You are on page 1of 10

GEOPHYSICS, VOL. 69, NO. 4 (JULY-AUGUST 2004); P. 958967, 8 FIGS. 10.1190/1.

1778239

Texture model regression for effective feature discrimination: Application to seismic facies visualization and interpretation

Dengliang Gao

ABSTRACT

The classical approach to feature discrimination requires extraction and classication of multiple attributes. Such an approach is expensive in terms of computational time and storage space, and the results are generally difcult to interpret. With increasing data size and dimensionality, along with demand for high performance and productivity, the effectiveness of a feature-discrimination methodology has become a critically important issue in many areas of science. To address such an issue, I developed a texture model regression (TMR) methodology. Unlike classical attribute extraction and classication algorithms, the TMR methodology uses an interpreter-dened texture model as a calibrating lter and regresses the model texture with the data texture at each sample location to create a regression-gradient volume. The new approach not only dramatically reduces computational cycle time and space but also creates betters results than those obtained from classical techniques, resulting in improved feature discrimination, visualization, and interpretation. Application of the TMR concept to reection seismic data demonstrates its value in seismic-facies analysis. In

order to characterize reection seismic images composed of wiggle traces with variable amplitude, frequency, and phase, I introduced two simple seismic-texture models in this application. The rst model is dened by a full cycle of a cosine function whose amplitude and frequency are the maximum amplitude and dominant frequency of wiggle traces in the interval of interest. The second model is dened by a specic reection pattern known to be associated with a geologic feature of interest, such as gas sand in a hydrocarbon reservoir. I applied both models to a submarine turbidite system offshore West Africa and to a gas eld in the deep-water Gulf of Mexico, respectively. Based on extensive experimentation and comparative analysis, I found that the TMR process with such simple texture models creates superior results, using minimal computational resources. The result is geologically intriguing, easily interpretable, and consistent with general depositional and reservoir-facies concepts. Such a successful application may be attributable to the sensitivity of image texture to physical texture in the Fresnel zone at an acoustic interface and therefore to lithology, depositional facies, and hydrocarbon saturation.

INTRODUCTION

Feature discrimination and visualization are fundamental to exploratory data analysis in many areas of science. Conventionally, feature discrimination and visualization involve multiple processes, computationally intensive algorithms, and high-dimensional data sets to reduce nonuniqueness (e.g., Reut et al., 1985; Richards, 1993; Sheriff and Geldart, 1995; Ahern, 1999; Roweis and Saul, 2000; Seung and Lee, 2000; Bertrand, 2001; Freedman et al., 2001). Basically, a typical workow is composed of three dependent components. First, a suite of algorithms performs attribute extraction by calculating multiple quantities from an original data set that represent observations from different perspectives (e.g., Haralick et al., 1973;

Taner and Sheriff, 1977; Taner et al., 1979; Reed and Hussong, 1989; Turcotte, 1992; Gao et al., 1998; Gao, 2002, 2003). Second, dimensionality reduction is achieved by transforming the highdimensional, observational attributes into low-dimensional, condensed attributes (Jolliffe, 1986; Cox and Cox, 1994). Third, condensed attributes are reduced further into a nal thematic set that groups features into a limited number of categories in a process called multivariate classication (Ritter and Hepner, 1990; Richards, 1993; Gurney, 1997). The workow from attribute extraction through dimensionality reduction to multivariate classication requires signicant amounts of computational time, storage space, and manual intervention, making the approach impractical for large volumes of data, such as three-dimensional (3D) seismic data sets that

Manuscript received by the Editor December 11, 2002; revised manuscript received March 15, 2004. Marathon Oil Company, P.O. Box 3128, Houston, Texas 77056. E-mail: dgao@marathonoil.com. c 2004 Society of Exploration Geophysicists. All rights reserved. 958

