Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
CHAPTER VI
INVERSION OF RESISTIVITY DATA BY ARTIFICIAL NEURAL
NETWORK (ANN)
6.1 Introduction
Interpretation of the Geophysical data is necessary to understand the concept of
reality behind the earth system and there should be an efficient tool to estimate and
predict the parameters which are relevant to the system. Thus computational method
involves the intelligent network to justify all the non linear problem is necessary. Neural
network can have the power to understand the factors related to the variations of the sub
surface geology. Conventional mathematical analysis is difficult when there are uncertain
and partially defined systems with a degree of vagueness. The use of intelligent systems
to evaluate such complex system is ideal.
Computational applications in the field of Geophysics is more advantageous
nowadays to understand the non linear behavior of the earth. Artificial neural network
(ANN) plays a major role to dig out the mystery of earth science by means of the
previous learned examples. An Artificial Neural Network (ANN) usually called Neural
Network (NN) is a mathematical model or computational model that tries to simulate the
structure and functional aspects of biological neural networks by training. It consists of
interconnected group of artificial neurons and also processes information using a
connectionist approach to computation. The concept of Neural network diagram is shown
in Figure (6.1).
Figure 6.1 Concept of Neural network
Neural networks are non-linear statistical data modeling tools. They can be used
to model complex relationships between inputs and outputs or to find patterns in data.
A computing system is built from a large number of simple processing elements dealing
individually with parts of a large problem. This net may have several layers of processing
elements that are massively interconnected. In this, adaptively adjusted weights are
applied to the inputs and to the connections between the processing elements. There are
many applications of ANN which includes oil and gas exploration, speech and pattern
recognition, financial forecasting and health care cost reduction. Thus ANN is a general-
purpose model with a limited set of variables and is used as a universal functional
approximation. Usually, ANNs are trained by adjusting the values of these connection
weights between the network elements. These networks have self-learning capability and
are fault-tolerant as well as noise-immune and have applications in various fields
(Sreekanth et al, 2009).
A notable feature of these networks is their adaptive nature, whose learning by
example replaces programming in solving problems. This feature makes such
computational models very appealing in application domains where one has little or
incomplete understanding of the problem to be solved but where training data is readily
available. ANNs are now being increasingly recognized in the area of classification and
prediction, where regression model and other related statistical techniques have
traditionally been employed. Introducing the layer of neurons involved in the training is
shown in Figure (6.2 and 6.3). The most widely used learning algorithm in an ANN is the
back propagation algorithm, others are multilayered perception, radial basis function and
kohonen networks. In fact, majority of the network are more closely related to the
traditional, mathematical and statistical models such as non-parametric pattern classifiers,
clustering algorithms, non-linear filters, and statistical regression models. ANNs have
been used for a wide variety of applications where statistical methods are traditionally
employed.
Figure 6.2 A neuron with a single scalar input and no bias appears on the left and
with bias on the right.
Figure 6.3 A neuron with a single R-element vector is shown above
In ANN, the sigmoid function is used as activation function. It can produce
outputs with reasonable discriminating power and its output functions are differentiable,
which is essential for the back propagation of errors (Yegnanarayana, 2005).Its Gaussian
shaped derivatives help to stabilize the network and compensate for over-correction of the
weights (Caudill, 1988). During ANN training, each hidden and output neuron process
inputs by multiplying each input by its weights. The products are summed and processed
using the activation function. The expression for logistic sigmoid function and its
derivative are given below.
The number of neurons in the hidden layer varies as per the requirement of
optimum performance, which could be decided on trial and error basis. Initial weights are
assigned at random in the range suitable for the activation function at neurons. (Jimmy
Stephen et al,2004).In this iterative non-linear feed forward network Back-Propagation
algorithm, uses learning rate (µ) for training the network; if µ is very high, the network
may not meet the desired minimum and if µ is very small, the network takes so much
time to meet the desired minimum. The mentioned algorithm adds the momentum for
decreasing the probability of the error accumulation and also reducing the number of
training.
An adaptive network consists of a structure of nodes and the directional links
through which these nodes are connected. The outputs of the nodes depend on the
parameters associated with these nodes. The change in these parameters is brought about
by a learning rule to minimize the error. The learning techniques can automate this
process and substantially reduce development time and cost while improving
performance. For neural networks, the knowledge is automatically acquired by the back
propagation algorithm and will be a more efficient tool used for the interpretation of
Geophysical data extensively.
