This is a postprint of: 

Fast non-linear gravity inversion in spherical coordinates 

with application to the South American Moho 
by Leonardo Uieda and Valéria C. F. Barbosa (2017) 
 
This article has been accepted for publication in ​Geophysical Journal International. 
©: 2016 The Authors. Published by Oxford University Press on behalf of the Royal 
Astronomical Society.  
 
Version of record:​ ​doi.org/10.1093/gji/ggw390 

Source code:​ The Python code that produces the results presented here is 
available under a BSD 3-clause open-source license at 
github.com/pinga-lab/paper-moho-inversion-tesseroids 

Moho model and data: ​The Moho depth model for South America and the data 
used to generate it are available under a Creative Commons Attribution 4.0 
International (CC-BY) license at d
​ oi.org/10.6084/m9.figshare.3987267 

Citation:  
Uieda, L., and V. C. F. Barbosa (2017), Fast nonlinear gravity inversion in spherical 
coordinates with application to the South American Moho, ​Geophysical Journal 
International​, 208(1), 162-176, doi:​10.1093/gji/ggw390 
 Geophysical Journal International
Geophys. J. Int. (2017) 208, 162–176 doi: 10.1093/gji/ggw390
Advance Access publication 2016 October 17
GJI Gravity, geodesy and tides

Fast nonlinear gravity inversion in spherical coordinates with

application to the South American Moho

Leonardo Uieda1,2 and Valéria C.F. Barbosa2

1 Universidade do Estado do Rio de Janeiro, Rio de Janeiro, Brazil. E-mail: [email protected]
2 Observatório Nacional, Rio de Janeiro, Brazil

Accepted 2016 October 13. Received 2016 October 7; in original form 2016 March 24

SUMMARY
Estimating the relief of the Moho from gravity data is a computationally intensive nonlinear
inverse problem. What is more, the modelling must take the Earths curvature into account
when the study area is of regional scale or greater. We present a regularized nonlinear gravity

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

inversion method that has a low computational footprint and employs a spherical Earth ap-
proximation. To achieve this, we combine the highly efficient Bott’s method with smoothness
regularization and a discretization of the anomalous Moho into tesseroids (spherical prisms).
The computational efficiency of our method is attained by harnessing the fact that all matrices
involved are sparse. The inversion results are controlled by three hyperparameters: the reg-
ularization parameter, the anomalous Moho density-contrast, and the reference Moho depth.
We estimate the regularization parameter using the method of hold-out cross-validation. Ad-
ditionally, we estimate the density-contrast and the reference depth using knowledge of the
Moho depth at certain points. We apply the proposed method to estimate the Moho depth for
the South American continent using satellite gravity data and seismological data. The final
Moho model is in accordance with previous gravity-derived models and seismological data.
The misfit to the gravity and seismological data is worse in the Andes and best in oceanic
areas, central Brazil and Patagonia, and along the Atlantic coast. Similarly to previous results,
the model suggests a thinner crust of 30–35 km under the Andean foreland basins. Discrepan-
cies with the seismological data are greatest in the Guyana Shield, the central Solimões and
Amazonas Basins, the Paraná Basin, and the Borborema province. These differences suggest
the existence of crustal or mantle density anomalies that were unaccounted for during gravity
data processing.
Key words: Inverse theory; Satellite gravity; Gravity anomalies and Earth structure; South
America.

eas, resulting in large areas (e.g. forests and mountains) devoid of

1 I N T RO D U C T I O N
data.
The Mohorovičić discontinuity (or Moho) that marks the transi- Estimating Moho depth from gravity data is a nonlinear in-
tion from the crust to the mantle, is studied almost exclusively verse problem. One can generalize this problem of estimating the
through indirect geophysical methods. The two main geophysi- depths of an interface separating two media, such as the sediment-
cal methods used to estimate the depth of the Moho are seis- basement interface of a sedimentary basin or the crust-mantle inter-
mology, with both natural and controlled sources, and gravime- face (Moho). Several methods have been developed over the years to
try. With the advent of satellite gravimetry missions like GRACE solve this inverse problem, for example, Bott (1960); Barbosa et al.
and GOCE, gravity-derived crustal models can be produced in re- (1997, 1999a,b); Barnes & Barraud (2012); Leão et al. (1996);
gional or global scales (e.g. Reguzzoni et al. 2013; van der Meijde Martins et al. (2010, 2011); Oldenburg (1974); Reguzzoni et al.
et al. 2013, 2015). New spherical harmonic gravity models that (2013); Santos et al. (2015); Silva et al. (2006, 2014), to name a few.
use these satellite observation, like GOCO5S (Mayer-Guerr et al. Solving the inverse problem is computationally demanding because
2015), provide almost homogeneous data coverage in difficult to it requires the construction of large dense matrices and the solution
access regions traditionally poor in terrestrial data. An example of large linear systems. As a result, some authors search for ways to
is South America, where seismological and terrestrial gravity data increase the computational efficiency of this class of inverse prob-
are traditionally concentrated around urban centres and coastal ar- lem. Bott (1960) proposed a method based on iteratively applying

162 
C The Authors 2016. Published by Oxford University Press on behalf of The Royal Astronomical Society.
 Fast nonlinear gravity inversion 163

corrections to a starting estimate based on the inversion residu-

als. The algorithm is fast because it bypasses the construction and
solution of linear systems and only involves forward modelling.
Oldenburg (1974) showed that the fast FFT-based forward mod-
elling of Parker (1973) could be rearranged to estimate the relief.
Barnes & Barraud (2012) use a form of adaptive discretization to
compute the Jacobian, or sensitivity, matrix. For each data point, the
discretization will be progressively coarser the further way from the
point. This reduces the matrix and, consequently, the linear systems
to a sparse form that can be solved efficiently. Recently, Silva et al.
(2014) extended and generalized the original method of Bott (1960)
and Santos et al. (2015) used this extension to estimate a basement
relief with sharp boundaries.
A spherical Earth approximation is preferred when estimating the
Moho depth from gravity data in continental and global scale stud-
ies. Wieczorek & Phillips (1998) developed a spherical harmonic
equivalent of the Parker-Oldenburg FFT algorithm and applied it to Figure 1. Sketch of the stages in gravity data correction and the discretiza-
estimate the crustal structure of the Moon. Reguzzoni et al. (2013) tion of the anomalous Moho relief using tesseroids. (a) The Earth and the
use a spherical Earth approximation to estimate the global Moho re- measured gravity at point P (g(P)). (b) The Normal Earth and the calculated
lief using data from the GOCE satellite mission. Another approach normal gravity at point P (γ (P)). zref is the depth of the Normal Earth Moho.
(c) The gravity disturbance (δ(P)) and the corresponding density anomalies

