You are on page 1of 12

Computers & Geosciences 42 (2012) 177–188

Contents lists available at SciVerse ScienceDirect

Computers & Geosciences


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

VFSARES—a very fast simulated annealing FORTRAN program for


interpretation of 1-D DC resistivity sounding data from various
electrode arrays
Shashi Prakash Sharma n
Department of Geology and Geophysics, Indian Institute of Technology, Kharagpur, W.B. 721302, India

a r t i c l e i n f o abstract

Article history: Employing the very fast simulated annealing (VFSA) global optimization technique, a FORTRAN
Received 30 June 2010 program is developed for the interpretation of one-dimensional direct current resistivity sounding
Received in revised form data from various electrode arrays. The VFSA optimization depicts various good fitting solutions
28 June 2011
(models) after analyzing a large number of models within a predefined model space. Various models
Accepted 24 August 2011
Available online 16 September 2011
that yield reasonably well fitting responses with the observed response lie along a narrow elongated
region of the model space. Therefore, instead of selecting the global model on the basis of the lowest
Keywords: misfit error, it is better to analyze histograms and probability density functions (PDFs) of such models
Global optimization for depicting the global model. In a multidimensional model space, the most appropriate region to
VFSA
select suitable models to compute the mean model is the one in which the PDF is larger in comparison
Resistivity sounding
to the other regions of the model space. Initially, accepted models with misfit errors less than the
Various electrode arrays
predefined threshold value are selected and lognormal PDFs for each model parameter are computed.
Subsequently, mean model and uncertainties are computed using the models in which each model
parameter has a PDF more than the defined threshold value ( 468.2%). The mean model computed from
such models is very close to the actual subsurface structure (global model). It is observed that the mean
model computed using models with a PDF more than 95% for each model parameters yields the actual
model. Moreover uncertainty computed using models with such a high PDF and lying in a small model
space will be small and it will not be considered as the actual global uncertainty. Resistivity sounding
(synthetic and field) data over different subsurface structures are optimized using the VFSA program
developed in the present study. Optimization results reveal that the actual model always locates within
the estimated uncertainty in the mean model. Since the approach requires much less computing time (a
few minutes) using an ordinary PC, results with smaller uncertainty can be obtained using repeated
computations with a smaller search range in comparison to the results obtained in a large search range.
The efficacy of the program is demonstrated by interpreting data from various layered earth structures.
Field examples associated with groundwater and mineral exploration are also presented. Interpreted
model parameters show excellent correlation with drilling results. The optimization program can be
used for various case studies like those associated with groundwater, mineral exploration, subsurface
pollution studies, and saline water incursion in coastal areas.
& 2011 Elsevier Ltd. All rights reserved.

1. Introduction sounding is still significant and widely used for different practical
applications such as groundwater, mineral exploration, and sub-
One-dimensional (1-D) inversion of direct current (DC) resis- surface pollution studies (Van Overmeeren, 1989; Barker, 1990;
tivity data sets is a simple and rapid tool for mapping vertical Benson et al., 1997; Ebraheem et al., 1997; Sharma and Baranwal,
variations in the electrical conductivity of the earth’s crust 2005; Baranwal and Sharma, 2006). Interpretation of geophysical
(Inman, 1975). In addition, 1-D interpretation is very useful in data is always a challenging task and is performed using auto-
constructing initial models for multidimensional interpretations matic inversion schemes. Various linearized and global inversion
(Loke and Barker, 1996). Layered earth interpretation in resistivity approaches have been developed for the interpretation of resis-
tivity sounding data (Zohdy, 1989; Bas- okur, 1990; Sen et al.,
1993; Vedanti et al., 2005; Fernández Martı́nez et al., 2010).
n
Tel.: þ91 3222 283386; fax: þ 91 3222 282268. Linearized inversion techniques yield a single solution and
E-mail address: spsharma@gg.iitkgp.ernet.in have limitations such as selection of the proper initial guess

0098-3004/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2011.08.029
178 S.P. Sharma / Computers & Geosciences 42 (2012) 177–188

model, unstable solutions, and non-uniqueness of the solution resistivity transform given by the recurrence relation
(Inman, 1975). Global optimization using genetic algorithms and Ti þ 1 þ ri tanhðlhi Þ
artificial neural networks also yields a single solution and pre- Ti ¼ ð5Þ
1 þ ðTi þ 1 tanhðlhi Þ=ri Þ
defined resolution (Singh et al., 2005). VFSA optimization, which
finds solutions in a predefined model space, does not require where i¼N 1,y,1, TN ¼ rN, ri and hi are the resistivity and
partial derivative computation and yields a number of solutions thickness of the ith layer, respectively. Using x ¼ln(s) and
to overcome the limitation of linearized inversions (Sen and y¼ln(1/l), Eqs. (1)–(4) become
Stoffa, 1995). Sharma and Kaikkonen (1999) have analyzed VFSA Z 1
studies for equivalence and suppression using joint inversion of raT ðxÞ ¼ TðyÞ½eðxyÞ J0 ðexy Þdy ð6Þ
1
resistivity and electromagnetic data. Bhattacharya et al. (2003)
Z 1
have used VFSA for analysis of resolution, sensitivity, and uncer-
raW ðxÞ ¼ TðyÞ½2eðxyÞ fJ0 ðexy ÞJ0 ð2exy Þgdy ð7Þ
tainty in resistivity and induced polarization data. It is essential to 1
study various solutions simultaneously in order to determine Z 1
geologically relevant models. In a set of solutions, the most raS ðxÞ ¼ TðyÞ½e2ðxyÞ J1 ðexy Þdy ð8Þ
appropriate direction for computing the mean model is the one 1
where maximum numbers of solution converge. The mean model Z 1 Z 1 
1
computed from such models will be closest to the actual subsur- raDr ðxÞ ¼ TðyÞfe2ðxyÞ J1 ðexy Þgdy TðyÞfe3ðxyÞ J0 ðexy Þgdy
face structure. 2 1 1

Although the Schlumberger electrode configuration is the most ð9Þ


commonly used array for resistivity sounding due to its simplicity Note that raDr represents dipole radial array in Eq. (9). The
in field operations, there are various other electrode configura- resistivity transform, T(y), is the input of the second term in Eqs.
tions having different depths of penetration and resolutions. (6)–(9), which is the filter function. Various researchers have
Verma and Sharma (1993) have presented linearized inversion developed different sets of filter coefficient to compute apparent
studies associated with different electrode configurations and resistivity for different electrode configurations (Ghosh, 1971b;
shown that dipole array has better resolution. Therefore, in the Das et al., 1974; Das and Ghosh, 1973; Johansen, 1977; Koefoed,
present study a common VFSA program is developed for the 1979). The expression for dipole perpendicular can be obtained by
interpretation of resistivity sounding data from two-electrode, putting the value of b as 1/3 in Eq. (4) and will look similar to Eq.
Wenner, Schlumberger, dipole perpendicular, and dipole radial (9). These convolution integrals, Eqs. (6)–(9), can be written in the
arrays. A number of good fitting models are optimized and a summation form and used for calculation of theoretical resistivity
statistical mean model is computed. Further, uncertainty in the sounding data for different electrode configurations.
mean model is minimized and the best fitting model with the
least uncertainty is investigated through independent resistivity
observations. 3. VFSA optimization method

