基于结构保护去噪神经网络的地震数据随机噪声压制  

Seismic random noise suppression based on SP-DnCNN neural network

作　　者：赵振聪 饶莹 ZHAO ZhenCong;RAO Ying(College of Geophysics,China University of Petroleum(Beijing),Beijing 102249,China;Centre for Reservoir Geophysics,Resource Geophysics Academy,Imperial College London,London SW72AZ,UK)

机构地区：[1]中国石油大学(北京)地球物理学院,北京102249 [2]帝国理工学院资源地球物理学院,油藏地球物理中心,伦敦SW72AZ

出　　处：《地球物理学报》2024年第10期3841-3850,共10页Chinese Journal of Geophysics

基　　金：国家自然科学基金(42055402,42404127);中国石油大学(北京)科研基金(2462021QNXZ005)资助。

摘　　要：随机噪声压制是地震数据处理中的重要环节,直接关系到后续地震资料处理和解释质量.与相干噪声不同,随机噪声具有频谱较宽、规律性差等特点,很难利用常规方法区分随机噪声与地震数据有效信号.相较于传统的基于稀疏域的噪声压制方法,基于神经网络深度学习的随机噪声压制方法具有自动化程度高的特点.现有的网络结构在压制地震数据随机噪声时,不可避免地破坏地震数据结构,尤其是对于含有复杂构造的地震数据.基于以上问题,本文在充分探索地震数据随机噪声特征的基础上,利用结构保护的深层卷积神经网络对野外采集的含噪地震数据进行随机噪声压制研究.研究中利用模拟数据与真实去噪地震数据作为标签数据对,结构保护的深层卷积神经网络可以学习含噪和去噪地震数据之间的内在特征联系.考虑到常规神经网络在地震数据随机噪声压制过程中不能有效保护地下复杂结构特征,文中采用地震数据局部倾角作为约束,并通过修改目标函数达到保护地震数据结构特征的目的.文中利用模拟数据与野外地震数据,对本文方法的噪声压制能力与常用的包括中值滤波、多道奇异谱分解方法在内的地震数据随机噪声压制方法进行了分析对比.数值结果表明,深层卷积神经网络可以有效压制地震数据中的随机噪声,地震倾角的加入可以有效保护地震数据中的复杂构造特征.Seismic random noise suppression has been a persistent challenge in seismic data processing. It can directly determine the quality of successive seismic data processing and interpretation. Unlike coherent noise, random noise carries the characteristic of wide frequency band, which makes it hard to be separated from effective signals in the seismic data. To solve the above-mentioned problems, we propose to perform random noise suppression with the help of deep convolutional neural networks. Compared with conventional random noise suppression methods, methods based on neural network are more efficient and automatic. However, random noise suppression with existing deep learning algorithms ignores the existence of complex structure, resulting in energy loss of effective signals. In this manuscript, we propose a novel random noise suppression network, which is referred to as structure-preserving deep denoising convolutional neural network. The training labels in the manuscript are denoised field and synthetic seismic data, the features of which are used to train the neural networks. And considering the drawback of conventional neural networks in preserving the structure of seismic signals, we introduce local seismic dip information into the network and incorporate local seismic dip information into the objective function. Numerical and field seismic data applications indicate that the neural network can effectively suppress random noise in seismic data. And structure of seismic data is preserved with the introduction of local dipping information.

关 键 词：叠后 随机噪声压制 神经网络 局部地震倾角 结构保护 

分 类 号：P631[天文地球—地质矿产勘探]