Real-world application of deep learning in large-scale seismic interference attenuation: A case study in the Camie field of Angola

INTRODUCTION

• SI varies in incidence and amplitude across sail lines and can arrive from ahead, broadside, or astern of the recording vessel.Its amplitude depends in part on the external source’s distance, while incidence varies with its orientation relative to receivers.Source: INTRODUCTION · pp. 2–3
• High-amplitude SI can overlap much weaker subsurface reflections, making it difficult to separate interference from desired seismic signals.SI can remain consistent over large distances, increasing concern when it overlaps the reflections.Source: INTRODUCTION · p. 3
• Geophysical SI attenuation methods broadly either break coherence through data resorting and denoising or model the noise for subtraction.The second group depends strongly on building an accurate SI model (Jansen 2013).Source: INTRODUCTION · pp. 5–6
• A shipping-lane grid method requires prior lane knowledge, can be time-consuming, and has accuracy limited by grid size.A method based on detected apices fails when apices are absent from the seismic record.Source: INTRODUCTION · p. 7
• Conventional methods may require labor-intensive manual parameter selection for varying SI moveout and struggle when SI comes from multiple directions or has a dip like the underlying signal.These conditions make separating SI from desired reflections especially challenging.Source: INTRODUCTION · p. 7

• Earlier DNN studies showed potential on field data, but had not validated SI attenuation at industrial standards or compared full-block processing with a field-tested geophysical algorithm.A prior workflow was tested on two sail lines, while no full-block large-scale comparison had been presented.Source: INTRODUCTION · p. 8
• The study applies Sun et al. (2023)’s DNN workflow to the entire 345 km2 Camie marine block, presenting a large-scale real-world case study.The introduction describes it as the first published case study covering an area of this scale.Source: INTRODUCTION · pp. 8–9
• The comparison covers common-shot and stacked CMP examples, high-frequency RMS measures across sail lines and the survey, and total processing times.The timing comparison is not fully controlled across a shared computational environment, but reflects production scenarios with diverse resources and project-specific optimization.Source: INTRODUCTION · p. 9
• SI from other seismic vessels and local fishing vessels had diverse characteristics, including simultaneous sources from two directions, flat profiles, and dips resembling desired signals.Moveout- or dip-based filters can leave residual noise or leak useful signals when SI resembles primary reflections.Source: INTRODUCTION · p. 11

THE EMPLOYED DNN-BASED WORKFLOW AND GEOPHYSICAL ALGORITHM

• Training data came from 5 sail lines × 6 cables, representing 6% of the survey’s SI-contaminated data and covering the previously observed SI types.A conventional geophysical algorithm processed these selected gathers to estimate SI models.Source: THE EMPLOYED DNN-BASED WORKFLOW AND… · p. 12
• The conventional algorithm preconditioned common-shot gathers, used rank-reduction denoising, and modelled SI with sparse tau-p inversion.Complex wavelet adaptation combined a complex Morlet wavelet frame with unary complex Wiener filters.Source: THE EMPLOYED DNN-BASED WORKFLOW AND… · p. 12
• The study shuffled the conventional-algorithm SI model and blended it with naturally SI-free shots from a survey with 5 sail lines × 3 cables, augmenting both via vertical translation and amplitude modulation.Source: THE EMPLOYED DNN-BASED WORKFLOW AND… · p. 12
• The DNN used adjacent shots on both sides of the to-be-processed shot as additional input channels and was trained to predict both the SI and a blend of the SI-free shot with the intact shot.It followed a standard U-Net encoder-decoder structure with skip connections.Source: THE EMPLOYED DNN-BASED WORKFLOW AND… · pp. 13, 15
• Once trained, the DNN can attenuate SI noise in the remaining SI-contaminated survey data (91% of this case-study dataset); reprocessing the 6% used to generate training pairs could be commercially justified, but raises research concerns.Source: THE EMPLOYED DNN-BASED WORKFLOW AND… · pp. 13–14
• Evaluation compared the trained U-Net with the conventional algorithm on unseen survey data in shot and stacked CMP domains, including processing time.The authors state that examples were not from the exact same cables as the training data, to avoid repeatedly using the same dataset for DNN training and testing.Source: THE EMPLOYED DNN-BASED WORKFLOW AND… · pp. 15–16