3.1. Basic concept


2. Direct current resistivity sounding
Global optimization methods such as simulated annealing
(SA), genetic algorithms (GAs), and artificial neural networks have
Apparent resistivity expressions over multilayered earth for
been applied to multiparametric optimization of various geophy-
different electrode configurations are described well by Koefoed
sical data sets (Rothman, 1985, 1986; Sen et al., 1993; Sen and
(1979) and Das and Verma (1981). A digital linear filtering
Stoffa, 1995; Sharma and Kaikkonen, 1998, 1999; Vedanti et al.,
technique (Ghosh, 1971a, 1971b) is used for the computation of
2005; Fernández Alvarez et al., 2008; Fernández Martı́nez et al.,
forward responses for different arrays. For completeness, relevant
2010). Conventional global optimization techniques (simulated
expressions for apparent resistivity for different arrays (two-
annealing using a heat-bath algorithm or a genetic algorithm)
electrode, Wenner, Schlumberger, Dipole arrays, respectively)
compute the misfit errors for a large number of models in the
are provided below:
model space. Subsequently they compute the probability of each
Z 1
model and try to concentrate in the region of high probability. In
raT ðsÞ ¼ s TðlÞJ0 ðlsÞdl ð1Þ
0 the present study, very fast simulated annealing (VFSA) is used,
which does not compute misfit errors for a large number of
Z 1 models at a time but it moves in the model space randomly. It
raW ðsÞ ¼ 2s TðlÞ½J0 ðlsÞJ0 ð2lsÞdl ð2Þ
selects a new model, computes misfit errors and probability for
0
this model, and then selects or rejects this model with respect to
Z 1 the previous model. Movement in the model space follows
raS ðsÞ ¼ s2 TðlÞlJ1 ðlsÞdl ð3Þ Cauchy probability distribution, which has a sharper peak than
0
Gaussian distribution. This allows the temperature to be lowered
 Z 1 Z 1  at a faster rate in VFSA than conventional SA approaches. Hence
2
raD ðsÞ ¼ s2 ð1bÞ TðlÞlJ1 ðlsÞdlbs TðlÞl J0 ðlsÞdl ð4Þ VFSA reaches the final temperature rapidly. This makes it more
0 0
efficient and it takes a lesser amount of computing time in
Here s is the electrode separation and varies for different comparison to other global algorithms (Sen and Stoffa, 1995,
arrays. s is the distance between two acting electrodes for two- pp. 267).
electrode array, one-third of current electrode separation for One of the salient features of VFSA is that it does not need to
Wenner array, half of the current electrode separation for store a vast number of models like other global converging
Schlumberger array, and distance between the centers of two approaches. It selects better models while moving in the multi-
dipoles for dipole arrays. b is a constant and it is 1/2 for dipole dimensional model space at different temperature levels and
radial and 1/3 for dipole perpendicular array. J0 and J1 are the finally yields a model with the lowest misfit error. However in a
Bessel functions of zero and first order, respectively. T(l) is the complex situation of equivalence in DC resistivity, if the VFSA
S.P. Sharma / Computers & Geosciences 42 (2012) 177–188 179

process is repeated several times then equivalent solutions are value depends on the magnitude of the objective function
obtained with similar misfit errors. Under such circumstances it is considered for the optimization; and M is the number of model
better to perform statistical analysis of various models showing a parameters. In the present study ci ¼1 is used and 1/M is replaced
misfit error lower than a predefined threshold value. Therefore in by the parameter CS. After several test runs in this study, T0i is
the present VFSA approach, all the accepted models are also found suitable at 0.01 and CS as 0.4 for resistivity data inter-
stored in the memory for the posterior analysis. A short descrip- pretation. There is no need to change these parameters in the
tion of the objective function and governing equations of the VFSA program as they are tested and they work efficiently.
process is given below. Once again the specified number of moves with the selection
Initially a model Pi (ri and hi, resistivity and thickness for criterion noted above is made at a lower temperature level.
various layers, respectively) is selected randomly in the model Subsequently, the temperature is lowered gradually using Eq.
space Pmini rPi rPmax
i . It is important to note that each model (13) to a sufficiently low value selecting a better and better model
min max
space Pi to Pi is transformed into log domain and then a at each temperature level. After completing the predefined itera-
sample Pi (which is actually log(Pi)) is drawn according to Cauchy tions (say 1000), one solution is obtained. The whole procedure is
probability distribution. Subsequently, log(Pi) is transformed back repeated several times to obtain various solutions and each time
to the actual domain to compute the model response. Since log– the process starts from a randomly selected model in the
log scale is utilized for the presentation of apparent resistivity predefined model space. The model parameters obtained in
data, discrepancy between the observed and the model data different runs could be the same for a well posed simple problem.
(visualized on the log–log plot) can be mathematically repre- However, they are different according to the physics of the
sented by log transformed observed and computed apparent problem for complex problems. For example, in the DC resistivity
resistivity. Therefore, the following objective function e is calcu- case, model parameters of intermediate layers follow the princi-
lated: ple of equivalence.
It is observed that initially at higher temperature levels,
1XN
accepted model parameters vary over a large range. This is due
e¼ ½lnðr0j Þlnfrcj ðPi Þg2 ð10Þ
Nj¼1 to the fact that in a highly complex multidimensional error
(misfit) surface, the PDF is also as complex as the misfit surface
where r0j and rcj ðPi Þ are jth observed and model responses,
at higher temperature levels. However, as the temperature is
respectively, N is the number of observation points. Model
lowered, the PDF becomes smoother and shows sharper peaks
parameters and the objective function of the above model are
(Sharma and Kaikkonen, 1998). Hence, the accepted model para-
kept in memory and each parameter is updated according to the
meter is localized near the well-defined peak and shows a
Cauchy probability distribution. The updating factor yi for the ith
tendency to cluster around the minimum. At sufficiently low
parameter is computed from the following equation:
" # temperatures, the algorithm accepts models in the vicinity of the