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

is to use non-spectral (space domain) gravity inversion methods.
after removal of the normal gravity: topography, oceans, crustal and man-
Many such methods were developed for estimating the basement
tle heterogeneities, and the anomalous Moho. (d) The Bouguer disturbance
relief of a sedimentary basin (e.g. Barbosa et al. 1997, 1999a,b; (δ bg (P)) after topographic correction and the remaining density anomalies.
Martins et al. 2010, 2011; Sun & Li 2014). These methods approx- (e) All density anomalies save the anomalous Moho are assumed to have
imate the sedimentary pack by a set of juxtaposed right-rectangular been removed before inversion. (f) The discretization of the anomalous
prisms. The top of the prisms coincide with the Earth’s surface Moho in tesseroids. Grey tesseroids will have a negative density contrast
and the prisms’ thicknesses represent the depths to the basement while red tesseroids will have a positive one.
and are the parameters to be estimated in the inversion. The use
of rectangular prisms implies a planar Earth approximation and
may not be adequate for depth-to-Moho estimates in continental- the Silva et al. (2014) Gauss-Newton formulation of the method of
or global-scale study. A straightforward way to circumvent this hin- Bott (1960). We use tesseroids to discretize the anomalous Moho
drance is to adapt one of the methods developed for rectangular and the adaptive discretization algorithm of Uieda et al. (2016) for
prisms to use tesseroids (spherical prisms). One of the difficul- the forward modelling. The stability of the inversion is achieved
ties of this approach is that the forward problem for a tesseroid through smoothness regularization. In order to maintain the com-
must be solved numerically. Two alternatives proposed in the litera- putational efficiency of Bott’s method, we exploit the sparse nature
ture to the numerical solution are Taylor series expansion (Heck & of all matrices involved in the computations. We employ a variant
Seitz 2007; Grombein et al. 2013) and the Gauss-Legendre Quadra- of GCV known as hold-out cross-validation (Kim 2009) to estimate
ture (Asgharzadeh et al. 2007). Numerical experiments by Wild- the regularization parameter. Additionally, we estimate the density-
Pfeiffer (2008) suggest that the Gauss-Legendre Quadrature (GLQ) contrast and reference level simultaneously in a second validation
offers superior results. However, the GLQ suffers from numeri- step using knowledge of the Moho depth at a few points, similarly
cal instability when the computation point is close to the tesseroid to Silva et al. (2006) and Martins et al. (2010). Finally, we apply the
(Asgharzadeh et al. 2007). To overcome the numerical instability, proposed method to estimate the Moho depth for South America
Li et al. (2011) proposed an adaptive discretization algorithm which using gravity data from the GOCO5S model (Mayer-Guerr et al.
was later improved upon by Uieda et al. (2016). 2015) and the seismological data of Assumpção et al. (2013).
In any gravity inversion for estimating the relief of an inter-
face, two hyperparameters control the inversion results: the density-
2 METHODOLOGY
contrast between the two media and the reference level around which
the interface undulates. The reference level is the constant depth of In potential field methods, we must isolate the target anomalous
the Normal Earth Moho in the case of the anomalous Moho. For density distribution before modelling and inversion. In our case, the
regularized inversions, an additional hyperparameter is the regular- target is the relief of the real Moho undulating around a reference
ization parameter that balances the relative importance between the Moho. We do this by removing all other effects from the gravity
data-misfit measure and the regularizing function. The two most observations. The first correction is to remove the scalar gravity of
commonly used methods for estimating the regularization param- an ellipsoidal reference Earth (the Normal Earth), hereafter denoted
eter are the L-curve criterion and Generalized Cross Validation as γ . This effect is calculated on the same point P where the gravity
(GCV). Farquharson & Oldenburg (2004) provide for a thorough observation was made (Figs 1a and b). γ (P) is calculated using the
comparison of both methods. Estimating the density-contrast in a closed-form solution presented by Li & Götze (2001). The differ-
sedimentary basin context has been tackled by Silva et al. (2006) ence between the observed gravity at point P (g(P)) and Normal
and Martins et al. (2010) when the basement depth is known at a gravity at the same point is known as the gravity disturbance,
few points. To the authors’ knowledge no attempt has been made to
δ(P) = g(P) − γ (P). (1)
estimate the reference level.
We present a nonlinear gravity inversion to estimate the Moho The disturbance contains only the gravitational effects of density
depth in a spherical Earth approximation. Our method is based on distributions that are anomalous with respect to the Normal Earth
164 L. Uieda and V.C.F. Barbosa

the Moho is a nonlinear function of the parameters zk , k = 1, . . . ,

M,
di = f i (p), (3)
in which di is the ith element of the N-dimensional predicted data
vector d, p is the M-dimensional parameter vector containing the M
Moho depths (zk ), and fi is the ith nonlinear function that maps the
parameters onto the data. The functions fi are the radial component
of the gravitational attraction of the tesseroid Moho model.

2.2 Inverse problem

We wish to estimate the parameter vector p from a set of observed
gravity data do . The least-squares estimate is the one that minimizes
the data-misfit function
φ(p) = [do − d(p)]T [do − d(p)]. (4)
Figure 2. Sketch of a tesseroid (spherical prism) in a geocentric coordinate
system (X, Y, Z). Observations are made at point P with respect to its local Function φ(p) is nonlinear with respect to p. Thus, we can de-
north-oriented coordinate system (x, y, z). After Uieda (2015). termine its minimum using gradient-based iterative optimization
methods like Gauss–Newton or Steepest Descent. Such methods

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

start from an initial approximation to the model parameter vector
p0 and estimate a parameter perturbation vector p0 . The perturba-
tion vector is used to update p0 to p1 = p0 + p0 . This procedure
(see Fig. 1c). This includes all masses above the surface of the
is repeated until a minimum of function φ(p) (eq. 4) is reached.
ellipsoid (the topography), the mass deficiency of the oceans, the
For the Gauss-Newton method, the parameter perturbation vector
mass deficiency of sedimentary basins, crustal sources (e.g. igneous
at the kth iteration pk is obtained by solving the linear system
intrusions, lateral density changes, etc.), heterogeneities below the
upper mantle, and the effect of the difference between the real Moho Hk pk = −∇φ k , (5)
topography and the Moho of the Normal Earth.
To estimate the anomalous Moho relief from gravity data, we in which ∇φ and H are, respectively, the gradient vector and the
k k

must first isolate its gravitational attraction. Thus, all other gravita- Hessian matrix of φ(p).
tional effects must be either removed or assumed negligible. Here, The gradient vector and the Gauss-Newton approximation of the
we will remove the gravitational effect produced by the known to- Hessian matrix of φ(p) are, respectively,
pography and ocean masses to obtain the full Bouguer disturbance T
∇φ k = −2Ak [do − d(pk )], (6)
(Fig. 1d),
and
δbg (P) = δ(P) − gtopo (P). (2) T
Hk ≈ 2Ak Ak , (7)
We will also remove the gravitational effect of know sedimentary
in which Ak is the N × M Jacobian or sensitivity matrix whose
basins but assume that the effects of other crustal and mantle sources
elements are
are negligible. Thus, the only effect left will be that of the anomalous
Moho relief (Fig. 1e). The gravitational attraction of the topography, ∂ fi k
Aikj = (p ). (8)
oceans, and basins are calculated in a spherical Earth approximation ∂pj
by forward modelling using tesseroids (Fig. 2). The tesseroid effects
are calculated numerically using GLQ integration (Asgharzadeh
et al. 2007). The accuracy of the GLQ integration is improved by 2.3 Regularization
the adaptive discretization scheme of Uieda et al. (2016).
Nonlinear gravity inversions for estimating the relief of an interface
separating two media (like the Moho) are ill-posed and require addi-
tional constraints in the form of regularization (Silva et al. 2001). A
2.1 Parametrization and the forward problem common approach is to use the first-order Tikhonov regularization
(Tikhonov & Arsenin 1977) to impose smoothness on the solution.
We parameterize the forward problem by discretizing the anomalous The cost function for smoothness regularization is given by
Moho into a grid of Mlon × Mlat = M juxtaposed tesseroids (Fig. 1f).
The true (real Earth) Moho varies in depth with respect to the Moho θ(p) = pT RT Rp, (9)
of the Normal Earth. Hereafter we will refer to the depth of the
where R is an L × M finite-difference matrix representing L first-
Normal Earth Moho as zref (see Fig. 1b). If the true Moho is above
order differences between the depths of adjacent tesseroids.
zref , the top of the kth tesseroid is the Moho depth zk , the bottom
To transform the ill-posed inverse problem into a well-posed one
is zref , and the density-contrast (ρ) is positive (red tesseroids in
via Tikhonov regularization, we adopted the well-established pro-
Fig. 1f). If the Moho is below zref , the top of the tesseroid is zref , the
cedure of formulating a constrained inverse problem that is solved
bottom is zk , and ρ is negative (grey tesseroids in Fig. 1f).
by minimizing an unconstrained goal function
Considering that the absolute value of the density-contrasts of
the tesseroids is a fixed parameter, the predicted gravity anomaly of (p) = φ(p) + μθ(p), (10)
 Fast nonlinear gravity inversion 165