Texture Model Regression for Facies Analysis

959

are typically on the order of 109 samples. Also, the result from such an approach is usually difcult to interpret and resistant to interactive quality control because the result is shown as a discrete number of classes on a classication map and relies on any number of input attributes whose physical implications are usually unknown. This deciency is particularly acute in frontier regions where little calibration data are available. Here, a fundamental issue is the effectiveness of an attribute or a combination of multiple attributes in discriminating features. Theoretically, each additional uncorrelated attribute adds a new dimension to the attribute space and, thus, contributes to better discrimination because features are more separable in a higher-dimensional attribute space than in a lower one. Practically, however, increasing data dimensionality requires more computation time and storage space and makes visualization/interpretation complicated. Therefore, many interpreters have difculty managing and interpreting a large number of seismic attributes, and some authors (e.g., Kalkomey, 1997) recommended avoiding the use of a large number of attributes to classify geologic features of interest to avoid false correlation. As the number of attributes and dimensionality of attribute space are reduced, the computational time and storage space decrease, and the interpretation and classication processes become easier. Ideally, if a single, powerful attribute could well-represent important features, the weight of that attribute relative to the others would be overwhelmingly high; thus, one such single attribute would contribute much more to a robust discrimination than a combination of many others. Therefore, there is a need to develop an algorithm that is computationally efcient, conceptually inclusive, functionally versatile, and physically meaningful. The texture model regression (TMR) methodology discussed in this paper satises these criteria. In the following sections, I introduce the TMR concept on a general basis to emphasize its broad applicability. Then, I apply the concept to reection seismology to demonstrate its value for seismic-facies analysis. I suggest that the successful application is attributable to a genetic link between image texture and physical texture in the Fresnel zone at an acoustic interface. Such a link is critical to better understand the geologic implications of the TMR methodology for seismic facies interpretation.
CONCEPTS AND METHODOLOGY

patterns such as gas cap and gas chimney, since the three word patterns are grouped into the same category. To make the process even more effective, assume that the word-search engine has an advanced functionality that calculates a continuum of pattern similarities relative to the model and color codes all word patterns based on their similarity values. Although gas sand is not identical to the model gas ????, all such word pairs should have the same similarity value (e.g., 0.55), permitting them to be distinguished and isolated from gas cap, gas chimney, and the rest of the text. Furthermore, I can still discriminate and isolate gas sand using a different word model as a lter, simply because it is the similarity difference, not the absolute value, that makes gas sand distinguishable from other word patterns. Similar to word-pattern recognition, amplitude-pattern recognition in a digital image domain is based primarily upon image texture that is dened by the magnitude, variation, and relationship of data samples at a given sample location in the image space. At each sample location, texture is characterized by using a small analysis window, which is usually called a texture element or texel by workers in this area to imply its functionality (e.g., Haralick et al., 1973; Reed and Hussong, 1989; Gao et al., 1998; Gao, 1999a, 1999b, 2001, 2002, 2003) (Figure 1). The work ow for the TMR is fundamentally a process that characterizes a texel at each sample location in the image space using a user-dened texel model as a regression lter (Figure 2). Unlike other digital images, a reection seismic image is composed of wiggle traces that appear to be similar to a cyclic trigonometric function with variable amplitude, phase, and frequency (Taner and Sheriff, 1977; Taner et al., 1979; Sheriff and

To make the TMR methodology simple and easy to comprehend, consider a familiar word-pattern search, for example gas sand in a word document. The use of just a single character g as a lter most likely results in a large number of words, most of which have no bearings on gas sand. Instead, if I use the eight-character gas sand as a lter, I can discriminate and isolate the word much more effectively because all characters and their lateral relationships are involved in the process to reduce nonuniqueness. To make this process more practical, consider a general case in which the word pattern to be searched is unknown or uncertain. In such a case, I can use the word model gas ???? as a lter for word patterns that begin with gas throughout the word document. As a result, all gas-related word patterns, including gas sand, are identied with enhanced efciency and productivity. However, a major limitation of this process is the difculty in distinguishing gas sand from other gas-related word

Figure 1. Hierarchy of visualization terms from a 2D pixel, through a 3D voxel, to a 3D texel. The texel is geometrically equivalent to a cubic window, but here it is so called to denote its functionality as a texture element and also for convenience of description in image texture analysis. In this example, the texel is composed of 405 (9 by 9 by 5) voxels (samples). At each of the sample locations in a 3D image space, a texel is characterized by analyzing and evaluating its internal amplitude conguration.

