GEOPHYSICS

Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Seismic attribute selection for unsupervised seismic facies

analysis using user guided data-adaptive weights

Journal: Geophysics

Manuscript ID GEO-2017-0192.R2

Manuscript Type: Technical Paper

Fo
Date Submitted by the Author: 12-Sep-2017

Complete List of Authors: Zhao, Tao; University of Oklahoma, ConocoPhillips School of Geology and
rP
Geophysics
Li, Fangyu; University of Oklahoma, ConocoPhillips School of Geology and
Geophysics
Marfurt, Kurt; University of Oklahoma, ConocoPhillips School of Geology
ee

and Geophysics

Keywords: artificial intelligence, interpretation, neural networks, seismic attributes

rR

Area of Expertise: Interpretation Methods, Seismic Attributes and Pattern Recognition

ev
iew

This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 1 of 61 GEOPHYSICS

1
2
3
4
5
6 Seismic attribute selection for unsupervised seismic facies analysis using user guided data-
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 adaptive weights
9
10 Tao Zhao1, Fangyu Li1, and Kurt J. Marfurt1
11
12
13
14 1
15 The University of Oklahoma, ConocoPhillips School of Geology and Geophysics
16
17 tao-zhao@ou.edu, fangyu.li@ ou.edu, and kmarfurt@ou.edu
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31
32
ev

33
34
35
36
Corresponding author:
iew

37
38 Tao Zhao
39
40 The University of Oklahoma, ConocoPhillips School of Geology and Geophysics
41
42
710 Sarkeys Energy Center
43
44
45 100 East Boyd Street
46
47 Norman, OK 73019
48
49 tao-zhao@ou.edu
50
51
52
53
54
55
56
57
58 1
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 2 of 61

1
2
3 ABSTRACT
4
5
6
7 With the rapid development in seismic attribute and interpretation techniques, interpreters
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 can be overwhelmed by the number of attributes at their disposal. Pattern recognition driven
10
11 seismic facies analysis provides a means to identify subtle variations across multiple attributes
12
13
14
that may only be partially defined on a single attribute. Typically, interpreters intuitively choose
15
16 input attributes for multiattribute facies analysis based on their experience and the geologic
17
18 target of interest. However, such an approach may overlook unsuspected or subtle features
19
Fo
20
hidden in the data. We therefore augment this qualitative attribute selection process with
21
22
23 quantitative measures of candidate attributes that best differentiate features of interest. Instead of
rP

24
25 selecting a group of attributes and assuming all the selected attributes contribute equally to the
26
ee

27 facies map, we weight the interpreter-selected input attributes based on both their response from
28
29
rR

30 the unsupervised learning algorithm and interpreter’s knowledge. In other words, we expect the
31
32 weights to represent both “which attribute is favored by an interpreter as input for unsupervised
ev

33
34 learning” from an interpretation perspective, and “which attribute is ‘favored’ by the learning
35
36
iew

37
algorithm” from a data-driven perspective. Therefore, we claim the weights are user guided and
38
39 data-adaptive, as the derivation of weight for each input attribute is embedded into the learning
40
41 algorithm, providing a specific measurement tailored to the selected learning algorithm, while
42
43
still taking interpreter’s knowledge into account. We illustrate our workflow using Barnett Shale
44
45
46 surveys and an unsupervised self-organizing map (SOM) seismic facies analysis algorithm. We
47
48 find that the proposed weighting-based attribute selection method better differentiates features of
49
50 interest than using equally weighted input attributes. Furthermore, the weight values provide
51
52
53 insights into dependency among input attributes.
54
55
56
57
58 2
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 3 of 61 GEOPHYSICS

1
2
3 Keywords: seismic attributes, self-organizing maps, seismic facies analysis, seismic
4
interpretation
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31
32
ev

33
34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58 3
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 4 of 61

1
2
3 INTRODUCTION
4
5
6
7 Seismic attributes are routinely used in the exploration, appraisal, and production stages
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 of an oilfield life cycle with great success. However, there are many redundant attributes, and
10
11 many attributes with vague geological meaning rendering them useless (Barnes, 2007).
12
13
14
Excluding such redundant and useless attributes from the seismic interpretation workflow is
15
16 important, but interpreters still face the challenge of selecting which of the remaining attributes
17
18 are appropriate for a given interpretation task. Interpreters have spent a considerable amount of
19
Fo
20
effort on how to select the most suitable attributes for both qualitative and quantitative facies and
21
22
23 reservoir property estimation. Chen and Sidney (1997) provide a comprehensive review on
rP

24
25 attribute selection for reservoir prediction and monitoring, dividing attributes into wave
26
ee

27 kinematics and reservoir feature categories, and further summarize the applicability for each
28
29
rR

30 attribute. Kalkomey (1997) discusses the risk of false correlation between seismic attributes and
31
32 reservoir properties, in which she suggests that special caution is needed when there are too few
ev

33
34 wells to correlate with too many attributes. She further suggests to use only those attributes that
35
36
iew

37
have a physically justifiable relationship with reservoir property as predictors. Hart and Balch
38
39 (2000) present a case study on predicting reservoir properties from seismic attributes with
40
41 limited well control, in which they propose a suite of visual correlation schemes to define the
42
43
attributes of choice. From a more quantitative aspect, Schuelke and Quirein (1998) propose to
44
45
46 use cross-validation as a measure of prediction performance, then select attributes that lead to
47
48 higher cross-validation. Since then, almost all the proposed alternative strategies have shared one
49
50 fundamental concept, which is to select attributes that lead to the lowest validation error. While
51
52
53 sequential correlation (e.g. Hampson et al., 2001; Leiphart and Hart, 2001) is the most
54
55 commonly used correlation method to determine the input attributes, Dorrington and Link (2004)
56
57
58 4
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 5 of 61 GEOPHYSICS

1
2
3 use a genetic nonlinear inversion algorithm to automatically determine which combination of
4
5
6 input attributes minimizes the error in neural network porosity prediction. More recently, for
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 porosity prediction purpose, researchers explored conducting input attribute selection by

9
10 incorporating model complexity into the analysis using group method of data handling (Ahmed
11
12
13
at al., 2010; AlBinHassan and Wang, 2011), while Wang et al. (2015) first use a rough set
14
15 method to find the attribute that best correlate with porosity, then use principal component
16
17 analysis to further reduce the number of input attributes to a subset of principal components.
18
19
Fo
20
Such prediction error based attribute selection methods require a relatively large number
21
22
23 of training samples, either from well control or from interpreter handpicked polygons defining
rP

24
25 seismic facies, providing supervision to the process. This is a supervised learning process, in
26
ee

27 which interpreters have training samples of “ground truth”, with the objective of recovering the
28
29
rR

30 link from input variables (e.g. seismic attributes) to a target property (e.g. lithofacies). In
31
32 contrast, interpreters may also wish to discover the natural facies distribution in the data in an
ev

33
34 unsupervised manner, without introducing any bias. In this case, interpreters then use
35
36
iew

37
unsupervised learning, in which there are no “ground truth” of target properties.
38
39
40 Barnes and Laughlin (2002) find that the selection of input attributes has a higher impact
41
42 on the facies map than the unsupervised learning algorithm used for classification. Zhao et al.
43
44
(2015) review several competing unsupervised learning algorithms and through iterative
45
46
47 experiments find a subset of attributes to highlight different architectural elements in a turbidite
48
49 system. Roden et al. (2015) use principal component analysis (PCA) to quantitatively measure
50
51 the contribution of each seismic attribute to the top principal components which in turn are used
52
53
54 in the subsequent facies analysis. Although PCA estimates the contribution of each attribute to
55
56 represent the data variability as a whole, it does not provide a means of determining which
57
58 5
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 6 of 61

1
2
3 attributes best differentiate a given facies from other facies and the background behavior. Amin
4
5
6 et al. (2017) propose a ranking framework to select the best attributes to augment supervised salt
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 detection based on information theory. Such information based method is independent from the
9
10 supervised learning process, therefore can be transplanted to unsupervised learning. However, it
11
12
13
still suffers the similar fact as PCA that it does not provide a direct measurement of attribute
14
15 importance to the unsupervised learning algorithm.
16
17
18 One of the most commonly used seismic facies analysis algorithms is the Kohonen
19
Fo
20
(1982) self-organizing map (SOM). SOM is an unsupervised learning technique that projects
21
22
23 higher dimensional multiattribute data onto a lower dimensional (usually 2D) space called a
rP

24
25 manifold in which clusters of seismic facies are more easily defined and interpreted. In perhaps
26
ee

27 the first application of SOM to seismic data, Poupon et al. (1999) use amplitude from a suite of
28
29
rR

30 30 phantom horizons as input projected to a 1D manifold, thereby generating a facies map along
31
32 a target horizon. Strecker and Uden (2002) and Coléou et al. (2003) extend these concepts to
ev

33
34 volumetric SOM analysis, in which interpreters take multiple attributes at each voxel as inputs.
35
36
iew

37
Since then, researchers have been investigating how to effectively recover and represent the
38
39 information buried in multiple seismic attributes. Matos et al. (2007) use a k-means clustering on
40
41 the SOM result to reduce the number of facies. To assist in presenting SOM facies, Matos et al.
42
43
(2009) display an SOM facies map using a 2D color map. Matos et al. (2011) adapt the 2D color
44
45
46 map and use gray-level co-occurrence (GLCM) texture attributes (Gao, 2003; Di and Gao, 2017)
47
48 to map Pennsylvanian channels in northern Oklahoma. Marroquín et al. (2009) introduce a visual
49
50 data-mining system that displays SOM facies map with U-matrix (distance among SOM nodes)
51
52
53 and parallel coordinates plot, both of which serve as quality measurements for the SOM facies
54
55 map. Hu et al. (2014) and Zhao et al. (2016) introduce improvements to preserve the Euclidean
56
57
58 6
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 7 of 61 GEOPHYSICS

