[Paper Review] An effective self-supervised learning method for various seismic noise attenuation

Shijun Cheng, Zhiyao Cheng+3 more|arXiv (Cornell University)|Nov 3, 2023

Seismic Imaging and Inversion Techniques4 citations

TL;DR

This paper proposes a self-supervised learning method for seismic noise attenuation that uses a two-phase training strategy—warm-up with synthetic noise augmentation and iterative data refinement (IDR)—to denoise real field data without requiring clean labels. The method achieves superior performance on field data compared to supervised learning baselines, especially under high-noise conditions, by progressively reducing the domain gap between noisy and clean data distributions through pseudo-labeling and iterative noise injection.

ABSTRACT

Faced with the scarcity of clean label data in real scenarios, seismic denoising methods based on supervised learning (SL) often encounter performance limitations. Specifically, when a model trained on synthetic data is directly applied to field data, its performance would drastically decline due to significant differences in feature distributions between the two. To address this challenge, we develop an effective self-supervised strategy. This strategy, while relying on a single denoising network model, adeptly attenuates various types of seismic noise. The strategy comprises two main phases: 1. The warm-up phase. By using prior knowledge or extracting information from real data, we introduce additional noise to the original noisy data, constructing a noisier data with intensified noise. This data serves as the input, with the original noisy data acting as pseudo-labels. This facilitates rapid pre-training of the network to capture a certain noise characteristics and boosts network stability, setting the stage for the subsequent phase. 2. Iterative data refinement (IDR) phase. During this phase, we use the predictions of the original noisy data from the network trained in the previous epoch as the pseudo-labels. We continue to add noise to the predictions, creating a new noisier-noisy dataset for the current epoch of network training. Through this iterative process, we progressively reduce the discrepancy between the original noisy data and the desired clean data. Ultimately, the network's predictions on the original noisy data become our denoised results. Validations under scenarios with random noise, backscattered noise, and blending noise reveal that our method not only matches the traditional SL techniques on synthetic data but significantly outperforms them on field data.

Motivation & Objective

• Address the performance drop of supervised deep learning models on real seismic field data due to domain shift between synthetic and real data.
• Develop a self-supervised learning framework that eliminates the need for clean labels in seismic denoising by leveraging pseudo-labels from network predictions.
• Improve denoising robustness across diverse noise types—random, backscattered, and blending noise—common in real-world seismic data.
• Minimize reliance on expert-tuned parameters by automating the denoising process through end-to-end neural network training with iterative refinement.
• Enable effective noise attenuation in scenarios where labeled clean data is unavailable or prohibitively expensive to obtain.

Proposed method

• In the warm-up phase, additional noise is injected into original noisy seismic data to create a 'noisier' input, with the original noisy data used as pseudo-labels for pre-training the network.
• The network is trained to map the noisier input to the original noisy data, enabling it to learn basic noise characteristics and stabilize before the iterative phase.
• In the iterative data refinement (IDR) phase, the network's predictions on the original noisy data from the previous epoch are used as pseudo-labels.
• A new noisier-noisy dataset is generated by adding noise to these predicted labels, forming the training target for the next epoch.
• This process is repeated across multiple epochs, progressively reducing the distributional gap between the noisier-noisy data and the ideal clean data.
• The final denoised output is the network's prediction on the original noisy data after full IDR training, with no need for ground-truth labels.

Experimental results

Research questions

• RQ1 Can a self-supervised learning approach effectively denoise seismic data without requiring clean labeled data?
• RQ2 How does iterative refinement of pseudo-labels improve generalization on real field data compared to supervised learning on synthetic data?
• RQ3 To what extent does the level of injected noise in the IDR process affect the final denoising performance?
• RQ4 Can the method generalize across diverse noise types, including random, backscattered, and blending noise, in real seismic data?
• RQ5 How does the method compare in performance to traditional supervised learning and curvelet-based denoising on real-world seismic datasets?

Key findings

• The proposed self-supervised method significantly outperforms supervised learning models trained on synthetic data when applied to real field seismic data, especially under high noise levels.
• The method achieves competitive denoising performance on synthetic data while demonstrating superior robustness and generalization on real data, indicating effective domain adaptation.
• The IDR phase effectively reduces the distributional shift between the training data and the desired clean data, leading to improved denoising quality over successive epochs.
• The performance of the method is sensitive to the noise level introduced during training; optimal results are achieved when the injected noise level matches the actual noise intensity in the data.
• The method minimizes signal leakage and preserves weak signals, as validated by both SNR and MAE metrics across multiple noise types.
• The approach remains effective even when noise is not easily extractable, provided prior knowledge or conventional denoising techniques are used to estimate noise characteristics for noise injection.