1 92ui 19 deep seated minimum where the PDF has a sharp peak and other
yi ¼ sgnðui 0:5ÞTi 1 þ 1 ð11Þ local minima are absolutely invisible. When the VFSA process is
Ti
repeated several times then it yields different good fitting models
such that it varies between 1 and þ1. In the above equation ui (different deep-seated minimum in the objective function) that
is a random number varying between 0 and 1, and Ti is the are spread over the multidimensional model space according to
temperature, which may be the same or different for various the physics of the problem.
parameters depending on the nature of the problem. Each para-
meter Pi is updated to Pmi
þ1
from its previous value Pm
i by the 3.2. Global model and uncertainty using all accepted models
following equation:
A single run of global converging algorithms is not sufficient to
Pim þ 1 ¼ Pim þ yi ðPi max Pi min Þ ð12Þ
find the global solution (Sen and Stoffa, 1995). Therefore, a
and thus a new model is obtained. Subsequently, the objective number of good fitting models are optimized (10 in the present
function for the new model is calculated and compared with the study). Model parameters (ri and hi) of these good fitting models
previous model. If the misfit error of the new model is less than may differ from each other and lie in a wide range in the
the misfit error of the previous model, then the new model is multidimensional model space. It is essential to sample the
selected with the probability exp(  De/T), where De is the models from the most appropriate region (where a large number
difference in the objective functions of the two models. When of models are located) of the model space. Different sampling
the misfit error of the new model is greater than that of the techniques have been used by different scientists (Mosegaard and
previous model, then a random number between 0 and 1 is drawn Tarantola, 1995; Sen and Stoffa, 1996) to obtain the global model
and compared with the probability. If the probability is greater and minimize uncertainty in the solution. Sampling in the model
than the random number then also the new model is selected space is based on different statistical distributions and could
with the same probability otherwise the new model is rejected differ from one geophysical method to other. In DC resistivity,
and an earlier model and its objective function are kept in the individual models lie along the equivalent regions of the model
memory. Next, specified numbers of moves are made at the same space and equivalent regions are different for resistive and
temperature level by accepting and rejecting the model according conducting layers. However, transformation of variables in loga-
to the above-noted criterion and this makes one iteration. rithmic domain yields equivalent regions in a common geome-
Repeated movement in the model space at one temperature level trical shape (a straight line) such that an individual model lies
yields a better and better model. After completing the specified along this line for resistive as well as conducting layers. Therefore,
number of moves at a particular temperature, the temperature is in the present study, sampling in the model space is performed
lowered according to the following cooling schedule: using lognormal probability distributions.
To obtain a best fitting model, computations are performed at
Ti ðkÞ ¼ T0i expðci k1=M Þ: ð13Þ
1000 different temperature levels with 20 moves at one tem-
Here k is the number of iteration; ci is a constant, which may perature level. The VFSA procedure is repeated 10 times and the
vary for different model parameters; T0i is the initial temperature, 10 best fitting solutions are obtained. Thus 200,000 models and
which may also be different for various parameters, its initial their misfit errors are stored in memory where misfit error varies
180 S.P. Sharma / Computers & Geosciences 42 (2012) 177–188

from a large value to a very small value. Out of these models, The mean model Pi is computed from the new best models
repeated models as well as models whose misfit error is higher where each model parameter has a PDF larger than the defined
than the defined threshold value are discarded. Therefore, models threshold value (say 68.2%) using the expression
that fit the observed response up to certain degree are selected for ( )
1 XNM
further analysis. Pi ¼ exp lnðPi,k Þ ð15Þ
First, the PDF for each model parameter for all selected models NM k ¼ 1
is computed. The probability density function fx ðx, m, sÞ of a In the above equation NM is the number of models satisfying
lognormal distribution of a variable x (ri and hi ) is given by the above-noted criterion of a higher PDF. Subsequently, the
covariance and correlation matrices are computed using the
1 2 2
fx ðx, m, sÞ ¼ pffiffiffiffiffiffi eððln xmÞ =2s Þ ð14Þ equations (Tarantola, 1987)
xs 2p
1 XNM
Cov Pði,jÞ ¼ ðP P ÞðP P Þ ð16Þ
In a lognormal distribution, parameters denoted as m and s are NM k ¼ 1 i,k i j,k j
the mean and standard deviation, respectively, of the variable’s
natural logarithm. and
Several approaches are analyzed to reach the global solution Cov Pði,jÞ
Cor Pði,jÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð17Þ
using lognormal probability distributions. First, a model is tried to Cov Pði,iÞ  Cov Pðj,jÞ
be designed by picking up the model parameter on the basis of
the highest PDF for individual model parameters. It is observed In Eqs. (16) and (17) i and j vary from 1 to M. The correlation
that such a model differs from the global model. Next, the global matrix formed with these solutions fitting well to the observed
model is also tried to be selected on the basis of sum of the response conveys the relationship of the parameters and asso-
normalized PDF (sum of the probability of the individual model ciated physics. The square roots of the diagonal elements of the
parameters normalized by the respective highest probability) for covariance matrix represent the uncertainties in the mean model
various models. The magnitude of the PDF for different variables parameters and it is used to estimate the uncertainties in the
(ri and hi) is at different levels; therefore, to make them at the mean model.
same level they are normalized by the respective maximum value
such that the PDF varies between 0 and 1 for each variable. The 3.3. Mean model and uncertainty using best fitted models
sums of normalized PDFs (all parameters) for all models are
computed and compared. Once again, it is emphasized that a As discussed in the previous section that 10 best fitting
model having the highest sum of normalized probability is the solutions (models) are derived after 20,000 forward computations
best model among all models with the lowest misfit error. for each run. Mean model and associated uncertainty are also
However, this is also not the global model. Such model can also computed using Eqs. (15) and (16) from these 10 best fitting
be picked up by simply looking for the lowest objective function solutions that lie along the equivalent region of the model space.
for all solutions. Further, a model can also be picked up by fitting These 10 best fitting models lie over a wider model space
a smooth curve through the PDF for various model parameters (ri compared to the models selected on the basis of PDFs ( 468.2%
and hi) using all selected models or a few selected variables for each parameter). Therefore, uncertainty computed using
whose probability is larger than a given threshold value. Fitting selected models on the basis of PDFs will be smaller compared
using lognormal distribution will produce a mean and standard to these 10 best fitting models derived from 10 runs of the VFSA
deviation for each model parameter (ri and hi) that can also be process.
defined as a global model and associated uncertainty. However, It is emphasized that the mean models in both cases are
this approach is also similar to an approach such as picking up similar but uncertainties will be different. Since the actual model
individual model parameters corresponding to the highest PDF. is located within the estimated uncertainty, smaller uncertainty
In the present study, global model and associated uncertainty will be better for accurate results, which will be obtained from
are obtained using the following approach. After computing the PDF the posterior analysis of the history of all models during the VFSA
for all selected models parameters using Eq. (14), the maximum process.
PDF for each parameter is determined. Subsequently, for selection
of new good models for the computation of the mean model a 3.4. Response correlation
68.2% limit for the PDF is set for each parameter. If any parameter
(ri and hi) of a model has a PDF lower than 68.2% then that model is To quantify the fitting between the mean model response and
located in the undesired region of the model space and discarded. the observed response, the response correlation a is computed
This results in sampling of the most appropriate region of the model PN 0 cn
i ¼ 1 ri ri
space of the high probability region in multidimensional model a ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN P ð18Þ
space. Finally, only those models in which all model parameters ð i ¼ 1 r0i r0i n Þð N c cn
i ¼ 1 ri ri Þ
have a PDF greater than 68.2% are selected for computation of the In Eq. (18) * represents the complex conjugate of the quantity;
mean model and uncertainty. It is observed that such a mean model however, it is the same for a real quantity. The quantity a yields a
is very close to the global model and the global model is always value near unity for a good matching of the responses and a lower
located within the uncertainty estimated in the mean model. The value for a poor matching.
program VFSARES.F is written in such a way that one can set any
desired cutoff limit for PDF in the input file.
It is emphasized that different limits in PDFs such as 60%, 75%, 4. Results
90%, and 95% are used to demonstrate how the selected models
are closer to the actual model parameters. Obviously, higher Initially, synthetic resistivity sounding data are generated for
threshold values resulted in selection of lesser models and small various models that represent different types of geological struc-
uncertainty. A lower threshold value of PDF allows more models tures. Actual model parameters for various synthetic data sets are
to be selected for the computation of the mean model and yields shown in respective tables. To demonstrate the efficacy of the
larger uncertainty in the mean model. developed program for delineation of the actual subsurface
S.P. Sharma / Computers & Geosciences 42 (2012) 177–188 181

