Acquisition & imaging artifacts in raw seismic images can result in exploration risks and inaccurate rock property predictions. In this work, 3D partial–full image pairs were used to train 3D convolutional neural networks (CNN) to simulate the stacking operation on a single 3D image. Compared to full images, partial images are overwhelmed by noise and poor illumination, so the learned mapping/function from partial to full images attenuates migration artifacts and partially compensates for fold/illuminated-related amplitude distortions. The trained model can be applied to either pre-stack or post-stack 3D image volumes to better reconstruct geological structures and simplify interpretation.
Presentation Date: Monday, September 16, 2019
Session Start Time: 1:50 PM
Presentation Time: 3:30 PM
Location: 221D
Presentation Type: Oral