960

Gao

Geldart, 1995). Wiggle traces are aligned in x and y directions to depict geologic features, with limited resolution. Therefore, recognizing seismic texture is actually characterizing the set of wiggle traces. To characterize seismic wiggle traces effectively, I experimented with two prototype models. First, I tested a standard model (model case 1) dened by a full cycle of a trigonometric cosine function (Figure 3a), whose amplitude and frequency are the maximum amplitude and the dominant frequency, respectively, of the wiggle traces in the interval of interest. The dominant frequency is obtained using a Fourier transform, but in practice, it is estimated simply by measuring the average travel time between neighboring peaks or troughs of wiggle traces within the interval of interest. Because of its afnity and sensitivity to seismic wiggle traces, I thought such a cosine model might be effective in characterizing reection seismic facies in frontier sedimentary basins where little well-bore information is available. Second, I tested a texture model derived from seismic data (model case 2) that represents a specic depositional or reservoir facies based on well-bore information (Figure 3b). Unlike model case 1, the model case 2 subroutine rst investigates the volume on a running-window basis to nd an object pattern of interest. Typical model patterns include, for example, high amplitude and low frequency, with trough over peak; low amplitude and high homogeneity; high amplitude and high contrast; and hummocky or chaotic amplitude and high randomness, etc. It can also be a synthetic seismogram computed from sonic and density logs. Specically, if an interpreter is interested in a seismic facies that is related to oil and gas accumulation, the algorithm extracts from the reservoir interval a texture model that is considered most likely to be the hydrocarbon-related feature, thus, helping to extrapolate and predict depositional and reservoir facies in well-developed sedimentary basins where extensive well data are available. With the model texel determined, I retrieve a data texel at each sample location that is set to be the same size as the model and plot a scattergram between model and data texels. Then I perform a linear regression analysis on the scattergram. Typi-

cally, I use a least-squares technique to calculate gradient ( ), intercept ( ), and scattering ( ) of the regression line using the following equations:
n

Ti D (x , y , z ) T D (x , y , z ) (Ti M T M ) =
i =1 n

, (1) Ti
M 2 TM

i =1

= T D (x , y , z ) T M ,
and
n

(2)

= 1/ n
i =1

Ti D (x , y , z ) T D (x , y , z ) + Ti M T M

(3)

where n is the number of samples in the texel, and T D (x , y , z ) and T M denote the mean values of the data texel T D (x , y , z ) and the model texel T M , respectively. The regression gradient ( ), which ranges from 0.0 to 1.0, is a measure of similarity of the data texture relative to the model texture. The linear, least-squares technique has limitations that may affect the effectiveness of the TMR concept. First, for equation (1) to be mathematically meaningful, its denominator must not be zero; that is, the amplitude of the model texture must not be constant (plain texture). Second, there are certain cases where different features may not be distinguishable based solely on the value. In such cases, the values of and may provide additional constraints. The value is a measure of amplitude bulk shift that usually results from amplitude scaling, whereas the value is a measure of the deviation of amplitude from the regression line, a measure that generally indicates regression quality. Although a nonlinear regression subroutine could make the TMR process more effective, the linear, least-squares subroutine has the advantage of computational efciency and, in most cases, achieves the objective. Last, I map the values of , , and in the highest resolution (32 bit). Although the values of and may be used as additional constraints to reduce nonuniqueness of the value, I usually elect to calculate the value alone and visualize it on horizon slices to facilitate interactive facies interpretation.
GEOLOGIC IMPLICATIONS

Figure 2. A owchart showing TMR methodology.

A reection seismic response is a composite function of many physical properties at an acoustic interface in the subsurface. Typically, a reection seismic response has been attributed to acoustic/elastic impedance and Poissons ratio changes at an interface. However, a seismic signal is reected from a Fresnel zone rather than a single point at an interface. Physical texture in the Fresnel zone at an interface, dened by morphological features such as roughness or rugosity, may contribute signicantly to seismic amplitude and waveform as a result of constructive and destructive interference by signals reected from contributing micromirrors in the Fresnel zone. Since physical texture at an acoustic interface is critical to partitioning of backscattered, reected, and transmitted acoustic energy, its acoustic impact and geologic implications cannot be ignored. Physical texture in the Fresnel zone at an acoustic interface is a function of morphological characteristics and is suggestive of lithology, depositional environment, and uid saturation.

