Accuracy and Resolution in 2D Resistivity Inversion
byJeffrey Z. Snyder
B.S. Geology
Duke University, 1996
Submitted to the Department of
Earth, Atmospheric, and Planetary Sciences
in partial fulfillment of the requirements for the degree of
Master of Science in Geosystems
at the
Massachusetts Institute of Technology
May 2001© Massachusetts Institute of Technology 2001. All rights reserved.
Signature of Author... . ... . . ... --Department of Earth, Atmospheric, and Planetary Sciences 11 May 2001
Certified by ... %
...--.-.--.-Frank Dale Morgan Professor of Geophysics Thesis Supervisor
Accepted by... ... ... ... .. ... Ronald G. Prinn Chairman, Department of Earth, Atmospheric, and Planetary Sciences
Accuracy and Resolution in 2D Resistivity Inversion
by
Jeffrey Z. Snyder
Submitted to the Department of
Earth, Atmospheric, and Planetary Sciences on 11 May 2001 in partial fulfillment of the requirements for the degree of
Master of Science in Geosystems
Abstract
Two-dimensional resistivity inversion employing regularization enforces a constraint of smoothness that minimizes error and avoids unrealistically complex solutions to the inverse problem. The insensitivity of surface
array data to deeper targets typically presents an under-determined problem for solution by the inversion algorithm, and the smoothing function within the algorithm tends to "smear" tomographic imagery. Together, the physical challenges of electrical resistivity as a geophysical method and the numerical challenges inherent in solving the inverse problem introduce errors in the accuracy of 2D resistivity imagery.
It is important to know what surface array geometry will provide data that yields the best results from the inversion. Testing this inversion algorithm on data from several array geometries provides the opportunity to determine the comparative performance of each survey method.
By introducing two types of resistive anomalies and varying their location within a synthetic homogeneous
half-space, it is possible to generate synthetic data with a forward model algorithm. The data from each array geometry is then inverted in order to illustrate the accuracy and resolution response of the inversion algorithm. The inversion images are converted into binary images after defining a critical resistivity that describes the contrast between background resistivity and target resistivity. The binary images are used as interpretive tools that allow the user to overcome the "smearing" introduced by the inversion.
Because of its consistent performance from the margins to the center of an array, a left-right sweep geometry combined with a pseudosection geometry appears to be the best choice for a surface array when there is no
knowledge of the subsurface structure or resistivity distribution.
The critical resistivity and the area of the anomaly are used to describe the performance of the inversion. When taken as functions of increasing depth, the critical resistivity decreases and the area of anomaly increases, providing a respective correlation with the current density and the degree of smoothness. Initial results by forming a product of critical resistivity and area suggest that it is possible to approximate the product from the original forward model, but further testing is warranted to provide more conclusive results.
Thesis Supervisor: F. Dale Morgan Title: Professor of Geophysics
Acknowledgements
I'd like to take the opportunity to thank a few folks who helped me complete this work, and this program:
Professor Dale Morgan, who offered me the latitude to define a problem and run with it. Even to the end, his efforts to excite me about the science involved in this project, and where it was leading, helped give me the added motivation to drive forward with the project.
John Sogade, who provided me with the forward and inversion codes, taught me how to use them, provided me with a sounding board for ideas about the project, and helped me to better frame the problem as I immersed myself in it. John's guidance and editing were invaluable in helping me to provide a cohesive framework for the written report.
Dan Bums, whose positive attitude and flexibility helped me to pull this together right at the end. Dan was also instrumental in getting me to look at the fundamental science in this project, which made final revisions and alterations so much easier.
My officemates, Jiganesh Patel and Stacy Archfield, who comprised the Geosystems team this year. Successful completion of a program like this requires teamwork, and their support and assistance were fantastic.
Darrell Coles, whose MatLab prowess got me out of trouble on more than one occasion, and whose experience with Geosystems and with ERL made life easier.
I could not have been at MIT without the support of my closest friends and family, so to them I must say thanks: To my parents, who supported me throughout this entire process. To my sister, Larissa, who kept an eye out for me in the last few weeks of school and knew how to keep things under control at home in South Jersey. And to the Breault family, who offered their home in Rhode Island as a safe haven that I could visit in order to escape MIT.
Very special, heartfelt thanks must go to my fiancee', Lorna. Her immeasurable strength and support has kept me in the game since I arrived at MIT. Though I'm sure that I pushed the limits of her flexibility and understanding, she stood by to provide a pillar for me to lean on. Thanks so much for being there.
Finally, thank you to the Exxon Mobil Corporation for providing MIT with funding that ultimately paid for my fellowship and allowed me to be here in the first place.
Contents
1 Introduction
1.1 Background 6
1.2 Objectives 10
2
Formulation of the Forward and Inversion Problems
2.1 Current density as a function of depth 12
2.2 The forward model 16
2.3 Selection of grid discretization 22
2.4 Formulation of the inverse problem 22
3
Accuracy and Resolution Testing
3.1 Array geometries to be tested 29
3.2 Determining the critical resistivity 35
3.2.1 Binary plotting of inversion images 35
3.3 Critical resistivity as a function of depth 39
3.4 Calculating the integrated area of the anomaly 47
3.5 Smoothness as a function of depth 54
3.6 The product of per and the integrated area 61
4
Summary and Conclusions
4.1 Summary of Testing Results 68
4.2 Future Work 70
Contents (cont.)
Appendices
Appendix 1: Imaging of a 10,000 f-m anomaly. Central location. Al-i
Appendix 2: Imaging of a 500 1-m anomaly. Central location. A2-1
Appendix 3: Imaging of a 10,000 A-m anomaly. Left flank location. A3-1 Appendix 4: Imaging of a 500 9-m anomaly. Left flank location. A4-1
Chapter 1
Introduction
1.1 BackgroundElectrical resistivity surveying techniques have emerged as useful methods for
determining earth structure. By applying a known electrical current at various point source
locations on the earth's surface, surveyors attempt to infer the subsurface resistivity distribution
based on the different voltage potentials measured at the surface. Due to the sometimes wide
range of electrical resistivity values that certain materials in the earth possess, this technique
provides the surveyor with an effective geophysical method for delineating high contrast features
including reservoirs, cavities, and structure in the subsurface.
