7.

Inversion of EM data

7 Basics concerning inversion of electromagnetic data

The process of finding a solution to a given model (i.e., calculate the responses like apparent
resistivities and phases) is termed forward modeling. The backward procedure, i.e., estimate
the model parameters (e.g., layer resistivities and thicknesses) is called inversion. Here we
give only a brief introduction into inversion theory and we refer to several excellent textbooks
on the subject 1.
For a moment we think of the geoelectric problem as a linear problem (which it's certainly
not). The forward problem is then written as:
= =
d Gm oder d i ∑G ij mj i = 1,...n , j = 1,...p . (7.1)

The model parameters are contained in the model vector m, the data in the data vector d. The
operator G, also called data kernel, delineates how to transform the parameter mj to the data
di. Equation (7.1) describes a linear problem, the data di are proportional to the parameters mj.
For instance,
=
yi a x j + b

is of course a linear problem – a straight line. A linear geophysical problem is, e.g., the
gravity of a body with density ρ(r0):
ρ(r 0 )
g(r) = ∫
2
dr .
S (r − r 0 ) 3

When we approximate the integral by a sum, we obtain an equation of type (7.9). On the other
hand, the travel time t of a seismic signal along a path L isn't linearly dependent on wave
velocity:
1
t=∫ dl ,
L v(x, z)

but on it' reciprocal, the slowness s = 1/v. In discrete form we obtain

t i = ∑ Lij c j ;

the non-linear problem has thus been linearized by transformation of the variables. Often the
parameters are available in the form exp(mj) (as in MT and VES!), we then achieve a
linearization by taking the logarithm:
xj = log mj and yj = log dj .
Often, however – also in our case – such a simple new parameterization is not sufficient; a
linear problem is then achieved by approximation with a Taylor series. The inversion is now
carried out iteratively.
The system of equations (7.1) has an unique solution only if N = M and matrix G is not
singular:
1
See, e.g., Hjelt, S.-E. (1992): Pragmatic Inversion of Geophysical Data, Springer-Verlag, Berlin, or Meju, M.A.
(1994): Geophysical Data Analysis: Understanding Inverse Problem Theory and Practice, Society of Explo-
ration Geophysicists, Tulsa, Oklahoma, or Asters, R.C., Borchers, B., and Thurber, C.H. (2005): Parameter
Estimation and Inverse Problems, International Geophysics Series, 90, Elsevier.

7-1
7. Inversion of EM data

−1
=
m G= d Hd (7.2)
If there are more observations then model parameters the system of equations (7.1) is
overdetermined 2, in the reverse case it is underdetermined. An overdetermined system can get
underdetermined if G is singular.
An overdetermined system may be solved, e.g., by the method of least squares. Instead of
(7.1) we now search a solution of
=
d Gm + e ; (7.3)
e is the vector of vector residuals. We solve by minimization of the squared errors:
2
 p
 !
n
S= ( d − Gm ) ( d − Gm ) =
e e= ∑  d i − ∑ G ij m j  =
T T
Min . (7.4)
=i 1 =  j 1 
This means that the derivative of S to the model parameters = 0:

∂ d d − d Gm − m G T d + m G T Gm 
T T T T
∂S
= 0 . (7.5)
∂m j ∂m j

From there it follows:

−d G − G T d + G T Gm + m G T G =
T T
0 (7.6a)

2G T Gm = 2G T d . (7.6b)
For the transition from 7.5 → 7.6a we applied the product rule. Equations (7.6b) are called the
normal equations. It follows as solution

m̂ = ( G T G ) G T d
−1
. (7.7)

The matrix H = [GTG]-1GT is termed generalized inverse in least square sense. Numerical
calculation of this matrix may be obtained, e.g., with the Gauss-Jordan- or the LU-methods.
A measure of quality of the solution is the resolution matrix
R = HG (7.8)
which becomes equal to the identity matrix in the case of a unique solution. Minor diagonals
deviating from 0 mark the connection of parameters, e.g., in the product ρi hi (see equivalence
principle).
The quality of data fit is often expressed via the RMS error (root mean square)

( d − F(m) ) −1
( d − F(m) )
T
R dd
RMS = (7.9a)
n
or

2
An example: For 24 ρa-values (8 per decade at periods between 10…10,000s) the system is overdetermined for
a 5-layer case (5 ρ−values and 4 thicknesses).

