Classification of seismic waveform attributes in a Bayesian framework
Abstract
The work conducted in this thesis is an extension of an existing methodology developed at Schlumberger Stavanger Research. The Extrema Classification method (Borgos, 2003) provides a framework for automated mapping of structures in seismic data with the purpose of being a helpful contribution to manual interpretation. The method employs a model-based clustering technique in an effort to detect surfaces of similar waveform within the seismic dataset. Polynomial reconstruction of the seismic traces enables usage of the derivatives calculated at the extrema-points to describe the seismic waveform. Moreover, the derivatives representing similar waveforms are assumed to be normally distributed, which allows clustering to be performed by a variant of the expectation-maximation algorithm. In this thesis an alternative model-based clustering procedure, which includes neighborhood, is suggested as a possible improvement of the Extrema Classification method. The alternative method is based on the same assumption of Gaussian clusters, but extends the methodology into a Bayesian framework with a Markov random field as prior model for the distribution of cluster indices in the seismic cube. By selecting appropriate non-informative prior models for the Gaussian parameters clustering can be performed via the conditional posterior distributions. The new model is successfully implemented and tests on real seismic data show that the revised method is capable of producing a smoother output with less small surfaces than the original method. Due to increased complexity the new method is however more costly in terms of computer resources. Further studies should thus include an analysis of the trade-off and possibly look at the possibilities for a more efficient implementation.