You are on page 1of 5

Data mining of 3D poststack seismic attribute volumes using Kohonen self-organizing maps

UWE STRECKER and RICHARD UDEN, Rock Solid Images, Houston, Texas, U.S.

in data reduction processes through classifications applied to a wide spectrum of aspectsfrom traffic solutions and medicinal purposes to geophysical interpretations. Here we use an unsupervised approach where the neural network is free to search, to recognize, and to classify structural patterns in an n-dimensional vector field spanning the entire 3D input seismic attribute data set (Taner et al., 2001; Walls et al., 2002). Within the data set, each data sample is defined by a unique combination of physical, geometric, and hybrid attributes and is treated as an n-dimensional vector (Carr et al., 2001). Data classification occurs when similar data are captured within a Euclidean distance of a neural node, thus providing data clusters or classes as an output data set. In this paper, an unsupervised artificial neural network using four different suites of poststack seismic attributes is employed to classify a 3D seismic data volume from Lafourche Parish, South Louisiana. Figure 1 identifies Cretaceous through Holocene paleoshelf distribution of major Cenozoic depocenters. The star indicates the study area which is associated with play trends of Miocene age. Figure 2 outlines a 3D seismic survey near Thibodaux, Lafourche Parish, South Louisiana, that was acquired and processed in 2001. This survey encompasses approximately 72 mi2 with a bin spacing of 33.5 m. Densely spaced lease acreage positions crossing the survey are those of the Atchafalaya river levee system. Kohonen self-organizing map (KSOM) clustering and topological organization. The main reason to use artificial neural networks is to distinguish populations of similar multiattribute response. Artificial neural networks add value when classifying raw input data would have proved too complex for conventional statistical approaches. Artificial neural networks draw on the mammal brains capability of adaptive learning. In particular, each neuron has as many input connections as there are attributes and within a given data set each data sample is defined by a specific combination of physical and geometric attributes. Thus, each sample is treated as an n-dimensional vector. Kohonen self-organizing maps consist of a single layer of neural nodes. Each neuron represents the center of a data cloud of similar attribute combinations. Thus, each neural node defines a class (Taner et al., 2001). Each sample in a 3D seis-

For several decades, artificial neural networks have assisted

Figure 1. Cenozoic depocenters (adapted and modified from Martin, 1978).

Figure 2. Thibodaux 3D seismic survey outline.

mic data volume has an unsupervised class assigned to it, therefore creating a volume of discrete multiattribute classes. Table 1 summarizes individual poststack seismic attribute combinations used in individual KSOM runs. Runs 1-4 represent different combinations of input attribute vol-

Table 1. Poststack seismic attribute combinations employed in different Kohonen self-organizing map runs

1032

THE LEADING EDGE

OCTOBER 2002

a)

c)

b)

d)

Figure 3. Four Kohonen self-organizing map runs emphasizing different poststack attributes (see Table 1).

umes. Input parameters include a suite of physical (trace envelope, bandwidth, wavelet envelope, wavelet bandwidth, AVO slope, impedance), geometric (similarity, dip of maximum similarity), and composite (shale indicator) poststack seismic attributes. Inspection of the input parameter matrix illustrates that geometric attributes are favored in run 3. Figure 3 depicts four KSOM runs emphasizing different suites of poststack seismic attributes. Emphasis in runs 1-2 slightly favors physical seismic attributes, whereas runs 34 emphasize geometric attributes. Emphasis on geometric attributes favors clustering of structural and stratigraphic morphologies, whereas dominance of physical attributes tends to underscore importance of lithological differentiation (Taner et al., 2001). Figure 4 displays an uninterpreted detail of KSOM run 3. The vertical front panel depicts numerous convex downward shapes in the Kohonen self-organizing map volume. These convex downward morphologies closely resemble meandering channel forms embedded in floodplain deposits. Indeed, by stripping away overburden a cut probe display verifies the presence of a meandering channel form in the subsurface (Figure 5). Figures 6a and 6b provide an oblique view of a meandering channel system. Classes comprising channel fill appear dominantly in shades of yellow and orange. Note
1034 THE LEADING EDGE OCTOBER 2002

Figure 4. Uninterpreted 3D detail of KSOM run 3. Convex downward shapes in Kohonen self-organizing map data are interpreted as channels meandering on a floodplain.

how the cut-bank margin of the channel is well defined in amplitude data but exhibits a gradational contact on the point bar side (Figure 6a). In contrast, the margin at the point bar side of the channel is much better defined in Kohonen self-organizing map data (Figure 6b). Subsequent combination of individual Kohonen classes encountered within the channel facilitates distillation of a contiguous meandering channel form from the KSOM run 3 data set (Figure 7).

Figure 5. Cut probe verifying meander channel form in the subsurface.

Figure 7. Combined selection of contiguous Kohonen classes allows distillation of meandering channel form from 3D seismic data set.

a) a)

b)

b)

Figure 6. (a) Kohonen self-organizing map data volume with oblique view of meandering channel system atop time slice. Note well-defined margins on both sides of channel. (b) Kohonen self-organizing time slice with superposed amplitude time slice data.

