See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/269066088

3D Seismic waveform classification

Conference Paper · September 2012

DOI: 10.1190/IST092012-001.92

CITATIONS READS

0 136

2 authors, including:

Ivan I Priezzhev
IPLAB LLC
44 PUBLICATIONS 50 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Well log or production rate prediction using set of seismic cubes based on deep learning algorithms View project

All content following this page was uploaded by Ivan I Priezzhev on 27 May 2016.

The user has requested enhancement of the downloaded file.

3D Seismic waveform classification
Ivan Priezzhev (Schlumberger Information Solution), Surender Manral (Schlumberger Information Solution)

Summary  Automatic definition by applying a “K-means”

algorithm (MacQueen, 1967) or SOM Kohonen
Pattern recognition have been used for years in (Kohonen, 1995), which allocates “clusters” with a
interpretation and reservoir characterization workflows, compact distribution of similar waveforms.
but with the advent of the technology it has transformed  Semi-controlled classification with definition of
from purely visual inspection based to, automated point sets which correspond to the exact location of
unsupervised and supervised classification which can 3D waveform. Usually these are well locations
help in identification of subtle patterns hence allowing where reservoir parameters are known. This is
Geoscientist in identification of subsurface characteristics allow to find similar places in the layer via seismic
and exploiting the information in the seismic data. 3D waveform similarity.
Seismic waveform shape and character define facies with  Supervised approach with predefined waveforms
significant detail compared to time and amplitude taken from an external source, for instance from a
mapping. A new 3D waveform seismic facies synthetic modeling package where the reservoir
classification scheme has been presented in this paper. parameters are known (Balz et al., 1999; Veeken
Unlike the traditional seismic facies classification 2007). The influence of many reservoir
methodology, classification using the 3D sub-cube is characteristics is modeled and various parameters
done using a K-means (MacQueen, 1967) or SOM can be changed at the same time to build a large
Kohonen (1995) algorithm. The clustering is done in an suite of scenarios. The class specific master trace is
n-dimensional attribute space. A cross distance matrix for described in a statistical way as several
all classes is computed to assess the amount of petrophysical models may give the same modeled
confidence for the classification results. The main seismic response. This supervised method is of
advantage of using 3D sub-cube like objects for great help for the interpretation of the classification
classification is ability to do classification for very thin output.
layers with vertical size less then vertical seismic
resolution. The use of computation sub-cubes stabilizes The second stage is the classification step. The similarity
the calculation results. The main output is a distribution (or probability) is calculated as a measure for the degree
map of the classes and maps of probabilities (or of resemblance of the master set of traces (sub-cube) to
similarity) for every class type. A class typical waveform every classified sub-cube. The highest probability range
and sub-cube for every class is automatically generated. determines the class category for the sub-cube under
The supervised mode allows probabilistic output that examination (Figure 1).
facilitates the interpretation of the end result.
During the classification, that uses the “K-means”
Introduction algorithm or SOM Kohonen, two objects (waveform
segment or set of waveform segments – sub-cube) will be
A fast 3D waveform classification scheme has been in the same class if their waveforms have some kind of
developed based on the SOM or K-Mean. The seismic similarity. A measure for the amount of “similarity” can
amplitude or any attribute be can be taken as input for the be calculated from the distance between the waveforms in
classification procedure. Seismic waveforms or segments a so-called n-dimensional cross plot. Mathematically
of traces are classified in order to define seismic facies every object can be represented by a point in a multi-
changes in a specific interval which can be two geologic dimensional space. In our seismic classification scheme
formations or constant interval. every dimension corresponds to the sample position on
Along with the facies distribution map of the classes, the waveform. The distance can now be taken for
probabilities maps (or similarity) for every class type is example as the square difference between two waveforms
also generated. Furthermore, a class typical waveform (L2 method). The mathematical distance in such a cross
and typical sub-cube for every class is generated. plot can be described as follows:

Method Let object Oi be a sub-cube where every trace have a set

Objects for the classification scheme are 3D waveforms  

of measurements X i  xi1 , xi2 ,..., xin ,
or seismic sub-cubes. The latter allows a certain degree of In our case every measurement is the sample value in the
averaging before the classification is done. Waveforms seismic trace. And object O j is the second sub-cube
in sub-cubes (3D) have the following advantages:
which have the same size and
 It allows to get classes with locally varying
measurements X j  x1j , x 2j ,..., x nj , it represents also
seismic signals – faults, turbidities etc.
 Geological objects with dipping seismic signal, sampled values in the seismic trace (Figure 1).
like clinoforms, are better resolved.
 Input data smoothing and more stable results.
The procedure consists of a learning phase with class
typical waveform computation and a classification step.
The first step can be done in several ways:

Istanbul International Geophysical Conference and Oil & Gas Exhibition, Istanbul,Turkey, 17-19 September 2012.
 3D Seismic waveform classification

All distance measurements are normalized to range

a) between 0 (non-similar) and 1 (similar). This normalized
distance value is closely correlated to “probability” or
degree of similarity. The similarity distance between
objects in the n-dimensional space plot is determined by
different calculation methods: Correlation + L2 (using
coefficient of correlation (3) and L2 (2) method (Neff et
al. 2001); L1 (1) (abs difference), L2 (2)(square
difference), Average difference (average for every trace
in sub-cube L2 between objects), Spectrum difference (4)
(amplitude spectrum L2 difference).
The layer or investigation zone can be defined in two
ways:
- An interval between top and bottom surfaces (see
Figure 2).
- A constant thickness interval below the top
surface.
Results of the classifications can be presented in surface
(Figure 3 –a) or in cross sections (Figure 3 – b).

b)
Top
surface

Bottom
surface

Figure 1: a) Two sub-cubes with two objects for calculating a

measure of similarity-distance in the multi- dimensional domain
Figure 2: Top and bottom surfaces will define variable interval
b) Typical waveform at central trace in sub-cube1 for one of the
for waveform classification.
classes computed by the waveform classification algorithm.

Self-Organizing Map (SOM) by Kohonen (1995) use

scaling multi dimensional data (sub cube traces data) to
two dimensional (in our case – one dimensional) map. A’
Every neuron can be use like typical object for
classification.

In this case distance (measure of similarity) between these

two objects can be calculated in different ways:
subcube
1 n k
M L1 (Oi , O j )  L1 ( X i , X j )  
trace
(  xi  x kj )
n k 1
(1)

a) A
subcube
1 n (2)
M L 2 (Oi , O j )  L2 ( X i , X j )  
trace
(
n
 (x
k 1
k
i x ) )
k
j
2

A A’
M COR (Oi , O j )  COR( X i , X j )
(3)

M FUR (Oi , O j )  L2 [ F ( X i ), F ( X j )]
(4)

M L 2COR (Oi , O j )  ( L2 ( X i , X j )  LCOR ( X i , X j )) / 2

(5)

where L1 and L2 are measure of similarity, MCOR is the b)

“correlation coefficient”. MFUR is a measure of similarity Figure 3: Classification results and possible visualization of it on
amplitude spectrum difference and F(Xi) is Fourier surface - a) and on cross-section - b).
transfer of function Xi and where (5) is a mixed measure
of similarity that is normally used by default.

Istanbul International Geophysical Conference and Oil & Gas Exhibition, Istanbul,Turkey, 17-19 September 2012.
 3D Seismic waveform classification

Examples Figure 7 shown classification results for thin continental

sedimentary layer. We use 6 milliseconds layer for cube
Seismic amplitude data is used as input with two horizons with 2 milliseconds sample rate. For this case we used
as the top and the base for unsupervised classification, two approaches for classification. First approach is
number of class were input being the zone of interest and conventional classification based on single trace
the other geologic understanding in the area, clearly waveform. It is clear to see that in this case the results
showing the channel in red facies (Figure 4 & 5). looks very similar to seismic amplitude slice. Also some
small details what can be seen on seismic slice (Figure 7,
a) is missing on the classification map (Figure 7, b).
Second approach is based on 3D waveform classification.
In this case the results (Figure 7, c) are much better and
allow to detect Paleo Rivers or other objects with
significant sand thickness like shown on Figure 8.

a)

Figure 4: Being sensitive to waveform seismic facies map

reveals deeper insights into the seismic response. (a) 3D map
view of output classes and (b) seismic cross-section with classes
in interval strips along section view A-A’ showing the variation
in the waveform

b)

Figure 5: a) Probability (or similarity) of all traces with the

master trace of the red class and (b) probability interval cross-
section along seismic line A-A’.