structure, initially noise free synthetic data for Schlumberger arrays Table 1
over three-layer models (H- and K-type) are used in the optimiza- Optimization results for H-type model.
tion. Subsequently, random noise added synthetic data for different
Parameters Actual Search Mean model Mean model
electrode configurations over a four-layer model (HK-type) are value range (final 10 sol.) (PDF 468.2%)
optimized. It should be noted that noise added synthetic data may
correspond to a structure that could be different compared to the r1 (O m) 100 50–200 100.07 0.02 100.07 0.09
actual subsurface structure for which synthetic data were generated. r2 (O m) 20 10–80 22.70 7 9.90 18.71 7 3.74
r3 (O m) 500 100–2000 500.8 71.58 499.9 7 3.01
Finally, a number of field data sets are optimized to show the
h1 (m) 50 10–80 48.83 7 1.93 50.047 0.57
applicability of the program for the interpretation of data associated h2 (m) 20 5–50 23.02 7 10.79 18.70 73.85
with groundwater and mineral exploration. Misfit error – – 1.22  10  6 3.22  10  8
After looking at the observed responses (apparent resistivity Correlation 1.00 1.00
for synthetic or field data), the number of layers and suitable
search ranges (lower and upper bound) for various model para-
meters (resistivity and thickness) are determined to perform the
VFSA optimization. Initially, it is better to obtain a single solution.
If search ranges are not selected properly then some of the model
parameters may lie on the lower or upper bound of the search
range. This is erroneous; optimized model parameter should
never lie on the lower or higher bounds of the search range. If
that is the case with any of the model parameters then the lower/
higher bound should be modified accordingly. For example, if the
lower bound of a parameter is set at 30 and after executing the
program that parameter is also found to be 30, then this implies
that the actual value of the parameter is lower than 30. The lower
limit of that parameter must be set less than 30. Similarly, if a
parameter touches the higher limit of the search range after
optimization then the higher limit must be increased. This is a
very important step of VFSA global optimization. An experienced
person is able to set these limits in one step; however, one can
quickly learn to set the appropriate limits for each model para-
meter. When all the optimized parameters lie within the defined
search range after executing the program, then fitting between
the observed and the model data will be accurate. After testing
this, one can perform the desired number of runs (say 10) to have
a large number of samples to find the appropriate region of the
model space (high PDF region in the multidimensional model
space) to compute, (i) mean model which necessarily fits the
observed data, (ii) uncertainty in the mean model parameters
using the covariance matrix, (iii) correlation matrix depicting the
dependence of model parameters on each other, (iv) response
correlation between field and model responses, and (v) misfit
error between the observed and the mean model data.
It is essential to keep the search range narrow for precise results.
On some trial runs one should restrict the search range to obtain a
good result with smaller uncertainty in the mean model para-
meters. For the movement (sampling) in model space, a switch 1 or
0 can be set for logarithmic and linear domain sampling for each
model parameter. Generally, resistivity varies over a large range
and requires logarithmic variation and thicknesses vary over a
smaller range and need linear variation. However, the nature of the
DC resistivity problem is such that it is better to optimize both in
the logarithmic domain. In the present study, samples are drawn in
the logarithmic domain according to Cauchy probability distribu-
tions for both resistivity and thicknesses of various layers.
The convergence pattern of various model parameters (for Fig. 1. Convergence pattern for various model parameters and misfit errors for a
VFSA solution.
H-type model given in Table 1) after VFSA optimization for a
single solution is shown in Fig. 1. Since the number of moves at
one temperature level (one iteration) is 20 and 1000 iterations are value. Fig. 1 depicts the pattern of convergence for one best fitting
performed, 20,000 forward computations in total are performed solution. Performing several runs (say 10) will yield 10 similar
to obtain a solution. It is observed that various model parameters figures. Moreover, the final model in all 10 runs will be different
vary over a large range initially (at higher temperature levels) and (specially r2 and h2 or intermediate layer parameters). Further,
subsequently start stabilizing around a solution. During this these 10 runs yield enough models (200,000 in the present study)
process misfit error reduces gradually from 0.01 to a very low for statistical analysis. A histogram and lognormal probability
value of 10  8. It can be seen from Fig. 1 that all parameters are distribution will be analyzed for each model parameter. Finally,
stabilized after 5000 computations but misfit error systematically all accepted moves (models), accepted models with misfit errors
reduces by refining the model parameters to a very accurate below a defined threshold, selected models after applying 60%,
182 S.P. Sharma / Computers & Geosciences 42 (2012) 177–188

75%, 90%, and 95% PDF cutoff values for each parameter are expected, PDFs for r1, r3, and h1 are rather sharp while those
presented for two primary models (H- and K-type) to show the for r2 and h2 are broad. Therefore uncertainties in r2 and h2 will
efficacy of the procedure in depicting the appropriate sampling of be larger in comparison to other model parameters.
the model space for computation of the mean model. Fig. 3 depicts the cross-plots of h2 verses r2 after different
stages of the selection process. Fig. 3a and b depicts all accepted
4.1. Theoretical data models and accepted models whose misfit error is lower than
10  4, respectively. Fig. 3c–f depicts models in which all para-
4.1.1. H-type model meters have a PDF greater than 60%, 75%, 90%, and 95%, respec-
Initially, a model with a conducting intermediate layer is tively. Fig. 3f shows that h2 and r2 are centered on a point but
selected, which represents a groundwater-bearing zone, miner- there are many models located in this small region. The global
alized zone, subsurface pollution, and saline water incursion uncertainty can be estimated from the covariance matrix of the
zones. Actual model parameters and their search ranges are models in a larger model space. Since one standard deviation
shown in Table 1. As discussed above, 10 VFSA runs are corresponds to a 68.2% PDF level, a 68.2% limit is set for each
performed, and model parameters and misfit errors for 200,000 parameter for the selection of models for computation of the
models are stored in the memory for statistical analysis. Next, mean model and uncertainty.
repeated models and the models with misfit errors higher than Table 1 depicts the optimized model parameters and uncer-
10  4 are removed. Remaining models are used to prepare tainties in the mean models computed using two approaches. The
histograms and to compute the PDF. The upper panel of Fig. 2 first mean model is computed on the basis of the 10 best models
presents the histogram of the accepted model parameters whose obtained after each VFSA run. The second mean model is com-
misfit error is less than the threshold misfit error (10  4 in this puted using models from the high probability region ( 468.2% PDF
example). The histogram reveals that r1, r3, and h1 show a for each model parameter). Both mean models are centered close
tendency to cluster around the actual value (100, 500, and 50, to the actual model with different uncertainties. It is interesting to
respectively). However, r2 and h2 do not show such a tendency. note that computation of the mean model is also performed using
The number of accepted model parameters for intermediate layer the models in the highest probability region (on the basis of 95%
varies over a broad range. This is due to equivalence associated limit for each model parameter) and resulting model parameters
with a conducting layer. The lower panel of Fig. 2 depicts the are exactly the same as the actual model parameters with
lognormal probability distribution for each model parameter. A negligible uncertainty. Since, negligible uncertainty is not proper,
peak in the PDF can be seen for each model parameter, which is mean models computed from the models with PDFs greater than
roughly centered close to the actual model parameters. As 68.2% are presented in the various tables.

