[Paper Review] A hybrid approach to seismic deblending: when physics meets self-supervision

Nick Luiken, Matteo Ravasi+1 more|arXiv (Cornell University)|May 30, 2022

Seismic Imaging and Inversion Techniques9 citations

TL;DR

This paper proposes a hybrid seismic deblending method that integrates physics-based modeling with a self-supervised blind-spot denoising network within a Plug-and-Play (PnP) framework. By training the denoiser on-the-fly without ground truth labels and leveraging time-coherent noise modeling, the approach outperforms a state-of-the-art sparsity-based method while maintaining computational efficiency and generalizing across marine and land acquisitions.

ABSTRACT

To limit the time, cost, and environmental impact associated with the acquisition of seismic data, in recent decades considerable effort has been put into so-called simultaneous shooting acquisitions, where seismic sources are fired at short time intervals between each other. As a consequence, waves originating from consecutive shots are entangled within the seismic recordings, yielding so-called blended data. For processing and imaging purposes, the data generated by each individual shot must be retrieved. This process, called deblending, is achieved by solving an inverse problem which is heavily underdetermined. Conventional approaches rely on transformations that render the blending noise into burst-like noise, whilst preserving the signal of interest. Compressed sensing type regularization is then applied, where sparsity in some domain is assumed for the signal of interest. The domain of choice depends on the geometry of the acquisition and the properties of seismic data within the chosen domain. In this work, we introduce a new concept that consists of embedding a self-supervised denoising network into the Plug-and-Play (PnP) framework. A novel network is introduced whose design extends the blind-spot network architecture of [28 ] for partially coherent noise (i.e., correlated in time). The network is then trained directly on the noisy input data at each step of the PnP algorithm. By leveraging both the underlying physics of the problem and the great denoising capabilities of our blind-spot network, the proposed algorithm is shown to outperform an industry-standard method whilst being comparable in terms of computational cost. Moreover, being independent on the acquisition geometry, our method can be easily applied to both marine and land data without any significant modification.

Motivation & Objective

• To address the challenge of deblending seismic data from simultaneous source acquisitions, which entangle signals from consecutive shots.
• To overcome the lack of ground truth data in seismic inverse problems by developing a self-supervised learning approach that bypasses the need for labeled training examples.
• To improve deblending performance by combining physical modeling of the blending operator with a deep denoiser tailored for trace-wise coherent noise.
• To create a geometry-agnostic method that generalizes across marine and land seismic data without retraining or major architectural changes.
• To ensure practical applicability by maintaining computational cost comparable to conventional methods while achieving superior reconstruction quality.

Proposed method

• The method employs a Plug-and-Play (PnP) framework, where the data misfit term enforces consistency with the physics of the blending operator, and a learned denoiser acts as a prior.
• A novel self-supervised denoising network is designed based on the blind-spot architecture but extended to model coherent noise along the time axis, enabling better handling of burst-like interference.
• The denoiser is trained in a self-supervised manner at each PnP iteration using the current noisy estimate as input, eliminating the need for clean ground truth data.
• The network is trained for a fixed number of epochs per iteration using a denoising autoencoder objective on the current iterate, with no external labels.
• Hyperparameters such as inner and outer iterations, penalty parameter ρ, and training epochs are tuned, with the authors suggesting fixed values may generalize well.
• The algorithm uses the distance between the current iterate and the denoised estimate as a convergence indicator, enabling adaptive stopping without relying on fixed iteration counts.

Experimental results

Research questions

• RQ1 Can a self-supervised deep denoiser be effectively integrated into a physics-based inverse problem framework like PnP for seismic deblending without requiring ground truth data?
• RQ2 Does a time-extended blind-spot network architecture improve deblending performance compared to standard denoisers when dealing with coherent, trace-wise noise in blended seismic data?
• RQ3 Can the proposed hybrid method outperform a conventional sparsity-based deblending approach (e.g., patched Fourier) in terms of reconstruction accuracy while maintaining similar computational cost?
• RQ4 How robust is the method across different seismic acquisition geometries, such as marine and land surveys, without architectural or hyperparameter re-tuning?
• RQ5 To what extent does network initialization affect the final deblending result, and can convergence be reliably monitored during the optimization process?

Key findings

• The proposed method achieves superior deblending quality compared to the conventional patched Fourier approach, with a median relative error across 10 random seeds that is consistently lower than the benchmark.
• In 9 out of 10 random weight initializations, the final deblending result was of higher quality than the conventional method, as shown in the box plot of final relative errors.
• The algorithm demonstrates convergence in practice, with the iterates xₖ and yₖ approaching each other under carefully selected hyperparameters, suggesting stable behavior despite theoretical convergence guarantees being unproven.
• The number of training epochs per PnP iteration significantly affects reconstruction quality and convergence, indicating that this hyperparameter requires tuning for optimal performance.
• The method is robust to network initialization and generalizes well across different acquisition types, including marine and land data, due to its physics-informed, geometry-agnostic design.
• The algorithm maintains computational cost comparable to the benchmark method, making it a viable alternative for real-world seismic processing pipelines.