Artificial neural networks have been successfully applied to a number of
geophysical problems, including parameter estimation, classification, filtering and
optimization (Poulton et al,1992;Sheen,1997;Vander Baan and Jutten, 2000;Calderon-
Macias et al, 2000; El-QadyandUshijima, 2001). The adoption of the neural network
approach holds several advantages over more conventional modeling techniques,
allowing the effective solution of non-linear problems from complex, incomplete and
noisy data. The computational expense of the inversion using neural networks, being
dependent upon the dimension of the space of unknown parameters rather than the
physical dimensions of the model space, is also comparatively low (Spichak and Popova,
2000).
6.2 Significance of Artificial Neural network
ANN finely tune to tasks other than number crunching; the human brain possesses
many advantages over a digital computer. One can quickly recognize a face, even when
seen from the side in bad lighting in a room full of other objects. One can easily
understand speech, even that of an unknown person in a noisy room. Despite years of
focused research, computers are far from performing at this level. The brain is also
remarkably robust; it does not stop working just because a few cells die. Compare this
with a computer, which will not normally survive any degradation of the central
processing unit( CPU). But perhaps the most fascinating aspect of the brain is that it can
learn. No programming is necessary; one does not need a software upgrade, just because
one wants to learn to ride a bicycle. (Yegnanarayana 2005 )The computations of the
brain are done by a highly interconnected network of neurons, which communicate by
sending electric pulses through the neural wiring consisting of axons, synapses and
dendrites.
In some cases, a neuron is modeled as a switch that receives input from other
neurons and, depending on the total weighted input, is either activated or remains
inactive. The weight, by which an input from another cell is multiplied, corresponds to
the strength of a synapse—the neural contacts between nerve cells. These weights can be
both positive (excitatory) and negative (inhibitory). In the 1960s, it was shown that
networks of such model neurons have properties similar to the brain: they can perform
sophisticated pattern recognition, and they can function even if some of the neurons are
destroyed. The demonstration, in particular by Rosenblatt, that simple networks of such
model neurons called perceptrons could learn from examples stimulated interest in the
field, showed that simple perceptrons could solve only the very limited class of linearly
separable problems, activity in the field diminished. Nonetheless, the error back-
propagation method, which can make fairly complex networks of simple neurons learn
from examples, showed that these networks could solve problems that were not linearly
separable. Net talk, an application of an artificial neural network for machine reading of
text, was one of the first widely known applications, and many followed soon after. In the
field of biology, the exact type of network as in Net talk was also applied to prediction of
protein secondary structure; in fact, some of the best predictors still use essentially the
same method. Another big wave of interest in artificial neural networks started, and led to
a fair deal of hype about magical learning and thinking machines.
Artificial neural networks are inspired by the early models of sensory processing
by the brain. An artificial neural network can be created by simulating a network of
model neurons in a computer. By applying algorithms that mimic the processes of real
neurons, one can make the network to ‘learn’ and to solve many types of problems It
receives input from a number of other units or external sources, weighs each input and
adds them up. If the total input is above a threshold, the output of the unit is one;
otherwise it is zero. Therefore, the output changes from 0 to 1 when the total weighted
sum of inputs is equal to the threshold. The points in input space satisfying this condition
define a scaled hyper plane. In two dimensions, a hyper plane is a line, whereas in three
dimensions, it is a normal plane. Points on one side of the hyper plane are classified as 0
and those on the other side as 1. It means that a classification problem can be solved by a
threshold unit if the two classes separated by a hyper plane, are said to be linearly
separable.
