applied

sciences
Article
Uncertainty Analysis of Two Methods in
Hydrocarbon Prediction under Different Water
Saturation and Noise Conditions
Changcheng Liu * , Deva Ghosh and Ahmed Mohamed Ahmed Salim
Centre of Excellence in Subsurface Seismic Imaging & Hydrocarbon Prediction, Universiti Teknologi Petronas,
Seri Iskandar 32610, Malaysia; ghosh_deva@yahoo.com (D.G.); mohamed.salim@utp.edu.my (A.M.A.S.)
* Correspondence: lccgreatwall@gmail.com




Received: 14 August 2019; Accepted: 25 September 2019; Published: 2 December 2019


Abstract: The uncertainty of two recently proposed methods, “new fluid factor” and “delta K”, is
analyzed under different water saturation and noise conditions through Monte Carlo modelling.
The new fluid factor performs reliably (all metric parameters are above 0.9) when the water saturation
is up to 95%. The delta K has better performance (all metric parameters are close to 1) such that
it is able to distinguish hydrocarbon from brine without the interference of high water saturation.
The results prove the performances of the two methods are stable in a high water-saturation scenario.
The analysis of noise indicates the methods are sensitive to noise in the input data in that the
performance is excellent when the noise is relatively low (−20 dB) and decreases with increasing
noise energy. The new fluid factor, which is in the interface domain, is more sensitive than delta K
in the impedance domain. The metric parameters of the new fluid factor and delta K are in the
range of 0.5 to 0.8 when the noise is high (−7 dB). High-quality input data and integration with other
geophysical methods can effectively reduce these risks. In addition, two widely used traditional
methods (fluid factor and Lambda-Rho) are analyzed as comparisons. It turns out the new fluid
factor and delta K have better performance than traditional methods in both high water saturation
and noise conditions.

Keywords: fluid indicator; hydrocarbon prediction; high water saturation; noise; seismic attribute;
uncertainty analysis

1. Introduction
The seismic attribute is a useful tool for hydrocarbon prediction, which is important in the
petroleum industry. Many methods are proposed for hydrocarbon prediction from both reflectivity
and impedance domains. The fluid factor in the reflectivity domain is one of the most popular and
useful tools although it was introduced for more than 20 years ago. The fluid factor is defined as the
difference between measured and estimated P wave velocity [1]. Its expressions are as follows:

∆V p ∆V s
FFSmith = −a /γ, (1)
Vp Vs

where a and γ are the coefficients related to the Mudrock line [2] and fluid substitution, respectively.
A recent work has been called “new fluid factor (FFnew )” [3]. It is based on the J attribute [4] which
is proposed to reduce the ambiguity using the amplitude versus offset (AVO) method in hydrocarbon
prediction. To relate seismic amplitudes to geology, it is necessary to understand all the physical factors
that influence seismic amplitudes [5]. Seismic amplitude is affected by pore fluid, rock matrix and
porosity. J attribute is proven to be able to eliminate the porosity effect and enhance the accuracy in

Appl. Sci. 2019, 9, 5239; doi:10.3390/app9235239 www.mdpi.com/journal/applsci

Appl. Sci. 2019, 9, 5239 2 of 14

hydrocarbon prediction. The new fluid factor modified the equation of J attribute with consideration
of the effect of the rock matrix. Mathematically, the new fluid factor is expressed as follows:

FFnew = J − JCT , (2)

where,
J = Jp sin α − Js cos α, (3)

JCT = Jp_matrix sin α − Js_matrix cos α, (4)

where FFnew is the new fluid factor, J is the J attribute, and JCT is the correct term related to the rock
matrix. The new fluid factor FFnew is improved for hydrocarbon prediction with lower uncertainty.
The brine responses are close to zero. Hydrocarbons are obviously separated from brine. The uncertainty
of the new fluid factor value in the various Vclay cases are reduced to an acceptable degree without
mixing between brine and hydrocarbons. The new fluid factor is stable in this scenario with various
porosity and Vclay [3].
In the impedance domain, there are two popular methods to utilize the well log data or the
inversion result. One is the cross plot of acoustic impedance (AI) and Vp /Vs , which is introduced
by Ødegaard and Avseth [6]. The brine sands and the shales in the sedimentary basin have a trend
with the depth: low AI and high Vp /Vs in the shallow, and high AI and high Vp /Vs in the shallow.
The response of brines and shales in the cross plot of AI vs. Vp /Vs is called the “background trend”.
Both AI and Vp /Vs shift towards lower values from the background trend when the rock contains
hydrocarbons. Thus, this cross plot can be used to identify the hydrocarbon by the anomaly selection.
Another technique is the Lambda–Mu–Rho (λ-µ-ρ) to improve fluid dictation and lithology
discrimination, where λ is 1st Lamé parameters, µ is 2nd Lamé parameters or shear modulus, and ρ is
density [7]. Mathematically, λ can be represented by bulk modulus K and shear modulus µ. The common
attributes in the λ-µ-ρ technique (λρ and µρ) can be linked with acoustic impedance and shear wave
impedance (SI). Generally, the hydrocarbon reservoir can be identified because it has lower λρ than
brine sands. λρ and µρ are defined as follows:

λρ = AI2 − 2SI2 , (5)

µρ = SI2 , (6)

The above methods work successfully in many hydrocarbon reservoirs, in particular, gas layers.
Although these two methods have been proposed for more than 10 years, they are still applied in
recent works [8–10].
A new method is named “delta K (∆K)” [11]. It is defined as the difference of bulk modulus
between the real case (Kreal ) and water-substituted rock (Kwet ) as follows:

∆K = Kreal − Kwet , (7)

The definition of conventional fluid factor contains the information of shear modulus and density,
which are affected by porosity, rock matrix and pore fluid, simultaneously. However, delta K only
focuses on bulk modulus, which is more sensitive to fluid changing. Furthermore, the consideration of
the water-substituted case in the definition makes the delta K more precise.
Uncertainty exists in the hydrocarbon prediction. The uncertainty is defined as the estimated
amount or percentage by which an observed or calculated value may differ from the true value [12].
In the fluid prediction, it refers to the proportion of match or not match between predicted or true fluid
types. Uncertainty is quantitatively analysed using five metric parameters (precision, recall, accuracy,
F-measure and the area under the curve (AUC) in this study.
Appl. Sci. 2019, 9, 5239 3 of 14

