can also
be written in the matrix form:

C g Gg C :

12

The cokriging systems for joint inversion are:



CT;T
Cg;T

CT;g
Cg;g



1
1


CT;
;
Cg;

13

and


1 T 1 g;


Cg;g
CT;g


Cg;T
CT;T



2
2

14



Cg;

;
CT;

15

2 g 2 T:

16

Note that the two cokriging systems are solved simultaneously as

the left member in the cokriging Eqs. (13) and (15) is identical:


CT;T
Cg;T

CT;g
Cg;g



1
1

2
2

CT;
Cg;


CT;
;
Cg;

17

In summary, after estimating susceptibility and density covariance

matrices C and C, the cross-covariance matrix C and noise
matrices C0, and after computing geometrical matrices GT and Gg we
can calculate all the required covariances in Eqs. (13) and (15) and
then compute the cokriging weights 1, 2, 1 and 2. Finally, we use
these weights in Eqs. (14) and (16) to estimate density and susceptibility distributions.
For each cokriging system, the vector of cokriging variances is
obtained from:

Tn1 GT nm m1 ;



2
2
T~
ck diag I~ C
R

g n1 Ggnm m1

h
i
where ~ T T ; T is the coefcient matrix, C~ R is the right member
matrix where I is the identity matrix and 2 is the variance of density
or susceptibility.
The inverse matrix calculation to solve Eqs. (13) and (15) can be
done by Gauss elimination and by singular value decomposition
(SVD) for small inverse problems or by preconditioned CGA for larger
problems.
Note that separate inversion of density or magnetic susceptibility
are also obtained by cokriging using the following reduced systems:

with GT and Gg, matrices of geometric terms (or kernel matrices).

2.2. Joint inversion using cokriging
We assume the susceptibility and density covariance matrices C
and C are known based on the parametrization described in
Section 2.3 and the estimation procedure sketched in Section 2.4.
From Eqs. (4) and (5), parameters and observation eld covariance
matrices are linearly related:
T

CTT GT C GT C0 ;

18

CTT T CT ;

19

Cgg g Cg

20

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

3. Synthetic cases

and the separate inversions are:



21

22

T T;
g g:

2.3. Linear model of coregionalization (LCM)

Matheron (1965) introduced the proportional covariance model
(which he named intrinsic model). This model assumes all crosscovariances and simple-covariances are proportional to an elementary
covariance function. The LCM is the sum of proportional covariance
models. We assume the model covariance matrices C,, C, and C,
can be parametrized as an LCM. Therefore, one has (Chils and Delner,
1999):


 X
s
C; h C; h

Bk C k h
C; h C; h

23

k1

with

Bk

b;;k
b;;k

b;;k
b;;k


24

In this model, the two simple covariances and the cross-covariance

at distance h are linear combinations of a set of s basic structures (e.g.
nugget effect, spherical structure, etc.) evaluated at the same distance
h and having a unit sill. The Bk matrices (k = 1s) are 2 2 matrices of
coefcients invariant with distance. A sufcient condition for the
model to be valid is that each matrix Bk be positive semidenite. The
LCM model we used has 2 basic structures, one nugget and one spherical. The range parameter of the spherical structure and the Bk matrices of coefcients are identied by the vv plot method described in
the next section. The extension of the LCM to more than 2 variables
is immediate. Note that the theoretical simple correlation coefcient
between density and susceptibility can be computed from the LCM coefcients:
sk1 b;;k
Corr ; r



sk1 b;;k sk1 b;;k

29

25

2.4. Multivariate vv plot

The multivariate vv plot is the extended version of the vv plot
method (Asli et al., 2000; Gloaguen et al., 2005; Shamsipour et al.,
2010, 2011). Here, the Bk matrices of coefcients and elementary
structures parameters Ck(h) are optimized iteratively such that the
theoretical covariances Cg,g, CT,T and Cg,T computed from the LCM
(Eqs. (23), (6)(8)) show good correlations with their experimental
counterparts obtained as TT T, gg T and gT T. For all pairs of data points,
the experimental cross-products are compared to the theoretical
covariances computed with the current LCM model. To gain stability
in the optimization, pairs of points are binned according to the theoretical values found in Cg,g, CT,T and Cg,T. The same pairs are regrouped
and averaged in the experimental cross-products' counterparts. In
essence, the multivariate vv plot approach follows the steps for variogram computation except that the binning is done directly on the
theoretical direct and cross-covariances instead of the euclidean distance as with standard variogram computation.

