Sei sulla pagina 1di 8

Interactive seismic facies classification using textural attributes and neural networks

BRIAN P. WEST, STEVE R. MAY, JOHN E. EASTWOOD, and CHRISTINE ROSSEN, ExxonMobil Upstream Research Company, Houston, Texas, U.S.
n this study, we present an application of textural analysis to 3D seismic volumes. Specifically, we combine image textural analysis with a neural network classification to quantitatively map seismic facies in three-dimensional data. Key advantages of this approach are: 1) it produces a detailed 3D facies classification volume (whereas manual seismic facies classifications are typically 2D maps), 2) it enables rapid and quantitative analysis of the increasingly large seismic volumes available to the interpreter, and 3) it eliminates many time-consuming tasks, thereby freeing the interpreter to focus on determining seismic facies and integrating them into a geologic framework. Finally, we extend our textural analysis-based seismic facies classification technique to interpretation of AVO attribute volumes, such as A + B (AVO intercept + gradient), to reduce the inherent nonuniqueness of seismic facies to geologic and lithologic facies, and simplify the facies analysis of complex, mixed-impedance reservoirs. Seismic facies analysis. Seismic facies analysis is a powerful qualitative technique used in stratigraphic analysis from seismic data and in hydrocarbon exploration. Seismic facies are groups of seismic reflections whose parameters (such as amplitude, continuity, reflection geometry, and frequency) differ from those of adjacent groups. Seismic facies analysis involves two key steps(1) seismic facies classification (i.e., seismic facies are defined, and lateral/vertical extents delineated) and (2) interpretation (i.e., analysis of vertical/lateral associations, map patterns, and calibration to wells) to produce a geologic and depositional interpretation. This interpretation step is required because there is a nonunique relationship between seismic data, seismic facies, and depositional environment or rock property relationships (Figure 1). In the past, the seismic facies mapping or classification step has occurred through time-consuming, manual methods. Seismic facies are conventionally delineated in the context of mapped horizons (i.e., the interpreter analyzes seismic facies that occur between mapped horizons). This is done by examining successive vertical sections through the seismic volume to determine the dominant seismic facies that occurs between the mapped horizons, and posting this information on a map. The output of this step is therefore a 2D map that generalizes the distribution of seismic facies vertically within a mapped interval. In large and complex areas, it may be difficult to map different seismic facies consistently. Manual seismic facies mapping, although time-consuming and qualitative, has proven extremely useful for hydrocarbon exploration and reservoir characterization, even when seismic facies cannot be uniquely related to physical properties. A skilled interpreters knowledge and experience contribute greatly to the success of seismic facies analysis. However, with increasingly large 3D seismic volumes, a more efficient and quantitative, 3D or volume-based approach is required but one which still incorporates inter1042 THE LEADING EDGE OCTOBER 2002

Figure 1. Examples of seismic facies and potential associated geologic fill. A seismic facies can be defined as a stratigraphic region in the seismic data volume that has a characteristic reflection pattern distinguishable from those of other areas on the basis of reflection amplitude, continuity, geometry, and/or internal configuration of reflectors. Inherent in a seismic facies analysis, however, is the nonunique relationship between seismic data, seismic facies, environment of deposition (EOD), and rock property relationships. Here, typical deepwater seismic facies have been interpreted to represent the stacking patterns illustrated beneath each seismic example.

Figure 2. GLCMs for identical subregions with differing azimuths results in very different GLCMs. When constructing GLCMs, comparison orientation is a significant variable that must be considered.

preter skill and experience. We believe that volume-based seismic facies mapping using textural image analysis and an interpreter-trained neural network is an important advance toward that goal.