2. Research Methodology
The Monte Carlo method is a useful algorithm that has been applied in many aspects. It consists of
repeated random sampling to generate a series of data. Then, these data are input into a model/system to
obtain a numerical result. In principle, the Monte Carlo method takes advantage of its stochastic nature
to solve complex problems. The typical application of the Monte Carlo method includes sampling,
estimation and optimization [13]. This method has been applied in the field of petroleum exploration,
including seismic processing, inversion, reservoir identification and uncertainty quantification [14–17].
The Monte Carlo method is used in the forwarding model in this study. The range of random
sampling can be adjusted conveniently, and the combination of the variables can cover most scenarios
when the number of samples is sufficient. The responses of these scenarios can be visualized and
analysed in the different methods of hydrocarbon prediction.
Some empirical equations are used for Monte Carlo modelling. Three empirical relationships [18]
are used to derive Vp , Vs , and the density of sand in this model for the sand layer:

Vp = 5.59 − 6.93 ∗ Φ − 2.18 ∗ Vclay , (8)

Vs = 3.52 − 4.91 ∗ Φ − 1.89 ∗ Vclay , (9)

ρ = −0.0115 ∗ Vp2 + 0.261 ∗ Vp + 1.515, (10)

In addition, the equations for the shale layer which is used for the analysis of the interface-domain
attribute are as follows:
Vs_shale = 0.862 ∗ Vp_shale − 1.172, (11)
2
ρshale = −0.0261 ∗ Vp_shale + 0.373 ∗ Vp_shale + 1.458, (12)

Hydrocarbon prediction can be regarded as binary classification which contains bool values, true
and false. One of the standard ways to judge the performance is the confusion matrix. A confusion
matrix consists of the counts of each predicted label: true positive (TP), false positive (FP), false negative
(FN) and true negative (TN). TP and TN refer to correct predictions, and FP and FN are incorrect
predictions. The metrics which are usually used are the precision, recall, accuracy and F-measure as
shown in Table 1. Each metric ranges from 0 to 1, with 1 representing best performance.

Table 1. Metrics and definitions.

Metrics Definition
Precision TP/(TP + FP)
Recall TP/(TP + FN)
Accuracy (TP + TN)/(TP + FP + FN + TN)
F1 2(Precision ∗ Recall)/( Precision + Recall)

The receiver operating characteristic (ROC) curve is usually used to evaluate the binary classifier.
The AUC is equal to the probability that a classifier will rank a randomly chosen positive instance
higher than a randomly chosen negative one (assuming ‘positive’ ranks higher than ‘negative’) [19].
The AUC which equals 1 means the best prediction. A good classifier has the ROC which is above the
line of no-discrimination (a line from the left bottom to the top right corner, also called ‘random guess
line’). The ROC and AUC analysis, which are based on the evaluation metrics and confusion matrix,
output a threshold for a specific binary classification problem.

3. Results
In previous works [3,11], the effectiveness of FFnew and ∆K in various porosity and Vclay scenarios
are well explained. However, the uncertainty, which is essential to the application, remains to be
discussed. Two scenarios are selected to be used to analyse the performance of the attributes. One is
Appl. Sci. 2019, 9, x FOR PEER REVIEW 4 of 14
Appl. Sci. 2019, 9, 5239 4 of 14
One is high-water saturation. The water saturation range, which is discussed in previous works is 0%
to 80%. The performance of the new attributes in higher saturation cases will be discussed. Another
onehigh-water saturation.
is noise, since The water
it is difficult for thesaturation
data to berange, whichThe
noise-free. is discussed in previous
uncertainty works
in different is level
noise 0% tois
80%. The performance of the new attributes in higher saturation cases will be
needed to be understood. The Monte Carlo model is established using the empirical equations discussed. Another
one is noise,
(equation 8–12). since it is difficult
The porosity and for
V thearedata to be noise-free.
assigned randomly The
in uncertainty in3%
the ranges of different
to 30%noise
and level
0% to
is needed to be understood. The Monte Carlo model is established using the empirical equations
40%, respectively. In addition, the traditional fluid factor and λρ are analyzed to compare with the
(Equations (8)–(12)). The porosity and V are assigned randomly in the ranges of 3% to 30% and 0%
new fluid factor and delta K, respectively.clay
to 40%, respectively. In addition, the traditional fluid factor and λρ are analyzed to compare with the
new fluid factor and delta K, respectively.
3.1. Uncertainty Analysis under Different Water Saturation Conditions
3.1. Uncertainty Analysis under Different Water Saturation Conditions
3.1.1. FFnew
3.1.1. FFnew
The sensitivity of water saturation in the new fluid factor (FF ) is analyzed as shown in Figure
FF sensitivity
1a. The The of water saturation
can distinguish in the new accurately
the hydrocarbons fluid factor (FF
when new )the
is analyzed as shown in
water saturation isFigure 1a.
less than
The FF can distinguish the hydrocarbons accurately when the water saturation
80%. The true prediction ratio (TPR) decreases with increasing water saturation. When the water
new is less than 80%.
The trueisprediction
saturation ratio is
99%, the TPR (TPR) decreases with
approximately increasing
0.75. water saturation.
The performances of the When
FF the water
from 0%saturation
to 95% of
is 99%, the TPR is approximately 0.75. The performances of the
water saturation do not differ significantly. The evaluation metrics of the FF FF new from 0% to 95% of water
(the AUC, the
saturation do not differ significantly. The evaluation metrics of
precision, the recall, the accuracy and F1) are shown in Figure 1b. Thesenew the FF (the AUC, the precision,
metrics are approximately
0.98the recall, the accuracy and F1 ) are shown in Figure 1b. These metrics are approximately 0.98 when the
when the water saturation is less than 80%. After that, they start to decrease. Their values are
water saturation is less than 80%. After that, they start to decrease. Their values are above 0.9, which
above 0.9, which indicates good performance when the water saturation is less than 90%. When the
indicates good performance when the water saturation is less than 90%. When the water saturation is
water saturation is higher than 90%, their values reduce rapidly. The ability of the hydrocarbon
higher than 90%, their values reduce rapidly. The ability of the hydrocarbon prediction is strong even
prediction is strong even for high water saturation, for example, 95%, which qualifies as good
for high water saturation, for example, 95%, which qualifies as good performance in the hydrocarbon
performance in the hydrocarbon prediction field. However, in another aspect, high water saturation
prediction field. However, in another aspect, high water saturation is usually not the target in the oil
is usually not the target in the oil and gas industry.
and gas industry.