When conducting 2D resistivity surveying, researchers use linear arrays that consist of
electrical current sources and voltage potential receivers. The application of current at the surface
interacts with materials, objects, and structure in the subsurface, each with unique resistivity, to
produce voltage potential differences along the array. The locations of the voltage potential
measurements relative to the current sources dictate what area of the subsurface may possess the
resistivity that is effecting the voltage potential measurement. Varying the distance between the
current source electrodes and the magnitude of the current can change the current density and the
depth of current penetration beneath the array geometry. This in turn can affect the ability of the
voltage potential measurements to reflect the true resistivity distribution beneath the array.
Modern interpretation techniques utilize the solution of resistivity forward and inverse
problems. In our case the forward problem is the solution to the two-dimensional differential
equations that govern the relationship between applied currents and measurable potentials at the
system of blocks, each with resistivity p, (i =1 to m). The resistivities are called the model
parameters. The sole purpose of the inverse problem is to resolve the model parameters that will
fit the measured data to a prescribed tolerance and in so doing determine the subsurface
resistivity distribution.
Because of the potential to diagnose structure and material properties in the subsurface,
resistivity inversion has been investigated for applicability and effectiveness. Mufti (1976)
conducted 2-D resistivity inversion using a finite difference model utilizing a non-uniform
discretization of the earth. Later, Pelton et al (1978) constructed a faster, more computationally
efficient inversion algorithm using ridge regression and least-squares to achieve fast convergence
on a solution model that accurately represented the subsurface resistivity distribution. Tripp et al
(1984) expanded work in two-dimensional resistivity inversion by addressing the inverse
problem with a non-linear inversion method.
Recent work has attempted to optimize the inversion of surface measurements for the 2-D
and 3-D resistivity problems by utilizing computational methods that further increase the
accuracy and efficiency of algorithms while solving the inverse problem. The transmission
network analogy first developed for use in geophysical applications by Madden (1972) and then
adapted by others (Pelton et al, 1978; Tripp et al 1984; Zhang et al 1995) as a numerical
modeling approach to the resistivity inverse problem has yielded a computationally efficient
inversion algorithm that reduces computing time and minimizes error to produce accurate results.
Unfortunately, several physical and numerical challenges to solving the inverse problem
accompany any application of the resistivity method. The decrease in current density with depth
leaves the surface data relatively insensitive to deeper targets. The equivalence problem,
solutions because several subsurface resistivity distributions may be capable of producing
equivalent voltage potential measurements at the surface. The non-unique nature of the
resistivity inverse problem is further compounded by noisy surface data. This is because a
multiplicity of models often fit the surface data to prescribed tolerances. Which of the prescribed
tolerances works best is an open question.
Oftentimes, the surface potential measurements (data) are outnumbered by the model
parameters. This leads to an underdetermined inverse problem. Though determined or
overdetermined problems (where the amount of data is greater than the amount of model
parameters) are preferred, the inversion algorithm should be robust enough to accommodate the
underdetermined problems. Constraints that enforce requirements for acceptable solutions can
also be applied in order to overcome the non-uniqueness of the possible solutions to the inverse
problem.
Sometimes data seems to be insensitive to certain parameters or combinations of
parameters, which means that these parameters are irrelevant in an inversion sense. Similarly,
certain data points seem not to be affected by any of the parameters, so these are unimportant
data points. While survey arrays must be optimized to avoid unimportant data points, and models
should be chosen to avoid irrelevant parameters, situations often arise where such data points or
parameters exist in the formulation of an inverse solution. This leads to an ill-posed inverse
problem with consequent numerical instability.
Calculation of the inverse problem can yield several non-unique solutions that minimize
error in the forward model, when only one unique resistivity distribution should exist for a given
area. Assuming that there is no a priori knowledge of the subsurface resistivity distribution,
converge on inappropriate local minimum error solutions rather than global minimum solutions.
As a result, it is possible for inversion methods to yield any number of "rough" solutions that are
far more complex than the data or the reality of the earth may demand.
One approach towards overcoming the problem of non-uniqueness has been to apply
regularization to enforce a constraint of minimum roughness between nearest neighbors for each
iterative solution. This regularization pushes the inversion algorithm towards the simplest model
demanded by the data. It is equivalent to selecting a certain class of solution with prescribed
properties from amongst a variety of non-unique solutions (Shi 1998). In electromagnetic
sounding inversion, Constable et al (1987) coined the term "Occam's inversion" because their
inversion algorithm sought the simplest possible model demanded by the data. Their justification
was that "the real profile must be at least as rich as the profile found, but never less complex in
structure." LaBrecque (1996) later adapted this concept to inverting crosshole resistivity data.
While regularization yields simple, low error solutions, the affect of smoothing tends to
"smear" the results such that sharp boundaries are avoided if at all possible. There is then a loss
of accuracy and resolution in the subsurface resistivity model that becomes evident in the
resulting imagery created by the inversion.
The spatial distortion that accompanies this technique hinders appropriate interpretation of
subsurface imagery. In order to more effectively utilize the inversion method and its output
imagery, users should possess a perspective whereby they can expect the accuracy and resolution
behavior of the method in response to data produced by different anomaly types and locations
1.2 Objectives
To date, there has been little attempt to test this inversion method and standardize an
understanding of its accuracy response to different targets and locations relative to the surface
survey. Because of the physical challenges inherent with the resistivity method and the side
effects of overcoming the numerical challenges in the inverse problem, there are two primary
sources of inaccuracy and resolution loss. First, the weakening current density as a function of
depth and the insensitivity that it invokes in surface potential measurements makes it difficult to
"illuminate" relatively deep resistive or conductive anomalies within the subsurface. Second, the
nearest-neighbor approximation and smoothing introduced by the regularization of the inverse
problem creates a smearing and homogenization of deep anomalies. Understanding the roles that
these two factors play in effecting an accurate and well-resolved response from the resistivity
inverse problem will provide a perspective with which to interpret resistivity imagery, and
perhaps provide an avenue for future optimization of the inversion method.
