Transient Electromagnetic Method Application

ph.d.
thesis
Application of airborne TEM methods in Denmark and layered 2D inversion of resistivity data
Anders Vest Christiansen
Department of Earth Sciences University of Aarhus Denmark August 2003
Application of airborne TEM methods in Denmark and layered 2D inversion of resistivity data
Anders Vest Christiansen Department of Earth Sciences University of Aarhus Denmark anders.vest@geo.au.dk http://www.geo.aau.dk/geophys/vest/index.htm
A TEX Prepared with L #pages: 144 12.08.03
Contents
Preface Background
1 The Danish groundwater project 1.1 Aquifer mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Geophysical methods applied . . . . . . . . . . . . . . . . . . . . . . . .
vii 1
1 1 1
Application of airborne TEM methods in Denmark

2 Introduction 2.1 Airborne TEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 The research eld 3.1 The TEM method . . . . . . . . 3.1.1 Basic equations . . . . . 3.1.2 Systems . . . . . . . . . 3.1.3 Sensitivity functions . . 3.2 Inversion . . . . . . . . . . . . 3.2.1 Noise model . . . . . . 3.2.2 Field amplitude data type 3.2.3 Forward modelling . . . 3.2.4 On-time modelling . . . 3.2.5 Model setup . . . . . . .
3
3 3 4 4 4 6 6 7 8 9 10 10 12 12 12 12 13 14 16 16 16 17
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
4 Results 4.1 Synthetic modelling . . . . . . . . . . . 4.2 Field data . . . . . . . . . . . . . . . . 4.2.1 Coupling . . . . . . . . . . . . 4.2.2 The results, Ringkjbing County
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
5 Discussion and conclusions 5.1 Coupling to man-made conductors . . . . . . . . . . . . . . . . . . . . . 5.2 New strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Layered 2D inversion of resistivity data

6 Introduction 7 Methodology 7.1 The Geoelectrical method . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.1 Basic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.2 4-electrode measurements . . . . . . . . . . . . . . . . . . . . .
19
19 20 20 20 21
iv
CONTENTS
7.2
7.1.3 Systems . . . . . . . . . . . . . . . . . . . . . . . Inversion setup . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Data and model . . . . . . . . . . . . . . . . . . . 7.2.2 Forward modelling . . . . . . . . . . . . . . . . . 7.2.3 Forward calculations using a sliding model window 7.2.4 Forward mapping . . . . . . . . . . . . . . . . . . 7.2.5 Lateral constraints . . . . . . . . . . . . . . . . . 7.2.6 Prior information . . . . . . . . . . . . . . . . . . 7.2.7 Inversion . . . . . . . . . . . . . . . . . . . . . . 7.2.8 Model analysis . . . . . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
22 23 23 24 24 26 26 28 28 29 29 29 29 30 31 33 34 35 36 36 37 37 37 39 40 40 41 41 41 41 43
8 Results 8.1 Synthetic models . . . . . . . . . . . . 8.1.1 Von Karman stochastic models . 8.1.2 Model 1 - PACES . . . . . . . . 8.1.3 Model 2 - Gradient array CVES 8.2 Statistical model comparison . . . . . . 8.3 Field example, CVES, Sweden . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
9 Optimising the 2D-LCI 9.1 Broydens update formula . . . . . . . . . . . . . 9.2 1D derivatives . . . . . . . . . . . . . . . . . . . 9.3 Combining Broydens update and 1D derivatives 9.4 Examples . . . . . . . . . . . . . . . . . . . . . 9.4.1 Simple dip-model, PACES data . . . . . 9.4.2 CVES, road construction, Sweden . . . . 10 Discussion 10.1 Smooth minimum-structure or layered? . . . . . 10.2 Using layered 2D inversion with other data types 10.3 Code optimising . . . . . . . . . . . . . . . . . . 10.4 Code improvements . . . . . . . . . . . . . . . . 11 Conclusions References
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
Appendices
Paper 1: Christiansen, A. V. and Christensen, N. B., 2001: A Quantitative appraisal of airborne and ground-based transient electromagnetic (TEM) measurements in Denmark. Geophysics, vol. 68, No. 2, 523-534, March-April 2003 Paper 2: 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
CONTENTS
Paper 3: Auken, E. and Christiansen, A. V., 2003: Layered and laterally constrained 2D inversion of resistivity data. With minor revision accepted to be published in Geophysics (revised version printed in this summary) Paper 4: Christiansen, A. V. and Auken, E., 2003: Layered 2-D inversion of prole data, evaluated using stochastic models. In print, 3DEM-3 proceedings volume, Adelaide, Australia Paper 5: Christiansen, A. V. and Auken, E., 2003: Optimizing a layered and laterally constrained 2D inversion of resistivity data using Broydens update and 1D derivatives. Submitted to Journal of Applied Geophysics
vi
CONTENTS
Preface
This thesis marks the conclusion of a ph.d. project at the Geophysical Laboratory, Department of Earth Sciences, University of Aarhus. The thesis is divided into two separate parts. The rst part is on the application of airborne TEM methods in Denmark. The work was connected with a report evaluating the use of airborne TEM methods in hydrogeological investigations. The report was carried out as a cooperation between the Geological Surveys of Denmark and Greenland (GEUS) and the University of Aarhus. The main results of the work with airborne TEM are presented in Paper 1. A small part was presented at the SAGEEP conference in Las Vegas, 2002, and included in the non-reviewed proceedings volume. The second part is on the development of a layered 2D inversion code for resistivity data. The inversion framework is presented thoroughly in Paper 2 dealing with the 1D case. Papers 3 and 4 deal with specic 2D applications. Finally, Paper 5 concentrates on various code optimisations. Papers 2, 3 and 5 have been edited to t the format of this thesis, and some black and white gures have been exchanged with colour gures. A few equations in Paper 2 have been modied due to some errors we were made aware of from the review of Paper 3.
Papers included in this thesis

1. Christiansen, A. V. and Christensen, N. B., 2001: A Quantitative appraisal of airborne and ground-based transient electromagnetic (TEM) measurements in Denmark. Geophysics, vol. 68, No. 2, 523-534, March-April 2003. 2. 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. 3. Auken, E. and Christiansen, A. V., 2003: Layered and laterally constrained 2D inversion of resistivity data. With minor revision accepted to be published in Geophysics (revised version printed in this summary). 4. Christiansen, A. V. and Auken, E., 2003: Layered 2-D inversion of prole data, evaluated using stochastic models. In print, 3DEM-3 proceedings volume, Adelaide, Australia. 5. Christiansen, A. V. and Auken, E., 2003: Optimizing a layered and laterally constrained 2D inversion of resistivity data using Broydens update and 1D derivatives. Submitted to Journal of Applied Geophysics. These publications and the present summary constitute the contribution to the Faculty of Science, the University of Aarhus, for the ph.d. degree in geology. Additional publications, selected conference proceedings and conference abstracts are listed below.
Selected additional publications

Wisen, R, Auken, E., Christiansen, A.V. and Dahlin, T: Combined interpretation of layered and smooth 2D inversion results: three case histories from geotechnical site
viii
investigations, In preparation. Niels B. Christensen, Anders V. Christiansen, Lene H. Poulsen, Thorkild M. Rasmussen and Kurt I. Srensen, 2000: En vurdering af yb arne transiente metoders anvendelse til hydrogeofysisk kortlgning i Danmark (An estimate of the application of airborne transient electromagnetic methods used in hydrogeological mapping in Denmark). Danmarks og Grnlands geologiske undersgelse (GEUS) rapport 2000/41. Danmarks og Grnlands geologiske undersgelse, Milj- og energiministeriet. Lene H. Poulsen, Anders V. Christiansen og Niels B. Christensen, 2001: Sikring af luftb arne elektromagnetiske data fra Fyns Amt (Securing airborne electromagnetic data from Funen County). Aarhus Universitet, Geologisk Institut. Lene H. Poulsen, Anders V. Christiansen og Niels B. Christensen, 2001: Sikring af luftb arne elektromagnetiske data fra Ringkjbing Amt (Securing airborne electromagnetic data from Ringkjbing County). Aarhus Universitet, Geologisk Institut.
Selected conference abstracts and proceedings not included in this thesis

Christiansen, A. V. and Auken, E., 2003: Optimizing the 2D laterally constrained inversion (2D-LCI) using a quasi-Newton method and 1D derivatives. Proceedings of the 9th meeting of EEGS-ES, Prague, Czech Republic 2003. Christiansen, A. V., Auken, E., Srensen, K. I. and Smith, T., 2002: 2D laterally constrained inversion (LCI) of resistivity data. Proceedings of the 8th meeting of EEGS-ES, Aveiro, Portugal 2002, New technologies and research trends session. Christiansen, A. V. and Christensen, N. B., 2002: Application and analysis of airborne transient electromagnetic methods in Denmark. The role of on-time data. Proceedings of the SAGEEP meeting, Las Vegas, Nevada. Christiansen, A. V. and Christensen, N. B., 2001: Quantitative interpretation and analysis of airborne transient electromagnetic data in Denmark, Proceedings of the 7th meeting of EEGS-ES, Birmingham, England 2001, Electrical and Electromagnetic methods session, 122-123. Christiansen, A. V. and Christensen, N. B., 2000: The sensitivity functions of TEM methods. Proceedings of the 6th meeting of EEGS-ES, Bochum, Germany, September 2000, EM09.
ix
Acknowledgements
First of all I wish to thank my supervisor Niels Bie Christensen, for interested supervision, and generous giving of time and guidance whenever needed. I also wish to thank co-supervisor Esben Auken for participating with great enthusiasm and support during the second part of my ph.d. I also wish to stress my gratitude to Kurt Srensen who also played an important role in the second part. I am grateful to Dr. Douglas W. Oldenburg, University of British Columbia (UBC) who kindly provided the 2D forward code used in the programming. Also, he, and many others at UBC, provided an inspiring environment at my 5-month stay in 2002. Friends and colleges at the Geophysical Laboratory in Arhus participated in numerous fruitful discussions, in particular, my ofce mate Rasmus J. Tlbll, Nikolaj Foged and Jens E. Danielsen from the HydroGeophysics group and my arcade game opponent Mads F. Knudsen. Jrgen Knudsen improved the English language. Roger Wisen from the Geotechnical University in Lund, Sweden tested the 2D code on eld examples, suggesting many improvements. Dr. Torleif Dahlin, also from Lund, was always helpful providing resistivity eld examples. Ringkjbing County supplied the airborne and ground-based TEM eld data, all enlarging the perspective of this study. Finally, I thank Betina for her love and patience over the years.
August 2003, Department of Earth Sciences University of Aarhus, Denmark anders.vest@geo.au.dk
Background
1 The Danish groundwater project
Denmark covers some 43 000 km2 and has nearly 500 islands. The subsurface is mainly Quaternary deposits overlying chalk, limestone and Tertiary sand and clay. The combination of low topography, an annual rainfall of 400-800 mm and widespread consolidated and unconsolidated aquifers ensure a plentiful and easily accessible supply of groundwater. However, the increasing problems with water quality due to urban growth and intensive farming in 1996 called upon a 10-year plan to improve groundwater protection. Through an agreement between the Danish counties and the government, it was decided to map more than 33% of Denmarks surface area, using geophysical methods to obtain a thorough knowledge of the groundwater resources. The estimated total cost of the project is around 120 million Euro, nanced through water taxes, a project unique in both intention and scope.
1.1
Aquifer mapping
Initiating the 10-year plan the recharge areas were grouped into three categories: those of major, those of moderate, and those of limited drinking water interests. On the basis of these classication zones a comprehensive geophysical, hydrological and hydraulic investigation was initiated to ascertain the location, structure, vulnerability, and water quality of aquifers. Based on the models derived from the mapping effort, water and land classication zones can be determined. Integration of the various data sets is extremely useful for the authorities to make water policy decisions, and future land planning will be established according to the classication of the area.
1.2
Geophysical methods applied
Among the methods applied in hydrogeophysical investigations, electrical and electromagnetic methods have gained a central position (Fitterman and Stewart 1986; McNeill 1990; Sandberg 1993; Meju et al. 1999; Srensen et al. 2003), because these methods allow distinction between formations of different electrical resistivity. Electrical resistivity is often strongly correlated with clay content; hence a distinction between permeable and nonpermeable formations can be made. Two ground-based methods have proved particularly useful in Denmark: transient electromagnetic (TEM) soundings and geoelectrical sounding/proling. The TEM method is primarily used to delineate the lower boundary of aquifers, and more than 30 000 soundings have been made in Denmark for this purpose (Poulsen and Christensen 1998). The most widely used geoelectrical methods employ the systems of continuous vertical electrical sounding (CVES) with multi-electrode systems (e.g. Bernstone and Dahlin 1999) and the pulled-array continuous vertical electrical sounding system (PACES) (Srensen 1996) which have been employed mainly for near-surface mapping to determine aquifer vulnerability.
The Danish groundwater project
Application of airborne TEM methods in Denmark
Introduction
This section deals with the work carried out comparing and analysing airborne and groundbased transient electromagnetic (TEM) methods, most of it presented in Paper 1. Part of the work has not been published as peer-reviewed. This is mainly the section on sensitivity functions and the section on on-time modelling that rounded off the project on airborne TEM. The latter was published in the proceedings volume of the SAGEEP conference (Christiansen and Christensen 2002), but not as a peer-reviewed publication due to lack of high-quality eld data.
2.1
Airborne TEM
The delineation of aquifers using the TEM method started in Denmark around 1990. Since then the mapping campaign has intensied which ve years ago drew attention to airborne TEM methods as a fast way to map large areas. Airborne TEM methods have traditionally been used in mineral exploration for mapping highly conductive bodies in a resistive background (e.g. Pedersen and Thompson 1991; Smith and Keating 1996), for qualitative more than quantitative purposes. 1D imaging techniques have so far been the most abundant tools for processing the airborne data (DeMoully and Becker 1984; Macnae and Lamontagne 1987; Macnae et al. 1991; Liu and Asten 1993). In the imaging techniques, the downward and outward diffusion of the induced current system in the earth is approximated by one or more image current loops that move downwards with time. The variation in the depth and speed of these current loops with time can be converted into a conductivity-depth section that gives a smooth approximation to the true conductivity structure. They can be thought of as sophisticated transformations of the observed data. The conductivity-depth transform (CDT) by Wolfgram and Karlik (1995) belongs to the group of methods mentioned above, and is used by Fugro Airborne Surveys Ltd. as the main tool for interpretation of data from the widely used GEOTEM system. However, the usefulness of a groundwater investigation in a sedimentary environment is substantially enhanced with a quantitative output model including quantitative analyses of model resolution (Srensen et al. 2003). This can be obtained from a full nonlinear inversion scheme, producing 1D layered earth models with a quantitative analysis of resolution. This is the main theme of Paper 1. To set up a quantitative inversion scheme special attention should be given to the description of the recording situation. In particular this includes the noise model, the transmitter and receiver characteristics. A thorough description of these parameters has been a major part of the work with the airborne TEM methods, and is also dealt with in detail in this summary. Introducing the main sections, I have included brief descriptions of the basic principle and the mathematics. Furthermore, a short section on sensitivity functions is included which visualises fundamental differences between a ground-based and an
The research eld
airborne TEM system. All analyses and inversions in the central sections compare the airborne TEM with a traditional ground-based TEM method. The study involves both synthetic models and eld case studies. The eld data are part of the rst two airborne TEM projects carried out in Denmark in Ringkjbing and Fyn Counties with the GEOTEM system. The ground-based PROTEM 47 equipment from Geonics Ltd. was used for comparison.
3
3.1
The research eld

The TEM method
The transient electromagnetic method is part of the time domain group, as opposed to the frequency domain group. Time domain methods use a direct current, usually passed through an ungrounded loop. The current is abruptly interrupted, and the rate of change of the secondary eld due to induced eddy currents is recorded in an induction coil. Ideally, the primary eld is absent while measuring. Figure 3.1 summarises the basic principles and nomenclature. The waveform depicted in Figure 3.1 is called square, whereas the airborne GEOTEM system uses a half-sine current pulse because rapid turn-off is difcult with high currents.
Current On-time Turn-on Turn-off Secondary magnetic field Off-time
Measurement in off-time
Time gates
Figure 3.1: The basic measuring principles of the TEM method. Note that this is for a central loop system and the relative lengths of the turn-on ramp and the turn off-ramp are out of scale.
The data recording is usually done in time windows (often called gates) with increasing length to improve the signal to noise ratio at late times. The shifts in current direction subdue the unwanted signal from power lines if the repetition frequency is chosen as a multiple of a subharmonic of the 50 Hz power line frequency. 3.1.1 Basic equations This section gives a short introduction to the basic theory using the notation of Ward and Hohmann (1987). Time-domain are lower-case letters, frequency domain are upper-case letters. An electromagnetic eld is dened by the ve vector functions: e (electric eld intensity), b (magnetic induction), d (dielectric displacement), h (magnetic eld intensity) and j (electric current density). The interaction between the elements is governed by Maxwells Equations, describing all electromagnetic phenomena. In the time domain they are given
3.1 The TEM method
as
e+
b = 0, t d = j, h t b = 0, d = ,
(3-1) (3-2) (3-3) (3-4)
where is electric charge density [C/m3 ]. Fourier transformation and application of the constitutive relations yield the Maxwell Equations in the frequency domain E+z H = 0, Hy E = 0, (3-5) (3-6)
where the impedivity, z = i and the admittivity, y = + i have been used. The homogeneous Maxwell Equations in (3-5) and (3-6) apply only in source-free regions. Otherwise they are replaced by the inhomogeneous equations: E+z H = JS m, Hy E = JS e, (3-7) (3-8)
S where JS m is magnetic source current, and Je is electric source current. Expressing E and H in terms of the Schelkunoff potentials A and F facilitates the derivation of E and H by differentiation. The TEM methods use a magnetic source (JS m ) transmitting a transverse electric eld. This simplies the situation because only the Schelkunoff potential F is necessary to express Em and Hm . Em is dened as
Em F, from which Hm ends up as: 1 F + ( F). Hm = y z
(3-9)
(3-10)
Assuming a circular loop with current I and radius a, the Schelkunoff potential between the source at P (x, y, z ) and the earth is F (, z ) = z 0 Ia 2
0
1 u0 (z +h) [e + rTE eu0 (z h) ]J1 (a)J0 ()d, u0
(3-11)
where Jp is the Bessel function of order p, =
2 + k 2 for k and k wavenumbers, kx x y y
2 . Finally, r = x2 + y 2 and un = 2 kn TE is the reection coefcient obtained by recursion from the bottom layer of the model. The eld value Hz is found using (3-10)
Hz =
Ia 2
[eu0 (z +h) + rTE eu0 (z h) ]
2 J1 (a)J0 ()d, u0
(3-12)
which in the general case can be evaluated only be numerical integration.
The research eld
Tx 120 m Rx 131 m 70 m
Tx
GEOTEM
GEOTEM
PROTEM 47
PROTEM 47
Rx
Figure 3.2: Schematic representation of PROTEM 47 and GEOTEM eld setups.
3.1.2 Systems The systems employed in the following sections are sketched in Figure 3.2. PROTEM 47 from Geonics Ltd is a standard, single-site, ground-based system. The most popular conguration in Denmark is a 4040 m2 loop, with the receiver coil at the centre of the loop measuring the vertical component of the dB/dt eld. The measurements are split into 3 segments, each with 20 time channels, distributed with 10 per decade in time between 7 s and 7 ms from the end of the transmitter pulse. GEOTEM from Fugro Airborne Surveys is a xed-wing TEM system with receiver coils in a trailing bird. The loop is strung around the airplane and the receiver consists of three mutually perpendicular induction coils measuring the three components of dB/dt. Data are sampled continuously and subsequently binned in 20 gates with only the last 15 being in the off-time (between 0.2 ms and 4.4 ms from the end of the transmitter pulse). 3.1.3 Sensitivity functions Mental modelling of electromagnetic phenomena is very difcult, and EM methods often seem to resist visualisation in terms of the physical processes in the ground which produce them. Introducing the sensitivity function facilitates a way of describing how a conguration responds to subsurface conductivity structures in time and space. Assuming that the quasi-static approximation is valid, it is possible to derive an expression for the electrical eld in the subsurface originating from a circular transmitter loop at a height h over a homogeneous halfspace. Applying Ohms Law of the constitutive relations, J = E, then yields the current density. The contribution of a current element in the subsurface to the total magnetic eld is described by Biot-Savarts law: dB = 0 I dl (r2 r1 ) 4 |r2 r1 |3 (3-13)
where I is the current of the current element, dl is a unit vector in the same direction as the current, r2 is a position vector to the observation point (the receiver) and r1 is a position vector to the current element. Vertical and horizontal slices of the 3D sensitivity functions for the PROTEM 47 system and the GEOTEM system, utilising the full waveform, are shown in gure 3.3. The examples are 30 m homogeneous halfspaces at time 0.3 ms. The ight direction is from left to right, meaning that the transmitter is at x = 0 m for both systems, and the GEOTEM receiver is at x = 131 m. The vertical section of z-component structure of the PROTEM 47 system (Figure 3.3a) is the well-known ear shape stretching outwards and downwards, with zero sensitivity directly underneath the receiver (Figure 3.3d). Compared to the GEOTEM z-component
3.2 Inversion
Depth [m]
0 100 200 300

0
(a)
0
(b)
0
(c)
0.8 0.6
Normalized values
-300 -150 -300 -200 -100 0 100 200 300
150 300
-300 -150
150 300
-300 -150
150 300
0.4 0.2
PROTEM47
GEOTEM, Z
GEOTEM, X
(d)
0
0
(e)
0
(f)
0 0.2 0.4 0.6 0.8 1
y [m]
-300 -150
150 300
-300 -150
150 300
-300 -150
150 300
x [m]
x [m]
x [m]
Figure 3.3: The sensitivity function of the z-component for the PROTEM 47 system and the z- and xcomponents for the GEOTEM system. The top row gives a vertical section and the bottom row a horizontal section. Contour intervals are 0%, 20% - 80% of the maximum value. The vertical PROTEM 47 section is scaled with a factor 10 due to very high sensitivities at the surface. The dotted line in the vertical slices indicate the intersection of the horizontal slices. Thin sheet responses at an arbitrary scale are shown above the GEOTEM responses.
(Figure 3.3b and e) the PROTEM 47 sensitivity function is relatively focused. The zcomponent GEOTEM sensitivity function exhibits sign shifts at the positions of the transmitter and the receiver with negative sensitivity in between. The ngerprint of the general shape is recognised in a response of a thin sheet plotted directly above the vertical sensitivity slice. The x-component structure of the GEOTEM system (Figure 3.3c and f) is much simpler, with a pronounced positive structure beneath the receiver, zero sensitivity at a perpendicular line cutting through the transmitter, and a weak negative structure ahead of the transmitter. Again, a ngerprint of the main structure is recognised in the thin sheet response above. The sensitivity sections underline the fact that the airborne system has a very large footprint compared to the ground-based system, predicting an inferior lateral resolution. It cannot be concluded directly that the vertical resolution is poorer for an airborne system, but effectively this is the case due to long turn-off times and low signal/noise ratio. It is important to notice that the sensitivity functions are not Frechet kernels. The Frechet kernel is in the frequency domain dened as the dot product of the electric eld originating from the transmitter and the electric eld from the receiver considered as a transmitter. Thus, for transmitter and receiver at the same height, the Frechet kernel is the same when interchanging the receiver and the transmitter. This, however, is not the case for the sensitivity function dened above. The calculation of sensitivities as presented here is a simple weighting of the current density distribution applying Biot-Savarts law. Hence, the distortion of the E-eld from current channelling when removing or adding conductivity to a very small element is not considered.
3.2
Inversion
To obtain meaningful models from the inversion of TEM data, and to make comparative analyses of data t and model reliability, it is essential to specify the recording situation.
The research eld
Special attention must be given to system specications and a thorough description of the noise model (Munkholm and Auken 1996). The systems were described in section 3.1.2. 3.2.1 Noise model The transient data value is an averaging of the induced electromotive force in the receiver coil within the gate. Ground-based methods normally use logarithmic gating, meaning that the length of a gate is proportional to delay time. If the surrounding noise is white (i.e., stochastic through all frequencies), logarithmic gating results in an effective noise decay of t1/2 . However, the surrounding noise is not white. The spectra of AM transmitters have high amplitudes at single frequencies. These thin spectral lines will overlay the white noise. Averaging and stacking such monochromatic signals results in an effective noise decay of t1 (Effers et al. 1999). This contribution will dominate at early times, whereas the stochastic noise will dominate at late times. In Denmark, the transition time of the PROTEM 47 system is 100 s and the value at this delay time is 10 nV /m2 , determined through several experiments as a good average. The total absolute noise is the sum of the two contributions. Figure 3.4 shows the total noise as a function of time.
10 Amplitude [V/m 2]
-6
10
-7
PROTEM 47, z 10
-8
10
-9
GEOTEM, z GEOTEM, x 10 -2 10 -1 Time [ms] 10 0 10 1
10 -3
Figure 3.4: Noise model off-time. The length of the lines indicates the time-interval in which data are recorded, measured from the end of the transmitter pulse. The noise levels are absolute and not normalised according to the effective transmitter current. The labels x and z refer to the x and z-components of the measured dB/dt eld.
Standard eld procedures with the GEOTEM system do not include actual measurements of the noise level. Ideally, the noise model would be calculated for each ight, based on; 1) high-altitude measurements with the transmitter turned on to get the system response and its variation1 , and 2) measurements at low altitude with the transmitter turned off, to determine the ambient noise, at the measuring site. The total noise would be a sum of the two. Normally, the only available data are the gated data from which the primary eld of the average waveform has been subtracted. The average waveform is also available. Hence, the only possible way to estimate the noise on GEOTEM data is by inspecting the regular data. Assuming that the data from the last gate contain no or little signal from the earth, the standard deviation (STD) on these data approximates the noise level. The noise levels
Standard eld procedures of high-altitude measurements are aimed at determining the average waveform and do not include a statistical analysis of its variation.
1
3.2 Inversion
on all other off-time gates are found by scaling with the relative length of the gates noise(n) = ST D(20) length(20) length(n) (3-14)
for the nth gate. This noise model for GEOTEM off-time data as found from the Ringkjbing eld data set is also illustrated in Figure 3.4. The presentation is a little awkward because the reference is the time of turn-off, which is not well-dened for the GEOTEM system. This issue will be addressed later in section 3.2.3. Note that all other noise sources are ignored here (i.e., roll, pitch, bird position uctuations etc.). The data are assigned another 5% relative noise due to the fact that the model assumption has a lower-ranking dimensionality than the targets, and to account for deviations from the nominal conguration (Smith 2001b). For the on-time gates the method described above cannot be used. The noise model for the on-time gates is determined by the variations in the transmitter current combined with the ambient noise. These data are only present if specially addressed in the recording situation. In one of the Danish surveys an on-time noise model was estimated from two data sets; one full data set of high-altitude measurements with the transmitter turned on to determine the total system response and its variation, and one data set from lowaltitude measurements with the transmitter turned off to determine the ambient noise at the measuring site. This noise model enables us to include on-time gates in the modelling. 3.2.2 Field amplitude data type
During ight the receiver bird position varies (Smith 2001b), and the motions tend to mix the three components of the measured eld, so that dBx /dt contains some dBz /dt and dBy /dt and correspondingly for the other components. Using the Field Amplitude (FA) as data deals with this problem (Christiansen and Christensen 2003). The FA is dened as FA = dBx dt
2
dBy dt
dBz dt
(3-15)
where dBx /dt, dBy /dt and dBz /dt are the measured time derivatives of the components of the magnetic eld. Figure 3.5a presents a small section of three selected channels (early-time, intermediate time and late-time) from the eld data set presented later. The oscillating character of the x- and z-component data is caused by the receiver bird motions. To a large extent the FA data type eliminates the effects of receiver bird movements. Figures 3.5b-d are the results of an inversion of one sounding, indicated by the dashed line in Figure 3.5a. Figures 3.5b and c illustrate the problem with a joint inversion of the x- and z-components. Compared with the predicted data, the x-component data are too high and the z-component data are too small, giving a misleading outcome of the joint inversion with a residual of 1.6 (equation (7-28)). Figure 3.5d is the result using the FA data type, with a residual of 0.07. The main reason to use the FA data type is the reduction of the effects of receiver bird motion, but additional advantages are achieved; 1) the FA enables inversion on raw data, which means less lateral smoothing, reducing the effects of coupling to man-made conductors as much as possible (see section 4.2.1). 2) the FA is less affected by possible 2D effects on a quasi-layered halfspace because the FA sensitivity function is smooth, whereas the x- and z-component sensitivity functions are very different
10
The research eld
10
(a)
7
(b)
(c)
(d)
Amplitude [V/m2]
10
FA z x
10
10
3000
3250
3500 101
10
10 10
10
10 10
10
10
Profile coordinate [m]
Time [ms]
Time [ms]
Time [ms]
Figure 3.5: Field Amplitude data type. Components x (green), z (blue) and FA (red) of channels 7, 11 and 18 from observed data are presented in (a), (b) shows predicted x-component data (dots) from a joint inversion of the x- and z-component data and observed data (solid), (c) is as (b) but for the z-component. Finally, (c) shows predicted data (dots) compared with observed data (solid) using the FA data type for inversion. The soundings shown in (b)-(d) are indicated by the dashed line in (a)
as depicted in Figure 3.3. This will inevitably disrupt a joint inversion in the presence of 2D structures. All the following inversions and analyses of airborne data are based on the FA data type. 3.2.3 Forward modelling All modelling and inversions are done using the 1D full non-linear inversion programme SELMA (Christensen and Auken 1992). SELMA approximates the actual geometry of the transmitter source with a circular current loop of the same area. For the airborne data, the actual heights of the transmitter and receiver are taken into account. Calculation of PROTEM 47 responses includes both a waveform with turn-on and turnoff ramps, and modelling of the effect of all system band-pass lters.The characteristics of the lters were measured in the laboratory. The incorporation of the band-limitation has been shown to be crucial for an accurate inversion (Effers et al. 1999). The GEOTEM response, however, is calculated applying the waveform convolution only. The average waveform is found from the transmitter signal as measured by the receiver coils in high-altitude measurements (Figure 3.6). This signal should encompass the full receiver transfer function with all effects due to induction in the aircraft and any bandpass ltering applied through the receivers. The ltering effect of the receiver is seen as a distinct time delay in Figure 3.6 comparing the transmitter waveform with the waveform measured by the receiver coils. The gure also outlines the fact that the waveform used for computation is drawn to zero and truncated well before the rst gate opens, in order to do off-time calculations only. Modelling results suggest that the error on the rst gate using the truncated waveform is approximately 2%, decreasing for the following gates. Bringing the truncation of the discretised waveform closer to the rst gate makes the relative error larger. The chosen truncation is made in a rather pragmatic way. 3.2.4 On-time modelling On-time measurements are nding increasing use for various purposes (Annan et al. 1996; Smith and Balch 2000; Smith 2000; Smith 2001a; Smith and Lee 2002) and ideally all
3.2 Inversion
11
0.1 1
(a)
Nominal amplitude
0.08 0.06
(b)
0.8 0.6 0.04 0.4 0.02 0.2 0 0 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 0.02 6.5 2 2.1 2.2 2.3 2.4 2.5
Time [ms]
Time [ms]
Figure 3.6: The GEOTEM transmitter waveform. (a) The integrated waveform as measured with the receiver coils (blue line), and as measured with a pick-up coil at the aircraft (red line). The dotted green line is the 10-point piecewise linear waveform used for computations, and the circles mark the off-time gate-centre positions. (b) A magnication of the framed area in (a).
data should of course be modelled as on-time. It is clear from Figure 3.6 that there is a signicant signal in what should be off -time. Also, there are ve gates in the on-time and it is important to retrieve all possible information from the already information-sparse airborne data. This motivated an attempt to do all the modelling as on-time modelling. Figure 3.7a is the full 128 point sampled dB/dt waveform as seen by the receivers. In Figure 3.7b the off-time is enlarged to give an idea of the present on-time signal. The methodology is as follows. Step response eld quantities are calculated in log(time), and then interpolated to the equally spaced 128 sample points per half-cycle. The value at t = 0 is the primary eld. The values are differentiated numerically to get the impulse response. The system response is then found through a discrete convolution between the impulse response and the actual 128-point sampled waveform. Then the primary eld is
x 10 3
5
x 10 0 0.2 0.4 0.6 0.8
Amplitude [V/m2]
2 1 0 1 2
(a)
3 0 20 40 60 80 100 120 1 60 80 100
(b)
120
Sample #
Sample #
Figure 3.7: dB/dt waveform. (a) is the full waveform as seen by the receivers for the x-component (green), the y-component (red) and the z-component (blue). The circles mark the gate-centre positions. (b) is a closeup from section (a) indicating the signal amplitude in the off-time.
subtracted from the secondary eld in exactly the same way as eld data are processed. The amount to subtract is found by minimising the resulting signal in a least squares sense
12
Results
through the -parameter: min |dmeasured dprimary |2 . (3-16)
Finally the 128 samples are binned in 20 gates. Results will be compared using the off-time approach with results using the on-time approach. 3.2.5 Model setup In this summary only few-layer inversions will be presented. The few-layer inversion minimises the data mist using the lowest number of layers while tting the data to an appropriate level. The inversion parameters are layer thicknesses and resistivities. Multilayer inversions are presented in Paper 1. The initial model is a homogeneous halfspace, which means that no qualied guess on the model is necessary. The basic inversion scheme is described in detail in section 7.2.5.
4
4.1
Results
Synthetic modelling
Working on the GEUS report, I inverted and analysed 693 different 1D models on four different TEM systems. Of these I shall show a sequence of 21 models comparing the GEOTEM and the PROTEM 47 systems that summarise many of the general conclusions. The synthetic data are produced from forward calculations with noise ascribed using the noise model from section 3.2.1. The base model is double descending with resistivities 200 m (layer one), 70 m (layer two) and 5 m (layer three) (Figure 4.1a). The thickness of the rst layer is 30 m. The thickness of the second layer is changed in 21 exponential steps from 3 to 300 m (10 per decade). The layers are intended to represent a sandy top-layer overlaying a possible aquifer (both Quaternary) and at the bottom a heavy Tertiary clay delineating the lower aquifer boundary. PROTEM 47 (Figures 4.1b and c) needs three layers whenever the equivalences are not too strong, in this case when the thickness of the second layer is comparable to that of the rst layer (column 10). Two layers are needed in the remaining cases. The PROTEM 47 system resolves layers to depths of approximately 200 m in the best case. GEOTEM offtime data (4.1d and e) contain information to resolve a two-layer model in most cases. Three layers are possible to resolve in some cases when on-time models are included (Figure 4.1f and g column 18-21). The GEOTEM system detects layers to depths of approximately 300 m, slightly deeper if including on-time information. In some models not displayed here three layers were needed also when including only off-time data. In most other models the GEOTEM system was able to resolve two layers. The PROTEM 47 system is superior in resolving the top layers, where GEOTEM data have very limited resolution.
4.2
Field data
In June 2000, two GEOTEM test ights were own in Denmark. The one addressed here was own in Ringkjbing County in western Jutland. About 500 line kilometres of data were collected altogether, 25 kilometres of which are presented here. The section is chosen
4.2 Field data
13
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
RES1 RES2 RES3 THK1 THK2
(a)
Analysis
1.1 1.2 1.5 2 3
Elevation [m]
Resistivities [m]
3 10 30 100 300
(b)
(d)
(f)
Elevation [m]
(c)
(e)
(g)
2 4 6 8 10 12 14 16 18 20
2 4 6 8 10 12 14 16 18 20
2 4 6 8 10 12 14 16 18 20
Figure 4.1: Double descending model. Panel (a), true models. Panel (b) is PROTEM 47 inversions with the analyses of the true models in panel (c). Panels (d) and (e) are the same for the GEOTEM system with offtime information only. Panels (f) and (g) GEOTEM including the on-time gates. The colours of the analyses indicate the relative uncertainty on each parameter from well determined (red) to undetermined (blue).
because ground-based data have been collected for part of the line as well. However, before presenting the results, the problem with coupling to man-made conductors needs to be addressed. 4.2.1 Coupling
Denmark is densely populated, and the part not covered by cities is mainly used for agricultural purposes. This means that roads with crash barriers and buried cables, power lines, animal fences, telephone cables, etc, are ubiquitous. All these installations give rise to coupling when the primary eld from the transmitter loop is imposed (Srensen et al. 2001). This disturbance is deterministic, arising at the same delay time for all decays summed in the stacking process. A general model for the disturbance from man-made structures is that of an oscillating circuit, and it is normally categorised into two types: galvanic and capacitive coupling. Figure 4.2 presents the two types of coupling taken from the line of data to be presented shortly. A galvanic-type coupling could arise from high-voltage power lines, grounded at each pylon. Animal fences and highway crash barriers are other examples. It can be very hard to recognise on single-site soundings because the whole sounding curve is shifted. An example of a galvanic coupling from a 10 kV power line is shown for a single sounding in Figure 4.2c. In a data sweep (4.2a) this type of coupling can be recognised because it often looks like a response from a vertical thin sheet, as in Figure 3.3. However, data weakly affected by galvanic coupling might be difcult to recognise in the model sections from inversions on the FA because the FA data type tends to subdue (not remove) the inuence
14
Results
of weak galvanic coupling (Figure 4.2b). Coupling effects in data cannot be accurately removed to provide a reliable interpretation. Therefore, soundings distorted by coupling effects should be culled from the data set before interpretation. Models that cannot be trusted are marked in the model section. The footprint or trend of a galvanic coupling can be quite large as indicated by the amount of removed data, but it closely reects the tail of the coupling in Figure 4.2a
10 Amplitude [V/m2]
6
(a)
(c)
(d)
(e)
10
10
10
80 40 0 40 80 120 160 200 240 280 0.4 0.8 1.2
(b)
10 10 Time [ms]
10
10 10 Time [ms]
Resistivities [m]
3 10 30 100 300
10 80 40 0 40 80 120 160 200 240 280
(f)
Elevation [m]
1.6
Elevation [m]
23.4 23.8 24.2 24.6
Profile coordinate [km]
Figure 4.2: Coupling to man-made conductors. (a) is a data sweep of selected gates with the corresponding inversions in (b). x-component data are blue, z-component data are red. (c) is a galvanic-coupled (solid) and uncoupled (solid-dotted) sounding from the data sweep in (a) marked by the vertical lines. Panels (d), (e) and (f) are a capacitive-coupled sounding, data sweep and inverted section, respectively.
A capacitive-type coupling could arise from buried polyurethane-isolated cables. This type of coupling is easily recognised because of its oscillating character as seen in Figures 4.2d and e. The data in Figure 4.2d are clearly not a geological response. The capacitivetype coupling is easily recognised in the model section as well (Figure 4.2f), in this case as an elevation of the good conductor. Altogether, coupling affected 2/3 of the collected data, which were manually identied and removed before interpretation. 4.2.2 The results, Ringkjbing County The eld data and inversions are presented in Figure 4.3. The ight direction is from right to left. On the data sweep (Figure 4.3a), the enlarged sections presented in Figure 4.2 can be identied at prole coordinates 1 km and 24 km. A number of other coupled data sets are easily identied even at this scale (e.g., at coordinates 5, 6.3, 15.2, 18, 19.5, 21.5 and 22.6 km), others are not identiable at this scale. The feature at coordinate 3 km is caused by the aircraft rising to a height of 200 m. The most distinctive feature in the inversion of off-time GEOTEM data (Figure 4.3b) is the conductive layer at large depth, only missing around coordinate 18.5 km. The models generally need two layers to t the data, with three layers in 29% of the models. The GEO-
4.2 Field data
15
Resistivities [m]
Flight direction 3 10 30 100 300
Elevation [m] Amplitude [V/m2]
10 10 10 10
-6
(a)
-7
-8
-9
60 0 -60 -120 -180 -240 -300 60 0 -60 -120 -180 -240 -300 60 0 -60 -120 60 0 -60 -120 -180 -240 -300 0 2 4 6 8 10 12 14 16 18 20 22
(b)
Elev [m]
Elevation [m]
(c)
(d)
Elevation [m]
(e)
24
Figure 4.3: Inversions of the eld data. (a) Full data prole, x-component in red, z-component in blue. (b) is a few-layer inversions of off-time airborne data in (a). (c) includes on-time data in the modelling. (d) is few-layer inversions of PROTEM 47 data from part of the prole and (e) is the model obtained using the CDT inversion scheme.
TEM inversion in which on-time data are included (Figure 4.3c) has a high resemblance to the off-time result. The models from inversions of PROTEM 47 data in Figure 4.3d all have high resistivities with a slightly more conductive feature at the top, especially around prole coordinates 15-22 km. Unfortunately, the data are of very poor quality, possibly affected by the coupling also affecting the GEOTEM data at that part of the line. The data are tted with 2 to 5 layers, but no common features seem to connect the separate soundings. The CDT model section (Figure 4.3e) presents data from the whole line, including the coupled soundings. This is not possible without a heavy smoothing of data before processing, and two undesirable effects arise from this; (1) coupled data, and thereby unpredictably distorted models, are not identied, and (2) the lateral smoothing smears out the distortions of the coupled data to the uncoupled data sets. A number of the distinct features on the CDT prole appear in areas where the data are known to be affected by coupling. Most prominent is the conductive near-surface feature in the prole interval 16-22 km, the part of the prole with the most severely coupled data. It has been shown that unstable deconvolution of slightly noisy data leads to false indications of both poor
16
Discussion and conclusions
and good conductors (Macnae et al. 1998) and this could be one of them. Fugro Airborne Surveys have truncated the CDT prole at the given depth based on an estimated diffusion depth of the last datum followed by a visual evaluation of the model section. The CDT does not see the good conductor. The average resistivities found for the upper parts are similar to the resistivities found in the other inversions. The conductive layer at depth is interpreted to be heavy Paleogene clay known to be at that approximate depth in the area. Seismic sections close to the line (but not on it) also reveal a depression in the clay surface in the area around coordinate 18.3 km (Friborg and Thomsen 1999), as seen on the few uncoupled GEOTEM soundings from that area. The layers above the heavy clay are interpreted to be various unconsolidated sandy and clayey Quaternary and Tertiary sediments. Correlation between the GEOTEM and the ground-based results gives little information on the validity of the inversion of the airborne data. The main depth interval of investigation of the PROTEM 47 is within the rst 100 m as almost all inversions have levelled off to a homogeneous halfspace at this depth. The GEOTEM system has only very limited resolution in the top 100 m, as shown in the synthetic examples.
5
5.1