Figure 3. GLCMs calculated from the extracted seismic data subregions for the four different seismic facies of Figure 1. The left GLCM for each seismic facies class is calculated along the direction of dip at a user-defined distance. The right GLCM is calculated in the direction of dip at twice that distance to effectively characterize reflection continuity. As illustrated in this figure, a number of trends emerge. First, continuous regions exhibit a tighter distribution along the upper left, lower right diagonal than do semicontinuous regions (the high amplitude continuous versus high amplitude semicontinuous examples). Second, higher amplitude regions have a greater extent along this same diagonal (the moderate amplitude versus high amplitude semicontinuous examples). These characteristics are quantified through textural attributes and used for the seismic facies classification. Finally, within a specific seismic facies, the GLCM resulting from greater comparison distance always exhibits a less continuous character (more dispersed along the lower left, upper right diagonal). This characteristic illustrates the similarity of a GLCM calculation to the production of seismically defined variograms.

ple, 8-bit seismic data have 256 gray levels, and a GLCM constructed from 8-bit data will have 256 rows and 256 columns. In many instances even less dynamic range is needed for efficient GLCM calculation. For example, in many cases 5-bit data (32 32 GLCM) are sufficient for seismic facies analysis. GLCMs are constructed by comparing pixel relationships within a subregion for a specific distance (e.g., 4 pixels apart), in a specific direction (e.g., subparallel to a datumned time surface). Each element of the GLCM represents the relative frequency of occurrence of two pixels, the co-occurance, within the subregion. If pixel A, for example, has value i and is at defined distance D, strike , and dip , from pixel B with value j, then the GLCM matrix location i,j (row, column) will be incremented by one. If another example within the subregion has the same relationship, then that GLCM element is incremented again. This process of tallying relative occurrence is performed for each existing pixel pair within the subregion to produce the final GLCM for that area. For a sufficient description of the data region, several GLCMs, with varying orientations and look distances, are constructed. Consider a two-dimensional image, taken from a synthetic checkerboard data volume (Figure 2). Constructing a GLCM, with a distance of two pixels and a diagonal comparison direction, begins with the comparison between pixel (A, 1, 1 - white) and pixel (B, 2, 2 - white). Pixel A has a value of 1 and pixel B also has a value of 1; thus the matrix element in the first row and first column will be incremented by one. This process continues until all possible transitions at this distance and azimuth are considered. Next, consider a different possible orientation, a distance of three pixels and a horizontal azimuth. In this case, pixel A (1, 1) again has a value of 1 while pixel B (1, 4) has a value of 8. With this transition, the last row and first column of the GLCM will be incremented by one and the process will again continue until all possible pixel transitions for that azimuth and distance have been recorded into the GLCM. Seismic-based GLCMs. The structure of seismic-based GLCMs is quite simple to understand. Homogeneous regions, or instances where the GLCM is calculated in direction of maximum continuity (e.g., parallel to continuous high signal-tonoise reflections) exhibit a tight distribution along the diagonal (top left to bottom right) of the GLCM. This results from comparison of pixels with similar values, effectively a cross-plot of correlated values. Less homogeneous or discontinuous regions will have more occurrences (counts) farther away from the GLCM diagonal, resulting from disparate pixel comparisons. Pixel value magnitude is also captured in the GLCM. Regions of low amplitude have GLCMs with values clustered near the center (effectively the zero crossing). Regions that include higher amplitudes will have more broadly distributed values within the GLCM either along the diagonal for continuous textures, or throughout the GLCM in more discontinuous textures (Figure 3). Careful inspection of the GLCMs in Figure 3 demonstrates that despite the variability in the character, there is always symmetry about the upper left-lower right diagonal. In the current study, GLCMs are symmetric matrices about the upper left-lower right diagonal due to transitions from pixel a to pixel b being considered equally important as a transition from b to a. Further research is exploring the possibility of exploiting anisotropic GLCMs in the context of depositional grains to extract geometries with geologic significance. Geometries such as onlap, mounding from differential compaction, and shingles representing progradation are of particular interest.

Textural analysis and grey-level co-occurrence matrices (GLCMs). Textural analysis can quantitatively describe many aspects of the classic seismic facies analysis performed by the interpreter. Stratigraphically-steered, seismic texture is a quantitative, multitrace attribute suite, that mimics the visual process of the interpreter more effectively than traditional, single trace-based attribute analyses. For textural analysis, the interpreter examines an ensemble of traces as an image to render a classification rather than examining only one or two adjacent traces at a time. The goal of textural analysis is to mathematically describe the distribution of pixel values within a subregion of data, effectively quantifying the spatial organization of seismic reflections. Just as the physical texture of a object can be loosely described as relatively smooth or rough, the visual texture of a region can be intuitively thought of as smooth or rough, continuous or discontinuous. The technique commonly used to quantify data texture employs a transformation that results in grey-level co-occurrence matrices or GLCMs. GLCMs, also referred to as transition probability matrices or tally matrices, quantitatively describe the spatial relationships and relative occurrence of pixel values within a defined region. GLCMs have dimensions N N; N is the number of gray levels (dynamic range) used to quantify the data. For exam-

