- The paper proposes using self-supervised blind-spot networks for seismic noise suppression, which uniquely eliminates the requirement for clean training data pairs typical of supervised methods.
- Numerical experiments on synthetic and field data demonstrate the network's capability to substantially reduce noise while preserving crucial signal integrity, outperforming some traditional techniques.
- This approach improves practical seismic data processing by maintaining more details for enhanced interpretability and suggests future work on correlated noise and structured noise suppression.
Overview of the Self-Supervised Networks for Random Noise Suppression in Seismic Data
The paper presented in "The potential of self-supervised networks for random noise suppression in seismic data" investigates the application of blind-spot networks for seismic noise suppression with a focus on overcoming the limitations of supervised learning methodologies that require noisy-clean pairs for training. This necessity is often impractical in seismic applications due to the complexity and randomness of noise in real-world datasets.
The pivotal innovation in this paper is redefining seismic denoising as a self-supervised procedure. The approach utilizes blind-spot networks whereby the network employs surrounding noisy samples to estimate a central sample's noise-free value by leveraging the uncorrelated nature of noise and the predictable characteristics of the seismic signal. This relies on the assumption that noise has statistical independence between samples, which challenges the network in differentiating noise due to its unpredictability while successfully predicting the coherent seismic signal.
Several key contributions of the paper commence with a theoretical foundation of blind-spot networks as per the Noise2Void (N2V) framework. This framework configures a neural network to predict a pixel's value using neighboring pixel data while corrupting certain pixels in the training phase to ensure the focus remains on signal prediction and not naive noise replication. The model presented uses a simplified UNet architecture to accommodate the seismic data's intricacies, followed by an explanation of performance metrics such as Peak Signal-to-Noise Ratio (PSNR), frequency correlation, and inversion PSNR to evaluate the denoising quality comprehensively.
Numerical experiments conducted with both synthetic datasets contaminated with white Gaussian noise and band-pass filtered noise demonstrate substantial denoising capability while maintaining signal integrity. Synthetic tests reveal the network's efficacy in reducing noise with minimal signal damage, marked by considerable improvements in image and inversion domains. A detailed hyperparameter search reveals the importance of configurations like a reduced number of epochs and augmented active pixel percentages to optimize training without compromising the network’s performance on correlated noise.
The application to a field dataset contaminated by noise from land acquisition shows that self-supervised networks can successfully suppress random noise compared to established techniques such as FX-deconvolution and Curvelet transform. Observations conclude that the self-supervised approach maintains more details than overly smoothed outputs from other methods, promising enhanced interpretability and subsurface inversion results.
The paper concludes with a discussion on the broader implications of blind-spot networks in seismic applications. It proposes future developments to address noise correlation within the dataset and potential extensions to structured noise suppression. By eliminating the constraint of having to obtain clean reference data, this approach leverages the ability to predict signal over noise, offering a significant advantage in practical seismic data processing tasks—especially where traditional techniques falter due to the absence of clean training datasets.
In summary, the use of self-supervised blind-spot networks provides a sophisticated alternative to conventional noise suppression techniques, offering improved denoising outcomes and potential extensions across seismic workflows. While these networks might not rival supervised methods equipped with extensive training data, their application to seismic data represents an innovative step towards harnessing machine learning for efficient data processing in challenging environments. Future work could aim at refining noise models and exploring structured noise suppression to further advance the approach in denoising complex seismic data.