The Prediction Error Filters Applied Seismic Interpretation

One of the main problems faced in reservoir characterization is the need to infer subsurface properties from seismic data. Due to scarcity of well-log information, seismic attributes can be applied as delimiters of zones with similar seismic response that may be due to a set of reservoir properties. These techniques are called “classification techniques” and are based on the fact that seismic waves collect information from the physical properties of the subsurface.

Currently, Self-Organizing Maps (SOM) and Applied Neural Networks (ANN) are the two most popular methods in classifying seismic stratigraphic patterns. We compare the relative power of these two methods against the relatively underutilized Prediction Error Filter (or PEF, also known as an Autoregressive Filter) to identify user-defined patterns across multiple attribute volumes. The “vector of reference” can either be a suite of seismic amplitudes (giving rise to horizon-controlled waveform classification), a vector of attributes such as impedance, coherence, curvature, texture, and amplitude curvature at each sample (giving rise to volumetric classification), or a combination of the two Whatever their design, these filters are also used to compare and classify the seismic response with respect to a vector of reference. Unlike most implementations of SOM and ANN, the main advantage of using the PEF is that it provides a measure of confidence in the classification, thereby providing a measure of uncertainty to the interpretation. In this study, we compare these three methods to a gas shale and Mississippi-lime targets from the Midcontinent of the United States

AAPG Search and Discovery Article #90142 © 2012 AAPG Annual Convention and Exhibition, April 22-25, 2012, Long Beach, California