基于Huber-U-Net网络的地震数据重建与去噪  

Seismic data reconstruction and de-noising based on Huber-U-Net network

作　　者：欧炳霖 曾同生[2] 柳天成 高好天 李钟晓 李振春[3] OU BingLin;ZENG TongSheng;LIU TianCheng;GAO HaoTian;LI ZhongXiao;LI ZhenChun(School of Electronic Information,Qingdao University,Qingdao 266071,China;Research Institute of Petroleum Exploration and Development,PetroChina,Beijing 100083,China;School of Geosciences,China,University of Petroleum(East China),Qingdao 266580,China)

机构地区：[1]青岛大学电子信息学院,青岛266071 [2]中国石油勘探开发研究院,北京100083 [3]中国石油大学(华东)地球科学与技术学院,青岛266580

出　　处：《地球物理学进展》2023年第6期2540-2552,共13页Progress in Geophysics

基　　金：中国博士后面上基金“基于闭环可解释FISTA-Net的地震多次波自适应相减方法”(2022M723127);中国石化地球物理重点实验室开放基金“模型驱动深度神经网络用于智能地震多次波压制”(33550006-22-FW0399-0020);国家超级计算天津中心“天河”青索计划-地球科学领域专项基金资助项目“基于可解释性闭环FISTA-Net的地震多次波自适应相减方法”;山东省高等学校青年创新团队发展计划(2022KJ141)联合资助。

摘　　要：实际地震数据通常存在缺失地震道、噪声,需对其进行重建与去噪处理.基于压缩感知理论的凸集投影(Projection onto Convex Sets, POCS)方法对重建误差施加Huber范数最小化约束,等价于对大重构误差(异常噪声)施加L1范数最小化约束、对小重构误差(高斯随机噪声)施加L_(2)范数最小化约束,可以对缺失含噪地震数据实现同时重建与去噪.但由于POCS方法只是一个单层的无监督学习方法,数学表征能力较弱,导致其重建效果较差且难以将噪声压制完全.基于深度学习的U-Net方法以L_(2)范数最小化约束为损失函数对含噪随机缺失地震数据进行重建与去噪,该方法假设重建误差满足高斯分布.因此,U-Net方法虽能有效压制高斯随机噪声却难以有效去除超高斯分布的异常强振幅噪声.本文创新性地将上述两种方法进行结合,使用Huber范数约束替代L_(2)范数约束作为U-Net网络的误差函数,建立Huber-U-Net网络.合成数据和实际数据的处理结果表明,Huber-U-Net方法相比于POCS方法,能够实现更好重建效果和噪声压制效果;相比于传统U-Net方法,具有更好的鲁棒性,能更有效地消除强振幅噪声并且减少了信息损失.Thereal seismic data usually have missing traces and noises,so it needs to be reconstructed and denoised.The Projection onto Convex Sets(POCS)method based on compressed sensing theory imposes Huber norm to minimize reconstruction error,which is equivalent to applying L1 norm minimization constraint to large reconstruction error(abnormal noise)and L_(2) norm minimization constraint to small reconstruction error(Gaussian random noise).It can reconstruct and de-noise the missing noisy seismic data at the same time.However,POCS is only a single-layer unsupervised learning method with weak mathematical representation ability,resulting in poor reconstruction effect and difficulty in suppressing noise completely.The U-Net method based on deep learning uses the L_(2) norm minimization constraint as the loss function to reconstruct and de-noise the noisy random missing seismic data.The method assumes that the reconstruction error satisfies the Gaussian distribution.Therefore,although the U-Net method can effectively suppress the Gaussian random noise,it is difficult to effectively remove the abnormal strong amplitude noise with super Gaussian distribution.This paper innovatively combines the above two methods,using Huber norm constraint instead of L_(2) norm constraint as the error function of the U-Net to establish Huber-U-Net network.The processing results of synthetic data and real data show that Huber-U-Net method can achieve better reconstruction effect and noise suppression effect than POCS method;Compared with the traditional U-Net method,it has better robustness,can more effectively eliminate strong amplitude noise and reduce information loss.

关 键 词：地震数据 重建与去噪 强振幅噪声 POCS Huber-U-Net 

分 类 号：P631[天文地球—地质矿产勘探]