Paper LPSS PDF

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Journal of Applied Geophysics 159 (2018) 511–521

Contents lists available at ScienceDirect

Journal of Applied Geophysics

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

Application of LP and ML sparse spike inversion with probabilistic neural


network to classify reservoir facies distribution - A case study from the
Blackfoot field, Canada
S.P. Maurya, N.P. Singh ⁎
Department of Geophysics, Institute of Science, Banaras Hindu University, Varanasi, 221005, India

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

Article history: Sparse-Spike inversion techniques are used to estimate distribution of acoustic impedance in inter well region,
Received 11 April 2018 the important parameters for characterizing the reservoir facies from seismic and well log data. The purpose of
Received in revised form 15 September 2018 sparse spike impedance inversion is to obtain high-resolution impedance profile of the subsurface from the
Accepted 19 September 2018
low resolution seismic data with the integration of well log data and enhance the interpretation of the prospec-
Available online 25 September 2018
tive zone. In the present study, two types of sparse spike inversion techniques, namely, Linear Programming (LP)
Keywords:
sparse spike inversion and Maximum Likelihood (ML) sparse spike inversion are applied to estimate acoustic im-
Linear programing pedance from seismic data of the Blackfoot region, Alberta, Canada. The principle objective of the study is
Maximum likelihood assessing the relative performance of these techniques for estimation of petrophysical parameters to identifica-
Neural network tion of prospective zones in the area. Initially, the inversion methods are applied to the composite trace near to
Petrophysical analysis. well locations and inverted for acoustic impedance and compared with the actual impedance derived from the
well log data. The result demonstrates that both curves are matching with each other very well. The correlation
is estimated to be 0.97 and 0.93, Synthetic relative error (SRE) 0.23 and 0.34 and Root mean square (RMS) errors
are 1125 m/s*g/cc and 1205 m/s*g/cc for Linear Programming sparse spike inversion (LPSSI) and Maximum Like-
lihood sparse spike inversion (MLSSI), respectively. The analysis for composite traces depicts the robustness and
performance of the algorithm. Thereafter, the techniques are applied to seismic volume to estimate variation of
acoustic impedance in inter well regions. The analysis of inverted impedance shows a low impedance anomaly in
between 1060 and 1075 ms time intervals which may be due to presence of reservoir facies (sand channel). The
analyses of the inverted results reiterate that both methods work satisfactory and show variation of reservoir fa-
cies in similar way. The results found by LPSSI shows slightly higher resolution compared to the MLSSI results.
Thereafter, to enhance reservoir facies more clearly, porosity is predicted in inter well region by using probabilis-
tic neural network (PNN) technique. The result shows very high porosity (N 15%) in between 1060 and 1075 ms
time interval which corroborated with the low impedance zone and confirms the presence of sand channel. The
qualitatively and quantitatively analysis of inversion results suggest that the LPSSI along with PNN provides bet-
ter reservoir characterization than MLSSI and PNN combination for the Blackfoot field, Alberta, Canada.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction methods assume simple reflectivity model of the subsurface from the
well log data and synthetic traces are generated using convolution the-
The sparse spike inversion methods though developed in 80s (Zhang ory. Thereafter, the error is minimized between synthetic trace and seis-
et al., 2016) have now become the common methods to estimate mic trace by adding more and more spikes in the reflectivity series
petrophysical parameters. The sparse spike inversion is based on as- (Debeye and Riel, 1990). The sparse spike inversion techniques are
sumption that the earth's reflectivity series are composed with large divided into two categories on the basis of minimization of error. The
spikes with small gaussian spikes in the background. The methods first method is called Linear Programming sparse spike inversion
also assume that these large spikes are only meaningful which corre- (LPSSI) methods that use l1 norm solution for its implementation and
sponds for deposition gap and ignore background spikes (Bosch et al., the second is called Maximum Likelihood sparse spike inversion
2010; Maurya and Sarkar, 2016; Maurya and Singh, 2015a). The (MLSSI) methods that use l2 norm solution for its implementation
(Russell, 1988; Sacchi and Ulrich, 1995; Zhang and Castagna, 2011).
⁎ Corresponding author. The sparse spike inversions aim to estimate acoustic impedance
E-mail address: singhnpbhu@yahoo.co.in (N.P. Singh). in inter well region from the seismic data with the integration of

https://doi.org/10.1016/j.jappgeo.2018.09.026
0926-9851/© 2018 Elsevier B.V. All rights reserved.
512 S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521