1
2
3 distance from seismic attribute space to 2D SOM latent space. Zhao et al. (2017) constrain SOM
4
5
6 analysis using a sedimentary cycle model to introduce spatial/temporal awareness.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 At present, all multiattribute SOM analyses share one implicit assumption: input
10
11 attributes are all independent and have the same contribution/importance to the SOM facies. This
12
13
14
assumption has greatly simplified the SOM analysis, as interpreters just need to select the
15
16 attributes to feed into a SOM, without specifying which of the selected attributes are more
17
18 important than the others. Treating all input attributes equally is the somewhat “normal” practice
19
Fo
20
in pattern recognition; unfortunately, seismic attributes are not created equally, and the
21
22
23 importance of an attribute varies greatly with the exploration perspective. Geometric attributes
rP

24
25 (e.g. coherence) are routinely used to map discontinuities (e.g. faults) and stratigraphic edges in
26
ee

27 seismic data. In the Barnett Shale area, interpreters find coherence and curvature extremely
28
29
rR

30 useful when mapping the karst collapse features in the Ellenburger dolomitic formation, which
31
32 pose drilling hazards to the overlaying Barnett Shale formation and therefore are crucial during
ev

33
34 well path design (Sullivan et al., 2006). In contrast, Qi et al. (2014) find that spectral
35
36
iew

37
decomposition attributes provide estimates of the lateral change in layer thickness and lithology,
38
39 which can also map karst features in the same region. If an interpreter uses both structural and
40
41 spectral attributes in SOM, with the goal of mapping the facies distribution in the highly karsted
42
43
Ellenburger formation, he/she would expect these attributes to have varying degrees of impact.
44
45
46
47 In the structurally less deformed Barnett Shale reservoirs, interpreters are more interested
48
49 in mapping different lithofacies, which may contain high total organic carbon (TOC), or may be
50
51 more susceptible to hydraulic fracturing. In this case, seismic inversion attributes (e.g.
ρ and
52

μρ) that are closely related to geomechanical behavior, are more favored by interpreters (Altamar
53
54
55
56 and Marfurt, 2015). If the interpreter uses a combination of structural attributes and inversion
57
58 7
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 8 of 61

1
2
3 attributes as inputs for SOM, weaker differences measured by inversion attributes may be
4
5
6 suppressed by the stronger differences in structural attributes. One possible solution for
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 investigating the ambiguity of the SOM facies when using suboptimal input attributes is to use a
9
10 fuzzy (soft) clustering scheme, which means instead of assigning hard class labels to each data
11
12
13
sample vector, one calculates the probabilities of that sample belonging to each of the classes.
14
15 Generative topographic mapping (GTM) is a statistical variant of SOM which has recently been
16
17 used on seismic facies analysis (Roy et al., 2014). However in this study, we stick with SOM as
18
19
it is the well-established algorithm for seismic facies analysis and focus on manipulating the
Fo
20
21
22 input attributes.
23
rP

24
25 The attribute selection system we use today is in fact simply a weighting system: if we
26
ee

27 use an attribute, its weight is one; if we reject it, its weight is zero. We therefore ask the question:
28
29
rR

30 instead of either selecting or rejecting an attribute, can we define weights that represent the
31
32 significance of each input attribute in differentiating facies of interest? SOM clusters data by
ev

33
34 using either a Euclidean distance,
35

 = ( − ) ( − ),
36
iew

37 (1)
38
39 or a Mahalanobis (1936) distance,
40

 = ( − )   ( − ),
41
42 (2)
43
44 where is an  ⨯ 1 multiattribute data vector of N attributes, I is a diagonal matrix of ones,
45
46
is another multiattribute data vector, C is the covariance matrix of multiattribute variables, and r
47
48
49 is the Euclidean or Mahalanobis distance between and . If we assume the input attributes to
50
51 be independent, the Mahalanobis distance then becomes z-score, where
52
53
 =  , (3)
54
55
56 and
57
58 8
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 9 of 61 GEOPHYSICS

1
2
3  = 0, ∀  ≠ . (4)
4

 is the standard deviation of the ith attribute. To emphasize and deemphasize the importance of
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 a given attribute, we define a weight matrix W when calculating distance in SOM:

9

 = ( − ) "  ( − ).
10
11 (5)
12
13 Now we have a diagonal matrix W along with the diagonal normalization matrix C used for z-
14
15 score. In case C is symmetric and positive definite (while not diagonal), its inverse C-1 can be
16
17 # # # #
18 written as  $  $ , and the term WC-1 in equation 5 should then be replaced by  $ " $ . A
19
Fo
20
larger value of Wii means the ith attribute is more important and is weighted more in SOM, and a
21
22
23 smaller value of Wii means the ith attribute is less important and therefore has less impact on the
rP

24
25 SOM facies map. If the diagonal contains only zeros or ones, equation 5 reduces to the
26
ee

27 traditional (binary attribute weighting) SOM where one selects (Wii = 1) or rejects (Wii = 0) an
28
29
rR

30 attribute. In Figure 1, we show an example of randomly selected 50 sample points of two

31
32 attributes from the field data used in this study to demonstrate how the weight matrix W changes
ev

33
34 clusters in the data. We use k-means clustering for demonstration purpose. In this example,
35
36
iew

37
giving equally weighted coherence and peak frequency attributes, k-means finds four clusters
38
39 represented by the four colors in Figure 1a. These clusters are mathematically valid but
40
41 somewhat less valuable geologically, because clusters mainly follow coherence variations, but
42
43
coherence is nothing but an on/off switch of discontinuities. Change in peak frequency may
44
45
46 present change in lithology and layer thickness, but the yellow cluster contains strong variation
47
48 in peak frequency, which almost certainly is a mix of different geologic facies. If we weight
49
50 coherence less than the peak frequency, we see a clustering that makes more geological sense
51
52
53 (Figure 1b), in which one cluster represents very low coherence (discontinuity), and three
54
55 clusters of varying peak frequency.
56
57
58 9
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 10 of 61

1
2
3 We organize the rest of this paper as follows. We first present the data used in this study,
4
5
6 which are from the Barnett Shale in northwest Texas, United States. We then introduce our
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 definition and derivation of the weight matrix to be used for attribute selection, and demonstrate
9
10 the proposed workflow by comparing with the traditional, input attributes equally weighted SOM.
11
12
13
We further provide a second, yet concise field example, complementing the first experiment with
14
15 another input attribute scenario. Finally, we sum up the study with conclusions.
16
17
18 DATA DESCRIPTION
19
Fo
20
21
In our field example, the Barnett Shale lies directly on top of the dolomitic Ellenburger
22
23
rP

24 formation in the western region of the Fort Worth Basin (Figure 2). The Mississippian (early
25
26 Carboniferous) Barnett Shale unconformably overlies the Ordovician Ellenburger Dolomite
ee

27
28 which leaves a hiatus in the Silurian and Devonian with plenty of time for karsting of the
29
rR

30
31 dolomites. Going through a series of diagenetic episodes (Kupecz and Land, 1991) which
32
ev

33 include at least five karst events ranging between post Ellenburger to Early Pennsylvanian
34
35 (Canter et al., 1993), the Ellenburger formation is highly deformed, with extensive development
36
iew

37
38
of karst and joints that extend upwards from the water-saturated Ellenburger into the Barnett
39
40 Shale (Pollastro et al., 2007). Hydraulic fracturing may open these zones of weakness, so that
41
42 wells intersecting karsts and diagenetically altered joints may produce a significant amount of
43
44
water and therefore have to be abandoned (Qi et al., 2014).
45
46
47
48 Karst-related fractures are common in the upper Ellenburger (Kerans, 1988). In this
49
50 study, we focus on the upper Ellenburger. Figure 3a provides seismic amplitude section at line
51
52 XX’ that goes through several highly karsted regions. Proposed karst features are marked with
53
54
55 white arrows. Figure 3b is the seismic amplitude from a time slice at 0.75 s, which is
56
57
58 10
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 11 of 61 GEOPHYSICS

1
2
3 approximately at the level of Ellenburger. Figure 3c gives the seismic amplitude from a phantom
4
5
6 horizon (Horizon A) 25 ms below the top of Ellenburger. We highlight the assumed highly
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 karsted regions with black dashed lines in Figure 3b and 3c. Discussed in Qi et al. (2014), the
9
10 seismic data are prestack time-migrated with two poststack data conditioning steps applied,
11
12
13
aiming to improving the resolution in the thin Barnett Shale and the Ellenburger Dolomite. Step
14
15 one is structure-oriented filtering (SOF) using a Karhunen–Loève (or principal component) filter
16
17 and Kuwahara searching window (Kuwahara et al., 1976), which rejects random noises and
18
19
preserves discontinuities along structural dip. To minimize the risk of suppressing geology, a
Fo
20
21
22 single time-variant spectral balancing operator is then applied to the entire volume. We use
23
rP

24 spectral decomposition, geometric, and texture attributes, which are sensitive to strata thickness,
25
26 lithology, and structural deformation, to illuminate the architectural elements presented in the
ee

27
28
29 shallow part of Ellenburger formation. Table 1 provides a list of attributes that are used in this
rR

30
31 study.
32
ev

33
34 To study the geomorphology of the highly karsted Ellenburger formation, we use co-
35
36
iew

37
rendered multiple attributes along Horizon A, which is 25 ms below the top of Ellenburger
38
39 formation (Figure 4 to 7). We provide a summary of the meaning of arrow notations used in the
40
41 subsequent figures in Figure 4. Karst features in the study area appear either as isolated circular
42
43
to oval shape, or in a cellular network of polygonal karst. The positive (red) on the perimeter and
44
45
46 negative (blue) in the center in structural curvature (Figure 4) define collapse features.
47
48 Amplitude curvature provides a complementary image, enhancing smaller collapse features (e.g.,
49
50 blue arrows in Figure 5). Polygonal karst and fractures are delineated in both structural and
51
52
53 amplitude curvatures (outlined by black dashed lines), allowing the identification of the major
54
55 faults (red arrows), which compartmentalize the karst into different clusters. Modulating peak
56
57
58 11
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 12 of 61