Fig. 2. Histogram (top panel) and probability density function (bottom panel) of selected models that have misfit errors below a defined threshold value for the
H-type model.

Fig. 3. Cross-plots of thickness and resistivity of the target layer for the H-type model showing (a) all accepted models, (b) selected models having misfit error below a
defined threshold value, (c) models in which all model parameters have a PDF greater than 60%, (d) models in which all model parameters have a PDF greater than 75%,
(e) models in which all model parameters have a PDF greater than 90%, and (f) models in which all model parameters have a PDF greater than 95%.
S.P. Sharma / Computers & Geosciences 42 (2012) 177–188 183

Table 1 reveals that uncertainties in the resistivity and thick-


ness of the second layer are larger than other model parameters.
This is due to the fact that resistivity and thickness of the
intermediate layer in H-type case are positively correlated and
have a wide range. Fig. 4 shows a perfect fit between the synthetic
and the model data. The misfit errors shown in Table 1 are also
very small and response correlation is 1 showing the proper fit
between the observed and model responses. The correlation
matrix presented in Table 2 shows that the resistivity of the first
layer exhibits a weak correlation with all other model parameters.
The resistivity of the second layer shows strong positive correla-
tion with thickness of the second layer. Further, resistivity of the
second layer also shows a negative correlation with the thickness
of the first layer. This is also in accordance with the physics of the
problem. To keep the model data the same when the resistivity of
the intermediate layer increases then the thickness of the first
layer must reduce. Resistivity of the third layer is uncorrelated.
Fig. 4. Fitting between observed and model data for the H-type model.
The thickness of the first layer depicts a negative correlation with
resistivity and thickness of the second layer.

Table 2 4.1.2. K-type model


Correlation matrix for H-type model using 68.2% PDF limit. A model with resistive intermediate layer is selected. This
could also represent a groundwater-bearing zone, mineralized
r1 r2 r3 h1 h2
zone, subsurface pollution, and saline water incursion like the
r1 1.000  0.031 0.015  0.115  0.022 previous H-type model but now lies below a resistive layer. This
r2 1.000 0.098  0.790 0.999 type of subsurface structure is also quite common in the field.
r3 1.000 0.068 0.114 Table 3 presents the actual model parameters and search ranges.
h1 1.000  0.793
VFSA optimization is performed using noise free synthetic data
h2 1.000
like the H-type model and the 10 best fitting models are derived.
Once again repeated models and models with misfit errors greater
than 10  4 are removed. Subsequently, a histogram is prepared
and a lognormal PDF is computed (Fig. 5) using the remaining
Table 3
Optimization results for K-type model. models. Histograms and PDFs reveal that the parameters of
intermediate layers (r2 and h2) are spread over a larger model
Parameters Actual Search Mean model Mean model space compared to other model parameters.
value range (final 10 sol.) (PDF 468.2%) Fig. 6 depicts the cross-plots h2 verses r2 after different stages.
20 10–50 207 0.00 207 0.00
Fig. 6a and b reveals all accepted models and accepted models with
r1 (O m)
r2 (O m) 500 50–1000 5137 265 5167 24 misfit errors less than 10  4, respectively. Fig. 6c–f depict models in
r3 (O m) 50 10–100 49.98 70.08 49.97 70.27 which all parameters have a PDF greater than 60%, 75%, 90%, and
h1 (m) 50 10–80 49.82 70.49 50.027 0.16 95%, respectively. Again, Fig. 6f shows that h2 and r2 are centered
h2 (m) 20 5–50 19.58 712.88 19.37 70.94 on a point but many models are located in this small region.
Misfit error – – 6.73  10  7 1.76  10  9
Table 3 presents the mean models and uncertainties computed
Correlation 1.0000 1.0000
from 10 solutions as well as models accepted on the basis of

Fig. 5. Histogram (top panel) and probability density function (bottom panel) of selected models that have misfit errors below a defined threshold value for the
K-type model.
184 S.P. Sharma / Computers & Geosciences 42 (2012) 177–188

Fig. 6. Cross-plots of thickness and resistivity of the target layer for the K-type model showing (a) all accepted models, (b) selected models having misfit errors below a
defined threshold value, (c) models in which all model parameters have a PDF greater than 60%, (d) models in which all model parameters have a PDF greater than 75%,
(e) models in which all model parameters have a PDF greater than 90%, and (f) models in which all model parameters have a PDF greater than 95%.

Fig. 7. Fitting between observed and model data for the K-type model. Fig. 8. Fitting between noisy synthetic data and model data for various arrays
(model data for different arrays are shown by a common solid line which fits the
noisy synthetic data for respective arrays).

Table 4
Correlation matrix for K-type model using 68.2% PDF limit.
in the apparent resistivity. Since apparent resistivity data are
r1 r2 r3 h1 h2 affected by various kinds of noises, a 10% random noise is added
to the synthetic data to simulate the actual field data. Subse-
r1 1.000 0.013 0.080 0.170  0.051
r2 1.000  0.131 0.318  0.927
quently, VFSA optimization is performed using the noisy synthetic
r3 1.000  0.639  0.204 data from different arrays and optimization results are compared
h1 1.000  0.053 for the same search range. Noise in the data increases the
h2 1.000 magnitude of misfit error at a higher level; hence in this case a
misfit error of 0.01 and a PDF of 68.2% are set for the selection of
appropriate models for the computation of the mean model.
PDFs greater than 68.2% for each model parameter. Table 3 Table 5 presents the mean model computed from the models
reveals that once again uncertainties in the mean model para- with a PDF greater than 68.2% for individual model parameters.
meters of intermediate layers are large compared to other model The efficacy of VFSA optimization using the resistivity sounding
parameters. The mean model data fit accurately with the data from different arrays is presented in Table 5. Once again, it is
observed data (Fig. 7). necessary to note that when noise is added to the synthetic data,
The correlation matrix (Table 4) computed from the models the actual model changes in comparison to the true model for
with PDFs greater than 68.2% shows that resistivity of the first which the synthetic data were generated. Therefore, the quality of
layer exhibits a weak correlation with all other model parameters the inversion results will be determined by the estimated uncer-
similar to the H-type case. Resistivity of the second layer shows tainties in the mean model of the same search range. Table 5
negative correlation with the thickness of the second layer. This is indicates that the dipole perpendicular array yields the best
again in accordance with the principle of equivalence for a K-type results. Fittings between the noisy synthetic data from different
model. Resistivity of the third layer does not show strong arrays and the corresponding model data are shown in Fig. 8. In
correlation with other model parameters. Fig. 8, a common solid line depicts the model data for different
arrays that fit the respective noisy synthetic data shown by
4.2. Noisy data—HK type model different solid symbols.