The proposed method is tested on two sets of synthetic data: i) a

compact model and ii) a stochastic model. In the rst synthetic example Corr(, )=1, in the second, Corr(, )=0.7.
3.1. Compact model
The model consists of two cubes of 3 3 2 m buried in a homogeneous background. The center of the cubes are at locations (x = 5.5,
y = 7.5, z = 4) and (x = 9.5, y = 7.5, z = 2). The 3D domain is divided
into 15 15 10 = 2250 cubic prisms with dimensions 1 1 1 m.
Therefore, each cube consists of 18 prisms. Distances are given in
meters but are arbitrary. The susceptibility of all the prisms inside
the cubes is 0.005 SI and their density is 1000 kg/m 3. The model density and susceptibility distributions at section y = 7.5 m are shown in
Fig. 2(a) and (b). Because of the constant densities and susceptibilities, r(, )=1. The surface gravity anomaly is shown in Fig. 1(a).
We assume that the total magnetic eld is only available in a borehole
at (x = 5.5, y = 7.5). With an Earth's inducing eld with a strength of
50 000 nT, inclination angle I = 75 and declination angle D = 45, the
susceptibility model produces the borehole total magnetic anomaly
shown in Fig. 1(b).
Using inversion by cokriging (Shamsipour et al., 2010), we invert
gravity data at the surface to estimate the density distribution. The
densities at section y = 7.5 m are shown in Fig. 2(c). Then, we invert
the total magnetic eld at the borehole to estimate the susceptibility
distribution. The results at section y = 7.5 m are shown in Fig. 2(d).
Shamsipour et al. (2010, 2011) have shown that the calculated gravity and total magnetic eld from the inverted models match perfectly
the observed values when C0=0 in Eqs. (6) and (7). Note that independent inversion of gravity data does not recover the depth of the
cubes. Also, as the total magnetic eld data is measured in a single
borehole, the inverted susceptibility eld shows radially symmetric
variations.
The joint inversion of gravity and magnetic data are shown in
Fig. 2(ef) for the density and the susceptibility respectively. Both
density and susceptibility distributions better reproduce the synthetic
model compared to the separate inversions. In particular, the depths
of the cubes are close to the model.
The observed gravity and total magnetic elds are perfectly reproduced by both the separate and joint inversion models.
3.2. Stochastic model
The 3D domain is divided into 15 15 10 = 2250 cubic prisms.
Susceptibilities and densities are generated on 1 m 1 m 1 m prisms
using LU simulation (Chils and Delner, 1999). The LCM has a single
spherical structure with range a = 5 m, density variance of 50000
(kg/m 3) 2 and susceptibility variance of 0.0001 (SI) 2. The correlation
Corr(, ) is 0.7, thus, the magnetic eld and gravity eld are not
caused by the exact same source.
The generated density and susceptibility distributions at section
y = 7.5 m are shown in Fig. 4(a) and (b) respectively. The Earth's inducing eld has a strength of 50 000 nT and inclination and declination are
I = 45 and D = 70 respectively. Using the susceptibility values, we calculate the synthetic total magnetic eld data at the surface from susceptibility values (Fig. 3(a)). Using the density values, we calculate the
synthetic gravity data along a borehole, at x = 7.5 m and y = 7.5 m,
shown in Fig. 3(b). We add data noises as nugget effects (gravity
C0 = .000001 and total magnetic C0 = 1000) to both total magnetic
eld data and gravity data. From now on, we assume the gravity and
total magnetic eld data are known and we invert them to estimate
the density and susceptibility distributions.
Fig. 4 shows the results of the separate (c)(d) and joint density
and susceptibility inversions (e)(f). Using the multivariate vv plot