1
2
3 spectral frequency by peak spectral magnitude (Figure 6), we observe that highly karsted regions
4
5
6 exhibit lower frequency compared to the surrounding area, possibly due to the non-specular
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 scattering from the chaotic reflectors. These regions are also low in peak spectral magnitude
9
10 (dimmer color in Figure 6), as a substantial amount of the reflected energy is not properly
11
12
13
received by the receivers within the migration aperture. Besides the highly karsted regions, we
14
15 also notice a shift in frequency in the less deformed surrounding area, indicating a change in
16
17 layer thickness. The yellow arrows indicate a thick layer which should contain lower
18
19
frequencies. In contrast, the orange arrows indicate a thin layer which contains higher
Fo
20
21
22 frequencies. Figure 7 co-renders GLCM homogeneity with energy ratio similarity. Visually,
23
rP

24 areas that are less coherent are also less homogeneous, suggesting a more rugose surface. The
25
26 low GLCM homogeneity outlines the regions of karst (in red), inside which the similarity
ee

27
28
29 attribute provides details of the karst geometry at different scales (in black). These visual
rR

30
31 correlations of mathematically independent attributes for a given geologic feature provide an
32
ev

33 interactive, interpreter-driven means of selecting the most appropriate attributes for subsequent
34
35
36
machine-driven classification. Naturally, we would then ask: in order to represent most of the
iew

37
38 attribute information in a seismic facies map, shall we just dump all these attributes into a pattern
39
40 recognition algorithm, or shall we choose to honor some of them more than the others? In the
41
42
following sections, we demonstrate how weighting interpreter-selected input attributes
43
44
45 differently recovers more information from the attributes.
46
47
48 WEIGHTING AS A MEANS FOR ATTRIBUTE SELECTION
49
50
51 After arriving at a reduced group of attributes, our goal is then to define the weight
52
53
54 matrix W for this group of attributes as a function of interpreter’s knowledge and attributes’
55
56 contribution to SOM. Before going into the details of deriving the weight matrix W, we offer a
57
58 12
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 13 of 61 GEOPHYSICS

1
2
3 brief review of the SOM implementation used in this study, which is built on Roy (2013). A
4
5
6 more detailed mathematical description on this implementation is provided in Appendix A.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 Starting with multiple attributes in the physical 3D space defined by x, y, and t (or z for
10
11 depth), we first “transform” (or re-sort, because we are not changing any values) the data to an
12
13
14
N-dimensional space, and each of the N axes is defined by an attribute. The most common

instance for this N-dimensional space is a 2D crossplot. For example, a
ρ- μρ crossplot that
15
16
17
18 interpreters use to determine geomechanical rock properties is indeed a 2D attribute space, and
19
each point in this space is a 2D vector of (
ρ, μρ). We then use a z-score normalization to bring
Fo
20
21
22
23 the attributes to the same scale. We form an N × N covariance matrix for the N attributes and
rP

24
25 calculate the eigenvalues and eigenvectors. If we project all multiattribute data vectors onto the
26
ee

27
kth eigenvector uk, we will have the kth principal component of the multiattribute data. The 2D
28
29
rR

30 SOM topological map is represented by SOM nodes (with two degrees of freedom), which are
31
32 initialized by uniformly sampling the 2D space defined by the first two principal components.
ev

33
34 We then project the SOM nodes back to the N-dimensional attribute space to initialize the
35
36
iew

37 prototype vectors (note there is information lost during this projection if N > 2, as we are
38
39 reconstructing N-dimensional data from the first two principal components). The objective of
40
41 SOM is approximating the multiattribute data using a limited number of prototype vectors (in
42
43
44
this case, 256 prototype vectors), and the prototype vectors are updated iteratively using the
45
46 nearest data vectors. After each iteration, we update the SOM nodes by projecting the prototype
47
48 vectors onto the first two eigenvectors. Finally, after a predefined number of iterations, we color
49
50
code the SOM nodes using a 2D color map based on their positions in the 2D space, find the
51
52
53 corresponding prototype vectors in the multiattribute space, and assign the same color to the
54
55
56
57
58 13
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 14 of 61

1
2
3 multiattribute data vectors that are closest to a prototype vector. In this way, seismic samples that
4
5
6 have similar attribute response have similar colors, forming a seismic facies map.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 As previously introduced, our goal is to define a weight matrix W to quantify the value of
10
11 each input attribute for SOM facies analysis. We as interpreters propose that the significance of
12
13
14
an attribute should consist of two parts, which are (1) attributes’ contribution to a SOM model,
15
16 and (2) what attributes are suitable as inputs for SOM. In other words, the weight of an attribute
17
18 is a function of learning algorithm response and interpreter’s preference. We introduce the
19
Fo
20
definition of attribute weight in the rest of this section.
21
22
23
rP

24 Inspired by Benabdeslem and Lebbah (2007), given N input attributes and J prototype
25
26 vectors (which are the proxies of the 2D SOM neurons in the attribute space), we define & , the
ee

27
28 ith attribute’s contribution to a SOM model, as:
29
rR

30
31 /
)* )
& = ' ( , (6)
32

∑- )*, )
ev

33
34 . ,.
35
36
iew

and
37

ℎ
38
39
( = , (7)
40 2
41
42 where hj is the number of multiattribute training samples that are nearest to the jth prototype
43
44
vector, M is the total number of multiattribute training samples, dj represents the density of
45
46
47 training samples assigned to the jth prototype vectors, and pjk is the value of the jth prototype
48
49 vector along dimension k (the dimension of the kth attribute). Physically, if a prototype vector
50
51 has a very large value in the dimension of the target attribute, and a large percentage of training
52
53
54 samples are close to this prototype vector, then the target attribute’s contribution at this prototype
55
56
57
58 14
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 15 of 61 GEOPHYSICS

1
2
3 vector is significant. Summing up over all the prototype vectors, we then arrive at the target
4
5
6 attribute’s contribution to the whole SOM model.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Unfortunately, after some tests, we conclude that this contribution 4 depends on the
8
9
10
11 combination of attributes, which means the rank of contribution for a given group of attributes
12
13
14
changes by adding a new attribute in the group. Such phenomenon is a drawback of using PCA
15
16 to initialize the SOM neurons. In our SOM implementation, the SOM neurons are initialized
17
18 uniformly in the 2D space defined by the first two principal components of the input attributes.
19
Fo
20
The first two principal components will change by adding or dropping attributes from the input
21
22
23 group; as a result, the projection from input attribute space to the 2D SOM space changes, while
rP

24
25 the contribution of each attribute changes as well. To ensure an overall optimal contribution
26
ee

27 measurement for a given group of attributes, we propose to use an exhaustive search over all S
28
29
rR

30 combinations of three or more attributes,

31
-
32 !
5=' , (8)
! ( − )!
ev

33
.7
34
35
36
iew

37
and then weight by the number of attributes in each combination and take the sum over all used
38
39 combinations:
40
41 <
42 &9  = ' : &
;: , (9)
43
:.
44
45
46 where ! denotes factorial operation, S is the total number of SOM models to be searched, Nl is
47
48 the number of attributes in the lth combination, &
;: is the contribution of the ith attribute to the
49
50 lth SOM model, and &9  is the final contribution of the ith attribute to SOM. Although the
51
52
53 proposed method involves running SOM multiple times with different input attribute
54
55 combinations, it is an embarrassingly parallel problem so that the increase in computation time
56
57
58 15
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 16 of 61

1
2
3 over the traditional SOM is negligible given sufficient amount of threads/processors. Table 2
4
5
6 provides computation cost for running with equally-weighted attributes (building one SOM
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 model), adaptively-weighted attributes in parallel mode (building S SOM models), and

9
10 adaptively-weighted attributes in sequential mode (building S SOM models).
11
12
13
14
To quantify which attributes are favored by interpreters to be used for SOM facies
15
16 analysis, we first need to define an ideal facies map. Practically, an interpreter can either use
17
18 different colors, or delineate boundaries to define different facies. The authors favor color
19
Fo
20
representation, because then the boundary information becomes complementary once a color
21
22
23 change is identified. In a facies map, one facies corresponds to one and only one label (color), so
rP

24
25 that geobodies can be extracted using these facies labels. In contrast, if we define a facies map in
26
ee

27 terms of boundaries compartmentalizing facies, we lose the information of facies themselves. In

28
29
rR

30 order to generate a map of facies instead of facies boundaries from pattern recognition based on
31
32 seismic attributes, the input attributes have to follow the trend on facies, but not facies
ev

33
34 boundaries, which are usually discontinuities. Sometimes, a facies is defined by texture of
35
36
iew

37
discontinuities, such as mass transport complex, syneresis, and karst. Then we may use texture
38
39 attributes to “group” such discontinuities and form a facies. In real world, discontinuities and
40
41 deformations consist only a small portion of the seismic data, therefore the data distribution of
42
43
attributes such as coherence and curvature are highly skewed and sharp, and they can dominate
44
45
46 in voxel based classification methods. A window based seismic facies classification method is
47
48 proposed by Song et al. (2017), however the authors use waveforms instead of multiple
49
50 attributes. Here we loosely define an “edge” attribute to be an attribute representing the variation
51
52
53 among neighboring seismic samples, and a “body” attribute to be an attribute representing a
54
55 property of a seismic sample or samples within a window, which is usually along structure dip
56
57
58 16
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 17 of 61 GEOPHYSICS