• Across four shot-gather examples, DNN attenuation left much less SI residual and extracted SI more continuously and completely, with less signal leakage than conventional processing.Source: Comparison in Processing Quality · p. 16, Figure 8 · p. 17, Figure 10 · p. 18
• In CMP stacks, the DNN workflow achieved more complete SI extraction and significantly reduced signal leakage compared with the conventional algorithm.Source: Comparison in Processing Quality · p. 19, Figure 11 · p. 20
• Both workflows showed comparable limitations in the most heavily SI-contaminated areas, including central and northeastern sail lines.Source: Comparison in Processing Quality · p. 22
• On two sail lines, DNN processing largely removed vertical SI artifacts and produced more uniform backgrounds, improving continuity and preserving coherent signals in deeper channels.Source: Comparison in Processing Quality · p. 21, Figure 12 · p. 23
• Across the survey, conventional processing left significant residuals in heavily contaminated areas, while DNN processing reduced SI more consistently, with lower RMS levels on most sail lines.Source: Comparison in Processing Quality · p. 22, Figure 13 · p. 24

Comparison in Processing Time

• Conventional processing took 3.5 hours per sail line, excluding parameter optimization; scaling this to 41 sail lines gives approximately 143.5 hours.Source: Comparison in Processing Time · p. 25
• After training, the DNN processed each sail line in 15 minutes, or 10.25 hours across the survey; SI-model preparation took 8.75 hours and data preparation plus training took 35 hours.The reported total DNN time accounts for potential reprocessing of the SI-contaminated sail line used to provide an SI model for training pairs.Source: Comparison in Processing Time · p. 25
• Approximately 54 hours for the DNN-based workflow versus 143.5 hours for the conventional algorithm—roughly one-third the time using the same computational resources.The survey covered 345 km2.Source: Comparison in Processing Time · p. 25
• For larger surveys, DNN training time is not expected to increase significantly, while application-phase efficiency benefits are expected to scale up.Source: Comparison in Processing Time · p. 25

DISCUSSION

• Less than half the conventional algorithm’s computational time was required, and the DNN generally improved processing quality, especially in poor-SnR regions or where SI dips resembled signal dips.The time includes data preparation and DNN training.Source: DISCUSSION · p. 26
• The DNN attenuated SI despite training on incomplete, discontinuous SI models because those models covered expected SI types; SI too weak for conventional detection was absent from training and could also evade detection.Source: DISCUSSION · pp. 26–27
• When the entire survey is SI-contaminated, conventional processing must produce both SI models and pseudo-SI-free training shots, introducing uncertainty into DNN performance.Source: DISCUSSION · p. 27
• The evaluation excluded the 6% of raw data used to construct training pairs because its SI models were derived from those shots, creating partial dependency and fairness concerns despite possible production use.Source: DISCUSSION · p. 27
• Overfitting remains a concern despite augmentation and pairing strategies, while training SI models may not capture field complexity; sampling multiple sail lines broadens coverage of SI types.Source: DISCUSSION · p. 28
• Survey-specific training limits generalization: models may require retraining, fine-tuning, or domain adaptation for new data distributions, and future work should test robustness and other coherent noise types.Source: DISCUSSION · p. 28

• The study applied a U-Net workflow to attenuate SI across the entire 345 km2 Camie Field survey, presenting the largest-ever real-world comparison of this kind.The survey was in Angola.Source: CONCLUSIONS · p. 29
• The DNN workflow outperformed the conventional algorithm in processing quality and time across shot gathers, CMP stacks, and deep high-frequency RMS analysis.The comparison covered the full marine seismic survey block.Source: CONCLUSIONS · p. 29
• Future work includes investigating robustness and capability, applying the workflow to other surveys, evaluating alternative architectures, and exploring physics-based constraints to enhance generalization.Source: CONCLUSIONS · p. 29
• The supervised workflow depends on geophysical algorithms for labels covering inference-time SI types; training-pair quality is crucial, and selecting good samples requires human effort and experience.The paper notes that developing better geophysical algorithms remains important.Source: CONCLUSIONS · p. 29
• For large-scale real-world processing, a solid conventional algorithm remains essential, and experienced practitioners must balance effort across geophysical and deep-learning components.The conventional algorithm provides the foundation for the DNN to further improve SI attenuation.Source: CONCLUSIONS · p. 29