基于迭代数据增强的自监督叠前地震随机噪声压制方法  

Random noise suppression for pre-stack seismic data based on self-supervised learning via iterative data refinement

作　　者：石战战 黄果[1] 陈庆利[1] 庞溯[3,4] 王元君 SHI ZhanZhan;HUANG Guo;CHEN QingLi;PANG Su;WANG YuanJun(School of Artificial Intelligence,Leshan Normal University,Leshan 614000,China;School of Land and Resources,China West Normal University,Nanchong 637002,China;The Engineering and Technical College,Chengdu University of Technology,Leshan 614000,China;College of Geophysics,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]乐山师范学院人工智能学院,乐山614000 [2]西华师范大学国土资源学院,南充637002 [3]成都理工大学工程技术学院,乐山614000 [4]成都理工大学地球物理学院,成都610059

出　　处：《地球物理学进展》2024年第5期1824-1837,共14页Progress in Geophysics

基　　金：国家科技重大专项课题子课题(2016ZX05026-001-005);四川省教育厅项目(16ZB0410);川西南空间效应探测与应用四川省高等学校重点实验室开放基金(YBXM202102001)联合资助。

摘　　要：构造含噪声-纯净训练样本集是制约深度学习去噪的主要瓶颈之一.提出一种基于迭代数据增强的自监督叠前地震随机噪声压制方法,仅利用含噪声样本训练深度神经网络.该方法首先利用多元回归算法估计共偏移距道集随机噪声,再与含噪声样本叠加构造强噪声样本.以skip网络为模型,每次迭代分为2步:(1)以强噪声样本为输入,由上一次迭代优化模型预测弱噪声样本;(2)构造新的强噪声-弱噪声样本集,以监督学习策略优化网络模型.算法的优势为:(1)利用共偏移距道集平缓同相轴特征,能有效估计与实际噪声近似同分布的噪声样本;(2)随着迭代次数增加,弱噪声样本更加接近实际纯净样本,去噪结果接近监督学习;(3)迭代数据增强策略实现了数据增广,增加样本数量,避免过拟合.模拟和实际地震数据试算结果表明,所提算法具有较好的应用效果.The requirement of a large amount of noisy-clean training dataset is one of the main bottlenecks restricting deep learning denoising.A self-supervised pre-stack seismic random noise suppression method based on iterative data refinement is proposed,which uses only noise samples to train the deep neural networks.The method firstly uses the multiple regression theory to estimate the random noise in the common offset gathers,and then superimposes the noise components to the noisy samples to construct the strong noise samples.Taking the skip network as the model,each iteration is divided into two steps:(1)Taking the strong noise samples as input,the weak noise samples are predicted by the last iteration optimization model;(2)Constructing the nosier-noise training dataset,the network model is optimized with a supervised learning strategy.The advantages of the algorithm are as follows:(1)The noise samples that are drawn from the distribution,which is approximately the same as the actual noise,are estimated by using the characteristics of the flat events of the common offset gathers;(2)With the increase of the number of iterations,the predicted weak noise samples are similar to the actual clean samples,it is feasible to learn a self-supervised network approximating the optimal parameters of a supervised model;(3)The iterative data refinement strategy achieves data augmentation,increases the number of samples,and avoids overfitting.Experiments on synthetic and realistic noise removal demonstrate that the iterative data refinement approach achieves state-of-the-art performance.

关 键 词：迭代数据增强 自监督学习 随机噪声压制 共偏移距道集 

分 类 号：P631[天文地球—地质矿产勘探]