well log data. The inverted impedance displays a blocky structure of 2. The study area - Blackfoot field
the subsurface in time domain, which is directly related to the lithol-
ogy of the subsurface (Goutsias and Mendel, 1986; Wang, 2010). The In the present study, 3D seismic data from the Blackfoot field,
bandwidth of seismic data is limited and has generally 12 to 80 Hz Alberta, Canada is used for the analysis. The Blackfoot field is located
and hence, does not contain lower and higher frequency content south-east direction of Strathmore city, Alberta, Canada. The data is
but these are very meaningful particularly lower frequency. To get acquired by Pan Canadian Petroleum and the CREWES (the Consortium
a broadband spectrum of the inverted impedance volume, additional for Research in Elastic-Wave Exploration Seismology) in 1995 (Dufour
information is added from the well log data (Maurya and Singh, et al., 2002; Farfour et al., 2016). The data is recorded by using 708
2015a). These inverted results are used to classify reservoir facies shots into a fixed recording spread of 690 channels (Margrave et al.,
of the region but for the confirmation of these facies some additional 1998). The principle objective of the acquisition is to demonstrate 3
parameters are needed. In present study, porosity is predicted in components recording in this area. The data is acquired in two patches,
inter well region for the confirmation of reservoir facies in the Black- the first patch targeted the clastic Glauconitic channel, and the second
foot field, Canada. patch went deeper to study the reef-prone Beaver hill lake carbonates
Porosity is one of the key parameters associated with oil and gas (Lawton et al., 1996). Thereafter, the seismic data processing has been
reservoirs. Determination of this petrophysical parameter is an performed by Pulsonic Geophysical and Sensor Geophysical in 1996.
essential step to classify reservoir facies distribution. Among differ- The bandwidth of the data is 12–90 Hz for the vertical-component
ent linear and nonlinear prediction tools such as multi-regression, and 12–50 Hz for the horizontal component. The present study focuses
polynomial curve fitting, artificial neural network and Probabilistic in clastic Glauconitic channel. The details of the processing steps used
Neural Network has gained the attention of researchers over the for the vertical and horizontal component data can be found in Simin
past few years. Probabilistic Neural Network (PNN) methods are et al. (Simin et al., 1996). Fig. 1 shows the stratigraphy of Alberta,
being routinely followed to predict various geophysical parameters Canada. The time slice and location of wells are shown in inset of figure
from seismic and well log data (Bhatt and Helle, 2002). The PNN and depicts high and low of the subsurface. For the analysis purpose, the
method uses LPSSI and MLSSI derived impedances as external 3D seismic data from the first patch (Glauconitic formation) and 13 well
attributes and seismic derived attributes as internal to predict log data are used.
petrophysical parameters in inter well region (Doyen, 1988). The The Glauconite formation is of the Lower Cretaceous age which rep-
PNN methods uses sample points taken at different locations and in- resents sediment filled incised valley of the upper Manville Group
terpolates in the seismic section where log data are not available. (Miller et al., 1995). The grain size is fine to medium of the Glauconite
These sample points are measurements of petrophysical parameters Formation. On the other hand, eroded Mississippian carbonates are cov-
in the boreholes and attributes estimated directly or indirectly from ered with Lower Cretaceous sediments. These sediments are the detrital
the seismic data (Haas and Dubrule, 1994; Hampson et al., 2001; member of variable thickness, while above this Detrital member there
Mahmood et al., 2017; Maurya and Sarkar, 2016; Russell et al., are sheet and ribbon Sunburst sands (Dufour et al., 2002). At the Creta-
1997). In the present study, the two major issues are addressed, the ceous time, marine transgression deposited brackish shales, limestone
first issue is finding reservoir facies distribution (which is productive and quartz sands and silts to build the Ostracod member. The Glauco-
sands channels), and the second issue is defining clearly the bound- nitic member consists of shales and sands of lacustrine and channel
aries of these facies using above discussed techniques. The above origin. Within the channel, the sediments are subdivided into three
discussed techniques are implemented on standard seismic data units corresponding to three phases of valley incision with different
interpretation software i.e. Hampson Russell and the results are quality of sand deposited. These three units may are not be encountered
presented using Matlab Programming language. everywhere in the area (Dufour et al., 1998). The detailed stratigraphic
column of the Blackfoot field can be found in Margrave et al., (1998).

Fig. 1. The figure describes the stratigraphic features of the Blackfoot field, Alberta. In inset the time slice at 1065 ms time interval shows high and low of the subsurface along with 13 well
locations.
S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521 513

3. Methodology If in the Eq. (6), one choose p = 1 or 2, the equation gives a solution
of l1 norm or l2 norm respectively. In this way one can maximize the
In the present study, two types of seismic inversions, namely Linear objective function E and can obtain a solution that gives least error
Programming sparse spike inversion (LPSSI) and Maximum Likelihood between the model parameters and the observation (Li, 2001). The
sparse spike inversion (MLSSI) techniques are utilized to estimate the solution of l1 norm can be written as follows.
distribution of acoustic impedance in inter well region and characterize
 Z t 
the reservoir facies. Thereafter, these acoustic impedances are trans-
ξðt Þ ¼ ξð0Þ exp 2 r ðt Þdt ð7Þ
formed into the porosity section using Probabilistic Neural Network 0
algorithm. These techniques are explained briefly in the following
sections. From the above relation, One may notice that the sparseness in r(t)
may result in a blocky structure in ξ(t) (Oldenburg et al., 1983).
3.1. Linear programming inversion These equations provide a linear relationship between the Fourier
transform of the reflectivity function, the wavelet and the reflection
The theory used in the present study of Linear Programming sparse coefficients. The algorithm minimizes a linear objective function with
spike inversion is developed by Li (Li, 2001). First, the seismic inversion l1 − norm subject to given linear constraints in equality or in inequality
estimate earth's reflectivity series r(t) by using seismic data s(t) and form. The above principle is applied to estimate variation of acoustic
wavelet w(t) as input. Thereafter, these reflectivity series are trans- impedance in the subsurface under LPSSI techniques.
formed for acoustic impedances ξ(t) which indicate the subsurface lith-
ological structure. The earth's model can be written as follows: 3.2. Maximum likelihood inversion

