Seismic images are often contaminated by migration noise. The noise attenuation process can take a lot of effort from the domain expert and, in many cases, it can be challenging to get the optimal result. In recent years it has been demonstrated that data-driven approaches can produce quality results with minimum effort. In digital image processing, convolutional neural networks (CNN) have gained a lot of popularity. When trained properly on carefully selected data, CNNs can potentially outperform traditional methods through task automation leading to reduced turnaround time of processing projects. In this work we propose to train a neural network, specifically a U-net architecture, to eliminate migration artifacts from seismic images. We explain the data preparation step and describe the model parameters and training process. Finally, we demonstrate the model performance on field data examples from three different geographical regions.

This content is only available via PDF.
You can access this article if you purchase or spend a download.