1
2
3 and of cylindrical or cubical shape. Figure 8 shows an example of two attributes along a horizon
4
5
6 slice from the Ellenburger formation, Fort Worth Basin. The GLCM homogeneity is a texture
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 attribute that provides a measure of the amplitude smoothness along a horizon, which is a “body”
9
10 attribute. In contrast, the energy ratio similarity highlights edges, but ignores changes outside the
11
12
13
edges. Researchers in geostatistics routinely inspect histograms to determine if variables possess
14
15 distribution that satisfies the assumption for subsequent analysis, and even transform a variable
16
17 using quantile representation if its distribution is highly skewed (Coburn et al., 2005). Figure 8c
18
19
and 8d show the histogram of these two attributes, from which we observe that the body attribute
Fo
20
21
22 exhibits a flatter and more symmetric histogram, whereas the edge attribute’s histogram is tighter
23
rP

24 and skewed. Based on this observation, we proposed to use skewness and kurtosis of attributes,
25
26 which measures the asymmetry and sharpness of data distribution, to quantify the interpreter’s
ee

27
28
29 preference of body attributes over edge attributes.
rR

30
31
32 Skewness, which is the third moment of the standard score of a variable x, is defined as:
ev

33
34 ? − ?̅ 7
>(?) = @ AB E F, (10)
D
35
36
iew

37
38 where ?̅ is the mean of variable x, D is the standard deviation, and @ represents expectation.
39
40
Similarly, kurtosis is the fourth moment of the standard score of a variable x and is defined as:
41

? − ?̅ H
42
G(?) = @ AB E F. (11)
43
44 D
45
46
47 In practice, the skewness and kurtosis are precomputed before determining the attribute
48
49 contribution 4. After the computation of 4, we further normalize both skewness and kurtosis to
50
51 range between zero and one. Weighting the previously defined 4 using skewness and kurtosis,
52
53
54 and normalize the raw weights wi using the z-score to arrive at relative weights (the absolute
55
56 value of wi is of less interest):
57
58 17
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 18 of 61

1
2
3 |> | − min |> | G − min G
. ,- . ,-
4
I = J3 − − R &9  , (12)
5 max |> | − min |> | max G − min G
6 . ,- . ,- . ,- . ,-

I − I
U
7
I
T = . (13)
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 V
10
Here, I is the weight of attribute i before z-score normalization, I
U is the mean of I , V is the
11
12

standard deviation, and I

T  is the relative weight of attribute i after the z-score. In equation 12,
13
14
15
16 because the skewness term and kurtosis term are both normalized to range between zero and one,
17
18
we assume an equal impact of skewness and kurtosis. Finally, we constrain the weight to range
19
Fo
20
21 from zero to two using a sigmoid function, and defining the elements of the diagonal weight
22
23
rP
matrix W to be:
24
25 2
W = . (14)
1 + Y VTZ
26
ee

27
28
29 Using this heuristic approach, we have defined a weighting parameter as a function of the
rR

30
31 interpreter’s knowledge and SOM response. We then use the distance definition in equation 5
32
ev

33
with the updated weight matrix W to run SOM a final time, and generate the facies map. We
34
35
36 provide a workflow summarizing the steps in Figure 9.
iew

37
38
39 FIELD EXAMPLE 1: MAPPING STRUCTURAL FEATURES
40
41
42 We perform a SOM analysis on attributes computed within a 50 ms time window below
43
44
45 the top Ellenburger surface in order to capture the structural deformation and diagenetic
46
47 alteration of the shallow Ellenburger formation, which has a direct impact on the Barnett Shale
48
49 reservoir lying unconformably above. While interaction with crossplots of two or three attributes
50
51
52
(against x-, y-, and an additional color-axis, if needed) is simple, interacting with eight attributes
53
54 and an 8D crossplot is intractable. In our implementation, SOM projects this intractable 8D
55
56 crossplot into a 2D space that can be visualized on a computer screen, while still preserving most
57
58 18
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 19 of 61 GEOPHYSICS

1
2
3 of the topological and distance relationship from the original 8D space. After precomputing the
4
5
6 attribute weight matrix as described previously, we obtain the weights of the eight input
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 attributes and display with their corresponding histogram in Figure 10. Note that highly skewed
9
10 attributes (curvature and similarity) are penalized; in contrast, “body” attributes are amplified.
11
12
13
With such a penalty on curvature and similarity attributes, one may suspect that the anomalies
14
15 from these attributes would have been overly mitigated. We assume that the contrast in curvature
16
17 and similarity between features of interest (karst and faults, versus planar features in this
18
19
example) still allows them to be classified by the SOM facies map. The majority of geology is
Fo
20
21
22 planar and continuous, such that the mean value of curvature is close to zero and the mean value
23
rP

24 of similarity close to one. In contrast, structural or curvature anomalies of interest have values
25
26 several standard deviations from the mean. Therefore, we still expect to observe the anomalies
ee

27
28
29 after penalizing on these attributes.
rR

30
31
32 The SOM facies map with equally weighed input attributes is shown in Figure 11, and
ev

33
34 from the proposed workflow using the same group of input attribute in Figure 12. In our
35
36
iew

37
implementation, we output the SOM facies map as two volumes, one for each axis of the
38
39 projected 2D space. We then blend these two volumes together to generate a SOM facies map as
40
41 shown in Figure 11 and 12. Therefore, every sample in these SOM facies map is no longer a
42
43
scalar, but a two dimensional vector that represents the coordinates of that sample in the
44
45
46 projected SOM 2D space. This vector representation enables us to plot each dimension against
47
48 two 1D color bars perpendicular to each other. By using commercial crossplotting tools available
49
50 in many interpretation packages provides the color bar shown in Figure 11 and 12. Such
51
52
53 representation has greatly enhanced the interpretation of SOM facies.
54
55
56
57
58 19
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 20 of 61

1
2
3 Comparing Figure 11 and 12, both SOM facies maps are able to delineate the karst,
4
5
6 faults, and fractures equally well. This observation verifies the assumption that we made earlier,
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 that adding a penalty weight does not significantly alter the curvature and similarity anomaly
9
10 contributions. The polygonal karst regions are characterized by purple and cyan facies, where
11
12
13
purple corresponds to anticlinal components and cyan synclinal components. Compared to the
14
15 co-rendered structural curvatures, both SOM facies maps provide details about smaller scaled
16
17 karst caves that are not identifiable on structural curvatures, most of which correspond to fracture
18
19
joints (blue arrows). We are also able to identify the major faults (red arrows) close to the
Fo
20
21
22 polygonal karst regions, suggesting a tectonic control of the karst development (Khatiwada et al.,
23
rP

24 2013, Qi at al., 2014). The main difference between Figure 11 and 12 comes from regions
25
26 marked with yellow and orange arrows. In Figure 12, the yellow arrow regions are in a lime
ee

27
28
29 green facies, where the orange arrow regions are in an orange facies. In contrast, these regions
rR

30
31 look nearly identical in Figure 11, all being brownish cellular textures that somehow follow the
32
ev

33 trend on the curvature attributes. The lime green versus orange facies change in Figure 12
34
35
36
reflects the frequency variation found in Figure 8, where low peak frequency regions are in lime
iew

37
38 green facies (yellow arrows), and high frequency regions are in orange facies (orange arrows).
39
40 The peak frequency provides information on tuning thickness, which adds another dimension
41
42
besides surface morphology. The SOM facies map from equally weighted attributes, on the other
43
44
45 hand, does not distinguish such frequency variation clearly. After z-score normalization, the
46
47 contrast in some attributes (such as curvature and similarity in this example) are much higher
48
49 than the others (such as spectral decomposition attributes) and dominates the distance
50
51
52 calculation. SOM is biased towards finding clusters in those attributes. After stretching and
53
54 squeezing the distance along each attribute dimension using the precomputed weights, we are
55
56
57
58 20
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 21 of 61 GEOPHYSICS

1
2
3 able to recover the variation in peak frequency that is otherwise buried in the strong curvature
4
5
6 contrast. Figure 13 provides a SOM result generated from a subset of equally weighted attributes.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 In this example, we only use structural curvatures, peak spectral frequency and magnitude, and
9
10 GLCM homogeneity as input attributes, with the objective to better define the extent of
11
12
13
polygonal karst while reducing the dominance of edge attributes that we previously observed
14
15 when there were three more edge attributes. By eliminating the overprint from amplitude
16
17 curvature and similarity, we recognize a cleaner silhouette of the polygonal karst; however, the
18
19
change in frequency in the less deformed regions is still difficult to identify. This facies map is
Fo
20
21
22 excellent in delineating structural deformations of karst, faults, and fractures, but fails to properly
23
rP

24 represent frequency information.