Figure 1. (a) Receiver operating characteristic (ROC) of the new fluid factor (FFnew ). (b) Evaluation
Figure 1. (a) Receiver operating characteristic (ROC) of the new fluid factor (FF
metrics of the FFnew vary with water saturation.
) (b) Evaluation
metrics of the FF vary with water saturation.
The AUC, which indicates the performance of a binary classifier, is plotted in Figure 2. Its value
isThe AUC,
higher which
than indicates
0.95 when the performance
the water of a binary
saturation ranges from classifier,
0% to 95%;isafter
plotted
that,inthe
Figure
AUC2. Its value
decreases
is higher than
rapidly. 0.95
The when the water
performance of the saturation ranges
FFnew is good when from
the0% to 95%;
water after that,
saturation of thethe AUC decreases
reservoirs is less
rapidly. The performance of the FF
than 95%. is good when the water saturation of the reservoirs is less
than 95%.
Appl. Sci. 2019, 9, x FOR PEER REVIEW 5 of 14
Appl.
Appl. Sci.
Sci. 2019,9,9,x 5239
2019, FOR PEER REVIEW 5 of514of 14

Figure
Figure2.2.The
Thearea
areaunder
underthe
thecurve
curve (AUC)
(AUC) of FFnew vary
the FF
of the varywith
with water
water saturation.
saturation.
Figure 2. The area under the curve (AUC) of the FF vary with water saturation.
3.1.2.Delta
3.1.2. DeltaKK
3.1.2. Delta K
TheROC
The ROCofofthe ∆K is
the ∆K is plotted
plotted in
in Figure
Figure 3a,
3a, where
wherethethecurves
curvesforfordifferent
different water
watersaturation
saturation
conditions
The overlap
ROC of the ∆K
entirely.
isThe results
plotted inshow that
Figure the
3a, performance
where the of
curves∆K is
for
conditions overlap entirely. The results show that the performance of ∆K is very stable very stable
different in the
water different
saturation
in the
water saturation
conditions scenarios.
overlap Regardless
entirely. The of theshow
results waterthat
saturation,
the ∆K can clearly
performance of ∆K
separate
is the hydrocarbon
very stable in the
the
different water saturation scenarios. Regardless of the water saturation, ∆K can clearly separate
and brine. Correspondingly,
different water saturation the evaluation metrics
scenarios. Regardless of ∆K (the
of the water precision, the
saturation, ∆K precision,
recall, the accuracy
can clearlythe and
separateF )the
hydrocarbon and brine. Correspondingly, the evaluation metrics of ∆K (the recall,1 the
are 1, as shown
hydrocarbon andinbrine.
FigureCorrespondingly,
3b. the evaluation metrics of ∆K (the precision, the recall, the
accuracy and F1) are 1, as shown in Figure 3b.
accuracy and F1) are 1, as shown in Figure 3b.

(a)ROC
Figure3.3.(a)
Figure ROCofofthe ∆K. (b)
the ∆K. (b) Evaluation
Evaluation metrics
metricsof the∆K
ofthe ∆Kvary
varywith
withwater
watersaturation.
saturation.
Figure 3. (a) ROC of the ∆K. (b) Evaluation metrics of the ∆K vary with water saturation.
The AUC curve is shown in Figure 4. Its value is 1 when the water saturation ranges from 0 to
The AUC curve is shown in Figure 4. Its value is 1 when the water saturation ranges from 0 to
99%. The ∆K iscurve
able to identify
shown the presence4. of hydrocarbon without
thethe effect of water saturation. It is0 to
99%. The
The AUC
∆K is able to isidentify in
theFigure
presence Its
ofvalue is 1 when
hydrocarbon withoutwater saturation
the effect ranges
of water from
saturation. It
determined
99%. by its
The ∆K isbyable definition which is the difference between the in situ pore fluid and the brine. Hence,
is determined its to identify which
definition the presence of hydrocarbon
is the difference without
between the inthe effect
situ poreoffluid
waterand
saturation.
the brine. It
is∆K can detectby
determined a hydrocarbon
its definitionreservoir
which is with
theadifference
high waterbetween
saturation.
Hence, ∆K can detect a hydrocarbon reservoir with a high water saturation. the in situ pore fluid and the brine.
Hence, ∆K can detect a hydrocarbon reservoir with a high water saturation.
Appl. Sci. 2019, 9, 5239 6 of 14
Appl. Sci. 2019, 9, x FOR PEER REVIEW 6 of 14

Appl. Sci. 2019, 9, x FOR PEER REVIEW 6 of 14

Appl. Sci. 2019, 9, x FOR PEER REVIEW 6 of 14

Figure 4.
Figure 4. The
The AUC
AUC of
of the ∆K varies
the ∆K varieswith
withwater
watersaturation.
saturation.
Figure 4. The AUC of the ∆K varies with water saturation.
3.1.3.
3.1.3. FF
FFSmith
Smith

3.1.3.
The FFSmith Figure 4. The AUC of the ∆K (FF
varies with water saturation.
Theperformances
performancesof ofthe
thetraditional
traditionalfluidfluidfactor
factor (FFSmith ))are areshown
shownin inFigure
Figure5.5. TheThe ROCROC curves
curves
of the FF in different water saturation scenarios is shown in Figure 5a. The performance is
3.1.3.FFFFSmithin different water saturation scenarios is shown in Figure 5a. The performance is
of the The performances
Smith of the traditional fluid factor (FF ) are shown in Figure 5. The ROC curves
significantly
of the FF reduced
significantly reduced compared
in different
comparedwaterwith the
the FF
saturation
with FFnew shown
scenarios
shown isin Figure
shown
in Figure in1a. The
The curves
Figure
1a. 5a. Thefor
curves different
different water
performance
for is
water
saturation The
significantly
saturation performances
plot near
plotreduced the of
random
near thecompared the traditional
random with guess line.
guessthe FF Theshown
fluid
line. The factor (FF
evaluation ) are shown
metrics,
in Figure
evaluation in
which
1a. The
metrics, which Figure
are
curves 5.
shown
are for
shown The ROC
in
different curves
Figure
water5b,
in Figure 5b,
of
indicate thethe
saturation FF best
plot in
near different
precision
the is
randomwater
less saturation
than
guess 0.8,
line. Thescenarios
while the is shown
recall,
evaluation the
metrics, in Figure
accuracy
which 5a.
and
are The
the
shown performance
F
in1 have
Figure worse
5b, is
indicate the best precision is less than 0.8, while the recall, the accuracy and the F1 have worse
significantly
performances
indicate
performances the than reduced
best
than the compared
theprecision.
precision is lesswith
precision. thanthe0.8,FFwhileshown in Figure
the recall, 1a. The curves
the accuracy and theforFdifferent
1 have worsewater
saturation
The AUCplot
performances isthannear
thethe
plotted inrandom
precision.
Figure 6. guess
Theline. Theare
values evaluation
around 0.65 metrics,
when which are shown
the water in Figure
saturation 5b,
ranges
indicate the best precision is less than 0.8, while the
from 0% to 90%. Then the AUC decreases with the increasing water saturation. recall, the accuracy and the F 1 have worse

performances than the precision.