Coupling to man-made conductors
Coupling is a serious problem when measuring in populated areas, and coupled data have to be removed before interpretation. If not removed, the inversion can be erroneous, and a parameter analysis will be misleading. Removing coupled data is manual work and very time-consuming. Hence, both clients and contractors must understand that the expenses of interpretation and evaluation of airborne data measured in populated areas are comparable to the survey expenses. An automated procedure identifying and removing coupled data sets is desirable, though it must be expected that a fail-safe procedure is not likely to be found and that manual intervention will always be needed. Programmes capable of simultaneously displaying data proles, individual soundings, the ight video, and a map of the survey location would be helpful.
5.2
New strategies
The Field Amplitude data used in this paper are calculated using the post-stack x, y and z-component data. A better solution would be calculation of a pre-stack Field Amplitude, with subsequent stacking as a fourth data type. This procedure would need to be incorporated in the survey design, because at present only post-stack values are saved. This procedure would further reduce the effective noise on the FA data. There is a need for more detailed noise description on airborne TEM data. At the moment no specic information on the noise is available. The following two procedures are suggested incorporated in the survey design when aiming at a full non-linear inversion of the data set, and especially if on-time gates are included; 1) measurements at high altitude with the transmitter turned on. This is standard procedure today, but the time series must be stored and saved to enable a determination of not only average waveform but also its variability. 2) measurement at survey altitude with the transmitter turned off, to estimate the ambient noise at the location. Again full time series are needed. The sum
5.3 Conclusions
17
of the noise estimated from the two contributions would be a good approximation to the total noise. However, two main problems with airborne systems like GEOTEM still exist: the altitude and the speed. This has recently led to the development of an airborne TEM system (SkyTEM) operating from a helicopter (Srensen and Auken 2003). The transmitter loop is towed some 30 metres below the helicopter, minimising unwanted induction effects close to the transmitting system. In the design of the SkyTEM system it has been a key point that the data quality must meet the same standards as data from the ground-based systems, i.e. the resolution of both shallow and deep Earth structures must be comparable. Transmitter waveform information (current, turn-on and turn-off ramp times) and other controlling parameters of the measuring process are recorded for each data subset. This ensures high data quality, as any erroneous readings can be easily identied during data processing. It would be interesting to see the laterally constrained inversion, as presented in the coming sections, applied to the airborne TEM data. The constraints would most likely smooth the general picture while enhancing resolution of weakly determined parameters.
5.3
Conclusions
For a xed wing airborne system it is preferable to use the transmitter waveform as measured by the receiver, as it encompasses all the effects of the system, including the effect of the induced currents in the aircraft and any ltering applied in the receiver system. There is a lot of residual on-time signal in the off-time and ideally, all data should be modelled as on-time to extract the maximum amount of information. Airborne TEM data can be interpreted with success using full nonlinear 1D inversion techniques. Joint inversion of the x- and z-component data turned out disadvantageous compared to the newly proposed strategy using the Field Amplitude. Interpretations and analyses of synthetic models show that airborne data can resolve up to three layers in a geophysical model. Three layers are more common when including the on-time data. Ground-based methods resolve three (or more) layers in most cases. However, the depth of investigation is larger for the GEOTEM system than for the PROTEM 47 system. These observations are supported by the eld data. The main obstacle in an interpretation of eld data is coupling affecting up to 70 % of the measurements in a densely populated area like Denmark. If not identied and removed before interpretation they will result in erroneous models.
18
Layered 2D inversion of resistivity data

6 Introduction
Modern electrical mapping methods are eld-efcient, continuous surface techniques that provide densely sampled data along proles. The systems are either multiple electrode systems (CVES) (e.g. Dahlin 1996; Bernstone and Dahlin 1999) or various pulled systems (Srensen 1996; Panissod et al. 1997). The data coverage provides large sensitivity overlaps between individual soundings which naturally invites for 2D interpretations. Over the last 10 years, 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 often hard to recognise. 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 smeared out. Contrary, in sedimentary environments or for most hydrogeological investigations (Christensen and Srensen 1998) the investigator often suspects a predominantly layered subsurface. This often leads to a situation as in Figure 6.1. An interpreter has an inversion result as presented in Figure 6.1a. Based on this result he has to make a geological interpretation. If he thinks the geology is layered in some sense he will have to visualise formation boundaries based on the rather smooth and smeared inversion result. A possible interpretation might look like Figure 6.1b. I shall treat this example in more detail in Section 8.1.2 and show that it is preferable to use an inversion scheme utilising a layered model description.
Resistivities [m]
15 20 30 0 5 10 15 20 25 200 50 100 170 300 500
Sand
0 5 10 15 20 25
Depth [m]
(a)
Clay ? Sand Sand

200 300
Sand ?
400
(b)
?
300
400
500
500
Figure 6.1: Model interpretation. Panel (a) is an L2-norm inversion of a continuous resistivity data set. Panel (b) is a possible geological interpretation drawn on top of the inversion result. The interpreter assumed the subsurface to be layered.
Often a section of stitched-together 1D models with lateral constraints is sufcient in quasi-layered sedimentary environments (Srensen et al. 2003). However, neotectonics, glaciotectonics or other geological phenomena may have disturbed the sub-horizontal layering disqualifying the 1D formulation to describe the geophysical model. So, a 2D formulation is needed to enable a more complex layered-earth solution. Olayinka and Yaramanci (2000) present a 2D block-type inversion scheme using polygons with homogeneous resistivity. They invert for positions of polygon vertices and for the polygon resistivities and, as is the case for most 2D programmes available, no sensitivity analysis of model parameters accompanies the inversion output. Smith et al. (1999) present a sharp boundary inversion for MT data using a 2D formulation. They use a layered earth descretized along a prole
20
Methodology
with lateral interpolation between neighbouring nodes. The model is regularised using lateral constraints on layer conductivities and depths. The 2D inversion programme for resistivity data presented in this summary is based on the model description of Smith et al. (1999). The inversion is founded on the algorithm developed for a 1D based inversion scheme using lateral constraints on the model parameters described in Paper 2.
FH BE A ?=JE I =JAH= ? IJH=E JI 4F 4D 4F 4D 4F + =OAOJE 4 F FHE =HOF=H= AJAHI 4 D IA? @=HOF=H= AJAHI 5= @ + =O
Figure 6.2: The LCI principle. Lateral constraints, Rp and Rh , connect separate model sets resembling actual geologic variances along the prole.
The fundamental principle of the LCI method is sketched in Figure 6.2. The idea is to control lateral model variations through lateral constraints on resistivities, thicknesses and depths. 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 inversion scheme is tested and compared to standard 2D smooth inversions on both synthetic data and eld data.
7
7.1
Methodology
The Geoelectrical method
7.1.1 Basic equations A geoelectrical experiment is carried out by measuring the electrical potential arising from a current input into the ground for the purpose of achieving information on the resistivity structure in the ground. The geoelectrical problem is a steady-state problem of the electromagnetic problem and thus also described by Maxwells Equations (3-1), (3-2), (3-3) and (3-4). The basic equations used to model the electrical potential are governed only by Faradays law, (3-1), expressed for a steady state problem E= Ohms law of the constitutive relations j = E, and the law of conservation of electrical charge in the quasi-static approximation j= f = 0, t (7-3) (7-2) B = 0, t (7-1)
where E is the electric eld intensity, B the magnetic induction, j the current density, the electrical conductivity in an isotropic medium and f is the free charge.
7.1 The Geoelectrical method
21
Applying the vector identity = 0 to equation (7-3), the electric eld intensity can be expressed as E = , (7-4)
where is the scalar potential eld. The minus sign in equation (7-4) indicates that the potential decreases away from the source. Equation (7-3) applies for source-free regions. If sources are present the current density, j(r), is given by j = I (r rs ), (7-5)
where I is the amount of current injected at rs . Combining equations (7-2) and (7-4) and inserting them in equation (7-5) yields [ (r)(r)] = I (r rs ), (7-6)
which is Poissons equation for the electrical potential. The boundary conditions are given by z = 0 at the earth surface and 0 for |r rs | . 7.1.2 4-electrode measurements
Commonly, geoelectrical eld measurements are made in 4-electrode congurations, as illustrated in Figure 7.1a. The data are often expressed as apparent resistivities a = V K I (7-7)
where V is the measured potential, I is the transmitted current, and K = 2 1 1 1 1 + |rA rM | |rA rN | |rA rM | |rB rN |
1
(7-8)
is dened so that for a homogeneous halfspace the apparent resistivity is equal to the halfspace resistivity. In (7-8), rX is the position of electrode X . Different electrode congura1 8 ) * 8 1
=
>
Figure 7.1: Geoelectrical measurements. (a) is the general situation for a 4-electrode conguration over a homogeneous halfspace. (b) is a sketch of the geometry dependent depth penetration over a layered halfspace.
tions have different sensitivity ngerprints, especially, longer layouts contain information on deeper structures as illustrated in Figure 7.1b.
22
Methodology
7.1.3 Systems The systems applied today all use continuous or semi-continuous measuring techniques in order to cover larger areas with dense measurements. In Denmark the CVES system and the PACES systems have been used extensively in the groundwater mapping campaign. The CVES systems consist of a number of steel electrodes manually forced into the ground at a regular electrode spacing, typically from 2 to 10 m (e.g. Dahlin 1996 or Bernstone and Dahlin 1999). The principle is illustrated in Figure 7.2. The electrodes function as both current and potential electrodes. The measurements are controlled by a computer and can be of any conguration as entered by the user. For example, a CVES layout with 5 m electrode spacing and 4 cables with each 21 electrodes, can do Wenner congurations ranging from 5 m to 130 m electrode spacing. When using the continuous gradient array (presented in section 8.1.3) the number of data is approximately twice the number of data collected with traditional Wenner arrays, and it has proved to be superior in resolving non-horizontal earth structures (Dahlin and Zhou 2002). It collects 3555 data points for a prole 1 km long with a minimum electrode distance of 5 m.
Measure point 1 Cable 1 0m 100 m Cable 2 200 m Cable 3 300 m Cable 4 400 m 500 m
Cable 2
Cable 3
Cable 4
Cable 1
Measure point 2
Figure 7.2: The CVES system. Every tick represents an electrode forced into the ground. The layout is connected to the measuring devices and a small laptop at the measuring points. the measuring is semicontinuous by a roll-along technique moving the backmost cable up front.
The PACES system consists of a small tractor, equipped with processing electronics, pulling electrodes mounted on a tail (Srensen 1996) as illustrated in Figure 7.3. The electrodes are cylindrical steel tubes with a weight of 10 - 20 kg. Two electrodes are maintained as current electrodes with a maximum current of 30 mA, the remaining are potential electrodes. The receiver system records the 8 channels in parallel using a synchronous detection technique with a repetition frequency of 20 Hz followed by a robust prediction ltering scheme (Srensen et al. 2003). The prediction lter subdues the strong noise voltages from the rapidly varying electrochemical activity on the electrode surfaces.
2 JA JE= A A?JH @A +KHHA JA A?JH @A 2K AH?=HHOE C E IJHK A JI ,/25 ,H=CCA@A A?JH @AI
Figure 7.3: PACES system. A small tractor with the electronics pulls a 100 m tail comprising 8 different electrode congurations.
7.2 Inversion setup
23
7.2
Inversion setup
This section will give a brief summary of the LCI methodology presenting the key equations. A more detailed description is found in Papers 2 and 3. 7.2.1 Data and model
Consider a resistivity prole with nx reference node points, xi , in the horizontal direction. For each node point we have an ordinary data vector, di , of apparent resistivity observations for several different electrode spreads. The whole prole section is to be modelled as one inverse problem, and hence the relevant data vector is the concatenation of the data at each node: (7-9) dobs = (d1 , d2 , . . . , dnx )T , where T indicates the vector transpose. For a simple CVES prole the grouping of the data vector as in (7-9) is exemplied in Figure 7.4.
x1 x2 x3
... ...
xnx Nodes Profile distance
Figure 7.4: Data grouping. The individual 4-electrode congurations are grouped into soundings based on the lateral position of their focus point.
To minimise non-linearity and to impose positivity, logarithmic data and logarithmic parameters are used (e.g. Johansen 1977; Ward and Hohmann 1987). Hence di = log(a1 ), log(a2 ), . . . , log(aNi )
T
Pseudo depth
(7-10)
where a denotes apparent resistivity, and Ni is the number of electrode congurations (i.e. number of data) measured at xi . At each surface node the subsurface model is represented by a logarithmic 1D model with nl layers mi = log(i1 ), log(i2 ), . . . , log(inl ), log(ti1 ), log(ti2 ), . . . , log(ti(nl 1) )
T
, (7-11)
where denotes interval resistivity and t denotes interval thickness. The total number of sub-models is nx corresponding to the number of individual 1D soundings. Each submodel is described by nl layers, so the full model m1 m2 m = . , (7-12) . . mnx to be determined has M = nx (2nl 1) parameters.
24
Methodology
1D
t1,1 t2,2 3
2D
t1,1 t2,2 3
Figure 7.5: Model description. The model is described with thicknesses and resistivities at a number of nodes along a prole. The parameters between neighbouring nodes are linearly interpolated to produce a 2D model.
To produce a 2D model from the composite 1D prole, parameters from neighbouring nodes are interpolated linearly as illustrated in Figure 7.5. Thus, locally we have a 1D model description which in combination builds the full 2D prole. The 1D-like grouping of the data and model vectors is not necessary for the 2D inversion, but it enables mixing of 1D and 2D calculations dealt with in Paper 5 and later in this summary. 7.2.2 Forward modelling The 2D resistivity forward modelling in the inversion routine is performed with a nite difference code from University of British Columbia (McGillivray 1992). The code uses a nite difference approach similar to the one described by Dey and Morrison (1979). The layered 2D model, as exemplied in Figure 7.6a is translated to the nite difference grid by superimposing the grid on the model and assigning a resistivity value to each cell based on an area-weighted average of the contributing elements in the underlying layered 2D model (Figure 7.6b). Electrodes are always placed on node points in the grid. Electrodes not coinciding with node points are interpolated linearly to the two nearest nodes.
0 2 4 6 8 10
Depth [m]
0 2 4 6 8 10 12 14 16 18 20
Depth [m] Profile coordinate [m]
(a)
0 2 4 6 8 10
(b)
0 2 4 6 8 10 12 14 16 18 20
Figure 7.6: Model translation. The layered model in (a) is translated to the model super-imposed on the nite difference grid in (b) using weighted averages.
7.2.3 Forward calculations using a sliding model window Long proles or systems with irregular electrode congurations inevitably mean large nite-difference grids. This can make it practically impossible to calculate full forward responses with one large grid. Instead the prole is split up in model windows and calculations are done for one window at the time sliding it along the prole. Afterwards the responses from the windows (dark gray) are concatenated to create the full prole, as illustrated in Figure 7.7a. Experimentally I found the computation time of forward responses to be proportional
7.2 Inversion setup

Start profile End profile
25
Coordinate of parameter
(a)
(b)
Figure 7.7: Sliding model window. (a) shows the sliding model windows for forward responses, whereas (b) is for derivatives. Light grey and dark grey are the ne parts of the grid. Electrodes need to be within the ne parts of the grid with the lateral focus point in the dark grey. The light grey are overlaps between the sliding model windows and have to be sufciently large to ensure continuous forward responses when concatenating the full response of the dark gray sections. White areas are padding.
to n2 log2 (n) for n cells1 , when the number of vertical cells have been chosen according to the given survey. Thus, an optimal choice of the grid size can reduce the computation time drastically. A large grid means fewer but computationally more expensive computations, whereas a small grid means more but faster computations. The grid size is based on the total size of the prole and the electrode congurations used, keeping in mind that the overlap between neighboring model windows always needs to be sufciently large in order to ensure continuous forward data along the prole. The sizes of the overlap, the padding zones (white) and the number of vertical nodes are xed when appropriate values have been determined for the actual conguration. Besides these xed conguration dependent parameters the grid can still be optimized in terms of the size of the central sliding window (SW) (dark gray in Figures 7.7 and 7.8). Figure 7.8 shows the relative timing of two different congurations on a 1100 m prole, assuming a n2 log2 (n) dependence.
10 10 13 3 1 0.3 0.1 0 200 (400,1) 70 SW 70 13 3 1 0.3 0.1 0 (140,0.36) 200 400 600 13 60 SW 60 13
Normalized time
a) CVES, 400 m layout

400 600 800 1000
b) PACES
800 1000
SW [m]
SW [m]
Figure 7.8: Optimising the sliding model window. (a) shows a 1100 m CVES prole with a 400 m spread and 5 m horisontal grid size. (b) shows a 1100 m PACES prole (approx. 90 m spread) with 1 m grid. Numbers in the boxes are the number of horisontal grid points in each separate part. The thick solid line is the normalized computation time, relative to the time using just one grid, for varying sizes of the central sliding window (SW). Only a whole number of sliding windows is permitted, explaining the jagged appearance. The thin solid line is the underlying function permitting fractions of sliding windows.
Figure 7.8a shows the relative timing of a 400 m CVES layout with 5 m between horisontal grid points in the ne part. In this case, the layout is quite large compared to the prole length and there is no time gained using the slider. The shortest computation time
1 This is an approximation, because the computation time depends differently on the number of cells in the two directions. Also the number of current electrodes needs to be taken into account, but for the array types described here the correlation is very good for proles up to 3 km.
26
Methodology
is achieved when choosing the SW to be 400 m bringing the total size of the ne grid to 1100 m (400 m + 2*70*5 m) equal to the prole length. In Figure 7.8b the same prole is measured with the PACES system (Srensen 1996) which has approximately 90 m spread and the horisontal grid size is 1 m. For the PACES system the optimal size of the SW is 140 m (i.e. 140 cells) which is almost three times as fast as if the full prole where computed with just one grid. A longer prole would have made the time saving even larger. The model window is also used to speed up calculation of forward responses for calculation of derivatives, taking advantage of the fact that derivatives for a certain parameter are close to zero for measurements that are relatively far away. Hence, forward responses for derivatives are calculated with only one window centred at the lateral position of the parameter in question, as sketched in Figure 7.7b. In this summary the sliding model windows is used throughout, both for full forward responses as well as for calculation of derivatives. 7.2.4 Forward mapping The dependence of apparent resistivities on subsurface parameters is in general described as a non-linear differentiable forward mapping. The observed data, dobs , are expressed following the established practice of linearised approximation by the rst term of the Taylor expansion dobs (7-13) = g(mref ) + G(mtrue mref ) + eobs , where g is the nonlinear forward mapping of the model to the data space. The true model, mtrue has to be sufciently close to some arbitrary reference model, mref , for the linear approximation to be a good one. In short it becomes dobs = G mtrue + eobs . The Jacobian, G, contains all the partial derivatives of the mapping Gst = mt as ds log(as ) = = , mt log(mt ) as mt (7-15) (7-14)
for the sth apparent resistivity in the data vector and the tth parameter in the model vector. 7.2.5 Lateral constraints The backbone of the LCI method is, naturally, the lateral constraints. The constraints tie the parameters together laterally which transfers information from model parameters with a low variance through the lateral constraints to model parameters with higher variance. The code utilises lateral constraints on resistivities, thicknesses and depths. Constraints on thicknesses are favourable whenever there is a possibility of discontinuous layer boundaries, but continuous thicknesses, e.g. across a fault. Constraints on depths are preferred in cases where there is a demand for continuity of layer boundaries, e.g. a Quaternary sequence with sand and clay layers on top of a relatively smooth pre-Quaternary surface. Because the inversion algorithm deals with both 1D and 2D applications the model parameters are divided into primary (resistivities and thicknesses) and secondary parameters (depths). The constraints are added separately for the two.
7.2 Inversion setup
27
Formally, the constraints on primary parameters are connected to the true model as Rp mtrue = rp + erp , where erp is the error on the constraints with 0 as expected value and rp = Rp mref , making the effective constraints Rp mtrue + erp = 0, (7-18) (7-17) (7-16)
where the roughening matrix Rp , contains 1 and -1s for the constrained parameters, 0 in all other places, e.g. Rp = 1 0 0 1 0 . . . 0 0 0 0 1 0 0 1 0 . . . 0 1 0 0 0 0 0 0 . . . . (7-19)
0 1
The variance, or strength of the constraints, is described in the covariance matrix CRp . For most applications only constraints between neighbouring models are used, but can be between any two sub-models. Normally we use only lateral constraints, but vertical constraints can be applied as well (Farquharson and Oldenburg 1998). The lateral constraints on depths are implemented similarly to the constraints on the primary parameters, but the roughening matrix, Rh , needs to be expressed in terms of the primary model parameters: Rh mtrue = rh + erh , where erh is the error on the depth constraints with 0 as expected value and rh = Rh mref . (7-21) (7-20)
The derivatives with respect to resistivities are zero leaving only derivatives with respect to thicknesses. The constraints on sub-models xk and xk+1 have derivatives with respect to thicknesses given as; (log(hk,l ) log(hk+1,l )) log(hk,l ) log(hk+1,l ) = log(ti,j ) log(ti,j ) log(ti,j ) l ti,j l ti,j s=1 tk,s s=1 tk+1,s = hk,l ti,j hk+1,l ti,j ti,j ti+1,j for i = k and j l hi,j hi+1,j = , 0 otherwise (7-22)
for the lth depth in sub-models xk and xk+1 with respect to the j th thickness in sub-
28
Methodology
models xk and xk+1 . The full matrix with derivatives of depth constraints then becomes 0 0 1 0 0 0 0 0 tk,1 tk,2 0 0 hk,2 hk,2 Rh = . . . . . . . . . . . . . . . tk,1 tk,2 tk,3 tk,n 0 0 hk,n hk,n hk,n hk,n 0 0 1 0 0 0 t +1,1 t +1,2 0 0 hk hk 0 0 k+1,2 k+1,2 . . . . . . . . . . . . . . . tk+1,1 tk+1,2 tk+1,3 tk+1,n 0 0 hk+1,n hk+1,n hk+1,n hk+1,n (7-23) The derivatives on depths with respect to resistivities are zero, indicated by the zero-valued columns in (7-23). The variances on the lateral constraints are given in CRh . 7.2.6 Prior information Prior information, which is used to resolve ambiguities and to add e.g. geological information, can be added at any point of the prole and migrates through the lateral constraints to parameters at adjacent models. It is added in a similar way that lateral constraints were added. However, prior information has not been a subject in my ph.d. work, and I shall not include the equations here. It should be noted, though, that because the model setup is simple and piecewise 1D, the inclusion of prior information is easy and powerful. Papers 2 and 3 have detailed sections on the inclusion of prior information. 7.2.7 Inversion Including both the prior terms and lateral constraints the inversion is written G Pp Ph Rp Rh mtrue = dobs mprior mhprior rp rh + eobs eprior ehprior erp erh , (7-24)
where Pp and Ph refer to the prior information on primary and secondary parameters. mprior = mprior mref and mhprior = hprior Ph mref with errors in eprior and ehprior . hprior contains the logarithm to the prior depth values. Equation (7-24) can be collapsed as + mtrue = d e, (7-25) G with covariance matrix C = C Cprior Chprior 0 CRp CRh 0 . (7-26)
Cobs
29
The model estimate (Menke 1989) 1 G TC mest = G minimises Q= 1 N +A+M

