Abstract:
Volcanic rocks with intermediate magma composition indicate distinctive patterns in seismic amplitude data. Depending on the processes by which they were extruded to the surface, these patterns may be chaotic, moderate-amplitude reflectors (indicative of pyroclastic flows) or continuous high-amplitude reflectors (indicative of lava flows). We have identified appropriate seismic attributes that highlight the characteristics of such patterns and use them as input to self-organizing maps to isolate these volcanic facies from their clastic counterpart. Our analysis indicates that such clustering is possible when the patterns are approximately self-similar, such that the appearance of objects does not change at different scales of observation. We adopt a workflow that can help interpreters to decide what methods and what attributes to use as an input for machine learning algorithms, depending on the nature of the target pattern of interest, and we apply it to the Kora 3D seismic survey acquired offshore in the Taranaki Basin, New Zealand. The resulting clusters are then interpreted using the limited well control and principles of seismic geomorphology. Introduction In today’s modern era, the most effective way to gain knowledge on how a certain geologic feature such as a lava flow appears in seismic data is to do a Google search and type a few key words such as “lava flow seismic” then go to the images section and even go through a couple of scientific publications, until we reach an “aha moment” when we find patterns that appear similar to those in our data set. This type of pattern recognition is easy for a human interpreter but is quite difficult for computers. The advantage of computers is that once such a task is well-defined, they can apply the analysis to every voxel in a large 3D seismic data volume, perhaps identifying subtle features that may have been overlooked by an overworked interpreter. Machine learning pattern recognition of seismic data goes beyond automation of time-consuming analysis tasks. Specifically, each prediction can be weighted by a confidence value, which can be used in subsequent risk analysis. Machine learning was first used by Alan Turing to decipher the Nazi “enigma” code (Gunderson, 1964). Since then, it has branched out to nearly all forms of language analysis, including voice recognition and translators, and it has expanded into visual communication, marketing, and social media. Early machine learning applications to seismic facies analysis include work by Meldahl et al. (2001) and West et al. (2002), who use multilinear feed-forward neural networks with seismic attributes to produce a probability volume of gas chimneys. Linari et al. (2003), Coleou et al. (2003), and Poupon et al. (2004) use seismic amplitude waveform and self-organizing maps (SOMs) to define zones of interest. Similarly, Verma et al., (2012), Roy et al. (2013), Roden et al. (2015), and Zhao et al. (2016) use volumetric seismic attributes fed into SOM algorithms to find different facies in shale resources plays. Qi et al. (2016) and Olorunsola et al. (2016) use generative topographic mapping (GTM) to try to separate salt from clastic, mass transport deposits (MTDs) from layered sediments in the Gulf of Mexico, and producing from tight facies in the Granite Wash in the Texas Panhandle, respectively.Lubo-Robles (2018) applies independent component analysis of spectral components to try to predict sandy facies in the Miocene Moki A Formation in the Taranaki Basin, New Zealand. Machine learning techniques are relatively simple mathematical algorithms that can learn from and generate clusters/classes based on patterns in (or interrelationships between) the data. Depending upon data availability, we can use either supervised or unsupervised algorithms. In supervised classification, the interpreter defines facies of interest, either by selecting specific voxels (Meldahl et al., 2001) or by drawing polygons around facies of interest (West et al., 2002; Qi et al., 2016), which serve as “training data” that are used to establish the relationship between input and output. Once trained, the algorithm is then applied to another subset of “validation data” (interpreted facies not used The University of Oklahoma, School of Geology and Geophysics, Norman, Oklahoma, USA. E-mail: lennoninfante@ou.edu; kmarfurt@ou.edu. Manuscript received by the Editor 1 June 2018; revised manuscript received 29 August 2018; published ahead of production 07 February 2019; published online 11 April 2019. This paper appears in Interpretation, Vol. 7, No. 3 (August 2019); p. SE1–SE18, 29 FIGS. http://dx.doi.org/10.1190/INT-2018-0096.1. © 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 SE1 D ow nl oa de d 07 /3 1/ 19 to 6 8. 97 .1 15 .2 6. 