OCTOBER 2002

THE LEADING EDGE

1043

Dip-steering the analysis. Seismic facies must be considered within the structural/stratigraphic framework of a data set and the desired scale for interpretation. Textural analysis applied to seismic data must follow the stratigraphic dip of the reflections to produce satisfactory results. Texture analysis and construction of a subregions GLCM is extremely sensitive to the look-direction or azimuth in which the pixels within the region are related. Following the stratigraphic dip in a GLCM calculation, dynamically changing depending on the stratigraphy, maximizes the continuity of the image as expressed in the GLCM. The process of guiding a calculation by stratigraphic dip is called dip-steering. Dip-steering of a GLCM calculation requires quantitative knowledge of the orientation of seismic reflections at all points within an image. Many methods for dip-steering calculations are available. Coherency or discontinuity based dip-steering methods follow stratrigraphic frameworks through calculated trace to trace shifts in the lag distance. Others are based on an initial skeletonization of the reflections within a data volume. To further exploit the image-based nature of textural analysis techniques, we adopt a gradient-based dip-steering algorithm. The multitrace nature of the technique is thus exploited, and dips within an image are estimated with high-accuracy. The first step in this process requires calculation of the horizontal (dx) and vertical (dy) gradient of pixel values within the image. The local apparent dip of the reflectors is then calculated via: With dy/dx having units of time per CDP; however, for convenience, these units can be ignored and the dip can be expressed in terms of pseudo-radians or pseudo-degrees relative to a horizontal time slice. For fully three-dimensional analyses, two orthogonal apparent dip calculations are performed and the results used to estimate the true dip and azimuth for dip and directional steering of the GLCM calculation. Once reflection dip and azimuth are known everywhere, this information guides the dip and azimuth of the look direction for the GLCM calculation. Due to the discrete nature of a seismic image (CDPs and time samples), a continuum of angular values is not feasible. Because the calculation is performed over multitraces, the potential angular resolution is inversely proportional to the distance over which the comparison is made, and the size of the analysis subvolume. For example, a dip-accuracy of 15 pseudodegrees, can be approximated if the width is equal to three pixels. In our application, the interpreter controls the look distance, the size of the analysis subvolume, and the length scale for dip-steering. The size of the subvolume depends on the scale of the features the user wishes to analyze, and the comparison distances depend on the length over which continuity must be judged. In reservoir characterization, a typical subvolume might be 325 m on a side and approximately 50 m vertically. Within this subvolume, we generally calculate three GLCMs. Two GLCMs will be calculated in the direction of maximum seismic continuity (dip-steered) at distances of approximately 40 and 80 m. The third GLCM is calculated orthogonal to the seismically defined dip and can be used as a metric for frequency content and shape of the seismic data within the analysis window. Statistical stability considerations also play a factor in deciding the comparison distances within the subvolume. Generally, comparison distances average 20% of the overall subvolume dimension to ensure enough
1044 THE LEADING EDGE OCTOBER 2002

Figure 4. Generalized diagram of the workflow for interactive seismic facies classification with iterative quality control.

valid comparisons to produce a representative GLCM. Textural attributes. GLCMs are not efficiently interpreted directly; instead summary statistics calculated from the GLCM are exploited. Textural attributes are divided into first- and second-order descriptors. First-order statistics quantify the global distribution of pixel values within an image and can be calculated directly using standard statistical techniques without an intermediate GLCM transformation. Within the region of interest, average absolute amplitude values and standard deviation of amplitude values can be considered firstorder textural attributes and useful in delineating amplitude anomalies and reflection strength. Derived attributes such as instantaneous amplitude, phase, and frequency can also be used to produce first-order statistics. Second-order statistics of an image quantify the spatial relationships of pixels within the image and are calculated via the intermediate transform to the GLCM. Second-order GLCM statistics can capture trace shape characteristics, reflection geometry, and reflection continuity, in addition to amplitude strength. Second-order statistics of a subregion are a multitrace attribute, which allows reflection geometry and continuity to be captured through analysis of the dip-steered GLCM. The utility of various textural attributes for seismic facies analysis continues to be a topic of research. A general textural description can be achieved with four commonly used textural attributes: homogeneity, inertia, entropy, and energy. The mathematical expression of these GLCM attributes is:

