event

Ph.D. Proposal Oral Exam - Yazeed Alaudah

Primary tabs

Title:  Weakly Supervised Semantic Labeling of Subsurface Structures

Committee: 

Dr. AlRegib, Advisor    

Dr. Davenport, Chair

Dr. McClellan

Abstract:

The objective of the proposed research is to develop a weakly-supervised framework for the semantic labeling of subsurface structures within large seismic volumes, and to establish the proposed labeling framework as a major component towards an automated workflow for structural seismic interpretation. To be able to develop and train state-of-the-art deep convolutional networks for segmenting and labeling various subsurface structures in seismic volumes, vast amounts of training data are required. However, labeled seismic data is extremely limited, and therefore this can be an exceptionally difficult task. To overcome this challenge, a weakly-supervised framework is developed for obtaining large amounts of pixel-level labeled training data, using very minimal input from a seismic interpreter. This framework involves developing an accurate seismic image similarity measure and deploying it for retrieving large amounts of images with high visual similarity to reference images selected by an interpreter. A weakly-supervised label mapping technique is then used to transform these image-level labels into pixel-level labels that encode the locations of the target classes within the image. Finally, these weakly-mapped labels are used to train deep convolutional networks for labeling subsurface structures within large seismic volumes. The benefit of this work is that it enables the training and deployment of deep learning models to new application domains, such as seismic interpretation, where sufficient quantities of labeled data are not available.

Status

  • Workflow Status:Published
  • Created By:Daniela Staiculescu
  • Created:04/04/2018
  • Modified By:Daniela Staiculescu
  • Modified:04/04/2018

Categories

Target Audience