25
26
ee

27 To look for evidence of the facies discovered in Figure 12, we take four composite
28
29
rR

30 vertical sections (a) – (d) from seismic amplitude, and co-render with the SOM facies (Figure
31
32 14). In order to allow the background seismic amplitude to come through, we modulate the
ev

33
34 opacity of the SOM facies, which leads to faded color on the vertical sections compared to on the
35
36
iew

37
map along Horizon A. Horizon A lies vertically at the center of the SOM analysis window, so we
38
39 expect to see the same facies in Figure 12 to appear at the center (vertically) of the colored zone
40
41 on vertical sections. In section (b), we use a red-white-blue colorbar for seismic amplitude to
42
43
better visualize the thickness change in reflectors, where the black curve marks Horizon A. We
44
45
46 see a change in layer thickness in the two red ovals, which verifies the facies change that we
47
48 previously interpreted. The karst features identified on Horizon A nicely match the synclinal
49
50 events on seismic amplitude profile (marked as red curves), with the perimeter in purple (which
51
52
53 fades to magenta on the vertical sections) delineating the extent. We identify a large dome
54
55 shaped karst developed in the center of a collapse in section (d), which may be related to the
56
57
58 21
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 22 of 61

1
2
3 compression that caused the two reverse faults on the east and west sides (blue curves). Most
4
5
6 polygonal karst features develop along or in the vicinity of faults, and many of the smaller scale,
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 isolated karst caves appear at joints of folds. The location of karst indicates that fractures of
9
10 different scales might have accelerated the weathering and dissolution process of the Ellenburger
11
12
13
formation.
14
15
16 FIELD EXAMPLE 2: MAPPING STRATIGRAPHIC FEATURES
17
18
19 By now we have demonstrated that by adaptively weighting the input attribute,
Fo
20
21
information from input attributes are more adequately presented in the SOM facies map. To
22
23
rP

24 further investigate the effectiveness of the proposed attribute weighting scheme, we provide a
25
26 second example from the Barnett Shale, in which the main target is to map the interbedded layers
ee

27
28 within the Barnett Shale formation overlying the Ellenburger formation. In this example, we use
29
rR

30
31 structural attributes and inversion attributes as inputs to SOM, where the weights calculated
32
ev

33 using the proposed approach are provided in Table 3. Inversion attributes are closely related to
34
35 the elastic properties and lithology of the rocks, and therefore are effectively used to characterize
36
iew

37
38
shale reservoirs. On the other hand, structural attributes are useful when mapping faults and folds
39
40 in the shale formation. Figure 15 shows SOM facies maps from both equally weighted (Figure
41
42 15a) and data-adaptively weighted (Figure 15b) input attributes. We observe that the result from
43
44
adaptively weighted attributes provides better defined layers, contain less “noise” associated with
45
46
47 the structural attributes (black ovals), and still shows the major fault in the west of the survey.
48
49 Furthermore, while increasing the weights of the inversion attributes, we also note the weights
50
51 are consistent with the hierarchy of attribute dependence. P and S impedances are two
52
53
54 independent attributes directly obtained from seismic inversion and have the highest weights;
55
56 Lambda/Mu and Vp/Vs are calculated from P and S impedances and have more modest weights;
57
58 22
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 23 of 61 GEOPHYSICS

1
2
3 Poisson’s Ratio is calculated from Vp/Vs and has the lowest weight. Such weight values
4
5
6 indirectly verify the dependency among input attributes.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 CONCLUSIONS
10
11
12 In this study, we augment the attribute selection process with an attribute weighting
13
14
15
process for seismic facies analysis. We define the attribute weight to be a function of both SOM
16
17 response and interpreter’s preference. By using such weights, information in the input attributes
18
19 is more adequately represented in the SOM facies map, and less dominated by attributes with
Fo
20
21
very high contrast. The weights also provide information on attribute inter-dependency, which
22
23
rP

24 may allow interpreters to understand the origin and behavior of different attributes, and in turn to
25
26 select the most appropriate group of input attributes for seismic facies analysis. Being a heuristic
ee

27
28 method in nature, the proposed approach is one of the many possible ways for augmenting
29
rR

30
31 attribute selection, however we believe it is the very first attempt towards using weights
32
ev

33 complimenting simple selection. We hope that the proposed approach will inspire others to add
34
35 further metrics to the quantitative attribute selection problems.
36
iew

37
38
39
ACKNOWLEDGEMENTS
40
41
42 We thank Marathon Oil and Devon Energy for providing the seismic data used in this
43
44 study. Financial support for this effort is provided by the industrial sponsors of the Attribute-
45
46
Assisted Seismic Processing and Interpretation (AASPI) consortium at the University of
47
48
49 Oklahoma. The prestack inversion was performed using licenses from HampsonRussell software,
50
51 provided to the University of Oklahoma for research and education courtesy of CGG
52
53 GeoSoftware, and all visualizations are from Petrel, courtesy of Schlumberger.
54
55
56
57
58 23
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 24 of 61

1
2
3 APPENDIX A
4
5
6
7 Mathematical details on the SOM implementation
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10 For completeness, we provide the mathematical details on the SOM implementation used
11
12 in this study. This implementation differs from the traditional Kohonen SOM in that the SOM
13
14
15
nodes in the 2D topological map are no longer restricted at the initial (regular grid) positions.
16
17 Instead, they form an irregular grid to mimic the distribution of corresponding prototype vectors
18
19 in the multiattribute space. To emphasize the fact that their positions are flexible, we name the
Fo
20
21
SOM nodes position vectors in this appendix. This improvement enables the color representation
22
23
rP

24 of facies honor the degree of similarity found in the input attributes.

25
26
ee

27 Let’s consider a 2D SOM represented by P prototype vectors pi, pi = [pi1, pi2, …, piN],
28
29 where i = 1, 2, …, P and N is the dimension of these vectors defined by the number of input
rR

30
31
32 attributes (or samples for waveform classification). We first form an N × N normalized
ev

33
34 covariance matrix for the N attributes:

 ⋯  -
35
36
iew

37 \- = J ⋮ ⋱ ⋮ R, (`1)
38 -  ⋯ - -
39
After calculating the eigenvalues,
i, and eigenvectors, ui, of the covariance matrix, we initialize
40
41
42
43 the elliptical 2D SOM topological map as the 2D space defined by the first two eigenvectors, and
44
45 assign the position vectors (SOM nodes), ri, proportionally to the first two eigenvalues. We then
46
47
48
initialize a prototype vector, pi, as:

a = (b b )c . (`2)
49
50
51
52
53
54
55
56
57
58 24
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 25 of 61 GEOPHYSICS

1
2
3 Now we have the initial SOM model with a set of prototype vectors p and corresponding
4
5
6 position vectors r. Given a multiattribute training data vector a, we compute the Euclidean
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 distance (or alternatively, Mahalanobis distance) between vector a and all prototype vectors. The
9
10 prototype vector pb, which has the minimum distance to the input vector a, is defined to be the
11
12
13
“winner” or the best matching unit (BMU), pb:

)| − ad |) = mine)| − a |)f. (`3)

14
15
16
17 We then update the BMU and its neighbors. The updating rule for the ith prototype vector inside
18

and outside the neighborhood radius g(h) is given by

19
Fo
20
21
a (h) + j(h)ℎ(h)k − a (h)l, ∀ )|c − cd |) ≤ g(h)
a (h + 1) = i , (`4)
22
a (h), ∀ )|c − cd |) > g(h)
23
rP

24
25
26 where the neighborhood radius defined as g(h) is predefined for a problem and decreases with
ee

27
28 each iteration (epoch) t. rb and ri are the position vectors of the BMU pb and the ith prototype
29
rR

vector pi, respectively. We also define ℎ(h) as the neighborhood function, and j(h) as the
30
31
32
exponential learning rate. ℎ(h) and j(h) decrease with each iteration in the learning process and
ev

33
34
35 they are defined as
36
iew

37 $
o)cZ pcq )o

38
39 ℎ(h) = Y $r$ (s) , (`5)
40
41 and
42
u.uuv y/
j(h) = ju ( ) , (`6)
43
wx
44
45

where ju is the initial learning rate, and T is the total number of iterations. After every iteration,
46
47
48
49 we update the position vectors r by projecting the prototype vectors into the 2D SOM
50
51 topological map space:
52
 (h + 1) = a  (h)b
53
54
i . (`7)
55  (h + 1) = a  (h)b
56
57
58 25
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 26 of 61

1
2
3 In our implementation, in every iteration, we add a step of adjusting the position of all SOM
4
5
6 nodes according to their distances from the current BMU in both input multiattribute space and
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 2D SOM space. The adjustment rule is (Shao and Yang, 2012):

9
{|
c (h + 1) = c (h) + j(h) B1 − E }c| (h) − c (h)~, ∀  ≠ . (`8)
10
(|
11
12

In equation A8, c (h) is a positon vector before adjustment; c (h + 1) is the position vector after
13
14
15
16 adjustment; c| (h) is the position vector of the current BMU; {| and (| are the distance between
17
18
19
a prototype vector and the current BMU in input multiattribute space and 2D SOM space,
Fo
20
21 respectively. The above summarizes a full iteration in our SOM learning process. After the final
22
23
rP
iteration, we color code the position vectors using a 2D color map, and assign a data vector the
24
25
color of the closest prototype vector’s corresponding position vector.
26
ee

27
28
29
rR

30
31
32
ev

33
34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58 26
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 27 of 61 GEOPHYSICS

1
2
3 REFERENCES
4
5
6
7 Ahmed, O., R. Abdel-Aal, and H. AlMustafa, 2010, Reservoir property prediction using
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 abductive networks: Geophysics, 75(1), P1–P9.
10
11
12 AlBinHassan, N., and Y. Wang, 2011, Porosity prediction using the group method of data
13
14
15
handling: Geophysics, 76(5), O15–O22.
16
17
18 Altamar, R. P., and K. J. Marfurt, 2015, Identification of brittle/ductile areas in unconventional
19
Fo
20 reservoirs using seismic and microseismic data: Application to the Barnett Shale:
21
22
Interpretation, 3, T233-T243.
23
rP

24
25
26 Amin, A., M. Deriche, M. A. Shafiq, Z. Wang, and G. AlRegib, 2017, Automated salt-dome
ee

27
28 detection using an attribute ranking framework with a dictionary-based classifier:
29
rR

30 Interpretation, 5, SJ61-SJ79.
31
32
ev

33
34 Barnes, A. E., 2007, Redundant and useless seismic attributes: Geophysics, 72(3), P33-P38.
35
36
iew

37 Barnes, A. E., and K. J. Laughlin, 2002, Investigation of methods for unsupervised classification
38
39 of seismic data: 72nd Annual International Meeting, SEG, Expanded Abstracts, 2221-
40
41
42
2224.
43
44
45 Benabdeslem, K., and M. Lebbah, 2007, Feature selection for self-organizing map: IEEE 29th
46
47 International Conference on Information Technology Interfaces, 45-50.
48
49
50
Canter, K. L., D. B. Stearns, R. C. Geesaman, and J. L. Wilson, 1993, Paleostructural and related
51
52
53 paleokarst controls on reservoir development in the Lower Ordovician Ellenburger
54
55
56
57
58 27
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 28 of 61