N + A+ M i=1 1
, TC 1 d G
1 2
(7-27)
d2 i var( ei )
(7-28)
where N is the number of data, A is the number of constraints and M is the number of model parameters, including depths. The full solution when expanding (7-27) with respect to (7-24) can be found in Papers 2 and 3. The target mist for the 2D-LCI is in principle zero, which in practice means as low as possible within the computational limits. This would be considered huge overtting for a underdetermined smooth inversion problem, but for a parameterized and overdetermined problem like the 2D-LCI, this is not the case. Due to the restricted number of parameters, the inversion scheme cannot make models arbitrarily complex, which is normally the problem with underdetermined problems. The model complexity is furthermore restricted by the lateral constraints. 7.2.8 Model analysis
The model parameter analysis is based on a linear approximation to the covariance of the estimation error, Cest (Tarantola and Valette 1982), TC 1 G Cest = G
1
(7-29)
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 qs , dened by: STDF(qs ) = exp Cest,ss . (7-30)
8
8.1
Results
Synthetic models
To demonstrate the need for a 2D inversion code utilising layered models, I shall compare results from two synthetic 3-layer models. The synthetic data are generated using the 2D forward code DCIP2D from University of British Colombia (McGillivray 1992). 8.1.1 Von Karman stochastic models
The theoretical models for which synthetic data are created are the result of a stationary stochastic process. The degree of spatial correlation is characterised by the von Karman covariance functions (Mller et al. 2001; Serban and Jacobsen 2001). These functions enable construction of complex models with variations at multiple scales resembling a sedimentary model well. Furthermore, the generated data reproduce eld data to a high degree (Auken et al. 2001), which is also important for the relevance of a synthetic model.
30
Results
Furthermore, the synthetic models are constructed in a sedimentary way, i.e.: 1) bottom layer is deposited as a homogeneous halfspace; 2) bottom layer is eroded; 3) 2nd layer is deposited (draped) on top of the eroded surface of the 1st layer; 4) Erosion of bottom and 2nd layer unit and 5) top layer is deposited on top of the eroded surface. 8.1.2 Model 1 - PACES The rst data set is generated using the PACES electrode conguration, with one sounding per metre, each comprising 8 data points. These data have then had 3 % noise added and been processed to one sounding for each 5 m. The processing is identical to eld data processing. The model (Figure 8.1a) represents a typical sedimentary environment with clay interbedded in sand. Mean layer resistivities are 300 m, 30 m and 300 m for the three layers, respectively. The internal standard deviation on resistivities is 0.5 times the logarithm to the resistivity.
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
Depth [m] Depth [m] Depth [m] Depth [m]
(a)
(b)
(c)
(d)
RES1 RES2 RES3 RES4 THK1 THK2 THK3 DPH3 DPH4
(e)
200 300 400 500 600 700 800 900
100
Resistivities [m]
15 20 30 50 100 170
Profile coordinate [m] Analysis

300 500 1.1 1.2 1.5 2 3
Figure 8.1: Synthetic model. Panel (a) is the true model, with a clay layer (30 m) interbedded in a sandy layer (300 m). Panel (b) shows the inversion result obtained with DCINV2D L2-norm. Panel (c) is the same but with an L1-norm solution. Panel (d) presents the 2D-LCI result with an analysis of the model parameters (resistivities and thicknesses) presented in panel (e). The analysis uses a 6-graded colour code ranging from red (well-determined) to blue (undetermined).
In this example I have chosen to use the DCINV2D programme for comparison (Oldenburg and Li 1994; Farquharson and Oldenburg 2003) mainly because it does strict L1 and L2-norm solutions. All the presented model sections produce data that t the observed data to an acceptable level according to the relevant inversion scheme. Panel (b) of Figure 8.1 presents an ordinary minimum-structure least squares solution (L2-norm) whereas panel (c) is the block-type (L1-norm) inversion result. The apparent
8.1 Synthetic models
31
discontinuities in the proles are due to a required splitting of the prole into smaller sections in order to do inversions in a reasonable time. They both nd the basic surface near features, but have difculties in tracking the deeper parts of the low-resistivity layer. The L2-norm solution is very smooth whereas the L1-norm has some sharp transitions both vertically and laterally. Generally, they both fail to recognise the resistivity transitions as sharp boundaries between the layers in the true model. This is most pronounced in the deeper parts of the model, where the resolution capabilities are modest, and it is more noticeable in the L2-norm result. The 2D-LCI inversion (Figure 8.1c) is carried out with equal-sized constraints on layer resistivities and on the layer boundaries, i.e., depths to layers. The constraint on resistivities is a factor of 1.1 and on depths a factor of 1.3 These values are fairly general and have been tested on a wide variety of models, but were initially established by trial-anderror. With these settings both the overall geometry and the layered nature of the model are found, also for the deeper parts of the model. The 4-layer analysis in panel (e) presents analyses of resistivities and thicknesses along the prole. The analysis is presented with a 6-graded colour-coding of the analyses from equation (7-30). Red corresponds to well-determined parameters and blue is poorly or unresolved parameters. The rst four rows are resistivities of layers 1-4, followed by thicknesses 1-3 and depths to layers 3 and 4. The resistivity of the top two layers is well-determined for most of the prole. The resistivities of the third and fourth layer are well-determined when they are close to the surface (coordinate 100 m to 400 m). The thickness of the rst layer is well-determined for parts of the prole (e.g. coordinate 250 m to 400 m), whereas the thicknesses of the second and third layer are poorly resolved for most of the prole. Although the thicknesses of layers two and three are poorly determined for xk < 500 m given an undoubtedly strong anti-correlation, we see that the depth to the fourth layer is fairly well-determined for that part of the prole. The strength of the lateral constraints, Cp and Ch matrices in equation (7-26), helps resolve the layer boundaries by constraining the depths along the prole. Making the constraints very loose would allow model complexity only bounded by the data. Making very tight constraints would make a very smooth model with slow variations. Thus, determining the constraint values is a trade-off between tting the data (complex models, no constraints) and tting the lateral constraints (smooth models, tight constraints). The starting model is in all cases a layered halfspace (in this case a 4-layer model with the same resistivity in all the layers), making the inversion scheme fairly robust. 8.1.3 Model 2 - Gradient array CVES
This second synthetic data set was used in a statistic model comparison between the layered 2D-LCI approach and ordinary minimum-structure inversion. This is the theme of Paper 4, but is also treated in the next section. The programme used in the comparison study was the Res2dinv (Loke and Barker 1996) in standard settings for the robust inversion1 . The model is shown in Figure 8.2a. The average resistivity values of the three layers in the minimum type models are (top to bottom) 300 m, 25 m and 100 m. This model class could represent a sandy top layer (300 m) overlaying a clay-rich till (25 m)
The programme manual claims that this is similar to an ordinary L1-norm inversion, which is doubtful when comparing to e.g. the DCINV2D programme.
1
32
Results
protecting a groundwater reservoir in a sandy formation (100 m). Again, the internal standard deviation on resistivities is 0.5 times the logarithm to the resistivity. Figure 8.2b presents the result obtained with the Res2dinv programme. The major units are detected, and for the resistive upper layer the layer boundary is identiable. 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 identication based on this prole would probably lead to an overestimate of its true thickness for some parts of the prole.
Depth [m] Depth [m] Depth [m]
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
RES1 RES2 RES3 RES4 THK1 THK2 THK3
(a)
(b)
(c)
(d)
200 300 400 500 600 700 800 900
100
Resistivities [m]
15 20 30 50 100 170

300 500 1.1 1.2 1.5 2 3
Figure 8.2: Minimum type model. The true von Karman 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 2DLCI result is shown in (d). Red colours of the analyses indicate well-resolved parameters, blue poorly and unresolved parameters.
Figure 8.2c presents the 2D LCI result based on a 4 layer model. 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 recognised and identied as a separate unit. Figure 8.2d is the parameter analysis of the model presented in Figure 8.2c. 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 prole coordinate 300-450 m, where the resolution of the thin conductive layer (resistivity and thickness) is quite low compared to other parts of the prole. Based on the synthetic models I conclude that the layered solution has clear advantages over both the L1 and L2-norm minimum-structure solutions when the geological model is overall quasi-layered. The result can be evaluated using the sensitivity analysis from the programme and the identication of layered units is straight forward.
8.2 Statistical model comparison
33
8.2
Statistical model comparison
Knowing the true model, as in the synthetic examples above, enables comparison of the inversion results directly in the model space. Based on the model comparison a statistically based result can be produced, which is the main theme of Paper 4. Here I shall present the statistics as bar diagrams for the model mist, comparing results from the Res2dinv programme with results from the 2D-LCI programme. This kind of evaluation is important in order to justify the use of a layered model as opposed to the existing smooth minimumstructure programmes. The inverted models are evaluated on a point-to-point basis in the model sections. The error, En , on the nth cell is En = exp |ln(true,n ) ln(model,n )| (8-1)
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 0.5 m0.5 m cells for the entire prole length to depths of 80 m. The result of individual cells is then divided into 4 intervals dependent on the size of En . The intervals are < 1.2 for a welldetermined model cell, 1.2 1.5 for a fair determined model cell, 1.5 2.0 for a poorly determined model cell and > 2.0 for an undetermined model cell. The models are of the von Karman type with a gradient array CVES data type as presented in section 8.1.3. A total of 11 km of prole and gradient array CVES data were generated and the forward data from the synthetic models were perturbed to a noise level of 3%. No data processing was performed. The models are grouped in two categories, each with 51.1 km of von Karman model realisations. The two types are a double descending type of model (resistivities decreasing with depth - not shown) and the minimum type model of section 8.1.3 (Figure 8.2). The mean resistivities of the double descending model are 300 m, 70 m and 15 m. As before, the internal standard deviation on resistivities is 0.5 times the logarithm to the resistivity. All models are generated from a three-layer model with the same thickness input for the stochastic modelling. Interval thicknesses of rst and second layers are 17 and 28 m, respectively. This means that, for very long realisations, the mean thickness of layer one will be 17 m, and for the second layer 28 m. The standard deviation on the thicknesses is equal to the mean value. The summarised point-to-point comparisons based on equation 8-1 are shown in Figure 8.3. The result states that the model obtained from the 2D-LCI is equally good or slightly better than the model obtained from the Res2dinv. In both cases, the 2D-LCI is better for the cumulated values below an error factor of 1.5, and both methods show the best result at the double descending models (Figure 8.3b). For error factors larger than 2.0 the 2D-LCI and the Res2dinv have 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. However, it is important to note that the result is highly dependent 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 quasi-layered structure, the Res2dinv is not capable of reproducing the layered-ness of the model, and identication of boundaries becomes difcult.
34
Results
70 60
70
(a)
46 47 71 74 87 87 100 100
60 50 40 30 20 10
52
56
(b)
Percentage
50 40 30 20 10 0
74
81 87 86 100 100
< 1.2
1.21.5
1.52
>2
< 1.2
1.21.5
1.52
>2
Model error (factor)
Model error (factor)
Figure 8.3: Summary of statistics. Red bars are 2D-LCI, blue bars are Res2dinv. Results from the minimum type models are summarised in (a), double descending models are summarised in (b). Bars indicate percentage of point-to-point comparisons in the given intervals. The numbers on top of the bars are the cumulated percentages.
8.3
Field example, CVES, Sweden
As a last example to illustrate the use of the 2D-LCI programme I shall show a CVES eld data set from Sweden, kindly provided by Prof. Torleif Dahlin. The 300 m prole is presented in Figure 8.4. The data are collected using Wenner arrays, with a-distances between 2 m and 48 m. The resistivity survey was carried out as part of the geotechnical investigations for road construction in connection with a lled basin structure in bedrock. The data set has previously been presented by Dahlin (1996). The data pseudo section in Figure 8.4a has relatively smooth transitions but with clear indications of 2D structures. The L2-norm model in Figure 8.4b picks up a basin structure with its thickest appearance around coordinate 120-130 m. The identication of all three layers as depicted in the drill holes is not possible and the depth to bedrock is not easily identied, though the overall model geometry reects the drill data to some degree. For the left half of the prole it seems that the rock identied in the drill holes is of a different character from the rock detected at the right half of the prole based on prominent differences in resistivity. This is most likely due to 3D structures (Dahlin 1996). The L1-norm model (Figure 8.4c) is much more blocky than the L2-norm model, but still, the formations depicted by the drill holes are not picked up. Figure 8.4d presents a 2D-LCI section with 4 layers. No prior information is added. The basin unit is now clearly separated from the bedrock, and for the right half of the prole (coordinates 100 m- 230 m) the till-unit is easily identied on top of the bedrock. The thicknesses of the clay layer and the till clay layer are fairly consistent with those found in the drill holes. For the left half of the prole there are inconsistencies, as was the case for the Res2dinv model. The analyses in Figure 8.4e present mainly well-determined resistivities for the entire prole for layers 1 and 2, whereas the bottom two layers have only moderately resolved resistivities. The thicknesses of the till layer (THK3, coordinates 100-230 m) are poorly resolved, but the depth to the top of the till (DPH3), i.e. the basin depth is well-determined. The model sections produced data tting the observed data to an acceptable level.
35
Elev [m] Depth [m]
0 10 20 30 35 25 15 5 35 25 15 5 35 25 15 5
(a)
(b)
Elev [m]
(c)
Elev [m]
(d)
(e)
20 0 20 40 60 80 100 120 140 160 180 200 220 240 260
Profile coordinate [m] Analysis Resistivities [m]

10 30 100 300 1000 3000 10000 1.1 1.2 1.5 2 3
Figure 8.4: Field example. Panel (a) is the data pseudo section, (b) is the result from an ordinary L2-norm inversion with the corresponding L1-norm inversion in (c). (d) is the 2D-LCI model. The model parameter analyses are presented in (e). The colour-coding of the analyses ranges from well-resolved (red) to poorly resolved (blue). Lithological logs from drillings are located at every 20 metres from coordinate 20 m to 200 m. The colours of the drill holes indicate from the bottom: rock (dark gray), till (light gray) and clay (white).
Optimising the 2D-LCI
Ensuring reasonable computation times is very important. This section deals with optimisations to the original 2D-LCI code presented in Paper 5. Approximations to a full solution for an inversion algorithm can be done on several stages: in the forward solution, in the calculation of derivatives or in the inverse matrix manipulations. However, the savings in computation time is obtained at the expense of the loss of accuracy which is the inevitable consequence of numerical approximations. Thus, there is a trade-of between the savings in computation time and the required accuracy. Loke and Barker (1996) suggested a quasi-Newton formulation which gave major time reductions using Broydens update formula (Broyden 1965). Oldenburg and Ellis (1991), introduced approximate inverse mappings in both the model space (AIM-MS) and in the data space (AIM-DS). They exemplied the AIM-DS with the 2D magneto-telluric (MT) problem, using a 1D formulation as an approximate inverse. This proved powerful with MT data. The Rapid Relaxation Inverse (RRI) by Smith and Booker (1991) also combined 2D and 1D formulations for the MT problem. In the RRI the derivatives are 1D except for the fact that they comply with the 2D elds. An extremely fast multi-channel deconvolution (MCD) on resistivity data was suggested by Mller et al. (2001) where the output is a smooth picture of the subsurface resistivity distribution obtained without direct computation of the 2D elds. Torres-Verdin et al. (2000) used a smaller grid as an approximate 2D nite-difference forward, incorporated in an auxiliary inversion scheme. This inver-
36
sion method is similar to the AIM-DS presented by Oldenburg and Ellis (1991) only the approximate inverse is not a 1D formulation but a 2D formulation using a smaller grid. This section evaluates the 2D-LCI, using Broydens update formula, and the 2D-LCI with 1D derivatives. Also, a combination of the two is tested.
9.1
Broydens update formula
Broydens update formula (Broyden 1965) has been used widely for 2D resistivity inversion. The update formula approximates the Jacobian matrix (equation (7-15)), Gn+1 , of iteration n + 1 using either the Jacobian matrix or the Broyden matrix of iteration n: Gn+1 where dn = dn dn1 mn = mn mn1 (9-2) Bn+1 = Bn + [dn Bn mn ] m T n , m T n mn (9-1)
The update formula is based on an assumption that the difference in forward mappings, d(mn ), changes linearly with respect to mn , in the direction of d(mn ). This is most likely to be the case close to the minimum of the least squares cost-function. Most of the previous implementations have been in smooth minimum-structure programmes (e.g. Loke and Barker 1996). The cell-based nite-difference or nite-element implementation of the forward problem in these programmes implies that only resistivities are inverted for. When parameterizing the problem into, for example, thicknesses and resistivities of discrete layers we introduce a mixture of physical parameters and structure. Most likely, the forward mapping is in this case more linearly dependent on the physical parameters than on the structure. Torres-Verdin et al. (2000), implemented Broydens update formula on a parameterised cross-borehole problem. They experienced a need to reset the entries in the Jacobian matrix after a few iterations, doing the full, time-consuming solution. Otherwise the problem did not converge. This is not the case with the minimum-structure solutions. Loke and Dahlin (2002) found that if the Jacobian matrix is initiated and perhaps fully recalculated for the rst two or three iterations the nal model will be very close to the full Gauss-Newton model. Based on these considerations I implemented Broydens update formula with an option to reset and recalculate the Jacobian if necessary. As the criterion to recalculate the full Jacobian matrix the change in residual between the latest two iterations is used. The value is dened by the user, but 5-10% seems to be suitable.
9.2
1D derivatives
The derivatives in the Jacobian matrix are approximated by a rst-order forward-difference formula.This implementation requires at least one forward calculation per model parameter and one forward calculation per model. In the parameterised 2D case this means M + 1 forward calculations, which is very time-consuming for large problems. Instead it is possible to calculate the entries to the Jacobian matrix with a 1D forward mapping. For the 1D resistivity case a forward calculation is extremely fast, practically eliminating the Jacobian calculation as a time factor. Only the data from individual soundings corresponding
9.3 Combining Broydens update and 1D derivatives
37
to the grouping of the model vector now contribute to the total G-matrix making this a block-diagonal: d 0 m x=x1 d m x = x 2 , G= (9-3) .. . d 0 m x=xn
x
for the nx models with nx data sets. Now, for one iteration only one full 2D calculation on the model from the previous iteration and nx 1D calculations is needed to ll the Jacobian matrix, G. For a subsurface resistivity structure close to a 1D model this will work ne. For very complex 2D (or 3D) structures the approximation will be poor, since the offdiagonal elements in the 2D Jacobian matrix will have non-negligible values compared to the block-diagonal elements.
9.3
Combining Broydens update and 1D derivatives
Since the full 1D solution, 1D derivatives, the full 2D solution and Broydens update are all incorporated in the inversion scheme it is obvious to test a hybrid for an optimal solution in terms of speed and accuracy. Normally the iterative solution is started with a homogeneous halfspace. This means that very little 2D information is present in the beginning. Therefore, I initiate the inversion procedure with 1D derivatives, at some point switching to doing full 2D solutions and subsequently to Broydens update. If the subsurface is thought to be fairly 1D, the 1D derivatives can be exchanged with the full 1D solution. Generally, this iteration scheme should minimise the number of full 2D solutions while keeping as much 2D information as possible. Also, the Broyden solution should be stabilised because the model mist will be closer to the desired mist minimum, minimising nonlinearity errors.
9.4
9.4.1
Examples
Simple dip-model, PACES data
The rst example in Figure 9.1a, is a very simple two-layer model with a 45-degree descending slope on the lower layer, assuming the PACES system setup. Data were generated every 1 m, then 5% noise were added to the data which were then processed to one sounding per 5 m as if they were eld data. The prole is 600 m long with 960 data points in 120 models (660 model parameters). Panel b) presents a full 1D inversion which does not resolve the 2D slope. Panel c) is the full 2D solution, which nds the correct geometry except for the wiggles on the deep parts of the conductive layer due to noisy data. Panel d) presents the solution using Broydens update for all iterations after initiating the inversion with two full 2D iterations. From the model section and the residual plot on the right it is clear that Broydens update as stand-alone does not converge to a satisfactory model. Panel e) is Broydens update with full 2D recalculation of the Jacobian whenever the change in residual is less than 5%. The model now converges towards the true model, but more iterations are needed compared to the full 2D solution. Panel f) is the solution using 1D derivatives for all iterations. Iteration for iteration the residual of this method follows the residual of the 1D solution
38
Depth [m]
0 10 20
30
(h)
(a) True model
Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20 10
(b) Full 1D
Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(c) Full 2D
Data residual
Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(d) Broyden
1
Depth [m ] Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(e) Broyden 5%
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(f) 1D Derivatives
Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
0.3
10
15
20
Iteration # Resistivities [m]

3 10 30 100 300
(g) 1D Derivatives & Broyden 5%

200 220 240 260 280 300 320 340 360 380 400
Figure 9.1: Synthetic PACES example. Panel a) is the true model. Panel b) is the model from a full 1D inversion; c) is the result using a full 2D solution. Panel d) presents an inversion using Broydens update formula for all iterations initiated with two full 2D iterations. Panel e) also uses Broydens update formula, but the full Jacobian is recalculated when the residual change is less than 5 %. The model in panel f) is the result of 1D derivatives throughout. Panel g) combines 1D derivatives, with the full 2D solution and Broydens update formula. Finally, the data residuals are presented in h) as a function of iteration number. Filled markers (for Broyden 5% and 1D derivatives+ Broyden 5%) indicate full 2D iterations.
(Figure 9.1h) except for the last two iterations. This indicates that the 2D structures are found at the very last iterations. The nal model is something in between the model from the full 2D solution (Figure 9.1c) and the model from the full 1D solution (Figure 9.1b). Finally, panel g) presents the combination of 1D derivatives with a Broydens update on successive iterations. 1D derivatives were carried out for the rst 4 iterations, then one full 2D solution and nally Broydens update with the 5% criterion. The timing of the examples in Figure 9.1 is summarised in Table 9.1. All inversions are tested on a 2.8 GHz Pentium-4 processor. The primary time consumer in this case is clearly the 2D iterations. For larger problems
9.4 Examples
39
Method (b) Full 1D (c) Full 2D (d) Broyden (e) Broyden 5% (f) 1D derivatives (g) 1D derivatives & Broyden 5%
Computation time 5s 8779 s 1627 s 4397 s 284 s 2415 s
# Iterations 6 11 13 18 6 9
# 2D iterations 0 11 2 5 0 3
Table 9.1: Summary of computation times and number of full 2D iterations for the inversions presented in Figure 9.1.
the time taken to do the matrix calculations can be signicant as well, although it can be minimised, taking advantage of fairly sparse matrices. From this simple synthetic example it is evident that Broydens update by itself is insufcient for this type of problem. If combined with full 2D updates of the Jacobian matrix, convergence is assured. Using 1D derivatives works quite well with this simple geometry, and combining them with a full 2D solution using Broydens update speeds up the computation time without losing 2D information. However, for more complex structures the 1D derivatives as stand-alone will be insufcient. 9.4.2 CVES, road construction, Sweden
This example further investigates the eld example of section 8.3. This example only presents the full 2D inversion (Figure 9.2b)1 , the Broyden solution with the 5% criterion on the 2D update (Figure 9.2c) and the combined Broyden and 1D derivatives (Figure 9.2d). The model results from the different inversions are practically the same. Minor differences are obtained in the outer lower parts of the prole, where practically no information is present, as depicted by the pseudo section (Figure 9.2a). All models show relatively good correlation with the information from the drillings, which identies three distinct units in the small basin. Method (c) Full 2D (e) Broyden 5% (g) 1D derivatives & Broyden 5% Computation time 9891 s 5074 s 3291 s # Iterations 10 13 13 # 2D iterations 10 5 3
Table 9.2: Summary of computation times and number of full 2D iterations for the inversions presented in Figure 9.2.
The timing in Table 9.2 shows that the combination of 1D derivatives with Broydens update brings the effective number of full 2D solutions down to only three, which is very acceptable compared to the 10 needed if no approximations are used. In the examples given here it makes good sense to use either of the approximations suggested, but in both cases the combination of 1D derivatives with a Broydens update had the shortest computation time and thus recommends itself. However, models with strong 2D features would most likely be difcult to optimize using the 1D derivatives
This 2D result is actually slightly different from the example presented previously due to some ne tuning of the code implemented in the mean time.
1
40
Discussion
Depth [m]
0 10 20 30
30
(e)
(a) Pseudo section

0 50 100 150 200 250 10
Elevation [m] Elevation [m] Elevation [m]
25 15 5 35 25 15 5 35 25 15 5
(b) Full 2D
0 50 100 150 200 250
Data residual
3 1 0
35
(c) Broyden 5%
0 50 100 150 200 250
Iteration # Resistivities [m]

30 100 300
10
15
(d) 1D Derivatives & Broyden 5%

0 50 100 150 200
250
10
1000
3000
10000
Profile coordinae [m]

