← Back to Sessions

Denoising Seismic Migration Images Using ConvNeXt-style Neural Networks

Description: 

The interpretation of seismic migration images is essential for fault and reservoir characterization in geological carbon storage and geothermal projects. However, migration images often contain random and migration-related image noise that is difficult and time-consuming to remove. Data-driven approaches, such as supervised training of convolutional neural networks (CNNs), have shown promising results but require large amounts of training data. Vision transformers (VTs) have also demonstrated comparable or superior performance to CNNs for computer vision tasks, though they require more data and longer training times because of their lack of implicit bias compared to CNNs. To address these challenges, we generate realistic synthetic seismic migration volumes that include random and migration-related noise. We use these images to train a ConvNeXt-style neural network, which incorporates modern, efficient VT-inspired convolutional blocks, improving both accuracy and training/inferencing speed. On migration images of field seismic data, we show that this ConvNeXt-style architecture outperforms traditional CNNs and VT networks with limited training data while maintaining the accuracy of VTs.

Session: Seismology for the Energy Transition [Poster]

Type: Poster

Date: 4/16/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Joseph

Student Presenter: Yes

Invited Presentation: 

Poster Number: 132

Authors