You are on page 1of 8

Layered 2-D inversion of profile oriented data, evaluated using

stochastic models
Anders Vest Christiansen

Esben Auken

M.Sc, PhD student


Hydrogeophysics Group
Dept. of Earth Sciences
University of Aarhus, Denmark
Finlandsgade 6-8, 8200 Aarhus N
anders.vest@geo.au.dk

PhD, Research Assistant


Hydrogeophysics Group
Dept. of Earth Sciences
University of Aarhus, Denmark
Finlandsgade 6-8, 8200 Aarhus N
esben.auken@geo.au.dk , www.hgg.au.dk

SUMMARY

INTRODUCTION

In a sedimentary environment, layered models are often


capable of reproducing the actual geology more accurately
than smooth minimum structure models. We present and
evaluate a 2D inversion scheme with lateral constraints
(2D-LCI) and sharp boundaries for continuous profile
oriented data sets. Here, we focus on resistivity data. All
data and models are inverted as one system, producing
layered sections with laterally smooth transitions. The
models are regularized through laterally equal constraints
that tie interface depths and resistivities of adjacent
layers. Prior information, originating from e.g. electrical
logs, migrates through the lateral constraints to the
adjacent models, making resolution of equivalences
possible. Information from areas with well resolved
parameters will, in a similar way, migrate through the
constraints to help resolve the poorly constrained
parameters.

In many geophysical surveys, the investigator expects a


predominantly layered subsurface, as is most often the case in
sedimentary environments or for most hydrogeological
investigations (Srensen et al., 2003). In these cases it can be
preferable to use an inversion scheme utilizing a layered or
blocky model description. In some cases a 1D solution with
lateral constraints is sufficient (Auken et al., 2002), but
neotectonics, glaciotectonics or other geological phenomena
disturb the subhorizontal layering. This disqualifies the 1D
formulation and calls for a 2D or 3D formulation to enable a
more complex layered earth description.

The estimated model is complemented by a full


sensitivity analysis of the model parameters supporting
quantitative evaluation of the inversion result.
For evaluation we use broad-banded von Krmn
covariance functions to create various geological realistic
models. We compare results from the 2D-LCI routine
with results from the widely used smooth minimum
structure program Res2dinv. The comparison is point-topoint on resistivities in the model space. The statistics
conclude that the 2D-LCI resolves the true models
slightly better than the Res2dinv.
The clear distinction of separate units and unit boundaries
is often critical in hydrogeophysical or geotechnical
applications, and hence we suggest using a layered
inversion scheme in these cases.
Key words:
geoelectric

stochastic

modelling, inversion, 2D,

3DEMIII Workshop, February 2003, Adelaide.

Over the last decade the resistivity method has developed into
a widely used tool with a wide range of off-the-shelf field
equipment and a number of inversion programs available. The
instrumentation allows a detailed mapping by gathering profile
oriented data continuously (Dahlin, 1996, Srensen, 1996). The
data coverage is dense and, most often, profile oriented with
large sensitivity overlaps between individual soundings inviting
for 2D interpretations. 2 D inversion algorithms have been
presented by e.g. Oldenburg and Li (1994) and Loke and Barker
(1996). These algorithms produce smooth minimum structure
models in which sharp formation boundaries are hardly
recognizable. Using a robust inversion scheme (L1-norm) tends
to give a more blocky appearance of the model section (Loke et
al., 2001), but layer boundaries are still significantly smeared
out.
This study will evaluate a laterally constrained 2D inversion
scheme (2D-LCI) with layered models using synthetic
resistivity models based on a stochastic approach. The
inversion methodology is developed for the 1D case and a
paper has been submitted to Geophysical Prospecting on this
issue (Auken et al., 2003). The basics of the 2D methodology
was presented at the EEGS-ES, 2002 (Christiansen et al., 2002)
and a full paper is submitted (February - M arch, 2003) to
Geophysics (Auken and Christiansen, 2003). The methodology
applied in this study is based on the submitted papers