Texture Model Regression for Facies Analysis

961

For example, a carbonate-shale interface has a rougher texture than a sand-shale interface; marine shale in a highstand systems tract has a smoother texture than channel sand and mass transport complex in a lowstand systems tract; similarly, a maximum ooding surface may have a texture distinct from a sequence boundary; a uid contact may have a smoother texture than a regular lithofacies contact; and an interface with an abrupt lithologic change has a sharper texture than one with a gradational lithologic change. Therefore, seismic texture characterization provides an important basis for predicting and inferring depositional environment and facies, although not in as much detail as from direct observation of outcrops. This situation is analogous to side-scan sonar imagery, where backscattered amplitude intensity and texture have been well-recognized as reliable indicators of sea-oor sediments and roughness (e.g., Reut et. al., 1985; Reed and Hussong, 1989; Gao et. al., 1998). Figures 4 and 5 illustrate the application of the TMR methodology to seismic-facies analysis in a submarine turbidite system offshore Angola, West Africa. Since little well data are available in the study area, I elected to use model case 1 as the

regression lter in the TMR process. The dominant frequency was obtained by measuring the average time between neighboring troughs in the interval of the middle Miocene channelfan deposits. The average time is 60 ms, which is equivalent to a frequency of 16.7 Hz for the model to be used in the regression process. After initial processing, the result obtained using this simple model signicantly helped delineate the geometry, areal extent, and lateral relationships of critical depositional facies, such as the bypassed channel ll, levee/overbank, and lobes in the turbidite system. Experiments and comparisons with classical attribute-extraction algorithms showed this simple modelregression process to be not only computationally efcient but to produce superior results that are more inter pretable than those obtained using classical attribute-extraction algorithms. Figure 4 shows a channel-fan system imaged at a depth of 4000 ft (1220 m) below the sea oor. The well-organized distribution pattern and details shown in the TMR result (Figure 4b) are most intriguing and very consistent with a typical turbidite depositional pattern. The meandering, migrating, and cross-cutting channels are easily recognizable and interpretable. On both sides of the channels are laterally more extensive facies, possibly levee/overbank deposits. Distributed lobes in the frontal portion of the laterally extensive fan deposits were apparently fed by a braid of linear channels upslope from the northeast. Although this result is not derived from a clustering or neural network-based classication algorithm, the well-organized pattern is physically meaningful and can be easily categorized into geologically distinctive and spatially associated facies classes. For example, I recognize at least ve different depositional facies (denoted as 1, 2, 3, 4, and 5). Based on their distribution, geometry, and spatial relationships, coupled with the subregional depositional setting and limited exploration well-bore information, I interpret the progression from facies 1 to 5 to be associated with decreasing sand content from a proximal levee to a distal overbank or marineshale environment. The channel-ll facies is distinct from the levee facies and possibly represents a bypassed channel that is lled with shale. In addition to the ve major facies, I am also able to recognize and map more detailed facies changes simply by ne-tuning the color-mapping function in an interactive manner. All of these interpretations are consistent, in this particular area, with a regional facies architecture and, in general, with depositional facies patterns that are typically observed from contemporary and ancient deep-water depositional systems. In contrast, it is difcult to visualize, map, and interpret detailed facies changes and lateral relationships from Figure 3. Two examples of texture-model construction scenarios in the TMR process. (a) A standard model consisting of a full cycle of trigonometric cosine waves. (b) A the regular amplitude data (Figure 4a), even wiggle-trace model consisting of data traces of gas sand in a reservoir interval drilled though the existence of a major channel-fan depositional system is recognizable. by exploration and production well bores.

