[S-IT19] Coupling of deep Earth and surface processes

[SIT19-P04] Deep Learning-Based 3D Marine Multi-Channel Seismic Data Interpolation: Improving Subsurface Imaging

*Nasrin Tavakolizadeh^1,2, Jin-Oh Park², Hamzeh Mohammadigheymasi², Ehsan Jamali Hondori³, Nuno Pombo¹ (1.University of Beria Interior, Covilha, Portugal, 2.Department of Ocean Floor Geoscience, Atmosphere and Ocean Research Institute, The University of Tokyo, 3.Geoscience Enterprise Inc. (GSE), Japan)

Keywords:Deep learning-based interpolation, Multichannel seismic data, Generative Adversarial Networks (GANs), Convolutional Neural Networks (CNNs), Subsurface imaging

Interpolation of 3D marine multichannel seismic (MCS) data is critical for addressing acquisition-related effects such as feathering, which can degrade spatial continuity and compromise subsurface imaging accuracy. It also plays a key role in balancing inline and crossline bin sizes, as streamer line spacing is typically larger than receiver spacing due to technical constraints and the high cost of MCS data acquisition. Interpolation serves multiple purposes, including data regularization, reconstruction, de-noising, and super-resolution. Traditional interpolation methods, such as nearest-neighbor, Fourier-based, and kriging techniques, often rely on assumptions of linearity, sparsity, or low-rank representations of seismic events assumptions that may not always hold in complex geological environments. As a result, these methods often struggle to preserve high-frequency details and can introduce artifacts, particularly when dealing with intricate subsurface structures. In contrast, deep learning-based interpolation methods leverage data-driven approaches to adaptively reconstruct missing seismic traces, offering enhanced spatial continuity, improved resolution, and better generalization to geological variability. By learning spatial dependencies directly from the data, these methods overcome the limitations of conventional techniques, providing more accurate and reliable seismic reconstructions for imaging and interpretation. In this study, we explore the application of deep learning-based interpolation methods, specifically convolutional neural networks (CNN) and generative adversarial networks (GANs), to enhance the quality of marine MCS data. Our approach involves training these models on high-resolution seismic datasets to learn spatial patterns and reconstruct missing traces while preserving geological continuity. The models are optimized using a combination of loss functions that balance data fidelity and structural coherence, ensuring realistic and high-quality interpolated results. Our ultimate goal is to apply these techniques to the 3D seismic data acquired during the R/V Hakuho-maru cruise (KH-25-JE01) in January 2025 in the northeastern region of the Noto Peninsula, improving subsurface imaging for enhanced geological interpretation.