where cij is the ith, and jth component of GLCM, c, and n is the size of the matrix (squared number of gray levels within the image). The input GLCM, c is normalized such that: The GLCM characteristics that these mathematical relationships quantify are simple to understand. Textural homogeneity measures the similarity of pixels. Homogeneity will be high for GLCMs with elements concentrated near the diagonal. Conversely, low homogeneity values result from highly contrasting pixel values in the comparison orientation. These characteristics make textural homogeneity particularly useful for quantifying reflection continuity. Textural inertia is also indicative of the contrast of the

Figure 5. (upper) A typical training session showing the definition of training polygons on a seismic section. The training regions are provided directly by the interpreter. (lower) Same seismic section showing local dip of the seismic. Yellow represents dips down to the left; grays represent dips down to the right.

Figure 7. During the training phase, summary characteristics of the digitized training polygons (denoted by numbers on the chart) are calculated from the full suite of textural attributes and are plotted to assist the interpreter in training the probabilistic NN. Figure 6. Examples of two seismic facies training polygons from Figure 5, representing high-amplitude continuous (HAC) seismic facies (left) and moderate amplitude semicontinuous (MASC) seismic facies (right). Using the digitized examples, gray-level co-occurrence matrices (GLCMs, lower panels) are extracted and, from these GLCMs, textural attributes are calculated. Note that the data in these figures were rescaled to 5 bit precision (32 gray levels), thereby reducing the overall size of the GLCM and expediting subsequent calculations. In this figure, three GLCMs are produced from the polygon area, two in the direction of dip at differing distances (upper subimages), and one orthogonal to the dip direction (lower subimage).

ment or set of elements, corresponding to a particular pixel to pixel transition, is favored. Textural energy (sometimes called uniformity) is also indicative of the spatial organization. Energy is lowest when all elements of the GLCM are equal, nearly opposite of the textural entropy and very useful for highlighting regions of reflector continuity and geometry. Although energy and entropy are metrics of similar characteristics, our experience has shown that their combined use leads to more satisfactory results. Classification of textural attributes: probabilistic neural networks. Attribute classification is the mechanism by which a continuum of quantitative seismic characteristics is related to a discrete number of classifications within the volume, effectively a dimension reduction exercise to simplify the interpretation of the attributes. Neural networks are often used for statistical analysis and data classification. A neural network is an interconnected assembly of simple processing elements. The processing ability of the network is stored in the connection strengths, or weights, obtained by a process of adaptation to, or learning from a set of examples. One advantage of neural networks is the ability to train or modOCTOBER 2002 THE LEADING EDGE 1045

GLCM. Whereas homogeneity will be low for a highly contrasted data, textural inertia will be high. Inertia is a measure of the distance-weighted density of points away from the GLCM diagonal, similar to the physical property inertia. Although textural inertia and homogeneity are related, experiments with dip-steered calculations demonstrate that together they provide more seismic facies discrimination than either alone. Textural entropy measures the organization of pixels. Entropy is large when the values of the GLCM are uniform, corresponding to a scenario when all transitions are equally probable. Textural entropy is low when a particular ele-

Figure 8. A typical seismic facies classification using the interpreter trained probabilistic neural network, where multiple seismic facies classes have been identified. The seismic classification scheme on the right consists of high amplitude (HA), moderate amplitude (MA), low amplitude (LA), continuous (C) and semicontinuous (SC) seismic facies.

