Você está na página 1de 12

t

Tools, techniques, and tutorials

Quantitative assessment of the seismic net-pay method: A case study


Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Ramses G. Meza1, Juan M. Florez1, Stanislav Kuzmin1, and John P. Castagna2


Abstract
We applied the seismic net-pay (SNP) method to an oil discovery and predicted thicknesses consistent with
the actual thicknesses at the wellbore locations. This was accomplished by applying the method in a self-calibrating mode that did not require the direct use of well information. For net-pay estimation under a self-calibration scenario, the SNP method thickness estimates proved to be more accurate (mean absolute prediction
error at well validation locations under 3.0  1.5 m) than estimates from a reflectivity-based detuning method
(4.0  2.7 m) or multiple linear regression (5.9  5.8 m). Statistical F-tests indicated that the correspondences
of the predicted thickness estimates with actual net-pay values for the SNP and reflectivity methods (F approximately 5.56 for both) were statistically significant, whereas the multiple regression results did not prove to be
statistically significant.

Introduction
Accurate prediction of layer thicknesses from surface seismic data away from well control is important
for volumetric calculations, reservoir characterization,
well planning, well steering, and identification of additional drilling opportunities. This is particularly challenging for thin layers as defined by Widess (1973).
In this case, the interpreted or apparent time thickness/isochron of the thin layer is not representative
of the actual time thickness of the layer, which means
that the apparent thickness can be misleading. For a
thin layer, the amplitude of the associated composite
reflection event is not only dependent on the impedance contrast but also on the actual layer thickness
(Widess, 1973; Kallweit and Wood, 1982) and provides
an avenue for thickness estimation.
The amplitude variation with thickness, or tuning
curve, is largely dependent on the wavelet embedded
on the seismic data (Brown et al., 1986) as well as
the sign and symmetry of the reflection coefficients
(RFC). Practitioners have deployed several approaches
to mitigate or decouple the effect of the wavelet from
the analysis of seismic data at a particular target, such
as spectral decomposition (Okaya, 1995; Partyka et al.,
1999), spectral inversion (Puryear and Castagna, 2008),
spectral shaping techniques such as colored inversion
(Lancaster and Whitcombe, 2000), and spectral blueing
(Neep, 2007). These approaches aim to reduce the tuning effect on subsets of the seismic volume, so the interpreter can map seismic events associated to top and

base of the target bed and extract amplitude and traveltime information more directly associated with the
layer properties.
Brown et al. (1984, 1986) combine mapping top and
base event information with a modeled tuning curve to
obtain thickness estimates for thin and thick layers
based on reflectivity data. In their method, an envelope
was fit to the scattered points on the composite amplitude versus isochron crossplot. This envelope represents the reflectivity of a clean sand with a 100% net
to gross (NTG) and defines the tuning curve that could
be calibrated using wellbore information, if well data
are available. The method also relies on the assumption
that for any apparent thickness value, the composite
amplitude to tuning curve ratio would be proportional
to the NTG.
Connolly (2007) proposes a similar map-based approach using a relative-impedance (after colored inversion) volume, instead of a reflectivity volume, as the
basis for mapping the top and base horizons of the target layer (zero-crossings). This is known as the seismic
net-pay (SNP) method. The detuning approach is also
very similar to that in the reflectivity domain, resulting
in calibrated relative-impedance values and reduced
tuning imprint on the outcome. The technique is based
on the following assumptions:

The section of interest is represented by a suite


of constant-impedance reservoir layers embedded
in a matrix of constant-impedance nonreservoir
rocks.

1
Quantitative Interpretation Team, BHP Billiton Petroleum, Houston, Texas, USA. E-mail: lordtable@yahoo.com; mauricio.florez@
bhpbilliton.com; stanislav.kuzmin@bhpbilliton.com.
2
University of Houston, Department of Earth and Atmospheric Sciences, Houston, Texas, USA. E-mail: jcastagnaou@yahoo.com.
Manuscript received by the Editor 31 October 2014; revised manuscript received 29 January 2015; published online 15 April 2015. This paper
appears in Interpretation, Vol. 3, No. 2 (May 2015); p. B25B36, 19 FIGS.

http://dx.doi.org/10.1190/INT-2014-0241.1. 2015 Society of Exploration Geophysicists and American Association of Petroleum Geologists. All rights reserved.

Interpretation / May 2015 B25

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

The apparent thickness is less than one halfcycle of the lowest frequency component of the
wavelet.
The reservoir must be seismically isolated.

Once these assumptions are met, then the average


relative impedance for a given apparent thickness is
proportional to the seismic NTG (true net pay divided
by apparent thickness). This variable is obtained from
the average relative impedance versus apparent isochron chart in a similar way to that of the reflectivity-based method mentioned earlier (Brown et al.,
1984, 1986).
The SNP methods binary impedance requirement
can be achieved using extended elastic impedance
(EEI) (Whitcombe et al., 2002), which is in essence a
coordinate rotation of the acoustic impedance (AI)
and gradient impedance (GI) volumes obtained from
colored inversion (Connolly, 1999) of the amplitudeversus-angle (AVA) intercept and gradient-amplitude
data, respectively. On relative-impedance domain, this
AI-GI coordinate rotation has the following form:
LnEEI LnAI cos LnGI sin.