962

Gao

Figure 5 shows the same channel-fan system from four difthan 20 wells are available in this eld, I constructed a texferent perspectives. It is difcult to recognize and visualize ture model using wiggle traces that are known to be associated the same level of detail as that shown in Figure 5a from avwith gas sand in the reservoir interval (model case 2). The reerage absolute amplitude (Figure 5b), coherence (Figure 5c), sults obtained (Figures 7b and 8a) using such a model help or instantaneous frequency (Figure 5d). Average absolute amdifferentiate upslope gas sand from downslope wet sand in all plitude (Figure 5b) indicates extensively distributed high amthe major reservoirs, and the interpretation is consistent with plitude but fails to distinguish the channel-ll facies from drilling results from exploration and production operations levee/overbank and lobes, and the whole image has a much (e.g., Figure 7b and 8a). However, it is difcult to recognize lower resolution. Coherence (Figure 5c) is effective in enhancing lateral discontinuities and is particularly useful for delineating outlines of channels or fault traces on time slices; however, facies variations across boundaries are totally invisible to the algorithm. Instantaneous frequency (Figure 5d) is not effective at discriminating different facies or even at demonstrating the overall channel-fan geometric pattern. Figure 6 illustrates a complex channel system imaged from four different perspectives in the same region as in Figure 5 but at a different stratigraphic level. The TMR result shown in Figure 6a delineates a channel system that has different geometry and facies than that shown in Figure 5. The increased sinuosity and value along the channel at this stratigraphic interval suggest sand-prone deposits in the downslope portion of a turbidite system. This vertical change in facies and geometry of channel systems may have important implications for understanding the depositional history and sequence stratigraphy of the system. In contrast, average absolute amplitude (Figure 6b) has limited resolution, particularly in the eastern portion of the channel system. Although coherence (Figure 6c) effectively delineates channel boundaries, it contributes very little to resolving lateral facies variations within and outside the channel. Instantaneous frequency (Figure 6d) poorly denes not only facies variations associated with channels but even the meandering pattern of channels. Figures 7 and 8 show an example from a gas eld in the deep-water Gulf of Figure 4. (a) Original amplitude of a channel-fan system in deep-water offshore Mexico, illustrating implications of the Angola, West Africa. (b) The TMR result derived from the original amplitude data TMR methodology for reservoir-facies disusing a standard cosine model as the regression lter (see Figure 3a). The color code crimination and visualization. The reserrepresents values of the regression gradient ( ) given by equation (1), indicating a continuum of textural similarity to the model that ranges from 0.0 to 1.0, with 0.0 voir, consisting of lower Pliocene highindicating minimal similarity and 1.0 maximum similarity to the model. It is not the quality sand, was formed in a steeply absolute value but its variations and well-organized patterns that are geologically dipping monocline that laterally pinches meaningful and interpretable. Although it is not a discrete classication map, it out against a salt body to the north. To the can be easily interpreted into distinctive facies categories (e.g., 1, 2, 3, 4, and 5) east and west, the monoclinal reservoir is based on their distribution pattern and lateral relationships. In this specic area, the high value (red) suggests sand-rich facies such as in the levee/overbank bounded by two major subvertical faults. deposits or lobes in the distal portion of the turbidite system, whereas the low These stratigraphic and structural relationvalue (black) suggests shale-rich facies, such as the bypassed channel-ll deposits ships are favorable for migration and enand the sheetlike deposits distant from the channels. For unbiased comparison, trapment of hydrocarbons updip along the both the original amplitude and the TMR result are displayed using the same monocline (Figures 7 and 8). Since move color-mapping function.

Texture Model Regression for Facies Analysis

963

such differences based on average absolute amplitude (Figures 7c and 8b), coherence (Figures 7d and 8c), or the combination of instantaneous amplitude and frequency (Figures 7e and 8d). For example, in the updip portion of the reservoir, which is expected to have high amplitude in response to gas accumulation, average absolute amplitude (Figures 7c and 8b) shows no hint of increased amplitude. Coherence (Figures 7d and 8c) provides little discrimination between gas sand and wet sand, because the trace-to-trace similarity is not sensitive to waveform characteristics, such as trace shape, amplitude, frequency, and phase. The combination attribute of amplitude and frequency (Figures 7e and 8d), which has been very popular in hydrocarbon exploration, is still not as effective as the TMR result.
DISCUSSION

The TMR process enables scientists to discriminate features more efciently and interactively from a different perspective than the classical multivariate extraction and classication approach. Mathematically, a feature vector at each sample location is represented by one scatterpoint in a feature vector space. In such a space, a classication algorithm groups clouds of scatterpoints into a given number of clusters or classes by evaluating iteratively their closeness and distribution in the abstract, high-dimensional feature space. Since the cluster structure is initially invisible to the classication algorithm, interpreters have to guess at the number of classes on a trial-and-error

basis. As a result, the data set may be either underclassied or overclassied. In other words, different objects may have been grouped into the same class, or similar objects may have been split into two or more different classes. Once the classes are dened, it is impossible to split a class into different ones, and it is often difcult to merge multiple classes into a single one based on the class indexes. Interpreters need to classify the same data set repeatedly with different numbers of classes until a meaningful result is achieved, a process that is time consuming and can be impractically expensive in the case of large 3D regional seismic surveys. Furthermore, in reality, features such as geologic facies generally have transitional properties and vague boundaries that may not be easily separable in a physical space. It may not be worthwhile to seek mathematical boundaries among these patterns in an attribute space in a deterministic manner using a computationally expensive technique. The TMR process, however, uses model texture as a reference to evaluate texture similarity at each sample location. Therefore, features at all sample locations are calibrated relative to the texture model and are identied independently, based on their respective similarities to the texture model. The result is, effectively, a classication that can be interpreted and quality-checked in a more interactive manner than that derived from a classication algorithm, thereby minimizing cycle time and space and streamlining the interpretation process. In general, the TMR technique characterizes texture with a user-dened, object-oriented texture model. In seismic exploration, the model can be single (1D) or multiple (2D or