In order to illustrate the accuracy and resolution behavior of this resistivity inversion
technique, this investigation will provide a catalog of images produced by the inversion
algorithm in response to synthetic data from two resistive anomaly types across a variety of
locations. After defining a critical threshold resistivity value in the imagery, we will use that
threshold value to create binary images as an added interpretive tool. The threshold value will
also be used to determine the integrated area of a resistive anomaly, enabling simultaneous
comparison of the current density, the degree of smoothing, and the integrated area of the
method and the numerical challenges of the inverse problem combine to affect the overall
accuracy performance of the inversion.
Finally, the testing regimen will be conducted for several surface array geometries,
expanding upon similar work conducted by Shi and Morgan (1997), in order to exhibit how
different data collection methods might provide more complete data to make the inverse problem
better determined. Because real-world applications provide little real data or imagery with which
to verify the accuracy of the inversion, the approach used here is to create synthetic surface data
using an algorithm that incorporates known electrical currents and a prescribed resistivity
distribution to solve the forward problem and produce surface potential measurements. The
synthetic data will then be utilized in the algorithm to solve the inverse problem, and the
resulting resistivity model can be compared with the initially prescribed model. By varying the
locations and the magnitudes of the prescribed resistivity anomalies, it is possible to assess some
standard behaviors of the algorithm so that we may gain a better understanding of the limitations
Chapter 2
Formulation of the Forward and Inverse Problems
2.1 Current density as a function of depth
The ability of current to penetrate the subsurface, travel through the earth, and return to the surface to yield surface data measurements seriously affects the ability of the any inversion
method to solve for deep resistive anomalies. A consideration of the 1-D case for a point current
source on the surface of a homogeneous isotropic earth illustrates how the voltage potential at a
location within the earth decreases with respect to increasing depth. Given Ohm's Law,
V= IR (2.1)
where V is the voltage potential at a given point, I is current flowing through the point, and R is
the resistance of the path. This can be expanded to consider the resistivity of the material in the
current path and the distance between the source and the point of interest, yielding
vIp 1
V = (2.2)
2;r r
Next, consider the 2-D case where there are two current source electrodes placed on the surface.
Figure 2.1(a) illustrates an example of a pair of current source electrodes and receiver potential
electrodes. The voltage difference between any two points on the surface, as a result of the
current flowing from one electrode to the other, can be given by
AV= - (2.3)
2rct r, r2 r3 r4
Where AV is the voltage potential difference between two points, P, and P2. I is the current
Power
Figure 2.1(a). A sample current electrode pair and voltage potential pair, used to illustrate equation (2.3). Reproduced
from
(Telford, 1990).Power
Figure 2.1 (b). Solving for current density at depth with a dipole source at the surface. Reproduced from (Telford 1990).
Given these relations, we can also determine the horizontal current density within a
homogeneous medium with resistivity p as
Ji = (-) (2.4)
P &
x= I ){x (x-L) (2.5)
2;r r,'r
(Refer to Figure 2.1(b) for a graphic representation of (2.5)) If we wish to find the current
density at a point P in the subsurface at a depth of z, and we assume that P lies on a line
perpendicular to the midpoint between the current sources, then (2.5) becomes
J = ( ) 3> (2.6)
2;
{(z2 +L)24.
where L represents the distance between the current sources and I represents the current. Holding
the current I constant, we can solve for J, (in amperes/m2) as a function of z for various values of
L. Figure 2.1(c) illustrates how the current density varies with L and z. The magnitude of the
current density at shallow depths is highest when the current source electrodes are closest
together. However, it is possible to increase the current density at depth by increasing the
distance between the current sources. The cost of achieving greater penetration is a loss of
current density in the shallow subsurface, and in all cases the general trend is for current density
Current Density I =I
10 10 10 10', 10
log Current Density, J (Mps/m2)
Figure 2.1(c). Current density as a function of depth, with I (current) being held constant.
-0
2.2 The forward model
Expansion of Ohm's Law into the three-dimensional case yields a system of nonlinear
differential equations that describe the relationship between applied surface currents, subsurface
resistivity, and surface voltage potential measurements. The system of equations is
V V(x, y, z) = -p(x, y, z)J(x, y, z) (2.7) V -J(x, y, z) = I(x, y, z) (2.8)
where again V represents the voltage potential, p is the resistivity, J is the current density, and I
is the current.
The resistivity forward model has a transmission network analog, first developed in
electrical engineering but later adapted to geophysical applications by Madden (1972). It consists
of network nodes, boundary nodes, and impedance branches. Based on this transmission network
model, the resistivity forward problem described by Equations (2.7) and (2.8) below can be
converted to a linear set of algebraic equations using Kirchoff s current law, which is then solved
using a bi-conjugate gradient algorithm. The schematic in Figure 2.2(a) exhibits the construction
of the transmission network. Note that for the geophysical resistivity problem, current sources
can be placed at network nodes, and voltage values are placed at nodes located on the top center
of each discretized block within the transmission network (Zhang et al 1995 and Shi 1998).
These expressions can then be solved numerically for the forward model by using a
discretized 3-D model of the subsurface. The 2-D forward model algorithm used in this
investigation uses a three-dimensional discretized subsurface region that is always three elements
thick in the y-direction while the x and z ranges are specified by the user and the middle slice is
z
* - network node
o - boundary node
Figure 2.2(a). A schematic of the transmission network analog as applied to the resistivity problem. Reproduced
from
(Zhang et al 1995).In order to perform the testing and analysis of the resistivity inversion algorithm, it is necessary to produce synthetic data from a known field of structure. In this case, the forward model utilizes a 15 x 30 grid of blocks to represent a discretized subsurface cross-section. Each block represents a spatial value of size and resistivity as prescribed by the user. For the purposes of this investigation, each grid column entry will be 1 m (meter) specifying the width of the block, and each row grid entry will be 1 m specifying the thickness of each block. However, the first and thirtieth columns will possess a width of 150 m and the second and twenty-ninth
columns will possess a 50 m width in order to maintain appropriate boundary conditions that will
improve the accuracy of the forward model results. Figure 2.2(b) illustrates this spatial grid. A
100 Q-m homogeneous body will represent the background resistivity of the field of interest. The forward model allows users to choose the number of anomalies to introduce into the
cross-section; however, this study will utilize a target block of dimensions 2 m high by 4 m wide with
varying resistivity and/or position.
