基于Swin-Transformer与生成对抗网络的地震随机噪声压制方法

目前深度学习类地震数据去噪方法大多基于卷积神经网络, 而此类方法受限于卷积核的局部操作, 缺少对地震数据全局特征的分析, 因而降低了去噪效果。另外, 以L₁, L₂损失函数为指标的网络模型容易出现过度平滑效应, 产生虚假同相轴以及虚高的峰值信噪比(PSNR)与结构相似性(SSIM)值。为此, 提出一种基于Swin-Transformer(Swin-T)和生成对抗网络的去噪方法(ST-GAN)。该方法以Swin-Transformer作为生成对抗网络中的生成网络对地震数据去噪, 判别网络基于卷积神经网络。Transformer的自注意力机制是全局操作, 可以有效提取地震数据的全局特征, 并能与卷积神经网络的局部操作互补, 提升网络模型的特征提取能力。而对抗损失则提升了网络模型的细节恢复能力, 有效避免因过度平滑效应产生的同相轴假象。将该方法应用于地震数据去噪, 并与现有去噪方法进行对比分析, 实验结果表明, 该方法具有更加优异的特征提取能力, 能在有效压制随机噪声的同时, 恢复和保留更多的细节信息, 提高了地震信号的信噪比。

Noise suppression using deep learning methods is mostly based on convolutional neural networks. The convolution operation using the convolution kernel extracts local features, instead of global features, of seismic data; thus, random noises could not be eliminated perfectly. In addition, L₁ and L₂ loss functions tend to generate an over-smoothed network model and consequent false events and erroneously high values of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). To address this issue, we develop a denoising method based on the Swin-Transformer and generative adversarial network (ST-GAN). The Swin-Transformer functions as the generative network in the GAN for denoising, and the discrimination network is based on a convolutional neural network. Global features of seismic data, which could be obtained owing to the self-attention mechanism of the Transformer, and local features derived from the convolutional neural network may complement each other for the better performance of the network model. The use of adversarial loss makes it possible to recover more details by applying the network model and mitigate artificial events caused by over-smoothing. The comparative analysis shows that our approach is superior to other denoising methods in feature extraction and signal-to-noise ratio because random noises are effectively reduced and meanwhile more details of seismic data are recovered and preserved.