TL;DR
This paper introduces a novel "Triply Laplacian Scale Mixture" (TLSM) model for seismic data noise suppression. By jointly estimating sparse tensor coefficients and their variances, TLSM significantly improves noise removal accuracy and detail preservation compared to existing methods, all while maintaining computational efficiency.
What Problem Does This Paper Address?
Seismic data is crucial for understanding subsurface geology, but it's often contaminated by acquisition footprint and Gaussian noise. While sparsity-based tensor recovery methods are effective at noise suppression, they struggle with the non-stationary nature of seismic data. Specifically, accurately estimating the variances of sparse tensor coefficients from noisy data is challenging, leading to suboptimal denoising performance. This paper aims to overcome this limitation by proposing a more robust statistical modeling approach.
Methods
The core of this work is the Triply Laplacian Scale Mixture (TLSM) model. Unlike traditional methods that use fixed sparsity priors (like soft- or hard-thresholding), TLSM models each sparse tensor coefficient as a Laplacian distribution with a positive scalar multiplier. This allows for the joint estimation of both the coefficient values and their variances directly from the noisy data.
To handle the resulting complex optimization problem, the authors employ the Alternating Direction Method of Multipliers (ADMM) algorithm. This approach breaks down the problem into a series of smaller, manageable subproblems that can be solved iteratively. The ADMM framework is applied to a reformulated objective function that incorporates the TLSM priors for different regularization terms (e.g., for the tensor itself and its derivatives).
Datasets
The authors evaluate their proposed TLSM algorithm on:
- Synthetic Seismic Data: This dataset is generated to include linear cross-events with added time-decaying footprint noise and Gaussian noise. They create 12 distinct noise configurations by varying the footprint amplitude (F) and Gaussian noise standard deviation (σ). The synthetic data is tested in various configurations, including sizes like 100x200x400.
- Field Seismic Datasets: Two publicly available field datasets are used:
- Penobscot-3D dataset (subset)
- Kerry-3D dataset (subset)
Key Insights and Ideas
- Joint Estimation of Coefficients and Variances: The most significant insight is the move from fixed sparsity priors to a model that jointly estimates both the sparse coefficients and their variances. This directly addresses the non-stationarity issue in seismic data.
- Laplacian Scale Mixture (LSM) for Robust Priors: The use of LSM provides a more flexible and accurate statistical prior for the sparse tensor coefficients, leading to better denoising.
- ADMM for Tractable Optimization: The successful application of ADMM makes the complex TLSM model computationally feasible and stable for practical seismic data processing.
- Preservation of Fine Details: The TLSM approach not only suppresses noise effectively but also demonstrates a superior ability to preserve subtle geological features compared to other methods.
Key Results
- Quantitative Performance: The TLSM algorithm consistently outperforms six state-of-the-art methods (DRR, SGRDL, FR-Net, DNLR, S2S-WTV, and UTV-LRTA) in terms of PSNR and SSIM on synthetic data. For example, it achieves gains of up to 18dB in PSNR and significant improvements in SSIM. On field data, TLSM also achieves the lowest scores (indicating better performance) for no-reference quality metrics like BRISQUE, NIQE, and PIQE.
- Qualitative Performance: Visual comparisons show that TLSM produces cleaner seismic data with better preservation of structural details, while other methods either leave residual noise or blur important features.
- Computational Efficiency: TLSM exhibits the fastest running time among the compared methods, even when implemented solely on a CPU, highlighting its practical applicability.
Limitations and Future Work
- Parameter Sensitivity: While the paper suggests the algorithm is relatively insensitive to parameter choices, extensive parameter tuning was performed for optimal results. Further research could explore more automated parameter selection strategies.
- Computational Cost: Although efficient, the algorithm still involves complex tensor operations. Future work could explore further optimizations or hardware acceleration (e.g., GPU) to speed up computation even more.
- Extension to Other Noise Types: The current work focuses on acquisition footprint and Gaussian noise. Future research could investigate the TLSM model's effectiveness against other types of seismic noise.
Summary updated