30

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

(b)

Gravity (mGal)

(a)

0
1

0.02

3
0.015

depth (m)

Gravity (mGal)

0.03
0.02
0.01

0.01

15

15

7
10

10

0.005

x(m)

y(m)
0

10

500

g (mGal)

0
500
Total field (nT)

Fig. 1. (a) Gravity data at the surface and (b) total magnetic eld in the borehole.

method, the adjusted LCM is spherical with a = 4.8 m and variances

C = 0.00012 (SI) 2 and C = 50000 (kg/m 3) 2 and correlation coefcient
0.75, values close to those of the parameters used to simulate the

(a)

data. In the separate density inversion (c), the model shows circular
symmetry around the borehole as this was the only available data.
Density and susceptibility distributions in the joint inversions

(b)

Initial (density)

Initial (susceptibility)

x 10

800
2
600

4
6

400

200

10
5

10

4
6

8
10

0
15 (kg/m3 )

x(m)
Inverted data (density)

(c)

depth(m)

depth(m)

10

15

(SI)

x(m)
Inverted data (susceptibility)

(d)

x 10

800
2
600

4
6

400

200

10
5

10

4
6

8
10

15 (kg/m 3 )

x(m)
Joint Inverted data (density)

(e)

depth(m)

depth(m)

10

15

800

400

200

10
5

10

x(m)

15 (kg/m3)

x 10

2
600

depth(m)

depth(m)

(SI)

x(m)
(f) Joint Inverted data (susceptibility)

4
6

8
10
5

10

15

(SI)

x(m)

Fig. 2. (a) Model density distribution, (b) model susceptibility distribution, (c) inverted density values using the surface gravity data, (d) inverted susceptibility values using the
surface total magnetic eld, (e) inverted density values using both the surface total magnetic eld and gravity data, and (f) inverted susceptibility values using both the surface
total magnetic eld and gravity data. All the results are shown in section y = 7.5 m.

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

(a)

31

(b)
Total field (nT)
0
200
1
150
2

100
50

200
50

0
200

z(m)

Total field (nT)

400

100
15

400
15

150

5
6
7

200

10

10
250

300

x(m)

y(m)

10
0.01

nT

0.01

Gravity (mGal)
Fig. 3. (a) Synthetic total magnetic eld data at surface and (b) synthetic gravity data collected in the borehole.

depth(m)

(b)

Initial (density)

6
200

8
10
10

6
0.01
0.02

10

15 (kg/m )

x(m)
Inverted data (density)

(c)

0.01

400
5

Initial (Susceptibility)
2

200

depth(m)

(a)

10

15

x(m)
Inverted data (Susceptibility)

(d)

(SI)

x 10

100

2
0

4
6

100

depth(m)

depth(m)

10

200
10

10
5

10

15 (kg/m3)

x(m)
joint Inverted data (density)

(e)

10

15

x(m)
Joint Inverted data (Susceptibility)

(f)

(SI)

x 10

100
0

4
6

100

10

200
10

depth(m)

depth(m)

10
5

10

x(m)

15 (kg/m3)

10

15

(SI)

x(m)

Fig. 4. (a) Model density distribution, (b) model susceptibility distribution, (c) inverted density distribution using borehole gravity data, (d) inverted susceptibility distribution
using surface total magnetic eld data, (e) inverted density distribution using joint inversion of borehole gravity data and surface total magnetic eld data, and (f) inverted susceptibility distribution using joint inversion of borehole gravity and surface total magnetic eld data. All the results are shown for section y = 7.5 m.

