一种基于地震图像深度语义分割的初至拾取方法

First break picking method based on deep learning semantic segmentation of seismic image

本文从图像处理角度出发,模拟专业技术人员拾取初至的认知过程,提出了一种基于深度语义分割的初至拾取技术.该方法以更合适计算机辨识的地震数据密度图像为数据源,充分发挥语义分割在图像处理中的优势,采用端到端的编码-解码的语义分割网络模型实现初至语义特征学习,这是一种集合了深层抽象语义信息和浅层精细表征信息的跨层架构的语义分割网络,用来解决初至波语义和位置的内在张力问题,在收缩路径中捕捉初至上下文的语义信息和特征抽象,解决初至是什么的问题,在对称扩展路径中实现初至相对精准定位,结合浅层局部信息解决初至更高精度定位的问题.学习过程中通过样本筛选实现无初至时窗的数据聚焦和小粒度下的精细学习,通过模拟地形变化来对样本进行扩展,以实现在小样本下的优质学习.对深度网络模型进行优化,提高学习训练的精度与速度.通过实际数据测试表明,本方法可实现较高精度的初至检测,同时具有较强的抗噪能力,在低信噪比、不同噪声背景下也下仍能够较为精确地拾取初至.

In this paper,from the perspective of image processing,a first break picking method based on deep learning semantic segmentation is provided by simulating the cognitive process that geophysical experts pick first breaks.This method uses seismic data density image that can be easily recognized by computer as data source,so the advantages of image processing from semantic segmentation is fully used.An end-to-end encoding-decoding semantic segmentation network model is designed to achieve first break semantic feature learning;it is a cross-layered semantic segmentation network that integrates deep and coarse layer semantic information and shallow and fine layer feature information;it is used to resolve the intrinsic tension problem of first break semantics and position,it can capture the semantic information and abstracted features of first break contexts in contraction paths to solve the problem of what first break connection line is;first breaks in symmetrical extended paths can be comparatively precisely localized and then when combining with shallow local information,more precise first breaks can be picked.Data focus without first break time-window and fine learning are realized by sample filtering before learning.Samples are increased by simulating terrain changes to achieve high-quality learning under the condition of limited samples.The deep network model is optimized to improve accuracy and speed of learning.The actual data tests show that this method can precisely detect first breaks and has strong anti-noise ability,so it can pick first breaks for data with low signal-to-noise and different noise background.