in which μ is the regularization parameter that controls the bal- 2.5 Combining Bott’s method, regularization
ance between fitting the observed data and obeying the smoothness and tesseroids
constraint imposed by the regularizing function θ (p) (eq. 9).
We propose a regularized version of Bott’s method to invert gravity
The goal function (p) is also nonlinear with respect to p and
data for estimating the depth of the Moho in spherical coordinates.
can be minimized using the Gauss-Newton method. The gradient
To adapt Bott’s method to spherical coordinates, we replace the
vector and Hessian matrix of the goal function are, respectively,
right-rectangular prisms in the forward modelling (d(pk ) in eq. 14)
T with tesseroids. The tesseroid forward modelling uses the adaptive
∇ k = −2Ak [do − d(pk )] + 2μRT Rpk , (11)
discretization algorithm of Uieda et al. (2016) to achieve accurate
and results. Furthermore, our formulation maintains the regularized so-
T lution for the Gauss-Newton method (eq. 13) but replaces the full
Hk = 2Ak Ak + 2μRT R. (12)
Jacobian matrix with the Bouguer plate approximation. Here, the
At the kth iteration, the parameter perturbation vector pk is Jacobian matrix is replaced by a diagonal matrix (eq. 15) whose
obtained by solving the linear equation system elements are invariant along successive iterations. Using this ap-
  proximation eliminates the cost of computing and storing the full
T T
Ak Ak + μRT R pk = Ak [do − d(pk )] − μRT Rpk . (13) N × M-dimensional Jacobian matrix Ak at each iteration (eq. 8).
Traditionally, the full Jacobian matrix is computed using a first-
Estimating the Moho depths using the above equations is compu- order finite difference scheme, which requires 2 × N × M for-
tationally costly because of two main factors: (1) the evaluation and ward modelling operations per iteration. Using eq. (15) requires N
storage of the dense N × M Jacobian matrix Ak and (2) the solution multiplications that need only be performed once. This provides a
of the resulting M × M equation system. In practice, the derivatives considerable speed gain.

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

in the Jacobian (eq. 8) are often calculated through a first-order Matrix arithmetic operations can be performed efficiently by
finite-difference approximation. Thus, evaluating Ak requires 2 × taking advantage of the sparse nature of matrices A and R (re-
M × N forward modelling operations for each iteration of the gradi- spectively, eqs 15 and 9). The same is true for solving the equa-
ent descent algorithm. These computations are performed for each tion system in the Gauss–Newton method (eq. 13). However, the
iteration of the optimization of the goal function (p). computational cost of forward modelling is still present. Particu-
larly, forward modelling using tesseroids is more computationally
intensive than using right-rectangular prisms because of the nu-
2.4 Bott’s method merical integration and adaptive discretization (Uieda et al. 2016).
Bott (1960) developed an efficient method to determine the depth of We show later in this article that sparse matrix multiplications
the basement of a sedimentary basin from gravity observations. The and solving the sparse linear system in eq. (13) account for less
method requires data on a regular grid of Nx × Ny = N observations. than 0.1 per cent of the computation time required for a single
The basement relief is then discretized into an equal grid of Mx × inversion.
My = M elements with Mx = Nx and My = Ny . Bott’s iterative The main advantage of our formulation is that it retains the effi-
method starts with an initial approximation of the basement depths ciency of Bott’s method while stabilizing the solution through the
p0 equal to the null vector. The method updates the approximation by well-established formalism of Tikhonov regularization. For exam-
calculating a parameter perturbation vector pk using the formula ple, the total variation approach used by Martins et al. (2011) could
potentially be implemented in a more straight forward manner than
do − d(pk ) done by Santos et al. (2015).
pk = , (14)
2π Gρ
in which G is the gravitational constant and ρ is the contrast
2.6 Estimating the inversion hyperparameters
between the density of the sediments and the reference density. The
iterative process stops when the inversion residuals rk = do − d(pk ) Parameters that influence the inversion result but are not estimated
fall below the assumed noise level of the data. directly in the inversion are known as hyperparameters. In the case
Silva et al. (2014) showed that Bott’s method can be formulated of our regularized Moho depth inversion, the hyperparameters are
as a special case of the Gauss–Newton method (eq. 5) by setting the the regularization parameter μ (eq. 10), the Moho density-contrast
Jacobian matrix (eq. 8) to ρ (eq. 15), and the depth of the Normal Earth Moho, or reference
level, zref (Fig. 1b).
A = 2π GρI, (15) We estimate these hyperparameters in two steps. First, we as-
where I is the identity matrix. In this framework, Bott’s method sume fixed values for zref and ρ and perform a hold-out cross-
uses a Bouguer plate approximation of the gravitational effect of validation procedure (Hansen 1992) to estimate an optimal value
the relief, di = 2π Gρ pi . The derivative of di with respect to for μ. Our investigations suggest that the optimal value of μ does
the parameter pi is 2π Gρ, thus linearizing the Jacobian matrix. not depend on the particular values of zref and ρ used. Second,
However, the nonlinearity of the predicted data d(pk ) is preserved. we use the estimated μ to perform a validation procedure to es-
One of the advantages of Bott’s method over the traditional timate zref and ρ. The final outcomes of both steps are the val-
Gauss–Newton or Steepest Descent is the elimination of the com- ues of the three hyperparameters and the final estimated Moho
putation and storage of the dense Jacobian matrix Ak . Furthermore, depths.
Bott’s method also does not require the solution of equation sys-
tems. However, a disadvantage of Bott’s method is that it suffers
2.6.1 Estimating the regularization parameter
from instability (Silva et al. 2014). A common approach to counter
this issue is to apply a smoothing filter after the inversion to the The regularization parameter μ controls how much smoothness
unstable estimate, as in Silva et al. (2014). is applied to the inversion result. An optimal value of μ will
166 L. Uieda and V.C.F. Barbosa

stabilize and smooth the solution while not compromising the fit
to the observed data. Two widely used methods to estimate an opti-
mal μ are the L-curve criterion and cross-validation (Hansen 1992).
Here, we will adopt the hold-out method of cross-validation (Kim
2009). The hold-out method consists of splitting the observed data
set into two independent parts: a training set doinv and a testing set
dotest . The training set is used in the inversion while the testing set is
kept back and used to judge the quality of the chosen value of μ. For
a value of the regularization parameter μn , the training set is inverted
using μn to obtain an estimate p̂n . This estimate is used to calculate
predicted data on the same points as the testing set via forward
modelling
dntest = f(p̂n ). (16)
The metric chosen to evaluate μn is the mean square error (MSE)
of the misfit between the observed and predicted testing data sets,
Figure 3. Sketch of a data grid separated into the training (open circles)
do − dntest 2 and testing (black dots) data sets. The training data set is still displayed on
MSEn = test , (17)
Ntest a regular grid but with twice the grid spacing of the original data grid.
in which Ntest is the number of data in the testing set. The optimal

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

value of μ will be the one that minimizes the MSE, that is, the one
point estimates of the Moho depth to determine the optimal values
that best predicts the testing data. We emphasize that the inversion
of zref and ρ. These points will generally come from seismologic
is performed on the training data set only.
studies, like receiver functions, surface wave dispersion, and deep
The algorithm for the hold-out cross-validation is summarized as
refraction experiments.
follows:
Let zos be a vector of Ns known Moho depths. We use the MSE as
(i) Divide the observed data into the training (doinv ) and testing a measure of how well a given inversion output p̂l,m fits the know
(dotest )sets. depths. The optimal values of zref and ρ are the ones that best
(ii) For each μn ∈ [μ1 , μ2 , . . . , μ Nμ ]: fit the independent known Moho depths (i.e. produce the smallest
MSE). However, the points do not necessarily coincide with the
(a) Estimate p̂n by inverting the training set doinv . model elements of the inversion. Before computing the MSE, we
(b) Use p̂n to calculate the predicted testing set dntest using interpolate p̂l,m on the known points to obtain the predicted depths
eq. (16). zl,m
s . The MSE is defined as
(c) Calculate the mean square error MSEn using eq. (17).
zos − zl,m
s 
2

