Full text loading...
-
Understanding How a Deep Neural Network Architecture Choice Can Be Related to a Seismic Processing Task
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, First EAGE Digitalization Conference and Exhibition, Nov 2020, Volume 2020, p.1 - 5
Abstract
One of the many challenges in the way of the adoption of Deep Learning (DL) for seismic processing is the understanding of deep neural network (DNN) architecture and components with the associated underlying physics involved in a specific processing task. In this article, we study how some convolutional DNN architectures can be naturally suited to given processing tasks, helping the interpretability and opening the door to meaningful QCs. For instance, we show that the Unet architecture ( Ronneberger et al., 2015 ) can naturally learn to “separate” the kinematics of seismic events from their amplitude variations and use both information efficiently; this is illustrated on the CIG (common image gathers) skeletonization (or picks probability computation) and muting task. We also illustrate that the Denet ( Remez et al, 2017 ) architecture can naturally learn to decompose a “noise” model into meaningful complementary contributions, with the receiver deghosting from variable depth streamer data example.