Figure 5. A turbidite channel-fan system in the same area as in Figure 4 imaged from four different perspectives, for comparison. (a) TMR. The color code and implications are the same as in Figure 4. (b) Average absolute amplitude. (c) Coherence. (d) Combination of amplitude and frequency. Arrows indicate features that are more easily recognizable from the TMR result than from the amplitude, coherence, and combination of amplitude and frequency. For unbiased comparison, attributes are computed using the same parameters and displayed using the same color-mapping function.

964

Gao

3D) wiggle traces and can be static or dynamic in nature. First, if the model is a 1D cosine with variable frequency, the TMR algorithm discriminates waveform features at different frequencies and, thus, degenerates into a volumetric spectraldecomposition algorithm, except that results are values instead of amplitude as in the regular spectral-decomposition technique (Partyka et al., 1999). Second, if the model is 1D and constant in nature, the TMR algorithm characterizes waveform shape and degenerates into a volumetric version of the waveform classication algorithm, except that TMR produces a continuum of values as opposed to the discrete number of classes as in the regular classication algorithm (Poupon et al., 1999). Third, if the model is data driven at each sample location, the TMR algorithm degenerates into a lateral, trace-correlation algorithm commonly used in seismic-horizon interpretation (e.g., Johnson et al., 2001) and coherence calculation (Bahorich and Farmer, 1995), in which case, the functionality changes from facies discrimination to structure characterization. Furthermore, since the TMR is sensitive to amplitude, frequency, and phase of wiggle traces, it also includes classical seismic attributes, such as amplitude and frequency, that have been commonly extracted using multiple, independent algorithms (Taner and Sheriff, 1977). In some sense, all these popular techniques may be considered to be end members, or specic cases, of the more general and exible TMR methodology. Even in these specic cases, the TMR technique has important advantages over the classical, popular algorithms. For example, the TMR approach avoids computationally iterative and intensive waveform classication. Instead of making classication maps with discrete waveform classes based on interpreted horizons, the TMR methodology creates a volume with a continuum of waveform classes ( values) that is in-

dependent of interpreted horizons and has the highest resolution. By simply applying horizon slicing or depth slicing at any stratigraphic interval throughout the volume, geologists are now able to make tens of facies maps at a continuum of stratigraphic levels in less than half an hour, thereby recognizing both lateral and vertical facies changes much more quickly and interactively than spending more than half an hour to make a single facies map using a conventional, horizon-based waveform classication algorithm. These advantages make the TMR a more plausible approach to seismic-facies analysis than traditional algorithms. The spectral-decomposition algorithm emphasizes amplitude variation with frequency by decomposing a wiggle trace into discrete frequency components, using a Fourier transform. It effectively identies amplitude anomalies at different frequencies, but it may not be as effective as the TMR algorithm using a variable-frequency model in characterizing wiggle-trace shapes that are critical to seismic facies discrimination. The ultimate goal of seismic-facies analysis is to tie seismic patterns to depositional facies. It is more an interpretive process driven primarily by a geologist than a computational process driven by a mathematician or a computer scientist. The latter plays an important role by creating useful quantitative information that is otherwise not easily obtainable, whereas the former plays a critical role by translating such intermediate quantitative information into geology. Thus, depositional-facies interpretation by a geologist, based on seismic texture and stratigraphic principles along with the regional depositional setting, may be geologically more reliable than a computer-generated facies-classication map based on a mathematically robust clustering and neural-network algorithm. This distinction is analogous to a structural contour map

Figure 6. A turbidite channel-levee system imaged from four different perspectives, for comparison. It is in the same area as in Figure 4 but in a different stratigraphic interval. (a) TMR. The color code and implications are the same as in Figure 4. (b) Average absolute amplitude. (c) Coherence. (d) Instantaneous frequency. Arrows indicate the features that are more easily recognizable from the TMR result than from the amplitude, coherence, and frequency. For unbiased comparison, attributes are calculated using the same parameters and displayed using the same color-mapping function.

