Abstract:
Automated seismic facies classification using machine-learning algorithms is becoming more common in the geophysics industry. Seismic attributes are frequently used as input because they may express geologic patterns or depositional environments better than the original seismic amplitude. Selecting appropriate attributes becomes a crucial part of the seismic facies classification analysis. For unsupervised learning, principal component analysis can reduce the dimensions of the data while maintaining the highest variance possible. For supervised learning, the best attribute subset can be built by selecting input attributes that are relevant to the output class and avoiding using redundant attributes that are similar to each other. Multiple attributes are tested to classify salt diapirs, mass transport deposits (MTDs), and the conformal reflector “background” for a 3D seismic marine survey acquired on the northern Gulf of Mexico shelf. We have analyzed attribute-toattribute correlation and the correlation between the input attributes to the output classes to understand which attributes are relevant and which attributes are redundant. We found that amplitude and texture attribute families are able to differentiate salt, MTDs, and conformal reflectors. Our attribute selection workflow is also applied to the Barnett Shale play to differentiate limestone and shale facies. Multivariate analysis using filter, wrapper, and embedded algorithms was used to rank attributes by importance, so then the best attribute subset for classification is chosen. We find that attribute selection algorithms for supervised learning not only reduce computational cost but also enhance the performance of the classification. Introduction In the exploration and production industry, automated seismic facies classification is gradually being integrated into common workflows. Several machine-learning algorithms, such as self-organizing maps (SOMs) and K-means clustering, have been applied to automate seismic facies classification, and they are available in several commercial interpretation software packages. A great number of different seismic attributes can be used as input to machine-learning algorithms for classification and pattern recognition. However, some attributes express geologic or depositional patterns more effectively than others. For instance, the envelope (reflection strength) is sensitive to changes in acoustic impedance and has long been correlated to changes in lithology and porosity (Chopra and Marfurt, 2005). In many cases, the instantaneous frequency enhances interpretation of vertical and lateral variations of layer thickness (Chopra and Marfurt, 2005). Coherence measures lateral changes in the seismic waveform, which in turn can be correlated to lateral changes in structure and stratigraphy (Marfurt et al., 1998). Exploration generates large amounts of seismic data, andmany attributes generatedmay be highly redundant. Adding to this problem, the original seismic amplitude data (and therefore the subsequently derived attributes) may contain significant noise (Coléou et al., 2003). Therefore, understanding the nature of seismic attributes is of crucial importance for providing the most reliable classifications. According to the Hughes phenomenon, adding attributes beyond a threshold value causes a classifier’s performance to degrade (Hughes, 1968). Several studies found that dimensionality reduction in machine-learning problems reduces computation time and storage space as well as having meaningful results for facies classification (Coléou et al., 2003; Roy et al., 2010; Roden et al., 2015). Principal component analysis (PCA) is one of the most popular methods, reducing a large multidimensional (multiattribute) data set into a lower dimensional data set spanned by composite (linear combinations of the original) attributes, while preserving variation. SOM also creates a lower dimensional representation of high-dimensional data to aid interpretation. PCA and SOM are types of unsupervised learning, in which the goal is to discover the underlying structure of the input data. The University of Oklahoma, ConocoPhillips School of Geology and Geophysics, Norman, Oklahoma, USA. E-mail: yuji.kim@ou.edu (corresponding author); bob@ou.edu; kmarfurt@ou.edu. Manuscript received by the Editor 17 December 2018; revised manuscript received 14 March 2019; published ahead of production 29 May 2019; published online 23 August 2019. This paper appears in Interpretation, Vol. 7, No. 3 (August 2019); p. SE281–SE297, 16 FIGS., 9 TABLES. http://dx.doi.org/10.1190/INT-2018-0246.1. © 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 SE281 D ow nl oa de d 05 /0 8/ 20 to 6 8. 22 8. 16 8. 19 0. 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 / Roden et al. (2015) use PCA to define a framework for multiattribute analysis to understand which seismic attributes are significant for unsupervised learning. In their study, the combination of attributes determined by PCA is used as input to SOM to identify geologic patterns and to define stratigraphy, seismic facies, and direct hydrocarbon indicators. Zhao et al. (2018) build on these ideas and suggest a weight matrix computed from the skewness and kurtosis of attribute histograms to improve SOM learning. In general, attribute selection in unsupervised learning relies on the data distribution of the input attributes and the correlation between input attributes. Supervised learning maps a relationship between input attributes and output using an interpreter-defined training data set. Several supervised learning studies introduced attribute selection methods, also known as feature selection or variable selection to reduce dimensionality (Jain and Zongker, 1997; Chandrashekar and Sahin, 2014). We present multiple strategies to select appropriate attributes for seismic facies classification with a case study. Our goals are to provide a good classification model in terms of validation accuracy, to avoid overfitting, and to reduce the computation and memory requirements needed for generating seismic attributes. A desirable attribute subset might be built by detecting relevant attributes and discarding the irrelevant ones (Sánchez-Maroño et al., 2007). Although relevant attributes are those that are highly correlated with the output classes, redundant attributes are highly correlated with each other. Barnes (2007) suggests that there are many redundant and useless attributes that breed confusion in seismic interpretation; we argue that these attributes also pose problems in machine-learning classification. To avoid building an unnecessarily complex model, we evaluate several attribute selection algorithms to maximize relevance and minimize redundancy to build an efficient subset of attributes for supervised facies classification analysis. Attribute selection methods can be classified into three groups: (1) a filter method that uses a correlation or dependency measure, (2) a wrapper method that applies a predictive model to evaluate the performance of an attribute subset, and (3) an embedded method, which measures the attribute importance during the training process. Because multiple attributes are analyzed simultaneously in the test, we consider our attribute selection algorithm to be a multivariate algorithm. We compare the three types of attribute selection algorithms to build an efficient subset to differentiate seismic facies in a Gulf of Mexico survey. We generate 20 attributes from amplitude, instantaneous, geometric, texture, and spectral categories. The aim of the case study is to classify the specific facies based on patterns from a labeled training data set. We define the target classes of training data as being the facies corresponding to salt diapirs, MTDs, and conformal reflectors, which are created from manual geologic and stratigraphic interpretation. Correlations between attributes and correlations between attributes and output classes are analyzed using different measures to investigate Figure 1. Different types of relationship between variables X and Y and their correlation coefficients and regression score. Each scatterplot describes a different relationship between X and Y : (a and c) linear and monotonic relationships, (b and e) nonlinear, monotonic relationship, and (c and f) nonlinear, nonmonotonic relationships. Gaussian noise of 10% has been added to variable Y in (d-f). Coefficients are computed using Pearson, rank, MI, and distance correlation methods. A regression score is computed for the linear Bayesian, NN, RF, and SVM repressor predictive algorithms. The best hyperparameters for each model were obtained using a grid-search algorithm. SE282 Interpretation / August 2019 D ow nl oa de d 05 /0 8/ 20 to 6 8. 22 8. 16 8. 19 0. 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 / the relevance and redundancy of each seismic attribute. The selected attributes are tested using a random forest (RF) algorithm, and the classification results are discussed. We also apply our workflow to the Barnett Shale play in the Fort Worth Basin to differentiate shale and limestone facies using inverted physical properties as input attributes. The output class is labeled based on stratigraphic interpretation aided by adjacent wireline logs. The classification results using different attribute subsets are discussed. Correlation measures to maximize relevance and minimize redundancy Finding an optimal subset can be achieved by maximizing the relevance between attributes and output classes, while minimizing redundancy among attributes (Yu and Liu, 2004; Peng et al., 2005). To maximize relevance, attributes