Vous êtes sur la page 1sur 29

See

discussions, stats, and author profiles for this publication at:


http://www.researchgate.net/publication/275670745

Combining Multiple Electrode Arrays for


Two-Dimensional Electrical Resistivity
Imaging Using the Unsupervised
Classification Technique
ARTICLE in PURE AND APPLIED GEOPHYSICS JUNE 2015
Impact Factor: 1.62 DOI: 10.1007/s00024-014-1007-4

READS

53

3 AUTHORS:
Kehinde S. Ishola

Mohd Nawawi

University of Science Malaysia

University of Science Malaysia

7 PUBLICATIONS 1 CITATION

55 PUBLICATIONS 70 CITATIONS

SEE PROFILE

SEE PROFILE

Khiruddin Abdullah
University of Science Malaysia
219 PUBLICATIONS 176 CITATIONS
SEE PROFILE

Available from: Mohd Nawawi


Retrieved on: 01 November 2015

Pure Appl. Geophys.


 2014 Springer Basel
DOI 10.1007/s00024-014-1007-4

Pure and Applied Geophysics

Combining Multiple Electrode Arrays for Two-Dimensional Electrical Resistivity Imaging


Using the Unsupervised Classification Technique
K. S. ISHOLA,1,2 M. N. M. NAWAWI,1 and K. ABDULLAH1
AbstractThis article describes the use of k-means clustering,
an unsupervised image classification technique, to help interpret
subsurface targets. The k-means algorithm is employed to combine
and classify the two-dimensional (2D) inverse resistivity models
obtained from three different electrode arrays. The algorithm is
initialized through the selection of the number of clusters, number of
iterations and other parameters such as stopping criteria. Automatically, it seeks to find groups of closely related resistivity values that
belong to the same cluster and are more similar to each other than
resistivity values belonging to other clusters. The approach is applied
to both synthetic and field data. The 2D postinversions of the
resistivity data were preprocessed by resampling and interpolating to
the same coordinate. Following the preprocessing, the three images
are combined into a single classified image. All the image preprocessing, manipulation and analysis are performed using the PCI
Geomatics software package. The results of the clustering and
classification are presented as classified images. An assessment of
the performance of the individual and combined images for the
synthetic models is carried out using an error matrix, mean absolute
error and mean absolute percent error. The estimated errors show
that images obtained from maximum values of the reconstructed
resistivity for the different models give the best representation of the
true models. Additionally, the overall accuracy and kappa values
show good agreement between the combined classified images and
true models. Depending on the model, the overall accuracy ranges
from 86 to 99 %, while the kappa coefficient is in the range of
5498 %. Classified images with kappa coefficients greater than 0.8
show strong agreement, while images with kappa coefficients greater
than 0.5 but less than 0.8 give moderate agreement. For the field data,
the k-mean classifier produces images that incorporate structural
features of the three electrode array configurations. Consequently,
some clusters that overwhelmingly correspond to the lithologic units
of the investigated areas are better identified than the tomographic
images of each data set considered separately, underscoring the
relevance of the unsupervised classification technique in this study.
Key words: Unsupervised classification, electrode arrays,
k-means clustering, clusters, overall accuracy, lithologic units.

Geophysics section, School of Physics, Universiti Sains


Malaysia, 11800 Pulau Pinang, Penang, Malaysia. E-mail: saidisho@yahoo.co.uk;
mnawawi_nordin@yahoo.com;
khir_deen@yahoo.com
2
Department of Geosciences, University of Lagos, Akoka,
Lagos, Nigeria.

1. Introduction
In geophysical investigations involving hydrogeology, subsurface exploration, mining, geotechnical
and archaeological prospecting, electrical resistivity
imaging has remained a vital tool for some decades
(DAHLIN 1996; SEATON and BURBEY 2000; CANDANSAYAR and BASOKUR 2001; LOKE et al. 2013). The
advancements made in the use of this tool by
allowing resistivity data to be collected and processed
in a short period of time using a suitable electrode
array are worthy of commendation. This is coupled
with the advent of multielectrode resistivity recording
systems that have popularized the use of electrical
resistivity imaging. However, one of the problems of
resistivity investigations is the choice of a suitable
electrode configuration that will give the best
response to the observed targets in the subsurface
(ZHOU and GREENHALGH 2000; ZHOU and DAHLIN
2003). Several electrode arrays have been used in
resistivity investigations, including, but not limited
to, the Wenner, Wenner-Schlumberger, dipoledipole, pole-pole and pole-dipole arrays (DAHLIN
1996; CHAMBERS et al. 1999; STORZ et al. 2000). Each
electrode array has its advantages and limitations
regarding field operations and interpretation capabilities (LOKE et al. 2003; DAHLIN and ZHOU 2004; ABER
and MESHIN CHIASL 2010). Also, the depth of investigations (ROY and APPARAO 1971; BARKER 1989;
OLDENBURG and LI 1999; SZALAI and SZARKA 2008),
sensitivity to horizontal or vertical variations, and
signal strength (LOKE 2001; SEATON and BURBEY
2002; DAHLIN and ZHOU 2004; CANDANSAYAR 2008)
are some of the other factors needing to be taken into
consideration in using these arrays. In addition, the
geological structures to be mapped, heterogeneities of
the subsurface, sensitivity of the resistivity meter

K. S. Ishola et al.

(FURMAN et al. 2003), sensitivity of the arrays to


vertical and lateral variations in the resistivity of the
subsurface, horizontal data coverage and signal
strength of the array (DAHLIN and ZHOU 2004) should
be considered before embarking on field surveys.
In conducting an electrical resistivity survey over
a noisy site, the Wenner array is preferred because of
its high signal strength and high resolution of horizontal structures, but it is less easily applicable when
mapping 3D structures where the dipole-dipole, poledipole and pole-pole arrays are given preference
(DAHLIN and LOKE 1997; ZHOU and DAHLIN 2003). The
Wenner array also has a good vertical resolution, thus
making it suitable for imaging horizontal structures
(IBRAHIM et al. 2003). The dipole-dipole array is most
sensitive to resistivity variations beneath the electrodes in each dipole length and is very sensitive to
horizontal variations. The dipole-dipole array is relatively insensitive to vertical variations in the
subsurface resistivity (GRIFFITHS and BARKER 1993).
As a result, the dipole-dipole array is given priority
for mapping of vertical structures such as dykes and
cavities (LOKE 2001; EL-QADY et al. 2005). In practice, electrical resistivity imaging involves the use of
either four electrode arrays (e.g., the Wenner Schlumberger and dipole-dipole) or the three electrode
type typical of the pole-dipole with a polarityreversed array for all possible currents and potential
electrode combinations. To gain knowledge of the
relative potentials of these electrode arrays, several
researchers have worked in this direction. SZALAI and
SZARKA (2008) classified about 100 electrode arrays
into eight classes based on three parameters; they are
superposition, focusing and collinearity. Also, STUMMER et al. (2004) opined that electrical resistivity data
acquired using a large number of four-point electrode
arrays gave substantial subsurface information compared to the data sets obtained from both individual
arrays, such as the Wenner, dipole-dipole or a combination of the Wenner and dipole-dipole arrays.
Moreover, DE LA VEGA et al. (2003) used a combined
inversion algorithm for different electrode array data
sets with the combined inversion images outperforming the inversion results obtained from separate
electrode arrays. ATHANASIOU et al. (2007) introduced
a weighting factor to the two-dimensional (2D)
combined inversion algorithm, which prevented the

Pure Appl. Geophys.

dominance of one array type over another as the


applied parameter allowed equal participation of the
data from an individual array.
In previous studies, attempts have been made at
comparing the ability of different electrode arrays to
resolve, map or identify subsurface targets (SEATON
and BURBEY 2002; DAHLIN and ZHOU 2004; CANDANSAYAR 2008). Several researchers have compared
different electrode arrays individually on the basis of
their sensitivity analysis, depth of investigations, and
responses to resolving vertical or horizontal structures (SASAKI 1992; DAHLIN 2001; BENTLEY and
GHARIBI 2004; FIANDACA et al. 2005; CAPIZZI et al.
2007; BERGE and DRAHOR 2009; MARTORANA et al.
2009; NEYAMADPOUR et al. 2010; RUCKER 2012).
Meanwhile, in related studies, the use of joint
inversion techniques have been introduced and used
for combining two or more geophysical data into a
single image for the cross gradients (GALLARDO and
MEJU 2003, 2004, 2007) and structural approach
(HABER and OLDENBURG 1997). The efficacy of the
joint inversion method hinges on the complementary
nature of the geophysical data sets (DOETSCH et al.
2010). In addition, concerted efforts have been geared
toward optimizing different electrode arrays for
electrical resistivity surveys in order to obtain as
much information as necessary in the detection and
imaging of subsurface structures within a short period
of time (STUMMER et al. 2002, 2004; WILKINSON et al.
2006a, b; COSCIA et al. 2008; LOKE et al. 2010; HAGREY and PETERSEN 2011; HAGREY 2012). In view of
this, COSCIA et al. (2008) proposed an experimental
design method involving an independent comparison
between the information provided by one electrode
array with others using crosshole electrical resistivity
imaging. In another work, LOKE et al. (2010) adopted
four methods to automatically select an optimal set of
array configurations that provided maximum subsurface model resolution for an electrical imaging
applied to both synthetic and field surveys.
In the present study, however, we focus on the
combination of the 2D post-inversion resistivity
models from three different electrode arrays, namely,
the dipole-dipole (Dpd), Wenner-Schlumberger
(Wsc) and pole-dipole (Pdp), reflecting a wide range
of geological situations, using an unsupervised image
classification technique. To this end, the capability

Combining Multiple Electrode Arrays

and functionality of an image processing technique


are explored to generate a single integrated image for
both synthetic and real-field examples.
In this article, we present a review of the different
electrode arrays used in electrical resistivity imaging
with emphasis on the advantages and their limitations. Also, a technique for combining the images
from some commonly used electrode array configurations is mentioned. The remainder of the article is
organized into the traditional theoretical background,
method, results and discussion followed by directions
for further work. In Sect. 2, we introduce the image
classification technique, which emphasizes two broad
groups of classification techniques. Section 3 discusses the methodology, which includes numerical
modeling, merging of post-resistivity inversion data
sets, clustering and classification procedures. Also,
the other aspects of the methodology presented are
the assignment of resistivity to the classification
images and criteria for evaluating the accuracy of the
clustering technique. Section 4 presents the results of
the application of the clustering technique to both
synthetic and field examples. In addition, the discussion section interprets the results, and suggestions
for continued research are provided.