(1)

This AI-GI coordinate rotation is controlled by the


angle , which is selected to optimize the discrimination
of a particular reservoir property, such as lithology
(sand versus shales) or fluid content (hydrocarbon versus brine) in a section of interest (Whitcombe and
Fletcher, 2001). The SNP method would yield net sand
or net pay outcomes depending on the property that the
input EEI volume has been optimized for, either lithology or fluid, respectively.
Simm (2009) makes a comparison of reflectivitybased and SNP methods using modeled data consisting

Figure 1. Interpreted horizon for the top of the reservoir target layer in depth, overlaid by production wells with complete
penetration on the reservoir (black stars).
B26 Interpretation / May 2015

of layers of sand with varying thickness embedded in a


shale background. He generates the characteristic seismic response of geologically plausible layering with
different NTG values. He finds that the SNP method
yields more accurate net-pay predictions than the reflectivity-based method, provided that the rock and
fluid properties do not deviate from the methods assumptions above.
The objective of this case study is to assess whether
the technical effort associated with yielding thickness
estimates via SNP may be justified by obtaining more
accurate and reliable net pay/sand predictions in exploration scenarios as claimed by Connolly (2007) and
Simm (2009). The technical effort required for carrying
out the SNP method consists mainly of performing colored inversions of prestack data, selecting the optimal
relative-impedance volume for either net sand or net
pay based on EEI, and the accurate horizon picking
of top and base of the layer of interest on the zero crossings of the corresponding relative-impedance volume.
The accuracy of the SNP method was compared with
conventional tuning analysis by evaluating the prediction accuracy of both methods using the actual net
pay penetrated in the wells, under the assumptions
stated above.
Methodology
Seismic analysis and rock physics
The hypothesis that the SNP method could accurately determine layer thickness from a measured seismic response was tested using data from an oil field
located on the northwest shelf of offshore Australia.
The reservoir target was a Cretaceous high-porosity,
low-impedance, deep marine turbidite sandstone embedded in a thick, high-impedance, and somewhat
homogeneous deepwater shale section. Normal faulting
is the dominant structural style of the hydrocarbon
traps, and the vast majority of the drilling targets in
the area for exploration and development were identified using anomalous seismic amplitude measurements
interpreted as direct hydrocarbon indicators. The interpreted depth horizon for the top of the reservoir layer is
shown in Figure 1, overlaid by seven production wells
(shown as black stars) with complete penetration of the
reservoir.
Semiregional rock-physics analysis (Duncan et al.,
2013) has established that the presence of hydrocarbon-bearing sands results in class III AVA behavior
(Rutherford and Williams, 1989; Castagna and Swan,
1997). Forward modeling indicates that the AVA gradient does not significantly change with fluid content
for given lithology and porosity conditions, which are
assumed to be fairly uniform in the area. The AVA intercept is significantly sensitive to the pore fluid change
as observed on the half-space modeling in Figure 2.
Six exploratory wells were used for regional rockphysics analysis, in addition to the seven production
wells within the oil field. The production wells all
had a complete penetration of the target bed for SNP