6.3 Back-propagation
Back-propagation learning algorithm works for feed-forward networks with
continuous output. Training starts by setting all the weights in the network to small
random numbers. Now, for each input example, the network gives an output, which starts
randomly. The sum of all these numbers over all training examples is called the total error
of the network. If this number is zero, the network would be perfect, and the smaller the
error, the better the network. By choosing the weights that minimize the total error, one
can obtain the neural network that best solves the problem at hand. This is the same as
linear regression, where the two parameters characterizing the line are chosen such that
the sum of squared differences between the line and the data point is minimal. This can be
done analytically in linear regression, but there is no analytical solution in a feed-forward
neural network with hidden units. In back-propagation, the weights and thresholds are
changed each time an example is presented, such that the error gradually becomes
smaller. This is repeated, often hundreds of times, until the error no longer changes. An
illustration can be found at the Neural Java site, following the link “Multi-layer
Perceptron (with neuron outputs in {0;1})”. In back-propagation, a numerical
optimization technique called gradient descent makes computing particularly simple; the
form of the equations gave rise to the name of this method. There are some learning
parameters called learning rate and momentum that need tuning when using back-
propagation, and there are other problems to consider. For instance, gradient descent is
not guaranteed to find the global minimum of the error, so the result of the training
depends on the initial values of the weights. However, one problem overshadows the
others: that of over-fitting. Over-fitting occurs when the network has too many
parameters to be learned from the number of examples available, that is, when few points
are fitted with a function with too many free parameters. Although this is true for any
method of classification or regression, neural networks seem especially prone to over
parameterization. For instance, a network with 10 hidden units for solving the example
problem would have 221 parameters: 20 weights and a threshold for the 10 hidden units
and 10 weights and a threshold for the output unit. That is too many parameters to be
learned from 100 examples. A network that over fits the training data is unlikely to
generalize well to inputs that are not in the training data. There are many ways to limit
over-fitting apart from simply making small networks, but the most common include
averaging over several networks, regularization and using methods from Bayesian
statistics. To estimate the generalization performance of the neural network, one needs to
test it on independent data, which have not been used to train the network. This is usually
done by cross-validation, where the data set is split into, for example, and ten sets of
equal size. The network is then trained on nine sets and tested on the tenth, and this is
repeated ten times, so all the sets are used for testing. This gives an estimate of the
generalization ability of the network; that is, its ability to classify inputs that it was not
trained on. To get an unbiased estimate, it is very important that the individual sets do
not contain examples that are very similar. Goh (1995) developed back propagation
neural network (BPNN) models to evaluate the shaft resistance of piles driven in cohesive
soil. Teh et al,(1997) developed a BPNN model to estimate the static pile capacity from
the dynamic stress wave data.
In an artificial neural network several processing units are interconnected
according to some topology to accomplish a pattern recognition task. Therefore, the
inputs to a processing unit may come from the outputs of other processing units, and from
external sources. The output of each unit may be given to several units including itself.
The amount of the output of one unit received by another unit depends on the strength of
the connection between the units, and it is reflected in the weight value associated with
the connection link. If there are N units in a given ANN, then at any instant of time each
unit will have a unique activation value and a unique output value. The set of the N
activation values of the network defines the activation state of the network at that instant.
Likewise, the set of the N output values of the network defines the output state of the
network at that instant. Depending on the discrete or continuous nature of the activation
and output values, the state of the network can be described by a discrete or continuous
point in an N-dimensional space. In operation, each unit of an ANN receives inputs from
other connected units and from an external source. A weighted sum of the inputs is
computed at a given instant of time. The activation value determines the actual output
from the output function unit and the output state of the unit. The output values and the
other external inputs in turn determine the activation and output states of the other units.
Activation Dynamics determines the activation values of all the units, and also the
activation state of the network as a function of time. The activation dynamics also
determines the dynamics of the output state of the network. The set of all output states
defines the output state space of the network. Activation dynamics determines the
trajectory of the path of the states in the state space of the network. For a given network,
defined by the units and their interconnections with appropriate weights, the activation
states determine the short term memory function of the network.
Generally, given an external input, the activation dynamics is followed to recall a
pattern stored in a network. In order to store a pattern in a network, it is necessary to
adjust the weights of the connections in the network. The set of all possible weight
vectors define the weight space. When the weights are changing, then the synaptic
dynamics of the network determines the weight vector as a function of time. Synaptic
dynamics is followed to adjust the weights in order to store the given patterns in the
network. The process of adjusting the weights is referred to as learning. Once the learning
process is completed, the final set of weight values corresponds to the long term memory
function of the network. The procedure to incrementally update each of the weights is
called a learning law or learning algorithm. Artificial neural networks are useful only
when the processing units are organized in a suitable manner to accomplish a given
recognition task (Anderson and Rosenfeld 1988). These networks are normally organized
into layers of processing units. The units of a layer are similar in the sense that they all
have the same activation dynamics and output function. Connections can be made either
from the units of one layer to the units of another layer (interlayer connections) or among
the units within the intralayer connections or both interlayer and interlayer connections.
