The potential of self-supervised networks for random noise suppression in seismic data

Highlights
1. Application of self-supervised deep learning for random noise suppression.
2. Blind-spot networks shown to effectively reduce white and coloured noise.
3. SNR improvements seen in the image, frequency and acoustic impedance domains.
4. Network successfully trained and applied to post-stack field data.
5. Blind-spot approach comparable to commonly used noise suppression procedures.

Abstract

Noise suppression is an essential step in many seismic processing workflows. A portion of this noise, particularly in land datasets, presents itself as random noise. In recent years, neural networks have been successfully used to denoise seismic data in a supervised fashion. However, supervised learning always comes with the often unachievable requirement of having noisy-clean data pairs for training. Using blind-spot networks, we redefine the denoising task as a self-supervised procedure where the network uses the surrounding noisy samples to estimate the noise-free value of a central sample. Based on the assumption that noise is statistically independent between samples, the network struggles to predict the noise component of the sample due to its randomicity, whilst the signal component is accurately predicted due to its spatio-temporal coherency.

Illustrated on synthetic examples, the blind-spot network is shown to be an efficient denoiser of seismic data contaminated by random noise with minimal damage to the signal; therefore, providing improvements in both the image domain and down-the-line tasks, such as post-stack inversion. To conclude our study, the suggested approach is applied to field data and the results are compared with two commonly used random denoising techniques: FX-deconvolution and sparsity-promoting inversion by Curvelet transform. By demonstrating that blind-spot networks are an efficient suppressor of random noise, we believe this is just the beginning of utilising self-supervised learning in seismic applications.

Theory
N2V works by replacing a set of non-adjacent pixels in an image with randomly selected pixels that pertain to the receptive field of the chosen network. The corrupted image becomes the input to a standard NN architecture (in our case a U-Net was used), whilst the original noisy image is used as the target. As opposed to standard NN image processing tasks, the loss function is not computed on every pixel in the input data, instead it is only evaluated for the blind-spotted pixels, i.e., those that were corrupted in the input image. Whilst originally proposed for the suppression of i.i.d. noise in natural images and microscopy data, in this work we show that N2V can be adapted into an efficient suppressor of band-passed random noise in seismic data, i.e. data that has some degree of correlation along the time axis. This however requires a careful consideration of the algorithm's hyper-parameters and their meaning with respect to the seismic scenario.

Results
Under a White, Gaussian Noise scenario, the self-supervised procedure is shown to efficiently suppress almost all noise with minimal signal leakage. For the scenario where the data is contaminated by bandpassed noise, the majority of the noise can be suppressed however not quite to the performance of the WGN case, given there now exists some spatial correlation in the noise. Nevertheless, the product of an acoustic impedance inversion shows the substantial gains available by the denoising procedure. Finally, the N2V workflow is applied to a field dataset with its performance benchmarked against two common suppression procedures: FK deconvolution and curvelets. The N2V procedure is shown to be that which suppresses the most noise with minimal signal leakage and introduction of artefacts.

Citation

If you found the paper useful, please cite it via:

                  
Birnie, C., Ravasi, M., Liu, S. and Alkhalifah, T., 2021. The potential of
self-supervised networks for random noise suppression in seismic data. Artificial
Intelligence in Geosciences, 2, pp.47-59


                  
@article{birnie2021potential,
title={The potential of self-supervised networks for random noise suppression in seismic data},
author={Birnie, Claire and Ravasi, Matteo and Liu, Sixiu and Alkhalifah, Tariq},
journal={Artificial Intelligence in Geosciences},
volume={2},
pages={47--59},
year={2021},
publisher={Elsevier}
}