DISCUSSION

We attempted to find a PE filter that removes the local, plane reflector component of an image, and thereby enhances and exposes any subsurface discontinuities, such as erosional surfaces or faults. Since a general PE filter removes any predictable part of the data and whitens its output's power spectrum, we had to restrict the PE filter so that the filter preserves the subsurface's discontinuities. To restrict the filter, we limited its shape to a set of planar 2-D filters which were two columns wide and which had only about ten free coefficients. We did not use the subcube size to tune which events are predictable and which are not. Instead, we strove for small subcubes that promise an optimal resolution.

Despite our efforts to restrict the PE filter, the filter usually removed most discontinuities that we were seeking to enhance. Our restrictions failed, because our assumptions about the subsurface were partially wrong: geological discontinuities in a seismic subsurface image are usually zones of rapid but continuous transition (often several pixels wide) and not singular discontinuous boundaries. Consequently, we need to develop ways of tuning the filter process to reject certain bandwidths while preserving others.

The PE filter framework itself offers several potential remedies: we can

• low-pass filter the residuals.
• introduce gaps in the PE filters which limits the filters' ability to predict and consequently remove rapidly varying image features.
• prewhiten the input data, before using the PE filter described in this article.

The PE filter approach was easy to test because we had developed the necessary software. Obviously, there are alternative filter techniques (such as traditional edge detection schemes) that we may want to pursue alternatively.

A working coherency measure could possibly be an important building block in a library of data visualization and processing tools. A natural extension of the image processing aspect of the coherency measure could be software that picks the enhanced surfaces and delivers the interpreter a proposed set of continuous surfaces (e.g., as a GOCAD model). The coherency image could be complemented by other local data attributes, such as instantaneous frequency or seismic amplitude. An ambitious data analysis and visualization software could combine statistics of the image subcube, of the migrated data that stacks to the image subcube, and of the raw data that contributes to the image subcube.