Abstract

In recent years, supervised-deep-learning methods have shown some advantages over conventional methods in seismic data denoising, such as higher signal-to-noise ratio after denoising, complete separation of signals and noise in shared frequency bands and intelligent denoising without artificial parameter tuning. However, the lack of real noise data matched with raw seismic data has greatly limited its further application. In this paper, we take the surface seismic shot gather as an example to explore the corresponding solutions and propose a novel supervised-deep-learning method with weak dependence on real noise data based on the data augmentation of a generative adversarial network. We utilize the generative adversarial network to augment the pre-arrival noise data acquired from the shot gather itself, thereby obtaining a large amount of synthetic noise data whose probability distribution is extremely similar to that of the real noise in shot gather; the augmented synthetic noise data and sufficient synthetic signal data obtained by forward modeling together form the augmented training dataset. Meanwhile, the dilated convolution and gradual denoising strategy are adopted to construct the basic architecture of denoising convolution neural network. Finally, the above augmented dataset is used to train the network, so as to establish a nonlinear and complex mapping relationship between raw seismic data and desired signals. Both synthetic and real experiments demonstrate that our method can realize the intelligent denoising of different common-shot-point records in shot gather with the help of limited pre-arrival noise data.

Article Highlights

• We introduce the data augmentation strategy into the field of deep-learning-based seismic denoising, thereby alleviating the dependence of supervised-deep-learning methods on real noise data

• We propose a novel denoising network architecture with strong recovery ability for weak desired signals by using the gradual denoising strategy and dilated convolution

• The augmented synthetic noise data can meet the requirement of supervised-deep-learning methods on the quantity and authenticity of training data, so this data augmentation strategy by using the Generative Adversarial Net (GAN) is a solution to the lack of real noise data

中文翻译：

---

使用对真实噪声数据弱依赖的监督深度学习方法进行地震炮集去噪：解决真实噪声数据缺乏的解决方案

摘要

近年来，有监督的深度学习方法在地震数据去噪方面表现出优于传统方法的一些优势，如去噪后的信噪比更高、共享频段的信噪比完全分离、无需人工参数的智能去噪等。调音。然而，由于缺乏与原始地震数据相匹配的真实噪声数据，极大地限制了其进一步的应用。在本文中，我们以地表地震炮集为例，探索相应的解决方案，并提出了一种基于生成对抗网络数据增强的对真实噪声数据具有弱依赖性的监督深度学习方法。我们利用生成对抗网络来增强从镜头收集本身获取的到达前噪声数据，从而获得大量合成噪声数据，其概率分布与炮集中真实噪声的概率分布极为相似；通过前向建模获得的增强的合成噪声数据和足够的合成信号数据共同形成增强的训练数据集。同时，采用空洞卷积和渐进去噪策略构建去噪卷积神经网络的基本架构。最后，利用上述增强数据集对网络进行训练，从而建立原始地震数据与期望信号之间的非线性复杂映射关系。综合实验和实际实验都表明，我们的方法可以在有限的到达前噪声数据的帮助下，实现炮集不同共炮点记录的智能去噪。

文章重点

• 我们将数据增强策略引入基于深度学习的地震去噪领域，从而减轻监督深度学习方法对真实噪声数据的依赖

• 我们通过使用渐进去噪策略和扩张卷积提出了一种新颖的去噪网络架构，该架构对弱期望信号具有很强的恢复能力

• 增强的合成噪声数据可以满足监督深度学习方法对训练数据的数量和真实性的要求，因此这种使用生成对抗网络（GAN）的数据增强策略是解决真实噪声数据缺乏的解决方案