Lm þ n ¼ d ð1Þ The basic theory of maximum likelihood inversion is depends on


maximum likelihood deconvolution that estimate earth's reflectivity
where L is the operator, the data point is written as d = (x1, x2, …) and, series from the seismic data and then it is transformed into the acoustic
the model parameters m = (r1, r2, …). n is the random noises in the data. impedance under Maximum Likelihood inversion. The acoustic imped-
The target of seismic inversion is to estimate subsurface impedance ance shows blocky structure of the subsurface and enhances seismic
model from the seismic reflection data. In Eq. (1), the parameter d is interpretation. The Maximum likelihood deconvolution assumes that
known, and then the subsurface model can be defined by the probability the earth's reflectivity is composed of series of large spike embedded
p(m| d) (Li, 2001; Sacchi and Ulrich, 1995). This probability can be small gaussian spikes in the background. These large spikes indicate
described by the Bayes formula as following: the deposition gap in the subsurface and the method target these
large spikes to estimate. The model of the seismic trace can be written
pðdjmÞpðmÞ
pðmjdÞ ¼ ð2Þ as:
pðdÞ
Sðt Þ ¼ r ðt Þ  W ðt Þ þ nðt Þ ð8Þ
where p(d| m) is the probability or likelihood of obtaining the data, p(m)
is the prior probability of the model, p(d) is the data likelihood and X
t
enters into the problem as a normalization factor, and p(m| d) is the pos- Sðt Þ ¼ r ð jÞW ðt−jÞ þ nðt Þ; t ¼ 1; 2; …; N ð9Þ
terior probability of the model. One could use the MAP solution, mMAP j¼1

that maximize a posterior probability p(m| d). The objective function


can be written as: where S(t) is seismic trace, r(t) is earth's reflectivity, W(t) is the source
wavelet and denotes convolution operator. From the assumption about
E ¼ − logðpðmjdÞÞ ¼ − logðpðdjmÞÞ− logðpðmÞÞ ð3Þ subsurface model, one can minimize the objective function for the opti-
mum solution i.e. earth reflectivity series. The objective function E can
In Eq. (2), p(d) is a constant and hence omitted for the simplicity. be written as
Our prior knowledge of the model is generally given as a global con-
straint S(m). Thereafter, we use the principle of maximum entropy to X
t
r 2 ð jÞ X
t
n2 ð jÞ
E¼ 2
þ −23 ln ðλÞ−2ðt−3Þ ln ð1−λÞ ð10Þ
compute the prior probability (Sacchi and Ulrich, 1995). J¼1 R J¼1 N2
Consider continuous model parameters m with probability density
function p(m). The entropy h given by where λ is the likelihood that a given sample has a reflection (Russell,
Z 1988; Zhang and Castagna, 2011). The solution of deconvolution with
h¼ pðmÞ log½pðmÞdm ð4Þ minimum number of spikes in its reflectivity series and lowest noise
component that matches well with the seismic traces is considered as
the optimum solution. There may be possible a number of solutions
Eq. (4) expresses the uncertainty associated with the distribution of
which could match with the seismic traces therefore single most likely
parameter p(m). If the information about the parameter m is available in
addition algorithm (SMLA) is used to constraints the solution.
the form of some global constraint S(m) then, the corresponding maxi-
mum entropy probability distribution will be given as follows:
3.3. Probabilistic neural network
−SðmÞ
pðmÞ ¼ Ae ð5Þ
The probabilistic neural network (PNN) is a mathematical interpola-
tion scheme which uses neural network architecture for its implemen-
where A is normalization constant. The generalized version of p(d| m)
tation. The methods analyze variety of attributes estimated directly
can be written as a function of discrepancy between the model and
from the seismic data and attributes that is estimated indirectly i.e. in-
the observation as follows:
version derived impedance and best attributes are selected for further
! analysis (Bosch et al., 2010). Thereafter, these selected attributes are
p
p1−1=p −1 jd−Lmj
pðdjmÞ ¼ exp  p ð6Þ cross plotted with the predicted well log i.e. porosity in this case. Fur-
2σ p Γð1=pÞ p σp ther, a best fit nonlinear equation gives the relation between porosity
and seismic attributes that is used for predicting porosity in the seismic
514 S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521

one output layer. So total 15 connections have been made to get output
logs and hence we have 15 weights to estimate.

