Blind Curvelet-Based Denoising of Seismic Surveys in Coherent and Incoherent Noise Environments

Naveed Iqbal, Mohamed Deriche, Ghassan AlRegib, Sikandar KhanArabian 2023 · DOI · arXiv

Introduction

Seismic imaging serves as the cornerstone of modern geophysical exploration, providing the primary means to map subsurface geological structures for oil and gas reservoir identification. However, the fidelity of these images is frequently compromised by noise. Raw seismic data recorded by geophones is often obscured by a combination of incoherent noise (random sensor fluctuations) and coherent noise, such as "ground roll"—low-frequency, high-amplitude waves that propagate along the surface. This paper presents a "blind" denoising framework that leverages the curvelet transform and pre-whitening techniques to recover clean seismic signals from complex noise environments.

The Challenge of Seismic Noise

The primary difficulty in seismic denoising lies in the overlap between signal and noise characteristics. Traditional methods, such as band-pass filtering, often fail when the frequency spectra of the reflections and the noise (like ground roll) coincide, leading to the inadvertent removal of critical geological information.

Seismic shot gather illustrating the distinction between geological reflections and ground roll noise.

As shown in Figure 1, seismic reflections appear as hyperbolic events in the time-distance domain, while ground roll noise presents as high-amplitude, low-velocity linear or dispersive waves. While the curvelet transform is theoretically well-suited for representing these curved reflections, its performance typically degrades when noise is "colored" or correlated, a common occurrence in field data.

Proposed Methodology

To address these limitations, the authors propose a four-stage automated workflow designed to handle correlated noise without prior knowledge of noise statistics.

1. Pre-Whitening

The core innovation is the application of Zero-Phase Component Analysis (ZCA) to whiten the data. By decorrelating noise samples, ZCA transforms complex "colored" noise into "white" noise. This step is crucial because it allows standard thresholding techniques, which are optimized for white noise, to be applied effectively to seismic data.

2. Blind Noise Estimation

Rather than requiring manual input for noise variance, the algorithm estimates it "blindly." By identifying a 20x20 "quiet" patch within the seismic image—an area devoid of signal—the system calculates noise statistics directly from the data.

3. Curvelet Denoising

The whitened data is then processed in the curvelet domain. Curvelets are particularly effective for seismic data because they are localized in time, frequency, and orientation.

The curvelet transform partitions the frequency domain into angular wedges, allowing for optimal representation of curved seismic events.

As illustrated in Figure 2, the curvelet transform partitions the frequency plane into angular wedges. This multi-scale and multi-directional decomposition allows the algorithm to apply hard-thresholding to zero out coefficients associated with noise while preserving the high-magnitude coefficients that represent geological structures.

4. Inverse Whitening

Finally, the denoised data is transformed back to the spatial domain, and the whitening process is reversed to restore the original amplitude and character of the seismic reflections.

Experimental Results

The framework was validated using both synthetic and field datasets.

Pseudo-Synthetic Data: Using a clean dataset from the Netherlands offshore F3 block, the authors added synthetic Brownian noise ($\alpha=2$). The proposed method achieved a remarkable improvement in Peak Signal-to-Noise Ratio (PSNR), rising from -4.93 dB to 17.64 dB. Notably, without the pre-whitening step, the standard curvelet transform failed to improve the PSNR at all.

Field Data: In a "stress test" using 2D land line data from East Texas, USA, the method successfully suppressed heavy ground roll noise. The proposed approach reached a PSNR of 22.26 dB, significantly outperforming Wavelet decomposition (11.36 dB) and Empirical Mode Decomposition (10.52 dB). Qualitative analysis confirmed that the method preserved faint, steeply dipping reflections that are often lost during conventional processing.

Conclusion

The integration of ZCA-based pre-whitening with the curvelet transform provides a robust solution for seismic denoising. By simplifying the statistical nature of the noise and automating variance estimation, the "blind" framework removes the need for manual tuning. While the computational cost of curvelets remains a consideration for large 3D volumes, the significant gains in signal clarity and geological fidelity make this a promising approach for modern seismic interpretation.

Summary updated

BibTeX
@inproceedings{iqbal2023blind,
  author = {Naveed Iqbal and Mohamed Deriche and Ghassan AlRegib and Sikandar Khan},
  title = {{Blind Curvelet-Based Denoising of Seismic Surveys in Coherent and Incoherent Noise Environments}},
  booktitle = {Arabian},
  year = {2023},
  doi = {10.1007/s13369-023-07836-3},
}
Figure