基于半监督V-Net和迁移学习的三维地震波阻抗反演  

3D seismic acoustic impedance inversion based on semi-supervised V-Net and transfer learning

作　　者：杨雨 汪玲玲 董志明 朱振宇 YANG Yu;WANG LingLing;DONG ZhiMing;ZHU ZhenYu(School of Geophysics and Geomatics,China University of Geosciences(Wuhan),Wuhan 430074,China;Research Institute of CNOOC,Beijing 100028,China)

机构地区：[1]中国地质大学(武汉)地球物理与空间信息学院,武汉430074 [2]中海油研究总院,北京100028

出　　处：《地球物理学报》2025年第2期730-745,共16页Chinese Journal of Geophysics

基　　金：国家重点研发计划(2020YFA0713400);国家自然科学基金项目(41874154)资助.

摘　　要：近年来,深度学习算法因其优越的非线性表征能力,在地球物理反演领域受到广泛关注.本文针对实际应用中可用于制作网络训练标签的测井资料量较少,以及以深度学习算法为基础的一维和二维地震波阻抗反演方法在处理三维地震数据时可能存在的空间连续性问题,提出一种基于半监督V-Net和迁移学习的三维地震波阻抗反演方法.该半监督V-Net网络由一个反演网络和一个正演网络组成.反演网络是一个具有三维卷积核和分段残差块连接的多分辨率V-Net网络,正演网络由五个三维卷积层构成.首先用具有较复杂地层结构的模型合成数据对反演网络进行预训练,接着用实际测井数据插值得到的井旁数据对反演网络进行迁移学习微调,使其更适用于目标实际数据,从而完成反演网络的初始化.正演网络也用井旁插值数据进行预训练初始化.然后在波阻抗低频模型的基础上,用测井标签和地震数据半监督训练整个神经网络.最后,把三维地震数据输入到训练好的反演网络,即可预测得到相应的三维波阻抗.模型和实际数据算例表明,与一维半监督卷积-门控循环神经网络(CNN-GRU)、二维半监督U-Net以及未使用迁移学习的半监督V-Net方法相比,本文方法反演得到的波阻抗的空间连续性很好、精度最高,且本文方法对噪声具有较好的鲁棒性.In recent years,deep learning algorithm has been widely concerned in the field of geophysical inversion due to its superior nonlinear representation ability.To address the limited amount of logging data that can be used to produce network training labels in practice,as well as the potential spatial continuity issues of 1D and 2D seismic acoustic impedance inversion methods based on deep learning algorithms when processing 3D seismic data,we propose a 3D seismic acoustic impedance inversion method based on semi-supervised V-Net and transfer learning.The semi-supervised V-Net network consists of an inverse network and a forward network.The inversion network is a multi-resolution V-Net with 3D convolutional kernels and piece-wise residual blocks,and the forward network is composed of five 3D convolutional layers.Firstly,the inversion network is pre-trained using synthetic data with relatively complex geological structures.Then,the well side data obtained by interpolating the logging data of the target data is used to fine tune the inversion network by transfer learning,making it more suitable for the target data,thus completing the initialization of the inversion network.The forward network is also pre-trained and initialized using well side interpolation data.Then,based on the low-frequency acoustic impedance model,the semi-supervised V-Net is trained using the logging labels and seismic data.Finally,the corresponding 3D acoustic impedance can be predicted by inputting 3D seismic data into the trained inversion network.Experiments on synthetic and field datasets demonstrate that compared with 1D semi-supervised CNN-GRU,2D semi-supervised U-Net,and semi-supervised V-Net method without transfer learning,the acoustic impedance obtained by the proposed method has good spatial continuity and the best prediction accuracy,and the proposed method has good robustness to noise.

关 键 词：三维 地震波阻抗反演 V-Net 深度学习 迁移学习 

分 类 号：P631[天文地球—地质矿产勘探]