Least-squares reverse-time migration (LSRTM) based on reflection demigration is a linear inversion-based imaging method toward true reflectivity. We develop a new preconditioner for LSRTM with a supervised machine learning (ML) method, i.e. Support Vector Machine (SVM), to automatically recognize and suppress possible imaging noise. We design a nonlinear SVM classifier to separate signal and noise on the shot-indexed common image gathers (CIGs) at each iteration. The proposed SVM preconditioner introduces only small additional computational cost: (1) The shot-indexed CIGs can be sorted conveniently; (2) only two efficient attributes, the local coherency by semblance analysis and the absolute amplitudes of CIG imaging points, are measured by the Gaussian kernel function of SVM; and (3) the time-consuming SVM training is completed prior to LSRTM and is not involved in the inversion procedure. During the iterative classification procedure, we use only one trained SVM model to define classifying functions to automatically remove different imaging artifacts. Numerical experiments show that the SVM preconditioned LSRTM can effectively suppress sparse acquisition footprints and is robust to some velocity errors.
Presentation Date: Monday, October 12, 2020
Session Start Time: 1:50 PM
Presentation Time: 3:30 PM
Location: Poster Station 7
Presentation Type: Poster