7-2
 7. Inversion of EM data

2
1 n  d i − Fi (m) 
RMS = ∑
n i=1  ∆d i
 . (7.9b)

−1
Here n is the number of data points (ρa,i) and R dd is the error covariance matrix and ∆di the
data error. Don't we have such an error available (which happens in some geoelectric
methods), the error is arbitrarily set equal to a constant (for instance = 1), alternatively we
may also set a minimal error (error floor). The RMS error describes the global fit of the
model response to the data. For an "ideal" fit, i.e., the deviation of model response from the
data is in the range of data errors (di – Fi(m) ≈ ∆di) the RMS would be ≈ 1; this is a value
which is difficult to achieve in practice 3. Often the fitting errors are normalized by the data
and one gets a relative RMS, mostly expressed in percent:
2
1 n  d i − Fi (m) 
RRMS = ∑ d 
n i=1 
= RRMS × 100
PRRMS (7.9c,d)
i 
It's clear that our EM or geoelectric problem is non-linear and that it can't be linearized by
simple re-parameterization (e.g., ρ → log ρ). How can we solve such non-linear problems
(which are even ill-posed, which means that large parameter variations may have a small
impact on the response, e.g., in the case of equivalent models or suppression of layers)?
A non-linear problem has the general form:
=d i f=
i (m1 ,...m m ) f i (m)=
oder d F(m) , (7.10)
with F as a functional which transforms the model parameters into the model response (like
the Wait-Lipskaya algorithm to calculate ρa,i from ρj, hj). Many inversion algorithms expect a
starting or initial model m0, which is somehow created (a homogeneous half space or
structures from a-priori information, for example if one knows a layer boundary from other
methods, etc.). The theoretical model response is then:
=d 0i f= 0 0 0 0
i (m1 ,...m p ) oder d F(m ) . (7.11)

Let F now be linear around m0, so that we may express the response of a small change around
m0 as a Taylor series:
= f i ( m10 + δm1 , m 02 + δm 2 ,...m 0p + δm p )
F(m)
∂f i ∂f ∂f
= f i (m ) + δm1 + i δm 2 + ... + i δm p + terms of higher order (7.12)
0

∂m1 ∂m 2 ∂m p

 p ∂f (m ) 
0

= F(m ) + ∑ i |m0 δm j  + 0 ( δm )
0

 j=1 ∂m j 
The observations d always contain errors, (7.10) has thus to be amended:
=d F(m) + e . (7.13)

3
The amount of the data error thus enters directly into the RMS; if one underestimates the measuring errors one
obtains a too large RMS and vice versa. A somehow arbitrarily chosen error floor may thus decrease the fitting
error. The value of the RMS has always to be regarded with respect to the specific problem. It should never be
used as a sole measure for successful modeling; of similar importance is the reproduction of the curve shape.

7-3
7. Inversion of EM data

To obtain a solution, we have to minimize this error e = d – F(m). It follows:

 p ∂f (m) 
d F(m ) − ∑ i
d − F(m) =− |m0 δm j  =−
0
y Ax (714)
 j =1 ∂m j 
and
e= y − Ax . (7.15)
A is the matrix of partial derivatives (Frechet derivatives), also called Jacobi or sensitivity
matrix; they describe the response of an – infinitesimal – model variation. The xj =δmj
contain the unknown corrections to m0, to be determined in order to minimize e.
We thus try to minimize the objective function
S=
S(m) =
eT e =
(d − F(m))T (d − F(m))
. (7.16)
=
(y − Ax)T (y − Ax)

This may be done with, e.g., the Gauss-Newton method; in the framework of this lecture we
cannot go into details here. The minimization is carried out iteratively, i.e., in a first step a
modification x = δm is applied to the starting model m0, this leads to a – hopefully improved
– model estimate m1 and so on:

m = m + ( AT A ) AT y
k +1 k −1
. (7.17)

The problems of such a least squares inversion are often found in the stability of the solution,
i.e., calculating the inverse (ATA)−1. To stabilize the inversion, in the Marquardt-Levenberg
method (also called ridge regression) weights are added to the main diagonal of ATA and
instead of S the new objective function

φ= S1 + αS2= e e + α x x − LT L
T
( T
) (7.18)

is minimized. α is a damping factor, describing the relative weight of S1 and S2, respectively;
L further bounds the parameter variation. One obtains the iterative solution

m = m + ( A T A + αI ) A T y
k +1 k −1
. (7.19)

An example of a resolution matrix is shown in table 7.1, which refers to a geoelectric data set
(see the following chapter) from Saxony, Germany. A 4-layer model was calculated, resulting
in 7 parameters (4 resistivities ρ1 − ρ4, 3 layer thicknesses d1 – d3). The percent RMS for this
inversion was 3.9%. Best resolved is the resistivity of the 2nd layer, also the conductance of
the 2nd layer is well resolved (0.89), badly resolved is the resistivity of the 3rd layer (0.09);
this reflects the principle of layer suppression so that one may perhaps ignore this layer. The
inversion with a 3-layer model yields a slightly worse fit with an RMS of 4.4%. For further
discussion see Fig. 7.8.