Figure 9. A relative confidence section resulting from the extraction of probabilities from the neural network output. Low confidence values can be observed in and near fault zones and in areas of facies transition.

analysis, a seismic facies classification can occur in principle with as little as one good example per facies class with the textural attributes of the training images supplying the weight vectors in the first layer of the network. When an input pattern is presented to the probabilistic neural network, the first layer computes distances from the input vector to the training input vectors, and produces a vector whose elements indicate how close the input is to a training input. The second layer sums these contributions for each class of inputs to produce, as its net output, a vector of probabilities. A second advantage of the PNN in pattern recognition is the ability to extract classification probabilities directly from the second, hidden layer, in addition to the classification of the maximum probability from the output layer. The resultant data quantifies the confidence of the classification output based on the input training set. Finally, a competitive transfer function (a function which simply selects the maximum value) on the output of the second layer picks the maximum of these probabilities, and produces a 1 for the chosen class and a 0 for the other classes. Interactive training and iterative QC. Figure 4 shows the workflow for seismic facies classification using textural analysis and neural networks. Two extremely important aspects of the technique are: (1) interactive training of multiple seismic facies classes simultaneously and (2) the capability to do iterative training and quality control checking of results between the seismic interpreter and the neural network classification prior to full data analysis. This iterative approach for multiple, user-defined classes enhances the ability of a probabilistic neural network to reproduce the geologically significant classifications. Training the neural network begins with the definition of textural analysis parameters, such as calculation volume size, dip-steering window size, bit depth of the data, and analysis distance. Specification of these parameters is based on the experience of the interpreter, and actual parameter values are dependent on the frequency and spacing of the data, as well as the scale of the analysis. To train the neural network, the interpreter selects one or several key seismic lines, and digitizes polygons of a consistent scale that are examples of different seismic facies (Figure 5, upper). The computer automatically calculates on-the-fly the data and dip/azimuth information associated with the training polygon (Figure 5, lower) and performs a textural analysis on these subregions (Figure 6). The interpreter inspects the resulting GLCMs and can either reject the training polygon or accept it and classify it as a specific seismic facies. The

Figure 10. Result of the textural analysis using neural networks is a seismic classification volume (upper left). This volume can be examined in vertical section or horizon slice mode, and compared to other seismic attribute volumes, in addition to available well or core data, in order to develop a geologic interpretation of the seismic facies classification. In this example from a channelized deepwater reservoir, a slice from the seismic facies classification volume (middle image) is compared to an equivalent slice from the seismic discontinuity volume (lower right). Calibration of these features to available well data results in the environment of deposition interpretation map shown in Figure 11.

ify the connection strengths within the network to produce desired results. Computationally, the connectivity of the nodes within a general neural network (the weights), modify an input vector of attributes, and pass the modified values on to the next layer of the network. Once sufficiently trained on a number of calibration images, the neural network can then be applied to the remaining images in a data volume. Probabilistic neural networks (PNNs) can efficiently perform pattern classification. Mathematically, these neural networks are similar to kriging, where proximity to known points (attribute values in the training set), guide the classification and prediction of unknown points. Probabilistic neural networks do not require extensive training. In our
1046 THE LEADING EDGE OCTOBER 2002

Figure 11. Environment of deposition and net:gross distribution map for slices shown in Figure 10. This discontinuity slice highlights the lateral edges of the channel (red lines), a broad, older sinuous element (1) and a narrower, younger sinuous element (2). Comparison with the seismic facies slice shows that sinuous element (1) is composed of HAC to MASC seismic facies, whereas sinuous element 2 is primarily composed of HAC seismic facies.Because the textural analysis seismic facies classification is a volume, this type of analysis can be applied at numerous stratigraphic levels within an interval of interest, whether or not these intervals are bounded by mapped horizons. Combining these results with the conceptual relationships illustrated in Figure 1, the net result of the analysis is effectively an "environment of deposition volume" where 3D regions can delineate differing depositional and geologic properties.

Figure 12. Example of seismic data for mixed impedance sands for near, far, full, and A+B volumes.

interpreter should specify several examples of each seismic facies class, and these examples should be distributed throughout the 3D seismic area of interest. In the initial training phase, plots of summary attributes, such as amplitude and reflection continuity, assist the interpreter in providing consistent training examples to the neural network (Figure 7). Once initial training is complete, a probabilistic neural network is constructed from the textural attributes and their associated classifications. This initial neural network is then used to classify a portion of the data set (typically several key seismic lines) to facilitate interactive quality control and analysis. Determining whether the PNN has been trained adequately occurs in the quality control phase of the workflow. In this step, the interpreter asks is the computer doing what I would have otherwise done manually and is it geologically reasonable? The interpreter can judge the performance of the result supplied by the neural network simply by examining representative classification and confidence profiles (Figures 8 and 9) prior to classification of larger data sets. If results are unsatisfactory, the training set can be

