TL;DR
This paper introduces a novel constrained optimization framework for separating primary seismic signals from unwanted "multiple" reflections. By reformulating the problem as a convex optimization task with carefully chosen constraints, the authors demonstrate improved accuracy in estimating both the primary signal and the filters responsible for generating the multiples, particularly by leveraging hyperplane constraints.
What Problem Does This Paper Address?
Seismic data acquisition, used for subsurface imaging, is plagued by noise. A significant challenge is the presence of "multiple" reflections, which are secondary wave paths that can obscure or mimic genuine primary reflections from deeper geological layers. These multiples share similar waveform and frequency characteristics with primaries, making their separation difficult. Traditional methods often struggle with weak decorrelation, non-stationary data, or high noise levels. The problem is inherently under-determined, meaning there are more unknowns than observations, necessitating the incorporation of prior knowledge.
Methods
The core of this paper is a constrained convex optimization approach. The authors model the observed seismic trace $z$ as the sum of the primary signal $y$, the multiples $Rh$ (where $R$ represents known templates and $h$ are unknown time-varying filters), and additive noise $b$. The problem is then formulated as minimizing an objective function that balances data fidelity (how well the model fits the observed data) with prior knowledge about the primary signal $y$ and the filters $h$.
What's novel here is the emphasis on hard constraints rather than traditional regularization. They propose two main types of constraints:
- Data fidelity: Using a norm (specifically the $l_1$-norm is highlighted) to measure the difference between the observed data and the model's prediction.
- Prior knowledge constraints: These are incorporated into convex sets $C$ (for filters $h$) and $D$ (for primaries $y$).
- For filters $h$, they impose constraints on the smoothness of their time-varying nature.
- For primaries $y$, they explore constraints based on wavelet frame representations. Crucially, they introduce and investigate the use of hyperplane constraints in addition to standard $l_p$-ball constraints. These hyperplanes, defined by linear operators, are shown to be more effective in guiding the optimization towards geophysically plausible solutions and are easier to project onto.
The optimization is solved using a primal-dual proximal algorithm, which is well-suited for problems involving sums of convex functions and indicator functions of convex sets.
Datasets
The authors use synthetic seismic data for their simulations. They generate observations based on a realistic primary signal and templates, with known time-varying filters and added Gaussian noise at different levels ($\sigma = 0.01$ and $\sigma = 0.04$). No new benchmark datasets are introduced.
Key Insights and Ideas
- Hyperplanes as powerful constraints: The paper strongly advocates for the use of hyperplane constraints for guiding the optimization. These constraints, which define linear subspaces, are shown to be more effective than traditional $l_p$-ball constraints in improving the estimation of the primary signal.
- Ease of projection: Hyperplane constraints are computationally advantageous because they allow for explicit and efficient projection steps within the optimization algorithm, simplifying implementation.
- Flexibility of the framework: The constrained optimization approach is highly flexible, allowing for the integration of various types of prior information about the signal and filters, making it adaptable to different seismic scenarios.
- Bridging regularization and constraints: By reformulating regularization into hard constraints, the authors aim to simplify hyperparameter selection, a common bottleneck in inverse problems.
Key Results
The simulations demonstrate that the proposed constrained optimization approach, particularly when using hyperplane constraints for the primary signal, leads to significant improvements in signal-to-noise ratio (SNR) for both the estimated primary signal ($y$) and the estimated multiples ($s$).
- In Table 1, the "I+sign" constraint (an intersection of two hyperplane constraints) consistently yields the highest SNRs across different noise levels and for both $y$ and $s$. For instance, at $\sigma = 0.01$, the SNRy reaches 26.40 dB, and SNRs reaches 15.56 dB.
- The paper shows that the estimated multiples are well-separated from the primary signal, and the stochastic noise is effectively removed, even for low-amplitude signal components.
- The introduction of prior information through constraints provides a gain of up to 1.4 dB in SNR compared to cases where no constraint is applied to the primary signal's representation.
Limitations and Future Work
- Hyperparameter tuning for hyperplanes: While the framework aims to reduce hyperparameter tuning, the specific parameters defining the hyperplanes (e.g., $\lambda_1, \lambda_2$ in equation (7)) still require careful selection.
- Extension to higher dimensions: The authors express interest in extending their study to higher dimensions and exploring the class of hyperplane constraints further, including their inner parameters.
- Real-world data validation: The current results are based on synthetic data. Future work would involve validating the approach on real seismic datasets to assess its performance in more complex and noisy environments.
- Generalization of filter constraints: While smooth variations of filters are considered, exploring other types of prior knowledge for filter estimation could be beneficial.
Summary updated