Depicting geologic sequences from three-dimensional seismic surveying is of wide applications to subsurface reservoir exploration. In this paper, we present an innovative workflow for seismic stratigraphy interpretation by utilizing the state-of-the-art deep convolutional neural networks (CNNs). Specifically, the workflow consists with two components: (a) seismic feature self-learning (SFSL) and (b) stratigraphy model building (SMB), each of which is achieved in a deep CNN. While the latter is supervised and of the typical network architecture used in image segmentation, we design the former as unsupervised and requiring no knowledge from domain experts. Compared to the convolutional approaches, the proposed workflow is superior in two aspects. First, by initializing the SMB network from the SFSL one, it successfully inherits the prior-knowledge for understanding the target seismic data, and therefore such supervised learning can be efficiently completed by only a small amount of training data. Second, for the convenience of seismic experts in providing training labels, we design our workflow applicable to three scenarios, trace-wise, paint-brush, and full-section annotation. The performance of our proposed workflow is verified through application to two real seismic datasets from the North Sea and the Solsikke. It is concluded that the new workflow is not only capable of providing reliable seismic stratigraphy interpretation but holds the potential for assisting other geophysical problems, such as geobody detection.
Presentation Date: Wednesday, September 18, 2019
Session Start Time: 8:30 AM
Presentation Time: 8:30 AM
Presentation Type: Oral