modified through deletion of existing polygons and/or addition of new polygons. The neural network is then recreated with the modified training set, and again checked, until the interpreter is satisfied with the results. After training and QC (i.e., the interpreter is satisfied that the algorithms are producing the facies classification which would have been arrived at through manual interpretation), the network and textural analysis algorithms can be applied to the rest of the data volume. The result is a seismic facies classification volume which has a seismic facies classification for every trace and sample within the volume (Figure 10) and an associated relative confidence volume. Key advantages of the neural net classification volume over a manual classification are: (1) the neural network classification is volume-based (rather than map-based for manual classifications), (2) the neural network classification is typically much more detailed than any manual classification and can be produced in a fraction of the time, (3) the classification is based on quantitative criteria and is reproducible given a particular training set. These characteristics promote improved efficiency in the analysis of large data volumes,
OCTOBER 2002 THE LEADING EDGE 1047

and improved capture of lateral and vertical heterogeneities in the seismic volume that are due to stratigraphic or other factors. Although seismic facies classification using textural analysis and neural networks techniques greatly improves the efficiency of seismic classification, the output classification volume is not the final product. This involves translation of the amplitude and geometric characteristics of the different seismic facies classes into a geologic, depositional environment, or rock property model. This stage of analysis still requires strong interpreter input, and typically involves integration of numerous techniques including: analysis of vertical and lateral seismic facies associations on vertical sections through the classification volume, determination of map patterns on horizontal and stratigraphic sections through the volume, comparison to the input seismic and other seismic attribute volumes, and finally, calibration to available well-log and core data (Figure 11). At the final integration stage, the improved efficiency and quality of the seismic facies classification volume produced using textural analysis and neural networks techniques translates into more time for the interpreter to focus on interpretation of the classification volume. This enables the interpreter to produce a more robust geologic analysis. Seismic facies interpretation applied to AVO attribute volumes. Mixed impedance reservoirs, typical of some deepwater depositional environments, are composed of sands and shales, which are not uniquely defined with only a near-offset volume or a full stack volume. In these environments, an integrated multidisciplinary workflow is especially important in delineating the seismic-scale distribution of sand and shale for reservoir characterization and input to volumebased geologic models. Seismic interpretation, including seismic facies analysis, remains a cornerstone of the workflow. However, with increasing reservoir complexity, reliance on sophisticated multitrace and multivolume (including AVO) attributes is necessary to capture this geologic complexity. We have extended the method for seismic facies analysis to these amplitude-versus-offset attribute volumes to demonstrate how the inherent subjectivity of a facies analysis, especially in mixed impedance deepwater environments, can be reduced with reliable prestack seismic data. In particular, the AVO properties of the interval, once fluid contacts have been identified, can be fundamental in the seismic-scale placement of sand and shale packages. Deterministic sand placement within facies-based net-togross packages can significantly improve the accuracy of a volume-based reservoir characterization workflow. Among the advantages of using a volume AVO attribute is the reduction in nonuniqueness in the seismic data. For example, using an A + B AVO attribute cube, high-amplitude seismic facies can be directly interpreted as likely HCcharged sands whereas low-amplitude facies can be interpreted as shale-prone and/or water-wet. Simplified well-to-facies ties are also a result of this type of application. We illustrate the predictive power of an AVO attribute facies analysis with an example from a stratigraphically complex, mixed-impedance clastic reservoir. In a mixed impedance scenario, the seismic expression of sand-prone intervals, in terms of amplitude and reflection continuity, will vary with AVO class of the sands and with the stack of data examined (Figure 12). Laterally amalgamated sands may also have varying responses depending on the stack of data examined (Figure 12). For example, a well-sorted, fine-grain class III sand may be associated with a low-amplitude-continuous
1048 THE LEADING EDGE OCTOBER 2002

