You are on page 1of 12

Journal of Applied Geophysics 105 (2014) 213–224

Contents lists available at ScienceDirect

Journal of Applied Geophysics


journal homepage: www.elsevier.com/locate/jappgeo

Efficient 2D inversion of long ERT sections


Panagiotis Tsourlos a, Nikos Papadopoulos b,⁎, Costas Papazachos a, Myeong-Jong Yi c, Jung-Ho Kim c
a
Department of Geophysics, School of Geology, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece
b
Laboratory of Geophysical-Satellite Remote Sensing and Archaeo-Environment, Institute for Mediterranean Studies, Foundation of Research and Technology,
Hellas (F.O.R.T.H.), P.O. Box 119, Rethymno 74 100, Crete, Greece
c
Korean Institute of Geosciences and Mineral Resources (KIGAM), Mineral Resources Research Division, Department of Exploration Geophysics and Mining Engineering, 124 Gwahang-no,
Yuseong-gu, Daejeon, South Korea

a r t i c l e i n f o a b s t r a c t

Article history: In this work a new algorithm for the efficient and fast two dimensional (2D) inversion of long electrical resistivity
Received 19 July 2013 tomography (ERT) sections is introduced. The algorithm is based on 2.5D finite element method (FEM) scheme to
Accepted 26 March 2014 solve Poisson's equation that describes the current flow into the earth's subsurface. The adjoint equation tech-
Available online 5 April 2014
nique was incorporated into the FEM framework to estimate the sensitivity values. The reconstructed 2D resis-
tivity models are recovered through an iterative, non-linear smoothness constrained least-squares approach.
Keywords:
Fast 2D inversion
The algorithm further incorporates an experimental procedure to avoid the calculation and storage of the entire
Jacobian matrix Jacobian matrix. The basic concept of this new algorithm relies on the fact that for every measurement there is a
Least-squares number of model parameters which are located in parts of the 2-D model at distant locations from potential and
LSQR current electrodes. The corresponding absolute Jacobian matrix values in such cases are very small (almost zero)
and can be omitted by the Jacobian calculation. Around every measurement a fixed rectangular threshold region
is defined a-priori based on geometrical criteria. The algorithm calculates only Jacobian matrix values for the
model parameters that are included in this threshold area omitting the calculation of the Jacobian entries related
to model parameters outside this region. This approach speeds up the Jacobian matrix calculations while the
efficient storage of the sparse Jacobian and Smoothness matrices and the inversion using an iterative routine
like LSQR method increase significantly the inversion speed and reduce the memory requirements. The new
algorithm is almost more than one order of magnitude (~30 times) faster and consumes one order of magnitude
(~90%) less storage memory than the original one based on full Jacobian calculations for typical applications. The
application of the new algorithm to synthetic and real data sets shows that the reconstructed models exhibit
comparable accuracy to the standard inversion approach.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction presented a “roll-along inversion” methodology by splitting up the


long ERT data into overlapping segments, which are subsequently treated
The electrical resistivity tomography (ERT) technique is increasingly in inversion by using sophisticated spatial constraints to account for the
used to solve a variety of engineering, environmental, hydrogeological, limited resolution in the boundaries of the separate models.
geological and archeological problems (Loke et al., 2013). The advent Despite the increasing CPU and memory computer resources, stan-
of automated multichannel resistivity meters has given the ability to dard 2D ERT inversion algorithms are still generally inefficient for deal-
perform ERT surveys in a limited time hence increasing the productivity ing with very large measured data-sets as the calculation, storage and
in field practice. Further, the combination of multichannel systems with inversion of the corresponding large Jacobian matrix render its usage
towed arrays for either land or marine applications has led to unprece- impractical. To partly alleviate this problem the use of Quasi-Newton
dented survey speeds and amounts of collected data. To efficiently cope techniques for approximately calculating (updating) the Jacobian ma-
with these long ERT lines, processing techniques which use one dimen- trix, rather than fully calculating it, has been proposed in literature
sional (1D) inversion approaches combined with lateral constrains have (Loke and Barker, 1996). These techniques are known to increase inver-
been proposed (Auken et al., 2005). However in cases of complicated sion speed yet they are less accurate when resistivity contrasts are high
subsurface structure, processing with full two dimensional (2D) inver- and further they do not reduce the memory storage space of the Jacobian
sion algorithms is still required. To this direction Günther (2007) matrix.
In the present work, a standard 2D ERT smoothness constrained in-
version algorithm (Tsourlos, 1995) was modified to perform efficient
⁎ Corresponding author. inversion of long ERT lines. The specific algorithm is based on the simple
E-mail address: nikos@ims.forth.gr (N. Papadopoulos). concept that in the case of long lines the majority of the Jacobian matrix