Next, synthetic resistivity sounding data are generated over a 4.3. Field data
typical four-layer model (HK-type) for different electrode config-
urations. The presence of all the layers can hardly be detected 4.3.1. Groundwater investigation
(Fig. 8) in the apparent resistivity data, specially for the two- Resistivity sounding data (VES-1) presented in Fig. 9 deal with
electrode array. Two-electrode array data show the least detect- groundwater investigation. The data are measured in a lateritic
ability while dipole radial array data depict the largest variation terrain of the Ganga basin, Eastern India. Despite a high amount of
S.P. Sharma / Computers & Geosciences 42 (2012) 177–188 185

Table 5
Optimization results using HK-type noisy data from various electrode configurations using models with PDF greater than 68.2% for each model parameter.

Parameters Actual value Search ranges Two-electrode Wenner Schlumberger Dip. perpendicular Dip. radial

r1 (O m) 150 100–200 149.6 71.8 149.8 70.9 150.07 0.3 150.07 0.0 149.6 7 0.2
r2 (O m) 35 1–100 38.4 78.1 32.4 77.3 33.5 74.6 36.27 3.8 34.5 7 4.7
r3 (O m) 450 100–1000 518 7105 486 788 433 734 4617 29 479 7 43
r4 (O m) 50 20–100 50.0 70.5 50.0 70.4 50.0 70.1 50.07 0.0 50.07 0.1
h1 (m) 10 1–20 10.2 70.4 10.1 70.2 10.0 70.2 9.97 0.3 10.07 0.4
h2 (m) 15 5–30 16.9 73.2 14.2 71.6 14.7 71.2 14.97 1.3 16.3 7 1.9
h3 (m) 25 5–30 21.8 74.6 22.4 73.6 27.3 72.8 26.17 2.2 24.2 7 3.4
Misfit 8.42  10  4 1.35  10  4 4.73  10  4 5.48  10  4 7.43  10  4

Table 7
Correlation matrix for VES-1 using 68.2% PDF limit.

r1 r2 r3 r4 h1 h2 h3

r1 1.000 0.092 0.037  0.162 0.806  0.099 0.001


r2 1.000  0.035  0.020 0.270  0.911  0.018
r3 1.000 0.034 0.039 0.031 0.969
r4 1.000  0.152 0.013 0.082
h1 1.000  0.184 0.027
h2 1.000 0.056
h3 1.000

Fig. 9. Fitting between field and model data for VES-1.

Table 6
Optimization results for VES-1 depicting groundwater investigation.

Parameters Search range Inversion result


(PDF 468.2%)

r1 (O m) 30–60 36.28 7 0.34


r2 (O m) 100–500 401 715.2
r3 (O m) 5–50 11.1 7 0.49
r4 (O m) 1000–5000 4590 7 87
h1 (m) 1–10 5.34 7 0.10
h2 (m) 5–30 9.42 7 0.34 Fig. 10. Fitting between field and model data for VES-2.
h3 (m) 5–30 12.12 7 1.58
Misfit error – 3.50  10  4
Correlation 0.9990
results shown in Table 6. The laterite layer was encountered in
the borehole at 5.24 m depth and continued upto13.25 m depth. A
annual rainfall in the region there is a scarcity of groundwater. sandy clay layer followed after this and continued up to 26.75 m
Finding a suitable aquifer in the hard lateritic layer is challenging. depth. Massive laterite formation occurs again beyond 26.75 m. A
The measured resistivity sounding data show a four-layer KH- very good fitting between observed and model data is presented
type structure (Fig. 9). Thick dry clay, which forms the surface soil in Fig. 9. The correlation matrix for this field data is shown in
is followed by a hard laterite layer. Below the laterite layer there Table 7. It correlates well with the correlation matrix obtained for
is a water-saturated sandy clay layer. The final layer is again a dry theoretical H- and K-type data optimization.
laterite formation. The second layer does not allow percolation of
rainwater for recharge and rainwater simply runs off from this 4.3.2. Mineral exploration
region. An artificial recharge by perforating the lateritic layer Resistivity sounding data presented for mineral exploration
needs precise information about the thickness of the lateritic are taken from resistivity surveys carried out for the delineation
layer as well as the rechargeable layer for the success of the of chromite-bearing formation. An integrated geological and
scheme. Interpretation of data from this region is very crucial geophysical survey was performed to assess the suitability of
because there is always a large uncertainty associated with the area for chromite mineralization. Some parts of the area were
sandwiched resistive and conducting layers for a KH-type struc- found geologically suitable and therefore, detailed gravity, mag-
ture. The developed VFSA program serves the purpose effectively netic, VLF electromagnetic and electrical resistivity surveys were
under such conditions. Table 6 presents the interpreted model carried out (Mohanty et al., 2010) in this region. Two zones of
parameters and associated uncertainty. Exploratory drilling con- interest were chosen for resistivity survey to constrain the model
ducted at the survey location matched well with the interpreted obtained from gravity data interpretation. The first area depicted
186 S.P. Sharma / Computers & Geosciences 42 (2012) 177–188

a positive gravity anomaly of a high magnitude while the second


area revealed a positive gravity anomaly of a smaller magnitude.
A resistivity sounding (VES-2) was performed exactly at the
center of the high gravity anomaly zone in the first area. Fig. 10
depicts the measured apparent resistivity data. Interpreted results
shown in Table 8 reveal top layer resistivity 17.35 O m and
thickness 1.71 m. This layer is water-saturated sandy clay. The
second layer with resistivity 62.25 O m and thickness 33.42 m is
chromite-bearing mafic/ultramafic rock. The third layer is sheared
granite, which forms the basement. Exploratory drilling has been
carried out and basement rock encountered in the borehole as
interpretation given in the Table 8. The second layer is confirmed
as a chromite-bearing layer from the drilling.
Resistivity soundings were also performed in the second area
that depicted a relatively small but a regular positive gravity
anomaly zone. This means that a possible chromite body could be
located at a deeper depth or it may have smaller thickness
Fig. 12. Fitting between field and model data for VES-4.
compared to the main chromite body in the first area. First, a
resistivity sounding using Schlumberger arrays was performed
with the current electrode spread in N–S direction (VES-3). This Table 10
Optimization results for VES-4 depicting mineral exploration.
sounding showed only a two-layer structure (Fig. 11). Table 9
Parameters Search range Inversion result
(PDF 468.2%)