2. Image Classification Techniques


Mechanistically, image classification is a process
that translates the raw data into more meaningful
and understandable forms (TSO and OLSEN 2005).
The image classification technique involves clustering the pixels of an image, where a pixel is the
smallest unit of digital image data that can be
individually processed, into a relatively small set of
clusters, such that pixels in the same class have
similar statistical properties. Image classification
relies on distinctive signatures or spectral contents
of the classes as well as the ability to reliably distinguish these signatures from others (EASTMAN
2003). The image classification techniques stand out
from other data integration methods, such as artificial neural networks, self-organizing maps and joint
inversion methods, to mention but a few. This is
because of its predictive abilities (KVAMME 2006) in
subsurface target detection using scattered fields

(CHATURVEDI and PLUMB 1995). In the joint inversion


method, an external constraint is imposed on the
models (e.g., PRIDE 1994; TRYGGVASON et al. 2002;
LINDE et al. 2006; LINDE and DOETSCH 2010) as the
approach seeks an existing empirical or mathematical relationship between the models, whereas the
image classification technique looks for a natural
grouping/pattern in the data sets.
Generally, image classification can be grouped as
either supervised or unsupervised classification, and
both groups are used especially in remote sensing
for image analyses. In supervised classification, the
pixel classes are controlled by the analyst through
the process of selecting training sites in advance in
order to train the classifier (CAMPBELL 2002; TSO and
OLSEN 2005). On the other hand, unsupervised
classification works by objectively extracting various
features of an image. The detailed information
derived from an image is guided by the number of
clusters into which the image is partitioned. The
preference of unsupervised over supervised classification schemes in this study is a result of knowledge
of the site to be classified being immaterial at the
initial separation of the image pixels. Meanwhile,
the classification process is less prone to human
error as an analyst is not at liberty to make as many
decisions during the classification process and the
classes in the data are not overlooked (LILLESAND
and KIEFER 1994; EASTMAN 1995; ENDERLE and WEIH
2005).
Among the commonly used algorithms for partitioning of image data sets in unsupervised
classification techniques are the k-means, iterative
self-organizing data analysis (ISODATA) and fuzzy
c-means. The ISODATA algorithm is similar to the kmeans with the distinct difference that it allows for
the minimum-maximum number of clusters, while
the k-means assumes that the number of clusters is
known a priori. On the other hand, the fuzzy c-means
is an alternative way of assigning data to specific
clusters with an absolute in or out value based on
degree of membership of the data into clusters (WARD
et al. 2014). Each of the clustering algorithms uses an
iterative procedure for organizing objects into a predefined fixed number of clusters while attempting to
optimize the distance between the observational data
set and cluster centers (CHENG 2003; KHAN and

K. S. Ishola et al.

AHMAD 2004). Thus, unsupervised classification


techniques partition data sets into clusters by ensuring
that intercluster variability is maximized and minimizing intraclass variability without recourse to field
knowledge (LILLESAND and KEIFER 2000).
2.1. The k-means Clustering Algorithm
The k-means is a clustering algorithm (MACQUEEN
1967) that partitions multidimensional pixels in the
model space of the image, usually in a vector form
into prespecified k clusters. In the context of this
study, the clusters consist of resistivity values, and
the grouping is performed such that pixels within a
cluster are similar (i.e., intracluster variability) and
are dissimilar from pixels belonging to another
cluster (i.e., intercluster) with each cluster represented by cluster centers. Thus, clustering techniques
are popularly used for finding interesting features or
patterns in a data set that may not be obvious (RANA
et al. 2010).
The theoretical background of the k-means algorithm is as follows: suppose there are N input data
points such that x1 ; x2 ; . . .; xm 2 Rn is a finite number
of patterns existing in the model space of the model;
then the k-clustering aims at partitioning these data
points into k disjoint subsets or clusters C1 ; . . .; Ck ;
with k \ N such that optimization is performed based
on the clustering criterion. The criterion used for the
optimization is the sum of the squared of Euclidean
distances between each data point, xi, and the cluster
center, mk, of the subset Ck, which contains xi. This
optimization criterion is called the clustering error,
and it depends on the cluster centers m1, , mk. At
the k-th iterative step, the N observations are
partitioned into k clusters using the relation,
x 2 Cj k,


providedx  mj k\kx  mi kk
1
for all i 1; 2; . . .; k; i 6 j, where Cj k denotes the
set of observations whose cluster center is mj(k). In
the next step, a new cluster center mj k 1; j
1; 2; . . .; k is computed such that the sum of the
squared distances from all points (pixels) in Cj(k) to
the new cluster center is minimized. A measure that
minimizes this is the observation mean of Cj(k). Thus,
the new cluster center formed is given by:

Pure Appl. Geophys.

mj k 1

1 X
x;
Nj x2C k

j 1; 2; . . .; k

where Nj is the number of pixels in the image. The


cluster centers are reference points used by the algorithm to group the observation data points (pixels) into
clusters. The cluster centers, also called centroids, are
computed as the average of the observations in a
particular cluster. Being a point-based algorithm, the
k-means starts with the cluster centers that are initially
placed at arbitrary positions and proceeds by moving
the cluster centers at each step so that it minimizes the
clustering error. For clustering of the data set, a large
number of dimensions contains more information, and
the clusters obtained tend to increase with increasing
dimensionality in the data set.
The k-means algorithm calculates its cluster
centers iteratively by initializing the centers in mk
and decides membership of the pixels in one of the kclusters according to the closest neighbor, i.e., the
minimum distance from the cluster center. Then, it
calculates the new mk by repeating the steps above
until there are no changes in the cluster centers. Thus,
the goal is to minimize the sum of squares error
within each cluster represented by its Euclidean
distance as:
dx; mk

N X
m 
X

xi  mj 2

i1 j1

where kk2 is a measure of the minimum distance


between a pixel, xi , and the cluster center, mj. In
Eq. 3, high membership values are assigned to pixels
that are close to the cluster center for a particular
cluster, while low membership values are assigned to
pixels that are far from the centroid (PHAM and PRINCE
1999).
2.2. Selection of the Number of Clusters
In most clustering situations, an important step in
the implementation of the clustering procedures is to
specify the number of clusters into which the image is
to be partitioned. In this regard, an analyst is faced
with the dilemma of selecting the number of clusters
or partitions in the final solution as there is no
existing benchmark in the literature regarding the
selection of an optimal number of clusters. To

Combining Multiple Electrode Arrays

overcome this nagging problem in clustering, a


variety of procedures has been proposed and used
for determining the number of clusters in data sets
(MILLIGAN and COOPER 1985; HALKIDI et al. 2001;
KVAMME 2006; WARD et al. 2014). Doing so will
assist in the final interpretation of the classification
image. In practice, the appropriate number of clusters
for a given data set is based on trial and error. The
subjective nature of deciding what constitutes correct
grouping by this method makes the selection difficult.
Another approach is to run the clustering algorithm
many times with the number of clusters gradually
being increased from a certain initial value to some
threshold values and the partitioning of data resulting
in the best validity measure being selected (HALKIDI
et al. 2001). WARD et al. (2014) used the number of
peaks obtained from the application of the kernel
density estimation method to the data set to provide a
statistically grounded analysis of the data that in turn
was assumed to represent the appropriate number of
clusters required to group the data.
Grouping of the pixels into a few prespecified
clusters could result in a loss of important details
about the models during clustering of the data sets. In
order to ensure at the outset that no important
information in the image data sets is ignored,
clustering is often prepared with a large number of
clusters to reflect the specific characteristics of the
data sets (CIHLAR et al. 1998; PHAM et al. 2004; GAO
2009). However, using a large number of clusters also
poses a problem, as too many details that could give
irrelevant interpretations of the targets may be
created (FIGUEIREDO and JAIN 2002; ERNENWEIN
2009). Therefore, finding an optimum number of
clusters in a data set could be very challenging since
it requires a priori knowledge of the data in some
cases. The performance of the k-means clustering
algorithm in terms of adequate interpretation of the
classification image depends on the selection of an
optimal number of clusters, types of attributes (i.e.,
measured parameters), types of data sets and scales of
attributes.
2.3. Stopping Rules of the k-means Algorithm
The convergence of the k-means algorithm is
guided by the cost function or sum of the square

error, which reduces to a local minimum. At this local


minimum, the correct number of cluster centers
converges into an actual cluster center as other initial
centers move away from the input data set (ZALIK
2008). A satisfactory clustering result is obtained
when the number of iterations as a prerequisite is
indicated before the user/analyst runs the algorithm.
A number of convergence conditions are possible in
the execution of clustering algorithms. These include:
(1) the search may stop when a given or defined
number of iterations is reached, (2) stopping when the
partitioning error is not reduced because of the
relocation of the centers is an indication that the
partition is locally optimal, (3) when there is no
exchange of data points between clusters, exceeding a
predefined number of iterations, and (4) when a
threshold value is attained. This convergence threshold corresponds to the maximum percentage of pixels
whose clusters remain unchanged from the previous
iteration (REIGBER et al. 2010; LI et al. 2013). At this
stage, the clusters are saved as the best cluster
solution for the clustering. Also, the threshold
parameter, T, typically a specific value, is stated or
selected by the user. The k-means algorithm based on
a user-supplied parameter can modify the threshold
value by either increases or decreases depending on
the number of clusters required to represent the data
set.

3. Methodology
Subsurface imaging over a target using different
electrode configurations produces tomographic models of the target. Each inverse resistivity image gives
different information about the investigated area.
This is due to nonuniqueness in the inversion process
leading to ambiguity in the interpretation of the
results. To reduce the ambiguity, there is a need for
all the information from the different electrode configuration images to be integrated into a single image.
To this end, an unsupervised classification technique
using the k-means algorithm that automatically partitions the 2D inverse resistivity model data sets into
some prespecified number of clusters with each
cluster linked to a particular geological or hydrogeological (i.e., electrofacies) unit that makes up the

K. S. Ishola et al.

resistivity image is employed. The procedure used in


this study can be grouped into four stages: (1)
numerical modeling, (2) clustering and classification
procedures, (3) assignment of resistivity and (4)
accuracy assessment.
3.1. Numerical Modeling of Synthetic Data
3.1.1 Forward Modeling
The synthetic models used in this study were created
through forward modeling. Forward modeling, also
known as the forward solver, involves a mapping
from the model space to the data space (SCHWARZBACH
et al. 2005). Forward modeling estimates/predicts
data on the basis of the known distribution of model
parameters and electrode configuration used. The
forward modeling for the DC potentials is accomplished using a finite difference (FD) technique to
solve the partial differential equation for charge
distribution. The program that performs this calculation is the RES2DMOD software package (GEOTOMO
2005). This forward solver is used to calculate the
apparent resistivity data for a user-defined 2D
subsurface model (i.e., it computes the electric
potential difference for a particular resistivity model).
In the FD approach, a resistivity model of the
subsurface is first discretized into several rectangular
mesh or grids. Then, the FD method determines the
potentials at the nodes of the rectangular mesh in both
directions. The DC resistivity forward modeling was
performed with the synthetic data shown in Fig. 1ae
generated over the 2D resistivity structures. The
potential differences were computed for the three
electrode arrays, namely the Wenner-Schlumberger
(Wsc), dipole-dipole (Dpd) and pole-dipole (Pdp)
arrays. In all the simulations, 40 surface electrodes at
an electrode spacing of 1 m were used.
In this study, five synthetic models were used to
simulate a typical field survey and test the image
classification approach for the integration of different
electrode arrays images. These include:
1. A resistivity block prism (target) with a resistivity
value of 500 X-m (light blue) embedded in a
homogeneous medium of 10 X-m (deep blue)
(Fig. 1a).