4. Results

In present study, two types of sparse spike inversion namely Linear


Programing and Maximum Likelihood inversions are utilized to esti-
mate acoustic impedance in inter well region using seismic and well
log data from the Blackfoot field, Alberta, Canada. Initially, the logs
from well 08–08 are analyzed and the special features are highlighted
in Fig. 3. The Glauconitic formation is highlighted at 1580 m depth
which is target zone of the present study. Apart from these, three
major coal formations near to 1500 m depth are also identified along
Fig. 2. Neural network architecture. with Mannville formation at 1430 m depth.
After Hampson et al. (Hampson et al., 2001). Thereafter, the composite seismic traces are extracted near to well
locations and the LPSSI and MLSSI techniques are applied to these traces
section (Bhatt and Helle, 2002). The entire process is carried out sepa- for optimizing parameters and comparing inversion results with actual
rately for all the sample points and porosity is predicted. The new log AI from the well log. The comparison of inverted impedance with actual
value can be estimated as: impedance is shown in Fig. 4. The inverted results for trace near to
a) well 01–17, b) well 08–08 and c) well 16–08 are shown. Although
X
n
the comparison is performed for all the thirteen wells but the results
Li expð−Dðx; xi ÞÞ
i¼1
are shown only for three wells for purpose of simplicity. From Fig. 4, it
LðxÞ ¼ P n ð11Þ is noticed that, the inverted AI are in well agreement with the actual
i¼1 expð−Dðx; xi ÞÞ
AI from the well log data. The only difference noticed is that the AI
n 
X  from LPSSI is more close to the actual AI compare to the MLSSI results.
x j −xij 2
Dðx; xi Þ ¼ ð12Þ Fig. 5 depicts cross plots of inverted AI with the actual AI from the
σj
i¼1 well log data. Fig. 5a shows cross plot for LPSSI results whereas Fig. 5b
shows cross plot for MLSSI results for well 08–08. The figure shows
The quantity D(x, xi) is the distance between the input point and that the distribution of data points are very close to the best fitted line
each of the training point's xi (Mahmood et al., 2017; Masters, 1995). for both the cases and hence both methods produces AI in similar way
The ‘distance’ is scaled by the quantity σj, called the smoothing param- with high accuracy. The only differences noticed is that the data points
eter, which may be different for each independent variable. The target of are more close to the best fit line for LPSSI case as compare to the
training process of the probabilistic neural network (PNN) is to estimate MLSSI case which suggests that the LPSSI techniques estimate AI more
the smoothing parameters σj. The basic idea behind the optimization accurately as compared to the MLSSI techniques. Fig. 6 shows error anal-
technique is to minimize validation error between predicted value and ysis of inverted results for all the thirteen wells. Fig. 6a shows compari-
actual value. The prediction can be defined as followings. son of synthetic correlation with wells. The Fig. 6a depicts that the
correlation coefficients for LPSSI techniques gives higher value com-
X
n
pared to the MLSSI case. Fig. 6b shows synthetic relative error (SRE)
yi expð−Dðxm ; xi ÞÞ
whereas Fig. 6c shows root mean square error (RMSE). The Fig. 6 dem-
y0m ðxm Þ ¼ P
i≠m
n ð13Þ onstrates that the SRE is lower for LPSSI case compare to the MLSSI case
i≠m expð−Dðxm ; xi ÞÞ
and the RMSE shows lower value for LPSSI case compare to the MLSSI
The predicted log value at the mth sample point is given by ym′ case. The total correlation coefficient for LPSSI technique is 0.97 whereas
(Eq. (13)). As the original value ym from the well log is known at mth for MLSSI technique, it is 0.83. The total RMSE for LPSSI technique is
sample point and hence the validation error can be calculate at the mth found to be 1125 m/s*g/cc while for MLSSI technique, it is 1205
sample point using following formula. m/s*g/cc for all the wells. These statistical parameters demonstrate
that the LPSSI is more accurate and produces higher resolution com-
 2 pared to the MLSSI case.
em ¼ ym −y0m ð14Þ
After getting satisfactory results for the inversion of composite trace
near to the well locations, both techniques (LPSSI and MLSSI) are ap-
If the data contains n sample then the total validation error can be plied to entire seismic section from the Blackfoot field, Alberta,
calculate as followings. Canada. Using these techniques, the AI is estimated in inter well region.
A cross-section of inverted acoustic impedance at inline 41 is shown in
X
n
em ¼ ðyi −y0i Þ2 ð15Þ Fig. 7. Fig. 7a shows inverted acoustic impedance estimated using LPSSI
i¼1 techniques whereas Fig. 7b shows inversion results for MLSSI tech-
niques. The analysis of these inverted results depicts that both the inver-
Thereafter, by using conjugate-gradient technique, the validation sion methods provide subsurface AI model in a similar way but the
error is minimized with respect to the smoothing parameters σj. The impedance contrast is greater for LPSSI results as compared to the
basic neural network architecture for the present study is shown in MLSSI results. A low-impedance anomaly is noticed in between
Fig. 2. The network consist one input layer, one or more hidden layer 1060 ms to 1075 ms time interval which may be due to presence of
and one output layer. The architecture consist a set of neurons that are sand channel in this zone. Thereafter, these impedances are trans-
arranged into two or more layers. The input and output layer each con- formed into the porosity section using probabilistic neural network
taining at least one neuron. Let's four attributes A1, A2, A3 and A4 are techniques.
used to predict log values than these attributes make input layer The probabilistic neural network is applied to the Blackfoot post
(Fig. 2) which is connected to the three neurons of hidden layer. Each stack seismic data using LPSSI and MLSSI derived acoustic impedances
input is connected to the each hidden neurons. Each connection repre- as external attributes and using direct derived attributes from the seis-
sents a weight. Further, each neurons of hidden layer is connected to the mic data as internal attributes to predict porosity in inter well region.
S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521 515

Fig. 3. Well log plot (Well 08–08) where the gamma ray log (GA), density log, velocity (VP and VS) log, porosity, impedance and shale volumes are shown from left to right. The special
features are highlighted.