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

validation purposes. Overall good-quality, AVA-complicollocated seismic traces for AI, GI, and EEI
ant prestack time-migrated (PSTM) seismic data sets
25, show that seismically EEI 25 does not imwere used in this study, along with a complete suite
prove the illumination of the pay layer sufficiently that
of wireline wellbore logs, which were used for performit can be interpreted as a unique event in the window
ing well-seismic ties, such as the example shown in Figof interest. Further angle scanning was performed
ure 3. This example is representative of data quality and
with no improvement on the pay illumination when
semiregional framework, depicting an isolated highNTG sand as the target bed (yellow arrow) with easily
identified top and base reflections at or below tuning.
The PSTM processing on the data set included postmigration coherent noise attenuation, spectral balancing,
and residual moveout (RMO), producing angle gathers
that yielded four angle stacks whose central angles
are 8, 18, 26, and 33. Gathers were not available at
the time of this study; therefore, pseudogathers were
constructed by resorting the traces of each angle stack
into a common depth point (CDP) gather ensemble of
four fold, with each trace corresponding to the central
angle cited above. Figure 4 shows a vertical cross section
containing these pseudogathers displayed every fifth
CDP, with the central angle of each trace going from left
to right on each pseudogather. Overlying a collocated
gather is the zero-offset synthetic trace previously shown
in Figure 3. The seismic events corresponding to the target layer are enclosed by the green polygon.
Figure 2. Modeled AVA half-space response for a shale overThe AVA intercept and gradient stacks were generlying brine-saturated (blue) and oil-saturated (green) reserated from the pseudogathers, and then tied and colored
voir sandstone (after Duncan et al., 2013).
inverted using an elastic-impedance approach (Connolly, 1999) to design a global inversion
operator per each stack as shown in Figure 5. The figure shows the amplitude
spectra of the AI logs from the seven
semiregional wells and the linear fit to
their mean. This fit represents the target
spectra for designing the colored inversion operator used to shape the intercept stack to match the band-limited
AI observed in the wellbore data. The
same procedure was performed to design the corresponding operator for
band-limited GI. These steps aim at
obtaining AI and GI volumes.
Rock-physics modeling suggested
that there will be no angle capable
of yielding a clear EEI volume for lithology discrimination; but a EEI 25
volume was expected to enhance the
discrimination between hydrocarbonbearing and brine-bearing sands, while
minimizing the sensitivity to porosity.
Figure 6 shows the same well log as
shown in Figure 3, this time depicting
the logs for AI, GI, and EEI 25, filFigure 3. (a) Semiregional wellbore S-1 depicting pertinent logs (shale volume
tered using a trapezoidal filter 10/1885
V sh , water saturation S w , total porosity PHIT, P-wave velocity V P , bulk density
\100 Hz to match the overall bandwidth
rhob, AI, and reflection coefficients RC), zero-offset synthetic trace (blue), and
of the seismic data. The band-limited
collocated actual near-angle stack trace (red). The polarity convention is SEG
logs show that the EEI 25 volume
normal. The target sand is highlighted by the yellow arrow. (b) Wavelet for well
should improve the illumination of the
tie in the time domain, wavelet spectrum, and crosscorrelation function indicating a correlation coefficient higher than 80% and no apparent time shift.
pay layer. However, the corresponding
Interpretation / May 2015 B27

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

compared with the AI volume (AI EEI 0). This


raised uncertainty about the quality of the AVA gradient
as derived from the pseudogathers. Because of the lack
of a good lithology projection and uncertainty in the
quality of the AVO gradient, it was decided to perform
the SNP exercise using the AI volume.
Seismic net-pay estimation
Connolly (2007) performs three 1D forward models
that hold sand and shale elastic properties constant,
only varying the total net sand thickness among the
models. In this study, we use the same net sand relative
distribution to build wedge models as seen in Figure 7,

to obtain a first-hand understanding of the actual implementation of the method under controlled conditions.
In this modeling case, the total true net sand thickness
changes with true gross thickness, but the models were
built in such way that the true NTG is held constant.
Model A (top) has NTG 1, model B (middle) has
four sands that combined yield a NTG 0.57, and
model C (bottom) has two sands that combined yield
a NTG 0.47.
The top and base horizons can be used to calculate
the average relative impedance and apparent time
thickness for each model (Figure 8). The average relative impedance is proportional to the underlying NTG

Figure 4. Vertical section of AVA pseudogathers, displayed every fifth CDP. Incidence angle (8, 18, 26, and 33) increases from
left to right on each pseudogather. Collocated zero-offset synthetic trace of well S-1 is shown in red. Seismic events associated to
target layer are enclosed by the green polygon. The data polarity convention is SEG normal.

Figure 5. (a) Spectral response of the AI logs


for all the semiregional wells over the window
of interest and (b) mean impedance spectral
response, target for colored-inversion operator design.

B28 Interpretation / May 2015

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

for any given time thickness up to 2530 ms, in agreedicted values closely match the actual ones for models
ment with the SNP method assumption related to the
A and B with reduction of accuracy for larger gross inmaximum thickness for which this proportionality is
tervals. In the case of model C, the accuracy is also
clearly met, in this case approximately 27.5 ms (oneseverely affected by the internal layering of the pay,
half wavelength of the lowest usable frequency on data,
which in this case becomes less evenly distributed
which is approximately 18 Hz).
The SNP method takes advantage
of this relationship between relative
impedance and NTG to estimate netpay thickness for varying true NTG. If
it is assumed that the envelope of scattered points on the relative impedanceapparent thickness plane represents the
response of an expected maximum NTG
of one, relative-impedance values between that envelope and zero will provide an estimate of the actual NTG.
Seismic NTG for wedge models A, B,
and C are shown in Figure 9, illustrating
convergence to the true NTG beyond
tuning thickness. The seismic NTG becomes considerably lower as the apparent time thickness becomes lower than
the tuning thickness. By interpreting
Figures 8 and 9, it can be concluded that
the net-pay estimation must include a
proportionality or scaling function to be
applied to the apparent thickness that
depends on the average relative impedFigure 6. The same well as shown in Figure 3, this time including band-limited
logs and the respective collocated seismic traces for AI, GI, and EEI 25.
ance (Connolly, 2007; Simm, 2009).
The target pay layer is highlighted by the yellow arrow.
Such a correction or scaling function
is obtained from wedge modeling for a
maximum NTG:
corrt

modeled seismic NTGt


.
modeled AIt

(2)

The correction function is dependent on the wavelet


and is approximately linear, as depicted in Figure 10
for model A. This function allows obtaining an estimation of net pay for thin and thick layers based on Simms
(2009) proposed SNP formula:
SNPx; y corrt AIx; y zx; y.