Figure 6 show classification result for thin layer (~ 6

millisecond) via sub-cubes (6ms x 7 x 7 traces). In this
case vertical wavelet is too small to do classification and c)
lateral signal used to define seismic facies.

sub-cubes
size 6msx7x7

Figure 6: Thin layer (6 milliseconds) classification example for

seismic facies detection for carbonate layer. Picture is shown 3d
view of the horizon with classes and sub-cubes locations for the Figure 7: Thin layer (6 milliseconds) classification example. a) -
class typical 3D waveform (sub-cube 6ms x 7x 7 traces) for Seismic amplitude slice via center of the layer. b) - classification
every class. result based on single trace waveform. c) - classification result
based on 3D waveform (sub-cube 6ms x 3 x 3 traces).

Istanbul International Geophysical Conference and Oil & Gas Exhibition, Istanbul,Turkey, 17-19 September 2012.
 3D Seismic waveform classification

Usefulness of the classification scheme always needs to

be verified on a case-by-case basis. Subtle differences in
a) the input data can be rapidly highlighted by this method
that would remain undetected by conventional data
analysis. Also other attributes than seismic data cubes can
be used as input.
The method is flexible, very fast, requires very limited
user intervention and a robust workflow is proposed. The
supervised option facilitates the interpretation of the
classification results considerably and makes a
probabilistic approach to the facies interpretation
possible.

Acknowledgement

The authors would like to thank Schlumberger for the

opportunity to spend the necessary time to develop this
technique, and also for allowing us to publish the results
of the work.

References

b) Balz, O., F. Pivot, and P. Veeken, 1999,

Reservoir characterization using neural networks
controlled by petrophysical and seismic modelling.
Extended Abstracts, 61th EAGE annual meeting, S015,
p. 1-4.

Correlation coeff. 0.877407 MacQueen, J. B., 1967, Some Methods for classification
and analysis of multivariate observations. Proceedings of
5th Berkeley Symposium on Mathematical Statistics and
Probability, University of California Press, p. 281–297.

Neff, D. B., S. A. Runnestrand, and E. L. Butler, 2001,

Multi-attribute seismic waveform classification United
States Phillips Petroleum Company USA Patent
6223126

Veeken, P.C.H., 2007, Seismic stratigraphy, basin

Figure 8: a) - fragment of classification result shown on Figure 6 analysis and reservoir characterisation. Handbook of
with well position and sand thickness in meters. b) – correlation Geophysical Exploration, Volume 37, eds prof K. Helbig
between sand thickness and classes. and S. Treitel, Elsevier Scientific Publisher, Amsterdam,
509 p.

Kohonen, T., 1995, Self-organizing maps: Springer-

Verlag NewYork, Inc.
Conclusion

The 3D waveform classification scheme, based on a K-

means or SOM Kohonen methods, is a powerful tool to
classify seismic data relatively fast and to delineate
seismic facies units in a true 3D sense.
The main advantage of using 3D sub-cube like objects
for classification is ability to do classification for very
thin layers with vertical size less then vertical seismic
resolution. In this case the technology can be used for
lateral facies detection like faults, fracture zones or Paleo
Rivers. The capabilities of a seismic reservoir
characterization workstation are exploited to visualize the
results in a convenient way and hence good connectivity
with standard computation functionalities is ensured. An
elementary processing rule applies: the cleaner the input
data, the more reliable the output will be. The use of
computation sub-cubes stabilizes the calculation results.

Istanbul International Geophysical Conference and Oil & Gas Exhibition, Istanbul,Turkey, 17-19 September 2012.

View publication stats