1
2
3 Group, Val Verde basin, in R. D. Fritz, J. L. Wilson, and D. A. Yurewicz, eds.,
4
5
6 Paleokarst Related Hydrocarbon Reservoirs: SEPM Core Workshop 18, 61-99.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 Chen, Q., and S. Sidney, 1997, Seismic attribute technology for reservoir forecasting and
10
11 monitoring: The Leading Edge, 16, 445-448.
12
13
14
15
Coburn, T. C., J. M. Yarus, and R. L. Chambers, 2005, Stochastic modeling and geostatistics:
16
17 principles, methods, and case studies, vol. II, AAPG computer applications in geology 5:
18
19 AAPG.
Fo
20
21
22
Coléou, T., M. Poupon, and K. Azbel, 2003, Unsupervised seismic facies classification: A
23
rP

24
25 review and comparison of techniques and implementation: The Leading Edge, 22, 942–
26
ee

27 953.
28
29
rR

30 Di, H. and D. Gao, 2017, Nonlinear gray-level co-occurrence matrix texture analysis for
31
32
improved seismic facies interpretation, Interpretation, 5, SJ31-SJ40.
ev

33
34
35
36
iew

Dorrington, K. P., and C. A. Link, 2004, Genetic-algorithm/neural-network approach to seismic

37
38 attribute selection for well-log prediction: Geophysics, 69(1), 212-221.
39
40
41
42
Gao, D., 2003, Volume texture extraction for 3D seismic visualization and interpretation:
43
44 Geophysics, 68(4), 1294-1302.
45
46
47 Hampson, D. P., J. S. Schuelke, and J. A. Quirein, 2001, Use of multiattribute transforms to
48
49
predict log properties from seismic data, Geophysics, 66(1), 220-236.
50
51
52
53
54
55
56
57
58 28
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 29 of 61 GEOPHYSICS

1
2
3 Hart, B. S., and R. S. Balch, 2000, Approaches to defining reservoir physical properties from 3-
4
5
6 D seismic attributes with limited well control: An example from the Jurassic Smackover
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 Formation, Alabama: Geophysics, 65(2), 368-376.

9
10
11 Hu, C., W. Lu, and Y. Zhang, 2014, Seismic facies classification using visualization-induced
12
13
14
self-organizing mapping: 76th EAGE Conference and Exhibition 2014.
15
16
17 Kalkomey, C. T., 1997, Potential risks when using seismic attributes as predictors of reservoir
18
19 properties: The Leading Edge, 16, 247-251.
Fo
20
21
22
Kerans, C., 1988, Karst-controlled reservoir heterogeneity in Ellenburger Group carbonates of
23
rP

24
25 West Texas: AAPG Bulletin, 72, 1160–1183.
26
ee

27
28 Khatiwada, M., G. R. Keller, and K. J. Marfurt, 2013, A window into the Proterozoic:
29
rR

30 Integrating 3D seismic, gravity and magnetic data to image subbasement structures in the
31
32
southeast Fort Worth basin: Interpretation, 1, T125–T141.
ev

33
34
35
36
iew

Kohonen, T., 1982, Self-organized formation of topologically correct feature maps: Biological
37
38 Cybernetics, 43, 59–69.
39
40
41
42
Kupecz, J.A., and L. S. Land, 1991, Late-stage dolomitization of the Lower Ordovician
43
44 Ellenburger Group, West Texas: Journal of Sedimentary Petrology, 61, 551-574.
45
46
47 Kuwahara, M., K. Hachimura, S. Eiho, and M. Kinoshita, 1976, Processing of RI
48
49
angiocardiographic images, in Preston, K., and M. Onoe, Eds., Digital processing of
50
51
52 biomedical images: Plenum Press, 187–202.
53
54
55
56
57
58 29
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 30 of 61

1
2
3 Leiphart, D. J., and B. S. Hart, 2001, Comparison of linear regression and a probabilistic neural
4
5
6 network to predict porosity from 3-D seismic attributes in lower Brushy Canyon
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 channeled sandstones, southeast New Mexico: Geophysics, 66(5), 1349–1358.

9
10
11 Mahalanobis, P. C., 1936, On the generalized distance in statistics: Proceedings of the National
12
13
14
Institute of Sciences of India, 2, 49–55.
15
16
17 Marroquín, I. D., J.-J. Brault and B. S. Hart, 2009, A visual data-mining methodology for
18
19 seismic facies analysis: Part 1 — Testing and comparison with other unsupervised
Fo
20
21
clustering methods: Geophysics, 74(1), P1–P11.
22
23
rP

24
25 Matos, M. C., K. J. Marfurt, and P. R. S. Johann, 2009, Seismic color self-organizing maps:
26
ee

27 Presented at 11th International Congress of the Brazilian Geophysical Society, Extended

28
29 Abstracts.
rR

30
31
32
Matos, M. C., P. L. Osório and P. R. Johann, 2007, Unsupervised seismic facies analysis using
ev

33
34
35 wavelet transform and self-organizing maps: Geophysics, 72(1), P9–P21.
36
iew

37
38 Matos, M. C., M. Yenugu, S. M. Angelo, and K. J. Marfurt, 2011, Integrated seismic texture
39
40
41
segmentation and cluster analysis applied to channel delineation and chert reservoir
42
43 characterization: Geophysics, 76(5), P11–P21.
44
45
46 Pollastro, R. M., D. M. Jarvie, R. J. Hill, and C. W. Adams, 2007, Geologic framework of the
47
48
Mississippian Barnett shale, Barnett-paleozoic total petroleum system, Bend arch — Fort
49
50
51 Worth Basin, Texas: AAPG Bulletin, 91, 405–436.
52
53
54
55
56
57
58 30
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 31 of 61 GEOPHYSICS

1
2
3 Poupon, M., K. Azbel, and G. Palmer, 1999, A new methodology based on seismic facies
4
5
6 analysis and litho-seismic modeling: The Elkhorn Slough field pilot project, Solano
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 County, California: 69th Annual International Meeting, SEG, Expanded Abstracts, 927–
9
10 930.
11
12
13
14
Qi, J., B. Zhang, H. Zhou, and K. Marfurt, 2014, Attribute expression of fault-controlled karst —
15
16 Fort Worth Basin, Texas: A tutorial: Interpretation, 2, SF91-SF110.
17
18
19 Roden, R., T. Smith, and D. Sacrey, 2015, Geologic pattern recognition from seismic attributes:
Fo
20
21
Principal component analysis and self-organizing maps: Interpretation, 4, SAE59-SAE83.
22
23
rP

24
25 Roy, A., 2013, Latent space classification of seismic facies: Ph.D. Dissertation, University of
26
ee

27 Oklahoma.
28
29
rR

30 Roy, A., B. L. Dowdell, and K. J. Marfurt, 2013, Characterizing a Mississippian tripolitic chert
31
32
reservoir using 3D unsupervised and supervised multiattribute seismic facies analysis: An
ev

33
34
35 example from Osage County, Oklahoma: Interpretation, 1, SB109–SB124.
36
iew

37
38 Roy, A., A. S. Romero-Peláez, T. J. Kwiatkowski, and K. J. Marfurt, 2014, Generative
39
40
41
topographic mapping for seismic facies estimation of a carbonate wash, Veracruz Basin,
42
43 southern Mexico: Interpretation, 2, SA31-SA47.
44
45
46 Schuelke, J. S., and J. A. Quirein, 1998, Validation: A technique for selecting seismic attributes
47
48
and verifying results: 68th Annual International Meeting, SEG, Expanded Abstracts, 936-
49
50
51 939.
52
53
54
55
56
57
58 31
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 32 of 61

1
2
3 Shao, C. and Y. Yang, 2012, Distance-preserving SOM: a new data visualization algorithm:
4
5
6 Journal of Software, 7, 196–203.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 Song, C., Z. Liu, Y. Wang, X. Li, and G. Hu, 2017, Multi-waveform classification for seismic
10
11 facies analysis: Computers & Geosciences, 101, 1-9.
12
13
14
15
Strecker, U., and R. Uden, 2002, Data mining of 3D post- stack attribute volumes using Kohonen
16
17 self-organizing maps: The Leading Edge, 21, 1032–1037.
18
19
Fo
20 Sullivan, E. C., K. J. Marfurt, A. Lacazette, and M. Ammerman, 2006, Application of new
21
22
seismic attributes to collapse chimneys in the Fort Worth Basin: Geophysics, 71(4),
23
rP

24
25 B111-B119.
26
ee

27
28 Wang, Z., C. Yin, X. Lei, F. Gu, and J. Gao, 2015, Joint rough sets and Karhunen-Loève
29
rR

30 transform approach to seismic attribute selection for porosity prediction in a Chinese

31
32
sandstone reservoir: Interpretation, 3, SAE19-SAE28.
ev

33
34
35
36
iew

Zhao, T., V. Jayaram, A. Roy, and K. J. Marfurt, 2015, A comparison of classification

37
38 techniques for seismic facies recognition: Interpretation, 3, SAE29-SAE58.
39
40
41
42
Zhao, T., J. Zhang, F. Li, and K. J. Marfurt, 2016, Characterizing a turbidite system in
43
44 Canterbury Basin, New Zealand, using seismic attributes and distance-preserving self-
45
46 organizing maps: Interpretation, 4, SB79-SB89.
47
48
49
Zhao, T., F. Li, and K. J. Marfurt, 2017, Constraining self-organizing map facies analysis with
50
51
52 stratigraphy: An approach to increase the credibility in automatic seismic facies
53
54 classification: Interpretation, 5, T163-T171.
55
56
57
58 32
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 33 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31
32
ev

33
34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58 33
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 34 of 61