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 / in the training step) to determine if the algorithm is sufficiently accurate. If the validation is successful, the algorithm is then applied to the entire seismic data volume. In principle, unsupervised classification requires no interpreter input. In practice, the interpreter strongly biases the results of the algorithm by choosing a suite of attributes that best differentiate facies of interest. In a seismic interpretation context, this machine learning technique extracts patterns that exhibit a similar attribute expression for similar geologic facies, displaying these similar expressions, or clusters, using a 2D color-coded palette to allow subtle patterns to be identified by the interpreter (Qi et al., 2016; Zhao et al., 2016, 2017). Depending on the objective, supervised and unsupervised techniques use seismic attributes as input, in which the impedance and anisotropy attributes provide critical information for geomechanical clustering. In the absence of sufficient well control, instantaneous, geometric, spectral, and texture attributes provide critical information for interpreting seismic geomorphology from clustering (Zhao et al., 2016; Infante-Paez and Marfurt, 2017; Infante-Paez, 2018). Most recent studies in seismic interpretation have been focused on applying and comparing different machine learning methods, such as the multilayer perceptron network, SOMs, the support vector machine,Kmeans, and GTMs (Meldahl et al., 2001; Roy et al., 2013; Qi et al., 2016; Snyder, 2016; Zhao et al., 2016). We begin this study by defining the nature of the seismic patterns represented by volcanics in our seismic volume. We then propose a workflow that will allow interpreters to decide what machine learning algorithm to use, depending on the nature of the target pattern (TP). Next, we compute mathematically independent candidate attributes that highlight the continuity (such as gray-level co-occurrence matrix [GLCM] entropy), amplitude (peak spectral magnitude), and frequency (peak spectral frequency) of these TPs, with the goal of determining which input attributes best differentiate the volcanics from the surrounding clastic sediments. Finally, we input the GLCM entropy, peak spectral magnitude, and frequency attributes into the SOM, to interpret the seismic geomorphology of the internal elements of the Kora volcano. Methodology Selection of the TPs The TPs in our study include some of the internal and external elements of the Kora volcano, as well as adjacent volcanics from the Mohakatino Volcanic Belt (MVB). These volcanics form potential analogs to the volcanics in the Songliao Basin, China (Figure 1) and andesites from the Jatibarang field in Java (Figure 2), which have produced more than 1.2 billion barrels of oil and >2.7 trillion cubic feet of gas between 1969 and 1990 (Kartanegara et al., 1996). Figure 3 displays a vertical slice through the Kora 3D survey, where multiple TPs are highlighted by yellow boxes. Seismic-to-well ties indicate that these patterns have been drilled by exploration wells (Figure 4) validating the presence of volcanics. Nature of the TPs We define our human interpretation patterns as “monogenetic,” “composite,” and “intricate” patterns in which the goal is to examine relationships that can be evaluated by a machine. Monogenetic seismic patterns We define a monogenetic pattern as a facies that consists of a single seismic pattern. This pattern is statistically consistent, translational vertically and horizontally. The pattern is also consistent at different scales, such as conformal or chaotic reflectors within a 20 × 20 × 20 versus a 5 × 5 × 5 voxel window. Monogenetic seismic patterns are related to physical self-similarity, where the appearance of objects does not change at different scales of observation (Lam and Quattrochi, 1992; Dimri et al., 2011; Dasgupta, 2013; Herrera et al., 2017). Examples of monogenetic seismic patterns are shown in Figure 5. Composite seismic patterns Composite patterns are those facies that consist of two or more simpler patterns. Composite patterns do not entirely Figure 1. (a) Three-dimensional map of buried volcanos in the Xujiaweizi graben in the Songliao Basin, China, showing several wells targeting the buried volcanoes (Wang and Chen, 2015). (b) Buried volcanoes in the Taranaki Basin, New Zealand (after Giba et al., 2013; Bischoff et al., 2017). SE2 Interpretation / August 2019 D ow nl oa de d 07 /3 1/ 19 to 6 8. 97 .1 15 .2 6. 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 / preserve their character laterally, vertically, or at different scales, but they can still be distinguished from surrounding patterns (e.g., Figures 6 and 7). Intricate seismic patterns Intricate patterns are those facies that dramatically change their character with scale and location, and they Figure 2. Comparison of