(3)

For each mapped x; y point on the seismic survey,


the SNP will depend on the arithmetic product of the
correction factor corr for the apparent time thickness
t at x; y, the corresponding average relative impedance AI and the apparent isopach z at the same
x; y pair.
If this scaling function based on wedge model A can
be calculated and calibrated to closely match predicted
versus actual net-pay thickness, then this correction
function for maximum seismic NTG can also be applied
to models B and C to verify the accuracy of the SNP
method for varying true NTG. Figure 11 depicts predicted (y-axis) versus actual (x-axis) net-pay thickness
for models A (top), B (middle), and C (bottom). Pre-

Figure 7. Wedge models. Model A NTG 1, model B


NTG 0.57, and model C NTG 0.47. The background gray
color represents higher impedance shales, in yellow the lower
impedance sandstones, and the overlying traces are the modeled band-limited impedance traces for the wedges. The red
and blue horizons represent the apparent top and base of the
wedges as interpreted on the zero crossing of the band-limited
impedance traces.
Interpretation / May 2015 B29

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 8. Average band-limited impedance versus apparent


time thickness for model A (red), model B (cyan), and model
C (green).

within the apparent gross interval (Connolly and Kemper, 2007).


The SNP method requires accurate interpretation of
top and base horizons of the target layer on the AI volume at the corresponding zero crossing as seen in Figure 12. These horizons form the basis for the input of
the SNP method because they will be used to extract
the average relative impedance, apparent isochron,
and isopach maps (using an interval velocity volume),
which are the basis of the impedance tuning curve for
seismic NTG, as defined by Connolly (2007) and as depicted in Figure 13. Similarly to the reflectivity-based
method, a relative-impedance tuning curve is obtained
from a wedge model based on a trapezoidal filter representative of the AI volume bandwidth and scaled to
the envelope of the scattered points on the average relative impedance-isochron plot. This tuning curve will

Figure 9. Seismic NTG (NTGs) versus apparent time thickness for model A (red), model B (cyan), and model C (green).

Figure 10. Correction or scaling factor versus apparent time


thickness for model A.
B30 Interpretation / May 2015

Figure 11. Predicted (y-axis) versus actual (x-axis) net-pay


values (ms) for models A (top), B (middle), and C (bottom)
based on calibration of model A.

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

represent the maximum seismic NTG at each apparent


time thickness value.
For any kth point on a crossplot such as the one
shown in Figure 13, the main component of the detuning is the calculation of its seismic NTG:
seismic NTGk

AIk
;
Btk

(4)

are present in some cases and some are not intercepted


by the perfect fit line as expected, which implies that
some flaws on the current deployment might be affecting the quality of the predictions. For self-calibration,
the mean absolute prediction error at well validation
locations is under 3.0  1.5 m. The parameterization
on this self-calibration tuning curve is then used to

where AIk is the measured relative AI and Btk is the


relative-impedance value of the modeled tuning curve
evaluated at tk . Once the seismic NTG is calculated,
then it undergoes the application of the apparent-thickness correction factor obtained from the modeled tuning curve.
In cases in which well data are available, the calibration process is basically controlled by adjusting the
scaling factor of the modeled tuning curve, such that
the AIk B ratio, as shown in Figure 13, becomes proportionately closer to the actual NTG at the well.
The general workflow for SNP based on Connolly
(2007) and Simm (2009) is depicted in Figure 14. Several of the steps can be easily performed on modern
workstation and commercial software applications already available for interpreters and seismic analysts.
Results
The SNP thickness estimates were validated using
the actual net pay measured at seven wellbores that
penetrated the target bed located within the oil-leg of
the reservoir. The first validation scenario assumed that
no well calibration to actual thickness has been performed during the detuning process on SNP, defined
by Connolly (2007) as self calibration. Where there is
no local well control, such as in near-field exploration
or early field appraisal projects, the application of the
SNP method can have a big impact on the net-pay estimation. Figure 15 shows such a validation scenario:
the average AI-apparent isochron plot (left) depicts
the scattering of points as plotted from the input attribute horizons. A modeled tuning curve for impedance
was created with the main controlling parameters being
the wavelet bandwidth (a 10/1885\100 Hz trapezoidal
filter) and a wavelet scaling factor; the latter is adjusted
in such a way that the modeled tuning curve represents
an overall envelope to the scattered data points and
also the maximum seismic NTG 1. Also in this plot,
the validation wells are plotted based on their input
attribute values. The predicted versus actual net pay
(right) provides a graphical insight of the quality of
the prediction at well locations. The y-axis error bars
were defined by the SNP range around each wellbores
neighboring CDPs (in this case, eight neighbor CDP
bins). The x-axis error bars were defined by the netpay error obtained from the petrophysical evaluation.
Neither over- nor underprediction of significance is observed, so it is reasonable to state that the overall relative net-pay estimation is capturing the actual net-pay
values. It can be observed that large y-axis error bars

Figure 12. Seismic section (westeast) of relative AI with


semiregional S-1 well and major bounding faults overlaid. Target layer is highlighted by the yellow arrow and embedded by
the interpreted top (green) and base (magenta) horizons.