http://dx.doi.org/10.1016/j.jappgeo.2014.03.022
0926-9851/© 2014 Elsevier B.V. All rights reserved.
214 P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224

entries, which correspond to each measurement, are effectively zero. where C is a second difference smoothness operator (Sasaki, 1992).
This observation reflects the fact that most of the parameters are Simultaneous solution of Eqs. (1) and (2) is usually performed by min-
spatially far away from the measurement that they practically do not imization of the objective function:
affect it.
2 2
The proposed algorithm automatically identifies very small Jacobian U ¼ kdy−Jdxk þ λkrk ; ð3Þ
matrix entries based on a-priori geometric criteria. These entries are re-
lated to distant model parameters in relation to a specific measurement where || . || denotes the Euclidean norm and λ is a Lagrange multiplier.
and they are not included within specific threshold dimensions. Subse- Minimization of Eq. (3) leads to the following set of modified normal
quently the algorithm excludes these entries from the sensitivity equations:
matrix calculation thus, saving time and memory resources. More-  
T T T
over, to take advantage of the sparseness of the resulting Jacobian J J þ λC C dx ¼ J dy: ð4Þ
matrix, inversion is performed using the LSQR (least squares regres-
sion) method, which results in faster calculation time and has limit- The solution to the Eq. (4) is equivalent to the least-squares solution
ed memory requirements. of the joint system of Eqs. (2) and (3), i.e. (Sasaki, 1992):
It is noted that the main idea behind this work is the same with
   
what Papadopoulos et al. (2011) followed for the three dimensional pffiffiffiJffi dx ¼
dy
: ð5Þ
(3D) resistivity problem, but the actual implementation is rather λ C 0
different. In our case part of the Jacobian matrix is never calculated
since it is excluded in advance on the basis of predefined geometri- We use Eq. (5) to obtain the correction vector dx. Solving directly
cal criteria. Conversely, for the 3D case, Papadopoulos et al. (2011) Eq. (5) avoids the calculation of the JTJ and CTC matrices, as well as the
applied a specific scheme to calculate a variable masking region of matrix inversion involved in Eq. (4). The vector dx is added to the initial
the significant sensitivity area based on the calculation of the entire vector x0 to obtain the updated resistivity parameters. Furthermore, the
Jacobian matrix prior to the inversion procedure. Such an approach use of an appropriate method on Eq. (5) (e.g. conjugate gradients) can
was also followed by Günther et al. (2006) who calculated the entire accelerate the inversion procedure and possibly save computer memo-
sensitivity matrix and subsequently stored only the values that ry. The procedure is repeated until the misfit between the measured and
exceed a specific predefined threshold value in order to save PC modeled data is reduced to an acceptable relative root mean square
memory. (RMS) level or a maximum number of the iterations has been reached.
In this work the particular algorithm is presented in detail and the re- The maximum number of iterations for the synthetic and field data in
sults are compared to a standard 2D inversion scheme for the case of this work was set equal to 7. The quality of the collected data dictates
synthetic data deriving from various models and different geoelectrical the choice of the RMS threshold and corresponds to an estimation (or
arrays. The algorithm is evaluated with respect to the quality of the calculation) of the apparent resistivity data measured error. The La-
inversion results and its efficiency in computing speed and resources. grange multiplier weights the model constraints against the data misfit.
Finally its application is also demonstrated with field data. The starting value of Lagrange multiplier was set to 0.1 for all the tested
models and real data. This value was halved in the later iterations and
2. Direct current resistivity inversion remained constant after the 5th inversion iteration.