(iii) The final solution is the p̂n corresponding to the smallest MSE = . (18)
Ns
MSEn .
The algorithm for estimating zref and ρ is:
The separation of the training and testing data sets is commonly
done by taking random samples from the full data set. However, we (i) For every combination of z ref ,l ∈ [z ref ,1 , z ref ,2 , . . . , z ref ,Nz ] and
cannot perform the separation in this way because Bott’s method ρm ∈ [ρ1 , ρ2 , . . . , ρ Nρ ]:
requires data on a regular grid as well as having model elements (a) Perform the inversion on the training data set doinv using zref, l ,
directly below each data point. Thus, we take as our training set the ρ m , and the previously estimated value of μ. The inversion output
points from the observed data grid that fall on a similar grid but is the vector p̂l,m .
with twice the grid spacing (open circles in Fig. 3). All other points (b) Interpolate p̂l,m on the known points to obtain the predicted
from the original data grid make up the testing data set (black dots depths zl,m
s .
in Fig. 3). This separation will lead to a testing data set with more (c) Calculate the MSE between zos and zl,m
s using eq. (18).
points than the training data set. A way to balance this loss of data in
the inversion is to generate a data grid with half of the desired grid (ii) The final solution is the p̂l,m corresponding to the smallest
spacing, either through interpolation or from a spherical harmonic MSE.
model. A similar approach was used by Silva et al. (2006) and Martins
et al. (2010) to estimate the parameters defining the density-contrast
variation with depth of a sedimentary basin. van der Meijde et al.
2.6.2 Estimating zref and ρ (2013) also had a similar methodology for dealing with the hyper-
The depth of the Normal Earth Moho (zref ) and the density-contrast parameters, though in a less formalized way.
of the anomalous Moho (ρ) are other hyperparameters of the
inversion. That is, their value influences the final solution but they
2.7 Software implementation
are not estimated during the inversion. Both hyperparameters cannot
be determined from the gravity data alone. Estimating zref and ρ The inversion method proposed here is implemented in the
requires information that is independent of the gravity data, such Python programming language. The software is freely avail-
as knowledge of the parameters (Moho depths) at certain points. able under the terms of the BSD 3-clause open-source soft-
This information can be used in a manner similar to the cross- ware license. Our implementation relies on the open-source li-
validation described in the previous section. In this study, we use braries scipy and numpy (Jones et al. 2001, http://scipy.org)
 Fast nonlinear gravity inversion 167

for array-based computations, matplotlib (Hunter 2007, http://

matplotlib.org) and seaborn (Waskom et al. 2015, http://stanford.
edu/∼mwaskom/software/seaborn) for plots and maps, and Fa-
tiando a Terra (Uieda et al. 2013, http://www.fatiando.org) for
geophysics specific tasks, particularly for forward modelling us-
ing tesseroids. We use the scipy.sparse package for sparse ma-
trix arithmetic and linear algebra. The sparse linear system in
eq. (13) is solved using the conjugate gradient method implemented
in scipy.sparse.
The computational experiments (e.g. data processing, synthetic
tests, real data application) were performed in Jupyter (formerly
IPython) notebooks (Pérez & Granger 2007, http://jupyter.org/). The
notebook files combine the source code used to run the experiments,
the results and figures generated by the code, and rich text to explain
and document the analysis.
All source code, Jupyter notebooks, data, and model re-
sults are made available through an online repository (Uieda
& Barbosa 2016, http://dx.doi.org/10.6084/m9.figshare.3987267
or https://github.com/pinga-lab/paper-moho-inversion-tesseroids).
The repository also contains instructions for replicating all results

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

presented here.

3 A P P L I C AT I O N T O S Y N T H E T I C D ATA
Figure 4. A simple Moho model made of tesseroids for synthetic data
We test and illustrate the proposed inversion method by applying it application. (a) The Moho depth of the model in kilometres. The model
to two noise-corrupted synthetic data sets. The first one is generated transitions from a deep Moho in the right to a shallow Moho in left, sim-
by a simple Moho model simulating the transition from a thicker ulating the transition between a continental and an oceanic Moho. Each
continental crust to a thinner oceanic crust. This application uses pixel in the pseudo-colour image corresponds to a tesseroid of the model.
cross-validation to estimate the regularizing parameter (μ) while (b) Noise-corrupted synthetic gravity data generated from the model shown
assuming that the anomalous Moho density-contrast (ρ) and the in (a).
Normal Earth Moho depth (zref ) are known quantities. This first
test is simplified in order to investigate solely the efficiency of the
inversion and the cross-validation procedure to estimate μ. The sec-
We separated the synthetic data into training and testing data sets
ond data set is generated by a more complex model derived from
following Fig. 3. The training data set is a regular grid of Nlat ×
the South American portion of the global CRUST1.0 model (Laske
Nlon = 40 × 50 points (a total of Ntrain = 2000). The testing data set
et al. 2013). This second application uses cross-validation to esti-
is composed of Ntest = 5821 observations. We used cross-validation
mate μ and the validation procedure using synthetic seismological
to estimate an optimal regularization parameter (μ) from a set of
data to estimate ρ and zref . The model and corresponding syn-
Nμ = 16 values equally spaced on a logarithmic scale between
thetic data are meant to simulate with more fidelity the real data
10−6 and 10−1 . We ran our regularized inversion on the training
application.
data set for each value of μ, obtaining 16 Moho depth estimates.
For all inversions, the initial Moho depth estimate used to start the
Gauss-Newton optimization was set to 60 km depth for all inversion
3.1 Simple model
parameters. Furthermore, zref and ρ are set to their respective true
We simulate the transition from a continental-type Moho to an values. Finally, we computed the MSE (eq. 17) for each estimate
oceanic-type Moho using a model composed of Mlat × Mlon = and chose as the final estimated Moho model the one that minimizes
40 × 50 grid of juxtaposed tesseroids (a total of M = 2000 model the MSE.
elements). The anomalous Moho density-contrast is ρ = 400 kg Fig. 5(a) shows the final estimated Moho depth after the cross-
m−3 and the Normal Earth Moho depth is zref = 30 km. Fig. 4(a) validation. The recovered model is smooth, indicating that the cross-
shows the model Moho depths where we can clearly see an eastward validation procedure was effective in estimating an optimal regu-
crustal thinning. In Fig. 4(a), each pixel in the pseudo-colour image larization parameter. Fig. 5(b) shows difference between the true
corresponds to a tesseroid of the model. Moho depth (Fig. 4a) and the estimated Moho depth. The differ-
The synthetic data were forward modelled on a regular grid of ences appear to be semi-randomly distributed with a maximum
Nlat × Nlon = 79 × 99 points (a total of N = 7821 observations) coinciding with a short-wavelength feature in the true model. The
at a constant height of 50 km. The data were contaminated with maximum and minimum differences are approximately 2.19 and
pseudo-random noise sampled from a normal distribution with zero −2.13 km, respectively. Fig. 5(c) shows inversion residuals, de-
mean and 5 mGal standard deviation. Fig. 4(b) shows the noise- fined as the difference between the observed and predicted data (in
corrupted full synthetic data set exhibiting an eastward increase mGal). The largest residual (in absolute value) coincides with the
due to the simulated eastward crustal thinning shown in Fig. 4(a). largest difference between the true model and the estimate. The
The data grid spacing is half the grid spacing of the tesseroid model inversion residuals are normally distributed, as shown in Fig. 5(d),
so that, when separating the training and testing data sets (Fig. 3), the with 0.02 mGal mean and a standard deviation of 3.63 mGal. The
training data set points will fall directly above each model element. cross-validation curve in Fig. 5(e) shows a clear minimum MSE
168 L. Uieda and V.C.F. Barbosa

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

Figure 5. Results from the inversion of the simple synthetic data. (a) The estimated Moho depth. (b) The Moho depth residuals (difference between the true
and estimated Moho depths). (c) The gravity residuals (difference between the observed and predicted gravity data). (d) Histogram of the gravity residuals
shown in panel (c), with the calculated mean and standard deviation (std) of the residuals in mGal. (e) Cross-validation curve used to determine the optimal
regularization parameter (eq. 10). Both axes are in logarithmic scale. The minimum mean square error (eq. 17) is found at μ = 0.00046 (red triangle). (f) Goal
function value (eq. 10) per Gauss–Newton iteration showing the convergence of the gradient descent. The y-axis is in logarithmic scale.