Texture Model Regression for Facies Analysis

965 CONCLUSIONS

made by a geologist based on principles of structural geology and regional structural style, vis-a-vis ` a computer-generated contour map based on a mathematically robust interpolation and extrapolation algorithm. In order to make geologically robust facies interpretation and prediction from seismic data with little well information, it is important to transform regular amplitude into textural similarity to allow interactive-facies visualization and interpretation. The facies cube obtained from the TMR methodology represents one such superior data set that enhances the visibility and resolution of critical geologic features, thereby facilitating interactive observation, robust interpretation, and successful hydrocarbon exploration.

The TMR methodology provides a long-sought-after solution to integrating classical techniques into a unied process that improves data visualization and interpretation. Different from classical attribute extraction/classication approaches, the TMR technique uses an interpreter-dened texture model as a calibration lter to discriminate features. In seismic application, two simple texture models demonstrate the value of the TMR concept in seismic-facies analysis. The rst model, which is a full cycle of a cosine function dened by the maximum amplitude and the dominant frequency in the interval of interest, helps identify detailed facies variations in a turbidite system offshore Angola in West Africa. The second model, which is based on seismic data coupled with well information, helps differentiate gas sand from wet sand in a gas eld in the deep-water Gulf of Mexico. Comparative analysis indicates that the TMR methodology provides a unied and general approach to seismic-facies discrimination, which includes functionalities of popular but computationally expensive techniques. Not only does the TMR save computational time and storage space, but it creates a superior result that enables exploration geologists to identify, visualize, and map critical depositional and reservoir facies accurately, quickly, and interactively. The effectiveness of seismic texture analysis and its successful application to facies discrimination may be attributable to a direct link between image texture and physical texture in the Fresnel zone at an acoustic interface because physical texture is generally suggestive of depositional environment and hydrocarbon saturation. Such a link provides an important basis for inferring geologic facies from seismic data, although, because of the limited seismic resolution, not in as much detail as from direct observation of outcrops. Therefore, seismic texture analysis has direct and important implications for facies analysis and hydrocarbon exploration.
ACKNOWLEDGMENTS

This paper is part of the 3D exploratory data visualization and interpretation technologies (patent pending) developed

Figure 7. A seismic line showing the major reservoir in a gas eld in the deep-water Gulf of Mexico from 5 different perspectives, for comparison. (a) Original seismic line from which all the others are derived. (b) The TMR result that is created using a wiggle-trace model (model case 2) (see Figure 3b) as a regression lter that is considered to be gas sand, based on seismic data at a well bore penetrating the reservoir interval. The model has 27 samples (108 ms) in vertical dimension. The color code represents values of the regression gradient ( ) given by equation (1), indicating a continuum of textural similarity to the model (gas sand) that ranges from 0.0 to 1.0, with 0.0 indicating minimal similarity and 1.0 maximum similarity to gas sand. (c) Average absolute amplitude. (d) Coherence. (e) Combination of amplitude and frequency. The vertical dashed line marks approximately the gas/water contact. Magenta color denotes wells that penetrate gas sand, whereas light blue color represents wells that penetrate water sand. Notice that the result from the TMR method better denes the boundary between gas sand and wet sand than do the amplitude, coherence, and amplitude/frequency combination. For unbiased comparison, attributes are calculated using the same parameters and displayed using the same color-mapping function.

966

Gao

Figure 8. A reservoir system imaged on a horizon slice from four different perspectives, for comparison, in the same area as in Figure 7 (see the location of the horizon at top of the reservoir in Figure 7). (a) TMR. The color code and implications are the same as in Figure 7. (b) Average absolute amplitude. (c) Coherence. (d) Combination of amplitude and frequency. Notice that the result obtained from the TMR processing better denes the distribution of gas sand and the boundary between gas sand and wet sand than do the average absolute amplitude, coherence, and amplitude/ frequency combination. For unbiased comparison, attributes are calculated using the same parameters and displayed using the same color-mapping function.

by the author at Marathon Oil Corporation. Acknowledgment is made to Marathon management for permission to publish this work. Thanks go to Sharon Crawford and Tom Evans for their support of this study. I also thank Marathon geologists and geophysicists for their cooperation in applying the technology to worldwide exploration and production projects. I used the API (Application Program Interface) functions from Paradigm Geophysical, Inc. and Magic Earth, Inc. in programming the TMR technology. I thank ChevronTexaco, Petroleum Geo-Services, and WesternGeco for using their 3D digital seismic data. I thank Michael Schoenberger for his constructive suggestions. Journal reviews by Associate Editor Kurt Marfurt and two anonymous reviewers helped improve the quality of the paper.