Figure
Figure5.5.The
Theperformances of FF
performances of FF different water
in different watersaturation
saturationscenarios:
scenarios:
(a)(a)
ROCROC curves,
curves, (b) (b)
evaluationmetrics.
evaluation metrics.
Figure 5. The performances of FF in different water saturation scenarios: (a) ROC curves, (b)
Figure 5. The performances of FFSmith in different water saturation scenarios: (a) ROC curves,
The
The AUCisisplotted
evaluation
AUC plottedin
metrics. in Figure
Figure 6.
6. The
The values
values are
arearound
around0.65
0.65when
whenthe water
the water saturation
saturationranges
ranges
(b) evaluation metrics.
from
from 0%toto90%.
0% 90%.Then
Thenthe
theAUC
AUC decreases
decreases with
with the
theincreasing
increasingwater
watersaturation.
saturation.
The AUC is plotted in Figure 6. The values are around 0.65 when the water saturation ranges
from 0% to 90%. Then the AUC decreases with the increasing water saturation.

Figure6.6.AUC
Figure AUC curve
curve of traditional fluidfactor
traditional fluid factor(FF
(FFSmith
). ).
Figure 6. AUC curve of traditional fluid factor (FF ).
Figure 6. AUC curve of traditional fluid factor (FF ).
Appl. Sci. 2019, 9, 5239 7 of 14
Appl. Sci.
Appl. Sci. 2019,
2019, 9,
9, xx FOR
FOR PEER
PEER REVIEW
REVIEW 77 of
of 14
14

3.1.4. λρ
λρ
3.1.4. λρ
3.1.4.
The performances of λρ are shown in Figure 7. The ROC curves of the FFSmith is shown in
The performances
The performances of λρ are
of λρ are shown
shown inin Figure
Figure 7.7. The
The ROC
ROC curves
curves of
of the FF
the FF is shown
is shown in
in
Figure 7a. The performance in 0% water saturation scenarios is better than others. The evaluation
Figure 7a.
Figure 7a. The
The performance
performance in in 0%
0% water
water saturation
saturation scenarios
scenarios is
is better
better than
than others.
others. The
The evaluation
evaluation
metrics (Figure 7b) shows the precision decrease gently from 0.85 to 0.75 when the water saturation is
metrics (Figure 7b) shows the precision decrease gently from 0.85 to 0.75 when the water
metrics (Figure 7b) shows the precision decrease gently from 0.85 to 0.75 when the water saturation saturation
changed from 0% to 95%. The values of the recall, the accuracy and the F1 are less than 0.7.
is changed from 0% to 95%. The values of the recall, the accuracy and the F are less than
is changed from 0% to 95%. The values of the recall, the accuracy and the F1 are less than 0.7.
1 0.7.

7.7.The
Figure 7.
Figure TheThe
performances
performances
performances of 𝜆𝜌
of 𝜆𝜌 inλρ
ofin different water saturation
in different
different water saturation
water scenarios:
saturation (a) ROC
ROC curves.
scenarios:
scenarios: (a) curves.
(a) ROC(b) curves.
(b) metric
metric
parameters.
(b) metric parameters.
parameters.

The AUC
The is plotted
AUC is in Figure
plotted in 8. Its
Figure 8. value is
Its value
value isfrom
is from 0.7
from 0.7 to
0.7 to0.6
to 0.6when
0.6 whenthe
when thewater
the watersaturation
water saturation ranges
saturation ranges from
from
0 to 95%. Then the AUC decreases with the increasing water saturation.
0 to 95%. Then the AUC decreases with the increasing water saturation.

Figure 8. AUC curve of λρ.

Figure 8.
Figure 8. AUC
AUC curve
curve of 𝜆𝜌.
of 𝜆𝜌.
3.2. Uncertainty Analysis under Different Noise Conditions
3.2. Uncertainty
3.2. Uncertainty Analysis
Analysis under
under Different
Different Noise
Noise Conditions
Conditions
Noise is unavoidable in the analysis of rock physics. Hence, the uncertainty under noise conditions
Noise
is anNoise
essentialis unavoidable
is unavoidable
ability in the
the analysis
of the proposed
in analysis
methods.of rock
of rock physics.
Noisephysics. Hence, or
can be ambient
Hence, thesource
the uncertainty under
generated,
uncertainty under noise
coherent
noise
conditions
conditions
or random is is an essential
an essential
[20]. In a broaderability
ability of the
of the
sense, proposed
proposed
noise methods.
can be methods. Noise can be ambient
Noise can beofambient
from the uncertainty or source generated,
or source generated,
the measurement or seismic
coherent or
coherent
inversion or random
random
results. [20].
[20].
The real In data,
In aa broader
broader sense,well
sense,
including noise
noise log can
can be from
be
data, from the data
the
seismic uncertainty
uncertainty of the
of
and inversionthe measurement
measurement
data, cannot
or seismic
or
be seismic inversion results.
inversion
noise-free. results. TheThe real
real data,
data, including
including well well log
log data,
data, seismic
seismic data
data and
and inversion
inversion data,
data,
cannot
cannot be noise-free.
be noise-free.
Seismic inversion, simultaneous inversion in particular, is a useful tool in quantitative
Seismic inversion,
Seismic
interpretation, inversion,
although it simultaneous
simultaneous
was not involved inversion
inversion
in the in in particular,
particular,
previous is aa useful
is
discussion. useful tool in
tool
This technique in converts
quantitative
quantitative
the
interpretation,
interpretation,
data although it
although
from the reflectivity it(or
was
was not involved
not involved
interface) domain in to
in the
the theprevious
previous discussion.
discussion.
impedance This
domain,Thiswhich technique
technique converts
is moreconverts the
the
geologically
data
data from the reflectivity
from theSimultaneous
meaningful. (or interface)
reflectivity (orinversion
interface)candomain
domain
provideto the
to the impedance
impedanceon
information domain,
domain, which
which
the P-wave is more
is more
and S-wavegeologically
geologically
velocities
meaningful.
meaningful.
and the density Simultaneous
Simultaneous inversion
inversion
in a large area. can provide
can provide
The accuracy information on
information
of simultaneous on the P-wave
the
inversionP-wave and S-wave
and
is limited,S-wave velocities
velocities
although it can
and the density in a large area. The accuracy of simultaneous inversion
and the density in a large area. The accuracy of simultaneous inversion is limited, although it can is limited, although it can
achieve high accuracy in theory. Its accuracy is dominated by the quality
achieve high accuracy in theory. Its accuracy is dominated by the quality of the input pre-stack of the input pre-stack
Appl. Sci. 2019, 9, 5239 8 of 14

