基于残差U-net网络的地震资料分辨率提高方法  

Resolution Improvement Method for Seismic Data Based on Residual U-Net Network

作　　者：董博艺 张进[1,2,3] Dong Boyi;Zhang Jin(College of Marine Geoscience,Ocean University of China,Qingdao 266100,China;Laboratory for Evaluation and Prospecting Technology of Marine Mineral Resources,Qingdao Marine Science and Technology Center,Qingdao 266237,China;Key Laboratory of Submarine Geosciences and Prospecting Techniques Ministry of Education,Ocean University of China,Qingdao 266100,China)

机构地区：[1]中国海洋大学海洋地球科学学院,山东青岛266100 [2]青岛海洋科技中心海洋矿产资源评价与探测技术功能实验室,山东青岛266237 [3]中国海洋大学海底科学与探测技术教育部重点实验室,山东青岛266100

出　　处：《中国海洋大学学报（自然科学版）》2025年第1期140-148,共9页Periodical of Ocean University of China

基　　金：国家自然科学基金重点项目(U23B20158)资助。

摘　　要：高分辨率地震资料处理是获取高品质地震资料、实现薄储层良好地震地质解释的关键。传统提高地震分辨率的方法应用条件苛刻,关键参数求取复杂,在实际应用中受到诸多限制。深度学习中的U-net网络以纯数据驱动的优势,可学习低分辨率地震记录到高分辨率标签的非线性关系,实现地震资料的高分辨率处理。本文设计了残差U-net网络结构,同时提出了基于概率密度函数控制的同分布反射系数集生成方法,将测井反射系数的概率密度函数作为一种先验约束信息融入训练样本,不仅保证了足够的同分布样本来训练网络,还确保了训练样本更符合工区实际情况,以此提高模型预测的准确性。模型测试和实际资料应用结果表明,本文提出的方法能够有效应用于地震资料分辨率的提高,同时拓宽频带。High resolution seismic data processing is the key to obtaining high-quality seismic data and achieving good seismic geological interpretation of thin reservoirs.The traditional methods for improving seismic resolution have strict application conditions,complex calculation of key parameters,and are subject to many limitations in practical applications.The U-net network in deep learning has the advantage of pure data-driven learning,which can learn the nonlinear relationship from low resolution seismic records to high-resolution labels,achieving high-resolution processing of seismic data.This article designs a residual U-net network structure and proposes a method for generating a set of identically distributed reflection coefficients based on probability density function control.The probability density function of logging reflection coefficients is incorporated into the training samples as a prior constraint information,which not only ensures sufficient identically distributed samples for training the network,but also ensures that the training samples are more in line with the actual situation of the work area,thereby improving the accuracy of model prediction.The results of model testing and practical data application show that the method proposed in this paper can effectively improve the resolution of seismic data and broaden the frequency band.

关 键 词：提高分辨率 U-net 残差结构 同分布 

分 类 号：P714.8[天文地球—海洋科学]