Pure Appl. Geophys.


Figure 1
c
Synthetic resistivity models used for stimulation of a a resistive
block, b two resistive blocks, c three resistive blocks, d a vertical
fault and e a resistive dyke

2. Two resistivity blocks having resistivity values of


100 X-m (green) and 300 X-m (orange) embedded
in a homogeneous medium of 10 X-m (deep blue)
(Fig. 1b).
3. Three blocks with resistivity values of 100 X-m
(left), 300 X-m (middle) and 500 X-m (right) all
embedded in a homogeneous background of 10 Xm (Fig. 1c).
4. A vertical fault juxtaposing a conductive top layer
(hanging wall) with resistivity of 10 X-m (blue)
and a resistive bottom layer (foot wall) with
resistivity of 100 X-m (green) (Fig. 1d)
5. A resistivity dyke of 500 X-m (light blue)
intruding in a homogeneous background 100 X-m
(green) (Fig. 1e).
3.1.2 Inverse Modeling
The determination of the model parameter (i.e.,
resistivity estimates) using the data space and model
space is known as inversion. Unlike forward modeling, inverse modeling is the mapping from the data
space to the model space (OLDENBURG and ELLIS
1991). Inversion is used to reconstruct the subsurface
resistivity distribution from the measurements of
voltage and current data. Inversions of the apparent
resistivity data sets from forward modeling were
carried out using the commercially available
RES2DINV software (GEOTOMO 2005). During the
inversion process, 5 % Gaussian noise was added to
reflect field conditions. The inversion of the apparent
resistivity data from forward modeling into 2D
resistivity models was carried out using RES2DINV,
a commercially available inversion program. This
software program uses two different inversion routines for the generation of earth models from the
resistivity data. The inversion routines are based on
the blocky or L1-norm optimization (LOKE et al.
2003) and Gauss-Newton smoothness-constrained
least-squares method for L2-norm optimization (DEGROOT-HEDLIN and CONSTABLE 1990; ELLIS and
OLDENBURG 1994). The L1-norm optimization method

Combining Multiple Electrode Arrays

K. S. Ishola et al.

allows earth resistivity models with relatively sharp


variations across boundary resistivity to be reconstructed as it tries to minimize the sum of absolute
values of the models resistivity. It is a better
optimization choice when geological discontinuities
are expected (SEATON and BURBEY 2002; LOKE et al.
2003). On the other hand, the L2-norm based on the
least-squares optimization method attempts to minimize the sum of squares of the spatial variations in
the resistivity of the models and the data misfit
leading to earth resistivity models with gradual
transitions across zones of different resistivities
(LOKE et al. 2003). Although the default inversion
routine used by this program is based on an L2-norm,
both optimization techniques were used initially, but
the L2-norm was observed to have smearing effects
on the boundaries (LOKE et al. 2003) of the models.
As a result, the L1-norm was preferred throughout the
inversion process in the synthetic examples.
3.1.3 Regularization Parameters
Inverse problems are ill posed because of the
nonuniqueness resulting from sparse noisy data and
the need to finely parameterize the earth so that
enough variation is allowed in the solution (MEJU
1994; HABER and OLDENBURG 2001). To overcome illposedness in the inversion, regularization schemes
are usually applied (DOESTCH et al. 2010). Examples
of the regularization constraints that allow for
stabilization of the inversion process are damping
and smoothing (BLOME et al. 2011). The common
perspectives in these regularization schemes are that
the data are given and inversion should produce a
single model that fits the data (ELLIS and OLDENBURG
1994). A 2D earth model is parameterized by means
of a grid of rectangular prisms, each having a uniform
resistivity. As a tradition, each block is made smaller
than the data resolution length so that the position of
the block boundaries does not affect the final model.
The regularization parameter is a trade-off between
data misfit and model roughness as a large value
results in a smooth model and weak data misfit, while
a small value leads to a highly rough model with
good data misfit (TIKHONOV and ARSENIN 1977; LOKE
et al. 2003). Several methods exist to compromise
data misfit and model constraints (FARQUHARSON and

Pure Appl. Geophys.

OLDENBURG 2004). In order to obtain a solution for the


inverse problem, an important factor to consider is
the damping factor (k), which depends among other
factors on the distribution of the resistivity data.
An estimation of the damping factor is based on
the trial and error method (LOKE and BARKER 1996;
OLAYINKA and YARAMANCI 2000). The selection of an
appropriate damping factor is important for subsurface imaging inversion in order to avoid local
minima. Using low damping values could lead to
low data misfit, but produce high instability in the
values of the estimated resistivity between adjacent
cells; this in turn could result in erroneous assumptions of the model resistivity in the area under study
(LOKE et al. 2010). To establish an optimum value of
the damping factor that gives the best inversion
results after successive iterations, OLAYINKA and
YARAMANCI (2000) suggest the use of a damping
factor in the range of 0.050.25 with a 20 %
increment between successive iterations. Also, Easton (1987) opined that higher damping values should
be used at the beginning and lower the damping
values on subsequent iterations.
In this article, the suggestion provided by OLAYINKA and YARAMANCI (2000) in respect to the
value(s) for the damping factor for models inversions
was adopted. The inversions were carried out with an
initial damping factor of k = 0.25, which was
reduced to a minimum of 0.015 after successive
iterations. Tables 1 and 2 display the data points for
the three electrode arrays used for providing information about the inverse models and parameters used
for the resistivity inversion with the RES2DINV
software package, respectively. The propagation
factor, n, for the dipole-dipole array increases from
1 to 6 (Table 1). It is advisable not to use n values
greater than 6 for the dipole-dipole array in order to
increase the depth of investigation because of the
weak signal strengths of this array (LOKE 2009).
Table 1
Modeling parameters for the three electrode arrays used
Parameter/electrode array

Dpd

Wsc

Pdp

Number of data points


Number of model layers
Number of blocks

308
6
178

253
8
204

319
8
240

Combining Multiple Electrode Arrays

Table 2
Summary of parameters used during 2D resistivity inversions
Initial damping factor
Minimum damping factor
Convergence limit
Minimum change in absolute error
Number of iterations
Jacobian matrix is recalculated for the first
two iterations
Increase of the damping factor with depth
Robust data inversion constraint is used with
the cutoff factor
Robust model inversion constraint is used with
the cutoff factor
Extended model is used
Effect of side blocks is not reduced
Normal mesh is used
Finite difference method is used
Number of nodes between adjacent electrodes
Logarithm of the apparent resistivity used
Reference resistivity used is the average of minimum
and maximum values

0.25
0.015
1
\10 %
38

1.05
0.05
0.005

3.2. Assignment of Resistivity Without Clustering


The reconstructed inverse resistivity values of the
three electrode array configurations were combined
using Eqs. 4, 5, 6 and 7. This combination yields four
new images (i.e., minimum, maximum, median and
average), which are illustrated as:
Rmin xi ; yi minRdpd ; Rwsc ; Rpdp

Rmax xi ; yi maxRdpd ; Rwsc ; Rpdp

Rmed xi ; yi medRdpd ; Rwsc ; Rpdp




Rdpd Rwsc Rpdp
Ravg xi ; yi
3

6
7

where Rdpd xi ; yi ; Rwsc xi ; yi and Rpdp xi ; yi are the


resistivity values for the dipole-dipole, Wenner-Schlumberger and pole-dipole electrode arrays. Also, for
the models images, the mean resistivity values
denoted by Rmax , Rmin , Rmed and Ravg for the maximum, minimum, median and average images,
respectively, were assigned to the blocks and background of the models.
3.3. Clustering and Classification Procedures
The 2D post-inverse resistivity models saved in
text files format, i.e., as xyz, where x represents the
electrode position, y is the depth of investigation, and

z represents the resistivity values, were exported into


an image-processing environment using the PCI
Geomatics software for further manipulations and
analysis. The PCI Geomatics used throughout this
article has a variety of functionality that enables users
or analysts to get more from imagery in support of a
wide range of geospatial applications. It is used in the
remote sensing environment for imagery processing,
geographic information systems and photogrammetry
(PCI 2001).
To implement the k-means algorithm, the following input parameters were used in addition to the
three input images from the electrode configurations:
(1) the number of clusters, which is usually a priori
defined, was 16. To avoid the difficulty in the
selection of an appropriate number of clusters, in this
article the prior information obtained during the 2D
inversion process was used to specify the number of
clusters, k, for the clustering. By this an objective
evaluation measure to suggest suitable values for the
number of clusters, thus avoiding the need for trial
and error, is implied. Moreover, the use of prior
knowledge from the inversion results provided an
automatic decision rule eliminating the problems of
human subjectivity. This was to ensure that adequate
information needed for the interpretation of the
model was obtained: (2) the stopping criteria (i.e.,
the number of iterations and the convergence threshold were set to 20 and 0.01, respectively). After
convergence, the k-means algorithm creates an output
image file with a thematic raster layer as a result of
clustering, and this is stored in the PCIDSK program.
The PCIDSK is a data structure for holding images
and related data. When this convergence threshold is
attained, the program terminates. Prior to classification of the combined images, we masked the region
of the models images that did not have data points
manually. This was necessary to ensure that only the
coregistered part of the images was classified.
3.3.1 Assignment of Resistivity to Clusters
Each cluster formed is associated with three resistivity values from the three electrode array
configurations. These resistivity values are denoted
by Rc1, Rc2 and Rc3 in contrast to the resistivity
assigned to the images obtained without clustering in

K. S. Ishola et al.

Sect. 3.2. To assign a single resistivity value, we used


the minimum, maximum, median and average of Rc1,
Rc2 and Rc3 represented by Rcmin ; Rcmax ; Rcmed and
Rcavg , respectively, using Eqs. 8, 9, 10 and 11.
Rcmin xi ; yi minRc1 ; Rc2 ; Rc3

Rcmax xi ; yi maxRc1 ; Rc2 ; Rc3

Rcmed xi ; yi medRc1 ; Rc2 ; Rc3

10

Rcavg xi ; yi Rc1 Rc2 Rc3 =3:

11

Also, overall resistivity values denoted by


Rcmin ; Rcmax ; Rcmed and Rcavg were assigned to the
blocks and background of the classified images. The
final step in the classification campaign was merging
of clusters.
Another important stage in image classification
process involves aggregation or merging of clusters in
the classified images and then ascribing the merged
clusters to desirable classes according to similarity in
their spectral properties. To this end, merging was
confined to spatially adjacent clusters by comparison
of one cluster with its neighbor(s) and assigning it to
the same class on the basis of their pixel values. This
was carried out through the class merging submenu
available at the Geomatica Focus window of the PCI
software. Also, the spatial context (i.e., boundary) of
the individual pixels can be used. This is particularly
relevant when the spectral signatures of the clusters are
reasonably similar. Thus, as the clusters were merged,
the data structure of the classified image was updated
by reestimating the resistivity value of each class.
3.4. Accuracy Assessment
After assigning resistivity values to the clusters,
an assessment of the accuracy/performance of the
classified images was carried out. This was carried
out relative to the true model images. There are many
criteria for assessing the performance of models
images in the literature. These include: R-square,
semi-partial R-square, absolute error (AE), rootmean-square, standard deviations, mean absolute
error (MAE) and mean absolute percentage error
(MAPE). In this article, however, accuracy assessment being an important step in the classification
process is necessary in order to evaluate the quality of