Table 1. Time spent on each function during a single inversion of simple depths shown in Fig. 6(a). In our model, the Normal Earth Moho
synthetic data. The inversion was performed on a laptop computer with an is zref = 30 km and the density-contrast is ρ = 350 kg m−3 . We
Intel(R) Core(TM) i7-3612QM CPU @ 2.10 GHz processor. The total time produce the synthetic data at a constant height of 50 km and on a
for the inversion was 42.133 s. regular grid of Nlat × Nlon = 159 × 119 points (a total of 18 921
Time Percentage of total observations). We contaminate the synthetic data with normally dis-
Function description (s) time (per cent) tributed pseudo-random noise with zero mean and 5 mGal standard
Sparse conjugate gradient 0.021 0.050 deviation (Fig. 6b).
Sparse dot product 0.007 0.017 The validation procedure to determine ρ and zref requires
Tesseroid forward modelling 42.059 99.824 knowledge of the Moho depth at certain points (zos in eq. 18), usually
from seismic experiments. Thus, we must also generate synthetic
seismic data about the Moho depth. We produce such data by inter-
polating the Moho depth shown in Fig. 6(a) on the same 937 geo-
at μ = 0.00046 (indicated by the red triangle). Fig. 5(f) shows the graphic coordinates pinpointed in the data set of Assumpção et al.
convergence of the Gauss-Newton optimization in eight iterations. (2013). The resulting synthetic seismic data is shown in Fig. 6(c).
We also investigated the computation time spent in each section We estimate the three hyperparameters in two parts. First, we run
of the inversion process using a source code profiler. The profiler the cross-validation to estimate an optimal regularization parameter
measures how much time is spent inside each function during the (μ). The starting estimate for all inversions is 60 km depth for
execution of a program. We ran the profiler on a single inversion of all model parameters. For this cross-validation, we keep zref and
the training data set using the estimated regularization parameter. ρ fixed to 20 km and 500 kg m−3 , respectively. Second, we use
We tracked the total time spent inside each of the three functions the estimated μ to run the validation procedure with the synthetic
that represent the potential bottlenecks of the inversion: solving seismological data to estimate zref and ρ, thus obtaining the final
the linear system in eq. (13) using the conjugate gradient method, estimated Moho depths. Fig. 7 summarizes the results.
performing the dot products required to compute the Hessian matrix For the cross-validation, we separate the synthetic data (Fig. 3)
(eq. 12) and the gradient vector (eq. 11), and forward modelling to into a training set with twice the grid spacing of the original data
calculate the predicted data (eq. 3). The profiling results presented (Nlat × Nlon = 80 × 60) and a testing set with 14 121 observations.
in Table 1 show that the time spent on forward modelling accounts We run the inversion for 16 different values of μ equally spaced
for approximately 99.8 per cent of the total computation time. in a logarithmic scale between 10−7 and 10−2 . For each of the 16
estimates we compute the MSE (eq. 17), shown in Fig. 7(a) as
function of μ. The optimal regularization parameter that minimizes
3.2 Model based on CRUST1.0 the MSE is μ = 10−4 (red triangle in Fig. 7a).
In the validation using seismological data, we use the esti-
In this test, we simulate the anomalous Moho of South America us-
mated value of μ in all inversions. We test seven values of zref
ing Moho depth information extracted from the CRUST1.0 model
from 20 to 35 km with 2.5 km intervals and seven values of ρ
(Laske et al. 2013). We construct a tesseroid model with Mlat ×
from 200 to 500 kg m−3 with 50 kg m−3 intervals. We run the
Mlon = 80 × 60 juxtaposed elements, 4800 in total, using the Moho
 Fast nonlinear gravity inversion 169

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

Figure 6. Synthetic data of a model derived from CRUST1.0. The model is made of tesseroids with an constant density-contrast of ρ = 350 kg m−3 and
assuming a reference level of zref = 30 km. (a) The Moho depth of the model in kilometres. Each pixel in the pseudo-colour image corresponds to a tesseroid
of the model. (b) Noise-corrupted synthetic gravity data generated from the model. (c) Simulated points where the Moho depths are known from seismological
estimates (colour dots). Here, these point were obtained by interpolating the Moho depth in (a).

inversion for every combination of zref and ρ, totalling 49 in- 4 A P P L I C AT I O N T O T H E S O U T H

versions. Finally, we calculate the MSE (eq. 18) for each of the AMERICAN MOHO
49 estimates and choose the values of zref and ρ that mini-
We apply the inversion method proposed here to invert for the
mize the MSE. Fig. 7(b) shows a pseudo-colour map of the MSE
Moho depth of the South American continent. We follow the appli-
with a minimum (marked by the red triangle) at zref = 30 km and
cation of van der Meijde et al. (2013) but with some differences,
ρ = 350 kg m−3 .
mainly using a different data set and performing all modelling in
Fig. 7(c) shows the final solution after cross-validation and val-
spherical coordinates using tesseroids. The data are corrected of
idation using seismological data. The recovered model is smooth,
the effects of topography and sedimentary basins. Crust and mantle
indicating that the cross-validation procedure was effective in es-
heterogeneities cannot be properly accounted for in regions where
timating an optimal regularization parameter. Fig. 7(d) shows the
information coverage is sparse and readily accessible models are
difference between the true Moho depths (Fig. 6a) and the esti-
not available, like in South America and Africa. Hence, for the pur-
mated depths (Fig. 7c). The maximum and minimum differences
poses of this study, we will assume to be negligible all other crustal
are, respectively, 9.8 and −8.2 km. The largest absolute differences
and mantle sources, including lateral variations in density along the
are located along the central and northern Andes, where there is
Moho. All tesseroid models are defined with respect to a spherical
a sharp increase in the true Moho depth (Fig. 6a). Positive differ-
Earth of radius 6378.137 km.
ences (indicating a too shallow estimate) appear along the central
portion of the Andes, flanked by regions of negative differences (in-
dicating a too deep estimate) on the continental and Pacific sides.
4.1 Gravity and seismological data
Figs 7(e) and (g) show the gravity residuals, defined as the differ-
ence between the observed and predicted gravity data. The residuals The raw gravity data are generated from the satellite only
appear normally distributed, with 0.03 mGal mean and a standard spherical harmonic model GOCO5S (Mayer-Guerr et al. 2015).
deviation of 4.10 mGal. The gravity residuals follow a similar, The GOCO5S model combines data from 15 satellites, includ-
though reversed, pattern to the differences shown in Fig. 7(d). The ing the complete mission data from the GOCE satellite. The
largest residuals (in absolute value) are along the Andes, with the data were downloaded from the International Centre for Global
central portion being dominated by negative residuals and flanked Earth Models (ICGEM) web-service (Barthelmes & Köhler 2012,
by positive residuals on both sides. Figs 7(f) and (h) show the http://icgem.gfz-potsdam.de/ICGEM/) in the form of the complete
differences between the synthetic seismic data (Fig. 6c) and the gravity field on a regular grid with 0.2◦ grid spacing at ellipsoidal
estimated Moho depths. Once more, the largest differences are con- height 50 km. We calculate the gravity disturbance (δ(P) in eq. 1)
centrated along the Andes, particularly in the central Andes and by subtracting from the raw data the normal gravity of the WGS84
near Ecuador and Colombia. The differences are smaller along the reference ellipsoid (γ (P)) using the formula of Li & Götze (2001).
Atlantic coast of South America, with notable larger differences in Fig. 8(a) show the calculated gravity disturbance of South America.
a few points of northeastern Brazil and along the Amazon river. In We remove the gravitational effect of the topography from the
general, large residuals are associated with sharp increases in Moho gravity disturbance by modelling the ETOPO1 digital terrain model
depth. (Amante & Eakins 2009, http://dx.doi.org/10.7289/V5C8276M)
170 L. Uieda and V.C.F. Barbosa

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

Figure 7. Inversion results from the CRUST1.0 synthetic data. (a) Cross-validation curve used to determine the regularization parameter (eq. 10). The minimum
MSE (eq. 17) is found at μ = 0.0001 (red triangle). (b) Validation results used to determine the reference level (zref ) and the density-contrast (ρ). The
colours represent the MSE (eq. 18) in km2 . The minimum (red triangle) is found at zref = 30 km and ρ = 350 kg m−3 . (c) The estimated Moho depth. (d)
Difference between the CRUST1.0 model depths (Fig. 6a) and the estimated depths. (e) Histogram of the inversion residuals (observed minus predicted data).
(f) Histogram of the differences between the synthetic seismic observations (Fig. 6c) and the estimated depths. (g) The inversion residuals. (h) Difference
between the seismic and the estimated depths.