Figure 13. Schematic chart depicting the main input elements involved on the SNP method: Modeled band-limited
impedance wedge model tuning curve for maximum NTG
(dashed cyan), apparent thickness, and average band-limited
impedance maps as obtained from the horizon picking process at the target event on the band-limited impedance volume.
Interpretation / May 2015 B31

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

yield seismic NTG and SNP maps from the input isochron, isopach, and average relative-impedance maps
as shown in Figure 16.
Comparative net-pay estimations
The SNP method combines two attributes (isochron
and average relative impedance) and yields an outcome
that correlates with the actual net-pay values observed
at the wellbores. However, the level of correlation is not
necessarily diagnostic of the reliability of the prediction
given that a spurious correlation may exist (Kalkomey,
1997). In other words, the predicted net-pay values using the SNP method may correlate with the actual netpay values not because there is an underlying causal
relationship but because a correlation occurs just by
chance. To estimate the spuriousness of the SNP outcomes, a statistical significance test known as the F-test
(McKillup and Darby, 2010) was performed to assess
the spuriousness of the net-pay predictions, according
to the following equation:
h i
Fh

R2
K

1R2
nK1

i;

(5)

where K represents the number of predictors, n represents the number of data points, and R is the correlation
coefficient of the linear fit between the observed and
predicted values as yielded by the combined predictors.
In this case, K is the number of seismic attributes used
for prediction of net pay and n is the number of wells
predicted. The larger the value of F, the more likely that
the results are statistically significant. As a rule of
thumb in such studies, as F falls below one, we lose
confidence in the predictive ability of the method.
To assess the value of the SNP predictions, we
benchmark against two standard methods: reflectivity
Figure 14. General workflow for SNP (modified after Condetuning and multivariable linear regression. The multinolly, 2007; Simm, 2009).
variable linear regression uses the same attributes (isochron and average relative impedance), calibrated with
all wells other than the predicted well. The multivariable linear regression is an empirical
method that ignores tuning effects and
has been occasionally used by interpreters and geomodelers. The measured
t and AI at each calibration well were
used to determine the multivariable linear regression coefficients that then
were used to predict the corresponding
net-pay values for the out-of-sample
wells. This crossvalidation involves predicting at each well location using the
other six wells to calibrate the regression coefficients. So each prediction at
each well is a result of a slightly different regression equations. For this
method, the mean absolute prediction
error at well validation locations
Figure 15. (a) Self-calibration tuning chart and (b) predicted versus actual net
pay at each validation wellbore location.
is 5.9  5.8 m.
B32 Interpretation / May 2015

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 17 shows the comparison of two statistics


metrics of this benchmark. Figure 17a depicts the correlation coefficient R2 yielded from the linear fit among
predicted and actual net-pay values using the SNP (blue
bars) and multivariable linear regression (red bars).
Figure 17b shows the F-test outcomes for the same
methods. For the multivariable regression, six wells
are used for calibration and the results are reported
for the calibration wells and an additional out-of-sample
well. The SNP method consistently yields a very high
correlation coefficient R2 of approximately 0.70.8;
the correlation coefficient for the multivariable linear
regression method is comparable, approximately 0.6.
However, the F-test outcomes for the multivariable linear regression method are significantly lower, almost
half of those of the SNP method, indicating that
the SNP method, for this case study, yields net-pay predictions that are close to the actual net-pay values. Furthermore, the accuracy of the predictions is statistically
significant.
Simm (2009) performs a modeling study comparing
the SNP method and reflectivity (RFC)-based method
(Brown et al., 1984, 1986). In our study, we compare
the two methods based on actual seismic and well data
only for a self-calibration scenario to assess the value of
performing SNP instead of reflectivity-based estimations in exploration scenarios. Figure 18 shows the
composite amplitude on the y-axis versus the isochron
on the x-axis as extracted from reflectivity or amplitude
data, in this case, the near-angle stack reflectivity. Sim-