1
2
3 LIST OF FIGURE CAPTIONS
4
5
6
7 Figure 1. An example of randomly selected 50 sample points of two attributes from the field
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 data used in this study to show the effect of attribute weights in k-means clustering. Before
10
11 clustering, the attributes are normalized using a z-score algorithm. (a) Giving equally weighted
12
13
14
coherence and peak frequency attributes, k-means finds four clusters represented by the four
15
16 colors. The yellow cluster contains strong variation in peak frequency, which almost certainly is
17
18 a mix of different geologic facies. (b) The same data samples but now with less weight applied to
19
Fo
20
coherence than the peak frequency. This clustering makes more geological sense, with one
21
22
23 cluster represents very low coherence (discontinuity), and three clusters of varying peak
rP

24
25 frequency. The circles denote cluster centers.
26
ee

27
28 Figure 2. Stratigraphic cross section of the Fort Worth Basin. In the study area, the Barnett Shale
29
rR

30
31 lies unconformably on top of the dolomitic Ellenburger formation. The seismic survey showing
32
ev

33 the Ellenburger karst is to the south of Young County in this plot (after Qi et al., 2014).
34
35
36
iew

Figure 3. (a) Seismic amplitude section from line XX’ (location marked in Figure 3b and 3c);
37
38
39
(b) seismic amplitude at t = 0.75 s; (c) seismic amplitude along phantom horizon A, which is 25
40
41 ms below the top of Ellenburger formation. White arrows mark some of the proposed karst
42
43 features. Green dashed lines outline the proposed highly karsted regions.
44
45
46
Figure 4. Co-rendered structural curvatures k1 and k2 along phantom Horizon A 25 ms below the
47
48
49 top of the Ellenburger formation. Black dashed lines outline the extent of highly karsted regions
50
51 (including polygonal karst). Red arrows can be interpreted as locations of large regional faults.
52
53 Blue arrows can be interpreted as small scale, isolated karst collapse features. Yellow arrows can
54
55
56
57
58 34
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 35 of 61 GEOPHYSICS

1
2
3 be interpreted as less deformed regions of relatively thick layers, while orange arrows less
4
5
6 deformed regions of relatively thin layers.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 Figure 5. Co-rendered amplitude curvatures epos and eneg along phantom Horizon A 25 ms below
10
11 the top of the Ellenburger formation Red arrows denote locations of large regional faults.
12
13
14
Features marked by colored arrows are discussed in Figure 4. Amplitude curvature shows more
15
16 details about the small scale karst.
17
18
19 Figure 6. Peak spectral frequency modulated by peak spectral magnitude along phantom
Fo
20
21
Horizon A 25 ms below the top of the Ellenburger formation. Features marked by colored arrows
22
23
rP

24 are discussed in Figure 4. We choose a solid gray colorbar with varying opacity for peak spectral
25
26 magnitude to modulate the peak spectral frequency. The polygonal karst regions are of low
ee

27
28 frequency and low magnitude, possibly due to the non-specular reflection from the highly
29
rR

30
31 deformed reflectors.
32
ev

33
34 Figure 7. Co-rendered GLCM homogeneity and energy ratio similarity along phantom Horizon
35
36
iew

A 25 ms below the top of Ellenburger formation. Features marked by colored arrows are
37
38
39
discussed in Figure 4.
40
41
42 Figure 8. Phantom Horizon A 25 ms below the top of Ellenburger formation through the (a)
43
44 GLCM homogeneity (with different colorbar from the display in Figure 7) and (b) energy ratio
45
46
similarity volumes. Features marked by colored arrows are discussed in Figure 4. We loosely
47
48
49 define the GLCM homogeneity as a “body” attribute as it represents information within a
50
51 window along structural dip. It is useful for mapping the extent of the highly deformed regions.
52
53 In contrast, the energy ratio similarity is an edge attribute that highlights the boundaries. (c) The
54
55
56 histogram of GLCM homogeneity within the analysis window. (d) The histogram of energy ratio
57
58 35
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 36 of 61

1
2
3 similarity within the analysis window. Comparing to GLCM homogeneity, the histogram of
4
5
6 energy ratio similarity is highly skewed and sharp.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 Figure 9. The proposed attribute weighting workflow. Mathematical notations are discussed in
10
11 the main text.
12
13
14
15
Figure 10. Histograms of the eight input attributes within the analysis window. Weights are
16
17 computed using the method described in this paper. Note that histograms exhibiting high kurtosis
18
19 and/or skewness are assigned lower weights, while those that are broader and more symmetric
Fo
20
21
are assigned higher weights. Values at the bottom of each histogram mark the z-scores, and
22
23
rP

24 dashed lines mark values at zero, one, and two standard deviations. Note curvature and
25
26 coherence attributes span several standard deviations.
ee

27
28
29 Figure 11. SOM facies map from equally weighted attributes along phantom Horizon A 25 ms
rR

30
31
32 below the top of the Ellenburger formation. Features marked by colored arrows are discussed in
ev

33
34 Figure 4. The dashed lines in the 2D color maps denotes the 1D colorbar used for each projection
35
36
iew

along a SOM axis in order to approximate the 2D color map.

37
38
39
40
Figure 12. SOM facies map from adaptively weighted attributes along phantom Horizon A 25
41
42 ms below the top of the Ellenburger formation. Features marked by colored arrows are discussed
43
44 in Figure 4. The dashed lines in the 2D color maps denotes the 1D colorbar used for each
45
46
projection along a SOM axis in order to approximate the 2D color map.
47
48
49
50 Figure 13. SOM facies map from a subset of equally weighted attributes along phantom Horizon
51
52 A 25 ms below the top of the Ellenburger formation. Features marked by colored arrows are
53
54
55
56
57
58 36
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 37 of 61 GEOPHYSICS

1
2
3 discussed in Figure 4. The dashed lines in the 2D color maps denotes the 1D colorbar used for
4
5
6 each projection along a SOM axis in order to approximate the 2D color map.
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9 Figure 14. SOM facies map from adaptively weighted attributes along phantom Horizon A 25
10
11 ms below the top of the Ellenburger formation. Features marked by colored arrows are discussed
12
13
14
in Figure 4. The dashed lines in the 2D color maps denotes the 1D colorbar used for each
15
16 projection along a SOM axis in order to approximate the 2D color map. Inserts (a) to (d) shows
17
18 seismic evidence of the interpreted facies.
19
Fo
20
21
Figure 15. Vertical sections at Line X from a second Barnett Shale seismic survey. (a) SOM
22
23
rP

24 facies map with equally weighted input attributes; (b) SOM facies map from adaptively weighed
25
26 input attributes. Weights for each attribute are provided in Table 3. We notice the layers within
ee

27
28 the shale formations are better defined in (b), and with less noise associated with structural
29
rR

30
31 attributes (black ovals).
32
ev

33
34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58 37
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 38 of 61

1
2
3 LIST OF TABLE CAPTIONS
4
5
6
7 Table 1. Attributes used in this study and why using them.
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10 Table 2. Computation cost for SOM running with (1) equally-weighted attributes in sequential
11
12 mode, (2) equally-weighted attributes in parallel mode, (3) adaptively-weighted attributes in
13
14
15
sequential mode, (4) adaptively-weighted attributes in partial parallel mode, and (5) adaptively-
16
17 weighted attributes in full parallel mode. Within the analysis window defined by two horizons,
18
19 the total amount of data vectors is about 35 million, from which we use 1% for building SOM
Fo
20
21
models. The suboptimal computation cost reduction in the final clustering stage is due to the fact
22
23
rP

24 that the parallelization conducts at trace level. In this study, we use a very small vertical analysis
25
26 window, which results in a considerable amount of communication overhead in the final
ee

27
28 clustering. The parallelization is implemented using Intel MPI.
29
rR

30
31
32 Table 3. Input attributes and their corresponding weights used in the second Barnett Shale
ev

33
34 example. Note the weight values of inversion attributes are consistent with attribute dependency.
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58 38
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 39 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 1 An example of randomly selected 50 sample points of two attributes from the field data used in this study
32 to show the effect of attribute weights in k-means clustering. Before clustering, the attributes are
ev

33 normalized using a z-score algorithm. (a) Giving equally weighted coherence and peak frequency attributes,
34 k-means finds four clusters represented by the four colors. The yellow cluster contains strong variation in
35 peak frequency, which almost certainly is a mix of different geologic facies. (b) The same data samples but
now with less weight applied to coherence than the peak frequency. This clustering makes more geological
36
iew

sense, with one cluster represents very low coherence (discontinuity), and three clusters of varying peak
37 frequency. The circles denote cluster centers.
38
39 254x190mm (300 x 300 DPI)
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 40 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 2 Stratigraphic cross section of the Fort Worth Basin. In the study area, the Barnett Shale lies
32 unconformably on top of the dolomitic Ellenburger formation. The seismic survey showing the Ellenburger
ev

33 karst is to the south of Young County in this plot (after Qi et al., 2014).
34
35 254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 41 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 3a. (a) Seismic amplitude section from line XX’ (location marked in Figure 3b and 3c); (b) seismic amplitude
32 at t = 0.75 s; (c) seismic amplitude along phantom horizon A, which is 25 ms below the top of Ellenburger
ev

33 formation. White arrows mark some of the proposed karst features. Green dashed lines outline the proposed
34 highly karsted regions.
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 42 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 3b. (a) Seismic amplitude section from line XX’ (location marked in Figure 3b and 3c); (b) seismic amplitude
32 at t = 0.75 s; (c) seismic amplitude along phantom horizon A, which is 25 ms below the top of Ellenburger
ev

33 formation. White arrows mark some of the proposed karst features. Green dashed lines outline the proposed
34 highly karsted regions.
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 43 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 3c. (a) Seismic amplitude section from line XX’ (location marked in Figure 3b and 3c); (b) seismic amplitude
32 at t = 0.75 s; (c) seismic amplitude along phantom horizon A, which is 25 ms below the top of Ellenburger
ev