using tesseroids (Fig. 8b). We used the standard densities of 2670 kg gravitational attraction of the sedimentary basin tesseroid model.
m−3 for continents and −1630 kg m−3 for the oceans. Fig. 8(c) We subtract the total effect of sediments from the Bouguer distur-
shows the calculated gravitational attraction of the topographic bance in Fig. 8(d) to obtain the sediment-free Bouguer disturbance
masses at 50 km height. Fig. 8(d) shows the Bouguer disturbance (Fig. 9a), which will be used as input for the inversion.
(eq. 2) obtained after subtracting the topographic effect from the Fig. 9(b) shows the 937 known Moho depths (coloured dots)
gravity disturbance. which were estimated from seismological data by Assumpção et al.
The effect of sedimentary basins is removed using tesseroid (2013). This data set is used in the validation procedure.
models of the three sedimentary layers present in the CRUST1.0
model (Laske et al. 2013, http://igppweb.ucsd.edu/˜gabi/rem.html).
Each sedimentary layer model includes the density of each 1◦ × 1◦ 4.2 Inversion, cross-validation, and validation
model cell. Figs 8(e)–(g) show the thickness of the upper, middle, using seismological data
and lower sedimentary layers, respectively. The density-contrasts of As in the CRUST1.0 synthetic data test (Section 3.2), we estimate
the tesseroid model are obtained by subtracting 2670 kg m−3 from the hyperparameters in two steps. First, we run the cross-validation
the density of each model element. Fig. 8(h) shows the combined to estimate an optimal regularization parameter (μ). The starting
 Fast nonlinear gravity inversion 171

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

Figure 8. Gravity data for South America and the models used in the data corrections. (a) The gravity disturbance (eq. 1) calculated from the raw gravity
data. (b) Topography from ETOPO1. (c) Gravitational attraction of the topography calculated at the observation height using tesseroids. (d) The Bouguer
disturbance (eq. 2) obtained by subtracting (c) from (a). The upper (e), middle (f) and lower (g) sediment layer thicknesses from the CRUST1.0 model. (h) The
total gravitational attraction of the sediment layers shown in (e), (f) and (g), calculated using tesseroids.

estimate for all inversions is 60 km depth for all model parameters. (Fig. 7b), which is expected because in reality ρ is not homoge-
For this cross-validation, we keep zref and ρ fixed to 20 km and neous across all of South America and the surrounding oceans.
500 kg m−3 , respectively. Second, we use the estimated μ to run the
validation using the seismological data of Assumpção et al. (2013)
to estimate zref and ρ, thus obtaining the final estimated Moho 4.3 Moho model for South America
depth model. The final Moho depth model for South America is shown as a
We split the sediment-free gravity disturbance (Fig. 9a) into the pseudo-colour map in Fig. 11. The model is available in the online
training and testing data sets. The training data set is a regular repository that accompanies this contribution (see Section 2.7).
grid with 0.4◦ grid spacing (twice the spacing of the original data Each model element is a 0.4◦ × 0.4◦ tesseroid, represented by the
grid) and Nlat × Nlon = 201 × 151 grid points, a total of 30 351 pixels in the pseudo-colour map.
observations. The remaining 90 350 points compose the testing Our model differs significantly from CRUST1.0 (Fig. 6a) but
data set. We test 16 values of the regularization parameter (μ) contains most of the large-scale features present in the GMSA12
equally spaced on a logarithmic scale between 10−10 and 10−2 . gravity-derived model of van der Meijde et al. (2013). The deepest
Fig. 10(a) shows the Mean Square Error (MSE) as a function of μ. Moho is along the central Andes, reaching depths upward of 70 km.
The minimum MSE is found at μ = 10−10 , the lowest value of μ The oceanic areas present the shallowest Moho, ranging approxi-
tested, suggesting that little or no regularization is required. mately from 7.5 to 20 km. The Brazilian and Guyana Shields have
We proceed with the validation using seismological data using a deeper Moho (greater than 35 km), with the deepest portions in
μ = 10−10 in all inversions. We test all combinations of nine val- the area around the São Francisco Craton and the northern border
ues of zref , from 20 to 40 km with 2.5 km intervals, and seven of the Parecis Basin. The Moho is shallower than 35 km along the
values of ρ, from 200 to 500 kg m−3 with 50 kg m−3 intervals. Guyana Basin, the Andean foreland basins, the Chaco Basin, and
Fig. 10(b) shows a pseudo-colour map of the MSE with respect to along the centres of the Solimões, Amazonas and Paraná Basins.
the Assumpção et al. (2013) data set. The MSE has a minimum, Fig. 12(a) shows the gravity residuals, defined as the differ-
indicated by the red triangle, at zref = 35 km and ρ = 400 kg m−3 . ence between the observed and predicted gravity data. Fig. 12(b)
The minimum is not as well-defined as for the CRUST1.0 synthetic shows the differences between the seismic-derived Moho depths of
172 L. Uieda and V.C.F. Barbosa

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

Figure 9. Input data for the South American Moho inversion. (a) Sediment-free Bouguer disturbance for South America. Obtained by subtracting the total
sediment gravitational effect (Fig. 8h) from the Bouguer disturbance (Fig. 8d). (b) Seismological Moho depth estimates from Assumpção et al. (2013).

Assumpção et al. (2013; Fig. 9b) and the depths of our gravity- (summarized in Fig. 1) are inadequate or where we have failed to
derived model (Fig. 11). The differences shown in Fig. 12(b) range correct for all crustal and mantle sources. The largest differences
from approximately −23 to 23 km and have a mean of 1.18 km and are seen along the Andean Province and are likely caused by the
a standard deviation of 6.84 km. The gravity residuals and Moho fact that our model does not include the subducting Nazca plate.
depth differences from seismic are smallest in the oceanic areas, Furthermore, the CRUST1.0 synthetic data test (Fig. 7) suggests that
southern Patagonia, and the eastern coast of the continent. The our inversion method is not able to fully recover deep Moho depths
largest gravity residuals are located along the Andes and correlate in the Andes, even without the effect of the subducting plate. In the
with the deepest Moho depths. These large residuals follow a pattern Guyana Basin, our model is able to fit the gravity data but differs
of a negative value in the centre flanked by positive values to the east from the seismological data by up to ±10 km with no clear pattern
and west. This same pattern is observed in the CRUST1.0 synthetic for the distribution of the differences. A possible explanation is an
test results (Fig. 7). In general, larger gravity residuals appear to inaccuracy in the CRUST1.0 sediment model (Laske et al. 2013)
be associated with sharp variations in the estimated Moho depth. used to correct our gravity data. The inversion results will be biased
Along the Andes, large differences with seismic data are corre- if the input data includes effects other than the anomalous Moho.
lated with the larger gravity residuals. Conversely, this correlation In the Amazon and Paraná Basins, our model fits the gravity data
is absent from the large differences seen in the Guyana, Paraná, but underestimates the seismological data by up to 15 km. This
and the Solimões Basins. In the Borborema province, northeastern indicates that a mass excess may be present in the crust or in the
Brazil, our model slightly overestimates the Moho depth. On the upper mantle. A body with positive density contrast whose grav-
other hand, our model underestimates the Moho depths in the Ama- itational effect was not removed from the data during processing
zonas, Solimões, and Paraná Basins. Particularly in the Amazonas will make the observed gravity disturbance greater than it would
and Solimões Basins, where our model predicts a Moho depth of ap- be otherwise. This will cause the inversion to produce a shallower
proximately 30 km, the differences with the seismological estimates Moho estimate. These discrepancies between gravity and seismo-
can reach 10 km or more. logical estimates have been noted before by Nunn & Aires (1988)
for the Amazon Basin and Mariani et al. (2013) for the Paraná
Basin. Both studies propose high density rocks in the lower crust as
probable causes for the discrepancies. Another possible cause for
4.4 Discussion the observed discrepancy in our model could be our failure to fully
Differences between our Moho depth model and the seismological remove from the data the effects of the igneous intrusions present
data (Fig. 12b) may indicate regions where our initial assumptions in both basins. Using a sediment model for South America more
 Fast nonlinear gravity inversion 173

data are at higher altitudes and the model is not outcropping, as

shown by the applications to synthetic data. We employ hold-out
cross-validation to estimate the regularization parameter and a val-
idation procedure using seismological data to estimate the Moho
density-contrast and the Normal Earth Moho depth.
There are two main advantages of the proposed method over pre-
vious works. First, unlike the Parker-Oldenburg method or methods
using rectangular prisms, our inversion method does not require the
data to be projected onto a plane. Second, the Parker-Oldenburg
method and methods derived from Bott’s method cannot apply
the traditional methodology for constraining inverse problems. Our
method has no such restriction because we use the formalism of tra-
ditional Tikhonov inversion. However, the proposed method is not
without limitations. It requires data on a regular grid and restricts
the model to be a regular mesh tied to the data grid. Like Bott’s
method, our method only works for gravity disturbances and not for
the gravity gradients.
The test on simple synthetic data shows that our inversion method
is able to recover a smooth Moho relief with a homogeneous density-
contrast distribution. The inversion was not able to fully recover