2D-laterally constrained inversion

mentioned above, but with focus on a statistical evaluation and


comparison with existing programs based on realistic synthetic
models.
The model description is adapted from Smith et al. (1999) and
implemented in the 2D-LCI algorithm. Prior information,
originating from e.g. electrical logs, can be added at any point of
the profile. This information migrates through the lateral
constraints to the adjacent nodes. The inversion result is
supported by a full sensitivity analysis of the model
parameters, which is essential in hydrogeophysical
investigations to ascertain the quality of the inversion result.
The LCI inversion kernel is general, and inclusion of other
forward routines, e.g. 2D or 3D MT, TEM or FEM , is straight
forward.
To evaluate the 2D LCI method we use models based on the
broad-banded von Krmn covariance functions (Serban and
Jacobsen, 2001, Mller et al., 2001). These functions enable
construction of complex models with variations at multiple
scales resembling a sedimentary model well. Furthermore, the
generated data reproduce field data to a high degree (Auken et
al., 2002).
Knowing the true model finally enables us to compare
inversion results directly in the model space to produce a
statistically based result. Here we present the statistics as bar
diagrams for the model misfit in different intervals. This study
compares results from the widely used Res2dinv program
(Loke and Barker, 1996) with results from the 2D-LCI
program.
METHODOLOGY
Data description
The CVES systems consist of a number of steel electrodes
(typically around 60, depending on the system type) manually
forced into the ground at a regular electrode spacing (Van
Overmeeren and Ritsema, 1988).

Christiansen and Auken

The electrodes function as current as well as potential


electrodes and measure in any configuration, as desired by the
user. The data collecting is semi-continuous by using a rollalong technique. In this study, we use a newly proposed
electrode configuration called Gradient array (Figure 1). The
number of data is approximately twice the number of data
collected with traditional Wenner arrays, and it has proven to
be superior in resolving non-horizontal earth structures
(Dahlin and Zhou, 2002). 3555 data points are collected for a
profile 1 km long with a minimum electrode distance of 5 m.
Inversion methodology
Data and model
Consider a CVES gradient array data set. To minimize
nonlinearity and to impose positivity we will apply
logarithmic data and logarithmic parameters (e.g. Johansen,
1977, Ward and Hohmann, 1987). Hence

d obs = log( a1 ) ,log( a2 ) ,K ,log( aN ) T .

(1)

We assume the observational errors to be uncorrelated so that


the corresponding covariance matrix, Cobs, is diagonal.
The model has nx surface node points, xi, in the horizontal
direction corresponding to the data profile described above.
This model description follows that of Smith et al. (1999). At
each surface node, xi, the subsurface model is represented by a
logarithmic model with nl layers.
mi = log( ?i1 ),log( ?i 2 ),K,log( ?inl ),K
,
(2)
K, log( ti1 ) ,log( t i 2 ) ,K ,log( t i( nl 1) ) T

where denotes interval resistivity, and t denotes interval


thickness. The full model
m1

m2
m=
,
(3)
M

mn
x
to be determined has M=nx *(2nl-1) parameters. The parameters
from neighbouring nodes are interpolated linearly to produce a
2D model, as illustrated in Figure 2.
The dependence of apparent resistivities on subsurface
parameters is in general described as a non-linear differentiable
forward mapping. We follow the established practice of
linearized approximation by the first term of the Taylor
expansion
d obs g (m ref ) + G (mtrue m ref ) + eobs ,
(4)

Figure 1: Sensitivity of the Gradient-array setup. P


denotes potential electrode, C1 and C2 are current
electrodes, which are fixed for a number of different
potential electrode positions. Blue and green colors are
negative sensitivities whereas red and yellow are positive
sensitivies. After Dahlin and Zhou (2002).

where g is the nonlinear mapping of the model to the data


space. mtrue has to be sufficiently close to some arbitrary
reference model, mref for the linear approximation to be valid. In
short, we write:
dobs = Gm true + eobs ,
(5)
The Jacobian matrix, G, contains the partial derivatives of the
mapping.

2D-laterally constrained inversion

Christiansen and Auken

Inversion
By joining equations (5), (6) and (7) we may write the
inversion problem as:
G
d obs e obs

(8)
P m true = m prior + e prior ,
R
r e r
or more compact:

G'mtrue = d '+e ' .

Figure 2: Model description. The model is described with


thicknesses and resistivities at a number of nodes along a
profile. The parameters between neighbouring nodes are
linearly interpolated to produce a 2D model.