ilar to the SNP method, self-calibration is performed by


building a tuning curve (red) from wedge modeling assumed to represent a maximum NTG of one and scaling
it to represent the envelope of the scattered points. Amplitude detuning is performed by applying an isochrondependent scalar that is a function of the ratio of the notuning baseline (blue) and the maximum NTG tuning
curve (red). The amplitude-based net-pay estimate
(RFC) is obtained by multiplying the detuned amplitude
map by the apparent thickness map. Figure 19 shows
the predicted versus actual net-pay values of the SNP
(Figure 19a), RFC (Figure 19b), and multivariable regression (Figure 19c) methods. No wells (i.e., self-calibration) were used for the SNP and RFC methods,
whereas six wells were used to calibrate the multivariable regression method. The results show that the RFCbased method, in this case, underestimates the net-pay
values as compared with the SNP outcome (the mean
absolute prediction error at well validation locations is
4.0  2.7 m for the RFC-based method), despite the fact
that both methods perform a detuning of the data by
applying isochron-dependent scalars derived from forward-modeled data. The differences in the methods are
due to (1) an inherent better detuning on relative impedance compared with the reflectivity volume and (2) the
modeled tuning curve on the reflectivity-based method
contains several peaks/valleys that affect the prediction of
values above tuning thickness (Simm, 2009). When well
calibration is available, a further linear fit between net
pay at the wells and the predicted net-pay estimates

Figure 16. (a) Apparent isochron, (b) apparent isopach, (c) average relative impedance input maps for the SNP method. (d) Selfcalibration seismic NTG, and (e) net pay output maps. Validation wells are represented by black stars.
Figure 17. (a) Square of correlation coefficient for the SNP method (blue) and multivariable linear regression (red) and (b) F-test for
the SNP method (blue) and multivariable linear regression (red).

Interpretation / May 2015 B33

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

can be used to remove bias or systematic error in the results and improve accuracy. The SNP method is apparently unbiased without well calibration: The only bias
of SNP on this self-calibration mode would consist of
the selection of the scaling factor applied to the modeled
tuning curve. In contrast, the multivariable regression has
wide scatter (mean error is 5.9  5.8 m) due to the nonlinearity of convolution with a wavelet for thin layers.

However, there is still a substantial uncertainty that


needs to be recognized in subsurface management decision making. It is important to highlight the possible
source(s) of uncertainty of the predictions, which for
this particular study were identified as (1) accuracy
in horizon picking, (2) limited-frequency bandwidth
of the seismic data and in particular deficiency in the
low-end frequencies, and (3) lateral variations in both
fluid properties and porosity.

Discussion
Documented in this paper is the application of SNP
technology (Connolly, 2007) to a geologic setting that
closely adheres to the assumptions of the method.
The more general inference from the results is that
the method may yield predicted values that are close
to the actual net-pay values, even without well calibration, which is of the utmost importance in exploration.

Horizon picking accuracy


Even though picking zero crossings on a relativeimpedance volume is significantly easier than picking
the corresponding amplitude onsets on a reflectivity
volume, it is still probable that inaccurate picks may
arise. This is especially the case when working with
a large volume of data that demand the intensive usage
of state-of-the-art automated picking tools, which was
the case in this project. In our experience with other
offshore settings, it has been found that for deepwater
sands visible on seismic data, the base reservoir reflection is commonly easier to pick and to extend surveywide than the top reservoir reflection because the target layer commonly shows a significant impedance contrast at the base. These kind of reservoir sandstones
usually have a fining-upward component or even minor
scour surfaces on top that affect the strength and
continuity of the top reflector. The horizon picking inaccuracy may affect the method in two ways. First, inaccuracy in the apparent isochron and isopach maps
calculated from the top and base horizons give rise
to inaccurate apparent thickness maps. Second, inaccuracy in the average band-limited impedance changes
the number of samples contributing to the mean that
must be calculated between the top and base of the
layer represented by the horizons picked on the zero
crossings. To the best of our knowledge, the method
does not explicitly account for this horizon picking inaccuracy, which may be difficult to estimate given the
nonobvious deterministic nature of it.

Figure 18. Composite amplitude versus apparent thickness


as measured from reflectivity data (black points), including
overlay of collocated well observations (green circles), wedge
model clean sand curve (red), and no-tuning baseline (blue).

Figure 19. (a) Predicted versus actual net-pay thickness at validation locations using the SNP method, (b) the reflectivity-based
method, and (c) multivariable linear regression. The SNP method mean error is 3.0 m, and its standard deviation is 1.5 m. The
reflectivity-based method mean error is 4.0 m, and its standard deviation is 2.7 m. The multivariable linear regression method mean
error is 5.9  5.8 m.
B34 Interpretation / May 2015

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Deficiency of low-end frequencies