Pure Appl. Geophys.

the classified images. With this end in view, the error


matrix or contingence tables were used for quantitative evaluation of the images. The error matrix table
displays statistics for assessing the accuracy of
classification images using the number of pixels that
represent the features in the image (SINGH 1989;
CONGALTON and GREEN 1999; LILLESAND and KEIFER
2000; LIU et al. 2003; MELGANI and BRUZZONE 2004).
To construct the error matrix table, every pixel in the
classified images was compared with the pixel in the
reference images (i.e., true models) for the blocks and
background of the models images. Then, percentages
obtained from the table gave the accuracies of the
classification (CONGALTON 1991; LILLESAND and KEIFER 2000; LIU et al. 2003; ENDERLE and WEIH 2005;
HASMADI et al. 2009; PERUMAL and BHASKARAN 2010).
In interpreting classification accuracy, it is important
not only to note the percentage of correctly classified
pixels but also to determine the nature of errors of
omission and commission on a pixel-by-pixel basis.
Producers accuracy (PA) is obtained by dividing
the number of correctly classified pixels in each
category by the total number of pixels in the
corresponding column. Users accuracy (UA) is
computed by dividing the number of correctly
classified pixels by the total number of pixels in the
corresponding row (i.e., it is concerned with what
percentage of the cluster has been correctly classified). The PA and UA were obtained using Eqs. 12
and 13, respectively, as follows:
PA % 100  error of omission %

12

UA % 100  error of commission %: 13


In these equations, both the PA and UA are
related to errors of omission and commission,
respectively. Pixels that are incorrectly excluded
from a particular cluster are defined as an error of
omission, while pixels that are incorrectly assigned to
a particular cluster that actually belong in other
classes are defined as an error of commission
(BOSCHETTI et al. 2004).
Also, the percentage of overall accuracy (OA)
was estimated using Eq. 14 given by:
O:A

U
 100 %
W

14

Combining Multiple Electrode Arrays

where U is the total number of correctly classified


pixels in the same category (i.e., sum of diagonal
pixels) and W is the total number of pixels (i.e., sum
of pixels in a row or column) in the classified image.
There are situations where the correct classification is
made by chance. In this case the most widely used
statistics for the estimation of the effect of change
agreement is known as the kappa coefficient (j). The
kappa coefficient was calculated using Eqs. 15 and
16a, 16b as follows:
j

a1  a2
1  a2

16a

N
P

16b

where a1 is the overall accuracy, a2 is the percentage


of items that has been classified correctly by chance,
xij is the number of pixels in the ijth cell of the
contingence table, yi marginal total of row,
zi = marginal total of column, N = the total number
of pixels in the image under consideration.
Furthermore, the AE was computed by making
pixel-by-pixel comparison between the true models
and the combined images. With this end in view,
both MAE and MAPE were calculated. First, the
absolute error (AE) of the model features (i.e., for
the blocks and background) was computed. This was
carried out by estimating the difference in resistivity
between the true model and classified image on a
pixel-by-pixel basis. For all the pixels within the
blocks, for instance, MAE was estimated as the
average of all AEs. Also, the MAPE was estimated
for the blocks and background of the models
images. AE, MAE and MAPE are given in Eqs. 17,
18, 19 as follows:
A:E xi ; yi jqi  q^i j

17

N
1X
jq  q^i j
N i1 i

18

MAE

19

where qi is the true resistivity of the block(s) and q^i is


the calculated resistivity of the block; N is the total
number of pixels in the image.

4. Results and Discussion


4.1. Synthetic Results

and
y i zi
a2 i1 2
N


N 

1X
qi  q^i   100

N i1
qi 

15

N
P

xij
a1 i1
N

MAPE

A tabulated set of reconstructed resistivity values


for each block and background of the models in
Fig. 2 are presented in Table 3. The images
obtained using Eqs. 4, 5, 6 and 7 to merge the
resistivity values of the three individual electrode
arrays (i.e., the dipole-dipole, pole-dipole and Wenner-Schlumberger) are shown in Fig. 2dg. It is
observed from the table that for all the models, the
reconstructed resistivity values of the blocks are
underestimated compared to the resistivity of the
true models, while for the background the reconstructed resistivity values are an overestimation. In
essence, the amplitude of resistivity has been
dampened and is a known feature inherent in
resistivity imaging (RUCKER 2012). For the individual electrode arrays, it is observed that the dipoledipole images give better results than WennerSchlumberger and pole-dipole images. This means
that the reconstructed blocks resistivity for dipoledipole images is closer to the true models resistivity
than the resistivity for Wenner-Schlumberger or
pole-dipole images. Also, for the combined images,
the reconstructed resistivity of the maximum images
is higher than that of the remaining images (i.e.,
minimum, median or average). Thus, the maximum
outperforms any of the other combination as well as
the individual images in attempting to reproduce the
true models.
The classification results after the pixels (i.e.,
resistivity values) had been grouped into 16 clusters
by the classifier algorithm together with their mean
resistivity values are presented in Table 4. It is seen
that each cluster has three mean resistivities associated with it. These resistivity values show that each

K. S. Ishola et al.

Pure Appl. Geophys.

Figure 2
Two-dimensional inverse images of a block model for individual arrays: a Dpd, b Pdp, c Wsc and combined images, d max, e min, f med and
g avg

cluster contains information about the three electrode


configurations used. Furthermore, a gradual increase
in resistivity from one cluster to another is noted.
This gradual increase might be due to the electrical
contrast between the blocks and the background of
the models images as the classifier recognizes
certain patterns in resistivity that are grouped to a
particular cluster and other patterns to another cluster.
Also, the classification results of the combined
classified images obtained for the model of a single
block embedded in a homogeneous background are
shown in Fig. 3ad. The rest of the classified images
containing thematic information are shown in Figs. 4,
5, 6 and 7.
The aggregation of the clusters into classes shows
that, for a single block model, cluster 1 is assigned to

the background and clusters 216 to the block leading


to a model with two classes. Class 1 corresponds to
the background and class 2 to the block (Fig. 3). For
the two-block model, clusters 24 are assigned to the
first block (left), clusters 516 to the second block
(right) and cluster 1 to the background. As a result
three classes are produced, and each class corresponds to a particular feature of the model (Fig. 4).
The three-block model images are such that cluster 1
is assigned to the background, clusters 24 are
assigned to the left block, while clusters 5 through
10 are assigned to the middle block; clusters 1116
are assigned to the right block with a total of four
classes representing this model. Class 1 represents the
background, while classes 24 represent the blocks of
the model (Fig. 5). In Fig. 6, clusters 13 are

59

58

170 130

84

180

83

140

130

240 150 120

250

130

150

180

18

16

14

18

13

15

16

13

12

13

14

12

13

13

95

93

92

100

86

92

93

470 460 460


100 100 100

470
100

460
100

460
100

460
100

assigned to the conductive top layer, and clusters


416 are assigned to the resistivity bottom layer. As a
result, two classes are obtained for the fault model:
class 1 is typical of the top layer, and class 2
corresponds to the bottom layer of the model. In the
resistivity dyke model, clusters 216 are assigned to
the intrusive block, while cluster 1 belongs to the
background. This implies that only two classes are
desirable to represent these models. Again, class 1
represents the background, while class 2 represents
the intruding block (Fig. 7). In consequence, the
number of clusters aggregated/merged into a particular class depends on the structural features identified
in the models image.
The overall resistivity obtained after assigning the
mean resistivity values obtained in Table 3 to the
blocks and background of the classified images is
summarized in Table 5. From this table, two deductions could be drawn. For instance, for a one-block
model type (1) the resistivity values range from a
minimum of 180 X-m to a maximum of 340 X-m.
With this range, it shows that after classification of
the images, the overall resistivity values obtained for

430
340
220
560
500
410
310
180
140
110
120
140
470
480
480
420
320
210
520
470
380
280
170
130
110
120
130
440
460
470
380
290
200
480
440
350
250
160
130
110
120
120
420
430
430
Dyke

55

Fault

61

Three
blocks

53

Two
blocks

57

340
260
180
450
410
310
220
140
120
100
110
110
390
400
400

59

390
240
160
400
370
290
210
160
100
110
95
91
360
370
370

17

340
220
160
360
340
250
180
120
110
99
86
82
340
350
350

17

300
170
130
320
300
220
170
130
79
91
76
73
310
310
320

15

220
170
120
280
260
190
140
100
77
82
67
64
290
290
290

19

190
140
100
230
220
160
120
82
61
72
59
56
260
260
270

15

150
120
85
190
180
120
90
60
41
62
51
48
230
230
240

18

120
92
72
140
130
100
68
63
59
52
42
40
200
200
210

17

90
68
56
110
100
76
52
43
34
41
33
32
180
170
180

320

58
45
41
71
72
51
29
25
21
30
25
24
150
150
150

330

9
9
8
9
10
9
10
10
10
11
12
11
100
100
99

260

Dpd
Pdp
Wsc
Dpd
Pdp
Wsc
Dpd
Pdp
Wsc
Dpd
Pdp
Wsc
Dpd
Pdp
Wsc

370

One block

360 330 260

Cluster
14

110

Cluster
13

110

Cluster
12

89

Cluster
11

120

Cluster
10

87

Cluster
9

110 100

Cluster
8

250
12

Cluster
7

260
12

Cluster
6

160
13

Cluster
5

330
11

Cluster
4

320 250 160


14 13 12

Cluster
3

Block (500)
Background
(10)
Two
Block 1
blocks
(100)
Block 2
(300)
Background
(10)
Three
Block 1
blocks
(100)
Block 2
(300)
Block 3
(500)
Background
(10)
Fault
Hanging wall
(10)
Foot wall
(100)
Dyke
Block (500)
Background
(100)

Cluster
2

Dpd Pdp Wsc Rmax Rmin Rmed Ravg

Cluster
1

Resistivity
(X-m)

Model type Array


type

Reconstructed mean resistivity (X-m)

Table 4

One
block

True model

Summary of clustering for synthetic models with mean resistivity in X-m

Model
type

Cluster
15

Summary of reconstructed resistivity for the different models


without classification

29
24
27
37
41
29
10
10
11
19
17
16
120
120
130

Cluster
16

Table 3

460
360
220
590
530
440
340
140
140
110
130
140
500
490
490

Combining Multiple Electrode Arrays

K. S. Ishola et al.

Pure Appl. Geophys.