7-4
 7. Inversion of EM data

ρ1 0.60
ρ2 -0.02 0.99
ρ3 -0.01 -0.01 0.09
ρ4 0.00 0.00 0.02 0.89
d1 0.14 0.01 0.01 0.00 0.94
d2 -0.03 -0.02 -0.20 -0.01 0.02 0.89
d3 0.02 0.01 -0.11 -0.09 -0.01 0.11 0.22
ρ1 ρ2 ρ3 ρ4 ρ1 ρ2 ρ3

Table 7.1: Resolution matrix for a geoelectric sounding in Saxony.

Abb. 7.8: Two examples for venturous modeling with under- and over-parameterization.
Taken from the online manuscript "How to graph badly", J. Boyd, Univ. Michigan.

Tikhonov Regularization
In earlier chapters it was already emphasized that geophysical problems are often “ill-posed”.
Many methods exist to stabilize a solution particularly in the presence of noise, e.g., damped
cutoff in singular value decomposition (SVD). For the time being, these methods are only
treated in the German version of the lecture notes.
Here we extend this regularization by introducing constraints which have somehow
something to do with the length of the model vector (the absolute value), i.e., its flatness,
smoothness and roughness, respectively 4. These methods are applicable for both linear data
kernels G as well as non-linear functionals G(m). Meaningful and stable solutions are often
obtained only by this kind of regularization.

4
An illustration: The Longitudinal Valley is flat, the Costal Cordillera smooth, the Eastern Cordillera rough.

7-5
7. Inversion of EM data

We first look at approaches which only incorporate the length ||m|| of the model vector in the
search for a stabilized solution. This may happen in various ways.
1) We look at solutions for which Gm − d ≤ δ is fulfilled and minimize ||m||:
m = Min! für Gm − d ≤ δ (7.20)

With growing δ the number of suitable models increases, too and the value of ||m|| decreases
(see Fig. 7.9).
2) We look at solutions for which m ≤ ε is fulfilled and minimize Gm − d :
Gm − d =
Min! für m ≤ε (7.21)
With growing ε the model increases and the number of suitable models increases, too.
3) Another option is the solution of the smoothed least squares problems

Gm − d + α2 m = Min! .
2 2
(7.22)

This is the original ansatz of Tikhonov regularization of zero order. The second term is a
penalty function, it penalizes deviations from a given model idea, here thus the shortness of
the model vector. One can show that by suitable choice of the constants δ, ε and α all three
approaches yield the same solution. If the values of m and Gm − d are plotted in a
diagram with α as parameter, one often obtains an L-shaped curve, the so-called “L-curve”.
An optimal model is obtained by choosing the regularization parameter α such that the
solution is close to the inflection point of the L-curve.
Until now we have only taken the length of the model vector as criterion for minimization.
This is, however, inadequate in many cases because we don’t want to find models which are
close to zero but which are close to another value (e.g., a homogeneous half space of constant
conductivity) or models which are simple in that they are particularly flat, i.e., the model
parameters vary only slowly in model space.
We generalize eq. (7.22) by introducing a differentiate operator L and don’t minimize now
||m|| any more, but instead ||Lm||:

Gm − d + α2 Lm = Min!
2 2
(7.23)

ε
||m||

||m||

||Gm - d|| δ ||Gm - d||

Fig. 7.9: Optimum values of ||m|| and ||Gm – d|| at variation of δ (left) and ε (right).

7-6
 7. Inversion of EM data

Fig 7.10: Example of an L-curve for a 2-D model of the Bolivian Altiplano (Brasse &
Eydam 2008).

Another option consists in penalizing the deviation from a given model by replacing m with
the deviation (m − ma) from an a-priori model ma:

Gm − d + α2 L(m − ma ) = Min!
2 2
. (D.30)

The magnitude of the regularization parameter α thus determines the smoothness of the
model. Rougher models often lead to a smaller fit δ; however, they often contain structures
which are not required by the data. The search of an “optimal” model – e.g., in the sense of
the L-curve criterion − thus always means the sampling of model space by variation of α.

7-7