(9)
The covariance matrix for the joint observation error, e,
becomes:
Cobs
0

C'=
C prior
(10)
.
0

C
R

Prior values and lateral constraints


The inclusion of prior values and lateral constraints is based on
an approach developed for 1D laterally constrained inversion
(LCI). Prior information helps resolve the non-uniqueness of
the model and is a way to include information not originating
from the resistivity data itself. Following Jackson (1979) prior
information on parameters (resistivities, thicknesses and
depths) is included as an extra data set, mprior,
m prior = P m true + e prior ,
(6)

The model estimate (Menke, 1989)

where mprior=mprior-mref and eprior is the error on the prior


model with 0 as the expected value, and P is the identity matrix
claiming identity between the prioir value and the model value.
The variance in the prior model is described in the covariance
matrix Cprior.

Analysis of model estimation uncertainty


The parameter sensitivity analysis of the final model is the
linearized approximation of the covariance of the estimation
error, Cest (e.g. Tarantola and Valette, 1982)

Then we add constraints to the solution. Formally, the


constraints are connected to the true model as
r = Rm true + e r ,
(7)
where er is the error on the constraints with 0 as expected
value, r is a zero vector claiming identity between the
parameters tied by constraints in the roughening matrix R,
containing 1 and 1s for the constrained parameters, 0 in all
other places. The variance, or strength of the constraints, is
described in the covariance matrix CR. For most applications,
constraints only between neighbouring models are used.
Constraints between any two sub-models add another sequence
of 1 and -1 terms in R. In the 2D-LCI method we only operate
with lateral constraints although vertical constraints can be
used as well.

mest = G 'T C ' 1 G ' G 'T C '1 d ' ,

(11)

minimizes

[(

)]

1
2
Q =
d 'T C ' 1 d ' ,
N + A+ M

(12)

where, A is the number of constraints, and M is the number of


model parameters including depths.

C est = G'T C 'G '

(13)

Standard deviations on model parameters are calculated as the


square root of the diagonal elements in Cest. For mildly
nonlinear problems, this is a good approximation. Because the
model parameters are represented as logarithms, the analysis
gives a standard deviation factor (STDF) on the parameter ms,
defined by:
STDF (m s ) = exp( Cest, ss ) ,

(14)

Thus, the theoretical case of perfect resolution has a STDF=1.


A factor of STDF=1.1 is approximately equivalent to an error
of 10%. Well resolved parameters have a STDF<1.2.
Moderately resolved parameters fall within 1.2<STDF<1.5,
poorly resolved parameters in 1.5<STDF<2, and, finally,
unresolved parameters have a STDF>2.
Forward modeling

For many applications, lateral constraints on depths are


advantageous to constraints on thicknesses. Constraints on
thicknesses are favourable whenever there is a possibility of
discontinuous layer boundaries, but continuous thicknesses, i.e.
across a fault. Constraints on depths are preferred in cases
where you have a demand for continuity of layer boundaries,
i.e. a Quaternary sequence with disturbed sand and clay layers
on top of a relatively flat pre-Quaternary surface.

The 2D DC forward modelling in the inversion routine is


performed using the finite difference code from University of
British Columbia (McGillivray, 1992). The code uses a finite
difference approach, as described by Dey and Morrison
(1979). The finite difference grid is superimposed on the
model, and a resistivity value is assigned to each cell based on
an area-weighted average of the contributing parts of the
layered 2D model, as shown in Figure 3.

2D-laterally constrained inversion

Christiansen and Auken

The layers are realizations of the covariance function in


equation (15), with larger lateral correlation lengths. Altogether
this produces models as presented in Figure 4.

Figure 4: Stochastic model realization using von Krmn


covariance functions. In this case the model is based on
three layers, each with internal variations.
Statistical evaluation
Figure 3: Model translation. The 2D model in (a) is
translated to the finite difference grid in (b) using area
weighted averages. The thick black line in (b) indicates
the layer boundary positions in (a).
Stochastic models
The theoretical models for which we will create synthetic data
are made using a stationary stochastic process. The degree of
spatial correlation is parameterized in terms of the fractal
dimension D of a self-affine process, which is characterized by
the von Krmn covariance functions (Mller et al., 2001,
Serban and Jacobsen, 2001).