Figure 3
Two-dimensional integrated classified images for a resistivity block model: a max, b min, c med and d avg

the maximum images are closer to the true models


resistivity (i.e., 500 X-m) than the resistivity values
obtained for the rest of the models images. Also, the
rest of the models images show similar trends, and
(2) in comparison with the models images obtained
without classification, it is noted that there is an
increase in resistivity values of the combined classified images resulting from the unsupervised
technique. This could suggest that the unsupervised
classification technique finds patterns or features
residing in the original models that might not be
obviously known by any of the individual electrode
array configurations before the unsupervised classification technique is utilized (ERNENWEIN 2009; RANA
et al. 2010).

Furthermore, the error matrices generated for


statistical evaluation of the performance of the
classified images relative to the true models are
presented in Tables 6, 7, 8, 9 and 10. PA indicates the
degree to which the reference pixels are classified
(LILLESAND and KEIFER 2000). UA indicates the
probability that a given cluster is actually present
on the ground (LILLESAND and KEIFER 2000). Generally, both the users and producers accuracies
indicate good classification of the models as the
errors of commission and omission during the
classification were less than 40 %. The accuracy
assessment from the unsupervised classification technique further shows that the overall percentage
accuracy is greater than 70 %. Overall accuracy

Combining Multiple Electrode Arrays

Figure 4
Two-dimensional integrated classified images for the resistivity two-block model: a max, b min, c med and d avg

evaluates the percentage of cases correctly classified,


i.e., the probability that a pixel is classified correctly
in the classified image. This indicates a good
agreement between the classified images and the true
model images used as reference data. According to
LANDIS and KOCH (1977), the ranking of the kappa
coefficient (j) ranges from 1 to 1. As a result, three
groups can be obtained: (1) j greater than 0.8
represents strong agreement between the classification and the true models images; (2) j between 0.4
and 0.8 represents moderate agreement; (3) j less
than 0.4 represents poor agreement between the

classified and the true model images. By this


standard, j shows that for the one-, two- and threeblock images, substantial agreement exists between
the classified images and true models with kappa
accuracy from 76 to 83 %, while for the fault and
dyke models, the j values are 90 and 93 %,
respectively. This implies that there is perfect agreement between the classified and true model. Thus, j
compensates for chance agreement in the classification and provides a measure of how much better the
classification performs in the probability of randomly
assigning pixels to their correct classes.

K. S. Ishola et al.

Pure Appl. Geophys.

Figure 5
Two-dimensional integrated classified images for the resistivity three-block model: a max, b min, c med and d avg

In addition, an assessment of the accuracy of the


combined images, employing both MAE and MAPE,
shows that for all the synthetic models, the maximum
resistivity values assigned to the models give the least
errors, as presented in Tables 11 and 12, respectively.
Thus, the maximum approach is considered the best
representative of the true models as it reproduces the
true model with the least possible errors compared to
the other models images. Although our focus in this
article was on classified images, more details on the
quality of images obtained without clustering in
terms of MAE and MAPE for the synthetic models
can be found in ISHOLA et al. (2014).

4.2. Application to Field Examples


As feasibility tests of the application of the
aforementioned unsupervised classification technique on synthetic data, we also extended the
clustering technique to real field data. In field
example 1, electrical resistivity imaging was conducted to delineate subsurface lithological units that
control the hydrogeology of the study area for
groundwater resources development. Also, in field
example 2, electrical resistivity imaging was applied
to demarcate regions of leachate infiltration in a
septic field.

Combining Multiple Electrode Arrays

Figure 6
Two-dimensional integrated classified images for a resistivity fault model: a max, b min, c med and d avg

4.2.1 Field Example 1


A 2D electrical resistivity survey was performed at
Universiti Sains Malaysia Engineering campus, Nibong Tebal, in Penang, Malaysia. The ABEM Lund
Imaging system comprising a Terrameter SAS 4000
supplemented with an automatic multielectrode
selector ES 10-64C system was used in the acquisition of electrical resistivity data. A resistivity profile
of about 420 m was established. Along this profile,
four cable reels, each 100 m long and trending in the
west-east direction, were used. Sixty-two takeouts
electrodes were deployed at 5-m spacing for the two
inner cable sets (100300 m), i.e., short (S) and 10-m
electrode spacing for the outer cable sets, i.e., long

(L) corresponded to distances of 0100 m and


300400 m on the standard 400 m (ADIAT et al.
2013). Two cable joints were used to link the cables to
ensure continuity. The first cable joint was used to
connect cables 1 and 2, while the second cable joint
connected cables 3 and 4. The other terminals of
cables 2 and 3 were connected to the Terrameter used
for the measurements. The three electrode array
configurations used were the dipole-dipole (LS),
pole-dipole (LS) and Schlumberger (LS). A forward
modeling subroutine was used to calculate the apparent resistivity values, and a nonlinear least-squares
optimization technique was used for the inversion.
Like in the synthetic example, the RES2DINV
program was employed. In the dipole-dipole array,

K. S. Ishola et al.

Pure Appl. Geophys.

Figure 7
Two-dimensional integrated classified images for a resistivity dyke model: a max, b min, c med and d avg

with a minimum electrode spacing of 5 m, a total of


770 data points were used for measurements of
resistivity (i.e., 414 data points for L, 356 data points
for S), for the pole-dipole array 792 data points (516
for L, 276 for S) and for the Schlumberger array 748
data points (418 for L, 330 S). For all three arrays,
after the seven iterations, the inversion converged
with misfit error less than 20 %. The misfit errors were
for the dipole-dipole (17 %), pole-dipole (13 %) and
Schlumberger (10 %) arrays. Inversion results for the
noise-contaminated data sets were saved in text file
format in order to allow for further manipulation and
analysis using the image processing program package
of the PCI Geomatics 10.3 software.

The reconstructed tomographic images of the


three electrode array configurations are shown in
Fig. 8ac. From a geophysical viewpoint, the geological structures delineated behave like a three-layer
model for all the three arrays. With the dipole-dipole
array, the inversion image (upper panel) shows that a
low resistivity layer is detected close to the surface.
The low resistivity zone has a surface layer thickness
between 3 and 10 m. This zone is underlain by a layer
with higher resistivity, while the third layer, which
seems to extend throughout the model, is characterized by an intermediate resistivity. Using the poledipole array, the inverse image (middle panel) reveals
that the first layer is a low-resistivity pocket-like

Combining Multiple Electrode Arrays

Table 5
Summary of overall resistivity for classified images
Model type

One block
Two blocks

Three blocks

Fault
Dyke

Model description
(X-m)
Block 1 (500)
Background (10)
Block 1 (100)
Block 2 (500)
Background (10)
Block 1 (100)
Block 2 (300)
Block 3 (500)
Background (10)
Hanging wall (10)
Footwall (100)
Block 1 (500)
Background (100)

Overall resistivity (X-m)


Rc max

Rc min

Rc med

Rc avg

340
14
130
380
19
74
200
250
12
14
100
480
100

180
11
92
260
15
59
90
150
5
11
86
460
100

260
12
120
330
17
67
140
160
8
12
93
470
100

260
13
110
320
16
67
150
190
8
13
93
470
100

Table 6
Error matrix/contingency table for a one-block classified image on
a pixel-by-pixel basis
j = 0.80

Block 1

Background

Row total

UA

Block1
Background
Column total
PA

108,658
51,351
160,009
0.68

6,568
751,003
757,571
0.99

115,226
802,354
917,580

0.94
0.99
O.A = 0.94

Table 7
Error matrix/contingency table for two-block classified image on a
pixel-by-pixel basis
j = 0.76

Block
1

Block
2

Background Row
Total

Block 1
Block 2
Background
Column
total
PA

54,965
0
2,035
57,000

11,775
56,136
89
68,000

37,360
10,298
690,060
737,718

0.96

0.83

0.94

UA

104,100 0.53
66,434 0.85
692,184 0.99
862,718
O.A = 0.93

structure haloed by a more resistive layer that extends


downward to a depth of about 60 m. This layer
appears to be separated by a less conductive layer
with depth from 15 to 40 m. The Schlumberger array
image (lower panel) shows that the first layer is a
conductive zone that extends through the model. It
surrounds a more conductive pocket-like structure.
Also, a high-resistivity structure of finite extent is

observed, while the deeper portion of the inverse


image is a zone that corresponds with an intermediate
resistivity. For resistivity imaging, the difference in
the imaging abilities of the three electrode arrays
when applied to the geological model is underscored.
The difference in the imaging capability is controlled
by the array parameters electrode spacing (a) and
propagation factor (n) (DAHLIN and ZHOU 2004). We
should also mention that the differing effects of the
electrode arrays on the 2D inverse images are due to
the arrangement of the electrodes, spreading patterns
and density of the data.
The aforementioned electrical imaging emphasizes the nonuniqueness of the inversion process
(TARANTOLA and VALETTE 1982; NARAYAN et al. 1994).
To reduce the ambiguity in the tomographic images,
the three 2D inverse resistivity images were combined using an unsupervised image classification
technique. The three images in Fig. 8ac were used
as inputs to the k-means algorithm alongside other
initial parameters used in the synthetic modeling
section. Also, like the synthetic examples, the number
of clusters into which the combined images should be
partitioned was set at 16, while the number of
iterations and convergence threshold where 20 and
0.01, respectively. Again, the prior knowledge of the
inversion results for these models guided the choice
of the number of clusters used. The clustering result
is presented in Table 13. Following the assignment of
resistivity values to the clusters, the clusters were
merged into six desirable classes. The resulting
classification image with a superimposed borehole
lithologic log for the assignment of lithologies is
shown in Fig. 9.
Cluster 1 was assigned to class 1; clusters 2 and 3
were merged to class 2; clusters 46 were merged to
class 3. Also, clusters 7 and 8 were merged to class 4,
and clusters 912 were merged to class 5, while
clusters 1316 were assigned to class 6. This
aggregation/merger of the clusters corresponded with
the subsurface geological units of the study area.
Though the unavailability of resistivity log data (i.e.,
as ground truth) made it impossible to evaluate the
accuracy of the classified image using the error
matrix table as in the synthetic examples, nevertheless a qualitative approach involving the available
borehole lithologic log was used. The electrical

K. S. Ishola et al.

Pure Appl. Geophys.

