基于深度学习的复杂地质条件下地震波初至拾取研究  

Research on seismic first break picking under complex geological conditions based on depth learning

作　　者：袁联生[1] YUAN LianSheng(SINOPEC Geophysical Research Institute,Nanjing 211103,China)

机构地区：[1]中国石油化工股份有限公司石油物探技术研究院,南京211103

出　　处：《地球物理学进展》2022年第4期1657-1668,共12页Progress in Geophysics

基　　金：中国石化地球物理重点实验室项目“基于深度学习的遥感图像障碍物识别研究”(33550006-20-ZC0699-0015);国家自然科学基金项目(41930112)联合资助。

摘　　要：随着地质勘探工作不断深入到西部山地、戈壁等复杂地质环境中,数据量的指数增长以及采集所得的地震数据信噪比较低,导致早先的地震初至波自动拾取方法效率低下,精度不高,必须通过专家拾取干预才能满足实际工程需求.本文提出一种可以解决复杂地质条件下低信噪比的地震波初至自动拾取方法,本方法在地震波的初至拾取时,对地震数据进行了特殊的特征工程处理,然后采用多种语义分割网络模型对处理后的小批量数据进行训练,并把训练得到的网络模型用于低信噪比的地震波初至拾取工作.方法具体步骤为,首先通过地震数据预处理,即将进行线性校正等步骤处理后的地震数据裁剪为合适的大小以达到网络数据输入要求;接着,用两种不同的标注方式标注样本,并进行分析对比,得到初至到来之前和初至到来之后(包含起跳点)的二分类问题;然后,选择不同的语义分割网络模型进行测试,并根据模型最终的拾取率和IoU评估指标对比结果,得到实验效果最佳的网络模型;最后,对于一些异常的初至点,选取异常点上下十个样本点,通过比较它们之间的振幅大小,判断振幅值最大的样本点为最终的初至点.预测结果表明,本文提出的初至拾取方法对低信噪比信号有更好的效果.With the geological exploration work going deep into the western mountainous,gobi area and other complex geological environments,the exponential growth of the amount of data and the low signal-to-noise ratio of the seismic data acquisition lead to the low efficiency and low precision of the previous automatic pickup method of seismic first break.It must be picked up by experts to meet the practical engineering needs.It is presented an automatic first break pick-up method of seismic wave which can solve the problem of low signal-to-noise ratio under complex geological conditions in this paper.When picking up the first break of seismic wave,this method carries out special characteristic engineering processing on seismic data.Then,a variety of semantic segmentation network models are used to train the processed small batch data.The trained network model is used to pick up the first break of seismic wave with low signal-to-noise ratio.The specific steps of the method are as follows:firstly,in seismic data preprocessing,the seismic data processed by linearity correction and other steps are clipped to an appropriate size to meet the requirements of network data input.Secondly,two different annotation methods are used to label the samples.After analysis and comparison,the first break is divided into two categories:before the arrival of the first break and after the arrival of the first break(including the take-off point).Thirdly,different semantic segmentation network models are selected for testing.According to the final pick-up rate of the model and the comparison results of IoU evaluation indicators,the network model with the best experimental effect is obtained.Lastly,for some abnormal first break points,the upper and lower ten sample points of the abnormal point are selected.By comparing the amplitude between them,it is concluded that the sample point with the largest amplitude value is the final first break.The prediction results show that the first break pickup method proposed in this paper has better effect on l

关 键 词：地震波初至拾取 语义分割网络 低信噪比 

分 类 号：P631[天文地球—地质矿产勘探]