基于标准化流的地震偏移噪声压制和不确定性量化  

Noise suppression and uncertainty quantification in seismic imaging using normalizing flows

作　　者：孙昌潇 毛伟建 石星辰 SUN Changxiao;MAO Weijian;SHI Xingchen(Center for Computational and Exploration Geophysics,Innovation Academy for Precision Measurement Science and Technology,Chinese Academy of Sciences,Wuhan 430077,China;State Key Laboratory of Geodesy and Earth s Dynamics,Wuhan 430077,China;University of Chinese Academy of Sciences,Beijing 100049,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,武汉430077 [2]大地测量与地球动力学国家重点实验室,武汉430077 [3]中国科学院大学,北京100049

出　　处：《世界地质》2024年第4期541-550,共10页World Geology

基　　金：国家自然科学基金重点项目(42130808)。

摘　　要：为了定量评估地震成像结果的不确定性,同时减少成像伪影、提高成像分辨率,笔者采用了机器学习领域中的一种可逆神经网络—标准化流,实现高精度成像,并进行成像结果的后验估计。通过对估计的后验分布进行采样及其统计分析,笔者量化了成像过程中存在的不确定性。与传统的抽样方法,例如马尔科夫链蒙特卡罗相比,标准化流法能够直接预测后验分布,而不是依赖于随机采样,因此在计算效率和适应性方面展现出显著的优势。笔者将该方法应用于稀疏数据成像的去噪与增强,并在合成数据上进行了测试。测试结果表明,稀疏海底节点采集数据的偏移成像结果得到了改善,成像噪声减少,同相轴的连续性以及深度照明得到了加强。另外,不确定性量化通过统计后验成像结果生成。To quantify the uncertainty of seismic imaging,reduce imaging artifacts and improve resolution,the authors use a technique known as normalizing flows,which employs invertible neural networks,to both enhance the images and estimate the posterior distribution of imaging results.By sampling from the estimated posterior distribution and conducting statistical analysis on these samples,the authors perform uncertainty quantification for seismic imaging.In comparison to traditional sampling methods like Markov Chain Monte Carlo,the normalizing flows method allows to predict the posterior distribution directly,instead of relying on random sampling.Therefore,it demonstrates significant advantages in terms of computational efficiency and adaptability.The authors apply this method to denoise and enhance imaging results from sparse acquisition.The tests on synthetic data show improvements in the migration imaging results of the sparse ocean bottom node acquisition data,with reduced noise,enhanced event continuity and improved depth illumination.Furthermore,by statistically analyzing the posteriors,the authors provide uncertainty quantification in the imaging results.

关 键 词：贝叶斯推断 标准化流 不确定性量化 地震成像 

分 类 号：P631.4[天文地球—地质矿产勘探]