应用STGCN时空建模的地震波阻抗反演方法  

Seismic impedance inversion method based on STGCN spatio-temporal modeling

作　　者：王泽峰 赵海波[4] 杨懋新 王团[4] 许辉群[6] 毛伟建 WANG Zefeng;ZHAO Haibo;YANG Maoxin;WANG Tuan;XU Huiqun;MAO Weijian(Research Center for Computational and Exploration Geophysics,Innovation Academy for Precision Measurement Science and Technology,Chinese Academy of Sciences,Wuhan,Hubei 430077,China;State Key Laboratory of Geodesy and Earth’s Dynamics,Wuhan,Hubei 430077,China;University of Chinese Academy of Sciences,Beijing 100049,China;Exploration and Development Research Institute,Daqing Oilfield Company Ltd.,PetroChina,Daqing,Heilongjiang 163712,China;Daqing Branch,Geophysical Research Institute,BGP Inc.,Daqing,Heilongjiang 173712,China;College of Geophysics and Petroleum Resources,Yangtze University,Wuhan,Hubei 430100,China)

机构地区：[1]中国科学院精密测量科学与技术创新研究院计算与勘探地球物理研究中心,湖北武汉430077 [2]大地测量与地球动力学国家重点实验室,湖北武汉430077 [3]中国科学院大学,北京100049 [4]大庆油田有限责任公司勘探开发研究院,黑龙江大庆163712 [5]东方地球物理公司研究院大庆物探研究院,黑龙江大庆173712 [6]长江大学地球物理与石油资源学院,湖北武汉430100

出　　处：《石油地球物理勘探》2025年第1期43-53,共11页Oil Geophysical Prospecting

基　　金：国家自然科学基金项目“海底四分量高精度地震偏移成像方法研究”(42130808)资助。

摘　　要：现今,深度学习地震波阻抗反演方法通常是通过低维度的时序建模,忽略了空间构造拓扑结构信息,导致反演精度较低。针对此问题,提出了一种基于STGCN(时空图卷积神经网络)时空建模的地震波阻抗反演方法。该方法考虑到地震数据的空间构造拓扑结构及互相关性,使用马氏距离对地震数据进行空间邻近度的加权处理建立邻接矩阵;进一步通过切比雪夫多项式扩大空间感受野的同时减少参数量,高效地提取地震数据的空间构造特征,同时利用门控循环单元捕获其时序相关性;最后构建时空图卷积单元实现基于STGCN的地震数据与波阻抗在时间和空间两个维度的映射。模型测试及实际资料反演结果表明,该方法在提高反演精度的同时对噪声具有一定的适应性,并可以很好的体现地层的横向变化。Nowadays,the seismic impedance inversion methods based on deep learning usually realize inversion through low‑dimensional time series modeling,ignoring the topological structure information of the spatial structure of seismic impedance and thus resulting in low inversion accuracy.To solve this problem,this paper proposes a seismic impedance inversion method based on spatio‑temporal modeling with spatio‑temporal graph convolutional networks(STGCN).Considering the topological structure and mutual correlation of seismic data,the method uses Mahalanobis distance to weight the spatial proximity of seismic data so that an adjacency matrix can be established.Furthermore,the Chebyshev polynomial is used to enlarge the spatial receptive field and reduce the number of parameters.The spatial structure features of seismic data are extracted efficiently,and the temporal correlation is captured by a gated recurrent unit.Finally,a spatio‑temporal graph convolution unit is constructed for the mapping of seismic data and wave impedance in both time and space based on STGCN.The results of the model test and actual data inversion show that the proposed method improves inversion accuracy,has certain adaptability to noise,and can well reflect the lateral changes of strata.

关 键 词：地震波阻抗反演 深度学习 时空建模 时空图卷积神经网络 

分 类 号：P631[天文地球—地质矿产勘探]