Seismic Data Enhancement and Representation by Self-Supervised Learning

Seismic acquisition is a costly and labor-intensive process. As such, advanced technologies have emerged in recent years to strike a balance between data quality and acquisition cost, including improved sensor technology, novel seismic acquisition strategies, and advanced data processing and enhancement algorithms. While some methods are designed to directly enhance data quality or acquisition efficiency, others prioritize one aspect over the other initially, followed by post-processing remediation to achieve the optimal balance between the two.

This dissertation primarily focuses on three critical aspects to address the efficiency-cost trade-off, namely seismic data deblending for simultaneous source acquisition, seismic data interpolation, and seismic data compression. Simultaneous source acquisition, or blended acquisition has been introduced to reduce acquisition costs and improve efficiency by firing denser sources within limited time frames. However, an effective deblending algorithm is necessary to separate the blended energy and eliminate the blending noises. Seismic data interpolation is another area of focus that aims to improve data resolution and restore missing or contaminated seismic traces. As the volume of seismic data continues to grow, data compression methods can alleviate storage, transmission, and computation costs.

Recent advances in deep learning and its data-driven approaches toward feature engineering have yielded many encouraging results for various seismic data processing tasks, including seismic deblending, interpolation, and compression. In this dissertation, we propose new approaches for these tasks that leverage cutting-edge deep learning techniques. Compared to the prevalent supervised learning scheme for seismic data processing, our work targets exploring self-supervised learning schemes to solve the problems. We present innovative networks and algorithms that are specifically tailored for seismic data processing, incorporating unique features and physics information of seismic data. To evaluate the effectiveness of our proposed methods, we conduct experiments using both synthetic and field datasets.