reflection (LAC) on the near offset volume and a high-amplitude-continuous (HAC) on the far offset volume. Conversely, a poorly sorted, coarse-grained class I sand may be highamplitude continuous on the near offsets and dim to a lowor moderate-amplitude, continuous package on the far-offset volume. The full stack data would be some average of the two end-members, complicating the interpretation. In a properly constructed AVO attribute volume, lateral transitions between classes I, II, and III can be mitigated. Using an AVO attribute volume as the basis for seismic facies classification can thus minimize the complexity of interpreting multiple stacks of data to identify sand-prone intervals. For the location described above, the A + B volume contains a single high-amplitude continuous reflection denoting a single sand body regardless of AVO class (Figure 12, lower). The inclusion of the AVO classification information can be exploited to further discriminate high-impedance sands from lowimpedance sands for rock property assignment. Potential pitfalls of quantitative facies analyses. As with any quantitative analysis, data quality weighs heavily into the confidence of the interpretation. In cases where data quality is suspect, the potential subjectivity of a traditional, manual seismic facies analysis can be an asset when interpreters are aware of the limitations of their data. For example, where data are known to be nonstationary across a reservoir interval, manual seismic interpretation can compensate for this variability. Automated seismic facies methods, however, may not necessarily take this data variability into account unless spatially variant training is employed. Further, the use of computer generated facies data without appropriate geologic insight and interpreter understanding of the technology is dangerous. Such application of technology can give automated methods a black-box reputation and lead to a misunderstanding of the technology being applied, its expected benefits, and its limitations. Fundamental knowledge of the geology and geophysics of an exploration or development target, as well as an appreciation of the technology applied, is critical to the successful application of any technique in an integrated workflow. Conclusions. This article demonstrated: 1) The combination of seismic textural analysis and neural network techniques can be successfully applied to automate the seismic classification step of traditional seismic facies analysis, with resulting improvements in the efficiency, degree of detail, and reproducibility of the seismic facies classification product. Time-efficiencies and quality improvements result in more time for the interpreter to interpret the classification results, and to translate it into a robust geologic framework. 2) Two- and three-dimensional textural analysis provides a significantly different approach to seismic facies classification over single, trace-based seismic attributes in that seismic texture is a multitrace, volume-based attribute that provides a quantitative measure of the reflection amplitude, continuity, and internal configuration of reflectors. Seismic textural analysis thus provides a means to quantify the description of elements contained in seismic facies analysis. 3) The use of neural networks to classify textural attributes provides an efficient method by which the seismic interpreter can interactively teach the computer the seismic characteristics of interest. Probabilistic neural networks also offer the advantage of providing a quantitative measure of the confidence placed in each facies classification,

thereby increasing the efficiency of iterative training. The confidence value can also be used to dynamically reclassify until a specified confidence value is achieved. 4) Seismic facies analysis of reliable AVO attribute volumes can significantly reduce the uncertainty and nonuniqueness of lithologic interpretations based on seismic facies interpretation. With appropriate data, an AVO facies analysis may also be used to position sands in a clastic reservoir model.
Suggested reading. Segmentation of the Mid-Atlantic Ridge south of the Azores, based on acoustic classification of TOBI data by Blondel (in Tectonic, Magmatic, Hydrothermal and Biological Segmentation of Mid-Ocean Ridges, Geological Society Special Publication 118, 1996). Interactive AVO time-alignment and neural network classification by Eastwood and West (SEG 2001 Expanded Abstracts). An Introduction to Neural Networks by Gurney (UCL Press Limited, London, 1997). Statistical and structural approaches to texture by Haralick (Proceedings of the IEEE, 1979). Three-dimensional texture attributes for seismic data analysis by Randen et al. (SEG 2000 Expanded Abstracts). Seismic stratigraphy and global changes in sea level, Part 6: Stratigraphic interpretation of seismic reflection patterns in depositional sequence by Mitchum (in AAPG Memoir 26, 1977). Digital imaging processing techniques for enhancement and classification of SeaMarc II side-scan sonar imagery by Reed and Hussong ( Journal of Geophysical Research , 1989). Probabilistic neural networks by Specht (Neural Networks, 1990). 3D seismic texture classification by Vinther et al. (SPE 35482, 1996). TLE
Acknowledgments: We thank the numerous ExxonMobil geoscientists who have contributed to the development and effective use of these techniques: in particular, J. Ardill, C. Dawson, D. Gao, L. Foreman, M. Porter, R. Hill, R. Stephens, and L. Magennis. We also thank ExxonMobils affiliates and partners for permission to publish this work. Corresponding author: brian.p.west@exxonmobil.com

OCTOBER 2002

THE LEADING EDGE

1049

Potrebbero piacerti anche