In this work the forward resistivity calculations were performed 3. Algorithm description
using a proven 2.5D finite element method (FEM) algorithm (Tsourlos,
1995). A 2.5D model suggests that the source is considered to be 3-D The basic concept of the present work relies on the fact that for every
but the subsurface resistivity is only 2-D. The technique has been exten- surface measurement there is a number of model parameters for which
sively described in many works (Coggon, 1971; Rijo, 1977) so no further their absolute Jacobian values are very small and they can be regarded
description of the modeling method is presented here. as almost equal to zero. These parameters are typically expected for
The DC linearized inverse problem can be expressed as parts of the 2D model space at distant locations from potential and cur-
rent electrodes. Thus, it is reasonable to assume that any larger or small-
dy ¼ J dx; ð1Þ er resistivity change of these parameters will not contribute any
additional useful information in the specific measurement. This can be
where dy = y − F(x) is the vector of differences between observed actually illustrated in the example of Fig. 1 for the sensitivity distribu-
(y) and calculated data (logarithm of apparent resistivities, F(x)), dx is tion of a specific surface dipole–dipole measurement with dipole spac-
the correction vector to the initial model parameters x0 and J is the ing 1a and separation between current and potential dipoles Nsep =
Jacobian matrix for the x parameter (logarithm of resistivity) distribu- 4a, from which it is clear that there are regions for which the calculated
tion. The Jacobian matrix links the variation of the apparent resistivity Jacobian matrix entries are practically zero (Fig. 1—bottom). In practice
data (ραi) with the change of model subsurface resistivity (ρj). The these parameters cannot be resolved from this measurement, as it does
adjoint equation technique was incorporated into the FEM framework not carry the necessary information content to reconstruct these far
to estimate the values of the Jacobian matrix (McGillivray and distant model parameters. In practice the existence of any resistivity
Oldenburg, 1990). A common approach for the model parameteriza- inhomogeneity in this distant model parameter will not affect the mea-
tion is to divide the earth into rectangular blocks of unknown constant sured resistivity value, as demonstrated by Papadopoulos et al. (2011)
resistivity, with cell dimensions (X, Z) closely related to the minimum in a similar case for the 3D resistivity problem.
electrode spacing, (a), [e.g. (X, Z) = (a, a/2)]. The problem now becomes how to define the corresponding impor-
Since the geoelectrical inverse problem is ill-conditioned, unstable tant sensitivity region for each different measurement i.e. the model
and non-unique (Tikhonov et al., 1998), it is necessary to impose addi- parameters for which the Jacobian matrix entries can be considered
tional constraints on Eq. (1) in order to obtain reliable results. A typical as significant for the inversion procedure and which needs to be calcu-
approach, also adopted here, is to incorporate additional smoothness lated. In the present work we defined this area experimentally, using
constraints on the model in Eq. (1) in the form: specific geometric criteria based on the relative position and distance
of each apparent resistivity measurement with respect to all model pa-
rameters that describe the subsurface resistivity variation. To this direc-
r ¼ C dx ¼ 0 ð2Þ tion, several sensitivity patterns of various standard arrays in
P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224 215

Fig. 1. 2D sensitivity distribution of a surface dipole–dipole configuration with Nsep = 4a, where a is the minimum dipole spacing (top). Graph of the respective sensitivity values plotted in
increasing order (bottom). The region of minimal sensitivity corresponds to the far distant model parameters with respect to the location of the electrodes that are used to collect the
apparent resistivity data.

combination to different resistivity distributions were observed to processing procedure a specific value is assigned to the parameter t.
obtain safe conclusions. Then, during the Jacobian calculation within the inversion program
Assuming that the distance between the external electrodes of the distance d of the external electrodes of every measurement is calcu-
one particular measurement obtained with a standard electrode array lated through their relative coordinates along the survey line and the
(e.g. Dipole–Dipole, Pole–Dipole, Pole–Pole, Wenner, Schlumberger, rectangular threshold region is defined for each specific measurement.
Gradient, etc.) with typical geometry (inter-electrode spacing a and Row entries of the Jacobian matrix associated with every measurement
Nsep = 1a, 2a…,Nmax a) is d, then we define a rectangular threshold are calculated only for those parameters which fall within the dimen-
region around the specific measurement (Fig. 2) with dimensions t.d sions of this rectangular threshold region, while for parameters outside
along the X axis and t.d/2 along the Y axis. The floating parameter t this area the Jacobian entry is neither calculated nor stored.
controls the horizontal and vertical extend of this rectangular region. The above strategy during the Jacobian calculation will finally result
A smaller high-sensitivity region is defined when assigning small values in obtaining a sparse sensitivity matrix. Moreover, the smoothness ma-
to the parameter t, which will finally lead in the actual calculation of trix that describes the relation of each parameter with its immediate
fewer entries for the Jacobian matrix. Assigning a large positive value neighbors is also sparse by its definition. Thus, the augmented matrix
to parameter t (e.g. 106) actually leads to the standard inversion in the left hand side of Eq. (5) will be eventually large and sparse. In
approach where the complete Jacobian matrix is calculated. order to reduce the total memory requirements, only the non-zero
The implementation of the above concept in the inversion algorithm elements of this augmented matrix are efficiently stored as vectors
is rather simple and straightforward. In the beginning of the inversion (Papadopoulos et al., 2011). Furthermore, iterative matrix solvers like

Fig. 2. Sensitivity distribution of a dipole–dipole measurement with Nsep = 3a, where a is the minimum dipole spacing. The area inside the red rectangular depicts the orthogonal threshold
region defined for this measurement, for which the Jacobian matrix entries will be calculated.
216 P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224
P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224 217