32

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

(e)(f) are closer to the models (a)(b) than the separate inversions
(c)(d). The estimated densities have especially been improved close
to the surface because of the extra information provided by the total
magnetic eld data. On the other hand, the susceptibilities have
been improved at depths because of the extra information from gravity data along the borehole. The joint density and susceptibility inversions are more correlated with the true model compared to separate
inversions (0.29 and 0.46 versus 0.12 and 0.41). Also, we note that
the correlation between the separate density and susceptibility
inverted elds is close to zero while the jointly inverted elds have
a correlation of 0.77, a value close to the model correlation (0.7).
In Fig. 5(a), we compare the observed gravity with the calculated
gravity using the estimated densities by joint inversion. They show a
correlation of 0.9985. In (b), the observed total magnetic eld versus
the calculated total magnetic eld using the estimated susceptibilities
by joint inversion is shown (correlation 0.9926). In both cases, we see
a good correlation between observed and calculated elds. The difference is because of the nugget effect we added to the observed data.
We repeat the same synthetic case but the inversion is done using
only gravity and magnetic anomalies on the surface. Fig. 6 shows the
inversion results obtained. The joint density and susceptibility inversions are slightly more correlated with the true model compared to
separate inversions (0.44 and 0.43 versus 0.43 and 0.39).The two
joint inversions show some differences compared to the separate
inversions. A low density and susceptibility zone appears in the joint inversion between x = 1 3 m and z = 3 8 m. Compared to the synthetic model in Fig. 4(a) and (b), one notes that this corresponds to a real
low density and average susceptibility zone found in the model.
4. Case study
4.1. Application to survey data
The survey data was collected over the area of the Perseverance
mine located in the Matagami region in Quebec, Canada. The area of
the 2001 total magnetic eld survey extends from longitude 77 47
46" W to 77 46 59" W and from latitude 49 45 13" N to 49 45
59" N, where 441 magnetic ground measurements with about 50-m
spacing are available. 44 ground gravity data were collected using a
Scintrex CG5 gravity meter (accuracy .005 mGal) in a restricted area
of the Perseverance mine (anomaly A in Fig. 7). The data were provided
by Xstrata Zinc inc., owner and operator of the Perseverance zinc mine.
The Perseverance mine was discovered in April 2000 and entered
production in 2008. It has three major massive sulde deposits

known as Perseverance, Perseverance west and Equinox. They have

been shown in Fig. 7 with A, B and C respectively. Magnetite and pyrrhotite are often linked with the VMS deposits. For Perseverance
lenses, even though pyrrhotite can have some local contribution,
the magnetic anomalies are mainly caused by the magnetite content.

4.1.1. Inversion
The residual magnetic eld anomaly was obtained by subtracting
the IGRF (International Geomagnetic Reference Field) from the measured total eld (Fig. 8(a)). The regional gravity eld was obtained
by calculating the rst order trend from the Bouguer anomaly data.
Subtracting the regional from the Bouguer anomaly produced the
residual anomaly (Fig. 8(b)) ranging between 0.04 and 0.4 mGal.
The inversion domain is divided into nx = 26 by ny = 26 by nz = 10
cubes of dimension 40 m 40 m 30 m. Therefore, the whole domain
is 1040 m 1040 m 300 m and the total number of prisms is m=6760.
Sections from separate and joint inversions for susceptibilities
are shown in Fig. 9(a)-(b) for an Eastwest vertical section at y =
5515450 m, and in Fig. 10(a(d) for horizontal sections at z = 150 m
and z = 200 m respectively. The LCM was adjusted using the vv
plot method. The basic structure is spherical with a = 130 m and sills
C =.0003 (SI)2 and C = 17000 (kg/m3) 2 respectively. The tted correlation obtained using multivariate vv plot is 0.89.
Two strong and one weak susceptibility highs are seen in the
plane sections (z = 150 and z = 200 m) for both separate and joint inversions. The highs are well correlated with the three Perseverance
deposits indicated by letters A, B and C. The largest susceptibility
high is correlated with Perseverance deposit (A), the smallest susceptibility high is correlated with Perseverance west deposit (C) and the
third susceptibility high is related to Equinox deposit (B). The susceptibilities at cross section y = 5515450 m (Fig. 9(b) show the extension
of Perseverance deposit (A) and Equinox deposit (B) at depth.
Depths of the deposits are the following: Perseverance (30
200 m), Perseverance-West (100200 m) and Equinoxe (90275 m)
(Allard, 2011, personal communication). The joint inversion better
recovers than separate inversions of the top and bottom location of
Perseverance A deposit in the susceptibility inverted elds.
Since the gravity data is available only on the top of deposit A, we
now perform inversion on a smaller region around this deposit (shown
in Fig. 8(b)). The results at sections x =299,380 m, y= 5,515,500 m
and z = 150 m are shown in Figs. 11, 12 and 13 respectively. The joint
inversions appear obviously deeper and stronger than the separate inversions, hence allowing a better delimitation of deposit A.

