A model‐based data‐driven dictionary learning for seismic data representation. (17th July 2017)

• Record Type: Journal Article
• Authors: Yarman, Can Evren; Kumar, Rajiv; Rickett, James
• Is Part Of: Geophysical prospecting. Volume 66:Number 1(2018)
• Page Start: 98
• Page End: 123
• Publication Date: 2017-07-17
• Subjects: Dictionary learning -- Kernel method -- Reproducing kernel -- Variational method -- non‐linear optimisation -- variable projection; Prospecting -- Geophysical methods -- Periodicals; 622.15
• DOI: 10.1111/1365-2478.12533
• Languages: English
• ISSNs: 0016-8025

Abstract

ABSTRACT: Planar waves events recorded in a seismic array can be represented as lines in the Fourier domain. However, in the real world, seismic events usually have curvature or amplitude variability, which means that their Fourier transforms are no longer strictly linear but rather occupy conic regions of the Fourier domain that are narrow at low frequencies but broaden at high frequencies where the effect of curvature becomes more pronounced. One can consider these regions as localised "signal cones". In this work, we consider a space–time variable signal cone to model the seismic data. The variability of the signal cone is obtained through scaling, slanting, and translation of the kernel for cone‐limited (C‐limited) functions (functions whose Fourier transform lives within a cone) or C‐Gaussian function (a multivariate function whose Fourier transform decays exponentially with respect to slowness and frequency), which constitutes our dictionary. We find a discrete number of scaling, slanting, and translation parameters from a continuum by optimally matching the data. This is a non‐linear optimisation problem, which we address by a fixed‐point method that utilises a variable projection method with ℓ1 constraints on the linear parameters and bound constraints on the non‐linear parameters. We observe that slow decay and oscillatory behaviour of the kernel for C‐limited functions constitute bottlenecks for the optimisation problem, which we partially overcome by the C‐Gaussian function. We demonstrate our method through an interpolation example. We present the interpolation result using the estimated parameters obtained from the proposed method and compare it with those obtained using sparsity‐promoting curvelet decomposition, matching pursuit Fourier interpolation, and sparsity‐promoting plane‐wave decomposition methods.