基于深度学习的地震速度谱自动拾取方法

为了克服人工拾取地震速度谱效率低、耗时长等缺点,提出了一种基于深度学习的地震叠加速度自动拾取方法。其核心是模仿地震数据处理人员在速度谱上拾取速度的行为和过程,实现叠加速度的自动拾取。将速度谱视为图像,并依据所拾取的“时间速度”对具有时间序列的特点,设计了一个复杂的能用于速度拾取的卷积神经网络(convolutional neural network,CNN)和长短期记忆(long-short term memory,LSTM)模型混合结构神经网络模型。该模型经过训练,可以对输入的速度谱进行自动拾取,并输出“时间速度”对序列。理论和实际地震数据测试结果表明,相对于基于反演过程的传统速度拾取算法,基于深度学习的地震速度谱自动拾取方法无需附加任何约束和干预,不仅实现了完全自动化的速度拾取,而且具有更高的拾取精度。

To overcome the drawbacks of inefficient and time-consuming manual picking of seismic velocity,an auto-picking stack velocity method based on deep learning is proposed in this paper.This method aims to achieve automatic picking stack velocity by imitating the behavior and process by which seismic processing engineers pick velocity on velocity spectra.We regarded the latter as images and designed a hybrid neural network structure that combines a convolutional neural network (CNN) and longshort term memory (LSTM) according to “time-velocity” pairs with the characteristics of a time series.After training,the model can auto-pick the stack velocity from velocity spectra and output the “time-velocity” pairs.Experimental results with synthetic and real seismic data showed that,compared with the conventional algorithm based on inversion,this method implemented full autopicking velocity without any constraints and manual interventions and with better accuracy at picking velocity.It provides an effective solution to automatic picking seismic stack velocity and liberates technicians from the hard manual labor of picking velocity.

国家重点基础研究发展计划(973计划)项目(2017YFC0601504,2017YFC0601500)资助。