(b)

(a)
0.01

300

Estimated magnetic field

0.008

Estimated gravity

0.006
0.004
0.002
0
0.002
0.004
0.006

100
0
100
200
300

0.008
0.01
0.01

200

0.005

0.005

Observed gravity

0.01

400
400 300 200 100

100

200

300

Observed magnetic field

Fig. 5. (a) Observed gravity versus the calculated gravity using the estimated densities by joint inversion and (b) observed total magnetic eld versus the calculated total magnetic
eld using the estimated susceptibilities by joint inversion.

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

(a)

(b)

Inverted data ()

33

3
x 10

Inverted data ()

100
0

depth(m)

depth(m)

100

200

300

10

10
5

10

15 (kg/m )

x(m)
joint Inverted data ()

(c)

5
10
5

10

15

x(m)
Joint Inverted data ()

(d)

(SI)

3
x 10

100
2
0

100

200

300

10
5

10

15 (kg/m3)

x(m)

depth(m)

depth(m)

6
5
8
10

10
5

10

15

(SI)

x(m)

Fig. 6. (a) Inverted density distribution using surface gravity data, (b) inverted susceptibility distribution using surface total magnetic eld data, (c) inverted density distribution
using joint inversion of surface gravity data and surface total magnetic eld data, and (d) inverted susceptibility distribution using joint inversion of surface gravity and surface total
magnetic eld data. All the results are shown for section y = 7.5 m.

5. Discussion
One advantage of the inversion by cokriging is that the magnetic
eld and gravity eld do not need to be caused by the exact same
source. In one of our synthetic case studies, a correlation of only 0.7
exists between the density and susceptibility variations. Similarly, in
the real case study, the correlation was found to be 0.89 between
the density and susceptibility elds. This correlation is obtained by
adjusting experimentally the gravity and total magnetic eld data
available, using the vv plot approach. Therefore, the joint inversion

will use the auxiliary information only when this correlation is present in the data. For example, Corr, = 0 would result in joint inversions identical to the separate inversions.
The proportional covariance model has the following general
screening property: when all the variables are measured at the same
points the cokriging estimator boils down to the kriging estimator
based on that variable alone. Translated in the present context, this
means that the joint inversion is identical to the separate inversions
when data are located at the same points, whatever the value for Corr,.
This property does not hold strictly when an LCM model different from

Fig. 7. Geology map of Matagami camp with Perseverance mine highlighted. Three major deposits are: A: Perseverance, B: Perseverance west and C: Equinox.

34

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

(a)

(b)

Total field (nT)

Residual gravity (mGal)

1200

0.3

Residual gravity (mGal)

Total field (nT)

1000

1000

800
C

500
600

0
5.516

2.996

5.5158

x 106

0.3
0.2
0.1

0.2

5.516
5.5158

0.15

x 106 5.5156

400

2.996
2.994

2.994

5.5156
5.5154

5.5154

2.992

y(m)

0.25

x 10

200

y(m)

2.99

5.5152

2.988

5.515 2.986

x 10

0.1

2.99

5.5152

2.988
5.515 2.986

x(m)

2.992

x(m)
0.05

g(mGal)

T(nT)

Fig. 8. (a) Total magnetic eld residual anomaly in the study area. (b) Residual gravity anomaly above Perseverance mine (anomaly A).