33 formation. White arrows mark some of the proposed karst features. Green dashed lines outline the proposed
34 highly karsted regions.
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 44 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 4 Co-rendered structural curvatures k1 and k2 along phantom Horizon A 25 ms below the top of the
32 Ellenburger formation. Black dashed lines outline the extent of highly karsted regions (including polygonal
ev

33 karst). Red arrows can be interpreted as locations of large regional faults. Blue arrows can be interpreted as
34 small scale, isolated karst collapse features. Yellow arrows can be interpreted as less deformed regions of
35 relatively thick layers, while orange arrows less deformed regions of relatively thin layers.
36
iew

254x190mm (300 x 300 DPI)

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 45 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 5 Co-rendered amplitude curvatures epos and eneg along phantom Horizon A 25 ms below the top of the
32 Ellenburger formation Red arrows denote locations of large regional faults. Features marked by colored
ev

33 arrows are discussed in Figure 4. Amplitude curvature shows more details about the small scale karst.
34
35 254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 46 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 6 Peak spectral frequency modulated by peak spectral magnitude along phantom Horizon A 25 ms below the
32 top of the Ellenburger formation. Features marked by colored arrows are discussed in Figure 4. We choose a
ev

33 solid gray colorbar with varying opacity for peak spectral magnitude to modulate the peak spectral
34 frequency. The polygonal karst regions are of low frequency and low magnitude, possibly due to the non-
35 specular reflection from the highly deformed reflectors.
36
iew

254x190mm (300 x 300 DPI)

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 47 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 7 Co-rendered GLCM homogeneity and energy ratio similarity along phantom Horizon A 25 ms below the top
32 of Ellenburger formation. Features marked by colored arrows are discussed in Figure 4.
ev

33
34 254x190mm (300 x 300 DPI)
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 48 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 8a Phantom Horizon A 25 ms below the top of Ellenburger formation through the (a) GLCM homogeneity
32 (with different colorbar from the display in Figure 7) and (b) energy ratio similarity volumes. Features
ev

33 marked by colored arrows are discussed in Figure 4. We loosely define the GLCM homogeneity as a “body”
34 attribute as it represents information within a window along structural dip. It is useful for mapping the
35 extent of the highly deformed regions. In contrast, the energy ratio similarity is an edge attribute that
highlights the boundaries. (c) The histogram of GLCM homogeneity within the analysis window. (d) The
36
iew

histogram of energy ratio similarity within the analysis window. Comparing to GLCM homogeneity, the
37 histogram of energy ratio similarity is highly skewed and sharp.
38
39 254x190mm (300 x 300 DPI)
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 49 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 8b Phantom Horizon A 25 ms below the top of Ellenburger formation through the (a) GLCM homogeneity
32 (with different colorbar from the display in Figure 7) and (b) energy ratio similarity volumes. Features
ev

33 marked by colored arrows are discussed in Figure 4. We loosely define the GLCM homogeneity as a “body”
34 attribute as it represents information within a window along structural dip. It is useful for mapping the
35 extent of the highly deformed regions. In contrast, the energy ratio similarity is an edge attribute that
highlights the boundaries. (c) The histogram of GLCM homogeneity within the analysis window. (d) The
36
iew

histogram of energy ratio similarity within the analysis window. Comparing to GLCM homogeneity, the
37 histogram of energy ratio similarity is highly skewed and sharp.
38
39 254x190mm (300 x 300 DPI)
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 50 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 8c Phantom Horizon A 25 ms below the top of Ellenburger formation through the (a) GLCM homogeneity
32 (with different colorbar from the display in Figure 7) and (b) energy ratio similarity volumes. Features
ev

33 marked by colored arrows are discussed in Figure 4. We loosely define the GLCM homogeneity as a “body”
34 attribute as it represents information within a window along structural dip. It is useful for mapping the
35 extent of the highly deformed regions. In contrast, the energy ratio similarity is an edge attribute that
highlights the boundaries. (c) The histogram of GLCM homogeneity within the analysis window. (d) The
36
iew

histogram of energy ratio similarity within the analysis window. Comparing to GLCM homogeneity, the
37 histogram of energy ratio similarity is highly skewed and sharp.
38
39 254x190mm (300 x 300 DPI)
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 51 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 8d Phantom Horizon A 25 ms below the top of Ellenburger formation through the (a) GLCM homogeneity
32 (with different colorbar from the display in Figure 7) and (b) energy ratio similarity volumes. Features
ev

33 marked by colored arrows are discussed in Figure 4. We loosely define the GLCM homogeneity as a “body”
34 attribute as it represents information within a window along structural dip. It is useful for mapping the
35 extent of the highly deformed regions. In contrast, the energy ratio similarity is an edge attribute that
highlights the boundaries. (c) The histogram of GLCM homogeneity within the analysis window. (d) The
36
iew

histogram of energy ratio similarity within the analysis window. Comparing to GLCM homogeneity, the
37 histogram of energy ratio similarity is highly skewed and sharp.
38
39 254x190mm (300 x 300 DPI)
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 52 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 9 The proposed attribute weighting workflow. Mathematical notations are discussed in the main text.
32
ev

33 254x190mm (300 x 300 DPI)

34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 53 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 10 Histograms of the eight input attributes within the analysis window. Weights are computed using the
32 method described in this paper. Note that histograms exhibiting high kurtosis and/or skewness are assigned
ev

33 lower weights, while those that are broader and more symmetric are assigned higher weights. Values at the
34 bottom of each histogram mark the z-scores, and dashed lines mark values at zero, one, and two standard
35 deviations. Note curvature and coherence attributes span several standard deviations.
36
iew

254x190mm (300 x 300 DPI)

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 54 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 11 SOM facies map from equally weighted attributes along phantom Horizon A 25 ms below the top of the
32 Ellenburger formation. Features marked by colored arrows are discussed in Figure 4. The dashed lines in the
ev

33 2D color maps denotes the 1D colorbar used for each projection along a SOM axis in order to approximate
34 the 2D color map.
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 55 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 12 SOM facies map from adaptively weighted attributes along phantom Horizon A 25 ms below the top of
32 the Ellenburger formation. Features marked by colored arrows are discussed in Figure 4. The dashed lines in
ev

33 the 2D color maps denotes the 1D colorbar used for each projection along a SOM axis in order to
34 approximate the 2D color map.
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 56 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 13 SOM facies map from a subset of equally weighted attributes along phantom Horizon A 25 ms below the
32 top of the Ellenburger formation. Features marked by colored arrows are discussed in Figure 4. The dashed
ev

33 lines in the 2D color maps denotes the 1D colorbar used for each projection along a SOM axis in order to
34 approximate the 2D color map.
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 57 of 61 GEOPHYSICS

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 14 SOM facies map from adaptively weighted attributes along phantom Horizon A 25 ms below the top of
32 the Ellenburger formation. Features marked by colored arrows are discussed in Figure 4. The dashed lines in
ev

33 the 2D color maps denotes the 1D colorbar used for each projection along a SOM axis in order to
34 approximate the 2D color map. Inserts (a) to (d) shows seismic evidence of the interpreted facies.
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 58 of 61

1
2
3
4
5
6
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
9
10
11
12
13
14
15
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31 15 Vertical sections at Line X from a second Barnett Shale seismic survey. (a) SOM facies map with equally
32 weighted input attributes; (b) SOM facies map from adaptively weighed input attributes. Weights for each
ev

33 attribute are provided in Table 3. We notice the layers within the shale formations are better defined in (b),
34 and with less noise associated with structural attributes (black ovals).
35
254x190mm (300 x 300 DPI)
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 59 of 61 GEOPHYSICS

1
2
3 Table 1
4
5 Attribute Why using this attribute
6
7 Due to non-specular reflection, substantial amount of the
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8 Peak spectral magnitude reflected energy is not properly received by the receivers within
9 the migration aperture.
10 Peak spectral frequency
Destructive interference at karsted and faulted regions exhibits
11 lower spectral frequency.
12 Most positive principal curvature Many karst collapse caves are in bowl shape which produce
13 strong positive anomaly along the perimeter and negative
14 Most negative principal curvature anomaly at the center.
15
16 Most positive amplitude curvature Karst collapse as well as diagenetically altered joints and faults
17
generate lateral changes in amplitude.
18 Most negative amplitude curvature
19
Fo
20 Energy ratio similarity (coherence) Karsted and faulted regions are less coherent.
21
22 GLCM homogeneity Karsted and faulted regions are less smooth.
23
rP

24
25
26
ee

27
28
29
rR

30
31
32
ev

33
34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 GEOPHYSICS Page 60 of 61

1
2
3 Table 2
4
5
Number Number of processors Time elapsed (in seconds)
6 Test Adaptive
7 of SOM Building SOM Final Model Final
Case weighting?
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

models Total*
8 model(s) clustering building clustering
9 1 No 1 1 1 12 572 792
10 2 No 1 1 24 12 220 444
11 3 Yes 219 1 1 2981 564 3723
12
4 Yes 219 1 24 2792 224 3217
13
14 5 Yes 219 24 24 137 217 550
15 *: Total time includes model building, final clustering, and data preparation stages.
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31
32
ev

33
34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.
 Page 61 of 61 GEOPHYSICS

1
2
3 Table 3
4
5 Attribute Weight
6 Coherent energy 0.52
7
Downloaded 12/19/17 to 128.227.24.141. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

8
Dip magnitude 0.69
9 Energy ratio similarity 0.52
10 P impedance 1.66
11 S impedance 1.73
12 Lambda/Mu 0.99
13
14
Vp/Vs 0.94
15 Poisson’s Ratio 0.72
16
17
18
19
Fo
20
21
22
23
rP

24
25
26
ee

27
28
29
rR

30
31
32
ev

33
34
35
36
iew

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
This paper presented here as accepted for publication in Geophysics prior to copyediting and composition.
© 2018 Society of Exploration Geophysicists.