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

the shortest wavelength feature in the model, possibly due to the
smoothness constraints which tend to soften high-frequency (sharp)
variations. The cross-validation Mean Square Error curve has a well-
defined minimum, indicating a value of the regularization parameter
(μ) whose corresponding estimate best predicts data that were not
included in the inversion. Using this value of μ in the inversion
Figure 10. Cross-validation results for the South American Moho inversion. leads to a stable solution characterized by a smooth Moho relief
(a) Cross-validation to determine the regularization parameter μ (eq. 10). with an acceptable data misfit.
The minimum Mean Square Error (eq. 17), shown as a red triangle, corre- The efficiency of the proposed method is because solving lin-
sponds to μ = 10−10 . (b) Validation to determine the reference level (zref ) ear systems and performing matrix multiplications together ac-
and the density-contrast (ρ). The colours represent the Mean Square Error count for a mere 0.067 per cent of the total computation time
(eq. 18). The minimum (red triangle) is found at zref = 35 km and ρ = required for a single inversion. The majority of the computation
400 kg m−3 .
time (99.824 per cent) is spent on forward modelling. Thus, we are
able to retain the high computational efficiency of Bott’s method and
use a classic Tikhonov regularization formulation. This approach
detailed than CRUST1.0 might lead to more accurate results for
could, in theory, be extended to other types of regularization (e.g.
these basins.
total variation) and misfit functions (e.g. re-weighted least squares)
Greater confidence in our Moho model can be had in areas where
already available in the literature.
it is able to fit both the gravity and the seismological data. In such
The more complex synthetic data test based on CRUST1.0 shows
places, a gravity-derived model can serve as an interpolator for the
that the validation using pointwise Moho depth information is
seismological point estimates. In general, our model fits both data
able to correctly estimate the density-contrast (ρ) and Normal
sets in the oceans, the Atlantic coast of the continent, and central
Earth Moho depth (zref ). This test indicates that the inversion nei-
Brazil. In the Parnaı́ba Basin, our model has small differences with
ther correctly estimates Moho depth nor adequately fits the grav-
the few seismological data points that fall inside and on the borders
ity and pointwise data when sharp variations in Moho depth oc-
of the basin. The Moho surrounding the basins continental borders
cur. This phenomenon is particularly strong in the region below
is deep, following inward with a shallower Moho, then a deeper step,
the Andes. A likely explanation is that the smoothness regulariza-
and finally a shallower part in the middle of the basin. Recent deep
tion is intrinsically unable to produce sharp variations in Moho
seismic reflection studies by Daly et al. (2014) that cross the basin
depth. These effects might be mitigated with the use of sharpness-
from west to east agree with our model. The Northern border of the
inducing regularization, like the weighted smoothness inversion,
Parecis Basin has a deep Moho in our model that is corroborated by
Cauchy norm regularization, entropic regularization, total varia-
a single seismological data point.
tion regularization, or an adaptive mixed smoothness-sharpness
regularization.
We applied the method proposed here to estimate the Moho depth
5 C O N C LU S I O N S
for South America. Our estimated Moho depth model is in accor-
We have developed a computationally efficient gravity inversion dance with previous results. The model fits well the gravity and
method in spherical coordinates. Our method extends the Gauss- seismic data in all oceanic regions, the central portion of the An-
Newton formulation of Bott’s method to use tesseroids as model el- dean foreland, Patagonia, and coastal and central parts of Brazil.
ements and Tikhonov regularization. The computational efficiency However, the model is unable to fit the gravity and seismic data in
of our method is due to using Bott’s approximation for the Jacobian places with sharp variations in Moho depth, particularly below the
matrix and using sparse matrix algorithms for arithmetic operations Andes and in the boundaries of the main geotectonic provinces of
and the solution of linear systems. This approximation for the Ja- the South American Plate, like the Borborema province, the Par-
cobian matrix is adequate and the method converges even when the naiba Basin, and the São Francisco Craton. This might indicate that
174 L. Uieda and V.C.F. Barbosa

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

Figure 11. The estimated Moho depth of South America. Dotted lines represent the boundaries between major geologic provinces (after Assumpção et al.
2013; Goutorbe et al. 2015); AD: Andean Province, AFB: Andean foreland basins, AM: Amazonas Basin, BR: Brazilian Shield, BO: Borborema province, CH:
Chaco Basin, GB: Guyana Basin, GU: Guyana Shield, PB: Parnaı́ba Basin, PC: Parecis Basin, PR: Paraná Basin, PT: Patagonia province, SF: São Francisco
Craton, SM: Solimões Basin. Solid orange lines mark the limits of the main lithospheric plates (Bird 2003); AF: Africa Plate, AN: Antarctica Plate, CA:
Caribbean Plate, CO: Cocos Plate, SA: South America Plate, SC: Scotia Plate, NZ: Nazca Plate. The solid light grey line is the 35 km Moho depth contour.

smoothness regularization should not be applied indiscriminately able to fit the gravity data but differs significantly from the seismic
to the whole model, as suggested by the CRUST1.0 synthetic data data. These discrepancies in the Paraná and Amazonas Basins are
test. Another reason for the observed misfit might be the presence interpreted in the literature as high density rocks in the lower crust.
of crustal or mantle density anomalies whose gravitational effects In general, differences between a gravity and a seismically derived
were not removed during the data corrections. In the Guyana Basin Moho model may indicate the presence of crustal or mantle density
on the coastal region of Venezuela, along the central Amazonas and anomalies that were unaccounted for in the data processing. Such
Solimões Basins, and in the Paraná Basin, our Moho depth model is locations warrant further detailed investigation.
 Fast nonlinear gravity inversion 175

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

Figure 12. Residuals for the estimated South American Moho depth in Fig. 11. (a) Gravity residuals, defined as the difference between the observed data
in Fig. 9(a) and the data predicted by the estimate in Fig. 11. (b) Differences between the seismological depth estimates of Assumpção et al. (2013) and our
gravity-derived Moho depth estimate. The inset in b shows a histogram of the differences along with their calculated mean and standard deviation (std). Dotted
lines mark the limits of major geologic provinces and lithospheric plates.