Table 8
r1 (O m) 20–100 34.82 7 1.56
Optimization results for VES-2 depicting mineral exploration.
r2 (O m) 100–5000 824 765
r3 (O m) 10–100 38.16 7 3.72
Parameters Search range Inversion result
r4 (O m) 1000–20000 8576 768
(PDF 468.2%)
h1 (m) 2–10 8.41 70.15
r1 (O m) 1–30 17.35 7 0.15 h2 (m) 5–30 22.85 7 2.14
20–200 65.25 7 4.62 h3 (m) 5–30 17.45 7 3.14
r2 (O m)
Misfit error – 5.68  10  3
r3 (O m) 50–1000 559 7 13.80
Correlation 0.9994
h1 (m) 1–10 1.71 7 0.04
h2 (m) 5–50 33.42 7 2.38
Misfit error – 4.28  10  4
Correlation 0.9995
depicts the interpreted model parameters. The top layer with
resistivity 33.54 O m and thickness 8.35 m is a thick dry surface
soil. This is followed by a resistive layer with resistivity 568 O m,
which could be massive laterite. The presence of a conducting
layer (possible chromite) cannot be seen in the sounding data.
Since positive gravity anomaly has been observed in this area, a
two-layer resistivity structure could not explain the reason for the
positive gravity anomaly. A Schlumberger sounding was per-
formed at the same location with spread in an E–W direction
(VES-4). Surprisingly, sounding in an E–W direction at the same
location depicted a four-layer structure (Fig. 12). Interpreted
results for VES-4 are shown in Table 10. The third layer in
Table 10 with thickness 17.45 m and resistivity 38.16 O m can
be considered as a conducting chromite body. Compared to the
chromite body at the first location which starts from very near the
surface and extends to 33 m depth, the conducting body at this
location has a smaller thickness of 17.45 m and it is located at a
larger depth (about 30 m depth); therefore, a weak positive
gravity anomaly can be explained by this model.

Fig. 11. Fitting between field and model data for VES-3.
5. Discussion

Table 9 To obtain the 10 best fitting models in a specified search range,


Optimization results for VES-3 depicting mineral exploration. a large number of models (200,000) are analyzed during 10 runs
of VFSA optimization. Subsequently, statistical analysis of models
Parameters Search range Inversion result
(PDF 468.2%)
that have misfit errors below a certain threshold is performed.
The most appropriate region for computing the mean model is the
r1 (O m) 1–50 33.54 7 1.57 one where models have high a PDF. The mean model computed
r2 (O m) 100–10,000 568.7 7 15.2 from such models will be closest to the actual subsurface
h1 (m) 1–10 8.35 7 0.78 structure. It is important to optimize models by searching in a
Misfit error – 4.79  10  3
Correlation 0.9985
large model space initially. Subsequently, the model space can be
reduced either according to the estimated uncertainly (which will
S.P. Sharma / Computers & Geosciences 42 (2012) 177–188 187

also be large) or according to the known geological information, obtained by repeating the process. Therefore, a total of 200,000
and computations are performed again for a better mean model forward computations were performed to find the mean model
and smaller uncertainty. However, uncertainly always exists in and the associated uncertainty. The required computation times
the inverse solution. It can only be minimized to a certain extent for three-layered and four-layered models having 20 data points
based on the noise present in the data. are about 36 and 48 s, respectively.
The program VFSARES.F yields mean models on the basis of
two methods of computations. The first mean model is computed
from the 10 best fitting solutions obtained from various VFSA 7. Conclusions
runs. These models lie along the flat region of the equivalent
model space (specially for intermediate layer parameters) and A FORTRAN program is developed to interpret one-dimen-
individual parameters of the layer may be located in a wider sional resistivity sounding data for various electrode configura-
range. Since these models are searched in a large model space, the tions using the VFSA approach. Initially, results are presented for
uncertainty estimated from the covariance matrix computed from Schlumberger sounding data as it is the most commonly used
these best fitting models can be considered as global uncertainty. array and applicability of the developed program is demonstrated
It is interesting to derive results with smaller uncertainty by using synthetic and field data representing different types of
restricting the search range to appropriate model space only. subsurface structures. A large number of forward models
Therefore, uncertainty computed from these 10 best fitting (200,000 from 10 runs in this study) are evaluated and models
models is a guide to restrict the search range for recomputation are accepted or rejected according to the VFSA algorithm. Subse-
and the global model is derived on the basis of analysis of the PDF. quently, models having misfit errors lower than a predefined
The second mean model is computed from the models selected on value are selected for statistical analysis. Statistical analysis of
the basis of PDFs after analysis of the history of all the models and selected models is performed without giving any importance to
their misfit errors. Even though a limit of 68.2% is used in the their misfit error. Therefore, statistical analysis is performed in a
selection of the most appropriate region of the model space but model domain. Lognormal probability distribution is used to
one can set any desired limit. The result obtained by the statistical eliminate the models with lower PDFs. A 68.2% cutoff in PDF is
analysis is more robust than the first method and the mean model set for each model parameter to select models from the high PDF
is reliable and close to the actual subsurface structure. region of the multidimensional model space to compute the mean
An important question may be asked: Why do we need model and uncertainty. This yields a better mean model with a
multiple VFSA runs? The answer lies in the fact that whatever smaller uncertainty in comparison to a mean model computed on
slow cooling schedule and number of iterations (temperature the basis of 10 best fitting models from 10 VFSA runs. It is
level) are used in finding the best fitting model, this model will observed that uncertainty in the model parameters of the sand-
simply lie at a point along the elongated zone of equivalence. wiched layer is larger in comparison to other model parameters.
Histograms and PDFs depicted from the history of accepted However, actual model parameters always lie within the uncer-
models obtained from one VFSA run will be biased toward this tainty estimated for the mean model parameters. Inversion
particular best fitting model and simply exhibit the characteris- results for noisy synthetic data from different electrode config-
tics of the linearized inversion. Therefore, multiple VFSA runs are urations are also presented. In the case of noisy synthetic data,
required. All accepted models whose computed response fits the the actual model is never known, and in such a situation only the
observed data reasonably well (magnitude of noise in the data) uncertainty in the mean model can suggest the quality of
from different VFSA runs lie along the region of equivalence. The the result. Study reveals that dipole perpendicular arrays yield
histograms and PDFs depicted from such models help in finding the best results in resolving various layers efficiently. Finally, the
the global model. efficacy of the program is demonstrated for field data dealing
Another important question may be raised: Why is lognormal with groundwater and mineral exploration. The interpreted
distribution used for statistical computations? One can see from results matched very well with the drilling results. The program
the histogram in Figs. 2 and 5 that the target layer parameters (r2 is very easy to handle and it can be used to interpret resistivity
and h2) do not exhibit the Gaussian distribution. It shows a sounding data related to a number of geological problems.
lognormal distribution. PDFs in Figs. 3 and 6 also show a similar
nature, exhibiting a lognormal distribution. Other model para-
meters exhibit a Gaussian distribution. In principle, a mixed Acknowledgments
distribution will be more appropriate. However, it will be rather
tricky to handle such situations of mixed distribution and may be I thank Dr. S.K. Verma and Prof. P. Kaikkonen for their valuable
model dependent. Since intermediate layers are tricky to resolve support during the development of the approach. I thank the
and follow lognormal distributions, a single approach of lognor- Editor and anonymous reviewers for comments and suggestions
mal distribution is used for all model parameters. for improving the manuscript. Thanks are due to Prof. J. Chakra-
borty, Prof. D. Sengupta, and Johanna for editing and improving
the English language of the paper.
6. Computation time