Further, the connections across the layers and among the units within the layer can be
organized either in a feed forward manner or in a feedback manner. In a feedback
network, the same processing unit may be visited more than once.
6.3.1 Feed forward Back propagation Algorithm
Back propagation learning involves propagation of the error backwards from the
output layer to the hidden layers in order to determine the update for the weights leading
to the units in a hidden layer. The error at the output layer itself is computed using the
difference between the desired output and the actual output at each of the output units.
The actual output for a given input training pattern is determined by computing the
outputs of units for each hidden layer in the forward pass of the input data
(Yegnanarayana, 2005). A trained multilayer feed forward neural network is expected to
capture the functional relationship between the input-output pattern pairs in the given
training data. It is implicitly assumed that the mapping function corresponding to the data
is a smooth one. The function approximation interpretation of a single layer feed forward
neural network enables one to view different hidden layers of the network performing
different functions. Feed forward back propagation algorithm is used here to invert the
Vertical Electrical Sounding (VES) Data.
6.4 Learning
Synaptic dynamics is attributed to learning in a biological neural network. The
synaptic weights are adjusted to learn the pattern information in the input samples.
Typically, learning is a slow process, and the samples containing a pattern may have to be
presented to the network several times before the pattern information is captured by the
weights of the network. A large number of samples are normally needed for the network
to learn the pattern implicit in the samples. Pattern information is distributed across all the
weights, and it is difficult to relate the weights directly to the training samples. The only
way to demonstrate the evidence of learning pattern information is that, given another
sample from the same pattern source, the network would classify the new sample into the
pattern class of the earlier trained samples. Another interesting feature of learning is that
the pattern information is slowly acquired by the network from the training samples, and
the training samples themselves are never stored in the network. That shows a preference
to learn from examples, not store the examples themselves.
The adjustment of synaptic weights is represented by a set of learning equations,
which describe the synaptic dynamics of the network. The learning equation describing a
synaptic dynamics model is given as an expression for the first derivative of the synaptic
weight wij connecting the unit j to the unit i. The set of equations for all the weights in the
network determine the trajectory of the weight states in the weight space from a given
initial weight state. Learning laws refer to the specific manner in which the learning
equations are implemented. Depending on the synaptic dynamics model and the manner
of implementation, several learning laws have been proposed.
6.4.1 Requirements of learning laws
a. The learning law should lead to convergence of weights.
b. The learning or training time for capturing the pattern information from samples
should be as small as possible.
c. An on-line learning is preferable than to an off-line learning. That is, the weights
should be adjusted on presentation of each sample containing the pattern information.
d. Learning should use only the local information as far as possible. That is, the change
in the weight on a connecting link between two units should depend on the states of
these two units only. In such a case, it is possible to implement the learning law in
parallel for all the weights, thus speeding up the learning process.
e. Learning should be able to capture complex nonlinear mapping between input-output
pattern pairs, as well as between adjustment patterns in a temporal sequence of
patterns.
f. Learning should be able to capture as many patterns as possible into the network. That
is, the pattern information storage capacity should be as large as possible for a given
network.
6.4.2 Categories of learning
Learning can be viewed as searching through the weight space in a systematic
manner to determine the weight vector that leads to an optimum (minimum or maximum)
value of an objective function. The search depends on the criterion used for learning.
They include minimization of mean squared error, relative entropy, maximum likelihood,
gradient descent, etc. There are several learning laws in use, and new laws are being
proposed to suit a given application and architecture. At first, the learning or weight
adjustment could be supervised or unsupervised. In supervised learning the weight
adjustment is determined based on the deviation of the desired output from the actual
output. Supervised learning may be used for structural learning or for temporal learning.
Structural learning is concerned with capturing in the weights the relationships between
the given input-output pattern pairs. Temporal learning is concerned with capturing in the
weights the relationship between neighboring patterns in a sequence of patterns.
Unsupervised learning discovers features in a given set of patterns, and organizes
the patterns accordingly. There is no externally specified desired output in the case.
Unsupervised learning uses mostly local information to update the weights. The local
information consists of signal or activation values of the units at either end of the
connection for which the weight update is being made.