A verity of internal attributes is generated and best attributes are se- and then validation is performed. We hide wells and use the trained
lected on the basis of correlation and errors. The desired attributes are neural network to predict their values. Since all thirteen wells are
shown in Table 1. Table 1 shows that four attributes are suitable for used for the training process, hence each training wells are kept hidden
the LPSSI case and five attributes are suitable for MLSSI case respec- one by one and predicted porosity values using remaining wells. Fig. 8
tively. First, the algorithm is trained using the above discussed attributes depicts comparison of predicted porosity for particular well by hiding

Fig. 4. The inverted acoustic impedance from LPSSI (in blue) and MLSSI (in red). The black curve shows original AI from a) well 01–08, b) well 08–08 and c) well 16–08. (For interpretation
of the references to colour in this figure legend, the reader is referred to the web version of this article.)
516 S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521

Fig. 5. Cross-plots between actual impedance and inverted impedance estimated by using a) LPSSI and b). MLSSI techniques for well 08–08.

Fig. 6. a) The variation of correlation coefficient, b) Synthetic relative error and c) total error with wells. The dark black histogram shows for MLSSI case and gray colour histograms shows
for LPSSI case.

them and using remaining wells. Fig. 8a shows comparison for LPSSI
case whereas Fig. 8b depicts comparison for MLSSI case. The correlation
is 0.76 and 0.70 and error is 3.82% and 4.03% for LPSSI and MLSSI cases
respectively. The analyses are shown for well 04–16, 05–16 and 08–08
only for simplicity. Fig. 9 shows comparison of validation plot for both
cases. The Validation Plot shows the RMS error as a function of the
wells. From both the case the errors are in acceptable accuracy for
almost all the wells.
Thereafter, the PNN is applied to the composite trace near to well lo-
cations to predict porosity and to optimize the parameters. Fig. 10
shows predicted porosity from LPSSI and MLSSI with PNN combination
with each other. The comparison of predicted porosity and well log
porosity from well 01–17, 08–08 and 16–08 are shown in Fig. 10a, b
and c, respectively. The Fig. 10 depicts that the predicted porosity
from LPSSI and PNN combination and MLSSI and PNN combinations
are matching very well with the well log porosity for all the wells. The
only difference is noticed that the predicted porosity from LPSSI and
PNN combinations are extremely in well accordance with the well log
porosity whereas the predicted porosity estimated by MLSSI and PNN
combination shows a small deviation with well log porosity, although
both techniques successfully predicted porosity which corroborated
well with the well log porosity. The cross plot of predicted porosity
with actual porosity is shown in Fig. 11 for well 08–08 to check the qual-
ity control of the inverted results. Fig. 11a shows cross plot of predicted
Fig. 7. Cross-section (inline 41) of acoustic impedance estimated using a) LPSSI and porosity generated by LPSSI and PNN combination whereas Fig. 11b
b) MLSSI techniques. The anomaly is highlighted by the rectangle. shows crossplot for predicted porosity estimated by using MLSSI and
S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521 517

Table 1
List of attributes used for the prediction of porosity using PNN as prediction tool.

Target Linear Programming (LP) Maximum Likelihood (ML)

Porosity 1/inverted Zp 5.230816 5.331997 1/inverted Zp 5.309830 5.456217


Porosity Time 5.103231 5.234985 Amp. Envelope 5.172046 5.371833
Porosity Int. Abs. Amp. 5.040044 5.198439 Filter 25/30–35/40 5.102282 5.331982
Porosity Filter15/20–25/30 5.013860 5.189903 X-coordinate 5.038378 5.333245
Porosity Int. Abs. Amp. 4.989610 5.308743

