基于深度学习的子波整形反褶积方法  

Wavelet shaping deconvolution based on deep learning

作　　者：倪文军 刘少勇[1] 王丽萍[1] 韩冰凯 盛燊 Ni Wenjun;Liu Shaoyong;Wang Liping;Han Bingkai;Sheng Shen(School of Geophysics and Geomatics,China University of Geosciences,Wuhan,Hubei 430074,China;School of Ocean and Earth Science,Tongji University,Shanghai 200092,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,湖北武汉430074 [2]同济大学海洋与地球科学学院,上海200092

出　　处：《石油地球物理勘探》2023年第6期1313-1321,共9页Oil Geophysical Prospecting

基　　金：国家自然科学基金项目“基于非平稳滤波算子的最小二乘反射系数估计及宽带波阻抗成像”(41974125);中石化地球物理重点实验室项目“数据驱动的像域地震数据高保真高分辨处理”(36750000-23-FW0399-0003)资助。

摘　　要：地震数据偏移成像是地下介质反射系数估计的重要方法之一,其结果通常受子波影响而波数带展布有限。有效拓展成像结果的波数带、提高空间分辨率是宽带反射系数估计的一个重要目的。为此,首先从反演成像的角度分析,指出子波和观测系统照明是影响成像结果分辨率的两个主要因素;其次,基于卷积神经网络(CNN),利用宽频子波构建标签,将常规成像结果作为输入,利用CNN挖掘其中的映射关系,提出了相应的深度学习算法子波整形反褶积方法;然后,针对反褶积中初始子波估计不准确的问题,设计了子波与反射系数串联、迭代、更新的实现方案,定制的宽频子波能兼顾低波数和高波数信息,用于训练网络时可以更好地恢复宽带的反射系数;最后,利用已知模型进行网络的预训练,将基于目标数据体提取的有效子波作为靶区数据反褶积的初始子波,进行子波整形反褶积处理,并通过薄层模型测试了该方法的正确性和可靠性。实际资料处理结果表明了该方法具有较好的应用潜力。Seismic data migration imaging is one of the important methods for estimating the reflectivity of under-ground media.However,the imaging results are often affected by the wavelet,with limited wavenumber band distri-bution.Effectively extending the wavenumber band of the imaging results to improve the spatial resolution is a key objective in broadband reflectivity estimation.To achieve this,we firstly point out that the wavelet and the illumina-tion of the geometry system are two important factors that affect the resolution of imaging results from an inversion imaging perspective.Then,based on convolutional neural networks(CNN),we use broadband wavelets to construct labels and employ conventional imaging results as input features to explore the mapping relationship using CNN.We also develop a corresponding deep learning algorithm,namely the wavelet shaping deconvolution method,and design a solution to the problem of inaccurate initial wavelet estimation in deconvolution by concatenating,iterating,and up-dating wavelets and reflectivity.Customized broadband wavelets can take into account both low-wavenumber and high-wavenumber information and can better restore broadband reflectivity during network training.Finally,we use a known model for network pre-training,extract effective wavelets based on the target data as the initial wavelets for deconvolution of the target data,carry out wavelet shaping deconvolution processing,and test the correctness and re-liability of the method through thin-layer model testing.The filed data processing results indicate that this method has great potential for practical applications.

关 键 词：图像反褶积 子波 卷积神经网络 深度学习 

分 类 号：P631[天文地球—地质矿产勘探]