By design, colored inversion shapes the average input
seismic spectra to match that of the average band-limited
impedance log (Lancaster and Whitcombe, 2000), which
yields two very attractive output characteristics: (1)
global matching of the impedance magnitudes between
inverted seismic and impedance logs in a band-limited
basis and (2) spectral shaping carried out by the colored
inversion operator yields an outcome with boosted lowend frequencies compared with the high-end frequencies
within the input seismic bandwidth, leading to a considerable attenuation of the side lobes with the subsequent
reduction of the tuning effect.
It is important to highlight the fact that our data set
bandwidth at target was approximately 2085 Hz, highly
deficient in low-end frequencies that are known to significantly drive the detectability of events from seismic
data. Kallweit and Wood (1982) perform forward modeling using two wavelets with equal high-end but different low-end frequencies, and show that the resolution
graph (apparent versus actual time thickness) for both
wavelets remained almost unchanged as a function of
the low-end frequency. In contrast, the detection (tuning) chart for the broader-bandwidth wavelet was quite
sensitive to the low-end frequencies, with the wavelet
side lobes attenuated when low-end frequencies are
added. Despite the good results obtained from SNP,
the deficiency in the low-end frequency bandwidth
on our data set may limit the ability of achieving a more
efficient detuning of the input maps, adversely affecting
the SNP outcomes (Connolly and Kemper, 2007).
Lateral changes in fluid phase and porosity
Wellbores S-3 and S-6 were consistent outliers for the
SNP predictions. Though not explicitly depicted in this
paper, these two wellbores lie very close to the oilwater contact (OWC), where lateral variations on water
saturation may be of enough magnitude to produce
more transitional rather than abrupt seismic responses
associated with the presence of the fluid contact within
the seismic resolution limits. In that case, although
these wellbores lie within the hydrocarbon imprint of
the field, they may not be totally correlated with the rest
of the wellbores in terms of fluid saturations. The determination of a good EEI lithology projection can be used
to avoid the effect of lateral changes in fluids; however,
in this case, there was not a good lithology projection.
A similar problem may arise when dealing with reservoirs with significant lateral variations in porosity because the SNP and RFC detuning methods assume
constant properties for the reservoir. This emphasizes
the importance of using a reliable EEI volume optimized for fluid illumination and able to minimize the
sensitivity to porosity variations. Also, the petrophysical
cut-offs for clay content, porosity, and water saturation
should be used to define the range of velocity variation
for reservoir sands, and these velocity ranges should be
used to establish the range of uncertainty in the tuning
curve, reservoir thickness, and scaling factor.

Conclusions
We have tested the SNP method on a data set that we
claim complies with the assumptions of the method. The
SNP method captures the relative net-pay trends as validated by wellbore data, especially when no well information is used for calibration of net-pay estimations,
which is very important in exploration scenarios where
well control is scarce or inexistent. The SNP method
yielded a mean absolute error of 3.0  1.5 m. In contrast,
the multivariable linear regression method yielded an error of 5.9  5.8 m and the reflectivity-based detuning
method yielded an error of 4.0  2.7 m. In terms of statistical significance of the predictions, the SNP method
yielded an F value of approximately 5.8, whereas the
F value for multivariable linear regression method
was less than half (approximately 2.8), indicating fewer
statistically significant results compared with the selfcalibrated SNP outcomes. When fewer wells were used
to calibrate the multiple regression results, the expected
statistical significance was poorer. The reflectivity-based
method yielded a robust statistical significance of 5.5, in
range with that of the SNP method, but it is more biased
without well calibration.
The main strength of the SNP method is the use of a
band-limited calibrated impedance volume obtained
from colored inversion that scales and partially detunes
the input amplitude data. However, the method may be
very sensitive to other variables beyond data quality
and rock-physics assumptions, such as accuracy of
the horizon interpretation. Although we did not test
the method in other geologic settings, the assumptions
of the method may restrict its deployment to a variety of
exploration settings. For instance, thick stacked sand
packages can cause seismic interference that can
clearly violate the requirement of seismic isolation of
the layer under study. The method may be also heavily
restricted on some fluvial and transitional environments where the binary impedance assumption may
not necessarily apply.
Even though the SNP and the reflectivity-based
methods are map-based detuning techniques, in the absence of well calibration SNP seems to provide more
accurate predictions than the reflectivity-based method.
However, further analysis may be needed to quantify
and compare the impact of each method on any subsurface management decision making. In any case, both
methods show the importance of considering tuning effects when considering amplitude strength for net-sand
and net-pay estimation.

Acknowledgments
The authors wish to thank BHP Billiton Petroleum
and Woodside Energy Ltd. for permission to publish
these results. We are particularly thankful to our colleagues G. Duncan, R. Hill, R. Keen, M. Gutierrez, B.
Asher, S. Misra, and S. Tadepalli for their accurate insights, support, and information during the project.
Interpretation / May 2015 B35

Downloaded 09/28/15 to 124.195.4.82. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

