Improving seismic fault detection by super-attribute-based classification

Abstract

Fault interpretation is one of the routine processes used for subsurface structure mapping and reservoir characterization from 3D seismic data. Various techniques have been developed for computer-aided fault imaging in the past few decades; for example, the conventional methods of edge detection, curvature analysis, redgreen-blue rendering, and the popular machine-learning methods such as the support vector machine (SVM), the multilayer perceptron (MLP), and the convolutional neural network (CNN). However, most of the conventional methods are performed at the sample level with the local reflection pattern ignored and are correspondingly sensitive to the coherent noises/processing artifacts present in seismic signals. The CNN has proven its efficiency in utilizing such local seismic patterns to assist seismic fault interpretation, but it is quite computationally intensive and often demands higher hardware configuration (e.g., graphics processing unit). We have developed an innovative scheme for improving seismic fault detection by integrating the computationally efficient SVM/MLP classification algorithms with local seismic attribute patterns, here denoted as the super-attribute-based classification. Its added values are verified through applications to the 3D seismic data set over the Great South Basin (GSB) in New Zealand, where the subsurface structure is dominated by polygonal faults. A good match is observed between the original seismic images and the detected lineaments, and the generated fault volume is tested usable to the existing advanced fault interpretation tools/modules, such as seeded picking and automatic extraction. It is concluded that the improved performance of our scheme results from its two components. First, the SVM/MLP classifier is computationally efficient in parsing as many seismic attributes as specified by interpreters and maximizing the contributions from each attribute, which helps minimize the negative effects from using a less useful or “wrong” attribute. Second, the use of super attributes incorporates local seismic patterns into training a fault classifier, which helps exclude the random noises and/or artifacts of distinct reflection patterns. Introduction Faults and fractures are important subsurface structures of significant geologic implications for hydrocarbon accumulation and migration in a petroleum reservoir, and the presence of a fault can be visually recognized as a lineament/plane of abrupt variations of the reflection signals in a 3D seismic data set. However, fault interpretation is a time-consuming and labor-intensive process, especially for an exploration area of a large number of faults and complicated faulting histories and distributions. In the past few decades, great efforts have been devoted into computer-aided fault interpretation by developing new attributes and methods/algorithms to help detect, depict, and extract the faults of interpretational interest from the surrounding nonfaulting features. Specifically, from the perspective of seismic attribute analysis, edge detection and reflector geometry estimation are applicable to the problem of fault mapping from 3D seismic data, owing to the lateral changes in seismic signals across a fault, including reflection waveform/amplitude and depth/two-way traveltime. Geoscientists have devoted substantial efforts for quantifying such changes and improving the resolution and noise robustness of fault detection (e.g., Bahorich and Farmer, 1995; Luo et al., 1996; Marfurt et al., 1998; Gersztenkorn and Marfurt, 1999; van Bemmel and Pepper, 2000; Cohen and Coifman, 2002; Tingdahl and de Rooij, 2005; Di and Gao, 2014a; Wang et al., 2016). For example, Bahorich and Farmer (1995) present the coherence attribute by estimating the crosscorrelation of two adjacent seismic traces to highlight the faults and stratigraphic features from a seismic cube. Marfurt et al. (1998) present the semblance attribute by estimating the amplitude variations in a horizontal window. Gersztenkorn and Marfurt (1999) perform principal component analysis on a local coherence cube for Formerly Georgia Institute of Technology, School of Electrical and Computer Engineering, Center for Energy and Geo Processing (CeGP), Atlanta, Georgia 30308, USA; presently Schlumberger, Houston, Texas 77056, USA. E-mail: [email protected] Georgia Institute of Technology, School of Electrical and Computer Engineering, Center for Energy and Geo Processing (CeGP), Atlanta, Georgia 30308, USA. E-mail: [email protected]; [email protected]; [email protected] Manuscript received by the Editor 18 October 2018; revised manuscript received 4 February 2019; published ahead of production 03 June 2019; published online 7 August 2019. This paper appears in Interpretation, Vol. 7, No. 3 (August 2019); p. SE251–SE267, 18 FIGS., 13 TABLES. http://dx.doi.org/10.1190/INT-2018-0188.1. © 2019 Society of Exploration Geophysicists and American