LSQR (Paige and Saunders, 1982) can take advantage of the sparseness the parameter t) can result in stable and successful inversions. However
of such matrices for the solution of the joint system of Eq. (5) and the given that the spatial sensitivity pattern may significantly change with
updating of the model resistivity during the different iterations. inhomogeneity, it was decided to finally adopt the value t = 2 as the
LSQR is based on the Golub–Kahan bidiagonalization (Golub and standard parameter to determine the dimensions of the threshold re-
Kahan, 1965) and it is a generalized conjugate gradient (CG) method gion for each different measurement. This choice is relatively conserva-
that operates directly on the augmented matrix of Eq. (5). LSQR reduces tive, but minimizes the possibility of removing significant information
monotonically the linear misfit (quantitative measure of the quality of from the Jacobian matrix during the later iterations of the inversion
the updated solution) during its iterations and approximates the solu- procedure.
tion of over determined and under determined linear systems of the
form Ax = b in an iterative manner. LSQR involves only a few multipli- 3.2. Efficiency in computational resources
cations of the augmented matrix of Eq. (5) with vectors, hence the
whole process can significantly speed up by taking advantage of any Adopting a constant value for parameter t (t = 2) means that the
matrix sparseness in these multiplications, such as the case of the actual percentage of the Jacobian matrix values that will be finally calcu-
 
J lated and stored depends on the number of electrodes that are used
matrix A ¼ pffiffiffiffi .
λ C along the ERT line, the total number of the apparent resistivity data
It is very important to set up suitable terminating criteria for the that will be captured and the number of model blocks that will be
LSQR due to its iterative nature. A comprehensive discussion on this used to discretize the subsurface. As we show in Fig. 4a the percentage
subject is provided for the 3D resistivity inversion problem by of the Jacobian matrix which is actually important and needs to be
Papadopoulos et al. (2011) and a similar strategy was followed in the updated and stored within the inversion procedure decreases exponen-
case of the 2D inversion algorithm. The convergence of the LSQR de- tially with the number of electrodes and the number of measurements.
pends on the singularity of the Jacobian matrix and the relative error At the same time the efficient storage of only the non-zero elements of
in the data. In general, there is not an objective criterion to determine the sparse Jacobian and smoothness matrices in vectors can reduce the
the optimum LSQR termination step, as this depends on the degree of memory requirements up to 96%, since the computer needs up to 30
the linearity of the problem. As a rule of thumb it is recommended to times less storage memory for these matrices (see Fig. 4b). In order
stop the LSQR steps when the linear misfit does not show significant to demonstrate the memory efficiency of the new algorithm, a dipole–
changes. This approach may result in a larger number of non-linear dipole ERT survey over a line of 100 electrodes is considered, assuming
iterations of the whole inverse problem, but allows the determination measurements with maximum dipole separation up to 6. The resulting
of robust, physically meaningful resistivity solutions. 580 measurements and 550 model parameters (unknown resistivities)
with the new scheme involve the calculation of only 10% of the full
Jacobian and requires just 15% of the storage memory (0.38 MB)
3.1. Deciding the threshold value compared to storing the full Jacobian (approximately 2.5 MB).
Furthermore the calculation of only the Jacobian matrix values for
An important issue concerning the new algorithm refers to the choice parameters included in the threshold rectangular region of each mea-
of the optimum value for the parameter t, which controls which part of surement results in a dramatic decrease (typically more than 90%) in
the model Jacobian matrix values will be actually calculated and subse- the time needed to perform the sparse Jacobian matrix calculation
quently stored. An empirical approach was followed, that did not employ (see Fig. 5a). Similar accelerations are also obtained by the use of the
any kind of an optimization workflow. This was accomplished by testing LSQR routine in the augmented matrix inversion (see Fig. 5b).
a number of different values of t with a variety of simple and more com- Fig. 6 illustrates the total time that it is consumed by the new and the
plicated resistivity models and electrode configurations. original algorithm to complete seven non-linear iterations for different
An indicative example is illustrated in Fig. 3 where we present the data sets with increasing number of electrodes and measurements.
resistivity distribution of the original model that was constructed to The tests were performed on a PC with Intel Core i5 CPU and 8 GB
simulate a resistive (1000 Ω-m) and a conductive (5 Ω-m) prism with RAM. It is obvious that in terms of overall speed the new algorithm is
similar dimensions (10 × 5 m) buried in a 12 meter thick layer more than one order of magnitude (~30 times) faster than the original
(100 Ω-m). A more conductive layer (10 Ω-m) is located below one especially in relatively large inversion problems.
the depth of 12 m. A long ERT line was simulated assuming the
Dipole–Dipole array with a = 5 meter inter electrode spacing and 4. Algorithm evaluation with synthetic data
maximum Nsep = 8a. The 2.5D forward FEM algorithm was used to
generate the 588 noise free synthetic data for the particular model Several tests were performed with synthetic data generated from
and survey parameters. The synthetic data were inverted with the 2D various resistivity models to evaluate the efficiency of the new
inversion algorithm using different values for the parameter t. The algorithm. Fig. 7(a) depicts a typical resistivity model representing a
model space was divided in a total number of 584 parameters. The in- common geological setting, namely a fault with a vertical depression
version model where the complete Jacobian matrix is calculated within of a resistive layer with resistivity 500 Ω-m with respect to an overlain
the inverse procedure (t = 106) is also shown for comparison. conductive layer with resistivity 30 Ω-m. A local resistivity inhomogene-
The final inversion models after seven iterations exhibit low RMS ity (1000 Ω-m) is attributed to the block with X = 165–180 m and Z =
errors (less than 0.2%). The proposed and the original algorithms result- 7.5–12.5 m. A relatively long ERT section with a typical configuration
ed in 2D inversion models of comparable accuracy. More specifically (inter line spacing a = 5 m, 50 electrodes) perpendicular to the fault
these inversion models exhibit a mean relative difference of less than was considered. The Dipole–Dipole array with a maximum Nsep = 8a
2.8% (in cases where t N 1.5). Even in the extreme case of calculating was used, resulting in 348 measurements. The forward response for
only the 7% of the entire Jacobian matrix elements, the algorithm man- this specific resistivity distribution and survey parameters was calculat-
aged to converge to an inversion model that is a very good representa- ed with the 2.5D FEM modeling algorithm. Realistic data errors were
tion of the original model. These results clearly show that the definition simulated by adding 3% Gaussian noise prior to their inversion. The re-
of a relatively small threshold region (by assigning a very small value to constructed 2D resistivity models are presented in Fig. 7(b,c) for the

