Seismic Data Processing and Denoising Techniques

Seismic data processing encompasses a sequence of computational steps designed to transform raw wavefield recordings into accurate images of subsurface structure and properties. Initial stages include trace editing, deconvolution and deghosting to correct for instrument and source effects, followed by noise attenuation to remove both coherent disturbances (such as ground roll and multiples) and random noise. Traditional methods exploit transform domains (Fourier, f-k, wavelet and curvelet) and multichannel coherency to separate signal from noise. More recently, rank-reduction techniques decompose multi-dimensional data matrices into low-rank signal and high-rank noise subspaces, enabling simultaneous reconstruction of missing traces and denoising. Parallel advances in machine learning have led to plug-and-play frameworks and convolutional neural networks (CNNs) that learn statistical features of seismic signals, facilitating interpolation and attenuation without explicit physical models. Self-supervised and unsupervised architectures further relax the need for paired clean data, instead exploiting spatial and temporal coherency to predict clean waveforms. Together, these developments enhance imaging resolution, improve reservoir characterisation and support global seismic hazard monitoring.

Research from Nature Portfolio

Recent studies have shown an iterative rank-reduction approach that reconstructs missing spatial traces while suppressing noise in dense array recordings, enhancing tomographic P-wave imaging through principal component analysis of three-dimensional wavefields. Another investigation applied encoder–decoder convolutional networks to transform incomplete to complete seismic datasets, demonstrating superior reconstruction of both irregularly and regularly missing traces via a plug-and-play architecture that obviates problem-specific retraining.

Research from all publishers

One approach integrates denoising CNNs trained on natural images into a project-onto-convex-set (POCS) framework for seismic interpolation, achieving high signal-to-noise recovery and faithful reconstruction of weak features without bespoke seismic labels. Unsupervised deep CNNs have been developed that learn noise characteristics directly from raw data via blind-spot training, effectively attenuating random noise while preserving coherent events. Self-supervised blind-spot networks further exploit the statistical independence of noise to suppress random disturbances in both synthetic and field datasets, improving image quality and downstream inversion tasks without access to clean reference data.

Seismic Data Processing and Denoising Techniques publication trend

The graph below shows the total number of articles in seismic data processing and denoising techniques across all publications each year (not limited to Nature Index journals).

Technical terms

Signal-to-noise ratio (SNR): measure of signal amplitude relative to background noise.

Rank-reduction method: decomposition of data matrices into low-rank signal and high-rank noise subspaces via singular value truncation.

Convolutional neural network (CNN): hierarchical deep learning model that extracts spatial features through convolutional layers.

Blind-spot network: self-supervised architecture that predicts clean signals by masking and reconstructing central samples using neighbouring data.

Project-onto-convex-set (POCS): iterative algorithm enforcing multiple constraints for signal recovery through successive projections.

References

1. Simultaneous denoising and reconstruction of 5-D seismic data via damped rank-reduction method. Geophysical Journal International (2016).
2. Deep Learning for Geophysics: Current and Future Trends. Reviews of Geophysics (2021).
3. Can learning from natural image denoising be used for seismic data interpolation?. Geophysics (2020).
4. Deep learning for irregularly and regularly missing data reconstruction. Scientific Reports (2020).
5. Unsupervised Seismic Random Noise Attenuation Based on Deep Convolutional Neural Network. IEEE Access (2019).
6. The potential of self-supervised networks for random noise suppression in seismic data. Artificial Intelligence in Geosciences (2021).