21 r
(15)
K (r L) ,
( ) L
where is the standard deviation, () is the gamma function,
C ( x, z, , ) = 2

r = x 2 + z 2 is spatial distance, L is the maximum correlation


length accounted for, and K() is the modified Bessel function
of second kind and order . The amplitude ratios between
short wavelength and long wavelength varations are controlled
by . Prior knowledge on the presence of horizontal layering
may be included by assuming a correlation length that is
greater laterally than vertically. This will increase lateral
coherence in the inversion result, but will also reduce lateral
resolution.
The synthetic models are made in a sedimentary way, i.e.:
1) 1st layer (bottom layer) is deposited as a homogeneous
halfspace
2) 1st layer is eroded
3) 2nd layer is deposited (draped) on top of the eroded
surface of the 1st layer
4) Erosion of 1st and 2nd layer unit
5) 3rd layer (top layer) deposited on top of eroded surface.

To evaluate the inverted models we make a point-to-point


comparison on the model sections. The error, En, on the nth
cell is calculated as:

E n = exp ln ( true ,n ) ln ( model,n ) .


(16)

Using the absolute difference on logarithms provides a


difference factor between the true model and the inverted
model. The point-to-point comparison is made with 0.5m x
0.5m cells for the entire profile length to depths of 80 m. The
result of individual cells is then divided into 4 intervals
dependent on the size of En. We use the intervals <1.2 for well
determined model cell, 1.2-1.5 for a fair determined model cell
and >2.0 for a poor determined model cell. The cumulated
result is presented as a bar diagram comparing the 2D-LCI
with a traditional minimum structure inversion.
RESULTS
Inversion results
We have generated 11 km of gradient array CVES data using
the setup described earlier. The forward data from the
synthetic models are perturbed to a noise level of 3 %. No data
processing is performed. The models are grouped in two
categories, each with 51.1 km of von Krmn model
realizations. The two types are a minimum type model and a
double descending model. All models are generated from a
three-layer model with the same thickness input for the
stochastic modelling. Interval thicknesses of first and second
layer are 17 and 28 m, respectively. This means that, for very
long realizations, the mean thickness of layer one will be 17 m,
and for the second layer 28 m.
Minimum type model
A representative minimum type model is shown in Figure 6a
at the end of this paper. The average resistivity values of the

2D-laterally constrained inversion

three layers in the minimum type models are (top to bottom)


300 ohm-m, 25 ohm-m and 100 ohm-m. This model class
could represent a sandy top layer (300 ohm-m) overlaying a
clay-rich till (25 ohm-m) protecting a groundwater reservoir in
a sandy formation (100 ohm-m).
Figure 6b presents the result obtained with the Res2dinv
program. The major units are detected, and for the resistive
upper layer the layer boundary is identifiable. However, the
lower boundary of the conductive layer is smeared out, and it
is not possible to read the thickness of the layer in the interval
270 m 530 m. An identification based on this profile would
probably lead to an overestimate of its true thickness for some
parts of the profile.
Figure 6c presents the 2D LCI result based on a 4 layer model.
4 layers are used because the data density is fairly large, and
hence 4 layers would probably have been chosen for field data.
The thickness of the conductive layer is now easily picked up,
even around coordinate 300-450 m where it is relatively thin.
The resistive top layer is also easily recognized and identified
as a separate unit.
Figure 6d is a parameter analysis of the model presented in
Figure 6c. The analysis is presented with a 6-graded colorcoding of the analyses from equation (14). Red corresponds to
well determined parameters, blue is poorly or unresolved
parameters. The first four rows are resistivities of layers 1-4,
below follow thicknesses 1-3. As expected, the parameters are
well determined when the units are thick and not too deep. A
poorer resolution inevitably follows for deeper or thinner
structures, as can be seen around profile coordinate 300-450 m
where the resolution of the thin conductive layer (resistivity
and thickness) is quite low compared to other parts of the
profile.
Double descending model
A double descending model is shown in Figure 7a at the end of
the paper. The average resistivity values of the three layers in
the double descending models are (top to bottom) 300 ohm-m,
70 ohm-m and 15 ohm-m. This model class could represent a
sandy top layer (300 ohm-m) overlaying either a sandy till or a
groundwater reservoir in a sandy formation (70 ohm-m) on top
of a fairly sticky clay (15 ohm-m).

