Semisupervised multiattribute seismic facies analysis

Abstract

One of the key components of traditional seismic interpretation is to associate or “label” a specific seismic amplitude package of reflectors with an appropriate seismic or geologic facies. The object of seismic clustering algorithms is to use a computer to accelerate this process, allowing one to generate interpreted facies for large 3D volumes. Determining which attributes best quantify a specific amplitude or morphology component seen by the human interpreter is critical to successful clustering. Unfortunately, many patterns, such as coherence images of salt domes, result in a salt-and-pepper classification. Application of 3D Kuwahara median filters smooths the interior attribute response and sharpens the contrast between neighboring facies, thereby preconditioning the attribute volumes for subsequent clustering. In our workflow, the interpreter manually painted [Formula: see text] target facies using traditional interpretation techniques, resulting in attribute training data for each facies. Candidate attributes were evaluated by crosscorrelating their histogram for each facies with low correlation implying good facies discrimination, and Kuwahara filtering significantly increased this discrimination. Multiattribute voxels for the [Formula: see text] interpreter-painted facies were projected against a generative topographical mapping manifold, resulting in [Formula: see text] probability density functions (PDFs). The Bhattacharyya distance between the PDF of each unlabeled voxel to each of [Formula: see text] facies PDFs resulted in a probability volume of each user-defined facies. We have determined the effectiveness of this workflow to a large 3D seismic volume acquired offshore Louisiana, USA.