基于TransUNet神经网络的叠后地震波阻抗反演方法  

Post-stack seismic impedance inversion method based on TransUNet neural network

作　　者：彭真 许辉群[1] PENG Zhen;XU HuiQun(School of Geophysics and Petroleum Resources,Yangtze University,Wuhan 430100,China)

机构地区：[1]长江大学地球物理与石油资源学院,武汉430100

出　　处：《地球物理学进展》2024年第2期704-715,共12页Progress in Geophysics

基　　金：中国石油天然气股份有限公司勘探开发研究院地球物理重点实验室开放基金资助课题“考虑标签与网络适应性的高分辨率地震波阻抗反演技术”(2022-KFKT-25)资助。

摘　　要：卷积网络是深度学习地震波阻抗反演方法的主流框架,然而卷积网络存在捕捉数据长期依赖能力的不足,导致针对具有长期任务特性的地震反演结果会受到较大的影响.鉴于Transformer网络能够关注地震数据全局特征的优势,它不仅可以提供地震数据特征的位置信息,还能弥补卷积对地震数据全局信息表征的不足.因此,笔者基于Transformer网络和UNet网络构建了一种既能刻画地震数据的局部细节,又能表征地震数据全局特征的网络结构,也就是在UNet框架内嵌入Transformer而形成的一种网络作为地震波阻抗反演的反演映射网络(TransUNet).TransUNet网络的优势在于它既利用了UNet可提取地震数据特征的功能,又利用Transformer对上述特征位置进行编码的作用,从而使得TransUNet具有捕获地震数据的全局信息能力,为地震数据和波阻抗之间的映射关系提供了一种新的方法和思路,且在模型和实际资料中得到有效验证.Convolutional networks are the mainstream framework for deep learning seismic impedance inversion method;however,convolutional networks have the deficiency of capturing the long-term data dependence ability,which leads to the inversion results for seismic with long-term task characteristics will be greatly affected.Given the advantages of Transformer's ability to focus on global features of seismic,it can not only provide the location information of seismic data features,but also compensate the deficiency of convolutional for global information characterization of seismic data.Therefore,the author constructs a network structure based on Transformer and UNet networks that can not only portray the local details of seismic data but also characterize the global features of seismic data,that is,a network formed by embedding Transformer within the UNet framework as an inverse mapping network for seismic impedance inversion(TransUNet).The advantage of TransUNet is that it utilizes both the function of UNet that can extract the features of seismic data and the role of Transformer to encode the location of the above features,thus making TransUNet have the ability to capture the global information of seismic data and provide a new method and idea for the mapping relationship between seismic data and impedance,and it has been effectively verified in model and actual It provides a new method and idea for the mapping relationship between seismic data and impedance,and is effectively verified in models and real data.

关 键 词：地震波阻抗反演 TRANSFORMER UNet TransUNet 

分 类 号：P631[天文地球—地质矿产勘探]