The forward model algorithm requires the user to provide an input data file that
represents the survey geometry. Figure 2.2(c) shows an example of a dipole-dipole
pseudosection data set, and Figure 2.2(d) is an example of a left-right sweep data set developed
at the Earth Resources Laboratory (MIT).
The user possesses the flexibility of creating synthetic data sets that mimic selected
survey geometries and of creating data sets in response to varying subsurface resistivity
distributions. It is important to note that, in the physical sense, a variety of subsurface resistivity
distributions could be constructed to yield identical surface data sets. This problem of
equivalence will take on an even greater role of importance when we consider the inverse
150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150 150 50 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 50 150
Voltage (mV) Current (mA)
Total number of electrode source pairs Initial source pair locations
Number of receiver pairs Receiver pair locations
3 4 -4.08416 1 4 5 -1.67314 1 5 6 -0.83572 1 6 7 -0.47804 1 7 8 -0.29642 1 8 9 -0.19191 1 9 10 -0.12718 1 2 3 7 4 5 -4.08416 1 5 6 -1.67314 1 6 7 -0.83572 1 7 8 -0.47804 1 8 9 -0.29642 1 9 10 -0.19191 1 10 11 -0.12718 1
Voltage (mV)
Total number of electrode source pairs 3 4 Initial source pair locations
Number of receiver pairs Receiver pair locations
Current (mA) 2 3 -4.08416 1 3 4 -1.67314 1 4 5 -0.83572 1 5 6 -0.47804 1 6 7 -0.29642 1 7 8 -0.19191 1 8 9 -0.12718 1 9 10 -0.10153 1 10 11 -4.08416 1 11 12 -1.67314 1 12 13 -0.83572 1 13 14 -0.47804 1 14 15 -0.29642 1 15 16 -0.19191 1 16 17 -0.12718 1
Figure 2.2(d). Example offorward model data format for dipole-dipole left-right sweep array. (Note: Potential values are not representative of actual data. Figure has been provided only to
2.3 Selection of grid discretization
In order to conduct tests on the behavior of the inversion, it was necessary to establish an appropriate grid discretization for use in the forward model. In this investigation, we have chosen the simplest approach by using equivalent grid discretizations in both the forward and inversion grids. The forward model produces data based on a 15 x 30 element grid as illustrated in Figure 2.2(b), where each grid element within the boundaries is 1 m by 1 m in size. The inversion algorithm solves for model parameters based on the same grid discretization.
2.4 Formulation of the inverse problem
For the purposes of numerical calculations, the subsurface of the earth is discretized into a grid similar to that shown in Figure 2.4(a). Each block in the grid represents a resistivity value and a spatial dimension, such that the discretized subsurface electrical resistivity represents the parameters of the model, m, that are to be inverted. The data, d, constitutes all of the measured voltage-current pairs that are taken along a survey line at the surface.
The non-linear resistivity inverse problem requires solution of an iterative linear system of equations. This system of linear equations may be ill-posed because the initial guess for a solution in the iterative inversion algorithm may be far from the truth, and some of the data and/or model parameters may be irrelevant or unimportant to the solution of the inverse problem.
Without a priori knowledge of the subsurface structure forcing the inversion to converge
on a particular solution, it would be quite possible for the inversion to yield an unrealistically
complex, or "rough" solution. Though such a solution may be a minimum error result, the
complexity or roughness may be unnatural or perhaps more complex than the original data
Surface data observations, d
I I I I I I I I I I I I I
- I -, -, -I- 1-I-u - I - ~- I-I-I-I - I
-model parameters, m representing resistive blocks within the discretized subsurface.
Figure 2.4(a). Schematic comparison between surface data measurements and model parameters. 4-- - - -- - m - -
--
-
-
-
-
---
4.
E~E E~E-seek a low error solution while also enforcing a constraint of smoothing on the inversion.
Assuming that simplest solutions and the structures they represent are appropriate for
geophysical interpretation of high contrast structure in the real earth, the nearest-neighbor
regularization imposes a condition of uniqueness on the solution to the inverse problem by
requiring that it be the simplest low-error solution demanded by the data (Shi 1998).
To see how this regularization can be imposed, first consider a generic linear
algebraic case:
d =Gm + e (2.9)
For the resistivity inverse problem, d represents the surface data measurements (voltage
potentials), G represents the forward model that relates current, voltage, and resistivity in space;
m represents the unknown resistivity distribution in the subsurface; and e is the misfit error.
The Tikhonov regularization requires the definition and minimization of a function that
combines data misfit and the roughness of the model:
T = (d -Gm)) T R -'(d - Gm))+ r(m - m
)T L L(m - mo) (2.10)
where ![ function that must be minimized, R is the data covariance matrix, r is the regularization
parameter, and L is a linear operator (Tikhonov 1977). In (2.10), the first term represents the data
misfit produced by the model and the second term represents the "stabilizing functional" that
defines the spatial roughness of the model parameters so that (2.10) can be redefined as
T = s, + rs2 (2.11)
The inversion algorithm employs a non-linear conjugate gradient method that iteratively
attempts to find a solution model to the inverse problem while minimizing the function V. Each
of .For relatively high values of , s, dominates the behavior of the algorithm in attempting to
minimize . For low values of r, sj dominates the inversion algorithm.
If [ approaches a minimum value for both the first and second terms, the algorithm will stop iterating in order to prevent it from relaxing the smoothness constraint while in pursuit of
even lower error solutions. This ensures that the algorithm converges on the smoothest, low-error
solution. Figure 2.4(b) illustrates the behavior of V, z , s1, and S2 with advancing iterations.
While the smoothing constraint yields a simple, low error solution, the effect of the
smoothing also tends to "smear" the results, such that there is a loss in accuracy and resolution
that is evident in the subsurface imagery attained by the inversion. Figure 2.4(c) is a comparison
between imagery from a forward model and the related inversion. The top image is a graphic
representation of a single resistive block anomaly that was used to create synthetic surface data
in the forward model algorithm. The bottom image is the graphic representation of the inversion
results. The inversion tends to smear the anomaly downward, and it changes the values of the
resistivity in the target. The smearing in the imagery illustrates the effect that the
nearest-neighbor smoothing constraint has when producing the simplest model. The decrease in
resistivity values can be attributed to both the nearest neighbor smoothing during the inversion
and the increasing insensitivity of the surface data to relatively deeper targets.