Figure 9.2: CVES eld example. Panel a) is the data pseudo section, panel b) is the full 2D solution, panel c) uses Broydens update formula with a 5% recalculation criterion. Panel d) combines 1D derivatives with a 5% Broydens update. Finally, the data residuals versus iteration number are plotted in e). Filled markers (for Broyden 5% and 1D derivatives+ Broyden 5%) indicate full 2D solution. Lithological logs from drillings are located at every 20 metres from coordinate 20 m to 200 m. The colours of the drill holes indicate from the bottom: rock (dark gray), till (light gray) and clay (white).
solution in which cases either the full 2D or the 2D with Broyden 5% should be used. More examples are shown in Paper 5. Extremely large data sets can be practically impossible to invert even with the approximate solutions suggested here. Normally a full 1D solution would be the alternate choice, but instead I suggest to use 1D derivatives throughout producing an intermediate result between the 1D and the 2D solutions.
10
10.1
Discussion
Smooth minimum-structure or layered?
This is really the key question one has to ask. Often an interpreter has some background knowledge of the area and often this knowledge comes in terms of distinct units with different electrical properties. Therefore, when inverting resistivity data the result is often visualised in terms of separate units. Unfortunately, this is difcult with a smooth minimum-structure formulation which has been the main interpretation tool over the last years. This initiated the development of the LCI formalism, rst in the 1D case, later incorporating the 2D case as well. I suggest the use of the 2D-LCI whenever these units take the form of a quasi-layered environment. Whenever they are not quasi-layered a minimum-structure solution should be used. Generally, it is important to consider the dimensionality of the underlying model before applying a layered inversion scheme and subsequently choose the appropriate interpretation tool.
10.2 Using layered 2D inversion with other data types
41
10.2
Using layered 2D inversion with other data types
The 2D-LCI presented in this paper is intended for resistivity data, but the methodology is general and it is straight forward to implement the solution for any data set for which a 2D or 3D forward code exists.
10.3
Code optimising
In order to make the code functional for large data sets it was optimised using Broydens update formula and 1D derivatives. However, it is easy to imagine models that would not benet from the approximate solutions. Models with strong 2D features would most likely be difcult to optimise using the 1D derivatives solution, although strong 2D features would cause problems for most 2D inversion schemes due to constraints. Finally, any 3D feature causes problems to a 2D inversion programme. The solution to this is to implement a 3D forward code in the LCI programme and then constrain layer parameters in both lateral directions. A number of additional approximations could be implemented. Torres-Verdin et al. (2000) used a dual-grid approach to speed up computations. I tried applying a ne grid for forward calculations and a coarser grid for calculation of derivatives, but accuracy was lost when reducing the number of cells, and I skipped the dual-grid approach though it might have interesting perspectives if used as an initiator similar to 1D derivatives described above. Common to many approximations is the objective to achieve a reasonable starting model before doing the time consuming full 2D solutions. The multi-channel deconvolution as suggested by Mller et al. (2001) is a fast tool to produce a smooth deconvolved model. Having produced a smooth 2D model we only need to nd the best-tting layered models along the prole and use those as starting models.
10.4
Code improvements
A few issues still remain in order to improve the 2D-LCI on some key points. Inclusion of a L1-norm solution would stabilise the inversion of noisy data sets, and would perhaps reduce the number of iterations because the L1-norm in nature is blocky. At this point the LCI operates with relative constraints. The option of absolute constraints on thicknesses and depths would be appreciated because they would be easier to control for layers covering large depth scales.
11
Conclusions
The LCI method utilises a layered model description for 1D and 2D resistivity inversion. The models and data are grouped into pairs of soundings and models along the prole. The parameters in the LCI are regularised using lateral constraints on layer resistivities, thicknesses and/or depths, producing laterally smooth models with discrete layers. The layered model description makes identication of formation boundaries easy, as compared to standard minimum-structure 2D algorithms, which produce a smeared picture of the geological model. The output from the LCI is supported by a full sensitivity analysis of the model parameters entering the inversion scheme giving the interpreter a chance to ascertain the
42
Conclusions
inversion result. Statistical comparison of model sections obtained from the 2D-LCI and a smooth minimum-structure programme suggested that equally good or better results are obtained with the 2D-LCI. Thus, if the geology is expected to be layered in some sense, I suggest using a layered inversion scheme in order to make identication of formation boundaries easy. An inversion scheme combining 1D derivatives with a full 2D solution and Broydens update speeded up the computations approximately by a factor of 3.
REFERENCES
43
References
Annan, A. P., R. S. Smith, J. Lemieux, M. D. OConnell, and R. N. Pedersen (1996). Resistive-limit, time-domain AEM apparent conductivity. Geophysics 61(01), 93 99. Auken, E., N. Foged, and K. I. Srensen (2001). Model recognition by 1-D laterally constrained inversion of resistivity data. See Hill (2001). Bernstone, C. and T. Dahlin (1999). Assessment of two automated DC resistivity data acquisition systems for landll location surveys: Two case studies. Journal of Environmental and Engineering Geophysics 4(2), 113121. Broyden, C. G. (1965). A class of methods for solving nonlinear simultaneous equations. Mathematics of Computation 19, 577593. Christensen, N. B. and E. Auken (1992). Simultaneous electromagnetic layered model analysis. In B. H. Jocobsen (Ed.), Proceedings of the Interdisciplinary Inversion Workshop 1, Geoskrifter 41, pp. 4956. Aarhus University. Christensen, N. B. and K. I. Srensen (1998). Surface and borehole electric and electromagnetic methods for hydrogeological investigations. European Journal of Environmental and Engineering Geophysics 3, 7590. Christiansen, A. V. and N. B. Christensen (2002). Application and analysis of airborne transient electromagnetic methods in Denmark. the role of on-time data. In Proceedings of the Symposium on the Application of Geophysics to Environmental and Engineering Problems (SAGEEP), Las Vegas. The Environmental and Engineering Geophysical Society. Christiansen, A. V. and N. B. Christensen (2003). A quantitative appraisal of airborne and ground-based transient electromagnetic (TEM) measurements in Denmark. Geophysics 68(2), 523534. Dahlin, T. (1996). 2D resistivity surveying for environmental and engineering applications. First Break 14(7), 275283. Dahlin, T. and B. Zhou (2002). Gradient and mid-point referred measurements for multi-channel 2D resistivity imaging. In M. S. Matias and C. Grangeia (Eds.), Proceedings of the 9th Meeting, Environmental and Engineering Geophysics, Aveiro, Portugal, pp. Integrated Case Histories session. Environmental and Engineering Geophysical Society - European Section. DeMoully, G. T. and A. Becker (1984). Automated interpretation of airborne electromagnetic data. Geophysics 49(8), 13011312. Dey, A. and H. F. Morrison (1979). Resistivity modeling for arbitrarily shaped twodimensional structures. Gophysical Prospecting 27, 106136. Effers, F., E. Auken, and K. I. Srensen (1999). Inversion of band-limited TEM responses. Geophysical Prospecting 47, 551564. Farquharson, C. and D. Oldenburg (2003). Constructing piece-wise-constant models using general measures in non-linear, minimum-structure inversion algorithms. In 6th International Symposium of the Society of Exploration Geophysicists of Japan, Tokyo. Society of Exploration Geophysicists of Japan.
44
REFERENCES
Farquharson, C. G. and D. W. Oldenburg (1998). Non-linear inversion using general measures of data mist and model structure. Geophysical Journal International 134, 213227. Fitterman, D. V. and M. T. Stewart (1986). Transient electromagnetic sounding for groundwater. Geophysics 51(04), 9951005. Friborg, R. and S. Thomsen (1999). (In Danish) Kortlgning af Ribe Formationen. Et fllesjydsk grundvandssamarbejde. (Mapping the Ribe Formation. A cooperate groundwater investigation project.). Populr rapport, Ribe, Ringkjbing, Viborg, Arhus, Vejle og Snderjyllands Amt. Hill, I. (Ed.) (2001). Proceedings of the 7th Meeting, Environmental and Engineering Geophysics, Birmingham, England. Environmental and Engineering Geophysical Society - European Section. Johansen, H. K. (1977). A man/computer interpretation system for resistivity soundings over a horizontally stratied earth. Geophysical Prospecting 25, 667691. Liu, G. and M. W. Asten (1993). Conductance-depth image of airborne TEM data. Exploration Geophysics 24, 655662. Loke, M. H., I. Acworth, and T. Dahlin (2001). A comparison of smooth and blocky inversion methods in 2-D electrical imaging surveys. Proceedings of the ASEG 15th geophysical conference and exhibition, Brisbane, August 2001. Loke, M. H. and R. D. Barker (1996). Rapid least-squares inversion of apparent resistivity pseudosections by a quasi-Newton method. Geophysical Prospecting 44, 131152. Loke, M. H. and T. Dahlin (2002). A comparison of the Gauss-Newton and quasi-Newton methods in resistivity imaging inversion. Journal of Applied Geophysics 49(3), 149162. Macnae, J., A. King, N. Stolz, A. Osmakoff, and A. Blaha (1998). Fast AEM data processing and inversion. Exploration Geophysics 29(1/2), 163169. Macnae, J. and Y. Lamontagne (1987). Imaging quasi-layered conductive structures by simple processing of transient electromagnetic data. Geophysics 52(04), 545554. Macnae, J. C., R. Smith, B. D. Polzer, Y. Lamontagne, and P. S. Klinkert (1991). Conductivity-depth imaging of airborne electromagnetic step-response data. Geophysics 56(01), 102114. McGillivray, P. (1992). Forward modeling and inversion of DC resistivity and MMR data. Ph.D. thesis, The University of British Columbia, Vancouver, Canada. McNeill, J. D. (1990). Use of electromagnetic methods for groundwater studies. In S. H. Ward (Ed.), Geotechnical and environmental geophysics, Volume 01, pp. 191218. Society Of Exploration Geophysicists. Meju, M. A., S. L. Fontes, M. F. B. Oliveira, J. P. R. Lima, E. U. Ulugergerli, and A. A. Carrasquilla (1999). Regional aquifer mapping using combined VES-TEMAMT/EMAP methods in the semiarid eastern margin of Parnaiba Basin, Brazil. Gephysics 64(2), 337356. Menke, W. (1989). Geophysical Data Analysis: Discrete Inverse Theory (Revised edn. ed.). Academic Press Inc.
REFERENCES
45
Mller, I., B. Jacobsen, and N. Christensen (2001). Rapid inversion of 2-D geoelectrical data by multichannel deconvolution. Geophysics 66(3), 800808. Munkholm, M. S. and E. Auken (1996). Electromagnetic noise contamination on transient electromagnetic soundings in culturally disturbed environments. Journal of Environmental and Engineering Geophysics 1, 119127. Nabighian, M. N. (Ed.) (1987). Investigations in Geophysics No.3. Society of Exploration Geophysics. Olayinka, A. I. and U. Yaramanci (2000). Use of block inversion in the 2-D interpretation of apparent resistivity data and its comparison with smooth inversion. Journal of Applied Geophysics 45, 6381. Oldenburg, D. W. and R. G. Ellis (1991). Inversion of geophysical data using an approximate inverse mapping. Geophysical Journal International 105, 325353. Oldenburg, D. W. and Y. Li (1994). Inversion of induced polarization data. Geophysics 59, 13271341. Panissod, C., M. Lajarthe, and A. Tabbagh (1997). Potential focusing: a new multielectrode array concept, simulating study, and eld tests in archaeological prospecting. Geophysics 38, 123. Pedersen, R. and S. Thompson (1991). GEOTEM case history: Appendix G of Time domain electromagnetic prospecting methods, by M. N. Nabighian and J. C. Macnae. See Nabighian (1987), pp. 509513. Poulsen, L. H. and N. B. Christensen (1998). Hydrogeophysical mapping with the transient electromagnetic sounding method. European Journal of Environmental and Engineering Geophysics 3, 201220. Sandberg, S. K. (1993). Examples of resolution improvement in geoelectrical soundings applied to groundwater investigations. Geophysical Prospecting 41, 207228. Serban, D. Z. and B. H. Jacobsen (2001). The use of broad-band prior covariance for inverse palaeoclimate estimation. Geophysical Journal International 147, 2940. Smith, J. T. and J. R. Booker (1991). Rapid inversion of two- and three-dimensional Magnetotelluric data. Journal of Geophysical Research 96(B3), 39053922. Smith, J. T., M. Hoversten, E. Gasperikova, and F. Morrison (1999). Sharp boundary inversion of 2D Magnetotelluric data. Geophysical Prospecting 47, 469486. Smith, R. S. (2000). The realizable resistive limit: A new concept for mapping geological features spanning a broad range of conductances. Geophysics 65(4), 11241127. Smith, R. S. (2001a). On removing the primary eld from xed-wing time-domain airborne electromagnetic data: some consequences for quantitative modelling, estimating bird position and detecting perfect conductors. Geophysical Prospecting 49, 405416. Smith, R. S. (2001b). Tracking the transmitting-receiving offset in xed-wing transient EM systems: methodology and applications. Exploration Geophysics 32, 1419. Smith, R. S. and S. J. Balch (2000). Robust estimation of the band-limited inductivelimit response from impulse-response TEM measurements taken during the transmitter switch-off and the transmitter off-time: Theory and an example from Voiseys Bay, Labrador, Canada. Geophysics 65(2), 476481.
46
REFERENCES
Smith, R. S. and P. B. Keating (1996). The usefulness of multicomponent, time-domain airborne electromagnetic measurements. Geophysics 61(01), 7481. Smith, R. S. and T. J. Lee (2002). Using the moments of a thick layer to map the conductance and conductivity from airborne electromagnetic data. Journal of Applied Geophysics 49(3), 173183. Srensen, K. and E. Auken (2003). New developments in high resolution airborne TEM instrumentation. In Proceedings of the ASEG 16th Geophysical Conference, Volume 26, 2/3. Australian Society of Exploration Geophysicists. Srensen, K., E. Auken, N. B. Christensen, and L. Pellerin (2003). An Integrated Approach for Hydrogeophysical Investigations: New technologies and a Case History, accepted for publication in SEG Special Publication, Near-Surface Geophysics, Volume 2. Society of Exploration Geophysicists. Srensen, K., P. Thomsen, E. Auken, and L. Pellerin (2001). Effect of coupling in electromagnetic data. See Hill (2001), pp. 108109. Srensen, K. I. (1996). Pulled array continuous electrical proling. First break 14, 85 90. Tarantola, A. and B. Valette (1982). Generalized non-linear inverse problems solved using the least squares criterion. Reviews of Geophysics and Space Physics 20(2), 219232. Torres-Verdin, C., V. L. Druskin, S. Fang, L. A. Knizhnerman, and A. Malinverno (2000). A dual-grid nonlinear inversion technique with applications to the interpretation of dc resistivity data. Geophysics 65(6), 17331745. Ward, S. H. and G. W. Hohmann (1987). Electromagnetic theory for geophysical applications. See Nabighian (1987), Chapter 4, pp. 285426. Wolfgram, P. and G. Karlik (1995). Conductivity-depth transform of GEOTEM data. In J. C. Dooley (Ed.), Proceedings of the ASEG 13th Geophysical Conference, Volume 26, 2/3, pp. 179185. Australian Society of Exploration Geophysicists.
PAPER 1
A Quantitative appraisal of airborne and ground-based transient electromagnetic (TEM) measurements in Denmark
by
Anders Vest Christiansen and Niels Bie Christensen
Department of Earth Sciences, University of Aarhus, Finlandsgade 6-8 DK-8200 Arhus N, Denmark
Geophysics, vol. 68, No. 2, 523-534, March-April 2003
GEOPHYSICS, VOL. 68, NO. 2 (MARCH-APRIL 2003); P. 523534, 9 FIGS., 1 TABLE. 10.1190/1.1567220
A quantitative appraisal of airborne and ground-based transient electromagnetic (TEM) measurements in Denmark
Anders Vest Christiansen and Niels Bie Christensen
ABSTRACT
The last decade has seen growing use of ground-based transient electromagnetic (TEM) methods in Denmark for hydrogeological purposes. Due to an intensied mapping campaign, airborne TEM methods were proposed as a possible tool for mapping large areas. The rst test ights were own in June 2000 using the GEOTEM system. Traditional approximate interpretation tools for airborne data are insufcient in hydrogeological investigations where a quantitative model specifying model parameter reliability is needed. We have carried out full nonlinear one-dimensional inversion on the eld amplitude of airborne synthetic and eld data and compared the airborne method with the traditional ground-based PROTEM 47 system that has found extensive use in Denmark. An improved measuring procedure for airborne systems is suggested to facilitate the
estimation of noise that is necessary in a quantitative inversion. The analyses of synthetic data demonstrate the differences in resolution capability between ground-based and airborne data. Ground-based data typically resolve three- or four-layer models and occasionally up to ve layers. Airborne data resolve three layers as a maximum, one or two layers being common. The airborne GEOTEM system detects layers to depths of more than 300 m, bearing only little information about the top 50 70 m. The ground-based PROTEM 47 system has a maximum penetration of approximately 170 m, with higher resolution capabilities in the top 100 m. Coupling to man-made conductors is a serious problem for all TEM methods in densely populated areas and results in distorted data. Coupling inuences the airborne data from Denmark on two-thirds of the area covered. These data must be eliminated to avoid misinterpretation.
INTRODUCTION
The last decade has seen growing use of geophysical methods for general geological mapping purposes as well as for detailed mapping of the extent and vulnerability of aquifers (Fitterman and Stewart, 1986; McNeill, 1990; Christensen and Srensen, 1998). Two ground-based methods have proved particularly useful in Denmark: transient electromagnetic (TEM) soundings and geoelectrical sounding/proling. The most widely used geoelectrical methods employ the systems of continuous vertical electrical sounding with multielectrode systems (CVES) and the pulled-array continuous vertical electrical sounding system (PA-CVES) (Srensen, 1995), which have been employed mainly for near-surface mapping to determine aquifer vulnerability. The TEM method is primarily used for delineating the lower boundary of aquifers, and more than
30,000 soundings have been made in Denmark for this purpose (Poulsen and Christensen, 1998). Recently, a pulled-array continuous TEM (PA-TEM) system has been developed allowing much faster data acquisition (Srensen et al., 2000). Through an agreement between the Danish counties and the national government, it was decided to map more than 20 000 km2 (about half the area of Denmark) over the next decade using geophysical methods to obtain a thorough knowledge of the groundwater resources. The estimated total cost is US$100 million, nanced through water taxes, a project unique in both intention and proportions. The project scope has drawn attention to the requisite geophysical methods. The continuous PA-TEM system is much faster than a single site stationary system, but is still insufcient for a project of this size, and airborne TEM methods, which offer high efciency, have been proposed.
Published on Geophysics Online November 15, 2002. Manuscript received by the Editor June 26, 2001; revised manuscript received May 14, 2002. University of Aarhus, Department of Earth Sciences, Finlandsgade 8, 8200 Aarhus N, Denmark. E-mail: anders.vest@geo.au.dk; nbc@geo.au.dk. c 2003 Society of Exploration Geophysicists. All rights reserved. 523
524
Christiansen and Christensen
Airborne TEM methods Airborne TEM methods have traditionally been used in mineral exploration for mapping highly conductive bodies in a resistive background (Pedersen and Thompson, 1991; Smith and Keating, 1996), with qualitative more than quantitative purposes. One-dimensional imaging techniques have so far been the most abundant tools for processing of the airborne data, and several techniques have been suggested based on the variation of the diffusion velocity with conductivity (DeMoully and Becker, 1984; Macnae and Lamontagne, 1987; Macnae et al., 1991; Liu and Asten, 1993). These algorithms nd the depth to an equivalent current lament as a function of time, from which the diffusion velocity and thereby the conductivity can be found. The conductivity is then ascribed to a depth equal to the image depth scaled with an ad hoc factor to produce the best results. The conductivity-depth transform (CDT) by Wolfgram and Karlik (1995) belongs to the group of methods mentioned above, and is used by Fugro Airborne surveys as a standard tool for interpretation of data from the widely used GEOTEM system. The CDT deconvolves the measured data with respect to the system response to produce the step response (repeated with alternating polarity). Based on the simplifying assumption that the step response of the earth can be represented as a series expansion into decaying exponential functions with different time constants (Stolz and Macnae, 1998), the measured data are expressed in terms of these basis functions convolved with the system response. A system of linear equations is then solved for the amplitudes of the basis functions. The deconvolution is nonunique and depends on the range and number of time constants preselected for the exponentials. This method results in smooth pictures of the subsurface conductivity. However, the usefulness of a groundwater investigation in a sedimentary environment is substantially enhanced with a quantitative output model including quantitative analyses of model resolution (Christensen et al., 2000). We present a full nonlinear inversion scheme, producing 1D layered earth models with a quantitative analysis of resolution. Instead of the inherently unstable deconvolution of data, the inversion convolves the 1D model response with the transmitter waveform, which is a stable process (Christensen, 2002). This technique is related to the Liu and Asten (1993) convolution. We compare airborne and traditional ground-based TEM methods. The study involves both synthetic models and eld case studies. The ight lines presented are part of the rst airborne TEM project carried out in Denmark in Ringkjbing County. The system used was the GEOTEM system from Fugro Airborne Surveys Ltd. The ground-based system used for comparison is the PROTEM 47 equipment from Geonics Ltd. The eld area The eld area is characterized by Quaternary and Tertiary sediments to quite large depths (>300 m). The area was not covered with ice during the last glacial period [Weichsel (Europe) or Wisconsin (North America)]. Therefore, the top layers are dominated by two features: tills older than the Weichsel and uvial deposits from the Weichsel glaciation. The Quaternary sediments vary in thickness from a few meters to more than 100 m,
and are often deformed by glacial activity. The formations below the Quaternary sequence contain important aquifers in uvial and deltaic Oligocene and Miocene deposits interbedded with mica clay. This sequence is underlain by Paleogene ne clay (Friis et al., 1998; Friborg and Thomsen, 1999). Figure 1 presents a schematic overview of the lithology in the area with approximate unit thicknesses.
METHODOLOGY
To obtain meaningful models from the inversion of TEM data and to make comparative analyses of data t and model reliability, it is essential to specify the recording situation. Special attention must be given to system specications and a thorough description of the noise model (Munkholm and Auken, 1996). System specications System specications are summarized in Table 1 and Figure 2. PROTEM 47 from Geonics Ltd is a standard, single-site, ground-based system. The current source is a 40 40 m2 loop, with the receiver coil at the center of the loop measuring the vertical component of the dB/dt eld. The measurements are split in three segments each with 20 time channels (gates), distributed with 10 per decade in time. GEOTEM from Fugro Airborne Surveys is a xed-wing TEM system with receiver coils in a trailing bird. The loop is strung around the airplane, and the receiver consists of three mutually perpendicular induction coils measuring the three components of dB/dt. Data are sampled continuously and subsequently binned in 20 gates with only the last 15 being in the off-time (distributed between 0.2 and 4.4 ms from the end of the transmitter pulse). For a more detailed description of both
FIG. 1. Schematic representation of the lithologies in western Jutland Formations. Depths are not to scale.
A Quantitative Appraisal of Airborne TEM
525
the TEM method in general and various systems, see Nabighian and Macnae (1987). Noise model The transient data value is an averaging of the induced electromotive force in the receiver coil within the gate. Groundbased methods normally use logarithmic gating, meaning that the length of a gate is proportional to delay time. If the surrounding noise is white (i.e., stochastic through all frequencies), logarithmic gating results in an effective noise decay of t 1/2 . However, the surrounding noise is not white. The spectra of AM transmitters have high amplitudes at single frequencies. These thin spectral lines will overlay the white noise. Averaging and stacking such monochromatic signals results in an effective noise decay of t 1 (Christensen et al., 2000). This contribution will dominate at early times, whereas the stochastic noise will dominate at late times. In Denmark, the transition time of the PROTEM 47 system is 100 s and the value at this delay time is 10 nV/m2 , determined through several experiments. The total absolute noise is the sum of the two contributions. Figure 3 shows the total noise as a function of time. Standard eld procedures with the GEOTEM system do not include actual measurements of the noise level because this is rather time consuming. Thus, the only way to estimate the noise Table 1. Summary of specications for the GEOTEM and PROTEM 47 systems. GEOTEM Area Max. moment Waveform Duty cycle Components Sampling Gates PROTEM 47
on GEOTEM data is by inspecting the regular data. Assuming that the data from the last gate contain no or little signal from the earth, the standard deviation (STD) on these data can be used to estimate the noise level. The noise levels on all other off-time gates are found by scaling with the relative length of the gates:
noise(n ) = STD(20)
length(20) length(n )
(1)
for the n th gate. This noise model as found from the Ringkjbing data set is also illustrated in Figure 3. Note that all other noise sources are ignored here (i.e., roll, pitch, bird position uctuations, etc.). The data are assigned another 5% relative noise due to the fact that the model assumption has a lower-ranking dimensionality than the targets, and to account for deviations from the nominal conguration (Smith, 2001b). The absolute noise described above and the 5% relative noise together describe the standard noise model. The GEOTEM system measures continuously and, thus, also has on-time gates. However, the standard eld procedures of high-altitude measurements are presently only aimed at determining the average waveform and does not include a statistical analysis of its variation. The primary eld of the average waveform is subtracted from the measured data, and the variation
Tx elevation Rx elevation Tx-Rx dist.
Transmitter (Tx) 230 m2 1600 m2 6.18 105 Am2 4800 Am2 2 ms half-sine square 2 s turnoff 33% 50% Receiver (Rx) x, y, and z z continuous 128 7 s7 ms points per from end of half-cycle turnoff 20 with 15 60 in in off-time 3 segments Geometry 120 m 0 m 70 m 0 m 131 m 0 m
FIG. 3. The absolute noise assumed for the TEM sounding data. The length of the lines indicate the time-interval in which data are recorded, measured from the end of the transmitter pulse. The noise levels are absolute and not normalized according to the effective transmitter current. The labels x and z refer to the x and z-components of the measured dB/dt eld.
FIG. 2. Schematic representation of the PROTEM 47 and GEOTEM systems eld setups.
526
of the primary eld thus contributes to the noise, particularly on the on-time measurements. The absence of an estimate on the variation of the primary eld makes it virtually impossible to estimate a noise model for on-time data. Therefore, we have focused on modeling only the off-time gates of both systems. On-time measurements are nding increasing use for various purposes (Annan et al., 1996; Smith, 2000, 2001a, Smith and Balch, 2000; Smith and Lee, 2002), and with an extended analysis of already existing eld procedures, including an estimate of the variability of the primary eld, the inclusion of on-time data in a full nonlinear quantitative inversion scheme is very interesting and an area of present research (Christiansen and Christensen, 2001). Forward calculations All modeling and inversion are done using the 1D full nonlinear inversion program SELMA (Christensen and Auken, 1992). SELMA approximates the actual geometry of the transmitter source with a circular current loop of the same area. For the airborne data, the actual height of the transmitter and receiver is taken into account. Computation of the eld quantities as functions of time and space, are done through calculations in the Laplace/wavenumber domain followed by an inverse Laplace transform using the Gaver-Stehfest algorithm (Knight and Raiche, 1982) and a subsequent Hankel transform using the digital lters of Christensen (1990). Calculation of PROTEM 47 responses includes both a waveform with turn-on and turn-off ramps, and modeling of the effect of all system band-pass lters. The characteristics of the lters were measured in the laboratory. The incorporation of the band limitation has been shown to be crucial for accuracy (Effers et al., 1999). The band limitation is modeled by multiplication with the pertinent lter in the Laplace domain before transforming to the time domain. The actual transmitter waveform is included as a convolution in the time domain by approximating the waveform with a piecewise linear function and calculating the induction coil response as a linear combination of step response values. The GEOTEM response, however, is calculated using the waveform convolution only. The waveform is found from the transmitter signal as measured by the receiver coils in highaltitude measurements (Figure 4). This signal comprises all effects due to induction in the aircraft and any band-pass ltering applied through the receivers. The ltering effect of the receiver is seen as a distinct time delay in Figure 4 comparing the transmitter waveform with the waveform measured by the receiver coils. The gure also outlines the fact that the waveform used for computation is drawn to zero and truncated well before the rst gate opens, in order to do off-time calculations only. At the time of this study, on-time modeling was not available, but present research includes modeling of the on-time signal as well (Christiansen and Christensen, 2001). Preliminary results suggest that the error on the rst gate using the truncated waveform is approximately 2%, decreasing for the following gates. Bringing the truncation of the discretized waveform closer to the rst gate makes the relative error larger. Ideally all data should of course be modeled on-time.
Inversion methodology Two extensively used inversion types were applied: few-layer and multilayer inversion. The few-layer (or block-type) inversion we dene as an inversion minimizing the data mist using the fewest number of layers. The inversion parameters are layer thicknesses and resistivities. In the multilayer inversion (underdetermined, minimum-structure), the only parameters free to vary are the layer resistivities, the layer boundaries being kept xed. To avoid erratic models, the inversion is regularized by claiming identity between the resistivities of neighboring layers within a certain relative uncertainty. The initial model is a homogeneous half-space, which means that no qualied guess is necessary. The inversion is realized as an iterative damped least-squares approach (Menke, 1989), formally written as the model update at the n th iteration:
T 1 1 mn +1 = mn + Gn Cobs Gn + BT C c B + I 1
T 1 1 Cobs (dobs dn ) + BT C Gn c (C0 Bmn ) ,
(2)
where m denotes the model vector, Gn is the Jacobian matrix, Cobs is the data error covariance matrix, B is the roughness
FIG. 4. The GEOTEM transmitter waveform. (a) The waveform as measured with the receiver coils (solid line) and as measured with a pick-up coil at the aircraft (dashed line). The dotted solid line is the 10-point piecewise linear waveform used for computations, and the circles mark the gate-center positions. (b) A magnication of the framed area in (a). The time delay between the transmitter waveform and the waveform picked up by the receivers is marked with the arrow in (b).
527
matrix containing 1 and 1 for the parameters that are tied together in the multilayer inversion, Cc is the the covariance matrix of the constraints between layers, is the damping factor, I is the identity matrix, dobs is the data vector, dn is the forward data vector based on the previous model vector mn , and nally C0 is a null vector claiming identity between the constrained parameters. The standard noise model enters the inversion procedure in the data covariance matrix assuming uncorrelated Gaussian noise. The model parameter analysis is based on a linear approximation to the covariance of the estimation error, Cest (Menke, 1989):
1 Cest = GT C obs G 1
are achieved: 1) The FA enables inversion on raw data, which means less lateral smoothing, reducing the effects of coupling to man-made conductors as much as possible (see later section). 2) The FA is less affected by possible 2D effects on a sublayered half-space. This is due to the fact that the lateral character of the FA sensitivity function is smooth, whereas the x- and z-component sensitivity functions are very different (Christiansen and Christensen, 2000), inevitably disrupting a joint inversion in the presence of 2D structures. We will base all interpretations and analyses of airborne data on the FA data type.
SYNTHETIC EXAMPLES
(3)
where G is based on the nal model obtained in the inversion. Inversion is carried out in the the log(data)-log(parameter) space, and the data t is determined by the residual, RES, given by
RES =
1 N
N n =1
( yn dn )2 , 2 n
The 1D earth model is in many cases inadequate to describe the complex resistivity structure of the earth, but due to the
(4)
where dn denotes the observed data, yn denotes the predicted data, n denotes the standard deviation, and N is the number of data points. The eld amplitude data type During ight, the receiver bird position varies (Smith, 2001b), and the motions tend to mix the three components of the measured eld, so that dBx /dt contains some dBz /dt and dB y /dt, and correspondingly for the other components. To deal with this problem, it is advantageous to introduce a new data type called the eld amplitude (FA) (e.g., Poulsen, 2000), dened as
FA =
dB2 dB2 dB2 y x z + + , dt dt dt
(5)
where dBx /dt, dB y /dt, and dBz /dt are the measured time derivatives of the components of the magnetic eld. Figure 5a presents a small section of three selected channels (early time, intermediate time and late time) from the eld data. The oscillating character of the x- and z-component data are caused by the receiver bird motions. The FA data type eliminates to a large extent the effects of receiver bird movements. Figure 5b is the result of an inversion on one sounding (indicated by the dashed line in Figure 5a) using the FA data type. The predicted data t the observed with a residual of 0.07 [equation (4)]. Figures 5b and 5c illustrates the problem with a joint inversion of the x- and z-components. Compared with the predicted data, the x-component data are too high and the z-component data are too small, giving a misleading outcome of the joint inversion with a residual of 1.6. The main reason to use the FA data type is the reduction of the effects of receiver bird motion, but additional advantages
FIG. 5. The advantages of the FA data type. (a) Components x, z, and FA of channels 7, 11, and 18 from observed data are presented. (b) Predicted data (dots) compared with observed data (solid) using the FA data type for inversion. (c) Predicted x-component data (dots) from a joint inversion of the x- and z-component data and observed data (solid). (d) As (c) but for the z-component. The data shown in (b)(d) are indicated by the vertical dashed line in (a).
528
prohibitively high computational costs of 2D and 3D models it is by far the most frequently used model in the interpretation of TEM data. We have chosen to analyze two models each of which is changed in 21 steps, giving rise to 42 1D models altogether. Both central models represent important settings for hydrogeological surveys in Denmark. The models include three layers. More layers could be relevant in describing the geological setting, and the ground-based systems are in many cases able to resolve more than three layers, but this is beyond the capabilities of an airborne system (Christensen et al., 2000). The synthetic data are produced using forward calculations, and noise is ascribed using the standard noise model described above. For both systems, Figures 6 and 7 present, from top to bottom, the true models, analyses, few-layer inversions, and multilayer inversions. The ve color-coded lines in the model analyses, as described in equation (3), represent the ve parameters in the true three-layer model, from the top: layer resistivities (RES1, RES2, and RES3) and layer thicknesses (THK1 and THK2). The colors indicate the relative uncertainty on each parameter from well determined (red) to
undetermined (blue). All the displayed models gave predicted data that t the observations to a level appropriate to the noise introduced. Model 1, double descending Model 1 in Figure 6a is a double-descending model with resistivities 200 ohm-m (layer one), 70 ohm-m (layer two), and 5 ohm-m (layer three). The thickness of the rst layer is 30 m. The thickness of the second layer is changed in 21 exponential steps from 3 m to 300 m (10 per decade). The layers are intended to represent a sandy top layer overlaying an aquifer (both Quaternary) and, at the bottom, a heavy Tertiary clay delineating the lower aquifer boundary. The true-model analyses for the PROTEM 47 system (see Figure 6b) present all well-determined parameters when the thickness of the second layer is about the thickness of the rst layer (Figure 6b, 10). This is the case until the second layer reaches considerable thickness (Figure 6b, 1819). The true-model analyses for the GEOTEM system (Figure 6c) has
FIG. 6. Model 1: 21 double-descending 1D models. (a) True models. (b) and (c) True model analyses. (d) and (e) Few-layer inversions. (f) and (g) Multilayer inversions. PROTEM 47 on the left side, GEOTEM on the right. The colors of the analyses indicate the relative uncertainty on each parameter from well determined (red) to undetermined (blue).
FIG. 7. Model 2: 21 low-contrast 1D models. (a) True models. (b) and (c) True model analyses. (d) and (e) Few-layer inversions. (f) and (g) Multilayer inversions. PROTEM 47 on the left side, GEOTEM on the right. The colors of the analyses indicate the relative uncertainty on each parameter from well determined (red) to undetermined (blue).
529
poorly determined parameters in all models, indicating that the models cannot be resolved. The few-layer inversions for the PROTEM 47 system (Figure 6d) reect the analysis above since all three layers are only resolved for the part where all the model parameters were well determined, but there is some problem nding the correct thicknesses and resistivities due to equivalences (Figure 6d, 10 14). The depth to the good conductor is resolved for all models until it reaches a depth of 200 m (Figure 6d, 19). The GEOTEM data (Figure 6e) are in all cases satisfactorily interpreted with a two-layer model, in agreement with the analyses. The good conductor is found for the entire depth range except the very last part where the conductor is at a depth of 330 m (Figure 6e, 21). Only very little new information appears when the PROTEM 47 data are interpreted with multilayer models (Figure 6f), though it seems that some information about the third layer is present when it is buried at great depths (Figure 6f, 19). The GEOTEM data (Figure 6g) introduce a little information on the second layer using multilayer inversions (e.g., Figure 6g, 1617) but, on the other hand, the top layer resistivity is erroneous at other models (Figure 6g, 112), indicating low resolution capability. Model 2, low contrast Model 2 (Figure 7a) is a low-contrast model with resistivities 70 ohm-m, 15 ohm-m, and 40 ohm-m. Again the thickness of layer one is 30 m, and the thickness of layer two is changed in 21 steps from 3 m to 300 m. In geological terms, this model could resemble a clayey and sandy Quaternary top layer overlaying a Tertiary mica clay, and a sandy aquifer at the bottom. The true-model analyses for the PROTEM 47 inversions (Figure 7b) reect well-determined parameters in the top layers except for a very thin (Figure 7b, 15) or very thick (Figure 7b, 1721) second layer. The bottom layer is well determined for depths less than 100 m (Figure 7b, 114); otherwise, it is undetermined. The GEOTEM system (Figure 7c) shows mainly poorly resolved parameters, with a little information on the third layer for intermediate depths (Figure 7c, 913), and on the second layer for large thicknesses of this layer (Figure 7c, 1921). The top layers are distinguishable in most cases with the few-layer inversions on the PROTEM 47 system (Figure 7d, 3 21), whereas the bottom layer can only be distinguished until depths of approximately 100 m (Figure 7d, 15). The few-layer inversions with the GEOTEM system (Figure 7e) invert this model with a homogeneous half-space in many cases (Figure 7e, 310). For considerable thicknesses of the second layer, a threelayer model is necessary (Figure 7e, 1117), but the thickness of the second layer is greatly underestimated, causing a lower resistivity due to equivalences. The multilayer inversions on PROTEM 47 data (Figure 7f) add some information on the deeper parts of the models (Figure 7f, 1618), whereas the second layer seems harder to recognize for small thicknesses (e.g., Figure 7f, 310). The GEOTEM multilayer inversions (Figure 7g) also bring some information to the deeper parts of the models (Figure 7g, 12 19), but no new information on the thickness of the second layer is introduced for the few-layer models that needed a three-layer model (Figure 7g, 1117).
Interpretations On the basis of the synthetic examples it can be concluded that: 1) PROTEM 47 needs three layers whenever the equivalences are not too strong, but always at least two layers, reected by both the analyses and the few-layer interpretations. 2) GEOTEM data contain information to resolve a twolayer model in most cases, with erroneous three-layer models in some cases, again in agreement with both analyses and few-layer inversions. 3) The PROTEM 47 system resolves layers to depths of approximately 200 m in the best case, (Figure 6d, 18). 4) The GEOTEM system detects layers to depths of approximately 300 m (Figure 6e, 20). 5) The PROTEM 47 system is superior in resolving the top layers, where GEOTEM data have very limited resolution (e.g., Figures 7d and 7e, 410). 6) Multilayer inversion in some cases enhances information from weak signals referring to the deeper parts of the models (Figure 6f, 19). 7) For the shallower parts, the few-layer inversion seems superior to multilayer inversion (Figures 7d and 7f, 5). 8) GEOTEM data seem to gain most by multilayer inversion.
FIELD EXAMPLES
In June 2000, two test ights were own in Denmark using the GEOTEM system from Fugro Airborne Surveys Ltd. The one addressed here was own in Ringkjbing Amt in western Jutland. About 500 line-km of data were collected altogether, of which we will present 25 km here. The section is chosen because ground-based data have been collected for part of the line as well. However, before presenting the results, the problem with coupling to man-made conductors needs to be addressed. Coupling Denmark is densely populated, and the part not covered by cities is used for agricultural purposes. This means that roads with crash barriers and buried cables, power lines, animal fences, telephone cables, etc. are ubiquitous. All these installations give rise to coupling when the primary eld from the transmitter loop is imposed (Srensen et al., 2001). This disturbance is deterministic, arising at the same delay time for all decays summed in the stacking process. A general model for the disturbance from man-made structures is that of an oscillating circuit, and is normally categorized into two types: galvanic and capacitive coupling. Figure 8 presents the two types of coupling, from the line of data to be presented below. A galvanic-type coupling could arise from high-voltage power lines, grounded at each pylon. Animal fences and highway crash barriers are other examples. The galvanic coupling is characterized by an L-R-circuit, with the nonoscillatory decay decreasing exponentially. The disturbance depends on the time constant of the circuit, and it can be very hard to recognize on single-site soundings because the whole sounding curve is shifted. An example of a galvanic coupling from a 10-kV power line is shown for a single sounding in Figure 8a. In
530
a data sweep, this type of coupling can be recognized because it often looks like a response from a vertical thin sheet, as seen in Figure 8c (Smith and Keating, 1996). However, data weakly affected by galvanic coupling might be difcult to recognize in the model sections from inversions on the FA, as seen in Figure 8e, in which the coupling is only visible as a dome structure from approximately coordinate 0.4 km to 1.4 km with a smaller structure superimposed at coordinate 0.8 km. Thus, using the FA data type tends to subdue (not remove) the inuence of weak galvanic coupling, because the x- and the z-component data prole have different shapes and are slightly out of phase across the coupling. Stronger galvanic coupling would inevitably disturb the model section. Because accuracy
can be crucial in a ground water survey, all articially coupled data need to be removed before the nal interpretation. Figure 8g shows model section (e) without the coupled data. The footprint or trend of a galvanic coupling can be quite large as indicated by the amount of removed data, but it closely reects the tail of the coupling in Figure 8c. A capacitive type coupling could arise from buried polyurethane insulated cables. The capacitive coupling is characterized by an L-C-R-circuit, and the disturbance depends on the time constant and the resonance frequency of the circuit. This type of coupling is most often easily recognized because of its oscillating character as seen in Figures 8b and 8d. The data in Figure 8b are clearly not a geological response. The capacitivetype coupling is easily recognized in the model section as well (Figure 8f), in this case as an elevation of the good conductor. The effect of the coupled data is again removed in the last panel (Figure 8h). Altogether, coupling affected two-thirds of the collected data, which had to be manually identied and removed before interpretation. All the model sections from the nonlinear inversion scheme are presented without the coupled data and thus appear with large white areas in between model subsections. The remaining subsections are the parts that can be trusted and on which a meaningful interpretation can be based. Results All the eld data are presented in Figure 9. The ying direction is from right to left, and the reference point is the position of the receiver. On the data sweep (Figure 9a), the enlarged sections presented in Figure 8 can be identied at prole coordinates 1 and 24 km. A number of other coupled data sets are easily identied even at this scale (e.g., at coordinates 5, 6.3, 15.2, 18, 19.5, 21.5, and 22.6 km); others are not identiable at this scale. The feature at coordinate 3 km is caused by the aircraft rising to a height of 200 m. The most distinctive feature in the few-layer inversion of GEOTEM data (Figure 9b) is the conductive layer at large depth, only missing around coordinate 18.5 km. The models generally need two layers to t the data, with three layers in 29% of the models. The models from inversions of PROTEM 47 data in Figure 9c all have high resistivities with a slightly more conductive feature at the top, especially around prole coordinates 1522 km. The data are unfortunately of very poor quality, possibly affected by the coupling also affecting the GEOTEM data at that part of the line. The data are tted with two to ve layers, but no common features seem to connect the separate soundings. The multilayer interpretations of GEOTEM data (Figure 9d) also have the conductive layer at depth as the most distinctive feature. The layer is missing around coordinate 3 km, but this is just where the aircraft rose to 200 m with a resulting lower signal-to-noise-ratio. Again the layer is also missing at coordinate 19 km. The layers above the conductive layer appear to have more structure than indicated by the few-layer inversions. The multilayer inversions on PROTEM 47 data (Figure 9e) are very similar to the few-layer inversions, revealing no new information. The CDT model section (Figure 9f) presents data from the whole line, including the coupled soundings. This is not possible
FIG. 8. Coupling to man-made conductors. (a) and (b) Coupled (solid) and uncoupled (solid, dotted) data. The blue lines are the x-component, the red lines are the z-component. Galvanic type coupling on the left, capacitive type on the right. (c) and (d) Selected gates as a data sweep along the prole. The position of the soundings displayed in (a) and (b) are indicated by the vertical gray lines. (e) and (f) Model sections from inversions on the FA. (g) and (h) The same prole without the coupled soundings.
531
FIG. 9. Inversions of the eld data. (a) Full data prole with x-component in red and z-component in blue. (b) Few-layer inversions of the airborne data in (a). (c) Few-layer inversions of PROTEM 47 data from part of the prole. (d) and (e) Inversions using a multilayer model. (f) The model obtained using the CDT inversion scheme.
532
without a heavy smoothing of data before processing, and two undesirable effects arise from this: (1) coupled data, and thereby unpredictably distorted models, are not identied, and (2) the lateral smoothing smears out the distortions of the coupled data to the uncoupled data sets. A number of the distinct features on the CDT prole appear in areas where the data are known to be affected by coupling. Most prominent is the conductive near-surface feature in the prole interval 1622 km, the part of the prole with the most severely coupled data. Fugro Airborne Surveys has truncated the CDT prole at the given depth based on an estimated diffusion depth of the last datum followed by a visual evaluation of the model section. The CDT does not see the good conductor. The average resistivities found for the upper parts are similar to the resistivities found in the few-layer and multilayer sections. Interpretations The conductive layer at depth is interpreted to be heavy Paleogene clay known to be at that approximate depth in the area (Figure 1). Seismic sections close to the line (but not on it) also reveal a depression in the clay surface in the area around coordinate 18.3 km (Friborg and Thomsen, 1999), as seen on the few uncoupled GEOTEM soundings from that area. The layers above the heavy clay are interpreted to be the various sandy and clayey Quaternary and Tertiary sediments corresponding to the log in Figure 1. The individual formations cannot be distinguished with either the ground-based data or airborne data. Correlation of the GEOTEM few-layer and multilayer inversions with the ground-based results gives only little information on the validity of the inversion of the airborne data. The main depth interval of investigation of the PROTEM 47 is within the rst 100 m because almost all inversions have leveled off to a homogeneous half-space at this depth. The GEOTEM system has only very limited resolution in the top 100 m, as shown with the synthetic examples. However, the comparison with a ground-based system highlights the differences in resolution capability. PROTEM 47 data can resolve between two and ve layers, whereas the GEOTEM system only resolves three layers as a maximum. The multilayer inversion of GEOTEM data has more structural information in the top layers than does the few-layer inversion, in agreement with observations on the synthetic data. The average resistivity of the CDT prole agrees with the the nonlinear inversion in the top 200 m. The only subsection with direct similarities is the deep conductor around coordinate 2325 km, but the capacitive coupling around 24 km (Figure 8) might be responsible for the feature in the CDT prole. The conductive top layer from prole coordinate 1522 km is also believed to be a coupling effect. It has been shown that unstable deconvolution of slightly noisy data leads to false indications of both poor and good conductors (Macnae et al., 1998). The feature around 1522 km is also reected in the PROTEM 47 section (Figures 9c, e), suggesting that these data are also affected by coupling. However, to positively identify a coupling, we need densely sampled prole data, which is not the case with the PROTEM 47 data, collected at approximately every 250 m.
For the rest of the prole, there are still prominent differences between the nonlinear inversions and the CDT. An example is seen around prole coordinates 610 km where the CDT has a near-surface conductive feature overlying a higher resistivity, whereas the opposite is the case for the few-layer and multilayer models.
DISCUSSION
Computation Until recently, computation time has been critical in doing quantitative interpretation of extensive airborne data sets. As far as 1D models are concerned, this is no longer a severe limitation using modern fast PCs, but 1D interpretation brings new problems into focus. The GEOTEM system samples continuously in both on- and off-time, but the details of the resolution improvements obtained by including on-time data has not yet been quantied. Recent investigations, although incomplete, have indicated near-surface resolution improvements, which is not surprising (Christiansen and Christensen, 2001). Transmitter waveform Effers et al. (1999) demonstrated the necessity of including the full system response in the calculation of responses from low-current ground-based methods. This was realized by using the transmitter waveform (nominal or measured close to the transmitter) and by modeling the effect of the lters of the receiver system. For airborne systems, it is preferable to use the transmitter waveform as measured by the receiver, because it encompasses all the effects of the system, including the effect of the induced currents in the aircraft, and, ideally, a program capable of on-time modeling is needed to extract the maximum amount of information. Coupling to man-made conductors Coupling is a serious problem when measuring in populated areas and has to be removed before interpretation. If not removed, the inversion can be erroneous, and a parameter analysis will be misleading. Removing coupled data is manual work and very time consuming. Hence, both clients and contractors must understand that the expenses of interpretation and evaluation of airborne data measured in populated areas are comparable to the survey expenses. An automated procedure identifying and removing coupled data sets is desirable, though it is must be expected that a fail-safe procedure is not likely to be found and that manual intervention will always be needed. Programs capable of simultaneously displaying data proles, individual soundings, the ight video, and a map of the survey location would be helpful. New strategies The FA data used in this paper are calculated using the poststack x, y, and z-component data. A better solution would be calculation of a prestack FA, with subsequent stacking as a fourth data type. This procedure would need to be incorporated in the survey design, because at present only poststack values
533
are saved. This procedure would further reduce the effective noise on the FA data. There is a need for more detailed noise description on airborne TEM data. At the moment no specic information on the noise is available. We suggest the following two procedures to be incorporated in the survey design: 1) Measurements at high altitude with the transmitter turned on. This is standard procedure today, but the time series must be stored and saved to enable a determination of not only average waveform but also its variability. 2) Measurement at survey altitude with the transmitter turned off, to estimate the ambient noise at the location. Again, full time series are needed. The sum of the noise estimated from the two contributions would be a good approximation to the total noise.
CONCLUSIONS
Airborne TEM data can be interpreted with success using full nonlinear 1D inversion techniques. Joint inversion of the x- and z-component data turned out disadvantageous compared to the newly proposed strategy using the FA. Interpretations and analyses of synthetic models show that airborne data can resolve up to three layers in a geophysical model. Ground-based methods resolve three (or more) layers in most cases. However, the depth of investigation is larger for the GEOTEM system than for the PROTEM 47 system. These observations are supported by the eld data. The main obstacle in an interpretation of eld data is coupling affecting up to 70% of the measurements in a densely populated area like Denmark. If not identied and removed before interpretation they will result in erroneous models.
ACKNOWLEDGMENTS
Ringkjbing County made the airborne and ground-based data available and were supportive in all matters. Without their help, this article would not have been written. Rambll and Fugro Airborne Surveys were very helpful when asked about survey or system details. Lene Hjelm Poulsen participated in numerous helpful discussions concerning inversion and interpretation of both the airborne and ground-based data. We also thank Colin Farquharson, Peter Wolfgram, an anonymous reviewer, and associate editor Randall Mackie for their helpful comments.
REFERENCES Annan, A. P., Smith, R. S., Lemieux, J., OConnell, M. D., and Pedersen, R. N., 1996, Resistive-limit, time-domain AEM apparent conductivity: Geophysics, 61, 9399. Christensen, N., 1990, Optimized fast Hankel transform lters: Geophys. Prosp., 38, 545568. 2002, A generic 1-D imaging method for transient electromagnetic data, Geophysics, 67, 438447. Christensen, N. B., and Auken, E., 1992, Simultaneous electromagnetic layered model analysis: Geoskrifter 41, 4956. Christensen, N. B., and Srensen, K. I., 1998, Surface and borehole electric and electromagnetic methods for hydrogeological investigations: Eur. J. Environmental and Eng. Geophys., 3, 7590. Christensen, N. B., Srensen, K. I., Christiansen, A. V., Rasmussen, T. M., and Poulsen, L. H., 2000, The use of airborne electromagnetic systems for hydrogeological investigations: Proc. Symp. on the Application of Geophysics to Environmental and Engineering Problems (SAGEEP), 7382.
Christiansen, A. V., and Christensen, N. B., 2000, The sensitivity functions of TEM methods: 6th Mtg., Environmental and Engineering Geophys. Soc. Euro. Section, Proc., EM09. 2001, Quantitative interpretation and analysis of airborne transient electromagnetic data in Denmark: 7th Mtg., Environmental and Engineering Geophys. Soci. Euro. Section, Proc., 122123. DeMoully, G. T., and Becker, A., 1984, Automated interpretation of airborne electromagnetic data: Geophysics, 49, 13011312. Effers, F., Auken, E., and Srensen, K. I., 1999, Inversion of bandlimited TEM responses: Geophys. Prosp., 47, 551564. Fitterman, D. V., and Stewart, M. T., 1986, Transient electromagnetic sounding for groundwater: Geophysics, 51, 9951005. Friborg, R., and Thomsen, S., 1999, Kortlgning af Ribe Formationen. Et fllesjydsk grundvandssamarbejde [Mapping the Ribe Formation. A cooperate groundwater investigation project]: Ribe, Ringkjbing, Viborg, Arhus, Vejle og Snderjyllands Amt Populr rapport. Friis, H., Mikkelsen, J., and Sandersen, P., 1998, Depositional environment of the Vejle Fjord Formation, upper Oligocenelower Miocene, Denmark: A barrier island/barrier protected depositional complex: Sedimentary Geol., 117, 221244. Knight, J. H., and Raiche, A., 1982, Transient electromagnetic calculations using the Gaver-Stehfest inverse Laplace transform method: Geophysics, 47, 4750. Liu, G., and Asten, M. W., 1993, Conductance-depth image of airborne TEM data: Expl. Geophys., 24, 655662. Macnae, J., King, A., Stolz, N., Osmakoff, A., and Blaha, A., 1998, Fast AEM data processing and inversion, Expl. Geophy., 29, 163 169. Macnae, J., and Lamontagne, Y., 1987, Imaging quasi-layered conductive structures by simple processing of transient electromagnetic data: Geophysics, 52, 545554. Macnae, J. C., Smith, R., Polzer, B. D., Lamontagne, Y., and Klinkert, P. S., 1991, Conductivity-depth imaging of airborne electromagnetic step-response data: Geophysics, 56, 102114. McNeill, J. D., 1990, Use of electromagnetic methods for groundwater studies in Ward, S. H., Ed., Geotechnical and environmental geophysics: Soci. Expl. Geophys., 191218. Menke, W., 1989, Geophysical data analysis, rev. edn.: Discrete inverse theory: Academic Press Inc. Munkholm, M. S., and Auken, E., 1996, Electromagnetic noise contamination on transient electromagnetic soundings in culturally disturbed environments: J. Environmental and Eng. Geophys., 1, 119 127. Nabighian, M. N., and Macnae, J. C., 1987, Time domain electromagnetic prospecting methods, in Nabighian, M. N., Ed., Electromagnetic methods in applied geophysics: Soc. Expl. Geophys., 427 520. Pedersen, R., and Thompson, S., 1991, GEOTEM case history: Appendix G of Time domain electromagnetic prospecting methods by M. N. Nabighian and J. C. Macnae, in Nabighian, M. N., Ed., Electromagnetic methods in applied geophysics: Soc. Expl. Geophys., 509513. Poulsen, L. H., and Christensen, N. B., 1998, Hydrogeophysical mapping with the transient electromagnetic sounding method: Eur. J. of Environmental and Eng. Geophys., 3, 201220. Poulsen, L. H., 2000, Inversion of airborne transient electromagnetic data: Geological Survey of Denmark and Greenland, Ministry of Environment and Energy, Rapport, 2000/82. Smith, R. S., 2000, The realizable resistive limit: A new concept for mapping geological features spanning a broad range of conductances, Geophysics: 65, 11241127. 2001a, On removing the primary eld from xed-wing timedomain airborne electromagnetic data: Some consequences for quantitative modelling, estimating bird position and detecting perfect conductors: Geophys. Prosp., 49, 405416. 2001b, Tracking the transmitting-receiving offset in xed-wing transient EM systems: Methodology and applications, Expl. Geophys., 32, 1419. Smith, R. S., and Balch, S. J., 2000, Robust estimation of the bandlimited inductive-response from impulse-response TEM measurements taken during the transmitter switch-off and the transmitter off-time: Theory and an example from Voiseys Bay, Labrador, Canada: Geophysics, 65, 476481. Smith, R. S., and Keating, P. B., 1996, The usefulness of multicomponent, time-domain airborne electromagnetic measurements, Geophysics, 61, 7481. Smith, R. S., and Lee, T. J., 2002, Using the moments of a thick layer to map the conductance and conductivity from airborne electromagnetic data: J. Appl. Geophys., 49, 173183. Srensen, K. I., 1995, Pulled array continuous vertical electrical sounding (PA-CVES): Proc. Symp. on the Application of Geophysics
534
Christiansen and Christensen ronmental and Eng. Geophys. Soc. Euro. Section, Proc., 108 109. Stolz, E. M., and Macnae, J. C., 1998, Evaluating EM waveforms by singular-value decomposition of exponential basis functions: Geophysics, 63, 6474. Wolfgram, P., and Karlik, G., 1995, Conductivity-depth transform of GEOTEM data: 13th Geophys., Conf., Australian Soc. Expl. Geophys., Proc., 179185.
to Environmental and Engineering Problems (SAGEEP), 893 898. Srensen, K. I., Auken, E., and Thomsen, P., 2000, TDEM in groundwater mappingA continuous approach.: Proc. Symp. on the Application of Geophysics to Environmental and Engineering Problems (SAGEEP), 485491. Srensen, K., Thomsen, P., Auken, E., and Pellerin, L., 2001, Effect of coupling in electromagnetic data: 7th Mtg., Envi-
PAPER 2
Piecewise 1D Laterally Constrained Inversion of resistivity data
by
Esben Auken, Anders Vest Christiansen, Bo Holm Jacobsen, Nikolaj Foged and Kurt I. Srensen
Geophysical Prospecting, submitted
Piecewise 1D Laterally Constrained Inversion of resistivity data