for any kriging or cokriging) in the presence of noise. When C0 is present in Eqs. (6) and (7), the computed gravity and magnetic eld
anomalies do not match perfectly the observed data, however the observed discrepancies are compatible with the noise level described by
C0. Note that when C0 is set to 0, the computed gravity and magnetic
eld match perfectly the observed data (as long as there are more
parameters in the model than observations), irrespective of the data
being really noise-free or not. Conversely, when C0 0, the computed
gravity and magnetic eld match the observed data more or less
depending of the noise variance level. The vv plot helps determine
the values of C0, but other knowledge like the expected level of
noise for a given data acquisition could also be used to estimate this
matrix.
The proposed non-iterative joint inversion method based on cokriging is computationally efcient. For mid-size problems like the
case study, the inversion is done on standard laptop in less than 10
minutes. For larger problems, the bottleneck is the available memory
to store the required covariance matrices. However, practical solutions exist. A simple one is to adopt an isotropic covariance function
with a nite range (e.g. spherical model) with the sparse matrix coding. Moreover, for an isotropic model, the parameter covariance

the proportional model is used (e.g. a spherical component plus a nugget effect). However, the gain expected from joint inversion in this case
is not expected to be important. When data are not exactly co-located,
joint inversions can improve over separate inversions even if both
data are taken at the surface as was shown in the second synthetic example (Fig. 6).
Obtaining a good t between the theoretical and experimental
covariances by the vv plot method is sometimes challenging. Nevertheless, it is possible to obtain from geological knowledge, at least
crude estimates of the variances of each variable and of the ranges
along the principal geologic directions. This could dictate good starting parameter values for the LCM model. As for any optimization, it
can be useful to consider various starting values to help identify the
presence of local minima. The adjustment is usually best done semiautomatically. The structure of LCM model (number and type of elementary structures) is selected rst manually after a few trials with
different elementary structures. Then, the LCM parameters are adjusted automatically and modied at the end, when judged necessary
(e.g. unrealistic parameter values upon stopping criteria). The method accommodates white noise (observation errors) in the form of a
nugget effect C0 matrix (see Eqs. (6) and (7)). Results are stable (as

5
0
300

400

600

x(m)

800

60

10

180

200

Joint Inverted data ()

x 10
A

60

depth(m)

(b)

Inverted data ()

5
1000 (SI)

depth(m)

(a)

x 10
A

10
5

180

0
300

200

400

600

800

5
1000 (SI)

x(m)

Fig. 9. (a) Inverted susceptibility distribution using surface total magnetic eld data and (b) inverted susceptibility distribution using joint inversion of gravity data and total magnetic eld
data. All the results are shown for an eastwest cross-section at y = 5,515,450 m from the 3D inverted models.

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

(a)

(b)
Inverted data ()

x 10
15

1000

10

800

600

400

10

y(m)

600
400

x 10
15

800

y(m)

Inverted data ()

1000

200

200
5
200 400 600 800 1000 (SI)

5
200 400 600 800 1000 (SI)

x(m)

x(m)

(c)

Joint Inverted data ()

x 10
15

1000

(d)

Joint Inverted data ()

800

y(m)

10

600
400

5
B

200

10

600
400

x 10
15

1000

800

y(m)

35

200
5
200 400 600 800 1000 (SI)

5
200 400 600 800 1000 (SI)

x(m)

x(m)

Fig. 10. Panels (a) and (b) show inverted susceptibility distribution using surface total magnetic eld data. Panels (c) and (d) show inverted susceptibility distribution using joint
inversion of gravity and total magnetic eld data. Right hand sub-plots results are shown for horizontal section z = 150 m and left hand sub-plots results are shown for horizontal
section z = 200 m.

matrices are all circulant matrices. Hence only the covariances of the
rst row need really be computed and stored (see Nowak et al.,
2003).
The approach uses simple cokriging. Therefore, it is based on the
modeling of variations of density and susceptibility relative to their
means. Hence, the effect of the mean density and susceptibility have
to be removed from the gravity and total magnetic eld before the