REFERENCES Ahern, K., 1999, The eyes have it: Science, 284, 454. Bahorich, M., and Farmer, S., 1995, 3-D seismic discontinuity for faults and stratigraphic features: The coherence cube: The Leading Edge, 14, 10531058. Bertrand, C. A., 2001, Color by number: Imaging large data: Science, 293, 13351336.

Cox, T., and Cox, M., 1994, Multidimensional scaling: Chapman & Hall. Gao, D., 1999a, The rst-order and the second-order seismic textures: American Association of Petroleum Geologists Annual Meeting Program and Abstracts, 8, A45. Gao, D., 1999b, 3-D VCM seismic textures: A new technology to quantify seismic interpretation: 69th Annual International Meeting, SEG, Expanded Abstracts, 10371039. Gao, D., 2001, Method for analyzing and classifying three dimensional seismic information: US Patent 6,226,596. Gao, D., 2002, Seismic textures aid exploration: Offshore, 62, no. 9, 65. Gao, D., 2003, Volume texture extraction for 3D seismic visualization and interpretation: Geophysics, 68, 12941302. Gao, D., Hurst, S. D., Karson, J. A., Delaney, J. R., and Spiess, F. W., 1998, Computer-aided interpretation of side-looking sonar images from the eastern intersection of the Mid-Atlantic Ridge with the Kane Transform: Journal of Geophysical Research, 103, 20997 21014. Gurney, K., 1997, An introduction to neural network: UCL Press. Haralick, R. M., Shanmugam, K., and Dinstein, I., 1973, Textural features for image classication: Institute of Electrical and Electronics Engineers Transations on Systems, Man, and Cybernetics, SMC-3, 610621. Johnson, W. E., Louden, R. O., Lehman, D. D., and Edwards, D. L., 2001, Seismic has its many attributes: American Association of Petroleum Geologists Explorer, http: //www.aapg.org/explorer/ geophysical corner/2001/03gpc.html. Jolliffe, I. T., 1986, Principal component analysis: Springer-Verlag. Kalkomey, C. T., 1997, Potential risks when using seismic attributes as predictors of reservoir properties: The Leading Edge, 16, 247251.

Texture Model Regression for Facies Analysis Partyka, G. A., Gridley, J. A., and Lopez, J. A., 1999, Interpretational aspects of spectral decomposition in reservoir characterization: The Leading Edge, 18, 353360. Poupon, M., Azbel, K., and Palmer, G., 1999, A new methodology based on seismic facies analysis and litho-seismic modeling, the Elkhorn Slough eld pilot project, Solano County, California: 69th Annual International Meeting, SEG, Expanded Abstracts, 927930. Reed, T. B., and Hussong, D., 1989, Digital image processing techniques for enhancement and classication of SeaMARC II side-scan sonar imagery: Journal of Geophysical Research, 94, 74697490. Reut, Z., Pace, N. G., and Heaton, M. J. P., 1985, Computer classication of seabeds by sonar: Nature, 314, 426428. Richards, J. A., 1993, Remote sensing digital image analysis: SpringerVerlag. Ritter, N. D., and Hepner, G. F., 1990, Application of an articial neural network to land-cover classication of thematic mapper imagery:

967

Computers & Geosciences, 16, 873880. Roweis, S., and Saul, L., 2000, Nonlinear dimensionality reduction by locally linear embedding: Science, 290, 23232326. Seung, H. S., and Lee, D. D., 2000, The manifold ways of perception: Science, 290, 22682269. Sheriff, R. E., and Geldart, L. P., 1995, Exploration seismology: Cambridge University Press. Taner, M. T., and Sheriff, R. E., 1977, Application of amplitude, frequency, and other attributes to stratigraphic and hydrocarbon determination, in C. E. Payton, ed., Seismic StratigraphyApplications to Hydrocarbon Exploration: American Association of Petroleum Geologists Memoir 26, 301327. Taner, M. T., Koehler, F., and Sheriff, R. E., 1979, Complex seismic trace analysis: Geophysics, 44, 10411063. Turcotte, D. L., 1992, Fractals and chaos in geology and geophysics: Cambridge University Press.

You might also like