Seismic Multichannel Reflectivity Inversion by Block Sparse Bayesian Learing
J.J. Wang, M. Ma and T.Y. Wang
Event name: 78th EAGE Conference and Exhibition 2016
Session: Broader Bandwidth Seismic Signal Processing
Publication date: 31 May 2016
Info: Extended abstract, PDF ( 1.1Mb )
Price: € 20
A Bayesian learning inversion of seismic reflectivity is achieved by employing the Relevance Vector Machine approach. However, it does not consider the spatial continuity of layers in adjacent traces. In this paper, we propose a new multichannel reflectivity inversion method by using the block sparse Bayesian learning (BSBL), which models the temporal correlation of adjacent nonzero solutions in each block. Unlike the conventional Bayesian inversion, this approach restore the sparse reflectivity series by iteratively estimating hyperparameters. A positive definite matrix is added to prior information, which is used to calculate the Mahalanobis distance and need to be estimated. Through Type-II maximum likelihood or Expectation-Maximization method, the global sparse resolution of reflectivity can be obtained. Synthetic examples illustrate the advantage of the presented method in recognizing weak signals and identifying thin beds, particularly in a low SNR cases. Meanwhile, the inversion result based on real post-stack data demonstrate the effectiveness and robustness of this improved technique.