Abstract: Unveiling meaningful geophysical information from seismic data requires to deal with both random and structured "noises". As their amplitude may be greater than signals of interest (primaries), additional prior information is especially important in performing efficient signal separation. We address here the problem of multiple reflections, caused by wave-field bouncing between layers. Since only approximate models of these phenomena are available, we propose a flexible framework for time-varying adaptive filtering of seismic signals, using sparse representations, based on inaccurate templates. We recast the joint estimation of adaptive filters and primaries in a new convex variational formulation. This approach allows us to incorporate plausible knowledge about noise statistics, data sparsity and slow filter variation in parsimony-promoting wavelet frames. The designed primal-dual algorithm solves a constrained minimization problem that alleviates standard regularization issues in finding hyperparameters. The approach demonstrates significantly good performance in low signal-to-noise ratio conditions, both for simulated and real field seismic data.
The Challenge of Seismic Multiples
In reflection seismology, sound waves are transmitted into the earth to map subsurface geology. The primary goal is to record primaries—waves that reflect exactly once off a geological interface and return to the surface. However, the resulting data is often heavily corrupted by multiples. These are reverberations that bounce back and forth between layers (such as the sea surface and the sea floor) before reaching the sensors.
Because multiples share similar frequency and waveform characteristics with primaries but arrive at different times, they can perfectly mask deeper, high-value geological targets. The industry standard for mitigation involves creating approximate "templates" of these multiples and using adaptive filters to subtract them. However, traditional least-squares adaptive filtering often fails when noise levels are high, templates are inaccurate, or when the filters accidentally suppress the primary signal itself.
A Convex Variational Framework
This paper proposes a more robust, mathematically principled approach by framing the separation of primaries and multiples as a constrained convex optimization problem. Instead of simple error minimization, the authors inject physical and statistical prior knowledge into the model through three key regularizers:
- Signal Sparsity: The primary signal is assumed to be sparse when transformed into a redundant wavelet frame domain.
- Filter Smoothness: Geological properties change gradually, so the time-varying Finite Impulse Response (FIR) filters applied to multiple templates are constrained to vary slowly between time steps.
- Filter Concentration: Filter taps are kept well-behaved using $\ell_1$, $\ell_2$, or mixed $\ell_{1,2}$ norms to prevent wild, non-physical fluctuations.
To solve this complex objective—which mixes smooth data-fidelity terms with non-smooth constraints—the authors employ a Monotone+Lipschitz Forward-Backward-Forward (M+L FBF) primal-dual algorithm. This allows the handling of constraints via proximity operators (geometric projections) rather than computationally expensive matrix inversions.
Key Insights: Frames and Constraints
A major finding of the research is that redundant wavelet frames (like Symmlets) significantly outperform standard orthogonal bases for preserving primary signals. The shift-invariance of frames prevents the introduction of unwanted artifacts during the subtraction process.
Furthermore, the authors argue that hard constraints are superior to standard regularization hyperparameters. By formulating prior information as strict bounds (e.g., "the filter cannot change by more than a specific threshold between steps"), the parameters become easier to estimate directly from the data using a fast, coarse first-pass algorithm. While an $\ell_2$ norm on filter taps yielded the highest SNR in low-noise scenarios, $\ell_1$ and mixed $\ell_{1,2}$ norms proved much more robust across varying and unknown noise levels.
Performance and Real-World Impact
The proposed method was tested against industry-standard algorithms on synthetic and real field data:
- Synthetic 2D Data: In high-noise conditions, the proposed method achieved a primary signal SNR of 16.77 dB, vastly outperforming a state-of-the-art industry algorithm which only achieved 3.11 dB.
- Real Field Data: When applied to marine seismic surveys, the algorithm successfully uncovered weak, flat primary signals that were entirely hidden beneath strong, sloping multiple events. Even when artificial wide-band noise was added to test robustness, the method maintained the integrity of the primary signal.
Conclusion and Future Directions
While the current methodology demonstrates exceptional performance, it relies on 1D wavelet frames applied trace-by-trace. Future work aims to extend this framework to 2D and 3D directional multiscale representations, such as curvelets or shearlets, to better capture the inherent spatial geometry of seismic waves. Beyond geophysics, this framework holds promise for other domains plagued by structured echoes, including acoustic echo cancellation in telecommunications and ultrasound non-destructive testing.
Summary updated