Esben Auken, Anders vest Christiansen, Bo Holm Jacobsen, Nikolaj Foged and Kurt I. Srensen
Department of Earth Sciences, University of Aarhus, Finlandsgade 6-8, DK-8200 Aarhus N, Denmark E-mails: esben@geo.au.dk; anders.vest@geo.au.dk; bo@geo.au.dk; nik.foged@geo.au.dk and kurt.sorensen@geo.au.dk
Abstract
In a sedimentary environment, layered models are often capable of representing the actual geology more accurately than smooth minimum structure models. Furthermore, interval thicknesses and resistivities are often the parameters to which non-geophysicists can relate and base decisions on when using them in waste site remediation, groundwater modelling and physical planning. We present a Laterally Constrained Inversion (LCI) scheme for continuous resistivity data based on a layered earth model (1D). All 1D data sets and models are inverted as one system, producing layered sections with lateral smooth transitions. The models are regularized through laterally equal constraints that tie interface depths and resistivities of adjacent layers. Prior information, e.g. originating from electric logs, migrates through the lateral constraints to the adjacent models, making resolution of equivalences possible to some extent. Information from areas with well resolved parameters will in a similar way migrate through the constraints to help resolve the poorly constrained parameters. The estimated model is complemented by a full sensitivity analysis of the model parameters supporting quantitative evaluation of the inversion result. Examples from synthetic 2D models show that the model recognition of a sub-layered 2D wedge model is improved using the LCI approach when compared to a section of stitched-together 1D models and when compared to a 2D minimum structure inversion. Case stories with data from two different continuous DC systems support the conclusions drawn from the synthetic example.
Introduction
Electrical methods have for a long time been used with success in environmental and hydro-geophysical studies (Albouy et al. 2001; Dodds and Ivic 1990; Fitterman 1987; Robineau et al. 1997; Sandberg and Hall 1990; Taylor et al. 1992). Nowadays standard methods allow a more detailed mapping by continuously gathering prole oriented data using either multiple electrode systems (CVES) (Dahlin 1996; Bernstone and Dahlin 1999) or systems like the Pulled Array Continuous Electrical Sounding system (PACES)
Paper 2
(Srensen 1996) and others (Panissod et al. 1997). Programs for inverting single sounding data with 1D models have been available for a number of years (e.g. Inman et al. 1975; Johansen 1977). Both the CVES and the PACES methods result in dense prole oriented data coverage with large sensitivity overlaps between individual soundings. This of course invites for 2D interpretations but unfortunately 2D inversion is still a relatively slow process, considering the amounts of data collected. Furthermore, the most widely used routines produce smooth earth models (Loke and Barker 1996; Oldenburg and Li 1994) in which formation boundaries are hard to recognize. This is not as severe when using a robust inversion scheme (L1-norm) as when using standard least squares (Loke et al. 2001), but layer boundaries are still smeared out. In many instances the investigator may suspect a predominantly layered geology, e.g. in hydrogeological investigations in a sedimentary environment (Christensen and Srensen 1998; Srensen et al. 2003). We present a piecewise 1D Laterally Constrained Inversion (LCI) scheme capable of performing inversion of very large data sets. The primary parameters of the earth model are resistivities and thicknesses. The models are tied together laterally by demanding approximate identity between neighbouring parameters, typically resistivities and depths, within a specied variance. A series of soundings is inverted as one system providing lateral smooth model sections. Prior information, originating from e.g. electrical logs, can be added at any point of the prole and the information migrates through the lateral constraints to the adjacent nodes. The inversion result is supported by a full sensitivity analysis of the model parameters. It is essential in hydrogeophysical investigations to ascertain the quality of the inversion result. The development of the piecewise 1D LCI formalism is closely linked to the development of instrumentation and eld methodologies to ensure that the interpretation tools used can handle the large data volumes and extract the maximum amount of information. The piecewise 1D formalism was developed based on the following considerations: In a sedimentary environment the subsurface is often sub-layered with relatively slow lateral variations. An inversion scheme must be fast and capable of handling large data sets. The inversion scheme must be robust to data noise and to different starting models. It should be possible to combine different data sets, e.g. transient electromagnetic data and resistivity data. There must be an option to include geophysical prior information. The output model must be accompanied by a sensitivity analysis of the model parameters. This paper demonstrates that the LCI algorithm furnishes a practical interpretation tool which meets these design criteria.
Description of measuring systems

The examples given in this paper are based on data from the CVES and the PACES systems and we shall therefore briey present these systems.
Manuscript submitted to Geophysical Prospecting
The CVES systems consist of a number of steel electrodes (typically about 60, depending on the system type) manually forced into the ground at a regular electrode spacing, typically from 5 to 12 m (Van Overmeeren and Ritsema 1988; Dahlin 1996; Dahlin 2001). A schematic representation of a CVES system is seen in Figure 1. The electrodes function as both current and potential electrodes. The measurements are controlled by a computer and can be of any conguration as entered by the user, but limited by the available electrodes. For example, a CVES layout with 5 m electrode spacing, 4 cables with each 21 electrodes, can do Wenner congurations ranging from 5 m to 130 m electrode spacing. The data collecting is semi-continuous by using a roll-along technique.
Measure point 1 Cable 1 0m 100 m Cable 2 200 m Cable 3 300 m Cable 4 400 m 500 m
Cable 2
Cable 3
Cable 4
Cable 1
Measure point 2
Figure 1: CVES system. The system shown uses 4 cables. The data are collected semi-continuous by moving the backmost cable to the front.
The PACES system consists of a small tractor, equipped with processing electronics, pulling electrodes mounted on a tail (Srensen 1996; Srensen et al. 2003). The electrodes are cylindrical steel tubes with a weight of 10 - 20 kg. Two electrodes are maintained as current electrodes with a maximum current of 30 mA. The remaining electrodes serve as potential electrodes in 8 different congurations. The positions of the electrodes together with a sketch of the system are shown in Figure 2. The receiver system records the 8 channels in parallel using a synchronous detection technique with a repetition frequency of 20 Hz followed by a robust prediction ltering scheme (Munkholm et al. 1995). The prediction lter subdues the strong noise voltages from the rapidly varying electrochemical activity on the electrode surfaces.
Inversion methodology
Data and model
Consider a CVES prole or PACES prole with nx reference node points, xi , in the horizontal direction. For each node point we have an ordinary VES data vector, di , of apparent resistivity observations for several different electrode spreads. The whole prole section shall be modelled as one inverse problem, and hence the relevant data vector is the concatenation of the VES data at each node: dobs = (d1 , d2 , . . . , dnx )T , (1)
where T indicates the vector transpose. Thus, dobs is a column vector. To minimize nonlinearity and to impose positivity, we apply logarithmic data and logarithmic parameters (e.g. Johansen 1977, Ward and Hohmann 1987). Hence di = (log(a1 ), log(a2 ), . . . , log(aN d ))T , (2)
Paper 2
90m Tail with electrodes Current electrodes Potential electrodes
GPS
1 2 3
Channel #
4 5 6 7 8 50 40 30 20 10 0 10 20 30 40 50
Current electrode Potential electrode
Layout (m)
Figure 2: The PACES system. The upper gure shows a sketch of the PACES system while the lower gure shows the eight electrode congurations. The total length of the electrode array is approx. 90 m.
where a denotes apparent resistivity, and N d is the number of electrode congurations measured at xi . The data vector has an observational error, eobs , which we suppose is unbiased, meaning that the expectation value is zero. Then the covariance matrix, Cobs , has the elements Cobs,st = cov(eobs,s , eobs,t ), (3)
The derivation of the inversion formalism applies for this general case. However, in the cases presented in this paper we assume the observational errors to be uncorrelated so that Cobs is a diagonal matrix. At each surface node, xi , the subsurface model is represented by a logarithmic 1D model mi = log(i1 ), log(i2 ), . . . , log(inl ), log(ti1 ), log(ti2 ), . . . , log(ti(nl 1) )
T
(4)
where denotes interval resistivity, and t denotes interval thickness. The total number of sub-models is nx corresponding to the number of individual 1D soundings. Each submodel is described by nl layers, so the full model m1 m2 m = . , (5) . . mnx to be determined has M = nx (2nl 1) parameters.
The dependence of apparent resistivities on subsurface parameters is generally described as a non-linear differentiable forward mapping. We follow the established practice of linearized approximation by the rst term of the Taylor expansion dobs = g(mref ) + G(mtrue mref ) + eobs , (6)
where g is the non-linear mapping of the model to the data space. mtrue has to be sufciently close to some arbitrary reference model, mref ,for the linear approximation to be a good one. In short we write: dobs = G mtrue + eobs . The Jacobian, G, contains all the partial derivatives of the mapping Gst = mt as ds log(as ) = = , mt log(mt ) as mt (8) (7)
for the sth apparent resistivity in the data vector and the tth parameter in the model vector. Only the data from individual soundings corresponding to the grouping of the model vector contribute to the total G-matrix making this a block-diagonal: d 0 m x=x1 d m x=x2 , G= (9) . .. d 0 m x=xn
x
for the nx models with nx data sets.
Prior and lateral constraints on primary parameters

Following Jackson (1979), prior information on primary parameters (resistivities and thicknesses) is included as an extra data set, mprior , Pp mtrue = mprior + eprior , (10)
where mprior = mprior mref and eprior is the error on the prior model with 0 as the expected value, and Pp is the identity matrix with the dimension of the model vector. Subscript p denotes primary parameters. The variance in the prior model is described in the covariance matrix Cprior . Prior information helps resolve the non-uniqueness of the model (Jackson 1979) and is a way to include information not originating from the resistivity data. A wrong guess of e.g. the rst layer thickness at one location may well be correlated with a wrong guess at the neighbouring locations. Therefore, it makes good sense that the error in prior guesses should have off diagonal elements in its covariance matrix. But how should we quantify it in practice? This we do indirectly by means of lateral constraints. Formally, the constraints are connected to the true model as Rp mtrue = rp + erp , (11)
Paper 2
where erp is the error on the constraints with 0 as expected value and rp = Rp mref , making the effective roughening Rp mtrue + erp = 0, (13) (12)
where the roughening matrix Rp , contains 1 and -1s for the constrained parameters, 0 in all other places, e.g. 1 0 0 1 0 0 0 0 0 1 0 0 1 0 0 0 Rp = . (14) . . . . . . . . . 0 0 0 0 1 0 0 1 The variance, or strength of the constraints, is described in the covariance matrix CRp . In this study CRp , is taken to be a diagonal matrix. For most applications, only constraints between neighbouring models, as shown in (14), are used. Constraints between any two sub-models add another sequence of 1 and -1 terms in Rp , and can be applied either laterally, vertically (Farquharson and Oldenburg 1998) or both. In the LCI method we only operate with lateral constraints.
Inversion
By joining equations (7), (10) and (11) we may write the inversion problem as: G eobs dobs Pp mtrue = mprior + eprior , Rp rp erp or more compact: G mtrue = d + e The covariance matrix for the joint observation error, e , becomes: 0 Cobs . Cprior C = 0 CRp The model estimate (Menke 1989) mest = G T C 1 G minimizes Q= 1 N +M +A
N +M +A i=1 1
(15)
(16)
(17)
G T C 1 d ,
1 2
(18)
( d T C 1 d )
(19)
where A is the number of constraints on primary parameters, and M is the number of primary model parameters. For the full non-linear iterative inversion scheme, please refer to appendix A. When only diagonal error covariances are used, as is the case in this paper, the mist criterion simplies to Q= 1 N +M +A
N +M +A i=1
d2 i var(ei )
1 2
(20)
Prior depth values and lateral constraints

For many applications, lateral constraints on depths are advantageous to constraints on thicknesses. Imagine a Quaternary sequence with disturbed sand and clay layers on top of a relatively at pre-Quaternary surface. A demand for continuity of thicknesses (lateral constraints on thicknesses) may cause discontinuity of the deeper interface, which is in contrast to the relevant geologically founded prior information. In this case, the demand for continuity of depths will be a more relevant constraint. However, the forward modelling is based on the logarithm of the primary parameters, resistivities and thicknesses, and
j
log(hj ) =
n=1
log(tn ),
(21)
for the depth, h, to the j th interface. Thus, we need to deal with depths separately. First we will add prior information on depths to the solution: Ph mtrue = mhprior + ehprior , where mhprior = hprior Ph mref , so that effectively Ph mtrue = hprior + ehprior . The vector, hprior contains the values to which we constrain individual depths: hprior = log(h1,1 ), . . . , log(hnx ,nl 1 )
T
(22)
(23) (24)
(25)
in which hij is the prior depth number j , in the model number i. The error on the prior data is ehprior with 0 as the expected value. The variance in the prior data is described in the covariance matrix Chprior . In this study, Chprior is taken to be a diagonal matrix. The matrix, Ph , contains derivatives on prior depths with respect to the primary model parameters. The derivatives with respect to resistivities are all zero. The derivatives with respect to thicknesses for the lth depth in sub-model number k are ti,j l for i = k and j l log(hk,l ) ti,j hk,l ti,j s=1 tk,s h = = = . (26) i,j 0 log(ti,j ) hk,l ti,j hk,l ti,j else Now, we write out the elements of the Ph matrix with respect to sub-model k ; 0 0 1 0 0 0 0 0 tk,1 tk,2 0 0 hk,2 hk,2 . Ph = . . . . . . . . . . . . . . . tk,1 tk,2 tk,3 tk,n 0 0 hk,n hk,n hk,n hk,n
(27)
for the n prior depths at sub-model k . The rst nl columns of zeros are the derivatives with respect to resistivities. The variance on the prior depths will enter as Chprior , which we take as a diagonal matrix.
Paper 2
The lateral constraints on depths are implemented similar to the constraints on the primary parameters, (28) Rh mtrue = rh + erh , where erh is the error on the depth constraints with 0 as expected value and rh = Rh mref , making the effective roughening Rh mtrue + erh = 0. (30) (29)
Similar to (26), the constraints on sub-models k and k +1 have derivatives with respect to thicknesses given as; (log(hk,l ) log(hk+1,l )) log(hk,l ) log(hk+1,l ) = log(ti,j ) log(ti,j ) log(ti,j ) l ti,j l ti,j s=1 tk,s s=1 tk+1,s hk,l ti,j hk+1,l ti,j ti,j ti+1,j for i = k and j l hi,j hi+1,j = , 0 else = (31)
for the lth depth in sub-models k and k + 1 with respect to the j th thickness in sub-models k and k + 1. Finally, we write the matrix with derivatives of depth constraints; 0 0 1 0 0 0 0 0 tk,1 tk,2 0 0 hk,2 hk,2 Rh = . . . . . . . . . . . . . . . tk,1 tk,2 tk,3 tk,n 0 0 hk,n hk,n hk,n hk,n 0 0 1 0 0 0 t +1,1 t +1,2 0 0 hk hk 0 0 k+1,2 k+1,2 . . . . . . . . . . . . . . . t +1,1 tk+1,2 tk+1,3 tk+1,n 0 0 hk hk+1,n hk+1,n hk+1,n k+1,n (32) Again, the rst columns with zeros are derivatives with respect to resistivities. The variances on the lateral constraints are given in CRh . Expanding (15) we now write the inversion problem as follows; G dobs eobs Pp mprior eprior Ph mtrue = mhprior + ehprior , (33) Rp erp rp Rh rh erh and similar to (16) collapse it as G mtrue = d + e (34)
The total covariance matrix for the system becomes Cobs Cprior 0 Chprior C = 0 CRp CRh
. (35)
The inverse solution is still written as (18), with the full iterative inversion scheme given in Appendix A.
Analysis of model estimation uncertainty

The parameter sensitivity analysis of the nal model is the linearized approximation to the covariance of the estimation error, Cest (e.g. Tarantola and Valette 1982) Cest = G T C 1 G Including all sub-terms this becomes, . (37) Standard deviations on model parameters are calculated as the square root of the diagonal elements in Cest . For mildly non-linear 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 , dened by: STDF(ms ) = exp Cest,ss (38)
1 1 T 1 T 1 T 1 Cest = GT C obs G + Cprior + Ph Chprior Ph + Rp CRp Rp + Rh CRh Rh 1 1
(36)
Hence, under a log-normal assumption, it is 68% likely that a given model parameter, m, falls in the interval, m < m < m STDFm (39) STDFm Thus, the impossible case of perfect resolution has a STDF = 1. A factor of STDF = 1.1 is approximately equivalent to an error of 10%. Moderate to well resolved parameters have a STDF < 1.5, poorly resolved parameters a STDF < 2 and, nally, mainly unresolved parameters a STDF > 2.
Implementation of the forward response

Forward responses are calculated as a summation of pole-pole responses over a layered earth as described by Telford et al. (1990). The potentials are computed using the Hankel transform lters of Johansen and Srensen (1979) as calculated by Christensen (1990).
Preparing the data for piecewise 1D LCI inversion