PNN combinations. The distribution of scattered points show that the Fig. 13 shows slices at 1060 ms two way travel time demonstrating
inverted porosity is very close to the actual porosity and shows high cor- the horizontal distribution of acoustic impedance and porosity.
relation value. The average correlation coefficient for LPSSI and PNN Fig. 13a shows impedance slice for LPSSI case whereas Fig. 13b shows
combination is 0.76 whereas; it is 0.70 for the MLSSI and PNN impedance slice for MLSSI case and Fig. 13c and d shows porosity slices
combination. estimated by LPSSI and PNN combinations and MLSSI and PNN combi-
After getting satisfactory results from the composite trace, the PNN nations, respectively. These slices show horizontal distribution of the
technique is applied to 3D seismic datasets to predict porosity in inter reservoir. The both impedance/porosity slices show approximately the
well region. Fig. 12 shows cross section of predicted porosity (inline same distribution of low/high values and hence variation of reservoir.
41). Fig. 12a shows cross-section of predicted porosity estimated using The only difference is that the contrast of the reservoir is poorer for
LPSSI and PNN combinations whereas the Fig. 12b shows cross-section MLSSI case compared to the LPSSI case. It is noticed that the reservoir
of predicted porosity estimated using MLSSI and PNN combinations. varies from southeast (SE) to northwest (NW) direction in the
The analysis of predicted porosity sections show a high porosity anom- subsurface.
aly in between 1060 ms to 1075 ms time interval which is corroborated
well with low impedance anomaly. This zone is characterized as reser- 5. Discussions
voir facies (sand channel). The maximum porosity in the reservoir
zone is estimated to be 21% for LPSSI and PNN combinations and 19% The results estimated by LPSSI and MLSSI along with PNN combina-
is estimated for MLSSI and PNN combinations. tions are similar but in some area both methods giving conflicting vari-
ations. To study this dissimilarity, difference between AI and porosity
estimated by LPSSI and MLSSI cases have been taken and shown in
Fig. 14. Fig. 14a shows impedance difference whereas Fig. 14b shows
porosity difference estimated by LPSSI and MLSSI with PNN combina-
tion, respectively. It is noticed that there are relatively huge differences
between AI volume estimated by LPSSI and MLSSI techniques particu-
larly after 1100 ms time interval. These differences arises due to depen-
dency of inversion methods on well log data and are not available
bellow 1100 ms time interval and hence both inversion methods giving
average AI contrast in this zone. The porosity section from both the
cases is also show similar results with large differences at later time
intervals. It is noticed that the impedance differences are higher than
the porosity differences. This is because the inversion directly utilized
seismic and well log data whereas PNN mostly uses seismic data infor-
mation in its implementations. The huge differences between these sec-
tions are highlighted by the ellipse in Fig. 14.
Further, the amplitude spectrum of inverted synthetic traces from
LPSSI and MLSSI are taken and plotted with amplitude spectrum of
Blackfoot seismic data (Fig. 15). The Fig. 15 depicts that the amplitude
spectrum of inverted synthetic trace follow the trend of Blackfoot seis-
mic amplitude spectrum very well for both the cases. The only differ-
ence observed is that the amplitude spectrum of inverted synthetic
trace from LPSSI technique is more accurate and almost overly to the
Blackfoot amplitude spectrum whereas for the MLSSI case, the ampli-
tude spectrum shows small deviation from the Blackfoot amplitude
spectrum particularly for larger frequency. This match demonstrate
that the LPSSI keeps all the frequency contents as it is during its process,
on the other hand the mismatch of amplitude spectrum for MLSSI case
shows that the methods distort the frequency contents during its imple-
mentation, particularly at larger frequencies. This is small disadvantages
of MLSSI methods as compared with LPSSI. The correlation coefficient of
amplitude spectrum for LPSSI case is 0.99 and RMS error is 0.025
whereas for MLSSI case the correlation coefficient is 0.93 and RMS
error is found to be 0.077.
The slice of volume difference has been generated and shown in
Fig. 16. Fig. 16a shows impedance slice of volume difference whereas
Fig. 16b shows porosity slice of volume difference. The larger differences
Fig. 8. Validation of PNN for a) LPSSI case using 4 attributes and b) MLSSI case using 5 between them are highlighted by the ellipse. LPSSI and MLSSI methods
attributes. estimated the low-impedance values ranging from 6500 to 9000 (g/cc)*
518 S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521

Fig. 9. Variation of validation error with well logs.

Fig. 10. Depicts the geostatistical analysis for well a) 01–16, b) 08–08 and c) 16–08. The estimated porosity using LPSSI and PNN combination is shown by blue solid line and the porosity
estimated by MLSSI and. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) PNN combination is shown by red solid
line.

Fig. 11. Cross plot of predicted porosity against actual porosity estimated using probabilistic neural network and a) LPSSI combination and b) MLSSI combination for well 08–08.
S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521 519

Fig. 12. Porosity cross-section (inline 41) estimated using probabilistic neural network
and a) LPSSI combination and b) MLSSI combination. The anomaly zone is highlighted
by the rectangle.

(m/s) within the time interval of 1060 - 1075 ms which is interpreted as


reservoir zone (sand channel). The estimated impedances using LPSSI
algorithm showed a higher resolution of the reservoir zone compared Fig. 14. Volume difference of a) inverted AI estimated using LPSSI and MLSSI and
to those derived from MLSSI algorithm. The average correlation coeffi- b) volume difference of predicted porosity estimated using LPSSI and PNN combination
and MLSSI and PNN combination. The larger differences are highlighted by the ellipse.
cients for all the thirteen wells are found to be 0.97 for LPSSI case and
0.93 for MLSSI approach. The RMS error varied from 600 (m/s)*(g/cc)
to 1500 (m/s)*(g/cc) with a mean value of 1125 (m/s)*(g/cc) for LPSSI plot for MLSSI and PNN combinations are shown in Fig. 17b. In these fig-
method and is 1205 (m/s)*(g/cc) for MLSSI method. Thereafter, a ures the reservoir zone is highlighted by ellipse which corresponds to
cross plot of inverted acoustic impedance and predicted porosities is low acoustic impedance and high porosity values. The only difference
generated and shown in Fig. 17. The AI and porosity cross plot for can be seen in these crossplots are the larger number of data points
LPSSI and PNN combinations are shown in Fig. 17a whereas the cross fall within the highlighted zone for MLSSI case compared to the LPSSI

Fig. 13. Slices at 1065 ms two way travel time of acoustic impedance calculated from a) LPSSI and b) MLSSI technique. The porosity slices at TWT 1065 ms derived using c) LPSSI and PNN
combination and d) MLSSI and PNN combination. The sand channel trend is highlighted by the ellipse.
520 S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521

techniques to predict petrophysical parameters in inter well region.