Learning methods may be off-line or on-line. In an off-line learning all the given
patterns are used together to determine the weights. On the other hand, in an on-line
learning the information in each new pattern is incorporated into the network by
incrementally adjusting the weights. Thus an on-line learning allows the neural network
to update the information continuously. In practice, the training patterns can be
considered as samples of random processes. Accordingly, the activation and output states
could also be considered as samples of random processes. Randomness in the output state
could also result if the output function is implemented in a probabilistic manner rather
than in a deterministic manner. These input, activation and output variables may also be
viewed as fuzzy quantities instead of crisp quantities. Thus we can view the learning
process as deterministic or stochastic or fuzzy or a combination of these characteristics.
Finally, in the implementation of the learning methods, the variables may be
discrete or continues. Likewise the update of weight values may be in discrete steps or in
continuous time. All these factors influence not only the convergence of weights, but also
the ability of the network to learn from the training samples. Yegnananarayana, (2005).
Neural network is also gaining importance in liquefaction susceptibility evaluation of
foundation soils (Goh 1995, Hanna et al, 2007).
6.5 Application of ANN on interpreting Vertical Electrical Sounding Data
Computational neural networks provide much more than just a set of novel, and
valuable data driven mathematical tools , with respect to geophysical analysis and
modeling.These tools provide an appropriate frame work for re-engineering many well
established spatial analysis and environmental modeling techniques to meet the large
scale data processing. Application of ANN models to spatial data set holds the potential
for fundamental advances in empirical understanding across a broad spectrum of sub-
surface features of the study area.
The study of 1D resistivity inversion procedure using the ANN system was carried
out because the procedure works well for the observed data at Abishekapatti study area.
Best training performance can be achieved after the iterations has been successfully
completed the goal using the number of epochs for the synthetic data. Whenever the goal
has been achieved in a particular number of epochs, the training stops and the plot
showing the performance and training state, synthetic and trained and the regression plot
data for the respective VES 2, 6, 8,14, as shown in Figure( 6.4 to 6.12). The test using
trained data indicates that the ANN system can converge to the target rapidly and
accurately. This inversion thus represents the layer parameters of subsurface geology like,
thickness and resistivity. A successful interpretation was produced only after the training
and determines the model parameters with the trained data. The results of interpretation of
the near-surface features by means of this ANN technique is satisfactory and are more
efficient and the error percentage before and after training is less than 11% (Table
6.1).Here, the adaptability of ANN in direct parameter estimated from VES data on hard
rock area involves a variety of curve types with a wide range of model parameter values.
Fig
ure
6.4
S
yn
thet
ic D
ata
Tra
inin
g a
nd
La
yer
mo
del
of
VE
S 2
Fig
ure
6.5
S
how
s th
e p
erfo
rma
nce
an
d T
rain
ing
sta
te o
f V
ES
2
Fig
ure
6.6
S
ho
ws
the
Reg
ress
ion
sta
te o
f V
ES
2
Fig
ure
6.7
Th
e T
rain
ing a
nd