Association of Petroleum Geologists. All rights reserved. t Special section: Machine learning in seismic data analysis Interpretation / August 2019 SE251 D ow nl oa de d 08 /2 0/ 19 to 1 36 .2 52 .1 29 .6 7. R ed is tr ib ut io n su bj ec t t o SE G li ce ns e or c op yr ig ht ; s ee T er m s of U se a t h ttp :// lib ra ry .s eg .o rg / improved coherence analysis. Cohen and Coifman (2002) propose using the local structural entropy for fault mapping. Tingdahl and de Rooij (2005) develop the similarity operator for measuring the differences between two seismic trace segments. Di and Gao (2014a) compare the performance of the common edge detectors, including the popular Canny detector, on seismic fault detection. A comprehensive summary of the edge-detection attributes can be found in Chopra (2002), Kington (2015), and Di and Gao (2017a). However, the conventional seismic edge detection is limited in its detection resolution for small-scale faults and fractures beyond the seismic scale and offers no physical link for predicting the fundamental fracture properties (e.g., intensity, orientation, and sense of displacement) either quantitatively or qualitatively (Gao, 2013). Then, the seismic geometric attributes are developed for more robust fault detection and fracture characterization by quantifying the lateral variations of the geometry of seismic reflectors, including the first-order dip, the second-order curvature, and the third-order flexure attributes. Specifically, the dip describes the local dipping of a reflector, and a fault is highlighted as a lineament of large dipping angle. The curvature describes the bending of a reflector, and a fault is highlighted as a juxtaposition of positive and negative curvatures (Roberts, 2001; Al-Dossary and Marfurt, 2006). The flexure describes the shearing of a reflector, and a fault is highlighted as a local peak accompanied with two subtle side lobes (Gao, 2013; Di and Gao, 2014b, 2016a, 2016b; Yu, 2014; Gao and Di, 2015; Qi and Marfurt, 2018). Comprehensive summaries of the curvature and flexure analysis can be found in Roberts (2001) and Di and Gao (2017b), respectively. From the perspective of fault-interpretation methods, manual picking is considered most reliable if performed by an experienced interpreter. However, it is limited by the interpretation efficiency especially for a large seismic data set with a complicated deformation history (e.g., folding and faulting). Correspondingly, computer-aided fault interpretation becomes the research focus with the progress in computer graphics and image processing since 2000, and various methods/algorithms have been developed for refining the edge-detection attributes and interpreting fault surfaces (e.g., Crawford and Medwedeff, 1999; Pedersen et al., 2002; Admasu et al., 2006; Barnes, 2006; Lavialle et al., 2007; Hale, 2013; Zhang et al., 2014; Machado et al., 2016; Wu and Hale, 2016; Wu and Fomel, 2018; Di and AlRegib, 2019). For example, Pedersen et al. (2002) introduce the concept of ant colony optimization from computer science and develop an ant-tracking algorithm for sharpening the lineaments in a variance volume. AlBinHassan and Marfurt (2003) apply the 2D Hough transform for enhancing the fault lines on time slices; later, Wang and AlRegib (2014) extend it to 3D space for fault surface extraction from a semblance volume. Barnes (2006) performs eigenvector analysis to a coherence volume and designs a discontinuity filter of three components for imaging the steeply dipping faults. Admasu et al. (2006) propose an autotracking method of fault line propagation from one vertical section to another throughout a seismic volume for extracting an individual fault patch. Lavialle et al. (2007) present a nonlinear filtering approach for noise suppression and fault enhancement based on 3D gradient structure tensor analysis. Hale (2013) proposes scanning over all possible fault orientations for computing fault likelihood (faultoriented semblance), strikes, and dips, constructing fault surfaces as quadmeshes from the three fault images, and further applying a dynamic time warping algorithm to estimate fault throws on the each fault surface. Zhang et al. (2014) first apply a biometric algorithm to the coherence attribute for fault skeletonization and then group discrete fault points into one fault patch under local planar constraints (Gibson et al., 2005). Wang et al. (2014a) borrow the ideas of motion vectors in video coding and processing to assist seismic fault extraction. Machado et al. (2016) perform volumetric fault imaging (VFI) by applying the directional Laplacian of a Gaussian filter to coherence anomalies along reflector dip and azimuth. Wu and Hale (2016) propose using the fault skin, a simple linked-data structure, to construct fault surfaces and fill holes. Wu and Fomel (2018) present an optimal surface voting algorithm to enhance a fault attribute image, estimate fault orientations, and construct complete fault surfaces. Di and AlRegib (2019) pro