Fig. 3. Original model (top) that was used to generate the synthetic apparent resistivity data for a long ERT line with 80 electrodes using the Dipole–Dipole array. The results from the data
processing with the 2D resistivity inversion algorithm are shown for different values of the parameter t that control the dimensions of the threshold region that is used to calculate only the
significant part of the Jacobian matrix. The percentage of the complete Jacobian (Jac) that was calculated is also shown for each case.
218 P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224

Fig. 4. Relation between the number of electrodes and measurements of a long ERT line as a function of the actual percentage of the Jacobian matrix elements that are calculated and stored
with the new algorithm (for a value of t = 2). The dashed line corresponds to the best-fit power law in the chart. The memory requirements (in Mb) for the original and the new algorithm
inversion matrix storage are presented in the bottom image. The relative decrease (bold numbers) of the memory requirements for the new algorithm due to the efficient storage of the
non-zero elements of the augmented matrix is also shown.

original and the new algorithm respectively. Both algorithms converged and they exhibit a mean relative difference of 1.2%. Fig. 8d shows that
to a final solution after 7 iterations with practically similar RMS (~1.5– the distribution of the relative differences between the inverted resistiv-
1.6%) recovering the main characteristics of the original model. The in- ity models resulted by the original and new algorithm lies within the
version models resulted by the original and the new algorithm are visu- range of +/− 5%. The results from this synthetic example verify that
ally identical and they exhibit a mean relative difference of 0.5%. Fig. 7d the proposed algorithm can produce reliable geoelectrical models,
shows that the distribution of the relative differences between the even in cases of demanding and complicated geological structures, with-
inverted resistivity models resulted by the original and new algorithm out information loss with respect to the standard ERT tomographic
lies within the range of +/−1.5%. approach.
The second test deals with a rather more complicated and demand- The third synthetic example (Fig. 9a) represents a typical scenario of
ing situation, where a 10 Ω-m conductive superficial layer with thick- recent valley deposits composed of different stratigraphy units. The
ness 5 m is overlain by a 100 Ω-m resistive half-space layer (Fig. 8a). blocks with very low resistivity values (10 Ω-m) simulate a clay horizon
Two small high resistive bodies (500 Ω-m) indicate local inhomogenei- below the depth of 4.5 m and a very-fine material that fills a superficial
ties close to the ground surface. The left prism is totally embedded in the ditch (horizontal distance 11–24 m). A sandy–clayey layer with resistiv-
first conductive layer, while the right body is situated in the contact be- ity 100 Ω-m (horizontal distance 31–95 m) is formulated from the sur-
tween the two main layers. A diagonal conductive dyke (10 Ω-m) at the face up to the depth of 1.5 m. The blocks with the higher resistivity value
center of the section divides the bottom layer into two main parts. Sim- (200 Ω-m) show lenses of a coarser material (gravels and sand) that
ilar survey parameters as in the case of the first model were used to gen- have been deposited within the silty–clayey horizon (resistivity 50 Ω-
erate the synthetic data (Dipole–Dipole array, 50 electrodes, a = 5 m, m).
Nsep = 8a, 3% Gaussian noise). Once again the new algorithm recon- The forward response of the specific model was calculated assuming a
structs the original model with great accuracy and similar RMS error layout of 96 equally spaced electrodes (a = 1 m) along the line. The sim-
with respect to the original algorithm (Fig. 8b,c). The inversion models ulated apparent resistivity data corresponding to Wenner–Schlumberger
resulted by the original and the new algorithm are visually identical (WS), Pole–Dipole (PD) and Dipole–Dipole (DD) array configurations,
P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224 219