Interactive tracking throughout the KSOM run 3 data volume allows the interpreter to delineate several meander channel systems in the subsurface using connectivity analysis (Figure 8a). Comparison of Figures 8a and Figure 8b demonstrates how an entire meander belt system, including an apparent oxbow meander cut-off, can be delineated in the subsurface. It is noteworthy that, for small vertical distances, classification remains constant, allowing seed picking of inclined channel forms through the data volume (i.e., tracking dipping geobodies is possible). A surface representing the channel base can be created by draping the lower portion (the hull) of the channel geobody (Figure 9). The resultant channel form is color-coded by depth with warm colors identifying the channel edges and cold colors marking the channel bottom. In maximum dimensions the channel is generally less than 400 m wide, 4 km long, and approximately 40 m deep. These dimensions

Figure 8. Multiple meandering channel systems (a) identified by seedpicked Kohonen classes representing channel forms. (b) Note inclination of channel feature on left-hand side.

are compatible with a paleovalley or a small-scale submarine canyon. However, without geologic calibration, no physical meaning can be assigned to the data classification. Nevertheless, internal channel stacking patterns appear
OCTOBER 2002 THE LEADING EDGE 1035

Figure 9. Enveloping channel base morphology distilled from Kohonen self-organizing map volumes.

Figure 12. Classes 1 (violet) and 2 (green).

Figure 10. Cross-section across channel form. Warm colors identify Kohonen classes interpreted as laterally accreting channel-fill. Class 1 (yellow), class 2 (ocher).

Figure 13. Possible horizontal well path exploiting point bar geometry of class 3.

Thus, in spite of the absence of a geological calibration that can be performed by calibrating Kohonen classes to lithologies from a well bore, data are nevertheless treated as if resulting from an incised paleovalley. Within the confines of the channel several Kohonen classes become vertically stacked. At the same time, individual Kohonen classes are intercalated along strike of the channel feature and occupy different segments of the channel feature. Tracking of the channel form in the subsurface was made possible by grouping various seed-picked, individual Kohonen classes. Figures 10-14 feature the different classes comprising channel fill. Vertical and lateral stacking patterns of classes 2 and 3 strongly resemble lateral accretion patterns. Figure 13 conceptually illustrates how channel fill resembling a point bar could potentially be exploited via placement of a horizontal well. Class 4 strongly resembles class 3 but it accumulated primarily adjacent to a neighboring meander bend. Discussion and conclusions. Data classification represents one principal use of Kohonen self-organizing maps. Application of neural network runs distills additional stratigraphic detail not ordinarily resolved in conventional amplitude data. In particular, individual, discrete Kohonen classes combine to form meandering channel forms in subsurface data from Lafourche

Figure 11. Class 1 (base of channel; yellow base in Figure 10).

slightly asymmetric and conform to a lateral accretion mode of sedimentation, as is frequently found in meandering river systems (Figure 10).
1036 THE LEADING EDGE OCTOBER 2002

within this feature into different classes of channel fill. This division into its constituting classes permits the analysis of potential reservoir from seismic attribute data. Thus, in conclusion, classified, yet uncalibrated, Kohonen self-organizing maps provide an opportunity for geologic interpretations of 3D seismic data volumes when additional stratigraphic detail is largely masked in conventional amplitude stack data. Indeed, interactive grouping and tracking of Kohonen classes allows identification of sinuous channel belts in the subsurface. Additionally, inspection of areal and lateral extent of individual Kohonen classes allows distinction of reservoir-scale physical features that underscore importance of reservoir heterogeneities and subsequent optimization of reservoir exploitation via, for instance, horizontal well bores.
Suggested reading. Well log calibration of Kohonen-classified seismic attributes using Bayesian logic by Taner et al. (SEG Expanded Abstracts, 2000). Seismic reservoir characterization of a U.S. midcontinent fluvial system using rock physics, poststack seismic attributes, and neural networks by Walls et al. (TLE, 2002). The integration of surface and seismic and log data to generate a rock and fluid properties volumea South Texas example by Carr et al. (43rd Annual PBGS Exploration Meeting, 2001). Northern and Eastern Gulf of Mexico continental margin: Stratigraphic and structural framework by Martin (in Framework, Facies, and Oil Trapping Characteristics of the Upper Continental Margin, AAPG Studies in Geology, No. 7, 1978). TLE
Acknowledgments: Greg Whitmire and Jason Tinder for helping draft some figures, Seitel for providing the original 3D seismic data. Corresponding author: u.strecker@rocksolidimages.com; r.uden@rocksolidimages.com

Figure 14. Channel class 4.

Parish, Louisiana. Individual classes accumulated along different segments of this subsurface feature but can also be differentiated vertically. Encountered sinuous geometry and lateral accretion patterns are compatible with a meandering river system/paleovalley, but a geologic calibration of the data is ultimately required. Nevertheless, application of a neural network to seismic attribute data succeeds in delineating a stratigraphically meaningful subsurface pattern and proceeds to break down subtle stratigraphic heterogeneities

OCTOBER 2002

THE LEADING EDGE

1037