A Self-Supervised Deep Learning Method for Seismic Data Deblending Using a Blind-Trace Network.

Area of Science:

• Geophysics
• Machine Learning
• Seismic Data Processing

Background:

• Simultaneous-source technology enhances seismic surveying efficiency and cost-effectiveness by recording blended subsurface responses from multiple sources.
• Deblending is crucial for separating individual source signals after simultaneous-source seismic acquisition.
• Deep learning offers potential for seismic processing but faces challenges with labeled data, overfitting, and generalization.

Purpose of the Study:

• To propose a novel self-supervised learning method for seismic deblending.
• To address the challenge of limited labeled data in deep learning for seismic processing.
• To achieve accurate separation of individual source-response pairs without labeled training datasets.

Main Methods:

• Development of a novel self-supervised learning approach for seismic deblending.
• Utilizing a blind-trace deep neural network architecture.
• Designing a specialized blending loss function to guide the deblending process.

Main Results:

• The proposed self-supervised method successfully separates individual source-response pairs.
• Accurate deblending was demonstrated across three distinct blended-acquisition designs.
• The method overcomes the need for labeled training datasets, mitigating overfitting and generalization issues.

Conclusions:

• Self-supervised learning offers a viable solution for seismic deblending in simultaneous-source acquisition.
• The developed blind-trace neural network and blending loss function are effective for signal separation.
• This approach enhances the practical application of deep learning in seismic data processing.