Table 8
Error matrix/contingency table for the three-block classification model on a pixel-by-pixel basis
j = 0.83

Block 1

Block 2

Block 3

Background

Row total

UA

Block 1
Block 2
Block 3
Background
Column total
PA

28,381
0
0
1,619
30,000
0.95

18,562
20,475
2,892
571
42,500
0.48

10,989
25,256
7,255
0
43,500
0.17

55,014
2,260
0
689,447
746,721
0.92

112,946
47,991
10,147
691,637
862,721

0.25
0.43
0.71
0.99
O.A = 0.86

Table 9

Table 12

Error matrix/contingency table for a fault classified image on a


pixel-by-pixel basis

Summary of mean absolute percentage error for the models


classified images

j = 0.90

Top layer Bottom layer Row total UA

Top layer
Bottom layer
Column total
PA

37,408
2,311
39,719
0.94

35,050
451,263
486,313
0.93

72,458
453,574
526,032

Model type

0.52
0.99
One block
O.A = 0.93
Two blocks

Table 10
Error matrix/contingency table for a dyke classified image on a
pixel-by-pixel basis

Three blocks

j = 0.93

Block 1

Background

Row total

UA

Fault

Block 1
Background
Column total
PA

108,658
3,383
112,041
0.96

0
751,003
751,003
1.0

108,658
754,386
863,044

1.0
0.99

Dyke

Summary of mean absolute error for classified images

One block

Two blocks

Three blocks

Fault
Dyke

MAPE (X-m)

Description

Max

Min

Med

Avg

Block 1
Background
Block 1
Block 2
Background
Block 1
Block 2
Block 3
Background
Hanging wall
Foot wall
Block 1
Background

37
11
29
34
25
56
49
56
84
32
21
6.4
4.4

67
33
21
48
65
71
73
78
100
17
19
6.6
4.3

51
41
25
35
89
63
64
72
97
26
23
6.6
4.3

52
42
24
37
82
63
62
69
98
28
20
6.1
3.4

O.A = 0.99

Table 11

Model type

Model

Model

MAE (X-m)

Description

Max

Min

Med

Avg

Block 1
Background
Block 1
Block 2
Background
Block 1
Block 2
Block 3
Background
Hanging wall
Foot wall
Block 1
Background

160
4.2
29
160
8.5
56
150
280
8.5
3.2
21
31
4.4

320
3.3
21
230
6.5
71
220
390
11
1.7
19
34
4.3

240
4.1
25
160
8.9
63
190
360
9.7
23
2.6
32
4.3

240
4.2
24
180
8.2
63
190
340
9.8
20
2.8
31
3.4

contrast in the geologic units delineated, on corroboration with the representative borehole lithologic log
information, showed that the geology of the area
consists of clay, sandy clay, clayey sand and sand.
The range of resistivity values to the depth of
investigation was between 5 and 55 X-m. The
relationship that links the various classes with the
litho-resistivity values for the subsurface characterization of the study area is summarized in Table 12.
A conductive anomalous body that presumes to be an
iron material was identified as class 1. This material
is buried in a less conductive layer that corresponds
with clay (class 2). Class 3 corresponds with sandy
clay. Because it is less cohesive, the groundwater
potential of this layer might be low. The promising
zones for probable groundwater exploitation in the
area are clayey sand (i.e., class 4) and sand units
belonging to classes 5 and 6. In general, the

Combining Multiple Electrode Arrays

Figure 8
Two-dimensional inverse resistivity models for field test 1: a dipole-dipole (Dpd), b pole-dipole (Pdp) and c Wenner-Schlumberger (Wsc)
arrays. Arrow showing the borehole (BH) position

classification image showed prominent clusters/classes patterns for the subsurface features delineated
with a clay layer overlying the aquifer units (i.e.,
clayey sand and sand sediments) (Table 14).
4.2.2 Field Example 2
In the second site, the electrical resistivity survey was
carried out in the center of the Tucson basin,
southeastern Arizona, USA. The study area lies
within a semiarid environment. The subsurface
geology of the area consists mainly of clay, sand
and gravels. The resistivity survey was conducted
over a septic leach field using the Superstring R8

(Advanced Geosciences, Inc.) resistivity instrument.


A total of 27 electrodes were used, at electrode
spacing of 1 m, in order to maximize the required
depth of the investigation (RUCKER et al. 2011a, b).
The three electrode array configurations employed
were the dipole-dipole, gradient and Schlumberger.
Resistivity measurements were carried out with a
minimum surface electrode spacing of 1 m. A total of
237 data points were used for the dipole-dipole array,
604 data points for the gradient and 169 data points
for the Schlumberger array. Also, the inversions of
the apparent resistivity data were carried out using
RE2DINV software. For the dipole-dipole array, after
the eight iterations the inversion converged with an

19
13
38
19
13
4.9

3.2
4.3
3.4

Figure 9
A 2D classified image with six classes for the field test 1

20
12
17

Cluster
15

Table 14

54
12
22

Cluster
12

Pure Appl. Geophys.

55
11
22
54
9.9
23
27
9.3
46
3.2
4.3
3.4
Dpd
Pdp
Sch

11
7.4
7.6

33
7.2
8.3

34
7.9
53

41
7.2
17

38
7.6
26

35
8.1
35

31
8.7
42

Cluster
10
Cluster
1

Cluster
2

Cluster
3

Cluster
4

Cluster
5

Cluster
6

Cluster
7

Cluster
8

Cluster
9

Cluster
11

Classes and litho-resistivity for the field test 1

Electrode array
type

Summary of clustering results with mean resistivity values in X-m for field test 1

Table 13

Cluster
13

Cluster
14

Cluster
16

K. S. Ishola et al.

Classes

Probable lithology

Resistivity
range (X-m)

1
2
3
4
5
6

Highly conductive (buried iron material)


Low-resistivity topsoil (clay)
Sandy clay (low yield)
Clayey sand (medium yield)
Sand (high yield)
Sand

15
813
1825
2830
3550
55

absolute error of 2.1 %. The misfit errors for the


gradient and Schlumberger arrays are 1.0 % and
1.5 %, respectively, after six iterations.
The inverted earth resistivity models of the three
electrode arrays reproduced by the image processing
software are shown in Fig. 10ac. It is observed that
the images of the three electrode array configurations
show substantial differences in the earth models
generated from data acquired by each of the electrode
arrays over the same subsurface conditions. Using the
dipole-dipole array, the inverted image (upper panel)
shows a structure with an intermediate resistivity
close to the ground surface overlying a low resistivity
material. This layer, however, halos an anomalous
body with high resistivity located at a depth between
1.5 and 3.0 m in the section. However, the Schlumberger array inverted earth model (middle panel)
shows that a low-resistivity material (blue) occupies
more of the section and extends to the ground surface
than in the dipole-dipole model. For the gradient
array (lower panel), the inverted data indicate a
model that has more homogeneous layers at the
surface and a deeper zone of the section. Also, two
zones of high-resistivity values overlying a low-

Combining Multiple Electrode Arrays

Figure 10
Two-dimensional inverse resistivity models for field test 2 for a dipole-dipole, b Schlumberger and c gradient arrays

resistivity zone are observed. Unlike the images of


the dipole-dipole and Schlumberger, the low-resistivity zone is delineated between the 3rd and 19th
electrode positions and from a depth of 3.5 m to the
extent of investigation.
Again, these tomographic images were used as
input parameters for the k-means algorithm for
clustering and classification campaigns. The primary
purpose of the clustering and classification is to
produce a unified classification image for the three
input images. The cluster results showing 16 clusters
with their mean resistivity values are presented in
Table 15. After clustering, the combined image that
was obtained following the merging some of the
clusters into five classes is shown in Fig. 11. The
classified image clearly shows the regions contamination by the leachate infiltration and the
uncontaminated layers with the classes corresponding

overwhelmingly to the local geology of the survey


area (Table 16). The most important observation is
that of a low earth resistivity (i.e., highly conductive
layer), which might be a manifestation of leachate
contamination. This conductive zone of the layered
earth corresponds to cluster 1 with resistivity values
ranging from 5 to 20 X-m. It extends up to 6 m from
the surface. However, caution was exercised in its
interpretation as clay units are also conductive and
may be misinterpreted as leachate contamination
(CARPENTER et al. 2012). The high-resistivity layers
represent zones free of contamination. Cluster 2 is a
region overlying the leachate infiltrated zone. The
resistivity is between 30 and 90 X-m (clay); in cluster
3, it is 100210 X-m (clayey sand); in cluster 4,
220350 X-m (sand); in cluster 5, 360400 X-m
(sand); in clusters 68, 410550 X-m (sand); in
clusters 916, 5801,000 X-m (gravely sand).

2,600
400
420
2,400
420
410
2,300
410
390
2,100
400
380

Cluster
16
Cluster
15
Cluster
14
Cluster
13

1,900
390
370
1,700
380
350
1,300
350
330
10
10
13
Dpd
Grd
Sch

94
90
87

160
180
170

420
380
520

240
230
230

530
280
270

810
330
330

1,100
340
330

1,500
370
350

Cluster
12
Cluster
11
Cluster
10
Cluster
9
Cluster
8
Cluster
7
Cluster
6
Cluster
5
Cluster
4
Cluster
3
Cluster
1

Cluster
2

Pure Appl. Geophys.

Figure 11
A 2D classified image with five classes for field test 2

Table 16

Electrode array
type

Summary of clustering results with mean resistivity values in X-m for field test 2

Table 15

K. S. Ishola et al.

Classes and litho-resistivity for field test 2


Classes

Probable lithology

Resistivity range (X-m)

1
2
3
4
5

Conductive (leachate)
Clay
Clayey sand
Sand
Gravelly sand

520
3090
100210
220550
5801,000

However, clusters 13 remain as individual classes


13, respectively. Clusters 48 were merged as class
4, and clusters 916 were merged as class 5. The
classes as well as the corresponding litho-resistivity
relationship established are summarized in Table 13.

5. Conclusions
In this study, we investigated the electrical imaging capabilities of different electrode array
configurations, taking advantage of the data sets from
both synthetic and real field examples. The use of a
clustering technique provides a unique opportunity to
improve our understanding of the subsurface condition characterization as the differing information from
the various electrode arrays is combined into a single
image in an automated way. By this approach, human
intervention or involvement is efficiently minimized
in the processing and interpretation of the images.
Also, the important results of our investigation
showed that for the accuracy of the classified images,
the estimated errors for both MAE and MAPE, the
maximum images provide a good representation of the
models. Furthermore, the overall accuracy and kappa
coefficients are indexes of good classification agreement between the classified and reference images. In
this way, the results obtained from this study show
that the integration of 2D electrical images by the

Combining Multiple Electrode Arrays

unsupervised image classification technique via the


k-means clustering algorithm can provide an inexpensive, efficient and reliable method for
characterization of the subsurface geology. Although
the proposed approach is applied only to electrical
resistivity data, there are still many open opportunities
for further studies or works based on the application of
the proposed technique in other geophysical methods.
The approach could be adopted where there are
challenges of de-risking the ambiguities in the interpretation of geophysical images from different
techniques. In addition, future works will include
more synthetic models, especially conductive blocks
embedded in a resistive background. Also, other
clustering algorithms and weighted clustering optimization algorithms are proposed for future research.
Acknowledgments
We would like to acknowledge hydroGEOHYSICS,
Inc., Tucson, AZ, USA, for providing part of the data
used in this work and Universiti Sains Malaysia
(USM) for funding this research under the short-term
research Grant No. 304/PFIZIK/6310057. The first
author expresses appreciation to University of Lagos,
Akoka, Nigeria for supports. We thank the associate
editor and anonymous reviewers for their comments,
which improved the quality of the manuscript
significantly.
REFERENCES
ABER, H., and MESHIN CHIASL, M. (2010). Present a Proper Pattern
for Choose Best Electrode Array based on Geological Structure
Investigating in Geoelectrical Tomography, in order to get
highest Resolution Image of the Subsurface. 1st International
Applied Geological Congress, Iran, 18031809.
ADIAT, K.A.N., NAWAWI, M.N.M., ABDULLAH, K., ISHOLA, K.S., and
ABDURAHMAN, A. (2013). Effects of electrode spacing and
inversion techniques on the efficacy of 2-D resistivity imaging to
delineate subsurface features. American Journal of Applied
Sciences, 10(1), 6472. doi:10.3844/ajassp.2013.64.72.
ATHANASIOU, E.N., TSOURLOS, P.I., PAPAZACHOS, C.N., and TSOKAS,
G.N. (2007). Combined weighted inversion of electrical resistivity data arising from different array types. Journal of Applied
Geophysics, 62, 124140.
BARKER, R. D. (1989). Depth of investigation of collinear symmetrical four electrode arrays. Geophysics. 54, 10311037.
BENTLEY, L. R., and GHARIBI, M. (2004). Two and three dimensional
electrical resistivity imaging at a heterogeneous remediation
site. Geophysics, 69(3), 674680, doi:10.1190/1.1759453.