Our gravity-derived Moho model for South America can Barbosa, V., Silva, J. & Medeiros, W., 1997. Gravity inversion of basement re-
be downloaded from the online repository http://dx.doi.org/ lief using approximate equality constraints on depths, Geophysics, 62(6),
10.6084/m9.figshare.3987267. 1745–1757.
Barbosa, V., Silva, J. & Medeiros, W., 1999a. Stable inversion of
gravity anomalies of sedimentary basins with nonsmooth basement
AC K N OW L E D G E M E N T S reliefs and arbitrary density contrast variations, Geophysics, 64(3),
754–764.
We are indebted to the developers and maintainers of the open- Barbosa, V.C.F., Silva, J.B.C. & Medeiros, W.E., 1999b. Gravity inversion
source software without which this work would not have been pos- of a discontinuous relief stabilized by weighted smoothness constraints
sible. We thank Marcelo Assumpção for providing the seismological on depth, Geophysics, 64(5), 1429–1437.
Moho depth estimates, Naomi Ussami, Julio C.S.O. Lyrio, Cosme Barnes, G. & Barraud, J., 2012. Imaging geologic surfaces by invert-
F.P. Neto, and Daniel R. Franco for their constructive feedback, ing gravity gradient data with depth horizons, Geophysics, 77(1),
and Editor Dr Gary Egbert and two anonymous reviewers for their G1–G11.
Barthelmes, F. & Köhler, W., 2012. International Centre for Global Earth
thoughtful comments that helped improve this work. L. Uieda was
Models (ICGEM), J. Geod., 86(10), 932–934.
supported by a scholarship from Coordenação de Aperfeiçoamento
Bird, P., 2003. An updated digital model of plate boundaries, Geochem.
de Pessoal de Nı́vel Superior (CAPES). V.C.F. Barbosa was sup- Geophys. Geosyst., 4(3), 1027, doi:10.1029/2001GC000252.
ported by a fellowship from Conselho Nacional de Desenvolvimento Bott, M.H.P., 1960. The use of rapid digital computing methods for direct
Cientı́fico e Tecnológico (CNPq). gravity interpretation of sedimentary basins, Geophys. J. Int., 3(1), 63–67.
Daly, M.C., Andrade, V., Barousse, C.A., Costa, R., McDowell, K., Piggott,
N. & Poole, A.J., 2014. Brasiliano crustal structure and the tectonic setting
REFERENCES of the Parnaı́ba basin of NE Brazil: results of a deep seismic reflection
Amante, C. & Eakins, B.W., 2009. ETOPO1 1 Arc-Minute Global Relief profile, Tectonics, 33(11), 2102–2120.
Model: Procedures, Data Sources and Analysis, NOAA Technical Memo- Farquharson, C.G. & Oldenburg, D.W., 2004. A comparison of automatic
randum NESDIS NGDC-24. National Geophysical Data Center, NOAA. techniques for estimating the regularization parameter in non-linear in-
Asgharzadeh, M.F., von Frese, R.R.B., Kim, H.R., Leftwich, T.E. & Kim, verse problems, Geophys. J. Int., 156(3), 411–425.
J.W., 2007. Spherical prism gravity effects by Gauss-Legendre quadrature Goutorbe, B., Coelho, D.L.d.O. & Drouet, S., 2015. Rayleigh wave group ve-
integration, Geophys. J. Int., 169(1), 1–11. locities at periods of 6–23 s across Brazil from ambient noise tomography,
Assumpção, M., Feng, M., Tassara, A. & Julià, J., 2013. Models of crustal Geophys. J. Int., 203(2), 869–882.
thickness for South America from seismic refraction, receiver functions Grombein, T., Seitz, K. & Heck, B., 2013. Optimized formulas for the
and surface wave tomography, Tectonophysics, 609, 82–96. gravitational field of a tesseroid, J. Geod., 87(7), 645–660.
176 L. Uieda and V.C.F. Barbosa

Hansen, P., 1992. Analysis of Discrete Ill-Posed Problems by Means of the Reguzzoni, M., Sampietro, D. & Sansò, F., 2013. Global Moho from the
L-Curve, SIAM Rev., 34(4), 561–580. combination of the CRUST2.0 model and GOCE data, Geophys. J. Int.,
Heck, B. & Seitz, K., 2007. A comparison of the tesseroid, prism and point- 195(1), 222–237.
mass approaches for mass reductions in gravity field modelling, J. Geod., Santos, D., Silva, J., Martins, C., dos Santos, R., Ramos, L. & de Araújo,
81(2), 121–136. A., 2015. Efficient gravity inversion of discontinuous basement relief,
Hunter, J.D., 2007. Matplotlib: A 2D graphics environment, Comput. Sci. Geophysics, pp. G95–G106.
Eng., 9(3), 90–95. Silva, J., Medeiros, W. & Barbosa, V., 2001. Potential-field inversion: choos-
Jones, E., Oliphant, T., Peterson, P. & others, 2001. SciPy: Open source ing the appropriate technique to solve a geologic problem, Geophysics,
scientific tools for Python, accessed August 2015. 66(2), 511–520.
Kim, J.H., 2009. Estimating classification error rate: repeated cross- Silva, J., Costa, D. & Barbosa, V., 2006. Gravity inversion of basement relief
validation, repeated hold-out and bootstrap, Comput. Stat. Data Anal., and estimation of density contrast variation with depth, Geophysics, 71(5),
53(11), 3735–3745. J51–J58.
Laske, G., Masters, G., Ma, Z. & Pasyanos, M., 2013. Update Silva, J., Santos, D. & Gomes, K., 2014. Fast gravity inversion of basement
on CRUST1.0—1-degree Global Model of Earth’s Crust, in EGU relief, Geophysics, 79(5), G79–G91.
General Assembly Conference Abstracts, vol. 15, pp. EGU2013– Sun, J. & Li, Y., 2014. Adaptive Lp inversion for simultaneous recovery of
2658. both blocky and smooth features in a geophysical model, Geophys. J. Int.,
Leão, J., Menezes, P., Beltrão, J. & Silva, J., 1996. Gravity inversion of 197, 882–899.
basement relief constrained by the knowledge of depth at isolated points, Tikhonov, A.N. & Arsenin, V.Y., 1977. Solutions of Ill-posed Problems,
Geophysics, 61(6), 1702–1714. Scripta Series in Mathematics, Winston.
Li, X. & Götze, H., 2001. Ellipsoid, geoid, gravity, geodesy, and geophysics, Uieda, L., 2015. A tesserioid (spherical prism) in a geocentric coordinate
Geophysics, 66(6), 1660–1668. system with a local-North-oriented coordinate system, figshare, Avail-
Li, Z., Hao, T., Xu, Y. & Xu, Y., 2011. An efficient and adaptive approach able at: https://dx.doi.org/10.6084/m9.figshare.1495525.v1, last accessed

Downloaded from http://gji.oxfordjournals.org/ by guest on November 11, 2016

for modeling gravity effects in spherical coordinates, J. appl. Geophys., November 2015.
73(3), 221–231. Uieda, L. & Barbosa, C.F.V., 2016. A gravity-derived moho model
Mariani, P., Braitenberg, C. & Ussami, N., 2013. Explaining the thick crust for south america: source code, data, and model results from
in Paraná basin, Brazil, with satellite GOCE gravity observations, J. South ‘Fast non-linear gravity inversion in spherical coordinates with ap-
Am. Earth Sci., 45, 209–223. plication to the South American Moho’, figshare, Available at:
Martins, C., Barbosa, V. & Silva, J., 2010. Simultaneous 3D depth-to- http://dx.doi.org/10.6084/m9.figshare.3987267, accessed October 2016.
basement and density-contrast estimates using gravity data and depth Uieda, L., Oliveira, V.C. Jr. & Barbosa, V.C.F., 2013. Modeling the Earth
control at few points, Geophysics, 75(3), I21–I28. with Fatiando a Terra, in Proceedings of the 12th Python in Science
Martins, C., Lima, W., Barbosa, V. & Silva, J., 2011. Total variation regu- Conference, Austin, Texas, USA, pp. 91–98.
larization for depth-to-basement estimate: Part 1 — Mathematical details Uieda, L., Barbosa, V. & Braitenberg, C., 2016. Tesseroids: forward-
and applications, Geophysics, 76(1), I1–I12. modeling gravitational fields in spherical coordinates, Geophysics, 81,
Mayer-Guerr, T. et al., 2015. The combined satellite gravity field model F41–F48.
GOCO05s, in EGU General Assembly Conference Abstracts, vol. 17, pp. van der Meijde, M., Julià, J. & Assumpção, M., 2013. Gravity derived Moho
EGU2015–12364. for South America, Tectonophysics, 609, 456–467.
Nunn, J.A. & Aires, J.R., 1988. Gravity anomalies and flexure of the litho- van der Meijde, M., Fadel, I., Ditmar, P. & Hamayun, M., 2015. Uncertainties
sphere at the Middle Amazon Basin, Brazil, J. geophys. Res., 93(B1), in crustal thickness models for data sparse environments: a review for
415–428. South America and Africa, J. Geodyn., 84, 1–18.
Oldenburg, D., 1974. The inversion and interpretation of gravity anomalies, Waskom, M. et al., 2015. seaborn: v0.6.0, accessed November 2015.
Geophysics, 39(4), 526–536. Wieczorek, M.A. & Phillips, R.J., 1998. Potential anomalies on a sphere:
Parker, R.L., 1973. The Rapid Calculation of Potential Anomalies, Geophys. applications to the thickness of the lunar crust, J. geophys. Res., 103(E1),
J. Int., 31(4), 447–455. 1715–1724.
Pérez, F. & Granger, B.E., 2007. IPython: A System for Interactive Scientific Wild-Pfeiffer, F., 2008. A comparison of different mass elements for use in
Computing, Comput. Sci. Eng., 9(3), 21–29. gravity gradiometry, J. Geod., 82(10), 637–653.