Christiansen and Auken

detail than was the case in the Res2dinv result. The


identification of the layer as actually present for the entire
profile length could be critical in e.g. geotechnical
investigations.
Figure 7d is the parameter analysis of the model presented in
Figure 7c. Again, we conclude that parameters are well
resolved as long as they are not too deep or the layers too thin.
The total result obtained in the two models shown depends
mostly on the variance in resistivity of the synthetic models.
A model with large internal variations in layer resistivities
would, of course, give a poorer result with the 2D-LCI because
the model structure then becomes more smooth than layered.
However, with a layered structure, the Res2dinv is not capable
of reproducing the layered-ness of the model and identification
of boundaries becomes difficult. The clear distinction of
separate units (layers) and unit boundaries is often critical in
hydrogeophysical or geotechnical applications, i.e. thickness
of a protective clay cover over a groundwater reservoir or an
identification of a peat unit in connection with road
construction.
Statistical evaluation
The summarized point-to-point comparisons based on
equation (16) for all 11 profile kilometers (minimum type and
double descending) are shown in Figure 5.
The result states that the model obtained from the 2D-LCI is
slightly better than the model obtained from the Res2dinv. In
both cases, the 2D-LCI is better for the cumulated errors
below a factor of 1.5, and both methods show the best result
at the double descending models. For error factors larger than
2.0 the 2D-LCI, and the Res2dinv has as approximately the
same number of occurrences. The cumulative sums on top of
the bars state that the 2D-LCI for both model types
reproduces 74% or more of the model within a factor of 1.5
from the true model.
Based on the statistics for the model types presented here, it
can be concluded that the 2D-LCI reproduces the model
equally good or better than the Res2dinv.
CONCLUSIONS

Figure 7b presents the result obtained with the Res2dinv


program. Again, the resistive upper layer is fairly accurately
reproduced. The second layer is not identifiable for the right
half of the profile (600-1000 m) because it is very thin.
Finally, the boundary to the lower conductive layer is
significantly smeared out with only rough information on its
true appearance.
Figure 7c presents the 2D-LCI result. The middle layer is
identified for the entire profile and the structures on the
boundary to the lower conductive layer are described in more

The 2D LCI inversion of continuous resistivity data has


proven to be a robust tool to obtain reliable inversion results in
sub-layered environments. The layered model description
makes identification of formation boundaries easy, as compared
to standard minimum structure 2D algorithms, which produce a
more smeared picture of the geological model. The output of
the 2D-LCI is supported by a full sensitivity analysis of the
model parameters entering the inversion scheme. Thus, the
interpreter is given a chance to ascertain the inversion result.

2D-laterally constrained inversion

Christiansen and Auken

REFERENCES
Auken, E., Foged, N, and Srensen, K. I., 2002, Model
recognition by 1-D laterally constrained inversion of
resistivity data: Proceedings - New Technologies and Research
Trends Session, 8th meeting EEGS-ES, Aveiro, Portugal,
EEGS-ES,
Auken, E., Christiansen, A. V., Jacobsen, B. H., Foged, N.,
and Srensen, K. I., 2003, Piecewise 1D Laterally Constrained
Inversion of resistivity data: Submitted to Geophysical
Prospecting,
Auken, E. and Christiansen, A. V., 2003, Sharp boundary,
laterally constrained 2D inversion (2D-LCI) of continuous
resistivity data: Submitted to Geophysics,
Christiansen, A. V, Auken, E., Srensen, K. I., and Smith, J. T,
2002, 2-D Laterally Constrained inversion (LCI) of resistivity
data: Proceedings - New Technologies and Research Trends
Session, 8th meeting EEGS-ES, Aveiro,Portugal, EEGS-ES,
Dahlin, T., 1996, 2D resistivity surveying for environmental
and engineering applications, First Break, vol 14, no 7, p 275283.

Figure 5: Summary of statistics. Red bars are 2D-LCI,


blue bars are Res2dinv. Results from the minimum type
models are summarized in (a), double descending models
are summarized in (b). Bars indicate percentage of pointto-point comparisons in the given intervals. The numbers
on top of the bars are the cumulated percentages.
We have demonstrated this on 11 km of data generated from
stochastic von Krmn models. The models have been inverted
using the 2D-LCI and the Res2dinv program. Statistical
comparison of the model sections obtained with the different
methods suggested that equally good or better results are
obtained. Thus, the choice of whether or not to use a layered
inversion scheme must be based on prior geological knowledge.
If the geology is expected to be layered in some sense, we
suggest using a layered inversion scheme in order to make
identification of formation boundaries easy. The clear
distinction of separate units (layers) and unit boundaries is
often critical in hydrogeophysical or geotechnical applications.
ACKNOWLEDGEMENTS
We would like to thank Prof. Douglas Oldenburg at University
of British Columbia for letting us use the 2D resistivity code
for forward calculations and inversions. Lone Davidsen and
Jens Danielsen helped clarifying the paper.