Fig. 5. Computation time of the new algorithm for the calculation of the sparse Jacobian matrix (up) and the inversion of the sparse augmented matrix with LSQR (down), with respect to
the original algorithm for one non-linear inversion iteration.

with multiple combinations of electrode spacing and N separations (1a, out at the embankment of the west bank of river Elba at the area of
2a, 3a, Nsep = 10a) were extracted and corrupted with 3% Gaussian Havelberg (Sachen Anhalt), Germany (Fig. 10a). The geological forma-
noise. The WS, PD and DD data sets contained 2160, 2490 and 2430 mea- tions in the region consist of sands, mixture of sand and clays and lenses
surements respectively. of shale. The target of the geophysical survey was to provide information
The WS, PD and DD were inverted with the original and the new algo- regarding the subsurface structure of the embankment and in particular
rithm and the corresponding resistivity inversion sections are shown in to identify lithological variations (i.e. shale lenses) possibly associated
Fig. 9(b1–b2; c1–c2; d1–d2). The new algorithm omitted the calculation with structural vulnerability zones (Sotiropoulos et al., 2002).
of 55%, 75% and 70% of the Jacobian matrix elements during the inversion The presented algorithm was applied to a long 2D ERT line (total
of the WS, PD and DD data respectively. The inversion models resulted by length of 1500 m) which was situated on the top of the embankment
the original and the new algorithm look practically identical for all the (Fig. 10b). Data were obtained using a dipole–dipole array in a roll-
three arrays. This is also verified by inspecting the mean relative differ- along mode with inter-electrode spacing of a = 5 m and a maximum
ences: Fig. 9(b3,c3,d3) shows that the distribution of the relative differ- Nsep = 6a. Inversion results using both the standard and the proposed
ences between the inverted resistivity models resulted from the original algorithm are depicted in Fig. 11. The resulting models show a top
and new algorithm for all the arrays lie within the range of +/−3.5%. more resistive layer representing unsaturated sand, which is followed
by a lower resistivity formation representing the transition between
5. Application of the algorithm to field data areas of almost pure sand (higher resistivities) and sand with higher
clay content (lower resistivities).
The presented algorithm was also tested with real data, which were Overall for both applications convergence was obtained after 6 iter-
obtained in the framework of a 2-D geoelectrical survey that was carried ations reaching similar %RMS errors (3.6% for the standard and 3.7% for
220 P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224

Fig. 6. Comparison of the total processing time of the original and new algorithm for seven non-linear inversion iterations, for different numbers of measurements and electrodes. Bold
numbers indicate the acceleration factor of the new algorithm with respect to the original one.

Fig. 7. Model 1 — Fault: (a) Original resistivity model. Inverted 2D resistivity sections resulted by the data processing with the original (b) and the proposed (c) inversion algorithm. A long ERT
dipole–dipole array was simulated with 50 electrodes, a = 5 m and Nsep = 8a. (d) Relative difference between the inversion models resulted by the standard and the new algorithm.
P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224 221

Fig. 8. Model 2 — Dyke: (a) Original resistivity model. b) Inverted 2D resistivity sections resulted by the data processing with the original (b) and the new (c) inversion algorithm (the
configuration is the same as in Fig. 7). (d) Relative difference between the inversion models resulted by the standard and the new algorithm.