achieve high
Appl. Sci. accuracy in theory.
2019, 9, x FOR Its accuracy is dominated by the quality of the input pre-stack seismic 8 of 14
PEER REVIEW
data and the well log, the seismic-well tie, and the extracted wavelet together. In a real application,
the correlation
seismic data coefficient
and thebetween
well log, thethe
inverted and real
seismic-well tie,logand
maybe
the is not as high
extracted as thetogether.
wavelet expectation.
In a real
Hence, there are errors
application, between the
the correlation invertedbetween
coefficient data and thethe real value.
inverted and Thereal errors can be
log maybe regarded
is not as high asas the
noise expectation.
when applying the inverted data as the input.
Hence, there are errors between the inverted data and the real value. The errors can be
The commonly
regarded as noiseused unitapplying
when of noise level is decibel
the inverted (dB).
data By definition,
as the input. the noise level (NL ) can be
Enoise Anoise
derived byThethe commonly
ratio of energy (
used Eunit ) or amplitude (
of noise level is Adecibel ) between noise and
(dB). By definition, signal as follows:
the noise level (𝑁 ) can be
signal signal
derived by the ratio of energy ( ) or amplitude
! ( ) between! noise and signal as follows:
Enoise Anoise
NL = 10 log10 = 20 log10 , (13)
𝑁 = 10E𝑙𝑜𝑔 signal = 20A𝑙𝑜𝑔
signal , (13)
In thisInsection, the noise
this section, the levels
noise levels at −20
are setare set at −10 dB−10dB
dB,−20dB, and −7 dB−7dB
and which correspond
which the energy
correspond the energy
ratio is 0.01, 0.1 and 0.2. The noise analysis is performed with different water saturation conditions.
ratio is 0.01, 0.1 and 0.2. The noise analysis is performed with different water saturation conditions.
3.2.1. FFnew
3.2.1. FFnew
The ROC curves of FFnew under different noise levels are shown in Figure 9. The performances of
The ROC curves of FF under different noise levels are shown in Figure 9. The performances
the FFnew are high when the noise is relatively low (−20 dB), as shown in Figure 9a. The true prediction
of the FF are high when the noise is relatively low (−20 dB), as shown in Figure 9a. The true
ratio decreases when the noise is −20 dB. Furthermore, the performances are worse with increasing
prediction ratio decreases when the noise is −20 dB. Furthermore, the performances are worse with
noise at −10 dB and −7 dB, as shown in Figure 9b,c. Compared with the −20 dB case, the ROC curves
increasing noise at −10 dB and −7 dB, as shown in Figure 9b and c. Compared with the −20 dB case,
shift towards the random guess line. The ROC curve is close to the random line when the noise reaches
the ROC curves shift towards the random guess line. The ROC curve is close to the random line when
−7 dB. which indicates that the FFnew performance in a high-noise situation is poor. The evaluation
the noise reaches −7 dB. which indicates that the FF performance in a high-noise situation is poor.
metrics of the FFnew are shown in Figure 9d–f. The parameters are above 0.8 when the noise is −20 dB
The evaluation metrics of the FF are shown in Figure 9d, e and f. The parameters are above 0.8
and the water saturation is less than 80%. When the noise is −10 dB, the metrics decrease: the precision
when the noise is −20 dB and the water saturation is less than 80%. When the noise is −10 dB, the
is approximately 0.8, whereas the recall, accuracy and F1 are approximately 0.6 to 0.7. When the noise
metrics decrease: the precision is approximately 0.8, whereas the recall, accuracy and F1 are
is −7 dB, the parameters reduce continuously. The precision remains at approximately 0.8, whereas the
approximately 0.6 to 0.7. When the noise is −7 dB, the parameters reduce continuously. The precision
other three parameters are approximately 0.6. Note that the recall is close to 0.5, which indicates that
remains at approximately 0.8, whereas the other three parameters are approximately 0.6. Note that
the performance is near that of the random guess.
the recall is close to 0.5, which indicates that the performance is near that of the random guess.

FigureFigure
9. The9.ROCs
The ROCs
of theofFF newFFin theindifferent
the the different
noisenoise scenarios:
scenarios: (a) −20
(a) −20 dB,dB, −10−10
(b)(b) dB,dB, −7−7dB.
(c)(c) dB. The
metrics
The metrics ofof
the FFFF
the new in in
the the different
different noise
noise scenarios:
scenarios: (d)(d)
−20−20
dB,dB,
(e)(e)
−10−10 dB,
dB, (f)(f)
−7 −7 dB.
dB.

The AUC curves are plotted in Figure 10. The performances in the cases of −20 dB, −10 dB and
−7 dB are represented in blue, red and yellow, respectively. The AUC reaches 0.9 when the noise is
 Appl. Sci. 2019, 9, 5239 9 of 14

Appl. Sci.The
2019,AUC
9, x FOR PEERare
curves REVIEW 9 of 14
plotted in Figure 10. The performances in the cases of −20 dB, −10 dB and
−7 dB are represented in blue, red and yellow, respectively. The AUC reaches 0.9 when the noise is
low.
low.The AUCsdecrease
The AUCs decrease with
with increasing
increasing noise. noise. Theare
The AUCs AUCs are approximately
approximately 0.7 and 0.65, 0.7 and 0.65,
respectively,
respectively, when
when the noise the noise
is −10 −10dB.
dB andis−7 dB and −7 dB.