Both CVES and PACES data are semi-continuous, describing lateral as well as vertical variations along the prole. For an ideal 1D model description, we would like the data to be sampled at the same location of which we want to describe the sub-surface. For 1D DC data this typically means that the model is assigned to the lateral focus point of
10
Paper 2
a sounding. So, e.g., a Schlumberger sounding is assigned at the center of the electrode layout. For a semi-continuous piecewise 1D model description, we typically divide the prole into an even distribution of 1D models. The data assigned to the different models are then chosen based on the location of their lateral focus point, as shown in Figure 3. PACES data are usually divided into soundings with a spacing of normally 5 or 10 m.
x1 x2 x3
... ...
Figure 3: The gure shows focus points for a data pseudo section. The data are marked with a dot. The prole is divided into sub-continuous 1D soundings based on the lateral focus point of each individual data point. In this example, the prole is divided into nx 1D soundings. If more than one data point, with the same electrode conguration, falls into a subinterval, the data points are not averaged.
For symmetric congurations (Wenner, Schlumberger, pole-pole), the lateral focus point is always at the centre of the conguration, whereas asymmetric congurations have their lateral focus point closer to the smallest dipole.
Computation times
A system with 100 separate 1D models with a total of 500 model parameters and 100 PACES data sets (800 data total) uses approximately 4 seconds for one iteration on a Pentium4, 2 GHz machine. Depending on the model, between 10-20 iterations are normally needed to converge to a satisfying mist level. Primary time-consuming are the matrix operations of the inversion scheme. Hence, inverting smaller data sets and model sections, cuts down computation time dramatically.
Synthetic example
Data processing - preparation of synthetic data
The synthetic data are calculated using the 2D nite difference forward program, DCIPF2D developed at the University of British Columbia (Dey and Morrison 1979; McGillivray 1992). Analyses on model parameters are calculated using equation (37) and presented together with the inversions. We use a colour-grading of the STDF in equation (38) of resistivities and thicknesses, ranging from well determined (red) to undetermined (blue). Data have been calculated for the PACES system for every 1 m and subsequently processed in a similar way to the actual processing of eld data (Srensen et al. 2003). The resulting distance between soundings is 5 m, and 5% Gaussian distributed noise has been added to the data. The example illustrates a typical hydrogeological model where the focus is the delineation of low resistivity clay or moraine layers. Surface-near clayey layers are important for the protection of underlying aquifers (Srensen et al. 2003). The models, Figure 4a and b, consist of a 4 m thick layer at a depth of 3 m. The layer is thinner
Pseudo depth
11
at the middle, and it vanishes in the central part of the section. The resistivities of the layer and the surroundings in Figure 4a are 40 ohmm and 200 ohmm, respectively. This is reversed for the model in Figure 4b. A stochastic variation is superimposed on the layer resistivities to resemble an actual geological formation more closely. The question is now if it is possible to map both the layer and its absence reasonably accurately, knowing that the second layer is partly equivalent (Fitterman et al. 1988). The stitched-together 1D inversions without constraints on parameters (Figure 4,c and d), nd the overall geometry of the models, but it is clear that especially the central part of the model shows high/low resistivity equivalence. This is conrmed when watching the poorly constrained parameters in the analyses in panels (e) and (f). The models close to where the second layer is missing are all affected by 2D effects from the edges, showing pant-leg effects. The somewhat jagged appearance of the section is due to the 5% noise added to the data. The laterally constrained models in panels (g) and (h) minimize the effects of both 2D effects and equivalences. The values of the constraints are in this case a factor of 1.14 on both depths to layer boundaries and on resistivities. Normally, the values of the constraints are chosen between 1.1 and 1.3, but this is not critical. However, the setting of the strength or variance of the constraints depends on the expected variation in the underlying geological model. In this case, we expect (we know) changes in both resistivities and boundary positions, hence the choice of equally valued constraints. The analyses on the model parameters in panels (i) and (j) reveal an improved resolution on especially layer one and layer three. Layer two with the equivalences is still poorly resolved, but slightly better than was the case with unconstrained models. For comparison we have inverted the section using a standard least-squares minimumstructure 2D inversion. For this purpose we have used the program DCINV2D (Oldenburg and Li 1994). The results obtained are comparable to results from the RES2DINV program (not shown) by Loke and Barker (1996). The inversion results are presented in Figure 4, panels (k) and (l). The inversion result shows a smeared image of the true model. The original layer boundaries are hard to recognize. For the central part where the layer is thin, it virtually disappears in the minimum structure 2D inversion. From a large number of numerical simulations, as the one shown above, it is our experience that, given a layered earth with relatively smooth transitions in resistivities and layer boundaries, the LCI produces results that resemble the actual model well (Foged 2001). Resolution of individual parameters is improved compared to ordinary 1D inversion, as is resolution of potential equivalences.
Field examples
Field examples of both the PACES and the CVES systems are plenty, since these systems have been used extensively over the last decade. We will present two examples, the rst from a CVES survey in southern Sweden, the second from a regional (100 line kilometres) PACES survey in Jutland, Denmark.
CVES example, southern Sweden

Figure 5, shows the interpretation of a 300 m prole from the southern part of Sweden. The resistivity survey was carried out as part of the geotechnical investigations for road construction in connection with the motorway connections to the Oresund bridge-tunnel
12
True model True model
Paper 2
Depth [m]
0 5 10 15 20 200 0 5 10 15 20
(a)
0 5 10 15
(b)
250
300
350
400
20 200 0 5 10
250
300
350
400
1D model, no constraints
(c)
1D model, no constraints
(d)
Analysis Depth [m]
(e)
15 20
(f)
200
250
300
350
400
200
250
300
350
400
Analysis Depth [m]
0 5 10 15 20
1D LCI model
(g)
0 5 10
1D LCI model
(h)
(i)
15 20
(j)
200
250
300
350
400
200
250
300
350
400
Depth [m]
0 5 10 15 20 200 250
Minimum structure 2D (k)
0 5 10 15
Minimum structure 2D (l)
300
350
400
20 200
250
300
350
400
Profile coordinate [m] Resistivities [ohmm]

10 30 100 300
1.1

1.2 1.5 2 3
Figure 4: Synthetic examples, PACES. Panels (a) and (b) present the true 2D model. Panels (c) and (d) are standard stitched-together 1D inversions with no constraints on the model parameters. Panels (e) and (f) are analyses of the primary parameters in the model, ranging from well resolved (red) to poorly resolved (blue). Panels (g) to (j) repeat panel (c) to (f), but now as LCI with lateral constraints on depths and resistivities. Finally panels (k) and (l) present a minimum structure 2D inversion of the proles. The sounding distance is 5 m.
between Denmark and Sweden. Panel (a) in Figure 5 shows the data pseudo section, panel (b) is a minimum-structure 2D inversion, and panel (d) is a stitched-together 1D model section with accompanying analysis in panel (d). Panel (e) and (f) shows the LCI model and the parameter analysis, respectively. The constraints between resistivities and depth interfaces used in the LCI inversion form a factor of 1.14. The data pseudo section in (a) presents relatively smooth transitions and there are no clear signs of characteristic 2D structures, although near surface resistivity variations can be recognized at e.g. prole coordinate 150 m. The minimum structure 2D inversion in (b) detects a number of near-surface inhomogeneities along the prole. Up to a depth of approximately 10 m, there seems to be a relatively horizontal layer with resistivities about 40 ohmm above a more resistive basement. Formation boundaries cannot be rec-
13
Depth [m]
0 10 20 30 13.95 14
Data pseudo section
(a)
0 10 20
(c)
Stitched together
14.05
14.1
14.15
14.2
14.25
30
(d)
13.95 14 14.05 14.1 14.15 14.2 14.25
Depth [m]
0 10 20 30 13.95 14
Minimum structure 2D
(b)
0 10 20
(e)
LCI
14.05
14.1
14.15
14.2
14.25
Profile coordinate [km] Resistivities [ohmm]

30 50 100 150 300
30
(f)
13.95 14 14.05 14.1 14.15 14.2 14.25
Analysis
1.1 1.2 1.5 2 3
Figure 5: Field example, CVES. Panel (a) is a data pseudo section, (b) is a minimum structure 2D inversion, (c) is a stitched-together section of 1D inversion with analyses in (d) and, nally, (e) and (f) are for the LCI section. The colour-coding of the analyses ranges from well-resolved (red) to poorly resolved (blue). Two drill holes are located at 14.1 km and 14.2 km. The colours from the hole located at 14.2 km indicate from the bottom: grey medium ne clay (grey), silty sand (dark yellow), brown medium ne clay (brown) and medium sands at the top (yellow). In the drill hole at 14.2 km the silty layer is missing, otherwise they are the same.
ognized from the inverted section. Panel (c) presents a section of stitched-together 1D inversions along the prole. Indications of formation boundaries are seen, but they have a geologically unrealistic appearance. The analysis in panel (d) presents poorly resolved parameters with only occasionally well determined parameters. Panel (e), with the LCI model section, presents a model using 4 layers with smooth transitions in resistivities and layer boundaries while still picking up the near-surface resistivity changes. We now see a layered structure with a bowl-shape on the bottom layer along the prole, and we are able to differentiate between the two clay layers depicted by the drill holes. The thicknesses of the top layer and the brown clay layer are consistent with those found in the drill holes. The silty layer is not found, either because it is a very local structure, as indicated by the drill holes, or because the resistivity of the layer is close to that of the clay layers. The analyses in (f) present mainly well determined parameters along the prole. Only the thickness of the second layer is rather poorly determined due to the low resistivity contrast to the third layer. All the model sections presented produce data that t the observed data to an acceptable level.
PACES example, Jutland Denmark

This is a typical example from a groundwater survey in Denmark. The purpose of these surveys is to determine absolute amounts of sand versus clay in the upper 15 - 20 m. Figure 6 presents the results comparing the LCI inversion (Figure 6b) and the minimum structure inversion from the DCINV2D program (Figure 6d). Data are shown in (Figure 6a) and the LCI model parameter analysis in (Figure 6c). The constraints between resistivities and depth interfaces used in the LCI inversion form a factor of 1.14.
14
Paper 2
Depth [m]
0 10 20 30 0 50 100 150 200 250
Pseudo section
(a)
300
350
400
450
500
550
600
650
Depth [m]
0 10 20 30 0 50 100 150 200 250
Minimum structure 2D
(b)
300
350
400
450
500
550
600
650
Depth [m]
0 10 20
LCI
(c)
30
(d)
50
100
150
200
250
300
350
400
450
500
550
600
650
Profile coordinate [m] Resistivities [ohm-m ]

10 30 100 300 1000
1.1 1.2
Analysis
1.5 2 3
Figure 6: Field example PACES. Panel (a) shows the data pseudo section, panel (b)shows the LCI inversion with analyses in (c), and panel (e) is the 2D minimum structure inversion from DCINV2D. The sounding distance is 5 m, and the total prole length is 700 m.
The data pseudo section in (a) presents relatively smooth transitions, but with some evidence of 2D structures from around prole coordinate 500 m. The minimum structure 2D inversion (Figure 6b) presents a smooth picture with a resistive top layer on top of a more conductive layer. From around coordinate 300 m, a resistive body is seen beneath the top resistive layer, but the extent of the body is hardly recognized. The LCI section (Figure 6c) shows a three-layer model with smooth lateral variations in all layers. The model gives a clear indication of a resistive top layer above a more conductive layer overlying a resistive bottom layer. The thickness of the conductive middle layer varies considerably along the prole from more than 20 m around coordinate 140 m to nothing around coordinate 550-600 m. The analyses in (d) reveal well determined parameters in most parts of the prole. The resistivity of the bottom layer is rather poorly determined at the beginning of the prole where the depth to the layer is considerable compared to the maximum layout of electrode congurations. The conductive middle layer is a clayey till, and it is extremely important for the hydraulic and chemical impact on a possible underlying aquifer. If, in this case, we had had only the minimum structure 2D inversion section, we would have lost the detailed information about the thickness of the clay layer, and the hydrological interpretation of the survey would have been signicantly different from the one obtained based on the piecewise 1D LCI interpretation.
15
Discussion
LCI or 2D minimum structure?
The near-surface geology in Denmark most often consists of tills or glacial sands. The sub-surface has an overall layering with minor resistivity variations within the individual layers. In this setting, it is rare that the 1D model assumption, in the sensitivity volume of the electrode congurations, is signicantly violated. However, it is important to consider the dimensionality of the underlying model before applying a layered inversion scheme. Synthetic modelling (Foged 2001), not shown here, has shown that the LCI is capable of identifying 2D slope structures dipping up to 20%. Above 20%, the 1D model is violated too much, and the slope is smoothened.
The use of the LCI algorithm for hydrogeophysical investigations

A few years ago it was decided to carry out a detailed hydrogeophysical mapping of half Denmark, and the PACES method is one of the primary methods to be used for mapping of the resistivity and thereby the clay content of the upper 15-30 m with the purpose of estimating the vulnerability of deeper lying aquifers to inltration of polluting substances. Since 1999 the piecewise 1D LCI algorithm has been used as the primary interpretation algorithm for more than 5000 km of PACES data in hydrogeophysical surveys in Denmark. The algorithm has proven to be robust to data noise and to the presence of the 2D resistivity structures to be expected in sedimentary environments. The interval thicknesses and resistivities achieved from the inversions form the basis for the physical planning of land use in many elds, such as urban development, location of polluting industry, permission for water abstraction, and lately permissions are given, based on the PACES models, for industrial farming and the amount of fertilizer and pesticides that can be used. Recently, CVES data collected in Denmark and Sweden have been experimentally interpreted using both the LCI algorithm and the smooth 2D minimum structure inversion produced by the RES2DINV program or the DCINV2D program. This combination of inversion strategies has proven to be a powerful combination as a basis for the geological/hydrogeological interpretation. The geological interpretation is primarily based on models produced by piecewise 1D LCI, and the 2D minimum structure inversion is used to reveal areas where lateral resistivity variations in the subsurface violate the 1D model assumption in the piecewise 1D LCI unacceptably. Present research aims at integrating a 2D forward code in the LCI inversion scheme enabling a layered 2D inversion in the cases where lateral variations are violating the 1D model assumption (Christiansen et al. 2002).
Conclusion
The piecewise 1D LCI method provides a robust and quick method to obtain reliable inversion results in semi-layered environments from continuous resistivity data. The layered model description makes identication of formation boundaries easy, as compared to standard minimum structure 2D algorithms, which produce a smeared picture of the geological model. The inclusion of lateral constraints improves the resolution of poorly resolved parameters. This is clearly demonstrated by model sections comparing 1D sections with and without lateral constraints. Prior knowledge can be added at any point along the model prole, and the output is supported by a full sensitivity analysis of the model parameters
16
Paper 2
entering the inversion scheme. Thus, the interpreter is given a chance to ascertain the inversion result. The combination of the LCI and traditional minimum structure 2D inversion schemes has proven to be a powerful tool for hydrogeophysical investigations in Denmark.
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. Prof. Torleif Dahlin kindly provided the CVES eld data. Prof. Niels B. Christensen helped us clarify the paper and has contributed with numerous helpful comments.
Appendix A
In equation (18) we wrote the solution to the inverse problem as: mest = G T C 1 G
1
G T C 1 d ,
(40)
with respect to some reference model, mref . Writing this as the model update at the nth iteration in an iterative inversion scheme, we get
T T mn+1 = mn + [Gn C 1 Gn + n I]1 [Gn C 1 dn ] ,
(41)
where is a Marquart damping parameter (Marquart 1963). Expanding (41) with respect to (33) we end up with: mn+1 = mn +
1 1 T 1 T 1 GT C obs G + Cprior + Ph Chprior Ph + Rp CRp Rp + 1 RT h CRh Rh + n I 1 1 GT C obs (dobs g(mn ))+
1 1 C h mn )+ prior (mprior mn ) + Ph Chprior (hprior P
(42)
1 T 1 RT p CRp (Rp mn ) + Rh CRh (Rh mn )
where g(mn ) is the non-linear forward response of the nth model. The convergence of the inversion process is stabilized by two processes: 1) the Marquart modication via the parameter n and 2) an adaptive damping on the step size for the model update based on the success of the previous iteration. These two factors make the inversion process very robust. In practice, we achieve safe convergence by starting the iteration from horizontal layers with equal resistivities, i.e. a layered homogeneous halfspace.
References
Albouy, Y., P. Andrieux, G. Rakotondrasoa, M. Ritz, M. Descloitres, J.-L. Join, and E. Rasolo-manana (2001). Mapping coastal aquifers by joint inversion of dc and tem soundings-three case histories. Ground Water 39, 8797. Bernstone, C. and T. Dahlin (1999). Assessment of two automated DC resistivity data acquisition systems for landll location surveys: Two case studies. Journal of Environmental and Engineering Geophysics 4(2), 113121.
17
Christensen, N. (1990). Optimized fast hankel transform lters. Geophysical Prospecting 38, 545568. Christensen, N. B. and K. I. Srensen (1998). Surface and borehole electric and electromagnetic methods for hydrogeological investigations. European Journal of Environmental and Engineering Geophysics 3, 7590. Christiansen, A. V., E. Auken, K. I. Srensen, and T. Smith (2002). 2-D laterally constrained inversion (2D-LCI) of resistivity data. In M. S. Matias and C. Grangeia (Eds.), Proceedings of the 9th Meeting, Environmental and Engineering Geophysics, Aveiro, Portugal. Environmental and Engineering Geophysical Society European Section. Dahlin, T. (1996). 2D resistivity surveying for environmental and engineering applications. First Break 14(7), 275283. Dahlin, T. (2001). The development of DC resistivity imaging techniques. Computers and Geosciences 27, 10191029. Dey, A. and H. F. Morrison (1979). Resistivity modeling for arbitrarily shaped twodimensional structures. Gophysical Prospecting 27, 106136. Dodds, A. R. and D. Ivic (1990). Integrated geophysical methods used for groundwater studies in the Murray Basin, South Australia, pp. 303310. In Ward Ward (1990). Farquharson, C. G. and D. W. Oldenburg (1998). Non-linear inversion using general measures of data mist and model structure. Geophysical Journal International 134, 213227. Fitterman, D. V. (1987). Examples of transient sounding for ground-water exploration in sedimentary aquifers. Ground Water 25, 684693. Fitterman, D. V., J. A. C. Meekes, and I. L. Ritsema (1988). Equivalence behavior of three electrical sounding methods as applied to hydrogeological problems. In 50th Annual Meeting and Technical Exhibition of the European Association of Exploration Geophysicist, EAGE. Foged, N. (2001). Laterally constrained inversion of 2-d stochastic earth structures. Masters thesis, University of Aarhus, Denmark. Inman, J. R., J. Ryu, and S. H. Ward (1975). Resistivity inversion. Geophysics 38, 10881108. Jackson, D. D. (1979). The use of a priori data to resolve non-uniqueness in linear inversion. Geophys. J. R.astr.Soc. 57, 137157. Johansen, H. K. (1977). A man/computer interpretation system for resistivity soundings over a horizontally stratied earth. Geophysical Prospecting 25, 667691. Johansen, H. K. and K. I. Srensen (1979). Fast hankel transforms. Geophysical Prospecting 27, 876901. Loke, M. H., I. Acworth, and T. Dahlin (2001). A comparison of smooth and blocky inversion methods in 2-D electrical imaging surveys. Proceedings of the ASEG 15th geophysical conference and exhibition, Brisbane, August 2001. Loke, M. H. and R. D. Barker (1996). Rapid least-squares inversion of apparent resistivity pseudosections by a quasi-Newton method. Geophysical Prospecting 44, 131152.
18
Paper 2
Marquart, D. (1963). An algorithm for least squares estimation of nonlinear parameters. SIAM, Journal of Applied Mathematics 11, 431441. McGillivray, P. (1992). Forward modeling and inversion of DC resistivity and MMR data. Ph.D. thesis, The University of British Columbia, Vancouver, Canada. Menke, W. (1989). Geophysical Data Analysis: Discrete Inverse Theory (Revised edn. ed.). Academic Press Inc. Munkholm, M. S., K. I. Srensen, and B. H. Jacobsen (1995). Characterization and in-eld suppresion of noise in hydrogeophysics. In Proceedings of the Symposium on the Application of Geophysics to Environmental and Engineering Problems (SAGEEP), Orlando, Florida, pp. 339347. The Environmental and Engineering Geophysical Society. Oldenburg, D. W. and Y. Li (1994). Inversion of induced polarization data. Geophysics 59, 13271341. Panissod, C., M. Lajarthe, and A. Tabbagh (1997). Potential focusing: a new multielectrode array concept, simulating study, and eld tests in archaeological prospecting. Geophysics 38, 123. Robineau, B., M. Ritz, M. Courteaud, and M. Descloitres (1997). Electromagnetic investigations of aquifers in the Grand Brul e coastal area of Piton de la Fournaise volcano, Reunion Island. Ground Water 35, 585592. Sandberg, S. K. and D. W. Hall (1990). Geophysical investigation of an unconsolidated coastal plain aquifer system and the underlying bedrock geology in Central New Jersey, pp. 311320. In Ward Ward (1990). Srensen, K., E. Auken, N. B. Christensen, and L. Pellerin (2003). An Integrated Approach for Hydrogeophysical Investigations: New technologies and a Case History, accepted for publication in SEG Special Publication, Near-Surface Geophysics, Volume 2. Society of Exploration Geophysicists. Srensen, K. I. (1996). Pulled array continuous electrical proling. First break 14, 85 90. Tarantola, A. and B. Valette (1982). Generalized non-linear inverse problems solved using the least squares criterion. Reviews of Geophysics and Space Physics 20(2), 219232. Taylor, K., M. Widmer, and M. Chesley (1992). Use of transient electromagnetics to dene local hydrogeology in an arid alluvial environment. Geophysics 57, 343352. Telford, W. M., L. P. Geldart, and R. E. Sheriff (1990). Applied geophysics (2nd ed.). Cam-bridge, England, Cambridge University Press. Van Overmeeren, R. A. and I. L. Ritsema (1988). Continuous vertical electrical sounding. First Break 6, 313324. Ward, S. H. (Ed.) (1990). Geotechnical and environmental geophysics. Society Of Exploration Geophysicists. Ward, S. H. and G. W. Hohmann (1987). Electromagnetic theory for geophysical applications. In M. N. Nabighian (Ed.), Electromagnetic methods in applied geophysics, Volume 1 of Investigations in Geophysics No.3, Chapter 4, pp. 285426. Society of Exploration Geophysics.
PAPER 3
Layered and laterally constrained 2D inversion of resistivity data
by
Esben Auken and Anders Vest Christiansen
Geophysics, accepted with minor revision (revised version)
Layered and laterally constrained 2D inversion of resistivity data

Esben Auken and Anders vest Christiansen
Department of Earth Sciences, University of Aarhus, Finlandsgade 6-8, DK-8200 Aarhus N, Denmark E-mails: esben@geo.au.dk and anders.vest@geo.au.dk
Abstract
In a sedimentary environment, quasi-layered models are often capable of representing the actual geology more accurately than smooth minimum structure models. We present a 2D inversion scheme (2D-LCI) with lateral constraints and sharp boundaries for continuous resistivity data. All data and models are inverted as one system, producing layered solutions with laterally smooth transitions. The models are regularized through lateral constraints that tie interface depths or thicknesses and resistivities of adjacent layers. Prior information, which is used to resolve ambiguities and to add e.g. geological information, can be added at any point of the prole and migrates through the lateral constraints to parameters at adjacent models. Similarly, information from areas with well resolved parameters migrates through the constraints to help resolve areas with poorly constrained parameters. The estimated model is complemented by a full sensitivity analysis of the model parameters supporting quantitative evaluation of the inversion result. A simple synthetic model proves the need for a quasi-layered 2D inversion when compared with a traditional 2D minimum structure inversion. A 2D minimum structure inversion produces models with spatially smooth resistivity transitions making identication of layer boundaries difcult. A CVES (Continuous Vertical Electrical Sounding) eld example from Sweden with a depression in the depth to bedrock supports the conclusions drawn from the synthetic example. In the eld example a till layer on top of the bedrock, hidden in the traditional inversion result is identied using the 2D-LCI. Furthermore, the depth to the bedrock surface is easily identied for most of the prole with the 2D-LCI, which is not the case with the traditional minimum structure inversion.
Introduction
Standard electrical methods allow for a detailed mapping by gathering prole-oriented data continuously using either multiple electrode systems (CVES) (e.g. Dahlin 1996; Bernstone and Dahlin 1999) or various pulled systems (Srensen 1996; Panissod et al. 1997). These systems provide a dense prole-oriented data coverage with large sensitivity overlaps between individual soundings. This naturally invites 2D interpretations. Over the last 10 years, 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
Paper 3
which sharp formation boundaries are hard often to recognize. 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 smeared out. In many instances the investigator may suspect a predominantly layered subsurface, as is most often the case in sedimentary environments or for most hydrogeological investigations (Christensen and Srensen 1998). The situation is often as presented in Figure 1. An interpreter has an inversion result as presented in Figure 1a. Based on this result he has to make a geological interpretation. If he thinks the geology is layered in some sense he will have to visualize formation boundaries based on the rather smooth and smeared inversion result. A possible interpretation might look like Figure 1b. We will later get back to this example and show that it is preferable to use an inversion scheme utilizing a layered model description.
Resistivities [ohm-m]
15 20 30 0 5 10 15 20 25 200 50 100 170 300 500
Depth [m]
(a)
300
400
500
0 5 10 15 20 25
Sand Clay ? Sand Sand

200 300
Sand ?
400
(b)
?
500
Figure 1: Model interpretation. Panel (a) is an L2-norm inversion (Oldenburg and Li 1994) of a continuous resistivity data set. Panel (b) is a possible geological interpretation drawn on top of the inversion result. The interpreter assumed the subsurface to be layered.
Often a 1D solution with lateral constraints is sufcient in quasi-layered sedimentary environments (Srensen et al. 2003). However, neotectonics, glaciotectonics or other geological phenomena may disturb the sub-horizontal layering disqualifying the 1D formulation to describe the geophysical model. So, a 2D formulation is needed to enable a more complex layered earth solution. Olayinka and Yaramanci (2000) present a 2D block- type inversion scheme using polygons of equal resistivity. The inversion algorithm is made commercially available by Interpex Ltd.. They invert for positions of polygon vertices and for the polygon resistivities and, as is the case for most 2D programs available, no sensitivity analysis on model parameters accompanies the inversion output. Smith et al. (1999) presented a sharp boundary inversion for MT data using a 2D formulation. They use a layered earth discretized along a prole with lateral interpolation between neighboring nodes. The model is regularized using lateral constraints on layer conductivities and depths. We have adapted the model description used by Smith et al. (1999) to be used in a 2D inversion program for resistivity data. The inversion is based on an algorithm developed for a 1D based inversion scheme using lateral constraints on resistivities, depths or thicknesses (1D-LCI). The 1D-LCI approach produces pseudo 2D models when lateral resistivity variations are smooth. A paper on this subject has been submitted to Geophysical Prospecting. Prior information can be added at any point of the prole. 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 inversion
Manuscript accepted for Geophysics
scheme is tested and compared to a standard 2D smooth inversion on both synthetic data and eld data.
Description of measuring systems

The examples given in this paper are based on data from CVES and the PACES (Pulled Array Continuous Electrical Sounding) systems and we shall therefore briey present these systems. The CVES systems consist of a number of steel electrodes manually forced into the ground at regular electrode spacing, typically from 2 to 12 m (Van Overmeeren and Ritsema 1988; Dahlin 1996). The electrodes function as both current and potential electrodes and can measure in any conguration desired by the user. The data collecting is semi-continuous using a roll-along technique. The PACES system consists of a small tractor, equipped with processing electronics, pulling the electrodes mounted on a tail (Srensen et al.,2003). The electrodes are cylindrical steel tubes with a weight of about 15 kg. Two electrodes are maintained as current electrodes, the remaining electrodes serve as potential electrodes in 8 different congurations. A sketch of the system is shown in Figure 3. The data collecting is continuous at approximately 1.5 m/s with one full sounding saved each second, later processed to one sounding for every 5.0 meter.
90m Tail with electrodes Current electrodes Potential electrodes
GPS
1 2 3
Channel #
4 5 6 7 8 50 40 30 20 10 0 10 20 30 40 50
Current electrode Potential electrode
Layout (m)
Figure 2: PACES system. On top is a sketch of the PACES system and at the bottom the layout of the eight electrode congurations is shown. The penetration depth of the electrode congurations is logarithmically distributed from about 1.5 m to 20 m. The total length of the electrode array is approx. 100 m.
Paper 3
Data and model
Consider a resistivity data set. The data set consists of all apparent resistivity data collected along the prole assembled in a data vector: dobs = (a1 , a2 , . . . , aN )T , (1)
where T indicates the vector transpose, and N is the number of data. Thus, dobs is a column vector. To minimize nonlinearity and to impose positivity we will apply logarithmic data and logarithmic parameters (e.g. Johansen 1977; Ward and Hohmann 1987). Hence dobs = (log(a1 ), log(a2 ), . . . , log(aN ))T . (2)
The data vector has an observational error, eobs , which we suppose is unbiased, meaning that the expectation value is zero. Then the covariance matrix, Cobs , has the elements Cobs,st = cov(eobs,s , eobs,t ). (3)
The derivation of the inversion formalism applies for this general case. However, in the cases presented in this paper we assume the observational errors to be uncorrelated so that Cobs is a diagonal matrix. The model has nx reference node points, xi , in the horizontal direction corresponding to the data prole described above. At each surface node, xi , the subsurface model is represented by a logarithmic model with nl layers. mi = log(i1 ), log(i2 ), . . . , log(inl ), log(ti1 ), log(ti2 ), . . . , log(ti(nl 1) )
T
(4)
where denotes interval resistivity and t denotes interval thickness. The full model m1 m2 m = . , . . mnx
(5)
to be determined has M = nx (2nl 1) parameters. The parameters from neighboring nodes are interpolated linearly to produce a 2D model as illustrated in Figure 3.
x1 11,t11 12,t12 1,nl x2 xi i1,ti1 i2,ti2
i,nl
Figure 3: Model description. The model is described with thicknesses and resistivities at a number of nodes along a prole. The parameters between neighboring nodes are linearly interpolated to produce a 2D model.
Forward modeling
The 2D DC forward modeling in the inversion routine is performed using the nite difference code from University of British Columbia (McGillivray 1992). The code uses a nite difference approach similar to the one described by Dey and Morrison (1979). First we need to translate the layered 2D model. To do this we super impose the nite difference grid on the model and assign a resistivity value to each cell based on an area weighted average of the contributing elements in the underlying layered 2D model as described in Figure 4.
0 2 4 6 8 10
Depth [m]
0 2 4 6 8 10 12 14 16 18 20
(a)
0 2 4 6 8 10
(b)
0 2 4 6 8 10 12 14 16 18 20
Figure 4: Model translation. The layered model in (a) is translated to the model super-imposed on the nite difference grid in (b) using weighted averages.
Then electrodes are placed on node points in the grid. For irregular electrode congurations we have implemented a linear interpolation on the electrode positions to the two nearest nodes to avoid rounding odd positions. Systems with irregular electrode congurations inevitably mean dense nite difference grids. This, in combination with proles of the order of kilometers, makes it practically impossible to calculate all forward responses with one large grid. Instead we split up the prole in pieces and do calculations for each piece on its own, afterwards combining them to create the full prole. The choice of the sub-grid sizes used in this splitting can reduce the computation time drastically. The choice is based on the total size of the prole and the smallest and largest electrode spreads. Because of computational costs, the grid size should be limited to 10,000 cells, keeping in mind that the overlap between neighboring pieces needs to be sufciently large in order to ensure continuous forward data along the prole.
Forward mapping
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 rst term of the Taylor expansion dobs = g(mref ) + G(mtrue mref ) + eobs , (6)
where g is the nonlinear mapping of the model to the data space. mtrue has to be sufciently close to some arbitrary reference model, mref , for the linear approximation to be a good one. In short we write: dobs = G mtrue + eobs . The Jacobian, G, contains all the partial derivatives of the mapping Gst = mt as ds log(as ) = = , mt log(mt ) as mt (8) (7)
Paper 3
for the sth apparent resistivity in the data vector and the tth parameter in the model vector.
Prior and lateral constraints on primary parameters

The inclusion of lateral constraints is based on an approach used for 1D-LCI. This paragraph will review the basic features of the approach. Because the methodology builds on the 1D case, we distinguish between primary parameters (thicknesses and resistivities) and secondary parameters, in this case depths. 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 primary parameters (resistivities and thicknesses) is included as an extra data set, mprior , Pp mtrue = mprior + eprior , where mprior = mprior mref , so that effectively Pp mtrue = mprior + eprior , (10) (9)
where eprior is the error on the prior model with 0 as the expected value, and Pp is the identity matrix with the dimension of the model vector. Subscript p denotes primary parameters. The variance in the prior model is described in the covariance matrix Cprior . Then we add roughening constraints to the solution. Formally, the constraints are connected to the true model as Rp mtrue = rp + erp , where erp is the error on the constraints with 0 as expected value and rp = Rp mref , making the effective roughening Rp mtrue + erp = 0, (13) (12) (11)
where the roughening matrix Rp , contains 1 and -1s for the constrained parameters, 0 in all other places, e.g. Rp = 1 0 0 1 0 . . . 0 0 0 0 1 0 0 1 0 . . . 0 1 0 0 0 0 0 0 . . . . (14)
0 1
The variance, or strength of the constraints, is described in the covariance matrix CRp . In this study CRp , is taken to be a diagonal matrix. For most applications only constraints between neighboring models are used, but can be between any two sub-models. Normally we use only lateral constrains but vertical constraints (Farquharson and Oldenburg 1998) can be applied as well. Applying both vertical and horizontal constraints results in a minimum structure model.
Prior depth values and lateral constraints