(a)

vv plot method is applied, i.e. we assume the data represents zeromean relative anomalies. It is possible to adapt the approach to
account, prior to the inversion, for a spatially variable mean. As examples, one could impose a vertical drift on the density mean, or simply
assume that different geological units have well distinct density and
susceptibility means. Because the inversion estimates contrasts for
the properties, one must ensure that adding back the mean provides

(b)

Inverted data ()

Inverted data ()
0.03

50

depth(m)

depth(m)

100
120

240

80

160

240

x(m)

0.01
240

320 (kg/m3)

(c)

0.02

120

0
80

160

240

320

(SI)

x(m)

(d)
joint Inverted data ()

Joint Inverted data ()

0.03

50
0

240

80

160

x(m)

240

320 (kg/m3)

depth(m)

depth(m)

100
120

0.02

120

0.01
240

80

160

240

320

(SI)

0.01

x(m)

Fig. 11. (a) Inverted density distribution using gravity data, (b) inverted susceptibility distribution using surface total magnetic eld data, (c) inverted density distribution using
joint inversion of gravity data and total magnetic eld data, and (d) inverted susceptibility distribution using joint inversion of gravity data and total magnetic eld data. All the
results are shown for an eastwest cross-section at X = 299,380 m from the 3D inverted models.

36

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

(a)

(b)

Inverted data ()

Inverted data ()
0.02

60

120

40
20

240

depth(m)

depth(m)

80
120

0.01
0

240

0
80

160

240

x(m)

(c)

320 (kg/m3)

80

240

320

(SI)

0.01

x(m)

(d)

joint Inverted data ()

160

Joint Inverted data ()

100

0.02

120

60
40
20

240

depth(m)

depth(m)

80
120

0.01
0

240

0
80

160

240

x(m)

80

320 (kg/m3)

160

240

320

(SI)

0.01

x(m)

Fig. 12. (a) Inverted density distribution using gravity data, (b) inverted susceptibility distribution using surface total magnetic eld data, (c) inverted density distribution using
joint inversion of gravity data and total magnetic eld data, and (d) inverted susceptibility distribution using joint inversion of gravity data and total magnetic eld data. All the
results are shown for an eastwest cross-section at Y = 5,515,500 m from the 3D inverted models.

positive values. This is of course the case in the synthetic examples

and the case study where the inverted contrasts represent only a
small fraction of the mean.
6. Conclusion
We present a joint inversion method based on the cokriging geostatistical approach for the joint 3D inversion of total magnetic eld and

(a)

gravity data. Tests on synthetic data show the potential advantages

of this method. The proposed joint inversion applied on data from
Matagami mining camp (Perseverance mine) proves better agreement
with the known geology. The method can be easily applied to other geophysical variables and it can accommodate more than two variables.
Greater gains from the joint inversion with respect to the separate
inversions are expected when the data are not co-located and when
the correlation between the physical properties is strong.

(b)

Inverted data ()
320

0.015
0.01

50

160

240

y(m)

240

y(m)

Inverted data ()
320

80

0.005

160

0
0.005

80
50
80

160

240

0.01

320 (kg/m3)

80

x(m)

(c)

240

320

(SI)

x(m)

(d)

joint Inverted data ()

Joint Inverted data ()

0.015

320

320

0.01

50

160

240

y(m)

240

y(m)

160

0.005

160

0
0.005

80

80
50
80

160

x(m)

240

320 (kg/m3)

0.01
80

160

240

320

(SI)

x(m)

Fig. 13. (a) Inverted density distribution using gravity data, (b) inverted susceptibility distribution using surface total magnetic eld data, (c) inverted density distribution using
joint inversion of gravity data and total magnetic eld data, and (d) inverted susceptibility distribution using joint inversion of gravity and total magnetic eld data. All the results
are shown for horizontal section z = 150 m.