Figure 10. The AUC curves of the FFnew in the different noise scenarios: −20 dB, −10 dB and
Figure
−7 dB,10. The AUC curves of the FF
respectively. in the different noise scenarios: −20 dB, −10 dB and −7 dB,
respectively.
3.2.2. Delta K
3.2.2. Delta K curves of ∆K are shown in Figure 11. This parameter performs well when the noise is low
The ROC
(−20
ThedB),
ROCas shown ∆K are
curvesinofFigure 11a.shown
The ROC curve is11.
in Figure close toparameter
This the top left performs
corner, which
wellindicates
when the that the is
noise
prediction reaches a high performance. The TPR is approximately 0.6 even when
low (−20 dB), as shown in Figure 11a. The ROC curve is close to the top left corner, which indicatesthe water saturation
is high
that (98% and 99%).
the prediction reaches When
a high noise is −10 dB,
theperformance. Thethe TPRperformance of ROC0.6
is approximately is worse,
even when as shown in
the water
Figure 11b. The lowest TPR is 0.75 when the fluid is saturated hydrocarbon, whereas
saturation is high (98% and 99%). When the noise is −10 dB, the performance of ROC is worse, as the TPR values
range from 0.4 to 0.6 when the water saturation is higher than 80%. The lowest TPR reduces to 0.6 in
shown in Figure 11b. The lowest TPR is 0.75 when the fluid is saturated hydrocarbon, whereas the
the −7 dB noise situation (shown in Figure 11c) when the fluid is saturated hydrocarbon. For cases
TPR values range from 0.4 to 0.6 when the water saturation is higher than 80%. The lowest TPR
of high water saturation, the TPR is only 0.2 to 0.4. The evaluation metrics are plotted in Figure 11.
reduces to 0.6 in the −7 dB noise situation (shown in Figure 11c) when the fluid is saturated
The parameters are close to one when the water saturation is low, and the noise is −20 dB, as shown in
hydrocarbon. For cases of high water saturation, the TPR is only 0.2 to 0.4. The evaluation metrics are
Figure 11d. The results illustrate that the ∆K parameter has good performance when the noise is low.
plotted in Figure 11. The parameters are close to one when the water saturation is low, and the noise
In the scenario where the noise is −10 dB, as shown in Figure 11e, the parameters reduce to different
is −20 dB, as shown in Figure 11d. The results illustrate that the ∆K parameter has good performance
degrees. The precision remains above 0.9, whereas the others range mainly between 0.7 and 0.9. When
when the noise
the noise is the
is −7 dB, low.parameters
In the scenario where as
reduce further, theshown −10 11f.
noiseinisFigure dB, The
as shown
precision inisFigure
between 11e,
0.85the
parameters reducethe
and 0.9, whereas to others
different degrees.
range Thetoprecision
from 0.65 0.75. remains above 0.9, whereas the others range
mainlyThe between 0.7 and 0.9. When the noise is −7
comparison of the AUC curves in the scenarios where dB, the parameters
the noise isreduce
−20 dB,further,
−10 dB andas shown
−7 dB in
Figure 11f. The
is shown precision
in Figure is between
12. The 0.85 and
performance 0.9,decreases
of ∆K whereas the withothers rangenoise.
increasing from 0.65
The to 0.75.
approximate
ranges of the AUC are 0.97 to 1, 0.8 to 0.9 and 0.75 to 0.8 when the noise is −20 dB (blue), −10 dB (red)
and −7 dB (yellow), respectively.
Appl. Sci. 2019, 9, 5239 10 of 14
Appl. Sci. 2019, 9, x FOR PEER REVIEW 10 of 14

Figure 11. The ROCs of the ∆K in the different noise scenarios: (a) −20 dB, (b) −10 dB, (c) −7 dB. The
metrics of the ∆K in the different noise scenarios: (d) −20 dB, (e) −10 dB, (f) −7 dB.

The comparison of the AUC curves in the scenarios where the noise is −20 dB, −10 dB and −7 dB
is shown in Figure 12. The performance of ∆K decreases with increasing noise. The approximate
ranges of the AUC are 0.97 to 1, 0.8 to 0.9 and 0.75 to 0.8 when the noise is −20 dB (blue), −10 dB (red)
Figure The
11.11.
Figure ROCs
The the∆K
ROCsofofthe ∆Kininthe
thedifferent
differentnoise
noisescenarios:
scenarios: (a) −20 dB,
(a) −20 (b) −10
dB, (b) −10dB, (c)−7−7dB.
dB,(c) dB.The
and −7 dB (yellow), respectively.
The metrics of the ∆K in the different noise scenarios: (d) −20 dB, (e) −10 dB, (f) −7 dB.
metrics of the ∆K in the different noise scenarios: (d) −20 dB, (e) −10 dB, (f) −7 dB.

The comparison of the AUC curves in the scenarios where the noise is −20 dB, −10 dB and −7 dB
is shown in Figure 12. The performance of ∆K decreases with increasing noise. The approximate
ranges of the AUC are 0.97 to 1, 0.8 to 0.9 and 0.75 to 0.8 when the noise is −20 dB (blue), −10 dB (red)
and −7 dB (yellow), respectively.

Figure 12. The AUC curves of the ∆K in the different noise scenarios: −20 dB, −10 dB and −7 dB.
Figure 12. The AUC curves of the ∆K in the different noise scenarios: −20 dB, −10 dB and −7 dB.
3.2.3. FFSmith
3.2.3. FFSmith
The ROC curves of FFSmith under different noise conditions (−20 dB, −10 dB and −7 dB) are shown
The ROC
in Figure 13a–c,curves of FF Theunder
respectively. shapesdifferent
of ROC noise
curvesconditions
are similar,(−20 −10 dB and
dB, indicates
which −7 dB)
FFSmith are
is less
shown in Figure 13a, b and c, respectively. The shapes of ROC curves are similar,
affected by noise. However, the performance of FFSmith is poor that all the curves gather near the which indicates
FF
random is less line.
Figure
guess affected
12. The by
TheAUC noise.
evaluation the ∆K in
curvesHowever,
ofmetrics the
the
are performance
different
plotted in noiseof
Figure FF
scenarios:
13d–f. −20
is poor
The dB, −10that all
dB and
precision −7 dB.
the
is close curves
to 0.8
gather near the random guess line. The evaluation metrics are plotted in Figure
when the noise is −20 dB, while the other metrics are less than 0.7 as shown in Figure 13d. The results 13d, e and f. The
3.2.3. FFisSmith
precision close to 0.8 when the noise is −20 dB, while the other metrics are less than 0.7 as shown
in −10 dB and −7 dB (Figure 13e and f) are similar to the −20 dB scenario.
in Figure
The 13d.
ROC Thecurves −10 dB and
resultsofin FF −7 different
under dB (Figure 13e and
noise f) are similar
conditions −10−20
to the
(−20 dB, dBdBandscenario.
−7 dB) are
shown in Figure 13a, b and c, respectively. The shapes of ROC curves are similar, which indicates
FF is less affected by noise. However, the performance of FF is poor that all the curves
gather near the random guess line. The evaluation metrics are plotted in Figure 13d, e and f. The
precision is close to 0.8 when the noise is −20 dB, while the other metrics are less than 0.7 as shown
in Figure 13d. The results in −10 dB and −7 dB (Figure 13e and f) are similar to the −20 dB scenario.
 Appl. Sci. 2019, 9, 5239 11 of 14
Appl.
Appl. Sci.
Sci. 2019,
2019, 9, 9, x FOR
x FOR PEER
PEER REVIEW
REVIEW 1111
ofof
1414

Figure
Figure 13.13.
Figure The
13. The
The ROCs
ROCs
ROCs ofof the
the
of the FFSmithinin
FFFF in
thethe
the different
different
different noise
noise
noise scenarios:
scenarios:
scenarios: −20
(a)
(a)(a) −20
−20 dB,
dB, (b)−10
(b)(b)
dB, −10
−10 dB,
dB,
dB, −7−7
(c)(c)
(c) dB.
dB.
−7 dB.
The
The
The metrics
metrics
metrics ofof the
the
of FFSmithinin
theFFFF in the
the
the different
different
different noise
noise
noise scenarios:
scenarios:
scenarios:(d)(d)
(d) −20
−20
−20 dB,
dB,
dB, (e)(e)
(e) −10
−10
−10 dB,
dB, (f)
(f)(f)
dB, −7
−7−7 dB.
dB.
dB.