Because of this spatial distortion, it is important for users of this inversion technique to
understand the accuracy and resolution response of the inversion based on the current density
provided by different survey geometries and the location of anomalies within the subsurface.
The design of this inversion algorithm for 2D resistivity produces a distinctive
manifestation in the output imagery, in that anomalies in the 'inverted' images are smeared as a
1 2 3 4 5 6 7 8 9 10 11
Iteration
Forward: Resistivity (Ohm-m) 0 10000 2 8000 4 E 6 6000 4000 10 12 2000 14 0 5 10 15 20 25 Distance (m)
inversion: Resistivity (Ohm-m) 0 7000 2 6000 4 5000 4000 'M 8 3000 10 2000 12 1000 14 0 5 10 15 20 25 Distance (m)
Figure 2.4(c). A comparison between a synthetic forward model (top) and the modelproduced by inversion (bottom) of the data generated by the forward model.
in this inversion method, the inversion creates a smearing of resistivity anomalies in the imagery
that can lead to misinterpretation by the user. This smearing effect is further compounded by the
decrease in data density and sensitivity for deeper targets. Due to the physics that govern the
flow of electrical current in the solid earth, the current density from surface sources decays with
depth. This combination of regularization and data density distorts accuracy and resolution with
increasing depth.
Sasaki (1992) undertook a resolution analysis of an inversion method that dealt with
borehole and surface potential measurements. The analysis he presented focused primarily on
factors internal to the inversion algorithm, such as grid discretization and damping coefficient
(T), with some investigation into electrode placement within the survey array. However,
systematic testing of the inversion in response to resistive anomalies that vary with location and
magnitude has yet to be done, and in this case may yield some consistent qualitative standards of
interpretation and performance, particularly when compared with the amount of smoothing
imposed by the inversion and the current density. Furthermore, there may also be ways to
increase the accuracy of the inversion by providing it better data in order to better determine the
inverse problem. Applying different array geometries may yield better current density or greater
sampling at depth. The remainder of this investigation will consist of a series of simple yet
elegant analytical tests that will qualify the accuracy behavior of the inversion in response to
Chapter 3
Accuracy And Resolution Testing
3.1 Array geometries to be tested
To illustrate the potential for improvement in inversion results, five surface array
geometries will be used to produce synthetic data from the forward model. Recall that our
forward model will utilize a 15 x 30 discretized grid to represent the subsurface, and the first, second, twenty-ninth, and thirtieth columns are use only to provide boundaries to the grid. Also,
referring back to the transmission network analogy reminds us that voltage values can be
assigned to the top center of each block and current values can be assigned to any network node.
For the purposes of creating synthetic data sets, current source locations and receiver locations
can be at any location between 3 and 28.
In this case, the dipole-dipole pseudosection array uses 23 current source pair locations.
Starting at the extreme left side of a field to be imaged, two adjacent electrodes (positions 3 and
4) are selected as the current source pair. Then, the receiver pairs are staggered across the
remainder of the electrodes along the survey line, with the terminus being the 2 8th position. The source pairs are then shifted one position to the right, and again the receiver pairs extend from
the source pair to the terminus. With this array geometry, a complete survey has a total of 276
potential measurements. Figure 3.1 displays a simple schematic of the survey geometry.
The left-right sweep begins with the current source pair placed at the extreme ends of the
survey line, such that one current electrode starts at the 3 position and the other begins at the 28
position. Potential measurements are then taken at all of the electrode positions between the
current source electrodes. After the measurements are taken, the left hand electrode is then
measurements are only taken between the source electrodes. This procedure continues until the
survey has utilized a total of 50 source pair locations for a total of 552 measurements. Figure 3.1(b) is a schematic of the left-right sweep.
The dipole-dipole middle sweep begins with the current source electrodes paired at the
extreme ends of the array line. The mid-point of the line is designated as an axis of symmetry.
All receiver electrodes are paired so that they are equidistant from the mid-point of the array. This procedure requires a total of 13 source pair locations for a total of 156 measurements.
Figure 3.1(c) is a schematic of the middle-sweep.
The final two arrays to be tested will be hybrids of those mentioned. One array will be a
left-right-pseudo, the other will be a left-right-middle sweep. For example, adding the data
collected from a left-right-sweep survey to a pseudosection survey gives us the left-right-pseudo
hybrid.
Figure 3.1(d) is a test matrix that describes the different parameters that will be varied in
r&"ti
4AW*b 4 4 V 4 A, V,
4 4 49,
49
.
4n=1
n=2
n=3
1*0Figure 3.1(a). The dipole-dipole pseudosection array geometry. The current source dipole is located at the top left hand side of the survey line, and the receiver dipoles are paired to the
right. Reprinted from (Shi 1998).
rtv V1
1 2 3
1 2 3
1 2 3
Figure 3.1(b). The left-right sweep. The current source electrodes begin at the extreme ends of the survey geometry, and all of the receiver pairs are between the source pairs. The left hand electrode is then swept to the right towards the right electrode. The process is repeated in order
Figure 3.1(c). The middle-sweep. The current source electrodes are paired, beginning at the extreme ends of the survey geometry. Receiver pairs are centered, symmetric to the mid-point of
Array Geometry Varied Pseudosection Left-Right Sweep Middle Sweep Left-Right-Pseudo Left-Right-Middle 2D Resistivity Inversion
Anomaly Type Anomaly Size
Varied Constant 10,000 92m 4 m wide x 2 m thick 500 Qm 1 Q-m Test Matrix Anomaly Depth Varied 1 m to 8 m depth Location Varied Center of grid Left flank of grid
Background Resistivity: Constant, 100 Q-m.
Figure 3.1(d). Matrix for varied parameters in 2D resistivity inversion accuracy testing. Note that the 1 -m has only been tested for centrally located anomalies.
