基于U形卷积神经网络的震相识别与到时拾取方法研究

精确获取震相到时是地震定位和地震走时成像等研究的重要基础.近年来，随着地震台站的不断加密，地震台网监测到的地震数量成倍增长，发展快速、准确、适用性强的震相到时自动拾取算法是地震行业的迫切需求.本文在前人工作基础上，发展了Pg、Sg震相自动识别与到时拾取的U网络算法（Unet_cea），使用汶川余震和首都圈地震台网记录的89344个不同震级、不同信噪比的样本进行训练和测试.研究表明，U网络能够较好地识别Pg、Sg震相类型和拾取到时，Pg、Sg震相的正确识别率分别为81%和79.1%，与人工标注到时的均方根误差分别为0.41 s和0.54 s.U网络在命中率、均方根误差等性能指标上均明显优于STA/LTA和峰度分析自动拾取方法.研究获得的最优模型可以为区域地震台网的自动处理提供辅助.

Earthquake phase arrival auto-picking based on U-shaped convolutional neural network

Accurate seismic phase arrival time picking is the basis for earthquake location and seismic travel time tomography. With the increase of seismic stations and the improvement of monitoring capabilities, it is an urgent need to develop fast, accurate and adaptable algorithms. Based on previous work, this paper developed a U-shaped neural network (Unet_cea) for Pg and Sg phase detection and arrival time picking, and trained model on 89344 waveform samples with different magnitudes and signal-to-noise ratio levels from Wenchuan aftershocks and the Beijing capital circle seismic network. The research shows that the U-net can recognize Pg and Sg phases with high correct rate, 81% and 79.1% respectively. Compared with manual picks, the mean square root errors of Pg and Sg are 0.41 s and 0.54 s, respectively. The U-net is superior to the STA/LTA and kurtosis analysis methods in performance and has higher hit rate and lower root mean square error. The optimal model obtained from the study can assist automatic cataloging work of the regional seismic networks.