Thecomparison
The comparison
comparisonofofthe
oftheAUC
AUC
the AUC curvesininthe
curves
curves the
in scenarios wherethe
the scenarios
scenarios where thenoise
where noise
isis
the −20dB,
−20
noise dB, −10
is−10
−20dBdB
dB,and
and −7−7dBdB
−10
isis
andshown
−7 in
shown dBin Figure
Figure 14.
is shown
14. The performance
in performance
The Figure ofofFF
14. The FF
performance decreases
decreases FFwith
of with increasing
decreases
increasing
Smith noise.
with
noise. The
The approximate
increasing
approximatenoise.
rangesofofthe
theAUC
The approximate
ranges AUC are0.6
ranges
are 0.6toto
of the0.7
0.7 when
AUC
when arethe
the 0.6 noise
to 0.7
noise −20dBdB
isis−20
when the (blue),
(blue), is−10
noise−10 dBdBdB
−20 (red)
(red) and−7
(blue),
and −7dB
−10 (yellow),
dB(yellow),
(red) and
respectively.
−7 dB (yellow), respectively.
respectively.

Figure 14. The AUC curves of the FFSmith in the different noise scenarios: −20dB, −10 dB and −7 dB.
Figure
Figure 14.14.
TheThe AUC
AUC curves
curves ofof theFFFF
the inin
thethe different
different noise
noise scenarios:
scenarios: −20dB,
−20dB, −10
−10 dBdB −7−7
and
and dB.
dB.
3.2.4. λρ
3.2.4. λρλρ
3.2.4.
The ROC curves of λρ under different noise conditions (−20 dB, −10 dB and −7 dB) are shown in
The
The 15a–c,
Figure ROC curvesofofλρλρunder
ROCcurves
respectively. under
The different
different
ROC noiseconditions
curvesnoise
shift conditions
towards (−20
(−20
the dB,
random −10
dB,−10 dBdBand
guess lineand−7−7dB)
with dB)
the are
are shown
shown
increasing
inin Figure
Figure
noise level.15a,
15a, b band
The andc, c,respectively.
respectively.
evaluation metricsTheTheROC
are ROCcurves
given curves
in shift
shift
Figure towards
towards
15d–f. the
Thethe random
random
precision guessline
isguess
close toline with
with
0.8 andthe the
increasing
increasing noise
noise
other metrics level.
level.
are lessTheThe evaluation
evaluation
than metricsare
0.7. Themetrics
results inare given
given
−10 dBinin Figure
Figure
and 15d,
−7 15d,
dB e eand
andf.15e,f)
(Figure f.The
Theprecision
precision
are similarisisclose
toclose
the
to
to−200.8
0.8 dBand the
andscenario. other metrics are less than 0.7. The results in −10 dB and −7
the other metrics are less than 0.7. The results in −10 dB and −7 dB (Figure 15e and f) are
dB (Figure 15e and f) are
similartotothe
similar −20dBdBscenario.
the−20 scenario.
Appl. Sci. 2019, 9, 5239 12 of 14
Appl. Sci. 2019, 9, x FOR PEER REVIEW 12 of 14

Figure
Figure 15. TheROCs
15. The theλρ
ROCsofofthe λρininthe
thedifferent
differentnoise
noisescenarios: −20−20
scenarios:(a)(a) dB,dB, −10−10
(b) (b) (c) −7
dB, dB, −7 The
(c)dB. dB.
The metrics
metrics of the λρ in
of the in the
λρthe different
different noise
noise scenarios:
scenarios: −20−20
(d) (d) (e) −10
dB, dB, (e) −10 (f) −7
dB, dB, −7 dB.
(f)dB.

The
The comparison
comparison of of the
theAUC
AUCcurves
curvesin
inthe
thescenarios
scenarioswhere
wherethe noiseisis−20
thenoise dB,−10
−20dB, −10 dB and−7
dB and −7 dB
dB
is
is shown in Figure 16. The performance decreases with increasing noise. The approximate ranges of
shown in Figure 16. The performance decreases with increasing noise. The approximate ranges of
the
the AUC
AUC are
are 0.6
0.6 to
to 0.7,
0.7, 0.58
0.58 to
to 0.68
0.68 and
and 0.55
0.55toto0.65
0.65when
whenthe noiseisis−20
thenoise −20 dB (blue), −10
dB (blue), −10 dB
dB (red)
(red) and
and
−7 dB (yellow),
−7 dB (yellow), respectively.
respectively.

Figure 16. The AUC curves of the λρ in the different noise scenarios: −20dB, −10 dB and −7 dB.
Figure 16. The AUC curves of the λρ in the different noise scenarios: −20dB, −10 dB and −7 dB.
4. Conclusions
4. Conclusions
Two methods (FFnew and ∆K) have recently been proposed for hydrocarbon prediction. This
studyTwo methods
analyzes (FF
the uncertainty ∆K) have
andunder recently been
different-water proposed
saturation for hydrocarbon
and noise conditions ofprediction. This
these methods,
study analyzes the uncertainty under different-water
which are not included in the original works [3,11]. saturation and noise conditions of these
methods,
Both which
FFnew are
andnot ∆Kincluded in the
keep good original works
performance [3,11].
when water saturation is changed from 0% to
95%. Both FF
The values and
of the∆K keep good
related metricperformance
parameters when waterrecall,
(precision, saturation is changed
accuracy, from 0%
F-measure andtoAUC)
95%.
The values of the related metric parameters (precision, recall, accuracy, F-measure and AUC)
are greater than 0.9. On the one hand, it illustrates the stability of these methods in hydrocarbon are
greater than 0.9. On the one hand, it illustrates the stability of these methods in hydrocarbon
prediction even in a high water saturation scenario. On the other hand, they cannot distinguish low
Appl. Sci. 2019, 9, 5239 13 of 14