3.2 Determining the critical resistivity
In order to determine the size and magnitude of the resistive anomaly produced by the
inversion algorithm, it is necessary to define a threshold resistivity value for each model that the
inversion produces. This critical resistivity, per, can be used as a baseline to create binary plots and integrated area plots, and it can also be used as a proxy for the decay of resistivity values that
can be witnessed for deeper targets in the forward model. Using the definition provided by Beard
and Morgan (1991) for resistive anomalies set against relatively conductive backgrounds,
x= log10 Pmax (3.1)
x= log10 Pmin (3.2)
Ax = (xh -x,)/3 (3.3)
Pcr =1 OXh (3.4)
For conductive anomalies set against relatively resistive backgrounds, the solution for critical
resistivity becomes
Pcr = lOx+ (3.5)
The calculation of the critical resistivity value contains information about the actual resistivity
contrast in the image that can be used as a simple descriptor of the accuracy in resistivity
magnitude that the inversion produces.
3.2.1 Binary plotting of inversion images
With our definition for the critical resistivity, we conduct an element-wise comparison of
the model parameters in the inversion grid to Per. Values in the inversion grid that are greater
than Per will are assigned the value of 1 (resistive), and those values that are less than Per received
the original forward model resistivity distribution that was used to create synthetic data, and the
solution model that was found by the inversion.
All of the figures for these tests can be found in the appendices. In each appendix, there are test results for each of the five arrays (pseudosection, right sweep, middle sweep,
left-right-pseudo sweep, and left-right-middle sweep). The appendices have been ordered according
to the following test examples: Appendix 1 possesses test results for a 10,000 Q-m target that
was located near the surface, along the midpoint of the array line. Appendix 2 is for a 500 Q m
anomaly located near the surface along the midpoint. Appendices 3 and 4 are for 10,000 2-m
and 500 Q-m targets, respectively, that were located near the left margin. Appendix 5 possesses
results from the 1 9-m case for centrally located anomalies.
Based on the inversion images and the binary plots in Appendices 1 and 2, it would
appear at first glance that the pseudosection array actually produced better imagery than the
left-right sweep, particularly for the anomalies on the centerline. The left-left-right-pseudo also produced
an accurate response that was comparable to the pseudosection alone. For a shallow target, the
inversion placed the anomaly at the appropriate depth with the appropriate size in response to
both the pseudosection data (Fig. Al-1) and the left-right-pseudo data (A 1-4), but there appears
to be some loss of resolution in the left-right sweep test case (Fig. A1-2). As evidenced in the
binary plots, the left-right-middle sweep (Fig. A 1-5) was nominally less effective than the
pseudosection and left-right-pseudo sweep, particularly for the 10,000 Q m target. The middle
sweep was very inaccurate for all centerline targets (Figs. Al-3, Al-8, A2-3, A2-8), producing
heavily smeared anomalies and significantly reduced resistivity values.
If we recall from Figure 2.1(c), the current density was greatest at the surface when L, the distance between the current source electrodes, was minimized. This explains the comparatively
better performance of the pseudosection array in response to centerline shallow targets, as the
distance between current source electrodes was always 1 meter. The increased current density in
shallow areas beneath the midpoint of the array made the surface potential measurements more
sensitive to resistive targets located there.
The images in Appendices 3 and 4 represent repetitions of the same experiments except
that the anomaly was located along the left margin of the subsurface grid. The utility of the
left-right sweep, and particularly the left-left-right-pseudo sweep becomes immediately apparent by their
ability to better resolve anomalies along the flanks. The left-right-pseudo sweep also produced a
closer approximation of the magnitude of resistivity.
The pseudosection array, with its apparently triangular capture zone, completely lost
accuracy and the ability to resolve any targets near the margins. The inversion rendered tear-drop
shaped anomalies smeared down and towards the center and much lower resistivity values in
response to data from the pseudosection array. The middle-sweep appeared to be very inaccurate
when applied alone; the results it provided permitted the inversion to render anomalies on both
flanks when only one target was introduced in the forward model. It would appear that the
left-right-middle sweep was not made more effective by the added presence of the middle-sweep
data.
By these images, we can see how the different array geometries used current source placement and voltage receiver placement to determine current density and subsurface sampling.
In the case of the pseudosection geometry, it accurately (relative to the other geometries)
captured shallow resistive targets near the centerline, but the geometry limited its ability to
Appendix 5 contains images from tests of a conductive (1 t-m) target set against a relatively resistive background. The imagery tends to be much more difficult to interpret, as the
inversion tended to return a more smeared anomaly than we saw in the cases of resistive targets.
While the resistive targets tended to be underestimated by the inversion, the conductive targets
tended to be overestimated in magnitude. However, the determination of the critical resistivity
value and the creation of the binary plot appeared to be a relatively robust approach towards
interpreting the imagery for the conductive targets. The performance of the inversion appeared to be consistent with that witnessed for the resistive targets with one notable exception: though the
left-right-pseudo sweep offered the best performance, the left-right sweep outperformed the
pseudosection for the conductive target.
To summarize, the performance of the inversion response to the following surveys
geometries can be ranked as follows:
For capturing a resistive anomaly near the array midpoint 1. Left-Right-Pseudo Sweep
2. Pseudosection
3. Left-Right Sweep
4. Left-Right-Middle Sweep
5. Middle Sweep
For capturing a resistive anomaly near the array margins 1. Left-Right-Pseudo Sweep
2. Left-Right Sweep
3. Left-Right-Middle Sweep
4. Pseudosection
5. Middle Sweep
For capturing a conductive anomaly near the array midpoint 1. Left-Right-Pseudo Sweep
2. Left-Right Sweep
3. Left-Right-Middle Sweep
4. Pseudosection
5. Middle Sweep
3.3 Critical resistivity as a function of depth
It is convenient that pc, is determined by the resistivity values within the solution to the inverse problem because the critical resistivity can also be used as proxy for the behavior of the
inversion. The response of the inversion tends to yield decaying resistivity values as a function of depth; the full field plots in the appendices exhibit this behavior. However, plotting per as a function of the depth of the anomaly can yield insight into the decay in resistivity that the
inversion produces. The decay in resistivity is in response to the lack of sensitivity in the data for
deeper targets. A plot of pr for a 10,000 Q-m and 500 &-m, centrally located anomaly in
response to pseudsection array data can be found in Figure 3.3(a). A similar plot for the inversion
results borne from left-right sweep data can be found in Figure 3.3(b), and 3.3(c) refers to the
left-right-pseudo sweep. All experiments were repeated for anomalies on the flanks, and the
results are plotted in Figures 3.3(d)-(f).