P. Shamsipour et al. / Journal of Applied Geophysics 79 (2012) 2737

Acknowledgments
We would like to express sincere thanks to Michel Allard from
Xstrata Zinc for providing the Matagami data, and for his guidance
and support. We also thank Colin G. Farquharson and the anonymous
reviewer for their comments which helped to improve the manuscript.
References
Asli, M., Marcotte, D., Chouteau, M., 2000. Direct inversion of gravity data by cokriging.
In: Kleingeld, W., Krige, D. (Eds.), 6th International Geostatistics Congress, Capetown,
South Africa, pp. 6473. Cape Town, South Africa.
Bhattacharyya, B.K., 1964. Magnetic anomalies due to prism-shaped bodies with arbitrary polarization. Geophysics 29, 517531.
Bosch, M., Meza, R., Jimnez, R., Hnig, A., 2006. Joint gravity and magnetic inversion in
3D using Monte Carlo methods. Geophysics 71, G153G156.
Chasseriau, P., Chouteau, M., 2003. 3D gravity inversion using a model of parameter
covariance. Journal of Applied Geophysics 52, 5974.
Chils, J.P., Delner, P., 1999. Geostatistics: Modeling Spatial Uncertainty. Wiley.
Franklin, J.N., 1970. Well posed stochastic extensions of Ill posed linear problems. Journal
of Mathematics 31, 682716.
Fregoso, E., Gallardo, L.A., 2009. Cross-gradients joint 3D inversion with applications to
gravity and magnetic data. Geophysics 74, L31L42.
Gallardo-Delgado, L., Perez-Flores, M., Gmez-Trevio, E., 2003. A versatile algorithm
for joint 3D inversion of gravity and magnetic data. Geophysics 68, 949959.
Giroux, B., Gloaguen, E., Chouteau, M., 2007. bh_tomo a Matlab borehole georadar 2D
tomography package. Computers and Geosciences 33, 126137.

37

Gloaguen, E., Marcotte, D., Chouteau, M., Perroud, H., 2005. Borehole radar velocity
inversion using cokriging and cosimulation. Journal of Applied Geophysics 57,
242259.
Gloaguen, E., Marcotte, D., Giroux, B., Dubreuil-Boisclair, C., Chouteau, M., Aubertin, M.,
2007. Stochastic borehole radar velocity and attenuation tomographies using
cokriging and cosimulation. Journal of Applied Geophysics 62, 141157.
Hansen, T.M., Journel, A.G., Tarantola, A., Mosegaard, K., 2006. Linear inverse Gaussian
theory and geostatistics. Geophysics 71, R101R111.
Li, X., Chouteau, M., 1998. Three-dimensional gravity modeling in all space. Surveys in
Geophysics 19, 339368.
Matheron, G., 1965. Les variables rgionalises et leur estimation. Une Application de la
Thorie des Functions Alatories aux Sciences de la Nature. Masson, Paris.
Nowak, W., Tenkleve, S., Cirpka, O., 2003. Efcient computation of linearized crosscovariance and auto-covariance matrices of interdependent quantities. Mathematical Geology 35, 5366.
Pilkington, M., 2006. Joint inversion of gravity and magnetic data for two-layer models.
Geophysics 71, L35L42.
Rao, D.B., Babu, N.R., 1991. A rapid method for three-dimensional modeling of magnetic anomalies. Geophysics 56, 17291737.
Shamsipour, P., Marcotte, D., Chouteau, M., Keating, P., 2010. 3D stochastic inversion of
gravity data using cokriging and cosimulation. Geophysics 75, I1I10.
Shamsipour, P., Chouteau, M., Marcotte, D., 2011. 3D stochastic inversion of magnetic
data. Journal of Applied Geophysics 73, 336347.
Tarantola, A., Valette, B., 1982. Generalized nonlinear inverse problems solved using
the least squares criterion. Reviews of Geophysics and Space Physics 20, 219232.
Zeyen, H., Pous, J., 1993. 3-D joint inversion of magnetic and gravimetric data with a
priori information. Geophysical Journal International 112, 244256.