Seismic interpretation with the aid of deep learning

Loading...
Thumbnail Image
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Alabama Libraries
Abstract

Nowadays one of the biggest challenges for geoscientists is effectively extracting the useful information from massive geo-datasets. Deep learning algorithms have become incredibly good at analyzing and identifying pieces of objects from massive data. The application of deep learning in seismic exploration has become one of the hottest research topics in recently two-years. My dissertation focuses on developing new workflows for seismic data processing and interpretation with the aid of deep learning algorithms. Picking the first arrival of seismic data is one of the most time-consuming tasks in the seismic data processing. The first arrival segments the seismic traces into two parts. Each part of the seismic traces can be viewed as a unique object. I automatically identify the two objects of the seismic trace by using a state-of-art pixel-wise convolutional image segmentation method. The boundary of the two objects is regarded as the first arrivals of seismic data. Noise filtering is another important step in the seismic data processing. I proposed to filter the noise in seismic data by integrating deep learning and variational mode decomposition. My new method does not require prior information about the noise which is one of the compulsory inputs for image de-noising using deep learning. My method not only effectively removes the random noise in the seismic image but also the coherence noise such as migration artifacts which is beyond the capability of current filtering methods. The process of seismic horizon interpretation can be treated as dividing the seismic traces into several segments. I proposed a workflow to perform semi-automated horizon interpretation method by using the encoder-decoder convolutional neural network. There are two main parts of my workflow. The first part is segmenting the seismic traces into different parts using deep learning and treat the boundary of two nearby parts as the horizon. The second part is refining the horizons using a two-step filtering. My method does not require seismic attributes such as the dip and azimuth of a seismic reflector as the inputs which are compulsory for current horizon picking algorithms.

Description
Electronic Thesis or Dissertation
Keywords
Geology, Geophysics
Citation