Fig. 17 also shows a relation between predicted porosity and predicted
acoustic impedance in the region. Both the methods give similar type
of relations with minor differences. The average relation for the Black-
foot field is given as follows.

AI ¼ 3:05ϕ2 −275ϕ þ 1315 ð16Þ

These equations can be used to estimate porosity from the AI and


vice versa for quick interpretation. Table 2 illustrates quantitative differ-
ences among the results of seismic inversions and PNN methods. In
Table 2, column 1 represents parameters to be compare, column 2
shows impedance analysis results, column 3 and column 4 depicts po-
rosity analysis results and amplitude spectrum for Linear Programming
Fig. 15. Amplitude spectra of seismic section (black) and the inverted impedance derived
sparse spike inversion whereas column 5, 6 and 7 depicts impedance,
synthetics section using LPSSI (in red) and MLSSI (in blue) algorithms. (For interpretation
of the references to colour in this figure legend, the reader is referred to the web version of porosity and Amplitude spectrum, respectively for Maximum Likeli-
this article.) hood inversion. The Table 2 implies significant quantitative difference
among LPSSI, MLSSI and Probabilistic Neural Network. The table implies
case which suggest that the LPSSI has higher contrast of reservoir. The that for most steps of seismic inversion and geostatistical methods,
highest porosity is estimated in the channel region and it is nearly 21% LPSSI and Probabilistic Neural Network display higher quantitative
for LPSSI and PNN combinations and 19% for MLSSI and PNN combina- values and hence indicate a better prediction technique together. This
tions. The analysis shows that the LPSSI and PNN together is the best is also concluded from the analysis that if the input seismic data

Fig. 16. slices of volume differences for a) inverted AI and b) porosity. The larger differences are highlighted by the ellipse.

Fig. 17. Crossplot of inverted impedance and predicted porosity estimated using a) LPSSI and PNN combination and b) MLSSI and PNN combination. The reservoir zone is highlighted by the
ellipse. The relation between AI and porosity is shown on top of the figure.
S.P. Maurya, N.P. Singh / Journal of Applied Geophysics 159 (2018) 511–521 521

Table 2
Quantitative comparison of inverted results (The PSNR stands for peak signal to noise ratio, MAE stands for mean absolute error, SAE stands for sum absolute error, RMSE stands for root
mean square error).

Parameters Linear programming (LP) Maximum likelihood (ML)

Impedance Porosity Amplitude spec. Impedance Porosity Amplitude spec.


3 3
Reservoir 6.5–8.5 × 10 21% NA 6.3–9.5 × 10 19% NA
Correlation 0.97 0.83 0.99 0.93 0.76 0.93
PSNR −13.18 38.35 89.65 −15.24 80.11 39.57
MAE 949.41 2.53 0.005 1172 2.18 0.015
SAE 6.17 × 104 164.16 1.37 7.62 × 104 141.84 3.86
Norm 9.41 × 103 24.96 0.14 1.19 × 104 21.68 0.41
RMSE 1.17 × 103 3.10 0.008 1.48 × 103 2.69 0.025

