基于U-Net网络的端到端地震高分辨率处理技术  

End-to-end high-resolution seismic processing method based on U-Net network

作　　者：孙永壮 黄鋆 俞伟哲 胡伟 SUN YongZhuang;HUANG Jun;YU WeiZhe;HU Wei(CNOOC China Limited Shanghai,Shanghai 200335,China)

机构地区：[1]中海石油(中国)有限公司上海分公司研究院,上海200335

出　　处：《地球物理学进展》2021年第3期1297-1305,共9页Progress in Geophysics

基　　金：国家科技重大专项课题“东海深层大型气田勘探评价技术”(2016ZX05027-02)资助。

摘　　要：拓展频带提高地震资料分辨率是薄储层预测和岩性尖灭点识别的关键手段.目前提升分辨率主要依靠反褶积谱蓝化、Q补偿等技术,这些方法存在假设条件苛刻、参数求取过程复杂、需要井提供额外信息等问题,为实际工作带来诸多不便.本文采用纯地震数据驱动的思想,通过构建大量三维地震伪反射系数模型,与不同主频的地震子波进行褶积获得不同分辨率的正演地震样本及标签数据,然后采用U-Net深度学习网络开展训练和测试并应用到东海某凹陷实际地震资料进行效果评估.这是一个从输入数据直接得到期望结果的"端到端"模型.结果表明,U-net网络高分辨率处理后地震带宽有效展宽了30%,主频从27.5 Hz提升到了37.5 Hz,对靶区主要目的层多期河道叠置关系的分辨效果提升明显.Expanding the frequency band to improve the resolution of seismic data is a key method for thin reservoir prediction and lithological pinch point identification. At present, the resolution improvement mainly relies on deconvolution, spectrum blueing, Q compensation and so on. Most of them have problems such as strict assumptions, complicated parameter solving processes, or wells dependent, which brings many inconveniences to actual work. This paper adopts the idea of seismic data driven. Firstly, a large number of 3 d seismic pseudo-reflection coefficient models are constructed and convolved with different seismic wavelets to obtain forward seismic records in different resolutions, then U-Net deep learning network was used for training and testing, and finally applied to the actual seismic data of a sag in East China Sea. This is an "end-to-end" model that gets the desired result directly from input data. The results show that the seismic bandwidth is effectively broadened by 30%, and the main frequency is increased from 27.5 Hz to 37.5 Hz. The high-resolution seismic data has a more obvious effect on multi-phase river channel identification, which shows a good application prospect.

关 键 词：U-Net网络 端到端 地震数据驱动 深度学习 高分辨率处理 

分 类 号：P631[天文地球—地质矿产勘探] P738[天文地球—地质学]