Lay
er m
od
el a
nd
per
form
an
ce p
lot
of
VE
S 6
Fig
ure
6.8
Sh
ow
s th
e T
rain
ing
sta
te a
nd
th
e R
egre
ssio
n s
tate
of
VE
S 6
Fig
ure
6.9
Th
e T
rain
ing a
nd
per
form
an
ce p
lot
of
VE
S 8
Fig
ure
6.1
0
Sh
ow
s th
e T
rain
ing
sta
te
an
d t
he
Reg
ress
ion
sta
te o
f V
ES
8
Fig
ure
6.1
1
Th
e T
rain
ing a
nd
per
form
an
ce m
od
el o
f V
ES
14
Fig
ure
6.1
2
Sh
ow
s th
e T
rain
ing s
tate
a
nd
th
e R
egre
ssio
n s
tate
of
VE
S 1
4
Tab
le 6
.1 I
nte
rpre
ted
La
yer
Mo
del
of
fiel
d d
ata
alo
ng w
ith
Err
or
Per
cen
t
S.N
o
DA
TA
M
OD
EL
PA
RA
ME
TE
RS
(ρ ρρρ=
Tru
e re
sist
ivit
y o
f su
b-s
urf
ace
ER
RO
R
PE
RC
EN
TA
GE
lay
ers,
T=
Th
ick
nes
s of
lay
ers)
1
VE
S 2
ρ ρρρ
1=
10
.38
ρ ρρρ2
=6
4.6
88
ρ ρρρ3
=7
66.2
T1
=1
5
T2
=1
4.4
41
2
10
.71
12
2
VE
S 6
ρ ρρρ
1=
18
.362
9
ρ ρρρ2
=9
4.8
32
9
ρ ρρρ3
=7
96
T1
=1
4
T2
=1
5
11
.43
3
VE
S 8
ρ ρρρ
1=
7.6
91
5
ρ ρρρ2
=6
1.9
ρ ρρρ3
=2
65
T1
=1
8
T2
=1
7
10
.4
4
VE
S 1
4
ρ ρρρ1
=2
5.1
60
4
ρ ρρρ2
=5
1.3
09
6
ρ ρρρ3
=7
52
T1
=2
9.0
9
T2
=2
8.0
0.1
73
3
Fig
ure
6.1
3
Th
e S
yn
thet
ic D
ata
Tra
inin
g a
nd
per
form
an
ce m
od
el o
f S
3
Fig
ure
6.1
4
Sh
ow
s th
e T
rain
ing
an
d t
he
Reg
ress
ion
sta
te o
f S
3
Fig
ure
6.1
5 T
he
Sy
nth
etic
Da
ta T
rain
ing a
nd
Per
form
an
ce m
od
el o
f S
7
Fig
ure
6.1
6
Sh
ow
s th
e T
rain
ing
sta
te a
nd
Reg
ress
ion
sta
te o
f S
7
Fig
ure
6
.17
T
he
Sy
nth
etic
Da
ta T
rain
ing
an
d L
ay
er m
od
el o
f S
11
Fig
ure
6
.18
Sh
ow
s th
e T
rain
ing
sta
te a
nd
Reg
ress
ion
sta
te o
f S
11
Fig
ure
6
.19
Th
e S
yn
thet
ic D
ata
Tra
inin
g a
nd
La
yer
mo
del
of
S13
Fig
ure
6.2
0
Sh
ow
s th
e T
rain
ing s
tate
an
d R
egre
ssio
n s
tate
of
S1
3
Tab
le
6.2
In
terp
ret
ed L
ay
er M
od
el o
f fi
eld
data
alo
ng w
ith
Err
or
Per
cen
t
S.N
o
Data
M
od
el
Pa
ram
eter
s
(ρ ρρρ=
Tru
e re
sist
ivit
y o
f su
b-s
urf
ace
la
yer
s, T
=T
hic
kn
ess
of
lay
ers)
Err
or
per
cen
t
1
S3
ρ1
=1
7.4
ρ2
=1
.2
ρ3
=0
.56
5
T1
=3
.6,
T2
=1.5
7
1.3
63
3
2
S7
ρ1
=1
7.4
ρ2
=3
.2
ρ3
=0
.56
50
T1
=1
.0
T2
=0
.36
1.2
39
4
3
S11
ρ1
=4
.18
69
ρ2
=5
0.1
1
ρ3
=7
51
.73
T1
=5
.6,
T2
=4
5.6
4.0
01
3
4
S13
ρ1
=2
7.1
ρ2
=8
4.1
ρ3
=7
83
T1
=2
3,
T2
=1
77
6.4
04
6
Similarly, the study of 1D resistivity inversion procedure using the ANN system
was carried out at Thuthukudi study area for the sounding data S1 to S20. Best training
performance can be achieved after the iterations has been successfully completed the goal
using the number of epochs for the synthetic data. Whenever the goal has been achieved
in a particular number of epochs, the training stops. The plot showing the performance
and training state and also the regression of the VES of S3 - S13 data as shown in Figure
(6.13 to 6.20) for the respective VES of S3 - S13. The test using trained data indicates that
the ANN system can converge to the target rapidly and accurately. The synthetic and
ANN trained synthetic data along with layer model is plotted and. This inversion thus
represents the layer parameters of subsurface geology. The interpretation was produced
with the trained data. The results of interpretation were satisfactory and efficient and
show less than 6% (Table 6.2) error percentage before and after training. The present
approach of two step net work parameter estimation performs effectively and once
trained, the model parameters are calculated fast on simulating the net work with saved
weights and bias on new examples. The network also achieved the generalization
capability but within the limits of the presented model parameters in the study area .This
will help to a greater extent for onsite interpretation of VES data.