Loading...
Data-driven and machine learning approaches for suppression of multiply reflected seismic waves
Kiraz, Mert Sinan Recep
Kiraz, Mert Sinan Recep
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2023
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
Imaging the Earth's interior requires seismic data to be free from the events which reflect more than once in the subsurface (also referred to as multiples). For most seismic imaging algorithms, multiples are considered noise since they shadow the useful information about the subsurface, and they need to be removed from the seismic data. However, the removal of multiples remains one of the most challenging tasks in seismic data processing to this day.
The occurrence of multiples depends on a strong impedance contrast between two adjacent subsurface structures and the strength of the multiples increases as the contrast increases. The presence of complex overburden structures (e.g., salt bodies) in the subsurface is one of the reasons for a strong impedance contrast. The solution of the Marchenko equation retrieves the up- and down-going focusing functions which allow one to account for the effects of complex overburden structures and enables the retrieval of the Green's functions for virtual source locations in the subsurface. I propose a two-dimensional (2-D) Marchenko equation where the up- and down-going focusing functions are not needed for the Green's function retrieval. The proposed Marchenko equation not only effectively retrieves the Green's function for a virtual source location in the subsurface, but also forms the basis for obtaining multiple-free images of the subsurface. I also investigate the role of the background velocity model for a successful Green's function retrieval with the Marchenko equation for the reflected and refracted waves. I show that a constant background model retrieves a high-accuracy Green's function as long as the average slowness can be obtained between the virtual source and receiver locations. I also show that the refracted waves can only be incorporated into the retrieved Green's function with a background model that is detailed enough to model the refracted waves during the Marchenko scheme.
The presence of a free surface (or air-water interface) is another reason for a strong impedance contrast. The free-surface effects can be seen in two different ways in the seismic data; (1) free-surface multiples and (2) ghost reflections. Free-surface multiples have at least one downward reflection at the free surface and are the most dominant set of multiples in marine seismic data. Ghost reflections create notches in the frequency content of the seismic data causing the low-frequency content to deteriorate. I develop a Convolutional Neural Network (CNN) (which is an important building block of artificial intelligence) solution to attenuate the free-surface multiples and remove ghost reflections from seismic data, simultaneously. The developed CNN algorithm works on a single seismic trace and does not need to interpolate, extrapolate, or regularize data gaps and near offsets, avoiding the costs and time associated with such procedures. It is designed for use in irregular surveys, such as ocean-bottom node seismic data or time-lapse monitoring studies. I present synthetic examples to demonstrate that the CNN-based solution removes free-surface effects from seismic data. I then apply the developed CNN-based simultaneous free-surface multiple attenuation and ghost removal method on a 2-D field data set acquired in the North Viking Graben. Lastly, I apply the developed CNN-based solution to the simultaneously acquired seismic data (also referred to as blended data). I first demonstrate the efficacy of the proposed algorithm on a 2-D synthetic blended data set. Then, I apply the proposed algorithm to a 2-D field data set acquired in the North Sea. The numerical examples demonstrate that the CNN algorithm effectively eliminates free-surface effects on blended data.
Associated Publications
Rights
Copyright of the original work is retained by the author.