prediction even in a high water saturation scenario. On the other hand, they cannot distinguish
low water-saturation reservoirs from high water-saturation reservoirs which is a problem that the
industry is facing. A solution of high-water saturation identification is the combination of seismic and
controlled-source electromagnetic (CSEM) methods [21].
Noise is another essential factor. In the analysis of this study, the noise levels are set to be −20 dB,
−10 dB and −7 dB. The ∆K has good performance (the parameters are generally above 0.8) when the
level of the noise is low (−20 dB). The AUCs decrease with increasing noise. The AUCs remain above
0.85 when the noise is −10 dB and remain above 0.75 when the noise is −7 dB. The FFnew , which is in
the interface domain, is more sensitive to noise than the impedance-domain methods. The AUC is
approximately 0.7 to 0.75 and 0.6 to 0.65 when the noise is −10 dB and −7 dB, respectively. Although
the two attributes have different values for noise sensitivity, the trend is consistent, that is, the stronger
the noise, the worse the performance. The attributes in the interface domain are more sensitive to noise.
Noise is required to be suppressed to get good results in the application.
In addition, two widely used traditional methods (FFSmith and λρ) are analyzed as comparisons in
the reflectivity and impedance domains, respectively. FFnew and ∆K have much higher precision, recall,
accuracy and F1 compared to the traditional methods. FFnew and ∆K under high noise condition (−7 dB)
are still better than the traditional methods, even though FFSmith and λρ are relatively insensitive
to noise.

Author Contributions: Conceptualization, C.L.; methodology, C.L.; code, C.L.; writing—original draft preparation,
C.L.; writing—review and editing, C.L.; visualization, C.L.; project administration, D.G. and A.M.A.S.
Funding: This research and the APC were funded by Universiti Teknologi Petronas, grant number 015LC0-075.
Acknowledgments: The authors would like to appreciate Universiti Teknologi PETRONAS (UTP) and Petroliam
Nasional Berhad (PETRONAS) for all the resources provided. In addition, many thanks are given to our colleagues
in the CSI group.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Smith, G.C.; Gidlow, P.M. Weighted stacking for rock property estimation and detection of gas.
Geophys. Prospect. 1987, 35, 993–1014. [CrossRef]
2. Castagna, J.P.; Batzle, M.L.; Eastwood, R.L. Relationships between compressional-wave and shear-wave
velocities in clastic silicate rocks. Geophysics 1985, 50, 571–581. [CrossRef]
3. Liu, C.; Ghosh, D.P.; Salim, A.M.A.; Chow, W.S. A new fluid factor and its application using a deep learning
approach: A new fluid factor and its application. Geophys. Prospect. 2019, 67, 140–149. [CrossRef]
4. Liu, C.; Ghosh, D. A new seismic attribute for ambiguity reduction in hydrocarbon prediction: Seismic
attribute for reducing ambiguity. Geophys. Prospect. 2017, 65, 229–239. [CrossRef]
5. Ghosh, D.; Sajid, M.; Ibrahim, N.A.; Viratno, B. Seismic attributes add a new dimension to prospect evaluation
and geomorphology offshore Malaysia. Lead. Edge 2014, 33, 536–545. [CrossRef]
6. Ødegaard, E.; Avseth, P. Interpretation of Elastic Inversion Results Using Rock Physics Templates.
In Proceedings of the 65th EAGE Conference & Exhibition, Stavanger, Norway, 2–5 June 2003.
7. Goodway, B.; Chen, T.; Downton, J. Improved AVO Fluid Detection and Lithology Discrimination Using Lamé
Petrophysical Parameters; “λρ”, “µρ” & “λ/µ Fluid Stack”, from P and S Inversions; SEG Technical Program
Expanded Abstracts 1997; Society of Exploration Geophysicists: New Orleans, LA, USA, 1997; pp. 183–186.
8. Jing, B.; Ren, J.; Qin, X. Inversion of Reservoir Properties: Quantitative Hydrocarbon Seismic Identification in Tight
Carbonate Reservoirs; SEG Technical Program Expanded Abstracts 2015; Society of Exploration Geophysicists:
New Orleans, LA, USA, 2015; pp. 2693–2697.
9. Chopra, S.; Sharma, R.K.; Grech, G.K.; Kjølhamar, B.E. Characterization of shallow high-amplitude seismic
anomalies in the Hoop Fault Complex, Barents Sea. Interpretation 2017, 5, 607–622. [CrossRef]
10. Liu, X.; Chen, Z.; Liu, H.; Zhang, W. Oil Detection Based on Layer Buried-Depth Corrected Elastic Inversion
in Deepwater of the Pearl River Basin. In Proceedings of the International Petroleum Technology Conference,
Beijing, China, 26–28 March 2019.
Appl. Sci. 2019, 9, 5239 14 of 14

11. Liu, C.; Ghosh, D.; Salim, A.M.A.; Chow, W.S. Fluid Discrimination Using Bulk Modulus and Neural Network.
In Proceedings of the International Petroleum Technology Conference, Beijing, China, 26–28 March 2019.
12. Freitas, C.J. The issue of numerical uncertainty. Appl. Math. Model. 2002, 26, 237–248. [CrossRef]
13. Kroese, D.P.; Brereton, T.; Taimre, T.; Botev, Z.I. Why the Monte Carlo method is so important today.
Wiley Interdiscip. Rev. Comput. Stat. 2014, 6, 386–392. [CrossRef]
14. Bosch, M.; Cara, L.; Rodrigues, J.; Navarro, A.; Díaz, M. A Monte Carlo approach to the joint estimation of
reservoir and elastic parameters from seismic amplitudes. Geophysics 2007, 72, O29–O39. [CrossRef]
15. Zunino, A.; Mosegaard, K.; Lange, K.; Melnikova, Y.; Hansen, T.M. Monte Carlo reservoir analysis combining
seismic reflection data and informed priors. Geophysics 2015, 80, R31–R41. [CrossRef]
16. Yu, S.; Ma, J.; Osher, S. Monte Carlo data-driven tight frame for seismic data recovery. Geophysics 2016, 81,
V327–V340. [CrossRef]
17. Zhu, D.; Gibson, R. Seismic inversion and uncertainty quantification using transdimensional Markov chain
Monte Carlo method. Geophysics 2018, 83, R321–R334. [CrossRef]
18. Mavko, G.; Mukerji, T.; Dvorkin, J. The Rock Physics Handbook: Tools for Seismic Analysis in Porous Media;
Cambridge University Press: Cambridge, UK, 2009.
19. Fawcett, T. An introduction to ROC analysis. Pattern Recognit. Lett. 2006, 27, 861–874. [CrossRef]
20. Avseth, P.; Mukerji, T.; Mavko, G. Quantitative Seismic Interpretation: Applying Rock Physics Tools to Reduce
Interpretation Risk; Cambridge University Press: Cambridge, UK, 2010.
21. Ghosh, D.; Halim, M.; Brewer, M.; Viratno, B.; Darman, N. Geophysical issues and challenges in Malay and
adjacent basins from an E & P perspective. Lead. Edge 2010, 29, 436–449.

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).