contains less number of reflectors (as Blackfoot data) then the LPSSI Bosch, M., Mukerji, T., Gonzalez, E.F., 2010. Seismic inversion for reservoir properties com-
bining statistical rock physics and geostatistics: a review. Geophysics 75 (5)
gives more accurate and high resolution image of the subsurface as (75A165-75A176).
compared to the MLSSI case. Debeye, H., Riel, V.P., 1990. Lp - norm deconvolution. Geophys. Prospect. 38, 381–403.
Doyen, P.M., 1988. Porosity from seismic data - a geostatistical approach. Geophysics 53,
1257–1263.
6. Conclusions Dufour, J., Goodway, B., Shook, I., Edmunds, A., et al., 1998. Avo analysis to extract rock pa-
rameters on the Blackfoot 3c-3d seismic data. 68th Ann. Int. SEG Mtg, pp. 174–177.
The present research describe, two type of Sparse Spike inversion Dufour, J., Squires, J., G, W., et al., 2002. Integrated geological and geophysical interpreta-
tion case study, and lame rock parameter extractions using AVO analysis on the
techniques, namely the Linear Programming and Maximum likelihood Blackfoot 3c-3d seismic data, southern Alberta, Canada. Geophysics 67 (1), 27–37.
inversion along with the Probabilistic Neural network to estimate the Farfour, M., Yoon, W.J., Jang, S., 2016. Energy-weighted amplitude variation with o set: a
AI and porosity in inter well region using seismic data with the integra- new AVO attribute for low impedance gas sands. J. Appl. Geophys. 129, 167–177.
Goutsias, J., Mendel, J.M., 1986. Maximum-likelihood deconvolution: an optimization the-
tion of well log data of the Blackfoot field, Alberta, Canada. The principle
ory perspective. Geophysics 51, 1206–1220.
objective of the study is assessing the relative performance of these Haas, A., Dubrule, O., 1994. Geostatistical inversion - a sequential method of stochastic
techniques for estimation of petrophysical parameters to identification reservoir modelling constrained by seismic data. First Break 12 (11), 561–569.
of prospective zones in the area. The results of seismic inversion sug- Hampson, D.P., Schuelke, J.S., Quirein, J.A., 2001. Use of multiattribute transforms to pre-
dict log properties from seismic data. Geophysics 66 (1), 220–236.
gested that the LPSSI and MLSSI techniques provide mutually consistent Lawton, D.C., Robert, R., Stewart, A.C., Hrycak, S., 1996. Design review of the Blackfoot 3c-
impedance contrast. Both methods estimated the low-impedance 3d seismic program. CREWES Ann. Res. Rep. 8 (39), 1–23.
values ranging from 6000 to 9000 (g/cc)*(m/s) within the time interval Li, Q., 2001. Lp Sparse Spike Inversion. Strata Technique Document, Hampson-Russell
Software Services Ltd.
of 1060 - 1075 ms which is interpreted as reservoir zone (sand chan- Mahmood, M.F., Shakir, U., Abuzar, M.K., Khan, M.A., Khattak, N., Hussain, H.S., Tahir, A.R.,
nel). The estimated acoustic impedances using LPSSI algorithm showed 2017. Probabilistic neural network approach for porosity prediction in balkassar area:
a higher impedance contrast of the reservoir zone compared to those a case study. J. Himal. Earth Sci. 50 (1).
Margrave, G.F., Lawton, D.C., Stewart, R.R., 1998. Interpreting channel sands with 3C–3D
derived from MLSSI algorithm. The average correlation coefficient is es- seismic data. Lead. Edge. 17 (4), 509–513.
timated to be 0.97 for LPSSI case and 0.93 for MLSSI case. Masters, T., 1995. Advanced Algorithms for Neural Networks: A C++ Sourcebook. John
Predicted porosities derived using probabilistic neural network pro- Wiley & Sons, Inc.
Maurya, S., Sarkar, P., 2016. Comparisons of post stack seismic inversion methods: a case
vides more accurate and high resolution porosity sections in inter well study from Blackfoot field, Canada. Int. J. Sci. Eng. Res. 7 (8), 1091–1101.
region for the cases when LPSSI derived acoustic impedances are used Maurya, S., Singh, K., 2015a. LP and ML sparse spike inversion for reservoir
as external attributes in comparison to the cases when MLSSI derived characterization-a case study from Blackfoot area, Alberta, Canada. 77th EAGE Con-
ference and Exhibition 2015.
acoustic impedances are used. The generated porosity cubes show sim-
Miller, S., Aydemir, E., Margrave, G., 1995. Preliminary Interpretation of PP and PS Seismic
ilar high porosity values that corroborate with the low impedance zone Data from the Blackfoot Broad-Band Survey (Technical Report. CREWES Research
derived from the seismic inversion. The maximum porosity in the sand Report).
channel is estimated to be 21% for the LPSSI and PNN combination Oldenburg, D., Scheuer, T., Levy, S., 1983. Recovery of the acoustic impedance from reflec-
tion seismograms. Geophysics 48 (10), 1318–1337.
whereas 19% is estimated for MLSSI and PNN combination. The results Russell, B.H., 1988. Introduction to Seismic Inversion Methods. Vol. 2. Society of Explora-
suggest that the combination of LPSSI and PNN approach together can tion Geophysicists, Tulsa.
generate a more reliable estimate of the petrophysical properties of Russell, B., Hampson, D., Schuelke, J., Quirein, J., 1997. Multiattribute seismic analysis.
Lead. Edge 16 (10), 1439–1444.
the subsurface. Sacchi, M.D., Ulrich, T.J., 1995. High resolution velocity gathers and o set space reconstruc-
tion. Geophysics 60 (4), 1169–1177.
Acknowledgements Simin, V., Harrison, M.P., Lorentz, G.A., 1996. Processing the Blackfoot 3c-3d seismic sur-
vey. CREWES Res. Rep. 8 (39), 1–11.
Wang, Y., 2010. Seismic impedance inversion using l 1-norm regularization and gradient
Author (SPM) is indebted to Science and Engineering Research descent methods. J. Inverse Ill-Posed Prob. 18 (7), 823–838.
Board, Department of Science and Technology, Government of India, Zhang, R., Castagna, J., 2011. Seismic sparse-layer reflectivity inversion using basis pursuit
decomposition. Geophysics 76 (6), R147–R158.
New Delhi for financial supports in form of research project (Grant no Zhang, Q., Yang, R., Meng, L., Zhang, T., Li, P., 2016. The description of reservoiring model
PDF/2016/000888) under National Post-doctoral Fellowship scheme. for gas hydrate based on the sparse spike inversion. 7th International Conference on
Authors also acknowledge the CGG Veritas for providing seismic and Environ-Mental and Engineering Geophysics and Summit Forum of Chinese Academy
of Engineering on Engineering Science and Technology. Atlantis Press, pp. 94–97.
well log data of Blackfoot field, Alberta, Canada.

References
Bhatt, A., Helle, H.B., 2002. Committee neural networks for porosity and permeability pre-
diction from well logs. Geophys. Prospect. 50 (6), 645–660.

You might also like