BERGE, M. A. and DRAHOR, M. G. (2009). ArcheoSciences. Optimum Electrical Resistivity Tomography (OERT) Approach using
combination of different arrays in archaeological investigations.
Suppl. 33, 263265.
BLOME, M., MAURER, H., and GREENHALGH, S. 2011. Geoelectric
experimental design-Efficient acquisition and exploitation of
complete pole-bipole data sets. Geophysics, 76(1), F15F26.
BOSCHETTI, L., FLASSE, S.P., and BRIVIO, P.A. (2004). Analysis of the
conflict between omission and commission in low spatial resolution dichotomic thematic products: The Pareto Boundary.
Remote Sensing of Environment, 91, 280292.
CAMPBELL, J.B. (2002). Introduction to Remote Sensing, 3rd Edition. Guilford Press, New York. 621630.
CANDANSAYAR, M.E. (2008). Two-dimensional individual and joint
inversion of three and four Electrode array dc resistivity data.
Journal of Geophysics and Engineering. 5, 290300. doi:10.
1088/1742-2132/5/3/005.
CANDANSAYAR, M. E., and BASOKUR, A. T. (2001). Detecting smallscale targets by the 2D inversion of two-sided three-electrode
data: application to an archaeological survey Geophys. Prospect, 49, 1325.
CAPIZZI, P., COSENTIO, P.L., FIANDACA, G., MARTORANA, R., MESSINA,
P., and VASSALLO, S. (2007). Geophysical investigations at the
Himera archaeological site, northern Sicily. Near Surface Geophysics. 5, 417426.
CHATURVEDI, P., and PLUMB, R.G. (1995). Electromagnetic imaging
of underground targets using constrained optimization. IEEE
Transactions on Geosciences and Remote Sensing. 33, 551561.
CARPENTER, P. J., DING, A., and CHENG, L. (2012). Identifying
Groundwater Contamination using Resistivity Surveys at a
Landfill near Maoming, China. Nature Education Knowledge. 3
(7), 2026.
CHAMBERS, J., OGILVY, R., MELDRUM, P., and NISSEN, J. (1999). 3-D
resistivity imaging of buried oil and tar contaminated waste
deposits. European Journal of Environmental and Engineering
Geophysics. 4, 315.
CHENG, Y.M. (2003). K-means: A new generalized k-means clustering algorithm. Pattern Recognition Letters, 24, 28832893.
CIHLAR, J., XIAO, Q., BEAUBIEN, J., FUNG, K. and LATIFOVIC, R.
(1998). Classification by progressive generalization: A new
automated methodology for remote sensing multichannel data.
International Journal of Remote Sensing. 19(14), 26852704.
doi:10.1080/014311698214451.
COSCIA, I., MARESCOT, L., MAURER, H., GREENHAULGH, S., and LINDE,
N. (2008). Experimental Design for Crosshole Electrical Resistivity Tomography Data Sets. 14th European Meeting of
Environmental and Engineering Geophysics. EAGE.
CONGALTON, R.G. (1991). A review of Assessing the Accuracy of
Classifications of Remotely sensed data. Remote Sensing Environment. 37, 3546.
CONGALTON, R. G. and GREEN, K. (1999). Assessing the accuracy of
remotely sensed data: Principles and Practices. Lewis Publisher,
Boca Raton.
DAHLIN, T. (1996). 2-D Resistvity surveying for environmental and
engineering applications. First Break. 14, 275283.
DAHLIN, T and LOKE, M.H. (1997). Quasi-3D resistivity imagingmapping 3D structures using 2D DC resistivity techniques.
Proceedings of the 3rd Meeting of the Environmental Engineering Geophysical Society, 143146.
DAHLIN, T. (2001). The development of DC resistivity imaging
techniques. Computers and Geosciences. 27(9), 10191029.

K. S. Ishola et al.
DAHLIN, T., and ZHOU, B. (2004). A numerical comparison of 2-D
resistivity imaging with 10 electrode arrays. Geophysical Prospecting. 52(5), 379398, doi:10.1111/j.1365-2478.2004.00423.x.
DEGROOT-HEDLIN, C., CONSTABLE, S.(1990). Occams inversion to
generate smooth, two-dimensional models from magnetotelluric
data. Geophysics, 55, 16131624.
DE LA VEGA, M., OSELLA, A., and LASCANO, E. (2003). Joint
inversion of Wenner and dipole-dipole data to study a gasoline
contaminated soil. Journal of Applied Geophysics, 54, 97109.
DOETSCH, J.A. COSCIA, I., GREENHALGH, S., LINDE, N., GREEN, A., and
GUNTHER, T. (2010). The borehole-fluid effect in electrical
resistivity imaging. GEOPHYSICS, 75, 4 F107F114,
EASTMAN, J.R. (1995). IDRIS for Windows, Users Guide, Clark
University, Worcester, MA, USA, 405408.
EASTMAN, J. R. (2003). Guide to GIS and image processing. 14,
239247. Clark University manual, USA.
ELLIS, R.G., and OLDENBURG, D.W. (1994). Applied geophysical
inversion. Geophysical Journal International, 116(1), 511, doi:
10.1111/j.1365-246X.1994.tb02122.x.
EL-QADY, G., HAFEZ, M., ABDALLAH, M.A., and USHIJIMA, K. (2005).
Imaging subsurface cavities using geoelectric tomography and
ground-penetrating radar. Journal of Cave and Karst Studies,
67(3), 174181.
ENDERLE, D.I.M., and WEIH, R.C. (2005). Integrating supervised
and unsupervised classification methods to develop a more
accurate land cover classification. Journal of the Arkanas,
Academy of Science. 59, 6573.
ERNENWEIN, E.G. (2009). Integration of multidimensional archaeogeophysical data using supervised and unsupervised
classifications. Near surface Geophysics. 7, 145158.
FARQUHARSON, C.G., and OLDENBURG, D.W. (2004). A comparison of
automatic techniques for estimating the regularization parameter
in non-linear inverse problems. Geophys. J. Int. 156, 411425.
FIANDACA, G., MARTORANA, R. and COSENTINO, P.L. (2005). Use of
the linear grid array in 2D resistivity tomography Proc. Near
Surface 11th European Meeting of Environmental and Engineering Geophysics.
FIGUEIREDO, M. and JAIN, A. K. (2002). Unsupervised learning of
finite mixture models. IEEE Transaction. Pattern Analysis and
machine intelligence, 24(3), 381396.
FURMAN, A., FERRE, P. A., and WARRICK, A.W. (2003). A Sensitivity Analysis of Electrical Resistivity Tomography Array Types
Using Analytical Element Modeling. Vadose Zone Journal. 2,
416423.
GALLARDO, L.A. and MEJU, M.A. (2003). Characterization of heterogeneous near-surface materials by joint 2D inversion of dc
resistivity and seismic data, Geophs. Res. Lett., 30, 1658, doi:10.
1029/2003GL017370.
GALLARDO, L.A. and MEJU, M.A. (2004). Joint two-dimensional DC
resistivity and seismic travel time inversion with cross-gradients
constraints, J. Geophys. Res-Solid Earth, 109, B03311, doi:10.
1029/2003JB002716.
GALLARDO, L.A. and MEJU, M.A. (2007). Joint two-dimensional
cross-gradient imaging of magnetotelluric and seismic traveltime
data for structural and lithological classification. Geophys.
J. Int. 169, 12611272.
GAO, J. (2009). Digital Analysis of Remotely sensed Imagery.
McGraw Hill Professional, New York.
GEOTOMO. (2005). Geotomo softwares, Minden Heights, Pinang,
Malaysia.

Pure Appl. Geophys.


GRIFFITHS, D.H., and BARKER, R.D. (1993). Two dimensional
resistivity imaging and modeling in areas of complex geology.
Journal of Applied Geophysics. 29, 211226.
HABER, E., and OLDENBURG, D. (1997). Joint inversion: a structural
approach. Inverse Problems, 13, 6377, doi:10.1088/0266-5611/
13/1/006.
HABER, E., and OLDENBURG, D. (2001). A GCV based methods for
nonlinear inverse problem. Comput. Geosci. 4, 4163.
HAGREY, S. A. AL and PETERSEN, T. (2011). Numerical and
experimental mapping of small root zones using optimized surface and borehole resistivity tomography. Geophysics 76(2),
G25G35.
HAGREY, S. A. AL (2012). 2-D Optimized Electrode Arrays for
Borehole Resistivity Tomography and CO2 Sequestration Modelling. Pure and Appl. Geophys. 169, 12831292.
HALKIDI, M., BATISTAKIS, Y., and VAZIRGIANNIS, M. (2001). On
Clustering Validation Techniques. Journal of Intelligent Information Systems, 17: 2/3, 107145.
HASMADI MOHD, I., PAKHIAZAD, H.Z., and SHAHRIN, M.F. (2009).
Evaluating supervised and unsupervised techniques for land
cover mapping using remote sensing data. Malaysian Journal of
Society and Space. 5(1), 110.
IBRAHIM, A.N., HARITH, Z.Z.T. and NAWAWI, M.N.M. (2003). Using
integrated 2-D Electrical imaging and seismic refraction methods
for study of groundwater aquifer in Brookland Plantation, Selangor, Malaysia, 6th SEG Journal International Symposium
Imaging Technology, Tokyo, Japan, 398404.
ISHOLA, K.S., NAWAWI, M.N.M., ABDULLAH, K., SABRI, A.I.A., and
ADIAT, K.A.N. (2014). Assessment of the reliability of reproducing two-dimensional resistivity models using an image
processing technique. Springerplus, 3:214, 112.
KHAN, S.S., and AHMAD, A. (2004). Cluster center initialization
algorithm for K-means clustering. Pattern Recognition Letters,
25(11), 1293-1302,
KVAMME, K. L. (2006). Integrating multidimensional geophysical
data. Archaeological Prospection. 13, 5772, doi:10.1002/arp.268.
LANDIS, J. R. and KOCH, G.G. (1977). The measurement of observer
agreement for categorical data, Biometrics. 33, 159174.
LI, Z., LIU, X., MA,T., KEJIA, D., ZHOU, Q.,YAO, B., and NIU, T.
(2013). Retrieval of the surface evapotranspiration patterns in
the alpine grasslandwetland ecosystem applying SEBAL model
in the sourceregion of the Yellow River China. Ecological
Modelling 270, 6475.
LILLESAND, T. M., and KIEFER, R.W. (1994). Remote Sensing and
Image Interpretation. (3rd Ed.) Wiley, New York.
LILLESAND, T.M. and KEIFER, R.W. (2000). Remote Sensing and
Image Interpretation, 4th Edition, John Wiley and Sons, New
York, 585608.
LINDE, N., and DOETSCH, J.A. (2010). Joint inversion of crosshole
GPR and seismic traveltime data in R. D. Miller, J. H. Bradford,
and K. Holliger, eds., Advances in nearsurface seismology and
ground-penetrating Radar: SEG Geophysical Developments
Series 15, 116.
LINDE, N., FINSTERLE, S., and HUBBARD, S. (2006). Inversion of
tracer test data using tomographic constraints: Water Resources
Research, 42, W04410, doi:10.1029/2004WR003806.
LIU, J.Y., ZHUANG, D.F., LUO, D. and XIAO, X. (2003). Land-cover
classification of China: integrated analysis of AVHRR imagery
and geophysical data. Int. Journal of Remote Sensing. 24(12),
24852500.