Dahlin, T. and Zhou, B., 2002, Gradient and mid-point


reffered measurements for multi-channel 2D resistivity
imaging: Proceedings, Integrated Case Histories session, 8th
meeting EEGS-ES, Aveiro,Portugal,
Dey, A. and Morrison, H. F., 1979, Resistivity Modeling for
Arbitrarily Shaped 2-Dimensional Structures: Geophysical
Prospecting, 27, 106-136.
Jackson, D. D., 1979, The use of a priori data to resolve nonuniqueness in linear inversion.: Geophys.J.R.astr.Soc., 57,
137-157.
Johansen, H. K., 1977, A Man/Computer Interpretation
System for Resistivity Soundings over a Horizontally
Stratified Earth: Geophysical Prospecting, 25, 667-691.
Loke, M. H. and Barker, R. D., 1996, Rapid least squares
inversion of apparent resistivity pseudosections by a quasiNewton method: Geophysical Prospecting, 44, 131-152.
Loke, M. H., Dahlin, Torleif, and Acworth, Ian, 2001, A
comparison of smooth and blocky inversion methods in 2-D
electrical imaging surveys: 15th Geophysical Conference and
Exhibition, August 2001, Brisbane., ASEG,
McGillivray, P. R., 1992, Forward modeling and inversion of
DC resistivity and MMR data: PhD thesis University of
British Columbia, Vancouver, Canada

2D-laterally constrained inversion

Menke, William. Geophysical Data Analysis - discrete inverse


theory. Rev. ed. --. 1989. San Diego, Academic Press.
International Geophysics Series.
Mller, I., Jacobsen, B. H., and Christensen, N. B., 2001,
Rapid inversion of 2-D geoelectrical data by multichannel
deconvolution: Geophysics, 66, 800-808.
Oldenburg, D. W. and Li.Y, 1994, Inversion of induced
polarization data: Geophysics, 59, 1327-1341.
Serban, D. Z. and Jacobsen, B. H., 2001, The use of broadband prioir covariance for inverse palaeoclimate estimation:
Geophys.J.Int., 147, 29-40.
Smith, T., Hoversten, M., Gasperikova, E., and Morrison, F.,
1999, Sharp boundary inversion of 2D magnetotelluric data:
Geophysical Prospecting, 47, 469-486.

Christiansen and Auken

Srensen, K.I., Auken, E., Christensen, N.B., Pellerin, L.,


2003, An Integrated Approach for Hydrogeophysical
Investigations: New Technologies and a Case History,
Accepted for publication in SEG, NSG Vol II: Applications
and Case Histories
Tarantola, A. and Valette, B., 1982, Generalized nonlinear
inverse problems solved using a least squares criterion:
Rewiews of Geophysics and Space Physics, 20, 219-232.
Van Overmeeren, R. A. and Ritsema, I. L., 1988, Continuous
vertical electrical sounding: First break, 06, 313-324.
Ward, S. H. and Hohmann, G. W. Electromagnetic theory for
geophysical applications. Nabighian, M. N. Electromagnetic
methods in applied geophysics. 1[4], 131-311. 1988. Tulsa,
Society of exploration geophysicists (SEG).

Figure 6: Minimum type model. The true von Krmn model is shown (a), the L1-norm Res2dinv inversion result in (b),
the 2D-LCI inversion result in (c) and the model parameter analysis accompanying the 2D-LCI result is shown in (d). Red
colors of the analyses indicate well resolved parameters, blue poorly and unresolved parameters.

Figure 7: Double descending model. The true von Krmn model is shown (a), the L1-norm Res2dinv inversion result in
(b), the 2D-LCI inversion result in (c) and the model parameter analysis accompanying the 2D-LCI result is shown in (d).
Red colors of the analyses indicate well resolved parameters, blue poorly and unresolved parameters.

You might also like