基于注意力机制U-Net的复杂地表地震初至拾取方法

Firstarrival traveltime pickup method for seismic data based on attention U-Net

地震数据的初至拾取是近地表初至波反演、静校正等工作的基础,是后续地震数据处理中不可或缺的一部分.随着地震数据量的不断增加,近地表条件更为复杂,信噪比越来越低,传统的自动拾取方法逐渐无法满足海量的拾取任务,需要大量的人工校正工作.因此需要一种更为精确的地震数据自动初至拾取方法.本文的贡献主要有以下几点.本文提出将一个注意力机制的U-Net网络应用于信噪比低、强近地表干扰的地区的初至拾取方法.首先将注意力门结构加入到一个标准的U-Net模型中,逐步抑制不相关的背景部分特征,提升拾取的准确性.其次本文提出使用地震波形数据与能量表象(Energy Semblance,ES)作为双通道数据体,可以使U-Net不仅关注相位信息,而且关注到初至的能量属性.此外,我们使用单通道的"0-1"标签作为网络的输出,可以大大缓解标签值分布不均衡的问题,提升预测精度.最后我们提出构建复杂近地表特征的合成数据集,包含复杂初至特征及复杂非均匀噪声的样本,验证方法的鲁棒性,并证明Dice-loss函数可以提升拾取精度.本文提出的方法对于强噪声干扰、缺道坏道及复杂近地表情况均有良好的适应性.

The first arrival pickup is the basis for near-surface first arrivals inversion and statics correction, an indispensable part of subsequent seismic data processing. With the ever increasing seismic data volumes, more complex near-surface conditions, and low Signal-to-Noise Ratio(SNR), traditional automatic methods can not meet the pickup requirements. So a more accurate automatic first arrival pickup method is crucial. The main contributions of this paper are as follows: This study proposed an attention-U-Net and applied it to the first arrival pickup of seismic data with a low signal-to-noise ratio and strong near-surface influence. Firstly, the attention gates were integrated into a standard U-Net model, which progressively suppressed features in irrelevant background parts and improved the picking accuracy. Secondly, this paper proposed to use seismic waveform data and energy semblance data as two channels data body, which makes the U-Net focus not only on the phase but also on the energy attribute of the first arrival traveltime. Thirdly, we used a single channel"0-1"label as the U-Net's output, which can alleviate the unbalanced distribution of label value and improve the prediction accuracy. Then, we proposed a scheme to construct a synthetic dataset containing characteristics of seismic field data to test the robustness of the U-Net and proved that the Dice-loss function could improve the pickup accuracy. At last, for mountain field, seismic data, the U-Net proposed in this paper has better adaptability to strong noise interference, missing and bad traces, and complex near-surface conditions, significantly reducing labor expense.