A Primal-Dual Proximal Algorithm for Sparse Template-Based Adaptive Filtering: Application to Seismic Multiple Removal

Mai Quyen Pham, Laurent Duval, Caroline Chaux, Jean-Christophe Pesquet

TL;DR

This paper introduces a novel convex optimization framework to separate true seismic signals ("primaries") from structured reverberations ("multiples") and random noise. By combining sparse wavelet representations with constrained, time-varying adaptive filters, the authors deploy a primal-dual algorithm to uncover weak geological signals that traditional industry methods often miss or distort.

What Problem Does This Paper Address?

In reflection seismology, sound waves are fired into the earth to map subsurface geology. The signals of interest are primaries—waves that bounce exactly once off a geological interface and return to the surface.

However, seismic data is heavily corrupted by multiples. These are waves that bounce back and forth between layers (like an echo in a canyon) before reaching the surface. Because multiples share the same frequency and waveform characteristics as primaries but arrive later, they often perfectly mask deeper, highly valuable geological targets.

The industry standard for removing multiples involves creating approximate "templates" of the multiples and using adaptive filters to subtract them from the recorded data. However, traditional least-squares adaptive filtering struggles when noise levels are high, when the templates are highly inaccurate, or when the filters accidentally adapt to and subtract the primary signal itself. This paper seeks a more robust, mathematically principled way to perform this adaptive subtraction.

Methods

The authors frame the separation of primaries and multiples as a constrained convex optimization problem. Instead of just minimizing the error between the data and the model, they inject heavy prior knowledge (regularization) into both the signal and the adaptive filters:

  1. Signal Sparsity: The primary signal is assumed to be sparse when transformed into a wavelet frame domain (enforced via an $\ell_1$ norm).
  2. Filter Smoothness: Because geological properties change gradually, the time-varying Finite Impulse Response (FIR) filters applied to the multiple templates are strictly constrained to vary slowly from one time step to the next.
  3. Filter Concentration: The filter taps are constrained to be well-behaved (preventing wild fluctuations) using $\ell_1$, $\ell_2$, or mixed $\ell_{1,2}$ norms.

To solve this complex objective function—which mixes smooth data-fidelity terms with non-smooth constraints and $\ell_1$ penalties—the authors use a Monotone+Lipschitz Forward-Backward-Forward (M+L FBF) primal-dual algorithm. This allows them to handle the constraints via proximity operators (geometric projections) without requiring massive, computationally prohibitive matrix inversions.

Datasets

  • Synthetic 1D Data: Modeled seismic traces containing known primaries, two distinct multiple templates, and varying levels of additive Gaussian noise.
  • Synthetic 2D Data: A 2D gather of traces showing bended hyperbolas, simulating a multi-event seismic shot with high noise.
  • Real Field Data: Actual marine seismic survey data, complete with recorded traces and industry-provided multiple wavefield templates.

Key Insights and Ideas

  • Frames Beat Bases: Using redundant wavelet frames (like Symmlets or Daubechies) is significantly better for preserving the primary signal than using orthogonal wavelet bases. The shift-invariance of frames prevents artifacts that bases introduce.
  • Hard Constraints > Hyperparameters: By formulating the prior information as hard constraints (e.g., "the filter cannot change by more than $x$ between steps") rather than standard regularization penalties, the parameters become easier to estimate directly from the data using a fast, coarse first-pass algorithm.
  • The Right Norm for the Job: While an $\ell_2$ norm on the filter taps yielded the highest absolute Signal-to-Noise Ratio (SNR) in low noise, the $\ell_1$ and mixed $\ell_{1,2}$ norms provided much more statistically consistent and robust results across varying and unknown noise levels.

Key Results

  • Quantitative (Synthetic 1D): The proposed frame-based method improved primary restoration SNR by up to ~2.2 dB compared to baseline orthogonal basis approaches.
  • Comparative (Synthetic 2D): On a highly noisy 2D synthetic dataset ($\sigma = 0.08$), the proposed method achieved an SNR of 16.77 dB for the separated primaries. In contrast, a state-of-the-art industry-standard algorithm (based on complex continuous wavelets) only achieved 3.11 dB.
  • Qualitative (Real Data): When applied to real marine data, the algorithm successfully uncovered a weak, flat primary signal that was entirely hidden beneath a strong, sloping multiple event. When artificial wide-band noise was added to the real data to test robustness, the proposed method maintained the primary signal much better than the industry baseline.

Limitations and Future Work

  • Limitations: The current methodology relies on 1D wavelet frames applied trace-by-trace. Seismic data is inherently 2D/3D, meaning 1D processing leaves spatial correlations on the table. Additionally, setting the bounds for the hard constraints still requires running a faster, cruder algorithm first to get baseline estimates.
  • Future Work: The authors suggest extending the framework to 2D directional multiscale representations (like curvelets or shearlets) to better capture the geometry of seismic waves. They also note this framework could be highly effective in other domains suffering from structured echoes, such as acoustic echo cancellation in audio or ultrasound non-destructive testing.

Summary updated

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.
BibTeX
@article{pham2014primal,
  author = {Mai Quyen Pham and Laurent Duval and Caroline Chaux and Jean-Christophe Pesquet},
  title = {{A Primal-Dual Proximal Algorithm for Sparse Template-Based Adaptive Filtering: Application to Seismic Multiple Removal}},
  journal = {IEEE Trans. Signal Process.},
  year = {2014},
  url = {https://openreview.net/forum?id=QORXntOLxp},
}

Loading related content...

Figure