(Further testing will be performed on the 10,000 Q-m and 500 Q-m targets. The middle
and left-right-middle sweeps will be omitted from further testing. The results in Section 3.2
suggested that the left-right-middle sweep was not competitive with the left-right-pseudo sweep, and the middle sweep is quite inaccurate as a stand-alone survey.)
Recall that in Figure 2.1(c), we illustrated how current density decays with increasing
depth. In response to data from each array geometry, the critical resistivity value decays with
increasing depth, though the 500 9-m cases decay more gradually than the 10,000 Q-m cases. It
appears that the critical resistivity shares a directly proportional relationship with the natural
behavior of current density in the subsurface, but the rate and magnitude of decay of the critical
In comparing the quality of the results for these three array types, it appears that the
left-right-pseudo sweep data produced consistently better results, particularly because pcr reached an absolute maximum in response to the left-right-pseudo sweep data and because per was
consistently higher as a function of depth. The significance of this higher critical resistivity value
is that it describes, essentially, a higher magnitude of contrast between the target anomaly and
the background and thus a more accurate response. Physically, this is most likely a function of
the increased current density that the larger distances between current source electrodes tend to
produce for deeper targets.
It is interesting to note that the pseudosection produced a critical resistivity value for the
500 Q-m at 1 m that was a near match for the critical resistivity value in the forward model (Figure 3.3(a), lower plate). These results agree with the concept that the shorter dipole lengths
in the pseudosection array produced higher current density at shallow depths, thus providing
more current to interact with resistive targets in the shallow subsurface.
When targets on the margins were considered, the results show that the pseudosection
data produced a rapid decay for per. The left-right sweep and the left-right-pseudo sweep
produced more gradual decay with increasing depth, consistent with the results from Section 3.3
that indicated better capture in the margins by the left-right sweep and its hybrid, the
10,000 ohm-m Anomaly 500 1000 1500 2000 Resistivity (Ohm-m) 500 ohm-m Anomaly 120 140 160 180 200 220 Resistivity (Ohm-m) 240 260 280
Figure 3.3(a). Critical resistivity as ajfunction of depth. The inversion algorithm was run using synthetic dipole-dipole pseudosection data generated in response to the two centrally located
anomaly types. Notice the behavior of the critical resistivity with respect to depth, and the similarity with current density as a function of depth (Figure 2.1(c)).
S-2 E 0 c -3 .- 6 IL --C, 2500 -1 E 0--3 0-5 .r -4 1L CD--8' 10 F - Inversion Rho ocr __Forward Rhoc 0 300
10,000 ohm-m Anomaly 500 1000 1500 2000 Resistivity (Ohm-m) 500 ohm-m Anomaly -1 $ -2-E 0 -4 0-5 o --8 0- -Q. -8 L 0 -1 E 0 -3 -l 4 0-5 10 180 200 220 Resistivity (Ohm-m) 240 260 280
Figure 3.3(b). Critical resistivity as afuinction of depth. The inversion algorithm was run using synthetic left-right sweep data generated in response to the two centrally located anomaly types. Notice the behavior of the critical resistivity with respect to depth, and the similarity with current
density as a
function
of depth (Figure 2.1(c)).160 2500 0 120 - --I --_-- Inversion Rhocr Forward Rhoor -r -140 300
10,000 ohm-m Anomaly 2500 500 1000 1500 2000 Resistivity (Ohm-m) 500 ohm-m Anomaly ___ Inversion Rhocr -- Forward Rhoor -r -160 180 200 220 Resistivity (Ohm-m) 240 260 280 300
Figure 3.3(c). Critical resistivity as a
function
of depth. The inversion algorithm was run using synthetic left-right-pseudo sweep data generated in response to the two centrally located anomaly types. Notice the behavior of the critical resistivity with respect to depth, and thesimilarity with current density as a function of depth (Figure 2.1(c)).
-1 -E 0 C -3 IL 5 0 c 6-o -7 -8 -0 -6 -7 -81
10
0 120 140 --10,000 ohm-m Anomaly 500 1000 1500 2000 2500 Resistivity (Ohm-m) 500 ohm-m Anomaly I I I F 180 200 220 Resistivity (Ohm-m) 240 260 280 300
Figure 3.3(d). Critical resistivity as afunction of depth. Dipole-dipole pseudosection data in response to margin-located anomalies.
' -2 E 0--3 o -4 I-5 5-6 . 0 --2 E 0--3 0-5 -6 8-7 -8 '0 Inversion Rhoc Forward Rho I-I a a 120 140 160
10,000 ohm-m Anomaly 500 1000 1500 2000 Resistivity (Ohm-m) 500 ohm-m Anomaly -1 S-2 E 0-c -3 -z 4 0-5 0 -6 (. o -7 -8 -1 S-2 E 0 C -3 -5 4 0 U) -6 o -7 -8
i
240 260 280Figure 3.3(e). Critical resistivity as a
function
of depth. Left-right sweep data in response to margin-located anomalies. ) 160 180 200 220 Resistivity (Ohm-m) 2500 0 120 -_~- Inversion Rhoor _ Forward Rho - or 140 30010,000 ohm-m Anomaly 500 1000 1500 2000 Resistivity (Ohm-m) 2500 500 ohm-m Anomaly 120 140 160 180 200 220 240 260 280 Resistivity (Ohm-m) 300
Figure 3.3(f). Critical resistivity as a finction of depth. Left-right-pseudo sweep data in response to margin-located anomalies. -1 E 0 C -3 0 -5 o. 0. o -7 -8 E-2 E 0 c -3 0-5 .- 6 0. c-7 -810 100
3.4 Calculating integrated area of anomaly
Using the critical resistivity values and the binary plots created from the inversion results, it is also possible to characterize the smoothing imposed by the algorithm by integrating the area
of the anomaly and plotting it as a function of depth. Figures 3.4(a)-(c) are integrated area plots
based for anomalies produced by the inversion in response to 10,000 Q-m and 500 Q-m,
centrally located anomalies that were used to produce synthetic pseudosection array, left-right
sweep, and left-right-pseudo sweep array data, respectively. Figures 3.4(d)-(f) are integrated area
plots based on targets located near the margins.
