Self-supervised learning-based seismic trace extrapolation for near-offset gap

Abstract

Marine seismic surveys, measured using a towed streamer system, inherently acquire data with missing traces in the near-offset range, approximately 100 to 200 meters from the shot location due to the limitations of the survey equipment. This near-offset gap presents a challenge in removing seismic multiples from seismic datasets, which ultimately affects the quality of the final migrated image. This thesis presents a deep learning (DL) based seismic trace extrapolation methodology that aims to reconstruct the near-offset gap. The restoration of near-offset data using DL techniques presents unique challenges because it is fundamentally impossible to learn from the label data typically used in the DL-based interpolation methods. To address this, the novel two-step approach involving self-supervised learning technique is proposed. In addition, this work analyzes and mitigates the dataset bias, drawing on the concept from the field of computer vision, in the construction of training dataset for near-field extrapolation. A two-step seismic trace extrapolation methodology forms the core of this work. The upstream task involves training various near-offset features using open synthetic datasets in the public domain. Subsequently, the downstream task improves the performance of near-offset gap extrapolation by tuning the pre-trained near-offset features to the target data, which has no labeled data, through transfer learning. The use of open synthetic seismic datasets for pre-training significantly reduces both the computational time and cost associated with generating synthetic data similar to the field data using conventional 2D or 3D modeling methods. In addition, simultaneous training on near-offset features derived from different datasets mitigates the selection bias associated with synthetic data tailored to specific models. At the same time, it mitigates the labeling bias in field data due to the lack of near-offset data. In this thesis, the effectiveness of the proposed near-offset extrapolation method was validated through numerical experiments using an open synthetic dataset which is not used for pre-training and calculated from complex geological structure. The reliability of the proposed approach was established through cross-validation, comparing its results with those obtained by the previous DL-based trace interpolation method and the pre-trained model. Then, in order to determine the field applicability of the proposed method, its performance has been tested by applying it to real field data. The results of field experiments confirmed that the proposed method can effectively extrapolate near-offset gap in real field data. The proposed self-supervised learning-based method may have a different final purpose depending on which task is defined in upstream and downstream tasks. Thus, the two-stage strategy outlined in this thesis is not exclusively limited to the domain of near-offset extrapolation; it inherently has the capacity for application and extension to various facets of seismic data processing.