Principal Component Analysis Applied to 3D Seismic Data for Reservoir Property Estimation
- J.R. Scheevel (Chevron Petroleum Technology Co.) | K. Payrazyan (Chevron Petroleum Technology Co.)
- Document ID
- Society of Petroleum Engineers
- SPE Reservoir Evaluation & Engineering
- Publication Date
- February 2001
- Document Type
- Journal Paper
- 64 - 72
- 2001. Society of Petroleum Engineers
- 4.1.5 Processing Equipment, 1.7.5 Well Control, 4.3.4 Scale, 5.1.2 Faults and Fracture Characterisation, 5.1.8 Seismic Modelling, 5.6.3 Deterministic Methods, 5.1.7 Seismic Processing and Interpretation, 1.6 Drilling Operations, 5.1.5 Geologic Modeling, 5.1 Reservoir Characterisation, 5.8.5 Oil Sand, Oil Shale, Bitumen, 2.4.3 Sand/Solids Control
- 0 in the last 30 days
- 536 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 10.00|
|SPE Non-Member Price:||USD 30.00|
We apply a common statistical tool, Principal Component Analysis (PCA) to the problem of direct property estimation from three-dimensional (3D) seismic-amplitude data. We use PCA in a novel way to successfully make detailed effective porosity predictions in channelized sand and shale.
The novelty of this use revolves around the sampling method, which consists of a small vertical sampling window applied by sliding along each vertical trace in a cube of seismic-amplitude data. The window captures multiple, vertically adjacent amplitude samples, which are then treated as vectors for purposes of the PCA analysis. All vectors from all sample window locations within the seismic-data volume form the set of input vectors for the PCA algorithm.
Final output from the PCA algorithm can be a cube of assigned classes, whose clustering is based on the values of the most significant principal components (PC's). The clusters are used as a categorical variable when predicting reservoir properties away from well control. The novelty in this approach is that PCA analysis is used to analyze covariance relationships between all vector elements (neighboring amplitude values) by using the statistical mass of the large number of vectors sampled in the seismic data set.
Our approach results in a powerful signal-analysis method that is statistical in nature. We believe it offers data-driven objectivity and a potential for property extraction not easily achieved in model-driven fourier-based time-series methods of analysis (digital signal processing).
We evaluate the effectiveness of our method by applying a cross-validation technique, alternately withholding each of the three wells drilled in the area and computing predicted effective porosity (PHIE) estimates at the withheld location by using the remaining two wells as hard data. This process is repeated three times, each time excluding only one of the wells as a blind control case. In each of the three blind control wells, our method predicts accurate estimates of sand/shale distribution in the well and effective porosity-thickness product values. The method properly predicts a low sand-to-shale ratio at the blind well location, even when the remaining two hard data wells contain only high sand-to-shale ratios.
Good predictive results from this study area make us optimistic that the method is valuable for general reservoir property prediction from 3D seismic data, especially in areas of rapid lateral variation of the reservoir. We feel that this method of predicting properties from the 3D seismic is preferable to traditional, solely variogram-based geostatistical estimation methods. Such methods have difficulty capturing the detailed lithology distribution when limited by the hard data control's sampling bias. This problem is especially acute in areas where rapid lateral geological variation is the rule. Our method effectively overcomes this limitation because it provides a deterministic soft template for reservoir-property distributions.
Reservoir Prediction from Seismic.
The use of the reflection seismic-attribute data for the prediction of detailed reservoir properties began at least as early as 1969.1 Use of seismic attributes for reservoir prediction has accelerated in recent years, especially with the advent of widely available high-quality 3D seismic data.
In practice, a seismic attribute is any property derived from the seismic reflection (amplitude) signal during or after final processing. Any attributes may be compared with a primary reservoir property or lithology in an attempt to devise a method of attribute-guided prediction of the primary property away from well control. The prediction method can vary from something as simple as a linear multiplier (single attribute) to multi-attribute analysis with canonical correlation techniques,2 geostatistical methods,3 or fully nonlinear, fuzzy methods.4
The pace of growth in prediction methodologies using seismic attributes seems to be outpaced only by the proliferation in the number and types of seismic attributes reported in the literature.5 As more researchers find predictive success with one or more new attributes, the list of viable reservoir-predictive attributes continues to grow. Chen and Sidney6 have cataloged more than 60 common seismic attributes along with a description of their apparent significance and use.
Despite the rich history of seismic attribute in reservoir prediction, the practice remains difficult and uncertain. The bulk of this uncertainty arises from the unclear nature of the physics connecting many demonstrably useful attributes to a corresponding reservoir property. Because of the complex and varied physical processes responsible for various attributes, the unambiguous use of attributes for direct reservoir prediction will likely remain a challenge for years to come.
In addition to the questions about the physical origin of some attributes, there is the possibility of encountering statistical pitfalls while using multiple attributes for empirical reservoir-property prediction. For example, it has been demonstrated that as the number of attributes used in an evaluation increases, the potential arises that one or more attributes will produce a false correlation with well data.7 Also, many attributes are derived with similar signal-processing methods and can, in some cases, be considered largely redundant with respect to their seismic-signal description. Lendzionowski et al.8 maintain that the maximum number of independent attributes required to fully describe a trace segment is a quantity 2BT, where B=bandwidth (Hz) and T=trace-segment length (sec). If this is supportable, it suggests that most of the more common attributes are at least partially redundant. The danger of such redundancy is that it falsely enhances statistical correlation with the well property. Doing so may suggest that many seemingly independent seismic attributes display similar well-property trends.
Finally, the use of a particular approach with attributes involves some subjectivity and prior experience on the part of the practitioner to be successful and reproducible. This is a source of potential error that cannot be quantified but also, in most cases, cannot be avoided. The most successful workers in the field of reservoir prediction from seismic, not coincidentally, are also the most experienced in the field.
|File Size||444 KB||Number of Pages||9|