For many applications, lateral constraints on depths are advantageous to constraints on thicknesses. Constraints on thicknesses are favorable whenever there is a possibility of discontinuous layer boundaries, but continuous thicknesses, e.g. across a fault. Constraints on depths are preferred in cases where we have a demand for continuity of layer boundaries, e.g. a Quaternary sequence with sand and clay layers on top of a relatively smooth pre-Quaternary surface. The inverse solution is formulated in terms of the primary model parameters. This means that prior information on depths is added with respect to the primary parameters in mtrue : Ph mtrue = mhprior + ehprior , (15) where mhprior = hprior Ph mref , so that effectively Ph mtrue = hprior + ehprior . (17) The matrix Ph is derived in Appendix A. The vector, hprior contains the values to which we constrain individual depths: hprior = log(h1,1 ), . . . , log(hnx ,nl 1 )
T
(16)
(18)
in which hij is the prior depth number j , in the model number i. The error on the prior data is ehprior with 0 as the expected value. The variance in the prior data is described in the covariance matrix Chprior . In this study, Chprior is taken to be a diagonal matrix. Also for the lateral constraints on depths we need to derive the equations with respect to the primary parameters in mtrue : Rh mtrue = rh + erh , where erh is the error on the depth constraints with 0 as expected value and rh = Rh mref , making the effective roughening Rh mtrue + erh = 0. The derivation of the matrix, Rh , can be found in Appendix A. (21) (20) (19)
Inversion
By joining equations (7), (9), (11), (15) and (19) we may write the inversion problem as: dobs G eobs mprior eprior Pp Ph mtrue = mhprior + ehprior , (22) Rp rp erp Rh rh erh Matrices Ph and Rh are derived in Appendix A.
Paper 3
We write this more compactly: G mtrue = d + e . The covariance matrix for the joint observation error, e , becomes: C = Cobs Cprior Chprior 0 CRp CRh The model estimate (Menke 1989) mest = G T C 1 G minimizes Q= 1 N +M +A
N +M +A i=1 1
(23)
0 . (24)
G T C 1 d ,
1 2
(25)
( d T C 1 d )
(26)
where N is the number of data, A is the number of constraints and M is the number of model parameters, including depths. When only diagonal error covariances are used, as is the case in this paper, the mist criterion simplies to: 1 N +M +A
N +M +A i=1
Q=
d2 i var(ei )
1 2
(27)
The target mist for the 2D-LCI is in principle zero, which in practice means as low as possible within the computational limits. This would be considered huge overtting for a underdetermined smooth inversion problem, but for a parameterized and overdetermined problem as the 2D-LCI, we believe that this is not the case. Due to the restricted number of parameters, the inversion scheme cannot make models arbitrarily complex. This is furthermore restricted by the lateral constraints. The full iterative inversion scheme is given in Appendix B.
Analysis of model estimation uncertainty

The parameter sensitivity analysis of the nal model is the linearized approximation to the covariance of the estimation error, Cest (e.g. Tarantola and Valette 1982) Cest = G T C 1 G Including all sub-terms this becomes, . (29) 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
1 1 T 1 T 1 T 1 Cest = GT C obs G + Cprior + Ph Chprior Ph + Rp CRp Rp + Rh CRh Rh 1 1
(28)
the model parameters are represented as logarithms, the analysis gives a standard deviation factor (STDF) on the parameter qs , dened by: STDF(qs ) = exp Cest,ss . (30)
Hence, under a log-normal assumption, it is 68% likely that a given model parameter, q , falls in the interval, q < q < q STDFq (31) STDFq Thus, the impossible case of perfect resolution has a STDF = 1. A factor of STDF = 1.1 is approximately equivalent to an error of 10%. Moderate to well resolved parameters have a STDF < 1.5, poorly resolved parameters a STDF < 2 and, nally, mainly unresolved parameters a STDF > 2.
Synthetic example
To demonstrate the need for a 2D inversion code utilizing layered models, we will compare results from a simple synthetic model. The data set are generated using the PACES electrode conguration, with one sounding per meter, each comprising 8 data points. These data have then had 3% noise added and processed to one sounding for each 5 m. The processing steps taken here are identical to the processing steps taken for real data. The model (Figure 5a) is a 3-layer earth representing a typical sedimentary environment with clay interbedded in sand. The model scenario is typical for groundwater surveys where we wish to map protective clay layers (Srensen et al. 2003). The layer boundaries and internal resistivity variations are set using the result of stationary stochastic processes characterized by the von Karman covariance functions (Mller et al. 2001). Mean layer resistivities are 300 ohmm, 30 ohmm and 300 ohmm for the three layers respectively. The internal standard deviation on resistivities is 0.5 times the logarithm to the resistivity. Mean interval thicknesses of rst and second layer are 5 and 10 m, respectively. The standard deviation on the thicknesses is equal to the mean value. We use the code by Oldenburg and Li (1994) for comparison. This code is internationally recognized and is commercially available. Also it is capable of performing strict L1-norm (and L2-norm) inversions, which is not possible with the widely used Res2dinv (Loke and Barker 1996). The L2-norm inversion (Figure 5b) nds the basic surface near features, but has difculties in tracking the deeper parts of the low-resistive layer. Also, it fails to recognize the resistivity transitions as sharp boundaries between the layers in the true model. This is most pronounced in the deeper parts of the model where the resolution capabilities are modest. When applying a L1-norm mist criterion (Figure 5c) instead of the usual L2-norm mist criterion the model appearance becomes more blocky with sharp resistivity transitions both laterally and vertically (Loke et al. 2001; Farquharson and Oldenburg 2003). Though, these sharp transitions does not reproduce the actual boundary transitions in the true model. The 2D-LCI inversion (Figure 5d) is carried out with equal sized constraints on layer resistivities and on the layer boundaries, i.e., depths to layers. The constraint on resistivities is a factor of 1.1 (matrix Cp ) and on depths a factor of 1.3 (matrix Ch ). These
10
Paper 3
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
Depth [m] Depth [m] Depth [m] Depth [m]
(a)
(b)
(c)
(d)
(e)
200 300 400 500 600 700 800 900
100
15 20 30 50 100 170
Analysis
1.2 1.5 2 3
300
500
1.1
Figure 5: Synthetic model. Panel (a) is the true model, with a clay layer (30 ohm-m) interbedded in a sandy layer (300 ohm-m). Panel (b) is the minimum structure L2-norm inversion result. Panel (c) is the corresponding L1-norm inversion result. Panel (d) presents the 2D-LCI result. An analysis of the model parameters (resistivities, RES1-RES4, thicknesses, THK1-THK3, and depths, DPH3-DPH4) is presented in panel (e). The analysis uses a 6-graded color code ranging from red (well determined) to blue (undetermined).
values are fairly general and has been tested on a wide variety of models, but was initially established by trial-and-error. With these settings we nd both the correct geometry and the layered nature of the model, also for the deeper parts of the model. The presented model sections produce data that t the observed data to an acceptable level according to the relevant inversion scheme. The 4-layer analysis in panel (e) presents analyses of resistivities, thicknesses and depths along the prole. DPH3 is the depth to the third layer, DPH4 the depth to the fourth layer (the depth to the second layer is equal to the thickness of the rst layer). We see that the resistivity of the top two layers is well determined for most of the prole. The resistivity of the third and fourth layer layers are well determined when they are close to the surface (coordinate 100 m to 400 m). The thickness of the rst layer is well determined for parts of the prole (e.g. coordinate 250 m to 400 m), whereas the thicknesses of the second and third layer are poorly resolved for most of the prole. Although the thickness of layers two and three are poorly determined for xk < 500 m given an undoubtably strong anti-correlation we see that the depth to the fourth layer is fairly well determined for that part of the prole. The strength of the lateral constraints, CRh and Cp matrices in equation (24), helps resolve the layer boundaries by constraining the depths and resistivities along the prole. Making the constraints very loose would allow model complexity only bounded by the data. Making very tight constraints would make a very smooth model with slow variations. Thus, determining the constraint values is a trade-off between tting the data (com-
11
plex models, no constraints) and tting the lateral constraints (smooth models, tight constraints). One might argue that we are cheating because in this case we know the actual geological variations and thus, chose the constraints accordingly. However, numerous experiments on far more complex synthetic models (not shown here) proved these settings to be good for a large fraction of the models. Finally, the starting model is in all cases a layered halfspace (in this case a 4-layer model with the same resistivity in all the layers), making the inversion scheme robust. The number of layers in the LCI section was chosen to be 4 as an inversion with 3 layers had a poorer data t and 5 layers did not result in a signicantly better data t (models not shown here). Based on this synthetic model we conclude that the layered solution has clear advantages over the smooth minimum structure solution when the geological model is quasilayered. The result can be evaluated using the sensitivity analysis from the program and the identication of layered units is straight forward.
Field example, CVES, Sweden

Figure 6 is a CVES eld example from Sweden. The prole is approximately 300 m. The resistivity survey was carried out as part of the geotechnical investigations for road construction in connection with a lled basin structure in bedrock. The data set has previously been presented by Dahlin (1996).
Elev [m] Depth [m]
0 10 20 30 35 25 15 5 35 25 15 5 35 25 15 5
(a)
(b)
Elev [m]
(c)
Elev [m]
(d)
(e)
-20 0 20 40 60 80 100 120 140 160 180 200 220 240 260
Profile coordinate [m] Resistivities [ohm-m] Analysis

10 30 100 300 1000 3000 10000 1.1 1.2 1.5 2 3
Figure 6: Field example. Panel (a) is the data pseudo section, Panel (b) is the minimum structure L2-norm inversion result. Panel (c) is the corresponding L1-norm inversion result. Panel (d) presents the 2D-LCI result with the parameter analyses in (e). The color-coding of the analyses ranges from well-resolved (red) to poorly resolved (blue). Lithological logs from drillings are located at every 20 meters from coordinate 20 m to 200 m. The colors of the drill holes indicate from the bottom: rock (dark gray), till (light gray) and clay (white).
12
Paper 3
The data pseudo section in Figure 6a presents relatively smooth transitions but with clear indications of 2D structures. The data are collected using Wenner arrays, with adistances between 2 m and 48 m. The L2-norm model in Figure 6b picks up a basin structure with its thickest appearance around coordinate 120-130 m. The identication of all three layers as depicted in the drill holes is not possible and the depth to bedrock is not easily identied, though the overall model geometry reects the drill data to some degree. For the left half of the prole it seems that the rock identied in the drill holes is of a different character from the rock detected at the right half of the prole based on prominent differences in resistivity. This is most likely due to 3D structures (Dahlin 1996). The L1-norm model (Figure 6c) is much more blocky than the L2-norm model, but still, the formations depicted by the drill holes are not picked up. Figure 6d presents a 2D-LCI section with 4 layers. No prior information is added. The basin unit is now clearly separated from the bedrock, and for the right half of the prole (coordinates 100-230 m) the till-unit is easily identied on top of the bedrock. The thicknesses of the clay layer and the till clay layer are fairly consistent with those found in the drill holes. For the left half of the prole there are inconsistencies, as was the case for the Res2dinv model. We pick up the till layer at coordinate 40 m, but it is not possible to track the layer between coordinates 40 m and 100 m. The analyses in Figure 6d present mainly well determined resistivities for the entire prole for layers 1 and 2, whereas the bottom two layers have only moderately resolved resistivities. The thicknesses of the till layer (THK3, coordinates 100-230 m) is poorly resolved, but the depth to the top of the till (DPH3), i.e. the basin depth is well determined. The presented model sections produce data that t the observed data to an acceptable level according to the relevant inversion scheme.
Discussion
Smooth minimum structure or discrete layers?
As seen from the examples, a smooth minimum structure L2-norm inversion produces smooth models even for geological models which are overall layered. For an L1-norm solution the models are more blocky, but still fails to identify quasi-layered units. The question is when to use a smooth minimum structure solution and when to use a solution with discrete layers? Often an interpreter has some background knowledge on the area and often this knowledge comes in terms of distinct units with different electrical properties. We suggest the use of the 2D-LCI whenever these units take form of a quasi-layered environment, and whenever they are not quasi-layered (e.g. a vertical structure or a very complex) a minimum structure solution should be used.
Using the 2D-LCI with other data types

The 2D-LCI presented in this paper is intended for resistivity data, but the methodology is general and it is straight forward to implement the solution for any data set for which a 2D or 3D forward code exists.
13
Code optimizing
The use of the 2D forward solution is signicantly slower than using the 1D forward solution. For long proles this difference becomes even more distinct. Therefore, it is preferable to use the 2D code only when the sub-surface is not approximately 1D. The 2D-LCI as presented here is implemented in an inversion kernel developed for 1D-LCI simply by adding the 2D resistivity forward routine to the other forward routines in the program. This feature enables us to play around with hybrid forms in which parts of the calculations are 1D and others are 2D. One possible hybrid is to do 1D calculations for the rst few iterations and then shift to 2D calculations when the basic structures have been build. Other considerations to make a faster code include dual grid methods (e.g. TorresVerdin et al. (2000)) or to use a form of Broydens update (quasi-Newton) for Jacobian calculation (e.g. Loke and Barker (1996)). Finally one could use a fast approximate 2D inversion to get a suitable starting model for the full non-linear inversion (e.g. Mller et al. (2001)). All the above items are subjects for on-going research.
Conclusion
The 2D-LCI inversion of continuous resistivity data has proven to be a robust tool to obtain reliable inversion results in quasi-layered environments. The layered model description makes identication of formation boundaries easy, as compared to a standard minimum structure 2D program, which produces a more smeared picture of the geological model. The inclusion of lateral constraints improves the resolution of poorly resolved parameters. We have demonstrated this on a synthetic model comparing inverted sections using the 2D-LCI method with 2D minimum structure models. In a eld example from Sweden a till layer hidden in the smooth minimum structure section was identied using the 2D-LCI method, and conrmed by drill holes. Prior knowledge can be added at any point along the model prole, and the output is sup-ported 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.
Acknowledgements
We would like to thank Prof. Douglas Oldenburg at University of British Columbia for letting us use their 2D resistivity code in our implementation. The eld data were kindly provided by Prof. Torleif Dahlin, University of Lund, Sweden. Prof. Torquil Smith, Lawrence Berkeley National Laboratory, Berkeley, helped us getting started on the project at the very beginning. Prof. Bo Holm Jacobsen and Prof. Niels B. Christensen helped clarifying the paper and have contributed with numerous helpful comments. Lone Davidsen helped clarifying the English language.
Appendix A - Derivatives on prior depths and depth constraints

In equation (15) we added prior information on depths to the solution. With the prior depths added to the system, we also have to add the derivatives of the depths from the
14
Paper 3
primary model parameters to build the matrix, Ph . The derivatives with respect to resistivities are all zero. The derivatives with respect to layer thicknesses at horizontal position xk are ti,j for i = k and j l t log(hk,l ) ti,j hk,l ti,j l s=1 k,s hi,j = = = , (32) 0 log(ti,j ) hk,l ti,j hk,l ti,j otherwise for the lth depth with respect to the j th thickness. Now, we write out the elements of the Ph matrix with respect to the sub-model at xk ; Ph = 0 0 . . . 0 0 0 . . . 0 1
tk,1 hk,2
0
tk,2 hk,2
0 0
tk,3 hk,n
. . .
. . .
0 0 . . .
tk,n hk,n
tk,1 hk,n
tk,2 hk,n
(33)
for the n prior depths using the identity t1,1 = h1,1 . The rst nl columns of zeros are the derivatives with respect to resistivities. The variance on the prior depths will enter as Chprior , which we take to be a diagonal matrix. Then we need to nd the derivatives with respect to depths to build the matrix Rh . Similar to (32), the constraints on sub-models at xk and xk+1 have derivatives with respect to thicknesses given as; (log(hk,l ) log(hk+1,l )) log(hk,l ) log(hk+1,l ) = log(ti,j ) log(ti,j ) log(ti,j ) l ti,j ti,j l s=1 tk,s s=1 tk+1,s = hk,l ti,j hk+1,l ti,j ti,j ti+1,j for i = k and j l hi,j hi+1,j = , 0 otherwise (34)
for the lth depth with respect to the j th thickness in sub-models k and k + 1. The full matrix with derivatives of depth constraints then becomes Rh = 0 0 . . . 0 0 0 . . . 0 1
tk,1 hk,2
0
tk,2 hk,2
0 0
tk,3 hk,n t
0
+1,2 hk k+1,2 . . . tk+1,2 hk+1,n
. . .
. . .
0 0 . . .
tk,n hk,n
0 0
t
tk,1 hk,n
tk,2 hk,n
0 0 . . . 0
0 1 tk+1,1 0 hk+1,2 . . . . . . tk+1,1 0 hk+1,n
0 0 . . .
+1,n hk k+1,n
+1,3 hk k+1,n
. (35)
Again, the rst columns with zeros are derivatives with respect to resistivities. The variances on the lateral constraints are given in CRh .
15
Appendix B - The full iterative inversion scheme

In equation (25) we wrote the solution to the inverse problem as: mest = G T C 1 G
1
G T C 1 d ,
(36)
with respect to some reference model, mref . Writing this as the model update at the nth iteration in an iterative inversion scheme, we get
T T mn+1 = mn + [Gn C 1 Gn + n I]1 [Gn C 1 dn ] ,
(37)
where is a Marquart damping parameter (Marquart 1963). Expanding (37) with respect to equation (22) and ((24) we end up with: mn+1 = mn +
1 1 T 1 T 1 GT C obs G + Cprior + Ph Chprior Ph + Rp CRp Rp + 1 RT h CRh Rh + n I 1 1 GT C obs (dobs g(mn ))+
1 1 C h mn )+ prior (mprior mn ) + Ph Chprior (hprior P 1 T 1 RT p CRp (Rp mn ) + Rh CRh (Rh mn )
(38)
where g(mn ) is the non-linear forward response of the nth model. The convergence of the inversion process is stabilized by two processes: 1) the Marquart modication via the parameter n and 2) an adaptive damping on the step size for the model update based on the success of the previous iteration. These two factors make the inversion process very robust. In practice, we achieve safe convergence by starting the iteration from horizontal layers with equal resistivities, i.e. a layered homogeneous halfspace.
References
Bernstone, C. and T. Dahlin (1999). Assessment of two automated DC resistivity data acquisition systems for landll location surveys: Two case studies. Journal of Environmental and Engineering Geophysics 4(2), 113121. Christensen, N. B. and K. I. Srensen (1998). Surface and borehole electric and electromagnetic methods for hydrogeological investigations. European Journal of Environmental and Engineering Geophysics 3, 7590. Dahlin, T. (1996). 2D resistivity surveying for environmental and engineering applications. First Break 14(7), 275283. Dey, A. and H. F. Morrison (1979). Resistivity modeling for arbitrarily shaped twodimensional structures. Gophysical Prospecting 27, 106136. Farquharson, C. and D. Oldenburg (2003). Constructing piece-wise-constant models using general measures in non-linear, minimum-structure inversion algorithms. In 6th International Symposium of the Society of Exploration Geophysicists of Japan, Tokyo. Society of Exploration Geophysicists of Japan. Farquharson, C. G. and D. W. Oldenburg (1998). Non-linear inversion using general measures of data mist and model structure. Geophysical Journal International 134, 213227.
16
Paper 3
Jackson, D. D. (1979). The use of a priori data to resolve non-uniqueness in linear inversion. Geophys. J. R.astr.Soc. 57, 137157. Johansen, H. K. (1977). A man/computer interpretation system for resistivity soundings over a horizontally stratied earth. Geophysical Prospecting 25, 667691. Loke, M. H., I. Acworth, and T. Dahlin (2001). A comparison of smooth and blocky inversion methods in 2-D electrical imaging surveys. Proceedings of the ASEG 15th geophysical conference and exhibition, Brisbane, August 2001. Loke, M. H. and R. D. Barker (1996). Rapid least-squares inversion of apparent resistivity pseudosections by a quasi-Newton method. Geophysical Prospecting 44, 131152. Marquart, D. (1963). An algorithm for least squares estimation of nonlinear parameters. SIAM, Journal of Applied Mathematics 11, 431441. McGillivray, P. (1992). Forward modeling and inversion of DC resistivity and MMR data. Ph.D. thesis, The University of British Columbia, Vancouver, Canada. Menke, W. (1989). Geophysical Data Analysis: Discrete Inverse Theory (Revised edn. ed.). Academic Press Inc. Mller, I., B. Jacobsen, and N. Christensen (2001). Rapid inversion of 2-D geoelectrical data by multichannel deconvolution. Geophysics 66(3), 800808. Olayinka, A. I. and U. Yaramanci (2000). Use of block inversion in the 2-D interpretation of apparent resistivity data and its comparison with smooth inversion. Journal of Applied Geophysics 45, 6381. Oldenburg, D. W. and Y. Li (1994). Inversion of induced polarization data. Geophysics 59, 13271341. Panissod, C., M. Lajarthe, and A. Tabbagh (1997). Potential focusing: a new multielectrode array concept, simulating study, and eld tests in archaeological prospecting. Geophysics 38, 123. Smith, J. T., M. Hoversten, E. Gasperikova, and F. Morrison (1999). Sharp boundary inversion of 2D Magnetotelluric data. Geophysical Prospecting 47, 469486. Srensen, K., E. Auken, N. B. Christensen, and L. Pellerin (2003). An Integrated Approach for Hydrogeophysical Investigations: New technologies and a Case History, accepted for publication in SEG Special Publication, Near-Surface Geophysics, Volume 2. Society of Exploration Geophysicists. Srensen, K. I. (1996). Pulled array continuous electrical proling. First break 14, 85 90. Tarantola, A. and B. Valette (1982). Generalized non-linear inverse problems solved using the least squares criterion. Reviews of Geophysics and Space Physics 20(2), 219232. Torres-Verdin, C., V. L. Druskin, S. Fang, L. A. Knizhnerman, and A. Malinverno (2000). A dual-grid nonlinear inversion technique with applications to the interpretation of dc resistivity data. Geophysics 65(6), 17331745. Van Overmeeren, R. A. and I. L. Ritsema (1988). Continuous vertical electrical sounding. First Break 6, 313324.
17
Ward, S. H. and G. W. Hohmann (1987). Electromagnetic theory for geophysical applications. In M. N. Nabighian (Ed.), Electromagnetic methods in applied geophysics, Volume 1 of Investigations in Geophysics No.3, Chapter 4, pp. 285426. Society of Exploration Geophysics.
18
Paper 3
PAPER 4
Layered 2-D inversion of prole data, evaluated using stochastic models
by
Anders Vest Christiansen and Esben Auken
3DEM-3 proceedings volume, In print
Layered 2-D inversion of profile data, evaluated using stochastic models

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
Esben Auken
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 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. 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 to the same level as 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: stochastic modelling, inversion, 2D, geoelectric INTRODUCTION 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 assumption and calls for a 2D or 3D formulation to enable a more complex layered earth description. 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 ideally suited for 2D interpretations. 2D 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 due to the regularization. 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 Auken et al., (2002) and expanded to cover the 2D case by Christiansen et al., (2002). The methodology applied in this study is the 2D LCI, 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 straightforward. 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
3DEMIII Workshop, February 2003, Adelaide.
2D-laterally constrained inversion
Christiansen and Auken
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 continuous vertical electrical sounding (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).
each surface node, xi, the subsurface model is represented by a logarithmic model with nl layers. mi = log( i1 ) ,log( i 2 ) , ,log( inl ),
, log(t
i1 ) ,log(t i 2 ) ,
,log(t
i ( nl 1) )
],
T
(2)
where denotes interval resistivity, and t denotes interval thickness. The full model m1 m2 m= (3) , 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.
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). 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 the continuous 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 ) ,
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) 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: d obs = Gmtrue + eobs , (5) The Jacobian matrix, G, contains the partial derivatives of the mapping.
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. 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 = Pm true + e prior , (6) 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.
,log( )) .
T aN
(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
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. For many applications, constraining lateral variations on depths may offer advantages to constraining 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 of continuous layer boundaries, i.e. a Quaternary sequence with disturbed sand and clay layers on top of a relatively flat pre-Quaternary surface. Inversion By joining equations (5), (6) and (7) we may write the inversion problem as: G d obs e obs = P m (8) true m prior + e prior , R r e r or more compact: G 'mtrue = d '+ e ' . (9) The covariance matrix for the joint observation error, e, becomes: C obs 0 C '= C prior (10) . 0 C R The model estimate (Menke, 1989)
STDF (ms ) = 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 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 areaweighted average of the contributing parts of the layered 2D model, as shown in Figure 3.
mest = G 'T C '1 G ' G 'T C '1 d ' ,

minimizes
(11)
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
1 2 (12) d 'T C '1 d ' , Q= N + A+M where, A is the number of constraints, and M is the number of model parameters including depths. 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) (13) C est = G 'T C ' G ' . 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:
[(
)]
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. 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.
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. Model class one - minimum type models A representative minimum type model is shown in Figure 6a at the end of this paper. The average resistivity values of the 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. Model class two - resistivities decreasing with depth A model with resistivities decreasing with depth is shown in Figure 7a at the end of the paper. The average resistivity values of the three layers 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). 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
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 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: . En = exp ln true,n ln mod el ,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. Model class one are minimum type models, model class two have resistivities decreasing with depth. 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,
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. Though, it must be noted that the layer is highly equivalent, which makes any quantitative statements about thickness or resistivity difficult. However, the 2D-LCI identifies it, even though the precise layer parameters can not be retrieved. 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. If an equivalent layer is present it can not always be identified and quantitative statements are always very difficult to make. However, the 2D-LCI often allows us to identify the existence of a layer because information migrates from non-equivalent parts of the profile to the equivalent parts through the lateral constraints. 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 well as the Res2dinv. CONCLUSIONS The 2D LCI inversion of continuous resistivity data has proven to be a robust tool to obtain reliable inversion results in sublayered 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.
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 point-topoint 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 with distinct boundaries, 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. A constructive review was given by Dr. Brian Spies.
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, 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. 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 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. 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.
PAPER 5
Optimizing a layered and laterally constrained 2D inversion of resistivity data using Broydens update and 1D derivatives
by
Anders Vest Christiansen and Esben Auken
Journal of Applied Geophysics, submitted
Optimizing a layered and laterally constrained 2D inversion of resistivity data using Broydens update and 1D derivatives
Anders vest Christiansen and Esben Auken
Department of Earth Sciences, University of Aarhus, Finlandsgade 6-8, DK-8200 Aarhus N, Denmark E-mails: anders.vest@geo.au.dk and esben@geo.au.dk
Abstract
Dense prole oriented resistivity data allows for 2D and 3D inversions. However, huge amounts of data make it practically impossible to do full 2D or 3D inversions on a routine basis. Therefore, a number of approximations have been suggested over the years to speed up computations. We suggest using a combination of Broydens update on the Jacobian matrix with derivatives calculated using a 1D formulation on a parameterized 2D model of locally 1D layered models. The approximations bring down the effective number of 2D forward responses to a minimum, which again gives us the ability to invert very large sections. Broydens update is not as useful with a parameterized problem as is the case with a smooth minimum structure problem that has been the usual application. 1D derivatives, however, seem to be very effective when initiating a full 2D solution with Broydens update. We compare the different methods using two different kinds of data on two synthetic models and on two eld examples. The most effective and reliable optimization combines 1D derivatives with a full 2D solution and Broydens update. When using Broydens update the Jacobian matrix needs to be reset every once in a while. We do this whenever the difference in data residual from the previous iteration is less than 5%. This combined inversion method reduces the computation time approximately a factor of 3 without loosing model resolution.
Introduction
Standard electrical methods allow for a detailed mapping of the subsurface by gathering prole-oriented data continuously with large sensitivity overlaps between individual data points (e.g. Dahlin 1996; Bernstone and Dahlin 1999; Srensen 1996; Panissod et al. 1997). Some of the systems in use produce huge amounts of data in a eld day (>20,000 data points). With the dense data coverage provided, 2D and 3D interpretations are of course desirable. However, most 2D and certainly any 3D inversion code utilizing full forward and inverse solutions become increasingly slower with larger data sets. Standard 2D inversion algorithms have been presented by e.g. Oldenburg and Li (1994) and Loke and Barker (1996), both producing smooth minimum structure models in which sharp formation boundaries are hard to recognize. A few inversion programs have
Paper 5
been suggested that produce blocky or layered models. Olayinka and Yaramanci (2000) presented a 2D block-type inversion scheme using polygons of equal resistivity. Smith et al. (1999) presented a layered inversion scheme with lateral constraints on resistivities and depths applied to magneto-telluric (MT) data. Auken and Christiansen (2003) have adapted the model description of Smith et al. (1999) and used it in a 2D inversion scheme for resistivity data which produces laterally smooth models with discrete layers. Approximations to a full solution for an inversion algorithm can be done on several stages: in the forward solution, in the calculation of derivatives or in the inverse matrix manipulations. However, the savings in computation time is obtained at the expense of the loss of accuracy which is the inevitable consequence of numerical approximations. Thus, there is a trade-of between the savings in computation time and the required accuracy. Loke and Barker (1996) suggested a quasi-Newton formulation which gave major time reductions using Broydens update formula (Broyden 1965). The quasi-Newton method has been widely used with success in smooth minimum structure 2D resistivity inversions. Oldenburg and Ellis (1991), introduced approximate inverse mappings in both the model space (AIM-MS) and in the data space (AIM-DS). They exemplied the AIM-DS with the 2D MT problem using a 1D formulation as an approximate inverse. This proved powerful with MT data. The Rapid Relaxation Inverse (RRI) by Smith and Booker (1991), also combined 2D and 1D formulations for the MT problem. In the RRI the derivatives are 1D except for the fact that they comply with the 2D elds. An extremely fast multi-channel deconvolution (MCD) on resistivity data was suggested by Mller et al. (2001), where the output is a smooth picture of the subsurface resistivity distribution obtained without direct computation of the 2D elds. Torres-Verdin et al. (2000) used a smaller grid as an approximate 2D nite-difference forward, incorporated in an auxiliary inversion scheme. This inversion method is in many ways similar to the AIM-DS presented by Oldenburg and Ellis (1991) only the approximate inverse is not a 1D formulation but a 2D formulation using a smaller grid. This paper combines Broydens update formula with derivatives calculated using a composite 1D formulation along the prole. The 1D derivatives are used for the rst few iterations while the model is still in is infant stages without too much 2D structure. Later a full 2D solution is computed followed by iterations taking advantage of Broydens update formula.
The inversion is based on an algorithm developed for a laterally constrained 1D inversion scheme (1D-LCI) (Auken et al. 2004). The 1D-LCI was expanded to cover the 2D case as well by including a 2D forward response in the inversion algorithm (Auken and Christiansen 2003). This section will give a brief summary of the methodology presenting the key concepts and basic equations.
Data and model

Consider a resistivity prole with nx reference node points, xi , in the horizontal direction. For each node point we have an ordinary data vector, di , of apparent resistivity observations for several different electrode spreads. The whole prole section shall be modelled as one inverse problem, and hence the relevant data vector is the concatenation of the data
Manuscript submitted to Journal of Applied Geophysics
at each node: dobs = (d1 , d2 , . . . , dnx )T , (1) where T indicates the vector transpose. For a simple continuous vertical electrical sounding (CVES) prole the grouping of the data vector is sketched in Figure 1.
x1 x2 x3
... ...
Figure 1: Data grouping. The individual 4-electrode congurations are grouped into soundings based on the lateral position of their focus point.
To minimize non-linearity and to impose positivity, we apply logarithmic data and logarithmic parameters (e.g. Johansen 1977; Ward and Hohmann 1987). Hence di = log(a1 ), log(a2 ), . . . , log(aNi )
T
Pseudo depth
(2)
where a denotes apparent resistivity, and Ni is the number of electrode congurations measured at xi . At each surface node, xi , the subsurface model is represented by a logarithmic 1D model with nl layers mi = log(i1 ), log(i2 ), . . . , log(inl ), log(ti1 ), log(ti2 ), . . . , log(ti(nl 1) )
T
(3)
where denotes interval resistivity, and t denotes interval thickness. The total number of sub-models is nx corresponding to the number of individual 1D soundings. Each submodel is described by nl layers, so the full model m1 m2 m = . , (4) . . mnx to be determined has M = nx (2nl 1) parameters. To produce the 2D model from the composite 1D prole, parameters from neighbouring nodes are interpolated linearly as illustrated in Figure 2. Thus, locally we have a 1D model description which in combination builds the full 2D prole. Grouping the data vector according to the model setup enables mixing of 1D and 2D calculations.
Forward modeling
The 2D resistivity forward modeling in the inversion routine is performed using the nite difference code from University of British Columbia (McGillivray 1992). The code uses a nite difference approach similar to the one described by Dey and Morrison (1979).
4
x1 11,t11 12,t12 1,nl x2 xi i1,ti1 i2,ti2
Paper 5
i,nl
Figure 2: Model description. The model is described with thicknesses and resistivities at a number of nodes along a prole. The parameters between neighboring nodes are linearly interpolated to produce a 2D model.
The layered 2D model is translated to the nite difference grid by superimposing the grid on the model and assigning a resistivity value to each cell based on an area weighted average of the contributing elements in the underlying layered 2D model, see Figure 3. Electrodes are placed on node points in the grid. For electrodes not on node points we
0 2 4 6 8 10
Depth [m]
0 2 4 6 8 10 12 14 16 18 20
(a)
0 2 4 6 8 10
(b)
0 2 4 6 8 10 12 14 16 18 20
Figure 3: Model translation. The layered model in (a) is translated to the model super-imposed on the nite difference grid in (b) using weighted averages.
have implemented a linear interpolation to the electrode positions involving the two nearest nodes.
Forward mapping
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 rst term of the Taylor expansion dobs = g(mref ) + G(mtrue mref ) + eobs , (5)
where g is the nonlinear forward mapping of the model to the data space. The true model, mtrue has to be sufciently close to some arbitrary reference model, mref , for the linear approximation to be a good one. In short we write: dobs = G mtrue + eobs . The Jacobian, G, contains all the partial derivatives of the mapping Gst = mt as ds log(as ) = = , mt log(mt ) as mt (7) (6)
for the sth apparent resistivity in the data vector and the tth parameter in the model vector.
Prior information and lateral constraints