the new algorithm). At the same time the data processing time with the potential electrodes in long ERT sections do not essentially influence
new algorithm was approximately 40% of the time required by the the behavior of this specific measurement. The algorithm incorporates
standard 2D inversion approach. The corresponding inversion results an empirical procedure verified by testing to avoid calculation and stor-
obtained from both techniques are very similar; with the mean relative age of the entire Jacobian matrix. Only the larger absolute Jacobian
difference between the two inversion models (standard and new) being values that correspond to the parameters which are included in a
below 1.6%. By inspecting the relative differences between the inverted threshold orthogonal area around each resistivity measurement are cal-
results (Fig. 11c) it can be noticed that the maximum differences are culated. The definition of the rectangular region is performed empirical-
concentrated at the bottom parameter layers of the inversion area. ly based on the maximum distance (d) of the outer electrodes that
This can be explained by considering that these layers suffer from low participate in obtaining each particular measurement with any kind
resolution and thus the data errors can “contaminate” more the bottom of configuration. The threshold area is defined as a multiple (t) of
part of the model. As the very low (almost zero) entries of the Jacobian this outer electrode distance (d) by using a simple formula of the form
matrix are zeroed by the “fast” algorithm this “contamination” is now [t d x t/2 d], where the parameter t controls the actual region dimensions
changed resulting into differences which are localized at the bottom in the X and Z directions.
layer parameters. An optimum value for t was defined through a number of tests with
different electrode arrays and synthetic models simulating diverse
6. Discussion and conclusions geological conditions and the value t = 2 is finally recommended.
Although this choice seems to be relatively conservative, it is derived
In this work a new algorithm for the efficient and fast 2D inversion of from the threshold dimensions determined by the range of Jacobian
long ERT lines is proposed. Its main concept relies on the assumption matrix values for homogeneous earth. Since the sensitivity values are
that far distant model parameters with respect to the current and strongly controlled by the spatial resistivity distribution, this value
222
a)

b1) c1) d1)

P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224


b2) c2) d2)

b3) c3) d3)

Fig. 9. Model 3 — Stratigraphy. a) Original resistivity model b) Inversion resistivity model resulted by the original (b1) and the new (b2) algorithm for the Wennner–Schlumberger data. c) Inversion resistivity model resulted by the original (c1) and
the new (c2) algorithm for the Pole–Dipole data. d) Inversion resistivity model resulted by the original (c1) and the new (c2) algorithm for the Dipole–Dipole data. The relative differences between the inversion models resulted by the original and the
new algorithm for the Wenner–Schlumberger, Pole–Dipole and Dipole–Dipole are shown in figures b3, c3 and d3 respectively.
P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224 223

Fig. 10. (a) Location of the survey area in Germany, (b) photo showing the field measuring arrangement and the SE–NW direction of the long ERT line.

will minimize the possibility to remove significant information from the It has been stressed that the definition of the threshold area to calcu-
Jacobian matrix during the later iterations of the non-linear inversion late the significant part of the Jacobian matrix is actually based on the
procedure. distance of the outermost electrodes used to capture a specific resistiv-
The resulting sparseness of the Jacobian and smoothness matrices ity measurement. It has been demonstrated that the algorithm can be
was appropriately taken into account saving substantial memory re- successfully implemented without any further modification to account
sources that can reach up to 96% based on the scale of the inversion for surface ERT data collected with any kind of electrode configurations
problem. The technique for calculating only the significant Jacobian other than Dipole–Dipole (e.g. Wenner–Schlumberger or Pole–Dipole
values as well as the use of the LSQR routine to iteratively invert the aug- arrays). Further tests with ERT lines that cannot be considered as
mented linear system and update the resistivity within the non-linear being long (viz. a large number of electrodes) but they are rather typical
iterations result in a fast and efficient algorithm able to cope with in the number of electrodes used (i.e. 50 electrodes in the example of
huge amount of data collected from long ERT lines. The use of different Figs. 7, 8) illustrate that the “fast” technique can be used to efficiently
spatial smoothness or other spatial correlation constrains (e.g. Blaschek invert “typical” ERT lines. This can be useful if there is a need to further
et al., 2008; Gallardo and Meju, 2003, 2004; Hermans et al., 2012; Loke reduce computational resources (speed, memory) such as the case for
et al., 2003) would not affect the algorithm efficiency, since almost all time-lapse monitoring.
such constrains involve sparse linear/linearized equations that are effi- The proposed algorithm exhibits a similar convergence pattern as
ciently handled by LSQR. The new algorithm is almost more than one the original one and its application to synthetic and real data sets result-
order of magnitude (~30 times) faster than the original one. ed in reconstructed models of comparable accuracy with the standard