Notice that as depth increases in each case, and as pcr decreases in Figures 3.3(a)-(f), the area of the anomaly increases due to smearing. (This inversely proportional relationship between
the critical resistivity and the integrated area will be addressed in section 3.6.)
This smearing is notably more erratic for the margin-located anomalies that were imaged
with pseudosection data, but such behavior is expected when compared with the results in
Sections 3.2 and 3.3, further reinforcing the concept that pseudosection subsurface illumination
10,000 ohm-m Anomaly 10 20 30 40 50 60 Area (m 3) 500 ohm-m Anomaly S2 -E 0-C -3 4 LL .- 6-Q. $ 78 -0 -1-0-3 -32 -C -6-a) o -7 -8 -0
Figure 3.4(a). Integrated area of anomaly. The anomaly size is calculated based on the size of the binary plot (created with the critical resistivity value Pcr ). The inversion was run in response
to synthetic dipole-dipole pseudoseclion array data generated for centrally located anomalies.
10 20 30 40 50 60
10,000 ohm-m Anomaly 10 15 20 25 30 35 40 45 50 55 Area (m 3) 500 ohm-rn Anomaly -1- S-2-E 0-C-3 0-5- 0-6- -7--8 L 5 -1 -$ 2-E 0-C -3 0-5- 0-6-o -7 -8 -0
Figure 3.4(b). Integrated area of anomaly. The anomaly size is calculated based on the size of the binary plot (created with the critical resistivity value p,). The inversion was run in response
to synthetic left-right sweep array data generated for centrally located anomalies.
10 20 30 40 50
10,000 ohm-rn Anomaly 10 20 30 40 50 Area (m 3) 500 ohm-m Anomaly -1 -N -2-E 0--3 -54 a. 0-5 U-o -7 -8 -0 -1 S2 -E 0--3 5-~ 6 a. 8 7--8 -0
Figure 3.4(c). Integrated area of anomaly. The anomaly size is calculated based on the size of the binary plot (created with the critical resistivity value Pcr). The inversion was run in response
to synthetic left-right-pseudo sweep array data generated for centrally located anomalies.
10 20 30 40 50 60 70
10,000 ohm-m Anomaly 5 10 15 20 25 30 35 Area (m 3) 500 ohm-m Anomaly 5 10 15 20 25 30 35 Area (m3 )
Figure 3.4(d). Integrated area ofanomaly. Pseudosection data. Margin-located anomalies. 1 - -2- -3- -4- -5- -6- -7--8 -0 -1 E 0 c -3 0-6-Q-7 -8 0 ' I I L~ I
10,000 ohm-m Anomaly 10 15 20 25 Area (m3) 500 ohm-m Anomaly 10 15 20 25 Area (m3)
Figure 3.4(e). Integrated area of anomaly. Left-right sweep data. Margin-located anomalies.
S-2 E 0--3 o-5 -6 o -7 -8 -1 -E 0--3 -4 r 6-C. o -7 -8 -5
10,000 ohm-m Anomaly -1 -@ -2 E 0--3 -4 0-5 U-0 -6 8 7 -8 10 15 20 25 Area (m3) 500 ohm-m Anomaly -1 E 05 -LL 5 10 15 20 25 Area (m3)
Figure 3.4(). Integrated area of anomaly. Left-right-pseudo sweep data. Margin-located anomalies.
3.5 Smoothness as a function of depth
Recall from equation (2.10), the functional f to be minimized. The second term of the
equation represents the roughness of the model and is given by
s2 = (m -m O)T LTL(m - mo) (3.6)
If we take the inverse of the roughness, then we can arrive at the smoothness of the model at the final iteration such that
Smoothness= (3.7)
S2
Plotting final iteration smoothness as a function of depth reveals how the smoothness
increases as the inversion algorithm attempts to impose minimal structure while solving an
increasingly underdetermined problem. Figure 3.5(a) is an example of this type of graph for
10,000 92-m and 500 -m anomalies that were used to produce forward model data analogous to a dipole-dipole pseudosection array geometry. Both the smoothness graph and the integrated area
graph tend to exhibit proportional behavior, such that they both appear to increase with depth.
This illustrates evidence of the smearing that the inversion algorithm imposes on anomalies.
Figure 3.5(b) and 3.5(c) are plots of inversion behavior in response to right sweep and
left-right-pseudo sweep data. Figures 3.5(d)-(f) repeat the tests in response to the geometries, but for
10,000 Ohm-m anomaly - -4 Smoothness 500 Ohm-m anomaly II I || 0 20 40 60 80 100 Smoothness 120 140 160 180
Figure 3.5(a). Final iteration smoothness as a function of depth. Pseudosection data. Centrally located anomalies. Notice that the trend is for the smoothness to increase with depth.
-6 -8 -10 -12 0 200
-10,000 Ohm-m anomaly -12 0 0 1 2 3 4 5 6 7 8 Smoothness 500 Ohm-m anomaly -2 -6 .1 - -10--12' 0 5 10 15 20 25 Smoothness
Figure 3.5(b). Final iteration smoothness as a
function
of depth. Left-right sweep data. Centrally located anomalies. Notice that the trend is for the smoothness to increase with depth.10,000 Ohm-m anomaly 0.5 1 1.5 2 2.5 Smoothness 500 Ohm-m anomaly 0.1 0.2 0.3 0.4 0.5 0.6 Smoothness
Figure 3.5(c). Final iteration smoothness as a function of depth. Left-right-pseudo sweep data. Centrally located anomalies. Notice that the trend is for the smoothness to increase with depth.
-2 E E -4 - -6--8 -0 -1 - -2--3 -4--5
-10,000 Ohm-rn anomaly I II I I--N 0 0.5 1 1.5 2 2.5 3 3.5 4 4. Smoothness 500 Ohm-m anomaly 5 10 15 20 Smoothness 25 30 35
Figure 3.5(d). Final iteration smoothness as a function of depth. Pseudosection data. Margin-located anomalies. 0 -2 -~-4 0 -2 E E -4 -6 I I I I I I I I I I