Prior information, which is used to resolve ambiguities and to add e.g. geological information to the data set, can be added at any point of the prole and migrates through the lateral constraints to adjacent models. The models are tied together laterally by introducing lateral constraints between neighboring parameters. This means that information from models with a small variance migrates through the lateral bands to models with higher variance. The code utilizes lateral constraints on resistivities, thicknesses and depths. For most applications, lateral constraints on depths are advantageous to constraints on thicknesses. Constraints on depths are preferred in cases where there is a demand for continuity of layer boundaries, e.g. a Quaternary sequence with sand and clay layers on top of a relatively smooth pre-Quaternary surface. Constraints on thicknesses are favourable whenever there is a possibility of discontinuous layer boundaries, but continuous thicknesses, e.g. across a fault. In this paper constraints are applied to resistivities and depths.
Inversion
We write the inversion compactly as: G mtrue = d + e (8)
where G contains the partial derivatives of the data and roughness information on the lateral constraints as well as prior information. Likewise, d combines the observed data with the data for the constraints and prior information. The joint observation error is e , with covariance matrix C . The model estimate (Menke 1989) mest = G T C 1 G minimizes Q= 1 N +A+M
N + A+ M i=1 1
G T C 1 d ,
1 2
(9)
d2 i var(ei )
(10)
where N is the number of data, A is the number of constraints and M is the number of model parameters, including depths. For large models the matrix calculations in Equation (9) can be very time consuming, although a lot of time is saved using sparse solutions when appropriate.
Optimizations
Forward calculations with a sliding model window
Long proles or systems with irregular electrode congurations inevitably mean large nite difference grids. This can make it practically impossible to calculate full forward responses with one large grid. Instead we split up the prole in model windows and do calculations for one window at the time sliding it along the prole. Afterwards the responses from the windows (dark gray) are concatenated to create the full prole, as illustrated in Figure 4a. Experimentally we found the computation time to depend linearly on n2 log2 (n) for n cells, when the number of vertical nodes has be chosen appropriately. This is an approximation, because the computation time depends differently on the number of cells in the
6
Start profile End profile
Paper 5
Coordinate of parameter
(a)
(b)
Figure 4: Sliding model window. (a) shows the sliding model windows for forward responses, whereas (b) is for derivatives. Light grey and dark grey are the ne parts of the grid. Electrodes need to be within the ne parts of the grid with the lateral focus point in the dark grey. The light grey are overlaps between the sliding model windows and have to be sufciently large to ensure continuous forward responses when concatenating the full response of the dark gray sections. White areas are padding.
two directions. Also the number of current electrodes needs to be taken into account, but for proles up to 3 km the correlation is very good. Thus, an optimal choice of the grid size can reduce the computation time drastically. A large grid means fewer but computationally more expensive computations, whereas a small grid means more but faster computations. The grid size is based on the total size of the prole and the electrode congurations used, keeping in mind that the overlap between neighboring model windows always needs to be sufciently large in order to ensure continuous forward data along the prole. The sizes of the overlap, the padding zones (white) and the number of vertical nodes are xed when appropriate values have been determined for the actual conguration. Besides these xed conguration dependent parameters the grid can still be optimized in terms of the size of the central sliding window (SW) (dark gray in Figures 4 and 5). Figure 5 shows the relative timing of two different congurations on a 1100 m prole, assuming a n2 log2 (n) dependence.
10 10 13 3 1 0.3 0.1 0 200 (400,1) 70 SW 70 13 3 1 0.3 0.1 0 (140,0.36) 200 400 600 13 60 SW 60 13
Normalized time
a) CVES, 400 m layout

400 600 800 1000
b) PACES
800 1000
SW [m]
SW [m]
Figure 5: Optimising the sliding model window. (a) shows a 1100 m CVES prole with a 400 m spread and 5 m horisontal grid size. (b) shows a 1100 m PACES prole (approx. 90 m spread) with 1 m grid. Numbers in the boxes are the number of horisontal grid points in each separate part. The thick solid line is the normalized computation time, relative to the time using just one grid, for varying sizes of the central sliding window (SW). Only a whole number of sliding windows is permitted, explaining the jagged appearance. The thin solid line is the underlying function permitting fractions of sliding windows.
Figure 5a shows the relative timing of a 400 m CVES layout with 5 m between horisontal grid points in the ne part. In this case, the layout is quite large compared to the prole length and there is no time gained using the slider. The shortest computation time is achieved when choosing the SW to be 400 m bringing the total size of the ne grid to 1100 m (400 m + 2*70*5 m) equal to the prole length. In Figure 5b the same prole is measured with the PACES system (Srensen 1996) which has approximately 90 m spread and the horisontal grid size is 1 m. For the PACES
system the optimal size of the SW is 140 m (i.e. 140 cells) which is almost three times as fast as if the full prole where computed with just one grid. A longer prole would have made the time saving even larger. The sliding model window can also be used to speed up calculations of derivatives taking advantage of the fact that derivatives for a certain parameter have the highest values for congurations close to the parameter itself. Thus, we calculate derivatives with only one window, centering this at the lateral position of the parameter in question as depicted in Figure 4b. Doing this we need to take special care designing the grid for the sliding window. If it is too small we leave out derivatives that would have contributed to the 2D description. Making the grid very large only enhances accuracy, but might bring unnecessary computations since the derivatives for data far away from the parameter is close to zero anyway. In this paper we have used the slider throughout, both for full forward responses as well as for calculation of derivatives.
Broydens update
Broydens update formula (Broyden 1965) has been used widely for 2D resistivity inversion. The update formula approximates the Jacobian matrix, Gn+1 , of iteration n +1 using either the Jacobian matrix or the Broyden matrix of iteration n: Gn+1 where dn = dn dn1 mn = mn mn1 (12) Bn+1 = Bn + [dn Bn mn ] m T n , m T n mn (11)
The update formula is based on an assumption that the difference in forward mappings, d(mn ), changes linearly with respect to mn , in the direction of mn . This is most likely to be the case close to the minimum of the least squares cost-function. Most of the previous implementations have been in smooth minimum structure programs (e.g. Loke and Barker 1996). The cell-based nite-difference or nite-element implementation of the forward problem in these programs implies that only resistivities are inverted for. When parameterizing the problem into, for example, thicknesses and resistivities of discrete layers we introduce a mixture of physical parameters and structure. Several authors have noted that the forward mapping in this case is more linearly dependent on the physical parameters than on the structure. Torres-Verdin et al. (2000), implemented Broydens update formula on a parameterized cross-borehole problem. They experienced a need to reset the entries in the Jacobian matrix after a few iterations, doing the full, time consuming solution. Otherwise the problem did not converge. This is not the case with the minimum structure solutions. Loke and Dahlin (2002) found that if the Jacobian matrix is initiated and perhaps fully recalculated for the rst two or three iterations the nal model will be very close to the full Gauss-Newton model. Based on these considerations we have implemented Broydens update formula with an option to reset and recalculate the Jacobian if necessary. As the criterion to recalculate the full Jacobian matrix we use the change in data residual between the latest two iterations. The value is dened by the user, but to our experience 5-10% seems to be suitable.
Paper 5
1D derivatives
The derivatives in the Jacobian matrix of equation (9) are approximated by a rst-order forward-difference formula: mt as mt gs (mt + mt ) gs (mt ) = (13) Gst = as mt a s mt where mt is chosen as small as possible (1-2% of m). This implementation requires at least one forward calculation per model parameter and one forward per model. In the parameterized 2D case this means M +1 forward calculations, which is very time consuming for large problems. Instead we suggest calculating the entries to the Jacobian matrix with a 1D forward mapping. For the 1D resistivity case a forward calculation is extremely fast, practically eliminating the Jacobian calculation as a time factor. Only the data from individual soundings corresponding to the grouping of the model vector now contribute to the total G-matrix making this a block-diagonal: d 0 m x=x1 d m x=x2 , G= (14) . . . d 0 m x=xn
x
for the nx models with nx data sets. For one iteration we now need to do only one full 2D calculation on the model from the previous iteration and nx 1D calculations to ll the Jacobian matrix, G. For a subsurface resistivity structure close to a 1D model this will work ne. For very complex 2D structures the approximation will be poor, since the off-diagonal elements in the 2D G-matrix will have non-negligible values compared to the bloc-diagonal elements.
Combining Broydens update and 1D derivatives

Since we now have the full 1D solution, 1D derivatives, the full 2D and Broydens update incorporated in the inversion scheme an obvious solution is to do a combination to make an optimal solution in terms of speed and accuracy. Normally the iterative solution is started with a homogeneous halfspace. This means that very little 2D information is present in the beginning. Based on this fact we have chosen to start the inversion procedure with 1D derivatives, at some point switching to doing full 2D solutions and subsequently to Broydens update. If the subsurface is thought to be fairly 1D, the 1D derivatives can be exchanged with the full 1D solution. Generally, this iteration scheme should minimize the number of full 2D solutions while keeping as much 2D information as possible. Also, the Broyden solution should be stabilized because the model mist will be closer to the desired mist minimum, minimizing nonlinearity errors.
Synthetic examples
Simple dip-model, PACES data
The rst example in Figure 6a, is a very simple two-layer model with a 45-degree descending slope on the lower layer. The data set is generated using the PACES electrode congu-
ration (Srensen 1996), with one sounding per meter, each comprising 8 data points. The forward code is the DCIP2D, from University of British Colombia (McGillivray 1992). 5% noise was added to the data, which were then processed to one sounding per 5 m as if they were eld data. The prole is 600 meters long with 960 data points in 120 models (660 model parameters). The PACES system consists of a small tractor, equipped with processing electronics, pulling the electrodes mounted on a tail. The system measures 8 different congurations continuously at approximately 1.5 m/s with one full sounding saved each second, later processed to one sounding for every 5.0 meter. The panels on the left side of Figure 6 (panels a-g) presents various inversions using 1D, 2D, Broydens update, 1D derivatives and a combination of Broydens update and 1D derivatives. The plots on the right side (Figure 6i) are the corresponding data residuals for each iteration. Panel c) presents a full 1D inversion which clearly does not resolve the 2D slope. Panel d) is the full 2D solution, which nds the correct geometry except for the wiggles on the deep parts of the conductive layer due to noisy data and decreasing resolution. Panel e) presents the solution using Broydens update for all iterations after initiating the inversion with two full 2D iterations. From the model section and the residual plot on the right it is clear that Broydens update as stand-alone does not converge to a satisfactory model. Panel f) is Broydens update with full 2D recalculation of the Jacobian whenever the change in data residual is less than 5%. The model now converges towards the true model, but more iterations are needed compared to the full 2D solution. Panel g) is the solution using 1D derivatives for all iterations. Iteration for iteration the residual of this method follows the residual 1D solution (Figure 6i) except for the last two iterations, indicating that the 2D structures are found at the very last iterations. The nal model is in between the model from the full 2D solution (Figure 6d) and the model from the full 1D solution (Figure 6c). Finally, panel h) presents the combination of 1D derivatives with a Broydens update on successive iterations. 1D derivatives were carried out for the rst 4 iterations, then one full 2D solution and nally Broydens update with the 5% criterion. The timing of the examples in Figure 6 is summarized in Table 1. The total CPU time is given in column 1, with the number of full 2D iterations in column two. All inversions are tested on a 2.8 GHz Pentium4 processor. Method (c) Full 1D (d) Full 2D (e) Broyden (f) Broyden 5% (g) 1D derivatives (h) 1D derivatives & Broyden 5% Computation time 5s 8779 s 1627 s 4397 s 284 s 2415 s # Iterations 6 11 13 18 6 9 # 2D iterations 0 11 2 5 0 3
Table 1: Summary of computation times and number of full 2D iterations for the inversions presented in Figure 6.
The primary time consumer in this case is clearly the 2D iterations. For larger problems the time taken to do the matrix calculations can be signicant as well, although it can be minimized, taking advantage of fairly sparse matrices. From this simple synthetic example we conclude that Broydens update by itself is not applicable for parameterized problems. If combined with full 2D updates of the Jacobian
10
Paper 5
Depth [m]
0 10 20
30
(i)
(a) True model
Depth [m]
200 220 240 260 280 300 320 340 360 380 400 0 10 20 10
(b) Pseudo section
Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(c) Full 1D
Data residual
Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(d) Full 2D
1
Depth [m ] Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(e) Broyden
200 220 240 260 280 300 320 340 360 380 400 0 10 20
(f) Broyden 5%
Depth [m ]
200 220 240 260 280 300 320 340 360 380 400 0 10 20
0.3
10
15
20
Iteration #
(g) 1D Derivatives
200 220 240 260 280 300 320 340 360 380 400 0
3 10 30 100 300
Depth [m
10 20
(h) 1D Derivatives & Broyden 5%

200 220 240 260 280 300 320 340 360 380 400
Figure 6: Synthetic PACES example. Panel a) is the true model. Panel b) is a pseudo section of the data. Panel c) is the model from a full 1D inversion; d) is the result using a full 2D solution. Panel e) presents an inversion using Broydens update formula for all iterations except the rst two. Panel f) also uses Broydens update formula, but the full Jacobian is recalculated when the residual change is less than 5 %. The model in panel g) is the result of 1D derivatives throughout. Panel h) combines 1D derivatives with the full 2D solution and Broydens update formula. Finally, the data residuals are presented in i) as a function of iteration number. Filled markers (for Broyden 5% and 1D derivatives+ Broyden 5%) indicate full 2D iteration.
11
matrix, convergence is assured. Using 1D derivatives works quite well with this simple geometry and combining them with a full 2D solution using Broydens update speeds up the computation time without losing 2D information. However, for more complex structures the 1D derivatives as stand-alone is insufcient and we will not show these results in the following examples. Results from the full 1D solution will not be shown either.
Von-Karman model, CVES, continuous gradient array

The 3-layered model (Figure 7a) is made from a stationary stochastic process. The degree of spatial correlation is parameterized in terms of the fractal dimension of a self-afne process, which is characterized by the von Karman covariance functions (Mller et al. 2001; Serban and Jacobsen 2001). The synthetic model is made in a sedimentary way, i.e., 1st layer (bottom layer) is deposited as a homogeneous halfspace; 1st layer is eroded; 2nd layer is deposited (draped) on top of the eroded surface of the 1st layer; Erosion of 1st and 2nd layer unit; 3rd layer (top layer) deposited on top of eroded surface. Average resistivities are 300 ohm-m (top layer), 30 ohm-m (middle layer) and 100 ohm-m. The standard deviation on the layer resistivities is 0.3 times the logarithm to the average values. The data are generated using the CVES system in a continuous gradient array conguration. 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 prole 1 km long with a minimum electrode distance of 5 m. The model has 215 surface node points each described by 3 layers, totally 1075 model parameters. More complex models would have needed four layers to produce a satisfying model, but in this case three layers are sufcient. Both the full 2D (Figure 7c), the 2D and Broyden 5% (Figure 7d) and the combination of 1D, 2D and Broyden 5% (Figure 7e) identies the major units and structures in the true model, but differs in the details. For some parts of the model more layers are present than is needed to describe the true model (e.g. around coordinate 200 m - 300 m). This causes soft transitions in both resistivities and layer boundaries. If any information on the prior geological appearance of the measuring site is present, these transitions can be controlled by the settings of the lateral constraints. Thus, if the lateral constraints on depths are tighter than the constraints on resistivities, transitions will mostly take place as changes in resistivity with constant layer boundaries and vice versa. In these examples we have tried to maintain average settings allowing both changes in resistivity and layer boundaries. The result from Res2dinv (Figure 7f) (Loke and Barker 1996) is rather similar to the layered results although the lower boundary is not accurately resolved. The residual curves (Figure 7g) follow very identical paths, but the Broyden 5% is somewhat slower to converge, and thus needs more iterations. All models t the data to the same degree. The timing of the models in Figure 7 are summarized in Table 2. This example is rather large in terms of data as well as model parameters explaining the rather large computation times. Also it is worth noticing that more iterations are needed for convergence compared to the previous simple dip model. We timed this data set with the Res2dinv as well. Setting recalculation of the Jacobian for the rst two iterations, this data set was inverted in 12532 s doing 8 iterations. This is comparable to the 12381 s we used in the optimized solution. Christiansen and Auken
12
Paper 5
Depth [m]
0 20 40 60 0
a) True model
50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Depth [m]
5 25 45 65 0 5 25 45 65 0
b) Pseudo section
50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Depth [m]
c) Full 2D
50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Depth [m]
5 25 45 65 0 5 25 45 65 0
d) Broyden 5%
50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Depth [m]
e) 1D Derivatives & Broyden 5%

50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Depth [m]
5 25 45 65 0 30
f) Res2dinv
50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
g)
10
Data residual
3 20 1 40 100 300 500
0.3 0 5 10 15 20 25
Iteration #
Figure 7: Von-Karman synthetic model. Panel a) is the true model,panel b) is the pseudo section, panel c) is the full 2D solution, panel d) uses Broydens update formula with a 5% recalculation criteria. Panel e) combines 1D derivatives with a 5% Broydens update. Panel f) shows the result from the Res2dinv program for comparison and nally, the data residuals versus iteration number are displayed in g). Filled markers (for Broyden 5% and 1D derivatives+ Broyden 5%) indicate full 2D iteration.
13
Method (c) Full 2D (d) Broyden 5% (e) 1D derivatives & Broyden 5% (f) Res2dinv
Computation time 32339 s 20158 s 12381 s 12532 s
# Iterations 17 23 16 8
# 2D iterations 17 11 7 2
(2003) showed that for a variety of synthetic examples the 2D-LCI performed equally well as Res2dinv comparing the inverted models point-to-point.
Field Examples
PACES, groundwater protection survey, Denmark
The PACES system has been used intensively in Denmark for hydrogeological mapping (Srensen et al. 2003). The line presented here is a small part of a 29 km2 campaign with a total of 113 km of PACES prole. The prole is 1 km long with one sounding containing 8 data points every 5 m. The model is described at 201 nodes, totally 1005 model parameters (3 layers). The resulting models (Figure 8b,c,d) have a high resistivity layer over a more conductive layer with a more resistive layer at the bottom. The model responses of the Broyden 5% (Figure 8c) and Broyden 5% combined with 1D derivatives (Figure 8d) are very similar and they are quite similar to the full 2D result (Figure 8b) as well. The thickness of the clay layer is the primary target of most PACES surveys because they form protective covers over potential groundwater reservoirs. Identifying windows in the clay layer is also very important. The timings of the PACES eld examples are summarized in Table 3. In this case the 1D derivatives + Broyden 5% reduces the computation time by approximately a factor of three. Method (c) Full 2D (d) Broyden 5% (e) 1D derivatives & Broyden 5% Computation time 14089 s 8301 s 4740 s # Iterations 9 14 11 # 2D iterations 9 5 5
CVES, road construction, Sweden

The last eld example is a CVES data set from Sweden. The prole is approximately 300 m. The resistivity survey was carried out as part of the geotechnical investigations for road construction in connection with a lled basin structure in bedrock. The data set consists of 536 data points measured with a CVES system using Wenner congurations with a-distances from 2m to 48 m. The data set has previously been presented by Dahlin
14
0
Paper 5
Depth [m]
5 10 15 20 0 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
a) Pseudo section
Depth [m]
5 10 15 20 0
b) Full 2D
50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Depth [m]
0 5 10 15 20 0
c) Broyden 5%
0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Depth [m]
5 10 15 20 0 10
d) 1D Derivatives & Broyden 5%

50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
e)
Data residual
3
10 30 100 300
0.3 0 5 10 15
Iteration #
Figure 8: PACES eld example. Panel a) is the data pseudo section, panel b) is the full 2D solution, panel c) uses Broydens update formula with a 5% recalculation criteria. Panel d) combines 1D derivatives with a 5% Broydens update. Finally, the data residuals versus iteration number are plotted in e). Filled markers (for Broyden 5% and 1D derivatives + Broyden 5%) indicate full 2D iterations.
(1996) and in (Auken and Christiansen 2003). The model has 71 horizontal nodes each with 4 layers, totally 497 model parameters The model results from the different inversions are practically the same. Minor differences are obtained in the outer lower parts of the prole, where practically no information is present, as depicted by the pseudo section (Figure 9a). All models correlated well with the information from the drillings, which identies three distinct units in the small peat basin. These units are very hard to distinguish using traditional smooth minimum structure inversion (Dahlin 1996). Between coordinate 20 and 100 there is less agreement between drill information and models, probably due to 3D effects associated with the basin edge (Dahlin 1996). The timing in Table 4 show that the combination of 1D derivatives with Broydens update brings the effective number of full 2D solutions down to only three which is very acceptable compared to the 10 needed if no approximations are used.
15
Depth [m]
0 10 20 30
30
(e)
(a) Pseudo section

0 50 100 150 200 250 10
Elevation [m] Elevation [m] Elevation [m]
25 15 5 35 25 15 5 35 25 15 5
(b) Full 2D
0 50 100 150 200 250
Data residual
3 1 0
35
(c) Broyden 5%
0 50 100 150 200 250
Iteration # Resistivities [ohm-m]

30 100 300 1000 3000
10
15
(d) 1D Derivatives & Broyden 5%

0 50 100 150 200
250
10
10000
Profile coordinae [m]

Figure 9: CVES eld example. Panel a) is the data pseudo section, panel b) is the full 2D solution, panel c) uses Broydens update formula with a 5% recalculation criterion. Panel d) combines 1D derivatives with a 5% Broydens update. Finally, the data residuals versus iteration number are plotted in e). Filled markers (for Broyden 5% and 1D derivatives+ Broyden 5%) indicate full 2D iteration. Lithological logs from drillings are located at every 20 meters from coordinate 20 m to 200 m. The colors of the drill holes indicate from the bottom: rock (dark gray), till (light gray) and clay (white).
Method (c) Full 2D (e) Broyden 5% (g) 1D derivatives & Broyden 5%
Computation time 9891 s 5074 s 3291 s
# Iterations 10 13 13
# 2D iterations 10 5 3
Discussion
In all the examples given here it makes good sense to use either of the approximations suggested, but for all cases the combination of 1D derivatives with a Broydens update had the shortest computation time and thus recommends itself. However, models with strong 2D features would most likely be difcult to optimize using the 1D derivatives solution in which cases either the full 2D or the 2D with Broyden 5% should be used. Extremely large data sets can be practically impossible to invert even with the approximate solutions suggested here. Normally a full 1D solution would be the alternate choice, but instead we suggest to use 1D derivatives throughout producing an intermediate result between the 1D and the 2D solutions. Another way to optimize the 2D-LCI is to shortcut the rst couple of iterations by producing a suitable starting model. This can be done with the multi-channel deconvolution as suggested by Mller et al. (2001). Having produced a smooth deconvolved 2D model we only need to nd the best tting layered models along the prole and use those
16
Paper 5
as starting models. Finally one could implement a dual-grid method because the size of the grid used in the nite element or nite difference calculations affects computation times dramatically. Thus, applying a ne grid for forward calculations and a coarser grid for calculation of derivatives could speed up computations (Torres-Verdin et al. 2000).
Conclusion
The 2D-LCI is a method utilizing a layered model description for 2D resistivity inversion. The models and data are grouped into pairs of soundings and models along the prole. This enables mixtures of 2D and 1D inversions. The parameters in the 2D-LCI are regularized using lateral constraints on layer resistivities, thicknesses and/or depths, producing laterally smooth models with discrete layers. We have presented the use of 1D derivatives and Broydens update with a 2D layered inversion of prole resistivity data. Broydens update as stand-alone is not usable with this kind of problem. To ensure convergence the Jacobian matrix needs to be reset using a full 2D solution every once in a while. We have used a 5% relative change in residual between the last two iterations as a criteria to update the Jacobian. Some of the speed gained using Broydens update is lost, but it is still approximately twice as fast as the full 2D solution. Calculation of the Jacobian matrix using a 1D mapping instead of a 2D forward mapping practically eliminates the Jacobian calculation as a time factor. However, the Jacobian becomes a block-diagonal containing no 2D information and it is only applicable for relatively smooth 2D earth structures. Used as a starter for a 2D Broydens update inversion scheme it proved very useful because very little 2D information is present in the rst few iterations. The inversions with 1D derivatives are very fast because only one full 2D forward calculation is needed for every iteration and they can be used as stand-alone for extremely large data sets where 2D solutions are not applicable. Whenever the 2D earth structures are not too strong we suggest using the inversion scheme combining 1D derivatives with a full 2D solution and Broydens update. The computation time for this inversion scheme were reduced by approximately a factor 3 compared to the full 2D solution.
Acknowledgements
The eld data were kindly provided by Prof. Torleif Dahlin, University of Lund, Sweden. We would also like to thank Prof. Douglas Oldenburg at University of British Columbia for letting us use their 2D resistivity code in our implementation.
References
Auken, E. and A. V. Christiansen (2003). Sharp boundary, laterally constrained 2D inversion (2D-LCI) of continuous resistivity data. Accepted for Geophysics, Online soon. Auken, E., A. V. Christiansen, B. H. Jacobsen, N. Foged, and K. I. Srensen (2004).
17
Piecewise 1D Laterally Constrained Inversion of resistivity data. Submitted to Geophysical Prospecting. Bernstone, C. and T. Dahlin (1999). Assessment of two automated DC resistivity data acquisition systems for landll location surveys: Two case studies. Journal of Environmental and Engineering Geophysics 4(2), 113121. Broyden, C. G. (1965). A class of methods for solving nonlinear simultaneous equations. Mathematics of Computation 19, 577593. Christiansen, A. V. and E. Auken (2003). Layered 2-D inversion of prole data, evaluated using stochastic models. Proceedings volume, 3DEMIII, February 2003, Adelaide. Dahlin, T. (1996). 2D resistivity surveying for environmental and engineering applications. First Break 14(7), 275283. Dahlin, T. and B. Zhou (2002). Gradient and mid-point referred measurements for multi-channel 2D resistivity imaging. In M. S. Matias and C. Grangeia (Eds.), Proceedings of the 9th Meeting, Environmental and Engineering Geophysics, Aveiro, Portugal, pp. Integrated Case Histories session. Environmental and Engineering Geophysical Society - European Section. Dey, A. and H. F. Morrison (1979). Resistivity modeling for arbitrarily shaped twodimensional structures. Gophysical Prospecting 27, 106136. Johansen, H. K. (1977). A man/computer interpretation system for resistivity soundings over a horizontally stratied earth. Geophysical Prospecting 25, 667691. Loke, M. H. and R. D. Barker (1996). Rapid least-squares inversion of apparent resistivity pseudosections by a quasi-Newton method. Geophysical Prospecting 44, 131152. Loke, M. H. and T. Dahlin (2002). A comparison of the Gauss-Newton and quasi-Newton methods in resistivity imaging inversion. Journal of Applied Geophysics 49(3), 149162. McGillivray, P. (1992). Forward modeling and inversion of DC resistivity and MMR data. Ph.D. thesis, The University of British Columbia, Vancouver, Canada. Menke, W. (1989). Geophysical Data Analysis: Discrete Inverse Theory (Revised edn. ed.). Academic Press Inc. Mller, I., B. Jacobsen, and N. Christensen (2001). Rapid inversion of 2-D geoelectrical data by multichannel deconvolution. Geophysics 66(3), 800808. Olayinka, A. I. and U. Yaramanci (2000). Use of block inversion in the 2-D interpretation of apparent resistivity data and its comparison with smooth inversion. Journal of Applied Geophysics 45, 6381. Oldenburg, D. W. and R. G. Ellis (1991). Inversion of geophysical data using an approximate inverse mapping. Geophysical Journal International 105, 325353. Oldenburg, D. W. and Y. Li (1994). Inversion of induced polarization data. Geophysics 59, 13271341. Panissod, C., M. Lajarthe, and A. Tabbagh (1997). Potential focusing: a new multielectrode array concept, simulating study, and eld tests in archaeological prospecting. Geophysics 38, 123.
18
Paper 5
Serban, D. Z. and B. H. Jacobsen (2001). The use of broad-band prior covariance for inverse palaeoclimate estimation. Geophysical Journal International 147, 2940. Smith, J. T. and J. R. Booker (1991). Rapid inversion of two- and three-dimensional Magnetotelluric data. Journal of Geophysical Research 96(B3), 39053922. Smith, J. T., M. Hoversten, E. Gasperikova, and F. Morrison (1999). Sharp boundary inversion of 2D Magnetotelluric data. Geophysical Prospecting 47, 469486. Srensen, K., E. Auken, N. B. Christensen, and L. Pellerin (2003). An Integrated Approach for Hydrogeophysical Investigations: New technologies and a Case History, accepted for publication in SEG Special Publication, Near-Surface Geophysics, Volume 2. Society of Exploration Geophysicists. Srensen, K. I. (1996). Pulled array continuous electrical proling. First break 14, 85 90. Torres-Verdin, C., V. L. Druskin, S. Fang, L. A. Knizhnerman, and A. Malinverno (2000). A dual-grid nonlinear inversion technique with applications to the interpretation of dc resistivity data. Geophysics 65(6), 17331745. Ward, S. H. and G. W. Hohmann (1987). Electromagnetic theory for geophysical applications. In M. N. Nabighian (Ed.), Electromagnetic methods in applied geophysics, Volume 1 of Investigations in Geophysics No.3, Chapter 4, pp. 285426. Society of Exploration Geophysics.

Transient Electromagnetic Method Application

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Transient Electromagnetic Method Application

Caricato da

Copyright:

Formati disponibili

ph.d.

Anders Vest Christiansen

Department of Earth Sciences University of Aarhus Denmark August 2003

Application of airborne TEM methods in Denmark

Layered 2D inversion of resistivity data

Papers included in this thesis

Selected additional publications

Selected conference abstracts and proceedings not included in this thesis

Anders Vest Christiansen

August 2003, Department of Earth Sciences University of Aarhus, Denmark anders.vest@geo.au.dk

Geophysical methods applied

The Danish groundwater project

Application of airborne TEM methods in Denmark

The research eld

The research eld

3.1 The TEM method

(3-1) (3-2) (3-3) (3-4)

Em F, from which Hm ends up as: 1 F + ( F). Hm = y z

1 u0 (z +h) [e + rTE eu0 (z h) ]J1 (a)J0 ()d, u0

where Jp is the Bessel function of order p, =

2 + k 2 for k and k wavenumbers, kx x y y

[eu0 (z +h) + rTE eu0 (z h) ]

which in the general case can be evaluated only be numerical integration.

The research eld

Figure 3.2: Schematic representation of PROTEM 47 and GEOTEM eld setups.

0 100 200 300

-300 -150 -300 -200 -100 0 100 200 300

0 0.2 0.4 0.6 0.8 1

The research eld

GEOTEM, z GEOTEM, x 10 -2 10 -1 Time [ms] 10 0 10 1

The research eld

Profile coordinate [m]

x 10 0 0.2 0.4 0.6 0.8

through the -parameter: min |dmeasured dprimary |2 . (3-16)

4.2 Field data

80 40 0 40 80 120 160 200 240 280 0.4 0.8 1.2

10 80 40 0 40 80 120 160 200 240 280

23.4 23.8 24.2 24.6

Profile coordinate [km]

Profile coordinate [km]

4.2 Field data

Elevation [m] Amplitude [V/m2]

Profile coordinate [km]

Discussion and conclusions

Discussion and conclusions

Discussion and conclusions

Layered 2D inversion of resistivity data

Clay ? Sand Sand

Profile coordinate [m]

Profile coordinate [m]

7.1 The Geoelectrical method

7.2 Inversion setup

xnx Nodes Profile distance

Depth [m] Profile coordinate [m]

Profile coordinate [m]

7.2 Inversion setup

a) CVES, 400 m layout

7.2 Inversion setup

The model estimate (Menke 1989) 1 G TC mest = G minimises Q= 1 N +A+M

Depth [m] Depth [m] Depth [m] Depth [m]

RES1 RES2 RES3 RES4 THK1 THK2 THK3 DPH3 DPH4

Profile coordinate [m] Analysis

8.1 Synthetic models

Depth [m] Depth [m] Depth [m]