References
Brown, A. R., R. M. Wright, K. D. Burkart, and W. L. Abriel,
1984, Interactive seismic mapping of net producible gas
sand in the Gulf of Mexico: Geophysics, 49, 686714,
doi: 10.1190/1.1441698.
Brown, A. R., R. M. Wright, K. D. Burkart, W. L. Abriel, and
R. G. McBeath, 1986, Tuning effects, lithological effects
and depositional effects in the seismic response of gas
reservoirs: Geophysical Prospecting, 34, 623647, doi:
10.1111/j.1365-2478.1986.tb00485.x.
Castagna, J. P., and H. Swan, 1997, Principles of AVO crossplotting: The Leading Edge, 16, 337344, doi: 10.1190/1
.1437626.
Connolly, P. A., 1999, Elastic impedance: The Leading
Edge, 18, 438452, doi: 10.1190/1.1438307.
Connolly, P. A., 2007, A simple, robust algorithm for seismic net pay estimation: The Leading Edge, 26, 1278
1282, doi: 10.1190/1.2794386.
Connolly, P. A., and M. Kemper, 2007, Statistical uncertainty for seismic net pay estimations: The Leading
Edge, 26, 12841289, doi: 10.1190/1.2794387.
Duncan, G., C. Hurren, R. Hill, M. Stanley, J. Woodward,
and D. Lumley, 2013, The Stybarrow field A 4D case
study: 70th Annual International Meeting, SEG, Expanded Abstracts, 15721575.
Kalkomey, C. T., 1997, Potential risks when using seismic
attributes as predictors of reservoir properties: The
Leading Edge, 16, 247251, doi: 10.1190/1.1437610.
Kallweit, R. S., and L. C. Wood, 1982, The limits of resolution of zero-phase wavelets: Geophysics, 47, 10351046,
doi: 10.1190/1.1441367.
Lancaster, S., and D. Whitcombe, 2000, Fast-track colored
inversion: 70th Annual International Meeting, SEG, Expanded Abstracts, 15721575.
McKillup, S., and M. Darby, 2010, Geostatistics explained:
An introductory guide for earth scientists: Cambridge
University Press.
Neep, J. P., 2007, Time variant colored inversion and spectral blueing: 69th Annual International Conference and
Exhibition, EAGE, Extended Abstracts, B009.
Okaya, D. A., 1995, Spectral properties of the earths contribution to seismic resolution: Geophysics, 60, 241
251, doi: 10.1190/1.1443752.
Partyka, G., J. Gridley, and J. Lopez, 1999, Interpretational
applications of spectral decomposition in reservoir
characterization: The Leading Edge, 18, 353360, doi:
10.1190/1.1438295.
Puryear, C. I., and J. P. Castagna, 2008, Layer-thickness determination and stratigraphic interpretation using spectral inversion: Theory and application: Geophysics, 73,
no. 2, R37R48, doi: 10.1190/1.2838274.
Rutherford, S. R., and R. H. Williams, 1989, Amplitude-versus-offset variations in gas sands: Geophysics, 54, 680
688, doi: 10.1190/1.1442696.
Simm, R., 2009, Simple net pay estimation from seismic: A
modeling study: First Break, 27, 4553.

B36 Interpretation / May 2015

Whitcombe, D. N., P. A. Connolly, R. L. Reagan, and T. C.


Redshaw, 2002, Extended elastic impedance for fluid
and lithology prediction: Geophysics, 67, 6367, doi:
10.1190/1.1451337.
Whitcombe, D. N., and J. G. Fletcher, 2001, The AIGI
crossplot as an aid to AVO analysis and calibration:
71st Annual International Meeting, SEG, Expanded Abstracts, 219222.
Widess, M. B., 1973, How thin is a thin bed?: Geophysics,
38, 11761180, doi: 10.1190/1.1440403.

Ramses G. Meza received a geophysical engineering degree from


the Universidad Simon Bolivar, Venezuela, an M.S. in geophysics from
the Colorado School of Mines, USA,
and he is pursuing a Ph.D. in geophysics from the University of Houston. He
is a geophysicist with BHP Billiton
Petroleum. Early in his career, he
worked as a reservoir geophysicist at PDVSA in Puerto
La Cruz, Venezuela, Harvest-Vinccler in Maturin, Venezuela, and ConocoPhillips in Houston, USA. His responsibilities included support in terms of quantitative seismic
interpretation for hydrocarbons exploration and production activities. Since 2012, he has been with the BHP Billitons quantitative interpretation (QI) team providing
support to all E&P assets with emphasis on integration
of QI products, visualization, seismic attributes and visualization, QI quality assurance, seismic reservoir characterization, DHI analysis, and risking. He is a member of SEG,
AAPG, EAGE, and SOVG.

Juan Mauricio Florez received a


B.S. in geology from the University
of Nacional of Columbia and a Ph.D.
in geophysics from Stanford University. He is a geophysicist at BHP Billiton Petroleum, currently manager of
the QI team. He worked as an exploration geologist in Colombia for about
eight years, and after finishing his
graduate studies in 2005, he has worked in reservoir characterization (iReservoir), rock physics, and AVO modeling
(BP America) and more recently quantitative seismic interpretation with BHP Billiton.

Stanislav Kuzmin received a Ph.D.


(2004) in physics from the University
of California, and he worked on a variety of exploration and development
projects in different basins. He is a
geophysicist at BHP Billiton Petroleum, and the main emphasis of his
work is QI and rock physics.

Você também pode gostar