Fig. 11. 2D inversion models of the LINE 1 from the data collected in Germany using: (a) the standard and (b) the new algorithm. (c) Relative difference between the inversion models
resulted by the two algorithms. The mean relative difference between two inversion models is 1.9%.
224 P. Tsourlos et al. / Journal of Applied Geophysics 105 (2014) 213–224

approach (mean relative difference between the resistivity inversion Günther, T., 2007. Roll-along-inversion — a new approach for very long DC resistivity
profiles, A14. Extended abstract book of EAGE Near Surface 2007. 13th European
models is on average less than 2%). Finally the rather simplified basic Meeting of Environmental and Engineering Geophysics, Istanbul, Turkey.
concept of this algorithm can be easily adopted and transferred Günther, T., Rücker, C., Spitzer, K., 2006. Three-dimensional modelling and inversion of
for the inversion of different kinds of tomographic geophysical data DC resistivity data incorporating topography-II. Inversion. Geophys. J. Int. 166,
506–517.
(e.g. induced polarization, controlled source electromagnetic, etc.). Hermans, T., Vandenbohede, A., Lebbe, L., Martin, R., Kemna, A., Beaujean, J., Nguyen, F.,
2012. Imaging artificial salt water infiltration using electrical resistivity tomography
constrained by geostatistical data. J. Hydrol. 438 (439), 168–180.
Acknowledgments Loke, M.H., Barker, R.D., 1996. Rapid least-squares inversion of apparent resistivity
pseudo-sections using Quasi-Newton method. Geophys. Prospect. 48, 131–152.
Loke, M.H., Acworth, I., Dahlin, T., 2003. A comparison of smooth and blocky inversion
This work was supported by the Basic Research Project of Korea methods in 2D electrical imaging surveys. Explor. Geophys. 34, 182–187.
Institute of Geoscience and Mineral Resources funded by the Ministry of Loke, M.H., Chambers, J.E., Rucker, D.F., Kuras, O., Wilkinson, P.B., 2013. Recent develop-
ments in the direct-current geoelectrical imaging method. J. Appl. Geophys. http://
Science, ICT & Future Planning, Republic of Korea. The authors would dx.doi.org/10.1016/j.jappgeo.2013.02.017.
like to thank two anonymous reviewers for their constructive criticism McGillivray, P., Oldenburg, D., 1990. Methods for calculating Frechet derivatives and
and comments that helped to significantly improve the manuscript. sensitivities for the non-linear inverse problem: a comparative study. Geophys. Prospect.
38, 499–524.
Paige, C., Saunders, M., 1982. LSQR: an algorithm for sparse linear equations and sparse
least squares. ACM Trans. Math. Softw. 8, 43–71.
References Papadopoulos, N.G., Tsourlos, P., Papazachos, C., Tsokas, G.N., Sarris, A., Kim, J.H., 2011. An
algorithm for the fast 3-D resistivity inversion of surface electrical resistivity Data:
Auken, E., Christiansen, A.V., Jacobsen, B.H., Foged, N., Sørensen, K., 2005. Piecewise 1D application on imaging buried antiquities. Geophys. Prospect. 59, 557–575.
laterally constrained inversion of resistivity data. Geophys. Prospect. 53 (4), 497–506. Rijo, L., 1977. Modelling of Electric and Electromagnetic Data. Ph.D. Thesis University of
Blaschek, R., Hördt, A., Kemna, A., 2008. A new sensitivity-controlled focusing regularization Utah.
scheme for the inversion of induced polarization data based on the minimum gradient Sasaki, Y., 1992. Resolution of resistivity tomography inferred from numerical simulation.
support. Geophysics 73 (2), F45–F54. Geophys. Prospect. 40, 453–463.
Coggon, J.H., 1971. Electromagnetic and electrical modelling by the finite element method. Sotiropoulos, P., Tsourlos, P., Buckup, K., Sideris, G., 2002. Geoelectrical survey at the water
Geophysics 36, 132–155. barrier of Elva river at the area of Havelberh (N. Germany). Proceedings of the 8th
Gallardo, L.A., Meju, M.A., 2003. Characterization of heterogeneous near-surface materials by EEGS-ES Meeting, Aveiro, Portugal, pp. 379–482.
joint 2D inversion of dc resistivity and seismic data. Geophys. Res. Lett. 30 (13), 1658. Tikhonov, A.N., Leonov, A.S., Yagola, A.G., 1998. Non-linear Ill-posed Problems. Chapman
Gallardo, L.A., Meju, M.A., 2004. Joint two-dimensional DC resistivity and seismic and Hall, London, UK.
traveltime inversion with cross-gradients constraints. J. Geophys. Res. 109, B03311. Tsourlos, P., 1995. Modeling Interpretation and Inversion of Multielectrode Resistivity
Golub, G.h., Kahan, W., 1965. Calculating the singular values and pseudo-inverse of a Survey Data. Ph.D. Thesis University of York, U.K.
matrix. J. Soc. Ind. Appl. Math. Ser. B Numer. Anal. 2, 205–224.

You might also like