Generally, a large number of forward calculations are per- Appendix A. Supplementary material
formed in a predefined model space in global optimization.
Therefore, global optimization is considered as a computationally Supplementary data associated with this article can be found
inefficient process. However, much less computing time is in the online version at doi:10.1016/j.cageo.2011.08.029.
required for analytical forward computations (1-D DC resistivity
in the present study). A desktop PC with Intel Core2Duo processor
is used to execute the developed program with MS FORTRAN References
Developer Studio. To obtain the best fitting model, computations
are performed at 1000 different temperature levels with 20 Baranwal, V.C., Sharma, S.P., 2006. Integrated geophysical studies in the East-
moves at one temperature level. Ten best fitting models are Indian geothermal province. Pure and Applied Geophysics 163, 206–227.
188 S.P. Sharma / Computers & Geosciences 42 (2012) 177–188

Barker, R.D., 1990. Investigation of groundwater salinity by geophysical methods. Loke, M.H., Barker, R.D., 1996. Practical techniques for 3D resistivity surveys and
In: Ward, S.H. (Ed.), Geotechnical and Environmental Geophysics, vol. 2. data inversion. Geophysical Prospecting 44, 499–523.
Society of Exploration Geophysicist, Tulsa, OK, pp. 201–211. Mohanty, W.K., Sharma, S.P., Gupta, S., 2010. Integrated Geological and Geophy-
Bas-okur, A.T., 1990. Microcomputer program for the direct interpretation of sical Study Around Tangarapada Area, Orissa. Final Technical Report IDCOL,
resistivity sounding data. Computers & Geosciences 16, 587–601. Government of Orissa, India, 123 pp.
Benson, A.K., Payne, K.L., Stubben, M.A., 1997. Mapping groundwater contamina- Mosegaard, K., Tarantola, A., 1995. Monte Carlo sampling of solutions to inverse
tion using DC resistivity and VLF geophysical methods—a case study. Geo- problems. Journal of Geophysical Research 100 (B7), 12431–12447.
physics 62, 80–86. Rothman, D.H., 1985. Nonlinear inversion, statistical mechanics and residual
Bhattacharya, B.B., Shalivahan, Sen, M.K., 2003. Use of VFSA for resolution, statics estimation. Geophysics 50, 2784–2796.
sensitivity and uncertainty analysis in 1D-DC resistivity and IP inversion. Rothman, D.H., 1986. Automatic estimation of large residual statics correction.
Geophysical Prospecting 51, 393–408. Geophysics 51, 337–346.
Das, U.C., Ghosh, D.P., 1973. A study on the direct interpretation of dipole sounding Sen, M.K., Bhattacharya, B.B., Stoffa, P.L., 1993. Nonlinear inversion of resistivity
resistivity measurements over layered earth. Geophysical Prospecting 21, sounding data. Geophysics 58, 496–507.
379–400. Sen, M.K., Stoffa, P.L., 1995. Global Optimization Methods in Geophysical Inver-
Das, U.C., Ghosh, D.P., Biewinga, D.T., 1974. Transformation of dipole resistivity sion. Elsevier Publishing Company 281 pp.
sounding measurements over layered earth by linear digital filtering. Geo- Sen, M.K., Stoffa, P.L., 1996. Bayesian inference, Gibbs’ sampler and uncer-
physical Prospecting 22, 476–489. tainty estimation in geophysical inversion. Geophysical Prospecting 44,
Das, U.C., Verma, S.K., 1981. The versatility of digital linear filters used in 313–350.
Sharma, S.P., Baranwal, V.C., 2005. Delineation of groundwater bearing fracture
computing resistivity and EM sounding curves. Geoexploration 18, 297–310.
zones in hard rock areas using integrated study of very low frequency
Ebraheem, A.M., Sensosy, M.M., Dahab, K.A., 1997. Geoelectrical and hydro-
electromagnetic and resistivity data. Journal of Applied Geophysics 57,
geochemical studies for delineating ground-water contamination due to salt-
155–166.
water intrusion in the northern part of the Nile delta, Egypt. Ground Water 35,
Sharma, S.P., Kaikkonen, P., 1998. Two-dimensional nonlinear inversion of VLF-R
216–222.
data using simulated annealing. Geophysical Journal International 133,
Fernández Alvarez, J.P., Fernández Martı́nez, J.L., Menéndez Pérez, C.O., 2008.
649–668.
Feasibility analysis of the use of binary genetic algorithms as importance
Sharma, S.P., Kaikkonen, P., 1999. Appraisal of equivalence and suppression
samplers—application to a 1-D DC resistivity inverse problem. Mathematical
problems in 1-D EM and DC measurements using global optimization and
Geosciences 40, 375–408. joint inversion. Geophysical Prospecting 47, 219–249.
Fernández Martı́nez, Juan L., Garcı́a Gonzalo, Esperanza, Fernández Álvarez, José P., Singh, U.K., Tiwari, R.K., Singh, S.B., 2005. One-dimensional inversion of geo-
Kuzma, Heidi A., Menéndez Pérez, César O., 2010. PSO: a powerful algorithm to electrical resistivity sounding data using artificial neural networks-a case
solve geophysical inverse problems: application to a 1D-DC resistivity case. study. Computers & Geosciences 31, 99–108.
Journal of Applied Geophysics 71, 13–25. Tarantola, A., 1987. Inverse Problem Theory, Methods of Data Fitting and Model
Ghosh, D.P., 1971a. The application of linear filter theory to the direct interpreta- Parameter Estimation. Elsevier Publishing Company 630 pp.
tion of geoelectrical resistivity sounding measurements. Geophysical Pro- Van Overmeeren, R.A., 1989. Aquifer boundaries explored by geoelectrical mea-
specting 19, 176–180. surements in the coastal plain of Yemen: a case of equivalence. Geophysics 54,
Ghosh, D.P., 1971b. Inverse filter coefficients for the computation of apparent 38–48.
resistivity standard curves for a horizontally stratified earth. Geophysical Vedanti, N., Srivastava, R., Sagode, J., Dimri, V.P., 2005. An efficient 1D Occam’s
Prospecting 19, 769–775. inversion algorithm using analytically computed first- and second-order
Inman, J.R., 1975. Resistivity inversion with ridge regression. Geophysics 40, derivatives for DC resistivity soundings. Computers & Geosciences 31,
798–817. 319–328.
Johansen, H.K., 1977. A man/computer interpretation system for resistivity Verma, S.K., Sharma, S.P., 1993. Resolution of thin layers using joint-inversion of
soundings over horizontally stratified earth. Geophysical Prospecting 25, electromagnetic and direct current resistivity sounding data. Journal of
667–691. Electromagnetic Waves and Applications 7, 443–479.
Koefoed, O., 1979. Resistivity Sounding Measurements, Methods in Geochemistry Zohdy, A.A.R., 1989. A new method for the automatic interpretation of Schlum-
and Geophysics 14A, Geosounding Principles 1Elsevier, Amsterdam 276 pp. berger and Wenner sounding curves. Geophysics 54, 245–253.

You might also like