Combining Multiple Electrode Arrays


LOKE, M.H. and BARKER, R. D. (1996). Rapid least squares inversion of apparent resistivity pseudo-sections by a quasi-Newton
method. Geophysical Prospecting. 44, 131152.
LOKE, M.H. (2001). Tutorial: 2-D and 3-D electrical imaging surveys. Course Notes for USGS Workshop 2-D and 3-D
Inversion and Modeling of Surface and Borehole Resistivity
Data, Storrs, CT, 1316.
LOKE, M.H., ACWORTH, I., and DAHLIN, T. (2003). A comparison of
smooth and blocky inversion methods in 2D electrical imaging
surveys. Exploration Geophysics. 34, 182187.
LOKE, M.H. (2009). Tutorial: 2-D and 3-D Electrical Imaging
Surveys.
(http://www-geo.phys.ualberta.ca/*unsworth/UAclasses/223/loke_course_notes.pdf).
LOKE, M.H., WILKINSON, P.B., and CHAMBERS, J.E. (2010). Fast
computation of optimized electrode arrays for 2D resistivity
surveys. Computers and Geosciences. 36, 14141426.
LOKE, M.H., CHAMBERS, J.E., RUCKER, D.F., KURAS, O., and WILKINSON, P.B. (2013). New developments in the direct-current
geoelectrical imaging method. Journal of Applied Geophysics
95, 135156.
MACQUEEN, J. B. (1967). Some methods for classification and
analysis of multivariate observations Proceedings of the Fifth
Symposium on Math, Statistics, and Probability, 281297.
MARTORANA, R., FIANDACA, G., CASAS, P. A., and COSENTINO, P. L.
(2009). Comparative tests on different muti-electrode arrays
using models in near-surface geophysics. Journal Geophysical
Engineering. 6, 120.
MEJU, M.A. (1994). Biased Estimation: A Simple Framework for
Inversion and Uncertainty Analysis with Prior Information.
Geophys. J. Int. 119, 521528.
MELGANI, F., and BRUZZONE, L. (2004). Classification of Hyperspectral Remote Sensing Images with Support Vector Machines.
IEEE Transactions on geosciences and remote sensing, 42(8),
17781790.
MILLIGAN, G.W and COOPER, M.C. (1985). An examination of
procedures for determining the number of clusters in a data set.
Psychometrika. 50(2), 159179.
NARAYAN, S., DUSSEAULT, B., and NOBES, D.C. (1994). Inversion
techniques applied to resistivity inverse problems. Inverse
Problems, 10, 669686.
NEYAMADPOUR, A., WAN ABDULLAH, W.A.T., TAIB, S., and NEYAMADPOUR, B. (2010). Comparison of Wenner and dipole-dipole
arrays in the study of an underground three-dimensional cavity.
Journal of Geophysics and Engineering, 7(1), 3040. doi:10.
1088/1742-2132/7/1/003.
OLAYINKA, A. I., and YARAMANCI, U. (2000). Assessment of the
reliability of 2D inversion of apparent resistivity data: Geophysical Prospecting, 48, 293316.
OLDENBURG, D.W., and ELLIS, R.G. (1991). Inversion of geophysical
data using an approximate inverse mapping. Geophys. J. Int.
105, 325353.
OLDENBURG, D.W. and LI, Y.G. (1999). Estimating depth of
investigation in dc resistivity and IP surveys. Geophysics, 64,
403416.
PERUMAL, K., and BHASKARAN, R. (2010). Supervised Classification
Performance of Multispectral Images. Journal of Computing,
2(2), 124129.
PHAM, D. L., and PRINCE, J. L. (1999). An adaptive fuzzy C-means
algorithm for image segmentation in the presence of intensity in
homogeneities. Pattern Recognition Letters 20, 5768.

PHAM, D.T., DIMOV, S.S., and NGUYEN, C. D. (2004). Selection of K


in k-means clustering. Proc. IMechE, 219 Part C: J. Mechanical
Engineering Science, doi:10.1243/095440605X8298.
PCI Geomatics (2001). X-Pace Reference Manual, Version 8.2.
PCI Geomatics, Ontario, Canada.
PRIDE, S. (1994). Governing equations for the coupled electromagnetics and acoustics of porous media: Physical Review B:
Condensed Matter and Materials Physics, 50(21), 1567815696,
doi:10.1103/PhysRevB.50.15678.
RANA, S., JASOLA, S. and KUMAR, R. (2010). A hybrid sequential
approach for data clustering using k-means and particle swarm
optimization algorithm. International Journal of Engineering,
Science and Technology. 2(6), 67176.
REIGBER, A., JAGER, M., NEUMANN, M., and FERRO-FAMIL, L. (2010).
Classifying polarimetric SAR data by combining expectation
methods with spatial context. International Journal of Remote
Sensing. 31(3), 727744.
ROY, A., and APPARAO, A. (1971). Depth of investigation in direct
current methods. Geophysics. 36, 943959.
RUCKER, D. F., NOONAN, G. E., and GREENWOOD, W. J. (2011).
Electrical resistivity in support of Geological mapping along the
Panama Canal. Engineering Geology. 117, 121133, doi:10.
1016/j.enggeo.2010.10.012.
RUCKER, D.F., FINK, J. B., and LOKE, M. H. (2011). Environmental
monitoring of leaks using time lapsed long electrode electrical
resistivity, Journal of Applied Geophysics. 74, 242254.
RUCKER, D.F. (2012). Enhanced resolution for long electrode ERT.
Geophysical Journal International. 191, 101111 doi:10.1111/j.
1365-246X.2012.05643.x.
SASAKI, Y. (1992). Resolution of resistivity tomography inferred
from numerical simulation Geophysical Prospecting. 40,
453463.
SCHWARZBACH, C., BORNER, R., and SPITZER, K. (2005). Twodimensional inversion of direct current resistivity data using a
parallel, multi-objective genetic algorithm. Geophys. J. Int. 162,
685695, doi:10.1111/j.1365-246X.2005.02702.x.
SEATON, W., and BURBEY, T. J. (2000). Aquifer characterization in
the Blue Ridge physiographic province using resistivity profiling
and borehole geophysics. Journal of Environmental and Engineering Geophysics 5(3), 4558.
SEATON, W. J.,and BURBEY, T. J. (2002). Evaluation of twodimensional resistivity methods in a fractured crystalline-rock
terrane. Journal of Applied Geophysics, 51(1), 2141.
SINGH, A. (1989). Review article digital changes detection using
remotely-sensed data. International Journal of Remote Sensing.
10(6), 9891003.
STORZ, H., STORZ, W., and JACOBS, F. (2000). Electrical resistivity
tomography to investigate geological structures of the earths
upper crust. Geophysical Prospecting. 48, 455471.
STUMMER, P., MAURER, H.R., HORSTMEYER, H., GREEN, A.G. (2002).
Optimization of DC resistivity data acquisition: real-time
experimental design and a new multi-electrode system. IEEE
Transactions on Geosciences and Remote Sensing 40(12),
27272736.
STUMMER, P., MAURER, H., GREEN, A. (2004). Experimental design:
electrical resistivity data sets that provide optimum subsurface
information. Geophysics 69, 120129.
SZALAI, S., and SZARKA, L. (2008). On the classification of surface
geoelectric arrays: Geophysical Prospecting, 56, 59175, doi:10.
1111/j.1365-2478.2007.00673.x.

K. S. Ishola et al.
TARANTOLA, A., and VALETTE, B. (1982). Generalized nonlinear
inverse problems solved using the least squares criterion.
Reviews of Geophysics, 20(2), 219232, doi:10.1029/
RG020i002p00219.
TIKHONOV, A. N., and ARSENIN, V. Y. (1977). Solutions of Ill-Posed
Problems, John Wiley, Hoboken, N. J.
. G. (2002).
TRYGGVASON, A., ROGNVALDSSON, S.T., and FLOVENZ, O
3-dimensional imaging of the P- and S-wave velocity structure
and earthquake locations beneath southwest Iceland: Geophysical Journal International, 151, 848866, doi:10.1046/j.1365246X.2002.01812.x.
TSO, B., and OLSEN, R. C. (2005). Combining spectral and spatial
information into hidden Markov models for unsupervised image
classification. International Journal of Remote Sensing. 26(10),
21132133.
WARD, W., WILKINSON, P., CHAMBERS, J., and BAI, Li. (2014).
Comparative Investigation of Guided Fuzzy Clustering and Mean
Shift Clustering for Edge Detection in Electrical Resistivity
Tomography Images of Mineral Deposits. Geophysical Research,
16, EGU2014EGU13558,

Pure Appl. Geophys.


WILKINSON, P.B., KURAS, O., MELDRUM, P.I., CHAMBERS, J.E., OGILVY, R.D. (2006a). Comparison of the spatial resolution of
standard and optimised electrical resistivity tomography arrays.
In: Proceedings of the 12th meeting of the European Association
of Geoscientists & Engineers (EAGE) Near Surface Geophysics
Conference, Helsinki, Finland, P040.
WILKINSON, P.B., MELDRUM, P.I., CHAMBERS, J.C., KURAS, O., OGILVY, R.D. (2006b). Improved strategies for the automatic
selection of optimized sets of electrical resistivity tomography
measurement configurations. Geophysical Journal International
167, 11191126.
ZALIK, K. R. (2008). An efficient k means clustering algorithm.
Pattern Recognition Letters, 29, 13851391.
ZHOU, B. and DAHLIN, T. (2003). Properties and effects of measurement errors on 2D resistivity imaging surveying. Near
surface Geophysics 1, 105117.
ZHOU, B. and GREENHALGH, S. A. (2000). Cross-hole resistivity
tomography using different electrode configurations, Geophysical Prospecting. 48, 887912.

(Received February 27, 2014, revised November 26, 2014, accepted December 4, 2014)

Vous aimerez peut-être aussi